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 if (nr_iowait_cpu(cpu) > 0)
162 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
164 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
165 ts->idle_entrytime = now;
168 if (last_update_time)
169 *last_update_time = ktime_to_us(now);
173 static void tick_nohz_stop_idle(int cpu, ktime_t now)
175 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
177 update_ts_time_stats(cpu, ts, now, NULL);
180 sched_clock_idle_wakeup_event(0);
183 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
189 update_ts_time_stats(cpu, ts, now, NULL);
191 ts->idle_entrytime = now;
193 sched_clock_idle_sleep_event();
198 * get_cpu_idle_time_us - get the total idle time of a cpu
199 * @cpu: CPU number to query
200 * @last_update_time: variable to store update time in. Do not update
203 * Return the cummulative idle time (since boot) for a given
204 * CPU, in microseconds.
206 * This time is measured via accounting rather than sampling,
207 * and is as accurate as ktime_get() is.
209 * This function returns -1 if NOHZ is not enabled.
211 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
213 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
216 if (!tick_nohz_enabled)
220 if (last_update_time) {
221 update_ts_time_stats(cpu, ts, now, last_update_time);
222 idle = ts->idle_sleeptime;
224 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
225 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
227 idle = ktime_add(ts->idle_sleeptime, delta);
229 idle = ts->idle_sleeptime;
233 return ktime_to_us(idle);
236 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
239 * get_cpu_iowait_time_us - get the total iowait time of a cpu
240 * @cpu: CPU number to query
241 * @last_update_time: variable to store update time in. Do not update
244 * Return the cummulative iowait time (since boot) for a given
245 * CPU, in microseconds.
247 * This time is measured via accounting rather than sampling,
248 * and is as accurate as ktime_get() is.
250 * This function returns -1 if NOHZ is not enabled.
252 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
254 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
257 if (!tick_nohz_enabled)
261 if (last_update_time) {
262 update_ts_time_stats(cpu, ts, now, last_update_time);
263 iowait = ts->iowait_sleeptime;
265 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
266 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
268 iowait = ktime_add(ts->iowait_sleeptime, delta);
270 iowait = ts->iowait_sleeptime;
274 return ktime_to_us(iowait);
276 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
279 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
281 * When the next event is more than a tick into the future, stop the idle tick
282 * Called either from the idle loop or from irq_exit() when an idle period was
283 * just interrupted by an interrupt which did not cause a reschedule.
285 void tick_nohz_stop_sched_tick(int inidle)
287 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
288 struct tick_sched *ts;
289 ktime_t last_update, expires, now;
290 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
294 local_irq_save(flags);
296 cpu = smp_processor_id();
297 ts = &per_cpu(tick_cpu_sched, cpu);
300 * Call to tick_nohz_start_idle stops the last_update_time from being
301 * updated. Thus, it must not be called in the event we are called from
302 * irq_exit() with the prior state different than idle.
304 if (!inidle && !ts->inidle)
308 * Set ts->inidle unconditionally. Even if the system did not
309 * switch to NOHZ mode the cpu frequency governers rely on the
310 * update of the idle time accounting in tick_nohz_start_idle().
314 now = tick_nohz_start_idle(cpu, ts);
317 * If this cpu is offline and it is the one which updates
318 * jiffies, then give up the assignment and let it be taken by
319 * the cpu which runs the tick timer next. If we don't drop
320 * this here the jiffies might be stale and do_timer() never
323 if (unlikely(!cpu_online(cpu))) {
324 if (cpu == tick_do_timer_cpu)
325 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
328 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
334 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
335 static int ratelimit;
337 if (ratelimit < 10) {
338 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
339 (unsigned int) local_softirq_pending());
346 /* Read jiffies and the time when jiffies were updated last */
348 seq = read_seqbegin(&xtime_lock);
349 last_update = last_jiffies_update;
350 last_jiffies = jiffies;
351 time_delta = timekeeping_max_deferment();
352 } while (read_seqretry(&xtime_lock, seq));
354 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
355 arch_needs_cpu(cpu)) {
356 next_jiffies = last_jiffies + 1;
359 /* Get the next timer wheel timer */
360 next_jiffies = get_next_timer_interrupt(last_jiffies);
361 delta_jiffies = next_jiffies - last_jiffies;
364 * Do not stop the tick, if we are only one off
365 * or if the cpu is required for rcu
367 if (!ts->tick_stopped && delta_jiffies == 1)
370 /* Schedule the tick, if we are at least one jiffie off */
371 if ((long)delta_jiffies >= 1) {
374 * If this cpu is the one which updates jiffies, then
375 * give up the assignment and let it be taken by the
376 * cpu which runs the tick timer next, which might be
377 * this cpu as well. If we don't drop this here the
378 * jiffies might be stale and do_timer() never
379 * invoked. Keep track of the fact that it was the one
380 * which had the do_timer() duty last. If this cpu is
381 * the one which had the do_timer() duty last, we
382 * limit the sleep time to the timekeeping
383 * max_deferement value which we retrieved
384 * above. Otherwise we can sleep as long as we want.
