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/module.h>
25 #include <linux/irq_work.h>
26 #include <linux/posix-timers.h>
27 #include <linux/context_tracking.h>
29 #include <asm/irq_regs.h>
31 #include "tick-internal.h"
33 #include <trace/events/timer.h>
36 * Per-CPU nohz control structure
38 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 struct tick_sched *tick_get_tick_sched(int cpu)
42 return &per_cpu(tick_cpu_sched, cpu);
45 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47 * The time, when the last jiffy update happened. Protected by jiffies_lock.
49 static ktime_t last_jiffies_update;
52 * Must be called with interrupts disabled !
54 static void tick_do_update_jiffies64(ktime_t now)
56 unsigned long ticks = 0;
60 * Do a quick check without holding jiffies_lock:
62 delta = ktime_sub(now, last_jiffies_update);
63 if (delta < tick_period)
66 /* Reevaluate with jiffies_lock held */
67 write_seqlock(&jiffies_lock);
69 delta = ktime_sub(now, last_jiffies_update);
70 if (delta >= tick_period) {
72 delta = ktime_sub(delta, tick_period);
73 last_jiffies_update = ktime_add(last_jiffies_update,
76 /* Slow path for long timeouts */
77 if (unlikely(delta >= tick_period)) {
78 s64 incr = ktime_to_ns(tick_period);
80 ticks = ktime_divns(delta, incr);
82 last_jiffies_update = ktime_add_ns(last_jiffies_update,
87 /* Keep the tick_next_period variable up to date */
88 tick_next_period = ktime_add(last_jiffies_update, tick_period);
90 write_sequnlock(&jiffies_lock);
93 write_sequnlock(&jiffies_lock);
98 * Initialize and return retrieve the jiffies update.
100 static ktime_t tick_init_jiffy_update(void)
104 write_seqlock(&jiffies_lock);
105 /* Did we start the jiffies update yet ? */
106 if (last_jiffies_update == 0)
107 last_jiffies_update = tick_next_period;
108 period = last_jiffies_update;
109 write_sequnlock(&jiffies_lock);
114 static void tick_sched_do_timer(ktime_t now)
116 int cpu = smp_processor_id();
118 #ifdef CONFIG_NO_HZ_COMMON
120 * Check if the do_timer duty was dropped. We don't care about
121 * concurrency: This happens only when the CPU in charge went
122 * into a long sleep. If two CPUs happen to assign themselves to
123 * this duty, then the jiffies update is still serialized by
126 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
127 && !tick_nohz_full_cpu(cpu))
128 tick_do_timer_cpu = cpu;
131 /* Check, if the jiffies need an update */
132 if (tick_do_timer_cpu == cpu)
133 tick_do_update_jiffies64(now);
136 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
138 #ifdef CONFIG_NO_HZ_COMMON
140 * When we are idle and the tick is stopped, we have to touch
141 * the watchdog as we might not schedule for a really long
142 * time. This happens on complete idle SMP systems while
143 * waiting on the login prompt. We also increment the "start of
144 * idle" jiffy stamp so the idle accounting adjustment we do
145 * when we go busy again does not account too much ticks.
147 if (ts->tick_stopped) {
148 touch_softlockup_watchdog_sched();
149 if (is_idle_task(current))
153 update_process_times(user_mode(regs));
154 profile_tick(CPU_PROFILING);
158 #ifdef CONFIG_NO_HZ_FULL
159 cpumask_var_t tick_nohz_full_mask;
160 cpumask_var_t housekeeping_mask;
161 bool tick_nohz_full_running;
162 static atomic_t tick_dep_mask;
164 static bool check_tick_dependency(atomic_t *dep)
166 int val = atomic_read(dep);
168 if (val & TICK_DEP_MASK_POSIX_TIMER) {
169 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
173 if (val & TICK_DEP_MASK_PERF_EVENTS) {
174 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
178 if (val & TICK_DEP_MASK_SCHED) {
179 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
183 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
184 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
191 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
193 WARN_ON_ONCE(!irqs_disabled());
195 if (unlikely(!cpu_online(cpu)))
198 if (check_tick_dependency(&tick_dep_mask))
201 if (check_tick_dependency(&ts->tick_dep_mask))
204 if (check_tick_dependency(¤t->tick_dep_mask))
207 if (check_tick_dependency(¤t->signal->tick_dep_mask))
213 static void nohz_full_kick_func(struct irq_work *work)
215 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
218 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
219 .func = nohz_full_kick_func,
223 * Kick this CPU if it's full dynticks in order to force it to
224 * re-evaluate its dependency on the tick and restart it if necessary.
