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, nsec_t 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 nsec_t 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 timer->state = HRTIMER_PENDING;
379 if (!base->first || timer->expires.tv64 <
380 rb_entry(base->first, struct hrtimer, node)->expires.tv64)
381 base->first = &timer->node;
385 * __remove_hrtimer - internal function to remove a timer
387 * Caller must hold the base lock.
389 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
392 * Remove the timer from the rbtree and replace the
393 * first entry pointer if necessary.
395 if (base->first == &timer->node)
396 base->first = rb_next(&timer->node);
397 rb_erase(&timer->node, &base->active);
401 * remove hrtimer, called with base lock held
404 remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
406 if (hrtimer_active(timer)) {
407 __remove_hrtimer(timer, base);
408 timer->state = HRTIMER_INACTIVE;
415 * hrtimer_start - (re)start an relative timer on the current CPU
417 * @timer: the timer to be added
419 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
423 * 1 when the timer was active
426 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
428 struct hrtimer_base *base, *new_base;
432 base = lock_hrtimer_base(timer, &flags);
434 /* Remove an active timer from the queue: */
435 ret = remove_hrtimer(timer, base);
437 /* Switch the timer base, if necessary: */
438 new_base = switch_hrtimer_base(timer, base);
440 if (mode == HRTIMER_REL) {
441 tim = ktime_add(tim, new_base->get_time());
443 * CONFIG_TIME_LOW_RES is a temporary way for architectures
444 * to signal that they simply return xtime in
445 * do_gettimeoffset(). In this case we want to round up by
446 * resolution when starting a relative timer, to avoid short
447 * timeouts. This will go away with the GTOD framework.
449 #ifdef CONFIG_TIME_LOW_RES
450 tim = ktime_add(tim, base->resolution);
453 timer->expires = tim;
455 enqueue_hrtimer(timer, new_base);
457 unlock_hrtimer_base(timer, &flags);
463 * hrtimer_try_to_cancel - try to deactivate a timer
465 * @timer: hrtimer to stop
468 * 0 when the timer was not active
469 * 1 when the timer was active
470 * -1 when the timer is currently excuting the callback function and
473 int hrtimer_try_to_cancel(struct hrtimer *timer)
475 struct hrtimer_base *base;
479 base = lock_hrtimer_base(timer, &flags);
481 if (base->curr_timer != timer)
482 ret = remove_hrtimer(timer, base);
484 unlock_hrtimer_base(timer, &flags);
491 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
493 * @timer: the timer to be cancelled
496 * 0 when the timer was not active
497 * 1 when the timer was active
499 int hrtimer_cancel(struct hrtimer *timer)
502 int ret = hrtimer_try_to_cancel(timer);
510 * hrtimer_get_remaining - get remaining time for the timer
512 * @timer: the timer to read
514 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
516 struct hrtimer_base *base;
520 base = lock_hrtimer_base(timer, &flags);
521 rem = ktime_sub(timer->expires, timer->base->get_time());
522 unlock_hrtimer_base(timer, &flags);
527 #ifdef CONFIG_NO_IDLE_HZ
529 * hrtimer_get_next_event - get the time until next expiry event
531 * Returns the delta to the next expiry event or KTIME_MAX if no timer
534 ktime_t hrtimer_get_next_event(void)
536 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
537 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
541 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
542 struct hrtimer *timer;
544 spin_lock_irqsave(&base->lock, flags);
546 spin_unlock_irqrestore(&base->lock, flags);
549 timer = rb_entry(base->first, struct hrtimer, node);
550 delta.tv64 = timer->expires.tv64;
551 spin_unlock_irqrestore(&base->lock, flags);
552 delta = ktime_sub(delta, base->get_time());
553 if (delta.tv64 < mindelta.tv64)
554 mindelta.tv64 = delta.tv64;
556 if (mindelta.tv64 < 0)
563 * hrtimer_init - initialize a timer to the given clock
565 * @timer: the timer to be initialized
566 * @clock_id: the clock to be used
567 * @mode: timer mode abs/rel
569 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
570 enum hrtimer_mode mode)
572 struct hrtimer_base *bases;
574 memset(timer, 0, sizeof(struct hrtimer));
576 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
578 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
579 clock_id = CLOCK_MONOTONIC;
581 timer->base = &bases[clock_id];
585 * hrtimer_get_res - get the timer resolution for a clock
587 * @which_clock: which clock to query
588 * @tp: pointer to timespec variable to store the resolution
590 * Store the resolution of the clock selected by which_clock in the
591 * variable pointed to by tp.
