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:
64 * The READ_ONCE() pairs with two updates done later in this function.
66 delta = ktime_sub(now, READ_ONCE(last_jiffies_update));
67 if (delta < tick_period)
70 /* Reevaluate with jiffies_lock held */
71 write_seqlock(&jiffies_lock);
73 delta = ktime_sub(now, last_jiffies_update);
74 if (delta >= tick_period) {
76 delta = ktime_sub(delta, tick_period);
77 /* Pairs with the lockless read in this function. */
78 WRITE_ONCE(last_jiffies_update,
79 ktime_add(last_jiffies_update, tick_period));
81 /* Slow path for long timeouts */
82 if (unlikely(delta >= tick_period)) {
83 s64 incr = ktime_to_ns(tick_period);
85 ticks = ktime_divns(delta, incr);
87 /* Pairs with the lockless read in this function. */
88 WRITE_ONCE(last_jiffies_update,
89 ktime_add_ns(last_jiffies_update,
94 /* Keep the tick_next_period variable up to date */
95 tick_next_period = ktime_add(last_jiffies_update, tick_period);
97 write_sequnlock(&jiffies_lock);
100 write_sequnlock(&jiffies_lock);
105 * Initialize and return retrieve the jiffies update.
107 static ktime_t tick_init_jiffy_update(void)
111 write_seqlock(&jiffies_lock);
112 /* Did we start the jiffies update yet ? */
113 if (last_jiffies_update == 0)
114 last_jiffies_update = tick_next_period;
115 period = last_jiffies_update;
116 write_sequnlock(&jiffies_lock);
120 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
122 int cpu = smp_processor_id();
124 #ifdef CONFIG_NO_HZ_COMMON
126 * Check if the do_timer duty was dropped. We don't care about
127 * concurrency: This happens only when the CPU in charge went
128 * into a long sleep. If two CPUs happen to assign themselves to
129 * this duty, then the jiffies update is still serialized by
132 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
133 && !tick_nohz_full_cpu(cpu))
134 tick_do_timer_cpu = cpu;
137 /* Check, if the jiffies need an update */
138 if (tick_do_timer_cpu == cpu)
139 tick_do_update_jiffies64(now);
142 ts->got_idle_tick = 1;
145 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
147 #ifdef CONFIG_NO_HZ_COMMON
149 * When we are idle and the tick is stopped, we have to touch
150 * the watchdog as we might not schedule for a really long
151 * time. This happens on complete idle SMP systems while
152 * waiting on the login prompt. We also increment the "start of
153 * idle" jiffy stamp so the idle accounting adjustment we do
154 * when we go busy again does not account too much ticks.
156 if (ts->tick_stopped) {
157 touch_softlockup_watchdog_sched();
158 if (is_idle_task(current))
161 * In case the current tick fired too early past its expected
162 * expiration, make sure we don't bypass the next clock reprogramming
163 * to the same deadline.
168 update_process_times(user_mode(regs));
169 profile_tick(CPU_PROFILING);
173 #ifdef CONFIG_NO_HZ_FULL
174 cpumask_var_t tick_nohz_full_mask;
175 bool tick_nohz_full_running;
176 static atomic_t tick_dep_mask;
178 static bool check_tick_dependency(atomic_t *dep)
180 int val = atomic_read(dep);
182 if (val & TICK_DEP_MASK_POSIX_TIMER) {
183 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
187 if (val & TICK_DEP_MASK_PERF_EVENTS) {
188 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
192 if (val & TICK_DEP_MASK_SCHED) {
193 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
197 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
198 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
205 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
207 lockdep_assert_irqs_disabled();
209 if (unlikely(!cpu_online(cpu)))
212 if (check_tick_dependency(&tick_dep_mask))
215 if (check_tick_dependency(&ts->tick_dep_mask))
218 if (check_tick_dependency(¤t->tick_dep_mask))
221 if (check_tick_dependency(¤t->signal->tick_dep_mask))
227 static void nohz_full_kick_func(struct irq_work *work)
229 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
232 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
233 .func = nohz_full_kick_func,
237 * Kick this CPU if it's full dynticks in order to force it to
238 * re-evaluate its dependency on the tick and restart it if necessary.
