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 #define MAX_STALLED_JIFFIES 5
174 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
176 int cpu = smp_processor_id();
178 #ifdef CONFIG_NO_HZ_COMMON
180 * Check if the do_timer duty was dropped. We don't care about
181 * concurrency: This happens only when the CPU in charge went
182 * into a long sleep. If two CPUs happen to assign themselves to
183 * this duty, then the jiffies update is still serialized by
186 * If nohz_full is enabled, this should not happen because the
187 * tick_do_timer_cpu never relinquishes.
189 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
190 #ifdef CONFIG_NO_HZ_FULL
191 WARN_ON_ONCE(tick_nohz_full_running);
193 tick_do_timer_cpu = cpu;
197 /* Check, if the jiffies need an update */
198 if (tick_do_timer_cpu == cpu)
199 tick_do_update_jiffies64(now);
202 * If jiffies update stalled for too long (timekeeper in stop_machine()
203 * or VMEXIT'ed for several msecs), force an update.
205 if (ts->last_tick_jiffies != jiffies) {
206 ts->stalled_jiffies = 0;
207 ts->last_tick_jiffies = READ_ONCE(jiffies);
209 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
210 tick_do_update_jiffies64(now);
211 ts->stalled_jiffies = 0;
212 ts->last_tick_jiffies = READ_ONCE(jiffies);
217 ts->got_idle_tick = 1;
220 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
222 #ifdef CONFIG_NO_HZ_COMMON
224 * When we are idle and the tick is stopped, we have to touch
225 * the watchdog as we might not schedule for a really long
226 * time. This happens on complete idle SMP systems while
227 * waiting on the login prompt. We also increment the "start of
228 * idle" jiffy stamp so the idle accounting adjustment we do
229 * when we go busy again does not account too much ticks.
231 if (ts->tick_stopped) {
232 touch_softlockup_watchdog_sched();
233 if (is_idle_task(current))
236 * In case the current tick fired too early past its expected
237 * expiration, make sure we don't bypass the next clock reprogramming
238 * to the same deadline.
243 update_process_times(user_mode(regs));
244 profile_tick(CPU_PROFILING);
248 #ifdef CONFIG_NO_HZ_FULL
249 cpumask_var_t tick_nohz_full_mask;
250 EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
251 bool tick_nohz_full_running;
252 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
253 static atomic_t tick_dep_mask;
255 static bool check_tick_dependency(atomic_t *dep)
257 int val = atomic_read(dep);
259 if (val & TICK_DEP_MASK_POSIX_TIMER) {
260 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
264 if (val & TICK_DEP_MASK_PERF_EVENTS) {
265 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
269 if (val & TICK_DEP_MASK_SCHED) {
270 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
274 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
275 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
279 if (val & TICK_DEP_MASK_RCU) {
280 trace_tick_stop(0, TICK_DEP_MASK_RCU);
284 if (val & TICK_DEP_MASK_RCU_EXP) {
285 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
292 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
294 lockdep_assert_irqs_disabled();
296 if (unlikely(!cpu_online(cpu)))
299 if (check_tick_dependency(&tick_dep_mask))
302 if (check_tick_dependency(&ts->tick_dep_mask))
305 if (check_tick_dependency(¤t->tick_dep_mask))
308 if (check_tick_dependency(¤t->signal->tick_dep_mask))
314 static void nohz_full_kick_func(struct irq_work *work)
316 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
319 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
320 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
323 * Kick this CPU if it's full dynticks in order to force it to
324 * re-evaluate its dependency on the tick and restart it if necessary.
325 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
328 static void tick_nohz_full_kick(void)
330 if (!tick_nohz_full_cpu(smp_processor_id()))
333 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
337 * Kick the CPU if it's full dynticks in order to force it to
338 * re-evaluate its dependency on the tick and restart it if necessary.
340 void tick_nohz_full_kick_cpu(int cpu)
342 if (!tick_nohz_full_cpu(cpu))
345 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
348 static void tick_nohz_kick_task(struct task_struct *tsk)
353 * If the task is not running, run_posix_cpu_timers()
354 * has nothing to elapse, IPI can then be spared.
