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 tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
213 * Re-evaluate the need for the tick on the current CPU
214 * and restart it if necessary.
216 void __tick_nohz_full_check(void)
218 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
220 if (tick_nohz_full_cpu(smp_processor_id())) {
221 if (ts->tick_stopped && !is_idle_task(current)) {
222 if (!can_stop_full_tick())
223 tick_nohz_restart_sched_tick(ts, ktime_get());
228 static void nohz_full_kick_work_func(struct irq_work *work)
230 __tick_nohz_full_check();
233 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
234 .func = nohz_full_kick_work_func,
235 .flags = IRQ_WORK_HARD_IRQ,
239 * Kick this CPU if it's full dynticks in order to force it to
240 * re-evaluate its dependency on the tick and restart it if necessary.
241 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
244 void tick_nohz_full_kick(void)
246 if (!tick_nohz_full_cpu(smp_processor_id()))
249 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
253 * Kick the CPU if it's full dynticks in order to force it to
254 * re-evaluate its dependency on the tick and restart it if necessary.
256 void tick_nohz_full_kick_cpu(int cpu)
258 if (!tick_nohz_full_cpu(cpu))
261 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
264 static void nohz_full_kick_ipi(void *info)
266 __tick_nohz_full_check();
270 * Kick all full dynticks CPUs in order to force these to re-evaluate
271 * their dependency on the tick and restart it if necessary.
273 void tick_nohz_full_kick_all(void)
275 if (!tick_nohz_full_running)
279 smp_call_function_many(tick_nohz_full_mask,
280 nohz_full_kick_ipi, NULL, false);
281 tick_nohz_full_kick();
286 * Re-evaluate the need for the tick as we switch the current task.
287 * It might need the tick due to per task/process properties:
288 * perf events, posix cpu timers, ...
290 void __tick_nohz_task_switch(struct task_struct *tsk)
294 local_irq_save(flags);
296 if (!tick_nohz_full_cpu(smp_processor_id()))
299 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
300 tick_nohz_full_kick();
303 local_irq_restore(flags);
306 /* Parse the boot-time nohz CPU list from the kernel parameters. */
307 static int __init tick_nohz_full_setup(char *str)
309 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
310 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
311 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
312 free_bootmem_cpumask_var(tick_nohz_full_mask);
315 tick_nohz_full_running = true;
319 __setup("nohz_full=", tick_nohz_full_setup);
321 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
322 unsigned long action,
325 unsigned int cpu = (unsigned long)hcpu;
327 switch (action & ~CPU_TASKS_FROZEN) {
328 case CPU_DOWN_PREPARE:
330 * If we handle the timekeeping duty for full dynticks CPUs,
331 * we can't safely shutdown that CPU.
333 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
340 static int tick_nohz_init_all(void)
344 #ifdef CONFIG_NO_HZ_FULL_ALL
345 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
346 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
350 cpumask_setall(tick_nohz_full_mask);
351 tick_nohz_full_running = true;
356 void __init tick_nohz_init(void)
360 if (!tick_nohz_full_running) {
361 if (tick_nohz_init_all() < 0)
365 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
366 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
367 cpumask_clear(tick_nohz_full_mask);
368 tick_nohz_full_running = false;
373 * Full dynticks uses irq work to drive the tick rescheduling on safe
374 * locking contexts. But then we need irq work to raise its own
375 * interrupts to avoid circular dependency on the tick
377 if (!arch_irq_work_has_interrupt()) {
378 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't "
379 "support irq work self-IPIs\n");
380 cpumask_clear(tick_nohz_full_mask);
381 cpumask_copy(housekeeping_mask, cpu_possible_mask);
382 tick_nohz_full_running = false;
386 cpu = smp_processor_id();
388 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
389 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
390 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
393 cpumask_andnot(housekeeping_mask,
394 cpu_possible_mask, tick_nohz_full_mask);
396 for_each_cpu(cpu, tick_nohz_full_mask)
397 context_tracking_cpu_set(cpu);
399 cpu_notifier(tick_nohz_cpu_down_callback, 0);
400 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
401 cpumask_pr_args(tick_nohz_full_mask));
406 * NOHZ - aka dynamic tick functionality
408 #ifdef CONFIG_NO_HZ_COMMON
412 static int tick_nohz_enabled __read_mostly = 1;
413 int tick_nohz_active __read_mostly;
415 * Enable / Disable tickless mode
417 static int __init setup_tick_nohz(char *str)
419 if (!strcmp(str, "off"))
420 tick_nohz_enabled = 0;
421 else if (!strcmp(str, "on"))
422 tick_nohz_enabled = 1;
428 __setup("nohz=", setup_tick_nohz);
430 int tick_nohz_tick_stopped(void)
432 return __this_cpu_read(tick_cpu_sched.tick_stopped);
436 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
438 * Called from interrupt entry when the CPU was idle
440 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
441 * must be updated. Otherwise an interrupt handler could use a stale jiffy
442 * value. We do this unconditionally on any cpu, as we don't know whether the
443 * cpu, which has the update task assigned is in a long sleep.
