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
102 * @ts: pointer to timespec variable
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.
108 void ktime_get_ts(struct timespec *ts)
110 struct timespec tomono;
114 seq = read_seqbegin(&xtime_lock);
116 tomono = wall_to_monotonic;
118 } while (read_seqretry(&xtime_lock, seq));
120 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
121 ts->tv_nsec + tomono.tv_nsec);
123 EXPORT_SYMBOL_GPL(ktime_get_ts);
126 * Get the coarse grained time at the softirq based on xtime and
129 static void hrtimer_get_softirq_time(struct hrtimer_base *base)
131 ktime_t xtim, tomono;
135 seq = read_seqbegin(&xtime_lock);
136 xtim = timespec_to_ktime(xtime);
137 tomono = timespec_to_ktime(wall_to_monotonic);
139 } while (read_seqretry(&xtime_lock, seq));
141 base[CLOCK_REALTIME].softirq_time = xtim;
142 base[CLOCK_MONOTONIC].softirq_time = ktime_add(xtim, tomono);
146 * Functions and macros which are different for UP/SMP systems are kept in a
151 #define set_curr_timer(b, t) do { (b)->curr_timer = (t); } while (0)
154 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
155 * means that all timers which are tied to this base via timer->base are
156 * locked, and the base itself is locked too.
158 * So __run_timers/migrate_timers can safely modify all timers which could
159 * be found on the lists/queues.
161 * When the timer's base is locked, and the timer removed from list, it is
162 * possible to set timer->base = NULL and drop the lock: the timer remains
165 static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
166 unsigned long *flags)
168 struct hrtimer_base *base;
172 if (likely(base != NULL)) {
173 spin_lock_irqsave(&base->lock, *flags);
174 if (likely(base == timer->base))
176 /* The timer has migrated to another CPU: */
177 spin_unlock_irqrestore(&base->lock, *flags);
184 * Switch the timer base to the current CPU when possible.
186 static inline struct hrtimer_base *
187 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
189 struct hrtimer_base *new_base;
191 new_base = &__get_cpu_var(hrtimer_bases[base->index]);
193 if (base != new_base) {
195 * We are trying to schedule the timer on the local CPU.
196 * However we can't change timer's base while it is running,
197 * so we keep it on the same CPU. No hassle vs. reprogramming
198 * the event source in the high resolution case. The softirq
199 * code will take care of this when the timer function has
200 * completed. There is no conflict as we hold the lock until
201 * the timer is enqueued.
203 if (unlikely(base->curr_timer == timer))
206 /* See the comment in lock_timer_base() */
208 spin_unlock(&base->lock);
209 spin_lock(&new_base->lock);
210 timer->base = new_base;
215 #else /* CONFIG_SMP */
217 #define set_curr_timer(b, t) do { } while (0)
219 static inline struct hrtimer_base *
220 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
222 struct hrtimer_base *base = timer->base;
224 spin_lock_irqsave(&base->lock, *flags);
229 #define switch_hrtimer_base(t, b) (b)
231 #endif /* !CONFIG_SMP */
234 * Functions for the union type storage format of ktime_t which are
235 * too large for inlining:
237 #if BITS_PER_LONG < 64
238 # ifndef CONFIG_KTIME_SCALAR
240 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
243 * @nsec: the scalar nsec value to add
245 * Returns the sum of kt and nsec in ktime_t format
247 ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
251 if (likely(nsec < NSEC_PER_SEC)) {
254 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
256 tmp = ktime_set((long)nsec, rem);
259 return ktime_add(kt, tmp);
262 #else /* CONFIG_KTIME_SCALAR */
264 # endif /* !CONFIG_KTIME_SCALAR */
267 * Divide a ktime value by a nanosecond value
269 static unsigned long ktime_divns(const ktime_t kt, s64 div)
274 dclc = dns = ktime_to_ns(kt);
276 /* Make sure the divisor is less than 2^32: */
282 do_div(dclc, (unsigned long) div);
284 return (unsigned long) dclc;
287 #else /* BITS_PER_LONG < 64 */
288 # define ktime_divns(kt, div) (unsigned long)((kt).tv64 / (div))
289 #endif /* BITS_PER_LONG >= 64 */
292 * Counterpart to lock_timer_base above:
295 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
297 spin_unlock_irqrestore(&timer->base->lock, *flags);
301 * hrtimer_forward - forward the timer expiry
303 * @timer: hrtimer to forward
304 * @now: forward past this time
305 * @interval: the interval to forward
307 * Forward the timer expiry so it will expire in the future.