356 * activate_task() STORE p->tick_dep_mask
358 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
359 * LOCK rq->lock LOAD p->on_rq
360 * smp_mb__after_spin_lock()
361 * tick_nohz_task_switch()
362 * LOAD p->tick_dep_mask
364 if (!sched_task_on_rq(tsk))
368 * If the task concurrently migrates to another CPU,
369 * we guarantee it sees the new tick dependency upon
372 * set_task_cpu(p, cpu);
373 * STORE p->cpu = @cpu
374 * __schedule() (switch to task 'p')
376 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
377 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
378 * LOAD p->tick_dep_mask LOAD p->cpu
384 tick_nohz_full_kick_cpu(cpu);
389 * Kick all full dynticks CPUs in order to force these to re-evaluate
390 * their dependency on the tick and restart it if necessary.
392 static void tick_nohz_full_kick_all(void)
396 if (!tick_nohz_full_running)
400 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
401 tick_nohz_full_kick_cpu(cpu);
405 static void tick_nohz_dep_set_all(atomic_t *dep,
406 enum tick_dep_bits bit)
410 prev = atomic_fetch_or(BIT(bit), dep);
412 tick_nohz_full_kick_all();
416 * Set a global tick dependency. Used by perf events that rely on freq and
419 void tick_nohz_dep_set(enum tick_dep_bits bit)
421 tick_nohz_dep_set_all(&tick_dep_mask, bit);
424 void tick_nohz_dep_clear(enum tick_dep_bits bit)
426 atomic_andnot(BIT(bit), &tick_dep_mask);
430 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
431 * manage events throttling.
433 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
436 struct tick_sched *ts;
438 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
440 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
443 /* Perf needs local kick that is NMI safe */
444 if (cpu == smp_processor_id()) {
445 tick_nohz_full_kick();
447 /* Remote irq work not NMI-safe */
448 if (!WARN_ON_ONCE(in_nmi()))
449 tick_nohz_full_kick_cpu(cpu);
454 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
456 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
458 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
460 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
462 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
465 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
466 * in order to elapse per task timers.
468 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
470 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
471 tick_nohz_kick_task(tsk);
473 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
475 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
477 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
479 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
482 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
483 * per process timers.
485 void tick_nohz_dep_set_signal(struct task_struct *tsk,
486 enum tick_dep_bits bit)
489 struct signal_struct *sig = tsk->signal;
491 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
493 struct task_struct *t;
495 lockdep_assert_held(&tsk->sighand->siglock);
496 __for_each_thread(sig, t)
497 tick_nohz_kick_task(t);
501 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
503 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
507 * Re-evaluate the need for the tick as we switch the current task.
508 * It might need the tick due to per task/process properties:
509 * perf events, posix CPU timers, ...
511 void __tick_nohz_task_switch(void)
513 struct tick_sched *ts;
515 if (!tick_nohz_full_cpu(smp_processor_id()))
518 ts = this_cpu_ptr(&tick_cpu_sched);
520 if (ts->tick_stopped) {
521 if (atomic_read(¤t->tick_dep_mask) ||
522 atomic_read(¤t->signal->tick_dep_mask))
523 tick_nohz_full_kick();
527 /* Get the boot-time nohz CPU list from the kernel parameters. */
528 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
530 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
531 cpumask_copy(tick_nohz_full_mask, cpumask);
532 tick_nohz_full_running = true;
535 bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
538 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
539 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
540 * CPUs. It must remain online when nohz full is enabled.
542 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
547 static int tick_nohz_cpu_down(unsigned int cpu)
549 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
552 void __init tick_nohz_init(void)
556 if (!tick_nohz_full_running)
560 * Full dynticks uses irq work to drive the tick rescheduling on safe
561 * locking contexts. But then we need irq work to raise its own
562 * interrupts to avoid circular dependency on the tick
564 if (!arch_irq_work_has_interrupt()) {
565 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
566 cpumask_clear(tick_nohz_full_mask);
567 tick_nohz_full_running = false;
571 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
572 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
573 cpu = smp_processor_id();
575 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
576 pr_warn("NO_HZ: Clearing %d from nohz_full range "
577 "for timekeeping\n", cpu);
578 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
582 for_each_cpu(cpu, tick_nohz_full_mask)
583 ct_cpu_track_user(cpu);
585 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
586 "kernel/nohz:predown", NULL,
589 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
590 cpumask_pr_args(tick_nohz_full_mask));
595 * NOHZ - aka dynamic tick functionality
597 #ifdef CONFIG_NO_HZ_COMMON
601 bool tick_nohz_enabled __read_mostly = true;
602 unsigned long tick_nohz_active __read_mostly;
604 * Enable / Disable tickless mode
606 static int __init setup_tick_nohz(char *str)
608 return (kstrtobool(str, &tick_nohz_enabled) == 0);
611 __setup("nohz=", setup_tick_nohz);
613 bool tick_nohz_tick_stopped(void)
615 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
617 return ts->tick_stopped;
620 bool tick_nohz_tick_stopped_cpu(int cpu)
622 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
624 return ts->tick_stopped;
628 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
630 * Called from interrupt entry when the CPU was idle
632 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
633 * must be updated. Otherwise an interrupt handler could use a stale jiffy
634 * value. We do this unconditionally on any CPU, as we don't know whether the
635 * CPU, which has the update task assigned is in a long sleep.
