projects/linux·blob v2.6.16 · refs

872 lines · 19.7 KiB

1/*
2 * linux/kernel/hrtimer.c
3 *
4 * Copyright(C) 2005, Thomas Gleixner <[email protected]>
5 * Copyright(C) 2005, Red Hat, Inc., Ingo Molnar
6 *
7 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API implemented in
10 * kernel/timer.c, hrtimers provide finer resolution and accuracy
11 * depending on system configuration and capabilities.
12 *
13 * These timers are currently used for:
14 * - itimers
15 * - POSIX timers
16 * - nanosleep
17 * - precise in-kernel timing
18 *
19 * Started by: Thomas Gleixner and Ingo Molnar
20 *
21 * Credits:
22 * based on kernel/timer.c
23 *
24 * Help, testing, suggestions, bugfixes, improvements were
25 * provided by:
26 *
27 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
28 * et. al.
29 *
30 * For licencing details see kernel-base/COPYING
31 */
32
33#include <linux/cpu.h>
34#include <linux/module.h>
35#include <linux/percpu.h>
36#include <linux/hrtimer.h>
37#include <linux/notifier.h>
38#include <linux/syscalls.h>
39#include <linux/interrupt.h>
40
41#include <asm/uaccess.h>
42
43/**
44 * ktime_get - get the monotonic time in ktime_t format
45 *
46 * returns the time in ktime_t format
47 */
48static ktime_t ktime_get(void)
49{
50 struct timespec now;
51
52 ktime_get_ts(&now);
53
54 return timespec_to_ktime(now);
55}
56
57/**
58 * ktime_get_real - get the real (wall-) time in ktime_t format
59 *
60 * returns the time in ktime_t format
61 */
62static ktime_t ktime_get_real(void)
63{
64 struct timespec now;
65
66 getnstimeofday(&now);
67
68 return timespec_to_ktime(now);
69}
70
71EXPORT_SYMBOL_GPL(ktime_get_real);
72
73/*
74 * The timer bases:
75 *
76 * Note: If we want to add new timer bases, we have to skip the two
77 * clock ids captured by the cpu-timers. We do this by holding empty
78 * entries rather than doing math adjustment of the clock ids.
79 * This ensures that we capture erroneous accesses to these clock ids
80 * rather than moving them into the range of valid clock id's.
81 */
82
83#define MAX_HRTIMER_BASES 2
84
85static DEFINE_PER_CPU(struct hrtimer_base, hrtimer_bases[MAX_HRTIMER_BASES]) =
86{
87 {
88 .index = CLOCK_REALTIME,
89 .get_time = &ktime_get_real,
90 .resolution = KTIME_REALTIME_RES,
91 },
92 {
93 .index = CLOCK_MONOTONIC,
94 .get_time = &ktime_get,
95 .resolution = KTIME_MONOTONIC_RES,
96 },
97};
98
99/**
100 * ktime_get_ts - get the monotonic clock in timespec format
101 *
102 * @ts: pointer to timespec variable
103 *
104 * The function calculates the monotonic clock from the realtime
105 * clock and the wall_to_monotonic offset and stores the result
106 * in normalized timespec format in the variable pointed to by ts.
107 */
108void ktime_get_ts(struct timespec *ts)
109{
110 struct timespec tomono;
111 unsigned long seq;
112
113 do {
114 seq = read_seqbegin(&xtime_lock);
115 getnstimeofday(ts);
116 tomono = wall_to_monotonic;
117
118 } while (read_seqretry(&xtime_lock, seq));
119
120 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
121 ts->tv_nsec + tomono.tv_nsec);
122}
123EXPORT_SYMBOL_GPL(ktime_get_ts);
124
125/*
126 * Functions and macros which are different for UP/SMP systems are kept in a
127 * single place
128 */
129#ifdef CONFIG_SMP
130
131#define set_curr_timer(b, t) do { (b)->curr_timer = (t); } while (0)
132
133/*
134 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
135 * means that all timers which are tied to this base via timer->base are
136 * locked, and the base itself is locked too.
137 *
138 * So __run_timers/migrate_timers can safely modify all timers which could
139 * be found on the lists/queues.
