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);
287 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
289 lockdep_assert_irqs_disabled();
291 if (unlikely(!cpu_online(cpu)))
294 if (check_tick_dependency(&tick_dep_mask))
297 if (check_tick_dependency(&ts->tick_dep_mask))
300 if (check_tick_dependency(¤t->tick_dep_mask))
303 if (check_tick_dependency(¤t->signal->tick_dep_mask))
309 static void nohz_full_kick_func(struct irq_work *work)
311 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
314 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
315 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
318 * Kick this CPU if it's full dynticks in order to force it to
319 * re-evaluate its dependency on the tick and restart it if necessary.
320 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
323 static void tick_nohz_full_kick(void)
325 if (!tick_nohz_full_cpu(smp_processor_id()))
328 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
332 * Kick the CPU if it's full dynticks in order to force it to
333 * re-evaluate its dependency on the tick and restart it if necessary.
335 void tick_nohz_full_kick_cpu(int cpu)
337 if (!tick_nohz_full_cpu(cpu))
340 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
343 static void tick_nohz_kick_task(struct task_struct *tsk)
348 * If the task is not running, run_posix_cpu_timers()
349 * has nothing to elapse, IPI can then be spared.
351 * activate_task() STORE p->tick_dep_mask
353 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
354 * LOCK rq->lock LOAD p->on_rq
355 * smp_mb__after_spin_lock()
356 * tick_nohz_task_switch()
357 * LOAD p->tick_dep_mask
359 if (!sched_task_on_rq(tsk))
363 * If the task concurrently migrates to another CPU,
364 * we guarantee it sees the new tick dependency upon
367 * set_task_cpu(p, cpu);
368 * STORE p->cpu = @cpu
369 * __schedule() (switch to task 'p')
371 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
372 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
373 * LOAD p->tick_dep_mask LOAD p->cpu
379 tick_nohz_full_kick_cpu(cpu);
384 * Kick all full dynticks CPUs in order to force these to re-evaluate
385 * their dependency on the tick and restart it if necessary.
387 static void tick_nohz_full_kick_all(void)
391 if (!tick_nohz_full_running)
395 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
396 tick_nohz_full_kick_cpu(cpu);
400 static void tick_nohz_dep_set_all(atomic_t *dep,
401 enum tick_dep_bits bit)
405 prev = atomic_fetch_or(BIT(bit), dep);
407 tick_nohz_full_kick_all();
411 * Set a global tick dependency. Used by perf events that rely on freq and
414 void tick_nohz_dep_set(enum tick_dep_bits bit)
416 tick_nohz_dep_set_all(&tick_dep_mask, bit);
419 void tick_nohz_dep_clear(enum tick_dep_bits bit)
421 atomic_andnot(BIT(bit), &tick_dep_mask);
425 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
426 * manage events throttling.
428 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
431 struct tick_sched *ts;
433 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
435 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
438 /* Perf needs local kick that is NMI safe */
439 if (cpu == smp_processor_id()) {
440 tick_nohz_full_kick();
442 /* Remote irq work not NMI-safe */
443 if (!WARN_ON_ONCE(in_nmi()))
444 tick_nohz_full_kick_cpu(cpu);
449 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
451 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
453 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
455 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
457 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
460 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
461 * in order to elapse per task timers.
463 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
465 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
466 tick_nohz_kick_task(tsk);
468 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
470 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
472 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
474 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
477 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
478 * per process timers.
480 void tick_nohz_dep_set_signal(struct task_struct *tsk,
481 enum tick_dep_bits bit)
484 struct signal_struct *sig = tsk->signal;
486 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
488 struct task_struct *t;
490 lockdep_assert_held(&tsk->sighand->siglock);
491 __for_each_thread(sig, t)
492 tick_nohz_kick_task(t);
496 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
498 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
502 * Re-evaluate the need for the tick as we switch the current task.
503 * It might need the tick due to per task/process properties:
504 * perf events, posix CPU timers, ...
506 void __tick_nohz_task_switch(void)
508 struct tick_sched *ts;
510 if (!tick_nohz_full_cpu(smp_processor_id()))
513 ts = this_cpu_ptr(&tick_cpu_sched);
515 if (ts->tick_stopped) {
516 if (atomic_read(¤t->tick_dep_mask) ||
517 atomic_read(¤t->signal->tick_dep_mask))
518 tick_nohz_full_kick();
522 /* Get the boot-time nohz CPU list from the kernel parameters. */
523 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
525 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
526 cpumask_copy(tick_nohz_full_mask, cpumask);
527 tick_nohz_full_running = true;
530 static int tick_nohz_cpu_down(unsigned int cpu)
533 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
534 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
535 * CPUs. It must remain online when nohz full is enabled.