637 static void tick_nohz_update_jiffies(ktime_t now)
641 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
643 local_irq_save(flags);
644 tick_do_update_jiffies64(now);
645 local_irq_restore(flags);
647 touch_softlockup_watchdog_sched();
651 * Updates the per-CPU time idle statistics counters
654 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
658 if (ts->idle_active) {
659 delta = ktime_sub(now, ts->idle_entrytime);
660 if (nr_iowait_cpu(cpu) > 0)
661 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
663 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
664 ts->idle_entrytime = now;
667 if (last_update_time)
668 *last_update_time = ktime_to_us(now);
672 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
674 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
677 sched_clock_idle_wakeup_event();
680 static void tick_nohz_start_idle(struct tick_sched *ts)
682 ts->idle_entrytime = ktime_get();
684 sched_clock_idle_sleep_event();
688 * get_cpu_idle_time_us - get the total idle time of a CPU
689 * @cpu: CPU number to query
690 * @last_update_time: variable to store update time in. Do not update
693 * Return the cumulative idle time (since boot) for a given
694 * CPU, in microseconds.
696 * This time is measured via accounting rather than sampling,
697 * and is as accurate as ktime_get() is.
699 * This function returns -1 if NOHZ is not enabled.
701 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
703 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
706 if (!tick_nohz_active)
710 if (last_update_time) {
711 update_ts_time_stats(cpu, ts, now, last_update_time);
712 idle = ts->idle_sleeptime;
714 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
715 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
717 idle = ktime_add(ts->idle_sleeptime, delta);
719 idle = ts->idle_sleeptime;
723 return ktime_to_us(idle);
726 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
729 * get_cpu_iowait_time_us - get the total iowait time of a CPU
730 * @cpu: CPU number to query
731 * @last_update_time: variable to store update time in. Do not update
734 * Return the cumulative iowait time (since boot) for a given
735 * CPU, in microseconds.
737 * This time is measured via accounting rather than sampling,
738 * and is as accurate as ktime_get() is.
740 * This function returns -1 if NOHZ is not enabled.
742 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
744 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
747 if (!tick_nohz_active)
751 if (last_update_time) {
752 update_ts_time_stats(cpu, ts, now, last_update_time);
753 iowait = ts->iowait_sleeptime;
755 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
756 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
758 iowait = ktime_add(ts->iowait_sleeptime, delta);
760 iowait = ts->iowait_sleeptime;
764 return ktime_to_us(iowait);
766 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
768 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
770 hrtimer_cancel(&ts->sched_timer);
771 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
773 /* Forward the time to expire in the future */
774 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
776 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
777 hrtimer_start_expires(&ts->sched_timer,
778 HRTIMER_MODE_ABS_PINNED_HARD);
780 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
784 * Reset to make sure next tick stop doesn't get fooled by past
785 * cached clock deadline.
790 static inline bool local_timer_softirq_pending(void)
792 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
795 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
797 u64 basemono, next_tick, delta, expires;
798 unsigned long basejiff;
801 /* Read jiffies and the time when jiffies were updated last */
803 seq = read_seqcount_begin(&jiffies_seq);
804 basemono = last_jiffies_update;
806 } while (read_seqcount_retry(&jiffies_seq, seq));
807 ts->last_jiffies = basejiff;
808 ts->timer_expires_base = basemono;
811 * Keep the periodic tick, when RCU, architecture or irq_work
813 * Aside of that check whether the local timer softirq is
814 * pending. If so its a bad idea to call get_next_timer_interrupt()
815 * because there is an already expired timer, so it will request
816 * immediate expiry, which rearms the hardware timer with a
817 * minimal delta which brings us back to this place
818 * immediately. Lather, rinse and repeat...
