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>
24 #include <asm/irq_regs.h>
26 #include "tick-internal.h"
29 * Per cpu nohz control structure
31 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
36 static ktime_t last_jiffies_update;
38 struct tick_sched *tick_get_tick_sched(int cpu)
40 return &per_cpu(tick_cpu_sched, cpu);
44 * Must be called with interrupts disabled !
46 static void tick_do_update_jiffies64(ktime_t now)
48 unsigned long ticks = 0;
52 * Do a quick check without holding xtime_lock:
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 < tick_period.tv64)
58 /* Reevalute with xtime_lock held */
59 write_seqlock(&xtime_lock);
61 delta = ktime_sub(now, last_jiffies_update);
62 if (delta.tv64 >= tick_period.tv64) {
64 delta = ktime_sub(delta, tick_period);
65 last_jiffies_update = ktime_add(last_jiffies_update,
68 /* Slow path for long timeouts */
69 if (unlikely(delta.tv64 >= tick_period.tv64)) {
70 s64 incr = ktime_to_ns(tick_period);
72 ticks = ktime_divns(delta, incr);
74 last_jiffies_update = ktime_add_ns(last_jiffies_update,
79 /* Keep the tick_next_period variable up to date */
80 tick_next_period = ktime_add(last_jiffies_update, tick_period);
82 write_sequnlock(&xtime_lock);
86 * Initialize and return retrieve the jiffies update.
88 static ktime_t tick_init_jiffy_update(void)
92 write_seqlock(&xtime_lock);
93 /* Did we start the jiffies update yet ? */
94 if (last_jiffies_update.tv64 == 0)
95 last_jiffies_update = tick_next_period;
96 period = last_jiffies_update;
97 write_sequnlock(&xtime_lock);
102 * NOHZ - aka dynamic tick functionality
108 static int tick_nohz_enabled __read_mostly = 1;
111 * Enable / Disable tickless mode
113 static int __init setup_tick_nohz(char *str)
115 if (!strcmp(str, "off"))
116 tick_nohz_enabled = 0;
117 else if (!strcmp(str, "on"))
118 tick_nohz_enabled = 1;
124 __setup("nohz=", setup_tick_nohz);
127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
129 * Called from interrupt entry when the CPU was idle
131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
133 * value. We do this unconditionally on any cpu, as we don't know whether the
134 * cpu, which has the update task assigned is in a long sleep.
136 static void tick_nohz_update_jiffies(ktime_t now)
138 int cpu = smp_processor_id();
139 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
142 ts->idle_waketime = now;
144 local_irq_save(flags);
145 tick_do_update_jiffies64(now);
146 local_irq_restore(flags);
148 touch_softlockup_watchdog();
152 * Updates the per cpu time idle statistics counters
155 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
159 if (ts->idle_active) {
160 delta = ktime_sub(now, ts->idle_entrytime);
161 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
162 if (nr_iowait_cpu(cpu) > 0)
163 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
164 ts->idle_entrytime = now;
167 if (last_update_time)
168 *last_update_time = ktime_to_us(now);
172 static void tick_nohz_stop_idle(int cpu, ktime_t now)
174 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
176 update_ts_time_stats(cpu, ts, now, NULL);
179 sched_clock_idle_wakeup_event(0);
182 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
188 update_ts_time_stats(cpu, ts, now, NULL);
190 ts->idle_entrytime = now;
192 sched_clock_idle_sleep_event();
197 * get_cpu_idle_time_us - get the total idle time of a cpu
198 * @cpu: CPU number to query
199 * @last_update_time: variable to store update time in
201 * Return the cummulative idle time (since boot) for a given
202 * CPU, in microseconds. The idle time returned includes
203 * the iowait time (unlike what "top" and co report).
205 * This time is measured via accounting rather than sampling,
206 * and is as accurate as ktime_get() is.
208 * This function returns -1 if NOHZ is not enabled.
210 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
212 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
214 if (!tick_nohz_enabled)
217 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
219 return ktime_to_us(ts->idle_sleeptime);
221 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
224 * get_cpu_iowait_time_us - get the total iowait time of a cpu
225 * @cpu: CPU number to query
226 * @last_update_time: variable to store update time in
228 * Return the cummulative iowait time (since boot) for a given
229 * CPU, in microseconds.
231 * This time is measured via accounting rather than sampling,
232 * and is as accurate as ktime_get() is.
234 * This function returns -1 if NOHZ is not enabled.
236 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
238 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
240 if (!tick_nohz_enabled)
243 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
245 return ktime_to_us(ts->iowait_sleeptime);
247 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
250 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
252 * When the next event is more than a tick into the future, stop the idle tick
253 * Called either from the idle loop or from irq_exit() when an idle period was
254 * just interrupted by an interrupt which did not cause a reschedule.
