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
156 static void update_ts_time_stats(struct tick_sched *ts, ktime_t now)
160 ts->idle_lastupdate = now;
161 if (ts->idle_active) {
162 delta = ktime_sub(now, ts->idle_entrytime);
163 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
167 static void tick_nohz_stop_idle(int cpu, ktime_t now)
169 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
171 update_ts_time_stats(ts, now);
174 sched_clock_idle_wakeup_event(0);
177 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
183 update_ts_time_stats(ts, now);
185 ts->idle_entrytime = now;
187 sched_clock_idle_sleep_event();
192 * get_cpu_idle_time_us - get the total idle time of a cpu
193 * @cpu: CPU number to query
194 * @last_update_time: variable to store update time in
196 * Return the cummulative idle time (since boot) for a given
197 * CPU, in microseconds. The idle time returned includes
198 * the iowait time (unlike what "top" and co report).
200 * This time is measured via accounting rather than sampling,
201 * and is as accurate as ktime_get() is.
203 * This function returns -1 if NOHZ is not enabled.
205 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
207 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
209 if (!tick_nohz_enabled)
213 *last_update_time = ktime_to_us(ts->idle_lastupdate);
215 *last_update_time = ktime_to_us(ktime_get());
217 return ktime_to_us(ts->idle_sleeptime);
219 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
222 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
224 * When the next event is more than a tick into the future, stop the idle tick
225 * Called either from the idle loop or from irq_exit() when an idle period was
226 * just interrupted by an interrupt which did not cause a reschedule.
228 void tick_nohz_stop_sched_tick(int inidle)
230 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
231 struct tick_sched *ts;
232 ktime_t last_update, expires, now;
233 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
237 local_irq_save(flags);
239 cpu = smp_processor_id();
240 ts = &per_cpu(tick_cpu_sched, cpu);
243 * Call to tick_nohz_start_idle stops the last_update_time from being
244 * updated. Thus, it must not be called in the event we are called from
245 * irq_exit() with the prior state different than idle.
247 if (!inidle && !ts->inidle)
251 * Set ts->inidle unconditionally. Even if the system did not
252 * switch to NOHZ mode the cpu frequency governers rely on the
253 * update of the idle time accounting in tick_nohz_start_idle().
257 now = tick_nohz_start_idle(ts);
260 * If this cpu is offline and it is the one which updates
261 * jiffies, then give up the assignment and let it be taken by
262 * the cpu which runs the tick timer next. If we don't drop
263 * this here the jiffies might be stale and do_timer() never
266 if (unlikely(!cpu_online(cpu))) {
267 if (cpu == tick_do_timer_cpu)
268 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
271 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
277 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
278 static int ratelimit;
280 if (ratelimit < 10) {
281 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
282 (unsigned int) local_softirq_pending());
288 if (nohz_ratelimit(cpu))
292 /* Read jiffies and the time when jiffies were updated last */
294 seq = read_seqbegin(&xtime_lock);
295 last_update = last_jiffies_update;
296 last_jiffies = jiffies;
297 time_delta = timekeeping_max_deferment();
298 } while (read_seqretry(&xtime_lock, seq));
300 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
301 arch_needs_cpu(cpu)) {
302 next_jiffies = last_jiffies + 1;
305 /* Get the next timer wheel timer */
306 next_jiffies = get_next_timer_interrupt(last_jiffies);
307 delta_jiffies = next_jiffies - last_jiffies;
310 * Do not stop the tick, if we are only one off
311 * or if the cpu is required for rcu
313 if (!ts->tick_stopped && delta_jiffies == 1)
316 /* Schedule the tick, if we are at least one jiffie off */
317 if ((long)delta_jiffies >= 1) {
320 * If this cpu is the one which updates jiffies, then
321 * give up the assignment and let it be taken by the
322 * cpu which runs the tick timer next, which might be
323 * this cpu as well. If we don't drop this here the
324 * jiffies might be stale and do_timer() never
325 * invoked. Keep track of the fact that it was the one
326 * which had the do_timer() duty last. If this cpu is
327 * the one which had the do_timer() duty last, we
328 * limit the sleep time to the timekeeping
329 * max_deferement value which we retrieved
330 * above. Otherwise we can sleep as long as we want.
