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>
26 #include <linux/context_tracking.h>
28 #include <asm/irq_regs.h>
30 #include "tick-internal.h"
32 #include <trace/events/timer.h>
35 * Per cpu nohz control structure
37 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 * The time, when the last jiffy update happened. Protected by jiffies_lock.
42 static ktime_t last_jiffies_update;
44 struct tick_sched *tick_get_tick_sched(int cpu)
46 return &per_cpu(tick_cpu_sched, cpu);
50 * Must be called with interrupts disabled !
52 static void tick_do_update_jiffies64(ktime_t now)
54 unsigned long ticks = 0;
58 * Do a quick check without holding jiffies_lock:
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta.tv64 < tick_period.tv64)
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
74 /* Slow path for long timeouts */
75 if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 s64 incr = ktime_to_ns(tick_period);
78 ticks = ktime_divns(delta, incr);
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
88 write_sequnlock(&jiffies_lock);
91 write_sequnlock(&jiffies_lock);
96 * Initialize and return retrieve the jiffies update.
98 static ktime_t tick_init_jiffy_update(void)
102 write_seqlock(&jiffies_lock);
103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update.tv64 == 0)
105 last_jiffies_update = tick_next_period;
106 period = last_jiffies_update;
107 write_sequnlock(&jiffies_lock);
112 static void tick_sched_do_timer(ktime_t now)
114 int cpu = smp_processor_id();
116 #ifdef CONFIG_NO_HZ_COMMON
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the cpu in charge went
120 * into a long sleep. If two cpus happen to assign themself to
121 * this duty, then the jiffies update is still serialized by
124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
125 && !tick_nohz_full_cpu(cpu))
126 tick_do_timer_cpu = cpu;
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu == cpu)
131 tick_do_update_jiffies64(now);
134 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
136 #ifdef CONFIG_NO_HZ_COMMON
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
145 if (ts->tick_stopped) {
146 touch_softlockup_watchdog();
147 if (is_idle_task(current))
151 update_process_times(user_mode(regs));
152 profile_tick(CPU_PROFILING);
155 #ifdef CONFIG_NO_HZ_FULL
156 cpumask_var_t tick_nohz_full_mask;
157 cpumask_var_t housekeeping_mask;
158 bool tick_nohz_full_running;
160 static bool can_stop_full_tick(void)
162 WARN_ON_ONCE(!irqs_disabled());
164 if (!sched_can_stop_tick()) {
165 trace_tick_stop(0, "more than 1 task in runqueue\n");
169 if (!posix_cpu_timers_can_stop_tick(current)) {
170 trace_tick_stop(0, "posix timers running\n");
174 if (!perf_event_can_stop_tick()) {
175 trace_tick_stop(0, "perf events running\n");
179 /* sched_clock_tick() needs us? */
180 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
182 * TODO: kick full dynticks CPUs when
183 * sched_clock_stable is set.
185 if (!sched_clock_stable()) {
186 trace_tick_stop(0, "unstable sched clock\n");
188 * Don't allow the user to think they can get
189 * full NO_HZ with this machine.
191 WARN_ONCE(tick_nohz_full_running,
192 "NO_HZ FULL will not work with unstable sched clock");
200 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
203 * Re-evaluate the need for the tick on the current CPU
204 * and restart it if necessary.
206 void __tick_nohz_full_check(void)
208 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
210 if (tick_nohz_full_cpu(smp_processor_id())) {
211 if (ts->tick_stopped && !is_idle_task(current)) {
212 if (!can_stop_full_tick())
213 tick_nohz_restart_sched_tick(ts, ktime_get());
218 static void nohz_full_kick_work_func(struct irq_work *work)
220 __tick_nohz_full_check();
223 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
224 .func = nohz_full_kick_work_func,
228 * Kick the current CPU if it's full dynticks in order to force it to
229 * re-evaluate its dependency on the tick and restart it if necessary.
231 void tick_nohz_full_kick(void)
233 if (tick_nohz_full_cpu(smp_processor_id()))
234 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
237 static void nohz_full_kick_ipi(void *info)
239 __tick_nohz_full_check();
243 * Kick all full dynticks CPUs in order to force these to re-evaluate
244 * their dependency on the tick and restart it if necessary.
