2 * linux/kernel/time/tick-sched.c
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
8 * No idle tick implementation for low and high resolution timers
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * Distribute under GPLv2.
14 #include <linux/cpu.h>
15 #include <linux/err.h>
16 #include <linux/hrtimer.h>
17 #include <linux/interrupt.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/percpu.h>
20 #include <linux/profile.h>
21 #include <linux/sched.h>
22 #include <linux/module.h>
24 #include <asm/irq_regs.h>
26 #include "tick-internal.h"
29 * Per cpu nohz control structure
31 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
36 static ktime_t last_jiffies_update;
38 struct tick_sched *tick_get_tick_sched(int cpu)
40 return &per_cpu(tick_cpu_sched, cpu);
44 * Must be called with interrupts disabled !
46 static void tick_do_update_jiffies64(ktime_t now)
48 unsigned long ticks = 0;
52 * Do a quick check without holding xtime_lock:
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 < tick_period.tv64)
58 /* Reevalute with xtime_lock held */
59 write_seqlock(&xtime_lock);
61 delta = ktime_sub(now, last_jiffies_update);
62 if (delta.tv64 >= tick_period.tv64) {
64 delta = ktime_sub(delta, tick_period);
65 last_jiffies_update = ktime_add(last_jiffies_update,
68 /* Slow path for long timeouts */
69 if (unlikely(delta.tv64 >= tick_period.tv64)) {
70 s64 incr = ktime_to_ns(tick_period);
72 ticks = ktime_divns(delta, incr);
74 last_jiffies_update = ktime_add_ns(last_jiffies_update,
79 /* Keep the tick_next_period variable up to date */
80 tick_next_period = ktime_add(last_jiffies_update, tick_period);
82 write_sequnlock(&xtime_lock);
86 * Initialize and return retrieve the jiffies update.
88 static ktime_t tick_init_jiffy_update(void)
92 write_seqlock(&xtime_lock);
93 /* Did we start the jiffies update yet ? */
94 if (last_jiffies_update.tv64 == 0)
95 last_jiffies_update = tick_next_period;
96 period = last_jiffies_update;
97 write_sequnlock(&xtime_lock);
102 * NOHZ - aka dynamic tick functionality
108 int tick_nohz_enabled __read_mostly = 1;
111 * Enable / Disable tickless mode
113 static int __init setup_tick_nohz(char *str)
115 if (!strcmp(str, "off"))
116 tick_nohz_enabled = 0;
117 else if (!strcmp(str, "on"))
118 tick_nohz_enabled = 1;
124 __setup("nohz=", setup_tick_nohz);
127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
129 * Called from interrupt entry when the CPU was idle
131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
133 * value. We do this unconditionally on any cpu, as we don't know whether the
134 * cpu, which has the update task assigned is in a long sleep.
136 static void tick_nohz_update_jiffies(ktime_t now)
138 int cpu = smp_processor_id();
139 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
142 ts->idle_waketime = now;
144 local_irq_save(flags);
145 tick_do_update_jiffies64(now);
146 local_irq_restore(flags);
148 touch_softlockup_watchdog();
152 * Updates the per cpu time idle statistics counters
155 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
159 if (ts->idle_active) {
160 delta = ktime_sub(now, ts->idle_entrytime);
161 if (nr_iowait_cpu(cpu) > 0)
162 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
164 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
165 ts->idle_entrytime = now;
168 if (last_update_time)
169 *last_update_time = ktime_to_us(now);
173 static void tick_nohz_stop_idle(int cpu, ktime_t now)
175 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
177 update_ts_time_stats(cpu, ts, now, NULL);
180 sched_clock_idle_wakeup_event(0);
183 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
185 ktime_t now = ktime_get();
187 ts->idle_entrytime = now;
189 sched_clock_idle_sleep_event();
194 * get_cpu_idle_time_us - get the total idle time of a cpu
195 * @cpu: CPU number to query
196 * @last_update_time: variable to store update time in. Do not update
199 * Return the cummulative idle time (since boot) for a given
200 * CPU, in microseconds.
202 * This time is measured via accounting rather than sampling,
203 * and is as accurate as ktime_get() is.
205 * This function returns -1 if NOHZ is not enabled.
