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 void __init tick_nohz_init(void)
412 if (!tick_nohz_full_running)
416 * Full dynticks uses irq work to drive the tick rescheduling on safe
417 * locking contexts. But then we need irq work to raise its own
418 * interrupts to avoid circular dependency on the tick
420 if (!arch_irq_work_has_interrupt()) {
421 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
422 cpumask_clear(tick_nohz_full_mask);
423 tick_nohz_full_running = false;
427 cpu = smp_processor_id();
429 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
430 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
432 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
435 for_each_cpu(cpu, tick_nohz_full_mask)
436 context_tracking_cpu_set(cpu);
438 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
439 "kernel/nohz:predown", NULL,
442 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
443 cpumask_pr_args(tick_nohz_full_mask));
448 * NOHZ - aka dynamic tick functionality
450 #ifdef CONFIG_NO_HZ_COMMON
454 bool tick_nohz_enabled __read_mostly = true;
455 unsigned long tick_nohz_active __read_mostly;
457 * Enable / Disable tickless mode
459 static int __init setup_tick_nohz(char *str)
461 return (kstrtobool(str, &tick_nohz_enabled) == 0);
464 __setup("nohz=", setup_tick_nohz);
466 bool tick_nohz_tick_stopped(void)
468 return __this_cpu_read(tick_cpu_sched.tick_stopped);
471 bool tick_nohz_tick_stopped_cpu(int cpu)
473 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
475 return ts->tick_stopped;
479 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
481 * Called from interrupt entry when the CPU was idle
483 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
484 * must be updated. Otherwise an interrupt handler could use a stale jiffy
485 * value. We do this unconditionally on any CPU, as we don't know whether the
486 * CPU, which has the update task assigned is in a long sleep.
488 static void tick_nohz_update_jiffies(ktime_t now)
492 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
494 local_irq_save(flags);
495 tick_do_update_jiffies64(now);
496 local_irq_restore(flags);
498 touch_softlockup_watchdog_sched();
502 * Updates the per-CPU time idle statistics counters
505 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
509 if (ts->idle_active) {
510 delta = ktime_sub(now, ts->idle_entrytime);
511 if (nr_iowait_cpu(cpu) > 0)
512 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
514 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
515 ts->idle_entrytime = now;
518 if (last_update_time)
519 *last_update_time = ktime_to_us(now);
523 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
525 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
528 sched_clock_idle_wakeup_event();
531 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
533 ktime_t now = ktime_get();
535 ts->idle_entrytime = now;
537 sched_clock_idle_sleep_event();
542 * get_cpu_idle_time_us - get the total idle time of a CPU
543 * @cpu: CPU number to query
544 * @last_update_time: variable to store update time in. Do not update
547 * Return the cumulative idle time (since boot) for a given
548 * CPU, in microseconds.
550 * This time is measured via accounting rather than sampling,
551 * and is as accurate as ktime_get() is.
553 * This function returns -1 if NOHZ is not enabled.
555 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
557 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
560 if (!tick_nohz_active)
564 if (last_update_time) {
565 update_ts_time_stats(cpu, ts, now, last_update_time);
566 idle = ts->idle_sleeptime;
568 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
569 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
571 idle = ktime_add(ts->idle_sleeptime, delta);
573 idle = ts->idle_sleeptime;
577 return ktime_to_us(idle);
580 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
583 * get_cpu_iowait_time_us - get the total iowait time of a CPU
584 * @cpu: CPU number to query
585 * @last_update_time: variable to store update time in. Do not update
588 * Return the cumulative iowait time (since boot) for a given
589 * CPU, in microseconds.
591 * This time is measured via accounting rather than sampling,
592 * and is as accurate as ktime_get() is.
594 * This function returns -1 if NOHZ is not enabled.
596 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
598 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
601 if (!tick_nohz_active)
605 if (last_update_time) {
606 update_ts_time_stats(cpu, ts, now, last_update_time);
607 iowait = ts->iowait_sleeptime;
609 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
610 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
612 iowait = ktime_add(ts->iowait_sleeptime, delta);
614 iowait = ts->iowait_sleeptime;
618 return ktime_to_us(iowait);
620 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
622 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
624 hrtimer_cancel(&ts->sched_timer);
625 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
627 /* Forward the time to expire in the future */
628 hrtimer_forward(&ts->sched_timer, now, tick_period);
630 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
631 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
633 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
636 * Reset to make sure next tick stop doesn't get fooled by past
637 * cached clock deadline.
