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>
23 #include <linux/irq_work.h>
25 #include <asm/irq_regs.h>
27 #include "tick-internal.h"
30 * Per cpu nohz control structure
32 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
35 * The time, when the last jiffy update happened. Protected by jiffies_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 jiffies_lock:
55 delta = ktime_sub(now, last_jiffies_update);
56 if (delta.tv64 < tick_period.tv64)
59 /* Reevalute with jiffies_lock held */
60 write_seqlock(&jiffies_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(&jiffies_lock);
87 * Initialize and return retrieve the jiffies update.
89 static ktime_t tick_init_jiffy_update(void)
93 write_seqlock(&jiffies_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(&jiffies_lock);
103 static void tick_sched_do_timer(ktime_t now)
105 int cpu = smp_processor_id();
107 #ifdef CONFIG_NO_HZ_COMMON
109 * Check if the do_timer duty was dropped. We don't care about
110 * concurrency: This happens only when the cpu in charge went
111 * into a long sleep. If two cpus happen to assign themself to
112 * this duty, then the jiffies update is still serialized by
115 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
116 && !tick_nohz_full_cpu(cpu))
117 tick_do_timer_cpu = cpu;
120 /* Check, if the jiffies need an update */
121 if (tick_do_timer_cpu == cpu)
122 tick_do_update_jiffies64(now);
125 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
127 #ifdef CONFIG_NO_HZ_COMMON
129 * When we are idle and the tick is stopped, we have to touch
130 * the watchdog as we might not schedule for a really long
131 * time. This happens on complete idle SMP systems while
132 * waiting on the login prompt. We also increment the "start of
133 * idle" jiffy stamp so the idle accounting adjustment we do
134 * when we go busy again does not account too much ticks.
136 if (ts->tick_stopped) {
137 touch_softlockup_watchdog();
138 if (is_idle_task(current))
142 update_process_times(user_mode(regs));
143 profile_tick(CPU_PROFILING);
146 #ifdef CONFIG_NO_HZ_FULL
147 static cpumask_var_t nohz_full_mask;
148 bool have_nohz_full_mask;
150 int tick_nohz_full_cpu(int cpu)
152 if (!have_nohz_full_mask)
155 return cpumask_test_cpu(cpu, nohz_full_mask);
158 /* Parse the boot-time nohz CPU list from the kernel parameters. */
159 static int __init tick_nohz_full_setup(char *str)
163 alloc_bootmem_cpumask_var(&nohz_full_mask);
164 if (cpulist_parse(str, nohz_full_mask) < 0) {
165 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
169 cpu = smp_processor_id();
170 if (cpumask_test_cpu(cpu, nohz_full_mask)) {
171 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
172 cpumask_clear_cpu(cpu, nohz_full_mask);
174 have_nohz_full_mask = true;
178 __setup("nohz_full=", tick_nohz_full_setup);
180 static int __cpuinit tick_nohz_cpu_down_callback(struct notifier_block *nfb,
181 unsigned long action,
184 unsigned int cpu = (unsigned long)hcpu;
186 switch (action & ~CPU_TASKS_FROZEN) {
187 case CPU_DOWN_PREPARE:
189 * If we handle the timekeeping duty for full dynticks CPUs,
190 * we can't safely shutdown that CPU.
192 if (have_nohz_full_mask && tick_do_timer_cpu == cpu)
200 * Worst case string length in chunks of CPU range seems 2 steps
201 * separations: 0,2,4,6,...
202 * This is NR_CPUS + sizeof('\0')
204 static char __initdata nohz_full_buf[NR_CPUS + 1];
206 static int __init init_tick_nohz_full(void)
208 if (have_nohz_full_mask)
209 cpu_notifier(tick_nohz_cpu_down_callback, 0);
211 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask);
212 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
216 core_initcall(init_tick_nohz_full);
218 #define have_nohz_full_mask (0)
222 * NOHZ - aka dynamic tick functionality
224 #ifdef CONFIG_NO_HZ_COMMON
228 int tick_nohz_enabled __read_mostly = 1;
231 * Enable / Disable tickless mode
233 static int __init setup_tick_nohz(char *str)
235 if (!strcmp(str, "off"))
236 tick_nohz_enabled = 0;
237 else if (!strcmp(str, "on"))
238 tick_nohz_enabled = 1;
244 __setup("nohz=", setup_tick_nohz);
247 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
249 * Called from interrupt entry when the CPU was idle
251 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
252 * must be updated. Otherwise an interrupt handler could use a stale jiffy
253 * value. We do this unconditionally on any cpu, as we don't know whether the
254 * cpu, which has the update task assigned is in a long sleep.