308 * Returns the number of overruns.
311 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
313 unsigned long orun = 1;
316 delta = ktime_sub(now, timer->expires);
321 if (interval.tv64 < timer->base->resolution.tv64)
322 interval.tv64 = timer->base->resolution.tv64;
324 if (unlikely(delta.tv64 >= interval.tv64)) {
325 s64 incr = ktime_to_ns(interval);
327 orun = ktime_divns(delta, incr);
328 timer->expires = ktime_add_ns(timer->expires, incr * orun);
329 if (timer->expires.tv64 > now.tv64)
332 * This (and the ktime_add() below) is the
333 * correction for exact:
337 timer->expires = ktime_add(timer->expires, interval);
343 * enqueue_hrtimer - internal function to (re)start a timer
345 * The timer is inserted in expiry order. Insertion into the
346 * red black tree is O(log(n)). Must hold the base lock.
348 static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
350 struct rb_node **link = &base->active.rb_node;
351 struct rb_node *parent = NULL;
352 struct hrtimer *entry;
355 * Find the right place in the rbtree:
359 entry = rb_entry(parent, struct hrtimer, node);
361 * We dont care about collisions. Nodes with
362 * the same expiry time stay together.
364 if (timer->expires.tv64 < entry->expires.tv64)
365 link = &(*link)->rb_left;
367 link = &(*link)->rb_right;
371 * Insert the timer to the rbtree and check whether it
372 * replaces the first pending timer
374 rb_link_node(&timer->node, parent, link);
375 rb_insert_color(&timer->node, &base->active);
377 if (!base->first || timer->expires.tv64 <
378 rb_entry(base->first, struct hrtimer, node)->expires.tv64)
379 base->first = &timer->node;
383 * __remove_hrtimer - internal function to remove a timer
385 * Caller must hold the base lock.
387 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
390 * Remove the timer from the rbtree and replace the
391 * first entry pointer if necessary.
393 if (base->first == &timer->node)
394 base->first = rb_next(&timer->node);
395 rb_erase(&timer->node, &base->active);
396 timer->node.rb_parent = HRTIMER_INACTIVE;
400 * remove hrtimer, called with base lock held
403 remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
405 if (hrtimer_active(timer)) {
406 __remove_hrtimer(timer, base);
413 * hrtimer_start - (re)start an relative timer on the current CPU
415 * @timer: the timer to be added
417 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
421 * 1 when the timer was active
424 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
426 struct hrtimer_base *base, *new_base;
430 base = lock_hrtimer_base(timer, &flags);
432 /* Remove an active timer from the queue: */
433 ret = remove_hrtimer(timer, base);
435 /* Switch the timer base, if necessary: */
436 new_base = switch_hrtimer_base(timer, base);
438 if (mode == HRTIMER_REL) {
439 tim = ktime_add(tim, new_base->get_time());
441 * CONFIG_TIME_LOW_RES is a temporary way for architectures
442 * to signal that they simply return xtime in
443 * do_gettimeoffset(). In this case we want to round up by
444 * resolution when starting a relative timer, to avoid short
445 * timeouts. This will go away with the GTOD framework.
