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>
23 #include <linux/irq_work.h>
24 #include <linux/posix-timers.h>
25 #include <linux/perf_event.h>
27 #include <asm/irq_regs.h>
29 #include "tick-internal.h"
31 #include <trace/events/timer.h>
33 #if defined(CONFIG_SYSTEM_LOAD_ANALYZER)
34 #include <linux/load_analyzer.h>
37 * Per cpu nohz control structure
39 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
42 * The time, when the last jiffy update happened. Protected by jiffies_lock.
44 static ktime_t last_jiffies_update;
46 struct tick_sched *tick_get_tick_sched(int cpu)
48 return &per_cpu(tick_cpu_sched, cpu);
52 * Must be called with interrupts disabled !
54 static void tick_do_update_jiffies64(ktime_t now)
56 unsigned long ticks = 0;
60 * Do a quick check without holding jiffies_lock:
62 delta = ktime_sub(now, last_jiffies_update);
63 if (delta.tv64 < tick_period.tv64)
66 /* Reevalute with jiffies_lock held */
67 write_seqlock(&jiffies_lock);
69 delta = ktime_sub(now, last_jiffies_update);
70 if (delta.tv64 >= tick_period.tv64) {
72 delta = ktime_sub(delta, tick_period);
73 last_jiffies_update = ktime_add(last_jiffies_update,
76 /* Slow path for long timeouts */
77 if (unlikely(delta.tv64 >= tick_period.tv64)) {
78 s64 incr = ktime_to_ns(tick_period);
80 ticks = ktime_divns(delta, incr);
82 last_jiffies_update = ktime_add_ns(last_jiffies_update,
87 /* Keep the tick_next_period variable up to date */
88 tick_next_period = ktime_add(last_jiffies_update, tick_period);
90 write_sequnlock(&jiffies_lock);
94 * Initialize and return retrieve the jiffies update.
96 static ktime_t tick_init_jiffy_update(void)
100 write_seqlock(&jiffies_lock);
101 /* Did we start the jiffies update yet ? */
102 if (last_jiffies_update.tv64 == 0)
103 last_jiffies_update = tick_next_period;
104 period = last_jiffies_update;
105 write_sequnlock(&jiffies_lock);
110 static void tick_sched_do_timer(ktime_t now)
112 int cpu = smp_processor_id();
114 #ifdef CONFIG_NO_HZ_COMMON
116 * Check if the do_timer duty was dropped. We don't care about
117 * concurrency: This happens only when the cpu in charge went
118 * into a long sleep. If two cpus happen to assign themself to
119 * this duty, then the jiffies update is still serialized by
122 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
123 && !tick_nohz_full_cpu(cpu))
124 tick_do_timer_cpu = cpu;
127 /* Check, if the jiffies need an update */
128 if (tick_do_timer_cpu == cpu)
129 tick_do_update_jiffies64(now);
132 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
134 #ifdef CONFIG_NO_HZ_COMMON
136 * When we are idle and the tick is stopped, we have to touch
137 * the watchdog as we might not schedule for a really long
138 * time. This happens on complete idle SMP systems while
139 * waiting on the login prompt. We also increment the "start of
140 * idle" jiffy stamp so the idle accounting adjustment we do
141 * when we go busy again does not account too much ticks.
143 if (ts->tick_stopped) {
144 touch_softlockup_watchdog();
145 if (is_idle_task(current))
149 update_process_times(user_mode(regs));
150 profile_tick(CPU_PROFILING);
153 #ifdef CONFIG_NO_HZ_FULL
154 static cpumask_var_t nohz_full_mask;
155 bool have_nohz_full_mask;
157 static bool can_stop_full_tick(void)
159 WARN_ON_ONCE(!irqs_disabled());
161 if (!sched_can_stop_tick()) {
162 trace_tick_stop(0, "more than 1 task in runqueue\n");
166 if (!posix_cpu_timers_can_stop_tick(current)) {
167 trace_tick_stop(0, "posix timers running\n");
171 if (!perf_event_can_stop_tick()) {
172 trace_tick_stop(0, "perf events running\n");
176 /* sched_clock_tick() needs us? */
177 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
179 * TODO: kick full dynticks CPUs when
180 * sched_clock_stable is set.