225 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
228 static void tick_nohz_full_kick(void)
230 if (!tick_nohz_full_cpu(smp_processor_id()))
233 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
237 * Kick the CPU if it's full dynticks in order to force it to
238 * re-evaluate its dependency on the tick and restart it if necessary.
240 void tick_nohz_full_kick_cpu(int cpu)
242 if (!tick_nohz_full_cpu(cpu))
245 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
249 * Kick all full dynticks CPUs in order to force these to re-evaluate
250 * their dependency on the tick and restart it if necessary.
252 static void tick_nohz_full_kick_all(void)
256 if (!tick_nohz_full_running)
260 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
261 tick_nohz_full_kick_cpu(cpu);
265 static void tick_nohz_dep_set_all(atomic_t *dep,
266 enum tick_dep_bits bit)
270 prev = atomic_fetch_or(BIT(bit), dep);
272 tick_nohz_full_kick_all();
276 * Set a global tick dependency. Used by perf events that rely on freq and
279 void tick_nohz_dep_set(enum tick_dep_bits bit)
281 tick_nohz_dep_set_all(&tick_dep_mask, bit);
284 void tick_nohz_dep_clear(enum tick_dep_bits bit)
286 atomic_andnot(BIT(bit), &tick_dep_mask);
290 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
291 * manage events throttling.
293 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
296 struct tick_sched *ts;
298 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
300 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
303 /* Perf needs local kick that is NMI safe */
304 if (cpu == smp_processor_id()) {
305 tick_nohz_full_kick();
307 /* Remote irq work not NMI-safe */
308 if (!WARN_ON_ONCE(in_nmi()))
309 tick_nohz_full_kick_cpu(cpu);
315 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
317 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
319 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
323 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
326 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
329 * We could optimize this with just kicking the target running the task
330 * if that noise matters for nohz full users.
332 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
335 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
337 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
341 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
342 * per process timers.
344 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
346 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
349 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
351 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
355 * Re-evaluate the need for the tick as we switch the current task.
356 * It might need the tick due to per task/process properties:
357 * perf events, posix CPU timers, ...
359 void __tick_nohz_task_switch(void)
362 struct tick_sched *ts;
364 local_irq_save(flags);
366 if (!tick_nohz_full_cpu(smp_processor_id()))
369 ts = this_cpu_ptr(&tick_cpu_sched);
371 if (ts->tick_stopped) {
372 if (atomic_read(¤t->tick_dep_mask) ||
373 atomic_read(¤t->signal->tick_dep_mask))
374 tick_nohz_full_kick();
377 local_irq_restore(flags);
380 /* Parse the boot-time nohz CPU list from the kernel parameters. */
381 static int __init tick_nohz_full_setup(char *str)
383 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
384 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
385 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
386 free_bootmem_cpumask_var(tick_nohz_full_mask);
389 tick_nohz_full_running = true;
393 __setup("nohz_full=", tick_nohz_full_setup);
395 static int tick_nohz_cpu_down(unsigned int cpu)
398 * The boot CPU handles housekeeping duty (unbound timers,
399 * workqueues, timekeeping, ...) on behalf of full dynticks
400 * CPUs. It must remain online when nohz full is enabled.
402 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
407 static int tick_nohz_init_all(void)
411 #ifdef CONFIG_NO_HZ_FULL_ALL
412 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
413 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
417 cpumask_setall(tick_nohz_full_mask);
418 tick_nohz_full_running = true;
423 void __init tick_nohz_init(void)
427 if (!tick_nohz_full_running) {
428 if (tick_nohz_init_all() < 0)
432 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
433 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
434 cpumask_clear(tick_nohz_full_mask);
435 tick_nohz_full_running = false;
440 * Full dynticks uses irq work to drive the tick rescheduling on safe
441 * locking contexts. But then we need irq work to raise its own
442 * interrupts to avoid circular dependency on the tick
444 if (!arch_irq_work_has_interrupt()) {
445 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
446 cpumask_clear(tick_nohz_full_mask);
447 cpumask_copy(housekeeping_mask, cpu_possible_mask);
448 tick_nohz_full_running = false;
452 cpu = smp_processor_id();
454 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
455 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
457 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
460 cpumask_andnot(housekeeping_mask,
461 cpu_possible_mask, tick_nohz_full_mask);
463 for_each_cpu(cpu, tick_nohz_full_mask)
464 context_tracking_cpu_set(cpu);
466 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
467 "kernel/nohz:predown", NULL,
470 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
471 cpumask_pr_args(tick_nohz_full_mask));
474 * We need at least one CPU to handle housekeeping work such
475 * as timekeeping, unbound timers, workqueues, ...