207 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
209 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
212 if (!tick_nohz_enabled)
216 if (last_update_time) {
217 update_ts_time_stats(cpu, ts, now, last_update_time);
218 idle = ts->idle_sleeptime;
220 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
221 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
223 idle = ktime_add(ts->idle_sleeptime, delta);
225 idle = ts->idle_sleeptime;
229 return ktime_to_us(idle);
232 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
235 * get_cpu_iowait_time_us - get the total iowait time of a cpu
236 * @cpu: CPU number to query
237 * @last_update_time: variable to store update time in. Do not update
240 * Return the cummulative iowait time (since boot) for a given
241 * CPU, in microseconds.
243 * This time is measured via accounting rather than sampling,
244 * and is as accurate as ktime_get() is.
246 * This function returns -1 if NOHZ is not enabled.
248 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
250 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
253 if (!tick_nohz_enabled)
257 if (last_update_time) {
258 update_ts_time_stats(cpu, ts, now, last_update_time);
259 iowait = ts->iowait_sleeptime;
261 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
262 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
264 iowait = ktime_add(ts->iowait_sleeptime, delta);
266 iowait = ts->iowait_sleeptime;
270 return ktime_to_us(iowait);
272 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
274 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
275 ktime_t now, int cpu)
277 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
278 ktime_t last_update, expires, ret = { .tv64 = 0 };
279 unsigned long rcu_delta_jiffies;
280 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
283 /* Read jiffies and the time when jiffies were updated last */
285 seq = read_seqbegin(&xtime_lock);
286 last_update = last_jiffies_update;
287 last_jiffies = jiffies;
288 time_delta = timekeeping_max_deferment();
289 } while (read_seqretry(&xtime_lock, seq));
291 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) ||
292 arch_needs_cpu(cpu)) {
293 next_jiffies = last_jiffies + 1;
296 /* Get the next timer wheel timer */
297 next_jiffies = get_next_timer_interrupt(last_jiffies);
298 delta_jiffies = next_jiffies - last_jiffies;
299 if (rcu_delta_jiffies < delta_jiffies) {
300 next_jiffies = last_jiffies + rcu_delta_jiffies;
301 delta_jiffies = rcu_delta_jiffies;
305 * Do not stop the tick, if we are only one off
306 * or if the cpu is required for rcu
308 if (!ts->tick_stopped && delta_jiffies == 1)
311 /* Schedule the tick, if we are at least one jiffie off */
312 if ((long)delta_jiffies >= 1) {
315 * If this cpu is the one which updates jiffies, then
316 * give up the assignment and let it be taken by the
317 * cpu which runs the tick timer next, which might be
318 * this cpu as well. If we don't drop this here the
319 * jiffies might be stale and do_timer() never
320 * invoked. Keep track of the fact that it was the one
321 * which had the do_timer() duty last. If this cpu is
322 * the one which had the do_timer() duty last, we
323 * limit the sleep time to the timekeeping
324 * max_deferement value which we retrieved
325 * above. Otherwise we can sleep as long as we want.
327 if (cpu == tick_do_timer_cpu) {
328 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
329 ts->do_timer_last = 1;
330 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
331 time_delta = KTIME_MAX;
332 ts->do_timer_last = 0;
333 } else if (!ts->do_timer_last) {
334 time_delta = KTIME_MAX;
338 * calculate the expiry time for the next timer wheel
339 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
340 * that there is no timer pending or at least extremely
341 * far into the future (12 days for HZ=1000). In this
342 * case we set the expiry to the end of time.
344 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
346 * Calculate the time delta for the next timer event.
347 * If the time delta exceeds the maximum time delta
348 * permitted by the current clocksource then adjust
349 * the time delta accordingly to ensure the
350 * clocksource does not wrap.
352 time_delta = min_t(u64, time_delta,
353 tick_period.tv64 * delta_jiffies);
356 if (time_delta < KTIME_MAX)
357 expires = ktime_add_ns(last_update, time_delta);
359 expires.tv64 = KTIME_MAX;
361 /* Skip reprogram of event if its not changed */
362 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
368 * nohz_stop_sched_tick can be called several times before
369 * the nohz_restart_sched_tick is called. This happens when
370 * interrupts arrive which do not cause a reschedule. In the
371 * first call we save the current tick time, so we can restart
372 * the scheduler tick in nohz_restart_sched_tick.
374 if (!ts->tick_stopped) {
375 nohz_balance_enter_idle(cpu);
376 calc_load_enter_idle();
378 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
379 ts->tick_stopped = 1;
383 * If the expiration time == KTIME_MAX, then
384 * in this case we simply stop the tick timer.