140 *
141 * When the timer's base is locked, and the timer removed from list, it is
142 * possible to set timer->base = NULL and drop the lock: the timer remains
143 * locked.
144 */
145static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
146 unsigned long *flags)
147{
148 struct hrtimer_base *base;
149
150 for (;;) {
151 base = timer->base;
152 if (likely(base != NULL)) {
153 spin_lock_irqsave(&base->lock, *flags);
154 if (likely(base == timer->base))
155 return base;
156 /* The timer has migrated to another CPU: */
157 spin_unlock_irqrestore(&base->lock, *flags);
158 }
159 cpu_relax();
160 }
161}
162
163/*
164 * Switch the timer base to the current CPU when possible.
165 */
166static inline struct hrtimer_base *
167switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
168{
169 struct hrtimer_base *new_base;
170
171 new_base = &__get_cpu_var(hrtimer_bases[base->index]);
172
173 if (base != new_base) {
174 /*
175 * We are trying to schedule the timer on the local CPU.
176 * However we can't change timer's base while it is running,
177 * so we keep it on the same CPU. No hassle vs. reprogramming
178 * the event source in the high resolution case. The softirq
179 * code will take care of this when the timer function has
180 * completed. There is no conflict as we hold the lock until
181 * the timer is enqueued.
182 */
183 if (unlikely(base->curr_timer == timer))
184 return base;
185
186 /* See the comment in lock_timer_base() */
187 timer->base = NULL;
188 spin_unlock(&base->lock);
189 spin_lock(&new_base->lock);
190 timer->base = new_base;
191 }
192 return new_base;
193}
194
195#else /* CONFIG_SMP */
196
197#define set_curr_timer(b, t) do { } while (0)
198
199static inline struct hrtimer_base *
200lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
201{
202 struct hrtimer_base *base = timer->base;
203
204 spin_lock_irqsave(&base->lock, *flags);
205
206 return base;
207}
208
209#define switch_hrtimer_base(t, b) (b)
210
211#endif /* !CONFIG_SMP */
212
213/*
214 * Functions for the union type storage format of ktime_t which are
215 * too large for inlining:
216 */
217#if BITS_PER_LONG < 64
218# ifndef CONFIG_KTIME_SCALAR
219/**
220 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
221 *
222 * @kt: addend
223 * @nsec: the scalar nsec value to add
224 *
225 * Returns the sum of kt and nsec in ktime_t format
226 */
227ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
228{
229 ktime_t tmp;
230
231 if (likely(nsec < NSEC_PER_SEC)) {
232 tmp.tv64 = nsec;
233 } else {
234 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
235
236 tmp = ktime_set((long)nsec, rem);
237 }
238
239 return ktime_add(kt, tmp);
240}
241
242#else /* CONFIG_KTIME_SCALAR */
243
244# endif /* !CONFIG_KTIME_SCALAR */
245
246/*
247 * Divide a ktime value by a nanosecond value
248 */
249static unsigned long ktime_divns(const ktime_t kt, nsec_t div)
250{
251 u64 dclc, inc, dns;
252 int sft = 0;
253
254 dclc = dns = ktime_to_ns(kt);
255 inc = div;
256 /* Make sure the divisor is less than 2^32: */
257 while (div >> 32) {
258 sft++;
259 div >>= 1;
260 }
261 dclc >>= sft;
262 do_div(dclc, (unsigned long) div);
263
264 return (unsigned long) dclc;
265}
266
267#else /* BITS_PER_LONG < 64 */
268# define ktime_divns(kt, div) (unsigned long)((kt).tv64 / (div))
269#endif /* BITS_PER_LONG >= 64 */
270
271/*
272 * Counterpart to lock_timer_base above:
273 */
274static inline
275void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
276{
277 spin_unlock_irqrestore(&timer->base->lock, *flags);
278}
279
280/**
281 * hrtimer_forward - forward the timer expiry
282 *
283 * @timer: hrtimer to forward
284 * @interval: the interval to forward
285 *
286 * Forward the timer expiry so it will expire in the future.
287 * Returns the number of overruns.
