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(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() > 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(ts, now, NULL);
181 sched_clock_idle_wakeup_event(0);
184 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
190 update_ts_time_stats(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(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(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(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());
318 if (nohz_ratelimit(cpu))
322 /* Read jiffies and the time when jiffies were updated last */
324 seq = read_seqbegin(&xtime_lock);
325 last_update = last_jiffies_update;
326 last_jiffies = jiffies;
327 time_delta = timekeeping_max_deferment();
328 } while (read_seqretry(&xtime_lock, seq));
330 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
331 arch_needs_cpu(cpu)) {
332 next_jiffies = last_jiffies + 1;
335 /* Get the next timer wheel timer */
336 next_jiffies = get_next_timer_interrupt(last_jiffies);
337 delta_jiffies = next_jiffies - last_jiffies;
340 * Do not stop the tick, if we are only one off
341 * or if the cpu is required for rcu
343 if (!ts->tick_stopped && delta_jiffies == 1)
346 /* Schedule the tick, if we are at least one jiffie off */
347 if ((long)delta_jiffies >= 1) {
350 * If this cpu is the one which updates jiffies, then
351 * give up the assignment and let it be taken by the
352 * cpu which runs the tick timer next, which might be
353 * this cpu as well. If we don't drop this here the
354 * jiffies might be stale and do_timer() never
355 * invoked. Keep track of the fact that it was the one
356 * which had the do_timer() duty last. If this cpu is
357 * the one which had the do_timer() duty last, we
358 * limit the sleep time to the timekeeping
359 * max_deferement value which we retrieved
360 * above. Otherwise we can sleep as long as we want.
362 if (cpu == tick_do_timer_cpu) {
363 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
364 ts->do_timer_last = 1;
365 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
366 time_delta = KTIME_MAX;
367 ts->do_timer_last = 0;
368 } else if (!ts->do_timer_last) {
369 time_delta = KTIME_MAX;
373 * calculate the expiry time for the next timer wheel
374 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
375 * that there is no timer pending or at least extremely
376 * far into the future (12 days for HZ=1000). In this
377 * case we set the expiry to the end of time.
379 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
381 * Calculate the time delta for the next timer event.
382 * If the time delta exceeds the maximum time delta
383 * permitted by the current clocksource then adjust
384 * the time delta accordingly to ensure the
385 * clocksource does not wrap.
387 time_delta = min_t(u64, time_delta,
388 tick_period.tv64 * delta_jiffies);
391 if (time_delta < KTIME_MAX)
392 expires = ktime_add_ns(last_update, time_delta);
394 expires.tv64 = KTIME_MAX;
396 if (delta_jiffies > 1)
397 cpumask_set_cpu(cpu, nohz_cpu_mask);
399 /* Skip reprogram of event if its not changed */
400 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
404 * nohz_stop_sched_tick can be called several times before
405 * the nohz_restart_sched_tick is called. This happens when
406 * interrupts arrive which do not cause a reschedule. In the
407 * first call we save the current tick time, so we can restart
408 * the scheduler tick in nohz_restart_sched_tick.
410 if (!ts->tick_stopped) {
411 select_nohz_load_balancer(1);
413 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
414 ts->tick_stopped = 1;
415 ts->idle_jiffies = last_jiffies;
422 ts->idle_expires = expires;
425 * If the expiration time == KTIME_MAX, then
426 * in this case we simply stop the tick timer.
