1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
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 nohz_full is enabled, this should not happen because the
126 * tick_do_timer_cpu never relinquishes.
128 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
129 #ifdef CONFIG_NO_HZ_FULL
130 WARN_ON(tick_nohz_full_running);
132 tick_do_timer_cpu = cpu;
136 /* Check, if the jiffies need an update */
137 if (tick_do_timer_cpu == cpu)
138 tick_do_update_jiffies64(now);
141 ts->got_idle_tick = 1;
144 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
146 #ifdef CONFIG_NO_HZ_COMMON
148 * When we are idle and the tick is stopped, we have to touch
149 * the watchdog as we might not schedule for a really long
150 * time. This happens on complete idle SMP systems while
151 * waiting on the login prompt. We also increment the "start of
152 * idle" jiffy stamp so the idle accounting adjustment we do
153 * when we go busy again does not account too much ticks.
155 if (ts->tick_stopped) {
156 touch_softlockup_watchdog_sched();
157 if (is_idle_task(current))
160 * In case the current tick fired too early past its expected
161 * expiration, make sure we don't bypass the next clock reprogramming
162 * to the same deadline.
167 update_process_times(user_mode(regs));
168 profile_tick(CPU_PROFILING);
172 #ifdef CONFIG_NO_HZ_FULL
173 cpumask_var_t tick_nohz_full_mask;
174 bool tick_nohz_full_running;
175 static atomic_t tick_dep_mask;
177 static bool check_tick_dependency(atomic_t *dep)
179 int val = atomic_read(dep);
181 if (val & TICK_DEP_MASK_POSIX_TIMER) {
182 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
186 if (val & TICK_DEP_MASK_PERF_EVENTS) {
187 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
191 if (val & TICK_DEP_MASK_SCHED) {
192 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
196 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
197 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
204 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
206 lockdep_assert_irqs_disabled();
208 if (unlikely(!cpu_online(cpu)))
211 if (check_tick_dependency(&tick_dep_mask))
214 if (check_tick_dependency(&ts->tick_dep_mask))
217 if (check_tick_dependency(¤t->tick_dep_mask))
220 if (check_tick_dependency(¤t->signal->tick_dep_mask))
226 static void nohz_full_kick_func(struct irq_work *work)
228 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
231 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
232 .func = nohz_full_kick_func,
236 * Kick this CPU if it's full dynticks in order to force it to
237 * re-evaluate its dependency on the tick and restart it if necessary.
238 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
241 static void tick_nohz_full_kick(void)
243 if (!tick_nohz_full_cpu(smp_processor_id()))
246 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
250 * Kick the CPU if it's full dynticks in order to force it to
251 * re-evaluate its dependency on the tick and restart it if necessary.
253 void tick_nohz_full_kick_cpu(int cpu)
255 if (!tick_nohz_full_cpu(cpu))
258 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
262 * Kick all full dynticks CPUs in order to force these to re-evaluate
263 * their dependency on the tick and restart it if necessary.
265 static void tick_nohz_full_kick_all(void)
269 if (!tick_nohz_full_running)
273 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
274 tick_nohz_full_kick_cpu(cpu);
278 static void tick_nohz_dep_set_all(atomic_t *dep,
279 enum tick_dep_bits bit)
283 prev = atomic_fetch_or(BIT(bit), dep);
285 tick_nohz_full_kick_all();
289 * Set a global tick dependency. Used by perf events that rely on freq and
292 void tick_nohz_dep_set(enum tick_dep_bits bit)
294 tick_nohz_dep_set_all(&tick_dep_mask, bit);
297 void tick_nohz_dep_clear(enum tick_dep_bits bit)
299 atomic_andnot(BIT(bit), &tick_dep_mask);
303 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
304 * manage events throttling.
