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>
27 #include <asm/irq_regs.h>
29 #include "tick-internal.h"
31 #include <trace/events/timer.h>
34 * Per cpu nohz control structure
36 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
39 * The time, when the last jiffy update happened. Protected by jiffies_lock.
41 static ktime_t last_jiffies_update;
43 struct tick_sched *tick_get_tick_sched(int cpu)
45 return &per_cpu(tick_cpu_sched, cpu);
49 * Must be called with interrupts disabled !
51 static void tick_do_update_jiffies64(ktime_t now)
53 unsigned long ticks = 0;
57 * Do a quick check without holding jiffies_lock:
59 delta = ktime_sub(now, last_jiffies_update);
60 if (delta.tv64 < tick_period.tv64)
63 /* Reevalute with jiffies_lock held */
64 write_seqlock(&jiffies_lock);
66 delta = ktime_sub(now, last_jiffies_update);
67 if (delta.tv64 >= tick_period.tv64) {
69 delta = ktime_sub(delta, tick_period);
70 last_jiffies_update = ktime_add(last_jiffies_update,
73 /* Slow path for long timeouts */
74 if (unlikely(delta.tv64 >= tick_period.tv64)) {
75 s64 incr = ktime_to_ns(tick_period);
77 ticks = ktime_divns(delta, incr);
79 last_jiffies_update = ktime_add_ns(last_jiffies_update,
84 /* Keep the tick_next_period variable up to date */
85 tick_next_period = ktime_add(last_jiffies_update, tick_period);
87 write_sequnlock(&jiffies_lock);
91 * Initialize and return retrieve the jiffies update.
93 static ktime_t tick_init_jiffy_update(void)
97 write_seqlock(&jiffies_lock);
98 /* Did we start the jiffies update yet ? */
99 if (last_jiffies_update.tv64 == 0)
100 last_jiffies_update = tick_next_period;
101 period = last_jiffies_update;
102 write_sequnlock(&jiffies_lock);
107 static void tick_sched_do_timer(ktime_t now)
109 int cpu = smp_processor_id();
111 #ifdef CONFIG_NO_HZ_COMMON
113 * Check if the do_timer duty was dropped. We don't care about
114 * concurrency: This happens only when the cpu in charge went
115 * into a long sleep. If two cpus happen to assign themself to
116 * this duty, then the jiffies update is still serialized by
119 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
120 && !tick_nohz_full_cpu(cpu))
121 tick_do_timer_cpu = cpu;
124 /* Check, if the jiffies need an update */
125 if (tick_do_timer_cpu == cpu)
126 tick_do_update_jiffies64(now);
129 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
131 #ifdef CONFIG_NO_HZ_COMMON
133 * When we are idle and the tick is stopped, we have to touch
134 * the watchdog as we might not schedule for a really long
135 * time. This happens on complete idle SMP systems while
136 * waiting on the login prompt. We also increment the "start of
137 * idle" jiffy stamp so the idle accounting adjustment we do
138 * when we go busy again does not account too much ticks.
140 if (ts->tick_stopped) {
141 touch_softlockup_watchdog();
142 if (is_idle_task(current))
146 update_process_times(user_mode(regs));
147 profile_tick(CPU_PROFILING);
150 #ifdef CONFIG_NO_HZ_FULL
151 static cpumask_var_t nohz_full_mask;
152 bool have_nohz_full_mask;
154 static bool can_stop_full_tick(void)
156 WARN_ON_ONCE(!irqs_disabled());
158 if (!sched_can_stop_tick()) {
159 trace_tick_stop(0, "more than 1 task in runqueue\n");
163 if (!posix_cpu_timers_can_stop_tick(current)) {
164 trace_tick_stop(0, "posix timers running\n");
168 if (!perf_event_can_stop_tick()) {
169 trace_tick_stop(0, "perf events running\n");
173 /* sched_clock_tick() needs us? */
174 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
176 * TODO: kick full dynticks CPUs when
177 * sched_clock_stable is set.
179 if (!sched_clock_stable) {
180 trace_tick_stop(0, "unstable sched clock\n");
182 * Don't allow the user to think they can get
183 * full NO_HZ with this machine.
185 WARN_ONCE(have_nohz_full_mask,
186 "NO_HZ FULL will not work with unstable sched clock");
194 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
197 * Re-evaluate the need for the tick on the current CPU
198 * and restart it if necessary.
