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) {
168 * Ensure that the tick is aligned to a multiple of
171 div_u64_rem(tick_next_period, TICK_NSEC, &rem);
173 tick_next_period += TICK_NSEC - rem;
175 last_jiffies_update = tick_next_period;
177 period = last_jiffies_update;
178 write_seqcount_end(&jiffies_seq);
179 raw_spin_unlock(&jiffies_lock);
183 #define MAX_STALLED_JIFFIES 5
185 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
187 int cpu = smp_processor_id();
189 #ifdef CONFIG_NO_HZ_COMMON
191 * Check if the do_timer duty was dropped. We don't care about
192 * concurrency: This happens only when the CPU in charge went
193 * into a long sleep. If two CPUs happen to assign themselves to
194 * this duty, then the jiffies update is still serialized by
197 * If nohz_full is enabled, this should not happen because the
198 * tick_do_timer_cpu never relinquishes.
200 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
201 #ifdef CONFIG_NO_HZ_FULL
202 WARN_ON_ONCE(tick_nohz_full_running);
204 tick_do_timer_cpu = cpu;
208 /* Check, if the jiffies need an update */
209 if (tick_do_timer_cpu == cpu)
210 tick_do_update_jiffies64(now);
213 * If jiffies update stalled for too long (timekeeper in stop_machine()
214 * or VMEXIT'ed for several msecs), force an update.
216 if (ts->last_tick_jiffies != jiffies) {
217 ts->stalled_jiffies = 0;
218 ts->last_tick_jiffies = READ_ONCE(jiffies);
220 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
221 tick_do_update_jiffies64(now);
222 ts->stalled_jiffies = 0;
223 ts->last_tick_jiffies = READ_ONCE(jiffies);
228 ts->got_idle_tick = 1;
231 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
233 #ifdef CONFIG_NO_HZ_COMMON
235 * When we are idle and the tick is stopped, we have to touch
236 * the watchdog as we might not schedule for a really long
237 * time. This happens on complete idle SMP systems while
238 * waiting on the login prompt. We also increment the "start of
239 * idle" jiffy stamp so the idle accounting adjustment we do
240 * when we go busy again does not account too much ticks.
242 if (ts->tick_stopped) {
243 touch_softlockup_watchdog_sched();
244 if (is_idle_task(current))
247 * In case the current tick fired too early past its expected
248 * expiration, make sure we don't bypass the next clock reprogramming
249 * to the same deadline.
254 update_process_times(user_mode(regs));
255 profile_tick(CPU_PROFILING);
259 #ifdef CONFIG_NO_HZ_FULL
260 cpumask_var_t tick_nohz_full_mask;
261 EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
262 bool tick_nohz_full_running;
263 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
264 static atomic_t tick_dep_mask;
266 static bool check_tick_dependency(atomic_t *dep)
268 int val = atomic_read(dep);
270 if (val & TICK_DEP_MASK_POSIX_TIMER) {
271 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
275 if (val & TICK_DEP_MASK_PERF_EVENTS) {
276 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
280 if (val & TICK_DEP_MASK_SCHED) {
281 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
285 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
286 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
290 if (val & TICK_DEP_MASK_RCU) {
291 trace_tick_stop(0, TICK_DEP_MASK_RCU);
295 if (val & TICK_DEP_MASK_RCU_EXP) {
296 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
303 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
305 lockdep_assert_irqs_disabled();
307 if (unlikely(!cpu_online(cpu)))
310 if (check_tick_dependency(&tick_dep_mask))
313 if (check_tick_dependency(&ts->tick_dep_mask))
316 if (check_tick_dependency(¤t->tick_dep_mask))
319 if (check_tick_dependency(¤t->signal->tick_dep_mask))
325 static void nohz_full_kick_func(struct irq_work *work)
327 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
330 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
331 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
334 * Kick this CPU if it's full dynticks in order to force it to
335 * re-evaluate its dependency on the tick and restart it if necessary.