477 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
482 * NOHZ - aka dynamic tick functionality
484 #ifdef CONFIG_NO_HZ_COMMON
488 bool tick_nohz_enabled __read_mostly = true;
489 unsigned long tick_nohz_active __read_mostly;
491 * Enable / Disable tickless mode
493 static int __init setup_tick_nohz(char *str)
495 return (kstrtobool(str, &tick_nohz_enabled) == 0);
498 __setup("nohz=", setup_tick_nohz);
500 int tick_nohz_tick_stopped(void)
502 return __this_cpu_read(tick_cpu_sched.tick_stopped);
506 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
508 * Called from interrupt entry when the CPU was idle
510 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
511 * must be updated. Otherwise an interrupt handler could use a stale jiffy
512 * value. We do this unconditionally on any CPU, as we don't know whether the
513 * CPU, which has the update task assigned is in a long sleep.
515 static void tick_nohz_update_jiffies(ktime_t now)
519 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
521 local_irq_save(flags);
522 tick_do_update_jiffies64(now);
523 local_irq_restore(flags);
525 touch_softlockup_watchdog_sched();
529 * Updates the per-CPU time idle statistics counters
532 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
536 if (ts->idle_active) {
537 delta = ktime_sub(now, ts->idle_entrytime);
538 if (nr_iowait_cpu(cpu) > 0)
539 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
541 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
542 ts->idle_entrytime = now;
545 if (last_update_time)
546 *last_update_time = ktime_to_us(now);
550 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
552 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
555 sched_clock_idle_wakeup_event(0);
558 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
560 ktime_t now = ktime_get();
562 ts->idle_entrytime = now;
564 sched_clock_idle_sleep_event();
569 * get_cpu_idle_time_us - get the total idle time of a CPU
570 * @cpu: CPU number to query
571 * @last_update_time: variable to store update time in. Do not update
574 * Return the cumulative idle time (since boot) for a given
575 * CPU, in microseconds.
577 * This time is measured via accounting rather than sampling,
578 * and is as accurate as ktime_get() is.
580 * This function returns -1 if NOHZ is not enabled.
582 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
584 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
587 if (!tick_nohz_active)
591 if (last_update_time) {
592 update_ts_time_stats(cpu, ts, now, last_update_time);
593 idle = ts->idle_sleeptime;
595 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
596 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
598 idle = ktime_add(ts->idle_sleeptime, delta);
600 idle = ts->idle_sleeptime;
604 return ktime_to_us(idle);
607 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
610 * get_cpu_iowait_time_us - get the total iowait time of a CPU
611 * @cpu: CPU number to query
612 * @last_update_time: variable to store update time in. Do not update
615 * Return the cumulative iowait time (since boot) for a given
616 * CPU, in microseconds.
618 * This time is measured via accounting rather than sampling,
619 * and is as accurate as ktime_get() is.
621 * This function returns -1 if NOHZ is not enabled.
623 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
625 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
628 if (!tick_nohz_active)
632 if (last_update_time) {
633 update_ts_time_stats(cpu, ts, now, last_update_time);
634 iowait = ts->iowait_sleeptime;
636 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
637 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
639 iowait = ktime_add(ts->iowait_sleeptime, delta);
641 iowait = ts->iowait_sleeptime;
645 return ktime_to_us(iowait);
647 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
649 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
651 hrtimer_cancel(&ts->sched_timer);
652 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
654 /* Forward the time to expire in the future */
655 hrtimer_forward(&ts->sched_timer, now, tick_period);
657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
658 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
660 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
663 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
664 ktime_t now, int cpu)
666 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
667 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
668 unsigned long seq, basejiff;
671 /* Read jiffies and the time when jiffies were updated last */
673 seq = read_seqbegin(&jiffies_lock);
674 basemono = last_jiffies_update;
676 } while (read_seqretry(&jiffies_lock, seq));
677 ts->last_jiffies = basejiff;
679 if (rcu_needs_cpu(basemono, &next_rcu) ||
680 arch_needs_cpu() || irq_work_needs_cpu()) {
681 next_tick = basemono + TICK_NSEC;
684 * Get the next pending timer. If high resolution
685 * timers are enabled this only takes the timer wheel
686 * timers into account. If high resolution timers are
687 * disabled this also looks at the next expiring
690 next_tmr = get_next_timer_interrupt(basejiff, basemono);
691 ts->next_timer = next_tmr;
692 /* Take the next rcu event into account */
693 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
697 * If the tick is due in the next period, keep it ticking or
698 * force prod the timer.
