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:
65 delta = ktime_sub(now, last_jiffies_update);
66 if (delta < tick_period)
69 /* Reevaluate with jiffies_lock held */
70 write_seqlock(&jiffies_lock);
72 delta = ktime_sub(now, last_jiffies_update);
73 if (delta >= tick_period) {
75 delta = ktime_sub(delta, tick_period);
76 last_jiffies_update = ktime_add(last_jiffies_update,
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 last_jiffies_update = ktime_add_ns(last_jiffies_update,
90 /* Keep the tick_next_period variable up to date */
91 tick_next_period = ktime_add(last_jiffies_update, tick_period);
93 write_sequnlock(&jiffies_lock);
96 write_sequnlock(&jiffies_lock);
101 * Initialize and return retrieve the jiffies update.
103 static ktime_t tick_init_jiffy_update(void)
107 write_seqlock(&jiffies_lock);
108 /* Did we start the jiffies update yet ? */
109 if (last_jiffies_update == 0)
110 last_jiffies_update = tick_next_period;
111 period = last_jiffies_update;
112 write_sequnlock(&jiffies_lock);
117 static void tick_sched_do_timer(ktime_t now)
119 int cpu = smp_processor_id();
121 #ifdef CONFIG_NO_HZ_COMMON
123 * Check if the do_timer duty was dropped. We don't care about
124 * concurrency: This happens only when the CPU in charge went
125 * into a long sleep. If two CPUs happen to assign themselves to
126 * this duty, then the jiffies update is still serialized by
129 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
130 && !tick_nohz_full_cpu(cpu))
131 tick_do_timer_cpu = cpu;
134 /* Check, if the jiffies need an update */
135 if (tick_do_timer_cpu == cpu)
136 tick_do_update_jiffies64(now);
139 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
141 #ifdef CONFIG_NO_HZ_COMMON
143 * When we are idle and the tick is stopped, we have to touch
144 * the watchdog as we might not schedule for a really long
145 * time. This happens on complete idle SMP systems while
146 * waiting on the login prompt. We also increment the "start of
147 * idle" jiffy stamp so the idle accounting adjustment we do
148 * when we go busy again does not account too much ticks.
150 if (ts->tick_stopped) {
151 touch_softlockup_watchdog_sched();
152 if (is_idle_task(current))
155 * In case the current tick fired too early past its expected
156 * expiration, make sure we don't bypass the next clock reprogramming
157 * to the same deadline.
162 update_process_times(user_mode(regs));
163 profile_tick(CPU_PROFILING);
167 #ifdef CONFIG_NO_HZ_FULL
168 cpumask_var_t tick_nohz_full_mask;
169 bool tick_nohz_full_running;
170 static atomic_t tick_dep_mask;
172 static bool check_tick_dependency(atomic_t *dep)
174 int val = atomic_read(dep);
176 if (val & TICK_DEP_MASK_POSIX_TIMER) {
177 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
181 if (val & TICK_DEP_MASK_PERF_EVENTS) {
182 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
186 if (val & TICK_DEP_MASK_SCHED) {
187 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
191 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
192 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
199 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
201 lockdep_assert_irqs_disabled();
203 if (unlikely(!cpu_online(cpu)))
206 if (check_tick_dependency(&tick_dep_mask))
209 if (check_tick_dependency(&ts->tick_dep_mask))
212 if (check_tick_dependency(¤t->tick_dep_mask))
215 if (check_tick_dependency(¤t->signal->tick_dep_mask))
221 static void nohz_full_kick_func(struct irq_work *work)
223 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
226 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
227 .func = nohz_full_kick_func,
231 * Kick this CPU if it's full dynticks in order to force it to
232 * re-evaluate its dependency on the tick and restart it if necessary.
233 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
236 static void tick_nohz_full_kick(void)
238 if (!tick_nohz_full_cpu(smp_processor_id()))
241 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
245 * Kick the CPU if it's full dynticks in order to force it to
246 * re-evaluate its dependency on the tick and restart it if necessary.
