2 * linux/kernel/time/tick-sched.c
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
8 * No idle tick implementation for low and high resolution timers
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * Distribute under GPLv2.
14 #include <linux/cpu.h>
15 #include <linux/err.h>
16 #include <linux/hrtimer.h>
17 #include <linux/interrupt.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/percpu.h>
20 #include <linux/nmi.h>
21 #include <linux/profile.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/clock.h>
24 #include <linux/sched/stat.h>
25 #include <linux/sched/nohz.h>
26 #include <linux/module.h>
27 #include <linux/irq_work.h>
28 #include <linux/posix-timers.h>
29 #include <linux/context_tracking.h>
32 #include <asm/irq_regs.h>
34 #include "tick-internal.h"
36 #include <trace/events/timer.h>
39 * Per-CPU nohz control structure
41 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
43 struct tick_sched *tick_get_tick_sched(int cpu)
45 return &per_cpu(tick_cpu_sched, cpu);
48 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
50 * The time, when the last jiffy update happened. Protected by jiffies_lock.
52 static ktime_t last_jiffies_update;
55 * Must be called with interrupts disabled !
57 static void tick_do_update_jiffies64(ktime_t now)
59 unsigned long ticks = 0;
63 * Do a quick check without holding jiffies_lock:
64 * The READ_ONCE() pairs with two updates done later in this function.
66 delta = ktime_sub(now, READ_ONCE(last_jiffies_update));
67 if (delta < tick_period)
70 /* Reevaluate with jiffies_lock held */
71 write_seqlock(&jiffies_lock);
73 delta = ktime_sub(now, last_jiffies_update);
74 if (delta >= tick_period) {
76 delta = ktime_sub(delta, tick_period);
77 /* Pairs with the lockless read in this function. */
78 WRITE_ONCE(last_jiffies_update,
79 ktime_add(last_jiffies_update, tick_period));
81 /* Slow path for long timeouts */
82 if (unlikely(delta >= tick_period)) {
83 s64 incr = ktime_to_ns(tick_period);
85 ticks = ktime_divns(delta, incr);
87 /* Pairs with the lockless read in this function. */
88 WRITE_ONCE(last_jiffies_update,
89 ktime_add_ns(last_jiffies_update,
94 /* Keep the tick_next_period variable up to date */
95 tick_next_period = ktime_add(last_jiffies_update, tick_period);
97 write_sequnlock(&jiffies_lock);
100 write_sequnlock(&jiffies_lock);
105 * Initialize and return retrieve the jiffies update.
107 static ktime_t tick_init_jiffy_update(void)
111 write_seqlock(&jiffies_lock);
112 /* Did we start the jiffies update yet ? */
113 if (last_jiffies_update == 0)
114 last_jiffies_update = tick_next_period;
115 period = last_jiffies_update;
116 write_sequnlock(&jiffies_lock);
120 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
122 int cpu = smp_processor_id();
124 #ifdef CONFIG_NO_HZ_COMMON
126 * Check if the do_timer duty was dropped. We don't care about
127 * concurrency: This happens only when the CPU in charge went
128 * into a long sleep. If two CPUs happen to assign themselves to
129 * this duty, then the jiffies update is still serialized by
132 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
133 && !tick_nohz_full_cpu(cpu))
134 tick_do_timer_cpu = cpu;
137 /* Check, if the jiffies need an update */
138 if (tick_do_timer_cpu == cpu)
139 tick_do_update_jiffies64(now);
142 ts->got_idle_tick = 1;
145 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
147 #ifdef CONFIG_NO_HZ_COMMON
149 * When we are idle and the tick is stopped, we have to touch
150 * the watchdog as we might not schedule for a really long
151 * time. This happens on complete idle SMP systems while
152 * waiting on the login prompt. We also increment the "start of
153 * idle" jiffy stamp so the idle accounting adjustment we do
154 * when we go busy again does not account too much ticks.
156 if (ts->tick_stopped) {
157 touch_softlockup_watchdog_sched();
158 if (is_idle_task(current))
161 * In case the current tick fired too early past its expected
162 * expiration, make sure we don't bypass the next clock reprogramming
163 * to the same deadline.
