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/perf_event.h>
26 #include <linux/context_tracking.h>
28 #include <asm/irq_regs.h>
30 #include "tick-internal.h"
32 #include <trace/events/timer.h>
35 * Per cpu nohz control structure
37 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 * The time, when the last jiffy update happened. Protected by jiffies_lock.
42 static ktime_t last_jiffies_update;
44 struct tick_sched *tick_get_tick_sched(int cpu)
46 return &per_cpu(tick_cpu_sched, cpu);
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 raw_spin_lock(&jiffies_lock);
66 write_seqcount_begin(&jiffies_seq);
68 delta = ktime_sub(now, last_jiffies_update);
69 if (delta.tv64 >= tick_period.tv64) {
71 delta = ktime_sub(delta, tick_period);
72 last_jiffies_update = ktime_add(last_jiffies_update,
75 /* Slow path for long timeouts */
76 if (unlikely(delta.tv64 >= tick_period.tv64)) {
77 s64 incr = ktime_to_ns(tick_period);
79 ticks = ktime_divns(delta, incr);
81 last_jiffies_update = ktime_add_ns(last_jiffies_update,
86 /* Keep the tick_next_period variable up to date */
87 tick_next_period = ktime_add(last_jiffies_update, tick_period);
89 write_seqcount_end(&jiffies_seq);
90 raw_spin_unlock(&jiffies_lock);
93 write_seqcount_end(&jiffies_seq);
94 raw_spin_unlock(&jiffies_lock);
99 * Initialize and return retrieve the jiffies update.
101 static ktime_t tick_init_jiffy_update(void)
105 raw_spin_lock(&jiffies_lock);
106 write_seqcount_begin(&jiffies_seq);
107 /* Did we start the jiffies update yet ? */
108 if (last_jiffies_update.tv64 == 0)
109 last_jiffies_update = tick_next_period;
110 period = last_jiffies_update;
111 write_seqcount_end(&jiffies_seq);
112 raw_spin_unlock(&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 themself 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();
152 if (is_idle_task(current))
156 update_process_times(user_mode(regs));
157 profile_tick(CPU_PROFILING);
160 #ifdef CONFIG_NO_HZ_FULL
161 cpumask_var_t tick_nohz_full_mask;
162 cpumask_var_t housekeeping_mask;
163 bool tick_nohz_full_running;
165 static bool can_stop_full_tick(void)
167 WARN_ON_ONCE(!irqs_disabled());
169 if (!sched_can_stop_tick()) {
170 trace_tick_stop(0, "more than 1 task in runqueue\n");
174 if (!posix_cpu_timers_can_stop_tick(current)) {
175 trace_tick_stop(0, "posix timers running\n");
179 if (!perf_event_can_stop_tick()) {
180 trace_tick_stop(0, "perf events running\n");
184 if (!arch_irq_work_has_interrupt()) {
185 trace_tick_stop(0, "missing irq work interrupt\n");
189 /* sched_clock_tick() needs us? */
190 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
192 * TODO: kick full dynticks CPUs when
193 * sched_clock_stable is set.
195 if (!sched_clock_stable()) {
196 trace_tick_stop(0, "unstable sched clock\n");
198 * Don't allow the user to think they can get
199 * full NO_HZ with this machine.
201 WARN_ONCE(tick_nohz_full_running,
202 "NO_HZ FULL will not work with unstable sched clock");
210 static void nohz_full_kick_work_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_work_func,
217 .flags = IRQ_WORK_HARD_IRQ,
221 * Kick this CPU if it's full dynticks in order to force it to
222 * re-evaluate its dependency on the tick and restart it if necessary.
223 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
226 void tick_nohz_full_kick(void)
228 if (!tick_nohz_full_cpu(smp_processor_id()))
231 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
235 * Kick the CPU if it's full dynticks in order to force it to
236 * re-evaluate its dependency on the tick and restart it if necessary.
238 void tick_nohz_full_kick_cpu(int cpu)
240 if (!tick_nohz_full_cpu(cpu))
243 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
246 static void nohz_full_kick_ipi(void *info)
248 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
252 * Kick all full dynticks CPUs in order to force these to re-evaluate
253 * their dependency on the tick and restart it if necessary.