248 void tick_nohz_full_kick_cpu(int cpu)
250 if (!tick_nohz_full_cpu(cpu))
253 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
257 * Kick all full dynticks CPUs in order to force these to re-evaluate
258 * their dependency on the tick and restart it if necessary.
260 static void tick_nohz_full_kick_all(void)
264 if (!tick_nohz_full_running)
268 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
269 tick_nohz_full_kick_cpu(cpu);
273 static void tick_nohz_dep_set_all(atomic_t *dep,
274 enum tick_dep_bits bit)
278 prev = atomic_fetch_or(BIT(bit), dep);
280 tick_nohz_full_kick_all();
284 * Set a global tick dependency. Used by perf events that rely on freq and
287 void tick_nohz_dep_set(enum tick_dep_bits bit)
289 tick_nohz_dep_set_all(&tick_dep_mask, bit);
292 void tick_nohz_dep_clear(enum tick_dep_bits bit)
294 atomic_andnot(BIT(bit), &tick_dep_mask);
298 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
299 * manage events throttling.
301 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
304 struct tick_sched *ts;
306 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
308 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
311 /* Perf needs local kick that is NMI safe */
312 if (cpu == smp_processor_id()) {
313 tick_nohz_full_kick();
315 /* Remote irq work not NMI-safe */
316 if (!WARN_ON_ONCE(in_nmi()))
317 tick_nohz_full_kick_cpu(cpu);
323 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
325 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
327 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
331 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
334 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
337 * We could optimize this with just kicking the target running the task
338 * if that noise matters for nohz full users.
340 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
343 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
345 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
349 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
350 * per process timers.
352 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
354 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
357 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
359 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
363 * Re-evaluate the need for the tick as we switch the current task.
364 * It might need the tick due to per task/process properties:
365 * perf events, posix CPU timers, ...
367 void __tick_nohz_task_switch(void)
370 struct tick_sched *ts;
372 local_irq_save(flags);
374 if (!tick_nohz_full_cpu(smp_processor_id()))
377 ts = this_cpu_ptr(&tick_cpu_sched);
379 if (ts->tick_stopped) {
380 if (atomic_read(¤t->tick_dep_mask) ||
381 atomic_read(¤t->signal->tick_dep_mask))
382 tick_nohz_full_kick();
385 local_irq_restore(flags);
388 /* Get the boot-time nohz CPU list from the kernel parameters. */
389 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
391 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
392 cpumask_copy(tick_nohz_full_mask, cpumask);
393 tick_nohz_full_running = true;
396 static int tick_nohz_cpu_down(unsigned int cpu)
399 * The boot CPU handles housekeeping duty (unbound timers,
400 * workqueues, timekeeping, ...) on behalf of full dynticks
401 * CPUs. It must remain online when nohz full is enabled.
403 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
408 static int tick_nohz_init_all(void)
412 #ifdef CONFIG_NO_HZ_FULL_ALL
413 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
414 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
418 cpumask_setall(tick_nohz_full_mask);
419 tick_nohz_full_running = true;
424 void __init tick_nohz_init(void)
428 if (!tick_nohz_full_running) {
429 if (tick_nohz_init_all() < 0)
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 int tick_nohz_tick_stopped(void)
486 return __this_cpu_read(tick_cpu_sched.tick_stopped);
490 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
492 * Called from interrupt entry when the CPU was idle
494 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
495 * must be updated. Otherwise an interrupt handler could use a stale jiffy
496 * value. We do this unconditionally on any CPU, as we don't know whether the
497 * CPU, which has the update task assigned is in a long sleep.
499 static void tick_nohz_update_jiffies(ktime_t now)
503 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
505 local_irq_save(flags);
506 tick_do_update_jiffies64(now);
507 local_irq_restore(flags);
509 touch_softlockup_watchdog_sched();
513 * Updates the per-CPU time idle statistics counters
516 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
520 if (ts->idle_active) {
521 delta = ktime_sub(now, ts->idle_entrytime);
522 if (nr_iowait_cpu(cpu) > 0)
523 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
525 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
526 ts->idle_entrytime = now;
529 if (last_update_time)
530 *last_update_time = ktime_to_us(now);
534 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
536 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
539 sched_clock_idle_wakeup_event();
542 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
544 ktime_t now = ktime_get();
546 ts->idle_entrytime = now;
548 sched_clock_idle_sleep_event();
553 * get_cpu_idle_time_us - get the total idle time of a CPU
554 * @cpu: CPU number to query
555 * @last_update_time: variable to store update time in. Do not update
558 * Return the cumulative idle time (since boot) for a given
559 * CPU, in microseconds.
