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:
61 * The READ_ONCE() pairs with two updates done later in this function.
63 delta = ktime_sub(now, READ_ONCE(last_jiffies_update));
64 if (delta < tick_period)
67 /* Reevaluate with jiffies_lock held */
68 raw_spin_lock(&jiffies_lock);
69 write_seqcount_begin(&jiffies_seq);
71 delta = ktime_sub(now, last_jiffies_update);
72 if (delta >= tick_period) {
74 delta = ktime_sub(delta, tick_period);
75 /* Pairs with the lockless read in this function. */
76 WRITE_ONCE(last_jiffies_update,
77 ktime_add(last_jiffies_update, tick_period));
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 /* Pairs with the lockless read in this function. */
86 WRITE_ONCE(last_jiffies_update,
87 ktime_add_ns(last_jiffies_update,
92 /* Keep the tick_next_period variable up to date */
93 tick_next_period = ktime_add(last_jiffies_update, tick_period);
95 write_seqcount_end(&jiffies_seq);
96 raw_spin_unlock(&jiffies_lock);
99 write_seqcount_end(&jiffies_seq);
100 raw_spin_unlock(&jiffies_lock);
105 * Initialize and return retrieve the jiffies update.
107 static ktime_t tick_init_jiffy_update(void)
111 raw_spin_lock(&jiffies_lock);
112 write_seqcount_begin(&jiffies_seq);
113 /* Did we start the jiffies update yet ? */
114 if (last_jiffies_update == 0)
115 last_jiffies_update = tick_next_period;
116 period = last_jiffies_update;
117 write_seqcount_end(&jiffies_seq);
118 raw_spin_unlock(&jiffies_lock);
122 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
124 int cpu = smp_processor_id();
126 #ifdef CONFIG_NO_HZ_COMMON
128 * Check if the do_timer duty was dropped. We don't care about
129 * concurrency: This happens only when the CPU in charge went
130 * into a long sleep. If two CPUs happen to assign themselves to
131 * this duty, then the jiffies update is still serialized by
134 * If nohz_full is enabled, this should not happen because the
135 * tick_do_timer_cpu never relinquishes.
137 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
138 #ifdef CONFIG_NO_HZ_FULL
139 WARN_ON(tick_nohz_full_running);
141 tick_do_timer_cpu = cpu;
145 /* Check, if the jiffies need an update */
146 if (tick_do_timer_cpu == cpu)
147 tick_do_update_jiffies64(now);
150 ts->got_idle_tick = 1;
153 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
155 #ifdef CONFIG_NO_HZ_COMMON
157 * When we are idle and the tick is stopped, we have to touch
158 * the watchdog as we might not schedule for a really long
159 * time. This happens on complete idle SMP systems while
160 * waiting on the login prompt. We also increment the "start of
161 * idle" jiffy stamp so the idle accounting adjustment we do
162 * when we go busy again does not account too much ticks.
164 if (ts->tick_stopped) {
165 touch_softlockup_watchdog_sched();
166 if (is_idle_task(current))
169 * In case the current tick fired too early past its expected
170 * expiration, make sure we don't bypass the next clock reprogramming
171 * to the same deadline.
176 update_process_times(user_mode(regs));
177 profile_tick(CPU_PROFILING);
181 #ifdef CONFIG_NO_HZ_FULL
182 cpumask_var_t tick_nohz_full_mask;
183 bool tick_nohz_full_running;
184 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
185 static atomic_t tick_dep_mask;
187 static bool check_tick_dependency(atomic_t *dep)
189 int val = atomic_read(dep);
191 if (val & TICK_DEP_MASK_POSIX_TIMER) {
192 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
196 if (val & TICK_DEP_MASK_PERF_EVENTS) {
197 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
201 if (val & TICK_DEP_MASK_SCHED) {
202 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
206 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
207 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
211 if (val & TICK_DEP_MASK_RCU) {
212 trace_tick_stop(0, TICK_DEP_MASK_RCU);
219 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
221 lockdep_assert_irqs_disabled();
223 if (unlikely(!cpu_online(cpu)))
226 if (check_tick_dependency(&tick_dep_mask))
229 if (check_tick_dependency(&ts->tick_dep_mask))
232 if (check_tick_dependency(¤t->tick_dep_mask))
235 if (check_tick_dependency(¤t->signal->tick_dep_mask))
241 static void nohz_full_kick_func(struct irq_work *work)
243 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
246 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
247 .func = nohz_full_kick_func,
248 .flags = ATOMIC_INIT(IRQ_WORK_HARD_IRQ),
252 * Kick this CPU if it's full dynticks in order to force it to
253 * re-evaluate its dependency on the tick and restart it if necessary.