182 if (!sched_clock_stable) {
183 trace_tick_stop(0, "unstable sched clock\n");
191 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
194 * Re-evaluate the need for the tick on the current CPU
195 * and restart it if necessary.
197 void tick_nohz_full_check(void)
199 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
201 if (tick_nohz_full_cpu(smp_processor_id())) {
202 if (ts->tick_stopped && !is_idle_task(current)) {
203 if (!can_stop_full_tick())
204 tick_nohz_restart_sched_tick(ts, ktime_get());
209 static void nohz_full_kick_work_func(struct irq_work *work)
211 tick_nohz_full_check();
214 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
215 .func = nohz_full_kick_work_func,
219 * Kick the current CPU if it's full dynticks in order to force it to
220 * re-evaluate its dependency on the tick and restart it if necessary.
222 void tick_nohz_full_kick(void)
224 if (tick_nohz_full_cpu(smp_processor_id()))
225 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
228 static void nohz_full_kick_ipi(void *info)
230 tick_nohz_full_check();
234 * Kick all full dynticks CPUs in order to force these to re-evaluate
235 * their dependency on the tick and restart it if necessary.
237 void tick_nohz_full_kick_all(void)
239 if (!have_nohz_full_mask)
243 smp_call_function_many(nohz_full_mask,
244 nohz_full_kick_ipi, NULL, false);
249 * Re-evaluate the need for the tick as we switch the current task.
250 * It might need the tick due to per task/process properties:
251 * perf events, posix cpu timers, ...
253 void tick_nohz_task_switch(struct task_struct *tsk)
257 local_irq_save(flags);
259 if (!tick_nohz_full_cpu(smp_processor_id()))
262 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
263 tick_nohz_full_kick();
266 local_irq_restore(flags);
269 int tick_nohz_full_cpu(int cpu)
271 if (!have_nohz_full_mask)
274 return cpumask_test_cpu(cpu, nohz_full_mask);
277 /* Parse the boot-time nohz CPU list from the kernel parameters. */
278 static int __init tick_nohz_full_setup(char *str)
282 alloc_bootmem_cpumask_var(&nohz_full_mask);
283 if (cpulist_parse(str, nohz_full_mask) < 0) {
284 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
288 cpu = smp_processor_id();
289 if (cpumask_test_cpu(cpu, nohz_full_mask)) {
290 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
291 cpumask_clear_cpu(cpu, nohz_full_mask);
293 have_nohz_full_mask = true;
297 __setup("nohz_full=", tick_nohz_full_setup);
299 static int __cpuinit tick_nohz_cpu_down_callback(struct notifier_block *nfb,
300 unsigned long action,
303 unsigned int cpu = (unsigned long)hcpu;
305 switch (action & ~CPU_TASKS_FROZEN) {
306 case CPU_DOWN_PREPARE:
308 * If we handle the timekeeping duty for full dynticks CPUs,
309 * we can't safely shutdown that CPU.
311 if (have_nohz_full_mask && tick_do_timer_cpu == cpu)
319 * Worst case string length in chunks of CPU range seems 2 steps
320 * separations: 0,2,4,6,...
