x86/xen: resume timer irqs early
[linux/fpc-iii.git] / kernel / time / tick-sched.c
blob29b063b32ff03d591f0b72fd0599689b31210015
1 /*
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>
24 #include <linux/posix-timers.h>
25 #include <linux/perf_event.h>
26 #include <linux/context_tracking.h>
28 #include <asm/irq_regs.h>
30 #include "tick-internal.h"
32 #include <trace/events/timer.h>
35 * Per cpu nohz control structure
37 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 * The time, when the last jiffy update happened. Protected by jiffies_lock.
42 static ktime_t last_jiffies_update;
44 struct tick_sched *tick_get_tick_sched(int cpu)
46 return &per_cpu(tick_cpu_sched, cpu);
50 * Must be called with interrupts disabled !
52 static void tick_do_update_jiffies64(ktime_t now)
54 unsigned long ticks = 0;
55 ktime_t delta;
58 * Do a quick check without holding jiffies_lock:
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta.tv64 < tick_period.tv64)
62 return;
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
72 tick_period);
74 /* Slow path for long timeouts */
75 if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 s64 incr = ktime_to_ns(tick_period);
78 ticks = ktime_divns(delta, incr);
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
81 incr * ticks);
83 do_timer(++ticks);
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
88 write_sequnlock(&jiffies_lock);
92 * Initialize and return retrieve the jiffies update.
94 static ktime_t tick_init_jiffy_update(void)
96 ktime_t period;
98 write_seqlock(&jiffies_lock);
99 /* Did we start the jiffies update yet ? */
100 if (last_jiffies_update.tv64 == 0)
101 last_jiffies_update = tick_next_period;
102 period = last_jiffies_update;
103 write_sequnlock(&jiffies_lock);
104 return period;
108 static void tick_sched_do_timer(ktime_t now)
110 int cpu = smp_processor_id();
112 #ifdef CONFIG_NO_HZ_COMMON
114 * Check if the do_timer duty was dropped. We don't care about
115 * concurrency: This happens only when the cpu in charge went
116 * into a long sleep. If two cpus happen to assign themself to
117 * this duty, then the jiffies update is still serialized by
118 * jiffies_lock.
120 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
121 && !tick_nohz_full_cpu(cpu))
122 tick_do_timer_cpu = cpu;
123 #endif
125 /* Check, if the jiffies need an update */
126 if (tick_do_timer_cpu == cpu)
127 tick_do_update_jiffies64(now);
130 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
132 #ifdef CONFIG_NO_HZ_COMMON
134 * When we are idle and the tick is stopped, we have to touch
135 * the watchdog as we might not schedule for a really long
136 * time. This happens on complete idle SMP systems while
137 * waiting on the login prompt. We also increment the "start of
138 * idle" jiffy stamp so the idle accounting adjustment we do
139 * when we go busy again does not account too much ticks.
141 if (ts->tick_stopped) {
142 touch_softlockup_watchdog();
143 if (is_idle_task(current))
144 ts->idle_jiffies++;
146 #endif
147 update_process_times(user_mode(regs));
148 profile_tick(CPU_PROFILING);
151 #ifdef CONFIG_NO_HZ_FULL
152 cpumask_var_t tick_nohz_full_mask;
153 bool tick_nohz_full_running;
155 static bool can_stop_full_tick(void)
157 WARN_ON_ONCE(!irqs_disabled());
159 if (!sched_can_stop_tick()) {
160 trace_tick_stop(0, "more than 1 task in runqueue\n");
161 return false;
164 if (!posix_cpu_timers_can_stop_tick(current)) {
165 trace_tick_stop(0, "posix timers running\n");
166 return false;
169 if (!perf_event_can_stop_tick()) {
170 trace_tick_stop(0, "perf events running\n");
171 return false;
174 /* sched_clock_tick() needs us? */
175 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
177 * TODO: kick full dynticks CPUs when
178 * sched_clock_stable is set.
180 if (!sched_clock_stable) {
181 trace_tick_stop(0, "unstable sched clock\n");
183 * Don't allow the user to think they can get
184 * full NO_HZ with this machine.