445 static void tick_nohz_update_jiffies(ktime_t now)
449 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
451 local_irq_save(flags);
452 tick_do_update_jiffies64(now);
453 local_irq_restore(flags);
455 touch_softlockup_watchdog();
459 * Updates the per cpu time idle statistics counters
462 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
466 if (ts->idle_active) {
467 delta = ktime_sub(now, ts->idle_entrytime);
468 if (nr_iowait_cpu(cpu) > 0)
469 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
471 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
472 ts->idle_entrytime = now;
475 if (last_update_time)
476 *last_update_time = ktime_to_us(now);
480 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
482 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
485 sched_clock_idle_wakeup_event(0);
488 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
490 ktime_t now = ktime_get();
492 ts->idle_entrytime = now;
494 sched_clock_idle_sleep_event();
499 * get_cpu_idle_time_us - get the total idle time of a cpu
500 * @cpu: CPU number to query
501 * @last_update_time: variable to store update time in. Do not update
504 * Return the cummulative idle time (since boot) for a given
505 * CPU, in microseconds.
507 * This time is measured via accounting rather than sampling,
508 * and is as accurate as ktime_get() is.
510 * This function returns -1 if NOHZ is not enabled.
512 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
514 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
517 if (!tick_nohz_active)
521 if (last_update_time) {
522 update_ts_time_stats(cpu, ts, now, last_update_time);
523 idle = ts->idle_sleeptime;
525 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
526 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
528 idle = ktime_add(ts->idle_sleeptime, delta);
530 idle = ts->idle_sleeptime;
534 return ktime_to_us(idle);
537 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
540 * get_cpu_iowait_time_us - get the total iowait time of a cpu
541 * @cpu: CPU number to query
542 * @last_update_time: variable to store update time in. Do not update
545 * Return the cummulative iowait time (since boot) for a given
546 * CPU, in microseconds.
548 * This time is measured via accounting rather than sampling,
549 * and is as accurate as ktime_get() is.
551 * This function returns -1 if NOHZ is not enabled.
553 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
555 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
558 if (!tick_nohz_active)
562 if (last_update_time) {
563 update_ts_time_stats(cpu, ts, now, last_update_time);
564 iowait = ts->iowait_sleeptime;
566 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
567 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
569 iowait = ktime_add(ts->iowait_sleeptime, delta);
571 iowait = ts->iowait_sleeptime;
575 return ktime_to_us(iowait);
577 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
579 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
581 hrtimer_cancel(&ts->sched_timer);
582 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
584 /* Forward the time to expire in the future */
585 hrtimer_forward(&ts->sched_timer, now, tick_period);
587 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
588 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
590 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
593 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
594 ktime_t now, int cpu)
596 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
597 ktime_t last_update, expires, ret = { .tv64 = 0 };
598 unsigned long rcu_delta_jiffies;
599 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
602 time_delta = timekeeping_max_deferment();
604 /* Read jiffies and the time when jiffies were updated last */
606 seq = read_seqcount_begin(&jiffies_seq);
607 last_update = last_jiffies_update;
608 last_jiffies = jiffies;
609 } while (read_seqcount_retry(&jiffies_seq, seq));
611 if (rcu_needs_cpu(&rcu_delta_jiffies) ||
612 arch_needs_cpu() || irq_work_needs_cpu()) {
613 next_jiffies = last_jiffies + 1;
616 /* Get the next timer wheel timer */
617 next_jiffies = get_next_timer_interrupt(last_jiffies);
618 delta_jiffies = next_jiffies - last_jiffies;
619 if (rcu_delta_jiffies < delta_jiffies) {
620 next_jiffies = last_jiffies + rcu_delta_jiffies;
621 delta_jiffies = rcu_delta_jiffies;
626 * Do not stop the tick, if we are only one off (or less)
627 * or if the cpu is required for RCU:
629 if (!ts->tick_stopped && delta_jiffies <= 1)
632 /* Schedule the tick, if we are at least one jiffie off */
633 if ((long)delta_jiffies >= 1) {
636 * If this cpu is the one which updates jiffies, then
637 * give up the assignment and let it be taken by the
638 * cpu which runs the tick timer next, which might be
639 * this cpu as well. If we don't drop this here the
640 * jiffies might be stale and do_timer() never
641 * invoked. Keep track of the fact that it was the one
642 * which had the do_timer() duty last. If this cpu is
643 * the one which had the do_timer() duty last, we
644 * limit the sleep time to the timekeeping
645 * max_deferement value which we retrieved
646 * above. Otherwise we can sleep as long as we want.
