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 * Functions and macros which are different for UP/SMP systems are kept in a
131 #define set_curr_timer(b, t) do { (b)->curr_timer = (t); } while (0)
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.
138 * So __run_timers/migrate_timers can safely modify all timers which could
139 * be found on the lists/queues.
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
145 static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
146 unsigned long *flags)
148 struct hrtimer_base *base;
152 if (likely(base != NULL)) {
153 spin_lock_irqsave(&base->lock, *flags);
154 if (likely(base == timer->base))
156 /* The timer has migrated to another CPU: */
157 spin_unlock_irqrestore(&base->lock, *flags);
164 * Switch the timer base to the current CPU when possible.
166 static inline struct hrtimer_base *
167 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
169 struct hrtimer_base *new_base;
171 new_base = &__get_cpu_var(hrtimer_bases[base->index]);
173 if (base != new_base) {
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.
183 if (unlikely(base->curr_timer == timer))
186 /* See the comment in lock_timer_base() */
188 spin_unlock(&base->lock);
189 spin_lock(&new_base->lock);
190 timer->base = new_base;
195 #else /* CONFIG_SMP */
197 #define set_curr_timer(b, t) do { } while (0)
199 static inline struct hrtimer_base *
200 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
202 struct hrtimer_base *base = timer->base;
204 spin_lock_irqsave(&base->lock, *flags);
209 #define switch_hrtimer_base(t, b) (b)
211 #endif /* !CONFIG_SMP */
214 * Functions for the union type storage format of ktime_t which are
215 * too large for inlining:
217 #if BITS_PER_LONG < 64
218 # ifndef CONFIG_KTIME_SCALAR
220 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
223 * @nsec: the scalar nsec value to add
225 * Returns the sum of kt and nsec in ktime_t format
227 ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
231 if (likely(nsec < NSEC_PER_SEC)) {
234 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
236 tmp = ktime_set((long)nsec, rem);
239 return ktime_add(kt, tmp);
242 #else /* CONFIG_KTIME_SCALAR */
244 # endif /* !CONFIG_KTIME_SCALAR */
247 * Divide a ktime value by a nanosecond value
249 static unsigned long ktime_divns(const ktime_t kt, nsec_t div)
254 dclc = dns = ktime_to_ns(kt);
256 /* Make sure the divisor is less than 2^32: */
262 do_div(dclc, (unsigned long) div);
264 return (unsigned long) dclc;
267 #else /* BITS_PER_LONG < 64 */
268 # define ktime_divns(kt, div) (unsigned long)((kt).tv64 / (div))
269 #endif /* BITS_PER_LONG >= 64 */
272 * Counterpart to lock_timer_base above:
275 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
277 spin_unlock_irqrestore(&timer->base->lock, *flags);
281 * hrtimer_forward - forward the timer expiry
283 * @timer: hrtimer to forward
284 * @interval: the interval to forward
286 * Forward the timer expiry so it will expire in the future.
287 * Returns the number of overruns.
290 hrtimer_forward(struct hrtimer *timer, ktime_t interval)
292 unsigned long orun = 1;
295 now = timer->base->get_time();
297 delta = ktime_sub(now, timer->expires);
302 if (interval.tv64 < timer->base->resolution.tv64)
303 interval.tv64 = timer->base->resolution.tv64;
305 if (unlikely(delta.tv64 >= interval.tv64)) {
306 nsec_t incr = ktime_to_ns(interval);
308 orun = ktime_divns(delta, incr);
309 timer->expires = ktime_add_ns(timer->expires, incr * orun);
310 if (timer->expires.tv64 > now.tv64)
313 * This (and the ktime_add() below) is the
314 * correction for exact:
318 timer->expires = ktime_add(timer->expires, interval);
324 * enqueue_hrtimer - internal function to (re)start a timer
326 * The timer is inserted in expiry order. Insertion into the
327 * red black tree is O(log(n)). Must hold the base lock.
329 static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
331 struct rb_node **link = &base->active.rb_node;
332 struct rb_node *parent = NULL;
333 struct hrtimer *entry;
336 * Find the right place in the rbtree:
340 entry = rb_entry(parent, struct hrtimer, node);
342 * We dont care about collisions. Nodes with
343 * the same expiry time stay together.
