1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 * No idle tick implementation for low and high resolution timers
9 * Started by: Thomas Gleixner and Ingo Molnar
11 #include <linux/cpu.h>
12 #include <linux/err.h>
13 #include <linux/hrtimer.h>
14 #include <linux/interrupt.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/percpu.h>
17 #include <linux/nmi.h>
18 #include <linux/profile.h>
19 #include <linux/sched/signal.h>
20 #include <linux/sched/clock.h>
21 #include <linux/sched/stat.h>
22 #include <linux/sched/nohz.h>
23 #include <linux/sched/loadavg.h>
24 #include <linux/module.h>
25 #include <linux/irq_work.h>
26 #include <linux/posix-timers.h>
27 #include <linux/context_tracking.h>
30 #include <asm/irq_regs.h>
32 #include "tick-internal.h"
34 #include <trace/events/timer.h>
37 * Per-CPU nohz control structure
39 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
41 struct tick_sched *tick_get_tick_sched(int cpu)
43 return &per_cpu(tick_cpu_sched, cpu);
46 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
48 * The time, when the last jiffy update happened. Write access must hold
49 * jiffies_lock and jiffies_seq. tick_nohz_next_event() needs to get a
50 * consistent view of jiffies and last_jiffies_update.
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 = 1;
63 * 64bit can do a quick check without holding jiffies lock and
64 * without looking at the sequence count. The smp_load_acquire()
65 * pairs with the update done later in this function.
67 * 32bit cannot do that because the store of tick_next_period
68 * consists of two 32bit stores and the first store could move it
69 * to a random point in the future.
71 if (IS_ENABLED(CONFIG_64BIT)) {
72 if (ktime_before(now, smp_load_acquire(&tick_next_period)))
78 * Avoid contention on jiffies_lock and protect the quick
79 * check with the sequence count.
82 seq = read_seqcount_begin(&jiffies_seq);
83 nextp = tick_next_period;
84 } while (read_seqcount_retry(&jiffies_seq, seq));
86 if (ktime_before(now, nextp))
90 /* Quick check failed, i.e. update is required. */
91 raw_spin_lock(&jiffies_lock);
93 * Reevaluate with the lock held. Another CPU might have done the
96 if (ktime_before(now, tick_next_period)) {
97 raw_spin_unlock(&jiffies_lock);
101 write_seqcount_begin(&jiffies_seq);
103 delta = ktime_sub(now, tick_next_period);
104 if (unlikely(delta >= TICK_NSEC)) {
105 /* Slow path for long idle sleep times */
106 s64 incr = TICK_NSEC;
108 ticks += ktime_divns(delta, incr);
110 last_jiffies_update = ktime_add_ns(last_jiffies_update,
113 last_jiffies_update = ktime_add_ns(last_jiffies_update,
117 /* Advance jiffies to complete the jiffies_seq protected job */
121 * Keep the tick_next_period variable up to date.
123 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
125 if (IS_ENABLED(CONFIG_64BIT)) {
127 * Pairs with smp_load_acquire() in the lockless quick
128 * check above and ensures that the update to jiffies_64 is
129 * not reordered vs. the store to tick_next_period, neither
130 * by the compiler nor by the CPU.
132 smp_store_release(&tick_next_period, nextp);
135 * A plain store is good enough on 32bit as the quick check
136 * above is protected by the sequence count.
138 tick_next_period = nextp;
142 * Release the sequence count. calc_global_load() below is not
143 * protected by it, but jiffies_lock needs to be held to prevent
144 * concurrent invocations.
146 write_seqcount_end(&jiffies_seq);
150 raw_spin_unlock(&jiffies_lock);
155 * Initialize and return retrieve the jiffies update.