386 if (cpu == tick_do_timer_cpu) {
387 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
388 ts->do_timer_last = 1;
389 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
390 time_delta = KTIME_MAX;
391 ts->do_timer_last = 0;
392 } else if (!ts->do_timer_last) {
393 time_delta = KTIME_MAX;
397 * calculate the expiry time for the next timer wheel
398 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
399 * that there is no timer pending or at least extremely
400 * far into the future (12 days for HZ=1000). In this
401 * case we set the expiry to the end of time.
403 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
405 * Calculate the time delta for the next timer event.
406 * If the time delta exceeds the maximum time delta
407 * permitted by the current clocksource then adjust
408 * the time delta accordingly to ensure the
409 * clocksource does not wrap.
411 time_delta = min_t(u64, time_delta,
412 tick_period.tv64 * delta_jiffies);
415 if (time_delta < KTIME_MAX)
416 expires = ktime_add_ns(last_update, time_delta);
418 expires.tv64 = KTIME_MAX;
420 /* Skip reprogram of event if its not changed */
421 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
425 * nohz_stop_sched_tick can be called several times before
426 * the nohz_restart_sched_tick is called. This happens when
427 * interrupts arrive which do not cause a reschedule. In the
428 * first call we save the current tick time, so we can restart
429 * the scheduler tick in nohz_restart_sched_tick.
431 if (!ts->tick_stopped) {
432 select_nohz_load_balancer(1);
434 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
435 ts->tick_stopped = 1;
436 ts->idle_jiffies = last_jiffies;
443 ts->idle_expires = expires;
446 * If the expiration time == KTIME_MAX, then
447 * in this case we simply stop the tick timer.
449 if (unlikely(expires.tv64 == KTIME_MAX)) {
450 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
451 hrtimer_cancel(&ts->sched_timer);
455 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
456 hrtimer_start(&ts->sched_timer, expires,
457 HRTIMER_MODE_ABS_PINNED);
458 /* Check, if the timer was already in the past */
459 if (hrtimer_active(&ts->sched_timer))
461 } else if (!tick_program_event(expires, 0))
464 * We are past the event already. So we crossed a
465 * jiffie boundary. Update jiffies and raise the
468 tick_do_update_jiffies64(ktime_get());
470 raise_softirq_irqoff(TIMER_SOFTIRQ);
472 ts->next_jiffies = next_jiffies;
473 ts->last_jiffies = last_jiffies;
474 ts->sleep_length = ktime_sub(dev->next_event, now);
476 local_irq_restore(flags);
480 * tick_nohz_get_sleep_length - return the length of the current sleep
482 * Called from power state control code with interrupts disabled
484 ktime_t tick_nohz_get_sleep_length(void)
486 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
488 return ts->sleep_length;
491 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
493 hrtimer_cancel(&ts->sched_timer);
494 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
497 /* Forward the time to expire in the future */
498 hrtimer_forward(&ts->sched_timer, now, tick_period);
500 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
501 hrtimer_start_expires(&ts->sched_timer,
502 HRTIMER_MODE_ABS_PINNED);
503 /* Check, if the timer was already in the past */
504 if (hrtimer_active(&ts->sched_timer))
507 if (!tick_program_event(
508 hrtimer_get_expires(&ts->sched_timer), 0))
511 /* Update jiffies and reread time */
512 tick_do_update_jiffies64(now);
518 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
520 * Restart the idle tick when the CPU is woken up from idle
522 void tick_nohz_restart_sched_tick(void)
524 int cpu = smp_processor_id();
525 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
526 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
532 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
536 tick_nohz_stop_idle(cpu, now);
538 if (!ts->inidle || !ts->tick_stopped) {
548 /* Update jiffies first */
549 select_nohz_load_balancer(0);
550 tick_do_update_jiffies64(now);
552 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
554 * We stopped the tick in idle. Update process times would miss the
555 * time we slept as update_process_times does only a 1 tick
556 * accounting. Enforce that this is accounted to idle !