186 WARN_ONCE(tick_nohz_full_running,
187 "NO_HZ FULL will not work with unstable sched clock");
188 return false;
190 #endif
192 return true;
195 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
198 * Re-evaluate the need for the tick on the current CPU
199 * and restart it if necessary.
201 void __tick_nohz_full_check(void)
203 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
205 if (tick_nohz_full_cpu(smp_processor_id())) {
206 if (ts->tick_stopped && !is_idle_task(current)) {
207 if (!can_stop_full_tick())
208 tick_nohz_restart_sched_tick(ts, ktime_get());
213 static void nohz_full_kick_work_func(struct irq_work *work)
215 __tick_nohz_full_check();
218 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
219 .func = nohz_full_kick_work_func,
223 * Kick the current CPU if it's full dynticks in order to force it to
224 * re-evaluate its dependency on the tick and restart it if necessary.
226 void tick_nohz_full_kick(void)
228 if (tick_nohz_full_cpu(smp_processor_id()))
229 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
232 static void nohz_full_kick_ipi(void *info)
234 __tick_nohz_full_check();
238 * Kick all full dynticks CPUs in order to force these to re-evaluate
239 * their dependency on the tick and restart it if necessary.
241 void tick_nohz_full_kick_all(void)
243 if (!tick_nohz_full_running)
244 return;
246 preempt_disable();
247 smp_call_function_many(tick_nohz_full_mask,
248 nohz_full_kick_ipi, NULL, false);
249 tick_nohz_full_kick();
250 preempt_enable();
254 * Re-evaluate the need for the tick as we switch the current task.
255 * It might need the tick due to per task/process properties:
256 * perf events, posix cpu timers, ...
258 void __tick_nohz_task_switch(struct task_struct *tsk)
260 unsigned long flags;
262 local_irq_save(flags);
264 if (!tick_nohz_full_cpu(smp_processor_id()))
265 goto out;
267 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
268 tick_nohz_full_kick();
270 out:
271 local_irq_restore(flags);
274 /* Parse the boot-time nohz CPU list from the kernel parameters. */
275 static int __init tick_nohz_full_setup(char *str)
277 int cpu;
279 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
280 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
281 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
282 return 1;
285 cpu = smp_processor_id();
286 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
287 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
288 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
290 tick_nohz_full_running = true;
292 return 1;
294 __setup("nohz_full=", tick_nohz_full_setup);
296 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
297 unsigned long action,
298 void *hcpu)
300 unsigned int cpu = (unsigned long)hcpu;
302 switch (action & ~CPU_TASKS_FROZEN) {
303 case CPU_DOWN_PREPARE:
305 * If we handle the timekeeping duty for full dynticks CPUs,
306 * we can't safely shutdown that CPU.
308 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
309 return NOTIFY_BAD;
310 break;
312 return NOTIFY_OK;
316 * Worst case string length in chunks of CPU range seems 2 steps
317 * separations: 0,2,4,6,...
318 * This is NR_CPUS + sizeof('\0')
320 static char __initdata nohz_full_buf[NR_CPUS + 1];
322 static int tick_nohz_init_all(void)
324 int err = -1;
326 #ifdef CONFIG_NO_HZ_FULL_ALL
327 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
328 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
329 return err;
331 err = 0;
332 cpumask_setall(tick_nohz_full_mask);
333 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask);
334 tick_nohz_full_running = true;
335 #endif
336 return err;
339 void __init tick_nohz_init(void)
341 int cpu;
343 if (!tick_nohz_full_running) {
344 if (tick_nohz_init_all() < 0)
345 return;
348 for_each_cpu(cpu, tick_nohz_full_mask)
349 context_tracking_cpu_set(cpu);
351 cpu_notifier(tick_nohz_cpu_down_callback, 0);
352 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
353 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
355 #endif
358 * NOHZ - aka dynamic tick functionality
360 #ifdef CONFIG_NO_HZ_COMMON
362 * NO HZ enabled ?