255 void tick_nohz_full_kick_all(void)
257 if (!tick_nohz_full_running)
261 smp_call_function_many(tick_nohz_full_mask,
262 nohz_full_kick_ipi, NULL, false);
263 tick_nohz_full_kick();
268 * Re-evaluate the need for the tick as we switch the current task.
269 * It might need the tick due to per task/process properties:
270 * perf events, posix cpu timers, ...
272 void __tick_nohz_task_switch(void)
276 local_irq_save(flags);
278 if (!tick_nohz_full_cpu(smp_processor_id()))
281 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
282 tick_nohz_full_kick();
285 local_irq_restore(flags);
288 /* Parse the boot-time nohz CPU list from the kernel parameters. */
289 static int __init tick_nohz_full_setup(char *str)
291 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
292 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
293 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
294 free_bootmem_cpumask_var(tick_nohz_full_mask);
297 tick_nohz_full_running = true;
301 __setup("nohz_full=", tick_nohz_full_setup);
303 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
304 unsigned long action,
307 unsigned int cpu = (unsigned long)hcpu;
309 switch (action & ~CPU_TASKS_FROZEN) {
310 case CPU_DOWN_PREPARE:
312 * The boot CPU handles housekeeping duty (unbound timers,
313 * workqueues, timekeeping, ...) on behalf of full dynticks
314 * CPUs. It must remain online when nohz full is enabled.
316 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
323 static int tick_nohz_init_all(void)
327 #ifdef CONFIG_NO_HZ_FULL_ALL
328 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
329 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
333 cpumask_setall(tick_nohz_full_mask);
334 tick_nohz_full_running = true;
339 void __init tick_nohz_init(void)
343 if (!tick_nohz_full_running) {
344 if (tick_nohz_init_all() < 0)
348 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
349 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
350 cpumask_clear(tick_nohz_full_mask);
351 tick_nohz_full_running = false;
356 * Full dynticks uses irq work to drive the tick rescheduling on safe
357 * locking contexts. But then we need irq work to raise its own
358 * interrupts to avoid circular dependency on the tick
360 if (!arch_irq_work_has_interrupt()) {
361 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't "
362 "support irq work self-IPIs\n");
363 cpumask_clear(tick_nohz_full_mask);
364 cpumask_copy(housekeeping_mask, cpu_possible_mask);
365 tick_nohz_full_running = false;
369 cpu = smp_processor_id();
371 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
372 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
373 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
376 cpumask_andnot(housekeeping_mask,
377 cpu_possible_mask, tick_nohz_full_mask);
379 for_each_cpu(cpu, tick_nohz_full_mask)
380 context_tracking_cpu_set(cpu);
382 cpu_notifier(tick_nohz_cpu_down_callback, 0);
383 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
384 cpumask_pr_args(tick_nohz_full_mask));
387 * We need at least one CPU to handle housekeeping work such
388 * as timekeeping, unbound timers, workqueues, ...
390 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
395 * NOHZ - aka dynamic tick functionality
397 #ifdef CONFIG_NO_HZ_COMMON
401 static int tick_nohz_enabled __read_mostly = 1;
402 unsigned long tick_nohz_active __read_mostly;
404 * Enable / Disable tickless mode
406 static int __init setup_tick_nohz(char *str)
408 if (!strcmp(str, "off"))
409 tick_nohz_enabled = 0;
410 else if (!strcmp(str, "on"))
411 tick_nohz_enabled = 1;
417 __setup("nohz=", setup_tick_nohz);
419 int tick_nohz_tick_stopped(void)
421 return __this_cpu_read(tick_cpu_sched.tick_stopped);
425 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
427 * Called from interrupt entry when the CPU was idle
429 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
430 * must be updated. Otherwise an interrupt handler could use a stale jiffy
431 * value. We do this unconditionally on any cpu, as we don't know whether the
432 * cpu, which has the update task assigned is in a long sleep.
