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/tick.h>
23 #include <linux/module.h>
25 #include <asm/irq_regs.h>
27 #include "tick-internal.h"
30 * Per cpu nohz control structure
32 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
35 * The time, when the last jiffy update happened. Protected by xtime_lock.
37 static ktime_t last_jiffies_update;
39 struct tick_sched *tick_get_tick_sched(int cpu)
41 return &per_cpu(tick_cpu_sched, cpu);
45 * Must be called with interrupts disabled !
47 static void tick_do_update_jiffies64(ktime_t now)
49 unsigned long ticks = 0;
53 * Do a quick check without holding xtime_lock:
55 delta = ktime_sub(now, last_jiffies_update);
56 if (delta.tv64 < tick_period.tv64)
59 /* Reevalute with xtime_lock held */
60 write_seqlock(&xtime_lock);
62 delta = ktime_sub(now, last_jiffies_update);
63 if (delta.tv64 >= tick_period.tv64) {
65 delta = ktime_sub(delta, tick_period);
66 last_jiffies_update = ktime_add(last_jiffies_update,
69 /* Slow path for long timeouts */
70 if (unlikely(delta.tv64 >= tick_period.tv64)) {
71 s64 incr = ktime_to_ns(tick_period);
73 ticks = ktime_divns(delta, incr);
75 last_jiffies_update = ktime_add_ns(last_jiffies_update,
80 /* Keep the tick_next_period variable up to date */
81 tick_next_period = ktime_add(last_jiffies_update, tick_period);
83 write_sequnlock(&xtime_lock);
87 * Initialize and return retrieve the jiffies update.
89 static ktime_t tick_init_jiffy_update(void)
93 write_seqlock(&xtime_lock);
94 /* Did we start the jiffies update yet ? */
95 if (last_jiffies_update.tv64 == 0)
96 last_jiffies_update = tick_next_period;
97 period = last_jiffies_update;
98 write_sequnlock(&xtime_lock);
103 * NOHZ - aka dynamic tick functionality
109 static int tick_nohz_enabled __read_mostly = 1;
112 * Enable / Disable tickless mode
114 static int __init setup_tick_nohz(char *str)
116 if (!strcmp(str, "off"))
117 tick_nohz_enabled = 0;
118 else if (!strcmp(str, "on"))
119 tick_nohz_enabled = 1;
125 __setup("nohz=", setup_tick_nohz);
128 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
130 * Called from interrupt entry when the CPU was idle
132 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
133 * must be updated. Otherwise an interrupt handler could use a stale jiffy
134 * value. We do this unconditionally on any cpu, as we don't know whether the
135 * cpu, which has the update task assigned is in a long sleep.
137 static void tick_nohz_update_jiffies(ktime_t now)
139 int cpu = smp_processor_id();
140 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
143 cpumask_clear_cpu(cpu, nohz_cpu_mask);
144 ts->idle_waketime = now;
146 local_irq_save(flags);
147 tick_do_update_jiffies64(now);
148 local_irq_restore(flags);
150 touch_softlockup_watchdog();
154 * Updates the per cpu time idle statistics counters
157 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
161 if (ts->idle_active) {
162 delta = ktime_sub(now, ts->idle_entrytime);
163 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
164 if (nr_iowait_cpu(cpu) > 0)
165 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
166 ts->idle_entrytime = now;
169 if (last_update_time)
170 *last_update_time = ktime_to_us(now);
174 static void tick_nohz_stop_idle(int cpu, ktime_t now)
176 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
178 update_ts_time_stats(cpu, ts, now, NULL);
181 sched_clock_idle_wakeup_event(0);
184 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
190 update_ts_time_stats(cpu, ts, now, NULL);
192 ts->idle_entrytime = now;
194 sched_clock_idle_sleep_event();
199 * get_cpu_idle_time_us - get the total idle time of a cpu
200 * @cpu: CPU number to query
201 * @last_update_time: variable to store update time in
203 * Return the cummulative idle time (since boot) for a given
204 * CPU, in microseconds. The idle time returned includes
205 * the iowait time (unlike what "top" and co report).