246 void tick_nohz_full_kick_all(void)
248 if (!tick_nohz_full_running)
252 smp_call_function_many(tick_nohz_full_mask,
253 nohz_full_kick_ipi, NULL, false);
254 tick_nohz_full_kick();
259 * Re-evaluate the need for the tick as we switch the current task.
260 * It might need the tick due to per task/process properties:
261 * perf events, posix cpu timers, ...
263 void __tick_nohz_task_switch(struct task_struct *tsk)
267 local_irq_save(flags);
269 if (!tick_nohz_full_cpu(smp_processor_id()))
272 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
273 tick_nohz_full_kick();
276 local_irq_restore(flags);
279 /* Parse the boot-time nohz CPU list from the kernel parameters. */
280 static int __init tick_nohz_full_setup(char *str)
284 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
285 alloc_bootmem_cpumask_var(&housekeeping_mask);
286 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
287 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
291 cpu = smp_processor_id();
292 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
293 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
294 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
296 cpumask_andnot(housekeeping_mask,
297 cpu_possible_mask, tick_nohz_full_mask);
298 tick_nohz_full_running = true;
302 __setup("nohz_full=", tick_nohz_full_setup);
304 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
305 unsigned long action,
308 unsigned int cpu = (unsigned long)hcpu;
310 switch (action & ~CPU_TASKS_FROZEN) {
311 case CPU_DOWN_PREPARE:
313 * If we handle the timekeeping duty for full dynticks CPUs,
314 * we can't safely shutdown that CPU.
316 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
324 * Worst case string length in chunks of CPU range seems 2 steps
325 * separations: 0,2,4,6,...
326 * This is NR_CPUS + sizeof('\0')
328 static char __initdata nohz_full_buf[NR_CPUS + 1];
330 static int tick_nohz_init_all(void)
334 #ifdef CONFIG_NO_HZ_FULL_ALL
335 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
336 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
339 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
340 pr_err("NO_HZ: Can't allocate not-full dynticks cpumask\n");
344 cpumask_setall(tick_nohz_full_mask);
345 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask);
346 cpumask_clear(housekeeping_mask);
347 cpumask_set_cpu(smp_processor_id(), housekeeping_mask);
348 tick_nohz_full_running = true;
353 void __init tick_nohz_init(void)
357 if (!tick_nohz_full_running) {
358 if (tick_nohz_init_all() < 0)
362 for_each_cpu(cpu, tick_nohz_full_mask)
363 context_tracking_cpu_set(cpu);
365 cpu_notifier(tick_nohz_cpu_down_callback, 0);
366 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
367 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
372 * NOHZ - aka dynamic tick functionality
374 #ifdef CONFIG_NO_HZ_COMMON
378 static int tick_nohz_enabled __read_mostly = 1;
379 int tick_nohz_active __read_mostly;
381 * Enable / Disable tickless mode
383 static int __init setup_tick_nohz(char *str)
385 if (!strcmp(str, "off"))
386 tick_nohz_enabled = 0;
387 else if (!strcmp(str, "on"))
388 tick_nohz_enabled = 1;
394 __setup("nohz=", setup_tick_nohz);
397 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
399 * Called from interrupt entry when the CPU was idle
401 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
402 * must be updated. Otherwise an interrupt handler could use a stale jiffy
403 * value. We do this unconditionally on any cpu, as we don't know whether the
404 * cpu, which has the update task assigned is in a long sleep.
406 static void tick_nohz_update_jiffies(ktime_t now)
410 __this_cpu_write(tick_cpu_sched.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(struct tick_sched *ts, ktime_t now)
443 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
446 sched_clock_idle_wakeup_event(0);
449 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
451 ktime_t now = ktime_get();
453 ts->idle_entrytime = now;
455 sched_clock_idle_sleep_event();
460 * get_cpu_idle_time_us - get the total idle time of a cpu
461 * @cpu: CPU number to query
462 * @last_update_time: variable to store update time in. Do not update
465 * Return the cummulative idle time (since boot) for a given
466 * CPU, in microseconds.