434 static void tick_nohz_update_jiffies(ktime_t now)
438 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
440 local_irq_save(flags);
441 tick_do_update_jiffies64(now);
442 local_irq_restore(flags);
444 touch_softlockup_watchdog();
448 * Updates the per cpu time idle statistics counters
451 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
455 if (ts->idle_active) {
456 delta = ktime_sub(now, ts->idle_entrytime);
457 if (nr_iowait_cpu(cpu) > 0)
458 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
460 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
461 ts->idle_entrytime = now;
464 if (last_update_time)
465 *last_update_time = ktime_to_us(now);
469 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
471 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
474 sched_clock_idle_wakeup_event(0);
477 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
479 ktime_t now = ktime_get();
481 ts->idle_entrytime = now;
483 sched_clock_idle_sleep_event();
488 * get_cpu_idle_time_us - get the total idle time of a cpu
489 * @cpu: CPU number to query
490 * @last_update_time: variable to store update time in. Do not update
493 * Return the cummulative idle time (since boot) for a given
494 * CPU, in microseconds.
496 * This time is measured via accounting rather than sampling,
497 * and is as accurate as ktime_get() is.
499 * This function returns -1 if NOHZ is not enabled.
501 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
503 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
506 if (!tick_nohz_active)
510 if (last_update_time) {
511 update_ts_time_stats(cpu, ts, now, last_update_time);
512 idle = ts->idle_sleeptime;
514 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
515 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
517 idle = ktime_add(ts->idle_sleeptime, delta);
519 idle = ts->idle_sleeptime;
523 return ktime_to_us(idle);
526 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
529 * get_cpu_iowait_time_us - get the total iowait time of a cpu
530 * @cpu: CPU number to query
531 * @last_update_time: variable to store update time in. Do not update
534 * Return the cummulative iowait time (since boot) for a given
535 * CPU, in microseconds.
537 * This time is measured via accounting rather than sampling,
538 * and is as accurate as ktime_get() is.
540 * This function returns -1 if NOHZ is not enabled.
542 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
544 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
547 if (!tick_nohz_active)
551 if (last_update_time) {
552 update_ts_time_stats(cpu, ts, now, last_update_time);
553 iowait = ts->iowait_sleeptime;
555 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
556 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
558 iowait = ktime_add(ts->iowait_sleeptime, delta);
560 iowait = ts->iowait_sleeptime;
564 return ktime_to_us(iowait);
566 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
568 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
570 hrtimer_cancel(&ts->sched_timer);
571 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
573 /* Forward the time to expire in the future */
574 hrtimer_forward(&ts->sched_timer, now, tick_period);
576 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
577 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
579 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
582 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
583 ktime_t now, int cpu)
585 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
586 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
587 unsigned long seq, basejiff;
590 /* Read jiffies and the time when jiffies were updated last */
592 seq = read_seqcount_begin(&jiffies_seq);
593 basemono = last_jiffies_update.tv64;
595 } while (read_seqcount_retry(&jiffies_seq, seq));
596 ts->last_jiffies = basejiff;
598 if (rcu_needs_cpu(basemono, &next_rcu) ||
599 arch_needs_cpu() || irq_work_needs_cpu()) {
600 next_tick = basemono + TICK_NSEC;
603 * Get the next pending timer. If high resolution
604 * timers are enabled this only takes the timer wheel
605 * timers into account. If high resolution timers are
606 * disabled this also looks at the next expiring
609 next_tmr = get_next_timer_interrupt(basejiff, basemono);
610 ts->next_timer = next_tmr;
611 /* Take the next rcu event into account */
612 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
616 * If the tick is due in the next period, keep it ticking or
619 delta = next_tick - basemono;
620 if (delta <= (u64)TICK_NSEC) {
622 if (!ts->tick_stopped)
625 /* Tick is stopped, but required now. Enforce it */
626 tick_nohz_restart(ts, now);
632 * If this cpu is the one which updates jiffies, then give up
633 * the assignment and let it be taken by the cpu which runs
634 * the tick timer next, which might be this cpu as well. If we
635 * don't drop this here the jiffies might be stale and
636 * do_timer() never invoked. Keep track of the fact that it
637 * was the one which had the do_timer() duty last. If this cpu
638 * is the one which had the do_timer() duty last, we limit the
639 * sleep time to the timekeeping max_deferement value.