648 if (cpu == tick_do_timer_cpu) {
649 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
650 ts->do_timer_last = 1;
651 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
652 time_delta = KTIME_MAX;
653 ts->do_timer_last = 0;
654 } else if (!ts->do_timer_last) {
655 time_delta = KTIME_MAX;
658 #ifdef CONFIG_NO_HZ_FULL
660 time_delta = min(time_delta,
661 scheduler_tick_max_deferment());
666 * calculate the expiry time for the next timer wheel
667 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
668 * that there is no timer pending or at least extremely
669 * far into the future (12 days for HZ=1000). In this
670 * case we set the expiry to the end of time.
672 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
674 * Calculate the time delta for the next timer event.
675 * If the time delta exceeds the maximum time delta
676 * permitted by the current clocksource then adjust
677 * the time delta accordingly to ensure the
678 * clocksource does not wrap.
680 time_delta = min_t(u64, time_delta,
681 tick_period.tv64 * delta_jiffies);
684 if (time_delta < KTIME_MAX)
685 expires = ktime_add_ns(last_update, time_delta);
687 expires.tv64 = KTIME_MAX;
689 /* Skip reprogram of event if its not changed */
690 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
696 * nohz_stop_sched_tick can be called several times before
697 * the nohz_restart_sched_tick is called. This happens when
698 * interrupts arrive which do not cause a reschedule. In the
699 * first call we save the current tick time, so we can restart
700 * the scheduler tick in nohz_restart_sched_tick.
702 if (!ts->tick_stopped) {
703 nohz_balance_enter_idle(cpu);
704 calc_load_enter_idle();
706 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
707 ts->tick_stopped = 1;
708 trace_tick_stop(1, " ");
712 * If the expiration time == KTIME_MAX, then
713 * in this case we simply stop the tick timer.
715 if (unlikely(expires.tv64 == KTIME_MAX)) {
716 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
717 hrtimer_cancel(&ts->sched_timer);
721 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
722 hrtimer_start(&ts->sched_timer, expires,
723 HRTIMER_MODE_ABS_PINNED);
725 tick_program_event(expires, 1);
727 /* Tick is stopped, but required now. Enforce it */
728 tick_nohz_restart(ts, now);
732 ts->next_jiffies = next_jiffies;
733 ts->last_jiffies = last_jiffies;
734 ts->sleep_length = ktime_sub(dev->next_event, now);
739 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
741 #ifdef CONFIG_NO_HZ_FULL
742 int cpu = smp_processor_id();
744 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
747 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
750 if (!can_stop_full_tick())
753 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
757 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
760 * If this cpu is offline and it is the one which updates
761 * jiffies, then give up the assignment and let it be taken by
762 * the cpu which runs the tick timer next. If we don't drop
763 * this here the jiffies might be stale and do_timer() never
766 if (unlikely(!cpu_online(cpu))) {
767 if (cpu == tick_do_timer_cpu)
768 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
772 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
773 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
780 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
781 softirq_check_pending_idle();
785 if (tick_nohz_full_enabled()) {
787 * Keep the tick alive to guarantee timekeeping progression
788 * if there are full dynticks CPUs around
790 if (tick_do_timer_cpu == cpu)
793 * Boot safety: make sure the timekeeping duty has been
794 * assigned before entering dyntick-idle mode,
796 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
803 static void __tick_nohz_idle_enter(struct tick_sched *ts)
805 ktime_t now, expires;
806 int cpu = smp_processor_id();
808 now = tick_nohz_start_idle(ts);
810 if (can_stop_idle_tick(cpu, ts)) {
811 int was_stopped = ts->tick_stopped;
815 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
816 if (expires.tv64 > 0LL) {
818 ts->idle_expires = expires;
821 if (!was_stopped && ts->tick_stopped)
822 ts->idle_jiffies = ts->last_jiffies;
827 * tick_nohz_idle_enter - stop the idle tick from the idle task
829 * When the next event is more than a tick into the future, stop the idle tick
830 * Called when we start the idle loop.