820 if (rcu_needs_cpu() || arch_needs_cpu() ||
821 irq_work_needs_cpu() || local_timer_softirq_pending()) {
822 next_tick = basemono + TICK_NSEC;
825 * Get the next pending timer. If high resolution
826 * timers are enabled this only takes the timer wheel
827 * timers into account. If high resolution timers are
828 * disabled this also looks at the next expiring
831 next_tick = get_next_timer_interrupt(basejiff, basemono);
832 ts->next_timer = next_tick;
836 * If the tick is due in the next period, keep it ticking or
837 * force prod the timer.
839 delta = next_tick - basemono;
840 if (delta <= (u64)TICK_NSEC) {
842 * Tell the timer code that the base is not idle, i.e. undo
843 * the effect of get_next_timer_interrupt():
847 * We've not stopped the tick yet, and there's a timer in the
848 * next period, so no point in stopping it either, bail.
850 if (!ts->tick_stopped) {
851 ts->timer_expires = 0;
857 * If this CPU is the one which had the do_timer() duty last, we limit
858 * the sleep time to the timekeeping max_deferment value.
859 * Otherwise we can sleep as long as we want.
861 delta = timekeeping_max_deferment();
862 if (cpu != tick_do_timer_cpu &&
863 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
866 /* Calculate the next expiry time */
867 if (delta < (KTIME_MAX - basemono))
868 expires = basemono + delta;
872 ts->timer_expires = min_t(u64, expires, next_tick);
875 return ts->timer_expires;
878 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
880 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
881 u64 basemono = ts->timer_expires_base;
882 u64 expires = ts->timer_expires;
883 ktime_t tick = expires;
885 /* Make sure we won't be trying to stop it twice in a row. */
886 ts->timer_expires_base = 0;
889 * If this CPU is the one which updates jiffies, then give up
890 * the assignment and let it be taken by the CPU which runs
891 * the tick timer next, which might be this CPU as well. If we
892 * don't drop this here the jiffies might be stale and
893 * do_timer() never invoked. Keep track of the fact that it
894 * was the one which had the do_timer() duty last.
896 if (cpu == tick_do_timer_cpu) {
897 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
898 ts->do_timer_last = 1;
899 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
900 ts->do_timer_last = 0;
903 /* Skip reprogram of event if its not changed */
904 if (ts->tick_stopped && (expires == ts->next_tick)) {
905 /* Sanity check: make sure clockevent is actually programmed */
906 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
910 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
911 basemono, ts->next_tick, dev->next_event,
912 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
916 * nohz_stop_sched_tick can be called several times before
917 * the nohz_restart_sched_tick is called. This happens when
918 * interrupts arrive which do not cause a reschedule. In the
919 * first call we save the current tick time, so we can restart
920 * the scheduler tick in nohz_restart_sched_tick.
922 if (!ts->tick_stopped) {
923 calc_load_nohz_start();
926 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
927 ts->tick_stopped = 1;
928 trace_tick_stop(1, TICK_DEP_MASK_NONE);
931 ts->next_tick = tick;
934 * If the expiration time == KTIME_MAX, then we simply stop
937 if (unlikely(expires == KTIME_MAX)) {
938 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
939 hrtimer_cancel(&ts->sched_timer);
941 tick_program_event(KTIME_MAX, 1);
945 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
946 hrtimer_start(&ts->sched_timer, tick,
947 HRTIMER_MODE_ABS_PINNED_HARD);
949 hrtimer_set_expires(&ts->sched_timer, tick);
950 tick_program_event(tick, 1);
954 static void tick_nohz_retain_tick(struct tick_sched *ts)
956 ts->timer_expires_base = 0;
959 #ifdef CONFIG_NO_HZ_FULL
960 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
962 if (tick_nohz_next_event(ts, cpu))
963 tick_nohz_stop_tick(ts, cpu);
965 tick_nohz_retain_tick(ts);
967 #endif /* CONFIG_NO_HZ_FULL */
969 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
971 /* Update jiffies first */
972 tick_do_update_jiffies64(now);
975 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
976 * the clock forward checks in the enqueue path:
980 calc_load_nohz_stop();
981 touch_softlockup_watchdog_sched();
983 * Cancel the scheduled timer and restore the tick
985 ts->tick_stopped = 0;
986 tick_nohz_restart(ts, now);
989 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
992 #ifdef CONFIG_NO_HZ_FULL
993 int cpu = smp_processor_id();
995 if (can_stop_full_tick(cpu, ts))
996 tick_nohz_stop_sched_tick(ts, cpu);
997 else if (ts->tick_stopped)
998 tick_nohz_restart_sched_tick(ts, now);
1002 static void tick_nohz_full_update_tick(struct tick_sched *ts)
1004 if (!tick_nohz_full_cpu(smp_processor_id()))
1007 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1010 __tick_nohz_full_update_tick(ts, ktime_get());
1014 * A pending softirq outside an IRQ (or softirq disabled section) context
1015 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1016 * reach here due to the need_resched() early check in can_stop_idle_tick().
