1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 * No idle tick implementation for low and high resolution timers
9 * Started by: Thomas Gleixner and Ingo Molnar
11 #include <linux/cpu.h>
12 #include <linux/err.h>
13 #include <linux/hrtimer.h>
14 #include <linux/interrupt.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/percpu.h>
17 #include <linux/nmi.h>
18 #include <linux/profile.h>
19 #include <linux/sched/signal.h>
20 #include <linux/sched/clock.h>
21 #include <linux/sched/stat.h>
22 #include <linux/sched/nohz.h>
23 #include <linux/module.h>
24 #include <linux/irq_work.h>
25 #include <linux/posix-timers.h>
26 #include <linux/context_tracking.h>
29 #include <asm/irq_regs.h>
31 #include "tick-internal.h"
33 #include <trace/events/timer.h>
36 * Per-CPU nohz control structure
38 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 struct tick_sched *tick_get_tick_sched(int cpu)
42 return &per_cpu(tick_cpu_sched, cpu);
45 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47 * The time, when the last jiffy update happened. Protected by jiffies_lock.
49 static ktime_t last_jiffies_update;
52 * Must be called with interrupts disabled !
54 static void tick_do_update_jiffies64(ktime_t now)
56 unsigned long ticks = 0;
60 * Do a quick check without holding jiffies_lock:
62 delta = ktime_sub(now, last_jiffies_update);
63 if (delta < tick_period)
66 /* Reevaluate with jiffies_lock held */
67 write_seqlock(&jiffies_lock);
69 delta = ktime_sub(now, last_jiffies_update);
70 if (delta >= tick_period) {
72 delta = ktime_sub(delta, tick_period);
73 last_jiffies_update = ktime_add(last_jiffies_update,
76 /* Slow path for long timeouts */
77 if (unlikely(delta >= tick_period)) {
78 s64 incr = ktime_to_ns(tick_period);
80 ticks = ktime_divns(delta, incr);
82 last_jiffies_update = ktime_add_ns(last_jiffies_update,
87 /* Keep the tick_next_period variable up to date */
88 tick_next_period = ktime_add(last_jiffies_update, tick_period);
90 write_sequnlock(&jiffies_lock);
93 write_sequnlock(&jiffies_lock);
98 * Initialize and return retrieve the jiffies update.
100 static ktime_t tick_init_jiffy_update(void)
104 write_seqlock(&jiffies_lock);
105 /* Did we start the jiffies update yet ? */
106 if (last_jiffies_update == 0)
107 last_jiffies_update = tick_next_period;
108 period = last_jiffies_update;
109 write_sequnlock(&jiffies_lock);
113 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
115 int cpu = smp_processor_id();
117 #ifdef CONFIG_NO_HZ_COMMON
119 * Check if the do_timer duty was dropped. We don't care about
120 * concurrency: This happens only when the CPU in charge went
121 * into a long sleep. If two CPUs happen to assign themselves to
122 * this duty, then the jiffies update is still serialized by
125 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
126 && !tick_nohz_full_cpu(cpu))
127 tick_do_timer_cpu = cpu;
130 /* Check, if the jiffies need an update */
131 if (tick_do_timer_cpu == cpu)
132 tick_do_update_jiffies64(now);
135 ts->got_idle_tick = 1;
138 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
140 #ifdef CONFIG_NO_HZ_COMMON
142 * When we are idle and the tick is stopped, we have to touch
143 * the watchdog as we might not schedule for a really long
144 * time. This happens on complete idle SMP systems while
145 * waiting on the login prompt. We also increment the "start of
146 * idle" jiffy stamp so the idle accounting adjustment we do
147 * when we go busy again does not account too much ticks.
149 if (ts->tick_stopped) {
150 touch_softlockup_watchdog_sched();
151 if (is_idle_task(current))
154 * In case the current tick fired too early past its expected
155 * expiration, make sure we don't bypass the next clock reprogramming
156 * to the same deadline.