254 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
257 static void tick_nohz_full_kick(void)
259 if (!tick_nohz_full_cpu(smp_processor_id()))
262 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
266 * Kick the CPU if it's full dynticks in order to force it to
267 * re-evaluate its dependency on the tick and restart it if necessary.
269 void tick_nohz_full_kick_cpu(int cpu)
271 if (!tick_nohz_full_cpu(cpu))
274 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
278 * Kick all full dynticks CPUs in order to force these to re-evaluate
279 * their dependency on the tick and restart it if necessary.
281 static void tick_nohz_full_kick_all(void)
285 if (!tick_nohz_full_running)
289 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
290 tick_nohz_full_kick_cpu(cpu);
294 static void tick_nohz_dep_set_all(atomic_t *dep,
295 enum tick_dep_bits bit)
299 prev = atomic_fetch_or(BIT(bit), dep);
301 tick_nohz_full_kick_all();
305 * Set a global tick dependency. Used by perf events that rely on freq and
308 void tick_nohz_dep_set(enum tick_dep_bits bit)
310 tick_nohz_dep_set_all(&tick_dep_mask, bit);
313 void tick_nohz_dep_clear(enum tick_dep_bits bit)
315 atomic_andnot(BIT(bit), &tick_dep_mask);
319 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
320 * manage events throttling.
322 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
325 struct tick_sched *ts;
327 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
329 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
332 /* Perf needs local kick that is NMI safe */
333 if (cpu == smp_processor_id()) {
334 tick_nohz_full_kick();
336 /* Remote irq work not NMI-safe */
337 if (!WARN_ON_ONCE(in_nmi()))
338 tick_nohz_full_kick_cpu(cpu);
343 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
345 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
347 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
349 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
351 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
354 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
355 * in order to elapse per task timers.
357 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
359 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask)) {
360 if (tsk == current) {
362 tick_nohz_full_kick();
366 * Some future tick_nohz_full_kick_task()
367 * should optimize this.
369 tick_nohz_full_kick_all();
373 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
375 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
377 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
379 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
382 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
383 * per process timers.
385 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
387 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
390 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
392 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
396 * Re-evaluate the need for the tick as we switch the current task.
397 * It might need the tick due to per task/process properties:
398 * perf events, posix CPU timers, ...
400 void __tick_nohz_task_switch(void)
403 struct tick_sched *ts;
405 local_irq_save(flags);
407 if (!tick_nohz_full_cpu(smp_processor_id()))
410 ts = this_cpu_ptr(&tick_cpu_sched);
412 if (ts->tick_stopped) {
413 if (atomic_read(¤t->tick_dep_mask) ||
414 atomic_read(¤t->signal->tick_dep_mask))
415 tick_nohz_full_kick();
418 local_irq_restore(flags);
421 /* Get the boot-time nohz CPU list from the kernel parameters. */
422 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
424 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
425 cpumask_copy(tick_nohz_full_mask, cpumask);
426 tick_nohz_full_running = true;
428 EXPORT_SYMBOL_GPL(tick_nohz_full_setup);
430 static int tick_nohz_cpu_down(unsigned int cpu)
433 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
434 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
435 * CPUs. It must remain online when nohz full is enabled.