239 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
242 static void tick_nohz_full_kick(void)
244 if (!tick_nohz_full_cpu(smp_processor_id()))
247 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
251 * Kick the CPU if it's full dynticks in order to force it to
252 * re-evaluate its dependency on the tick and restart it if necessary.
254 void tick_nohz_full_kick_cpu(int cpu)
256 if (!tick_nohz_full_cpu(cpu))
259 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
263 * Kick all full dynticks CPUs in order to force these to re-evaluate
264 * their dependency on the tick and restart it if necessary.
266 static void tick_nohz_full_kick_all(void)
270 if (!tick_nohz_full_running)
274 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
275 tick_nohz_full_kick_cpu(cpu);
279 static void tick_nohz_dep_set_all(atomic_t *dep,
280 enum tick_dep_bits bit)
284 prev = atomic_fetch_or(BIT(bit), dep);
286 tick_nohz_full_kick_all();
290 * Set a global tick dependency. Used by perf events that rely on freq and
293 void tick_nohz_dep_set(enum tick_dep_bits bit)
295 tick_nohz_dep_set_all(&tick_dep_mask, bit);
298 void tick_nohz_dep_clear(enum tick_dep_bits bit)
300 atomic_andnot(BIT(bit), &tick_dep_mask);
304 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
305 * manage events throttling.
307 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
310 struct tick_sched *ts;
312 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
314 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
317 /* Perf needs local kick that is NMI safe */
318 if (cpu == smp_processor_id()) {
319 tick_nohz_full_kick();
321 /* Remote irq work not NMI-safe */
322 if (!WARN_ON_ONCE(in_nmi()))
323 tick_nohz_full_kick_cpu(cpu);
329 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
331 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
333 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
337 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
340 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
343 * We could optimize this with just kicking the target running the task
344 * if that noise matters for nohz full users.
346 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
349 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
351 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
355 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
356 * per process timers.
358 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
360 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
363 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
365 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
369 * Re-evaluate the need for the tick as we switch the current task.
370 * It might need the tick due to per task/process properties:
371 * perf events, posix CPU timers, ...
373 void __tick_nohz_task_switch(void)
376 struct tick_sched *ts;
378 local_irq_save(flags);
380 if (!tick_nohz_full_cpu(smp_processor_id()))
383 ts = this_cpu_ptr(&tick_cpu_sched);
385 if (ts->tick_stopped) {
386 if (atomic_read(¤t->tick_dep_mask) ||
387 atomic_read(¤t->signal->tick_dep_mask))
388 tick_nohz_full_kick();
391 local_irq_restore(flags);
394 /* Get the boot-time nohz CPU list from the kernel parameters. */
395 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
397 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
398 cpumask_copy(tick_nohz_full_mask, cpumask);
399 tick_nohz_full_running = true;
402 static int tick_nohz_cpu_down(unsigned int cpu)
405 * The boot CPU handles housekeeping duty (unbound timers,
406 * workqueues, timekeeping, ...) on behalf of full dynticks
407 * CPUs. It must remain online when nohz full is enabled.
