Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / kernel / time / tick-sched.c
blob29a5733eff83ec9e40432a2e9353bf19253d10b7
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/nmi.h>
21 #include <linux/profile.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/clock.h>
24 #include <linux/sched/stat.h>
25 #include <linux/sched/nohz.h>
26 #include <linux/module.h>
27 #include <linux/irq_work.h>
28 #include <linux/posix-timers.h>
29 #include <linux/context_tracking.h>
30 #include <linux/mm.h>
32 #include <asm/irq_regs.h>
34 #include "tick-internal.h"
36 #include <trace/events/timer.h>
39 * Per-CPU nohz control structure
41 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
43 struct tick_sched *tick_get_tick_sched(int cpu)
45 return &per_cpu(tick_cpu_sched, cpu);
48 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
50 * The time, when the last jiffy update happened. Protected by jiffies_lock.
52 static ktime_t last_jiffies_update;
55 * Must be called with interrupts disabled !
57 static void tick_do_update_jiffies64(ktime_t now)
59 unsigned long ticks = 0;
60 ktime_t delta;
63 * Do a quick check without holding jiffies_lock:
65 delta = ktime_sub(now, last_jiffies_update);
66 if (delta < tick_period)
67 return;
69 /* Reevaluate with jiffies_lock held */
70 write_seqlock(&jiffies_lock);
72 delta = ktime_sub(now, last_jiffies_update);
73 if (delta >= tick_period) {
75 delta = ktime_sub(delta, tick_period);
76 last_jiffies_update = ktime_add(last_jiffies_update,
77 tick_period);
79 /* Slow path for long timeouts */
80 if (unlikely(delta >= tick_period)) {
81 s64 incr = ktime_to_ns(tick_period);
83 ticks = ktime_divns(delta, incr);
85 last_jiffies_update = ktime_add_ns(last_jiffies_update,
86 incr * ticks);
88 do_timer(++ticks);
90 /* Keep the tick_next_period variable up to date */
91 tick_next_period = ktime_add(last_jiffies_update, tick_period);
92 } else {
93 write_sequnlock(&jiffies_lock);
94 return;
96 write_sequnlock(&jiffies_lock);
97 update_wall_time();
101 * Initialize and return retrieve the jiffies update.
103 static ktime_t tick_init_jiffy_update(void)
105 ktime_t period;
107 write_seqlock(&jiffies_lock);
108 /* Did we start the jiffies update yet ? */
109 if (last_jiffies_update == 0)
110 last_jiffies_update = tick_next_period;
111 period = last_jiffies_update;
112 write_sequnlock(&jiffies_lock);
113 return period;
117 static void tick_sched_do_timer(ktime_t now)
119 int cpu = smp_processor_id();
121 #ifdef CONFIG_NO_HZ_COMMON
123 * Check if the do_timer duty was dropped. We don't care about
124 * concurrency: This happens only when the CPU in charge went
125 * into a long sleep. If two CPUs happen to assign themselves to
126 * this duty, then the jiffies update is still serialized by
127 * jiffies_lock.
129 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
130 && !tick_nohz_full_cpu(cpu))
131 tick_do_timer_cpu = cpu;
132 #endif
134 /* Check, if the jiffies need an update */
135 if (tick_do_timer_cpu == cpu)
136 tick_do_update_jiffies64(now);
139 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
141 #ifdef CONFIG_NO_HZ_COMMON
143 * When we are idle and the tick is stopped, we have to touch
144 * the watchdog as we might not schedule for a really long
145 * time. This happens on complete idle SMP systems while
146 * waiting on the login prompt. We also increment the "start of
147 * idle" jiffy stamp so the idle accounting adjustment we do
148 * when we go busy again does not account too much ticks.
150 if (ts->tick_stopped) {
151 touch_softlockup_watchdog_sched();
152 if (is_idle_task(current))
153 ts->idle_jiffies++;
155 * In case the current tick fired too early past its expected
156 * expiration, make sure we don't bypass the next clock reprogramming
157 * to the same deadline.
