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/nmi.h>
21 #include <linux/profile.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/clock.h>
24 #include <linux/sched/stat.h>
25 #include <linux/sched/nohz.h>
26 #include <linux/module.h>
27 #include <linux/irq_work.h>
28 #include <linux/posix-timers.h>
29 #include <linux/context_tracking.h>
32 #include <asm/irq_regs.h>
34 #include "tick-internal.h"
36 #include <trace/events/timer.h>
39 * Per-CPU nohz control structure
41 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
43 struct tick_sched *tick_get_tick_sched(int cpu)
45 return &per_cpu(tick_cpu_sched, cpu);
48 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
50 * The time, when the last jiffy update happened. Protected by jiffies_lock.
52 static ktime_t last_jiffies_update;
55 * Must be called with interrupts disabled !
57 static void tick_do_update_jiffies64(ktime_t now)
59 unsigned long ticks = 0;
63 * Do a quick check without holding jiffies_lock:
65 delta = ktime_sub(now, last_jiffies_update);
66 if (delta < tick_period)
69 /* Reevaluate with jiffies_lock held */
70 write_seqlock(&jiffies_lock);
72 delta = ktime_sub(now, last_jiffies_update);
73 if (delta >= tick_period) {
75 delta = ktime_sub(delta, tick_period);
76 last_jiffies_update = ktime_add(last_jiffies_update,
79 /* Slow path for long timeouts */
80 if (unlikely(delta >= tick_period)) {
81 s64 incr = ktime_to_ns(tick_period);
83 ticks = ktime_divns(delta, incr);
85 last_jiffies_update = ktime_add_ns(last_jiffies_update,
90 /* Keep the tick_next_period variable up to date */
91 tick_next_period = ktime_add(last_jiffies_update, tick_period);
93 write_sequnlock(&jiffies_lock);
96 write_sequnlock(&jiffies_lock);
101 * Initialize and return retrieve the jiffies update.
103 static ktime_t tick_init_jiffy_update(void)
107 write_seqlock(&jiffies_lock);
108 /* Did we start the jiffies update yet ? */
109 if (last_jiffies_update == 0)
110 last_jiffies_update = tick_next_period;
111 period = last_jiffies_update;
112 write_sequnlock(&jiffies_lock);
117 static void tick_sched_do_timer(ktime_t now)
119 int cpu = smp_processor_id();
121 #ifdef CONFIG_NO_HZ_COMMON
123 * Check if the do_timer duty was dropped. We don't care about
124 * concurrency: This happens only when the CPU in charge went
125 * into a long sleep. If two CPUs happen to assign themselves to
126 * this duty, then the jiffies update is still serialized by
129 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
130 && !tick_nohz_full_cpu(cpu))
131 tick_do_timer_cpu = cpu;
134 /* Check, if the jiffies need an update */
135 if (tick_do_timer_cpu == cpu)
136 tick_do_update_jiffies64(now);
139 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
141 #ifdef CONFIG_NO_HZ_COMMON
143 * When we are idle and the tick is stopped, we have to touch
144 * the watchdog as we might not schedule for a really long
145 * time. This happens on complete idle SMP systems while
146 * waiting on the login prompt. We also increment the "start of
147 * idle" jiffy stamp so the idle accounting adjustment we do
148 * when we go busy again does not account too much ticks.
150 if (ts->tick_stopped) {
151 touch_softlockup_watchdog_sched();
152 if (is_idle_task(current))
155 * In case the current tick fired too early past its expected
156 * expiration, make sure we don't bypass the next clock reprogramming
157 * to the same deadline.
162 update_process_times(user_mode(regs));
163 profile_tick(CPU_PROFILING);
167 #ifdef CONFIG_NO_HZ_FULL
168 cpumask_var_t tick_nohz_full_mask;
169 bool tick_nohz_full_running;
170 static atomic_t tick_dep_mask;
172 static bool check_tick_dependency(atomic_t *dep)
174 int val = atomic_read(dep);
176 if (val & TICK_DEP_MASK_POSIX_TIMER) {
177 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
181 if (val & TICK_DEP_MASK_PERF_EVENTS) {
182 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
186 if (val & TICK_DEP_MASK_SCHED) {
187 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
191 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
192 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
199 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
201 WARN_ON_ONCE(!irqs_disabled());
203 if (unlikely(!cpu_online(cpu)))
206 if (check_tick_dependency(&tick_dep_mask))
209 if (check_tick_dependency(&ts->tick_dep_mask))
212 if (check_tick_dependency(¤t->tick_dep_mask))
215 if (check_tick_dependency(¤t->signal->tick_dep_mask))
221 static void nohz_full_kick_func(struct irq_work *work)
223 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
226 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
227 .func = nohz_full_kick_func,
231 * Kick this CPU if it's full dynticks in order to force it to
232 * re-evaluate its dependency on the tick and restart it if necessary.
