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/profile.h>
21 #include <linux/sched.h>
22 #include <linux/module.h>
23 #include <linux/irq_work.h>
24 #include <linux/posix-timers.h>
25 #include <linux/context_tracking.h>
27 #include <asm/irq_regs.h>
29 #include "tick-internal.h"
31 #include <trace/events/timer.h>
34 * Per cpu nohz control structure
36 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
38 struct tick_sched *tick_get_tick_sched(int cpu)
40 return &per_cpu(tick_cpu_sched, cpu);
43 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
45 * The time, when the last jiffy update happened. Protected by jiffies_lock.
47 static ktime_t last_jiffies_update;
50 * Must be called with interrupts disabled !
52 static void tick_do_update_jiffies64(ktime_t now)
54 unsigned long ticks = 0;
58 * Do a quick check without holding jiffies_lock:
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta.tv64 < tick_period.tv64)
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
74 /* Slow path for long timeouts */
75 if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 s64 incr = ktime_to_ns(tick_period);
78 ticks = ktime_divns(delta, incr);
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
88 write_sequnlock(&jiffies_lock);
91 write_sequnlock(&jiffies_lock);
96 * Initialize and return retrieve the jiffies update.
98 static ktime_t tick_init_jiffy_update(void)
102 write_seqlock(&jiffies_lock);
103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update.tv64 == 0)
105 last_jiffies_update = tick_next_period;
106 period = last_jiffies_update;
107 write_sequnlock(&jiffies_lock);
112 static void tick_sched_do_timer(ktime_t now)
114 int cpu = smp_processor_id();
116 #ifdef CONFIG_NO_HZ_COMMON
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the cpu in charge went
120 * into a long sleep. If two cpus happen to assign themself to
121 * this duty, then the jiffies update is still serialized by
124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
125 && !tick_nohz_full_cpu(cpu))
126 tick_do_timer_cpu = cpu;
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu == cpu)
131 tick_do_update_jiffies64(now);
134 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
136 #ifdef CONFIG_NO_HZ_COMMON
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
145 if (ts->tick_stopped) {
146 touch_softlockup_watchdog_sched();
147 if (is_idle_task(current))
151 update_process_times(user_mode(regs));
152 profile_tick(CPU_PROFILING);
156 #ifdef CONFIG_NO_HZ_FULL
157 cpumask_var_t tick_nohz_full_mask;
158 cpumask_var_t housekeeping_mask;
159 bool tick_nohz_full_running;
160 static unsigned long tick_dep_mask;
162 static void trace_tick_dependency(unsigned long dep)
164 if (dep & TICK_DEP_MASK_POSIX_TIMER) {
165 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
169 if (dep & TICK_DEP_MASK_PERF_EVENTS) {
170 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
174 if (dep & TICK_DEP_MASK_SCHED) {
175 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
179 if (dep & TICK_DEP_MASK_CLOCK_UNSTABLE)
180 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
183 static bool can_stop_full_tick(struct tick_sched *ts)
185 WARN_ON_ONCE(!irqs_disabled());
188 trace_tick_dependency(tick_dep_mask);
192 if (ts->tick_dep_mask) {
193 trace_tick_dependency(ts->tick_dep_mask);
197 if (current->tick_dep_mask) {
198 trace_tick_dependency(current->tick_dep_mask);
202 if (current->signal->tick_dep_mask) {
203 trace_tick_dependency(current->signal->tick_dep_mask);
210 static void nohz_full_kick_func(struct irq_work *work)
212 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
215 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
216 .func = nohz_full_kick_func,
220 * Kick this CPU if it's full dynticks in order to force it to
221 * re-evaluate its dependency on the tick and restart it if necessary.
222 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
225 static void tick_nohz_full_kick(void)
227 if (!tick_nohz_full_cpu(smp_processor_id()))
230 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
234 * Kick the CPU if it's full dynticks in order to force it to
235 * re-evaluate its dependency on the tick and restart it if necessary.
237 void tick_nohz_full_kick_cpu(int cpu)
239 if (!tick_nohz_full_cpu(cpu))
242 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
246 * Kick all full dynticks CPUs in order to force these to re-evaluate
247 * their dependency on the tick and restart it if necessary.
