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);
93 * Initialize and return retrieve the jiffies update.
95 static ktime_t tick_init_jiffy_update(void)
99 write_seqlock(&jiffies_lock);
100 /* Did we start the jiffies update yet ? */
101 if (last_jiffies_update.tv64 == 0)
102 last_jiffies_update = tick_next_period;
103 period = last_jiffies_update;
104 write_sequnlock(&jiffies_lock);
109 static void tick_sched_do_timer(ktime_t now)
111 int cpu = smp_processor_id();
113 #ifdef CONFIG_NO_HZ_COMMON
115 * Check if the do_timer duty was dropped. We don't care about
116 * concurrency: This happens only when the cpu in charge went
117 * into a long sleep. If two cpus happen to assign themself to
118 * this duty, then the jiffies update is still serialized by
121 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
122 && !tick_nohz_full_cpu(cpu))
123 tick_do_timer_cpu = cpu;
126 /* Check, if the jiffies need an update */
127 if (tick_do_timer_cpu == cpu)
128 tick_do_update_jiffies64(now);
131 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
133 #ifdef CONFIG_NO_HZ_COMMON
135 * When we are idle and the tick is stopped, we have to touch
136 * the watchdog as we might not schedule for a really long
137 * time. This happens on complete idle SMP systems while
138 * waiting on the login prompt. We also increment the "start of
139 * idle" jiffy stamp so the idle accounting adjustment we do
140 * when we go busy again does not account too much ticks.
142 if (ts->tick_stopped) {
143 touch_softlockup_watchdog();
144 if (is_idle_task(current))
148 update_process_times(user_mode(regs));
149 profile_tick(CPU_PROFILING);
152 #ifdef CONFIG_NO_HZ_FULL
153 cpumask_var_t tick_nohz_full_mask;
154 bool tick_nohz_full_running;
156 static bool can_stop_full_tick(void)
158 WARN_ON_ONCE(!irqs_disabled());
160 if (!sched_can_stop_tick()) {
161 trace_tick_stop(0, "more than 1 task in runqueue\n");
165 if (!posix_cpu_timers_can_stop_tick(current)) {
166 trace_tick_stop(0, "posix timers running\n");
170 if (!perf_event_can_stop_tick()) {
171 trace_tick_stop(0, "perf events running\n");
175 /* sched_clock_tick() needs us? */
176 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
178 * TODO: kick full dynticks CPUs when
179 * sched_clock_stable is set.
181 if (!sched_clock_stable) {
182 trace_tick_stop(0, "unstable sched clock\n");
184 * Don't allow the user to think they can get
185 * full NO_HZ with this machine.
187 WARN_ONCE(tick_nohz_full_running,
188 "NO_HZ FULL will not work with unstable sched clock");
196 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
199 * Re-evaluate the need for the tick on the current CPU
200 * and restart it if necessary.
202 void __tick_nohz_full_check(void)
204 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
206 if (tick_nohz_full_cpu(smp_processor_id())) {
207 if (ts->tick_stopped && !is_idle_task(current)) {
208 if (!can_stop_full_tick())
209 tick_nohz_restart_sched_tick(ts, ktime_get());
214 static void nohz_full_kick_work_func(struct irq_work *work)
216 __tick_nohz_full_check();
219 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
220 .func = nohz_full_kick_work_func,
224 * Kick the current CPU if it's full dynticks in order to force it to
225 * re-evaluate its dependency on the tick and restart it if necessary.
227 void tick_nohz_full_kick(void)
229 if (tick_nohz_full_cpu(smp_processor_id()))
230 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
233 static void nohz_full_kick_ipi(void *info)
235 __tick_nohz_full_check();
239 * Kick all full dynticks CPUs in order to force these to re-evaluate
240 * their dependency on the tick and restart it if necessary.
242 void tick_nohz_full_kick_all(void)
244 if (!tick_nohz_full_running)
248 smp_call_function_many(tick_nohz_full_mask,
249 nohz_full_kick_ipi, NULL, false);
250 tick_nohz_full_kick();
255 * Re-evaluate the need for the tick as we switch the current task.