233 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
236 static void tick_nohz_full_kick(void)
238 if (!tick_nohz_full_cpu(smp_processor_id()))
241 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
245 * Kick the CPU if it's full dynticks in order to force it to
246 * re-evaluate its dependency on the tick and restart it if necessary.
248 void tick_nohz_full_kick_cpu(int cpu)
250 if (!tick_nohz_full_cpu(cpu))
253 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
257 * Kick all full dynticks CPUs in order to force these to re-evaluate
258 * their dependency on the tick and restart it if necessary.
260 static void tick_nohz_full_kick_all(void)
264 if (!tick_nohz_full_running)
268 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
269 tick_nohz_full_kick_cpu(cpu);
273 static void tick_nohz_dep_set_all(atomic_t *dep,
274 enum tick_dep_bits bit)
278 prev = atomic_fetch_or(BIT(bit), dep);
280 tick_nohz_full_kick_all();
284 * Set a global tick dependency. Used by perf events that rely on freq and
287 void tick_nohz_dep_set(enum tick_dep_bits bit)
289 tick_nohz_dep_set_all(&tick_dep_mask, bit);
292 void tick_nohz_dep_clear(enum tick_dep_bits bit)
294 atomic_andnot(BIT(bit), &tick_dep_mask);
298 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
299 * manage events throttling.
301 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
304 struct tick_sched *ts;
306 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
308 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
311 /* Perf needs local kick that is NMI safe */
312 if (cpu == smp_processor_id()) {
313 tick_nohz_full_kick();
315 /* Remote irq work not NMI-safe */
316 if (!WARN_ON_ONCE(in_nmi()))
317 tick_nohz_full_kick_cpu(cpu);
323 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
325 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
327 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
331 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
334 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
337 * We could optimize this with just kicking the target running the task
338 * if that noise matters for nohz full users.
340 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
343 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
345 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
349 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
350 * per process timers.
352 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
354 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
357 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
359 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
363 * Re-evaluate the need for the tick as we switch the current task.
364 * It might need the tick due to per task/process properties:
365 * perf events, posix CPU timers, ...
367 void __tick_nohz_task_switch(void)
370 struct tick_sched *ts;
372 local_irq_save(flags);
374 if (!tick_nohz_full_cpu(smp_processor_id()))
377 ts = this_cpu_ptr(&tick_cpu_sched);
379 if (ts->tick_stopped) {
380 if (atomic_read(¤t->tick_dep_mask) ||
381 atomic_read(¤t->signal->tick_dep_mask))
382 tick_nohz_full_kick();
385 local_irq_restore(flags);
388 /* Parse the boot-time nohz CPU list from the kernel parameters. */
389 static int __init tick_nohz_full_setup(char *str)
391 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
392 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
393 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
394 free_bootmem_cpumask_var(tick_nohz_full_mask);
397 tick_nohz_full_running = true;
401 __setup("nohz_full=", tick_nohz_full_setup);
403 static int tick_nohz_cpu_down(unsigned int cpu)
406 * The boot CPU handles housekeeping duty (unbound timers,
407 * workqueues, timekeeping, ...) on behalf of full dynticks
408 * CPUs. It must remain online when nohz full is enabled.