336 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
339 static void tick_nohz_full_kick(void)
341 if (!tick_nohz_full_cpu(smp_processor_id()))
344 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
348 * Kick the CPU if it's full dynticks in order to force it to
349 * re-evaluate its dependency on the tick and restart it if necessary.
351 void tick_nohz_full_kick_cpu(int cpu)
353 if (!tick_nohz_full_cpu(cpu))
356 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
359 static void tick_nohz_kick_task(struct task_struct *tsk)
364 * If the task is not running, run_posix_cpu_timers()
365 * has nothing to elapse, IPI can then be spared.
367 * activate_task() STORE p->tick_dep_mask
369 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
370 * LOCK rq->lock LOAD p->on_rq
371 * smp_mb__after_spin_lock()
372 * tick_nohz_task_switch()
373 * LOAD p->tick_dep_mask
375 if (!sched_task_on_rq(tsk))
379 * If the task concurrently migrates to another CPU,
380 * we guarantee it sees the new tick dependency upon
383 * set_task_cpu(p, cpu);
384 * STORE p->cpu = @cpu
385 * __schedule() (switch to task 'p')
387 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
388 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
389 * LOAD p->tick_dep_mask LOAD p->cpu
395 tick_nohz_full_kick_cpu(cpu);
400 * Kick all full dynticks CPUs in order to force these to re-evaluate
401 * their dependency on the tick and restart it if necessary.
403 static void tick_nohz_full_kick_all(void)
407 if (!tick_nohz_full_running)
411 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
412 tick_nohz_full_kick_cpu(cpu);
416 static void tick_nohz_dep_set_all(atomic_t *dep,
417 enum tick_dep_bits bit)
421 prev = atomic_fetch_or(BIT(bit), dep);
423 tick_nohz_full_kick_all();
427 * Set a global tick dependency. Used by perf events that rely on freq and
430 void tick_nohz_dep_set(enum tick_dep_bits bit)
432 tick_nohz_dep_set_all(&tick_dep_mask, bit);
435 void tick_nohz_dep_clear(enum tick_dep_bits bit)
437 atomic_andnot(BIT(bit), &tick_dep_mask);
441 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
442 * manage events throttling.
444 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
447 struct tick_sched *ts;
449 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
451 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
454 /* Perf needs local kick that is NMI safe */
455 if (cpu == smp_processor_id()) {
456 tick_nohz_full_kick();
458 /* Remote irq work not NMI-safe */
459 if (!WARN_ON_ONCE(in_nmi()))
460 tick_nohz_full_kick_cpu(cpu);
465 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
467 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
469 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
471 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
473 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
476 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
477 * in order to elapse per task timers.
479 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
481 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
482 tick_nohz_kick_task(tsk);
484 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
486 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
488 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
490 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
493 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
494 * per process timers.
496 void tick_nohz_dep_set_signal(struct task_struct *tsk,
497 enum tick_dep_bits bit)
500 struct signal_struct *sig = tsk->signal;
502 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
504 struct task_struct *t;
506 lockdep_assert_held(&tsk->sighand->siglock);
507 __for_each_thread(sig, t)
508 tick_nohz_kick_task(t);
512 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
514 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
518 * Re-evaluate the need for the tick as we switch the current task.
519 * It might need the tick due to per task/process properties:
520 * perf events, posix CPU timers, ...
522 void __tick_nohz_task_switch(void)
524 struct tick_sched *ts;
526 if (!tick_nohz_full_cpu(smp_processor_id()))
529 ts = this_cpu_ptr(&tick_cpu_sched);
531 if (ts->tick_stopped) {
532 if (atomic_read(¤t->tick_dep_mask) ||
533 atomic_read(¤t->signal->tick_dep_mask))
534 tick_nohz_full_kick();
538 /* Get the boot-time nohz CPU list from the kernel parameters. */
539 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
541 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
542 cpumask_copy(tick_nohz_full_mask, cpumask);
543 tick_nohz_full_running = true;
546 bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
549 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
550 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
551 * CPUs. It must remain online when nohz full is enabled.
