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)
396 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
398 local_irq_save(flags);
399 tick_do_update_jiffies64(now);
400 local_irq_restore(flags);
402 touch_softlockup_watchdog();
406 * Updates the per cpu time idle statistics counters
409 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
413 if (ts->idle_active) {
414 delta = ktime_sub(now, ts->idle_entrytime);
415 if (nr_iowait_cpu(cpu) > 0)
416 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
418 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
419 ts->idle_entrytime = now;
422 if (last_update_time)
423 *last_update_time = ktime_to_us(now);
427 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
429 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
432 sched_clock_idle_wakeup_event(0);
435 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
437 ktime_t now = ktime_get();
439 ts->idle_entrytime = now;
441 sched_clock_idle_sleep_event();
446 * get_cpu_idle_time_us - get the total idle time of a cpu
447 * @cpu: CPU number to query
448 * @last_update_time: variable to store update time in. Do not update
451 * Return the cummulative idle time (since boot) for a given
452 * CPU, in microseconds.
454 * This time is measured via accounting rather than sampling,
455 * and is as accurate as ktime_get() is.
457 * This function returns -1 if NOHZ is not enabled.
459 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
461 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
464 if (!tick_nohz_enabled)
468 if (last_update_time) {
469 update_ts_time_stats(cpu, ts, now, last_update_time);
470 idle = ts->idle_sleeptime;
472 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
473 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
475 idle = ktime_add(ts->idle_sleeptime, delta);
477 idle = ts->idle_sleeptime;
481 return ktime_to_us(idle);
484 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
487 * get_cpu_iowait_time_us - get the total iowait time of a cpu
488 * @cpu: CPU number to query
489 * @last_update_time: variable to store update time in. Do not update
492 * Return the cummulative iowait time (since boot) for a given
493 * CPU, in microseconds.
495 * This time is measured via accounting rather than sampling,
496 * and is as accurate as ktime_get() is.
498 * This function returns -1 if NOHZ is not enabled.
500 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
502 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
505 if (!tick_nohz_enabled)
509 if (last_update_time) {
510 update_ts_time_stats(cpu, ts, now, last_update_time);
511 iowait = ts->iowait_sleeptime;
513 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
514 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
516 iowait = ktime_add(ts->iowait_sleeptime, delta);
518 iowait = ts->iowait_sleeptime;
522 return ktime_to_us(iowait);
524 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
526 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
527 ktime_t now, int cpu)
529 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
530 ktime_t last_update, expires, ret = { .tv64 = 0 };
531 unsigned long rcu_delta_jiffies;
532 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
535 /* Read jiffies and the time when jiffies were updated last */
537 seq = read_seqbegin(&jiffies_lock);
538 last_update = last_jiffies_update;
539 last_jiffies = jiffies;
540 time_delta = timekeeping_max_deferment();
541 } while (read_seqretry(&jiffies_lock, seq));
543 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
544 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
545 next_jiffies = last_jiffies + 1;
548 /* Get the next timer wheel timer */
549 next_jiffies = get_next_timer_interrupt(last_jiffies);
550 delta_jiffies = next_jiffies - last_jiffies;
551 if (rcu_delta_jiffies < delta_jiffies) {
552 next_jiffies = last_jiffies + rcu_delta_jiffies;
553 delta_jiffies = rcu_delta_jiffies;
558 * Do not stop the tick, if we are only one off (or less)
559 * or if the cpu is required for RCU:
561 if (!ts->tick_stopped && delta_jiffies <= 1)
564 /* Schedule the tick, if we are at least one jiffie off */
565 if ((long)delta_jiffies >= 1) {
568 * If this cpu is the one which updates jiffies, then
569 * give up the assignment and let it be taken by the
570 * cpu which runs the tick timer next, which might be
571 * this cpu as well. If we don't drop this here the
572 * jiffies might be stale and do_timer() never
573 * invoked. Keep track of the fact that it was the one
574 * which had the do_timer() duty last. If this cpu is
575 * the one which had the do_timer() duty last, we
576 * limit the sleep time to the timekeeping
577 * max_deferement value which we retrieved
578 * above. Otherwise we can sleep as long as we want.
580 if (cpu == tick_do_timer_cpu) {
581 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
582 ts->do_timer_last = 1;
583 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
584 time_delta = KTIME_MAX;
585 ts->do_timer_last = 0;
586 } else if (!ts->do_timer_last) {
587 time_delta = KTIME_MAX;
590 #ifdef CONFIG_NO_HZ_FULL
592 time_delta = min(time_delta,
593 scheduler_tick_max_deferment());
598 * calculate the expiry time for the next timer wheel
599 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
600 * that there is no timer pending or at least extremely
601 * far into the future (12 days for HZ=1000). In this
602 * case we set the expiry to the end of time.