640 * Otherwise we can sleep as long as we want.
642 delta = timekeeping_max_deferment();
643 if (cpu == tick_do_timer_cpu) {
644 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
645 ts->do_timer_last = 1;
646 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
648 ts->do_timer_last = 0;
649 } else if (!ts->do_timer_last) {
653 #ifdef CONFIG_NO_HZ_FULL
654 /* Limit the tick delta to the maximum scheduler deferment */
656 delta = min(delta, scheduler_tick_max_deferment());
659 /* Calculate the next expiry time */
660 if (delta < (KTIME_MAX - basemono))
661 expires = basemono + delta;
665 expires = min_t(u64, expires, next_tick);
668 /* Skip reprogram of event if its not changed */
669 if (ts->tick_stopped && (expires == dev->next_event.tv64))
673 * nohz_stop_sched_tick can be called several times before
674 * the nohz_restart_sched_tick is called. This happens when
675 * interrupts arrive which do not cause a reschedule. In the
676 * first call we save the current tick time, so we can restart
677 * the scheduler tick in nohz_restart_sched_tick.
679 if (!ts->tick_stopped) {
680 nohz_balance_enter_idle(cpu);
681 calc_load_enter_idle();
683 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
684 ts->tick_stopped = 1;
685 trace_tick_stop(1, " ");
689 * If the expiration time == KTIME_MAX, then we simply stop
692 if (unlikely(expires == KTIME_MAX)) {
693 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
694 hrtimer_cancel(&ts->sched_timer);
698 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
699 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
701 tick_program_event(tick, 1);
703 /* Update the estimated sleep length */
704 ts->sleep_length = ktime_sub(dev->next_event, now);
708 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
710 /* Update jiffies first */
711 tick_do_update_jiffies64(now);
712 update_cpu_load_nohz();
714 calc_load_exit_idle();
715 touch_softlockup_watchdog();
717 * Cancel the scheduled timer and restore the tick
719 ts->tick_stopped = 0;
720 ts->idle_exittime = now;
722 tick_nohz_restart(ts, now);
725 static void tick_nohz_full_update_tick(struct tick_sched *ts)
727 #ifdef CONFIG_NO_HZ_FULL
728 int cpu = smp_processor_id();
730 if (!tick_nohz_full_cpu(cpu))
733 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
736 if (can_stop_full_tick())
737 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
738 else if (ts->tick_stopped)
739 tick_nohz_restart_sched_tick(ts, ktime_get());
743 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
746 * If this cpu is offline and it is the one which updates
747 * jiffies, then give up the assignment and let it be taken by
748 * the cpu which runs the tick timer next. If we don't drop
749 * this here the jiffies might be stale and do_timer() never
752 if (unlikely(!cpu_online(cpu))) {
753 if (cpu == tick_do_timer_cpu)
754 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
758 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
759 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
766 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
767 softirq_check_pending_idle();
771 if (tick_nohz_full_enabled()) {
773 * Keep the tick alive to guarantee timekeeping progression
774 * if there are full dynticks CPUs around
776 if (tick_do_timer_cpu == cpu)
779 * Boot safety: make sure the timekeeping duty has been
780 * assigned before entering dyntick-idle mode,
782 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
789 static void __tick_nohz_idle_enter(struct tick_sched *ts)
791 ktime_t now, expires;
792 int cpu = smp_processor_id();
794 now = tick_nohz_start_idle(ts);
796 if (can_stop_idle_tick(cpu, ts)) {
797 int was_stopped = ts->tick_stopped;
801 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
802 if (expires.tv64 > 0LL) {
804 ts->idle_expires = expires;
807 if (!was_stopped && ts->tick_stopped)
808 ts->idle_jiffies = ts->last_jiffies;
813 * tick_nohz_idle_enter - stop the idle tick from the idle task
815 * When the next event is more than a tick into the future, stop the idle tick
816 * Called when we start the idle loop.
