2 * linux/kernel/hrtimer.c
4 * Copyright(C) 2005, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005, Red Hat, Inc., Ingo Molnar
7 * High-resolution kernel timers
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.
13 * These timers are currently used for:
17 * - precise in-kernel timing
19 * Started by: Thomas Gleixner and Ingo Molnar
22 * based on kernel/timer.c
24 * Help, testing, suggestions, bugfixes, improvements were
27 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
30 * For licencing details see kernel-base/COPYING
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>
41 #include <asm/uaccess.h>
44 * ktime_get - get the monotonic time in ktime_t format
46 * returns the time in ktime_t format
48 static ktime_t ktime_get(void)
54 return timespec_to_ktime(now);
58 * ktime_get_real - get the real (wall-) time in ktime_t format
60 * returns the time in ktime_t format
62 static ktime_t ktime_get_real(void)
68 return timespec_to_ktime(now);
71 EXPORT_SYMBOL_GPL(ktime_get_real);
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.
83 #define MAX_HRTIMER_BASES 2
85 static DEFINE_PER_CPU(struct hrtimer_base, hrtimer_bases[MAX_HRTIMER_BASES]) =
88 .index = CLOCK_REALTIME,
89 .get_time = &ktime_get_real,
90 .resolution = KTIME_REALTIME_RES,
93 .index = CLOCK_MONOTONIC,
94 .get_time = &ktime_get,
95 .resolution = KTIME_MONOTONIC_RES,
100 * ktime_get_ts - get the monotonic clock in timespec format
101 * @ts: pointer to timespec variable
103 * The function calculates the monotonic clock from the realtime
104 * clock and the wall_to_monotonic offset and stores the result
105 * in normalized timespec format in the variable pointed to by ts.
107 void ktime_get_ts(struct timespec *ts)
109 struct timespec tomono;
113 seq = read_seqbegin(&xtime_lock);
115 tomono = wall_to_monotonic;
117 } while (read_seqretry(&xtime_lock, seq));
119 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
120 ts->tv_nsec + tomono.tv_nsec);
122 EXPORT_SYMBOL_GPL(ktime_get_ts);
125 * Get the coarse grained time at the softirq based on xtime and
128 static void hrtimer_get_softirq_time(struct hrtimer_base *base)
130 ktime_t xtim, tomono;
134 seq = read_seqbegin(&xtime_lock);
135 xtim = timespec_to_ktime(xtime);
136 tomono = timespec_to_ktime(wall_to_monotonic);
138 } while (read_seqretry(&xtime_lock, seq));
140 base[CLOCK_REALTIME].softirq_time = xtim;
141 base[CLOCK_MONOTONIC].softirq_time = ktime_add(xtim, tomono);
145 * Functions and macros which are different for UP/SMP systems are kept in a
150 #define set_curr_timer(b, t) do { (b)->curr_timer = (t); } while (0)
153 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
154 * means that all timers which are tied to this base via timer->base are
155 * locked, and the base itself is locked too.
157 * So __run_timers/migrate_timers can safely modify all timers which could
158 * be found on the lists/queues.
160 * When the timer's base is locked, and the timer removed from list, it is
161 * possible to set timer->base = NULL and drop the lock: the timer remains
164 static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
165 unsigned long *flags)
167 struct hrtimer_base *base;
171 if (likely(base != NULL)) {
172 spin_lock_irqsave(&base->lock, *flags);
173 if (likely(base == timer->base))
175 /* The timer has migrated to another CPU: */
176 spin_unlock_irqrestore(&base->lock, *flags);
183 * Switch the timer base to the current CPU when possible.
185 static inline struct hrtimer_base *
186 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
188 struct hrtimer_base *new_base;
190 new_base = &__get_cpu_var(hrtimer_bases[base->index]);
192 if (base != new_base) {
194 * We are trying to schedule the timer on the local CPU.
195 * However we can't change timer's base while it is running,
196 * so we keep it on the same CPU. No hassle vs. reprogramming
197 * the event source in the high resolution case. The softirq
198 * code will take care of this when the timer function has
199 * completed. There is no conflict as we hold the lock until
200 * the timer is enqueued.