168 update_process_times(user_mode(regs));
169 profile_tick(CPU_PROFILING);
173 #ifdef CONFIG_NO_HZ_FULL
174 cpumask_var_t tick_nohz_full_mask;
175 bool tick_nohz_full_running;
176 static atomic_t tick_dep_mask;
178 static bool check_tick_dependency(atomic_t *dep)
180 int val = atomic_read(dep);
182 if (val & TICK_DEP_MASK_POSIX_TIMER) {
183 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
187 if (val & TICK_DEP_MASK_PERF_EVENTS) {
188 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
192 if (val & TICK_DEP_MASK_SCHED) {
193 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
197 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
198 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
202 if (val & TICK_DEP_MASK_RCU) {
203 trace_tick_stop(0, TICK_DEP_MASK_RCU);
210 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
212 lockdep_assert_irqs_disabled();
214 if (unlikely(!cpu_online(cpu)))
217 if (check_tick_dependency(&tick_dep_mask))
220 if (check_tick_dependency(&ts->tick_dep_mask))
223 if (check_tick_dependency(¤t->tick_dep_mask))
226 if (check_tick_dependency(¤t->signal->tick_dep_mask))
232 static void nohz_full_kick_func(struct irq_work *work)
234 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
237 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
238 .func = nohz_full_kick_func,
242 * Kick this CPU if it's full dynticks in order to force it to
243 * re-evaluate its dependency on the tick and restart it if necessary.
244 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
247 static void tick_nohz_full_kick(void)
249 if (!tick_nohz_full_cpu(smp_processor_id()))
252 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
256 * Kick the CPU if it's full dynticks in order to force it to
257 * re-evaluate its dependency on the tick and restart it if necessary.
259 void tick_nohz_full_kick_cpu(int cpu)
261 if (!tick_nohz_full_cpu(cpu))
264 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
268 * Kick all full dynticks CPUs in order to force these to re-evaluate
269 * their dependency on the tick and restart it if necessary.
271 static void tick_nohz_full_kick_all(void)
275 if (!tick_nohz_full_running)
279 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
280 tick_nohz_full_kick_cpu(cpu);
284 static void tick_nohz_dep_set_all(atomic_t *dep,
285 enum tick_dep_bits bit)
289 prev = atomic_fetch_or(BIT(bit), dep);
291 tick_nohz_full_kick_all();
295 * Set a global tick dependency. Used by perf events that rely on freq and
298 void tick_nohz_dep_set(enum tick_dep_bits bit)
300 tick_nohz_dep_set_all(&tick_dep_mask, bit);
303 void tick_nohz_dep_clear(enum tick_dep_bits bit)
305 atomic_andnot(BIT(bit), &tick_dep_mask);
309 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
310 * manage events throttling.
312 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
315 struct tick_sched *ts;
317 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
319 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
322 /* Perf needs local kick that is NMI safe */
323 if (cpu == smp_processor_id()) {
324 tick_nohz_full_kick();
326 /* Remote irq work not NMI-safe */
327 if (!WARN_ON_ONCE(in_nmi()))
328 tick_nohz_full_kick_cpu(cpu);
333 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
335 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
337 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
339 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
341 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
344 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
347 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
350 * We could optimize this with just kicking the target running the task
351 * if that noise matters for nohz full users.
353 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
356 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
358 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
362 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
363 * per process timers.
365 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
367 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
370 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
372 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
376 * Re-evaluate the need for the tick as we switch the current task.
377 * It might need the tick due to per task/process properties:
378 * perf events, posix CPU timers, ...
380 void __tick_nohz_task_switch(void)
383 struct tick_sched *ts;
385 local_irq_save(flags);
387 if (!tick_nohz_full_cpu(smp_processor_id()))
390 ts = this_cpu_ptr(&tick_cpu_sched);
392 if (ts->tick_stopped) {
393 if (atomic_read(¤t->tick_dep_mask) ||
394 atomic_read(¤t->signal->tick_dep_mask))
395 tick_nohz_full_kick();
398 local_irq_restore(flags);
401 /* Get the boot-time nohz CPU list from the kernel parameters. */
402 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
404 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
405 cpumask_copy(tick_nohz_full_mask, cpumask);
406 tick_nohz_full_running = true;
409 bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
412 * The boot CPU handles housekeeping duty (unbound timers,
413 * workqueues, timekeeping, ...) on behalf of full dynticks
414 * CPUs. It must remain online when nohz full is enabled.