410 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
415 static int tick_nohz_init_all(void)
419 #ifdef CONFIG_NO_HZ_FULL_ALL
420 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
421 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
425 cpumask_setall(tick_nohz_full_mask);
426 tick_nohz_full_running = true;
431 void __init tick_nohz_init(void)
435 if (!tick_nohz_full_running) {
436 if (tick_nohz_init_all() < 0)
441 * Full dynticks uses irq work to drive the tick rescheduling on safe
442 * locking contexts. But then we need irq work to raise its own
443 * interrupts to avoid circular dependency on the tick
445 if (!arch_irq_work_has_interrupt()) {
446 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
447 cpumask_clear(tick_nohz_full_mask);
448 tick_nohz_full_running = false;
452 cpu = smp_processor_id();
454 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
455 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
457 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
460 for_each_cpu(cpu, tick_nohz_full_mask)
461 context_tracking_cpu_set(cpu);
463 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
464 "kernel/nohz:predown", NULL,
467 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
468 cpumask_pr_args(tick_nohz_full_mask));
473 * NOHZ - aka dynamic tick functionality
475 #ifdef CONFIG_NO_HZ_COMMON
479 bool tick_nohz_enabled __read_mostly = true;
480 unsigned long tick_nohz_active __read_mostly;
482 * Enable / Disable tickless mode
484 static int __init setup_tick_nohz(char *str)
486 return (kstrtobool(str, &tick_nohz_enabled) == 0);
489 __setup("nohz=", setup_tick_nohz);
491 int tick_nohz_tick_stopped(void)
493 return __this_cpu_read(tick_cpu_sched.tick_stopped);
497 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
499 * Called from interrupt entry when the CPU was idle
501 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
502 * must be updated. Otherwise an interrupt handler could use a stale jiffy
503 * value. We do this unconditionally on any CPU, as we don't know whether the
504 * CPU, which has the update task assigned is in a long sleep.
506 static void tick_nohz_update_jiffies(ktime_t now)
510 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
512 local_irq_save(flags);
513 tick_do_update_jiffies64(now);
514 local_irq_restore(flags);
516 touch_softlockup_watchdog_sched();
520 * Updates the per-CPU time idle statistics counters
523 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
527 if (ts->idle_active) {
528 delta = ktime_sub(now, ts->idle_entrytime);
529 if (nr_iowait_cpu(cpu) > 0)
530 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
532 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
533 ts->idle_entrytime = now;
536 if (last_update_time)
537 *last_update_time = ktime_to_us(now);
541 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
543 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
546 sched_clock_idle_wakeup_event();
549 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
551 ktime_t now = ktime_get();
553 ts->idle_entrytime = now;
555 sched_clock_idle_sleep_event();
560 * get_cpu_idle_time_us - get the total idle time of a CPU
561 * @cpu: CPU number to query
562 * @last_update_time: variable to store update time in. Do not update
565 * Return the cumulative idle time (since boot) for a given
566 * CPU, in microseconds.
568 * This time is measured via accounting rather than sampling,
569 * and is as accurate as ktime_get() is.
571 * This function returns -1 if NOHZ is not enabled.
573 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
575 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
578 if (!tick_nohz_active)
582 if (last_update_time) {
583 update_ts_time_stats(cpu, ts, now, last_update_time);
584 idle = ts->idle_sleeptime;
586 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
587 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
589 idle = ktime_add(ts->idle_sleeptime, delta);
591 idle = ts->idle_sleeptime;
595 return ktime_to_us(idle);
598 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
601 * get_cpu_iowait_time_us - get the total iowait time of a CPU
602 * @cpu: CPU number to query
603 * @last_update_time: variable to store update time in. Do not update
606 * Return the cumulative iowait time (since boot) for a given
607 * CPU, in microseconds.
609 * This time is measured via accounting rather than sampling,
610 * and is as accurate as ktime_get() is.
612 * This function returns -1 if NOHZ is not enabled.
614 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
616 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
619 if (!tick_nohz_active)
623 if (last_update_time) {
624 update_ts_time_stats(cpu, ts, now, last_update_time);
625 iowait = ts->iowait_sleeptime;
627 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
628 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
630 iowait = ktime_add(ts->iowait_sleeptime, delta);
632 iowait = ts->iowait_sleeptime;
636 return ktime_to_us(iowait);
638 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
640 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
642 hrtimer_cancel(&ts->sched_timer);
643 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
645 /* Forward the time to expire in the future */
646 hrtimer_forward(&ts->sched_timer, now, tick_period);
648 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
649 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
651 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
654 * Reset to make sure next tick stop doesn't get fooled by past
655 * cached clock deadline.
660 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
661 ktime_t now, int cpu)
663 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
664 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
665 unsigned long seq, basejiff;
668 /* Read jiffies and the time when jiffies were updated last */
670 seq = read_seqbegin(&jiffies_lock);
671 basemono = last_jiffies_update;
673 } while (read_seqretry(&jiffies_lock, seq));
674 ts->last_jiffies = basejiff;
676 if (rcu_needs_cpu(basemono, &next_rcu) ||
677 arch_needs_cpu() || irq_work_needs_cpu()) {
678 next_tick = basemono + TICK_NSEC;
681 * Get the next pending timer. If high resolution
682 * timers are enabled this only takes the timer wheel
683 * timers into account. If high resolution timers are
684 * disabled this also looks at the next expiring
687 next_tmr = get_next_timer_interrupt(basejiff, basemono);
688 ts->next_timer = next_tmr;
689 /* Take the next rcu event into account */
690 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
694 * If the tick is due in the next period, keep it ticking or
695 * force prod the timer.