256 void tick_nohz_stop_sched_tick(int inidle)
258 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
259 struct tick_sched *ts;
260 ktime_t last_update, expires, now;
261 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
265 local_irq_save(flags);
267 cpu = smp_processor_id();
268 ts = &per_cpu(tick_cpu_sched, cpu);
271 * Call to tick_nohz_start_idle stops the last_update_time from being
272 * updated. Thus, it must not be called in the event we are called from
273 * irq_exit() with the prior state different than idle.
275 if (!inidle && !ts->inidle)
279 * Set ts->inidle unconditionally. Even if the system did not
280 * switch to NOHZ mode the cpu frequency governers rely on the
281 * update of the idle time accounting in tick_nohz_start_idle().
285 now = tick_nohz_start_idle(cpu, ts);
288 * If this cpu is offline and it is the one which updates
289 * jiffies, then give up the assignment and let it be taken by
290 * the cpu which runs the tick timer next. If we don't drop
291 * this here the jiffies might be stale and do_timer() never
294 if (unlikely(!cpu_online(cpu))) {
295 if (cpu == tick_do_timer_cpu)
296 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
299 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
305 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
306 static int ratelimit;
308 if (ratelimit < 10) {
309 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
310 (unsigned int) local_softirq_pending());
317 /* Read jiffies and the time when jiffies were updated last */
319 seq = read_seqbegin(&xtime_lock);
320 last_update = last_jiffies_update;
321 last_jiffies = jiffies;
322 time_delta = timekeeping_max_deferment();
323 } while (read_seqretry(&xtime_lock, seq));
325 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
326 arch_needs_cpu(cpu)) {
327 next_jiffies = last_jiffies + 1;
330 /* Get the next timer wheel timer */
331 next_jiffies = get_next_timer_interrupt(last_jiffies);
332 delta_jiffies = next_jiffies - last_jiffies;
335 * Do not stop the tick, if we are only one off
336 * or if the cpu is required for rcu
338 if (!ts->tick_stopped && delta_jiffies == 1)
341 /* Schedule the tick, if we are at least one jiffie off */
342 if ((long)delta_jiffies >= 1) {
345 * If this cpu is the one which updates jiffies, then
346 * give up the assignment and let it be taken by the
347 * cpu which runs the tick timer next, which might be
348 * this cpu as well. If we don't drop this here the
349 * jiffies might be stale and do_timer() never
350 * invoked. Keep track of the fact that it was the one
351 * which had the do_timer() duty last. If this cpu is
352 * the one which had the do_timer() duty last, we
353 * limit the sleep time to the timekeeping
354 * max_deferement value which we retrieved
355 * above. Otherwise we can sleep as long as we want.
357 if (cpu == tick_do_timer_cpu) {
358 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
359 ts->do_timer_last = 1;
360 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
361 time_delta = KTIME_MAX;
362 ts->do_timer_last = 0;
363 } else if (!ts->do_timer_last) {
364 time_delta = KTIME_MAX;
368 * calculate the expiry time for the next timer wheel
369 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
370 * that there is no timer pending or at least extremely
371 * far into the future (12 days for HZ=1000). In this
372 * case we set the expiry to the end of time.
374 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
376 * Calculate the time delta for the next timer event.
377 * If the time delta exceeds the maximum time delta
378 * permitted by the current clocksource then adjust
379 * the time delta accordingly to ensure the
380 * clocksource does not wrap.
382 time_delta = min_t(u64, time_delta,
383 tick_period.tv64 * delta_jiffies);
386 if (time_delta < KTIME_MAX)
387 expires = ktime_add_ns(last_update, time_delta);
389 expires.tv64 = KTIME_MAX;
391 /* Skip reprogram of event if its not changed */
392 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
396 * nohz_stop_sched_tick can be called several times before
397 * the nohz_restart_sched_tick is called. This happens when
398 * interrupts arrive which do not cause a reschedule. In the
399 * first call we save the current tick time, so we can restart
400 * the scheduler tick in nohz_restart_sched_tick.
402 if (!ts->tick_stopped) {
403 select_nohz_load_balancer(1);
405 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
406 ts->tick_stopped = 1;
407 ts->idle_jiffies = last_jiffies;
414 ts->idle_expires = expires;
417 * If the expiration time == KTIME_MAX, then
418 * in this case we simply stop the tick timer.