249 static void tick_nohz_full_kick_all(void)
253 if (!tick_nohz_full_running)
257 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
258 tick_nohz_full_kick_cpu(cpu);
262 static void tick_nohz_dep_set_all(unsigned long *dep,
263 enum tick_dep_bits bit)
267 prev = fetch_or(dep, BIT_MASK(bit));
269 tick_nohz_full_kick_all();
273 * Set a global tick dependency. Used by perf events that rely on freq and
276 void tick_nohz_dep_set(enum tick_dep_bits bit)
278 tick_nohz_dep_set_all(&tick_dep_mask, bit);
281 void tick_nohz_dep_clear(enum tick_dep_bits bit)
283 clear_bit(bit, &tick_dep_mask);
287 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
288 * manage events throttling.
290 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
293 struct tick_sched *ts;
295 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
297 prev = fetch_or(&ts->tick_dep_mask, BIT_MASK(bit));
300 /* Perf needs local kick that is NMI safe */
301 if (cpu == smp_processor_id()) {
302 tick_nohz_full_kick();
304 /* Remote irq work not NMI-safe */
305 if (!WARN_ON_ONCE(in_nmi()))
306 tick_nohz_full_kick_cpu(cpu);
312 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
314 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
316 clear_bit(bit, &ts->tick_dep_mask);
320 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
323 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
326 * We could optimize this with just kicking the target running the task
327 * if that noise matters for nohz full users.
329 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
332 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
334 clear_bit(bit, &tsk->tick_dep_mask);
338 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
339 * per process timers.
341 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
343 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
346 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
348 clear_bit(bit, &sig->tick_dep_mask);
352 * Re-evaluate the need for the tick as we switch the current task.
353 * It might need the tick due to per task/process properties:
354 * perf events, posix cpu timers, ...
356 void __tick_nohz_task_switch(void)
359 struct tick_sched *ts;
361 local_irq_save(flags);
363 if (!tick_nohz_full_cpu(smp_processor_id()))
366 ts = this_cpu_ptr(&tick_cpu_sched);
368 if (ts->tick_stopped) {
369 if (current->tick_dep_mask || current->signal->tick_dep_mask)
370 tick_nohz_full_kick();
373 local_irq_restore(flags);
376 /* Parse the boot-time nohz CPU list from the kernel parameters. */
377 static int __init tick_nohz_full_setup(char *str)
379 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
380 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
381 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
382 free_bootmem_cpumask_var(tick_nohz_full_mask);
385 tick_nohz_full_running = true;
389 __setup("nohz_full=", tick_nohz_full_setup);
391 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
392 unsigned long action,
395 unsigned int cpu = (unsigned long)hcpu;
397 switch (action & ~CPU_TASKS_FROZEN) {
398 case CPU_DOWN_PREPARE:
400 * The boot CPU handles housekeeping duty (unbound timers,
401 * workqueues, timekeeping, ...) on behalf of full dynticks
402 * CPUs. It must remain online when nohz full is enabled.
404 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
411 static int tick_nohz_init_all(void)
415 #ifdef CONFIG_NO_HZ_FULL_ALL
416 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
417 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
421 cpumask_setall(tick_nohz_full_mask);
422 tick_nohz_full_running = true;
427 void __init tick_nohz_init(void)
431 if (!tick_nohz_full_running) {
432 if (tick_nohz_init_all() < 0)
436 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
437 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
438 cpumask_clear(tick_nohz_full_mask);
439 tick_nohz_full_running = false;
444 * Full dynticks uses irq work to drive the tick rescheduling on safe
445 * locking contexts. But then we need irq work to raise its own
446 * interrupts to avoid circular dependency on the tick
448 if (!arch_irq_work_has_interrupt()) {
449 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
450 cpumask_clear(tick_nohz_full_mask);
451 cpumask_copy(housekeeping_mask, cpu_possible_mask);
452 tick_nohz_full_running = false;
456 cpu = smp_processor_id();
458 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
459 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
461 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
464 cpumask_andnot(housekeeping_mask,
465 cpu_possible_mask, tick_nohz_full_mask);
467 for_each_cpu(cpu, tick_nohz_full_mask)
468 context_tracking_cpu_set(cpu);
470 cpu_notifier(tick_nohz_cpu_down_callback, 0);
471 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
472 cpumask_pr_args(tick_nohz_full_mask));
475 * We need at least one CPU to handle housekeeping work such
476 * as timekeeping, unbound timers, workqueues, ...