604 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
606 * Calculate the time delta for the next timer event.
607 * If the time delta exceeds the maximum time delta
608 * permitted by the current clocksource then adjust
609 * the time delta accordingly to ensure the
610 * clocksource does not wrap.
612 time_delta = min_t(u64, time_delta,
613 tick_period.tv64 * delta_jiffies);
616 if (time_delta < KTIME_MAX)
617 expires = ktime_add_ns(last_update, time_delta);
619 expires.tv64 = KTIME_MAX;
621 /* Skip reprogram of event if its not changed */
622 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
628 * nohz_stop_sched_tick can be called several times before
629 * the nohz_restart_sched_tick is called. This happens when
630 * interrupts arrive which do not cause a reschedule. In the
631 * first call we save the current tick time, so we can restart
632 * the scheduler tick in nohz_restart_sched_tick.
634 if (!ts->tick_stopped) {
635 nohz_balance_enter_idle(cpu);
636 calc_load_enter_idle();
638 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
639 ts->tick_stopped = 1;
640 trace_tick_stop(1, " ");
644 * If the expiration time == KTIME_MAX, then
645 * in this case we simply stop the tick timer.
647 if (unlikely(expires.tv64 == KTIME_MAX)) {
648 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
649 hrtimer_cancel(&ts->sched_timer);
653 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
654 hrtimer_start(&ts->sched_timer, expires,
655 HRTIMER_MODE_ABS_PINNED);
656 /* Check, if the timer was already in the past */
657 if (hrtimer_active(&ts->sched_timer))
659 } else if (!tick_program_event(expires, 0))
662 * We are past the event already. So we crossed a
663 * jiffie boundary. Update jiffies and raise the
666 tick_do_update_jiffies64(ktime_get());
668 raise_softirq_irqoff(TIMER_SOFTIRQ);
670 ts->next_jiffies = next_jiffies;
671 ts->last_jiffies = last_jiffies;
672 ts->sleep_length = ktime_sub(dev->next_event, now);
677 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
679 #ifdef CONFIG_NO_HZ_FULL
680 int cpu = smp_processor_id();
682 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
685 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
688 if (!can_stop_full_tick())
691 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
695 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
698 * If this cpu is offline and it is the one which updates
699 * jiffies, then give up the assignment and let it be taken by
700 * the cpu which runs the tick timer next. If we don't drop
701 * this here the jiffies might be stale and do_timer() never
704 if (unlikely(!cpu_online(cpu))) {
705 if (cpu == tick_do_timer_cpu)
706 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
710 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
716 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
717 static int ratelimit;
719 if (ratelimit < 10 &&
720 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
721 pr_warn("NOHZ: local_softirq_pending %02x\n",
722 (unsigned int) local_softirq_pending());
728 if (tick_nohz_full_enabled()) {
730 * Keep the tick alive to guarantee timekeeping progression
731 * if there are full dynticks CPUs around
733 if (tick_do_timer_cpu == cpu)
736 * Boot safety: make sure the timekeeping duty has been
737 * assigned before entering dyntick-idle mode,
739 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
746 static void __tick_nohz_idle_enter(struct tick_sched *ts)
748 ktime_t now, expires;
749 int cpu = smp_processor_id();
751 now = tick_nohz_start_idle(ts);
753 if (can_stop_idle_tick(cpu, ts)) {
754 int was_stopped = ts->tick_stopped;
758 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
759 if (expires.tv64 > 0LL) {
761 ts->idle_expires = expires;
764 if (!was_stopped && ts->tick_stopped)
765 ts->idle_jiffies = ts->last_jiffies;
770 * tick_nohz_idle_enter - stop the idle tick from the idle task
772 * When the next event is more than a tick into the future, stop the idle tick
773 * Called when we start the idle loop.
775 * The arch is responsible of calling:
777 * - rcu_idle_enter() after its last use of RCU before the CPU is put
779 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
781 void tick_nohz_idle_enter(void)
783 struct tick_sched *ts;
785 WARN_ON_ONCE(irqs_disabled());
788 * Update the idle state in the scheduler domain hierarchy
789 * when tick_nohz_stop_sched_tick() is called from the idle loop.