1018 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1019 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1020 * triggering the below since wakep_softirqd() is ignored.
1023 static bool report_idle_softirq(void)
1025 static int ratelimit;
1026 unsigned int pending = local_softirq_pending();
1028 if (likely(!pending))
1031 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1032 if (!cpu_active(smp_processor_id())) {
1033 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1041 /* On RT, softirqs handling may be waiting on some lock */
1042 if (!local_bh_blocked())
1045 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1052 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1055 * If this CPU is offline and it is the one which updates
1056 * jiffies, then give up the assignment and let it be taken by
1057 * the CPU which runs the tick timer next. If we don't drop
1058 * this here the jiffies might be stale and do_timer() never
1061 if (unlikely(!cpu_online(cpu))) {
1062 if (cpu == tick_do_timer_cpu)
1063 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1065 * Make sure the CPU doesn't get fooled by obsolete tick
1066 * deadline if it comes back online later.
1072 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1078 if (unlikely(report_idle_softirq()))
1081 if (tick_nohz_full_enabled()) {
1083 * Keep the tick alive to guarantee timekeeping progression
1084 * if there are full dynticks CPUs around
1086 if (tick_do_timer_cpu == cpu)
1089 /* Should not happen for nohz-full */
1090 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1097 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1100 int cpu = smp_processor_id();
1103 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1104 * tick timer expiration time is known already.
1106 if (ts->timer_expires_base)
1107 expires = ts->timer_expires;
1108 else if (can_stop_idle_tick(cpu, ts))
1109 expires = tick_nohz_next_event(ts, cpu);
1115 if (expires > 0LL) {
1116 int was_stopped = ts->tick_stopped;
1118 tick_nohz_stop_tick(ts, cpu);
1121 ts->idle_expires = expires;
1123 if (!was_stopped && ts->tick_stopped) {
1124 ts->idle_jiffies = ts->last_jiffies;
1125 nohz_balance_enter_idle(cpu);
1128 tick_nohz_retain_tick(ts);
1133 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1135 * When the next event is more than a tick into the future, stop the idle tick
1137 void tick_nohz_idle_stop_tick(void)
1139 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1142 void tick_nohz_idle_retain_tick(void)
1144 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1146 * Undo the effect of get_next_timer_interrupt() called from
1147 * tick_nohz_next_event().
1153 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1155 * Called when we start the idle loop.
1157 void tick_nohz_idle_enter(void)
1159 struct tick_sched *ts;
1161 lockdep_assert_irqs_enabled();
1163 local_irq_disable();
1165 ts = this_cpu_ptr(&tick_cpu_sched);
1167 WARN_ON_ONCE(ts->timer_expires_base);
1170 tick_nohz_start_idle(ts);
1176 * tick_nohz_irq_exit - update next tick event from interrupt exit
1178 * When an interrupt fires while we are idle and it doesn't cause
1179 * a reschedule, it may still add, modify or delete a timer, enqueue
1180 * an RCU callback, etc...
1181 * So we need to re-calculate and reprogram the next tick event.
1183 void tick_nohz_irq_exit(void)
1185 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1188 tick_nohz_start_idle(ts);
1190 tick_nohz_full_update_tick(ts);
1194 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1196 bool tick_nohz_idle_got_tick(void)
1198 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1200 if (ts->got_idle_tick) {
1201 ts->got_idle_tick = 0;
1208 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1209 * or the tick, whatever that expires first. Note that, if the tick has been
1210 * stopped, it returns the next hrtimer.