642 static inline bool local_timer_softirq_pending(void)
644 return local_softirq_pending() & TIMER_SOFTIRQ;
647 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
648 ktime_t now, int cpu)
650 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
651 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
652 unsigned long seq, basejiff;
655 /* Read jiffies and the time when jiffies were updated last */
657 seq = read_seqbegin(&jiffies_lock);
658 basemono = last_jiffies_update;
660 } while (read_seqretry(&jiffies_lock, seq));
661 ts->last_jiffies = basejiff;
664 * Keep the periodic tick, when RCU, architecture or irq_work
666 * Aside of that check whether the local timer softirq is
667 * pending. If so its a bad idea to call get_next_timer_interrupt()
668 * because there is an already expired timer, so it will request
669 * immeditate expiry, which rearms the hardware timer with a
670 * minimal delta which brings us back to this place
671 * immediately. Lather, rinse and repeat...
673 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
674 irq_work_needs_cpu() || local_timer_softirq_pending()) {
675 next_tick = basemono + TICK_NSEC;
678 * Get the next pending timer. If high resolution
679 * timers are enabled this only takes the timer wheel
680 * timers into account. If high resolution timers are
681 * disabled this also looks at the next expiring
684 next_tmr = get_next_timer_interrupt(basejiff, basemono);
685 ts->next_timer = next_tmr;
686 /* Take the next rcu event into account */
687 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
691 * If the tick is due in the next period, keep it ticking or
692 * force prod the timer.
694 delta = next_tick - basemono;
695 if (delta <= (u64)TICK_NSEC) {
697 * Tell the timer code that the base is not idle, i.e. undo
698 * the effect of get_next_timer_interrupt():
702 * We've not stopped the tick yet, and there's a timer in the
703 * next period, so no point in stopping it either, bail.
705 if (!ts->tick_stopped) {
712 * If this CPU is the one which updates jiffies, then give up
713 * the assignment and let it be taken by the CPU which runs
714 * the tick timer next, which might be this CPU as well. If we
715 * don't drop this here the jiffies might be stale and
716 * do_timer() never invoked. Keep track of the fact that it
717 * was the one which had the do_timer() duty last. If this CPU
718 * is the one which had the do_timer() duty last, we limit the
719 * sleep time to the timekeeping max_deferment value.
720 * Otherwise we can sleep as long as we want.
722 delta = timekeeping_max_deferment();
723 if (cpu == tick_do_timer_cpu) {
724 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
725 ts->do_timer_last = 1;
726 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
728 ts->do_timer_last = 0;
729 } else if (!ts->do_timer_last) {
733 /* Calculate the next expiry time */
734 if (delta < (KTIME_MAX - basemono))
735 expires = basemono + delta;
739 expires = min_t(u64, expires, next_tick);
742 /* Skip reprogram of event if its not changed */
743 if (ts->tick_stopped && (expires == ts->next_tick)) {
744 /* Sanity check: make sure clockevent is actually programmed */
745 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
749 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
750 basemono, ts->next_tick, dev->next_event,
751 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
755 * nohz_stop_sched_tick can be called several times before
756 * the nohz_restart_sched_tick is called. This happens when
757 * interrupts arrive which do not cause a reschedule. In the
758 * first call we save the current tick time, so we can restart
759 * the scheduler tick in nohz_restart_sched_tick.
761 if (!ts->tick_stopped) {
762 calc_load_nohz_start();
763 cpu_load_update_nohz_start();
766 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
767 ts->tick_stopped = 1;
768 trace_tick_stop(1, TICK_DEP_MASK_NONE);
771 ts->next_tick = tick;
774 * If the expiration time == KTIME_MAX, then we simply stop
777 if (unlikely(expires == KTIME_MAX)) {
778 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
779 hrtimer_cancel(&ts->sched_timer);
783 hrtimer_set_expires(&ts->sched_timer, tick);
785 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
786 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
788 tick_program_event(tick, 1);
791 * Update the estimated sleep length until the next timer
792 * (not only the tick).
794 ts->sleep_length = ktime_sub(dev->next_event, now);
798 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
800 /* Update jiffies first */
801 tick_do_update_jiffies64(now);
802 cpu_load_update_nohz_stop();
805 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
806 * the clock forward checks in the enqueue path:
810 calc_load_nohz_stop();
811 touch_softlockup_watchdog_sched();
813 * Cancel the scheduled timer and restore the tick
815 ts->tick_stopped = 0;
816 ts->idle_exittime = now;
818 tick_nohz_restart(ts, now);
821 static void tick_nohz_full_update_tick(struct tick_sched *ts)
823 #ifdef CONFIG_NO_HZ_FULL
824 int cpu = smp_processor_id();
826 if (!tick_nohz_full_cpu(cpu))
829 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
832 if (can_stop_full_tick(cpu, ts))
833 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
834 else if (ts->tick_stopped)
835 tick_nohz_restart_sched_tick(ts, ktime_get());
839 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
842 * If this CPU is offline and it is the one which updates
843 * jiffies, then give up the assignment and let it be taken by
844 * the CPU which runs the tick timer next. If we don't drop
845 * this here the jiffies might be stale and do_timer() never
848 if (unlikely(!cpu_online(cpu))) {
849 if (cpu == tick_do_timer_cpu)
850 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
852 * Make sure the CPU doesn't get fooled by obsolete tick
853 * deadline if it comes back online later.