157 static ktime_t tick_init_jiffy_update(void)
161 raw_spin_lock(&jiffies_lock);
162 write_seqcount_begin(&jiffies_seq);
163 /* Did we start the jiffies update yet ? */
164 if (last_jiffies_update == 0)
165 last_jiffies_update = tick_next_period;
166 period = last_jiffies_update;
167 write_seqcount_end(&jiffies_seq);
168 raw_spin_unlock(&jiffies_lock);
172 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
174 int cpu = smp_processor_id();
176 #ifdef CONFIG_NO_HZ_COMMON
178 * Check if the do_timer duty was dropped. We don't care about
179 * concurrency: This happens only when the CPU in charge went
180 * into a long sleep. If two CPUs happen to assign themselves to
181 * this duty, then the jiffies update is still serialized by
184 * If nohz_full is enabled, this should not happen because the
185 * tick_do_timer_cpu never relinquishes.
187 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
188 #ifdef CONFIG_NO_HZ_FULL
189 WARN_ON(tick_nohz_full_running);
191 tick_do_timer_cpu = cpu;
195 /* Check, if the jiffies need an update */
196 if (tick_do_timer_cpu == cpu)
197 tick_do_update_jiffies64(now);
200 ts->got_idle_tick = 1;
203 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
205 #ifdef CONFIG_NO_HZ_COMMON
207 * When we are idle and the tick is stopped, we have to touch
208 * the watchdog as we might not schedule for a really long
209 * time. This happens on complete idle SMP systems while
210 * waiting on the login prompt. We also increment the "start of
211 * idle" jiffy stamp so the idle accounting adjustment we do
212 * when we go busy again does not account too much ticks.
214 if (ts->tick_stopped) {
215 touch_softlockup_watchdog_sched();
216 if (is_idle_task(current))
219 * In case the current tick fired too early past its expected
220 * expiration, make sure we don't bypass the next clock reprogramming
221 * to the same deadline.
226 update_process_times(user_mode(regs));
227 profile_tick(CPU_PROFILING);
231 #ifdef CONFIG_NO_HZ_FULL
232 cpumask_var_t tick_nohz_full_mask;
233 bool tick_nohz_full_running;
234 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
235 static atomic_t tick_dep_mask;
237 static bool check_tick_dependency(atomic_t *dep)
239 int val = atomic_read(dep);
241 if (val & TICK_DEP_MASK_POSIX_TIMER) {
242 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
246 if (val & TICK_DEP_MASK_PERF_EVENTS) {
247 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
251 if (val & TICK_DEP_MASK_SCHED) {
252 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
256 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
257 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
261 if (val & TICK_DEP_MASK_RCU) {
262 trace_tick_stop(0, TICK_DEP_MASK_RCU);
269 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
271 lockdep_assert_irqs_disabled();
273 if (unlikely(!cpu_online(cpu)))
276 if (check_tick_dependency(&tick_dep_mask))
279 if (check_tick_dependency(&ts->tick_dep_mask))
282 if (check_tick_dependency(¤t->tick_dep_mask))
285 if (check_tick_dependency(¤t->signal->tick_dep_mask))
291 static void nohz_full_kick_func(struct irq_work *work)
293 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
296 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
297 .func = nohz_full_kick_func,
298 .flags = ATOMIC_INIT(IRQ_WORK_HARD_IRQ),
302 * Kick this CPU if it's full dynticks in order to force it to
303 * re-evaluate its dependency on the tick and restart it if necessary.
304 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
307 static void tick_nohz_full_kick(void)
309 if (!tick_nohz_full_cpu(smp_processor_id()))
312 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
316 * Kick the CPU if it's full dynticks in order to force it to
317 * re-evaluate its dependency on the tick and restart it if necessary.
319 void tick_nohz_full_kick_cpu(int cpu)
321 if (!tick_nohz_full_cpu(cpu))
324 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
328 * Kick all full dynticks CPUs in order to force these to re-evaluate
329 * their dependency on the tick and restart it if necessary.