256 * It might need the tick due to per task/process properties:
257 * perf events, posix cpu timers, ...
259 void __tick_nohz_task_switch(struct task_struct *tsk)
263 local_irq_save(flags);
265 if (!tick_nohz_full_cpu(smp_processor_id()))
268 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
269 tick_nohz_full_kick();
272 local_irq_restore(flags);
275 /* Parse the boot-time nohz CPU list from the kernel parameters. */
276 static int __init tick_nohz_full_setup(char *str)
280 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
281 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
282 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
286 cpu = smp_processor_id();
287 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
288 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
289 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
291 tick_nohz_full_running = true;
295 __setup("nohz_full=", tick_nohz_full_setup);
297 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
298 unsigned long action,
301 unsigned int cpu = (unsigned long)hcpu;
303 switch (action & ~CPU_TASKS_FROZEN) {
304 case CPU_DOWN_PREPARE:
306 * If we handle the timekeeping duty for full dynticks CPUs,
307 * we can't safely shutdown that CPU.
309 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
317 * Worst case string length in chunks of CPU range seems 2 steps
318 * separations: 0,2,4,6,...
319 * This is NR_CPUS + sizeof('\0')
321 static char __initdata nohz_full_buf[NR_CPUS + 1];
323 static int tick_nohz_init_all(void)
327 #ifdef CONFIG_NO_HZ_FULL_ALL
328 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
329 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
333 cpumask_setall(tick_nohz_full_mask);
334 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask);
335 tick_nohz_full_running = true;
340 void __init tick_nohz_init(void)
344 if (!tick_nohz_full_running) {
345 if (tick_nohz_init_all() < 0)
349 for_each_cpu(cpu, tick_nohz_full_mask)
350 context_tracking_cpu_set(cpu);
352 cpu_notifier(tick_nohz_cpu_down_callback, 0);
353 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
354 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
359 * NOHZ - aka dynamic tick functionality
361 #ifdef CONFIG_NO_HZ_COMMON
365 int tick_nohz_enabled __read_mostly = 1;
368 * Enable / Disable tickless mode
370 static int __init setup_tick_nohz(char *str)
372 if (!strcmp(str, "off"))
373 tick_nohz_enabled = 0;
374 else if (!strcmp(str, "on"))
375 tick_nohz_enabled = 1;
381 __setup("nohz=", setup_tick_nohz);
384 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
386 * Called from interrupt entry when the CPU was idle
388 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
389 * must be updated. Otherwise an interrupt handler could use a stale jiffy
390 * value. We do this unconditionally on any cpu, as we don't know whether the
391 * cpu, which has the update task assigned is in a long sleep.
393 static void tick_nohz_update_jiffies(ktime_t now)
397 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
399 local_irq_save(flags);
400 tick_do_update_jiffies64(now);
401 local_irq_restore(flags);
403 touch_softlockup_watchdog();
407 * Updates the per cpu time idle statistics counters
410 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
414 if (ts->idle_active) {
415 delta = ktime_sub(now, ts->idle_entrytime);
416 if (nr_iowait_cpu(cpu) > 0)
417 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
419 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
420 ts->idle_entrytime = now;
423 if (last_update_time)
424 *last_update_time = ktime_to_us(now);
428 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
430 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
433 sched_clock_idle_wakeup_event(0);
436 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
438 ktime_t now = ktime_get();
440 ts->idle_entrytime = now;
442 sched_clock_idle_sleep_event();
447 * get_cpu_idle_time_us - get the total idle time of a cpu
448 * @cpu: CPU number to query
449 * @last_update_time: variable to store update time in. Do not update
452 * Return the cummulative idle time (since boot) for a given
453 * CPU, in microseconds.
455 * This time is measured via accounting rather than sampling,
456 * and is as accurate as ktime_get() is.
458 * This function returns -1 if NOHZ is not enabled.