478 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
483 * NOHZ - aka dynamic tick functionality
485 #ifdef CONFIG_NO_HZ_COMMON
489 bool tick_nohz_enabled __read_mostly = true;
490 unsigned long tick_nohz_active __read_mostly;
492 * Enable / Disable tickless mode
494 static int __init setup_tick_nohz(char *str)
496 return (kstrtobool(str, &tick_nohz_enabled) == 0);
499 __setup("nohz=", setup_tick_nohz);
501 int tick_nohz_tick_stopped(void)
503 return __this_cpu_read(tick_cpu_sched.tick_stopped);
507 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
509 * Called from interrupt entry when the CPU was idle
511 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
512 * must be updated. Otherwise an interrupt handler could use a stale jiffy
513 * value. We do this unconditionally on any cpu, as we don't know whether the
514 * cpu, which has the update task assigned is in a long sleep.
516 static void tick_nohz_update_jiffies(ktime_t now)
520 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
522 local_irq_save(flags);
523 tick_do_update_jiffies64(now);
524 local_irq_restore(flags);
526 touch_softlockup_watchdog_sched();
530 * Updates the per cpu time idle statistics counters
533 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
537 if (ts->idle_active) {
538 delta = ktime_sub(now, ts->idle_entrytime);
539 if (nr_iowait_cpu(cpu) > 0)
540 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
542 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
543 ts->idle_entrytime = now;
546 if (last_update_time)
547 *last_update_time = ktime_to_us(now);
551 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
553 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
556 sched_clock_idle_wakeup_event(0);
559 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
561 ktime_t now = ktime_get();
563 ts->idle_entrytime = now;
565 sched_clock_idle_sleep_event();
570 * get_cpu_idle_time_us - get the total idle time of a cpu
571 * @cpu: CPU number to query
572 * @last_update_time: variable to store update time in. Do not update
575 * Return the cummulative idle time (since boot) for a given
576 * CPU, in microseconds.
578 * This time is measured via accounting rather than sampling,
579 * and is as accurate as ktime_get() is.
581 * This function returns -1 if NOHZ is not enabled.
583 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
585 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
588 if (!tick_nohz_active)
592 if (last_update_time) {
593 update_ts_time_stats(cpu, ts, now, last_update_time);
594 idle = ts->idle_sleeptime;
596 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
597 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
599 idle = ktime_add(ts->idle_sleeptime, delta);
601 idle = ts->idle_sleeptime;
605 return ktime_to_us(idle);
608 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
611 * get_cpu_iowait_time_us - get the total iowait time of a cpu
612 * @cpu: CPU number to query
613 * @last_update_time: variable to store update time in. Do not update
616 * Return the cummulative iowait time (since boot) for a given
617 * CPU, in microseconds.
619 * This time is measured via accounting rather than sampling,
620 * and is as accurate as ktime_get() is.
622 * This function returns -1 if NOHZ is not enabled.
624 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
626 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
629 if (!tick_nohz_active)
633 if (last_update_time) {
634 update_ts_time_stats(cpu, ts, now, last_update_time);
635 iowait = ts->iowait_sleeptime;
637 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
638 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
640 iowait = ktime_add(ts->iowait_sleeptime, delta);
642 iowait = ts->iowait_sleeptime;
646 return ktime_to_us(iowait);
648 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
650 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
652 hrtimer_cancel(&ts->sched_timer);
653 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
655 /* Forward the time to expire in the future */
656 hrtimer_forward(&ts->sched_timer, now, tick_period);
658 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
659 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
661 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
664 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
665 ktime_t now, int cpu)
667 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
668 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
669 unsigned long seq, basejiff;
672 /* Read jiffies and the time when jiffies were updated last */
674 seq = read_seqbegin(&jiffies_lock);
675 basemono = last_jiffies_update.tv64;
677 } while (read_seqretry(&jiffies_lock, seq));
678 ts->last_jiffies = basejiff;
680 if (rcu_needs_cpu(basemono, &next_rcu) ||
681 arch_needs_cpu() || irq_work_needs_cpu()) {
682 next_tick = basemono + TICK_NSEC;
685 * Get the next pending timer. If high resolution
686 * timers are enabled this only takes the timer wheel
687 * timers into account. If high resolution timers are
688 * disabled this also looks at the next expiring
691 next_tmr = get_next_timer_interrupt(basejiff, basemono);
692 ts->next_timer = next_tmr;
693 /* Take the next rcu event into account */
694 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
698 * If the tick is due in the next period, keep it ticking or
699 * force prod the timer.