790 * State will be updated to busy during the first busy tick after
793 set_cpu_sd_state_idle();
797 ts = &__get_cpu_var(tick_cpu_sched);
799 * set ts->inidle unconditionally. even if the system did not
800 * switch to nohz mode the cpu frequency governers rely on the
801 * update of the idle time accounting in tick_nohz_start_idle().
804 __tick_nohz_idle_enter(ts);
808 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
811 * tick_nohz_irq_exit - update next tick event from interrupt exit
813 * When an interrupt fires while we are idle and it doesn't cause
814 * a reschedule, it may still add, modify or delete a timer, enqueue
815 * an RCU callback, etc...
816 * So we need to re-calculate and reprogram the next tick event.
818 void tick_nohz_irq_exit(void)
820 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
823 __tick_nohz_idle_enter(ts);
825 tick_nohz_full_stop_tick(ts);
829 * tick_nohz_get_sleep_length - return the length of the current sleep
831 * Called from power state control code with interrupts disabled
833 ktime_t tick_nohz_get_sleep_length(void)
835 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
837 return ts->sleep_length;
840 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
842 hrtimer_cancel(&ts->sched_timer);
843 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
846 /* Forward the time to expire in the future */
847 hrtimer_forward(&ts->sched_timer, now, tick_period);
849 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
850 hrtimer_start_expires(&ts->sched_timer,
851 HRTIMER_MODE_ABS_PINNED);
852 /* Check, if the timer was already in the past */
853 if (hrtimer_active(&ts->sched_timer))
856 if (!tick_program_event(
857 hrtimer_get_expires(&ts->sched_timer), 0))
860 /* Reread time and update jiffies */
862 tick_do_update_jiffies64(now);
866 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
868 /* Update jiffies first */
869 tick_do_update_jiffies64(now);
870 update_cpu_load_nohz();
872 calc_load_exit_idle();
873 touch_softlockup_watchdog();
875 * Cancel the scheduled timer and restore the tick
877 ts->tick_stopped = 0;
878 ts->idle_exittime = now;
880 tick_nohz_restart(ts, now);
883 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
885 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
888 if (vtime_accounting_enabled())
891 * We stopped the tick in idle. Update process times would miss the
892 * time we slept as update_process_times does only a 1 tick
893 * accounting. Enforce that this is accounted to idle !
895 ticks = jiffies - ts->idle_jiffies;
897 * We might be one off. Do not randomly account a huge number of ticks!
899 if (ticks && ticks < LONG_MAX)
900 account_idle_ticks(ticks);
905 * tick_nohz_idle_exit - restart the idle tick from the idle task
907 * Restart the idle tick when the CPU is woken up from idle
908 * This also exit the RCU extended quiescent state. The CPU
909 * can use RCU again after this function is called.
911 void tick_nohz_idle_exit(void)
913 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
918 WARN_ON_ONCE(!ts->inidle);
922 if (ts->idle_active || ts->tick_stopped)
926 tick_nohz_stop_idle(ts, now);
928 if (ts->tick_stopped) {
929 tick_nohz_restart_sched_tick(ts, now);
930 tick_nohz_account_idle_ticks(ts);
935 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
937 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
939 hrtimer_forward(&ts->sched_timer, now, tick_period);
940 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
944 * The nohz low res interrupt handler
946 static void tick_nohz_handler(struct clock_event_device *dev)
948 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
949 struct pt_regs *regs = get_irq_regs();
950 ktime_t now = ktime_get();
952 dev->next_event.tv64 = KTIME_MAX;
954 tick_sched_do_timer(now);
955 tick_sched_handle(ts, regs);
957 while (tick_nohz_reprogram(ts, now)) {
959 tick_do_update_jiffies64(now);
964 * tick_nohz_switch_to_nohz - switch to nohz mode
966 static void tick_nohz_switch_to_nohz(void)
968 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
971 if (!tick_nohz_enabled)
975 if (tick_switch_to_oneshot(tick_nohz_handler)) {
980 ts->nohz_mode = NOHZ_MODE_LOWRES;
983 * Recycle the hrtimer in ts, so we can share the
984 * hrtimer_forward with the highres code.
986 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
987 /* Get the next period */
988 next = tick_init_jiffy_update();
991 hrtimer_set_expires(&ts->sched_timer, next);
992 if (!tick_program_event(next, 0))
994 next = ktime_add(next, tick_period);
1000 * When NOHZ is enabled and the tick is stopped, we need to kick the
1001 * tick timer from irq_enter() so that the jiffies update is kept
1002 * alive during long running softirqs. That's ugly as hell, but
1003 * correctness is key even if we need to fix the offending softirq in
1006 * Note, this is different to tick_nohz_restart. We just kick the
1007 * timer and do not touch the other magic bits which need to be done
1008 * when idle is left.