288 */
289unsigned long
290hrtimer_forward(struct hrtimer *timer, ktime_t interval)
291{
292 unsigned long orun = 1;
293 ktime_t delta, now;
294
295 now = timer->base->get_time();
296
297 delta = ktime_sub(now, timer->expires);
298
299 if (delta.tv64 < 0)
300 return 0;
301
302 if (interval.tv64 < timer->base->resolution.tv64)
303 interval.tv64 = timer->base->resolution.tv64;
304
305 if (unlikely(delta.tv64 >= interval.tv64)) {
306 nsec_t incr = ktime_to_ns(interval);
307
308 orun = ktime_divns(delta, incr);
309 timer->expires = ktime_add_ns(timer->expires, incr * orun);
310 if (timer->expires.tv64 > now.tv64)
311 return orun;
312 /*
313 * This (and the ktime_add() below) is the
314 * correction for exact:
315 */
316 orun++;
317 }
318 timer->expires = ktime_add(timer->expires, interval);
319
320 return orun;
321}
322
323/*
324 * enqueue_hrtimer - internal function to (re)start a timer
325 *
326 * The timer is inserted in expiry order. Insertion into the
327 * red black tree is O(log(n)). Must hold the base lock.
328 */
329static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
330{
331 struct rb_node **link = &base->active.rb_node;
332 struct rb_node *parent = NULL;
333 struct hrtimer *entry;
334
335 /*
336 * Find the right place in the rbtree:
337 */
338 while (*link) {
339 parent = *link;
340 entry = rb_entry(parent, struct hrtimer, node);
341 /*
342 * We dont care about collisions. Nodes with
343 * the same expiry time stay together.
344 */
345 if (timer->expires.tv64 < entry->expires.tv64)
346 link = &(*link)->rb_left;
347 else
348 link = &(*link)->rb_right;
349 }
350
351 /*
352 * Insert the timer to the rbtree and check whether it
353 * replaces the first pending timer
354 */
355 rb_link_node(&timer->node, parent, link);
356 rb_insert_color(&timer->node, &base->active);
357
358 timer->state = HRTIMER_PENDING;
359
360 if (!base->first || timer->expires.tv64 <
361 rb_entry(base->first, struct hrtimer, node)->expires.tv64)
362 base->first = &timer->node;
363}
364
365/*
366 * __remove_hrtimer - internal function to remove a timer
367 *
368 * Caller must hold the base lock.
369 */
370static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
371{
372 /*
373 * Remove the timer from the rbtree and replace the
374 * first entry pointer if necessary.
375 */
376 if (base->first == &timer->node)
377 base->first = rb_next(&timer->node);
378 rb_erase(&timer->node, &base->active);
379}
380
381/*
382 * remove hrtimer, called with base lock held
383 */
384static inline int
385remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
386{
387 if (hrtimer_active(timer)) {
388 __remove_hrtimer(timer, base);
389 timer->state = HRTIMER_INACTIVE;
390 return 1;
391 }
392 return 0;
393}
394
395/**
396 * hrtimer_start - (re)start an relative timer on the current CPU
397 *
398 * @timer: the timer to be added
399 * @tim: expiry time
400 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
401 *
402 * Returns:
403 * 0 on success
404 * 1 when the timer was active
405 */
406int
407hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
408{
409 struct hrtimer_base *base, *new_base;
410 unsigned long flags;
411 int ret;
412
413 base = lock_hrtimer_base(timer, &flags);
414
415 /* Remove an active timer from the queue: */
416 ret = remove_hrtimer(timer, base);
417
418 /* Switch the timer base, if necessary: */
419 new_base = switch_hrtimer_base(timer, base);
420
421 if (mode == HRTIMER_REL) {
422 tim = ktime_add(tim, new_base->get_time());
423 /*
424 * CONFIG_TIME_LOW_RES is a temporary way for architectures
425 * to signal that they simply return xtime in
426 * do_gettimeoffset(). In this case we want to round up by
427 * resolution when starting a relative timer, to avoid short
428 * timeouts. This will go away with the GTOD framework.