207 * This time is measured via accounting rather than sampling,
208 * and is as accurate as ktime_get() is.
210 * This function returns -1 if NOHZ is not enabled.
212 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
214 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
216 if (!tick_nohz_enabled)
219 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
221 return ktime_to_us(ts->idle_sleeptime);
223 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
226 * get_cpu_iowait_time_us - get the total iowait time of a cpu
227 * @cpu: CPU number to query
228 * @last_update_time: variable to store update time in
230 * Return the cummulative iowait time (since boot) for a given
231 * CPU, in microseconds.
233 * This time is measured via accounting rather than sampling,
234 * and is as accurate as ktime_get() is.
236 * This function returns -1 if NOHZ is not enabled.
238 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
240 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
242 if (!tick_nohz_enabled)
245 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
247 return ktime_to_us(ts->iowait_sleeptime);
249 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
252 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
254 * When the next event is more than a tick into the future, stop the idle tick
255 * Called either from the idle loop or from irq_exit() when an idle period was
256 * just interrupted by an interrupt which did not cause a reschedule.
258 void tick_nohz_stop_sched_tick(int inidle)
260 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
261 struct tick_sched *ts;
262 ktime_t last_update, expires, now;
263 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
267 local_irq_save(flags);
269 cpu = smp_processor_id();
270 ts = &per_cpu(tick_cpu_sched, cpu);
273 * Call to tick_nohz_start_idle stops the last_update_time from being
274 * updated. Thus, it must not be called in the event we are called from
275 * irq_exit() with the prior state different than idle.
277 if (!inidle && !ts->inidle)
281 * Set ts->inidle unconditionally. Even if the system did not
282 * switch to NOHZ mode the cpu frequency governers rely on the
283 * update of the idle time accounting in tick_nohz_start_idle().
287 now = tick_nohz_start_idle(cpu, ts);
290 * If this cpu is offline and it is the one which updates
291 * jiffies, then give up the assignment and let it be taken by
292 * the cpu which runs the tick timer next. If we don't drop
293 * this here the jiffies might be stale and do_timer() never
296 if (unlikely(!cpu_online(cpu))) {
297 if (cpu == tick_do_timer_cpu)
298 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
301 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
307 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
308 static int ratelimit;
310 if (ratelimit < 10) {
311 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
312 (unsigned int) local_softirq_pending());
319 /* Read jiffies and the time when jiffies were updated last */
321 seq = read_seqbegin(&xtime_lock);
322 last_update = last_jiffies_update;
323 last_jiffies = jiffies;
324 time_delta = timekeeping_max_deferment();
325 } while (read_seqretry(&xtime_lock, seq));
327 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
328 arch_needs_cpu(cpu) || nohz_ratelimit(cpu)) {
329 next_jiffies = last_jiffies + 1;
332 /* Get the next timer wheel timer */
333 next_jiffies = get_next_timer_interrupt(last_jiffies);
334 delta_jiffies = next_jiffies - last_jiffies;
337 * Do not stop the tick, if we are only one off
338 * or if the cpu is required for rcu
340 if (!ts->tick_stopped && delta_jiffies == 1)
343 /* Schedule the tick, if we are at least one jiffie off */
344 if ((long)delta_jiffies >= 1) {
347 * If this cpu is the one which updates jiffies, then
348 * give up the assignment and let it be taken by the
349 * cpu which runs the tick timer next, which might be
350 * this cpu as well. If we don't drop this here the
351 * jiffies might be stale and do_timer() never
352 * invoked. Keep track of the fact that it was the one
353 * which had the do_timer() duty last. If this cpu is
354 * the one which had the do_timer() duty last, we
355 * limit the sleep time to the timekeeping
356 * max_deferement value which we retrieved
357 * above. Otherwise we can sleep as long as we want.