161 update_process_times(user_mode(regs));
162 profile_tick(CPU_PROFILING);
166 #ifdef CONFIG_NO_HZ_FULL
167 cpumask_var_t tick_nohz_full_mask;
168 bool tick_nohz_full_running;
169 static atomic_t tick_dep_mask;
171 static bool check_tick_dependency(atomic_t *dep)
173 int val = atomic_read(dep);
175 if (val & TICK_DEP_MASK_POSIX_TIMER) {
176 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
180 if (val & TICK_DEP_MASK_PERF_EVENTS) {
181 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
185 if (val & TICK_DEP_MASK_SCHED) {
186 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
190 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
191 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
198 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
200 lockdep_assert_irqs_disabled();
202 if (unlikely(!cpu_online(cpu)))
205 if (check_tick_dependency(&tick_dep_mask))
208 if (check_tick_dependency(&ts->tick_dep_mask))
211 if (check_tick_dependency(¤t->tick_dep_mask))
214 if (check_tick_dependency(¤t->signal->tick_dep_mask))
220 static void nohz_full_kick_func(struct irq_work *work)
222 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
225 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
226 .func = nohz_full_kick_func,
230 * Kick this CPU if it's full dynticks in order to force it to
231 * re-evaluate its dependency on the tick and restart it if necessary.
232 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
235 static void tick_nohz_full_kick(void)
237 if (!tick_nohz_full_cpu(smp_processor_id()))
240 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
244 * Kick the CPU if it's full dynticks in order to force it to
245 * re-evaluate its dependency on the tick and restart it if necessary.
247 void tick_nohz_full_kick_cpu(int cpu)
249 if (!tick_nohz_full_cpu(cpu))
252 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
256 * Kick all full dynticks CPUs in order to force these to re-evaluate
257 * their dependency on the tick and restart it if necessary.
259 static void tick_nohz_full_kick_all(void)
263 if (!tick_nohz_full_running)
267 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
268 tick_nohz_full_kick_cpu(cpu);
272 static void tick_nohz_dep_set_all(atomic_t *dep,
273 enum tick_dep_bits bit)
277 prev = atomic_fetch_or(BIT(bit), dep);
279 tick_nohz_full_kick_all();
283 * Set a global tick dependency. Used by perf events that rely on freq and
286 void tick_nohz_dep_set(enum tick_dep_bits bit)
288 tick_nohz_dep_set_all(&tick_dep_mask, bit);
291 void tick_nohz_dep_clear(enum tick_dep_bits bit)
293 atomic_andnot(BIT(bit), &tick_dep_mask);
297 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
298 * manage events throttling.
300 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
303 struct tick_sched *ts;
305 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
307 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
310 /* Perf needs local kick that is NMI safe */
311 if (cpu == smp_processor_id()) {
312 tick_nohz_full_kick();
314 /* Remote irq work not NMI-safe */
315 if (!WARN_ON_ONCE(in_nmi()))
316 tick_nohz_full_kick_cpu(cpu);
322 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
324 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
326 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
330 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
333 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
336 * We could optimize this with just kicking the target running the task
337 * if that noise matters for nohz full users.
339 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
342 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
344 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
348 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
349 * per process timers.
351 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
353 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
356 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
358 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
362 * Re-evaluate the need for the tick as we switch the current task.
363 * It might need the tick due to per task/process properties:
364 * perf events, posix CPU timers, ...
366 void __tick_nohz_task_switch(void)
369 struct tick_sched *ts;
371 local_irq_save(flags);
373 if (!tick_nohz_full_cpu(smp_processor_id()))
376 ts = this_cpu_ptr(&tick_cpu_sched);
378 if (ts->tick_stopped) {
379 if (atomic_read(¤t->tick_dep_mask) ||
380 atomic_read(¤t->signal->tick_dep_mask))
381 tick_nohz_full_kick();
384 local_irq_restore(flags);
387 /* Get the boot-time nohz CPU list from the kernel parameters. */
388 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
390 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
391 cpumask_copy(tick_nohz_full_mask, cpumask);
392 tick_nohz_full_running = true;
395 static int tick_nohz_cpu_down(unsigned int cpu)
398 * The boot CPU handles housekeeping duty (unbound timers,
399 * workqueues, timekeeping, ...) on behalf of full dynticks
400 * CPUs. It must remain online when nohz full is enabled.