200 void tick_nohz_full_check(void)
202 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
204 if (tick_nohz_full_cpu(smp_processor_id())) {
205 if (ts->tick_stopped && !is_idle_task(current)) {
206 if (!can_stop_full_tick())
207 tick_nohz_restart_sched_tick(ts, ktime_get());
212 static void nohz_full_kick_work_func(struct irq_work *work)
214 tick_nohz_full_check();
217 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
218 .func = nohz_full_kick_work_func,
222 * Kick the current CPU if it's full dynticks in order to force it to
223 * re-evaluate its dependency on the tick and restart it if necessary.
225 void tick_nohz_full_kick(void)
227 if (tick_nohz_full_cpu(smp_processor_id()))
228 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
231 static void nohz_full_kick_ipi(void *info)
233 tick_nohz_full_check();
237 * Kick all full dynticks CPUs in order to force these to re-evaluate
238 * their dependency on the tick and restart it if necessary.
240 void tick_nohz_full_kick_all(void)
242 if (!have_nohz_full_mask)
246 smp_call_function_many(nohz_full_mask,
247 nohz_full_kick_ipi, NULL, false);
252 * Re-evaluate the need for the tick as we switch the current task.
253 * It might need the tick due to per task/process properties:
254 * perf events, posix cpu timers, ...
256 void tick_nohz_task_switch(struct task_struct *tsk)
260 local_irq_save(flags);
262 if (!tick_nohz_full_cpu(smp_processor_id()))
265 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
266 tick_nohz_full_kick();
269 local_irq_restore(flags);
272 int tick_nohz_full_cpu(int cpu)
274 if (!have_nohz_full_mask)
277 return cpumask_test_cpu(cpu, nohz_full_mask);
280 /* Parse the boot-time nohz CPU list from the kernel parameters. */
281 static int __init tick_nohz_full_setup(char *str)
285 alloc_bootmem_cpumask_var(&nohz_full_mask);
286 if (cpulist_parse(str, nohz_full_mask) < 0) {
287 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
291 cpu = smp_processor_id();
292 if (cpumask_test_cpu(cpu, nohz_full_mask)) {
293 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
294 cpumask_clear_cpu(cpu, nohz_full_mask);
296 have_nohz_full_mask = true;
300 __setup("nohz_full=", tick_nohz_full_setup);
302 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
303 unsigned long action,
306 unsigned int cpu = (unsigned long)hcpu;
308 switch (action & ~CPU_TASKS_FROZEN) {
309 case CPU_DOWN_PREPARE:
311 * If we handle the timekeeping duty for full dynticks CPUs,
312 * we can't safely shutdown that CPU.
314 if (have_nohz_full_mask && tick_do_timer_cpu == cpu)
322 * Worst case string length in chunks of CPU range seems 2 steps
323 * separations: 0,2,4,6,...
324 * This is NR_CPUS + sizeof('\0')
326 static char __initdata nohz_full_buf[NR_CPUS + 1];
328 static int tick_nohz_init_all(void)
332 #ifdef CONFIG_NO_HZ_FULL_ALL
333 if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) {
334 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
338 cpumask_setall(nohz_full_mask);
339 cpumask_clear_cpu(smp_processor_id(), nohz_full_mask);
340 have_nohz_full_mask = true;
345 void __init tick_nohz_init(void)
347 if (!have_nohz_full_mask) {
348 if (tick_nohz_init_all() < 0)
352 cpu_notifier(tick_nohz_cpu_down_callback, 0);
353 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask);
354 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
357 #define have_nohz_full_mask (0)
361 * NOHZ - aka dynamic tick functionality
363 #ifdef CONFIG_NO_HZ_COMMON
367 int tick_nohz_enabled __read_mostly = 1;
370 * Enable / Disable tickless mode
372 static int __init setup_tick_nohz(char *str)
374 if (!strcmp(str, "off"))
375 tick_nohz_enabled = 0;
376 else if (!strcmp(str, "on"))
377 tick_nohz_enabled = 1;
383 __setup("nohz=", setup_tick_nohz);
386 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
388 * Called from interrupt entry when the CPU was idle
390 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
391 * must be updated. Otherwise an interrupt handler could use a stale jiffy
392 * value. We do this unconditionally on any cpu, as we don't know whether the
393 * cpu, which has the update task assigned is in a long sleep.