553 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
558 static int tick_nohz_cpu_down(unsigned int cpu)
560 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
563 void __init tick_nohz_init(void)
567 if (!tick_nohz_full_running)
571 * Full dynticks uses irq work to drive the tick rescheduling on safe
572 * locking contexts. But then we need irq work to raise its own
573 * interrupts to avoid circular dependency on the tick
575 if (!arch_irq_work_has_interrupt()) {
576 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
577 cpumask_clear(tick_nohz_full_mask);
578 tick_nohz_full_running = false;
582 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
583 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
584 cpu = smp_processor_id();
586 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
587 pr_warn("NO_HZ: Clearing %d from nohz_full range "
588 "for timekeeping\n", cpu);
589 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
593 for_each_cpu(cpu, tick_nohz_full_mask)
594 ct_cpu_track_user(cpu);
596 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
597 "kernel/nohz:predown", NULL,
600 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
601 cpumask_pr_args(tick_nohz_full_mask));
606 * NOHZ - aka dynamic tick functionality
608 #ifdef CONFIG_NO_HZ_COMMON
612 bool tick_nohz_enabled __read_mostly = true;
613 unsigned long tick_nohz_active __read_mostly;
615 * Enable / Disable tickless mode
617 static int __init setup_tick_nohz(char *str)
619 return (kstrtobool(str, &tick_nohz_enabled) == 0);
622 __setup("nohz=", setup_tick_nohz);
624 bool tick_nohz_tick_stopped(void)
626 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
628 return ts->tick_stopped;
631 bool tick_nohz_tick_stopped_cpu(int cpu)
633 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
635 return ts->tick_stopped;
639 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
641 * Called from interrupt entry when the CPU was idle
643 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
644 * must be updated. Otherwise an interrupt handler could use a stale jiffy
645 * value. We do this unconditionally on any CPU, as we don't know whether the
646 * CPU, which has the update task assigned is in a long sleep.
648 static void tick_nohz_update_jiffies(ktime_t now)
652 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
654 local_irq_save(flags);
655 tick_do_update_jiffies64(now);
656 local_irq_restore(flags);
658 touch_softlockup_watchdog_sched();
662 * Updates the per-CPU time idle statistics counters
665 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
669 if (ts->idle_active) {
670 delta = ktime_sub(now, ts->idle_entrytime);
671 if (nr_iowait_cpu(cpu) > 0)
672 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
674 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
675 ts->idle_entrytime = now;
678 if (last_update_time)
679 *last_update_time = ktime_to_us(now);
683 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
685 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
688 sched_clock_idle_wakeup_event();
691 static void tick_nohz_start_idle(struct tick_sched *ts)
693 ts->idle_entrytime = ktime_get();
695 sched_clock_idle_sleep_event();
699 * get_cpu_idle_time_us - get the total idle time of a CPU
700 * @cpu: CPU number to query
701 * @last_update_time: variable to store update time in. Do not update
704 * Return the cumulative idle time (since boot) for a given
705 * CPU, in microseconds.
707 * This time is measured via accounting rather than sampling,
708 * and is as accurate as ktime_get() is.
710 * This function returns -1 if NOHZ is not enabled.
712 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
714 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
717 if (!tick_nohz_active)
721 if (last_update_time) {
722 update_ts_time_stats(cpu, ts, now, last_update_time);
723 idle = ts->idle_sleeptime;
725 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
726 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
728 idle = ktime_add(ts->idle_sleeptime, delta);
730 idle = ts->idle_sleeptime;
734 return ktime_to_us(idle);
737 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
740 * get_cpu_iowait_time_us - get the total iowait time of a CPU
741 * @cpu: CPU number to query
742 * @last_update_time: variable to store update time in. Do not update
745 * Return the cumulative iowait time (since boot) for a given
746 * CPU, in microseconds.