428 if (unlikely(expires.tv64 == KTIME_MAX)) {
429 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
430 hrtimer_cancel(&ts->sched_timer);
434 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
435 hrtimer_start(&ts->sched_timer, expires,
436 HRTIMER_MODE_ABS_PINNED);
437 /* Check, if the timer was already in the past */
438 if (hrtimer_active(&ts->sched_timer))
440 } else if (!tick_program_event(expires, 0))
443 * We are past the event already. So we crossed a
444 * jiffie boundary. Update jiffies and raise the
447 tick_do_update_jiffies64(ktime_get());
448 cpumask_clear_cpu(cpu, nohz_cpu_mask);
450 raise_softirq_irqoff(TIMER_SOFTIRQ);
452 ts->next_jiffies = next_jiffies;
453 ts->last_jiffies = last_jiffies;
454 ts->sleep_length = ktime_sub(dev->next_event, now);
456 local_irq_restore(flags);
460 * tick_nohz_get_sleep_length - return the length of the current sleep
462 * Called from power state control code with interrupts disabled
464 ktime_t tick_nohz_get_sleep_length(void)
466 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
468 return ts->sleep_length;
471 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
473 hrtimer_cancel(&ts->sched_timer);
474 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
477 /* Forward the time to expire in the future */
478 hrtimer_forward(&ts->sched_timer, now, tick_period);
480 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
481 hrtimer_start_expires(&ts->sched_timer,
482 HRTIMER_MODE_ABS_PINNED);
483 /* Check, if the timer was already in the past */
484 if (hrtimer_active(&ts->sched_timer))
487 if (!tick_program_event(
488 hrtimer_get_expires(&ts->sched_timer), 0))
491 /* Update jiffies and reread time */
492 tick_do_update_jiffies64(now);
498 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
500 * Restart the idle tick when the CPU is woken up from idle
502 void tick_nohz_restart_sched_tick(void)
504 int cpu = smp_processor_id();
505 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
506 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
512 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
516 tick_nohz_stop_idle(cpu, now);
518 if (!ts->inidle || !ts->tick_stopped) {
528 /* Update jiffies first */
529 select_nohz_load_balancer(0);
530 tick_do_update_jiffies64(now);
531 cpumask_clear_cpu(cpu, nohz_cpu_mask);
533 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
535 * We stopped the tick in idle. Update process times would miss the
536 * time we slept as update_process_times does only a 1 tick
537 * accounting. Enforce that this is accounted to idle !
539 ticks = jiffies - ts->idle_jiffies;
541 * We might be one off. Do not randomly account a huge number of ticks!
543 if (ticks && ticks < LONG_MAX)
544 account_idle_ticks(ticks);
547 touch_softlockup_watchdog();
549 * Cancel the scheduled timer and restore the tick
551 ts->tick_stopped = 0;
552 ts->idle_exittime = now;
554 tick_nohz_restart(ts, now);
559 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
561 hrtimer_forward(&ts->sched_timer, now, tick_period);
562 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
566 * The nohz low res interrupt handler
568 static void tick_nohz_handler(struct clock_event_device *dev)
570 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
571 struct pt_regs *regs = get_irq_regs();
572 int cpu = smp_processor_id();
573 ktime_t now = ktime_get();
575 dev->next_event.tv64 = KTIME_MAX;
578 * Check if the do_timer duty was dropped. We don't care about
579 * concurrency: This happens only when the cpu in charge went
580 * into a long sleep. If two cpus happen to assign themself to
581 * this duty, then the jiffies update is still serialized by
584 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
585 tick_do_timer_cpu = cpu;
587 /* Check, if the jiffies need an update */
588 if (tick_do_timer_cpu == cpu)
589 tick_do_update_jiffies64(now);
592 * When we are idle and the tick is stopped, we have to touch
593 * the watchdog as we might not schedule for a really long
594 * time. This happens on complete idle SMP systems while
595 * waiting on the login prompt. We also increment the "start
596 * of idle" jiffy stamp so the idle accounting adjustment we
597 * do when we go busy again does not account too much ticks.
599 if (ts->tick_stopped) {
600 touch_softlockup_watchdog();
604 update_process_times(user_mode(regs));
605 profile_tick(CPU_PROFILING);
607 while (tick_nohz_reprogram(ts, now)) {
609 tick_do_update_jiffies64(now);
614 * tick_nohz_switch_to_nohz - switch to nohz mode
616 static void tick_nohz_switch_to_nohz(void)
618 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
621 if (!tick_nohz_enabled)
625 if (tick_switch_to_oneshot(tick_nohz_handler)) {
630 ts->nohz_mode = NOHZ_MODE_LOWRES;
633 * Recycle the hrtimer in ts, so we can share the
634 * hrtimer_forward with the highres code.
636 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
637 /* Get the next period */
638 next = tick_init_jiffy_update();
641 hrtimer_set_expires(&ts->sched_timer, next);
642 if (!tick_program_event(next, 0))
644 next = ktime_add(next, tick_period);
648 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
653 * When NOHZ is enabled and the tick is stopped, we need to kick the
654 * tick timer from irq_enter() so that the jiffies update is kept
655 * alive during long running softirqs. That's ugly as hell, but
656 * correctness is key even if we need to fix the offending softirq in
659 * Note, this is different to tick_nohz_restart. We just kick the
660 * timer and do not touch the other magic bits which need to be done
663 static void tick_nohz_kick_tick(int cpu, ktime_t now)
666 /* Switch back to 2.6.27 behaviour */
668 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
672 * Do not touch the tick device, when the next expiry is either
673 * already reached or less/equal than the tick period.