395 static void tick_nohz_update_jiffies(ktime_t now)
397 int cpu = smp_processor_id();
398 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
401 ts->idle_waketime = now;
403 local_irq_save(flags);
404 tick_do_update_jiffies64(now);
405 local_irq_restore(flags);
407 touch_softlockup_watchdog();
411 * Updates the per cpu time idle statistics counters
414 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
418 if (ts->idle_active) {
419 delta = ktime_sub(now, ts->idle_entrytime);
420 if (nr_iowait_cpu(cpu) > 0)
421 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
423 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
424 ts->idle_entrytime = now;
427 if (last_update_time)
428 *last_update_time = ktime_to_us(now);
432 static void tick_nohz_stop_idle(int cpu, ktime_t now)
434 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
436 update_ts_time_stats(cpu, ts, now, NULL);
439 sched_clock_idle_wakeup_event(0);
442 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
444 ktime_t now = ktime_get();
446 ts->idle_entrytime = now;
448 sched_clock_idle_sleep_event();
453 * get_cpu_idle_time_us - get the total idle time of a cpu
454 * @cpu: CPU number to query
455 * @last_update_time: variable to store update time in. Do not update
458 * Return the cummulative idle time (since boot) for a given
459 * CPU, in microseconds.
461 * This time is measured via accounting rather than sampling,
462 * and is as accurate as ktime_get() is.
464 * This function returns -1 if NOHZ is not enabled.
466 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
468 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
471 if (!tick_nohz_enabled)
475 if (last_update_time) {
476 update_ts_time_stats(cpu, ts, now, last_update_time);
477 idle = ts->idle_sleeptime;
479 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
480 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
482 idle = ktime_add(ts->idle_sleeptime, delta);
484 idle = ts->idle_sleeptime;
488 return ktime_to_us(idle);
491 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
494 * get_cpu_iowait_time_us - get the total iowait time of a cpu
495 * @cpu: CPU number to query
496 * @last_update_time: variable to store update time in. Do not update
499 * Return the cummulative iowait time (since boot) for a given
500 * CPU, in microseconds.
502 * This time is measured via accounting rather than sampling,
503 * and is as accurate as ktime_get() is.
505 * This function returns -1 if NOHZ is not enabled.
507 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
509 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
512 if (!tick_nohz_enabled)
516 if (last_update_time) {
517 update_ts_time_stats(cpu, ts, now, last_update_time);
518 iowait = ts->iowait_sleeptime;
520 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
521 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
523 iowait = ktime_add(ts->iowait_sleeptime, delta);
525 iowait = ts->iowait_sleeptime;
529 return ktime_to_us(iowait);
531 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
533 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
534 ktime_t now, int cpu)
536 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
537 ktime_t last_update, expires, ret = { .tv64 = 0 };
538 unsigned long rcu_delta_jiffies;
539 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
542 /* Read jiffies and the time when jiffies were updated last */
544 seq = read_seqbegin(&jiffies_lock);
545 last_update = last_jiffies_update;
546 last_jiffies = jiffies;
547 time_delta = timekeeping_max_deferment();
548 } while (read_seqretry(&jiffies_lock, seq));
550 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
551 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
552 next_jiffies = last_jiffies + 1;
555 /* Get the next timer wheel timer */
556 next_jiffies = get_next_timer_interrupt(last_jiffies);
557 delta_jiffies = next_jiffies - last_jiffies;
558 if (rcu_delta_jiffies < delta_jiffies) {
559 next_jiffies = last_jiffies + rcu_delta_jiffies;
560 delta_jiffies = rcu_delta_jiffies;
565 * Do not stop the tick, if we are only one off (or less)
566 * or if the cpu is required for RCU:
568 if (!ts->tick_stopped && delta_jiffies <= 1)
571 /* Schedule the tick, if we are at least one jiffie off */
572 if ((long)delta_jiffies >= 1) {
575 * If this cpu is the one which updates jiffies, then
576 * give up the assignment and let it be taken by the
577 * cpu which runs the tick timer next, which might be
578 * this cpu as well. If we don't drop this here the
579 * jiffies might be stale and do_timer() never
580 * invoked. Keep track of the fact that it was the one
581 * which had the do_timer() duty last. If this cpu is
582 * the one which had the do_timer() duty last, we
583 * limit the sleep time to the timekeeping
584 * max_deferement value which we retrieved
585 * above. Otherwise we can sleep as long as we want.
587 if (cpu == tick_do_timer_cpu) {
588 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
589 ts->do_timer_last = 1;
590 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
591 time_delta = KTIME_MAX;
592 ts->do_timer_last = 0;
593 } else if (!ts->do_timer_last) {
594 time_delta = KTIME_MAX;
597 #ifdef CONFIG_NO_HZ_FULL
599 time_delta = min(time_delta,
600 scheduler_tick_max_deferment());
605 * calculate the expiry time for the next timer wheel
606 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
607 * that there is no timer pending or at least extremely
608 * far into the future (12 days for HZ=1000). In this
609 * case we set the expiry to the end of time.
