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);
92 * Initialize and return retrieve the jiffies update.
94 static ktime_t tick_init_jiffy_update(void)
98 write_seqlock(&jiffies_lock);
99 /* Did we start the jiffies update yet ? */
100 if (last_jiffies_update.tv64 == 0)
101 last_jiffies_update = tick_next_period;
102 period = last_jiffies_update;
103 write_sequnlock(&jiffies_lock);
108 static void tick_sched_do_timer(ktime_t now)
110 int cpu = smp_processor_id();
112 #ifdef CONFIG_NO_HZ_COMMON
114 * Check if the do_timer duty was dropped. We don't care about
115 * concurrency: This happens only when the cpu in charge went
116 * into a long sleep. If two cpus happen to assign themself to
117 * this duty, then the jiffies update is still serialized by
120 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
121 && !tick_nohz_full_cpu(cpu))
122 tick_do_timer_cpu = cpu;
125 /* Check, if the jiffies need an update */
126 if (tick_do_timer_cpu == cpu)
127 tick_do_update_jiffies64(now);
130 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
132 #ifdef CONFIG_NO_HZ_COMMON
134 * When we are idle and the tick is stopped, we have to touch
135 * the watchdog as we might not schedule for a really long
136 * time. This happens on complete idle SMP systems while
137 * waiting on the login prompt. We also increment the "start of
138 * idle" jiffy stamp so the idle accounting adjustment we do
139 * when we go busy again does not account too much ticks.
141 if (ts->tick_stopped) {
142 touch_softlockup_watchdog();
143 if (is_idle_task(current))
147 update_process_times(user_mode(regs));
148 profile_tick(CPU_PROFILING);
151 #ifdef CONFIG_NO_HZ_FULL
152 cpumask_var_t tick_nohz_full_mask;
153 bool tick_nohz_full_running;
155 static bool can_stop_full_tick(void)
157 WARN_ON_ONCE(!irqs_disabled());
159 if (!sched_can_stop_tick()) {
160 trace_tick_stop(0, "more than 1 task in runqueue\n");
164 if (!posix_cpu_timers_can_stop_tick(current)) {
165 trace_tick_stop(0, "posix timers running\n");
169 if (!perf_event_can_stop_tick()) {
170 trace_tick_stop(0, "perf events running\n");
174 /* sched_clock_tick() needs us? */
175 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
177 * TODO: kick full dynticks CPUs when
178 * sched_clock_stable is set.
180 if (!sched_clock_stable) {
181 trace_tick_stop(0, "unstable sched clock\n");
183 * Don't allow the user to think they can get
184 * full NO_HZ with this machine.
186 WARN_ONCE(tick_nohz_full_running,
187 "NO_HZ FULL will not work with unstable sched clock");
195 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
198 * Re-evaluate the need for the tick on the current CPU
199 * and restart it if necessary.
201 void __tick_nohz_full_check(void)
203 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
205 if (tick_nohz_full_cpu(smp_processor_id())) {
206 if (ts->tick_stopped && !is_idle_task(current)) {
207 if (!can_stop_full_tick())
208 tick_nohz_restart_sched_tick(ts, ktime_get());
213 static void nohz_full_kick_work_func(struct irq_work *work)
215 __tick_nohz_full_check();
218 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
219 .func = nohz_full_kick_work_func,
223 * Kick the current CPU if it's full dynticks in order to force it to
224 * re-evaluate its dependency on the tick and restart it if necessary.
226 void tick_nohz_full_kick(void)
228 if (tick_nohz_full_cpu(smp_processor_id()))
229 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
232 static void nohz_full_kick_ipi(void *info)
234 __tick_nohz_full_check();
238 * Kick all full dynticks CPUs in order to force these to re-evaluate
239 * their dependency on the tick and restart it if necessary.
241 void tick_nohz_full_kick_all(void)
243 if (!tick_nohz_full_running)
247 smp_call_function_many(tick_nohz_full_mask,
248 nohz_full_kick_ipi, NULL, false);
249 tick_nohz_full_kick();
254 * Re-evaluate the need for the tick as we switch the current task.
255 * It might need the tick due to per task/process properties:
256 * perf events, posix cpu timers, ...