700 delta = next_tick - basemono;
701 if (delta <= (u64)TICK_NSEC) {
705 * Tell the timer code that the base is not idle, i.e. undo
706 * the effect of get_next_timer_interrupt():
710 * We've not stopped the tick yet, and there's a timer in the
711 * next period, so no point in stopping it either, bail.
713 if (!ts->tick_stopped)
717 * If, OTOH, we did stop it, but there's a pending (expired)
718 * timer reprogram the timer hardware to fire now.
720 * We will not restart the tick proper, just prod the timer
721 * hardware into firing an interrupt to process the pending
722 * timers. Just like tick_irq_exit() will not restart the tick
723 * for 'normal' interrupts.
725 * Only once we exit the idle loop will we re-enable the tick,
726 * see tick_nohz_idle_exit().
729 tick_nohz_restart(ts, now);
735 * If this CPU is the one which updates jiffies, then give up
736 * the assignment and let it be taken by the CPU which runs
737 * the tick timer next, which might be this CPU as well. If we
738 * don't drop this here the jiffies might be stale and
739 * do_timer() never invoked. Keep track of the fact that it
740 * was the one which had the do_timer() duty last. If this CPU
741 * is the one which had the do_timer() duty last, we limit the
742 * sleep time to the timekeeping max_deferment value.
743 * Otherwise we can sleep as long as we want.
745 delta = timekeeping_max_deferment();
746 if (cpu == tick_do_timer_cpu) {
747 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
748 ts->do_timer_last = 1;
749 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
751 ts->do_timer_last = 0;
752 } else if (!ts->do_timer_last) {
756 #ifdef CONFIG_NO_HZ_FULL
757 /* Limit the tick delta to the maximum scheduler deferment */
759 delta = min(delta, scheduler_tick_max_deferment());
762 /* Calculate the next expiry time */
763 if (delta < (KTIME_MAX - basemono))
764 expires = basemono + delta;
768 expires = min_t(u64, expires, next_tick);
771 /* Skip reprogram of event if its not changed */
772 if (ts->tick_stopped && (expires == dev->next_event))
776 * nohz_stop_sched_tick can be called several times before
777 * the nohz_restart_sched_tick is called. This happens when
778 * interrupts arrive which do not cause a reschedule. In the
779 * first call we save the current tick time, so we can restart
780 * the scheduler tick in nohz_restart_sched_tick.
782 if (!ts->tick_stopped) {
783 nohz_balance_enter_idle(cpu);
784 calc_load_enter_idle();
785 cpu_load_update_nohz_start();
787 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
788 ts->tick_stopped = 1;
789 trace_tick_stop(1, TICK_DEP_MASK_NONE);
793 * If the expiration time == KTIME_MAX, then we simply stop
796 if (unlikely(expires == KTIME_MAX)) {
797 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
798 hrtimer_cancel(&ts->sched_timer);
802 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
803 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
805 tick_program_event(tick, 1);
807 /* Update the estimated sleep length */
808 ts->sleep_length = ktime_sub(dev->next_event, now);
812 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
814 /* Update jiffies first */
815 tick_do_update_jiffies64(now);
816 cpu_load_update_nohz_stop();
819 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
820 * the clock forward checks in the enqueue path:
824 calc_load_exit_idle();
825 touch_softlockup_watchdog_sched();
827 * Cancel the scheduled timer and restore the tick
829 ts->tick_stopped = 0;
830 ts->idle_exittime = now;
832 tick_nohz_restart(ts, now);
835 static void tick_nohz_full_update_tick(struct tick_sched *ts)
837 #ifdef CONFIG_NO_HZ_FULL
838 int cpu = smp_processor_id();
840 if (!tick_nohz_full_cpu(cpu))
843 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
846 if (can_stop_full_tick(cpu, ts))
847 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
848 else if (ts->tick_stopped)
849 tick_nohz_restart_sched_tick(ts, ktime_get());
853 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
856 * If this CPU is offline and it is the one which updates
857 * jiffies, then give up the assignment and let it be taken by
858 * the CPU which runs the tick timer next. If we don't drop
859 * this here the jiffies might be stale and do_timer() never
862 if (unlikely(!cpu_online(cpu))) {
863 if (cpu == tick_do_timer_cpu)
864 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
868 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
869 ts->sleep_length = NSEC_PER_SEC / HZ;
876 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
877 static int ratelimit;
879 if (ratelimit < 10 &&
880 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