321 * This is NR_CPUS + sizeof('\0')
323 static char __initdata nohz_full_buf[NR_CPUS + 1];
325 static int tick_nohz_init_all(void)
329 #ifdef CONFIG_NO_HZ_FULL_ALL
330 if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) {
331 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
335 cpumask_setall(nohz_full_mask);
336 cpumask_clear_cpu(smp_processor_id(), nohz_full_mask);
337 have_nohz_full_mask = true;
342 void __init tick_nohz_init(void)
346 if (!have_nohz_full_mask) {
347 if (tick_nohz_init_all() < 0)
351 cpu_notifier(tick_nohz_cpu_down_callback, 0);
353 /* Make sure full dynticks CPU are also RCU nocbs */
354 for_each_cpu(cpu, nohz_full_mask) {
355 if (!rcu_is_nocb_cpu(cpu)) {
356 pr_warning("NO_HZ: CPU %d is not RCU nocb: "
357 "cleared from nohz_full range", cpu);
358 cpumask_clear_cpu(cpu, nohz_full_mask);
362 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask);
363 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
366 #define have_nohz_full_mask (0)
370 * NOHZ - aka dynamic tick functionality
372 #ifdef CONFIG_NO_HZ_COMMON
376 int tick_nohz_enabled __read_mostly = 1;
379 * Enable / Disable tickless mode
381 static int __init setup_tick_nohz(char *str)
383 if (!strcmp(str, "off"))
384 tick_nohz_enabled = 0;
385 else if (!strcmp(str, "on"))
386 tick_nohz_enabled = 1;
392 __setup("nohz=", setup_tick_nohz);
395 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
397 * Called from interrupt entry when the CPU was idle
399 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
400 * must be updated. Otherwise an interrupt handler could use a stale jiffy
401 * value. We do this unconditionally on any cpu, as we don't know whether the
402 * cpu, which has the update task assigned is in a long sleep.
404 static void tick_nohz_update_jiffies(ktime_t now)
406 int cpu = smp_processor_id();
407 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
410 ts->idle_waketime = now;
412 local_irq_save(flags);
413 tick_do_update_jiffies64(now);
414 local_irq_restore(flags);
416 touch_softlockup_watchdog();
420 * Updates the per cpu time idle statistics counters
423 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
427 if (ts->idle_active) {
428 delta = ktime_sub(now, ts->idle_entrytime);
429 if (nr_iowait_cpu(cpu) > 0)
430 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
432 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
433 ts->idle_entrytime = now;
436 if (last_update_time)
437 *last_update_time = ktime_to_us(now);
441 static void tick_nohz_stop_idle(int cpu, ktime_t now)
443 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
445 update_ts_time_stats(cpu, ts, now, NULL);
448 sched_clock_idle_wakeup_event(0);
451 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
453 ktime_t now = ktime_get();
455 ts->idle_entrytime = now;
457 sched_clock_idle_sleep_event();
462 * get_cpu_idle_time_us - get the total idle time of a cpu
463 * @cpu: CPU number to query
464 * @last_update_time: variable to store update time in. Do not update
467 * Return the cummulative idle time (since boot) for a given
468 * CPU, in microseconds.
470 * This time is measured via accounting rather than sampling,
471 * and is as accurate as ktime_get() is.
473 * This function returns -1 if NOHZ is not enabled.
475 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
477 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
480 if (!tick_nohz_enabled)
484 if (last_update_time) {
485 update_ts_time_stats(cpu, ts, now, last_update_time);
486 idle = ts->idle_sleeptime;
488 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
489 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
491 idle = ktime_add(ts->idle_sleeptime, delta);
493 idle = ts->idle_sleeptime;
497 return ktime_to_us(idle);
500 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
503 * get_cpu_iowait_time_us - get the total iowait time of a cpu
504 * @cpu: CPU number to query
505 * @last_update_time: variable to store update time in. Do not update
508 * Return the cummulative iowait time (since boot) for a given
509 * CPU, in microseconds.
511 * This time is measured via accounting rather than sampling,
512 * and is as accurate as ktime_get() is.
514 * This function returns -1 if NOHZ is not enabled.