306 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
309 struct tick_sched *ts;
311 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
313 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
316 /* Perf needs local kick that is NMI safe */
317 if (cpu == smp_processor_id()) {
318 tick_nohz_full_kick();
320 /* Remote irq work not NMI-safe */
321 if (!WARN_ON_ONCE(in_nmi()))
322 tick_nohz_full_kick_cpu(cpu);
328 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
330 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
332 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
336 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
339 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
342 * We could optimize this with just kicking the target running the task
343 * if that noise matters for nohz full users.
345 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
348 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
350 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
354 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
355 * per process timers.
357 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
359 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
362 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
364 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
368 * Re-evaluate the need for the tick as we switch the current task.
369 * It might need the tick due to per task/process properties:
370 * perf events, posix CPU timers, ...
372 void __tick_nohz_task_switch(void)
375 struct tick_sched *ts;
377 local_irq_save(flags);
379 if (!tick_nohz_full_cpu(smp_processor_id()))
382 ts = this_cpu_ptr(&tick_cpu_sched);
384 if (ts->tick_stopped) {
385 if (atomic_read(¤t->tick_dep_mask) ||
386 atomic_read(¤t->signal->tick_dep_mask))
387 tick_nohz_full_kick();
390 local_irq_restore(flags);
393 /* Get the boot-time nohz CPU list from the kernel parameters. */
394 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
396 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
397 cpumask_copy(tick_nohz_full_mask, cpumask);
398 tick_nohz_full_running = true;
401 static int tick_nohz_cpu_down(unsigned int cpu)
404 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
405 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
406 * CPUs. It must remain online when nohz full is enabled.
408 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
413 void __init tick_nohz_init(void)
417 if (!tick_nohz_full_running)
421 * Full dynticks uses irq work to drive the tick rescheduling on safe
422 * locking contexts. But then we need irq work to raise its own
423 * interrupts to avoid circular dependency on the tick
425 if (!arch_irq_work_has_interrupt()) {
426 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
427 cpumask_clear(tick_nohz_full_mask);
428 tick_nohz_full_running = false;
432 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
433 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
434 cpu = smp_processor_id();
436 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
437 pr_warn("NO_HZ: Clearing %d from nohz_full range "
438 "for timekeeping\n", cpu);
439 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
443 for_each_cpu(cpu, tick_nohz_full_mask)
444 context_tracking_cpu_set(cpu);
446 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
447 "kernel/nohz:predown", NULL,
450 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
451 cpumask_pr_args(tick_nohz_full_mask));
456 * NOHZ - aka dynamic tick functionality
458 #ifdef CONFIG_NO_HZ_COMMON
462 bool tick_nohz_enabled __read_mostly = true;
463 unsigned long tick_nohz_active __read_mostly;
465 * Enable / Disable tickless mode
467 static int __init setup_tick_nohz(char *str)
469 return (kstrtobool(str, &tick_nohz_enabled) == 0);
472 __setup("nohz=", setup_tick_nohz);
474 bool tick_nohz_tick_stopped(void)
476 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
478 return ts->tick_stopped;
481 bool tick_nohz_tick_stopped_cpu(int cpu)
483 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
485 return ts->tick_stopped;
489 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
491 * Called from interrupt entry when the CPU was idle
493 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
494 * must be updated. Otherwise an interrupt handler could use a stale jiffy
495 * value. We do this unconditionally on any CPU, as we don't know whether the
496 * CPU, which has the update task assigned is in a long sleep.
498 static void tick_nohz_update_jiffies(ktime_t now)
502 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
504 local_irq_save(flags);
505 tick_do_update_jiffies64(now);
506 local_irq_restore(flags);
508 touch_softlockup_watchdog_sched();
512 * Updates the per-CPU time idle statistics counters
515 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
519 if (ts->idle_active) {
520 delta = ktime_sub(now, ts->idle_entrytime);
521 if (nr_iowait_cpu(cpu) > 0)
522 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
524 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
525 ts->idle_entrytime = now;
528 if (last_update_time)
529 *last_update_time = ktime_to_us(now);
533 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
535 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
538 sched_clock_idle_wakeup_event();
541 static void tick_nohz_start_idle(struct tick_sched *ts)
543 ts->idle_entrytime = ktime_get();
545 sched_clock_idle_sleep_event();
549 * get_cpu_idle_time_us - get the total idle time of a CPU
550 * @cpu: CPU number to query
551 * @last_update_time: variable to store update time in. Do not update
554 * Return the cumulative idle time (since boot) for a given
555 * CPU, in microseconds.