416 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
421 static int tick_nohz_cpu_down(unsigned int cpu)
423 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
426 void __init tick_nohz_init(void)
430 if (!tick_nohz_full_running)
434 * Full dynticks uses irq work to drive the tick rescheduling on safe
435 * locking contexts. But then we need irq work to raise its own
436 * interrupts to avoid circular dependency on the tick
438 if (!arch_irq_work_has_interrupt()) {
439 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
440 cpumask_clear(tick_nohz_full_mask);
441 tick_nohz_full_running = false;
445 cpu = smp_processor_id();
447 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
448 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
450 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
453 for_each_cpu(cpu, tick_nohz_full_mask)
454 context_tracking_cpu_set(cpu);
456 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
457 "kernel/nohz:predown", NULL,
460 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
461 cpumask_pr_args(tick_nohz_full_mask));
466 * NOHZ - aka dynamic tick functionality
468 #ifdef CONFIG_NO_HZ_COMMON
472 bool tick_nohz_enabled __read_mostly = true;
473 unsigned long tick_nohz_active __read_mostly;
475 * Enable / Disable tickless mode
477 static int __init setup_tick_nohz(char *str)
479 return (kstrtobool(str, &tick_nohz_enabled) == 0);
482 __setup("nohz=", setup_tick_nohz);
484 bool tick_nohz_tick_stopped(void)
486 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
488 return ts->tick_stopped;
491 bool tick_nohz_tick_stopped_cpu(int cpu)
493 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
495 return ts->tick_stopped;
499 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
501 * Called from interrupt entry when the CPU was idle
503 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
504 * must be updated. Otherwise an interrupt handler could use a stale jiffy
505 * value. We do this unconditionally on any CPU, as we don't know whether the
506 * CPU, which has the update task assigned is in a long sleep.
508 static void tick_nohz_update_jiffies(ktime_t now)
512 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
514 local_irq_save(flags);
515 tick_do_update_jiffies64(now);
516 local_irq_restore(flags);
518 touch_softlockup_watchdog_sched();
522 * Updates the per-CPU time idle statistics counters
525 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
529 if (ts->idle_active) {
530 delta = ktime_sub(now, ts->idle_entrytime);
531 if (nr_iowait_cpu(cpu) > 0)
532 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
534 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
535 ts->idle_entrytime = now;
538 if (last_update_time)
539 *last_update_time = ktime_to_us(now);
543 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
545 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
548 sched_clock_idle_wakeup_event();
551 static void tick_nohz_start_idle(struct tick_sched *ts)
553 ts->idle_entrytime = ktime_get();
555 sched_clock_idle_sleep_event();
559 * get_cpu_idle_time_us - get the total idle time of a CPU
560 * @cpu: CPU number to query
561 * @last_update_time: variable to store update time in. Do not update
564 * Return the cumulative idle time (since boot) for a given
565 * CPU, in microseconds.
567 * This time is measured via accounting rather than sampling,
568 * and is as accurate as ktime_get() is.
570 * This function returns -1 if NOHZ is not enabled.
572 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
574 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
577 if (!tick_nohz_active)
581 if (last_update_time) {
582 update_ts_time_stats(cpu, ts, now, last_update_time);
583 idle = ts->idle_sleeptime;
585 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
586 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
588 idle = ktime_add(ts->idle_sleeptime, delta);
590 idle = ts->idle_sleeptime;
594 return ktime_to_us(idle);
597 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
600 * get_cpu_iowait_time_us - get the total iowait time of a CPU
601 * @cpu: CPU number to query
602 * @last_update_time: variable to store update time in. Do not update
605 * Return the cumulative iowait time (since boot) for a given
606 * CPU, in microseconds.
608 * This time is measured via accounting rather than sampling,
609 * and is as accurate as ktime_get() is.
611 * This function returns -1 if NOHZ is not enabled.