359 if (cpu == tick_do_timer_cpu) {
360 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
361 ts->do_timer_last = 1;
362 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
363 time_delta = KTIME_MAX;
364 ts->do_timer_last = 0;
365 } else if (!ts->do_timer_last) {
366 time_delta = KTIME_MAX;
370 * calculate the expiry time for the next timer wheel
371 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
372 * that there is no timer pending or at least extremely
373 * far into the future (12 days for HZ=1000). In this
374 * case we set the expiry to the end of time.
376 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
378 * Calculate the time delta for the next timer event.
379 * If the time delta exceeds the maximum time delta
380 * permitted by the current clocksource then adjust
381 * the time delta accordingly to ensure the
382 * clocksource does not wrap.
384 time_delta = min_t(u64, time_delta,
385 tick_period.tv64 * delta_jiffies);
388 if (time_delta < KTIME_MAX)
389 expires = ktime_add_ns(last_update, time_delta);
391 expires.tv64 = KTIME_MAX;
393 if (delta_jiffies > 1)
394 cpumask_set_cpu(cpu, nohz_cpu_mask);
396 /* Skip reprogram of event if its not changed */
397 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
401 * nohz_stop_sched_tick can be called several times before
402 * the nohz_restart_sched_tick is called. This happens when
403 * interrupts arrive which do not cause a reschedule. In the
404 * first call we save the current tick time, so we can restart
405 * the scheduler tick in nohz_restart_sched_tick.
407 if (!ts->tick_stopped) {
408 if (select_nohz_load_balancer(1)) {
410 * sched tick not stopped!
412 cpumask_clear_cpu(cpu, nohz_cpu_mask);
416 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
417 ts->tick_stopped = 1;
418 ts->idle_jiffies = last_jiffies;
425 ts->idle_expires = expires;
428 * If the expiration time == KTIME_MAX, then
429 * in this case we simply stop the tick timer.
431 if (unlikely(expires.tv64 == KTIME_MAX)) {
432 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
433 hrtimer_cancel(&ts->sched_timer);
437 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
438 hrtimer_start(&ts->sched_timer, expires,
439 HRTIMER_MODE_ABS_PINNED);
440 /* Check, if the timer was already in the past */
441 if (hrtimer_active(&ts->sched_timer))
443 } else if (!tick_program_event(expires, 0))
446 * We are past the event already. So we crossed a
447 * jiffie boundary. Update jiffies and raise the
450 tick_do_update_jiffies64(ktime_get());
451 cpumask_clear_cpu(cpu, nohz_cpu_mask);
453 raise_softirq_irqoff(TIMER_SOFTIRQ);
455 ts->next_jiffies = next_jiffies;
456 ts->last_jiffies = last_jiffies;
457 ts->sleep_length = ktime_sub(dev->next_event, now);
459 local_irq_restore(flags);
463 * tick_nohz_get_sleep_length - return the length of the current sleep
465 * Called from power state control code with interrupts disabled
467 ktime_t tick_nohz_get_sleep_length(void)
469 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
471 return ts->sleep_length;
474 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
476 hrtimer_cancel(&ts->sched_timer);
477 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
480 /* Forward the time to expire in the future */
481 hrtimer_forward(&ts->sched_timer, now, tick_period);
483 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
484 hrtimer_start_expires(&ts->sched_timer,
485 HRTIMER_MODE_ABS_PINNED);
486 /* Check, if the timer was already in the past */
487 if (hrtimer_active(&ts->sched_timer))
490 if (!tick_program_event(
491 hrtimer_get_expires(&ts->sched_timer), 0))
494 /* Update jiffies and reread time */
495 tick_do_update_jiffies64(now);
501 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
503 * Restart the idle tick when the CPU is woken up from idle
505 void tick_nohz_restart_sched_tick(void)
507 int cpu = smp_processor_id();
508 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
509 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
515 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
519 tick_nohz_stop_idle(cpu, now);
521 if (!ts->inidle || !ts->tick_stopped) {
531 /* Update jiffies first */
532 select_nohz_load_balancer(0);
533 tick_do_update_jiffies64(now);
534 cpumask_clear_cpu(cpu, nohz_cpu_mask);
536 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
538 * We stopped the tick in idle. Update process times would miss the
539 * time we slept as update_process_times does only a 1 tick
540 * accounting. Enforce that this is accounted to idle !