557 * This time is measured via accounting rather than sampling,
558 * and is as accurate as ktime_get() is.
560 * This function returns -1 if NOHZ is not enabled.
562 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
564 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
567 if (!tick_nohz_active)
571 if (last_update_time) {
572 update_ts_time_stats(cpu, ts, now, last_update_time);
573 idle = ts->idle_sleeptime;
575 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
576 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
578 idle = ktime_add(ts->idle_sleeptime, delta);
580 idle = ts->idle_sleeptime;
584 return ktime_to_us(idle);
587 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
590 * get_cpu_iowait_time_us - get the total iowait time of a CPU
591 * @cpu: CPU number to query
592 * @last_update_time: variable to store update time in. Do not update
595 * Return the cumulative iowait time (since boot) for a given
596 * CPU, in microseconds.
598 * This time is measured via accounting rather than sampling,
599 * and is as accurate as ktime_get() is.
601 * This function returns -1 if NOHZ is not enabled.
603 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
605 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
608 if (!tick_nohz_active)
612 if (last_update_time) {
613 update_ts_time_stats(cpu, ts, now, last_update_time);
614 iowait = ts->iowait_sleeptime;
616 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
617 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
619 iowait = ktime_add(ts->iowait_sleeptime, delta);
621 iowait = ts->iowait_sleeptime;
625 return ktime_to_us(iowait);
627 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
629 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
631 hrtimer_cancel(&ts->sched_timer);
632 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
634 /* Forward the time to expire in the future */
635 hrtimer_forward(&ts->sched_timer, now, tick_period);
637 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
638 hrtimer_start_expires(&ts->sched_timer,
639 HRTIMER_MODE_ABS_PINNED_HARD);
641 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
645 * Reset to make sure next tick stop doesn't get fooled by past
646 * cached clock deadline.
651 static inline bool local_timer_softirq_pending(void)
653 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
656 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
658 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
659 unsigned long basejiff;
662 /* Read jiffies and the time when jiffies were updated last */
664 seq = read_seqbegin(&jiffies_lock);
665 basemono = last_jiffies_update;
667 } while (read_seqretry(&jiffies_lock, seq));
668 ts->last_jiffies = basejiff;
669 ts->timer_expires_base = basemono;
672 * Keep the periodic tick, when RCU, architecture or irq_work
674 * Aside of that check whether the local timer softirq is
675 * pending. If so its a bad idea to call get_next_timer_interrupt()
676 * because there is an already expired timer, so it will request
677 * immeditate expiry, which rearms the hardware timer with a
678 * minimal delta which brings us back to this place
679 * immediately. Lather, rinse and repeat...
681 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
682 irq_work_needs_cpu() || local_timer_softirq_pending()) {
683 next_tick = basemono + TICK_NSEC;
686 * Get the next pending timer. If high resolution
687 * timers are enabled this only takes the timer wheel
688 * timers into account. If high resolution timers are
689 * disabled this also looks at the next expiring
692 next_tmr = get_next_timer_interrupt(basejiff, basemono);
693 ts->next_timer = next_tmr;
694 /* Take the next rcu event into account */
695 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
699 * If the tick is due in the next period, keep it ticking or
700 * force prod the timer.
702 delta = next_tick - basemono;
703 if (delta <= (u64)TICK_NSEC) {
705 * Tell the timer code that the base is not idle, i.e. undo
706 * the effect of get_next_timer_interrupt():
710 * We've not stopped the tick yet, and there's a timer in the
711 * next period, so no point in stopping it either, bail.