881 pr_warn("NOHZ: local_softirq_pending %02x\n",
882 (unsigned int) local_softirq_pending());
888 if (tick_nohz_full_enabled()) {
890 * Keep the tick alive to guarantee timekeeping progression
891 * if there are full dynticks CPUs around
893 if (tick_do_timer_cpu == cpu)
896 * Boot safety: make sure the timekeeping duty has been
897 * assigned before entering dyntick-idle mode,
899 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
906 static void __tick_nohz_idle_enter(struct tick_sched *ts)
908 ktime_t now, expires;
909 int cpu = smp_processor_id();
911 now = tick_nohz_start_idle(ts);
913 if (can_stop_idle_tick(cpu, ts)) {
914 int was_stopped = ts->tick_stopped;
918 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
921 ts->idle_expires = expires;
924 if (!was_stopped && ts->tick_stopped)
925 ts->idle_jiffies = ts->last_jiffies;
930 * tick_nohz_idle_enter - stop the idle tick from the idle task
932 * When the next event is more than a tick into the future, stop the idle tick
933 * Called when we start the idle loop.
935 * The arch is responsible of calling:
937 * - rcu_idle_enter() after its last use of RCU before the CPU is put
939 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
941 void tick_nohz_idle_enter(void)
943 struct tick_sched *ts;
945 WARN_ON_ONCE(irqs_disabled());
948 * Update the idle state in the scheduler domain hierarchy
949 * when tick_nohz_stop_sched_tick() is called from the idle loop.
950 * State will be updated to busy during the first busy tick after
953 set_cpu_sd_state_idle();
957 ts = this_cpu_ptr(&tick_cpu_sched);
959 __tick_nohz_idle_enter(ts);
965 * tick_nohz_irq_exit - update next tick event from interrupt exit
967 * When an interrupt fires while we are idle and it doesn't cause
968 * a reschedule, it may still add, modify or delete a timer, enqueue
969 * an RCU callback, etc...
970 * So we need to re-calculate and reprogram the next tick event.
972 void tick_nohz_irq_exit(void)
974 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
977 __tick_nohz_idle_enter(ts);
979 tick_nohz_full_update_tick(ts);
983 * tick_nohz_get_sleep_length - return the length of the current sleep
985 * Called from power state control code with interrupts disabled
987 ktime_t tick_nohz_get_sleep_length(void)
989 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
991 return ts->sleep_length;
994 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
996 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
999 if (vtime_accounting_cpu_enabled())
1002 * We stopped the tick in idle. Update process times would miss the
1003 * time we slept as update_process_times does only a 1 tick
1004 * accounting. Enforce that this is accounted to idle !
1006 ticks = jiffies - ts->idle_jiffies;
1008 * We might be one off. Do not randomly account a huge number of ticks!
1010 if (ticks && ticks < LONG_MAX)
1011 account_idle_ticks(ticks);
1016 * tick_nohz_idle_exit - restart the idle tick from the idle task
1018 * Restart the idle tick when the CPU is woken up from idle
1019 * This also exit the RCU extended quiescent state. The CPU
1020 * can use RCU again after this function is called.
1022 void tick_nohz_idle_exit(void)
1024 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1027 local_irq_disable();
1029 WARN_ON_ONCE(!ts->inidle);
1033 if (ts->idle_active || ts->tick_stopped)
1036 if (ts->idle_active)
1037 tick_nohz_stop_idle(ts, now);
1039 if (ts->tick_stopped) {
1040 tick_nohz_restart_sched_tick(ts, now);
1041 tick_nohz_account_idle_ticks(ts);
1048 * The nohz low res interrupt handler
1050 static void tick_nohz_handler(struct clock_event_device *dev)
1052 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1053 struct pt_regs *regs = get_irq_regs();
1054 ktime_t now = ktime_get();
1056 dev->next_event = KTIME_MAX;
1058 tick_sched_do_timer(now);
1059 tick_sched_handle(ts, regs);
1061 /* No need to reprogram if we are running tickless */
1062 if (unlikely(ts->tick_stopped))
1065 hrtimer_forward(&ts->sched_timer, now, tick_period);
1066 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1069 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1071 if (!tick_nohz_enabled)
1073 ts->nohz_mode = mode;
1074 /* One update is enough */
1075 if (!test_and_set_bit(0, &tick_nohz_active))
1076 timers_update_migration(true);
1080 * tick_nohz_switch_to_nohz - switch to nohz mode
1082 static void tick_nohz_switch_to_nohz(void)
1084 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1087 if (!tick_nohz_enabled)
1090 if (tick_switch_to_oneshot(tick_nohz_handler))
1094 * Recycle the hrtimer in ts, so we can share the
1095 * hrtimer_forward with the highres code.