748 * This time is measured via accounting rather than sampling,
749 * and is as accurate as ktime_get() is.
751 * This function returns -1 if NOHZ is not enabled.
753 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
755 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
758 if (!tick_nohz_active)
762 if (last_update_time) {
763 update_ts_time_stats(cpu, ts, now, last_update_time);
764 iowait = ts->iowait_sleeptime;
766 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
767 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
769 iowait = ktime_add(ts->iowait_sleeptime, delta);
771 iowait = ts->iowait_sleeptime;
775 return ktime_to_us(iowait);
777 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
779 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
781 hrtimer_cancel(&ts->sched_timer);
782 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
784 /* Forward the time to expire in the future */
785 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
787 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
788 hrtimer_start_expires(&ts->sched_timer,
789 HRTIMER_MODE_ABS_PINNED_HARD);
791 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
795 * Reset to make sure next tick stop doesn't get fooled by past
796 * cached clock deadline.
801 static inline bool local_timer_softirq_pending(void)
803 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
806 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
808 u64 basemono, next_tick, delta, expires;
809 unsigned long basejiff;
812 /* Read jiffies and the time when jiffies were updated last */
814 seq = read_seqcount_begin(&jiffies_seq);
815 basemono = last_jiffies_update;
817 } while (read_seqcount_retry(&jiffies_seq, seq));
818 ts->last_jiffies = basejiff;
819 ts->timer_expires_base = basemono;
822 * Keep the periodic tick, when RCU, architecture or irq_work
824 * Aside of that check whether the local timer softirq is
825 * pending. If so its a bad idea to call get_next_timer_interrupt()
826 * because there is an already expired timer, so it will request
827 * immediate expiry, which rearms the hardware timer with a
828 * minimal delta which brings us back to this place
829 * immediately. Lather, rinse and repeat...
831 if (rcu_needs_cpu() || arch_needs_cpu() ||
832 irq_work_needs_cpu() || local_timer_softirq_pending()) {
833 next_tick = basemono + TICK_NSEC;
836 * Get the next pending timer. If high resolution
837 * timers are enabled this only takes the timer wheel
838 * timers into account. If high resolution timers are
839 * disabled this also looks at the next expiring
842 next_tick = get_next_timer_interrupt(basejiff, basemono);
843 ts->next_timer = next_tick;
847 * If the tick is due in the next period, keep it ticking or
848 * force prod the timer.
850 delta = next_tick - basemono;
851 if (delta <= (u64)TICK_NSEC) {
853 * Tell the timer code that the base is not idle, i.e. undo
854 * the effect of get_next_timer_interrupt():
858 * We've not stopped the tick yet, and there's a timer in the
859 * next period, so no point in stopping it either, bail.
861 if (!ts->tick_stopped) {
862 ts->timer_expires = 0;
868 * If this CPU is the one which had the do_timer() duty last, we limit
869 * the sleep time to the timekeeping max_deferment value.
870 * Otherwise we can sleep as long as we want.
872 delta = timekeeping_max_deferment();
873 if (cpu != tick_do_timer_cpu &&
874 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
877 /* Calculate the next expiry time */
878 if (delta < (KTIME_MAX - basemono))
879 expires = basemono + delta;
883 ts->timer_expires = min_t(u64, expires, next_tick);
886 return ts->timer_expires;
889 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
891 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
892 u64 basemono = ts->timer_expires_base;
893 u64 expires = ts->timer_expires;
894 ktime_t tick = expires;
896 /* Make sure we won't be trying to stop it twice in a row. */
897 ts->timer_expires_base = 0;
900 * If this CPU is the one which updates jiffies, then give up
901 * the assignment and let it be taken by the CPU which runs
902 * the tick timer next, which might be this CPU as well. If we
903 * don't drop this here the jiffies might be stale and
904 * do_timer() never invoked. Keep track of the fact that it
905 * was the one which had the do_timer() duty last.