429 */
430#ifdef CONFIG_TIME_LOW_RES
431 tim = ktime_add(tim, base->resolution);
432#endif
433 }
434 timer->expires = tim;
435
436 enqueue_hrtimer(timer, new_base);
437
438 unlock_hrtimer_base(timer, &flags);
439
440 return ret;
441}
442
443/**
444 * hrtimer_try_to_cancel - try to deactivate a timer
445 *
446 * @timer: hrtimer to stop
447 *
448 * Returns:
449 * 0 when the timer was not active
450 * 1 when the timer was active
451 * -1 when the timer is currently excuting the callback function and
452 * can not be stopped
453 */
454int hrtimer_try_to_cancel(struct hrtimer *timer)
455{
456 struct hrtimer_base *base;
457 unsigned long flags;
458 int ret = -1;
459
460 base = lock_hrtimer_base(timer, &flags);
461
462 if (base->curr_timer != timer)
463 ret = remove_hrtimer(timer, base);
464
465 unlock_hrtimer_base(timer, &flags);
466
467 return ret;
468
469}
470
471/**
472 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
473 *
474 * @timer: the timer to be cancelled
475 *
476 * Returns:
477 * 0 when the timer was not active
478 * 1 when the timer was active
479 */
480int hrtimer_cancel(struct hrtimer *timer)
481{
482 for (;;) {
483 int ret = hrtimer_try_to_cancel(timer);
484
485 if (ret >= 0)
486 return ret;
487 }
488}
489
490/**
491 * hrtimer_get_remaining - get remaining time for the timer
492 *
493 * @timer: the timer to read
494 */
495ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
496{
497 struct hrtimer_base *base;
498 unsigned long flags;
499 ktime_t rem;
500
501 base = lock_hrtimer_base(timer, &flags);
502 rem = ktime_sub(timer->expires, timer->base->get_time());
503 unlock_hrtimer_base(timer, &flags);
504
505 return rem;
506}
507
508#ifdef CONFIG_NO_IDLE_HZ
509/**
510 * hrtimer_get_next_event - get the time until next expiry event
511 *
512 * Returns the delta to the next expiry event or KTIME_MAX if no timer
513 * is pending.
514 */
515ktime_t hrtimer_get_next_event(void)
516{
517 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
518 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
519 unsigned long flags;
520 int i;
521
522 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
523 struct hrtimer *timer;
524
525 spin_lock_irqsave(&base->lock, flags);
526 if (!base->first) {
527 spin_unlock_irqrestore(&base->lock, flags);
528 continue;
529 }
530 timer = rb_entry(base->first, struct hrtimer, node);
531 delta.tv64 = timer->expires.tv64;
532 spin_unlock_irqrestore(&base->lock, flags);
533 delta = ktime_sub(delta, base->get_time());
534 if (delta.tv64 < mindelta.tv64)
535 mindelta.tv64 = delta.tv64;
536 }
537 if (mindelta.tv64 < 0)
538 mindelta.tv64 = 0;
539 return mindelta;
540}
541#endif
542
543/**
544 * hrtimer_init - initialize a timer to the given clock
545 *
546 * @timer: the timer to be initialized
547 * @clock_id: the clock to be used
548 * @mode: timer mode abs/rel
549 */
550void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
551 enum hrtimer_mode mode)
552{
553 struct hrtimer_base *bases;
554
555 memset(timer, 0, sizeof(struct hrtimer));
556
557 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
558
559 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
560 clock_id = CLOCK_MONOTONIC;
561
562 timer->base = &bases[clock_id];
563}
564
565/**
566 * hrtimer_get_res - get the timer resolution for a clock
567 *
568 * @which_clock: which clock to query
569 * @tp: pointer to timespec variable to store the resolution
570 *
571 * Store the resolution of the clock selected by which_clock in the
572 * variable pointed to by tp.