386 if (unlikely(expires.tv64 == KTIME_MAX)) {
387 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
388 hrtimer_cancel(&ts->sched_timer);
392 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
393 hrtimer_start(&ts->sched_timer, expires,
394 HRTIMER_MODE_ABS_PINNED);
395 /* Check, if the timer was already in the past */
396 if (hrtimer_active(&ts->sched_timer))
398 } else if (!tick_program_event(expires, 0))
401 * We are past the event already. So we crossed a
402 * jiffie boundary. Update jiffies and raise the
405 tick_do_update_jiffies64(ktime_get());
407 raise_softirq_irqoff(TIMER_SOFTIRQ);
409 ts->next_jiffies = next_jiffies;
410 ts->last_jiffies = last_jiffies;
411 ts->sleep_length = ktime_sub(dev->next_event, now);
416 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
419 * If this cpu is offline and it is the one which updates
420 * jiffies, then give up the assignment and let it be taken by
421 * the cpu which runs the tick timer next. If we don't drop
422 * this here the jiffies might be stale and do_timer() never
425 if (unlikely(!cpu_online(cpu))) {
426 if (cpu == tick_do_timer_cpu)
427 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
430 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
436 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
437 static int ratelimit;
439 if (ratelimit < 10 &&
440 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
441 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
442 (unsigned int) local_softirq_pending());
451 static void __tick_nohz_idle_enter(struct tick_sched *ts)
453 ktime_t now, expires;
454 int cpu = smp_processor_id();
456 now = tick_nohz_start_idle(cpu, ts);
458 if (can_stop_idle_tick(cpu, ts)) {
459 int was_stopped = ts->tick_stopped;
463 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
464 if (expires.tv64 > 0LL) {
466 ts->idle_expires = expires;
469 if (!was_stopped && ts->tick_stopped)
470 ts->idle_jiffies = ts->last_jiffies;
475 * tick_nohz_idle_enter - stop the idle tick from the idle task
477 * When the next event is more than a tick into the future, stop the idle tick
478 * Called when we start the idle loop.
480 * The arch is responsible of calling:
482 * - rcu_idle_enter() after its last use of RCU before the CPU is put
484 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
486 void tick_nohz_idle_enter(void)
488 struct tick_sched *ts;
490 WARN_ON_ONCE(irqs_disabled());
493 * Update the idle state in the scheduler domain hierarchy
494 * when tick_nohz_stop_sched_tick() is called from the idle loop.
495 * State will be updated to busy during the first busy tick after
498 set_cpu_sd_state_idle();
502 ts = &__get_cpu_var(tick_cpu_sched);
504 * set ts->inidle unconditionally. even if the system did not
505 * switch to nohz mode the cpu frequency governers rely on the
506 * update of the idle time accounting in tick_nohz_start_idle().
509 __tick_nohz_idle_enter(ts);
515 * tick_nohz_irq_exit - update next tick event from interrupt exit
517 * When an interrupt fires while we are idle and it doesn't cause
518 * a reschedule, it may still add, modify or delete a timer, enqueue
519 * an RCU callback, etc...
520 * So we need to re-calculate and reprogram the next tick event.
522 void tick_nohz_irq_exit(void)
524 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
529 /* Cancel the timer because CPU already waken up from the C-states*/
530 menu_hrtimer_cancel();
531 __tick_nohz_idle_enter(ts);
535 * tick_nohz_get_sleep_length - return the length of the current sleep
537 * Called from power state control code with interrupts disabled
539 ktime_t tick_nohz_get_sleep_length(void)
541 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
543 return ts->sleep_length;
546 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
548 hrtimer_cancel(&ts->sched_timer);
549 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
552 /* Forward the time to expire in the future */
553 hrtimer_forward(&ts->sched_timer, now, tick_period);
555 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
556 hrtimer_start_expires(&ts->sched_timer,
557 HRTIMER_MODE_ABS_PINNED);
558 /* Check, if the timer was already in the past */
559 if (hrtimer_active(&ts->sched_timer))
562 if (!tick_program_event(
563 hrtimer_get_expires(&ts->sched_timer), 0))
566 /* Reread time and update jiffies */
568 tick_do_update_jiffies64(now);
572 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
574 /* Update jiffies first */
575 tick_do_update_jiffies64(now);
576 update_cpu_load_nohz();
578 calc_load_exit_idle();
579 touch_softlockup_watchdog();
581 * Cancel the scheduled timer and restore the tick
583 ts->tick_stopped = 0;
584 ts->idle_exittime = now;
586 tick_nohz_restart(ts, now);
589 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
591 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
594 * We stopped the tick in idle. Update process times would miss the
595 * time we slept as update_process_times does only a 1 tick
596 * accounting. Enforce that this is accounted to idle !