611 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
613 * Calculate the time delta for the next timer event.
614 * If the time delta exceeds the maximum time delta
615 * permitted by the current clocksource then adjust
616 * the time delta accordingly to ensure the
617 * clocksource does not wrap.
619 time_delta = min_t(u64, time_delta,
620 tick_period.tv64 * delta_jiffies);
623 if (time_delta < KTIME_MAX)
624 expires = ktime_add_ns(last_update, time_delta);
626 expires.tv64 = KTIME_MAX;
628 /* Skip reprogram of event if its not changed */
629 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
635 * nohz_stop_sched_tick can be called several times before
636 * the nohz_restart_sched_tick is called. This happens when
637 * interrupts arrive which do not cause a reschedule. In the
638 * first call we save the current tick time, so we can restart
639 * the scheduler tick in nohz_restart_sched_tick.
641 if (!ts->tick_stopped) {
642 nohz_balance_enter_idle(cpu);
643 calc_load_enter_idle();
645 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
646 ts->tick_stopped = 1;
647 trace_tick_stop(1, " ");
651 * If the expiration time == KTIME_MAX, then
652 * in this case we simply stop the tick timer.
654 if (unlikely(expires.tv64 == KTIME_MAX)) {
655 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
656 hrtimer_cancel(&ts->sched_timer);
660 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
661 hrtimer_start(&ts->sched_timer, expires,
662 HRTIMER_MODE_ABS_PINNED);
663 /* Check, if the timer was already in the past */
664 if (hrtimer_active(&ts->sched_timer))
666 } else if (!tick_program_event(expires, 0))
669 * We are past the event already. So we crossed a
670 * jiffie boundary. Update jiffies and raise the
673 tick_do_update_jiffies64(ktime_get());
675 raise_softirq_irqoff(TIMER_SOFTIRQ);
677 ts->next_jiffies = next_jiffies;
678 ts->last_jiffies = last_jiffies;
679 ts->sleep_length = ktime_sub(dev->next_event, now);
684 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
686 #ifdef CONFIG_NO_HZ_FULL
687 int cpu = smp_processor_id();
689 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
692 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
695 if (!can_stop_full_tick())
698 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
702 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
705 * If this cpu is offline and it is the one which updates
706 * jiffies, then give up the assignment and let it be taken by
707 * the cpu which runs the tick timer next. If we don't drop
708 * this here the jiffies might be stale and do_timer() never
711 if (unlikely(!cpu_online(cpu))) {
712 if (cpu == tick_do_timer_cpu)
713 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
717 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
723 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
724 static int ratelimit;
726 if (ratelimit < 10 &&
727 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
728 pr_warn("NOHZ: local_softirq_pending %02x\n",
729 (unsigned int) local_softirq_pending());
735 if (have_nohz_full_mask) {
737 * Keep the tick alive to guarantee timekeeping progression
738 * if there are full dynticks CPUs around
740 if (tick_do_timer_cpu == cpu)
743 * Boot safety: make sure the timekeeping duty has been
744 * assigned before entering dyntick-idle mode,
746 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
753 static void __tick_nohz_idle_enter(struct tick_sched *ts)
755 ktime_t now, expires;
756 int cpu = smp_processor_id();
758 now = tick_nohz_start_idle(cpu, ts);
760 if (can_stop_idle_tick(cpu, ts)) {
761 int was_stopped = ts->tick_stopped;
765 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
766 if (expires.tv64 > 0LL) {
768 ts->idle_expires = expires;
771 if (!was_stopped && ts->tick_stopped)
772 ts->idle_jiffies = ts->last_jiffies;
777 * tick_nohz_idle_enter - stop the idle tick from the idle task
779 * When the next event is more than a tick into the future, stop the idle tick
780 * Called when we start the idle loop.
782 * The arch is responsible of calling:
784 * - rcu_idle_enter() after its last use of RCU before the CPU is put
786 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
788 void tick_nohz_idle_enter(void)
790 struct tick_sched *ts;
792 WARN_ON_ONCE(irqs_disabled());
795 * Update the idle state in the scheduler domain hierarchy
796 * when tick_nohz_stop_sched_tick() is called from the idle loop.