409 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
414 void __init tick_nohz_init(void)
418 if (!tick_nohz_full_running)
422 * Full dynticks uses irq work to drive the tick rescheduling on safe
423 * locking contexts. But then we need irq work to raise its own
424 * interrupts to avoid circular dependency on the tick
426 if (!arch_irq_work_has_interrupt()) {
427 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
428 cpumask_clear(tick_nohz_full_mask);
429 tick_nohz_full_running = false;
433 cpu = smp_processor_id();
435 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
436 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
438 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
441 for_each_cpu(cpu, tick_nohz_full_mask)
442 context_tracking_cpu_set(cpu);
444 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
445 "kernel/nohz:predown", NULL,
448 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
449 cpumask_pr_args(tick_nohz_full_mask));
454 * NOHZ - aka dynamic tick functionality
456 #ifdef CONFIG_NO_HZ_COMMON
460 bool tick_nohz_enabled __read_mostly = true;
461 unsigned long tick_nohz_active __read_mostly;
463 * Enable / Disable tickless mode
465 static int __init setup_tick_nohz(char *str)
467 return (kstrtobool(str, &tick_nohz_enabled) == 0);
470 __setup("nohz=", setup_tick_nohz);
472 bool tick_nohz_tick_stopped(void)
474 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
476 return ts->tick_stopped;
479 bool tick_nohz_tick_stopped_cpu(int cpu)
481 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
483 return ts->tick_stopped;
487 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
489 * Called from interrupt entry when the CPU was idle
491 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
492 * must be updated. Otherwise an interrupt handler could use a stale jiffy
493 * value. We do this unconditionally on any CPU, as we don't know whether the
494 * CPU, which has the update task assigned is in a long sleep.
496 static void tick_nohz_update_jiffies(ktime_t now)
500 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
502 local_irq_save(flags);
503 tick_do_update_jiffies64(now);
504 local_irq_restore(flags);
506 touch_softlockup_watchdog_sched();
510 * Updates the per-CPU time idle statistics counters
513 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
517 if (ts->idle_active) {
518 delta = ktime_sub(now, ts->idle_entrytime);
519 if (nr_iowait_cpu(cpu) > 0)
520 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
522 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
523 ts->idle_entrytime = now;
526 if (last_update_time)
527 *last_update_time = ktime_to_us(now);
531 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
533 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
536 sched_clock_idle_wakeup_event();
539 static void tick_nohz_start_idle(struct tick_sched *ts)
541 ts->idle_entrytime = ktime_get();
543 sched_clock_idle_sleep_event();
547 * get_cpu_idle_time_us - get the total idle time of a CPU
548 * @cpu: CPU number to query
549 * @last_update_time: variable to store update time in. Do not update
552 * Return the cumulative idle time (since boot) for a given
553 * CPU, in microseconds.
555 * This time is measured via accounting rather than sampling,
556 * and is as accurate as ktime_get() is.
558 * This function returns -1 if NOHZ is not enabled.
560 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
562 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
565 if (!tick_nohz_active)
569 if (last_update_time) {
570 update_ts_time_stats(cpu, ts, now, last_update_time);
571 idle = ts->idle_sleeptime;
573 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
574 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
576 idle = ktime_add(ts->idle_sleeptime, delta);
578 idle = ts->idle_sleeptime;
582 return ktime_to_us(idle);
585 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
588 * get_cpu_iowait_time_us - get the total iowait time of a CPU
589 * @cpu: CPU number to query
590 * @last_update_time: variable to store update time in. Do not update
593 * Return the cumulative iowait time (since boot) for a given
594 * CPU, in microseconds.
596 * This time is measured via accounting rather than sampling,
597 * and is as accurate as ktime_get() is.
599 * This function returns -1 if NOHZ is not enabled.
601 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
603 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
606 if (!tick_nohz_active)
610 if (last_update_time) {
611 update_ts_time_stats(cpu, ts, now, last_update_time);
612 iowait = ts->iowait_sleeptime;
614 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
615 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
617 iowait = ktime_add(ts->iowait_sleeptime, delta);
619 iowait = ts->iowait_sleeptime;
623 return ktime_to_us(iowait);
625 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
627 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
629 hrtimer_cancel(&ts->sched_timer);
630 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
632 /* Forward the time to expire in the future */
633 hrtimer_forward(&ts->sched_timer, now, tick_period);
635 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
636 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
638 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
641 * Reset to make sure next tick stop doesn't get fooled by past
642 * cached clock deadline.