598 ticks = jiffies - ts->idle_jiffies;
600 * We might be one off. Do not randomly account a huge number of ticks!
602 if (ticks && ticks < LONG_MAX)
603 account_idle_ticks(ticks);
608 * tick_nohz_idle_exit - restart the idle tick from the idle task
610 * Restart the idle tick when the CPU is woken up from idle
611 * This also exit the RCU extended quiescent state. The CPU
612 * can use RCU again after this function is called.
614 void tick_nohz_idle_exit(void)
616 int cpu = smp_processor_id();
617 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
622 WARN_ON_ONCE(!ts->inidle);
626 /* Cancel the timer because CPU already waken up from the C-states*/
627 menu_hrtimer_cancel();
628 if (ts->idle_active || ts->tick_stopped)
632 tick_nohz_stop_idle(cpu, now);
634 if (ts->tick_stopped) {
635 tick_nohz_restart_sched_tick(ts, now);
636 tick_nohz_account_idle_ticks(ts);
642 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
644 hrtimer_forward(&ts->sched_timer, now, tick_period);
645 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
649 * The nohz low res interrupt handler
651 static void tick_nohz_handler(struct clock_event_device *dev)
653 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
654 struct pt_regs *regs = get_irq_regs();
655 int cpu = smp_processor_id();
656 ktime_t now = ktime_get();
658 dev->next_event.tv64 = KTIME_MAX;
661 * Check if the do_timer duty was dropped. We don't care about
662 * concurrency: This happens only when the cpu in charge went
663 * into a long sleep. If two cpus happen to assign themself to
664 * this duty, then the jiffies update is still serialized by
667 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
668 tick_do_timer_cpu = cpu;
670 /* Check, if the jiffies need an update */
671 if (tick_do_timer_cpu == cpu)
672 tick_do_update_jiffies64(now);
675 * When we are idle and the tick is stopped, we have to touch
676 * the watchdog as we might not schedule for a really long
677 * time. This happens on complete idle SMP systems while
678 * waiting on the login prompt. We also increment the "start
679 * of idle" jiffy stamp so the idle accounting adjustment we
680 * do when we go busy again does not account too much ticks.
682 if (ts->tick_stopped) {
683 touch_softlockup_watchdog();
687 update_process_times(user_mode(regs));
688 profile_tick(CPU_PROFILING);
690 while (tick_nohz_reprogram(ts, now)) {
692 tick_do_update_jiffies64(now);
697 * tick_nohz_switch_to_nohz - switch to nohz mode
699 static void tick_nohz_switch_to_nohz(void)
701 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
704 if (!tick_nohz_enabled)
708 if (tick_switch_to_oneshot(tick_nohz_handler)) {
713 ts->nohz_mode = NOHZ_MODE_LOWRES;
716 * Recycle the hrtimer in ts, so we can share the
717 * hrtimer_forward with the highres code.
719 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
720 /* Get the next period */
721 next = tick_init_jiffy_update();
724 hrtimer_set_expires(&ts->sched_timer, next);
725 if (!tick_program_event(next, 0))
727 next = ktime_add(next, tick_period);
733 * When NOHZ is enabled and the tick is stopped, we need to kick the
734 * tick timer from irq_enter() so that the jiffies update is kept
735 * alive during long running softirqs. That's ugly as hell, but
736 * correctness is key even if we need to fix the offending softirq in
739 * Note, this is different to tick_nohz_restart. We just kick the
740 * timer and do not touch the other magic bits which need to be done
743 static void tick_nohz_kick_tick(int cpu, ktime_t now)
746 /* Switch back to 2.6.27 behaviour */
748 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
752 * Do not touch the tick device, when the next expiry is either
753 * already reached or less/equal than the tick period.