364 static int tick_nohz_enabled __read_mostly = 1;
365 int tick_nohz_active __read_mostly;
367 * Enable / Disable tickless mode
369 static int __init setup_tick_nohz(char *str)
371 if (!strcmp(str, "off"))
372 tick_nohz_enabled = 0;
373 else if (!strcmp(str, "on"))
374 tick_nohz_enabled = 1;
375 else
376 return 0;
377 return 1;
380 __setup("nohz=", setup_tick_nohz);
383 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
385 * Called from interrupt entry when the CPU was idle
387 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
388 * must be updated. Otherwise an interrupt handler could use a stale jiffy
389 * value. We do this unconditionally on any cpu, as we don't know whether the
390 * cpu, which has the update task assigned is in a long sleep.
392 static void tick_nohz_update_jiffies(ktime_t now)
394 int cpu = smp_processor_id();
395 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
396 unsigned long flags;
398 ts->idle_waketime = now;
400 local_irq_save(flags);
401 tick_do_update_jiffies64(now);
402 local_irq_restore(flags);
404 touch_softlockup_watchdog();
408 * Updates the per cpu time idle statistics counters
410 static void
411 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
413 ktime_t delta;
415 if (ts->idle_active) {
416 delta = ktime_sub(now, ts->idle_entrytime);
417 if (nr_iowait_cpu(cpu) > 0)
418 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
419 else
420 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
421 ts->idle_entrytime = now;
424 if (last_update_time)
425 *last_update_time = ktime_to_us(now);
429 static void tick_nohz_stop_idle(int cpu, ktime_t now)
431 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
433 update_ts_time_stats(cpu, ts, now, NULL);
434 ts->idle_active = 0;
436 sched_clock_idle_wakeup_event(0);
439 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
441 ktime_t now = ktime_get();
443 ts->idle_entrytime = now;
444 ts->idle_active = 1;
445 sched_clock_idle_sleep_event();
446 return now;
450 * get_cpu_idle_time_us - get the total idle time of a cpu
451 * @cpu: CPU number to query
452 * @last_update_time: variable to store update time in. Do not update
453 * counters if NULL.
455 * Return the cummulative idle time (since boot) for a given
456 * CPU, in microseconds.
458 * This time is measured via accounting rather than sampling,
459 * and is as accurate as ktime_get() is.
461 * This function returns -1 if NOHZ is not enabled.
463 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
465 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
466 ktime_t now, idle;
468 if (!tick_nohz_active)
469 return -1;
471 now = ktime_get();
472 if (last_update_time) {
473 update_ts_time_stats(cpu, ts, now, last_update_time);
474 idle = ts->idle_sleeptime;
475 } else {
476 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
477 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
479 idle = ktime_add(ts->idle_sleeptime, delta);
480 } else {
481 idle = ts->idle_sleeptime;
485 return ktime_to_us(idle);
488 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
491 * get_cpu_iowait_time_us - get the total iowait time of a cpu
492 * @cpu: CPU number to query
493 * @last_update_time: variable to store update time in. Do not update
494 * counters if NULL.
496 * Return the cummulative iowait time (since boot) for a given
497 * CPU, in microseconds.
499 * This time is measured via accounting rather than sampling,
500 * and is as accurate as ktime_get() is.
502 * This function returns -1 if NOHZ is not enabled.