516 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
518 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
521 if (!tick_nohz_enabled)
525 if (last_update_time) {
526 update_ts_time_stats(cpu, ts, now, last_update_time);
527 iowait = ts->iowait_sleeptime;
529 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
530 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
532 iowait = ktime_add(ts->iowait_sleeptime, delta);
534 iowait = ts->iowait_sleeptime;
538 return ktime_to_us(iowait);
540 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
542 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
543 ktime_t now, int cpu)
545 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
546 ktime_t last_update, expires, ret = { .tv64 = 0 };
547 unsigned long rcu_delta_jiffies;
548 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
551 /* Read jiffies and the time when jiffies were updated last */
553 seq = read_seqbegin(&jiffies_lock);
554 last_update = last_jiffies_update;
555 last_jiffies = jiffies;
556 time_delta = timekeeping_max_deferment();
557 } while (read_seqretry(&jiffies_lock, seq));
559 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
560 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
561 next_jiffies = last_jiffies + 1;
564 /* Get the next timer wheel timer */
565 next_jiffies = get_next_timer_interrupt(last_jiffies);
566 delta_jiffies = next_jiffies - last_jiffies;
567 if (rcu_delta_jiffies < delta_jiffies) {
568 next_jiffies = last_jiffies + rcu_delta_jiffies;
569 delta_jiffies = rcu_delta_jiffies;
574 * Do not stop the tick, if we are only one off (or less)
575 * or if the cpu is required for RCU:
577 if (!ts->tick_stopped && delta_jiffies <= 1)
580 /* Schedule the tick, if we are at least one jiffie off */
581 if ((long)delta_jiffies >= 1) {
584 * If this cpu is the one which updates jiffies, then
585 * give up the assignment and let it be taken by the
586 * cpu which runs the tick timer next, which might be
587 * this cpu as well. If we don't drop this here the
588 * jiffies might be stale and do_timer() never
589 * invoked. Keep track of the fact that it was the one
590 * which had the do_timer() duty last. If this cpu is
591 * the one which had the do_timer() duty last, we
592 * limit the sleep time to the timekeeping
593 * max_deferement value which we retrieved
594 * above. Otherwise we can sleep as long as we want.
596 if (cpu == tick_do_timer_cpu) {
597 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
598 ts->do_timer_last = 1;
599 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
600 time_delta = KTIME_MAX;
601 ts->do_timer_last = 0;
602 } else if (!ts->do_timer_last) {
603 time_delta = KTIME_MAX;
606 #ifdef CONFIG_NO_HZ_FULL
608 time_delta = min(time_delta,
609 scheduler_tick_max_deferment());
614 * calculate the expiry time for the next timer wheel
615 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
616 * that there is no timer pending or at least extremely
617 * far into the future (12 days for HZ=1000). In this
618 * case we set the expiry to the end of time.
620 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
622 * Calculate the time delta for the next timer event.
623 * If the time delta exceeds the maximum time delta
624 * permitted by the current clocksource then adjust
625 * the time delta accordingly to ensure the
626 * clocksource does not wrap.
628 time_delta = min_t(u64, time_delta,
629 tick_period.tv64 * delta_jiffies);
632 if (time_delta < KTIME_MAX)
633 expires = ktime_add_ns(last_update, time_delta);
635 expires.tv64 = KTIME_MAX;
637 /* Skip reprogram of event if its not changed */
638 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
644 * nohz_stop_sched_tick can be called several times before
645 * the nohz_restart_sched_tick is called. This happens when
646 * interrupts arrive which do not cause a reschedule. In the
647 * first call we save the current tick time, so we can restart
648 * the scheduler tick in nohz_restart_sched_tick.
650 if (!ts->tick_stopped) {
651 nohz_balance_enter_idle(cpu);
652 calc_load_enter_idle();
654 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
655 ts->tick_stopped = 1;
656 trace_tick_stop(1, " ");
660 * If the expiration time == KTIME_MAX, then
661 * in this case we simply stop the tick timer.