593 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
595 struct hrtimer_base *bases;
597 bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
598 *tp = ktime_to_timespec(bases[which_clock].resolution);
604 * Expire the per base hrtimer-queue:
606 static inline void run_hrtimer_queue(struct hrtimer_base *base)
608 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;
621 timer = rb_entry(node, struct hrtimer, node);
622 if (base->softirq_time.tv64 <= timer->expires.tv64)
625 fn = timer->function;
627 set_curr_timer(base, timer);
628 timer->state = HRTIMER_RUNNING;
629 __remove_hrtimer(timer, base);
630 spin_unlock_irq(&base->lock);
633 * fn == NULL is special case for the simplest timer
634 * variant - wake up process and do not restart:
637 wake_up_process(data);
638 restart = HRTIMER_NORESTART;
642 spin_lock_irq(&base->lock);
644 /* Another CPU has added back the timer */
645 if (timer->state != HRTIMER_RUNNING)
648 if (restart == HRTIMER_RESTART)
649 enqueue_hrtimer(timer, base);
651 timer->state = HRTIMER_EXPIRED;
653 set_curr_timer(base, NULL);
654 spin_unlock_irq(&base->lock);
658 * Called from timer softirq every jiffy, expire hrtimers:
660 void hrtimer_run_queues(void)
662 struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
665 hrtimer_get_softirq_time(base);
667 for (i = 0; i < MAX_HRTIMER_BASES; i++)
668 run_hrtimer_queue(&base[i]);
672 * Sleep related functions:
676 * schedule_hrtimer - sleep until timeout
678 * @timer: hrtimer variable initialized with the correct clock base
679 * @mode: timeout value is abs/rel
681 * Make the current task sleep until @timeout is
684 * You can set the task state as follows -
686 * %TASK_UNINTERRUPTIBLE - at least @timeout is guaranteed to
687 * pass before the routine returns. The routine will return 0
689 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
690 * delivered to the current task. In this case the remaining time
693 * The current task state is guaranteed to be TASK_RUNNING when this
696 static ktime_t __sched
697 schedule_hrtimer(struct hrtimer *timer, const enum hrtimer_mode mode)
699 /* fn stays NULL, meaning single-shot wakeup: */
700 timer->data = current;
702 hrtimer_start(timer, timer->expires, mode);
705 hrtimer_cancel(timer);
707 /* Return the remaining time: */
708 if (timer->state != HRTIMER_EXPIRED)
709 return ktime_sub(timer->expires, timer->base->get_time());
711 return (ktime_t) {.tv64 = 0 };
714 static inline ktime_t __sched
715 schedule_hrtimer_interruptible(struct hrtimer *timer,
716 const enum hrtimer_mode mode)
718 set_current_state(TASK_INTERRUPTIBLE);
720 return schedule_hrtimer(timer, mode);
723 static long __sched nanosleep_restart(struct restart_block *restart)
725 struct timespec __user *rmtp;
727 void *rfn_save = restart->fn;
728 struct hrtimer timer;
731 restart->fn = do_no_restart_syscall;
733 hrtimer_init(&timer, (clockid_t) restart->arg3, HRTIMER_ABS);
735 timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
737 rem = schedule_hrtimer_interruptible(&timer, HRTIMER_ABS);
742 rmtp = (struct timespec __user *) restart->arg2;
743 tu = ktime_to_timespec(rem);
744 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
747 restart->fn = rfn_save;
749 /* The other values in restart are already filled in */
750 return -ERESTART_RESTARTBLOCK;
753 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
754 const enum hrtimer_mode mode, const clockid_t clockid)
756 struct restart_block *restart;
757 struct hrtimer timer;
761 hrtimer_init(&timer, clockid, mode);
763 timer.expires = timespec_to_ktime(*rqtp);
765 rem = schedule_hrtimer_interruptible(&timer, mode);
769 /* Absolute timers do not update the rmtp value and restart: */
770 if (mode == HRTIMER_ABS)
771 return -ERESTARTNOHAND;
773 tu = ktime_to_timespec(rem);
775 if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
778 restart = ¤t_thread_info()->restart_block;
779 restart->fn = nanosleep_restart;
780 restart->arg0 = timer.expires.tv64 & 0xFFFFFFFF;
781 restart->arg1 = timer.expires.tv64 >> 32;
782 restart->arg2 = (unsigned long) rmtp;
783 restart->arg3 = (unsigned long) timer.base->index;
785 return -ERESTART_RESTARTBLOCK;
789 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
793 if (copy_from_user(&tu, rqtp, sizeof(tu)))
796 if (!timespec_valid(&tu))
799 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
803 * Functions related to boot-time initialization:
805 static void __devinit init_hrtimers_cpu(int cpu)
807 struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
810 for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
811 spin_lock_init(&base->lock);
814 #ifdef CONFIG_HOTPLUG_CPU
816 static void migrate_hrtimer_list(struct hrtimer_base *old_base,
817 struct hrtimer_base *new_base)
819 struct hrtimer *timer;
820 struct rb_node *node;
822 while ((node = rb_first(&old_base->active))) {
823 timer = rb_entry(node, struct hrtimer, node);
824 __remove_hrtimer(timer, old_base);
825 timer->base = new_base;
826 enqueue_hrtimer(timer, new_base);
830 static void migrate_hrtimers(int cpu)
832 struct hrtimer_base *old_base, *new_base;
835 BUG_ON(cpu_online(cpu));
836 old_base = per_cpu(hrtimer_bases, cpu);
837 new_base = get_cpu_var(hrtimer_bases);
841 for (i = 0; i < MAX_HRTIMER_BASES; i++) {
843 spin_lock(&new_base->lock);
844 spin_lock(&old_base->lock);
846 BUG_ON(old_base->curr_timer);
848 migrate_hrtimer_list(old_base, new_base);
850 spin_unlock(&old_base->lock);
851 spin_unlock(&new_base->lock);
857 put_cpu_var(hrtimer_bases);
859 #endif /* CONFIG_HOTPLUG_CPU */
861 static int __devinit hrtimer_cpu_notify(struct notifier_block *self,
862 unsigned long action, void *hcpu)
864 long cpu = (long)hcpu;
869 init_hrtimers_cpu(cpu);
872 #ifdef CONFIG_HOTPLUG_CPU
874 migrate_hrtimers(cpu);
885 static struct notifier_block __devinitdata hrtimers_nb = {
886 .notifier_call = hrtimer_cpu_notify,
889 void __init hrtimers_init(void)
891 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
892 (void *)(long)smp_processor_id());
893 register_cpu_notifier(&hrtimers_nb);