256 static void tick_nohz_update_jiffies(ktime_t now)
258 int cpu = smp_processor_id();
259 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
262 ts->idle_waketime = now;
264 local_irq_save(flags);
265 tick_do_update_jiffies64(now);
266 local_irq_restore(flags);
268 touch_softlockup_watchdog();
272 * Updates the per cpu time idle statistics counters
275 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
279 if (ts->idle_active) {
280 delta = ktime_sub(now, ts->idle_entrytime);
281 if (nr_iowait_cpu(cpu) > 0)
282 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
284 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
285 ts->idle_entrytime = now;
288 if (last_update_time)
289 *last_update_time = ktime_to_us(now);
293 static void tick_nohz_stop_idle(int cpu, ktime_t now)
295 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
297 update_ts_time_stats(cpu, ts, now, NULL);
300 sched_clock_idle_wakeup_event(0);
303 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
305 ktime_t now = ktime_get();
307 ts->idle_entrytime = now;
309 sched_clock_idle_sleep_event();
314 * get_cpu_idle_time_us - get the total idle time of a cpu
315 * @cpu: CPU number to query
316 * @last_update_time: variable to store update time in. Do not update
319 * Return the cummulative idle time (since boot) for a given
320 * CPU, in microseconds.
322 * This time is measured via accounting rather than sampling,
323 * and is as accurate as ktime_get() is.
325 * This function returns -1 if NOHZ is not enabled.
327 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
329 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
332 if (!tick_nohz_enabled)
336 if (last_update_time) {
337 update_ts_time_stats(cpu, ts, now, last_update_time);
338 idle = ts->idle_sleeptime;
340 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
341 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
343 idle = ktime_add(ts->idle_sleeptime, delta);
345 idle = ts->idle_sleeptime;
349 return ktime_to_us(idle);
352 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
355 * get_cpu_iowait_time_us - get the total iowait time of a cpu
356 * @cpu: CPU number to query
357 * @last_update_time: variable to store update time in. Do not update
360 * Return the cummulative iowait time (since boot) for a given
361 * CPU, in microseconds.
363 * This time is measured via accounting rather than sampling,
364 * and is as accurate as ktime_get() is.
366 * This function returns -1 if NOHZ is not enabled.
368 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
370 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
373 if (!tick_nohz_enabled)
377 if (last_update_time) {
378 update_ts_time_stats(cpu, ts, now, last_update_time);
379 iowait = ts->iowait_sleeptime;
381 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
382 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
384 iowait = ktime_add(ts->iowait_sleeptime, delta);
386 iowait = ts->iowait_sleeptime;
390 return ktime_to_us(iowait);
392 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
394 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
395 ktime_t now, int cpu)
397 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
398 ktime_t last_update, expires, ret = { .tv64 = 0 };
399 unsigned long rcu_delta_jiffies;
400 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
403 /* Read jiffies and the time when jiffies were updated last */
405 seq = read_seqbegin(&jiffies_lock);
406 last_update = last_jiffies_update;
407 last_jiffies = jiffies;
408 time_delta = timekeeping_max_deferment();
409 } while (read_seqretry(&jiffies_lock, seq));
411 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
412 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
413 next_jiffies = last_jiffies + 1;
416 /* Get the next timer wheel timer */
417 next_jiffies = get_next_timer_interrupt(last_jiffies);
418 delta_jiffies = next_jiffies - last_jiffies;
419 if (rcu_delta_jiffies < delta_jiffies) {
420 next_jiffies = last_jiffies + rcu_delta_jiffies;
421 delta_jiffies = rcu_delta_jiffies;
425 * Do not stop the tick, if we are only one off
426 * or if the cpu is required for rcu
428 if (!ts->tick_stopped && delta_jiffies == 1)
431 /* Schedule the tick, if we are at least one jiffie off */
432 if ((long)delta_jiffies >= 1) {
435 * If this cpu is the one which updates jiffies, then
436 * give up the assignment and let it be taken by the
437 * cpu which runs the tick timer next, which might be
438 * this cpu as well. If we don't drop this here the
439 * jiffies might be stale and do_timer() never
440 * invoked. Keep track of the fact that it was the one
441 * which had the do_timer() duty last. If this cpu is
442 * the one which had the do_timer() duty last, we
443 * limit the sleep time to the timekeeping
444 * max_deferement value which we retrieved
445 * above. Otherwise we can sleep as long as we want.