613 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
615 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
618 if (!tick_nohz_active)
622 if (last_update_time) {
623 update_ts_time_stats(cpu, ts, now, last_update_time);
624 iowait = ts->iowait_sleeptime;
626 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
627 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
629 iowait = ktime_add(ts->iowait_sleeptime, delta);
631 iowait = ts->iowait_sleeptime;
635 return ktime_to_us(iowait);
637 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
639 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
641 hrtimer_cancel(&ts->sched_timer);
642 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
644 /* Forward the time to expire in the future */
645 hrtimer_forward(&ts->sched_timer, now, tick_period);
647 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
648 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
650 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
653 * Reset to make sure next tick stop doesn't get fooled by past
654 * cached clock deadline.
659 static inline bool local_timer_softirq_pending(void)
661 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
664 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
666 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
667 unsigned long seq, basejiff;
669 /* Read jiffies and the time when jiffies were updated last */
671 seq = read_seqbegin(&jiffies_lock);
672 basemono = last_jiffies_update;
674 } while (read_seqretry(&jiffies_lock, seq));
675 ts->last_jiffies = basejiff;
676 ts->timer_expires_base = basemono;
679 * Keep the periodic tick, when RCU, architecture or irq_work
681 * Aside of that check whether the local timer softirq is
682 * pending. If so its a bad idea to call get_next_timer_interrupt()
683 * because there is an already expired timer, so it will request
684 * immeditate expiry, which rearms the hardware timer with a
685 * minimal delta which brings us back to this place
686 * immediately. Lather, rinse and repeat...
688 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
689 irq_work_needs_cpu() || local_timer_softirq_pending()) {
690 next_tick = basemono + TICK_NSEC;
693 * Get the next pending timer. If high resolution
694 * timers are enabled this only takes the timer wheel
695 * timers into account. If high resolution timers are
696 * disabled this also looks at the next expiring
699 next_tmr = get_next_timer_interrupt(basejiff, basemono);
700 ts->next_timer = next_tmr;
701 /* Take the next rcu event into account */
702 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
706 * If the tick is due in the next period, keep it ticking or
707 * force prod the timer.
709 delta = next_tick - basemono;
710 if (delta <= (u64)TICK_NSEC) {
712 * Tell the timer code that the base is not idle, i.e. undo
713 * the effect of get_next_timer_interrupt():
717 * We've not stopped the tick yet, and there's a timer in the
718 * next period, so no point in stopping it either, bail.
720 if (!ts->tick_stopped) {
721 ts->timer_expires = 0;
727 * If this CPU is the one which had the do_timer() duty last, we limit
728 * the sleep time to the timekeeping max_deferment value.
729 * Otherwise we can sleep as long as we want.
731 delta = timekeeping_max_deferment();
732 if (cpu != tick_do_timer_cpu &&
733 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
736 /* Calculate the next expiry time */
737 if (delta < (KTIME_MAX - basemono))
738 expires = basemono + delta;
742 ts->timer_expires = min_t(u64, expires, next_tick);
745 return ts->timer_expires;
748 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
750 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
751 u64 basemono = ts->timer_expires_base;
752 u64 expires = ts->timer_expires;
753 ktime_t tick = expires;
755 /* Make sure we won't be trying to stop it twice in a row. */
756 ts->timer_expires_base = 0;
759 * If this CPU is the one which updates jiffies, then give up
760 * the assignment and let it be taken by the CPU which runs
761 * the tick timer next, which might be this CPU as well. If we
762 * don't drop this here the jiffies might be stale and
763 * do_timer() never invoked. Keep track of the fact that it
764 * was the one which had the do_timer() duty last.
766 if (cpu == tick_do_timer_cpu) {
767 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
768 ts->do_timer_last = 1;
769 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
770 ts->do_timer_last = 0;
773 /* Skip reprogram of event if its not changed */
774 if (ts->tick_stopped && (expires == ts->next_tick)) {
775 /* Sanity check: make sure clockevent is actually programmed */
776 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
780 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
781 basemono, ts->next_tick, dev->next_event,
782 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
786 * nohz_stop_sched_tick can be called several times before
787 * the nohz_restart_sched_tick is called. This happens when
788 * interrupts arrive which do not cause a reschedule. In the
789 * first call we save the current tick time, so we can restart
790 * the scheduler tick in nohz_restart_sched_tick.