159 ts->next_tick = 0;
161 #endif
162 update_process_times(user_mode(regs));
163 profile_tick(CPU_PROFILING);
165 #endif
167 #ifdef CONFIG_NO_HZ_FULL
168 cpumask_var_t tick_nohz_full_mask;
169 bool tick_nohz_full_running;
170 static atomic_t tick_dep_mask;
172 static bool check_tick_dependency(atomic_t *dep)
174 int val = atomic_read(dep);
176 if (val & TICK_DEP_MASK_POSIX_TIMER) {
177 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
178 return true;
181 if (val & TICK_DEP_MASK_PERF_EVENTS) {
182 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
183 return true;
186 if (val & TICK_DEP_MASK_SCHED) {
187 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
188 return true;
191 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
192 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
193 return true;
196 return false;
199 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
201 lockdep_assert_irqs_disabled();
203 if (unlikely(!cpu_online(cpu)))
204 return false;
206 if (check_tick_dependency(&tick_dep_mask))
207 return false;
209 if (check_tick_dependency(&ts->tick_dep_mask))
210 return false;
212 if (check_tick_dependency(&current->tick_dep_mask))
213 return false;
215 if (check_tick_dependency(&current->signal->tick_dep_mask))
216 return false;
218 return true;
221 static void nohz_full_kick_func(struct irq_work *work)
223 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
226 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
227 .func = nohz_full_kick_func,
231 * Kick this CPU if it's full dynticks in order to force it to
232 * re-evaluate its dependency on the tick and restart it if necessary.
233 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
234 * is NMI safe.
236 static void tick_nohz_full_kick(void)
238 if (!tick_nohz_full_cpu(smp_processor_id()))
239 return;
241 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
245 * Kick the CPU if it's full dynticks in order to force it to
246 * re-evaluate its dependency on the tick and restart it if necessary.
248 void tick_nohz_full_kick_cpu(int cpu)
250 if (!tick_nohz_full_cpu(cpu))
251 return;
253 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
257 * Kick all full dynticks CPUs in order to force these to re-evaluate
258 * their dependency on the tick and restart it if necessary.
260 static void tick_nohz_full_kick_all(void)
262 int cpu;
264 if (!tick_nohz_full_running)
265 return;
267 preempt_disable();
268 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
269 tick_nohz_full_kick_cpu(cpu);
270 preempt_enable();
273 static void tick_nohz_dep_set_all(atomic_t *dep,
274 enum tick_dep_bits bit)
276 int prev;
278 prev = atomic_fetch_or(BIT(bit), dep);
279 if (!prev)
280 tick_nohz_full_kick_all();
284 * Set a global tick dependency. Used by perf events that rely on freq and
285 * by unstable clock.
287 void tick_nohz_dep_set(enum tick_dep_bits bit)
289 tick_nohz_dep_set_all(&tick_dep_mask, bit);
292 void tick_nohz_dep_clear(enum tick_dep_bits bit)
294 atomic_andnot(BIT(bit), &tick_dep_mask);
298 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
299 * manage events throttling.
301 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
303 int prev;
304 struct tick_sched *ts;
306 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
308 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
309 if (!prev) {
310 preempt_disable();
311 /* Perf needs local kick that is NMI safe */
312 if (cpu == smp_processor_id()) {
313 tick_nohz_full_kick();
314 } else {
315 /* Remote irq work not NMI-safe */
316 if (!WARN_ON_ONCE(in_nmi()))
317 tick_nohz_full_kick_cpu(cpu);
319 preempt_enable();
323 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
325 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
327 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
331 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
332 * per task timers.
334 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
337 * We could optimize this with just kicking the target running the task
338 * if that noise matters for nohz full users.
340 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
343 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
345 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
349 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
350 * per process timers.