907 if (cpu == tick_do_timer_cpu) {
908 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
909 ts->do_timer_last = 1;
910 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
911 ts->do_timer_last = 0;
914 /* Skip reprogram of event if its not changed */
915 if (ts->tick_stopped && (expires == ts->next_tick)) {
916 /* Sanity check: make sure clockevent is actually programmed */
917 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
921 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
922 basemono, ts->next_tick, dev->next_event,
923 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
927 * nohz_stop_sched_tick can be called several times before
928 * the nohz_restart_sched_tick is called. This happens when
929 * interrupts arrive which do not cause a reschedule. In the
930 * first call we save the current tick time, so we can restart
931 * the scheduler tick in nohz_restart_sched_tick.
933 if (!ts->tick_stopped) {
934 calc_load_nohz_start();
937 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
938 ts->tick_stopped = 1;
939 trace_tick_stop(1, TICK_DEP_MASK_NONE);
942 ts->next_tick = tick;
945 * If the expiration time == KTIME_MAX, then we simply stop
948 if (unlikely(expires == KTIME_MAX)) {
949 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
950 hrtimer_cancel(&ts->sched_timer);
952 tick_program_event(KTIME_MAX, 1);
956 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
957 hrtimer_start(&ts->sched_timer, tick,
958 HRTIMER_MODE_ABS_PINNED_HARD);
960 hrtimer_set_expires(&ts->sched_timer, tick);
961 tick_program_event(tick, 1);
965 static void tick_nohz_retain_tick(struct tick_sched *ts)
967 ts->timer_expires_base = 0;
970 #ifdef CONFIG_NO_HZ_FULL
971 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
973 if (tick_nohz_next_event(ts, cpu))
974 tick_nohz_stop_tick(ts, cpu);
976 tick_nohz_retain_tick(ts);
978 #endif /* CONFIG_NO_HZ_FULL */
980 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
982 /* Update jiffies first */
983 tick_do_update_jiffies64(now);
986 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
987 * the clock forward checks in the enqueue path:
991 calc_load_nohz_stop();
992 touch_softlockup_watchdog_sched();
994 * Cancel the scheduled timer and restore the tick
996 ts->tick_stopped = 0;
997 tick_nohz_restart(ts, now);
1000 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
1003 #ifdef CONFIG_NO_HZ_FULL
1004 int cpu = smp_processor_id();
1006 if (can_stop_full_tick(cpu, ts))
1007 tick_nohz_stop_sched_tick(ts, cpu);
1008 else if (ts->tick_stopped)
1009 tick_nohz_restart_sched_tick(ts, now);
1013 static void tick_nohz_full_update_tick(struct tick_sched *ts)
1015 if (!tick_nohz_full_cpu(smp_processor_id()))
1018 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1021 __tick_nohz_full_update_tick(ts, ktime_get());
1025 * A pending softirq outside an IRQ (or softirq disabled section) context
1026 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1027 * reach here due to the need_resched() early check in can_stop_idle_tick().
1029 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1030 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1031 * triggering the below since wakep_softirqd() is ignored.
1034 static bool report_idle_softirq(void)
1036 static int ratelimit;
1037 unsigned int pending = local_softirq_pending();
1039 if (likely(!pending))
1042 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1043 if (!cpu_active(smp_processor_id())) {
1044 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1049 if (ratelimit >= 10)
1052 /* On RT, softirqs handling may be waiting on some lock */
1053 if (!local_bh_blocked())
1056 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1063 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1066 * If this CPU is offline and it is the one which updates
1067 * jiffies, then give up the assignment and let it be taken by
1068 * the CPU which runs the tick timer next. If we don't drop
1069 * this here the jiffies might be stale and do_timer() never
1072 if (unlikely(!cpu_online(cpu))) {
1073 if (cpu == tick_do_timer_cpu)
1074 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1076 * Make sure the CPU doesn't get fooled by obsolete tick
1077 * deadline if it comes back online later.