675 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
676 if (delta.tv64 <= tick_period.tv64)
679 tick_nohz_restart(ts, now);
683 static inline void tick_check_nohz(int cpu)
685 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
688 if (!ts->idle_active && !ts->tick_stopped)
692 tick_nohz_stop_idle(cpu, now);
693 if (ts->tick_stopped) {
694 tick_nohz_update_jiffies(now);
695 tick_nohz_kick_tick(cpu, now);
701 static inline void tick_nohz_switch_to_nohz(void) { }
702 static inline void tick_check_nohz(int cpu) { }
707 * Called from irq_enter to notify about the possible interruption of idle()
709 void tick_check_idle(int cpu)
711 tick_check_oneshot_broadcast(cpu);
712 tick_check_nohz(cpu);
716 * High resolution timer specific code
718 #ifdef CONFIG_HIGH_RES_TIMERS
720 * We rearm the timer until we get disabled by the idle code.
721 * Called with interrupts disabled and timer->base->cpu_base->lock held.
723 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
725 struct tick_sched *ts =
726 container_of(timer, struct tick_sched, sched_timer);
727 struct pt_regs *regs = get_irq_regs();
728 ktime_t now = ktime_get();
729 int cpu = smp_processor_id();
733 * Check if the do_timer duty was dropped. We don't care about
734 * concurrency: This happens only when the cpu in charge went
735 * into a long sleep. If two cpus happen to assign themself to
736 * this duty, then the jiffies update is still serialized by
739 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
740 tick_do_timer_cpu = cpu;
743 /* Check, if the jiffies need an update */
744 if (tick_do_timer_cpu == cpu)
745 tick_do_update_jiffies64(now);
748 * Do not call, when we are not in irq context and have
749 * no valid regs pointer
753 * When we are idle and the tick is stopped, we have to touch
754 * the watchdog as we might not schedule for a really long
755 * time. This happens on complete idle SMP systems while
756 * waiting on the login prompt. We also increment the "start of
757 * idle" jiffy stamp so the idle accounting adjustment we do
758 * when we go busy again does not account too much ticks.
760 if (ts->tick_stopped) {
761 touch_softlockup_watchdog();
764 update_process_times(user_mode(regs));
765 profile_tick(CPU_PROFILING);
768 hrtimer_forward(timer, now, tick_period);
770 return HRTIMER_RESTART;
774 * tick_setup_sched_timer - setup the tick emulation timer
776 void tick_setup_sched_timer(void)
778 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
779 ktime_t now = ktime_get();
783 * Emulate tick processing via per-CPU hrtimers:
785 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
786 ts->sched_timer.function = tick_sched_timer;
788 /* Get the next period (per cpu) */
789 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
790 offset = ktime_to_ns(tick_period) >> 1;
791 do_div(offset, num_possible_cpus());
792 offset *= smp_processor_id();
793 hrtimer_add_expires_ns(&ts->sched_timer, offset);
796 hrtimer_forward(&ts->sched_timer, now, tick_period);
797 hrtimer_start_expires(&ts->sched_timer,
798 HRTIMER_MODE_ABS_PINNED);
799 /* Check, if the timer was already in the past */
800 if (hrtimer_active(&ts->sched_timer))
806 if (tick_nohz_enabled)
807 ts->nohz_mode = NOHZ_MODE_HIGHRES;
810 #endif /* HIGH_RES_TIMERS */
812 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
813 void tick_cancel_sched_timer(int cpu)
815 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
817 # ifdef CONFIG_HIGH_RES_TIMERS
818 if (ts->sched_timer.base)
819 hrtimer_cancel(&ts->sched_timer);
822 ts->nohz_mode = NOHZ_MODE_INACTIVE;
827 * Async notification about clocksource changes
829 void tick_clock_notify(void)
833 for_each_possible_cpu(cpu)
834 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
838 * Async notification about clock event changes
840 void tick_oneshot_notify(void)
842 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
844 set_bit(0, &ts->check_clocks);
848 * Check, if a change happened, which makes oneshot possible.
850 * Called cyclic from the hrtimer softirq (driven by the timer
851 * softirq) allow_nohz signals, that we can switch into low-res nohz
852 * mode, because high resolution timers are disabled (either compile
855 int tick_check_oneshot_change(int allow_nohz)
857 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
859 if (!test_and_clear_bit(0, &ts->check_clocks))
862 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
865 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
871 tick_nohz_switch_to_nohz();