663 if (unlikely(expires.tv64 == KTIME_MAX)) {
664 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
665 hrtimer_cancel(&ts->sched_timer);
669 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
670 hrtimer_start(&ts->sched_timer, expires,
671 HRTIMER_MODE_ABS_PINNED);
672 /* Check, if the timer was already in the past */
673 if (hrtimer_active(&ts->sched_timer))
675 } else if (!tick_program_event(expires, 0))
678 * We are past the event already. So we crossed a
679 * jiffie boundary. Update jiffies and raise the
682 tick_do_update_jiffies64(ktime_get());
684 raise_softirq_irqoff(TIMER_SOFTIRQ);
686 ts->next_jiffies = next_jiffies;
687 ts->last_jiffies = last_jiffies;
688 ts->sleep_length = ktime_sub(dev->next_event, now);
693 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
695 #ifdef CONFIG_NO_HZ_FULL
696 int cpu = smp_processor_id();
698 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
701 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
704 if (!can_stop_full_tick())
707 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
711 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
714 * If this cpu is offline and it is the one which updates
715 * jiffies, then give up the assignment and let it be taken by
716 * the cpu which runs the tick timer next. If we don't drop
717 * this here the jiffies might be stale and do_timer() never
720 if (unlikely(!cpu_online(cpu))) {
721 if (cpu == tick_do_timer_cpu)
722 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
726 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
727 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
734 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
735 static int ratelimit;
737 if (ratelimit < 10 &&
738 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
739 pr_warn("NOHZ: local_softirq_pending %02x\n",
740 (unsigned int) local_softirq_pending());
746 if (have_nohz_full_mask) {
748 * Keep the tick alive to guarantee timekeeping progression
749 * if there are full dynticks CPUs around
751 if (tick_do_timer_cpu == cpu)
754 * Boot safety: make sure the timekeeping duty has been
755 * assigned before entering dyntick-idle mode,
757 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
764 static void __tick_nohz_idle_enter(struct tick_sched *ts)
766 ktime_t now, expires;
767 int cpu = smp_processor_id();
769 now = tick_nohz_start_idle(cpu, ts);
771 if (can_stop_idle_tick(cpu, ts)) {
772 int was_stopped = ts->tick_stopped;
776 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
777 if (expires.tv64 > 0LL) {
779 ts->idle_expires = expires;
782 if (!was_stopped && ts->tick_stopped)
783 ts->idle_jiffies = ts->last_jiffies;
788 * tick_nohz_idle_enter - stop the idle tick from the idle task
790 * When the next event is more than a tick into the future, stop the idle tick
791 * Called when we start the idle loop.
793 * The arch is responsible of calling:
795 * - rcu_idle_enter() after its last use of RCU before the CPU is put
797 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
799 void tick_nohz_idle_enter(void)
801 struct tick_sched *ts;
803 WARN_ON_ONCE(irqs_disabled());
806 * Update the idle state in the scheduler domain hierarchy
807 * when tick_nohz_stop_sched_tick() is called from the idle loop.
808 * State will be updated to busy during the first busy tick after
811 set_cpu_sd_state_idle();
815 ts = &__get_cpu_var(tick_cpu_sched);
817 * set ts->inidle unconditionally. even if the system did not
818 * switch to nohz mode the cpu frequency governers rely on the
819 * update of the idle time accounting in tick_nohz_start_idle().
822 __tick_nohz_idle_enter(ts);
826 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
829 * tick_nohz_irq_exit - update next tick event from interrupt exit
831 * When an interrupt fires while we are idle and it doesn't cause
832 * a reschedule, it may still add, modify or delete a timer, enqueue
833 * an RCU callback, etc...
834 * So we need to re-calculate and reprogram the next tick event.