402 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
407 void __init tick_nohz_init(void)
411 if (!tick_nohz_full_running)
415 * Full dynticks uses irq work to drive the tick rescheduling on safe
416 * locking contexts. But then we need irq work to raise its own
417 * interrupts to avoid circular dependency on the tick
419 if (!arch_irq_work_has_interrupt()) {
420 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
421 cpumask_clear(tick_nohz_full_mask);
422 tick_nohz_full_running = false;
426 cpu = smp_processor_id();
428 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
429 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
431 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
434 for_each_cpu(cpu, tick_nohz_full_mask)
435 context_tracking_cpu_set(cpu);
437 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
438 "kernel/nohz:predown", NULL,
441 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
442 cpumask_pr_args(tick_nohz_full_mask));
447 * NOHZ - aka dynamic tick functionality
449 #ifdef CONFIG_NO_HZ_COMMON
453 bool tick_nohz_enabled __read_mostly = true;
454 unsigned long tick_nohz_active __read_mostly;
456 * Enable / Disable tickless mode
458 static int __init setup_tick_nohz(char *str)
460 return (kstrtobool(str, &tick_nohz_enabled) == 0);
463 __setup("nohz=", setup_tick_nohz);
465 bool tick_nohz_tick_stopped(void)
467 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
469 return ts->tick_stopped;
472 bool tick_nohz_tick_stopped_cpu(int cpu)
474 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
476 return ts->tick_stopped;
480 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
482 * Called from interrupt entry when the CPU was idle
484 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
485 * must be updated. Otherwise an interrupt handler could use a stale jiffy
486 * value. We do this unconditionally on any CPU, as we don't know whether the
487 * CPU, which has the update task assigned is in a long sleep.
489 static void tick_nohz_update_jiffies(ktime_t now)
493 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
495 local_irq_save(flags);
496 tick_do_update_jiffies64(now);
497 local_irq_restore(flags);
499 touch_softlockup_watchdog_sched();
503 * Updates the per-CPU time idle statistics counters
506 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
510 if (ts->idle_active) {
511 delta = ktime_sub(now, ts->idle_entrytime);
512 if (nr_iowait_cpu(cpu) > 0)
513 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
515 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
516 ts->idle_entrytime = now;
519 if (last_update_time)
520 *last_update_time = ktime_to_us(now);
524 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
526 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
529 sched_clock_idle_wakeup_event();
532 static void tick_nohz_start_idle(struct tick_sched *ts)
534 ts->idle_entrytime = ktime_get();
536 sched_clock_idle_sleep_event();
540 * get_cpu_idle_time_us - get the total idle time of a CPU
541 * @cpu: CPU number to query
542 * @last_update_time: variable to store update time in. Do not update
545 * Return the cumulative idle time (since boot) for a given
546 * CPU, in microseconds.
548 * This time is measured via accounting rather than sampling,
549 * and is as accurate as ktime_get() is.
551 * This function returns -1 if NOHZ is not enabled.
553 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
555 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
558 if (!tick_nohz_active)
562 if (last_update_time) {
563 update_ts_time_stats(cpu, ts, now, last_update_time);
564 idle = ts->idle_sleeptime;
566 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
567 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
569 idle = ktime_add(ts->idle_sleeptime, delta);
571 idle = ts->idle_sleeptime;
575 return ktime_to_us(idle);
578 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
581 * get_cpu_iowait_time_us - get the total iowait time of a CPU
582 * @cpu: CPU number to query
583 * @last_update_time: variable to store update time in. Do not update
586 * Return the cumulative iowait time (since boot) for a given
587 * CPU, in microseconds.
589 * This time is measured via accounting rather than sampling,
590 * and is as accurate as ktime_get() is.
592 * This function returns -1 if NOHZ is not enabled.
594 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
596 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
599 if (!tick_nohz_active)
603 if (last_update_time) {
604 update_ts_time_stats(cpu, ts, now, last_update_time);
605 iowait = ts->iowait_sleeptime;
607 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
608 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
610 iowait = ktime_add(ts->iowait_sleeptime, delta);
612 iowait = ts->iowait_sleeptime;
616 return ktime_to_us(iowait);
618 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
620 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
622 hrtimer_cancel(&ts->sched_timer);
623 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
625 /* Forward the time to expire in the future */
626 hrtimer_forward(&ts->sched_timer, now, tick_period);
628 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
629 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
631 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
634 * Reset to make sure next tick stop doesn't get fooled by past
635 * cached clock deadline.