437 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
442 void __init tick_nohz_init(void)
446 if (!tick_nohz_full_running)
450 * Full dynticks uses irq work to drive the tick rescheduling on safe
451 * locking contexts. But then we need irq work to raise its own
452 * interrupts to avoid circular dependency on the tick
454 if (!arch_irq_work_has_interrupt()) {
455 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
456 cpumask_clear(tick_nohz_full_mask);
457 tick_nohz_full_running = false;
461 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
462 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
463 cpu = smp_processor_id();
465 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
466 pr_warn("NO_HZ: Clearing %d from nohz_full range "
467 "for timekeeping\n", cpu);
468 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
472 for_each_cpu(cpu, tick_nohz_full_mask)
473 context_tracking_cpu_set(cpu);
475 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
476 "kernel/nohz:predown", NULL,
479 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
480 cpumask_pr_args(tick_nohz_full_mask));
485 * NOHZ - aka dynamic tick functionality
487 #ifdef CONFIG_NO_HZ_COMMON
491 bool tick_nohz_enabled __read_mostly = true;
492 unsigned long tick_nohz_active __read_mostly;
494 * Enable / Disable tickless mode
496 static int __init setup_tick_nohz(char *str)
498 return (kstrtobool(str, &tick_nohz_enabled) == 0);
501 __setup("nohz=", setup_tick_nohz);
503 bool tick_nohz_tick_stopped(void)
505 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
507 return ts->tick_stopped;
510 bool tick_nohz_tick_stopped_cpu(int cpu)
512 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
514 return ts->tick_stopped;
518 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
520 * Called from interrupt entry when the CPU was idle
522 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
523 * must be updated. Otherwise an interrupt handler could use a stale jiffy
524 * value. We do this unconditionally on any CPU, as we don't know whether the
525 * CPU, which has the update task assigned is in a long sleep.
527 static void tick_nohz_update_jiffies(ktime_t now)
531 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
533 local_irq_save(flags);
534 tick_do_update_jiffies64(now);
535 local_irq_restore(flags);
537 touch_softlockup_watchdog_sched();
541 * Updates the per-CPU time idle statistics counters
544 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
548 if (ts->idle_active) {
549 delta = ktime_sub(now, ts->idle_entrytime);
550 if (nr_iowait_cpu(cpu) > 0)
551 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
553 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
554 ts->idle_entrytime = now;
557 if (last_update_time)
558 *last_update_time = ktime_to_us(now);
562 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
564 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
567 sched_clock_idle_wakeup_event();
570 static void tick_nohz_start_idle(struct tick_sched *ts)
572 ts->idle_entrytime = ktime_get();
574 sched_clock_idle_sleep_event();
578 * get_cpu_idle_time_us - get the total idle time of a CPU
579 * @cpu: CPU number to query
580 * @last_update_time: variable to store update time in. Do not update
583 * Return the cumulative idle time (since boot) for a given
584 * CPU, in microseconds.
586 * This time is measured via accounting rather than sampling,
587 * and is as accurate as ktime_get() is.
589 * This function returns -1 if NOHZ is not enabled.
591 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
593 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
596 if (!tick_nohz_active)
600 if (last_update_time) {
601 update_ts_time_stats(cpu, ts, now, last_update_time);
602 idle = ts->idle_sleeptime;
604 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
605 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
607 idle = ktime_add(ts->idle_sleeptime, delta);
609 idle = ts->idle_sleeptime;
613 return ktime_to_us(idle);
616 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
619 * get_cpu_iowait_time_us - get the total iowait time of a CPU
620 * @cpu: CPU number to query
621 * @last_update_time: variable to store update time in. Do not update
624 * Return the cumulative iowait time (since boot) for a given
625 * CPU, in microseconds.
627 * This time is measured via accounting rather than sampling,
628 * and is as accurate as ktime_get() is.
630 * This function returns -1 if NOHZ is not enabled.
632 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
634 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
637 if (!tick_nohz_active)
641 if (last_update_time) {
642 update_ts_time_stats(cpu, ts, now, last_update_time);
643 iowait = ts->iowait_sleeptime;
645 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
646 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
648 iowait = ktime_add(ts->iowait_sleeptime, delta);
650 iowait = ts->iowait_sleeptime;
654 return ktime_to_us(iowait);
656 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
658 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
660 hrtimer_cancel(&ts->sched_timer);
661 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
663 /* Forward the time to expire in the future */
664 hrtimer_forward(&ts->sched_timer, now, tick_period);
666 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
667 hrtimer_start_expires(&ts->sched_timer,
668 HRTIMER_MODE_ABS_PINNED_HARD);
670 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
674 * Reset to make sure next tick stop doesn't get fooled by past
675 * cached clock deadline.