1083 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1089 if (unlikely(report_idle_softirq()))
1092 if (tick_nohz_full_enabled()) {
1094 * Keep the tick alive to guarantee timekeeping progression
1095 * if there are full dynticks CPUs around
1097 if (tick_do_timer_cpu == cpu)
1100 /* Should not happen for nohz-full */
1101 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1108 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1111 int cpu = smp_processor_id();
1114 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1115 * tick timer expiration time is known already.
1117 if (ts->timer_expires_base)
1118 expires = ts->timer_expires;
1119 else if (can_stop_idle_tick(cpu, ts))
1120 expires = tick_nohz_next_event(ts, cpu);
1126 if (expires > 0LL) {
1127 int was_stopped = ts->tick_stopped;
1129 tick_nohz_stop_tick(ts, cpu);
1132 ts->idle_expires = expires;
1134 if (!was_stopped && ts->tick_stopped) {
1135 ts->idle_jiffies = ts->last_jiffies;
1136 nohz_balance_enter_idle(cpu);
1139 tick_nohz_retain_tick(ts);
1144 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1146 * When the next event is more than a tick into the future, stop the idle tick
1148 void tick_nohz_idle_stop_tick(void)
1150 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1153 void tick_nohz_idle_retain_tick(void)
1155 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1157 * Undo the effect of get_next_timer_interrupt() called from
1158 * tick_nohz_next_event().
1164 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1166 * Called when we start the idle loop.
1168 void tick_nohz_idle_enter(void)
1170 struct tick_sched *ts;
1172 lockdep_assert_irqs_enabled();
1174 local_irq_disable();
1176 ts = this_cpu_ptr(&tick_cpu_sched);
1178 WARN_ON_ONCE(ts->timer_expires_base);
1181 tick_nohz_start_idle(ts);
1187 * tick_nohz_irq_exit - update next tick event from interrupt exit
1189 * When an interrupt fires while we are idle and it doesn't cause
1190 * a reschedule, it may still add, modify or delete a timer, enqueue
1191 * an RCU callback, etc...
1192 * So we need to re-calculate and reprogram the next tick event.
1194 void tick_nohz_irq_exit(void)
1196 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1199 tick_nohz_start_idle(ts);
1201 tick_nohz_full_update_tick(ts);
1205 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1207 bool tick_nohz_idle_got_tick(void)
1209 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1211 if (ts->got_idle_tick) {
1212 ts->got_idle_tick = 0;
1219 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1220 * or the tick, whatever that expires first. Note that, if the tick has been
1221 * stopped, it returns the next hrtimer.
1223 * Called from power state control code with interrupts disabled
1225 ktime_t tick_nohz_get_next_hrtimer(void)
1227 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1231 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1232 * @delta_next: duration until the next event if the tick cannot be stopped
1234 * Called from power state control code with interrupts disabled.
1236 * The return value of this function and/or the value returned by it through the
1237 * @delta_next pointer can be negative which must be taken into account by its
1240 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1242 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1243 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1244 int cpu = smp_processor_id();
1246 * The idle entry time is expected to be a sufficient approximation of
1247 * the current time at this point.
1249 ktime_t now = ts->idle_entrytime;
1252 WARN_ON_ONCE(!ts->inidle);
1254 *delta_next = ktime_sub(dev->next_event, now);
1256 if (!can_stop_idle_tick(cpu, ts))
1259 next_event = tick_nohz_next_event(ts, cpu);
1264 * If the next highres timer to expire is earlier than next_event, the
1265 * idle governor needs to know that.