797 * State will be updated to busy during the first busy tick after
800 set_cpu_sd_state_idle();
804 ts = &__get_cpu_var(tick_cpu_sched);
806 * set ts->inidle unconditionally. even if the system did not
807 * switch to nohz mode the cpu frequency governers rely on the
808 * update of the idle time accounting in tick_nohz_start_idle().
811 __tick_nohz_idle_enter(ts);
815 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
818 * tick_nohz_irq_exit - update next tick event from interrupt exit
820 * When an interrupt fires while we are idle and it doesn't cause
821 * a reschedule, it may still add, modify or delete a timer, enqueue
822 * an RCU callback, etc...
823 * So we need to re-calculate and reprogram the next tick event.
825 void tick_nohz_irq_exit(void)
827 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
830 __tick_nohz_idle_enter(ts);
832 tick_nohz_full_stop_tick(ts);
836 * tick_nohz_get_sleep_length - return the length of the current sleep
838 * Called from power state control code with interrupts disabled
840 ktime_t tick_nohz_get_sleep_length(void)
842 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
844 return ts->sleep_length;
847 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
849 hrtimer_cancel(&ts->sched_timer);
850 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
853 /* Forward the time to expire in the future */
854 hrtimer_forward(&ts->sched_timer, now, tick_period);
856 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
857 hrtimer_start_expires(&ts->sched_timer,
858 HRTIMER_MODE_ABS_PINNED);
859 /* Check, if the timer was already in the past */
860 if (hrtimer_active(&ts->sched_timer))
863 if (!tick_program_event(
864 hrtimer_get_expires(&ts->sched_timer), 0))
867 /* Reread time and update jiffies */
869 tick_do_update_jiffies64(now);
873 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
875 /* Update jiffies first */
876 tick_do_update_jiffies64(now);
877 update_cpu_load_nohz();
879 calc_load_exit_idle();
880 touch_softlockup_watchdog();
882 * Cancel the scheduled timer and restore the tick
884 ts->tick_stopped = 0;
885 ts->idle_exittime = now;
887 tick_nohz_restart(ts, now);
890 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
892 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
895 if (vtime_accounting_enabled())
898 * We stopped the tick in idle. Update process times would miss the
899 * time we slept as update_process_times does only a 1 tick
900 * accounting. Enforce that this is accounted to idle !
902 ticks = jiffies - ts->idle_jiffies;
904 * We might be one off. Do not randomly account a huge number of ticks!
906 if (ticks && ticks < LONG_MAX)
907 account_idle_ticks(ticks);
912 * tick_nohz_idle_exit - restart the idle tick from the idle task
914 * Restart the idle tick when the CPU is woken up from idle
915 * This also exit the RCU extended quiescent state. The CPU
916 * can use RCU again after this function is called.
918 void tick_nohz_idle_exit(void)
920 int cpu = smp_processor_id();
921 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
926 WARN_ON_ONCE(!ts->inidle);
930 if (ts->idle_active || ts->tick_stopped)
934 tick_nohz_stop_idle(cpu, now);
936 if (ts->tick_stopped) {
937 tick_nohz_restart_sched_tick(ts, now);
938 tick_nohz_account_idle_ticks(ts);
943 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
945 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
947 hrtimer_forward(&ts->sched_timer, now, tick_period);
948 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
952 * The nohz low res interrupt handler
954 static void tick_nohz_handler(struct clock_event_device *dev)
956 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
957 struct pt_regs *regs = get_irq_regs();
958 ktime_t now = ktime_get();
960 dev->next_event.tv64 = KTIME_MAX;
962 tick_sched_do_timer(now);
963 tick_sched_handle(ts, regs);
965 while (tick_nohz_reprogram(ts, now)) {
967 tick_do_update_jiffies64(now);
972 * tick_nohz_switch_to_nohz - switch to nohz mode
974 static void tick_nohz_switch_to_nohz(void)
976 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
979 if (!tick_nohz_enabled)
983 if (tick_switch_to_oneshot(tick_nohz_handler)) {
988 ts->nohz_mode = NOHZ_MODE_LOWRES;
991 * Recycle the hrtimer in ts, so we can share the
992 * hrtimer_forward with the highres code.
994 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
995 /* Get the next period */
996 next = tick_init_jiffy_update();
999 hrtimer_set_expires(&ts->sched_timer, next);
1000 if (!tick_program_event(next, 0))
1002 next = ktime_add(next, tick_period);
1008 * When NOHZ is enabled and the tick is stopped, we need to kick the
1009 * tick timer from irq_enter() so that the jiffies update is kept
1010 * alive during long running softirqs. That's ugly as hell, but
1011 * correctness is key even if we need to fix the offending softirq in
1014 * Note, this is different to tick_nohz_restart. We just kick the
1015 * timer and do not touch the other magic bits which need to be done
1016 * when idle is left.