447 if (cpu == tick_do_timer_cpu) {
448 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
449 ts->do_timer_last = 1;
450 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
451 time_delta = KTIME_MAX;
452 ts->do_timer_last = 0;
453 } else if (!ts->do_timer_last) {
454 time_delta = KTIME_MAX;
458 * calculate the expiry time for the next timer wheel
459 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
460 * that there is no timer pending or at least extremely
461 * far into the future (12 days for HZ=1000). In this
462 * case we set the expiry to the end of time.
464 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
466 * Calculate the time delta for the next timer event.
467 * If the time delta exceeds the maximum time delta
468 * permitted by the current clocksource then adjust
469 * the time delta accordingly to ensure the
470 * clocksource does not wrap.
472 time_delta = min_t(u64, time_delta,
473 tick_period.tv64 * delta_jiffies);
476 if (time_delta < KTIME_MAX)
477 expires = ktime_add_ns(last_update, time_delta);
479 expires.tv64 = KTIME_MAX;
481 /* Skip reprogram of event if its not changed */
482 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
488 * nohz_stop_sched_tick can be called several times before
489 * the nohz_restart_sched_tick is called. This happens when
490 * interrupts arrive which do not cause a reschedule. In the
491 * first call we save the current tick time, so we can restart
492 * the scheduler tick in nohz_restart_sched_tick.
494 if (!ts->tick_stopped) {
495 nohz_balance_enter_idle(cpu);
496 calc_load_enter_idle();
498 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
499 ts->tick_stopped = 1;
503 * If the expiration time == KTIME_MAX, then
504 * in this case we simply stop the tick timer.
506 if (unlikely(expires.tv64 == KTIME_MAX)) {
507 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
508 hrtimer_cancel(&ts->sched_timer);
512 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
513 hrtimer_start(&ts->sched_timer, expires,
514 HRTIMER_MODE_ABS_PINNED);
515 /* Check, if the timer was already in the past */
516 if (hrtimer_active(&ts->sched_timer))
518 } else if (!tick_program_event(expires, 0))
521 * We are past the event already. So we crossed a
522 * jiffie boundary. Update jiffies and raise the
525 tick_do_update_jiffies64(ktime_get());
527 raise_softirq_irqoff(TIMER_SOFTIRQ);
529 ts->next_jiffies = next_jiffies;
530 ts->last_jiffies = last_jiffies;
531 ts->sleep_length = ktime_sub(dev->next_event, now);
536 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
539 * If this cpu is offline and it is the one which updates
540 * jiffies, then give up the assignment and let it be taken by
541 * the cpu which runs the tick timer next. If we don't drop
542 * this here the jiffies might be stale and do_timer() never
545 if (unlikely(!cpu_online(cpu))) {
546 if (cpu == tick_do_timer_cpu)
547 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
550 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
556 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
557 static int ratelimit;
559 if (ratelimit < 10 &&
560 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
561 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
562 (unsigned int) local_softirq_pending());
568 if (have_nohz_full_mask) {
570 * Keep the tick alive to guarantee timekeeping progression
571 * if there are full dynticks CPUs around
573 if (tick_do_timer_cpu == cpu)
576 * Boot safety: make sure the timekeeping duty has been
577 * assigned before entering dyntick-idle mode,
579 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
586 static void __tick_nohz_idle_enter(struct tick_sched *ts)
588 ktime_t now, expires;
589 int cpu = smp_processor_id();
591 now = tick_nohz_start_idle(cpu, ts);
593 if (can_stop_idle_tick(cpu, ts)) {
594 int was_stopped = ts->tick_stopped;
598 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
599 if (expires.tv64 > 0LL) {
601 ts->idle_expires = expires;
604 if (!was_stopped && ts->tick_stopped)
605 ts->idle_jiffies = ts->last_jiffies;
610 * tick_nohz_idle_enter - stop the idle tick from the idle task
612 * When the next event is more than a tick into the future, stop the idle tick
613 * Called when we start the idle loop.
615 * The arch is responsible of calling:
617 * - rcu_idle_enter() after its last use of RCU before the CPU is put
619 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
621 void tick_nohz_idle_enter(void)
623 struct tick_sched *ts;
625 WARN_ON_ONCE(irqs_disabled());
628 * Update the idle state in the scheduler domain hierarchy
629 * when tick_nohz_stop_sched_tick() is called from the idle loop.