332 if (cpu == tick_do_timer_cpu) {
333 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
334 ts->do_timer_last = 1;
335 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
336 time_delta = KTIME_MAX;
337 ts->do_timer_last = 0;
338 } else if (!ts->do_timer_last) {
339 time_delta = KTIME_MAX;
343 * calculate the expiry time for the next timer wheel
344 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
345 * that there is no timer pending or at least extremely
346 * far into the future (12 days for HZ=1000). In this
347 * case we set the expiry to the end of time.
349 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
351 * Calculate the time delta for the next timer event.
352 * If the time delta exceeds the maximum time delta
353 * permitted by the current clocksource then adjust
354 * the time delta accordingly to ensure the
355 * clocksource does not wrap.
357 time_delta = min_t(u64, time_delta,
358 tick_period.tv64 * delta_jiffies);
361 if (time_delta < KTIME_MAX)
362 expires = ktime_add_ns(last_update, time_delta);
364 expires.tv64 = KTIME_MAX;
366 if (delta_jiffies > 1)
367 cpumask_set_cpu(cpu, nohz_cpu_mask);
369 /* Skip reprogram of event if its not changed */
370 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
374 * nohz_stop_sched_tick can be called several times before
375 * the nohz_restart_sched_tick is called. This happens when
376 * interrupts arrive which do not cause a reschedule. In the
377 * first call we save the current tick time, so we can restart
378 * the scheduler tick in nohz_restart_sched_tick.
380 if (!ts->tick_stopped) {
381 if (select_nohz_load_balancer(1)) {
383 * sched tick not stopped!
385 cpumask_clear_cpu(cpu, nohz_cpu_mask);
389 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
390 ts->tick_stopped = 1;
391 ts->idle_jiffies = last_jiffies;
398 ts->idle_expires = expires;
401 * If the expiration time == KTIME_MAX, then
402 * in this case we simply stop the tick timer.
404 if (unlikely(expires.tv64 == KTIME_MAX)) {
405 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
406 hrtimer_cancel(&ts->sched_timer);
410 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
411 hrtimer_start(&ts->sched_timer, expires,
412 HRTIMER_MODE_ABS_PINNED);
413 /* Check, if the timer was already in the past */
414 if (hrtimer_active(&ts->sched_timer))
416 } else if (!tick_program_event(expires, 0))
419 * We are past the event already. So we crossed a
420 * jiffie boundary. Update jiffies and raise the
423 tick_do_update_jiffies64(ktime_get());
424 cpumask_clear_cpu(cpu, nohz_cpu_mask);
426 raise_softirq_irqoff(TIMER_SOFTIRQ);
428 ts->next_jiffies = next_jiffies;
429 ts->last_jiffies = last_jiffies;
430 ts->sleep_length = ktime_sub(dev->next_event, now);
432 local_irq_restore(flags);
436 * tick_nohz_get_sleep_length - return the length of the current sleep
438 * Called from power state control code with interrupts disabled
440 ktime_t tick_nohz_get_sleep_length(void)
442 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
444 return ts->sleep_length;
447 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
449 hrtimer_cancel(&ts->sched_timer);
450 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
453 /* Forward the time to expire in the future */
454 hrtimer_forward(&ts->sched_timer, now, tick_period);
456 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
457 hrtimer_start_expires(&ts->sched_timer,
458 HRTIMER_MODE_ABS_PINNED);
459 /* Check, if the timer was already in the past */
460 if (hrtimer_active(&ts->sched_timer))
463 if (!tick_program_event(
464 hrtimer_get_expires(&ts->sched_timer), 0))
467 /* Update jiffies and reread time */
468 tick_do_update_jiffies64(now);
474 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
476 * Restart the idle tick when the CPU is woken up from idle
478 void tick_nohz_restart_sched_tick(void)
480 int cpu = smp_processor_id();
481 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
482 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
488 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
492 tick_nohz_stop_idle(cpu, now);
494 if (!ts->inidle || !ts->tick_stopped) {
504 /* Update jiffies first */
505 select_nohz_load_balancer(0);
506 tick_do_update_jiffies64(now);
507 cpumask_clear_cpu(cpu, nohz_cpu_mask);
509 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
511 * We stopped the tick in idle. Update process times would miss the
512 * time we slept as update_process_times does only a 1 tick
513 * accounting. Enforce that this is accounted to idle !