258 void __tick_nohz_task_switch(struct task_struct *tsk)
262 local_irq_save(flags);
264 if (!tick_nohz_full_cpu(smp_processor_id()))
267 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
268 tick_nohz_full_kick();
271 local_irq_restore(flags);
274 /* Parse the boot-time nohz CPU list from the kernel parameters. */
275 static int __init tick_nohz_full_setup(char *str)
279 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
280 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
281 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
285 cpu = smp_processor_id();
286 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
287 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
288 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
290 tick_nohz_full_running = true;
294 __setup("nohz_full=", tick_nohz_full_setup);
296 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
297 unsigned long action,
300 unsigned int cpu = (unsigned long)hcpu;
302 switch (action & ~CPU_TASKS_FROZEN) {
303 case CPU_DOWN_PREPARE:
305 * If we handle the timekeeping duty for full dynticks CPUs,
306 * we can't safely shutdown that CPU.
308 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
316 * Worst case string length in chunks of CPU range seems 2 steps
317 * separations: 0,2,4,6,...
318 * This is NR_CPUS + sizeof('\0')
320 static char __initdata nohz_full_buf[NR_CPUS + 1];
322 static int tick_nohz_init_all(void)
326 #ifdef CONFIG_NO_HZ_FULL_ALL
327 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
328 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
332 cpumask_setall(tick_nohz_full_mask);
333 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask);
334 tick_nohz_full_running = true;
339 void __init tick_nohz_init(void)
343 if (!tick_nohz_full_running) {
344 if (tick_nohz_init_all() < 0)
348 for_each_cpu(cpu, tick_nohz_full_mask)
349 context_tracking_cpu_set(cpu);
351 cpu_notifier(tick_nohz_cpu_down_callback, 0);
352 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
353 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
358 * NOHZ - aka dynamic tick functionality
360 #ifdef CONFIG_NO_HZ_COMMON
364 int tick_nohz_enabled __read_mostly = 1;
367 * Enable / Disable tickless mode
369 static int __init setup_tick_nohz(char *str)
371 if (!strcmp(str, "off"))
372 tick_nohz_enabled = 0;
373 else if (!strcmp(str, "on"))
374 tick_nohz_enabled = 1;
380 __setup("nohz=", setup_tick_nohz);
383 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
385 * Called from interrupt entry when the CPU was idle
387 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
388 * must be updated. Otherwise an interrupt handler could use a stale jiffy
389 * value. We do this unconditionally on any cpu, as we don't know whether the
390 * cpu, which has the update task assigned is in a long sleep.
392 static void tick_nohz_update_jiffies(ktime_t now)
394 int cpu = smp_processor_id();
395 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
398 ts->idle_waketime = now;
400 local_irq_save(flags);
401 tick_do_update_jiffies64(now);
402 local_irq_restore(flags);
404 touch_softlockup_watchdog();
408 * Updates the per cpu time idle statistics counters
411 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
415 if (ts->idle_active) {
416 delta = ktime_sub(now, ts->idle_entrytime);
417 if (nr_iowait_cpu(cpu) > 0)
418 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
420 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
421 ts->idle_entrytime = now;
424 if (last_update_time)
425 *last_update_time = ktime_to_us(now);
429 static void tick_nohz_stop_idle(int cpu, ktime_t now)
431 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
433 update_ts_time_stats(cpu, ts, now, NULL);
436 sched_clock_idle_wakeup_event(0);
439 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
441 ktime_t now = ktime_get();
443 ts->idle_entrytime = now;
445 sched_clock_idle_sleep_event();
450 * get_cpu_idle_time_us - get the total idle time of a cpu
451 * @cpu: CPU number to query
452 * @last_update_time: variable to store update time in. Do not update
455 * Return the cummulative idle time (since boot) for a given
456 * CPU, in microseconds.
458 * This time is measured via accounting rather than sampling,
459 * and is as accurate as ktime_get() is.
461 * This function returns -1 if NOHZ is not enabled.