680 static inline bool local_timer_softirq_pending(void)
682 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
685 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
687 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
688 unsigned long basejiff;
691 /* Read jiffies and the time when jiffies were updated last */
693 seq = read_seqcount_begin(&jiffies_seq);
694 basemono = last_jiffies_update;
696 } while (read_seqcount_retry(&jiffies_seq, seq));
697 ts->last_jiffies = basejiff;
698 ts->timer_expires_base = basemono;
701 * Keep the periodic tick, when RCU, architecture or irq_work
703 * Aside of that check whether the local timer softirq is
704 * pending. If so its a bad idea to call get_next_timer_interrupt()
705 * because there is an already expired timer, so it will request
706 * immeditate expiry, which rearms the hardware timer with a
707 * minimal delta which brings us back to this place
708 * immediately. Lather, rinse and repeat...
710 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
711 irq_work_needs_cpu() || local_timer_softirq_pending()) {
712 next_tick = basemono + TICK_NSEC;
715 * Get the next pending timer. If high resolution
716 * timers are enabled this only takes the timer wheel
717 * timers into account. If high resolution timers are
718 * disabled this also looks at the next expiring
721 next_tmr = get_next_timer_interrupt(basejiff, basemono);
722 ts->next_timer = next_tmr;
723 /* Take the next rcu event into account */
724 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
728 * If the tick is due in the next period, keep it ticking or
729 * force prod the timer.
731 delta = next_tick - basemono;
732 if (delta <= (u64)TICK_NSEC) {
734 * Tell the timer code that the base is not idle, i.e. undo
735 * the effect of get_next_timer_interrupt():
739 * We've not stopped the tick yet, and there's a timer in the
740 * next period, so no point in stopping it either, bail.
742 if (!ts->tick_stopped) {
743 ts->timer_expires = 0;
749 * If this CPU is the one which had the do_timer() duty last, we limit
750 * the sleep time to the timekeeping max_deferment value.
751 * Otherwise we can sleep as long as we want.
753 delta = timekeeping_max_deferment();
754 if (cpu != tick_do_timer_cpu &&
755 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
758 /* Calculate the next expiry time */
759 if (delta < (KTIME_MAX - basemono))
760 expires = basemono + delta;
764 ts->timer_expires = min_t(u64, expires, next_tick);
767 return ts->timer_expires;
770 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
772 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
773 u64 basemono = ts->timer_expires_base;
774 u64 expires = ts->timer_expires;
775 ktime_t tick = expires;
777 /* Make sure we won't be trying to stop it twice in a row. */
778 ts->timer_expires_base = 0;
781 * If this CPU is the one which updates jiffies, then give up
782 * the assignment and let it be taken by the CPU which runs
783 * the tick timer next, which might be this CPU as well. If we
784 * don't drop this here the jiffies might be stale and
785 * do_timer() never invoked. Keep track of the fact that it
786 * was the one which had the do_timer() duty last.
788 if (cpu == tick_do_timer_cpu) {
789 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
790 ts->do_timer_last = 1;
791 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
792 ts->do_timer_last = 0;
795 /* Skip reprogram of event if its not changed */
796 if (ts->tick_stopped && (expires == ts->next_tick)) {
797 /* Sanity check: make sure clockevent is actually programmed */
798 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
802 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
803 basemono, ts->next_tick, dev->next_event,
804 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
808 * nohz_stop_sched_tick can be called several times before
809 * the nohz_restart_sched_tick is called. This happens when
810 * interrupts arrive which do not cause a reschedule. In the
811 * first call we save the current tick time, so we can restart
812 * the scheduler tick in nohz_restart_sched_tick.