504 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
506 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
507 ktime_t now, iowait;
509 if (!tick_nohz_active)
510 return -1;
512 now = ktime_get();
513 if (last_update_time) {
514 update_ts_time_stats(cpu, ts, now, last_update_time);
515 iowait = ts->iowait_sleeptime;
516 } else {
517 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
518 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
520 iowait = ktime_add(ts->iowait_sleeptime, delta);
521 } else {
522 iowait = ts->iowait_sleeptime;
526 return ktime_to_us(iowait);
528 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
530 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
531 ktime_t now, int cpu)
533 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
534 ktime_t last_update, expires, ret = { .tv64 = 0 };
535 unsigned long rcu_delta_jiffies;
536 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
537 u64 time_delta;
539 /* Read jiffies and the time when jiffies were updated last */
540 do {
541 seq = read_seqbegin(&jiffies_lock);
542 last_update = last_jiffies_update;
543 last_jiffies = jiffies;
544 time_delta = timekeeping_max_deferment();
545 } while (read_seqretry(&jiffies_lock, seq));
547 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
548 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
549 next_jiffies = last_jiffies + 1;
550 delta_jiffies = 1;
551 } else {
552 /* Get the next timer wheel timer */
553 next_jiffies = get_next_timer_interrupt(last_jiffies);
554 delta_jiffies = next_jiffies - last_jiffies;
555 if (rcu_delta_jiffies < delta_jiffies) {
556 next_jiffies = last_jiffies + rcu_delta_jiffies;
557 delta_jiffies = rcu_delta_jiffies;
562 * Do not stop the tick, if we are only one off (or less)
563 * or if the cpu is required for RCU:
565 if (!ts->tick_stopped && delta_jiffies <= 1)
566 goto out;
568 /* Schedule the tick, if we are at least one jiffie off */
569 if ((long)delta_jiffies >= 1) {
572 * If this cpu is the one which updates jiffies, then
573 * give up the assignment and let it be taken by the
574 * cpu which runs the tick timer next, which might be
575 * this cpu as well. If we don't drop this here the
576 * jiffies might be stale and do_timer() never
577 * invoked. Keep track of the fact that it was the one
578 * which had the do_timer() duty last. If this cpu is
579 * the one which had the do_timer() duty last, we
580 * limit the sleep time to the timekeeping
581 * max_deferement value which we retrieved
582 * above. Otherwise we can sleep as long as we want.
584 if (cpu == tick_do_timer_cpu) {
585 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
586 ts->do_timer_last = 1;
587 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
588 time_delta = KTIME_MAX;
589 ts->do_timer_last = 0;
590 } else if (!ts->do_timer_last) {
591 time_delta = KTIME_MAX;
594 #ifdef CONFIG_NO_HZ_FULL
595 if (!ts->inidle) {
596 time_delta = min(time_delta,
597 scheduler_tick_max_deferment());
599 #endif
602 * calculate the expiry time for the next timer wheel
603 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
604 * that there is no timer pending or at least extremely
605 * far into the future (12 days for HZ=1000). In this
606 * case we set the expiry to the end of time.
608 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
610 * Calculate the time delta for the next timer event.
611 * If the time delta exceeds the maximum time delta
612 * permitted by the current clocksource then adjust
613 * the time delta accordingly to ensure the
614 * clocksource does not wrap.
616 time_delta = min_t(u64, time_delta,
617 tick_period.tv64 * delta_jiffies);
620 if (time_delta < KTIME_MAX)
621 expires = ktime_add_ns(last_update, time_delta);
622 else
623 expires.tv64 = KTIME_MAX;
625 /* Skip reprogram of event if its not changed */
626 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
627 goto out;
629 ret = expires;
632 * nohz_stop_sched_tick can be called several times before
633 * the nohz_restart_sched_tick is called. This happens when
634 * interrupts arrive which do not cause a reschedule. In the
635 * first call we save the current tick time, so we can restart
636 * the scheduler tick in nohz_restart_sched_tick.
638 if (!ts->tick_stopped) {
639 nohz_balance_enter_idle(cpu);
640 calc_load_enter_idle();
642 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
643 ts->tick_stopped = 1;
644 trace_tick_stop(1, " ");
648 * If the expiration time == KTIME_MAX, then
649 * in this case we simply stop the tick timer.
651 if (unlikely(expires.tv64 == KTIME_MAX)) {
652 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
653 hrtimer_cancel(&ts->sched_timer);
654 goto out;
657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
658 hrtimer_start(&ts->sched_timer, expires,
659 HRTIMER_MODE_ABS_PINNED);
660 /* Check, if the timer was already in the past */
661 if (hrtimer_active(&ts->sched_timer))
662 goto out;
663 } else if (!tick_program_event(expires, 0))
664 goto out;
666 * We are past the event already. So we crossed a
667 * jiffie boundary. Update jiffies and raise the
668 * softirq.
670 tick_do_update_jiffies64(ktime_get());
672 raise_softirq_irqoff(TIMER_SOFTIRQ);
673 out:
674 ts->next_jiffies = next_jiffies;
675 ts->last_jiffies = last_jiffies;
676 ts->sleep_length = ktime_sub(dev->next_event, now);
678 return ret;
681 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
683 #ifdef CONFIG_NO_HZ_FULL
684 int cpu = smp_processor_id();
686 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
687 return;
689 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
690 return;
692 if (!can_stop_full_tick())
693 return;
695 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
696 #endif
699 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
702 * If this cpu is offline and it is the one which updates
703 * jiffies, then give up the assignment and let it be taken by
704 * the cpu which runs the tick timer next. If we don't drop
705 * this here the jiffies might be stale and do_timer() never
706 * invoked.