331 static void tick_nohz_full_kick_all(void)
335 if (!tick_nohz_full_running)
339 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
340 tick_nohz_full_kick_cpu(cpu);
344 static void tick_nohz_dep_set_all(atomic_t *dep,
345 enum tick_dep_bits bit)
349 prev = atomic_fetch_or(BIT(bit), dep);
351 tick_nohz_full_kick_all();
355 * Set a global tick dependency. Used by perf events that rely on freq and
358 void tick_nohz_dep_set(enum tick_dep_bits bit)
360 tick_nohz_dep_set_all(&tick_dep_mask, bit);
363 void tick_nohz_dep_clear(enum tick_dep_bits bit)
365 atomic_andnot(BIT(bit), &tick_dep_mask);
369 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
370 * manage events throttling.
372 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
375 struct tick_sched *ts;
377 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
379 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
382 /* Perf needs local kick that is NMI safe */
383 if (cpu == smp_processor_id()) {
384 tick_nohz_full_kick();
386 /* Remote irq work not NMI-safe */
387 if (!WARN_ON_ONCE(in_nmi()))
388 tick_nohz_full_kick_cpu(cpu);
393 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
395 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
397 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
399 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
401 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
404 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
405 * in order to elapse per task timers.
407 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
409 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask)) {
410 if (tsk == current) {
412 tick_nohz_full_kick();
416 * Some future tick_nohz_full_kick_task()
417 * should optimize this.
419 tick_nohz_full_kick_all();
423 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
425 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
427 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
429 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
432 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
433 * per process timers.
435 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
437 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
440 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
442 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
446 * Re-evaluate the need for the tick as we switch the current task.
447 * It might need the tick due to per task/process properties:
448 * perf events, posix CPU timers, ...
450 void __tick_nohz_task_switch(void)
453 struct tick_sched *ts;
455 local_irq_save(flags);
457 if (!tick_nohz_full_cpu(smp_processor_id()))
460 ts = this_cpu_ptr(&tick_cpu_sched);
462 if (ts->tick_stopped) {
463 if (atomic_read(¤t->tick_dep_mask) ||
464 atomic_read(¤t->signal->tick_dep_mask))
465 tick_nohz_full_kick();
468 local_irq_restore(flags);
471 /* Get the boot-time nohz CPU list from the kernel parameters. */
472 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
474 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
475 cpumask_copy(tick_nohz_full_mask, cpumask);
476 tick_nohz_full_running = true;
478 EXPORT_SYMBOL_GPL(tick_nohz_full_setup);
480 static int tick_nohz_cpu_down(unsigned int cpu)
483 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
484 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
485 * CPUs. It must remain online when nohz full is enabled.
487 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
492 void __init tick_nohz_init(void)
496 if (!tick_nohz_full_running)
500 * Full dynticks uses irq work to drive the tick rescheduling on safe
501 * locking contexts. But then we need irq work to raise its own
502 * interrupts to avoid circular dependency on the tick
504 if (!arch_irq_work_has_interrupt()) {
505 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
506 cpumask_clear(tick_nohz_full_mask);
507 tick_nohz_full_running = false;
511 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
512 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
513 cpu = smp_processor_id();
515 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
516 pr_warn("NO_HZ: Clearing %d from nohz_full range "
517 "for timekeeping\n", cpu);
518 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
522 for_each_cpu(cpu, tick_nohz_full_mask)
523 context_tracking_cpu_set(cpu);
525 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
526 "kernel/nohz:predown", NULL,
529 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
530 cpumask_pr_args(tick_nohz_full_mask));
535 * NOHZ - aka dynamic tick functionality
537 #ifdef CONFIG_NO_HZ_COMMON
541 bool tick_nohz_enabled __read_mostly = true;
542 unsigned long tick_nohz_active __read_mostly;
544 * Enable / Disable tickless mode
546 static int __init setup_tick_nohz(char *str)
548 return (kstrtobool(str, &tick_nohz_enabled) == 0);
551 __setup("nohz=", setup_tick_nohz);
553 bool tick_nohz_tick_stopped(void)
555 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
557 return ts->tick_stopped;
560 bool tick_nohz_tick_stopped_cpu(int cpu)
562 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
564 return ts->tick_stopped;
568 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
570 * Called from interrupt entry when the CPU was idle
572 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
573 * must be updated. Otherwise an interrupt handler could use a stale jiffy
574 * value. We do this unconditionally on any CPU, as we don't know whether the
575 * CPU, which has the update task assigned is in a long sleep.