647 static inline bool local_timer_softirq_pending(void)
649 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
652 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
654 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
655 unsigned long seq, basejiff;
657 /* Read jiffies and the time when jiffies were updated last */
659 seq = read_seqbegin(&jiffies_lock);
660 basemono = last_jiffies_update;
662 } while (read_seqretry(&jiffies_lock, seq));
663 ts->last_jiffies = basejiff;
664 ts->timer_expires_base = basemono;
667 * Keep the periodic tick, when RCU, architecture or irq_work
669 * Aside of that check whether the local timer softirq is
670 * pending. If so its a bad idea to call get_next_timer_interrupt()
671 * because there is an already expired timer, so it will request
672 * immeditate expiry, which rearms the hardware timer with a
673 * minimal delta which brings us back to this place
674 * immediately. Lather, rinse and repeat...
676 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
677 irq_work_needs_cpu() || local_timer_softirq_pending()) {
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) {
709 ts->timer_expires = 0;
715 * If this CPU is the one which had the do_timer() duty last, we limit
716 * the sleep time to the timekeeping max_deferment value.
717 * Otherwise we can sleep as long as we want.
719 delta = timekeeping_max_deferment();
720 if (cpu != tick_do_timer_cpu &&
721 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
724 /* Calculate the next expiry time */
725 if (delta < (KTIME_MAX - basemono))
726 expires = basemono + delta;
730 ts->timer_expires = min_t(u64, expires, next_tick);
733 return ts->timer_expires;
736 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
738 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
739 u64 basemono = ts->timer_expires_base;
740 u64 expires = ts->timer_expires;
741 ktime_t tick = expires;
743 /* Make sure we won't be trying to stop it twice in a row. */
744 ts->timer_expires_base = 0;
747 * If this CPU is the one which updates jiffies, then give up
748 * the assignment and let it be taken by the CPU which runs
749 * the tick timer next, which might be this CPU as well. If we
750 * don't drop this here the jiffies might be stale and
751 * do_timer() never invoked. Keep track of the fact that it
752 * was the one which had the do_timer() duty last.
754 if (cpu == tick_do_timer_cpu) {
755 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
756 ts->do_timer_last = 1;
757 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
758 ts->do_timer_last = 0;
761 /* Skip reprogram of event if its not changed */
762 if (ts->tick_stopped && (expires == ts->next_tick)) {
763 /* Sanity check: make sure clockevent is actually programmed */
764 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
768 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
769 basemono, ts->next_tick, dev->next_event,
770 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
774 * nohz_stop_sched_tick can be called several times before
775 * the nohz_restart_sched_tick is called. This happens when
776 * interrupts arrive which do not cause a reschedule. In the
777 * first call we save the current tick time, so we can restart
778 * the scheduler tick in nohz_restart_sched_tick.
780 if (!ts->tick_stopped) {
781 calc_load_nohz_start();
782 cpu_load_update_nohz_start();
785 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
786 ts->tick_stopped = 1;
787 trace_tick_stop(1, TICK_DEP_MASK_NONE);
790 ts->next_tick = tick;
793 * If the expiration time == KTIME_MAX, then we simply stop
796 if (unlikely(expires == KTIME_MAX)) {
797 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
798 hrtimer_cancel(&ts->sched_timer);
802 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
803 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
805 hrtimer_set_expires(&ts->sched_timer, tick);
806 tick_program_event(tick, 1);
810 static void tick_nohz_retain_tick(struct tick_sched *ts)
812 ts->timer_expires_base = 0;
815 #ifdef CONFIG_NO_HZ_FULL
816 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
818 if (tick_nohz_next_event(ts, cpu))
819 tick_nohz_stop_tick(ts, cpu);
821 tick_nohz_retain_tick(ts);
823 #endif /* CONFIG_NO_HZ_FULL */
825 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
827 /* Update jiffies first */
828 tick_do_update_jiffies64(now);
829 cpu_load_update_nohz_stop();
832 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