558 ticks = jiffies - ts->idle_jiffies;
560 * We might be one off. Do not randomly account a huge number of ticks!
562 if (ticks && ticks < LONG_MAX)
563 account_idle_ticks(ticks);
566 touch_softlockup_watchdog();
568 * Cancel the scheduled timer and restore the tick
570 ts->tick_stopped = 0;
571 ts->idle_exittime = now;
573 tick_nohz_restart(ts, now);
578 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
580 hrtimer_forward(&ts->sched_timer, now, tick_period);
581 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
585 * The nohz low res interrupt handler
587 static void tick_nohz_handler(struct clock_event_device *dev)
589 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
590 struct pt_regs *regs = get_irq_regs();
591 int cpu = smp_processor_id();
592 ktime_t now = ktime_get();
594 dev->next_event.tv64 = KTIME_MAX;
597 * Check if the do_timer duty was dropped. We don't care about
598 * concurrency: This happens only when the cpu in charge went
599 * into a long sleep. If two cpus happen to assign themself to
600 * this duty, then the jiffies update is still serialized by
603 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
604 tick_do_timer_cpu = cpu;
606 /* Check, if the jiffies need an update */
607 if (tick_do_timer_cpu == cpu)
608 tick_do_update_jiffies64(now);
611 * When we are idle and the tick is stopped, we have to touch
612 * the watchdog as we might not schedule for a really long
613 * time. This happens on complete idle SMP systems while
614 * waiting on the login prompt. We also increment the "start
615 * of idle" jiffy stamp so the idle accounting adjustment we
616 * do when we go busy again does not account too much ticks.
618 if (ts->tick_stopped) {
619 touch_softlockup_watchdog();
623 update_process_times(user_mode(regs));
624 profile_tick(CPU_PROFILING);
626 while (tick_nohz_reprogram(ts, now)) {
628 tick_do_update_jiffies64(now);
633 * tick_nohz_switch_to_nohz - switch to nohz mode
635 static void tick_nohz_switch_to_nohz(void)
637 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
640 if (!tick_nohz_enabled)
644 if (tick_switch_to_oneshot(tick_nohz_handler)) {
649 ts->nohz_mode = NOHZ_MODE_LOWRES;
652 * Recycle the hrtimer in ts, so we can share the
653 * hrtimer_forward with the highres code.
655 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
656 /* Get the next period */
657 next = tick_init_jiffy_update();
660 hrtimer_set_expires(&ts->sched_timer, next);
661 if (!tick_program_event(next, 0))
663 next = ktime_add(next, tick_period);
669 * When NOHZ is enabled and the tick is stopped, we need to kick the
670 * tick timer from irq_enter() so that the jiffies update is kept
671 * alive during long running softirqs. That's ugly as hell, but
672 * correctness is key even if we need to fix the offending softirq in
675 * Note, this is different to tick_nohz_restart. We just kick the
676 * timer and do not touch the other magic bits which need to be done
679 static void tick_nohz_kick_tick(int cpu, ktime_t now)
682 /* Switch back to 2.6.27 behaviour */
684 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
688 * Do not touch the tick device, when the next expiry is either
689 * already reached or less/equal than the tick period.