577 static void tick_nohz_update_jiffies(ktime_t now)
581 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
583 local_irq_save(flags);
584 tick_do_update_jiffies64(now);
585 local_irq_restore(flags);
587 touch_softlockup_watchdog_sched();
591 * Updates the per-CPU time idle statistics counters
594 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
598 if (ts->idle_active) {
599 delta = ktime_sub(now, ts->idle_entrytime);
600 if (nr_iowait_cpu(cpu) > 0)
601 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
603 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
604 ts->idle_entrytime = now;
607 if (last_update_time)
608 *last_update_time = ktime_to_us(now);
612 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
614 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
617 sched_clock_idle_wakeup_event();
620 static void tick_nohz_start_idle(struct tick_sched *ts)
622 ts->idle_entrytime = ktime_get();
624 sched_clock_idle_sleep_event();
628 * get_cpu_idle_time_us - get the total idle time of a CPU
629 * @cpu: CPU number to query
630 * @last_update_time: variable to store update time in. Do not update
633 * Return the cumulative idle time (since boot) for a given
634 * CPU, in microseconds.
636 * This time is measured via accounting rather than sampling,
637 * and is as accurate as ktime_get() is.
639 * This function returns -1 if NOHZ is not enabled.
641 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
643 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
646 if (!tick_nohz_active)
650 if (last_update_time) {
651 update_ts_time_stats(cpu, ts, now, last_update_time);
652 idle = ts->idle_sleeptime;
654 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
655 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
657 idle = ktime_add(ts->idle_sleeptime, delta);
659 idle = ts->idle_sleeptime;
663 return ktime_to_us(idle);
666 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
669 * get_cpu_iowait_time_us - get the total iowait time of a CPU
670 * @cpu: CPU number to query
671 * @last_update_time: variable to store update time in. Do not update
674 * Return the cumulative iowait time (since boot) for a given
675 * CPU, in microseconds.
677 * This time is measured via accounting rather than sampling,
678 * and is as accurate as ktime_get() is.
680 * This function returns -1 if NOHZ is not enabled.
682 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
684 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
687 if (!tick_nohz_active)
691 if (last_update_time) {
692 update_ts_time_stats(cpu, ts, now, last_update_time);
693 iowait = ts->iowait_sleeptime;
695 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
696 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
698 iowait = ktime_add(ts->iowait_sleeptime, delta);
700 iowait = ts->iowait_sleeptime;
704 return ktime_to_us(iowait);
706 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
708 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
710 hrtimer_cancel(&ts->sched_timer);
711 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
713 /* Forward the time to expire in the future */
714 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
716 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
717 hrtimer_start_expires(&ts->sched_timer,
718 HRTIMER_MODE_ABS_PINNED_HARD);
720 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
724 * Reset to make sure next tick stop doesn't get fooled by past
725 * cached clock deadline.
730 static inline bool local_timer_softirq_pending(void)
732 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
735 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
737 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
738 unsigned long basejiff;
741 /* Read jiffies and the time when jiffies were updated last */
743 seq = read_seqcount_begin(&jiffies_seq);
744 basemono = last_jiffies_update;
746 } while (read_seqcount_retry(&jiffies_seq, seq));
747 ts->last_jiffies = basejiff;
748 ts->timer_expires_base = basemono;
751 * Keep the periodic tick, when RCU, architecture or irq_work
753 * Aside of that check whether the local timer softirq is
754 * pending. If so its a bad idea to call get_next_timer_interrupt()
755 * because there is an already expired timer, so it will request
756 * immeditate expiry, which rearms the hardware timer with a
757 * minimal delta which brings us back to this place
758 * immediately. Lather, rinse and repeat...