420 if (unlikely(expires.tv64 == KTIME_MAX)) {
421 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
422 hrtimer_cancel(&ts->sched_timer);
426 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
427 hrtimer_start(&ts->sched_timer, expires,
428 HRTIMER_MODE_ABS_PINNED);
429 /* Check, if the timer was already in the past */
430 if (hrtimer_active(&ts->sched_timer))
432 } else if (!tick_program_event(expires, 0))
435 * We are past the event already. So we crossed a
436 * jiffie boundary. Update jiffies and raise the
439 tick_do_update_jiffies64(ktime_get());
441 raise_softirq_irqoff(TIMER_SOFTIRQ);
443 ts->next_jiffies = next_jiffies;
444 ts->last_jiffies = last_jiffies;
445 ts->sleep_length = ktime_sub(dev->next_event, now);
447 local_irq_restore(flags);
451 * tick_nohz_get_sleep_length - return the length of the current sleep
453 * Called from power state control code with interrupts disabled
455 ktime_t tick_nohz_get_sleep_length(void)
457 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
459 return ts->sleep_length;
462 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
464 hrtimer_cancel(&ts->sched_timer);
465 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
468 /* Forward the time to expire in the future */
469 hrtimer_forward(&ts->sched_timer, now, tick_period);
471 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
472 hrtimer_start_expires(&ts->sched_timer,
473 HRTIMER_MODE_ABS_PINNED);
474 /* Check, if the timer was already in the past */
475 if (hrtimer_active(&ts->sched_timer))
478 if (!tick_program_event(
479 hrtimer_get_expires(&ts->sched_timer), 0))
482 /* Update jiffies and reread time */
483 tick_do_update_jiffies64(now);
489 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
491 * Restart the idle tick when the CPU is woken up from idle
493 void tick_nohz_restart_sched_tick(void)
495 int cpu = smp_processor_id();
496 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
497 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
503 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
507 tick_nohz_stop_idle(cpu, now);
509 if (!ts->inidle || !ts->tick_stopped) {
519 /* Update jiffies first */
520 select_nohz_load_balancer(0);
521 tick_do_update_jiffies64(now);
523 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
525 * We stopped the tick in idle. Update process times would miss the
526 * time we slept as update_process_times does only a 1 tick
527 * accounting. Enforce that this is accounted to idle !
529 ticks = jiffies - ts->idle_jiffies;
531 * We might be one off. Do not randomly account a huge number of ticks!
533 if (ticks && ticks < LONG_MAX)
534 account_idle_ticks(ticks);
537 touch_softlockup_watchdog();
539 * Cancel the scheduled timer and restore the tick
541 ts->tick_stopped = 0;
542 ts->idle_exittime = now;
544 tick_nohz_restart(ts, now);
549 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
551 hrtimer_forward(&ts->sched_timer, now, tick_period);
552 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
556 * The nohz low res interrupt handler
558 static void tick_nohz_handler(struct clock_event_device *dev)
560 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
561 struct pt_regs *regs = get_irq_regs();
562 int cpu = smp_processor_id();
563 ktime_t now = ktime_get();
565 dev->next_event.tv64 = KTIME_MAX;
568 * Check if the do_timer duty was dropped. We don't care about
569 * concurrency: This happens only when the cpu in charge went
570 * into a long sleep. If two cpus happen to assign themself to
571 * this duty, then the jiffies update is still serialized by
574 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
575 tick_do_timer_cpu = cpu;
577 /* Check, if the jiffies need an update */
578 if (tick_do_timer_cpu == cpu)
579 tick_do_update_jiffies64(now);
582 * When we are idle and the tick is stopped, we have to touch
583 * the watchdog as we might not schedule for a really long
584 * time. This happens on complete idle SMP systems while
585 * waiting on the login prompt. We also increment the "start
586 * of idle" jiffy stamp so the idle accounting adjustment we
587 * do when we go busy again does not account too much ticks.
589 if (ts->tick_stopped) {
590 touch_softlockup_watchdog();
594 update_process_times(user_mode(regs));
595 profile_tick(CPU_PROFILING);
597 while (tick_nohz_reprogram(ts, now)) {
599 tick_do_update_jiffies64(now);
604 * tick_nohz_switch_to_nohz - switch to nohz mode
606 static void tick_nohz_switch_to_nohz(void)
608 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
611 if (!tick_nohz_enabled)
615 if (tick_switch_to_oneshot(tick_nohz_handler)) {
620 ts->nohz_mode = NOHZ_MODE_LOWRES;
623 * Recycle the hrtimer in ts, so we can share the
624 * hrtimer_forward with the highres code.