1018 static void tick_nohz_kick_tick(int cpu, ktime_t now)
1021 /* Switch back to 2.6.27 behaviour */
1023 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1027 * Do not touch the tick device, when the next expiry is either
1028 * already reached or less/equal than the tick period.
1030 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1031 if (delta.tv64 <= tick_period.tv64)
1034 tick_nohz_restart(ts, now);
1038 static inline void tick_check_nohz(int cpu)
1040 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1043 if (!ts->idle_active && !ts->tick_stopped)
1046 if (ts->idle_active)
1047 tick_nohz_stop_idle(cpu, now);
1048 if (ts->tick_stopped) {
1049 tick_nohz_update_jiffies(now);
1050 tick_nohz_kick_tick(cpu, now);
1056 static inline void tick_nohz_switch_to_nohz(void) { }
1057 static inline void tick_check_nohz(int cpu) { }
1059 #endif /* CONFIG_NO_HZ_COMMON */
1062 * Called from irq_enter to notify about the possible interruption of idle()
1064 void tick_check_idle(int cpu)
1066 tick_check_oneshot_broadcast(cpu);
1067 tick_check_nohz(cpu);
1071 * High resolution timer specific code
1073 #ifdef CONFIG_HIGH_RES_TIMERS
1075 * We rearm the timer until we get disabled by the idle code.
1076 * Called with interrupts disabled.
1078 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1080 struct tick_sched *ts =
1081 container_of(timer, struct tick_sched, sched_timer);
1082 struct pt_regs *regs = get_irq_regs();
1083 ktime_t now = ktime_get();
1085 tick_sched_do_timer(now);
1088 * Do not call, when we are not in irq context and have
1089 * no valid regs pointer
1092 tick_sched_handle(ts, regs);
1094 hrtimer_forward(timer, now, tick_period);
1096 return HRTIMER_RESTART;
1099 static int sched_skew_tick;
1101 static int __init skew_tick(char *str)
1103 get_option(&str, &sched_skew_tick);
1107 early_param("skew_tick", skew_tick);
1110 * tick_setup_sched_timer - setup the tick emulation timer
1112 void tick_setup_sched_timer(void)
1114 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1115 ktime_t now = ktime_get();
1118 * Emulate tick processing via per-CPU hrtimers:
1120 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1121 ts->sched_timer.function = tick_sched_timer;
1123 /* Get the next period (per cpu) */
1124 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1126 /* Offset the tick to avert jiffies_lock contention. */
1127 if (sched_skew_tick) {
1128 u64 offset = ktime_to_ns(tick_period) >> 1;
1129 do_div(offset, num_possible_cpus());
1130 offset *= smp_processor_id();
1131 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1135 hrtimer_forward(&ts->sched_timer, now, tick_period);
1136 hrtimer_start_expires(&ts->sched_timer,
1137 HRTIMER_MODE_ABS_PINNED);
1138 /* Check, if the timer was already in the past */
1139 if (hrtimer_active(&ts->sched_timer))
1144 #ifdef CONFIG_NO_HZ_COMMON
1145 if (tick_nohz_enabled)
1146 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1149 #endif /* HIGH_RES_TIMERS */
1151 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1152 void tick_cancel_sched_timer(int cpu)
1154 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1156 # ifdef CONFIG_HIGH_RES_TIMERS
1157 if (ts->sched_timer.base)
1158 hrtimer_cancel(&ts->sched_timer);
1161 memset(ts, 0, sizeof(*ts));
1166 * Async notification about clocksource changes
1168 void tick_clock_notify(void)
1172 for_each_possible_cpu(cpu)
1173 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1177 * Async notification about clock event changes
1179 void tick_oneshot_notify(void)
1181 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1183 set_bit(0, &ts->check_clocks);
1187 * Check, if a change happened, which makes oneshot possible.
1189 * Called cyclic from the hrtimer softirq (driven by the timer
1190 * softirq) allow_nohz signals, that we can switch into low-res nohz
1191 * mode, because high resolution timers are disabled (either compile
1194 int tick_check_oneshot_change(int allow_nohz)
1196 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1198 if (!test_and_clear_bit(0, &ts->check_clocks))
1201 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1204 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1210 tick_nohz_switch_to_nohz();