468 * This time is measured via accounting rather than sampling,
469 * and is as accurate as ktime_get() is.
471 * This function returns -1 if NOHZ is not enabled.
473 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
475 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
478 if (!tick_nohz_active)
482 if (last_update_time) {
483 update_ts_time_stats(cpu, ts, now, last_update_time);
484 idle = ts->idle_sleeptime;
486 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
487 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
489 idle = ktime_add(ts->idle_sleeptime, delta);
491 idle = ts->idle_sleeptime;
495 return ktime_to_us(idle);
498 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
501 * get_cpu_iowait_time_us - get the total iowait time of a cpu
502 * @cpu: CPU number to query
503 * @last_update_time: variable to store update time in. Do not update
506 * Return the cummulative iowait time (since boot) for a given
507 * CPU, in microseconds.
509 * This time is measured via accounting rather than sampling,
510 * and is as accurate as ktime_get() is.
512 * This function returns -1 if NOHZ is not enabled.
514 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
516 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
519 if (!tick_nohz_active)
523 if (last_update_time) {
524 update_ts_time_stats(cpu, ts, now, last_update_time);
525 iowait = ts->iowait_sleeptime;
527 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
528 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
530 iowait = ktime_add(ts->iowait_sleeptime, delta);
532 iowait = ts->iowait_sleeptime;
536 return ktime_to_us(iowait);
538 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
540 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
541 ktime_t now, int cpu)
543 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
544 ktime_t last_update, expires, ret = { .tv64 = 0 };
545 unsigned long rcu_delta_jiffies;
546 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
549 time_delta = timekeeping_max_deferment();
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 } while (read_seqretry(&jiffies_lock, seq));
558 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
559 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
560 next_jiffies = last_jiffies + 1;
563 /* Get the next timer wheel timer */
564 next_jiffies = get_next_timer_interrupt(last_jiffies);
565 delta_jiffies = next_jiffies - last_jiffies;
566 if (rcu_delta_jiffies < delta_jiffies) {
567 next_jiffies = last_jiffies + rcu_delta_jiffies;
568 delta_jiffies = rcu_delta_jiffies;
573 * Do not stop the tick, if we are only one off (or less)
574 * or if the cpu is required for RCU:
576 if (!ts->tick_stopped && delta_jiffies <= 1)
579 /* Schedule the tick, if we are at least one jiffie off */
580 if ((long)delta_jiffies >= 1) {
583 * If this cpu is the one which updates jiffies, then
584 * give up the assignment and let it be taken by the
585 * cpu which runs the tick timer next, which might be
586 * this cpu as well. If we don't drop this here the
587 * jiffies might be stale and do_timer() never
588 * invoked. Keep track of the fact that it was the one
589 * which had the do_timer() duty last. If this cpu is
590 * the one which had the do_timer() duty last, we
591 * limit the sleep time to the timekeeping
592 * max_deferement value which we retrieved
593 * above. Otherwise we can sleep as long as we want.
595 if (cpu == tick_do_timer_cpu) {
596 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
597 ts->do_timer_last = 1;
598 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
599 time_delta = KTIME_MAX;
600 ts->do_timer_last = 0;
601 } else if (!ts->do_timer_last) {
602 time_delta = KTIME_MAX;
605 #ifdef CONFIG_NO_HZ_FULL
607 time_delta = min(time_delta,
608 scheduler_tick_max_deferment());
613 * calculate the expiry time for the next timer wheel
614 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
615 * that there is no timer pending or at least extremely
616 * far into the future (12 days for HZ=1000). In this
617 * case we set the expiry to the end of time.
619 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
621 * Calculate the time delta for the next timer event.
622 * If the time delta exceeds the maximum time delta
623 * permitted by the current clocksource then adjust
624 * the time delta accordingly to ensure the
625 * clocksource does not wrap.
627 time_delta = min_t(u64, time_delta,
628 tick_period.tv64 * delta_jiffies);
631 if (time_delta < KTIME_MAX)
632 expires = ktime_add_ns(last_update, time_delta);
634 expires.tv64 = KTIME_MAX;
636 /* Skip reprogram of event if its not changed */
637 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
643 * nohz_stop_sched_tick can be called several times before
644 * the nohz_restart_sched_tick is called. This happens when
645 * interrupts arrive which do not cause a reschedule. In the
646 * first call we save the current tick time, so we can restart
647 * the scheduler tick in nohz_restart_sched_tick.