697 delta = next_tick - basemono;
698 if (delta <= (u64)TICK_NSEC) {
700 * Tell the timer code that the base is not idle, i.e. undo
701 * the effect of get_next_timer_interrupt():
705 * We've not stopped the tick yet, and there's a timer in the
706 * next period, so no point in stopping it either, bail.
708 if (!ts->tick_stopped) {
715 * If this CPU is the one which updates jiffies, then give up
716 * the assignment and let it be taken by the CPU which runs
717 * the tick timer next, which might be this CPU as well. If we
718 * don't drop this here the jiffies might be stale and
719 * do_timer() never invoked. Keep track of the fact that it
720 * was the one which had the do_timer() duty last. If this CPU
721 * is the one which had the do_timer() duty last, we limit the
722 * sleep time to the timekeeping max_deferment value.
723 * Otherwise we can sleep as long as we want.
725 delta = timekeeping_max_deferment();
726 if (cpu == tick_do_timer_cpu) {
727 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
728 ts->do_timer_last = 1;
729 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
731 ts->do_timer_last = 0;
732 } else if (!ts->do_timer_last) {
736 #ifdef CONFIG_NO_HZ_FULL
737 /* Limit the tick delta to the maximum scheduler deferment */
739 delta = min(delta, scheduler_tick_max_deferment());
742 /* Calculate the next expiry time */
743 if (delta < (KTIME_MAX - basemono))
744 expires = basemono + delta;
748 expires = min_t(u64, expires, next_tick);
751 /* Skip reprogram of event if its not changed */
752 if (ts->tick_stopped && (expires == ts->next_tick)) {
753 /* Sanity check: make sure clockevent is actually programmed */
754 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
758 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
759 basemono, ts->next_tick, dev->next_event,
760 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
764 * nohz_stop_sched_tick can be called several times before
765 * the nohz_restart_sched_tick is called. This happens when
766 * interrupts arrive which do not cause a reschedule. In the
767 * first call we save the current tick time, so we can restart
768 * the scheduler tick in nohz_restart_sched_tick.
770 if (!ts->tick_stopped) {
771 calc_load_nohz_start();
772 cpu_load_update_nohz_start();
775 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
776 ts->tick_stopped = 1;
777 trace_tick_stop(1, TICK_DEP_MASK_NONE);
780 ts->next_tick = tick;
783 * If the expiration time == KTIME_MAX, then we simply stop
786 if (unlikely(expires == KTIME_MAX)) {
787 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
788 hrtimer_cancel(&ts->sched_timer);
792 hrtimer_set_expires(&ts->sched_timer, tick);
794 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
795 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
797 tick_program_event(tick, 1);
800 * Update the estimated sleep length until the next timer
801 * (not only the tick).
803 ts->sleep_length = ktime_sub(dev->next_event, now);
807 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
809 /* Update jiffies first */
810 tick_do_update_jiffies64(now);
811 cpu_load_update_nohz_stop();
814 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
815 * the clock forward checks in the enqueue path:
819 calc_load_nohz_stop();
820 touch_softlockup_watchdog_sched();
822 * Cancel the scheduled timer and restore the tick
824 ts->tick_stopped = 0;
825 ts->idle_exittime = now;
827 tick_nohz_restart(ts, now);
830 static void tick_nohz_full_update_tick(struct tick_sched *ts)
832 #ifdef CONFIG_NO_HZ_FULL
833 int cpu = smp_processor_id();
835 if (!tick_nohz_full_cpu(cpu))
838 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
841 if (can_stop_full_tick(cpu, ts))
842 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
843 else if (ts->tick_stopped)
844 tick_nohz_restart_sched_tick(ts, ktime_get());
848 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
851 * If this CPU is offline and it is the one which updates
852 * jiffies, then give up the assignment and let it be taken by
853 * the CPU which runs the tick timer next. If we don't drop
854 * this here the jiffies might be stale and do_timer() never
857 if (unlikely(!cpu_online(cpu))) {
858 if (cpu == tick_do_timer_cpu)
859 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
861 * Make sure the CPU doesn't get fooled by obsolete tick
862 * deadline if it comes back online later.