708 if (unlikely(!cpu_online(cpu))) {
709 if (cpu == tick_do_timer_cpu)
710 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
711 return false;
714 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
715 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
716 return false;
719 if (need_resched())
720 return false;
722 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
723 static int ratelimit;
725 if (ratelimit < 10 &&
726 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
727 pr_warn("NOHZ: local_softirq_pending %02x\n",
728 (unsigned int) local_softirq_pending());
729 ratelimit++;
731 return false;
734 if (tick_nohz_full_enabled()) {
736 * Keep the tick alive to guarantee timekeeping progression
737 * if there are full dynticks CPUs around
739 if (tick_do_timer_cpu == cpu)
740 return false;
742 * Boot safety: make sure the timekeeping duty has been
743 * assigned before entering dyntick-idle mode,
745 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
746 return false;
749 return true;
752 static void __tick_nohz_idle_enter(struct tick_sched *ts)
754 ktime_t now, expires;
755 int cpu = smp_processor_id();
757 now = tick_nohz_start_idle(cpu, ts);
759 if (can_stop_idle_tick(cpu, ts)) {
760 int was_stopped = ts->tick_stopped;
762 ts->idle_calls++;
764 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
765 if (expires.tv64 > 0LL) {
766 ts->idle_sleeps++;
767 ts->idle_expires = expires;
770 if (!was_stopped && ts->tick_stopped)
771 ts->idle_jiffies = ts->last_jiffies;
776 * tick_nohz_idle_enter - stop the idle tick from the idle task
778 * When the next event is more than a tick into the future, stop the idle tick
779 * Called when we start the idle loop.
781 * The arch is responsible of calling:
783 * - rcu_idle_enter() after its last use of RCU before the CPU is put
784 * to sleep.
785 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
787 void tick_nohz_idle_enter(void)
789 struct tick_sched *ts;
791 WARN_ON_ONCE(irqs_disabled());
794 * Update the idle state in the scheduler domain hierarchy
795 * when tick_nohz_stop_sched_tick() is called from the idle loop.
796 * State will be updated to busy during the first busy tick after
797 * exiting idle.
799 set_cpu_sd_state_idle();
801 local_irq_disable();
803 ts = &__get_cpu_var(tick_cpu_sched);
804 ts->inidle = 1;
805 __tick_nohz_idle_enter(ts);
807 local_irq_enable();
809 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
812 * tick_nohz_irq_exit - update next tick event from interrupt exit
814 * When an interrupt fires while we are idle and it doesn't cause
815 * a reschedule, it may still add, modify or delete a timer, enqueue
816 * an RCU callback, etc...
817 * So we need to re-calculate and reprogram the next tick event.
819 void tick_nohz_irq_exit(void)
821 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
823 if (ts->inidle)
824 __tick_nohz_idle_enter(ts);
825 else
826 tick_nohz_full_stop_tick(ts);
830 * tick_nohz_get_sleep_length - return the length of the current sleep
832 * Called from power state control code with interrupts disabled
834 ktime_t tick_nohz_get_sleep_length(void)
836 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
838 return ts->sleep_length;
841 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
843 hrtimer_cancel(&ts->sched_timer);
844 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
846 while (1) {
847 /* Forward the time to expire in the future */
848 hrtimer_forward(&ts->sched_timer, now, tick_period);
850 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
851 hrtimer_start_expires(&ts->sched_timer,
852 HRTIMER_MODE_ABS_PINNED);
853 /* Check, if the timer was already in the past */
854 if (hrtimer_active(&ts->sched_timer))
855 break;
856 } else {
857 if (!tick_program_event(
858 hrtimer_get_expires(&ts->sched_timer), 0))
859 break;
861 /* Reread time and update jiffies */
862 now = ktime_get();
863 tick_do_update_jiffies64(now);
867 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
869 /* Update jiffies first */
870 tick_do_update_jiffies64(now);
871 update_cpu_load_nohz();
873 calc_load_exit_idle();
874 touch_softlockup_watchdog();
876 * Cancel the scheduled timer and restore the tick
878 ts->tick_stopped = 0;
879 ts->idle_exittime = now;
881 tick_nohz_restart(ts, now);
884 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
886 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
887 unsigned long ticks;
889 if (vtime_accounting_enabled())
890 return;
892 * We stopped the tick in idle. Update process times would miss the
893 * time we slept as update_process_times does only a 1 tick
894 * accounting. Enforce that this is accounted to idle !