463 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
465 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
468 if (!tick_nohz_enabled)
472 if (last_update_time) {
473 update_ts_time_stats(cpu, ts, now, last_update_time);
474 idle = ts->idle_sleeptime;
476 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
477 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
479 idle = ktime_add(ts->idle_sleeptime, delta);
481 idle = ts->idle_sleeptime;
485 return ktime_to_us(idle);
488 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
491 * get_cpu_iowait_time_us - get the total iowait time of a cpu
492 * @cpu: CPU number to query
493 * @last_update_time: variable to store update time in. Do not update
496 * Return the cummulative iowait time (since boot) for a given
497 * CPU, in microseconds.
499 * This time is measured via accounting rather than sampling,
500 * and is as accurate as ktime_get() is.
502 * This function returns -1 if NOHZ is not enabled.
504 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
506 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
509 if (!tick_nohz_enabled)
513 if (last_update_time) {
514 update_ts_time_stats(cpu, ts, now, last_update_time);
515 iowait = ts->iowait_sleeptime;
517 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
518 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
520 iowait = ktime_add(ts->iowait_sleeptime, delta);
522 iowait = ts->iowait_sleeptime;
526 return ktime_to_us(iowait);
528 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
530 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
531 ktime_t now, int cpu)
533 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
534 ktime_t last_update, expires, ret = { .tv64 = 0 };
535 unsigned long rcu_delta_jiffies;
536 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
539 /* Read jiffies and the time when jiffies were updated last */
541 seq = read_seqbegin(&jiffies_lock);
542 last_update = last_jiffies_update;
543 last_jiffies = jiffies;
544 time_delta = timekeeping_max_deferment();
545 } while (read_seqretry(&jiffies_lock, seq));
547 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
548 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
549 next_jiffies = last_jiffies + 1;
552 /* Get the next timer wheel timer */
553 next_jiffies = get_next_timer_interrupt(last_jiffies);
554 delta_jiffies = next_jiffies - last_jiffies;
555 if (rcu_delta_jiffies < delta_jiffies) {
556 next_jiffies = last_jiffies + rcu_delta_jiffies;
557 delta_jiffies = rcu_delta_jiffies;
562 * Do not stop the tick, if we are only one off (or less)
563 * or if the cpu is required for RCU:
565 if (!ts->tick_stopped && delta_jiffies <= 1)
568 /* Schedule the tick, if we are at least one jiffie off */
569 if ((long)delta_jiffies >= 1) {
572 * If this cpu is the one which updates jiffies, then
573 * give up the assignment and let it be taken by the
574 * cpu which runs the tick timer next, which might be
575 * this cpu as well. If we don't drop this here the
576 * jiffies might be stale and do_timer() never
577 * invoked. Keep track of the fact that it was the one
578 * which had the do_timer() duty last. If this cpu is
579 * the one which had the do_timer() duty last, we
580 * limit the sleep time to the timekeeping
581 * max_deferement value which we retrieved
582 * above. Otherwise we can sleep as long as we want.
584 if (cpu == tick_do_timer_cpu) {
585 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
586 ts->do_timer_last = 1;
587 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
588 time_delta = KTIME_MAX;
589 ts->do_timer_last = 0;
590 } else if (!ts->do_timer_last) {
591 time_delta = KTIME_MAX;
594 #ifdef CONFIG_NO_HZ_FULL
596 time_delta = min(time_delta,
597 scheduler_tick_max_deferment());
602 * calculate the expiry time for the next timer wheel
603 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
604 * that there is no timer pending or at least extremely
605 * far into the future (12 days for HZ=1000). In this
606 * case we set the expiry to the end of time.
608 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
610 * Calculate the time delta for the next timer event.
611 * If the time delta exceeds the maximum time delta
612 * permitted by the current clocksource then adjust
613 * the time delta accordingly to ensure the
614 * clocksource does not wrap.
616 time_delta = min_t(u64, time_delta,
617 tick_period.tv64 * delta_jiffies);
620 if (time_delta < KTIME_MAX)
621 expires = ktime_add_ns(last_update, time_delta);
623 expires.tv64 = KTIME_MAX;
625 /* Skip reprogram of event if its not changed */
626 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
632 * nohz_stop_sched_tick can be called several times before
633 * the nohz_restart_sched_tick is called. This happens when
634 * interrupts arrive which do not cause a reschedule. In the
635 * first call we save the current tick time, so we can restart
636 * the scheduler tick in nohz_restart_sched_tick.