713 if (!ts->tick_stopped) {
714 ts->timer_expires = 0;
720 * If this CPU is the one which had the do_timer() duty last, we limit
721 * the sleep time to the timekeeping max_deferment value.
722 * Otherwise we can sleep as long as we want.
724 delta = timekeeping_max_deferment();
725 if (cpu != tick_do_timer_cpu &&
726 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
729 /* Calculate the next expiry time */
730 if (delta < (KTIME_MAX - basemono))
731 expires = basemono + delta;
735 ts->timer_expires = min_t(u64, expires, next_tick);
738 return ts->timer_expires;
741 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
743 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
744 u64 basemono = ts->timer_expires_base;
745 u64 expires = ts->timer_expires;
746 ktime_t tick = expires;
748 /* Make sure we won't be trying to stop it twice in a row. */
749 ts->timer_expires_base = 0;
752 * If this CPU is the one which updates jiffies, then give up
753 * the assignment and let it be taken by the CPU which runs
754 * the tick timer next, which might be this CPU as well. If we
755 * don't drop this here the jiffies might be stale and
756 * do_timer() never invoked. Keep track of the fact that it
757 * was the one which had the do_timer() duty last.
759 if (cpu == tick_do_timer_cpu) {
760 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
761 ts->do_timer_last = 1;
762 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
763 ts->do_timer_last = 0;
766 /* Skip reprogram of event if its not changed */
767 if (ts->tick_stopped && (expires == ts->next_tick)) {
768 /* Sanity check: make sure clockevent is actually programmed */
769 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
773 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
774 basemono, ts->next_tick, dev->next_event,
775 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
779 * nohz_stop_sched_tick can be called several times before
780 * the nohz_restart_sched_tick is called. This happens when
781 * interrupts arrive which do not cause a reschedule. In the
782 * first call we save the current tick time, so we can restart
783 * the scheduler tick in nohz_restart_sched_tick.
785 if (!ts->tick_stopped) {
786 calc_load_nohz_start();
789 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
790 ts->tick_stopped = 1;
791 trace_tick_stop(1, TICK_DEP_MASK_NONE);
794 ts->next_tick = tick;
797 * If the expiration time == KTIME_MAX, then we simply stop
800 if (unlikely(expires == KTIME_MAX)) {
801 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
802 hrtimer_cancel(&ts->sched_timer);
806 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
807 hrtimer_start(&ts->sched_timer, tick,
808 HRTIMER_MODE_ABS_PINNED_HARD);
810 hrtimer_set_expires(&ts->sched_timer, tick);
811 tick_program_event(tick, 1);
815 static void tick_nohz_retain_tick(struct tick_sched *ts)
817 ts->timer_expires_base = 0;
820 #ifdef CONFIG_NO_HZ_FULL
821 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
823 if (tick_nohz_next_event(ts, cpu))
824 tick_nohz_stop_tick(ts, cpu);
826 tick_nohz_retain_tick(ts);
828 #endif /* CONFIG_NO_HZ_FULL */
830 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
832 /* Update jiffies first */
833 tick_do_update_jiffies64(now);
836 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
837 * the clock forward checks in the enqueue path:
841 calc_load_nohz_stop();
842 touch_softlockup_watchdog_sched();
844 * Cancel the scheduled timer and restore the tick
846 ts->tick_stopped = 0;
847 ts->idle_exittime = now;
849 tick_nohz_restart(ts, now);
852 static void tick_nohz_full_update_tick(struct tick_sched *ts)
854 #ifdef CONFIG_NO_HZ_FULL
855 int cpu = smp_processor_id();
857 if (!tick_nohz_full_cpu(cpu))
860 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
863 if (can_stop_full_tick(cpu, ts))
864 tick_nohz_stop_sched_tick(ts, cpu);
865 else if (ts->tick_stopped)
866 tick_nohz_restart_sched_tick(ts, ktime_get());
870 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
873 * If this CPU is offline and it is the one which updates
874 * jiffies, then give up the assignment and let it be taken by
875 * the CPU which runs the tick timer next. If we don't drop
876 * this here the jiffies might be stale and do_timer() never
879 if (unlikely(!cpu_online(cpu))) {
880 if (cpu == tick_do_timer_cpu)
881 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
883 * Make sure the CPU doesn't get fooled by obsolete tick
884 * deadline if it comes back online later.