868 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
869 ts->sleep_length = NSEC_PER_SEC / HZ;
876 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
877 static int ratelimit;
879 if (ratelimit < 10 &&
880 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
881 pr_warn("NOHZ: local_softirq_pending %02x\n",
882 (unsigned int) local_softirq_pending());
888 if (tick_nohz_full_enabled()) {
890 * Keep the tick alive to guarantee timekeeping progression
891 * if there are full dynticks CPUs around
893 if (tick_do_timer_cpu == cpu)
896 * Boot safety: make sure the timekeeping duty has been
897 * assigned before entering dyntick-idle mode,
899 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
906 static void __tick_nohz_idle_enter(struct tick_sched *ts)
908 ktime_t now, expires;
909 int cpu = smp_processor_id();
911 now = tick_nohz_start_idle(ts);
913 if (can_stop_idle_tick(cpu, ts)) {
914 int was_stopped = ts->tick_stopped;
918 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
921 ts->idle_expires = expires;
924 if (!was_stopped && ts->tick_stopped) {
925 ts->idle_jiffies = ts->last_jiffies;
926 nohz_balance_enter_idle(cpu);
932 * tick_nohz_idle_enter - stop the idle tick from the idle task
934 * When the next event is more than a tick into the future, stop the idle tick
935 * Called when we start the idle loop.
937 * The arch is responsible of calling:
939 * - rcu_idle_enter() after its last use of RCU before the CPU is put
941 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
943 void tick_nohz_idle_enter(void)
945 struct tick_sched *ts;
947 WARN_ON_ONCE(irqs_disabled());
950 * Update the idle state in the scheduler domain hierarchy
951 * when tick_nohz_stop_sched_tick() is called from the idle loop.
952 * State will be updated to busy during the first busy tick after
955 set_cpu_sd_state_idle();
959 ts = this_cpu_ptr(&tick_cpu_sched);
961 __tick_nohz_idle_enter(ts);
967 * tick_nohz_irq_exit - update next tick event from interrupt exit
969 * When an interrupt fires while we are idle and it doesn't cause
970 * a reschedule, it may still add, modify or delete a timer, enqueue
971 * an RCU callback, etc...
972 * So we need to re-calculate and reprogram the next tick event.
974 void tick_nohz_irq_exit(void)
976 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
979 __tick_nohz_idle_enter(ts);
981 tick_nohz_full_update_tick(ts);
985 * tick_nohz_get_sleep_length - return the length of the current sleep
987 * Called from power state control code with interrupts disabled
989 ktime_t tick_nohz_get_sleep_length(void)
991 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
993 return ts->sleep_length;
997 * tick_nohz_get_idle_calls - return the current idle calls counter value
999 * Called from the schedutil frequency scaling governor in scheduler context.
1001 unsigned long tick_nohz_get_idle_calls(void)
1003 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1005 return ts->idle_calls;
1008 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1010 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1011 unsigned long ticks;
1013 if (vtime_accounting_cpu_enabled())
1016 * We stopped the tick in idle. Update process times would miss the
1017 * time we slept as update_process_times does only a 1 tick
1018 * accounting. Enforce that this is accounted to idle !
1020 ticks = jiffies - ts->idle_jiffies;
1022 * We might be one off. Do not randomly account a huge number of ticks!
1024 if (ticks && ticks < LONG_MAX)
1025 account_idle_ticks(ticks);
1030 * tick_nohz_idle_exit - restart the idle tick from the idle task
1032 * Restart the idle tick when the CPU is woken up from idle
1033 * This also exit the RCU extended quiescent state. The CPU
1034 * can use RCU again after this function is called.
1036 void tick_nohz_idle_exit(void)
1038 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1041 local_irq_disable();
1043 WARN_ON_ONCE(!ts->inidle);
1047 if (ts->idle_active || ts->tick_stopped)
1050 if (ts->idle_active)
1051 tick_nohz_stop_idle(ts, now);
1053 if (ts->tick_stopped) {
1054 tick_nohz_restart_sched_tick(ts, now);
1055 tick_nohz_account_idle_ticks(ts);
1062 * The nohz low res interrupt handler
1064 static void tick_nohz_handler(struct clock_event_device *dev)
1066 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1067 struct pt_regs *regs = get_irq_regs();
1068 ktime_t now = ktime_get();
1070 dev->next_event = KTIME_MAX;
1072 tick_sched_do_timer(now);
1073 tick_sched_handle(ts, regs);
1075 /* No need to reprogram if we are running tickless */
1076 if (unlikely(ts->tick_stopped))
1079 hrtimer_forward(&ts->sched_timer, now, tick_period);
1080 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1083 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1085 if (!tick_nohz_enabled)
1087 ts->nohz_mode = mode;
1088 /* One update is enough */
1089 if (!test_and_set_bit(0, &tick_nohz_active))
1090 timers_update_migration(true);
1094 * tick_nohz_switch_to_nohz - switch to nohz mode
1096 static void tick_nohz_switch_to_nohz(void)
1098 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1101 if (!tick_nohz_enabled)
1104 if (tick_switch_to_oneshot(tick_nohz_handler))
1108 * Recycle the hrtimer in ts, so we can share the
1109 * hrtimer_forward with the highres code.