1267 next_event = min_t(u64, next_event,
1268 hrtimer_next_event_without(&ts->sched_timer));
1270 return ktime_sub(next_event, now);
1274 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1275 * for a particular CPU.
1277 * Called from the schedutil frequency scaling governor in scheduler context.
1279 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1281 struct tick_sched *ts = tick_get_tick_sched(cpu);
1283 return ts->idle_calls;
1287 * tick_nohz_get_idle_calls - return the current idle calls counter value
1289 * Called from the schedutil frequency scaling governor in scheduler context.
1291 unsigned long tick_nohz_get_idle_calls(void)
1293 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1295 return ts->idle_calls;
1298 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1301 unsigned long ticks;
1303 ts->idle_exittime = now;
1305 if (vtime_accounting_enabled_this_cpu())
1308 * We stopped the tick in idle. Update process times would miss the
1309 * time we slept as update_process_times does only a 1 tick
1310 * accounting. Enforce that this is accounted to idle !
1312 ticks = jiffies - ts->idle_jiffies;
1314 * We might be one off. Do not randomly account a huge number of ticks!
1316 if (ticks && ticks < LONG_MAX)
1317 account_idle_ticks(ticks);
1320 void tick_nohz_idle_restart_tick(void)
1322 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1324 if (ts->tick_stopped) {
1325 ktime_t now = ktime_get();
1326 tick_nohz_restart_sched_tick(ts, now);
1327 tick_nohz_account_idle_time(ts, now);
1331 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1333 if (tick_nohz_full_cpu(smp_processor_id()))
1334 __tick_nohz_full_update_tick(ts, now);
1336 tick_nohz_restart_sched_tick(ts, now);
1338 tick_nohz_account_idle_time(ts, now);
1342 * tick_nohz_idle_exit - restart the idle tick from the idle task
1344 * Restart the idle tick when the CPU is woken up from idle
1345 * This also exit the RCU extended quiescent state. The CPU
1346 * can use RCU again after this function is called.
1348 void tick_nohz_idle_exit(void)
1350 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1351 bool idle_active, tick_stopped;
1354 local_irq_disable();
1356 WARN_ON_ONCE(!ts->inidle);
1357 WARN_ON_ONCE(ts->timer_expires_base);
1360 idle_active = ts->idle_active;
1361 tick_stopped = ts->tick_stopped;
1363 if (idle_active || tick_stopped)
1367 tick_nohz_stop_idle(ts, now);
1370 tick_nohz_idle_update_tick(ts, now);
1376 * The nohz low res interrupt handler
1378 static void tick_nohz_handler(struct clock_event_device *dev)
1380 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1381 struct pt_regs *regs = get_irq_regs();
1382 ktime_t now = ktime_get();
1384 dev->next_event = KTIME_MAX;
1386 tick_sched_do_timer(ts, now);
1387 tick_sched_handle(ts, regs);
1389 if (unlikely(ts->tick_stopped)) {
1391 * The clockevent device is not reprogrammed, so change the
1392 * clock event device to ONESHOT_STOPPED to avoid spurious
1393 * interrupts on devices which might not be truly one shot.
1395 tick_program_event(KTIME_MAX, 1);
1399 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1400 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1403 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1405 if (!tick_nohz_enabled)
1407 ts->nohz_mode = mode;
1408 /* One update is enough */
1409 if (!test_and_set_bit(0, &tick_nohz_active))
1410 timers_update_nohz();
1414 * tick_nohz_switch_to_nohz - switch to nohz mode
1416 static void tick_nohz_switch_to_nohz(void)
1418 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1421 if (!tick_nohz_enabled)
1424 if (tick_switch_to_oneshot(tick_nohz_handler))
1428 * Recycle the hrtimer in ts, so we can share the
1429 * hrtimer_forward with the highres code.