202 if (unlikely(base->curr_timer == timer))
205 /* See the comment in lock_timer_base() */
207 spin_unlock(&base->lock);
208 spin_lock(&new_base->lock);
209 timer->base = new_base;
214 #else /* CONFIG_SMP */
216 #define set_curr_timer(b, t) do { } while (0)
218 static inline struct hrtimer_base *
219 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
221 struct hrtimer_base *base = timer->base;
223 spin_lock_irqsave(&base->lock, *flags);
228 #define switch_hrtimer_base(t, b) (b)
230 #endif /* !CONFIG_SMP */
233 * Functions for the union type storage format of ktime_t which are
234 * too large for inlining:
236 #if BITS_PER_LONG < 64
237 # ifndef CONFIG_KTIME_SCALAR
239 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
241 * @nsec: the scalar nsec value to add
243 * Returns the sum of kt and nsec in ktime_t format
245 ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
249 if (likely(nsec < NSEC_PER_SEC)) {
252 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
254 tmp = ktime_set((long)nsec, rem);
257 return ktime_add(kt, tmp);
260 #else /* CONFIG_KTIME_SCALAR */
262 # endif /* !CONFIG_KTIME_SCALAR */
265 * Divide a ktime value by a nanosecond value
267 static unsigned long ktime_divns(const ktime_t kt, s64 div)
272 dclc = dns = ktime_to_ns(kt);
274 /* Make sure the divisor is less than 2^32: */
280 do_div(dclc, (unsigned long) div);
282 return (unsigned long) dclc;
285 #else /* BITS_PER_LONG < 64 */
286 # define ktime_divns(kt, div) (unsigned long)((kt).tv64 / (div))
287 #endif /* BITS_PER_LONG >= 64 */
290 * Counterpart to lock_timer_base above:
293 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
295 spin_unlock_irqrestore(&timer->base->lock, *flags);
299 * hrtimer_forward - forward the timer expiry
300 * @timer: hrtimer to forward
301 * @now: forward past this time
302 * @interval: the interval to forward
304 * Forward the timer expiry so it will expire in the future.
305 * Returns the number of overruns.
308 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
310 unsigned long orun = 1;
313 delta = ktime_sub(now, timer->expires);
318 if (interval.tv64 < timer->base->resolution.tv64)
319 interval.tv64 = timer->base->resolution.tv64;
321 if (unlikely(delta.tv64 >= interval.tv64)) {
322 s64 incr = ktime_to_ns(interval);
324 orun = ktime_divns(delta, incr);
325 timer->expires = ktime_add_ns(timer->expires, incr * orun);
326 if (timer->expires.tv64 > now.tv64)
329 * This (and the ktime_add() below) is the
330 * correction for exact:
334 timer->expires = ktime_add(timer->expires, interval);
340 * enqueue_hrtimer - internal function to (re)start a timer
342 * The timer is inserted in expiry order. Insertion into the
343 * red black tree is O(log(n)). Must hold the base lock.
345 static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
347 struct rb_node **link = &base->active.rb_node;
348 struct rb_node *parent = NULL;
349 struct hrtimer *entry;
352 * Find the right place in the rbtree:
356 entry = rb_entry(parent, struct hrtimer, node);
358 * We dont care about collisions. Nodes with
359 * the same expiry time stay together.
361 if (timer->expires.tv64 < entry->expires.tv64)
362 link = &(*link)->rb_left;
364 link = &(*link)->rb_right;
368 * Insert the timer to the rbtree and check whether it
369 * replaces the first pending timer
371 rb_link_node(&timer->node, parent, link);
372 rb_insert_color(&timer->node, &base->active);
374 if (!base->first || timer->expires.tv64 <
375 rb_entry(base->first, struct hrtimer, node)->expires.tv64)
376 base->first = &timer->node;
380 * __remove_hrtimer - internal function to remove a timer
382 * Caller must hold the base lock.
384 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
387 * Remove the timer from the rbtree and replace the
388 * first entry pointer if necessary.
390 if (base->first == &timer->node)
391 base->first = rb_next(&timer->node);
392 rb_erase(&timer->node, &base->active);
393 rb_set_parent(&timer->node, &timer->node);
397 * remove hrtimer, called with base lock held
400 remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
402 if (hrtimer_active(timer)) {
403 __remove_hrtimer(timer, base);
410 * hrtimer_start - (re)start an relative timer on the current CPU
411 * @timer: the timer to be added
413 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
417 * 1 when the timer was active
420 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
422 struct hrtimer_base *base, *new_base;
426 base = lock_hrtimer_base(timer, &flags);
428 /* Remove an active timer from the queue: */
429 ret = remove_hrtimer(timer, base);
431 /* Switch the timer base, if necessary: */
432 new_base = switch_hrtimer_base(timer, base);
434 if (mode == HRTIMER_REL) {
435 tim = ktime_add(tim, new_base->get_time());
437 * CONFIG_TIME_LOW_RES is a temporary way for architectures
438 * to signal that they simply return xtime in
439 * do_gettimeoffset(). In this case we want to round up by
440 * resolution when starting a relative timer, to avoid short
441 * timeouts. This will go away with the GTOD framework.