1097 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1098 /* Get the next period */
1099 next = tick_init_jiffy_update();
1101 hrtimer_set_expires(&ts->sched_timer, next);
1102 hrtimer_forward_now(&ts->sched_timer, tick_period);
1103 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1104 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1107 static inline void tick_nohz_irq_enter(void)
1109 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1112 if (!ts->idle_active && !ts->tick_stopped)
1115 if (ts->idle_active)
1116 tick_nohz_stop_idle(ts, now);
1117 if (ts->tick_stopped)
1118 tick_nohz_update_jiffies(now);
1123 static inline void tick_nohz_switch_to_nohz(void) { }
1124 static inline void tick_nohz_irq_enter(void) { }
1125 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1127 #endif /* CONFIG_NO_HZ_COMMON */
1130 * Called from irq_enter to notify about the possible interruption of idle()
1132 void tick_irq_enter(void)
1134 tick_check_oneshot_broadcast_this_cpu();
1135 tick_nohz_irq_enter();
1139 * High resolution timer specific code
1141 #ifdef CONFIG_HIGH_RES_TIMERS
1143 * We rearm the timer until we get disabled by the idle code.
1144 * Called with interrupts disabled.
1146 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1148 struct tick_sched *ts =
1149 container_of(timer, struct tick_sched, sched_timer);
1150 struct pt_regs *regs = get_irq_regs();
1151 ktime_t now = ktime_get();
1153 tick_sched_do_timer(now);
1156 * Do not call, when we are not in irq context and have
1157 * no valid regs pointer
1160 tick_sched_handle(ts, regs);
1162 /* No need to reprogram if we are in idle or full dynticks mode */
1163 if (unlikely(ts->tick_stopped))
1164 return HRTIMER_NORESTART;
1166 hrtimer_forward(timer, now, tick_period);
1168 return HRTIMER_RESTART;
1171 static int sched_skew_tick;
1173 static int __init skew_tick(char *str)
1175 get_option(&str, &sched_skew_tick);
1179 early_param("skew_tick", skew_tick);
1182 * tick_setup_sched_timer - setup the tick emulation timer
1184 void tick_setup_sched_timer(void)
1186 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1187 ktime_t now = ktime_get();
1190 * Emulate tick processing via per-CPU hrtimers:
1192 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1193 ts->sched_timer.function = tick_sched_timer;
1195 /* Get the next period (per-CPU) */
1196 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1198 /* Offset the tick to avert jiffies_lock contention. */
1199 if (sched_skew_tick) {
1200 u64 offset = ktime_to_ns(tick_period) >> 1;
1201 do_div(offset, num_possible_cpus());
1202 offset *= smp_processor_id();
1203 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1206 hrtimer_forward(&ts->sched_timer, now, tick_period);
1207 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1208 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1210 #endif /* HIGH_RES_TIMERS */
1212 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1213 void tick_cancel_sched_timer(int cpu)
1215 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1217 # ifdef CONFIG_HIGH_RES_TIMERS
1218 if (ts->sched_timer.base)
1219 hrtimer_cancel(&ts->sched_timer);
1222 memset(ts, 0, sizeof(*ts));
1227 * Async notification about clocksource changes
1229 void tick_clock_notify(void)
1233 for_each_possible_cpu(cpu)
1234 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1238 * Async notification about clock event changes
1240 void tick_oneshot_notify(void)
1242 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1244 set_bit(0, &ts->check_clocks);
1248 * Check, if a change happened, which makes oneshot possible.
1250 * Called cyclic from the hrtimer softirq (driven by the timer
1251 * softirq) allow_nohz signals, that we can switch into low-res nohz
1252 * mode, because high resolution timers are disabled (either compile
1253 * or runtime). Called with interrupts disabled.
1255 int tick_check_oneshot_change(int allow_nohz)
1257 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1259 if (!test_and_clear_bit(0, &ts->check_clocks))
1262 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1265 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1271 tick_nohz_switch_to_nohz();