649 if (!ts->tick_stopped) {
650 nohz_balance_enter_idle(cpu);
651 calc_load_enter_idle();
653 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
654 ts->tick_stopped = 1;
655 trace_tick_stop(1, " ");
659 * If the expiration time == KTIME_MAX, then
660 * in this case we simply stop the tick timer.
662 if (unlikely(expires.tv64 == KTIME_MAX)) {
663 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
664 hrtimer_cancel(&ts->sched_timer);
668 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
669 hrtimer_start(&ts->sched_timer, expires,
670 HRTIMER_MODE_ABS_PINNED);
671 /* Check, if the timer was already in the past */
672 if (hrtimer_active(&ts->sched_timer))
674 } else if (!tick_program_event(expires, 0))
677 * We are past the event already. So we crossed a
678 * jiffie boundary. Update jiffies and raise the
681 tick_do_update_jiffies64(ktime_get());
683 raise_softirq_irqoff(TIMER_SOFTIRQ);
685 ts->next_jiffies = next_jiffies;
686 ts->last_jiffies = last_jiffies;
687 ts->sleep_length = ktime_sub(dev->next_event, now);
692 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
694 #ifdef CONFIG_NO_HZ_FULL
695 int cpu = smp_processor_id();
697 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
700 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
703 if (!can_stop_full_tick())
706 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
710 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
713 * If this cpu is offline and it is the one which updates
714 * jiffies, then give up the assignment and let it be taken by
715 * the cpu which runs the tick timer next. If we don't drop
716 * this here the jiffies might be stale and do_timer() never
719 if (unlikely(!cpu_online(cpu))) {
720 if (cpu == tick_do_timer_cpu)
721 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
725 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
726 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
733 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
734 static int ratelimit;
736 if (ratelimit < 10 &&
737 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
738 pr_warn("NOHZ: local_softirq_pending %02x\n",
739 (unsigned int) local_softirq_pending());
745 if (tick_nohz_full_enabled()) {
747 * Keep the tick alive to guarantee timekeeping progression
748 * if there are full dynticks CPUs around
750 if (tick_do_timer_cpu == cpu)
753 * Boot safety: make sure the timekeeping duty has been
754 * assigned before entering dyntick-idle mode,
756 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
763 static void __tick_nohz_idle_enter(struct tick_sched *ts)
765 ktime_t now, expires;
766 int cpu = smp_processor_id();
768 now = tick_nohz_start_idle(ts);
770 if (can_stop_idle_tick(cpu, ts)) {
771 int was_stopped = ts->tick_stopped;
775 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
776 if (expires.tv64 > 0LL) {
778 ts->idle_expires = expires;
781 if (!was_stopped && ts->tick_stopped)
782 ts->idle_jiffies = ts->last_jiffies;
787 * tick_nohz_idle_enter - stop the idle tick from the idle task
789 * When the next event is more than a tick into the future, stop the idle tick
790 * Called when we start the idle loop.
792 * The arch is responsible of calling:
794 * - rcu_idle_enter() after its last use of RCU before the CPU is put
796 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
798 void tick_nohz_idle_enter(void)
800 struct tick_sched *ts;
802 WARN_ON_ONCE(irqs_disabled());
805 * Update the idle state in the scheduler domain hierarchy
806 * when tick_nohz_stop_sched_tick() is called from the idle loop.
807 * State will be updated to busy during the first busy tick after
810 set_cpu_sd_state_idle();
814 ts = &__get_cpu_var(tick_cpu_sched);
816 __tick_nohz_idle_enter(ts);
820 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
823 * tick_nohz_irq_exit - update next tick event from interrupt exit
825 * When an interrupt fires while we are idle and it doesn't cause
826 * a reschedule, it may still add, modify or delete a timer, enqueue
827 * an RCU callback, etc...
828 * So we need to re-calculate and reprogram the next tick event.