859 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
860 ts->sleep_length = NSEC_PER_SEC / HZ;
867 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
868 static int ratelimit;
870 if (ratelimit < 10 &&
871 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
872 pr_warn("NOHZ: local_softirq_pending %02x\n",
873 (unsigned int) local_softirq_pending());
879 if (tick_nohz_full_enabled()) {
881 * Keep the tick alive to guarantee timekeeping progression
882 * if there are full dynticks CPUs around
884 if (tick_do_timer_cpu == cpu)
887 * Boot safety: make sure the timekeeping duty has been
888 * assigned before entering dyntick-idle mode,
890 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
897 static void __tick_nohz_idle_enter(struct tick_sched *ts)
899 ktime_t now, expires;
900 int cpu = smp_processor_id();
902 now = tick_nohz_start_idle(ts);
904 if (can_stop_idle_tick(cpu, ts)) {
905 int was_stopped = ts->tick_stopped;
909 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
912 ts->idle_expires = expires;
915 if (!was_stopped && ts->tick_stopped) {
916 ts->idle_jiffies = ts->last_jiffies;
917 nohz_balance_enter_idle(cpu);
923 * tick_nohz_idle_enter - stop the idle tick from the idle task
925 * When the next event is more than a tick into the future, stop the idle tick
926 * Called when we start the idle loop.
928 * The arch is responsible of calling:
930 * - rcu_idle_enter() after its last use of RCU before the CPU is put
932 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
934 void tick_nohz_idle_enter(void)
936 struct tick_sched *ts;
938 lockdep_assert_irqs_enabled();
942 ts = this_cpu_ptr(&tick_cpu_sched);
944 __tick_nohz_idle_enter(ts);
950 * tick_nohz_irq_exit - update next tick event from interrupt exit
952 * When an interrupt fires while we are idle and it doesn't cause
953 * a reschedule, it may still add, modify or delete a timer, enqueue
954 * an RCU callback, etc...
955 * So we need to re-calculate and reprogram the next tick event.
957 void tick_nohz_irq_exit(void)
959 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
962 __tick_nohz_idle_enter(ts);
964 tick_nohz_full_update_tick(ts);
968 * tick_nohz_get_sleep_length - return the length of the current sleep
970 * Called from power state control code with interrupts disabled
972 ktime_t tick_nohz_get_sleep_length(void)
974 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
976 return ts->sleep_length;
980 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
981 * for a particular CPU.
983 * Called from the schedutil frequency scaling governor in scheduler context.
985 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
987 struct tick_sched *ts = tick_get_tick_sched(cpu);
989 return ts->idle_calls;
993 * tick_nohz_get_idle_calls - return the current idle calls counter value
995 * Called from the schedutil frequency scaling governor in scheduler context.
997 unsigned long tick_nohz_get_idle_calls(void)
999 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1001 return ts->idle_calls;
1004 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1006 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1007 unsigned long ticks;
1009 if (vtime_accounting_cpu_enabled())
1012 * We stopped the tick in idle. Update process times would miss the
1013 * time we slept as update_process_times does only a 1 tick
1014 * accounting. Enforce that this is accounted to idle !
1016 ticks = jiffies - ts->idle_jiffies;
1018 * We might be one off. Do not randomly account a huge number of ticks!
1020 if (ticks && ticks < LONG_MAX)
1021 account_idle_ticks(ticks);
1026 * tick_nohz_idle_exit - restart the idle tick from the idle task
1028 * Restart the idle tick when the CPU is woken up from idle
1029 * This also exit the RCU extended quiescent state. The CPU
1030 * can use RCU again after this function is called.
1032 void tick_nohz_idle_exit(void)
1034 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1037 local_irq_disable();
1039 WARN_ON_ONCE(!ts->inidle);
1043 if (ts->idle_active || ts->tick_stopped)
1046 if (ts->idle_active)
1047 tick_nohz_stop_idle(ts, now);
1049 if (ts->tick_stopped) {
1050 tick_nohz_restart_sched_tick(ts, now);
1051 tick_nohz_account_idle_ticks(ts);
1058 * The nohz low res interrupt handler
1060 static void tick_nohz_handler(struct clock_event_device *dev)
1062 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1063 struct pt_regs *regs = get_irq_regs();
1064 ktime_t now = ktime_get();
1066 dev->next_event = KTIME_MAX;
1068 tick_sched_do_timer(now);
1069 tick_sched_handle(ts, regs);
1071 /* No need to reprogram if we are running tickless */
1072 if (unlikely(ts->tick_stopped))
1075 hrtimer_forward(&ts->sched_timer, now, tick_period);
1076 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1079 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1081 if (!tick_nohz_enabled)
1083 ts->nohz_mode = mode;
1084 /* One update is enough */
1085 if (!test_and_set_bit(0, &tick_nohz_active))
1086 timers_update_nohz();
1090 * tick_nohz_switch_to_nohz - switch to nohz mode
1092 static void tick_nohz_switch_to_nohz(void)
1094 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1097 if (!tick_nohz_enabled)
1100 if (tick_switch_to_oneshot(tick_nohz_handler))
1104 * Recycle the hrtimer in ts, so we can share the
1105 * hrtimer_forward with the highres code.