460 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
462 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
465 if (!tick_nohz_enabled)
469 if (last_update_time) {
470 update_ts_time_stats(cpu, ts, now, last_update_time);
471 idle = ts->idle_sleeptime;
473 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
474 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
476 idle = ktime_add(ts->idle_sleeptime, delta);
478 idle = ts->idle_sleeptime;
482 return ktime_to_us(idle);
485 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
488 * get_cpu_iowait_time_us - get the total iowait time of a cpu
489 * @cpu: CPU number to query
490 * @last_update_time: variable to store update time in. Do not update
493 * Return the cummulative iowait time (since boot) for a given
494 * CPU, in microseconds.
496 * This time is measured via accounting rather than sampling,
497 * and is as accurate as ktime_get() is.
499 * This function returns -1 if NOHZ is not enabled.
501 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
503 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
506 if (!tick_nohz_enabled)
510 if (last_update_time) {
511 update_ts_time_stats(cpu, ts, now, last_update_time);
512 iowait = ts->iowait_sleeptime;
514 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
515 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
517 iowait = ktime_add(ts->iowait_sleeptime, delta);
519 iowait = ts->iowait_sleeptime;
523 return ktime_to_us(iowait);
525 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
527 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
528 ktime_t now, int cpu)
530 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
531 ktime_t last_update, expires, ret = { .tv64 = 0 };
532 unsigned long rcu_delta_jiffies;
533 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
536 /* Read jiffies and the time when jiffies were updated last */
538 seq = read_seqbegin(&jiffies_lock);
539 last_update = last_jiffies_update;
540 last_jiffies = jiffies;
541 time_delta = timekeeping_max_deferment();
542 } while (read_seqretry(&jiffies_lock, seq));
544 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
545 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
546 next_jiffies = last_jiffies + 1;
549 /* Get the next timer wheel timer */
550 next_jiffies = get_next_timer_interrupt(last_jiffies);
551 delta_jiffies = next_jiffies - last_jiffies;
552 if (rcu_delta_jiffies < delta_jiffies) {
553 next_jiffies = last_jiffies + rcu_delta_jiffies;
554 delta_jiffies = rcu_delta_jiffies;
559 * Do not stop the tick, if we are only one off (or less)
560 * or if the cpu is required for RCU:
562 if (!ts->tick_stopped && delta_jiffies <= 1)
565 /* Schedule the tick, if we are at least one jiffie off */
566 if ((long)delta_jiffies >= 1) {
569 * If this cpu is the one which updates jiffies, then
570 * give up the assignment and let it be taken by the
571 * cpu which runs the tick timer next, which might be
572 * this cpu as well. If we don't drop this here the
573 * jiffies might be stale and do_timer() never
574 * invoked. Keep track of the fact that it was the one
575 * which had the do_timer() duty last. If this cpu is
576 * the one which had the do_timer() duty last, we
577 * limit the sleep time to the timekeeping
578 * max_deferement value which we retrieved
579 * above. Otherwise we can sleep as long as we want.
581 if (cpu == tick_do_timer_cpu) {
582 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
583 ts->do_timer_last = 1;
584 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
585 time_delta = KTIME_MAX;
586 ts->do_timer_last = 0;
587 } else if (!ts->do_timer_last) {
588 time_delta = KTIME_MAX;
591 #ifdef CONFIG_NO_HZ_FULL
593 time_delta = min(time_delta,
594 scheduler_tick_max_deferment());
599 * calculate the expiry time for the next timer wheel
600 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
601 * that there is no timer pending or at least extremely
602 * far into the future (12 days for HZ=1000). In this
603 * case we set the expiry to the end of time.
605 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
607 * Calculate the time delta for the next timer event.
608 * If the time delta exceeds the maximum time delta
609 * permitted by the current clocksource then adjust
610 * the time delta accordingly to ensure the
611 * clocksource does not wrap.
613 time_delta = min_t(u64, time_delta,
614 tick_period.tv64 * delta_jiffies);
617 if (time_delta < KTIME_MAX)
618 expires = ktime_add_ns(last_update, time_delta);
620 expires.tv64 = KTIME_MAX;
622 /* Skip reprogram of event if its not changed */
623 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
629 * nohz_stop_sched_tick can be called several times before
630 * the nohz_restart_sched_tick is called. This happens when
631 * interrupts arrive which do not cause a reschedule. In the
632 * first call we save the current tick time, so we can restart
633 * the scheduler tick in nohz_restart_sched_tick.