814 if (!ts->tick_stopped) {
815 calc_load_nohz_start();
818 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
819 ts->tick_stopped = 1;
820 trace_tick_stop(1, TICK_DEP_MASK_NONE);
823 ts->next_tick = tick;
826 * If the expiration time == KTIME_MAX, then we simply stop
829 if (unlikely(expires == KTIME_MAX)) {
830 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
831 hrtimer_cancel(&ts->sched_timer);
835 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
836 hrtimer_start(&ts->sched_timer, tick,
837 HRTIMER_MODE_ABS_PINNED_HARD);
839 hrtimer_set_expires(&ts->sched_timer, tick);
840 tick_program_event(tick, 1);
844 static void tick_nohz_retain_tick(struct tick_sched *ts)
846 ts->timer_expires_base = 0;
849 #ifdef CONFIG_NO_HZ_FULL
850 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
852 if (tick_nohz_next_event(ts, cpu))
853 tick_nohz_stop_tick(ts, cpu);
855 tick_nohz_retain_tick(ts);
857 #endif /* CONFIG_NO_HZ_FULL */
859 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
861 /* Update jiffies first */
862 tick_do_update_jiffies64(now);
865 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
866 * the clock forward checks in the enqueue path:
870 calc_load_nohz_stop();
871 touch_softlockup_watchdog_sched();
873 * Cancel the scheduled timer and restore the tick
875 ts->tick_stopped = 0;
876 ts->idle_exittime = now;
878 tick_nohz_restart(ts, now);
881 static void tick_nohz_full_update_tick(struct tick_sched *ts)
883 #ifdef CONFIG_NO_HZ_FULL
884 int cpu = smp_processor_id();
886 if (!tick_nohz_full_cpu(cpu))
889 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
892 if (can_stop_full_tick(cpu, ts))
893 tick_nohz_stop_sched_tick(ts, cpu);
894 else if (ts->tick_stopped)
895 tick_nohz_restart_sched_tick(ts, ktime_get());
899 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
902 * If this CPU is offline and it is the one which updates
903 * jiffies, then give up the assignment and let it be taken by
904 * the CPU which runs the tick timer next. If we don't drop
905 * this here the jiffies might be stale and do_timer() never
908 if (unlikely(!cpu_online(cpu))) {
909 if (cpu == tick_do_timer_cpu)
910 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
912 * Make sure the CPU doesn't get fooled by obsolete tick
913 * deadline if it comes back online later.
919 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
925 if (unlikely(local_softirq_pending())) {
926 static int ratelimit;
928 if (ratelimit < 10 &&
929 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
930 pr_warn("NOHZ: local_softirq_pending %02x\n",
931 (unsigned int) local_softirq_pending());
937 if (tick_nohz_full_enabled()) {
939 * Keep the tick alive to guarantee timekeeping progression
940 * if there are full dynticks CPUs around
942 if (tick_do_timer_cpu == cpu)
945 * Boot safety: make sure the timekeeping duty has been
946 * assigned before entering dyntick-idle mode,
947 * tick_do_timer_cpu is TICK_DO_TIMER_BOOT
949 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_BOOT))
952 /* Should not happen for nohz-full */
953 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
960 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
963 int cpu = smp_processor_id();
966 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
967 * tick timer expiration time is known already.
969 if (ts->timer_expires_base)
970 expires = ts->timer_expires;
971 else if (can_stop_idle_tick(cpu, ts))
972 expires = tick_nohz_next_event(ts, cpu);
979 int was_stopped = ts->tick_stopped;
981 tick_nohz_stop_tick(ts, cpu);
984 ts->idle_expires = expires;
986 if (!was_stopped && ts->tick_stopped) {
987 ts->idle_jiffies = ts->last_jiffies;
988 nohz_balance_enter_idle(cpu);
991 tick_nohz_retain_tick(ts);
996 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
998 * When the next event is more than a tick into the future, stop the idle tick
1000 void tick_nohz_idle_stop_tick(void)
1002 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1005 void tick_nohz_idle_retain_tick(void)
1007 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1009 * Undo the effect of get_next_timer_interrupt() called from
1010 * tick_nohz_next_event().
1016 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1018 * Called when we start the idle loop.
1020 void tick_nohz_idle_enter(void)
1022 struct tick_sched *ts;
1024 lockdep_assert_irqs_enabled();
1026 local_irq_disable();
1028 ts = this_cpu_ptr(&tick_cpu_sched);
1030 WARN_ON_ONCE(ts->timer_expires_base);
1033 tick_nohz_start_idle(ts);
1039 * tick_nohz_irq_exit - update next tick event from interrupt exit
1041 * When an interrupt fires while we are idle and it doesn't cause
1042 * a reschedule, it may still add, modify or delete a timer, enqueue
1043 * an RCU callback, etc...
1044 * So we need to re-calculate and reprogram the next tick event.