443 #ifdef CONFIG_TIME_LOW_RES
444 tim = ktime_add(tim, base->resolution);
447 timer->expires = tim;
449 enqueue_hrtimer(timer, new_base);
451 unlock_hrtimer_base(timer, &flags);
455 EXPORT_SYMBOL_GPL(hrtimer_start);
458 * hrtimer_try_to_cancel - try to deactivate a timer
459 * @timer: hrtimer to stop
462 * 0 when the timer was not active
463 * 1 when the timer was active
464 * -1 when the timer is currently excuting the callback function and
467 int hrtimer_try_to_cancel(struct hrtimer *timer)
469 struct hrtimer_base *base;
473 base = lock_hrtimer_base(timer, &flags);
475 if (base->curr_timer != timer)
476 ret = remove_hrtimer(timer, base);
478 unlock_hrtimer_base(timer, &flags);
483 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
486 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
487 * @timer: the timer to be cancelled
490 * 0 when the timer was not active
491 * 1 when the timer was active
493 int hrtimer_cancel(struct hrtimer *timer)
496 int ret = hrtimer_try_to_cancel(timer);
503 EXPORT_SYMBOL_GPL(hrtimer_cancel);
506 * hrtimer_get_remaining - get remaining time for the timer
507 * @timer: the timer to read
509 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
511 struct hrtimer_base *base;
515 base = lock_hrtimer_base(timer, &flags);
516 rem = ktime_sub(timer->expires, timer->base->get_time());
517 unlock_hrtimer_base(timer, &flags);
521 EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
523 #ifdef CONFIG_NO_IDLE_HZ
525 * hrtimer_get_next_event - get the time until next expiry event
527 * Returns the delta to the next expiry event or KTIME_MAX if no timer
530 ktime_t hrtimer_get_next_event(void)
532 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
533 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
537 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
538 struct hrtimer *timer;
540 spin_lock_irqsave(&base->lock, flags);
542 spin_unlock_irqrestore(&base->lock, flags);
545 timer = rb_entry(base->first, struct hrtimer, node);
546 delta.tv64 = timer->expires.tv64;
547 spin_unlock_irqrestore(&base->lock, flags);
548 delta = ktime_sub(delta, base->get_time());
549 if (delta.tv64 < mindelta.tv64)
550 mindelta.tv64 = delta.tv64;
552 if (mindelta.tv64 < 0)
559 * hrtimer_init - initialize a timer to the given clock
560 * @timer: the timer to be initialized
561 * @clock_id: the clock to be used
562 * @mode: timer mode abs/rel
564 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
565 enum hrtimer_mode mode)
567 struct hrtimer_base *bases;
569 memset(timer, 0, sizeof(struct hrtimer));
571 bases = __raw_get_cpu_var(hrtimer_bases);
573 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
574 clock_id = CLOCK_MONOTONIC;
576 timer->base = &bases[clock_id];
577 rb_set_parent(&timer->node, &timer->node);
579 EXPORT_SYMBOL_GPL(hrtimer_init);
582 * hrtimer_get_res - get the timer resolution for a clock
583 * @which_clock: which clock to query
584 * @tp: pointer to timespec variable to store the resolution
586 * Store the resolution of the clock selected by which_clock in the
587 * variable pointed to by tp.
589 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
591 struct hrtimer_base *bases;
593 bases = __raw_get_cpu_var(hrtimer_bases);
594 *tp = ktime_to_timespec(bases[which_clock].resolution);
598 EXPORT_SYMBOL_GPL(hrtimer_get_res);
601 * Expire the per base hrtimer-queue:
603 static inline void run_hrtimer_queue(struct hrtimer_base *base)
605 struct rb_node *node;
610 if (base->get_softirq_time)
611 base->softirq_time = base->get_softirq_time();
613 spin_lock_irq(&base->lock);
615 while ((node = base->first)) {
616 struct hrtimer *timer;
617 int (*fn)(struct hrtimer *);
620 timer = rb_entry(node, struct hrtimer, node);
621 if (base->softirq_time.tv64 <= timer->expires.tv64)
624 fn = timer->function;
625 set_curr_timer(base, timer);
626 __remove_hrtimer(timer, base);
627 spin_unlock_irq(&base->lock);
631 spin_lock_irq(&base->lock);
633 if (restart != HRTIMER_NORESTART) {
634 BUG_ON(hrtimer_active(timer));
635 enqueue_hrtimer(timer, base);
638 set_curr_timer(base, NULL);
639 spin_unlock_irq(&base->lock);
643 * Called from timer softirq every jiffy, expire hrtimers:
645 void hrtimer_run_queues(void)