896 ticks = jiffies - ts->idle_jiffies;
898 * We might be one off. Do not randomly account a huge number of ticks!
900 if (ticks && ticks < LONG_MAX)
901 account_idle_ticks(ticks);
902 #endif
906 * tick_nohz_idle_exit - restart the idle tick from the idle task
908 * Restart the idle tick when the CPU is woken up from idle
909 * This also exit the RCU extended quiescent state. The CPU
910 * can use RCU again after this function is called.
912 void tick_nohz_idle_exit(void)
914 int cpu = smp_processor_id();
915 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
916 ktime_t now;
918 local_irq_disable();
920 WARN_ON_ONCE(!ts->inidle);
922 ts->inidle = 0;
924 if (ts->idle_active || ts->tick_stopped)
925 now = ktime_get();
927 if (ts->idle_active)
928 tick_nohz_stop_idle(cpu, now);
930 if (ts->tick_stopped) {
931 tick_nohz_restart_sched_tick(ts, now);
932 tick_nohz_account_idle_ticks(ts);
935 local_irq_enable();
937 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
939 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
941 hrtimer_forward(&ts->sched_timer, now, tick_period);
942 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
946 * The nohz low res interrupt handler
948 static void tick_nohz_handler(struct clock_event_device *dev)
950 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
951 struct pt_regs *regs = get_irq_regs();
952 ktime_t now = ktime_get();
954 dev->next_event.tv64 = KTIME_MAX;
956 tick_sched_do_timer(now);
957 tick_sched_handle(ts, regs);
959 while (tick_nohz_reprogram(ts, now)) {
960 now = ktime_get();
961 tick_do_update_jiffies64(now);
966 * tick_nohz_switch_to_nohz - switch to nohz mode
968 static void tick_nohz_switch_to_nohz(void)
970 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
971 ktime_t next;
973 if (!tick_nohz_enabled)
974 return;
976 local_irq_disable();
977 if (tick_switch_to_oneshot(tick_nohz_handler)) {
978 local_irq_enable();
979 return;
981 tick_nohz_active = 1;
982 ts->nohz_mode = NOHZ_MODE_LOWRES;
985 * Recycle the hrtimer in ts, so we can share the
986 * hrtimer_forward with the highres code.
988 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
989 /* Get the next period */
990 next = tick_init_jiffy_update();
992 for (;;) {
993 hrtimer_set_expires(&ts->sched_timer, next);
994 if (!tick_program_event(next, 0))
995 break;
996 next = ktime_add(next, tick_period);
998 local_irq_enable();
1002 * When NOHZ is enabled and the tick is stopped, we need to kick the
1003 * tick timer from irq_enter() so that the jiffies update is kept
1004 * alive during long running softirqs. That's ugly as hell, but
1005 * correctness is key even if we need to fix the offending softirq in
1006 * the first place.
1008 * Note, this is different to tick_nohz_restart. We just kick the
1009 * timer and do not touch the other magic bits which need to be done
1010 * when idle is left.
1012 static void tick_nohz_kick_tick(int cpu, ktime_t now)
1014 #if 0
1015 /* Switch back to 2.6.27 behaviour */
1017 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1018 ktime_t delta;
1021 * Do not touch the tick device, when the next expiry is either
1022 * already reached or less/equal than the tick period.