352 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
354 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
357 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
359 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
363 * Re-evaluate the need for the tick as we switch the current task.
364 * It might need the tick due to per task/process properties:
365 * perf events, posix CPU timers, ...
367 void __tick_nohz_task_switch(void)
369 unsigned long flags;
370 struct tick_sched *ts;
372 local_irq_save(flags);
374 if (!tick_nohz_full_cpu(smp_processor_id()))
375 goto out;
377 ts = this_cpu_ptr(&tick_cpu_sched);
379 if (ts->tick_stopped) {
380 if (atomic_read(&current->tick_dep_mask) ||
381 atomic_read(&current->signal->tick_dep_mask))
382 tick_nohz_full_kick();
384 out:
385 local_irq_restore(flags);
388 /* Get the boot-time nohz CPU list from the kernel parameters. */
389 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
391 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
392 cpumask_copy(tick_nohz_full_mask, cpumask);
393 tick_nohz_full_running = true;
396 static int tick_nohz_cpu_down(unsigned int cpu)
399 * The boot CPU handles housekeeping duty (unbound timers,
400 * workqueues, timekeeping, ...) on behalf of full dynticks
401 * CPUs. It must remain online when nohz full is enabled.
403 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
404 return -EBUSY;
405 return 0;
408 static int tick_nohz_init_all(void)
410 int err = -1;
412 #ifdef CONFIG_NO_HZ_FULL_ALL
413 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
414 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
415 return err;
417 err = 0;
418 cpumask_setall(tick_nohz_full_mask);
419 tick_nohz_full_running = true;
420 #endif
421 return err;
424 void __init tick_nohz_init(void)
426 int cpu, ret;
428 if (!tick_nohz_full_running) {
429 if (tick_nohz_init_all() < 0)
430 return;
434 * Full dynticks uses irq work to drive the tick rescheduling on safe
435 * locking contexts. But then we need irq work to raise its own
436 * interrupts to avoid circular dependency on the tick
438 if (!arch_irq_work_has_interrupt()) {
439 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
440 cpumask_clear(tick_nohz_full_mask);
441 tick_nohz_full_running = false;
442 return;
445 cpu = smp_processor_id();
447 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
448 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
449 cpu);
450 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
453 for_each_cpu(cpu, tick_nohz_full_mask)
454 context_tracking_cpu_set(cpu);
456 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
457 "kernel/nohz:predown", NULL,
458 tick_nohz_cpu_down);
459 WARN_ON(ret < 0);
460 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
461 cpumask_pr_args(tick_nohz_full_mask));
463 #endif
466 * NOHZ - aka dynamic tick functionality
468 #ifdef CONFIG_NO_HZ_COMMON
470 * NO HZ enabled ?
472 bool tick_nohz_enabled __read_mostly = true;
473 unsigned long tick_nohz_active __read_mostly;
475 * Enable / Disable tickless mode
477 static int __init setup_tick_nohz(char *str)
479 return (kstrtobool(str, &tick_nohz_enabled) == 0);
482 __setup("nohz=", setup_tick_nohz);
484 int tick_nohz_tick_stopped(void)
486 return __this_cpu_read(tick_cpu_sched.tick_stopped);
490 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
492 * Called from interrupt entry when the CPU was idle
494 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
495 * must be updated. Otherwise an interrupt handler could use a stale jiffy
496 * value. We do this unconditionally on any CPU, as we don't know whether the
497 * CPU, which has the update task assigned is in a long sleep.