792 if (!ts->tick_stopped) {
793 calc_load_nohz_start();
794 cpu_load_update_nohz_start();
797 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
798 ts->tick_stopped = 1;
799 trace_tick_stop(1, TICK_DEP_MASK_NONE);
802 ts->next_tick = tick;
805 * If the expiration time == KTIME_MAX, then we simply stop
808 if (unlikely(expires == KTIME_MAX)) {
809 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
810 hrtimer_cancel(&ts->sched_timer);
814 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
815 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
817 hrtimer_set_expires(&ts->sched_timer, tick);
818 tick_program_event(tick, 1);
822 static void tick_nohz_retain_tick(struct tick_sched *ts)
824 ts->timer_expires_base = 0;
827 #ifdef CONFIG_NO_HZ_FULL
828 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
830 if (tick_nohz_next_event(ts, cpu))
831 tick_nohz_stop_tick(ts, cpu);
833 tick_nohz_retain_tick(ts);
835 #endif /* CONFIG_NO_HZ_FULL */
837 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
839 /* Update jiffies first */
840 tick_do_update_jiffies64(now);
841 cpu_load_update_nohz_stop();
844 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
845 * the clock forward checks in the enqueue path:
849 calc_load_nohz_stop();
850 touch_softlockup_watchdog_sched();
852 * Cancel the scheduled timer and restore the tick
854 ts->tick_stopped = 0;
855 ts->idle_exittime = now;
857 tick_nohz_restart(ts, now);
860 static void tick_nohz_full_update_tick(struct tick_sched *ts)
862 #ifdef CONFIG_NO_HZ_FULL
863 int cpu = smp_processor_id();
865 if (!tick_nohz_full_cpu(cpu))
868 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
871 if (can_stop_full_tick(cpu, ts))
872 tick_nohz_stop_sched_tick(ts, cpu);
873 else if (ts->tick_stopped)
874 tick_nohz_restart_sched_tick(ts, ktime_get());
878 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
881 * If this CPU is offline and it is the one which updates
882 * jiffies, then give up the assignment and let it be taken by
883 * the CPU which runs the tick timer next. If we don't drop
884 * this here the jiffies might be stale and do_timer() never
887 if (unlikely(!cpu_online(cpu))) {
888 if (cpu == tick_do_timer_cpu)
889 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
891 * Make sure the CPU doesn't get fooled by obsolete tick
892 * deadline if it comes back online later.
898 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
904 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
905 static int ratelimit;
907 if (ratelimit < 10 &&
908 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
909 pr_warn("NOHZ: local_softirq_pending %02x\n",
910 (unsigned int) local_softirq_pending());
916 if (tick_nohz_full_enabled()) {
918 * Keep the tick alive to guarantee timekeeping progression
919 * if there are full dynticks CPUs around
921 if (tick_do_timer_cpu == cpu)
924 * Boot safety: make sure the timekeeping duty has been
925 * assigned before entering dyntick-idle mode,
927 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
934 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
937 int cpu = smp_processor_id();
940 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
941 * tick timer expiration time is known already.
943 if (ts->timer_expires_base)
944 expires = ts->timer_expires;
945 else if (can_stop_idle_tick(cpu, ts))
946 expires = tick_nohz_next_event(ts, cpu);
953 int was_stopped = ts->tick_stopped;
955 tick_nohz_stop_tick(ts, cpu);
958 ts->idle_expires = expires;
960 if (!was_stopped && ts->tick_stopped) {
961 ts->idle_jiffies = ts->last_jiffies;
962 nohz_balance_enter_idle(cpu);
965 tick_nohz_retain_tick(ts);
970 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
972 * When the next event is more than a tick into the future, stop the idle tick
974 void tick_nohz_idle_stop_tick(void)
976 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
979 void tick_nohz_idle_retain_tick(void)
981 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
983 * Undo the effect of get_next_timer_interrupt() called from
984 * tick_nohz_next_event().
990 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
992 * Called when we start the idle loop.
994 void tick_nohz_idle_enter(void)
996 struct tick_sched *ts;
998 lockdep_assert_irqs_enabled();
1000 local_irq_disable();
1002 ts = this_cpu_ptr(&tick_cpu_sched);
1004 WARN_ON_ONCE(ts->timer_expires_base);
1007 tick_nohz_start_idle(ts);
1013 * tick_nohz_irq_exit - update next tick event from interrupt exit
1015 * When an interrupt fires while we are idle and it doesn't cause
1016 * a reschedule, it may still add, modify or delete a timer, enqueue
1017 * an RCU callback, etc...