760 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
761 irq_work_needs_cpu() || local_timer_softirq_pending()) {
762 next_tick = basemono + TICK_NSEC;
765 * Get the next pending timer. If high resolution
766 * timers are enabled this only takes the timer wheel
767 * timers into account. If high resolution timers are
768 * disabled this also looks at the next expiring
771 next_tmr = get_next_timer_interrupt(basejiff, basemono);
772 ts->next_timer = next_tmr;
773 /* Take the next rcu event into account */
774 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
778 * If the tick is due in the next period, keep it ticking or
779 * force prod the timer.
781 delta = next_tick - basemono;
782 if (delta <= (u64)TICK_NSEC) {
784 * Tell the timer code that the base is not idle, i.e. undo
785 * the effect of get_next_timer_interrupt():
789 * We've not stopped the tick yet, and there's a timer in the
790 * next period, so no point in stopping it either, bail.
792 if (!ts->tick_stopped) {
793 ts->timer_expires = 0;
799 * If this CPU is the one which had the do_timer() duty last, we limit
800 * the sleep time to the timekeeping max_deferment value.
801 * Otherwise we can sleep as long as we want.
803 delta = timekeeping_max_deferment();
804 if (cpu != tick_do_timer_cpu &&
805 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
808 /* Calculate the next expiry time */
809 if (delta < (KTIME_MAX - basemono))
810 expires = basemono + delta;
814 ts->timer_expires = min_t(u64, expires, next_tick);
817 return ts->timer_expires;
820 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
822 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
823 u64 basemono = ts->timer_expires_base;
824 u64 expires = ts->timer_expires;
825 ktime_t tick = expires;
827 /* Make sure we won't be trying to stop it twice in a row. */
828 ts->timer_expires_base = 0;
831 * If this CPU is the one which updates jiffies, then give up
832 * the assignment and let it be taken by the CPU which runs
833 * the tick timer next, which might be this CPU as well. If we
834 * don't drop this here the jiffies might be stale and
835 * do_timer() never invoked. Keep track of the fact that it
836 * was the one which had the do_timer() duty last.
838 if (cpu == tick_do_timer_cpu) {
839 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
840 ts->do_timer_last = 1;
841 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
842 ts->do_timer_last = 0;
845 /* Skip reprogram of event if its not changed */
846 if (ts->tick_stopped && (expires == ts->next_tick)) {
847 /* Sanity check: make sure clockevent is actually programmed */
848 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
852 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
853 basemono, ts->next_tick, dev->next_event,
854 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
858 * nohz_stop_sched_tick can be called several times before
859 * the nohz_restart_sched_tick is called. This happens when
860 * interrupts arrive which do not cause a reschedule. In the
861 * first call we save the current tick time, so we can restart
862 * the scheduler tick in nohz_restart_sched_tick.
864 if (!ts->tick_stopped) {
865 calc_load_nohz_start();
868 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
869 ts->tick_stopped = 1;
870 trace_tick_stop(1, TICK_DEP_MASK_NONE);
873 ts->next_tick = tick;
876 * If the expiration time == KTIME_MAX, then we simply stop
879 if (unlikely(expires == KTIME_MAX)) {
880 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
881 hrtimer_cancel(&ts->sched_timer);
885 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
886 hrtimer_start(&ts->sched_timer, tick,
887 HRTIMER_MODE_ABS_PINNED_HARD);
889 hrtimer_set_expires(&ts->sched_timer, tick);
890 tick_program_event(tick, 1);
894 static void tick_nohz_retain_tick(struct tick_sched *ts)
896 ts->timer_expires_base = 0;
899 #ifdef CONFIG_NO_HZ_FULL
900 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
902 if (tick_nohz_next_event(ts, cpu))
903 tick_nohz_stop_tick(ts, cpu);
905 tick_nohz_retain_tick(ts);
907 #endif /* CONFIG_NO_HZ_FULL */
909 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
911 /* Update jiffies first */
912 tick_do_update_jiffies64(now);
915 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
916 * the clock forward checks in the enqueue path:
920 calc_load_nohz_stop();
921 touch_softlockup_watchdog_sched();
923 * Cancel the scheduled timer and restore the tick
925 ts->tick_stopped = 0;
926 ts->idle_exittime = now;
928 tick_nohz_restart(ts, now);
931 static void tick_nohz_full_update_tick(struct tick_sched *ts)
933 #ifdef CONFIG_NO_HZ_FULL
934 int cpu = smp_processor_id();
936 if (!tick_nohz_full_cpu(cpu))
939 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
942 if (can_stop_full_tick(cpu, ts))
943 tick_nohz_stop_sched_tick(ts, cpu);
944 else if (ts->tick_stopped)
945 tick_nohz_restart_sched_tick(ts, ktime_get());
949 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
952 * If this CPU is offline and it is the one which updates
953 * jiffies, then give up the assignment and let it be taken by
954 * the CPU which runs the tick timer next. If we don't drop
955 * this here the jiffies might be stale and do_timer() never
958 if (unlikely(!cpu_online(cpu))) {
959 if (cpu == tick_do_timer_cpu)
960 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
962 * Make sure the CPU doesn't get fooled by obsolete tick
963 * deadline if it comes back online later.