499 static void tick_nohz_update_jiffies(ktime_t now)
501 unsigned long flags;
503 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
505 local_irq_save(flags);
506 tick_do_update_jiffies64(now);
507 local_irq_restore(flags);
509 touch_softlockup_watchdog_sched();
513 * Updates the per-CPU time idle statistics counters
515 static void
516 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
518 ktime_t delta;
520 if (ts->idle_active) {
521 delta = ktime_sub(now, ts->idle_entrytime);
522 if (nr_iowait_cpu(cpu) > 0)
523 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
524 else
525 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
526 ts->idle_entrytime = now;
529 if (last_update_time)
530 *last_update_time = ktime_to_us(now);
534 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
536 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
537 ts->idle_active = 0;
539 sched_clock_idle_wakeup_event();
542 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
544 ktime_t now = ktime_get();
546 ts->idle_entrytime = now;
547 ts->idle_active = 1;
548 sched_clock_idle_sleep_event();
549 return now;
553 * get_cpu_idle_time_us - get the total idle time of a CPU
554 * @cpu: CPU number to query
555 * @last_update_time: variable to store update time in. Do not update
556 * counters if NULL.
558 * Return the cumulative idle time (since boot) for a given
559 * CPU, in microseconds.
561 * This time is measured via accounting rather than sampling,
562 * and is as accurate as ktime_get() is.
564 * This function returns -1 if NOHZ is not enabled.
566 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
568 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
569 ktime_t now, idle;
571 if (!tick_nohz_active)
572 return -1;
574 now = ktime_get();
575 if (last_update_time) {
576 update_ts_time_stats(cpu, ts, now, last_update_time);
577 idle = ts->idle_sleeptime;
578 } else {
579 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
580 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
582 idle = ktime_add(ts->idle_sleeptime, delta);
583 } else {
584 idle = ts->idle_sleeptime;
588 return ktime_to_us(idle);
591 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
594 * get_cpu_iowait_time_us - get the total iowait time of a CPU
595 * @cpu: CPU number to query
596 * @last_update_time: variable to store update time in. Do not update
597 * counters if NULL.
599 * Return the cumulative iowait time (since boot) for a given
600 * CPU, in microseconds.
602 * This time is measured via accounting rather than sampling,
603 * and is as accurate as ktime_get() is.
605 * This function returns -1 if NOHZ is not enabled.
607 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
609 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
610 ktime_t now, iowait;
612 if (!tick_nohz_active)
613 return -1;
615 now = ktime_get();
616 if (last_update_time) {
617 update_ts_time_stats(cpu, ts, now, last_update_time);
618 iowait = ts->iowait_sleeptime;
619 } else {
620 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
621 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
623 iowait = ktime_add(ts->iowait_sleeptime, delta);
624 } else {
625 iowait = ts->iowait_sleeptime;
629 return ktime_to_us(iowait);
631 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
633 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
635 hrtimer_cancel(&ts->sched_timer);
636 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
638 /* Forward the time to expire in the future */
639 hrtimer_forward(&ts->sched_timer, now, tick_period);
641 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
642 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
643 else
644 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
647 * Reset to make sure next tick stop doesn't get fooled by past
648 * cached clock deadline.
650 ts->next_tick = 0;
653 static inline bool local_timer_softirq_pending(void)
655 return local_softirq_pending() & TIMER_SOFTIRQ;
658 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
659 ktime_t now, int cpu)
661 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
662 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
663 unsigned long seq, basejiff;
664 ktime_t tick;
666 /* Read jiffies and the time when jiffies were updated last */
667 do {
668 seq = read_seqbegin(&jiffies_lock);
669 basemono = last_jiffies_update;
670 basejiff = jiffies;
671 } while (read_seqretry(&jiffies_lock, seq));
672 ts->last_jiffies = basejiff;
675 * Keep the periodic tick, when RCU, architecture or irq_work
676 * requests it.
677 * Aside of that check whether the local timer softirq is
678 * pending. If so its a bad idea to call get_next_timer_interrupt()
679 * because there is an already expired timer, so it will request
680 * immeditate expiry, which rearms the hardware timer with a
681 * minimal delta which brings us back to this place
682 * immediately. Lather, rinse and repeat...