701 delta = next_tick - basemono;
702 if (delta <= (u64)TICK_NSEC) {
705 * We've not stopped the tick yet, and there's a timer in the
706 * next period, so no point in stopping it either, bail.
708 if (!ts->tick_stopped)
712 * If, OTOH, we did stop it, but there's a pending (expired)
713 * timer reprogram the timer hardware to fire now.
715 * We will not restart the tick proper, just prod the timer
716 * hardware into firing an interrupt to process the pending
717 * timers. Just like tick_irq_exit() will not restart the tick
718 * for 'normal' interrupts.
720 * Only once we exit the idle loop will we re-enable the tick,
721 * see tick_nohz_idle_exit().
724 tick_nohz_restart(ts, now);
730 * If this cpu is the one which updates jiffies, then give up
731 * the assignment and let it be taken by the cpu which runs
732 * the tick timer next, which might be this cpu as well. If we
733 * don't drop this here the jiffies might be stale and
734 * do_timer() never invoked. Keep track of the fact that it
735 * was the one which had the do_timer() duty last. If this cpu
736 * is the one which had the do_timer() duty last, we limit the
737 * sleep time to the timekeeping max_deferement value.
738 * Otherwise we can sleep as long as we want.
740 delta = timekeeping_max_deferment();
741 if (cpu == tick_do_timer_cpu) {
742 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
743 ts->do_timer_last = 1;
744 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
746 ts->do_timer_last = 0;
747 } else if (!ts->do_timer_last) {
751 #ifdef CONFIG_NO_HZ_FULL
752 /* Limit the tick delta to the maximum scheduler deferment */
754 delta = min(delta, scheduler_tick_max_deferment());
757 /* Calculate the next expiry time */
758 if (delta < (KTIME_MAX - basemono))
759 expires = basemono + delta;
763 expires = min_t(u64, expires, next_tick);
766 /* Skip reprogram of event if its not changed */
767 if (ts->tick_stopped && (expires == dev->next_event.tv64))
771 * nohz_stop_sched_tick can be called several times before
772 * the nohz_restart_sched_tick is called. This happens when
773 * interrupts arrive which do not cause a reschedule. In the
774 * first call we save the current tick time, so we can restart
775 * the scheduler tick in nohz_restart_sched_tick.
777 if (!ts->tick_stopped) {
778 nohz_balance_enter_idle(cpu);
779 calc_load_enter_idle();
781 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
782 ts->tick_stopped = 1;
783 trace_tick_stop(1, TICK_DEP_MASK_NONE);
787 * If the expiration time == KTIME_MAX, then we simply stop
790 if (unlikely(expires == KTIME_MAX)) {
791 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
792 hrtimer_cancel(&ts->sched_timer);
796 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
797 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
799 tick_program_event(tick, 1);
801 /* Update the estimated sleep length */
802 ts->sleep_length = ktime_sub(dev->next_event, now);
806 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active)
808 /* Update jiffies first */
809 tick_do_update_jiffies64(now);
810 update_cpu_load_nohz(active);
812 calc_load_exit_idle();
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(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(), 1);
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;
856 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
857 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
864 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
865 static int ratelimit;
867 if (ratelimit < 10 &&
868 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
869 pr_warn("NOHZ: local_softirq_pending %02x\n",
870 (unsigned int) local_softirq_pending());
876 if (tick_nohz_full_enabled()) {
878 * Keep the tick alive to guarantee timekeeping progression
879 * if there are full dynticks CPUs around
881 if (tick_do_timer_cpu == cpu)
884 * Boot safety: make sure the timekeeping duty has been
885 * assigned before entering dyntick-idle mode,
887 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
894 static void __tick_nohz_idle_enter(struct tick_sched *ts)
896 ktime_t now, expires;
897 int cpu = smp_processor_id();
899 now = tick_nohz_start_idle(ts);
901 if (can_stop_idle_tick(cpu, ts)) {
902 int was_stopped = ts->tick_stopped;
906 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
907 if (expires.tv64 > 0LL) {
909 ts->idle_expires = expires;
912 if (!was_stopped && ts->tick_stopped)
913 ts->idle_jiffies = ts->last_jiffies;
918 * tick_nohz_idle_enter - stop the idle tick from the idle task
920 * When the next event is more than a tick into the future, stop the idle tick
921 * Called when we start the idle loop.