1010 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1013 /* Switch back to 2.6.27 behaviour */
1017 * Do not touch the tick device, when the next expiry is either
1018 * already reached or less/equal than the tick period.
1020 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1021 if (delta.tv64 <= tick_period.tv64)
1024 tick_nohz_restart(ts, now);
1028 static inline void tick_check_nohz_this_cpu(void)
1030 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1033 if (!ts->idle_active && !ts->tick_stopped)
1036 if (ts->idle_active)
1037 tick_nohz_stop_idle(ts, now);
1038 if (ts->tick_stopped) {
1039 tick_nohz_update_jiffies(now);
1040 tick_nohz_kick_tick(ts, now);
1046 static inline void tick_nohz_switch_to_nohz(void) { }
1047 static inline void tick_check_nohz_this_cpu(void) { }
1049 #endif /* CONFIG_NO_HZ_COMMON */
1052 * Called from irq_enter to notify about the possible interruption of idle()
1054 void tick_check_idle(void)
1056 tick_check_oneshot_broadcast_this_cpu();
1057 tick_check_nohz_this_cpu();
1061 * High resolution timer specific code
1063 #ifdef CONFIG_HIGH_RES_TIMERS
1065 * We rearm the timer until we get disabled by the idle code.
1066 * Called with interrupts disabled.
1068 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1070 struct tick_sched *ts =
1071 container_of(timer, struct tick_sched, sched_timer);
1072 struct pt_regs *regs = get_irq_regs();
1073 ktime_t now = ktime_get();
1075 tick_sched_do_timer(now);
1078 * Do not call, when we are not in irq context and have
1079 * no valid regs pointer
1082 tick_sched_handle(ts, regs);
1084 hrtimer_forward(timer, now, tick_period);
1086 return HRTIMER_RESTART;
1089 static int sched_skew_tick;
1091 static int __init skew_tick(char *str)
1093 get_option(&str, &sched_skew_tick);
1097 early_param("skew_tick", skew_tick);
1100 * tick_setup_sched_timer - setup the tick emulation timer
1102 void tick_setup_sched_timer(void)
1104 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1105 ktime_t now = ktime_get();
1108 * Emulate tick processing via per-CPU hrtimers:
1110 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1111 ts->sched_timer.function = tick_sched_timer;
1113 /* Get the next period (per cpu) */
1114 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1116 /* Offset the tick to avert jiffies_lock contention. */
1117 if (sched_skew_tick) {
1118 u64 offset = ktime_to_ns(tick_period) >> 1;
1119 do_div(offset, num_possible_cpus());
1120 offset *= smp_processor_id();
1121 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1125 hrtimer_forward(&ts->sched_timer, now, tick_period);
1126 hrtimer_start_expires(&ts->sched_timer,
1127 HRTIMER_MODE_ABS_PINNED);
1128 /* Check, if the timer was already in the past */
1129 if (hrtimer_active(&ts->sched_timer))
1134 #ifdef CONFIG_NO_HZ_COMMON
1135 if (tick_nohz_enabled)
1136 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1139 #endif /* HIGH_RES_TIMERS */
1141 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1142 void tick_cancel_sched_timer(int cpu)
1144 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1146 # ifdef CONFIG_HIGH_RES_TIMERS
1147 if (ts->sched_timer.base)
1148 hrtimer_cancel(&ts->sched_timer);
1151 memset(ts, 0, sizeof(*ts));
1156 * Async notification about clocksource changes
1158 void tick_clock_notify(void)
1162 for_each_possible_cpu(cpu)
1163 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1167 * Async notification about clock event changes
1169 void tick_oneshot_notify(void)
1171 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1173 set_bit(0, &ts->check_clocks);
1177 * Check, if a change happened, which makes oneshot possible.
1179 * Called cyclic from the hrtimer softirq (driven by the timer
1180 * softirq) allow_nohz signals, that we can switch into low-res nohz
1181 * mode, because high resolution timers are disabled (either compile
1184 int tick_check_oneshot_change(int allow_nohz)
1186 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1188 if (!test_and_clear_bit(0, &ts->check_clocks))
1191 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1194 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1200 tick_nohz_switch_to_nohz();