1046 void tick_nohz_irq_exit(void)
1048 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1051 tick_nohz_start_idle(ts);
1053 tick_nohz_full_update_tick(ts);
1057 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1059 bool tick_nohz_idle_got_tick(void)
1061 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1063 if (ts->got_idle_tick) {
1064 ts->got_idle_tick = 0;
1071 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1072 * or the tick, whatever that expires first. Note that, if the tick has been
1073 * stopped, it returns the next hrtimer.
1075 * Called from power state control code with interrupts disabled
1077 ktime_t tick_nohz_get_next_hrtimer(void)
1079 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1083 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1084 * @delta_next: duration until the next event if the tick cannot be stopped
1086 * Called from power state control code with interrupts disabled
1088 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1090 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1091 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1092 int cpu = smp_processor_id();
1094 * The idle entry time is expected to be a sufficient approximation of
1095 * the current time at this point.
1097 ktime_t now = ts->idle_entrytime;
1100 WARN_ON_ONCE(!ts->inidle);
1102 *delta_next = ktime_sub(dev->next_event, now);
1104 if (!can_stop_idle_tick(cpu, ts))
1107 next_event = tick_nohz_next_event(ts, cpu);
1112 * If the next highres timer to expire is earlier than next_event, the
1113 * idle governor needs to know that.
1115 next_event = min_t(u64, next_event,
1116 hrtimer_next_event_without(&ts->sched_timer));
1118 return ktime_sub(next_event, now);
1122 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1123 * for a particular CPU.
1125 * Called from the schedutil frequency scaling governor in scheduler context.
1127 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1129 struct tick_sched *ts = tick_get_tick_sched(cpu);
1131 return ts->idle_calls;
1135 * tick_nohz_get_idle_calls - return the current idle calls counter value
1137 * Called from the schedutil frequency scaling governor in scheduler context.
1139 unsigned long tick_nohz_get_idle_calls(void)
1141 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1143 return ts->idle_calls;
1146 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1148 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1149 unsigned long ticks;
1151 if (vtime_accounting_enabled_this_cpu())
1154 * We stopped the tick in idle. Update process times would miss the
1155 * time we slept as update_process_times does only a 1 tick
1156 * accounting. Enforce that this is accounted to idle !
1158 ticks = jiffies - ts->idle_jiffies;
1160 * We might be one off. Do not randomly account a huge number of ticks!
1162 if (ticks && ticks < LONG_MAX)
1163 account_idle_ticks(ticks);
1167 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1169 tick_nohz_restart_sched_tick(ts, now);
1170 tick_nohz_account_idle_ticks(ts);
1173 void tick_nohz_idle_restart_tick(void)
1175 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1177 if (ts->tick_stopped)
1178 __tick_nohz_idle_restart_tick(ts, ktime_get());
1182 * tick_nohz_idle_exit - restart the idle tick from the idle task
1184 * Restart the idle tick when the CPU is woken up from idle
1185 * This also exit the RCU extended quiescent state. The CPU
1186 * can use RCU again after this function is called.
1188 void tick_nohz_idle_exit(void)
1190 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1191 bool idle_active, tick_stopped;
1194 local_irq_disable();
1196 WARN_ON_ONCE(!ts->inidle);
1197 WARN_ON_ONCE(ts->timer_expires_base);
1200 idle_active = ts->idle_active;
1201 tick_stopped = ts->tick_stopped;
1203 if (idle_active || tick_stopped)
1207 tick_nohz_stop_idle(ts, now);
1210 __tick_nohz_idle_restart_tick(ts, now);
1216 * The nohz low res interrupt handler
1218 static void tick_nohz_handler(struct clock_event_device *dev)
1220 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1221 struct pt_regs *regs = get_irq_regs();
1222 ktime_t now = ktime_get();
1224 dev->next_event = KTIME_MAX;
1226 tick_sched_do_timer(ts, now);
1227 tick_sched_handle(ts, regs);
1229 /* No need to reprogram if we are running tickless */
1230 if (unlikely(ts->tick_stopped))
1233 hrtimer_forward(&ts->sched_timer, now, tick_period);
1234 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1237 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1239 if (!tick_nohz_enabled)
1241 ts->nohz_mode = mode;
1242 /* One update is enough */
1243 if (!test_and_set_bit(0, &tick_nohz_active))
1244 timers_update_nohz();
1248 * tick_nohz_switch_to_nohz - switch to nohz mode
1250 static void tick_nohz_switch_to_nohz(void)
1252 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1255 if (!tick_nohz_enabled)
1258 if (tick_switch_to_oneshot(tick_nohz_handler))
1262 * Recycle the hrtimer in ts, so we can share the
1263 * hrtimer_forward with the highres code.