573 */
574int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
575{
576 struct hrtimer_base *bases;
577
578 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
579 *tp = ktime_to_timespec(bases[which_clock].resolution);
580
581 return 0;
582}
583
584/*
585 * Expire the per base hrtimer-queue:
586 */
587static inline void run_hrtimer_queue(struct hrtimer_base *base)
588{
589 ktime_t now = base->get_time();
590 struct rb_node *node;
591
592 spin_lock_irq(&base->lock);
593
594 while ((node = base->first)) {
595 struct hrtimer *timer;
596 int (*fn)(void *);
597 int restart;
598 void *data;
599
600 timer = rb_entry(node, struct hrtimer, node);
601 if (now.tv64 <= timer->expires.tv64)
602 break;
603
604 fn = timer->function;
605 data = timer->data;
606 set_curr_timer(base, timer);
607 timer->state = HRTIMER_RUNNING;
608 __remove_hrtimer(timer, base);
609 spin_unlock_irq(&base->lock);
610
611 /*
612 * fn == NULL is special case for the simplest timer
613 * variant - wake up process and do not restart:
614 */
615 if (!fn) {
616 wake_up_process(data);
617 restart = HRTIMER_NORESTART;
618 } else
619 restart = fn(data);
620
621 spin_lock_irq(&base->lock);
622
623 /* Another CPU has added back the timer */
624 if (timer->state != HRTIMER_RUNNING)
625 continue;
626
627 if (restart == HRTIMER_RESTART)
628 enqueue_hrtimer(timer, base);
629 else
630 timer->state = HRTIMER_EXPIRED;
631 }
632 set_curr_timer(base, NULL);
633 spin_unlock_irq(&base->lock);
634}
635
636/*
637 * Called from timer softirq every jiffy, expire hrtimers:
638 */
639void hrtimer_run_queues(void)
640{
641 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
642 int i;
643
644 for (i = 0; i < MAX_HRTIMER_BASES; i++)
645 run_hrtimer_queue(&base[i]);
646}
647
648/*
649 * Sleep related functions:
650 */
651
652/**
653 * schedule_hrtimer - sleep until timeout
654 *
655 * @timer: hrtimer variable initialized with the correct clock base
656 * @mode: timeout value is abs/rel
657 *
658 * Make the current task sleep until @timeout is
659 * elapsed.
660 *
661 * You can set the task state as follows -
662 *
663 * %TASK_UNINTERRUPTIBLE - at least @timeout is guaranteed to
664 * pass before the routine returns. The routine will return 0
665 *
666 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
667 * delivered to the current task. In this case the remaining time
668 * will be returned
669 *
670 * The current task state is guaranteed to be TASK_RUNNING when this
671 * routine returns.
672 */
673static ktime_t __sched
674schedule_hrtimer(struct hrtimer *timer, const enum hrtimer_mode mode)
675{
676 /* fn stays NULL, meaning single-shot wakeup: */
677 timer->data = current;
678
679 hrtimer_start(timer, timer->expires, mode);
680
681 schedule();
682 hrtimer_cancel(timer);
683
684 /* Return the remaining time: */
685 if (timer->state != HRTIMER_EXPIRED)
686 return ktime_sub(timer->expires, timer->base->get_time());
687 else
688 return (ktime_t) {.tv64 = 0 };
689}
690
691static inline ktime_t __sched
692schedule_hrtimer_interruptible(struct hrtimer *timer,
693 const enum hrtimer_mode mode)
694{
695 set_current_state(TASK_INTERRUPTIBLE);
696
697 return schedule_hrtimer(timer, mode);
698}
699
700static long __sched nanosleep_restart(struct restart_block *restart)
701{
702 struct timespec __user *rmtp;
703 struct timespec tu;
704 void *rfn_save = restart->fn;
705 struct hrtimer timer;
706 ktime_t rem;
707
708 restart->fn = do_no_restart_syscall;
709
710 hrtimer_init(&timer, (clockid_t) restart->arg3, HRTIMER_ABS);
711
712 timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
713
714 rem = schedule_hrtimer_interruptible(&timer, HRTIMER_ABS);
715
716 if (rem.tv64 <= 0)
717 return 0;
718