923 * The arch is responsible of calling:
925 * - rcu_idle_enter() after its last use of RCU before the CPU is put
927 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
929 void tick_nohz_idle_enter(void)
931 struct tick_sched *ts;
933 WARN_ON_ONCE(irqs_disabled());
936 * Update the idle state in the scheduler domain hierarchy
937 * when tick_nohz_stop_sched_tick() is called from the idle loop.
938 * State will be updated to busy during the first busy tick after
941 set_cpu_sd_state_idle();
945 ts = this_cpu_ptr(&tick_cpu_sched);
947 __tick_nohz_idle_enter(ts);
953 * tick_nohz_irq_exit - update next tick event from interrupt exit
955 * When an interrupt fires while we are idle and it doesn't cause
956 * a reschedule, it may still add, modify or delete a timer, enqueue
957 * an RCU callback, etc...
958 * So we need to re-calculate and reprogram the next tick event.
960 void tick_nohz_irq_exit(void)
962 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
965 __tick_nohz_idle_enter(ts);
967 tick_nohz_full_update_tick(ts);
971 * tick_nohz_get_sleep_length - return the length of the current sleep
973 * Called from power state control code with interrupts disabled
975 ktime_t tick_nohz_get_sleep_length(void)
977 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
979 return ts->sleep_length;
982 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
984 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
987 if (vtime_accounting_cpu_enabled())
990 * We stopped the tick in idle. Update process times would miss the
991 * time we slept as update_process_times does only a 1 tick
992 * accounting. Enforce that this is accounted to idle !
994 ticks = jiffies - ts->idle_jiffies;
996 * We might be one off. Do not randomly account a huge number of ticks!
998 if (ticks && ticks < LONG_MAX)
999 account_idle_ticks(ticks);
1004 * tick_nohz_idle_exit - restart the idle tick from the idle task
1006 * Restart the idle tick when the CPU is woken up from idle
1007 * This also exit the RCU extended quiescent state. The CPU
1008 * can use RCU again after this function is called.
1010 void tick_nohz_idle_exit(void)
1012 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1015 local_irq_disable();
1017 WARN_ON_ONCE(!ts->inidle);
1021 if (ts->idle_active || ts->tick_stopped)
1024 if (ts->idle_active)
1025 tick_nohz_stop_idle(ts, now);
1027 if (ts->tick_stopped) {
1028 tick_nohz_restart_sched_tick(ts, now, 0);
1029 tick_nohz_account_idle_ticks(ts);
1036 * The nohz low res interrupt handler
1038 static void tick_nohz_handler(struct clock_event_device *dev)
1040 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1041 struct pt_regs *regs = get_irq_regs();
1042 ktime_t now = ktime_get();
1044 dev->next_event.tv64 = KTIME_MAX;
1046 tick_sched_do_timer(now);
1047 tick_sched_handle(ts, regs);
1049 /* No need to reprogram if we are running tickless */
1050 if (unlikely(ts->tick_stopped))
1053 hrtimer_forward(&ts->sched_timer, now, tick_period);
1054 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1057 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1059 if (!tick_nohz_enabled)
1061 ts->nohz_mode = mode;
1062 /* One update is enough */
1063 if (!test_and_set_bit(0, &tick_nohz_active))
1064 timers_update_migration(true);
1068 * tick_nohz_switch_to_nohz - switch to nohz mode
1070 static void tick_nohz_switch_to_nohz(void)
1072 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1075 if (!tick_nohz_enabled)
1078 if (tick_switch_to_oneshot(tick_nohz_handler))
1082 * Recycle the hrtimer in ts, so we can share the
1083 * hrtimer_forward with the highres code.
1085 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1086 /* Get the next period */
1087 next = tick_init_jiffy_update();
1089 hrtimer_set_expires(&ts->sched_timer, next);
1090 hrtimer_forward_now(&ts->sched_timer, tick_period);
1091 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1092 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1096 * When NOHZ is enabled and the tick is stopped, we need to kick the
1097 * tick timer from irq_enter() so that the jiffies update is kept
1098 * alive during long running softirqs. That's ugly as hell, but
1099 * correctness is key even if we need to fix the offending softirq in
1102 * Note, this is different to tick_nohz_restart. We just kick the
1103 * timer and do not touch the other magic bits which need to be done
1104 * when idle is left.