833 * the clock forward checks in the enqueue path:
837 calc_load_nohz_stop();
838 touch_softlockup_watchdog_sched();
840 * Cancel the scheduled timer and restore the tick
842 ts->tick_stopped = 0;
843 ts->idle_exittime = now;
845 tick_nohz_restart(ts, now);
848 static void tick_nohz_full_update_tick(struct tick_sched *ts)
850 #ifdef CONFIG_NO_HZ_FULL
851 int cpu = smp_processor_id();
853 if (!tick_nohz_full_cpu(cpu))
856 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
859 if (can_stop_full_tick(cpu, ts))
860 tick_nohz_stop_sched_tick(ts, cpu);
861 else if (ts->tick_stopped)
862 tick_nohz_restart_sched_tick(ts, ktime_get());
866 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
869 * If this CPU is offline and it is the one which updates
870 * jiffies, then give up the assignment and let it be taken by
871 * the CPU which runs the tick timer next. If we don't drop
872 * this here the jiffies might be stale and do_timer() never
875 if (unlikely(!cpu_online(cpu))) {
876 if (cpu == tick_do_timer_cpu)
877 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
879 * Make sure the CPU doesn't get fooled by obsolete tick
880 * deadline if it comes back online later.
886 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
892 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
893 static int ratelimit;
895 if (ratelimit < 10 &&
896 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
897 pr_warn("NOHZ: local_softirq_pending %02x\n",
898 (unsigned int) local_softirq_pending());
904 if (tick_nohz_full_enabled()) {
906 * Keep the tick alive to guarantee timekeeping progression
907 * if there are full dynticks CPUs around
909 if (tick_do_timer_cpu == cpu)
912 * Boot safety: make sure the timekeeping duty has been
913 * assigned before entering dyntick-idle mode,
915 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
922 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
925 int cpu = smp_processor_id();
928 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
929 * tick timer expiration time is known already.
931 if (ts->timer_expires_base)
932 expires = ts->timer_expires;
933 else if (can_stop_idle_tick(cpu, ts))
934 expires = tick_nohz_next_event(ts, cpu);
941 int was_stopped = ts->tick_stopped;
943 tick_nohz_stop_tick(ts, cpu);
946 ts->idle_expires = expires;
948 if (!was_stopped && ts->tick_stopped) {
949 ts->idle_jiffies = ts->last_jiffies;
950 nohz_balance_enter_idle(cpu);
953 tick_nohz_retain_tick(ts);
958 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
960 * When the next event is more than a tick into the future, stop the idle tick
962 void tick_nohz_idle_stop_tick(void)
964 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
967 void tick_nohz_idle_retain_tick(void)
969 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
971 * Undo the effect of get_next_timer_interrupt() called from
972 * tick_nohz_next_event().
978 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
980 * Called when we start the idle loop.
982 void tick_nohz_idle_enter(void)
984 struct tick_sched *ts;
986 lockdep_assert_irqs_enabled();
990 ts = this_cpu_ptr(&tick_cpu_sched);
992 WARN_ON_ONCE(ts->timer_expires_base);
995 tick_nohz_start_idle(ts);
1001 * tick_nohz_irq_exit - update next tick event from interrupt exit
1003 * When an interrupt fires while we are idle and it doesn't cause
1004 * a reschedule, it may still add, modify or delete a timer, enqueue
1005 * an RCU callback, etc...
1006 * So we need to re-calculate and reprogram the next tick event.
1008 void tick_nohz_irq_exit(void)
1010 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1013 tick_nohz_start_idle(ts);
1015 tick_nohz_full_update_tick(ts);
1019 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1021 bool tick_nohz_idle_got_tick(void)
1023 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1025 if (ts->got_idle_tick) {
1026 ts->got_idle_tick = 0;
1033 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1034 * @delta_next: duration until the next event if the tick cannot be stopped
1036 * Called from power state control code with interrupts disabled
1038 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1040 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1041 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1042 int cpu = smp_processor_id();
1044 * The idle entry time is expected to be a sufficient approximation of
1045 * the current time at this point.