640 static inline bool local_timer_softirq_pending(void)
642 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
645 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
647 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
648 unsigned long seq, basejiff;
650 /* Read jiffies and the time when jiffies were updated last */
652 seq = read_seqbegin(&jiffies_lock);
653 basemono = last_jiffies_update;
655 } while (read_seqretry(&jiffies_lock, seq));
656 ts->last_jiffies = basejiff;
657 ts->timer_expires_base = basemono;
660 * Keep the periodic tick, when RCU, architecture or irq_work
662 * Aside of that check whether the local timer softirq is
663 * pending. If so its a bad idea to call get_next_timer_interrupt()
664 * because there is an already expired timer, so it will request
665 * immeditate expiry, which rearms the hardware timer with a
666 * minimal delta which brings us back to this place
667 * immediately. Lather, rinse and repeat...
669 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
670 irq_work_needs_cpu() || local_timer_softirq_pending()) {
671 next_tick = basemono + TICK_NSEC;
674 * Get the next pending timer. If high resolution
675 * timers are enabled this only takes the timer wheel
676 * timers into account. If high resolution timers are
677 * disabled this also looks at the next expiring
680 next_tmr = get_next_timer_interrupt(basejiff, basemono);
681 ts->next_timer = next_tmr;
682 /* Take the next rcu event into account */
683 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
687 * If the tick is due in the next period, keep it ticking or
688 * force prod the timer.
690 delta = next_tick - basemono;
691 if (delta <= (u64)TICK_NSEC) {
693 * Tell the timer code that the base is not idle, i.e. undo
694 * the effect of get_next_timer_interrupt():
698 * We've not stopped the tick yet, and there's a timer in the
699 * next period, so no point in stopping it either, bail.
701 if (!ts->tick_stopped) {
702 ts->timer_expires = 0;
708 * If this CPU is the one which had the do_timer() duty last, we limit
709 * the sleep time to the timekeeping max_deferment value.
710 * Otherwise we can sleep as long as we want.
712 delta = timekeeping_max_deferment();
713 if (cpu != tick_do_timer_cpu &&
714 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
717 /* Calculate the next expiry time */
718 if (delta < (KTIME_MAX - basemono))
719 expires = basemono + delta;
723 ts->timer_expires = min_t(u64, expires, next_tick);
726 return ts->timer_expires;
729 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
731 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
732 u64 basemono = ts->timer_expires_base;
733 u64 expires = ts->timer_expires;
734 ktime_t tick = expires;
736 /* Make sure we won't be trying to stop it twice in a row. */
737 ts->timer_expires_base = 0;
740 * If this CPU is the one which updates jiffies, then give up
741 * the assignment and let it be taken by the CPU which runs
742 * the tick timer next, which might be this CPU as well. If we
743 * don't drop this here the jiffies might be stale and
744 * do_timer() never invoked. Keep track of the fact that it
745 * was the one which had the do_timer() duty last.
747 if (cpu == tick_do_timer_cpu) {
748 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
749 ts->do_timer_last = 1;
750 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
751 ts->do_timer_last = 0;
754 /* Skip reprogram of event if its not changed */
755 if (ts->tick_stopped && (expires == ts->next_tick)) {
756 /* Sanity check: make sure clockevent is actually programmed */
757 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
761 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
762 basemono, ts->next_tick, dev->next_event,
763 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
767 * nohz_stop_sched_tick can be called several times before
768 * the nohz_restart_sched_tick is called. This happens when
769 * interrupts arrive which do not cause a reschedule. In the
770 * first call we save the current tick time, so we can restart
771 * the scheduler tick in nohz_restart_sched_tick.