542 ticks = jiffies - ts->idle_jiffies;
544 * We might be one off. Do not randomly account a huge number of ticks!
546 if (ticks && ticks < LONG_MAX)
547 account_idle_ticks(ticks);
550 touch_softlockup_watchdog();
552 * Cancel the scheduled timer and restore the tick
554 ts->tick_stopped = 0;
555 ts->idle_exittime = now;
557 tick_nohz_restart(ts, now);
562 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
564 hrtimer_forward(&ts->sched_timer, now, tick_period);
565 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
569 * The nohz low res interrupt handler
571 static void tick_nohz_handler(struct clock_event_device *dev)
573 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
574 struct pt_regs *regs = get_irq_regs();
575 int cpu = smp_processor_id();
576 ktime_t now = ktime_get();
578 dev->next_event.tv64 = KTIME_MAX;
581 * Check if the do_timer duty was dropped. We don't care about
582 * concurrency: This happens only when the cpu in charge went
583 * into a long sleep. If two cpus happen to assign themself to
584 * this duty, then the jiffies update is still serialized by
587 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
588 tick_do_timer_cpu = cpu;
590 /* Check, if the jiffies need an update */
591 if (tick_do_timer_cpu == cpu)
592 tick_do_update_jiffies64(now);
595 * When we are idle and the tick is stopped, we have to touch
596 * the watchdog as we might not schedule for a really long
597 * time. This happens on complete idle SMP systems while
598 * waiting on the login prompt. We also increment the "start
599 * of idle" jiffy stamp so the idle accounting adjustment we
600 * do when we go busy again does not account too much ticks.
602 if (ts->tick_stopped) {
603 touch_softlockup_watchdog();
607 update_process_times(user_mode(regs));
608 profile_tick(CPU_PROFILING);
610 while (tick_nohz_reprogram(ts, now)) {
612 tick_do_update_jiffies64(now);
617 * tick_nohz_switch_to_nohz - switch to nohz mode
619 static void tick_nohz_switch_to_nohz(void)
621 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
624 if (!tick_nohz_enabled)
628 if (tick_switch_to_oneshot(tick_nohz_handler)) {
633 ts->nohz_mode = NOHZ_MODE_LOWRES;
636 * Recycle the hrtimer in ts, so we can share the
637 * hrtimer_forward with the highres code.
639 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
640 /* Get the next period */
641 next = tick_init_jiffy_update();
644 hrtimer_set_expires(&ts->sched_timer, next);
645 if (!tick_program_event(next, 0))
647 next = ktime_add(next, tick_period);
651 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
656 * When NOHZ is enabled and the tick is stopped, we need to kick the
657 * tick timer from irq_enter() so that the jiffies update is kept
658 * alive during long running softirqs. That's ugly as hell, but
659 * correctness is key even if we need to fix the offending softirq in
662 * Note, this is different to tick_nohz_restart. We just kick the
663 * timer and do not touch the other magic bits which need to be done
666 static void tick_nohz_kick_tick(int cpu, ktime_t now)
669 /* Switch back to 2.6.27 behaviour */
671 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
675 * Do not touch the tick device, when the next expiry is either
676 * already reached or less/equal than the tick period.
678 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
679 if (delta.tv64 <= tick_period.tv64)
682 tick_nohz_restart(ts, now);
686 static inline void tick_check_nohz(int cpu)
688 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
691 if (!ts->idle_active && !ts->tick_stopped)
695 tick_nohz_stop_idle(cpu, now);
696 if (ts->tick_stopped) {
697 tick_nohz_update_jiffies(now);
698 tick_nohz_kick_tick(cpu, now);
704 static inline void tick_nohz_switch_to_nohz(void) { }
705 static inline void tick_check_nohz(int cpu) { }
710 * Called from irq_enter to notify about the possible interruption of idle()
712 void tick_check_idle(int cpu)
714 tick_check_oneshot_broadcast(cpu);
715 tick_check_nohz(cpu);
719 * High resolution timer specific code
721 #ifdef CONFIG_HIGH_RES_TIMERS
723 * We rearm the timer until we get disabled by the idle code.