1047 ktime_t now = ts->idle_entrytime;
1050 WARN_ON_ONCE(!ts->inidle);
1052 *delta_next = ktime_sub(dev->next_event, now);
1054 if (!can_stop_idle_tick(cpu, ts))
1057 next_event = tick_nohz_next_event(ts, cpu);
1062 * If the next highres timer to expire is earlier than next_event, the
1063 * idle governor needs to know that.
1065 next_event = min_t(u64, next_event,
1066 hrtimer_next_event_without(&ts->sched_timer));
1068 return ktime_sub(next_event, now);
1072 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1073 * for a particular CPU.
1075 * Called from the schedutil frequency scaling governor in scheduler context.
1077 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1079 struct tick_sched *ts = tick_get_tick_sched(cpu);
1081 return ts->idle_calls;
1085 * tick_nohz_get_idle_calls - return the current idle calls counter value
1087 * Called from the schedutil frequency scaling governor in scheduler context.
1089 unsigned long tick_nohz_get_idle_calls(void)
1091 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1093 return ts->idle_calls;
1096 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1098 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1099 unsigned long ticks;
1101 if (vtime_accounting_cpu_enabled())
1104 * We stopped the tick in idle. Update process times would miss the
1105 * time we slept as update_process_times does only a 1 tick
1106 * accounting. Enforce that this is accounted to idle !
1108 ticks = jiffies - ts->idle_jiffies;
1110 * We might be one off. Do not randomly account a huge number of ticks!
1112 if (ticks && ticks < LONG_MAX)
1113 account_idle_ticks(ticks);
1117 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1119 tick_nohz_restart_sched_tick(ts, now);
1120 tick_nohz_account_idle_ticks(ts);
1123 void tick_nohz_idle_restart_tick(void)
1125 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1127 if (ts->tick_stopped)
1128 __tick_nohz_idle_restart_tick(ts, ktime_get());
1132 * tick_nohz_idle_exit - restart the idle tick from the idle task
1134 * Restart the idle tick when the CPU is woken up from idle
1135 * This also exit the RCU extended quiescent state. The CPU
1136 * can use RCU again after this function is called.
1138 void tick_nohz_idle_exit(void)
1140 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1141 bool idle_active, tick_stopped;
1144 local_irq_disable();
1146 WARN_ON_ONCE(!ts->inidle);
1147 WARN_ON_ONCE(ts->timer_expires_base);
1150 idle_active = ts->idle_active;
1151 tick_stopped = ts->tick_stopped;
1153 if (idle_active || tick_stopped)
1157 tick_nohz_stop_idle(ts, now);
1160 __tick_nohz_idle_restart_tick(ts, now);
1166 * The nohz low res interrupt handler
1168 static void tick_nohz_handler(struct clock_event_device *dev)
1170 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1171 struct pt_regs *regs = get_irq_regs();
1172 ktime_t now = ktime_get();
1174 dev->next_event = KTIME_MAX;
1176 tick_sched_do_timer(ts, now);
1177 tick_sched_handle(ts, regs);
1179 /* No need to reprogram if we are running tickless */
1180 if (unlikely(ts->tick_stopped))
1183 hrtimer_forward(&ts->sched_timer, now, tick_period);
1184 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1187 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1189 if (!tick_nohz_enabled)
1191 ts->nohz_mode = mode;
1192 /* One update is enough */
1193 if (!test_and_set_bit(0, &tick_nohz_active))
1194 timers_update_nohz();
1198 * tick_nohz_switch_to_nohz - switch to nohz mode
1200 static void tick_nohz_switch_to_nohz(void)
1202 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1205 if (!tick_nohz_enabled)
1208 if (tick_switch_to_oneshot(tick_nohz_handler))
1212 * Recycle the hrtimer in ts, so we can share the
1213 * hrtimer_forward with the highres code.