755 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
756 if (delta.tv64 <= tick_period.tv64)
759 tick_nohz_restart(ts, now);
763 static inline void tick_check_nohz(int cpu)
765 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
768 if (!ts->idle_active && !ts->tick_stopped)
772 tick_nohz_stop_idle(cpu, now);
773 if (ts->tick_stopped) {
774 tick_nohz_update_jiffies(now);
775 tick_nohz_kick_tick(cpu, now);
781 static inline void tick_nohz_switch_to_nohz(void) { }
782 static inline void tick_check_nohz(int cpu) { }
787 * Called from irq_enter to notify about the possible interruption of idle()
789 void tick_check_idle(int cpu)
791 tick_check_oneshot_broadcast(cpu);
792 tick_check_nohz(cpu);
796 * High resolution timer specific code
798 #ifdef CONFIG_HIGH_RES_TIMERS
800 * We rearm the timer until we get disabled by the idle code.
801 * Called with interrupts disabled.
803 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
805 struct tick_sched *ts =
806 container_of(timer, struct tick_sched, sched_timer);
807 struct pt_regs *regs = get_irq_regs();
808 ktime_t now = ktime_get();
809 int cpu = smp_processor_id();
813 * Check if the do_timer duty was dropped. We don't care about
814 * concurrency: This happens only when the cpu in charge went
815 * into a long sleep. If two cpus happen to assign themself to
816 * this duty, then the jiffies update is still serialized by
819 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
820 tick_do_timer_cpu = cpu;
823 /* Check, if the jiffies need an update */
824 if (tick_do_timer_cpu == cpu)
825 tick_do_update_jiffies64(now);
828 * Do not call, when we are not in irq context and have
829 * no valid regs pointer
833 * When we are idle and the tick is stopped, we have to touch
834 * the watchdog as we might not schedule for a really long
835 * time. This happens on complete idle SMP systems while
836 * waiting on the login prompt. We also increment the "start of
837 * idle" jiffy stamp so the idle accounting adjustment we do
838 * when we go busy again does not account too much ticks.
840 if (ts->tick_stopped) {
841 touch_softlockup_watchdog();
842 if (is_idle_task(current))
845 update_process_times(user_mode(regs));
846 profile_tick(CPU_PROFILING);
849 hrtimer_forward(timer, now, tick_period);
851 return HRTIMER_RESTART;
854 static int sched_skew_tick;
856 static int __init skew_tick(char *str)
858 get_option(&str, &sched_skew_tick);
862 early_param("skew_tick", skew_tick);
865 * tick_setup_sched_timer - setup the tick emulation timer
867 void tick_setup_sched_timer(void)
869 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
870 ktime_t now = ktime_get();
873 * Emulate tick processing via per-CPU hrtimers:
875 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
876 ts->sched_timer.function = tick_sched_timer;
878 /* Get the next period (per cpu) */
879 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
881 /* Offset the tick to avert xtime_lock contention. */
882 if (sched_skew_tick) {
883 u64 offset = ktime_to_ns(tick_period) >> 1;
884 do_div(offset, num_possible_cpus());
885 offset *= smp_processor_id();
886 hrtimer_add_expires_ns(&ts->sched_timer, offset);
890 hrtimer_forward(&ts->sched_timer, now, tick_period);
891 hrtimer_start_expires(&ts->sched_timer,
892 HRTIMER_MODE_ABS_PINNED);
893 /* Check, if the timer was already in the past */
894 if (hrtimer_active(&ts->sched_timer))
900 if (tick_nohz_enabled)
901 ts->nohz_mode = NOHZ_MODE_HIGHRES;
904 #endif /* HIGH_RES_TIMERS */
906 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
907 void tick_cancel_sched_timer(int cpu)
909 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
911 # ifdef CONFIG_HIGH_RES_TIMERS
912 if (ts->sched_timer.base)
913 hrtimer_cancel(&ts->sched_timer);
916 ts->nohz_mode = NOHZ_MODE_INACTIVE;
921 * Async notification about clocksource changes
923 void tick_clock_notify(void)
927 for_each_possible_cpu(cpu)
928 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
932 * Async notification about clock event changes
934 void tick_oneshot_notify(void)
936 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
938 set_bit(0, &ts->check_clocks);
942 * Check, if a change happened, which makes oneshot possible.
944 * Called cyclic from the hrtimer softirq (driven by the timer
945 * softirq) allow_nohz signals, that we can switch into low-res nohz
946 * mode, because high resolution timers are disabled (either compile
949 int tick_check_oneshot_change(int allow_nohz)
951 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
953 if (!test_and_clear_bit(0, &ts->check_clocks))
956 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
959 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
965 tick_nohz_switch_to_nohz();