818 * The arch is responsible of calling:
820 * - rcu_idle_enter() after its last use of RCU before the CPU is put
822 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
824 void tick_nohz_idle_enter(void)
826 struct tick_sched *ts;
828 WARN_ON_ONCE(irqs_disabled());
831 * Update the idle state in the scheduler domain hierarchy
832 * when tick_nohz_stop_sched_tick() is called from the idle loop.
833 * State will be updated to busy during the first busy tick after
836 set_cpu_sd_state_idle();
840 ts = this_cpu_ptr(&tick_cpu_sched);
842 __tick_nohz_idle_enter(ts);
848 * tick_nohz_irq_exit - update next tick event from interrupt exit
850 * When an interrupt fires while we are idle and it doesn't cause
851 * a reschedule, it may still add, modify or delete a timer, enqueue
852 * an RCU callback, etc...
853 * So we need to re-calculate and reprogram the next tick event.
855 void tick_nohz_irq_exit(void)
857 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
860 __tick_nohz_idle_enter(ts);
862 tick_nohz_full_update_tick(ts);
866 * tick_nohz_get_sleep_length - return the length of the current sleep
868 * Called from power state control code with interrupts disabled
870 ktime_t tick_nohz_get_sleep_length(void)
872 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
874 return ts->sleep_length;
877 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
879 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
882 if (vtime_accounting_enabled())
885 * We stopped the tick in idle. Update process times would miss the
886 * time we slept as update_process_times does only a 1 tick
887 * accounting. Enforce that this is accounted to idle !
889 ticks = jiffies - ts->idle_jiffies;
891 * We might be one off. Do not randomly account a huge number of ticks!
893 if (ticks && ticks < LONG_MAX)
894 account_idle_ticks(ticks);
899 * tick_nohz_idle_exit - restart the idle tick from the idle task
901 * Restart the idle tick when the CPU is woken up from idle
902 * This also exit the RCU extended quiescent state. The CPU
903 * can use RCU again after this function is called.
905 void tick_nohz_idle_exit(void)
907 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
912 WARN_ON_ONCE(!ts->inidle);
916 if (ts->idle_active || ts->tick_stopped)
920 tick_nohz_stop_idle(ts, now);
922 if (ts->tick_stopped) {
923 tick_nohz_restart_sched_tick(ts, now);
924 tick_nohz_account_idle_ticks(ts);
931 * The nohz low res interrupt handler
933 static void tick_nohz_handler(struct clock_event_device *dev)
935 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
936 struct pt_regs *regs = get_irq_regs();
937 ktime_t now = ktime_get();
939 dev->next_event.tv64 = KTIME_MAX;
941 tick_sched_do_timer(now);
942 tick_sched_handle(ts, regs);
944 /* No need to reprogram if we are running tickless */
945 if (unlikely(ts->tick_stopped))
948 hrtimer_forward(&ts->sched_timer, now, tick_period);
949 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
952 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
954 if (!tick_nohz_enabled)
956 ts->nohz_mode = mode;
957 /* One update is enough */
958 if (!test_and_set_bit(0, &tick_nohz_active))
959 timers_update_migration(true);
963 * tick_nohz_switch_to_nohz - switch to nohz mode
965 static void tick_nohz_switch_to_nohz(void)
967 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
970 if (!tick_nohz_enabled)
973 if (tick_switch_to_oneshot(tick_nohz_handler))
977 * Recycle the hrtimer in ts, so we can share the
978 * hrtimer_forward with the highres code.
980 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
981 /* Get the next period */
982 next = tick_init_jiffy_update();
984 hrtimer_set_expires(&ts->sched_timer, next);
985 hrtimer_forward_now(&ts->sched_timer, tick_period);
986 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
987 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
991 * When NOHZ is enabled and the tick is stopped, we need to kick the
992 * tick timer from irq_enter() so that the jiffies update is kept
993 * alive during long running softirqs. That's ugly as hell, but
994 * correctness is key even if we need to fix the offending softirq in
997 * Note, this is different to tick_nohz_restart. We just kick the
998 * timer and do not touch the other magic bits which need to be done
1001 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1004 /* Switch back to 2.6.27 behaviour */
1008 * Do not touch the tick device, when the next expiry is either
1009 * already reached or less/equal than the tick period.