890 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
896 if (unlikely(local_softirq_pending())) {
897 static int ratelimit;
899 if (ratelimit < 10 &&
900 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
901 pr_warn("NOHZ: local_softirq_pending %02x\n",
902 (unsigned int) local_softirq_pending());
908 if (tick_nohz_full_enabled()) {
910 * Keep the tick alive to guarantee timekeeping progression
911 * if there are full dynticks CPUs around
913 if (tick_do_timer_cpu == cpu)
916 * Boot safety: make sure the timekeeping duty has been
917 * assigned before entering dyntick-idle mode,
918 * tick_do_timer_cpu is TICK_DO_TIMER_BOOT
920 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_BOOT))
923 /* Should not happen for nohz-full */
924 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
931 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
934 int cpu = smp_processor_id();
937 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
938 * tick timer expiration time is known already.
940 if (ts->timer_expires_base)
941 expires = ts->timer_expires;
942 else if (can_stop_idle_tick(cpu, ts))
943 expires = tick_nohz_next_event(ts, cpu);
950 int was_stopped = ts->tick_stopped;
952 tick_nohz_stop_tick(ts, cpu);
955 ts->idle_expires = expires;
957 if (!was_stopped && ts->tick_stopped) {
958 ts->idle_jiffies = ts->last_jiffies;
959 nohz_balance_enter_idle(cpu);
962 tick_nohz_retain_tick(ts);
967 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
969 * When the next event is more than a tick into the future, stop the idle tick
971 void tick_nohz_idle_stop_tick(void)
973 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
976 void tick_nohz_idle_retain_tick(void)
978 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
980 * Undo the effect of get_next_timer_interrupt() called from
981 * tick_nohz_next_event().
987 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
989 * Called when we start the idle loop.
991 void tick_nohz_idle_enter(void)
993 struct tick_sched *ts;
995 lockdep_assert_irqs_enabled();
999 ts = this_cpu_ptr(&tick_cpu_sched);
1001 WARN_ON_ONCE(ts->timer_expires_base);
1004 tick_nohz_start_idle(ts);
1010 * tick_nohz_irq_exit - update next tick event from interrupt exit
1012 * When an interrupt fires while we are idle and it doesn't cause
1013 * a reschedule, it may still add, modify or delete a timer, enqueue
1014 * an RCU callback, etc...
1015 * So we need to re-calculate and reprogram the next tick event.
1017 void tick_nohz_irq_exit(void)
1019 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1022 tick_nohz_start_idle(ts);
1024 tick_nohz_full_update_tick(ts);
1028 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1030 bool tick_nohz_idle_got_tick(void)
1032 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1034 if (ts->got_idle_tick) {
1035 ts->got_idle_tick = 0;
1042 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1043 * or the tick, whatever that expires first. Note that, if the tick has been
1044 * stopped, it returns the next hrtimer.
1046 * Called from power state control code with interrupts disabled
1048 ktime_t tick_nohz_get_next_hrtimer(void)
1050 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1054 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1055 * @delta_next: duration until the next event if the tick cannot be stopped
1057 * Called from power state control code with interrupts disabled
1059 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1061 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1062 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1063 int cpu = smp_processor_id();
1065 * The idle entry time is expected to be a sufficient approximation of
1066 * the current time at this point.