836 void tick_nohz_irq_exit(void)
838 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
841 __tick_nohz_idle_enter(ts);
843 tick_nohz_full_stop_tick(ts);
847 * tick_nohz_get_sleep_length - return the length of the current sleep
849 * Called from power state control code with interrupts disabled
851 ktime_t tick_nohz_get_sleep_length(void)
853 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
855 return ts->sleep_length;
858 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
860 hrtimer_cancel(&ts->sched_timer);
861 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
864 /* Forward the time to expire in the future */
865 hrtimer_forward(&ts->sched_timer, now, tick_period);
867 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
868 hrtimer_start_expires(&ts->sched_timer,
869 HRTIMER_MODE_ABS_PINNED);
870 /* Check, if the timer was already in the past */
871 if (hrtimer_active(&ts->sched_timer))
874 if (!tick_program_event(
875 hrtimer_get_expires(&ts->sched_timer), 0))
878 /* Reread time and update jiffies */
880 tick_do_update_jiffies64(now);
884 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
886 /* Update jiffies first */
887 tick_do_update_jiffies64(now);
888 update_cpu_load_nohz();
890 calc_load_exit_idle();
891 touch_softlockup_watchdog();
893 * Cancel the scheduled timer and restore the tick
895 ts->tick_stopped = 0;
896 ts->idle_exittime = now;
898 tick_nohz_restart(ts, now);
901 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
903 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
906 if (vtime_accounting_enabled())
909 * We stopped the tick in idle. Update process times would miss the
910 * time we slept as update_process_times does only a 1 tick
911 * accounting. Enforce that this is accounted to idle !
913 ticks = jiffies - ts->idle_jiffies;
915 * We might be one off. Do not randomly account a huge number of ticks!
917 if (ticks && ticks < LONG_MAX)
918 account_idle_ticks(ticks);
923 * tick_nohz_idle_exit - restart the idle tick from the idle task
925 * Restart the idle tick when the CPU is woken up from idle
926 * This also exit the RCU extended quiescent state. The CPU
927 * can use RCU again after this function is called.
929 void tick_nohz_idle_exit(void)
931 int cpu = smp_processor_id();
932 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
937 WARN_ON_ONCE(!ts->inidle);
941 if (ts->idle_active || ts->tick_stopped)
945 tick_nohz_stop_idle(cpu, now);
947 if (ts->tick_stopped) {
948 tick_nohz_restart_sched_tick(ts, now);
949 tick_nohz_account_idle_ticks(ts);
954 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
956 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
958 hrtimer_forward(&ts->sched_timer, now, tick_period);
959 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
963 * The nohz low res interrupt handler
965 static void tick_nohz_handler(struct clock_event_device *dev)
967 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
968 struct pt_regs *regs = get_irq_regs();
969 ktime_t now = ktime_get();
971 dev->next_event.tv64 = KTIME_MAX;
973 tick_sched_do_timer(now);
974 tick_sched_handle(ts, regs);
976 while (tick_nohz_reprogram(ts, now)) {
978 tick_do_update_jiffies64(now);
983 * tick_nohz_switch_to_nohz - switch to nohz mode
985 static void tick_nohz_switch_to_nohz(void)
987 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
990 if (!tick_nohz_enabled)
994 if (tick_switch_to_oneshot(tick_nohz_handler)) {
999 ts->nohz_mode = NOHZ_MODE_LOWRES;
1002 * Recycle the hrtimer in ts, so we can share the
1003 * hrtimer_forward with the highres code.
1005 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1006 /* Get the next period */
1007 next = tick_init_jiffy_update();
1010 hrtimer_set_expires(&ts->sched_timer, next);
1011 if (!tick_program_event(next, 0))
1013 next = ktime_add(next, tick_period);
1019 * When NOHZ is enabled and the tick is stopped, we need to kick the
1020 * tick timer from irq_enter() so that the jiffies update is kept
1021 * alive during long running softirqs. That's ugly as hell, but
1022 * correctness is key even if we need to fix the offending softirq in
1025 * Note, this is different to tick_nohz_restart. We just kick the
1026 * timer and do not touch the other magic bits which need to be done
1027 * when idle is left.
1029 static void tick_nohz_kick_tick(int cpu, ktime_t now)
1032 /* Switch back to 2.6.27 behaviour */
1034 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1038 * Do not touch the tick device, when the next expiry is either
1039 * already reached or less/equal than the tick period.