832 * The arch is responsible of calling:
834 * - rcu_idle_enter() after its last use of RCU before the CPU is put
836 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
838 void tick_nohz_idle_enter(void)
840 struct tick_sched *ts;
842 WARN_ON_ONCE(irqs_disabled());
845 * Update the idle state in the scheduler domain hierarchy
846 * when tick_nohz_stop_sched_tick() is called from the idle loop.
847 * State will be updated to busy during the first busy tick after
850 set_cpu_sd_state_idle();
854 ts = this_cpu_ptr(&tick_cpu_sched);
856 __tick_nohz_idle_enter(ts);
862 * tick_nohz_irq_exit - update next tick event from interrupt exit
864 * When an interrupt fires while we are idle and it doesn't cause
865 * a reschedule, it may still add, modify or delete a timer, enqueue
866 * an RCU callback, etc...
867 * So we need to re-calculate and reprogram the next tick event.
869 void tick_nohz_irq_exit(void)
871 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
874 __tick_nohz_idle_enter(ts);
876 tick_nohz_full_stop_tick(ts);
880 * tick_nohz_get_sleep_length - return the length of the current sleep
882 * Called from power state control code with interrupts disabled
884 ktime_t tick_nohz_get_sleep_length(void)
886 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
888 return ts->sleep_length;
891 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
893 /* Update jiffies first */
894 tick_do_update_jiffies64(now);
895 update_cpu_load_nohz();
897 calc_load_exit_idle();
898 touch_softlockup_watchdog();
900 * Cancel the scheduled timer and restore the tick
902 ts->tick_stopped = 0;
903 ts->idle_exittime = now;
905 tick_nohz_restart(ts, now);
908 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
910 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
913 if (vtime_accounting_enabled())
916 * We stopped the tick in idle. Update process times would miss the
917 * time we slept as update_process_times does only a 1 tick
918 * accounting. Enforce that this is accounted to idle !
920 ticks = jiffies - ts->idle_jiffies;
922 * We might be one off. Do not randomly account a huge number of ticks!
924 if (ticks && ticks < LONG_MAX)
925 account_idle_ticks(ticks);
930 * tick_nohz_idle_exit - restart the idle tick from the idle task
932 * Restart the idle tick when the CPU is woken up from idle
933 * This also exit the RCU extended quiescent state. The CPU
934 * can use RCU again after this function is called.
936 void tick_nohz_idle_exit(void)
938 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
943 WARN_ON_ONCE(!ts->inidle);
947 if (ts->idle_active || ts->tick_stopped)
951 tick_nohz_stop_idle(ts, now);
953 if (ts->tick_stopped) {
954 tick_nohz_restart_sched_tick(ts, now);
955 tick_nohz_account_idle_ticks(ts);
961 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
963 hrtimer_forward(&ts->sched_timer, now, tick_period);
964 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
968 * The nohz low res interrupt handler
970 static void tick_nohz_handler(struct clock_event_device *dev)
972 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
973 struct pt_regs *regs = get_irq_regs();
974 ktime_t now = ktime_get();
976 dev->next_event.tv64 = KTIME_MAX;
978 tick_sched_do_timer(now);
979 tick_sched_handle(ts, regs);
981 /* No need to reprogram if we are running tickless */
982 if (unlikely(ts->tick_stopped))
985 while (tick_nohz_reprogram(ts, now)) {
987 tick_do_update_jiffies64(now);
992 * tick_nohz_switch_to_nohz - switch to nohz mode
994 static void tick_nohz_switch_to_nohz(void)
996 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
999 if (!tick_nohz_enabled)
1002 local_irq_disable();
1003 if (tick_switch_to_oneshot(tick_nohz_handler)) {
1007 tick_nohz_active = 1;
1008 ts->nohz_mode = NOHZ_MODE_LOWRES;
1011 * Recycle the hrtimer in ts, so we can share the
1012 * hrtimer_forward with the highres code.
1014 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1015 /* Get the next period */
1016 next = tick_init_jiffy_update();
1019 hrtimer_set_expires(&ts->sched_timer, next);
1020 if (!tick_program_event(next, 0))
1022 next = ktime_add(next, tick_period);
1028 * When NOHZ is enabled and the tick is stopped, we need to kick the
1029 * tick timer from irq_enter() so that the jiffies update is kept
1030 * alive during long running softirqs. That's ugly as hell, but
1031 * correctness is key even if we need to fix the offending softirq in
1034 * Note, this is different to tick_nohz_restart. We just kick the
1035 * timer and do not touch the other magic bits which need to be done
1036 * when idle is left.