1212 * Called from power state control code with interrupts disabled
1214 ktime_t tick_nohz_get_next_hrtimer(void)
1216 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1220 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1221 * @delta_next: duration until the next event if the tick cannot be stopped
1223 * Called from power state control code with interrupts disabled.
1225 * The return value of this function and/or the value returned by it through the
1226 * @delta_next pointer can be negative which must be taken into account by its
1229 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1231 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1232 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1233 int cpu = smp_processor_id();
1235 * The idle entry time is expected to be a sufficient approximation of
1236 * the current time at this point.
1238 ktime_t now = ts->idle_entrytime;
1241 WARN_ON_ONCE(!ts->inidle);
1243 *delta_next = ktime_sub(dev->next_event, now);
1245 if (!can_stop_idle_tick(cpu, ts))
1248 next_event = tick_nohz_next_event(ts, cpu);
1253 * If the next highres timer to expire is earlier than next_event, the
1254 * idle governor needs to know that.
1256 next_event = min_t(u64, next_event,
1257 hrtimer_next_event_without(&ts->sched_timer));
1259 return ktime_sub(next_event, now);
1263 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1264 * for a particular CPU.
1266 * Called from the schedutil frequency scaling governor in scheduler context.
1268 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1270 struct tick_sched *ts = tick_get_tick_sched(cpu);
1272 return ts->idle_calls;
1276 * tick_nohz_get_idle_calls - return the current idle calls counter value
1278 * Called from the schedutil frequency scaling governor in scheduler context.
1280 unsigned long tick_nohz_get_idle_calls(void)
1282 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1284 return ts->idle_calls;
1287 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1290 unsigned long ticks;
1292 ts->idle_exittime = now;
1294 if (vtime_accounting_enabled_this_cpu())
1297 * We stopped the tick in idle. Update process times would miss the
1298 * time we slept as update_process_times does only a 1 tick
1299 * accounting. Enforce that this is accounted to idle !
1301 ticks = jiffies - ts->idle_jiffies;
1303 * We might be one off. Do not randomly account a huge number of ticks!
1305 if (ticks && ticks < LONG_MAX)
1306 account_idle_ticks(ticks);
1309 void tick_nohz_idle_restart_tick(void)
1311 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1313 if (ts->tick_stopped) {
1314 ktime_t now = ktime_get();
1315 tick_nohz_restart_sched_tick(ts, now);
1316 tick_nohz_account_idle_time(ts, now);
1320 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1322 if (tick_nohz_full_cpu(smp_processor_id()))
1323 __tick_nohz_full_update_tick(ts, now);
1325 tick_nohz_restart_sched_tick(ts, now);
1327 tick_nohz_account_idle_time(ts, now);
1331 * tick_nohz_idle_exit - restart the idle tick from the idle task
1333 * Restart the idle tick when the CPU is woken up from idle
1334 * This also exit the RCU extended quiescent state. The CPU
1335 * can use RCU again after this function is called.
1337 void tick_nohz_idle_exit(void)
1339 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1340 bool idle_active, tick_stopped;
1343 local_irq_disable();
1345 WARN_ON_ONCE(!ts->inidle);
1346 WARN_ON_ONCE(ts->timer_expires_base);
1349 idle_active = ts->idle_active;
1350 tick_stopped = ts->tick_stopped;
1352 if (idle_active || tick_stopped)
1356 tick_nohz_stop_idle(ts, now);
1359 tick_nohz_idle_update_tick(ts, now);
1365 * The nohz low res interrupt handler
1367 static void tick_nohz_handler(struct clock_event_device *dev)
1369 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1370 struct pt_regs *regs = get_irq_regs();
1371 ktime_t now = ktime_get();
1373 dev->next_event = KTIME_MAX;
1375 tick_sched_do_timer(ts, now);
1376 tick_sched_handle(ts, regs);
1378 if (unlikely(ts->tick_stopped)) {
1380 * The clockevent device is not reprogrammed, so change the
1381 * clock event device to ONESHOT_STOPPED to avoid spurious
1382 * interrupts on devices which might not be truly one shot.
1384 tick_program_event(KTIME_MAX, 1);
1388 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1389 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1392 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1394 if (!tick_nohz_enabled)
1396 ts->nohz_mode = mode;
1397 /* One update is enough */
1398 if (!test_and_set_bit(0, &tick_nohz_active))
1399 timers_update_nohz();
1403 * tick_nohz_switch_to_nohz - switch to nohz mode
1405 static void tick_nohz_switch_to_nohz(void)
1407 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1410 if (!tick_nohz_enabled)
1413 if (tick_switch_to_oneshot(tick_nohz_handler))
1417 * Recycle the hrtimer in ts, so we can share the
1418 * hrtimer_forward with the highres code.