969 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
975 if (unlikely(local_softirq_pending())) {
976 static int ratelimit;
978 if (ratelimit < 10 &&
979 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
980 pr_warn("NOHZ tick-stop error: Non-RCU local softirq work is pending, handler #%02x!!!\n",
981 (unsigned int) local_softirq_pending());
987 if (tick_nohz_full_enabled()) {
989 * Keep the tick alive to guarantee timekeeping progression
990 * if there are full dynticks CPUs around
992 if (tick_do_timer_cpu == cpu)
995 /* Should not happen for nohz-full */
996 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1003 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1006 int cpu = smp_processor_id();
1009 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1010 * tick timer expiration time is known already.
1012 if (ts->timer_expires_base)
1013 expires = ts->timer_expires;
1014 else if (can_stop_idle_tick(cpu, ts))
1015 expires = tick_nohz_next_event(ts, cpu);
1021 if (expires > 0LL) {
1022 int was_stopped = ts->tick_stopped;
1024 tick_nohz_stop_tick(ts, cpu);
1027 ts->idle_expires = expires;
1029 if (!was_stopped && ts->tick_stopped) {
1030 ts->idle_jiffies = ts->last_jiffies;
1031 nohz_balance_enter_idle(cpu);
1034 tick_nohz_retain_tick(ts);
1039 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1041 * When the next event is more than a tick into the future, stop the idle tick
1043 void tick_nohz_idle_stop_tick(void)
1045 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1048 void tick_nohz_idle_retain_tick(void)
1050 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1052 * Undo the effect of get_next_timer_interrupt() called from
1053 * tick_nohz_next_event().
1059 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1061 * Called when we start the idle loop.
1063 void tick_nohz_idle_enter(void)
1065 struct tick_sched *ts;
1067 lockdep_assert_irqs_enabled();
1069 local_irq_disable();
1071 ts = this_cpu_ptr(&tick_cpu_sched);
1073 WARN_ON_ONCE(ts->timer_expires_base);
1076 tick_nohz_start_idle(ts);
1082 * tick_nohz_irq_exit - update next tick event from interrupt exit
1084 * When an interrupt fires while we are idle and it doesn't cause
1085 * a reschedule, it may still add, modify or delete a timer, enqueue
1086 * an RCU callback, etc...
1087 * So we need to re-calculate and reprogram the next tick event.
1089 void tick_nohz_irq_exit(void)
1091 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1094 tick_nohz_start_idle(ts);
1096 tick_nohz_full_update_tick(ts);
1100 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1102 bool tick_nohz_idle_got_tick(void)
1104 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1106 if (ts->got_idle_tick) {
1107 ts->got_idle_tick = 0;
1114 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1115 * or the tick, whatever that expires first. Note that, if the tick has been
1116 * stopped, it returns the next hrtimer.