630 * State will be updated to busy during the first busy tick after
633 set_cpu_sd_state_idle();
637 ts = &__get_cpu_var(tick_cpu_sched);
639 * set ts->inidle unconditionally. even if the system did not
640 * switch to nohz mode the cpu frequency governers rely on the
641 * update of the idle time accounting in tick_nohz_start_idle().
644 __tick_nohz_idle_enter(ts);
648 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
651 * tick_nohz_irq_exit - update next tick event from interrupt exit
653 * When an interrupt fires while we are idle and it doesn't cause
654 * a reschedule, it may still add, modify or delete a timer, enqueue
655 * an RCU callback, etc...
656 * So we need to re-calculate and reprogram the next tick event.
658 void tick_nohz_irq_exit(void)
660 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
665 /* Cancel the timer because CPU already waken up from the C-states*/
666 menu_hrtimer_cancel();
667 __tick_nohz_idle_enter(ts);
671 * tick_nohz_get_sleep_length - return the length of the current sleep
673 * Called from power state control code with interrupts disabled
675 ktime_t tick_nohz_get_sleep_length(void)
677 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
679 return ts->sleep_length;
682 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
684 hrtimer_cancel(&ts->sched_timer);
685 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
688 /* Forward the time to expire in the future */
689 hrtimer_forward(&ts->sched_timer, now, tick_period);
691 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
692 hrtimer_start_expires(&ts->sched_timer,
693 HRTIMER_MODE_ABS_PINNED);
694 /* Check, if the timer was already in the past */
695 if (hrtimer_active(&ts->sched_timer))
698 if (!tick_program_event(
699 hrtimer_get_expires(&ts->sched_timer), 0))
702 /* Reread time and update jiffies */
704 tick_do_update_jiffies64(now);
708 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
710 /* Update jiffies first */
711 tick_do_update_jiffies64(now);
712 update_cpu_load_nohz();
714 calc_load_exit_idle();
715 touch_softlockup_watchdog();
717 * Cancel the scheduled timer and restore the tick
719 ts->tick_stopped = 0;
720 ts->idle_exittime = now;
722 tick_nohz_restart(ts, now);
725 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
727 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
730 if (vtime_accounting_enabled())
733 * We stopped the tick in idle. Update process times would miss the
734 * time we slept as update_process_times does only a 1 tick
735 * accounting. Enforce that this is accounted to idle !
737 ticks = jiffies - ts->idle_jiffies;
739 * We might be one off. Do not randomly account a huge number of ticks!
741 if (ticks && ticks < LONG_MAX)
742 account_idle_ticks(ticks);
747 * tick_nohz_idle_exit - restart the idle tick from the idle task
749 * Restart the idle tick when the CPU is woken up from idle
750 * This also exit the RCU extended quiescent state. The CPU
751 * can use RCU again after this function is called.
753 void tick_nohz_idle_exit(void)
755 int cpu = smp_processor_id();
756 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
761 WARN_ON_ONCE(!ts->inidle);
765 /* Cancel the timer because CPU already waken up from the C-states*/
766 menu_hrtimer_cancel();
767 if (ts->idle_active || ts->tick_stopped)
771 tick_nohz_stop_idle(cpu, now);
773 if (ts->tick_stopped) {
774 tick_nohz_restart_sched_tick(ts, now);
775 tick_nohz_account_idle_ticks(ts);
780 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
782 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
784 hrtimer_forward(&ts->sched_timer, now, tick_period);
785 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
789 * The nohz low res interrupt handler
791 static void tick_nohz_handler(struct clock_event_device *dev)
793 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
794 struct pt_regs *regs = get_irq_regs();
795 ktime_t now = ktime_get();
797 dev->next_event.tv64 = KTIME_MAX;
799 tick_sched_do_timer(now);
800 tick_sched_handle(ts, regs);
802 while (tick_nohz_reprogram(ts, now)) {
804 tick_do_update_jiffies64(now);
809 * tick_nohz_switch_to_nohz - switch to nohz mode
811 static void tick_nohz_switch_to_nohz(void)
813 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
816 if (!tick_nohz_enabled)
820 if (tick_switch_to_oneshot(tick_nohz_handler)) {
825 ts->nohz_mode = NOHZ_MODE_LOWRES;
828 * Recycle the hrtimer in ts, so we can share the
829 * hrtimer_forward with the highres code.