1068 ktime_t now = ts->idle_entrytime;
1071 WARN_ON_ONCE(!ts->inidle);
1073 *delta_next = ktime_sub(dev->next_event, now);
1075 if (!can_stop_idle_tick(cpu, ts))
1078 next_event = tick_nohz_next_event(ts, cpu);
1083 * If the next highres timer to expire is earlier than next_event, the
1084 * idle governor needs to know that.
1086 next_event = min_t(u64, next_event,
1087 hrtimer_next_event_without(&ts->sched_timer));
1089 return ktime_sub(next_event, now);
1093 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1094 * for a particular CPU.
1096 * Called from the schedutil frequency scaling governor in scheduler context.
1098 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1100 struct tick_sched *ts = tick_get_tick_sched(cpu);
1102 return ts->idle_calls;
1106 * tick_nohz_get_idle_calls - return the current idle calls counter value
1108 * Called from the schedutil frequency scaling governor in scheduler context.
1110 unsigned long tick_nohz_get_idle_calls(void)
1112 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1114 return ts->idle_calls;
1117 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1119 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1120 unsigned long ticks;
1122 if (vtime_accounting_cpu_enabled())
1125 * We stopped the tick in idle. Update process times would miss the
1126 * time we slept as update_process_times does only a 1 tick
1127 * accounting. Enforce that this is accounted to idle !
1129 ticks = jiffies - ts->idle_jiffies;
1131 * We might be one off. Do not randomly account a huge number of ticks!
1133 if (ticks && ticks < LONG_MAX)
1134 account_idle_ticks(ticks);
1138 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1140 tick_nohz_restart_sched_tick(ts, now);
1141 tick_nohz_account_idle_ticks(ts);
1144 void tick_nohz_idle_restart_tick(void)
1146 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1148 if (ts->tick_stopped)
1149 __tick_nohz_idle_restart_tick(ts, ktime_get());
1153 * tick_nohz_idle_exit - restart the idle tick from the idle task
1155 * Restart the idle tick when the CPU is woken up from idle
1156 * This also exit the RCU extended quiescent state. The CPU
1157 * can use RCU again after this function is called.
1159 void tick_nohz_idle_exit(void)
1161 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1162 bool idle_active, tick_stopped;
1165 local_irq_disable();
1167 WARN_ON_ONCE(!ts->inidle);
1168 WARN_ON_ONCE(ts->timer_expires_base);
1171 idle_active = ts->idle_active;
1172 tick_stopped = ts->tick_stopped;
1174 if (idle_active || tick_stopped)
1178 tick_nohz_stop_idle(ts, now);
1181 __tick_nohz_idle_restart_tick(ts, now);
1187 * The nohz low res interrupt handler
1189 static void tick_nohz_handler(struct clock_event_device *dev)
1191 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1192 struct pt_regs *regs = get_irq_regs();
1193 ktime_t now = ktime_get();
1195 dev->next_event = KTIME_MAX;
1197 tick_sched_do_timer(ts, now);
1198 tick_sched_handle(ts, regs);
1200 /* No need to reprogram if we are running tickless */
1201 if (unlikely(ts->tick_stopped))
1204 hrtimer_forward(&ts->sched_timer, now, tick_period);
1205 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1208 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1210 if (!tick_nohz_enabled)
1212 ts->nohz_mode = mode;
1213 /* One update is enough */
1214 if (!test_and_set_bit(0, &tick_nohz_active))
1215 timers_update_nohz();
1219 * tick_nohz_switch_to_nohz - switch to nohz mode
1221 static void tick_nohz_switch_to_nohz(void)
1223 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1226 if (!tick_nohz_enabled)
1229 if (tick_switch_to_oneshot(tick_nohz_handler))
1233 * Recycle the hrtimer in ts, so we can share the
1234 * hrtimer_forward with the highres code.