537 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
542 void __init tick_nohz_init(void)
546 if (!tick_nohz_full_running)
550 * Full dynticks uses irq work to drive the tick rescheduling on safe
551 * locking contexts. But then we need irq work to raise its own
552 * interrupts to avoid circular dependency on the tick
554 if (!arch_irq_work_has_interrupt()) {
555 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
556 cpumask_clear(tick_nohz_full_mask);
557 tick_nohz_full_running = false;
561 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
562 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
563 cpu = smp_processor_id();
565 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
566 pr_warn("NO_HZ: Clearing %d from nohz_full range "
567 "for timekeeping\n", cpu);
568 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
572 for_each_cpu(cpu, tick_nohz_full_mask)
573 context_tracking_cpu_set(cpu);
575 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
576 "kernel/nohz:predown", NULL,
579 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
580 cpumask_pr_args(tick_nohz_full_mask));
585 * NOHZ - aka dynamic tick functionality
587 #ifdef CONFIG_NO_HZ_COMMON
591 bool tick_nohz_enabled __read_mostly = true;
592 unsigned long tick_nohz_active __read_mostly;
594 * Enable / Disable tickless mode
596 static int __init setup_tick_nohz(char *str)
598 return (kstrtobool(str, &tick_nohz_enabled) == 0);
601 __setup("nohz=", setup_tick_nohz);
603 bool tick_nohz_tick_stopped(void)
605 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
607 return ts->tick_stopped;
610 bool tick_nohz_tick_stopped_cpu(int cpu)
612 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
614 return ts->tick_stopped;
618 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
620 * Called from interrupt entry when the CPU was idle
622 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
623 * must be updated. Otherwise an interrupt handler could use a stale jiffy
624 * value. We do this unconditionally on any CPU, as we don't know whether the
625 * CPU, which has the update task assigned is in a long sleep.
627 static void tick_nohz_update_jiffies(ktime_t now)
631 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
633 local_irq_save(flags);
634 tick_do_update_jiffies64(now);
635 local_irq_restore(flags);
637 touch_softlockup_watchdog_sched();
641 * Updates the per-CPU time idle statistics counters
644 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
648 if (ts->idle_active) {
649 delta = ktime_sub(now, ts->idle_entrytime);
650 if (nr_iowait_cpu(cpu) > 0)
651 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
653 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
654 ts->idle_entrytime = now;
657 if (last_update_time)
658 *last_update_time = ktime_to_us(now);
662 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
664 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
667 sched_clock_idle_wakeup_event();
670 static void tick_nohz_start_idle(struct tick_sched *ts)
672 ts->idle_entrytime = ktime_get();
674 sched_clock_idle_sleep_event();
678 * get_cpu_idle_time_us - get the total idle time of a CPU
679 * @cpu: CPU number to query
680 * @last_update_time: variable to store update time in. Do not update
683 * Return the cumulative idle time (since boot) for a given
684 * CPU, in microseconds.
686 * This time is measured via accounting rather than sampling,
687 * and is as accurate as ktime_get() is.
689 * This function returns -1 if NOHZ is not enabled.
691 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
693 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
696 if (!tick_nohz_active)
700 if (last_update_time) {
701 update_ts_time_stats(cpu, ts, now, last_update_time);
702 idle = ts->idle_sleeptime;
704 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
705 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
707 idle = ktime_add(ts->idle_sleeptime, delta);
709 idle = ts->idle_sleeptime;
713 return ktime_to_us(idle);
716 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
719 * get_cpu_iowait_time_us - get the total iowait time of a CPU
720 * @cpu: CPU number to query
721 * @last_update_time: variable to store update time in. Do not update
724 * Return the cumulative iowait time (since boot) for a given
725 * CPU, in microseconds.
727 * This time is measured via accounting rather than sampling,
728 * and is as accurate as ktime_get() is.
730 * This function returns -1 if NOHZ is not enabled.
732 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
734 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
737 if (!tick_nohz_active)
741 if (last_update_time) {
742 update_ts_time_stats(cpu, ts, now, last_update_time);
743 iowait = ts->iowait_sleeptime;
745 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
746 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
748 iowait = ktime_add(ts->iowait_sleeptime, delta);
750 iowait = ts->iowait_sleeptime;
754 return ktime_to_us(iowait);
756 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
758 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
760 hrtimer_cancel(&ts->sched_timer);
761 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
763 /* Forward the time to expire in the future */
764 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
766 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
767 hrtimer_start_expires(&ts->sched_timer,
768 HRTIMER_MODE_ABS_PINNED_HARD);
770 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
774 * Reset to make sure next tick stop doesn't get fooled by past
775 * cached clock deadline.