447 #ifdef CONFIG_TIME_LOW_RES
448 tim = ktime_add(tim, base->resolution);
451 timer->expires = tim;
453 enqueue_hrtimer(timer, new_base);
455 unlock_hrtimer_base(timer, &flags);
461 * hrtimer_try_to_cancel - try to deactivate a timer
463 * @timer: hrtimer to stop
466 * 0 when the timer was not active
467 * 1 when the timer was active
468 * -1 when the timer is currently excuting the callback function and
471 int hrtimer_try_to_cancel(struct hrtimer *timer)
473 struct hrtimer_base *base;
477 base = lock_hrtimer_base(timer, &flags);
479 if (base->curr_timer != timer)
480 ret = remove_hrtimer(timer, base);
482 unlock_hrtimer_base(timer, &flags);
489 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
491 * @timer: the timer to be cancelled
494 * 0 when the timer was not active
495 * 1 when the timer was active
497 int hrtimer_cancel(struct hrtimer *timer)
500 int ret = hrtimer_try_to_cancel(timer);
508 * hrtimer_get_remaining - get remaining time for the timer
510 * @timer: the timer to read
512 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
514 struct hrtimer_base *base;
518 base = lock_hrtimer_base(timer, &flags);
519 rem = ktime_sub(timer->expires, timer->base->get_time());
520 unlock_hrtimer_base(timer, &flags);
525 #ifdef CONFIG_NO_IDLE_HZ
527 * hrtimer_get_next_event - get the time until next expiry event
529 * Returns the delta to the next expiry event or KTIME_MAX if no timer
532 ktime_t hrtimer_get_next_event(void)
534 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
535 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
539 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
540 struct hrtimer *timer;
542 spin_lock_irqsave(&base->lock, flags);
544 spin_unlock_irqrestore(&base->lock, flags);
547 timer = rb_entry(base->first, struct hrtimer, node);
548 delta.tv64 = timer->expires.tv64;
549 spin_unlock_irqrestore(&base->lock, flags);
550 delta = ktime_sub(delta, base->get_time());
551 if (delta.tv64 < mindelta.tv64)
552 mindelta.tv64 = delta.tv64;
554 if (mindelta.tv64 < 0)
561 * hrtimer_init - initialize a timer to the given clock
563 * @timer: the timer to be initialized
564 * @clock_id: the clock to be used
565 * @mode: timer mode abs/rel
567 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
568 enum hrtimer_mode mode)
570 struct hrtimer_base *bases;
572 memset(timer, 0, sizeof(struct hrtimer));
574 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
576 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
577 clock_id = CLOCK_MONOTONIC;
579 timer->base = &bases[clock_id];
580 timer->node.rb_parent = HRTIMER_INACTIVE;
584 * hrtimer_get_res - get the timer resolution for a clock
586 * @which_clock: which clock to query
587 * @tp: pointer to timespec variable to store the resolution
589 * Store the resolution of the clock selected by which_clock in the
590 * variable pointed to by tp.
592 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
594 struct hrtimer_base *bases;
596 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
597 *tp = ktime_to_timespec(bases[which_clock].resolution);
603 * Expire the per base hrtimer-queue:
605 static inline void run_hrtimer_queue(struct hrtimer_base *base)
607 struct rb_node *node;
609 if (base->get_softirq_time)
610 base->softirq_time = base->get_softirq_time();
612 spin_lock_irq(&base->lock);
614 while ((node = base->first)) {
615 struct hrtimer *timer;
616 int (*fn)(struct hrtimer *);
619 timer = rb_entry(node, struct hrtimer, node);
620 if (base->softirq_time.tv64 <= timer->expires.tv64)
623 fn = timer->function;
624 set_curr_timer(base, timer);
625 __remove_hrtimer(timer, base);
626 spin_unlock_irq(&base->lock);
630 spin_lock_irq(&base->lock);
632 if (restart != HRTIMER_NORESTART) {
633 BUG_ON(hrtimer_active(timer));
634 enqueue_hrtimer(timer, base);
637 set_curr_timer(base, NULL);
638 spin_unlock_irq(&base->lock);
642 * Called from timer softirq every jiffy, expire hrtimers:
644 void hrtimer_run_queues(void)
646 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
649 hrtimer_get_softirq_time(base);
651 for (i = 0; i < MAX_HRTIMER_BASES; i++)