1236 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1237 /* Get the next period */
1238 next = tick_init_jiffy_update();
1240 hrtimer_set_expires(&ts->sched_timer, next);
1241 hrtimer_forward_now(&ts->sched_timer, tick_period);
1242 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1243 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1246 static inline void tick_nohz_irq_enter(void)
1248 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1251 if (!ts->idle_active && !ts->tick_stopped)
1254 if (ts->idle_active)
1255 tick_nohz_stop_idle(ts, now);
1256 if (ts->tick_stopped)
1257 tick_nohz_update_jiffies(now);
1262 static inline void tick_nohz_switch_to_nohz(void) { }
1263 static inline void tick_nohz_irq_enter(void) { }
1264 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1266 #endif /* CONFIG_NO_HZ_COMMON */
1269 * Called from irq_enter to notify about the possible interruption of idle()
1271 void tick_irq_enter(void)
1273 tick_check_oneshot_broadcast_this_cpu();
1274 tick_nohz_irq_enter();
1278 * High resolution timer specific code
1280 #ifdef CONFIG_HIGH_RES_TIMERS
1282 * We rearm the timer until we get disabled by the idle code.
1283 * Called with interrupts disabled.
1285 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1287 struct tick_sched *ts =
1288 container_of(timer, struct tick_sched, sched_timer);
1289 struct pt_regs *regs = get_irq_regs();
1290 ktime_t now = ktime_get();
1292 tick_sched_do_timer(ts, now);
1295 * Do not call, when we are not in irq context and have
1296 * no valid regs pointer
1299 tick_sched_handle(ts, regs);
1303 /* No need to reprogram if we are in idle or full dynticks mode */
1304 if (unlikely(ts->tick_stopped))
1305 return HRTIMER_NORESTART;
1307 hrtimer_forward(timer, now, tick_period);
1309 return HRTIMER_RESTART;
1312 static int sched_skew_tick;
1314 static int __init skew_tick(char *str)
1316 get_option(&str, &sched_skew_tick);
1320 early_param("skew_tick", skew_tick);
1323 * tick_setup_sched_timer - setup the tick emulation timer
1325 void tick_setup_sched_timer(void)
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1328 ktime_t now = ktime_get();
1331 * Emulate tick processing via per-CPU hrtimers:
1333 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1334 ts->sched_timer.function = tick_sched_timer;
1336 /* Get the next period (per-CPU) */
1337 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1339 /* Offset the tick to avert jiffies_lock contention. */
1340 if (sched_skew_tick) {
1341 u64 offset = ktime_to_ns(tick_period) >> 1;
1342 do_div(offset, num_possible_cpus());
1343 offset *= smp_processor_id();
1344 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1347 hrtimer_forward(&ts->sched_timer, now, tick_period);
1348 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1349 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1351 #endif /* HIGH_RES_TIMERS */
1353 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1354 void tick_cancel_sched_timer(int cpu)
1356 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1358 # ifdef CONFIG_HIGH_RES_TIMERS
1359 if (ts->sched_timer.base)
1360 hrtimer_cancel(&ts->sched_timer);
1363 memset(ts, 0, sizeof(*ts));
1368 * Async notification about clocksource changes
1370 void tick_clock_notify(void)
1374 for_each_possible_cpu(cpu)
1375 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1379 * Async notification about clock event changes
1381 void tick_oneshot_notify(void)
1383 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1385 set_bit(0, &ts->check_clocks);
1389 * Check, if a change happened, which makes oneshot possible.
1391 * Called cyclic from the hrtimer softirq (driven by the timer
1392 * softirq) allow_nohz signals, that we can switch into low-res nohz
1393 * mode, because high resolution timers are disabled (either compile
1394 * or runtime). Called with interrupts disabled.
1396 int tick_check_oneshot_change(int allow_nohz)
1398 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1400 if (!test_and_clear_bit(0, &ts->check_clocks))
1403 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1406 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1412 tick_nohz_switch_to_nohz();