515 ticks = jiffies - ts->idle_jiffies;
517 * We might be one off. Do not randomly account a huge number of ticks!
519 if (ticks && ticks < LONG_MAX)
520 account_idle_ticks(ticks);
523 touch_softlockup_watchdog();
525 * Cancel the scheduled timer and restore the tick
527 ts->tick_stopped = 0;
528 ts->idle_exittime = now;
530 tick_nohz_restart(ts, now);
535 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
537 hrtimer_forward(&ts->sched_timer, now, tick_period);
538 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
542 * The nohz low res interrupt handler
544 static void tick_nohz_handler(struct clock_event_device *dev)
546 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
547 struct pt_regs *regs = get_irq_regs();
548 int cpu = smp_processor_id();
549 ktime_t now = ktime_get();
551 dev->next_event.tv64 = KTIME_MAX;
554 * Check if the do_timer duty was dropped. We don't care about
555 * concurrency: This happens only when the cpu in charge went
556 * into a long sleep. If two cpus happen to assign themself to
557 * this duty, then the jiffies update is still serialized by
560 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
561 tick_do_timer_cpu = cpu;
563 /* Check, if the jiffies need an update */
564 if (tick_do_timer_cpu == cpu)
565 tick_do_update_jiffies64(now);
568 * When we are idle and the tick is stopped, we have to touch
569 * the watchdog as we might not schedule for a really long
570 * time. This happens on complete idle SMP systems while
571 * waiting on the login prompt. We also increment the "start
572 * of idle" jiffy stamp so the idle accounting adjustment we
573 * do when we go busy again does not account too much ticks.
575 if (ts->tick_stopped) {
576 touch_softlockup_watchdog();
580 update_process_times(user_mode(regs));
581 profile_tick(CPU_PROFILING);
583 while (tick_nohz_reprogram(ts, now)) {
585 tick_do_update_jiffies64(now);
590 * tick_nohz_switch_to_nohz - switch to nohz mode
592 static void tick_nohz_switch_to_nohz(void)
594 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
597 if (!tick_nohz_enabled)
601 if (tick_switch_to_oneshot(tick_nohz_handler)) {
606 ts->nohz_mode = NOHZ_MODE_LOWRES;
609 * Recycle the hrtimer in ts, so we can share the
610 * hrtimer_forward with the highres code.
612 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
613 /* Get the next period */
614 next = tick_init_jiffy_update();
617 hrtimer_set_expires(&ts->sched_timer, next);
618 if (!tick_program_event(next, 0))
620 next = ktime_add(next, tick_period);
624 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
629 * When NOHZ is enabled and the tick is stopped, we need to kick the
630 * tick timer from irq_enter() so that the jiffies update is kept
631 * alive during long running softirqs. That's ugly as hell, but
632 * correctness is key even if we need to fix the offending softirq in
635 * Note, this is different to tick_nohz_restart. We just kick the
636 * timer and do not touch the other magic bits which need to be done
639 static void tick_nohz_kick_tick(int cpu, ktime_t now)
642 /* Switch back to 2.6.27 behaviour */
644 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
648 * Do not touch the tick device, when the next expiry is either
649 * already reached or less/equal than the tick period.