635 if (!ts->tick_stopped) {
636 nohz_balance_enter_idle(cpu);
637 calc_load_enter_idle();
639 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
640 ts->tick_stopped = 1;
641 trace_tick_stop(1, " ");
645 * If the expiration time == KTIME_MAX, then
646 * in this case we simply stop the tick timer.
648 if (unlikely(expires.tv64 == KTIME_MAX)) {
649 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
650 hrtimer_cancel(&ts->sched_timer);
654 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
655 hrtimer_start(&ts->sched_timer, expires,
656 HRTIMER_MODE_ABS_PINNED);
657 /* Check, if the timer was already in the past */
658 if (hrtimer_active(&ts->sched_timer))
660 } else if (!tick_program_event(expires, 0))
663 * We are past the event already. So we crossed a
664 * jiffie boundary. Update jiffies and raise the
667 tick_do_update_jiffies64(ktime_get());
669 raise_softirq_irqoff(TIMER_SOFTIRQ);
671 ts->next_jiffies = next_jiffies;
672 ts->last_jiffies = last_jiffies;
673 ts->sleep_length = ktime_sub(dev->next_event, now);
678 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
680 #ifdef CONFIG_NO_HZ_FULL
681 int cpu = smp_processor_id();
683 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
686 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
689 if (!can_stop_full_tick())
692 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
696 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
699 * If this cpu is offline and it is the one which updates
700 * jiffies, then give up the assignment and let it be taken by
701 * the cpu which runs the tick timer next. If we don't drop
702 * this here the jiffies might be stale and do_timer() never
705 if (unlikely(!cpu_online(cpu))) {
706 if (cpu == tick_do_timer_cpu)
707 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
711 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
717 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
718 static int ratelimit;
720 if (ratelimit < 10 &&
721 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
722 pr_warn("NOHZ: local_softirq_pending %02x\n",
723 (unsigned int) local_softirq_pending());
729 if (tick_nohz_full_enabled()) {
731 * Keep the tick alive to guarantee timekeeping progression
732 * if there are full dynticks CPUs around
734 if (tick_do_timer_cpu == cpu)
737 * Boot safety: make sure the timekeeping duty has been
738 * assigned before entering dyntick-idle mode,
740 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
747 static void __tick_nohz_idle_enter(struct tick_sched *ts)
749 ktime_t now, expires;
750 int cpu = smp_processor_id();
752 now = tick_nohz_start_idle(ts);
754 if (can_stop_idle_tick(cpu, ts)) {
755 int was_stopped = ts->tick_stopped;
759 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
760 if (expires.tv64 > 0LL) {
762 ts->idle_expires = expires;
765 if (!was_stopped && ts->tick_stopped)
766 ts->idle_jiffies = ts->last_jiffies;
771 * tick_nohz_idle_enter - stop the idle tick from the idle task
773 * When the next event is more than a tick into the future, stop the idle tick
774 * Called when we start the idle loop.
776 * The arch is responsible of calling:
778 * - rcu_idle_enter() after its last use of RCU before the CPU is put
780 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
782 void tick_nohz_idle_enter(void)
784 struct tick_sched *ts;
786 WARN_ON_ONCE(irqs_disabled());
789 * Update the idle state in the scheduler domain hierarchy
790 * when tick_nohz_stop_sched_tick() is called from the idle loop.
791 * State will be updated to busy during the first busy tick after
794 set_cpu_sd_state_idle();
798 ts = &__get_cpu_var(tick_cpu_sched);
800 * set ts->inidle unconditionally. even if the system did not
801 * switch to nohz mode the cpu frequency governers rely on the
802 * update of the idle time accounting in tick_nohz_start_idle().
805 __tick_nohz_idle_enter(ts);
809 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
812 * tick_nohz_irq_exit - update next tick event from interrupt exit
814 * When an interrupt fires while we are idle and it doesn't cause
815 * a reschedule, it may still add, modify or delete a timer, enqueue
816 * an RCU callback, etc...
817 * So we need to re-calculate and reprogram the next tick event.