638 if (!ts->tick_stopped) {
639 nohz_balance_enter_idle(cpu);
640 calc_load_enter_idle();
642 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
643 ts->tick_stopped = 1;
644 trace_tick_stop(1, " ");
648 * If the expiration time == KTIME_MAX, then
649 * in this case we simply stop the tick timer.
651 if (unlikely(expires.tv64 == KTIME_MAX)) {
652 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
653 hrtimer_cancel(&ts->sched_timer);
657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
658 hrtimer_start(&ts->sched_timer, expires,
659 HRTIMER_MODE_ABS_PINNED);
660 /* Check, if the timer was already in the past */
661 if (hrtimer_active(&ts->sched_timer))
663 } else if (!tick_program_event(expires, 0))
666 * We are past the event already. So we crossed a
667 * jiffie boundary. Update jiffies and raise the
670 tick_do_update_jiffies64(ktime_get());
672 raise_softirq_irqoff(TIMER_SOFTIRQ);
674 ts->next_jiffies = next_jiffies;
675 ts->last_jiffies = last_jiffies;
676 ts->sleep_length = ktime_sub(dev->next_event, now);
681 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
683 #ifdef CONFIG_NO_HZ_FULL
684 int cpu = smp_processor_id();
686 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
689 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
692 if (!can_stop_full_tick())
695 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
699 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
702 * If this cpu is offline and it is the one which updates
703 * jiffies, then give up the assignment and let it be taken by
704 * the cpu which runs the tick timer next. If we don't drop
705 * this here the jiffies might be stale and do_timer() never
708 if (unlikely(!cpu_online(cpu))) {
709 if (cpu == tick_do_timer_cpu)
710 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
714 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
720 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
721 static int ratelimit;
723 if (ratelimit < 10 &&
724 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
725 pr_warn("NOHZ: local_softirq_pending %02x\n",
726 (unsigned int) local_softirq_pending());
732 if (tick_nohz_full_enabled()) {
734 * Keep the tick alive to guarantee timekeeping progression
735 * if there are full dynticks CPUs around
737 if (tick_do_timer_cpu == cpu)
740 * Boot safety: make sure the timekeeping duty has been
741 * assigned before entering dyntick-idle mode,
743 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
750 static void __tick_nohz_idle_enter(struct tick_sched *ts)
752 ktime_t now, expires;
753 int cpu = smp_processor_id();
755 now = tick_nohz_start_idle(cpu, ts);
757 if (can_stop_idle_tick(cpu, ts)) {
758 int was_stopped = ts->tick_stopped;
762 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
763 if (expires.tv64 > 0LL) {
765 ts->idle_expires = expires;
768 if (!was_stopped && ts->tick_stopped)
769 ts->idle_jiffies = ts->last_jiffies;
774 * tick_nohz_idle_enter - stop the idle tick from the idle task
776 * When the next event is more than a tick into the future, stop the idle tick
777 * Called when we start the idle loop.
779 * The arch is responsible of calling:
781 * - rcu_idle_enter() after its last use of RCU before the CPU is put
783 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
785 void tick_nohz_idle_enter(void)
787 struct tick_sched *ts;
789 WARN_ON_ONCE(irqs_disabled());
792 * Update the idle state in the scheduler domain hierarchy
793 * when tick_nohz_stop_sched_tick() is called from the idle loop.
794 * State will be updated to busy during the first busy tick after
797 set_cpu_sd_state_idle();
801 ts = &__get_cpu_var(tick_cpu_sched);
803 * set ts->inidle unconditionally. even if the system did not
804 * switch to nohz mode the cpu frequency governers rely on the
805 * update of the idle time accounting in tick_nohz_start_idle().
808 __tick_nohz_idle_enter(ts);
812 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
815 * tick_nohz_irq_exit - update next tick event from interrupt exit
817 * When an interrupt fires while we are idle and it doesn't cause
818 * a reschedule, it may still add, modify or delete a timer, enqueue
819 * an RCU callback, etc...
820 * So we need to re-calculate and reprogram the next tick event.