561 * This time is measured via accounting rather than sampling,
562 * and is as accurate as ktime_get() is.
564 * This function returns -1 if NOHZ is not enabled.
566 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
568 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
571 if (!tick_nohz_active)
575 if (last_update_time) {
576 update_ts_time_stats(cpu, ts, now, last_update_time);
577 idle = ts->idle_sleeptime;
579 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
580 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
582 idle = ktime_add(ts->idle_sleeptime, delta);
584 idle = ts->idle_sleeptime;
588 return ktime_to_us(idle);
591 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
594 * get_cpu_iowait_time_us - get the total iowait time of a CPU
595 * @cpu: CPU number to query
596 * @last_update_time: variable to store update time in. Do not update
599 * Return the cumulative iowait time (since boot) for a given
600 * CPU, in microseconds.
602 * This time is measured via accounting rather than sampling,
603 * and is as accurate as ktime_get() is.
605 * This function returns -1 if NOHZ is not enabled.
607 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
609 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
612 if (!tick_nohz_active)
616 if (last_update_time) {
617 update_ts_time_stats(cpu, ts, now, last_update_time);
618 iowait = ts->iowait_sleeptime;
620 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
621 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
623 iowait = ktime_add(ts->iowait_sleeptime, delta);
625 iowait = ts->iowait_sleeptime;
629 return ktime_to_us(iowait);
631 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
633 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
635 hrtimer_cancel(&ts->sched_timer);
636 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
638 /* Forward the time to expire in the future */
639 hrtimer_forward(&ts->sched_timer, now, tick_period);
641 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
642 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
644 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
647 * Reset to make sure next tick stop doesn't get fooled by past
648 * cached clock deadline.
653 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
654 ktime_t now, int cpu)
656 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
657 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
658 unsigned long seq, basejiff;
661 /* Read jiffies and the time when jiffies were updated last */
663 seq = read_seqbegin(&jiffies_lock);
664 basemono = last_jiffies_update;
666 } while (read_seqretry(&jiffies_lock, seq));
667 ts->last_jiffies = basejiff;
669 if (rcu_needs_cpu(basemono, &next_rcu) ||
670 arch_needs_cpu() || irq_work_needs_cpu()) {
671 next_tick = basemono + TICK_NSEC;
674 * Get the next pending timer. If high resolution
675 * timers are enabled this only takes the timer wheel
676 * timers into account. If high resolution timers are
677 * disabled this also looks at the next expiring
680 next_tmr = get_next_timer_interrupt(basejiff, basemono);
681 ts->next_timer = next_tmr;
682 /* Take the next rcu event into account */
683 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
687 * If the tick is due in the next period, keep it ticking or
688 * force prod the timer.
690 delta = next_tick - basemono;
691 if (delta <= (u64)TICK_NSEC) {
693 * Tell the timer code that the base is not idle, i.e. undo
694 * the effect of get_next_timer_interrupt():
698 * We've not stopped the tick yet, and there's a timer in the
699 * next period, so no point in stopping it either, bail.
701 if (!ts->tick_stopped) {
708 * If this CPU is the one which updates jiffies, then give up
709 * the assignment and let it be taken by the CPU which runs
710 * the tick timer next, which might be this CPU as well. If we
711 * don't drop this here the jiffies might be stale and
712 * do_timer() never invoked. Keep track of the fact that it
713 * was the one which had the do_timer() duty last. If this CPU
714 * is the one which had the do_timer() duty last, we limit the
715 * sleep time to the timekeeping max_deferment value.
716 * Otherwise we can sleep as long as we want.