1215 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1216 /* Get the next period */
1217 next = tick_init_jiffy_update();
1219 hrtimer_set_expires(&ts->sched_timer, next);
1220 hrtimer_forward_now(&ts->sched_timer, tick_period);
1221 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1222 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1225 static inline void tick_nohz_irq_enter(void)
1227 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1230 if (!ts->idle_active && !ts->tick_stopped)
1233 if (ts->idle_active)
1234 tick_nohz_stop_idle(ts, now);
1235 if (ts->tick_stopped)
1236 tick_nohz_update_jiffies(now);
1241 static inline void tick_nohz_switch_to_nohz(void) { }
1242 static inline void tick_nohz_irq_enter(void) { }
1243 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1245 #endif /* CONFIG_NO_HZ_COMMON */
1248 * Called from irq_enter to notify about the possible interruption of idle()
1250 void tick_irq_enter(void)
1252 tick_check_oneshot_broadcast_this_cpu();
1253 tick_nohz_irq_enter();
1257 * High resolution timer specific code
1259 #ifdef CONFIG_HIGH_RES_TIMERS
1261 * We rearm the timer until we get disabled by the idle code.
1262 * Called with interrupts disabled.
1264 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1266 struct tick_sched *ts =
1267 container_of(timer, struct tick_sched, sched_timer);
1268 struct pt_regs *regs = get_irq_regs();
1269 ktime_t now = ktime_get();
1271 tick_sched_do_timer(ts, now);
1274 * Do not call, when we are not in irq context and have
1275 * no valid regs pointer
1278 tick_sched_handle(ts, regs);
1282 /* No need to reprogram if we are in idle or full dynticks mode */
1283 if (unlikely(ts->tick_stopped))
1284 return HRTIMER_NORESTART;
1286 hrtimer_forward(timer, now, tick_period);
1288 return HRTIMER_RESTART;
1291 static int sched_skew_tick;
1293 static int __init skew_tick(char *str)
1295 get_option(&str, &sched_skew_tick);
1299 early_param("skew_tick", skew_tick);
1302 * tick_setup_sched_timer - setup the tick emulation timer
1304 void tick_setup_sched_timer(void)
1306 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1307 ktime_t now = ktime_get();
1310 * Emulate tick processing via per-CPU hrtimers:
1312 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1313 ts->sched_timer.function = tick_sched_timer;
1315 /* Get the next period (per-CPU) */
1316 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1318 /* Offset the tick to avert jiffies_lock contention. */
1319 if (sched_skew_tick) {
1320 u64 offset = ktime_to_ns(tick_period) >> 1;
1321 do_div(offset, num_possible_cpus());
1322 offset *= smp_processor_id();
1323 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1326 hrtimer_forward(&ts->sched_timer, now, tick_period);
1327 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1328 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1330 #endif /* HIGH_RES_TIMERS */
1332 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1333 void tick_cancel_sched_timer(int cpu)
1335 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1337 # ifdef CONFIG_HIGH_RES_TIMERS
1338 if (ts->sched_timer.base)
1339 hrtimer_cancel(&ts->sched_timer);
1342 memset(ts, 0, sizeof(*ts));
1347 * Async notification about clocksource changes
1349 void tick_clock_notify(void)
1353 for_each_possible_cpu(cpu)
1354 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1358 * Async notification about clock event changes
1360 void tick_oneshot_notify(void)
1362 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1364 set_bit(0, &ts->check_clocks);
1368 * Check, if a change happened, which makes oneshot possible.
1370 * Called cyclic from the hrtimer softirq (driven by the timer
1371 * softirq) allow_nohz signals, that we can switch into low-res nohz
1372 * mode, because high resolution timers are disabled (either compile
1373 * or runtime). Called with interrupts disabled.
1375 int tick_check_oneshot_change(int allow_nohz)
1377 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1379 if (!test_and_clear_bit(0, &ts->check_clocks))
1382 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1385 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1391 tick_nohz_switch_to_nohz();