822 void tick_nohz_irq_exit(void)
824 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
827 __tick_nohz_idle_enter(ts);
829 tick_nohz_full_stop_tick(ts);
833 * tick_nohz_get_sleep_length - return the length of the current sleep
835 * Called from power state control code with interrupts disabled
837 ktime_t tick_nohz_get_sleep_length(void)
839 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
841 return ts->sleep_length;
844 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
846 hrtimer_cancel(&ts->sched_timer);
847 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
850 /* Forward the time to expire in the future */
851 hrtimer_forward(&ts->sched_timer, now, tick_period);
853 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
854 hrtimer_start_expires(&ts->sched_timer,
855 HRTIMER_MODE_ABS_PINNED);
856 /* Check, if the timer was already in the past */
857 if (hrtimer_active(&ts->sched_timer))
860 if (!tick_program_event(
861 hrtimer_get_expires(&ts->sched_timer), 0))
864 /* Reread time and update jiffies */
866 tick_do_update_jiffies64(now);
870 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
872 /* Update jiffies first */
873 tick_do_update_jiffies64(now);
874 update_cpu_load_nohz();
876 calc_load_exit_idle();
877 touch_softlockup_watchdog();
879 * Cancel the scheduled timer and restore the tick
881 ts->tick_stopped = 0;
882 ts->idle_exittime = now;
884 tick_nohz_restart(ts, now);
887 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
889 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
892 if (vtime_accounting_enabled())
895 * We stopped the tick in idle. Update process times would miss the
896 * time we slept as update_process_times does only a 1 tick
897 * accounting. Enforce that this is accounted to idle !
899 ticks = jiffies - ts->idle_jiffies;
901 * We might be one off. Do not randomly account a huge number of ticks!
903 if (ticks && ticks < LONG_MAX)
904 account_idle_ticks(ticks);
909 * tick_nohz_idle_exit - restart the idle tick from the idle task
911 * Restart the idle tick when the CPU is woken up from idle
912 * This also exit the RCU extended quiescent state. The CPU
913 * can use RCU again after this function is called.
915 void tick_nohz_idle_exit(void)
917 int cpu = smp_processor_id();
918 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
923 WARN_ON_ONCE(!ts->inidle);
927 if (ts->idle_active || ts->tick_stopped)
931 tick_nohz_stop_idle(cpu, now);
933 if (ts->tick_stopped) {
934 tick_nohz_restart_sched_tick(ts, now);
935 tick_nohz_account_idle_ticks(ts);
940 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
942 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
944 hrtimer_forward(&ts->sched_timer, now, tick_period);
945 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
949 * The nohz low res interrupt handler
951 static void tick_nohz_handler(struct clock_event_device *dev)
953 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
954 struct pt_regs *regs = get_irq_regs();
955 ktime_t now = ktime_get();
957 dev->next_event.tv64 = KTIME_MAX;
959 tick_sched_do_timer(now);
960 tick_sched_handle(ts, regs);
962 while (tick_nohz_reprogram(ts, now)) {
964 tick_do_update_jiffies64(now);
969 * tick_nohz_switch_to_nohz - switch to nohz mode
971 static void tick_nohz_switch_to_nohz(void)
973 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
976 if (!tick_nohz_enabled)
980 if (tick_switch_to_oneshot(tick_nohz_handler)) {
985 ts->nohz_mode = NOHZ_MODE_LOWRES;
988 * Recycle the hrtimer in ts, so we can share the
989 * hrtimer_forward with the highres code.
991 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
992 /* Get the next period */
993 next = tick_init_jiffy_update();
996 hrtimer_set_expires(&ts->sched_timer, next);
997 if (!tick_program_event(next, 0))
999 next = ktime_add(next, tick_period);
1005 * When NOHZ is enabled and the tick is stopped, we need to kick the
1006 * tick timer from irq_enter() so that the jiffies update is kept
1007 * alive during long running softirqs. That's ugly as hell, but
1008 * correctness is key even if we need to fix the offending softirq in
1011 * Note, this is different to tick_nohz_restart. We just kick the
1012 * timer and do not touch the other magic bits which need to be done
1013 * when idle is left.