773 if (!ts->tick_stopped) {
774 calc_load_nohz_start();
775 cpu_load_update_nohz_start();
778 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
779 ts->tick_stopped = 1;
780 trace_tick_stop(1, TICK_DEP_MASK_NONE);
783 ts->next_tick = tick;
786 * If the expiration time == KTIME_MAX, then we simply stop
789 if (unlikely(expires == KTIME_MAX)) {
790 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
791 hrtimer_cancel(&ts->sched_timer);
795 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
796 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
798 hrtimer_set_expires(&ts->sched_timer, tick);
799 tick_program_event(tick, 1);
803 static void tick_nohz_retain_tick(struct tick_sched *ts)
805 ts->timer_expires_base = 0;
808 #ifdef CONFIG_NO_HZ_FULL
809 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
811 if (tick_nohz_next_event(ts, cpu))
812 tick_nohz_stop_tick(ts, cpu);
814 tick_nohz_retain_tick(ts);
816 #endif /* CONFIG_NO_HZ_FULL */
818 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
820 /* Update jiffies first */
821 tick_do_update_jiffies64(now);
822 cpu_load_update_nohz_stop();
825 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
826 * the clock forward checks in the enqueue path:
830 calc_load_nohz_stop();
831 touch_softlockup_watchdog_sched();
833 * Cancel the scheduled timer and restore the tick
835 ts->tick_stopped = 0;
836 ts->idle_exittime = now;
838 tick_nohz_restart(ts, now);
841 static void tick_nohz_full_update_tick(struct tick_sched *ts)
843 #ifdef CONFIG_NO_HZ_FULL
844 int cpu = smp_processor_id();
846 if (!tick_nohz_full_cpu(cpu))
849 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
852 if (can_stop_full_tick(cpu, ts))
853 tick_nohz_stop_sched_tick(ts, cpu);
854 else if (ts->tick_stopped)
855 tick_nohz_restart_sched_tick(ts, ktime_get());
859 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
862 * If this CPU is offline and it is the one which updates
863 * jiffies, then give up the assignment and let it be taken by
864 * the CPU which runs the tick timer next. If we don't drop
865 * this here the jiffies might be stale and do_timer() never
868 if (unlikely(!cpu_online(cpu))) {
869 if (cpu == tick_do_timer_cpu)
870 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
872 * Make sure the CPU doesn't get fooled by obsolete tick
873 * deadline if it comes back online later.
879 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
885 if (unlikely(local_softirq_pending())) {
886 static int ratelimit;
888 if (ratelimit < 10 &&
889 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
890 pr_warn("NOHZ: local_softirq_pending %02x\n",
891 (unsigned int) local_softirq_pending());
897 if (tick_nohz_full_enabled()) {
899 * Keep the tick alive to guarantee timekeeping progression
900 * if there are full dynticks CPUs around
902 if (tick_do_timer_cpu == cpu)
905 * Boot safety: make sure the timekeeping duty has been
906 * assigned before entering dyntick-idle mode,
908 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
915 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
918 int cpu = smp_processor_id();
921 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
922 * tick timer expiration time is known already.
924 if (ts->timer_expires_base)
925 expires = ts->timer_expires;
926 else if (can_stop_idle_tick(cpu, ts))
927 expires = tick_nohz_next_event(ts, cpu);
934 int was_stopped = ts->tick_stopped;
936 tick_nohz_stop_tick(ts, cpu);
939 ts->idle_expires = expires;
941 if (!was_stopped && ts->tick_stopped) {
942 ts->idle_jiffies = ts->last_jiffies;
943 nohz_balance_enter_idle(cpu);
946 tick_nohz_retain_tick(ts);
951 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
953 * When the next event is more than a tick into the future, stop the idle tick
955 void tick_nohz_idle_stop_tick(void)
957 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
960 void tick_nohz_idle_retain_tick(void)
962 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
964 * Undo the effect of get_next_timer_interrupt() called from
965 * tick_nohz_next_event().
971 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
973 * Called when we start the idle loop.
975 void tick_nohz_idle_enter(void)
977 struct tick_sched *ts;
979 lockdep_assert_irqs_enabled();
983 ts = this_cpu_ptr(&tick_cpu_sched);
985 WARN_ON_ONCE(ts->timer_expires_base);
988 tick_nohz_start_idle(ts);
994 * tick_nohz_irq_exit - update next tick event from interrupt exit
996 * When an interrupt fires while we are idle and it doesn't cause
997 * a reschedule, it may still add, modify or delete a timer, enqueue
998 * an RCU callback, etc...
999 * So we need to re-calculate and reprogram the next tick event.