651 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
652 if (delta.tv64 <= tick_period.tv64)
655 tick_nohz_restart(ts, now);
659 static inline void tick_check_nohz(int cpu)
661 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
664 if (!ts->idle_active && !ts->tick_stopped)
668 tick_nohz_stop_idle(cpu, now);
669 if (ts->tick_stopped) {
670 tick_nohz_update_jiffies(now);
671 tick_nohz_kick_tick(cpu, now);
677 static inline void tick_nohz_switch_to_nohz(void) { }
678 static inline void tick_check_nohz(int cpu) { }
683 * Called from irq_enter to notify about the possible interruption of idle()
685 void tick_check_idle(int cpu)
687 tick_check_oneshot_broadcast(cpu);
688 tick_check_nohz(cpu);
692 * High resolution timer specific code
694 #ifdef CONFIG_HIGH_RES_TIMERS
696 * We rearm the timer until we get disabled by the idle code.
697 * Called with interrupts disabled and timer->base->cpu_base->lock held.
699 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
701 struct tick_sched *ts =
702 container_of(timer, struct tick_sched, sched_timer);
703 struct pt_regs *regs = get_irq_regs();
704 ktime_t now = ktime_get();
705 int cpu = smp_processor_id();
709 * Check if the do_timer duty was dropped. We don't care about
710 * concurrency: This happens only when the cpu in charge went
711 * into a long sleep. If two cpus happen to assign themself to
712 * this duty, then the jiffies update is still serialized by
715 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
716 tick_do_timer_cpu = cpu;
719 /* Check, if the jiffies need an update */
720 if (tick_do_timer_cpu == cpu)
721 tick_do_update_jiffies64(now);
724 * Do not call, when we are not in irq context and have
725 * no valid regs pointer
729 * When we are idle and the tick is stopped, we have to touch
730 * the watchdog as we might not schedule for a really long
731 * time. This happens on complete idle SMP systems while
732 * waiting on the login prompt. We also increment the "start of
733 * idle" jiffy stamp so the idle accounting adjustment we do
734 * when we go busy again does not account too much ticks.
736 if (ts->tick_stopped) {
737 touch_softlockup_watchdog();
740 update_process_times(user_mode(regs));
741 profile_tick(CPU_PROFILING);
744 hrtimer_forward(timer, now, tick_period);
746 return HRTIMER_RESTART;
750 * tick_setup_sched_timer - setup the tick emulation timer
752 void tick_setup_sched_timer(void)
754 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
755 ktime_t now = ktime_get();
759 * Emulate tick processing via per-CPU hrtimers:
761 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
762 ts->sched_timer.function = tick_sched_timer;
764 /* Get the next period (per cpu) */
765 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
766 offset = ktime_to_ns(tick_period) >> 1;
767 do_div(offset, num_possible_cpus());
768 offset *= smp_processor_id();
769 hrtimer_add_expires_ns(&ts->sched_timer, offset);
772 hrtimer_forward(&ts->sched_timer, now, tick_period);
773 hrtimer_start_expires(&ts->sched_timer,
774 HRTIMER_MODE_ABS_PINNED);
775 /* Check, if the timer was already in the past */
776 if (hrtimer_active(&ts->sched_timer))
782 if (tick_nohz_enabled)
783 ts->nohz_mode = NOHZ_MODE_HIGHRES;
786 #endif /* HIGH_RES_TIMERS */
788 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
789 void tick_cancel_sched_timer(int cpu)
791 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
793 # ifdef CONFIG_HIGH_RES_TIMERS
794 if (ts->sched_timer.base)
795 hrtimer_cancel(&ts->sched_timer);
798 ts->nohz_mode = NOHZ_MODE_INACTIVE;
803 * Async notification about clocksource changes
805 void tick_clock_notify(void)
809 for_each_possible_cpu(cpu)
810 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
814 * Async notification about clock event changes
816 void tick_oneshot_notify(void)
818 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
820 set_bit(0, &ts->check_clocks);
824 * Check, if a change happened, which makes oneshot possible.
826 * Called cyclic from the hrtimer softirq (driven by the timer
827 * softirq) allow_nohz signals, that we can switch into low-res nohz
828 * mode, because high resolution timers are disabled (either compile
831 int tick_check_oneshot_change(int allow_nohz)
833 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
835 if (!test_and_clear_bit(0, &ts->check_clocks))
838 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
841 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
847 tick_nohz_switch_to_nohz();