1118 * Called from power state control code with interrupts disabled
1120 ktime_t tick_nohz_get_next_hrtimer(void)
1122 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1126 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1127 * @delta_next: duration until the next event if the tick cannot be stopped
1129 * Called from power state control code with interrupts disabled
1131 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1133 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1134 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1135 int cpu = smp_processor_id();
1137 * The idle entry time is expected to be a sufficient approximation of
1138 * the current time at this point.
1140 ktime_t now = ts->idle_entrytime;
1143 WARN_ON_ONCE(!ts->inidle);
1145 *delta_next = ktime_sub(dev->next_event, now);
1147 if (!can_stop_idle_tick(cpu, ts))
1150 next_event = tick_nohz_next_event(ts, cpu);
1155 * If the next highres timer to expire is earlier than next_event, the
1156 * idle governor needs to know that.
1158 next_event = min_t(u64, next_event,
1159 hrtimer_next_event_without(&ts->sched_timer));
1161 return ktime_sub(next_event, now);
1165 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1166 * for a particular CPU.
1168 * Called from the schedutil frequency scaling governor in scheduler context.
1170 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1172 struct tick_sched *ts = tick_get_tick_sched(cpu);
1174 return ts->idle_calls;
1178 * tick_nohz_get_idle_calls - return the current idle calls counter value
1180 * Called from the schedutil frequency scaling governor in scheduler context.
1182 unsigned long tick_nohz_get_idle_calls(void)
1184 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1186 return ts->idle_calls;
1189 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1191 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1192 unsigned long ticks;
1194 if (vtime_accounting_enabled_this_cpu())
1197 * We stopped the tick in idle. Update process times would miss the
1198 * time we slept as update_process_times does only a 1 tick
1199 * accounting. Enforce that this is accounted to idle !
1201 ticks = jiffies - ts->idle_jiffies;
1203 * We might be one off. Do not randomly account a huge number of ticks!
1205 if (ticks && ticks < LONG_MAX)
1206 account_idle_ticks(ticks);
1210 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1212 tick_nohz_restart_sched_tick(ts, now);
1213 tick_nohz_account_idle_ticks(ts);
1216 void tick_nohz_idle_restart_tick(void)
1218 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1220 if (ts->tick_stopped)
1221 __tick_nohz_idle_restart_tick(ts, ktime_get());
1225 * tick_nohz_idle_exit - restart the idle tick from the idle task
1227 * Restart the idle tick when the CPU is woken up from idle
1228 * This also exit the RCU extended quiescent state. The CPU
1229 * can use RCU again after this function is called.
1231 void tick_nohz_idle_exit(void)
1233 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1234 bool idle_active, tick_stopped;
1237 local_irq_disable();
1239 WARN_ON_ONCE(!ts->inidle);
1240 WARN_ON_ONCE(ts->timer_expires_base);
1243 idle_active = ts->idle_active;
1244 tick_stopped = ts->tick_stopped;
1246 if (idle_active || tick_stopped)
1250 tick_nohz_stop_idle(ts, now);
1253 __tick_nohz_idle_restart_tick(ts, now);
1259 * The nohz low res interrupt handler
1261 static void tick_nohz_handler(struct clock_event_device *dev)
1263 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1264 struct pt_regs *regs = get_irq_regs();
1265 ktime_t now = ktime_get();
1267 dev->next_event = KTIME_MAX;
1269 tick_sched_do_timer(ts, now);
1270 tick_sched_handle(ts, regs);
1272 /* No need to reprogram if we are running tickless */
1273 if (unlikely(ts->tick_stopped))
1276 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1277 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1280 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1282 if (!tick_nohz_enabled)
1284 ts->nohz_mode = mode;
1285 /* One update is enough */
1286 if (!test_and_set_bit(0, &tick_nohz_active))
1287 timers_update_nohz();
1291 * tick_nohz_switch_to_nohz - switch to nohz mode
1293 static void tick_nohz_switch_to_nohz(void)
1295 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1298 if (!tick_nohz_enabled)
1301 if (tick_switch_to_oneshot(tick_nohz_handler))
1305 * Recycle the hrtimer in ts, so we can share the
1306 * hrtimer_forward with the highres code.