780 static inline bool local_timer_softirq_pending(void)
782 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
785 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
787 u64 basemono, next_tick, delta, expires;
788 unsigned long basejiff;
791 /* Read jiffies and the time when jiffies were updated last */
793 seq = read_seqcount_begin(&jiffies_seq);
794 basemono = last_jiffies_update;
796 } while (read_seqcount_retry(&jiffies_seq, seq));
797 ts->last_jiffies = basejiff;
798 ts->timer_expires_base = basemono;
801 * Keep the periodic tick, when RCU, architecture or irq_work
803 * Aside of that check whether the local timer softirq is
804 * pending. If so its a bad idea to call get_next_timer_interrupt()
805 * because there is an already expired timer, so it will request
806 * immediate expiry, which rearms the hardware timer with a
807 * minimal delta which brings us back to this place
808 * immediately. Lather, rinse and repeat...
810 if (rcu_needs_cpu() || arch_needs_cpu() ||
811 irq_work_needs_cpu() || local_timer_softirq_pending()) {
812 next_tick = basemono + TICK_NSEC;
815 * Get the next pending timer. If high resolution
816 * timers are enabled this only takes the timer wheel
817 * timers into account. If high resolution timers are
818 * disabled this also looks at the next expiring
821 next_tick = get_next_timer_interrupt(basejiff, basemono);
822 ts->next_timer = next_tick;
826 * If the tick is due in the next period, keep it ticking or
827 * force prod the timer.
829 delta = next_tick - basemono;
830 if (delta <= (u64)TICK_NSEC) {
832 * Tell the timer code that the base is not idle, i.e. undo
833 * the effect of get_next_timer_interrupt():
837 * We've not stopped the tick yet, and there's a timer in the
838 * next period, so no point in stopping it either, bail.
840 if (!ts->tick_stopped) {
841 ts->timer_expires = 0;
847 * If this CPU is the one which had the do_timer() duty last, we limit
848 * the sleep time to the timekeeping max_deferment value.
849 * Otherwise we can sleep as long as we want.
851 delta = timekeeping_max_deferment();
852 if (cpu != tick_do_timer_cpu &&
853 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
856 /* Calculate the next expiry time */
857 if (delta < (KTIME_MAX - basemono))
858 expires = basemono + delta;
862 ts->timer_expires = min_t(u64, expires, next_tick);
865 return ts->timer_expires;
868 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
870 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
871 u64 basemono = ts->timer_expires_base;
872 u64 expires = ts->timer_expires;
873 ktime_t tick = expires;
875 /* Make sure we won't be trying to stop it twice in a row. */
876 ts->timer_expires_base = 0;
879 * If this CPU is the one which updates jiffies, then give up
880 * the assignment and let it be taken by the CPU which runs
881 * the tick timer next, which might be this CPU as well. If we
882 * don't drop this here the jiffies might be stale and
883 * do_timer() never invoked. Keep track of the fact that it
884 * was the one which had the do_timer() duty last.
886 if (cpu == tick_do_timer_cpu) {
887 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
888 ts->do_timer_last = 1;
889 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
890 ts->do_timer_last = 0;
893 /* Skip reprogram of event if its not changed */
894 if (ts->tick_stopped && (expires == ts->next_tick)) {
895 /* Sanity check: make sure clockevent is actually programmed */
896 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
900 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
901 basemono, ts->next_tick, dev->next_event,
902 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
906 * nohz_stop_sched_tick can be called several times before
907 * the nohz_restart_sched_tick is called. This happens when
908 * interrupts arrive which do not cause a reschedule. In the
909 * first call we save the current tick time, so we can restart
910 * the scheduler tick in nohz_restart_sched_tick.
912 if (!ts->tick_stopped) {
913 calc_load_nohz_start();
916 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
917 ts->tick_stopped = 1;
918 trace_tick_stop(1, TICK_DEP_MASK_NONE);
921 ts->next_tick = tick;
924 * If the expiration time == KTIME_MAX, then we simply stop
927 if (unlikely(expires == KTIME_MAX)) {
928 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
929 hrtimer_cancel(&ts->sched_timer);
931 tick_program_event(KTIME_MAX, 1);
935 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
936 hrtimer_start(&ts->sched_timer, tick,
937 HRTIMER_MODE_ABS_PINNED_HARD);
939 hrtimer_set_expires(&ts->sched_timer, tick);
940 tick_program_event(tick, 1);
944 static void tick_nohz_retain_tick(struct tick_sched *ts)
946 ts->timer_expires_base = 0;
949 #ifdef CONFIG_NO_HZ_FULL
950 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
952 if (tick_nohz_next_event(ts, cpu))
953 tick_nohz_stop_tick(ts, cpu);
955 tick_nohz_retain_tick(ts);
957 #endif /* CONFIG_NO_HZ_FULL */
959 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
961 /* Update jiffies first */
962 tick_do_update_jiffies64(now);
965 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
966 * the clock forward checks in the enqueue path:
970 calc_load_nohz_stop();
971 touch_softlockup_watchdog_sched();
973 * Cancel the scheduled timer and restore the tick
975 ts->tick_stopped = 0;
976 tick_nohz_restart(ts, now);
979 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
982 #ifdef CONFIG_NO_HZ_FULL
983 int cpu = smp_processor_id();
985 if (can_stop_full_tick(cpu, ts))
986 tick_nohz_stop_sched_tick(ts, cpu);
987 else if (ts->tick_stopped)
988 tick_nohz_restart_sched_tick(ts, now);
992 static void tick_nohz_full_update_tick(struct tick_sched *ts)
994 if (!tick_nohz_full_cpu(smp_processor_id()))
997 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1000 __tick_nohz_full_update_tick(ts, ktime_get());
1004 * A pending softirq outside an IRQ (or softirq disabled section) context
1005 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1006 * reach here due to the need_resched() early check in can_stop_idle_tick().