626 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
627 /* Get the next period */
628 next = tick_init_jiffy_update();
631 hrtimer_set_expires(&ts->sched_timer, next);
632 if (!tick_program_event(next, 0))
634 next = ktime_add(next, tick_period);
638 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", smp_processor_id());
642 * When NOHZ is enabled and the tick is stopped, we need to kick the
643 * tick timer from irq_enter() so that the jiffies update is kept
644 * alive during long running softirqs. That's ugly as hell, but
645 * correctness is key even if we need to fix the offending softirq in
648 * Note, this is different to tick_nohz_restart. We just kick the
649 * timer and do not touch the other magic bits which need to be done
652 static void tick_nohz_kick_tick(int cpu, ktime_t now)
655 /* Switch back to 2.6.27 behaviour */
657 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
661 * Do not touch the tick device, when the next expiry is either
662 * already reached or less/equal than the tick period.
664 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
665 if (delta.tv64 <= tick_period.tv64)
668 tick_nohz_restart(ts, now);
672 static inline void tick_check_nohz(int cpu)
674 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
677 if (!ts->idle_active && !ts->tick_stopped)
681 tick_nohz_stop_idle(cpu, now);
682 if (ts->tick_stopped) {
683 tick_nohz_update_jiffies(now);
684 tick_nohz_kick_tick(cpu, now);
690 static inline void tick_nohz_switch_to_nohz(void) { }
691 static inline void tick_check_nohz(int cpu) { }
696 * Called from irq_enter to notify about the possible interruption of idle()
698 void tick_check_idle(int cpu)
700 tick_check_oneshot_broadcast(cpu);
701 tick_check_nohz(cpu);
705 * High resolution timer specific code
707 #ifdef CONFIG_HIGH_RES_TIMERS
709 * We rearm the timer until we get disabled by the idle code.
710 * Called with interrupts disabled and timer->base->cpu_base->lock held.
712 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
714 struct tick_sched *ts =
715 container_of(timer, struct tick_sched, sched_timer);
716 struct pt_regs *regs = get_irq_regs();
717 ktime_t now = ktime_get();
718 int cpu = smp_processor_id();
722 * Check if the do_timer duty was dropped. We don't care about
723 * concurrency: This happens only when the cpu in charge went
724 * into a long sleep. If two cpus happen to assign themself to
725 * this duty, then the jiffies update is still serialized by
728 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
729 tick_do_timer_cpu = cpu;
732 /* Check, if the jiffies need an update */
733 if (tick_do_timer_cpu == cpu)
734 tick_do_update_jiffies64(now);
737 * Do not call, when we are not in irq context and have
738 * no valid regs pointer
742 * When we are idle and the tick is stopped, we have to touch
743 * the watchdog as we might not schedule for a really long
744 * time. This happens on complete idle SMP systems while
745 * waiting on the login prompt. We also increment the "start of
746 * idle" jiffy stamp so the idle accounting adjustment we do
747 * when we go busy again does not account too much ticks.
749 if (ts->tick_stopped) {
750 touch_softlockup_watchdog();
753 update_process_times(user_mode(regs));
754 profile_tick(CPU_PROFILING);
757 hrtimer_forward(timer, now, tick_period);
759 return HRTIMER_RESTART;
763 * tick_setup_sched_timer - setup the tick emulation timer
765 void tick_setup_sched_timer(void)
767 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
768 ktime_t now = ktime_get();
771 * Emulate tick processing via per-CPU hrtimers:
773 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
774 ts->sched_timer.function = tick_sched_timer;
776 /* Get the next period (per cpu) */
777 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
780 hrtimer_forward(&ts->sched_timer, now, tick_period);
781 hrtimer_start_expires(&ts->sched_timer,
782 HRTIMER_MODE_ABS_PINNED);
783 /* Check, if the timer was already in the past */
784 if (hrtimer_active(&ts->sched_timer))
790 if (tick_nohz_enabled) {
791 ts->nohz_mode = NOHZ_MODE_HIGHRES;
792 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", smp_processor_id());
796 #endif /* HIGH_RES_TIMERS */
798 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
799 void tick_cancel_sched_timer(int cpu)
801 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
803 # ifdef CONFIG_HIGH_RES_TIMERS
804 if (ts->sched_timer.base)
805 hrtimer_cancel(&ts->sched_timer);
808 ts->nohz_mode = NOHZ_MODE_INACTIVE;
813 * Async notification about clocksource changes
815 void tick_clock_notify(void)
819 for_each_possible_cpu(cpu)
820 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
824 * Async notification about clock event changes
826 void tick_oneshot_notify(void)
828 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
830 set_bit(0, &ts->check_clocks);
834 * Check, if a change happened, which makes oneshot possible.
836 * Called cyclic from the hrtimer softirq (driven by the timer
837 * softirq) allow_nohz signals, that we can switch into low-res nohz
838 * mode, because high resolution timers are disabled (either compile
841 int tick_check_oneshot_change(int allow_nohz)
843 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
845 if (!test_and_clear_bit(0, &ts->check_clocks))
848 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
851 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
857 tick_nohz_switch_to_nohz();