718 delta = timekeeping_max_deferment();
719 if (cpu == tick_do_timer_cpu) {
720 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
721 ts->do_timer_last = 1;
722 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
724 ts->do_timer_last = 0;
725 } else if (!ts->do_timer_last) {
729 #ifdef CONFIG_NO_HZ_FULL
730 /* Limit the tick delta to the maximum scheduler deferment */
732 delta = min(delta, scheduler_tick_max_deferment());
735 /* Calculate the next expiry time */
736 if (delta < (KTIME_MAX - basemono))
737 expires = basemono + delta;
741 expires = min_t(u64, expires, next_tick);
744 /* Skip reprogram of event if its not changed */
745 if (ts->tick_stopped && (expires == ts->next_tick)) {
746 /* Sanity check: make sure clockevent is actually programmed */
747 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
751 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
752 basemono, ts->next_tick, dev->next_event,
753 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
757 * nohz_stop_sched_tick can be called several times before
758 * the nohz_restart_sched_tick is called. This happens when
759 * interrupts arrive which do not cause a reschedule. In the
760 * first call we save the current tick time, so we can restart
761 * the scheduler tick in nohz_restart_sched_tick.
763 if (!ts->tick_stopped) {
764 calc_load_nohz_start();
765 cpu_load_update_nohz_start();
768 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
769 ts->tick_stopped = 1;
770 trace_tick_stop(1, TICK_DEP_MASK_NONE);
773 ts->next_tick = tick;
776 * If the expiration time == KTIME_MAX, then we simply stop
779 if (unlikely(expires == KTIME_MAX)) {
780 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
781 hrtimer_cancel(&ts->sched_timer);
785 hrtimer_set_expires(&ts->sched_timer, tick);
787 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
788 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
790 tick_program_event(tick, 1);
793 * Update the estimated sleep length until the next timer
794 * (not only the tick).
796 ts->sleep_length = ktime_sub(dev->next_event, now);
800 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
802 /* Update jiffies first */
803 tick_do_update_jiffies64(now);
804 cpu_load_update_nohz_stop();
807 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
808 * the clock forward checks in the enqueue path:
812 calc_load_nohz_stop();
813 touch_softlockup_watchdog_sched();
815 * Cancel the scheduled timer and restore the tick
817 ts->tick_stopped = 0;
818 ts->idle_exittime = now;
820 tick_nohz_restart(ts, now);
823 static void tick_nohz_full_update_tick(struct tick_sched *ts)
825 #ifdef CONFIG_NO_HZ_FULL
826 int cpu = smp_processor_id();
828 if (!tick_nohz_full_cpu(cpu))
831 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
834 if (can_stop_full_tick(cpu, ts))
835 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
836 else if (ts->tick_stopped)
837 tick_nohz_restart_sched_tick(ts, ktime_get());
841 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
844 * If this CPU is offline and it is the one which updates
845 * jiffies, then give up the assignment and let it be taken by
846 * the CPU which runs the tick timer next. If we don't drop
847 * this here the jiffies might be stale and do_timer() never
850 if (unlikely(!cpu_online(cpu))) {
851 if (cpu == tick_do_timer_cpu)
852 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
854 * Make sure the CPU doesn't get fooled by obsolete tick
855 * deadline if it comes back online later.
861 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
862 ts->sleep_length = NSEC_PER_SEC / HZ;
869 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
870 static int ratelimit;
872 if (ratelimit < 10 &&
873 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
874 pr_warn("NOHZ: local_softirq_pending %02x\n",
875 (unsigned int) local_softirq_pending());
881 if (tick_nohz_full_enabled()) {
883 * Keep the tick alive to guarantee timekeeping progression
884 * if there are full dynticks CPUs around
886 if (tick_do_timer_cpu == cpu)
889 * Boot safety: make sure the timekeeping duty has been
890 * assigned before entering dyntick-idle mode,
892 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
899 static void __tick_nohz_idle_enter(struct tick_sched *ts)
901 ktime_t now, expires;
902 int cpu = smp_processor_id();
904 now = tick_nohz_start_idle(ts);
906 if (can_stop_idle_tick(cpu, ts)) {
907 int was_stopped = ts->tick_stopped;
911 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
914 ts->idle_expires = expires;
917 if (!was_stopped && ts->tick_stopped) {
918 ts->idle_jiffies = ts->last_jiffies;
919 nohz_balance_enter_idle(cpu);
925 * tick_nohz_idle_enter - stop the idle tick from the idle task
927 * When the next event is more than a tick into the future, stop the idle tick
928 * Called when we start the idle loop.