819 void tick_nohz_irq_exit(void)
821 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
824 __tick_nohz_idle_enter(ts);
826 tick_nohz_full_stop_tick(ts);
830 * tick_nohz_get_sleep_length - return the length of the current sleep
832 * Called from power state control code with interrupts disabled
834 ktime_t tick_nohz_get_sleep_length(void)
836 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
838 return ts->sleep_length;
841 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
843 hrtimer_cancel(&ts->sched_timer);
844 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
847 /* Forward the time to expire in the future */
848 hrtimer_forward(&ts->sched_timer, now, tick_period);
850 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
851 hrtimer_start_expires(&ts->sched_timer,
852 HRTIMER_MODE_ABS_PINNED);
853 /* Check, if the timer was already in the past */
854 if (hrtimer_active(&ts->sched_timer))
857 if (!tick_program_event(
858 hrtimer_get_expires(&ts->sched_timer), 0))
861 /* Reread time and update jiffies */
863 tick_do_update_jiffies64(now);
867 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
869 /* Update jiffies first */
870 tick_do_update_jiffies64(now);
871 update_cpu_load_nohz();
873 calc_load_exit_idle();
874 touch_softlockup_watchdog();
876 * Cancel the scheduled timer and restore the tick
878 ts->tick_stopped = 0;
879 ts->idle_exittime = now;
881 tick_nohz_restart(ts, now);
884 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
886 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
889 if (vtime_accounting_enabled())
892 * We stopped the tick in idle. Update process times would miss the
893 * time we slept as update_process_times does only a 1 tick
894 * accounting. Enforce that this is accounted to idle !
896 ticks = jiffies - ts->idle_jiffies;
898 * We might be one off. Do not randomly account a huge number of ticks!
900 if (ticks && ticks < LONG_MAX)
901 account_idle_ticks(ticks);
906 * tick_nohz_idle_exit - restart the idle tick from the idle task
908 * Restart the idle tick when the CPU is woken up from idle
909 * This also exit the RCU extended quiescent state. The CPU
910 * can use RCU again after this function is called.
912 void tick_nohz_idle_exit(void)
914 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
919 WARN_ON_ONCE(!ts->inidle);
923 if (ts->idle_active || ts->tick_stopped)
927 tick_nohz_stop_idle(ts, now);
929 if (ts->tick_stopped) {
930 tick_nohz_restart_sched_tick(ts, now);
931 tick_nohz_account_idle_ticks(ts);
936 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
938 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
940 hrtimer_forward(&ts->sched_timer, now, tick_period);
941 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
945 * The nohz low res interrupt handler
947 static void tick_nohz_handler(struct clock_event_device *dev)
949 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
950 struct pt_regs *regs = get_irq_regs();
951 ktime_t now = ktime_get();
953 dev->next_event.tv64 = KTIME_MAX;
955 tick_sched_do_timer(now);
956 tick_sched_handle(ts, regs);
958 while (tick_nohz_reprogram(ts, now)) {
960 tick_do_update_jiffies64(now);
965 * tick_nohz_switch_to_nohz - switch to nohz mode
967 static void tick_nohz_switch_to_nohz(void)
969 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
972 if (!tick_nohz_enabled)
976 if (tick_switch_to_oneshot(tick_nohz_handler)) {
981 ts->nohz_mode = NOHZ_MODE_LOWRES;
984 * Recycle the hrtimer in ts, so we can share the
985 * hrtimer_forward with the highres code.
987 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
988 /* Get the next period */
989 next = tick_init_jiffy_update();
992 hrtimer_set_expires(&ts->sched_timer, next);
993 if (!tick_program_event(next, 0))
995 next = ktime_add(next, tick_period);
1001 * When NOHZ is enabled and the tick is stopped, we need to kick the
1002 * tick timer from irq_enter() so that the jiffies update is kept
1003 * alive during long running softirqs. That's ugly as hell, but
1004 * correctness is key even if we need to fix the offending softirq in
1007 * Note, this is different to tick_nohz_restart. We just kick the
1008 * timer and do not touch the other magic bits which need to be done
1009 * when idle is left.