1111 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1112 /* Get the next period */
1113 next = tick_init_jiffy_update();
1115 hrtimer_set_expires(&ts->sched_timer, next);
1116 hrtimer_forward_now(&ts->sched_timer, tick_period);
1117 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1118 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1121 static inline void tick_nohz_irq_enter(void)
1123 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1126 if (!ts->idle_active && !ts->tick_stopped)
1129 if (ts->idle_active)
1130 tick_nohz_stop_idle(ts, now);
1131 if (ts->tick_stopped)
1132 tick_nohz_update_jiffies(now);
1137 static inline void tick_nohz_switch_to_nohz(void) { }
1138 static inline void tick_nohz_irq_enter(void) { }
1139 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1141 #endif /* CONFIG_NO_HZ_COMMON */
1144 * Called from irq_enter to notify about the possible interruption of idle()
1146 void tick_irq_enter(void)
1148 tick_check_oneshot_broadcast_this_cpu();
1149 tick_nohz_irq_enter();
1153 * High resolution timer specific code
1155 #ifdef CONFIG_HIGH_RES_TIMERS
1157 * We rearm the timer until we get disabled by the idle code.
1158 * Called with interrupts disabled.
1160 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1162 struct tick_sched *ts =
1163 container_of(timer, struct tick_sched, sched_timer);
1164 struct pt_regs *regs = get_irq_regs();
1165 ktime_t now = ktime_get();
1167 tick_sched_do_timer(now);
1170 * Do not call, when we are not in irq context and have
1171 * no valid regs pointer
1174 tick_sched_handle(ts, regs);
1178 /* No need to reprogram if we are in idle or full dynticks mode */
1179 if (unlikely(ts->tick_stopped))
1180 return HRTIMER_NORESTART;
1182 hrtimer_forward(timer, now, tick_period);
1184 return HRTIMER_RESTART;
1187 static int sched_skew_tick;
1189 static int __init skew_tick(char *str)
1191 get_option(&str, &sched_skew_tick);
1195 early_param("skew_tick", skew_tick);
1198 * tick_setup_sched_timer - setup the tick emulation timer
1200 void tick_setup_sched_timer(void)
1202 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1203 ktime_t now = ktime_get();
1206 * Emulate tick processing via per-CPU hrtimers:
1208 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1209 ts->sched_timer.function = tick_sched_timer;
1211 /* Get the next period (per-CPU) */
1212 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1214 /* Offset the tick to avert jiffies_lock contention. */
1215 if (sched_skew_tick) {
1216 u64 offset = ktime_to_ns(tick_period) >> 1;
1217 do_div(offset, num_possible_cpus());
1218 offset *= smp_processor_id();
1219 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1222 hrtimer_forward(&ts->sched_timer, now, tick_period);
1223 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1224 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1226 #endif /* HIGH_RES_TIMERS */
1228 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1229 void tick_cancel_sched_timer(int cpu)
1231 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1233 # ifdef CONFIG_HIGH_RES_TIMERS
1234 if (ts->sched_timer.base)
1235 hrtimer_cancel(&ts->sched_timer);
1238 memset(ts, 0, sizeof(*ts));
1243 * Async notification about clocksource changes
1245 void tick_clock_notify(void)
1249 for_each_possible_cpu(cpu)
1250 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1254 * Async notification about clock event changes
1256 void tick_oneshot_notify(void)
1258 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1260 set_bit(0, &ts->check_clocks);
1264 * Check, if a change happened, which makes oneshot possible.
1266 * Called cyclic from the hrtimer softirq (driven by the timer
1267 * softirq) allow_nohz signals, that we can switch into low-res nohz
1268 * mode, because high resolution timers are disabled (either compile
1269 * or runtime). Called with interrupts disabled.
1271 int tick_check_oneshot_change(int allow_nohz)
1273 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1275 if (!test_and_clear_bit(0, &ts->check_clocks))
1278 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1281 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1287 tick_nohz_switch_to_nohz();