1001 void tick_nohz_irq_exit(void)
1003 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1006 tick_nohz_start_idle(ts);
1008 tick_nohz_full_update_tick(ts);
1012 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1014 bool tick_nohz_idle_got_tick(void)
1016 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1018 if (ts->got_idle_tick) {
1019 ts->got_idle_tick = 0;
1026 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1027 * @delta_next: duration until the next event if the tick cannot be stopped
1029 * Called from power state control code with interrupts disabled
1031 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1033 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1034 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1035 int cpu = smp_processor_id();
1037 * The idle entry time is expected to be a sufficient approximation of
1038 * the current time at this point.
1040 ktime_t now = ts->idle_entrytime;
1043 WARN_ON_ONCE(!ts->inidle);
1045 *delta_next = ktime_sub(dev->next_event, now);
1047 if (!can_stop_idle_tick(cpu, ts))
1050 next_event = tick_nohz_next_event(ts, cpu);
1055 * If the next highres timer to expire is earlier than next_event, the
1056 * idle governor needs to know that.
1058 next_event = min_t(u64, next_event,
1059 hrtimer_next_event_without(&ts->sched_timer));
1061 return ktime_sub(next_event, now);
1065 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1066 * for a particular CPU.
1068 * Called from the schedutil frequency scaling governor in scheduler context.
1070 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1072 struct tick_sched *ts = tick_get_tick_sched(cpu);
1074 return ts->idle_calls;
1078 * tick_nohz_get_idle_calls - return the current idle calls counter value
1080 * Called from the schedutil frequency scaling governor in scheduler context.
1082 unsigned long tick_nohz_get_idle_calls(void)
1084 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1086 return ts->idle_calls;
1089 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1091 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1092 unsigned long ticks;
1094 if (vtime_accounting_cpu_enabled())
1097 * We stopped the tick in idle. Update process times would miss the
1098 * time we slept as update_process_times does only a 1 tick
1099 * accounting. Enforce that this is accounted to idle !
1101 ticks = jiffies - ts->idle_jiffies;
1103 * We might be one off. Do not randomly account a huge number of ticks!
1105 if (ticks && ticks < LONG_MAX)
1106 account_idle_ticks(ticks);
1110 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1112 tick_nohz_restart_sched_tick(ts, now);
1113 tick_nohz_account_idle_ticks(ts);
1116 void tick_nohz_idle_restart_tick(void)
1118 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1120 if (ts->tick_stopped)
1121 __tick_nohz_idle_restart_tick(ts, ktime_get());
1125 * tick_nohz_idle_exit - restart the idle tick from the idle task
1127 * Restart the idle tick when the CPU is woken up from idle
1128 * This also exit the RCU extended quiescent state. The CPU
1129 * can use RCU again after this function is called.
1131 void tick_nohz_idle_exit(void)
1133 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1134 bool idle_active, tick_stopped;
1137 local_irq_disable();
1139 WARN_ON_ONCE(!ts->inidle);
1140 WARN_ON_ONCE(ts->timer_expires_base);
1143 idle_active = ts->idle_active;
1144 tick_stopped = ts->tick_stopped;
1146 if (idle_active || tick_stopped)
1150 tick_nohz_stop_idle(ts, now);
1153 __tick_nohz_idle_restart_tick(ts, now);
1159 * The nohz low res interrupt handler
1161 static void tick_nohz_handler(struct clock_event_device *dev)
1163 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1164 struct pt_regs *regs = get_irq_regs();
1165 ktime_t now = ktime_get();
1167 dev->next_event = KTIME_MAX;
1169 tick_sched_do_timer(ts, now);
1170 tick_sched_handle(ts, regs);
1172 /* No need to reprogram if we are running tickless */
1173 if (unlikely(ts->tick_stopped))
1176 hrtimer_forward(&ts->sched_timer, now, tick_period);
1177 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1180 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1182 if (!tick_nohz_enabled)
1184 ts->nohz_mode = mode;
1185 /* One update is enough */
1186 if (!test_and_set_bit(0, &tick_nohz_active))
1187 timers_update_nohz();
1191 * tick_nohz_switch_to_nohz - switch to nohz mode
1193 static void tick_nohz_switch_to_nohz(void)
1195 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1198 if (!tick_nohz_enabled)
1201 if (tick_switch_to_oneshot(tick_nohz_handler))
1205 * Recycle the hrtimer in ts, so we can share the
1206 * hrtimer_forward with the highres code.