830 void tick_nohz_irq_exit(void)
832 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
835 __tick_nohz_idle_enter(ts);
837 tick_nohz_full_stop_tick(ts);
841 * tick_nohz_get_sleep_length - return the length of the current sleep
843 * Called from power state control code with interrupts disabled
845 ktime_t tick_nohz_get_sleep_length(void)
847 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
849 return ts->sleep_length;
852 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
854 hrtimer_cancel(&ts->sched_timer);
855 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
858 /* Forward the time to expire in the future */
859 hrtimer_forward(&ts->sched_timer, now, tick_period);
861 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
862 hrtimer_start_expires(&ts->sched_timer,
863 HRTIMER_MODE_ABS_PINNED);
864 /* Check, if the timer was already in the past */
865 if (hrtimer_active(&ts->sched_timer))
868 if (!tick_program_event(
869 hrtimer_get_expires(&ts->sched_timer), 0))
872 /* Reread time and update jiffies */
874 tick_do_update_jiffies64(now);
878 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
880 /* Update jiffies first */
881 tick_do_update_jiffies64(now);
882 update_cpu_load_nohz();
884 calc_load_exit_idle();
885 touch_softlockup_watchdog();
887 * Cancel the scheduled timer and restore the tick
889 ts->tick_stopped = 0;
890 ts->idle_exittime = now;
892 tick_nohz_restart(ts, now);
895 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
897 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
900 if (vtime_accounting_enabled())
903 * We stopped the tick in idle. Update process times would miss the
904 * time we slept as update_process_times does only a 1 tick
905 * accounting. Enforce that this is accounted to idle !
907 ticks = jiffies - ts->idle_jiffies;
909 * We might be one off. Do not randomly account a huge number of ticks!
911 if (ticks && ticks < LONG_MAX)
912 account_idle_ticks(ticks);
917 * tick_nohz_idle_exit - restart the idle tick from the idle task
919 * Restart the idle tick when the CPU is woken up from idle
920 * This also exit the RCU extended quiescent state. The CPU
921 * can use RCU again after this function is called.
923 void tick_nohz_idle_exit(void)
925 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
930 WARN_ON_ONCE(!ts->inidle);
934 if (ts->idle_active || ts->tick_stopped)
938 tick_nohz_stop_idle(ts, now);
940 if (ts->tick_stopped) {
941 tick_nohz_restart_sched_tick(ts, now);
942 tick_nohz_account_idle_ticks(ts);
947 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
949 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
951 hrtimer_forward(&ts->sched_timer, now, tick_period);
952 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
956 * The nohz low res interrupt handler
958 static void tick_nohz_handler(struct clock_event_device *dev)
960 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
961 struct pt_regs *regs = get_irq_regs();
962 ktime_t now = ktime_get();
964 dev->next_event.tv64 = KTIME_MAX;
966 tick_sched_do_timer(now);
967 tick_sched_handle(ts, regs);
969 while (tick_nohz_reprogram(ts, now)) {
971 tick_do_update_jiffies64(now);
976 * tick_nohz_switch_to_nohz - switch to nohz mode
978 static void tick_nohz_switch_to_nohz(void)
980 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
983 if (!tick_nohz_enabled)
987 if (tick_switch_to_oneshot(tick_nohz_handler)) {
991 tick_nohz_active = 1;
992 ts->nohz_mode = NOHZ_MODE_LOWRES;
995 * Recycle the hrtimer in ts, so we can share the
996 * hrtimer_forward with the highres code.
998 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
999 /* Get the next period */
1000 next = tick_init_jiffy_update();
1003 hrtimer_set_expires(&ts->sched_timer, next);
1004 if (!tick_program_event(next, 0))
1006 next = ktime_add(next, tick_period);
1012 * When NOHZ is enabled and the tick is stopped, we need to kick the
1013 * tick timer from irq_enter() so that the jiffies update is kept
1014 * alive during long running softirqs. That's ugly as hell, but
1015 * correctness is key even if we need to fix the offending softirq in
1018 * Note, this is different to tick_nohz_restart. We just kick the
1019 * timer and do not touch the other magic bits which need to be done
1020 * when idle is left.