345 if (timer->expires.tv64 < entry->expires.tv64)
346 link = &(*link)->rb_left;
348 link = &(*link)->rb_right;
352 * Insert the timer to the rbtree and check whether it
353 * replaces the first pending timer
355 rb_link_node(&timer->node, parent, link);
356 rb_insert_color(&timer->node, &base->active);
358 timer->state = HRTIMER_PENDING;
360 if (!base->first || timer->expires.tv64 <
361 rb_entry(base->first, struct hrtimer, node)->expires.tv64)
362 base->first = &timer->node;
366 * __remove_hrtimer - internal function to remove a timer
368 * Caller must hold the base lock.
370 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
373 * Remove the timer from the rbtree and replace the
374 * first entry pointer if necessary.
376 if (base->first == &timer->node)
377 base->first = rb_next(&timer->node);
378 rb_erase(&timer->node, &base->active);
382 * remove hrtimer, called with base lock held
385 remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
387 if (hrtimer_active(timer)) {
388 __remove_hrtimer(timer, base);
389 timer->state = HRTIMER_INACTIVE;
396 * hrtimer_start - (re)start an relative timer on the current CPU
398 * @timer: the timer to be added
400 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
404 * 1 when the timer was active
407 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
409 struct hrtimer_base *base, *new_base;
413 base = lock_hrtimer_base(timer, &flags);
415 /* Remove an active timer from the queue: */
416 ret = remove_hrtimer(timer, base);
418 /* Switch the timer base, if necessary: */
419 new_base = switch_hrtimer_base(timer, base);
421 if (mode == HRTIMER_REL) {
422 tim = ktime_add(tim, new_base->get_time());
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.
430 #ifdef CONFIG_TIME_LOW_RES
431 tim = ktime_add(tim, base->resolution);
434 timer->expires = tim;
436 enqueue_hrtimer(timer, new_base);
438 unlock_hrtimer_base(timer, &flags);
444 * hrtimer_try_to_cancel - try to deactivate a timer
446 * @timer: hrtimer to stop
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
454 int hrtimer_try_to_cancel(struct hrtimer *timer)
456 struct hrtimer_base *base;
460 base = lock_hrtimer_base(timer, &flags);
462 if (base->curr_timer != timer)
463 ret = remove_hrtimer(timer, base);
465 unlock_hrtimer_base(timer, &flags);
472 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
474 * @timer: the timer to be cancelled
477 * 0 when the timer was not active
478 * 1 when the timer was active
480 int hrtimer_cancel(struct hrtimer *timer)
483 int ret = hrtimer_try_to_cancel(timer);
491 * hrtimer_get_remaining - get remaining time for the timer
493 * @timer: the timer to read
495 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
497 struct hrtimer_base *base;
501 base = lock_hrtimer_base(timer, &flags);
502 rem = ktime_sub(timer->expires, timer->base->get_time());
503 unlock_hrtimer_base(timer, &flags);
508 #ifdef CONFIG_NO_IDLE_HZ
510 * hrtimer_get_next_event - get the time until next expiry event
512 * Returns the delta to the next expiry event or KTIME_MAX if no timer
515 ktime_t hrtimer_get_next_event(void)
517 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
518 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
522 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
523 struct hrtimer *timer;
525 spin_lock_irqsave(&base->lock, flags);
527 spin_unlock_irqrestore(&base->lock, flags);
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;
537 if (mindelta.tv64 < 0)
544 * hrtimer_init - initialize a timer to the given clock
546 * @timer: the timer to be initialized
547 * @clock_id: the clock to be used
548 * @mode: timer mode abs/rel
550 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
551 enum hrtimer_mode mode)
553 struct hrtimer_base *bases;
555 memset(timer, 0, sizeof(struct hrtimer));
557 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
559 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
560 clock_id = CLOCK_MONOTONIC;
562 timer->base = &bases[clock_id];
566 * hrtimer_get_res - get the timer resolution for a clock
568 * @which_clock: which clock to query
569 * @tp: pointer to timespec variable to store the resolution
571 * Store the resolution of the clock selected by which_clock in the
572 * variable pointed to by tp.