647 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
650 hrtimer_get_softirq_time(base);
652 for (i = 0; i < MAX_HRTIMER_BASES; i++)
653 run_hrtimer_queue(&base[i]);
657 * Sleep related functions:
659 static int hrtimer_wakeup(struct hrtimer *timer)
661 struct hrtimer_sleeper *t =
662 container_of(timer, struct hrtimer_sleeper, timer);
663 struct task_struct *task = t->task;
667 wake_up_process(task);
669 return HRTIMER_NORESTART;
672 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
674 sl->timer.function = hrtimer_wakeup;
678 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
680 hrtimer_init_sleeper(t, current);
683 set_current_state(TASK_INTERRUPTIBLE);
684 hrtimer_start(&t->timer, t->timer.expires, mode);
688 hrtimer_cancel(&t->timer);
691 } while (t->task && !signal_pending(current));
693 return t->task == NULL;
696 static long __sched nanosleep_restart(struct restart_block *restart)
698 struct hrtimer_sleeper t;
699 struct timespec __user *rmtp;
703 restart->fn = do_no_restart_syscall;
705 hrtimer_init(&t.timer, restart->arg3, HRTIMER_ABS);
706 t.timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
708 if (do_nanosleep(&t, HRTIMER_ABS))
711 rmtp = (struct timespec __user *) restart->arg2;
713 time = ktime_sub(t.timer.expires, t.timer.base->get_time());
716 tu = ktime_to_timespec(time);
717 if (copy_to_user(rmtp, &tu, sizeof(tu)))
721 restart->fn = nanosleep_restart;
723 /* The other values in restart are already filled in */
724 return -ERESTART_RESTARTBLOCK;
727 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
728 const enum hrtimer_mode mode, const clockid_t clockid)
730 struct restart_block *restart;
731 struct hrtimer_sleeper t;
735 hrtimer_init(&t.timer, clockid, mode);
736 t.timer.expires = timespec_to_ktime(*rqtp);
737 if (do_nanosleep(&t, mode))
740 /* Absolute timers do not update the rmtp value and restart: */
741 if (mode == HRTIMER_ABS)
742 return -ERESTARTNOHAND;
745 rem = ktime_sub(t.timer.expires, t.timer.base->get_time());
748 tu = ktime_to_timespec(rem);
749 if (copy_to_user(rmtp, &tu, sizeof(tu)))
753 restart = ¤t_thread_info()->restart_block;
754 restart->fn = nanosleep_restart;
755 restart->arg0 = t.timer.expires.tv64 & 0xFFFFFFFF;
756 restart->arg1 = t.timer.expires.tv64 >> 32;
757 restart->arg2 = (unsigned long) rmtp;
758 restart->arg3 = (unsigned long) t.timer.base->index;
760 return -ERESTART_RESTARTBLOCK;
764 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
768 if (copy_from_user(&tu, rqtp, sizeof(tu)))
771 if (!timespec_valid(&tu))
774 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
778 * Functions related to boot-time initialization:
780 static void __devinit init_hrtimers_cpu(int cpu)
782 struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
785 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
786 spin_lock_init(&base->lock);
787 lockdep_set_class(&base->lock, &base->lock_key);
791 #ifdef CONFIG_HOTPLUG_CPU
793 static void migrate_hrtimer_list(struct hrtimer_base *old_base,
794 struct hrtimer_base *new_base)
796 struct hrtimer *timer;
797 struct rb_node *node;
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);
807 static void migrate_hrtimers(int cpu)
809 struct hrtimer_base *old_base, *new_base;
812 BUG_ON(cpu_online(cpu));
813 old_base = per_cpu(hrtimer_bases, cpu);
814 new_base = get_cpu_var(hrtimer_bases);
818 for (i = 0; i < MAX_HRTIMER_BASES; i++) {
820 spin_lock(&new_base->lock);
821 spin_lock(&old_base->lock);
823 BUG_ON(old_base->curr_timer);
825 migrate_hrtimer_list(old_base, new_base);
827 spin_unlock(&old_base->lock);
828 spin_unlock(&new_base->lock);
834 put_cpu_var(hrtimer_bases);
836 #endif /* CONFIG_HOTPLUG_CPU */
838 static int __devinit hrtimer_cpu_notify(struct notifier_block *self,
839 unsigned long action, void *hcpu)
841 long cpu = (long)hcpu;
846 init_hrtimers_cpu(cpu);
849 #ifdef CONFIG_HOTPLUG_CPU
851 migrate_hrtimers(cpu);
862 static struct notifier_block __devinitdata hrtimers_nb = {
863 .notifier_call = hrtimer_cpu_notify,
866 void __init hrtimers_init(void)
868 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
869 (void *)(long)smp_processor_id());
870 register_cpu_notifier(&hrtimers_nb);