1041 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1042 if (delta.tv64 <= tick_period.tv64)
1045 tick_nohz_restart(ts, now);
1049 static inline void tick_check_nohz(int cpu)
1051 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1054 if (!ts->idle_active && !ts->tick_stopped)
1057 if (ts->idle_active)
1058 tick_nohz_stop_idle(cpu, now);
1059 if (ts->tick_stopped) {
1060 tick_nohz_update_jiffies(now);
1061 tick_nohz_kick_tick(cpu, now);
1067 static inline void tick_nohz_switch_to_nohz(void) { }
1068 static inline void tick_check_nohz(int cpu) { }
1070 #endif /* CONFIG_NO_HZ_COMMON */
1073 * Called from irq_enter to notify about the possible interruption of idle()
1075 void tick_check_idle(int cpu)
1077 tick_check_oneshot_broadcast(cpu);
1078 tick_check_nohz(cpu);
1082 * High resolution timer specific code
1084 #ifdef CONFIG_HIGH_RES_TIMERS
1086 * We rearm the timer until we get disabled by the idle code.
1087 * Called with interrupts disabled.
1089 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1091 struct tick_sched *ts =
1092 container_of(timer, struct tick_sched, sched_timer);
1093 struct pt_regs *regs = get_irq_regs();
1094 ktime_t now = ktime_get();
1096 tick_sched_do_timer(now);
1099 * Do not call, when we are not in irq context and have
1100 * no valid regs pointer
1103 tick_sched_handle(ts, regs);
1105 #if defined(CONFIG_SYSTEM_LOAD_ANALYZER)
1106 store_external_load_factor(NR_RUNNING_TASK,nr_running());
1109 hrtimer_forward(timer, now, tick_period);
1111 return HRTIMER_RESTART;
1114 static int sched_skew_tick;
1116 static int __init skew_tick(char *str)
1118 get_option(&str, &sched_skew_tick);
1122 early_param("skew_tick", skew_tick);
1125 * tick_setup_sched_timer - setup the tick emulation timer
1127 void tick_setup_sched_timer(void)
1129 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1130 ktime_t now = ktime_get();
1133 * Emulate tick processing via per-CPU hrtimers:
1135 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1136 ts->sched_timer.function = tick_sched_timer;
1138 /* Get the next period (per cpu) */
1139 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1141 /* Offset the tick to avert jiffies_lock contention. */
1142 if (sched_skew_tick) {
1143 u64 offset = ktime_to_ns(tick_period) >> 1;
1144 do_div(offset, num_possible_cpus());
1145 offset *= smp_processor_id();
1146 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1150 hrtimer_forward(&ts->sched_timer, now, tick_period);
1151 hrtimer_start_expires(&ts->sched_timer,
1152 HRTIMER_MODE_ABS_PINNED);
1153 /* Check, if the timer was already in the past */
1154 if (hrtimer_active(&ts->sched_timer))
1159 #ifdef CONFIG_NO_HZ_COMMON
1160 if (tick_nohz_enabled)
1161 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1164 #endif /* HIGH_RES_TIMERS */
1166 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1167 void tick_cancel_sched_timer(int cpu)
1169 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1171 # ifdef CONFIG_HIGH_RES_TIMERS
1172 if (ts->sched_timer.base)
1173 hrtimer_cancel(&ts->sched_timer);
1176 memset(ts, 0, sizeof(*ts));
1181 * Async notification about clocksource changes
1183 void tick_clock_notify(void)
1187 for_each_possible_cpu(cpu)
1188 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1192 * Async notification about clock event changes
1194 void tick_oneshot_notify(void)
1196 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1198 set_bit(0, &ts->check_clocks);
1202 * Check, if a change happened, which makes oneshot possible.
1204 * Called cyclic from the hrtimer softirq (driven by the timer
1205 * softirq) allow_nohz signals, that we can switch into low-res nohz
1206 * mode, because high resolution timers are disabled (either compile
1209 int tick_check_oneshot_change(int allow_nohz)
1211 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1213 if (!test_and_clear_bit(0, &ts->check_clocks))
1216 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1219 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1225 tick_nohz_switch_to_nohz();