1018 * So we need to re-calculate and reprogram the next tick event.
1020 void tick_nohz_irq_exit(void)
1022 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1025 tick_nohz_start_idle(ts);
1027 tick_nohz_full_update_tick(ts);
1031 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1033 bool tick_nohz_idle_got_tick(void)
1035 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1037 if (ts->got_idle_tick) {
1038 ts->got_idle_tick = 0;
1045 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1046 * @delta_next: duration until the next event if the tick cannot be stopped
1048 * Called from power state control code with interrupts disabled
1050 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1052 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1053 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1054 int cpu = smp_processor_id();
1056 * The idle entry time is expected to be a sufficient approximation of
1057 * the current time at this point.
1059 ktime_t now = ts->idle_entrytime;
1062 WARN_ON_ONCE(!ts->inidle);
1064 *delta_next = ktime_sub(dev->next_event, now);
1066 if (!can_stop_idle_tick(cpu, ts))
1069 next_event = tick_nohz_next_event(ts, cpu);
1074 * If the next highres timer to expire is earlier than next_event, the
1075 * idle governor needs to know that.
1077 next_event = min_t(u64, next_event,
1078 hrtimer_next_event_without(&ts->sched_timer));
1080 return ktime_sub(next_event, now);
1084 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1085 * for a particular CPU.
1087 * Called from the schedutil frequency scaling governor in scheduler context.
1089 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1091 struct tick_sched *ts = tick_get_tick_sched(cpu);
1093 return ts->idle_calls;
1097 * tick_nohz_get_idle_calls - return the current idle calls counter value
1099 * Called from the schedutil frequency scaling governor in scheduler context.
1101 unsigned long tick_nohz_get_idle_calls(void)
1103 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1105 return ts->idle_calls;
1108 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1110 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1111 unsigned long ticks;
1113 if (vtime_accounting_cpu_enabled())
1116 * We stopped the tick in idle. Update process times would miss the
1117 * time we slept as update_process_times does only a 1 tick
1118 * accounting. Enforce that this is accounted to idle !
1120 ticks = jiffies - ts->idle_jiffies;
1122 * We might be one off. Do not randomly account a huge number of ticks!
1124 if (ticks && ticks < LONG_MAX)
1125 account_idle_ticks(ticks);
1129 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1131 tick_nohz_restart_sched_tick(ts, now);
1132 tick_nohz_account_idle_ticks(ts);
1135 void tick_nohz_idle_restart_tick(void)
1137 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1139 if (ts->tick_stopped)
1140 __tick_nohz_idle_restart_tick(ts, ktime_get());
1144 * tick_nohz_idle_exit - restart the idle tick from the idle task
1146 * Restart the idle tick when the CPU is woken up from idle
1147 * This also exit the RCU extended quiescent state. The CPU
1148 * can use RCU again after this function is called.
1150 void tick_nohz_idle_exit(void)
1152 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1153 bool idle_active, tick_stopped;
1156 local_irq_disable();
1158 WARN_ON_ONCE(!ts->inidle);
1159 WARN_ON_ONCE(ts->timer_expires_base);
1162 idle_active = ts->idle_active;
1163 tick_stopped = ts->tick_stopped;
1165 if (idle_active || tick_stopped)
1169 tick_nohz_stop_idle(ts, now);
1172 __tick_nohz_idle_restart_tick(ts, now);
1178 * The nohz low res interrupt handler
1180 static void tick_nohz_handler(struct clock_event_device *dev)
1182 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1183 struct pt_regs *regs = get_irq_regs();
1184 ktime_t now = ktime_get();
1186 dev->next_event = KTIME_MAX;
1188 tick_sched_do_timer(ts, now);
1189 tick_sched_handle(ts, regs);
1191 /* No need to reprogram if we are running tickless */
1192 if (unlikely(ts->tick_stopped))
1195 hrtimer_forward(&ts->sched_timer, now, tick_period);
1196 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1199 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1201 if (!tick_nohz_enabled)
1203 ts->nohz_mode = mode;
1204 /* One update is enough */
1205 if (!test_and_set_bit(0, &tick_nohz_active))
1206 timers_update_nohz();
1210 * tick_nohz_switch_to_nohz - switch to nohz mode
1212 static void tick_nohz_switch_to_nohz(void)
1214 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1217 if (!tick_nohz_enabled)
1220 if (tick_switch_to_oneshot(tick_nohz_handler))
1224 * Recycle the hrtimer in ts, so we can share the
1225 * hrtimer_forward with the highres code.