1431 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1432 /* Get the next period */
1433 next = tick_init_jiffy_update();
1435 hrtimer_set_expires(&ts->sched_timer, next);
1436 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1437 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1438 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1441 static inline void tick_nohz_irq_enter(void)
1443 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1446 if (!ts->idle_active && !ts->tick_stopped)
1449 if (ts->idle_active)
1450 tick_nohz_stop_idle(ts, now);
1452 * If all CPUs are idle. We may need to update a stale jiffies value.
1453 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1454 * alive but it might be busy looping with interrupts disabled in some
1455 * rare case (typically stop machine). So we must make sure we have a
1458 if (ts->tick_stopped)
1459 tick_nohz_update_jiffies(now);
1464 static inline void tick_nohz_switch_to_nohz(void) { }
1465 static inline void tick_nohz_irq_enter(void) { }
1466 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1468 #endif /* CONFIG_NO_HZ_COMMON */
1471 * Called from irq_enter to notify about the possible interruption of idle()
1473 void tick_irq_enter(void)
1475 tick_check_oneshot_broadcast_this_cpu();
1476 tick_nohz_irq_enter();
1480 * High resolution timer specific code
1482 #ifdef CONFIG_HIGH_RES_TIMERS
1484 * We rearm the timer until we get disabled by the idle code.
1485 * Called with interrupts disabled.
1487 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1489 struct tick_sched *ts =
1490 container_of(timer, struct tick_sched, sched_timer);
1491 struct pt_regs *regs = get_irq_regs();
1492 ktime_t now = ktime_get();
1494 tick_sched_do_timer(ts, now);
1497 * Do not call, when we are not in irq context and have
1498 * no valid regs pointer
1501 tick_sched_handle(ts, regs);
1505 /* No need to reprogram if we are in idle or full dynticks mode */
1506 if (unlikely(ts->tick_stopped))
1507 return HRTIMER_NORESTART;
1509 hrtimer_forward(timer, now, TICK_NSEC);
1511 return HRTIMER_RESTART;
1514 static int sched_skew_tick;
1516 static int __init skew_tick(char *str)
1518 get_option(&str, &sched_skew_tick);
1522 early_param("skew_tick", skew_tick);
1525 * tick_setup_sched_timer - setup the tick emulation timer
1527 void tick_setup_sched_timer(void)
1529 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1530 ktime_t now = ktime_get();
1533 * Emulate tick processing via per-CPU hrtimers:
1535 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1536 ts->sched_timer.function = tick_sched_timer;
1538 /* Get the next period (per-CPU) */
1539 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1541 /* Offset the tick to avert jiffies_lock contention. */
1542 if (sched_skew_tick) {
1543 u64 offset = TICK_NSEC >> 1;
1544 do_div(offset, num_possible_cpus());
1545 offset *= smp_processor_id();
1546 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1549 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1550 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1551 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1553 #endif /* HIGH_RES_TIMERS */
1555 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1556 void tick_cancel_sched_timer(int cpu)
1558 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1560 # ifdef CONFIG_HIGH_RES_TIMERS
1561 if (ts->sched_timer.base)
1562 hrtimer_cancel(&ts->sched_timer);
1565 memset(ts, 0, sizeof(*ts));
1570 * Async notification about clocksource changes
1572 void tick_clock_notify(void)
1576 for_each_possible_cpu(cpu)
1577 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1581 * Async notification about clock event changes
1583 void tick_oneshot_notify(void)
1585 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1587 set_bit(0, &ts->check_clocks);
1591 * Check, if a change happened, which makes oneshot possible.
1593 * Called cyclic from the hrtimer softirq (driven by the timer
1594 * softirq) allow_nohz signals, that we can switch into low-res nohz
1595 * mode, because high resolution timers are disabled (either compile
1596 * or runtime). Called with interrupts disabled.
1598 int tick_check_oneshot_change(int allow_nohz)
1600 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1602 if (!test_and_clear_bit(0, &ts->check_clocks))
1605 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1608 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1614 tick_nohz_switch_to_nohz();