684 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
685 irq_work_needs_cpu() || local_timer_softirq_pending()) {
686 next_tick = basemono + TICK_NSEC;
687 } else {
689 * Get the next pending timer. If high resolution
690 * timers are enabled this only takes the timer wheel
691 * timers into account. If high resolution timers are
692 * disabled this also looks at the next expiring
693 * hrtimer.
695 next_tmr = get_next_timer_interrupt(basejiff, basemono);
696 ts->next_timer = next_tmr;
697 /* Take the next rcu event into account */
698 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
702 * If the tick is due in the next period, keep it ticking or
703 * force prod the timer.
705 delta = next_tick - basemono;
706 if (delta <= (u64)TICK_NSEC) {
708 * Tell the timer code that the base is not idle, i.e. undo
709 * the effect of get_next_timer_interrupt():
711 timer_clear_idle();
713 * We've not stopped the tick yet, and there's a timer in the
714 * next period, so no point in stopping it either, bail.
716 if (!ts->tick_stopped) {
717 tick = 0;
718 goto out;
723 * If this CPU is the one which updates jiffies, then give up
724 * the assignment and let it be taken by the CPU which runs
725 * the tick timer next, which might be this CPU as well. If we
726 * don't drop this here the jiffies might be stale and
727 * do_timer() never invoked. Keep track of the fact that it
728 * was the one which had the do_timer() duty last. If this CPU
729 * is the one which had the do_timer() duty last, we limit the
730 * sleep time to the timekeeping max_deferment value.
731 * Otherwise we can sleep as long as we want.
733 delta = timekeeping_max_deferment();
734 if (cpu == tick_do_timer_cpu) {
735 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
736 ts->do_timer_last = 1;
737 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
738 delta = KTIME_MAX;
739 ts->do_timer_last = 0;
740 } else if (!ts->do_timer_last) {
741 delta = KTIME_MAX;
744 #ifdef CONFIG_NO_HZ_FULL
745 /* Limit the tick delta to the maximum scheduler deferment */
746 if (!ts->inidle)
747 delta = min(delta, scheduler_tick_max_deferment());
748 #endif
750 /* Calculate the next expiry time */
751 if (delta < (KTIME_MAX - basemono))
752 expires = basemono + delta;
753 else
754 expires = KTIME_MAX;
756 expires = min_t(u64, expires, next_tick);
757 tick = expires;
759 /* Skip reprogram of event if its not changed */
760 if (ts->tick_stopped && (expires == ts->next_tick)) {
761 /* Sanity check: make sure clockevent is actually programmed */
762 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
763 goto out;
765 WARN_ON_ONCE(1);
766 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
767 basemono, ts->next_tick, dev->next_event,
768 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
772 * nohz_stop_sched_tick can be called several times before
773 * the nohz_restart_sched_tick is called. This happens when
774 * interrupts arrive which do not cause a reschedule. In the
775 * first call we save the current tick time, so we can restart
776 * the scheduler tick in nohz_restart_sched_tick.
778 if (!ts->tick_stopped) {
779 calc_load_nohz_start();
780 cpu_load_update_nohz_start();
781 quiet_vmstat();
783 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
784 ts->tick_stopped = 1;
785 trace_tick_stop(1, TICK_DEP_MASK_NONE);
788 ts->next_tick = tick;
791 * If the expiration time == KTIME_MAX, then we simply stop
792 * the tick timer.
794 if (unlikely(expires == KTIME_MAX)) {
795 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
796 hrtimer_cancel(&ts->sched_timer);
797 goto out;
800 hrtimer_set_expires(&ts->sched_timer, tick);
802 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
803 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
804 else
805 tick_program_event(tick, 1);
806 out:
808 * Update the estimated sleep length until the next timer
809 * (not only the tick).