1308 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1309 /* Get the next period */
1310 next = tick_init_jiffy_update();
1312 hrtimer_set_expires(&ts->sched_timer, next);
1313 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1314 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1315 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1318 static inline void tick_nohz_irq_enter(void)
1320 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1323 if (!ts->idle_active && !ts->tick_stopped)
1326 if (ts->idle_active)
1327 tick_nohz_stop_idle(ts, now);
1328 if (ts->tick_stopped)
1329 tick_nohz_update_jiffies(now);
1334 static inline void tick_nohz_switch_to_nohz(void) { }
1335 static inline void tick_nohz_irq_enter(void) { }
1336 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1338 #endif /* CONFIG_NO_HZ_COMMON */
1341 * Called from irq_enter to notify about the possible interruption of idle()
1343 void tick_irq_enter(void)
1345 tick_check_oneshot_broadcast_this_cpu();
1346 tick_nohz_irq_enter();
1350 * High resolution timer specific code
1352 #ifdef CONFIG_HIGH_RES_TIMERS
1354 * We rearm the timer until we get disabled by the idle code.
1355 * Called with interrupts disabled.
1357 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1359 struct tick_sched *ts =
1360 container_of(timer, struct tick_sched, sched_timer);
1361 struct pt_regs *regs = get_irq_regs();
1362 ktime_t now = ktime_get();
1364 tick_sched_do_timer(ts, now);
1367 * Do not call, when we are not in irq context and have
1368 * no valid regs pointer
1371 tick_sched_handle(ts, regs);
1375 /* No need to reprogram if we are in idle or full dynticks mode */
1376 if (unlikely(ts->tick_stopped))
1377 return HRTIMER_NORESTART;
1379 hrtimer_forward(timer, now, TICK_NSEC);
1381 return HRTIMER_RESTART;
1384 static int sched_skew_tick;
1386 static int __init skew_tick(char *str)
1388 get_option(&str, &sched_skew_tick);
1392 early_param("skew_tick", skew_tick);
1395 * tick_setup_sched_timer - setup the tick emulation timer
1397 void tick_setup_sched_timer(void)
1399 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1400 ktime_t now = ktime_get();
1403 * Emulate tick processing via per-CPU hrtimers:
1405 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1406 ts->sched_timer.function = tick_sched_timer;
1408 /* Get the next period (per-CPU) */
1409 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1411 /* Offset the tick to avert jiffies_lock contention. */
1412 if (sched_skew_tick) {
1413 u64 offset = TICK_NSEC >> 1;
1414 do_div(offset, num_possible_cpus());
1415 offset *= smp_processor_id();
1416 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1419 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1420 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1421 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1423 #endif /* HIGH_RES_TIMERS */
1425 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1426 void tick_cancel_sched_timer(int cpu)
1428 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1430 # ifdef CONFIG_HIGH_RES_TIMERS
1431 if (ts->sched_timer.base)
1432 hrtimer_cancel(&ts->sched_timer);
1435 memset(ts, 0, sizeof(*ts));
1440 * Async notification about clocksource changes
1442 void tick_clock_notify(void)
1446 for_each_possible_cpu(cpu)
1447 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1451 * Async notification about clock event changes
1453 void tick_oneshot_notify(void)
1455 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1457 set_bit(0, &ts->check_clocks);
1461 * Check, if a change happened, which makes oneshot possible.
1463 * Called cyclic from the hrtimer softirq (driven by the timer
1464 * softirq) allow_nohz signals, that we can switch into low-res nohz
1465 * mode, because high resolution timers are disabled (either compile
1466 * or runtime). Called with interrupts disabled.
1468 int tick_check_oneshot_change(int allow_nohz)
1470 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1472 if (!test_and_clear_bit(0, &ts->check_clocks))
1475 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1478 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1484 tick_nohz_switch_to_nohz();