1024 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1025 if (delta.tv64 <= tick_period.tv64)
1026 return;
1028 tick_nohz_restart(ts, now);
1029 #endif
1032 static inline void tick_check_nohz(int cpu)
1034 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1035 ktime_t now;
1037 if (!ts->idle_active && !ts->tick_stopped)
1038 return;
1039 now = ktime_get();
1040 if (ts->idle_active)
1041 tick_nohz_stop_idle(cpu, now);
1042 if (ts->tick_stopped) {
1043 tick_nohz_update_jiffies(now);
1044 tick_nohz_kick_tick(cpu, now);
1048 #else
1050 static inline void tick_nohz_switch_to_nohz(void) { }
1051 static inline void tick_check_nohz(int cpu) { }
1053 #endif /* CONFIG_NO_HZ_COMMON */
1056 * Called from irq_enter to notify about the possible interruption of idle()
1058 void tick_check_idle(int cpu)
1060 tick_check_oneshot_broadcast(cpu);
1061 tick_check_nohz(cpu);
1065 * High resolution timer specific code
1067 #ifdef CONFIG_HIGH_RES_TIMERS
1069 * We rearm the timer until we get disabled by the idle code.
1070 * Called with interrupts disabled.
1072 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1074 struct tick_sched *ts =
1075 container_of(timer, struct tick_sched, sched_timer);
1076 struct pt_regs *regs = get_irq_regs();
1077 ktime_t now = ktime_get();
1079 tick_sched_do_timer(now);
1082 * Do not call, when we are not in irq context and have
1083 * no valid regs pointer
1085 if (regs)
1086 tick_sched_handle(ts, regs);
1088 hrtimer_forward(timer, now, tick_period);
1090 return HRTIMER_RESTART;
1093 static int sched_skew_tick;
1095 static int __init skew_tick(char *str)
1097 get_option(&str, &sched_skew_tick);
1099 return 0;
1101 early_param("skew_tick", skew_tick);
1104 * tick_setup_sched_timer - setup the tick emulation timer
1106 void tick_setup_sched_timer(void)
1108 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1109 ktime_t now = ktime_get();
1112 * Emulate tick processing via per-CPU hrtimers:
1114 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1115 ts->sched_timer.function = tick_sched_timer;
1117 /* Get the next period (per cpu) */
1118 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1120 /* Offset the tick to avert jiffies_lock contention. */
1121 if (sched_skew_tick) {
1122 u64 offset = ktime_to_ns(tick_period) >> 1;
1123 do_div(offset, num_possible_cpus());
1124 offset *= smp_processor_id();
1125 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1128 for (;;) {
1129 hrtimer_forward(&ts->sched_timer, now, tick_period);
1130 hrtimer_start_expires(&ts->sched_timer,
1131 HRTIMER_MODE_ABS_PINNED);
1132 /* Check, if the timer was already in the past */
1133 if (hrtimer_active(&ts->sched_timer))
1134 break;
1135 now = ktime_get();
1138 #ifdef CONFIG_NO_HZ_COMMON
1139 if (tick_nohz_enabled) {
1140 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1141 tick_nohz_active = 1;
1143 #endif
1145 #endif /* HIGH_RES_TIMERS */
1147 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1148 void tick_cancel_sched_timer(int cpu)
1150 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1152 # ifdef CONFIG_HIGH_RES_TIMERS
1153 if (ts->sched_timer.base)
1154 hrtimer_cancel(&ts->sched_timer);
1155 # endif
1157 memset(ts, 0, sizeof(*ts));
1159 #endif
1162 * Async notification about clocksource changes
1164 void tick_clock_notify(void)
1166 int cpu;
1168 for_each_possible_cpu(cpu)
1169 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1173 * Async notification about clock event changes
1175 void tick_oneshot_notify(void)
1177 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1179 set_bit(0, &ts->check_clocks);
1183 * Check, if a change happened, which makes oneshot possible.
1185 * Called cyclic from the hrtimer softirq (driven by the timer
1186 * softirq) allow_nohz signals, that we can switch into low-res nohz
1187 * mode, because high resolution timers are disabled (either compile
1188 * or runtime).
1190 int tick_check_oneshot_change(int allow_nohz)
1192 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1194 if (!test_and_clear_bit(0, &ts->check_clocks))
1195 return 0;
1197 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1198 return 0;
1200 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1201 return 0;
1203 if (!allow_nohz)
1204 return 1;
1206 tick_nohz_switch_to_nohz();
1207 return 0;