1107 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1108 /* Get the next period */
1109 next = tick_init_jiffy_update();
1111 hrtimer_set_expires(&ts->sched_timer, next);
1112 hrtimer_forward_now(&ts->sched_timer, tick_period);
1113 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1114 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1117 static inline void tick_nohz_irq_enter(void)
1119 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1122 if (!ts->idle_active && !ts->tick_stopped)
1125 if (ts->idle_active)
1126 tick_nohz_stop_idle(ts, now);
1127 if (ts->tick_stopped)
1128 tick_nohz_update_jiffies(now);
1133 static inline void tick_nohz_switch_to_nohz(void) { }
1134 static inline void tick_nohz_irq_enter(void) { }
1135 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1137 #endif /* CONFIG_NO_HZ_COMMON */
1140 * Called from irq_enter to notify about the possible interruption of idle()
1142 void tick_irq_enter(void)
1144 tick_check_oneshot_broadcast_this_cpu();
1145 tick_nohz_irq_enter();
1149 * High resolution timer specific code
1151 #ifdef CONFIG_HIGH_RES_TIMERS
1153 * We rearm the timer until we get disabled by the idle code.
1154 * Called with interrupts disabled.
1156 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1158 struct tick_sched *ts =
1159 container_of(timer, struct tick_sched, sched_timer);
1160 struct pt_regs *regs = get_irq_regs();
1161 ktime_t now = ktime_get();
1163 tick_sched_do_timer(now);
1166 * Do not call, when we are not in irq context and have
1167 * no valid regs pointer
1170 tick_sched_handle(ts, regs);
1174 /* No need to reprogram if we are in idle or full dynticks mode */
1175 if (unlikely(ts->tick_stopped))
1176 return HRTIMER_NORESTART;
1178 hrtimer_forward(timer, now, tick_period);
1180 return HRTIMER_RESTART;
1183 static int sched_skew_tick;
1185 static int __init skew_tick(char *str)
1187 get_option(&str, &sched_skew_tick);
1191 early_param("skew_tick", skew_tick);
1194 * tick_setup_sched_timer - setup the tick emulation timer
1196 void tick_setup_sched_timer(void)
1198 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1199 ktime_t now = ktime_get();
1202 * Emulate tick processing via per-CPU hrtimers:
1204 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1205 ts->sched_timer.function = tick_sched_timer;
1207 /* Get the next period (per-CPU) */
1208 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1210 /* Offset the tick to avert jiffies_lock contention. */
1211 if (sched_skew_tick) {
1212 u64 offset = ktime_to_ns(tick_period) >> 1;
1213 do_div(offset, num_possible_cpus());
1214 offset *= smp_processor_id();
1215 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1218 hrtimer_forward(&ts->sched_timer, now, tick_period);
1219 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1220 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1222 #endif /* HIGH_RES_TIMERS */
1224 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1225 void tick_cancel_sched_timer(int cpu)
1227 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1229 # ifdef CONFIG_HIGH_RES_TIMERS
1230 if (ts->sched_timer.base)
1231 hrtimer_cancel(&ts->sched_timer);
1234 memset(ts, 0, sizeof(*ts));
1239 * Async notification about clocksource changes
1241 void tick_clock_notify(void)
1245 for_each_possible_cpu(cpu)
1246 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1250 * Async notification about clock event changes
1252 void tick_oneshot_notify(void)
1254 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1256 set_bit(0, &ts->check_clocks);
1260 * Check, if a change happened, which makes oneshot possible.
1262 * Called cyclic from the hrtimer softirq (driven by the timer
1263 * softirq) allow_nohz signals, that we can switch into low-res nohz
1264 * mode, because high resolution timers are disabled (either compile
1265 * or runtime). Called with interrupts disabled.
1267 int tick_check_oneshot_change(int allow_nohz)
1269 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1271 if (!test_and_clear_bit(0, &ts->check_clocks))
1274 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1277 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1283 tick_nohz_switch_to_nohz();