1208 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1209 /* Get the next period */
1210 next = tick_init_jiffy_update();
1212 hrtimer_set_expires(&ts->sched_timer, next);
1213 hrtimer_forward_now(&ts->sched_timer, tick_period);
1214 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1215 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1218 static inline void tick_nohz_irq_enter(void)
1220 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 if (!ts->idle_active && !ts->tick_stopped)
1226 if (ts->idle_active)
1227 tick_nohz_stop_idle(ts, now);
1228 if (ts->tick_stopped)
1229 tick_nohz_update_jiffies(now);
1234 static inline void tick_nohz_switch_to_nohz(void) { }
1235 static inline void tick_nohz_irq_enter(void) { }
1236 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1238 #endif /* CONFIG_NO_HZ_COMMON */
1241 * Called from irq_enter to notify about the possible interruption of idle()
1243 void tick_irq_enter(void)
1245 tick_check_oneshot_broadcast_this_cpu();
1246 tick_nohz_irq_enter();
1250 * High resolution timer specific code
1252 #ifdef CONFIG_HIGH_RES_TIMERS
1254 * We rearm the timer until we get disabled by the idle code.
1255 * Called with interrupts disabled.
1257 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1259 struct tick_sched *ts =
1260 container_of(timer, struct tick_sched, sched_timer);
1261 struct pt_regs *regs = get_irq_regs();
1262 ktime_t now = ktime_get();
1264 tick_sched_do_timer(ts, now);
1267 * Do not call, when we are not in irq context and have
1268 * no valid regs pointer
1271 tick_sched_handle(ts, regs);
1275 /* No need to reprogram if we are in idle or full dynticks mode */
1276 if (unlikely(ts->tick_stopped))
1277 return HRTIMER_NORESTART;
1279 hrtimer_forward(timer, now, tick_period);
1281 return HRTIMER_RESTART;
1284 static int sched_skew_tick;
1286 static int __init skew_tick(char *str)
1288 get_option(&str, &sched_skew_tick);
1292 early_param("skew_tick", skew_tick);
1295 * tick_setup_sched_timer - setup the tick emulation timer
1297 void tick_setup_sched_timer(void)
1299 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1300 ktime_t now = ktime_get();
1303 * Emulate tick processing via per-CPU hrtimers:
1305 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1306 ts->sched_timer.function = tick_sched_timer;
1308 /* Get the next period (per-CPU) */
1309 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1311 /* Offset the tick to avert jiffies_lock contention. */
1312 if (sched_skew_tick) {
1313 u64 offset = ktime_to_ns(tick_period) >> 1;
1314 do_div(offset, num_possible_cpus());
1315 offset *= smp_processor_id();
1316 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1319 hrtimer_forward(&ts->sched_timer, now, tick_period);
1320 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1321 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1323 #endif /* HIGH_RES_TIMERS */
1325 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1326 void tick_cancel_sched_timer(int cpu)
1328 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1330 # ifdef CONFIG_HIGH_RES_TIMERS
1331 if (ts->sched_timer.base)
1332 hrtimer_cancel(&ts->sched_timer);
1335 memset(ts, 0, sizeof(*ts));
1340 * Async notification about clocksource changes
1342 void tick_clock_notify(void)
1346 for_each_possible_cpu(cpu)
1347 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1351 * Async notification about clock event changes
1353 void tick_oneshot_notify(void)
1355 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1357 set_bit(0, &ts->check_clocks);
1361 * Check, if a change happened, which makes oneshot possible.
1363 * Called cyclic from the hrtimer softirq (driven by the timer
1364 * softirq) allow_nohz signals, that we can switch into low-res nohz
1365 * mode, because high resolution timers are disabled (either compile
1366 * or runtime). Called with interrupts disabled.
1368 int tick_check_oneshot_change(int allow_nohz)
1370 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1372 if (!test_and_clear_bit(0, &ts->check_clocks))
1375 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1378 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1384 tick_nohz_switch_to_nohz();