930 * The arch is responsible of calling:
932 * - rcu_idle_enter() after its last use of RCU before the CPU is put
934 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
936 void tick_nohz_idle_enter(void)
938 struct tick_sched *ts;
940 lockdep_assert_irqs_enabled();
942 * Update the idle state in the scheduler domain hierarchy
943 * when tick_nohz_stop_sched_tick() is called from the idle loop.
944 * State will be updated to busy during the first busy tick after
947 set_cpu_sd_state_idle();
951 ts = this_cpu_ptr(&tick_cpu_sched);
953 __tick_nohz_idle_enter(ts);
959 * tick_nohz_irq_exit - update next tick event from interrupt exit
961 * When an interrupt fires while we are idle and it doesn't cause
962 * a reschedule, it may still add, modify or delete a timer, enqueue
963 * an RCU callback, etc...
964 * So we need to re-calculate and reprogram the next tick event.
966 void tick_nohz_irq_exit(void)
968 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
971 __tick_nohz_idle_enter(ts);
973 tick_nohz_full_update_tick(ts);
977 * tick_nohz_get_sleep_length - return the length of the current sleep
979 * Called from power state control code with interrupts disabled
981 ktime_t tick_nohz_get_sleep_length(void)
983 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
985 return ts->sleep_length;
989 * tick_nohz_get_idle_calls - return the current idle calls counter value
991 * Called from the schedutil frequency scaling governor in scheduler context.
993 unsigned long tick_nohz_get_idle_calls(void)
995 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
997 return ts->idle_calls;
1000 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1002 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1003 unsigned long ticks;
1005 if (vtime_accounting_cpu_enabled())
1008 * We stopped the tick in idle. Update process times would miss the
1009 * time we slept as update_process_times does only a 1 tick
1010 * accounting. Enforce that this is accounted to idle !
1012 ticks = jiffies - ts->idle_jiffies;
1014 * We might be one off. Do not randomly account a huge number of ticks!
1016 if (ticks && ticks < LONG_MAX)
1017 account_idle_ticks(ticks);
1022 * tick_nohz_idle_exit - restart the idle tick from the idle task
1024 * Restart the idle tick when the CPU is woken up from idle
1025 * This also exit the RCU extended quiescent state. The CPU
1026 * can use RCU again after this function is called.
1028 void tick_nohz_idle_exit(void)
1030 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1033 local_irq_disable();
1035 WARN_ON_ONCE(!ts->inidle);
1039 if (ts->idle_active || ts->tick_stopped)
1042 if (ts->idle_active)
1043 tick_nohz_stop_idle(ts, now);
1045 if (ts->tick_stopped) {
1046 tick_nohz_restart_sched_tick(ts, now);
1047 tick_nohz_account_idle_ticks(ts);
1054 * The nohz low res interrupt handler
1056 static void tick_nohz_handler(struct clock_event_device *dev)
1058 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1059 struct pt_regs *regs = get_irq_regs();
1060 ktime_t now = ktime_get();
1062 dev->next_event = KTIME_MAX;
1064 tick_sched_do_timer(now);
1065 tick_sched_handle(ts, regs);
1067 /* No need to reprogram if we are running tickless */
1068 if (unlikely(ts->tick_stopped))
1071 hrtimer_forward(&ts->sched_timer, now, tick_period);
1072 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1075 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1077 if (!tick_nohz_enabled)
1079 ts->nohz_mode = mode;
1080 /* One update is enough */
1081 if (!test_and_set_bit(0, &tick_nohz_active))
1082 timers_update_migration(true);
1086 * tick_nohz_switch_to_nohz - switch to nohz mode
1088 static void tick_nohz_switch_to_nohz(void)
1090 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1093 if (!tick_nohz_enabled)
1096 if (tick_switch_to_oneshot(tick_nohz_handler))
1100 * Recycle the hrtimer in ts, so we can share the
1101 * hrtimer_forward with the highres code.