719 rmtp = (struct timespec __user *) restart->arg2;
720 tu = ktime_to_timespec(rem);
721 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
722 return -EFAULT;
723
724 restart->fn = rfn_save;
725
726 /* The other values in restart are already filled in */
727 return -ERESTART_RESTARTBLOCK;
728}
729
730long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
731 const enum hrtimer_mode mode, const clockid_t clockid)
732{
733 struct restart_block *restart;
734 struct hrtimer timer;
735 struct timespec tu;
736 ktime_t rem;
737
738 hrtimer_init(&timer, clockid, mode);
739
740 timer.expires = timespec_to_ktime(*rqtp);
741
742 rem = schedule_hrtimer_interruptible(&timer, mode);
743 if (rem.tv64 <= 0)
744 return 0;
745
746 /* Absolute timers do not update the rmtp value and restart: */
747 if (mode == HRTIMER_ABS)
748 return -ERESTARTNOHAND;
749
750 tu = ktime_to_timespec(rem);
751
752 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
753 return -EFAULT;
754
755 restart = &current_thread_info()->restart_block;
756 restart->fn = nanosleep_restart;
757 restart->arg0 = timer.expires.tv64 & 0xFFFFFFFF;
758 restart->arg1 = timer.expires.tv64 >> 32;
759 restart->arg2 = (unsigned long) rmtp;
760 restart->arg3 = (unsigned long) timer.base->index;
761
762 return -ERESTART_RESTARTBLOCK;
763}
764
765asmlinkage long
766sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
767{
768 struct timespec tu;
769
770 if (copy_from_user(&tu, rqtp, sizeof(tu)))
771 return -EFAULT;
772
773 if (!timespec_valid(&tu))
774 return -EINVAL;
775
776 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
777}
778
779/*
780 * Functions related to boot-time initialization:
781 */
782static void __devinit init_hrtimers_cpu(int cpu)
783{
784 struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
785 int i;
786
787 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
788 spin_lock_init(&base->lock);
789}
790
791#ifdef CONFIG_HOTPLUG_CPU
792
793static void migrate_hrtimer_list(struct hrtimer_base *old_base,
794 struct hrtimer_base *new_base)
795{
796 struct hrtimer *timer;
797 struct rb_node *node;
798
799 while ((node = rb_first(&old_base->active))) {
800 timer = rb_entry(node, struct hrtimer, node);
801 __remove_hrtimer(timer, old_base);
802 timer->base = new_base;
803 enqueue_hrtimer(timer, new_base);
804 }
805}
806
807static void migrate_hrtimers(int cpu)
808{
809 struct hrtimer_base *old_base, *new_base;
810 int i;
811
812 BUG_ON(cpu_online(cpu));
813 old_base = per_cpu(hrtimer_bases, cpu);
814 new_base = get_cpu_var(hrtimer_bases);
815
816 local_irq_disable();
817
818 for (i = 0; i < MAX_HRTIMER_BASES; i++) {
819
820 spin_lock(&new_base->lock);
821 spin_lock(&old_base->lock);
822
823 BUG_ON(old_base->curr_timer);
824
825 migrate_hrtimer_list(old_base, new_base);
826
827 spin_unlock(&old_base->lock);
828 spin_unlock(&new_base->lock);
829 old_base++;
830 new_base++;
831 }
832
833 local_irq_enable();
834 put_cpu_var(hrtimer_bases);
835}
836#endif /* CONFIG_HOTPLUG_CPU */
837
838static int __devinit hrtimer_cpu_notify(struct notifier_block *self,
839 unsigned long action, void *hcpu)
840{
841 long cpu = (long)hcpu;
842
843 switch (action) {
844
845 case CPU_UP_PREPARE:
846 init_hrtimers_cpu(cpu);
847 break;
848
849#ifdef CONFIG_HOTPLUG_CPU
850 case CPU_DEAD:
851 migrate_hrtimers(cpu);
852 break;
853#endif
854
855 default:
856 break;
857 }
858
859 return NOTIFY_OK;
860}
861
862static struct notifier_block __devinitdata hrtimers_nb = {
863 .notifier_call = hrtimer_cpu_notify,
864};
865
866void __init hrtimers_init(void)
867{
868 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
869 (void *)(long)smp_processor_id());
870 register_cpu_notifier(&hrtimers_nb);
871}
872