1022 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1025 /* Switch back to 2.6.27 behaviour */
1029 * Do not touch the tick device, when the next expiry is either
1030 * already reached or less/equal than the tick period.
1032 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1033 if (delta.tv64 <= tick_period.tv64)
1036 tick_nohz_restart(ts, now);
1040 static inline void tick_nohz_irq_enter(void)
1042 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1045 if (!ts->idle_active && !ts->tick_stopped)
1048 if (ts->idle_active)
1049 tick_nohz_stop_idle(ts, now);
1050 if (ts->tick_stopped) {
1051 tick_nohz_update_jiffies(now);
1052 tick_nohz_kick_tick(ts, now);
1058 static inline void tick_nohz_switch_to_nohz(void) { }
1059 static inline void tick_nohz_irq_enter(void) { }
1061 #endif /* CONFIG_NO_HZ_COMMON */
1064 * Called from irq_enter to notify about the possible interruption of idle()
1066 void tick_irq_enter(void)
1068 tick_check_oneshot_broadcast_this_cpu();
1069 tick_nohz_irq_enter();
1073 * High resolution timer specific code
1075 #ifdef CONFIG_HIGH_RES_TIMERS
1077 * We rearm the timer until we get disabled by the idle code.
1078 * Called with interrupts disabled.
1080 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1082 struct tick_sched *ts =
1083 container_of(timer, struct tick_sched, sched_timer);
1084 struct pt_regs *regs = get_irq_regs();
1085 ktime_t now = ktime_get();
1087 tick_sched_do_timer(now);
1090 * Do not call, when we are not in irq context and have
1091 * no valid regs pointer
1094 tick_sched_handle(ts, regs);
1096 hrtimer_forward(timer, now, tick_period);
1098 return HRTIMER_RESTART;
1101 static int sched_skew_tick;
1103 static int __init skew_tick(char *str)
1105 get_option(&str, &sched_skew_tick);
1109 early_param("skew_tick", skew_tick);
1112 * tick_setup_sched_timer - setup the tick emulation timer
1114 void tick_setup_sched_timer(void)
1116 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1117 ktime_t now = ktime_get();
1120 * Emulate tick processing via per-CPU hrtimers:
1122 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1123 ts->sched_timer.function = tick_sched_timer;
1125 /* Get the next period (per cpu) */
1126 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1128 /* Offset the tick to avert jiffies_lock contention. */
1129 if (sched_skew_tick) {
1130 u64 offset = ktime_to_ns(tick_period) >> 1;
1131 do_div(offset, num_possible_cpus());
1132 offset *= smp_processor_id();
1133 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1137 hrtimer_forward(&ts->sched_timer, now, tick_period);
1138 hrtimer_start_expires(&ts->sched_timer,
1139 HRTIMER_MODE_ABS_PINNED);
1140 /* Check, if the timer was already in the past */
1141 if (hrtimer_active(&ts->sched_timer))
1146 #ifdef CONFIG_NO_HZ_COMMON
1147 if (tick_nohz_enabled) {
1148 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1149 tick_nohz_active = 1;
1153 #endif /* HIGH_RES_TIMERS */
1155 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1156 void tick_cancel_sched_timer(int cpu)
1158 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1160 # ifdef CONFIG_HIGH_RES_TIMERS
1161 if (ts->sched_timer.base)
1162 hrtimer_cancel(&ts->sched_timer);
1165 memset(ts, 0, sizeof(*ts));
1170 * Async notification about clocksource changes
1172 void tick_clock_notify(void)
1176 for_each_possible_cpu(cpu)
1177 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1181 * Async notification about clock event changes
1183 void tick_oneshot_notify(void)
1185 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1187 set_bit(0, &ts->check_clocks);
1191 * Check, if a change happened, which makes oneshot possible.
1193 * Called cyclic from the hrtimer softirq (driven by the timer
1194 * softirq) allow_nohz signals, that we can switch into low-res nohz
1195 * mode, because high resolution timers are disabled (either compile
1198 int tick_check_oneshot_change(int allow_nohz)
1200 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1202 if (!test_and_clear_bit(0, &ts->check_clocks))
1205 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1208 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1214 tick_nohz_switch_to_nohz();