811 ts->sleep_length = ktime_sub(dev->next_event, now);
812 return tick;
815 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
817 /* Update jiffies first */
818 tick_do_update_jiffies64(now);
819 cpu_load_update_nohz_stop();
822 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
823 * the clock forward checks in the enqueue path:
825 timer_clear_idle();
827 calc_load_nohz_stop();
828 touch_softlockup_watchdog_sched();
830 * Cancel the scheduled timer and restore the tick
832 ts->tick_stopped = 0;
833 ts->idle_exittime = now;
835 tick_nohz_restart(ts, now);
838 static void tick_nohz_full_update_tick(struct tick_sched *ts)
840 #ifdef CONFIG_NO_HZ_FULL
841 int cpu = smp_processor_id();
843 if (!tick_nohz_full_cpu(cpu))
844 return;
846 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
847 return;
849 if (can_stop_full_tick(cpu, ts))
850 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
851 else if (ts->tick_stopped)
852 tick_nohz_restart_sched_tick(ts, ktime_get());
853 #endif
856 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
859 * If this CPU is offline and it is the one which updates
860 * jiffies, then give up the assignment and let it be taken by
861 * the CPU which runs the tick timer next. If we don't drop
862 * this here the jiffies might be stale and do_timer() never
863 * invoked.
865 if (unlikely(!cpu_online(cpu))) {
866 if (cpu == tick_do_timer_cpu)
867 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
869 * Make sure the CPU doesn't get fooled by obsolete tick
870 * deadline if it comes back online later.
872 ts->next_tick = 0;
873 return false;
876 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
877 ts->sleep_length = NSEC_PER_SEC / HZ;
878 return false;
881 if (need_resched())
882 return false;
884 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
885 static int ratelimit;
887 if (ratelimit < 10 &&
888 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
889 pr_warn("NOHZ: local_softirq_pending %02x\n",
890 (unsigned int) local_softirq_pending());
891 ratelimit++;
893 return false;
896 if (tick_nohz_full_enabled()) {
898 * Keep the tick alive to guarantee timekeeping progression
899 * if there are full dynticks CPUs around
901 if (tick_do_timer_cpu == cpu)
902 return false;
904 * Boot safety: make sure the timekeeping duty has been
905 * assigned before entering dyntick-idle mode,
907 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
908 return false;
911 return true;
914 static void __tick_nohz_idle_enter(struct tick_sched *ts)
916 ktime_t now, expires;
917 int cpu = smp_processor_id();
919 now = tick_nohz_start_idle(ts);
921 if (can_stop_idle_tick(cpu, ts)) {
922 int was_stopped = ts->tick_stopped;
924 ts->idle_calls++;
926 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
927 if (expires > 0LL) {
928 ts->idle_sleeps++;
929 ts->idle_expires = expires;
932 if (!was_stopped && ts->tick_stopped) {
933 ts->idle_jiffies = ts->last_jiffies;
934 nohz_balance_enter_idle(cpu);
940 * tick_nohz_idle_enter - stop the idle tick from the idle task
942 * When the next event is more than a tick into the future, stop the idle tick
943 * Called when we start the idle loop.
945 * The arch is responsible of calling:
947 * - rcu_idle_enter() after its last use of RCU before the CPU is put
948 * to sleep.
949 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
951 void tick_nohz_idle_enter(void)
953 struct tick_sched *ts;
955 lockdep_assert_irqs_enabled();
957 * Update the idle state in the scheduler domain hierarchy
958 * when tick_nohz_stop_sched_tick() is called from the idle loop.
959 * State will be updated to busy during the first busy tick after
960 * exiting idle.
962 set_cpu_sd_state_idle();
964 local_irq_disable();
966 ts = this_cpu_ptr(&tick_cpu_sched);
967 ts->inidle = 1;
968 __tick_nohz_idle_enter(ts);
970 local_irq_enable();
974 * tick_nohz_irq_exit - update next tick event from interrupt exit
976 * When an interrupt fires while we are idle and it doesn't cause
977 * a reschedule, it may still add, modify or delete a timer, enqueue
978 * an RCU callback, etc...
979 * So we need to re-calculate and reprogram the next tick event.