1420 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1421 /* Get the next period */
1422 next = tick_init_jiffy_update();
1424 hrtimer_set_expires(&ts->sched_timer, next);
1425 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1426 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1427 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1430 static inline void tick_nohz_irq_enter(void)
1432 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1435 if (!ts->idle_active && !ts->tick_stopped)
1438 if (ts->idle_active)
1439 tick_nohz_stop_idle(ts, now);
1441 * If all CPUs are idle. We may need to update a stale jiffies value.
1442 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1443 * alive but it might be busy looping with interrupts disabled in some
1444 * rare case (typically stop machine). So we must make sure we have a
1447 if (ts->tick_stopped)
1448 tick_nohz_update_jiffies(now);
1453 static inline void tick_nohz_switch_to_nohz(void) { }
1454 static inline void tick_nohz_irq_enter(void) { }
1455 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1457 #endif /* CONFIG_NO_HZ_COMMON */
1460 * Called from irq_enter to notify about the possible interruption of idle()
1462 void tick_irq_enter(void)
1464 tick_check_oneshot_broadcast_this_cpu();
1465 tick_nohz_irq_enter();
1469 * High resolution timer specific code
1471 #ifdef CONFIG_HIGH_RES_TIMERS
1473 * We rearm the timer until we get disabled by the idle code.
1474 * Called with interrupts disabled.
1476 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1478 struct tick_sched *ts =
1479 container_of(timer, struct tick_sched, sched_timer);
1480 struct pt_regs *regs = get_irq_regs();
1481 ktime_t now = ktime_get();
1483 tick_sched_do_timer(ts, now);
1486 * Do not call, when we are not in irq context and have
1487 * no valid regs pointer
1490 tick_sched_handle(ts, regs);
1494 /* No need to reprogram if we are in idle or full dynticks mode */
1495 if (unlikely(ts->tick_stopped))
1496 return HRTIMER_NORESTART;
1498 hrtimer_forward(timer, now, TICK_NSEC);
1500 return HRTIMER_RESTART;
1503 static int sched_skew_tick;
1505 static int __init skew_tick(char *str)
1507 get_option(&str, &sched_skew_tick);
1511 early_param("skew_tick", skew_tick);
1514 * tick_setup_sched_timer - setup the tick emulation timer
1516 void tick_setup_sched_timer(void)
1518 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1519 ktime_t now = ktime_get();
1522 * Emulate tick processing via per-CPU hrtimers:
1524 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1525 ts->sched_timer.function = tick_sched_timer;
1527 /* Get the next period (per-CPU) */
1528 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1530 /* Offset the tick to avert jiffies_lock contention. */
1531 if (sched_skew_tick) {
1532 u64 offset = TICK_NSEC >> 1;
1533 do_div(offset, num_possible_cpus());
1534 offset *= smp_processor_id();
1535 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1538 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1539 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1540 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1542 #endif /* HIGH_RES_TIMERS */
1544 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1545 void tick_cancel_sched_timer(int cpu)
1547 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1549 # ifdef CONFIG_HIGH_RES_TIMERS
1550 if (ts->sched_timer.base)
1551 hrtimer_cancel(&ts->sched_timer);
1554 memset(ts, 0, sizeof(*ts));
1559 * Async notification about clocksource changes
1561 void tick_clock_notify(void)
1565 for_each_possible_cpu(cpu)
1566 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1570 * Async notification about clock event changes
1572 void tick_oneshot_notify(void)
1574 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1576 set_bit(0, &ts->check_clocks);
1580 * Check, if a change happened, which makes oneshot possible.
1582 * Called cyclic from the hrtimer softirq (driven by the timer
1583 * softirq) allow_nohz signals, that we can switch into low-res nohz
1584 * mode, because high resolution timers are disabled (either compile
1585 * or runtime). Called with interrupts disabled.
1587 int tick_check_oneshot_change(int allow_nohz)
1589 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1591 if (!test_and_clear_bit(0, &ts->check_clocks))
1594 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1597 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1603 tick_nohz_switch_to_nohz();