831 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
832 /* Get the next period */
833 next = tick_init_jiffy_update();
836 hrtimer_set_expires(&ts->sched_timer, next);
837 if (!tick_program_event(next, 0))
839 next = ktime_add(next, tick_period);
845 * When NOHZ is enabled and the tick is stopped, we need to kick the
846 * tick timer from irq_enter() so that the jiffies update is kept
847 * alive during long running softirqs. That's ugly as hell, but
848 * correctness is key even if we need to fix the offending softirq in
851 * Note, this is different to tick_nohz_restart. We just kick the
852 * timer and do not touch the other magic bits which need to be done
855 static void tick_nohz_kick_tick(int cpu, ktime_t now)
858 /* Switch back to 2.6.27 behaviour */
860 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
864 * Do not touch the tick device, when the next expiry is either
865 * already reached or less/equal than the tick period.
867 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
868 if (delta.tv64 <= tick_period.tv64)
871 tick_nohz_restart(ts, now);
875 static inline void tick_check_nohz(int cpu)
877 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
880 if (!ts->idle_active && !ts->tick_stopped)
884 tick_nohz_stop_idle(cpu, now);
885 if (ts->tick_stopped) {
886 tick_nohz_update_jiffies(now);
887 tick_nohz_kick_tick(cpu, now);
893 static inline void tick_nohz_switch_to_nohz(void) { }
894 static inline void tick_check_nohz(int cpu) { }
896 #endif /* CONFIG_NO_HZ_COMMON */
899 * Called from irq_enter to notify about the possible interruption of idle()
901 void tick_check_idle(int cpu)
903 tick_check_oneshot_broadcast(cpu);
904 tick_check_nohz(cpu);
908 * High resolution timer specific code
910 #ifdef CONFIG_HIGH_RES_TIMERS
912 * We rearm the timer until we get disabled by the idle code.
913 * Called with interrupts disabled.
915 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
917 struct tick_sched *ts =
918 container_of(timer, struct tick_sched, sched_timer);
919 struct pt_regs *regs = get_irq_regs();
920 ktime_t now = ktime_get();
922 tick_sched_do_timer(now);
925 * Do not call, when we are not in irq context and have
926 * no valid regs pointer
929 tick_sched_handle(ts, regs);
931 hrtimer_forward(timer, now, tick_period);
933 return HRTIMER_RESTART;
936 static int sched_skew_tick;
938 static int __init skew_tick(char *str)
940 get_option(&str, &sched_skew_tick);
944 early_param("skew_tick", skew_tick);
947 * tick_setup_sched_timer - setup the tick emulation timer
949 void tick_setup_sched_timer(void)
951 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
952 ktime_t now = ktime_get();
955 * Emulate tick processing via per-CPU hrtimers:
957 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
958 ts->sched_timer.function = tick_sched_timer;
960 /* Get the next period (per cpu) */
961 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
963 /* Offset the tick to avert jiffies_lock contention. */
964 if (sched_skew_tick) {
965 u64 offset = ktime_to_ns(tick_period) >> 1;
966 do_div(offset, num_possible_cpus());
967 offset *= smp_processor_id();
968 hrtimer_add_expires_ns(&ts->sched_timer, offset);
972 hrtimer_forward(&ts->sched_timer, now, tick_period);
973 hrtimer_start_expires(&ts->sched_timer,
974 HRTIMER_MODE_ABS_PINNED);
975 /* Check, if the timer was already in the past */
976 if (hrtimer_active(&ts->sched_timer))
981 #ifdef CONFIG_NO_HZ_COMMON
982 if (tick_nohz_enabled)
983 ts->nohz_mode = NOHZ_MODE_HIGHRES;
986 #endif /* HIGH_RES_TIMERS */
988 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
989 void tick_cancel_sched_timer(int cpu)
991 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
993 # ifdef CONFIG_HIGH_RES_TIMERS
994 if (ts->sched_timer.base)
995 hrtimer_cancel(&ts->sched_timer);
998 ts->nohz_mode = NOHZ_MODE_INACTIVE;
1003 * Async notification about clocksource changes
1005 void tick_clock_notify(void)
1009 for_each_possible_cpu(cpu)
1010 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1014 * Async notification about clock event changes
1016 void tick_oneshot_notify(void)
1018 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1020 set_bit(0, &ts->check_clocks);
1024 * Check, if a change happened, which makes oneshot possible.
1026 * Called cyclic from the hrtimer softirq (driven by the timer
1027 * softirq) allow_nohz signals, that we can switch into low-res nohz
1028 * mode, because high resolution timers are disabled (either compile
1031 int tick_check_oneshot_change(int allow_nohz)
1033 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1035 if (!test_and_clear_bit(0, &ts->check_clocks))
1038 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1041 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1047 tick_nohz_switch_to_nohz();