652 run_hrtimer_queue(&base[i]);
656 * Sleep related functions:
659 struct sleep_hrtimer {
660 struct hrtimer timer;
661 struct task_struct *task;
665 static int nanosleep_wakeup(struct hrtimer *timer)
667 struct sleep_hrtimer *t =
668 container_of(timer, struct sleep_hrtimer, timer);
671 wake_up_process(t->task);
673 return HRTIMER_NORESTART;
676 static int __sched do_nanosleep(struct sleep_hrtimer *t, enum hrtimer_mode mode)
678 t->timer.function = nanosleep_wakeup;
683 set_current_state(TASK_INTERRUPTIBLE);
684 hrtimer_start(&t->timer, t->timer.expires, mode);
688 if (unlikely(!t->expired)) {
689 hrtimer_cancel(&t->timer);
692 } while (!t->expired && !signal_pending(current));
697 static long __sched nanosleep_restart(struct restart_block *restart)
699 struct sleep_hrtimer t;
700 struct timespec __user *rmtp;
704 restart->fn = do_no_restart_syscall;
706 hrtimer_init(&t.timer, restart->arg3, HRTIMER_ABS);
707 t.timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
709 if (do_nanosleep(&t, HRTIMER_ABS))
712 rmtp = (struct timespec __user *) restart->arg2;
714 time = ktime_sub(t.timer.expires, t.timer.base->get_time());
717 tu = ktime_to_timespec(time);
718 if (copy_to_user(rmtp, &tu, sizeof(tu)))
722 restart->fn = nanosleep_restart;
724 /* The other values in restart are already filled in */
725 return -ERESTART_RESTARTBLOCK;
728 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
729 const enum hrtimer_mode mode, const clockid_t clockid)
731 struct restart_block *restart;
732 struct sleep_hrtimer t;
736 hrtimer_init(&t.timer, clockid, mode);
737 t.timer.expires = timespec_to_ktime(*rqtp);
738 if (do_nanosleep(&t, mode))
741 /* Absolute timers do not update the rmtp value and restart: */
742 if (mode == HRTIMER_ABS)
743 return -ERESTARTNOHAND;
746 rem = ktime_sub(t.timer.expires, t.timer.base->get_time());
749 tu = ktime_to_timespec(rem);
750 if (copy_to_user(rmtp, &tu, sizeof(tu)))
754 restart = ¤t_thread_info()->restart_block;
755 restart->fn = nanosleep_restart;
756 restart->arg0 = t.timer.expires.tv64 & 0xFFFFFFFF;
757 restart->arg1 = t.timer.expires.tv64 >> 32;
758 restart->arg2 = (unsigned long) rmtp;
759 restart->arg3 = (unsigned long) t.timer.base->index;
761 return -ERESTART_RESTARTBLOCK;
765 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
769 if (copy_from_user(&tu, rqtp, sizeof(tu)))
772 if (!timespec_valid(&tu))
775 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
779 * Functions related to boot-time initialization:
781 static void __devinit init_hrtimers_cpu(int cpu)
783 struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
786 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
787 spin_lock_init(&base->lock);
790 #ifdef CONFIG_HOTPLUG_CPU
792 static void migrate_hrtimer_list(struct hrtimer_base *old_base,
793 struct hrtimer_base *new_base)
795 struct hrtimer *timer;
796 struct rb_node *node;
798 while ((node = rb_first(&old_base->active))) {
799 timer = rb_entry(node, struct hrtimer, node);
800 __remove_hrtimer(timer, old_base);
801 timer->base = new_base;
802 enqueue_hrtimer(timer, new_base);
806 static void migrate_hrtimers(int cpu)
808 struct hrtimer_base *old_base, *new_base;
811 BUG_ON(cpu_online(cpu));
812 old_base = per_cpu(hrtimer_bases, cpu);
813 new_base = get_cpu_var(hrtimer_bases);
817 for (i = 0; i < MAX_HRTIMER_BASES; i++) {
819 spin_lock(&new_base->lock);
820 spin_lock(&old_base->lock);
822 BUG_ON(old_base->curr_timer);
824 migrate_hrtimer_list(old_base, new_base);
826 spin_unlock(&old_base->lock);
827 spin_unlock(&new_base->lock);
833 put_cpu_var(hrtimer_bases);
835 #endif /* CONFIG_HOTPLUG_CPU */
837 static int __devinit hrtimer_cpu_notify(struct notifier_block *self,
838 unsigned long action, void *hcpu)
840 long cpu = (long)hcpu;
845 init_hrtimers_cpu(cpu);
848 #ifdef CONFIG_HOTPLUG_CPU
850 migrate_hrtimers(cpu);
861 static struct notifier_block __devinitdata hrtimers_nb = {
862 .notifier_call = hrtimer_cpu_notify,
865 void __init hrtimers_init(void)
867 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
868 (void *)(long)smp_processor_id());
869 register_cpu_notifier(&hrtimers_nb);