724 * Called with interrupts disabled and timer->base->cpu_base->lock held.
726 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
728 struct tick_sched *ts =
729 container_of(timer, struct tick_sched, sched_timer);
730 struct pt_regs *regs = get_irq_regs();
731 ktime_t now = ktime_get();
732 int cpu = smp_processor_id();
736 * Check if the do_timer duty was dropped. We don't care about
737 * concurrency: This happens only when the cpu in charge went
738 * into a long sleep. If two cpus happen to assign themself to
739 * this duty, then the jiffies update is still serialized by
742 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
743 tick_do_timer_cpu = cpu;
746 /* Check, if the jiffies need an update */
747 if (tick_do_timer_cpu == cpu)
748 tick_do_update_jiffies64(now);
751 * Do not call, when we are not in irq context and have
752 * no valid regs pointer
756 * When we are idle and the tick is stopped, we have to touch
757 * the watchdog as we might not schedule for a really long
758 * time. This happens on complete idle SMP systems while
759 * waiting on the login prompt. We also increment the "start of
760 * idle" jiffy stamp so the idle accounting adjustment we do
761 * when we go busy again does not account too much ticks.
763 if (ts->tick_stopped) {
764 touch_softlockup_watchdog();
767 update_process_times(user_mode(regs));
768 profile_tick(CPU_PROFILING);
771 hrtimer_forward(timer, now, tick_period);
773 return HRTIMER_RESTART;
777 * tick_setup_sched_timer - setup the tick emulation timer
779 void tick_setup_sched_timer(void)
781 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
782 ktime_t now = ktime_get();
786 * Emulate tick processing via per-CPU hrtimers:
788 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
789 ts->sched_timer.function = tick_sched_timer;
791 /* Get the next period (per cpu) */
792 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
793 offset = ktime_to_ns(tick_period) >> 1;
794 do_div(offset, num_possible_cpus());
795 offset *= smp_processor_id();
796 hrtimer_add_expires_ns(&ts->sched_timer, offset);
799 hrtimer_forward(&ts->sched_timer, now, tick_period);
800 hrtimer_start_expires(&ts->sched_timer,
801 HRTIMER_MODE_ABS_PINNED);
802 /* Check, if the timer was already in the past */
803 if (hrtimer_active(&ts->sched_timer))
809 if (tick_nohz_enabled)
810 ts->nohz_mode = NOHZ_MODE_HIGHRES;
813 #endif /* HIGH_RES_TIMERS */
815 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
816 void tick_cancel_sched_timer(int cpu)
818 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
820 # ifdef CONFIG_HIGH_RES_TIMERS
821 if (ts->sched_timer.base)
822 hrtimer_cancel(&ts->sched_timer);
825 ts->nohz_mode = NOHZ_MODE_INACTIVE;
830 * Async notification about clocksource changes
832 void tick_clock_notify(void)
836 for_each_possible_cpu(cpu)
837 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
841 * Async notification about clock event changes
843 void tick_oneshot_notify(void)
845 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
847 set_bit(0, &ts->check_clocks);
851 * Check, if a change happened, which makes oneshot possible.
853 * Called cyclic from the hrtimer softirq (driven by the timer
854 * softirq) allow_nohz signals, that we can switch into low-res nohz
855 * mode, because high resolution timers are disabled (either compile
858 int tick_check_oneshot_change(int allow_nohz)
860 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
862 if (!test_and_clear_bit(0, &ts->check_clocks))
865 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
868 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
874 tick_nohz_switch_to_nohz();