691 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
692 if (delta.tv64 <= tick_period.tv64)
695 tick_nohz_restart(ts, now);
699 static inline void tick_check_nohz(int cpu)
701 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
704 if (!ts->idle_active && !ts->tick_stopped)
708 tick_nohz_stop_idle(cpu, now);
709 if (ts->tick_stopped) {
710 tick_nohz_update_jiffies(now);
711 tick_nohz_kick_tick(cpu, now);
717 static inline void tick_nohz_switch_to_nohz(void) { }
718 static inline void tick_check_nohz(int cpu) { }
723 * Called from irq_enter to notify about the possible interruption of idle()
725 void tick_check_idle(int cpu)
727 tick_check_oneshot_broadcast(cpu);
728 tick_check_nohz(cpu);
732 * High resolution timer specific code
734 #ifdef CONFIG_HIGH_RES_TIMERS
736 * We rearm the timer until we get disabled by the idle code.
737 * Called with interrupts disabled and timer->base->cpu_base->lock held.
739 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
741 struct tick_sched *ts =
742 container_of(timer, struct tick_sched, sched_timer);
743 struct pt_regs *regs = get_irq_regs();
744 ktime_t now = ktime_get();
745 int cpu = smp_processor_id();
749 * Check if the do_timer duty was dropped. We don't care about
750 * concurrency: This happens only when the cpu in charge went
751 * into a long sleep. If two cpus happen to assign themself to
752 * this duty, then the jiffies update is still serialized by
755 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
756 tick_do_timer_cpu = cpu;
759 /* Check, if the jiffies need an update */
760 if (tick_do_timer_cpu == cpu)
761 tick_do_update_jiffies64(now);
764 * Do not call, when we are not in irq context and have
765 * no valid regs pointer
769 * When we are idle and the tick is stopped, we have to touch
770 * the watchdog as we might not schedule for a really long
771 * time. This happens on complete idle SMP systems while
772 * waiting on the login prompt. We also increment the "start of
773 * idle" jiffy stamp so the idle accounting adjustment we do
774 * when we go busy again does not account too much ticks.
776 if (ts->tick_stopped) {
777 touch_softlockup_watchdog();
780 update_process_times(user_mode(regs));
781 profile_tick(CPU_PROFILING);
784 hrtimer_forward(timer, now, tick_period);
786 return HRTIMER_RESTART;
790 * tick_setup_sched_timer - setup the tick emulation timer
792 void tick_setup_sched_timer(void)
794 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
795 ktime_t now = ktime_get();
798 * Emulate tick processing via per-CPU hrtimers:
800 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
801 ts->sched_timer.function = tick_sched_timer;
803 /* Get the next period (per cpu) */
804 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
807 hrtimer_forward(&ts->sched_timer, now, tick_period);
808 hrtimer_start_expires(&ts->sched_timer,
809 HRTIMER_MODE_ABS_PINNED);
810 /* Check, if the timer was already in the past */
811 if (hrtimer_active(&ts->sched_timer))
817 if (tick_nohz_enabled)
818 ts->nohz_mode = NOHZ_MODE_HIGHRES;
821 #endif /* HIGH_RES_TIMERS */
823 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
824 void tick_cancel_sched_timer(int cpu)
826 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
828 # ifdef CONFIG_HIGH_RES_TIMERS
829 if (ts->sched_timer.base)
830 hrtimer_cancel(&ts->sched_timer);
833 ts->nohz_mode = NOHZ_MODE_INACTIVE;
838 * Async notification about clocksource changes
840 void tick_clock_notify(void)
844 for_each_possible_cpu(cpu)
845 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
849 * Async notification about clock event changes
851 void tick_oneshot_notify(void)
853 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
855 set_bit(0, &ts->check_clocks);
859 * Check, if a change happened, which makes oneshot possible.
861 * Called cyclic from the hrtimer softirq (driven by the timer
862 * softirq) allow_nohz signals, that we can switch into low-res nohz
863 * mode, because high resolution timers are disabled (either compile
866 int tick_check_oneshot_change(int allow_nohz)
868 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
870 if (!test_and_clear_bit(0, &ts->check_clocks))
873 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
876 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
882 tick_nohz_switch_to_nohz();