574 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
576 struct hrtimer_base *bases;
578 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
579 *tp = ktime_to_timespec(bases[which_clock].resolution);
585 * Expire the per base hrtimer-queue:
587 static inline void run_hrtimer_queue(struct hrtimer_base *base)
589 ktime_t now = base->get_time();
590 struct rb_node *node;
592 spin_lock_irq(&base->lock);
594 while ((node = base->first)) {
595 struct hrtimer *timer;
600 timer = rb_entry(node, struct hrtimer, node);
601 if (now.tv64 <= timer->expires.tv64)
604 fn = timer->function;
606 set_curr_timer(base, timer);
607 timer->state = HRTIMER_RUNNING;
608 __remove_hrtimer(timer, base);
609 spin_unlock_irq(&base->lock);
612 * fn == NULL is special case for the simplest timer
613 * variant - wake up process and do not restart:
616 wake_up_process(data);
617 restart = HRTIMER_NORESTART;
621 spin_lock_irq(&base->lock);
623 /* Another CPU has added back the timer */
624 if (timer->state != HRTIMER_RUNNING)
627 if (restart == HRTIMER_RESTART)
628 enqueue_hrtimer(timer, base);
630 timer->state = HRTIMER_EXPIRED;
632 set_curr_timer(base, NULL);
633 spin_unlock_irq(&base->lock);
637 * Called from timer softirq every jiffy, expire hrtimers:
639 void hrtimer_run_queues(void)
641 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
644 for (i = 0; i < MAX_HRTIMER_BASES; i++)
645 run_hrtimer_queue(&base[i]);
649 * Sleep related functions:
653 * schedule_hrtimer - sleep until timeout
655 * @timer: hrtimer variable initialized with the correct clock base
656 * @mode: timeout value is abs/rel
658 * Make the current task sleep until @timeout is
661 * You can set the task state as follows -
663 * %TASK_UNINTERRUPTIBLE - at least @timeout is guaranteed to
664 * pass before the routine returns. The routine will return 0
666 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
667 * delivered to the current task. In this case the remaining time
670 * The current task state is guaranteed to be TASK_RUNNING when this
673 static ktime_t __sched
674 schedule_hrtimer(struct hrtimer *timer, const enum hrtimer_mode mode)
676 /* fn stays NULL, meaning single-shot wakeup: */
677 timer->data = current;
679 hrtimer_start(timer, timer->expires, mode);
682 hrtimer_cancel(timer);
684 /* Return the remaining time: */
685 if (timer->state != HRTIMER_EXPIRED)
686 return ktime_sub(timer->expires, timer->base->get_time());
688 return (ktime_t) {.tv64 = 0 };
691 static inline ktime_t __sched
692 schedule_hrtimer_interruptible(struct hrtimer *timer,
693 const enum hrtimer_mode mode)
695 set_current_state(TASK_INTERRUPTIBLE);
697 return schedule_hrtimer(timer, mode);
700 static long __sched nanosleep_restart(struct restart_block *restart)
702 struct timespec __user *rmtp;
704 void *rfn_save = restart->fn;
705 struct hrtimer timer;
708 restart->fn = do_no_restart_syscall;
710 hrtimer_init(&timer, (clockid_t) restart->arg3, HRTIMER_ABS);
712 timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
714 rem = schedule_hrtimer_interruptible(&timer, HRTIMER_ABS);
719 rmtp = (struct timespec __user *) restart->arg2;
720 tu = ktime_to_timespec(rem);
721 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
724 restart->fn = rfn_save;
726 /* The other values in restart are already filled in */
727 return -ERESTART_RESTARTBLOCK;
730 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
731 const enum hrtimer_mode mode, const clockid_t clockid)
733 struct restart_block *restart;
734 struct hrtimer timer;
738 hrtimer_init(&timer, clockid, mode);
740 timer.expires = timespec_to_ktime(*rqtp);
742 rem = schedule_hrtimer_interruptible(&timer, mode);
746 /* Absolute timers do not update the rmtp value and restart: */
747 if (mode == HRTIMER_ABS)
748 return -ERESTARTNOHAND;
750 tu = ktime_to_timespec(rem);
752 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
755 restart = ¤t_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;
762 return -ERESTART_RESTARTBLOCK;
766 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
770 if (copy_from_user(&tu, rqtp, sizeof(tu)))
773 if (!timespec_valid(&tu))
776 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
780 * Functions related to boot-time initialization:
782 static void __devinit init_hrtimers_cpu(int cpu)
784 struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
787 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
788 spin_lock_init(&base->lock);
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);