1103 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1104 /* Get the next period */
1105 next = tick_init_jiffy_update();
1107 hrtimer_set_expires(&ts->sched_timer, next);
1108 hrtimer_forward_now(&ts->sched_timer, tick_period);
1109 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1110 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1113 static inline void tick_nohz_irq_enter(void)
1115 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1118 if (!ts->idle_active && !ts->tick_stopped)
1121 if (ts->idle_active)
1122 tick_nohz_stop_idle(ts, now);
1123 if (ts->tick_stopped)
1124 tick_nohz_update_jiffies(now);
1129 static inline void tick_nohz_switch_to_nohz(void) { }
1130 static inline void tick_nohz_irq_enter(void) { }
1131 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1133 #endif /* CONFIG_NO_HZ_COMMON */
1136 * Called from irq_enter to notify about the possible interruption of idle()
1138 void tick_irq_enter(void)
1140 tick_check_oneshot_broadcast_this_cpu();
1141 tick_nohz_irq_enter();
1145 * High resolution timer specific code
1147 #ifdef CONFIG_HIGH_RES_TIMERS
1149 * We rearm the timer until we get disabled by the idle code.
1150 * Called with interrupts disabled.
1152 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1154 struct tick_sched *ts =
1155 container_of(timer, struct tick_sched, sched_timer);
1156 struct pt_regs *regs = get_irq_regs();
1157 ktime_t now = ktime_get();
1159 tick_sched_do_timer(now);
1162 * Do not call, when we are not in irq context and have
1163 * no valid regs pointer
1166 tick_sched_handle(ts, regs);
1170 /* No need to reprogram if we are in idle or full dynticks mode */
1171 if (unlikely(ts->tick_stopped))
1172 return HRTIMER_NORESTART;
1174 hrtimer_forward(timer, now, tick_period);
1176 return HRTIMER_RESTART;
1179 static int sched_skew_tick;
1181 static int __init skew_tick(char *str)
1183 get_option(&str, &sched_skew_tick);
1187 early_param("skew_tick", skew_tick);
1190 * tick_setup_sched_timer - setup the tick emulation timer
1192 void tick_setup_sched_timer(void)
1194 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1195 ktime_t now = ktime_get();
1198 * Emulate tick processing via per-CPU hrtimers:
1200 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1201 ts->sched_timer.function = tick_sched_timer;
1203 /* Get the next period (per-CPU) */
1204 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1206 /* Offset the tick to avert jiffies_lock contention. */
1207 if (sched_skew_tick) {
1208 u64 offset = ktime_to_ns(tick_period) >> 1;
1209 do_div(offset, num_possible_cpus());
1210 offset *= smp_processor_id();
1211 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1214 hrtimer_forward(&ts->sched_timer, now, tick_period);
1215 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1216 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1218 #endif /* HIGH_RES_TIMERS */
1220 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1221 void tick_cancel_sched_timer(int cpu)
1223 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1225 # ifdef CONFIG_HIGH_RES_TIMERS
1226 if (ts->sched_timer.base)
1227 hrtimer_cancel(&ts->sched_timer);
1230 memset(ts, 0, sizeof(*ts));
1235 * Async notification about clocksource changes
1237 void tick_clock_notify(void)
1241 for_each_possible_cpu(cpu)
1242 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1246 * Async notification about clock event changes
1248 void tick_oneshot_notify(void)
1250 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1252 set_bit(0, &ts->check_clocks);
1256 * Check, if a change happened, which makes oneshot possible.
1258 * Called cyclic from the hrtimer softirq (driven by the timer
1259 * softirq) allow_nohz signals, that we can switch into low-res nohz
1260 * mode, because high resolution timers are disabled (either compile
1261 * or runtime). Called with interrupts disabled.
1263 int tick_check_oneshot_change(int allow_nohz)
1265 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1267 if (!test_and_clear_bit(0, &ts->check_clocks))
1270 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1273 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1279 tick_nohz_switch_to_nohz();