1011 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1012 if (delta.tv64 <= tick_period.tv64)
1015 tick_nohz_restart(ts, now);
1019 static inline void tick_nohz_irq_enter(void)
1021 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1024 if (!ts->idle_active && !ts->tick_stopped)
1027 if (ts->idle_active)
1028 tick_nohz_stop_idle(ts, now);
1029 if (ts->tick_stopped) {
1030 tick_nohz_update_jiffies(now);
1031 tick_nohz_kick_tick(ts, now);
1037 static inline void tick_nohz_switch_to_nohz(void) { }
1038 static inline void tick_nohz_irq_enter(void) { }
1039 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1041 #endif /* CONFIG_NO_HZ_COMMON */
1044 * Called from irq_enter to notify about the possible interruption of idle()
1046 void tick_irq_enter(void)
1048 tick_check_oneshot_broadcast_this_cpu();
1049 tick_nohz_irq_enter();
1053 * High resolution timer specific code
1055 #ifdef CONFIG_HIGH_RES_TIMERS
1057 * We rearm the timer until we get disabled by the idle code.
1058 * Called with interrupts disabled.
1060 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1062 struct tick_sched *ts =
1063 container_of(timer, struct tick_sched, sched_timer);
1064 struct pt_regs *regs = get_irq_regs();
1065 ktime_t now = ktime_get();
1067 tick_sched_do_timer(now);
1070 * Do not call, when we are not in irq context and have
1071 * no valid regs pointer
1074 tick_sched_handle(ts, regs);
1076 /* No need to reprogram if we are in idle or full dynticks mode */
1077 if (unlikely(ts->tick_stopped))
1078 return HRTIMER_NORESTART;
1080 hrtimer_forward(timer, now, tick_period);
1082 return HRTIMER_RESTART;
1085 static int sched_skew_tick;
1087 static int __init skew_tick(char *str)
1089 get_option(&str, &sched_skew_tick);
1093 early_param("skew_tick", skew_tick);
1096 * tick_setup_sched_timer - setup the tick emulation timer
1098 void tick_setup_sched_timer(void)
1100 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1101 ktime_t now = ktime_get();
1104 * Emulate tick processing via per-CPU hrtimers:
1106 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1107 ts->sched_timer.irqsafe = 1;
1108 ts->sched_timer.function = tick_sched_timer;
1110 /* Get the next period (per cpu) */
1111 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1113 /* Offset the tick to avert jiffies_lock contention. */
1114 if (sched_skew_tick) {
1115 u64 offset = ktime_to_ns(tick_period) >> 1;
1116 do_div(offset, num_possible_cpus());
1117 offset *= smp_processor_id();
1118 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1121 hrtimer_forward(&ts->sched_timer, now, tick_period);
1122 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1123 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1125 #endif /* HIGH_RES_TIMERS */
1127 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1128 void tick_cancel_sched_timer(int cpu)
1130 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1132 # ifdef CONFIG_HIGH_RES_TIMERS
1133 if (ts->sched_timer.base)
1134 hrtimer_cancel(&ts->sched_timer);
1137 memset(ts, 0, sizeof(*ts));
1142 * Async notification about clocksource changes
1144 void tick_clock_notify(void)
1148 for_each_possible_cpu(cpu)
1149 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1153 * Async notification about clock event changes
1155 void tick_oneshot_notify(void)
1157 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1159 set_bit(0, &ts->check_clocks);
1163 * Check, if a change happened, which makes oneshot possible.
1165 * Called cyclic from the hrtimer softirq (driven by the timer
1166 * softirq) allow_nohz signals, that we can switch into low-res nohz
1167 * mode, because high resolution timers are disabled (either compile
1168 * or runtime). Called with interrupts disabled.
1170 int tick_check_oneshot_change(int allow_nohz)
1172 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1174 if (!test_and_clear_bit(0, &ts->check_clocks))
1177 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1180 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1186 tick_nohz_switch_to_nohz();