1227 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1228 /* Get the next period */
1229 next = tick_init_jiffy_update();
1231 hrtimer_set_expires(&ts->sched_timer, next);
1232 hrtimer_forward_now(&ts->sched_timer, tick_period);
1233 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1234 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1237 static inline void tick_nohz_irq_enter(void)
1239 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1242 if (!ts->idle_active && !ts->tick_stopped)
1245 if (ts->idle_active)
1246 tick_nohz_stop_idle(ts, now);
1247 if (ts->tick_stopped)
1248 tick_nohz_update_jiffies(now);
1253 static inline void tick_nohz_switch_to_nohz(void) { }
1254 static inline void tick_nohz_irq_enter(void) { }
1255 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1257 #endif /* CONFIG_NO_HZ_COMMON */
1260 * Called from irq_enter to notify about the possible interruption of idle()
1262 void tick_irq_enter(void)
1264 tick_check_oneshot_broadcast_this_cpu();
1265 tick_nohz_irq_enter();
1269 * High resolution timer specific code
1271 #ifdef CONFIG_HIGH_RES_TIMERS
1273 * We rearm the timer until we get disabled by the idle code.
1274 * Called with interrupts disabled.
1276 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1278 struct tick_sched *ts =
1279 container_of(timer, struct tick_sched, sched_timer);
1280 struct pt_regs *regs = get_irq_regs();
1281 ktime_t now = ktime_get();
1283 tick_sched_do_timer(ts, now);
1286 * Do not call, when we are not in irq context and have
1287 * no valid regs pointer
1290 tick_sched_handle(ts, regs);
1294 /* No need to reprogram if we are in idle or full dynticks mode */
1295 if (unlikely(ts->tick_stopped))
1296 return HRTIMER_NORESTART;
1298 hrtimer_forward(timer, now, tick_period);
1300 return HRTIMER_RESTART;
1303 static int sched_skew_tick;
1305 static int __init skew_tick(char *str)
1307 get_option(&str, &sched_skew_tick);
1311 early_param("skew_tick", skew_tick);
1314 * tick_setup_sched_timer - setup the tick emulation timer
1316 void tick_setup_sched_timer(void)
1318 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1319 ktime_t now = ktime_get();
1322 * Emulate tick processing via per-CPU hrtimers:
1324 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1325 ts->sched_timer.function = tick_sched_timer;
1327 /* Get the next period (per-CPU) */
1328 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1330 /* Offset the tick to avert jiffies_lock contention. */
1331 if (sched_skew_tick) {
1332 u64 offset = ktime_to_ns(tick_period) >> 1;
1333 do_div(offset, num_possible_cpus());
1334 offset *= smp_processor_id();
1335 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1338 hrtimer_forward(&ts->sched_timer, now, tick_period);
1339 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1340 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1342 #endif /* HIGH_RES_TIMERS */
1344 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1345 void tick_cancel_sched_timer(int cpu)
1347 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1349 # ifdef CONFIG_HIGH_RES_TIMERS
1350 if (ts->sched_timer.base)
1351 hrtimer_cancel(&ts->sched_timer);
1354 memset(ts, 0, sizeof(*ts));
1359 * Async notification about clocksource changes
1361 void tick_clock_notify(void)
1365 for_each_possible_cpu(cpu)
1366 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1370 * Async notification about clock event changes
1372 void tick_oneshot_notify(void)
1374 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1376 set_bit(0, &ts->check_clocks);
1380 * Check, if a change happened, which makes oneshot possible.
1382 * Called cyclic from the hrtimer softirq (driven by the timer
1383 * softirq) allow_nohz signals, that we can switch into low-res nohz
1384 * mode, because high resolution timers are disabled (either compile
1385 * or runtime). Called with interrupts disabled.
1387 int tick_check_oneshot_change(int allow_nohz)
1389 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1391 if (!test_and_clear_bit(0, &ts->check_clocks))
1394 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1397 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1403 tick_nohz_switch_to_nohz();