981 void tick_nohz_irq_exit(void)
983 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
985 if (ts->inidle)
986 __tick_nohz_idle_enter(ts);
987 else
988 tick_nohz_full_update_tick(ts);
992 * tick_nohz_get_sleep_length - return the length of the current sleep
994 * Called from power state control code with interrupts disabled
996 ktime_t tick_nohz_get_sleep_length(void)
998 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1000 return ts->sleep_length;
1004 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1005 * for a particular CPU.
1007 * Called from the schedutil frequency scaling governor in scheduler context.
1009 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1011 struct tick_sched *ts = tick_get_tick_sched(cpu);
1013 return ts->idle_calls;
1017 * tick_nohz_get_idle_calls - return the current idle calls counter value
1019 * Called from the schedutil frequency scaling governor in scheduler context.
1021 unsigned long tick_nohz_get_idle_calls(void)
1023 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1025 return ts->idle_calls;
1028 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1030 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1031 unsigned long ticks;
1033 if (vtime_accounting_cpu_enabled())
1034 return;
1036 * We stopped the tick in idle. Update process times would miss the
1037 * time we slept as update_process_times does only a 1 tick
1038 * accounting. Enforce that this is accounted to idle !
1040 ticks = jiffies - ts->idle_jiffies;
1042 * We might be one off. Do not randomly account a huge number of ticks!
1044 if (ticks && ticks < LONG_MAX)
1045 account_idle_ticks(ticks);
1046 #endif
1050 * tick_nohz_idle_exit - restart the idle tick from the idle task
1052 * Restart the idle tick when the CPU is woken up from idle
1053 * This also exit the RCU extended quiescent state. The CPU
1054 * can use RCU again after this function is called.
1056 void tick_nohz_idle_exit(void)
1058 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1059 ktime_t now;
1061 local_irq_disable();
1063 WARN_ON_ONCE(!ts->inidle);
1065 ts->inidle = 0;
1067 if (ts->idle_active || ts->tick_stopped)
1068 now = ktime_get();
1070 if (ts->idle_active)
1071 tick_nohz_stop_idle(ts, now);
1073 if (ts->tick_stopped) {
1074 tick_nohz_restart_sched_tick(ts, now);
1075 tick_nohz_account_idle_ticks(ts);
1078 local_irq_enable();
1082 * The nohz low res interrupt handler
1084 static void tick_nohz_handler(struct clock_event_device *dev)
1086 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1087 struct pt_regs *regs = get_irq_regs();
1088 ktime_t now = ktime_get();
1090 dev->next_event = KTIME_MAX;
1092 tick_sched_do_timer(now);
1093 tick_sched_handle(ts, regs);
1095 /* No need to reprogram if we are running tickless */
1096 if (unlikely(ts->tick_stopped))
1097 return;
1099 hrtimer_forward(&ts->sched_timer, now, tick_period);
1100 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1103 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1105 if (!tick_nohz_enabled)
1106 return;
1107 ts->nohz_mode = mode;
1108 /* One update is enough */
1109 if (!test_and_set_bit(0, &tick_nohz_active))
1110 timers_update_nohz();
1114 * tick_nohz_switch_to_nohz - switch to nohz mode
1116 static void tick_nohz_switch_to_nohz(void)
1118 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1119 ktime_t next;
1121 if (!tick_nohz_enabled)
1122 return;
1124 if (tick_switch_to_oneshot(tick_nohz_handler))
1125 return;
1128 * Recycle the hrtimer in ts, so we can share the
1129 * hrtimer_forward with the highres code.