1015 static void tick_nohz_kick_tick(int cpu, ktime_t now)
1018 /* Switch back to 2.6.27 behaviour */
1020 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1024 * Do not touch the tick device, when the next expiry is either
1025 * already reached or less/equal than the tick period.
1027 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1028 if (delta.tv64 <= tick_period.tv64)
1031 tick_nohz_restart(ts, now);
1035 static inline void tick_check_nohz(int cpu)
1037 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1040 if (!ts->idle_active && !ts->tick_stopped)
1043 if (ts->idle_active)
1044 tick_nohz_stop_idle(cpu, now);
1045 if (ts->tick_stopped) {
1046 tick_nohz_update_jiffies(now);
1047 tick_nohz_kick_tick(cpu, now);
1053 static inline void tick_nohz_switch_to_nohz(void) { }
1054 static inline void tick_check_nohz(int cpu) { }
1056 #endif /* CONFIG_NO_HZ_COMMON */
1059 * Called from irq_enter to notify about the possible interruption of idle()
1061 void tick_check_idle(int cpu)
1063 tick_check_oneshot_broadcast(cpu);
1064 tick_check_nohz(cpu);
1068 * High resolution timer specific code
1070 #ifdef CONFIG_HIGH_RES_TIMERS
1072 * We rearm the timer until we get disabled by the idle code.
1073 * Called with interrupts disabled.
1075 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1077 struct tick_sched *ts =
1078 container_of(timer, struct tick_sched, sched_timer);
1079 struct pt_regs *regs = get_irq_regs();
1080 ktime_t now = ktime_get();
1082 tick_sched_do_timer(now);
1085 * Do not call, when we are not in irq context and have
1086 * no valid regs pointer
1089 tick_sched_handle(ts, regs);
1091 hrtimer_forward(timer, now, tick_period);
1093 return HRTIMER_RESTART;
1096 static int sched_skew_tick;
1098 static int __init skew_tick(char *str)
1100 get_option(&str, &sched_skew_tick);
1104 early_param("skew_tick", skew_tick);
1107 * tick_setup_sched_timer - setup the tick emulation timer
1109 void tick_setup_sched_timer(void)
1111 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1112 ktime_t now = ktime_get();
1115 * Emulate tick processing via per-CPU hrtimers:
1117 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1118 ts->sched_timer.function = tick_sched_timer;
1120 /* Get the next period (per cpu) */
1121 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1123 /* Offset the tick to avert jiffies_lock contention. */
1124 if (sched_skew_tick) {
1125 u64 offset = ktime_to_ns(tick_period) >> 1;
1126 do_div(offset, num_possible_cpus());
1127 offset *= smp_processor_id();
1128 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1132 hrtimer_forward(&ts->sched_timer, now, tick_period);
1133 hrtimer_start_expires(&ts->sched_timer,
1134 HRTIMER_MODE_ABS_PINNED);
1135 /* Check, if the timer was already in the past */
1136 if (hrtimer_active(&ts->sched_timer))
1141 #ifdef CONFIG_NO_HZ_COMMON
1142 if (tick_nohz_enabled)
1143 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1146 #endif /* HIGH_RES_TIMERS */
1148 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1149 void tick_cancel_sched_timer(int cpu)
1151 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1153 # ifdef CONFIG_HIGH_RES_TIMERS
1154 if (ts->sched_timer.base)
1155 hrtimer_cancel(&ts->sched_timer);
1158 memset(ts, 0, sizeof(*ts));
1163 * Async notification about clocksource changes
1165 void tick_clock_notify(void)
1169 for_each_possible_cpu(cpu)
1170 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1174 * Async notification about clock event changes
1176 void tick_oneshot_notify(void)
1178 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1180 set_bit(0, &ts->check_clocks);
1184 * Check, if a change happened, which makes oneshot possible.
1186 * Called cyclic from the hrtimer softirq (driven by the timer
1187 * softirq) allow_nohz signals, that we can switch into low-res nohz
1188 * mode, because high resolution timers are disabled (either compile
1191 int tick_check_oneshot_change(int allow_nohz)
1193 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1195 if (!test_and_clear_bit(0, &ts->check_clocks))
1198 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1201 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1207 tick_nohz_switch_to_nohz();