1265 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1266 /* Get the next period */
1267 next = tick_init_jiffy_update();
1269 hrtimer_set_expires(&ts->sched_timer, next);
1270 hrtimer_forward_now(&ts->sched_timer, tick_period);
1271 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1272 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1275 static inline void tick_nohz_irq_enter(void)
1277 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1280 if (!ts->idle_active && !ts->tick_stopped)
1283 if (ts->idle_active)
1284 tick_nohz_stop_idle(ts, now);
1285 if (ts->tick_stopped)
1286 tick_nohz_update_jiffies(now);
1291 static inline void tick_nohz_switch_to_nohz(void) { }
1292 static inline void tick_nohz_irq_enter(void) { }
1293 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1295 #endif /* CONFIG_NO_HZ_COMMON */
1298 * Called from irq_enter to notify about the possible interruption of idle()
1300 void tick_irq_enter(void)
1302 tick_check_oneshot_broadcast_this_cpu();
1303 tick_nohz_irq_enter();
1307 * High resolution timer specific code
1309 #ifdef CONFIG_HIGH_RES_TIMERS
1311 * We rearm the timer until we get disabled by the idle code.
1312 * Called with interrupts disabled.
1314 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1316 struct tick_sched *ts =
1317 container_of(timer, struct tick_sched, sched_timer);
1318 struct pt_regs *regs = get_irq_regs();
1319 ktime_t now = ktime_get();
1321 tick_sched_do_timer(ts, now);
1324 * Do not call, when we are not in irq context and have
1325 * no valid regs pointer
1328 tick_sched_handle(ts, regs);
1332 /* No need to reprogram if we are in idle or full dynticks mode */
1333 if (unlikely(ts->tick_stopped))
1334 return HRTIMER_NORESTART;
1336 hrtimer_forward(timer, now, tick_period);
1338 return HRTIMER_RESTART;
1341 static int sched_skew_tick;
1343 static int __init skew_tick(char *str)
1345 get_option(&str, &sched_skew_tick);
1349 early_param("skew_tick", skew_tick);
1352 * tick_setup_sched_timer - setup the tick emulation timer
1354 void tick_setup_sched_timer(void)
1356 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1357 ktime_t now = ktime_get();
1360 * Emulate tick processing via per-CPU hrtimers:
1362 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1363 ts->sched_timer.function = tick_sched_timer;
1365 /* Get the next period (per-CPU) */
1366 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1368 /* Offset the tick to avert jiffies_lock contention. */
1369 if (sched_skew_tick) {
1370 u64 offset = ktime_to_ns(tick_period) >> 1;
1371 do_div(offset, num_possible_cpus());
1372 offset *= smp_processor_id();
1373 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1376 hrtimer_forward(&ts->sched_timer, now, tick_period);
1377 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1378 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1380 #endif /* HIGH_RES_TIMERS */
1382 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1383 void tick_cancel_sched_timer(int cpu)
1385 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1387 # ifdef CONFIG_HIGH_RES_TIMERS
1388 if (ts->sched_timer.base)
1389 hrtimer_cancel(&ts->sched_timer);
1392 memset(ts, 0, sizeof(*ts));
1397 * Async notification about clocksource changes
1399 void tick_clock_notify(void)
1403 for_each_possible_cpu(cpu)
1404 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1408 * Async notification about clock event changes
1410 void tick_oneshot_notify(void)
1412 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1414 set_bit(0, &ts->check_clocks);
1418 * Check, if a change happened, which makes oneshot possible.
1420 * Called cyclic from the hrtimer softirq (driven by the timer
1421 * softirq) allow_nohz signals, that we can switch into low-res nohz
1422 * mode, because high resolution timers are disabled (either compile
1423 * or runtime). Called with interrupts disabled.
1425 int tick_check_oneshot_change(int allow_nohz)
1427 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1429 if (!test_and_clear_bit(0, &ts->check_clocks))
1432 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1435 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1441 tick_nohz_switch_to_nohz();