1131 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1132 /* Get the next period */
1133 next = tick_init_jiffy_update();
1135 hrtimer_set_expires(&ts->sched_timer, next);
1136 hrtimer_forward_now(&ts->sched_timer, tick_period);
1137 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1138 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1141 static inline void tick_nohz_irq_enter(void)
1143 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1144 ktime_t now;
1146 if (!ts->idle_active && !ts->tick_stopped)
1147 return;
1148 now = ktime_get();
1149 if (ts->idle_active)
1150 tick_nohz_stop_idle(ts, now);
1151 if (ts->tick_stopped)
1152 tick_nohz_update_jiffies(now);
1155 #else
1157 static inline void tick_nohz_switch_to_nohz(void) { }
1158 static inline void tick_nohz_irq_enter(void) { }
1159 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1161 #endif /* CONFIG_NO_HZ_COMMON */
1164 * Called from irq_enter to notify about the possible interruption of idle()
1166 void tick_irq_enter(void)
1168 tick_check_oneshot_broadcast_this_cpu();
1169 tick_nohz_irq_enter();
1173 * High resolution timer specific code
1175 #ifdef CONFIG_HIGH_RES_TIMERS
1177 * We rearm the timer until we get disabled by the idle code.
1178 * Called with interrupts disabled.
1180 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1182 struct tick_sched *ts =
1183 container_of(timer, struct tick_sched, sched_timer);
1184 struct pt_regs *regs = get_irq_regs();
1185 ktime_t now = ktime_get();
1187 tick_sched_do_timer(now);
1190 * Do not call, when we are not in irq context and have
1191 * no valid regs pointer
1193 if (regs)
1194 tick_sched_handle(ts, regs);
1195 else
1196 ts->next_tick = 0;
1198 /* No need to reprogram if we are in idle or full dynticks mode */
1199 if (unlikely(ts->tick_stopped))
1200 return HRTIMER_NORESTART;
1202 hrtimer_forward(timer, now, tick_period);
1204 return HRTIMER_RESTART;
1207 static int sched_skew_tick;
1209 static int __init skew_tick(char *str)
1211 get_option(&str, &sched_skew_tick);
1213 return 0;
1215 early_param("skew_tick", skew_tick);
1218 * tick_setup_sched_timer - setup the tick emulation timer
1220 void tick_setup_sched_timer(void)
1222 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 ktime_t now = ktime_get();
1226 * Emulate tick processing via per-CPU hrtimers:
1228 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1229 ts->sched_timer.function = tick_sched_timer;
1231 /* Get the next period (per-CPU) */
1232 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1234 /* Offset the tick to avert jiffies_lock contention. */
1235 if (sched_skew_tick) {
1236 u64 offset = ktime_to_ns(tick_period) >> 1;
1237 do_div(offset, num_possible_cpus());
1238 offset *= smp_processor_id();
1239 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1242 hrtimer_forward(&ts->sched_timer, now, tick_period);
1243 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1244 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1246 #endif /* HIGH_RES_TIMERS */
1248 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1249 void tick_cancel_sched_timer(int cpu)
1251 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1253 # ifdef CONFIG_HIGH_RES_TIMERS
1254 if (ts->sched_timer.base)
1255 hrtimer_cancel(&ts->sched_timer);
1256 # endif
1258 memset(ts, 0, sizeof(*ts));
1260 #endif
1263 * Async notification about clocksource changes
1265 void tick_clock_notify(void)
1267 int cpu;
1269 for_each_possible_cpu(cpu)
1270 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1274 * Async notification about clock event changes
1276 void tick_oneshot_notify(void)
1278 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1280 set_bit(0, &ts->check_clocks);
1284 * Check, if a change happened, which makes oneshot possible.
1286 * Called cyclic from the hrtimer softirq (driven by the timer
1287 * softirq) allow_nohz signals, that we can switch into low-res nohz
1288 * mode, because high resolution timers are disabled (either compile
1289 * or runtime). Called with interrupts disabled.
1291 int tick_check_oneshot_change(int allow_nohz)
1293 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1295 if (!test_and_clear_bit(0, &ts->check_clocks))
1296 return 0;
1298 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1299 return 0;
1301 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1302 return 0;
1304 if (!allow_nohz)
1305 return 1;
1307 tick_nohz_switch_to_nohz();
1308 return 0;