1038 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1041 /* Switch back to 2.6.27 behaviour */
1045 * Do not touch the tick device, when the next expiry is either
1046 * already reached or less/equal than the tick period.
1048 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1049 if (delta.tv64 <= tick_period.tv64)
1052 tick_nohz_restart(ts, now);
1056 static inline void tick_nohz_irq_enter(void)
1058 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1061 if (!ts->idle_active && !ts->tick_stopped)
1064 if (ts->idle_active)
1065 tick_nohz_stop_idle(ts, now);
1066 if (ts->tick_stopped) {
1067 tick_nohz_update_jiffies(now);
1068 tick_nohz_kick_tick(ts, now);
1074 static inline void tick_nohz_switch_to_nohz(void) { }
1075 static inline void tick_nohz_irq_enter(void) { }
1077 #endif /* CONFIG_NO_HZ_COMMON */
1080 * Called from irq_enter to notify about the possible interruption of idle()
1082 void tick_irq_enter(void)
1084 tick_check_oneshot_broadcast_this_cpu();
1085 tick_nohz_irq_enter();
1089 * High resolution timer specific code
1091 #ifdef CONFIG_HIGH_RES_TIMERS
1093 * We rearm the timer until we get disabled by the idle code.
1094 * Called with interrupts disabled.
1096 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1098 struct tick_sched *ts =
1099 container_of(timer, struct tick_sched, sched_timer);
1100 struct pt_regs *regs = get_irq_regs();
1101 ktime_t now = ktime_get();
1103 tick_sched_do_timer(now);
1106 * Do not call, when we are not in irq context and have
1107 * no valid regs pointer
1110 tick_sched_handle(ts, regs);
1112 /* No need to reprogram if we are in idle or full dynticks mode */
1113 if (unlikely(ts->tick_stopped))
1114 return HRTIMER_NORESTART;
1116 hrtimer_forward(timer, now, tick_period);
1118 return HRTIMER_RESTART;
1121 static int sched_skew_tick;
1123 static int __init skew_tick(char *str)
1125 get_option(&str, &sched_skew_tick);
1129 early_param("skew_tick", skew_tick);
1132 * tick_setup_sched_timer - setup the tick emulation timer
1134 void tick_setup_sched_timer(void)
1136 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1137 ktime_t now = ktime_get();
1140 * Emulate tick processing via per-CPU hrtimers:
1142 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1143 ts->sched_timer.irqsafe = 1;
1144 ts->sched_timer.function = tick_sched_timer;
1146 /* Get the next period (per cpu) */
1147 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1149 /* Offset the tick to avert jiffies_lock contention. */
1150 if (sched_skew_tick) {
1151 u64 offset = ktime_to_ns(tick_period) >> 1;
1152 do_div(offset, num_possible_cpus());
1153 offset *= smp_processor_id();
1154 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1158 hrtimer_forward(&ts->sched_timer, now, tick_period);
1159 hrtimer_start_expires(&ts->sched_timer,
1160 HRTIMER_MODE_ABS_PINNED);
1161 /* Check, if the timer was already in the past */
1162 if (hrtimer_active(&ts->sched_timer))
1167 #ifdef CONFIG_NO_HZ_COMMON
1168 if (tick_nohz_enabled) {
1169 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1170 tick_nohz_active = 1;
1174 #endif /* HIGH_RES_TIMERS */
1176 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1177 void tick_cancel_sched_timer(int cpu)
1179 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1181 # ifdef CONFIG_HIGH_RES_TIMERS
1182 if (ts->sched_timer.base)
1183 hrtimer_cancel(&ts->sched_timer);
1186 memset(ts, 0, sizeof(*ts));
1191 * Async notification about clocksource changes
1193 void tick_clock_notify(void)
1197 for_each_possible_cpu(cpu)
1198 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1202 * Async notification about clock event changes
1204 void tick_oneshot_notify(void)
1206 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1208 set_bit(0, &ts->check_clocks);
1212 * Check, if a change happened, which makes oneshot possible.
1214 * Called cyclic from the hrtimer softirq (driven by the timer
1215 * softirq) allow_nohz signals, that we can switch into low-res nohz
1216 * mode, because high resolution timers are disabled (either compile
1219 int tick_check_oneshot_change(int allow_nohz)
1221 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 if (!test_and_clear_bit(0, &ts->check_clocks))
1226 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1229 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1235 tick_nohz_switch_to_nohz();