1011 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1014 /* Switch back to 2.6.27 behaviour */
1018 * Do not touch the tick device, when the next expiry is either
1019 * already reached or less/equal than the tick period.
1021 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1022 if (delta.tv64 <= tick_period.tv64)
1025 tick_nohz_restart(ts, now);
1029 static inline void tick_check_nohz_this_cpu(void)
1031 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1034 if (!ts->idle_active && !ts->tick_stopped)
1037 if (ts->idle_active)
1038 tick_nohz_stop_idle(ts, now);
1039 if (ts->tick_stopped) {
1040 tick_nohz_update_jiffies(now);
1041 tick_nohz_kick_tick(ts, now);
1047 static inline void tick_nohz_switch_to_nohz(void) { }
1048 static inline void tick_check_nohz_this_cpu(void) { }
1050 #endif /* CONFIG_NO_HZ_COMMON */
1053 * Called from irq_enter to notify about the possible interruption of idle()
1055 void tick_check_idle(void)
1057 tick_check_oneshot_broadcast_this_cpu();
1058 tick_check_nohz_this_cpu();
1062 * High resolution timer specific code
1064 #ifdef CONFIG_HIGH_RES_TIMERS
1066 * We rearm the timer until we get disabled by the idle code.
1067 * Called with interrupts disabled.
1069 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1071 struct tick_sched *ts =
1072 container_of(timer, struct tick_sched, sched_timer);
1073 struct pt_regs *regs = get_irq_regs();
1074 ktime_t now = ktime_get();
1076 tick_sched_do_timer(now);
1079 * Do not call, when we are not in irq context and have
1080 * no valid regs pointer
1083 tick_sched_handle(ts, regs);
1085 hrtimer_forward(timer, now, tick_period);
1087 return HRTIMER_RESTART;
1090 static int sched_skew_tick;
1092 static int __init skew_tick(char *str)
1094 get_option(&str, &sched_skew_tick);
1098 early_param("skew_tick", skew_tick);
1101 * tick_setup_sched_timer - setup the tick emulation timer
1103 void tick_setup_sched_timer(void)
1105 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1106 ktime_t now = ktime_get();
1109 * Emulate tick processing via per-CPU hrtimers:
1111 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1112 ts->sched_timer.function = tick_sched_timer;
1114 /* Get the next period (per cpu) */
1115 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1117 /* Offset the tick to avert jiffies_lock contention. */
1118 if (sched_skew_tick) {
1119 u64 offset = ktime_to_ns(tick_period) >> 1;
1120 do_div(offset, num_possible_cpus());
1121 offset *= smp_processor_id();
1122 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1126 hrtimer_forward(&ts->sched_timer, now, tick_period);
1127 hrtimer_start_expires(&ts->sched_timer,
1128 HRTIMER_MODE_ABS_PINNED);
1129 /* Check, if the timer was already in the past */
1130 if (hrtimer_active(&ts->sched_timer))
1135 #ifdef CONFIG_NO_HZ_COMMON
1136 if (tick_nohz_enabled)
1137 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1140 #endif /* HIGH_RES_TIMERS */
1142 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1143 void tick_cancel_sched_timer(int cpu)
1145 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1147 # ifdef CONFIG_HIGH_RES_TIMERS
1148 if (ts->sched_timer.base)
1149 hrtimer_cancel(&ts->sched_timer);
1152 memset(ts, 0, sizeof(*ts));
1157 * Async notification about clocksource changes
1159 void tick_clock_notify(void)
1163 for_each_possible_cpu(cpu)
1164 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1168 * Async notification about clock event changes
1170 void tick_oneshot_notify(void)
1172 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1174 set_bit(0, &ts->check_clocks);
1178 * Check, if a change happened, which makes oneshot possible.
1180 * Called cyclic from the hrtimer softirq (driven by the timer
1181 * softirq) allow_nohz signals, that we can switch into low-res nohz
1182 * mode, because high resolution timers are disabled (either compile
1185 int tick_check_oneshot_change(int allow_nohz)
1187 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1189 if (!test_and_clear_bit(0, &ts->check_clocks))
1192 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1195 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1201 tick_nohz_switch_to_nohz();