1106 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1109 /* Switch back to 2.6.27 behaviour */
1113 * Do not touch the tick device, when the next expiry is either
1114 * already reached or less/equal than the tick period.
1116 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1117 if (delta.tv64 <= tick_period.tv64)
1120 tick_nohz_restart(ts, now);
1124 static inline void tick_nohz_irq_enter(void)
1126 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1129 if (!ts->idle_active && !ts->tick_stopped)
1132 if (ts->idle_active)
1133 tick_nohz_stop_idle(ts, now);
1134 if (ts->tick_stopped) {
1135 tick_nohz_update_jiffies(now);
1136 tick_nohz_kick_tick(ts, now);
1142 static inline void tick_nohz_switch_to_nohz(void) { }
1143 static inline void tick_nohz_irq_enter(void) { }
1144 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1146 #endif /* CONFIG_NO_HZ_COMMON */
1149 * Called from irq_enter to notify about the possible interruption of idle()
1151 void tick_irq_enter(void)
1153 tick_check_oneshot_broadcast_this_cpu();
1154 tick_nohz_irq_enter();
1158 * High resolution timer specific code
1160 #ifdef CONFIG_HIGH_RES_TIMERS
1162 * We rearm the timer until we get disabled by the idle code.
1163 * Called with interrupts disabled.
1165 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1167 struct tick_sched *ts =
1168 container_of(timer, struct tick_sched, sched_timer);
1169 struct pt_regs *regs = get_irq_regs();
1170 ktime_t now = ktime_get();
1172 tick_sched_do_timer(now);
1175 * Do not call, when we are not in irq context and have
1176 * no valid regs pointer
1179 tick_sched_handle(ts, regs);
1181 /* No need to reprogram if we are in idle or full dynticks mode */
1182 if (unlikely(ts->tick_stopped))
1183 return HRTIMER_NORESTART;
1185 hrtimer_forward(timer, now, tick_period);
1187 return HRTIMER_RESTART;
1190 static int sched_skew_tick;
1192 static int __init skew_tick(char *str)
1194 get_option(&str, &sched_skew_tick);
1198 early_param("skew_tick", skew_tick);
1201 * tick_setup_sched_timer - setup the tick emulation timer
1203 void tick_setup_sched_timer(void)
1205 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1206 ktime_t now = ktime_get();
1209 * Emulate tick processing via per-CPU hrtimers:
1211 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1212 ts->sched_timer.function = tick_sched_timer;
1214 /* Get the next period (per cpu) */
1215 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1217 /* Offset the tick to avert jiffies_lock contention. */
1218 if (sched_skew_tick) {
1219 u64 offset = ktime_to_ns(tick_period) >> 1;
1220 do_div(offset, num_possible_cpus());
1221 offset *= smp_processor_id();
1222 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1225 hrtimer_forward(&ts->sched_timer, now, tick_period);
1226 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1227 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1229 #endif /* HIGH_RES_TIMERS */
1231 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1232 void tick_cancel_sched_timer(int cpu)
1234 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1236 # ifdef CONFIG_HIGH_RES_TIMERS
1237 if (ts->sched_timer.base)
1238 hrtimer_cancel(&ts->sched_timer);
1241 memset(ts, 0, sizeof(*ts));
1246 * Async notification about clocksource changes
1248 void tick_clock_notify(void)
1252 for_each_possible_cpu(cpu)
1253 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1257 * Async notification about clock event changes
1259 void tick_oneshot_notify(void)
1261 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1263 set_bit(0, &ts->check_clocks);
1267 * Check, if a change happened, which makes oneshot possible.
1269 * Called cyclic from the hrtimer softirq (driven by the timer
1270 * softirq) allow_nohz signals, that we can switch into low-res nohz
1271 * mode, because high resolution timers are disabled (either compile
1272 * or runtime). Called with interrupts disabled.
1274 int tick_check_oneshot_change(int allow_nohz)
1276 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1278 if (!test_and_clear_bit(0, &ts->check_clocks))
1281 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1284 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1290 tick_nohz_switch_to_nohz();