1008 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1009 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1010 * triggering the below since wakep_softirqd() is ignored.
1013 static bool report_idle_softirq(void)
1015 static int ratelimit;
1016 unsigned int pending = local_softirq_pending();
1018 if (likely(!pending))
1021 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1022 if (!cpu_active(smp_processor_id())) {
1023 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1031 /* On RT, softirqs handling may be waiting on some lock */
1032 if (!local_bh_blocked())
1035 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1042 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1045 * If this CPU is offline and it is the one which updates
1046 * jiffies, then give up the assignment and let it be taken by
1047 * the CPU which runs the tick timer next. If we don't drop
1048 * this here the jiffies might be stale and do_timer() never
1051 if (unlikely(!cpu_online(cpu))) {
1052 if (cpu == tick_do_timer_cpu)
1053 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1055 * Make sure the CPU doesn't get fooled by obsolete tick
1056 * deadline if it comes back online later.
1062 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1068 if (unlikely(report_idle_softirq()))
1071 if (tick_nohz_full_enabled()) {
1073 * Keep the tick alive to guarantee timekeeping progression
1074 * if there are full dynticks CPUs around
1076 if (tick_do_timer_cpu == cpu)
1079 /* Should not happen for nohz-full */
1080 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1087 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1090 int cpu = smp_processor_id();
1093 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1094 * tick timer expiration time is known already.
1096 if (ts->timer_expires_base)
1097 expires = ts->timer_expires;
1098 else if (can_stop_idle_tick(cpu, ts))
1099 expires = tick_nohz_next_event(ts, cpu);
1105 if (expires > 0LL) {
1106 int was_stopped = ts->tick_stopped;
1108 tick_nohz_stop_tick(ts, cpu);
1111 ts->idle_expires = expires;
1113 if (!was_stopped && ts->tick_stopped) {
1114 ts->idle_jiffies = ts->last_jiffies;
1115 nohz_balance_enter_idle(cpu);
1118 tick_nohz_retain_tick(ts);
1123 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1125 * When the next event is more than a tick into the future, stop the idle tick
1127 void tick_nohz_idle_stop_tick(void)
1129 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1132 void tick_nohz_idle_retain_tick(void)
1134 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1136 * Undo the effect of get_next_timer_interrupt() called from
1137 * tick_nohz_next_event().
1143 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1145 * Called when we start the idle loop.
1147 void tick_nohz_idle_enter(void)
1149 struct tick_sched *ts;
1151 lockdep_assert_irqs_enabled();
1153 local_irq_disable();
1155 ts = this_cpu_ptr(&tick_cpu_sched);
1157 WARN_ON_ONCE(ts->timer_expires_base);
1160 tick_nohz_start_idle(ts);
1166 * tick_nohz_irq_exit - update next tick event from interrupt exit
1168 * When an interrupt fires while we are idle and it doesn't cause
1169 * a reschedule, it may still add, modify or delete a timer, enqueue
1170 * an RCU callback, etc...
1171 * So we need to re-calculate and reprogram the next tick event.
1173 void tick_nohz_irq_exit(void)
1175 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1178 tick_nohz_start_idle(ts);
1180 tick_nohz_full_update_tick(ts);
1184 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1186 bool tick_nohz_idle_got_tick(void)
1188 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1190 if (ts->got_idle_tick) {
1191 ts->got_idle_tick = 0;
1198 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1199 * or the tick, whatever that expires first. Note that, if the tick has been
1200 * stopped, it returns the next hrtimer.
1202 * Called from power state control code with interrupts disabled
1204 ktime_t tick_nohz_get_next_hrtimer(void)
1206 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1210 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1211 * @delta_next: duration until the next event if the tick cannot be stopped
1213 * Called from power state control code with interrupts disabled.
1215 * The return value of this function and/or the value returned by it through the
1216 * @delta_next pointer can be negative which must be taken into account by its
1219 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1221 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1222 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 int cpu = smp_processor_id();
1225 * The idle entry time is expected to be a sufficient approximation of
1226 * the current time at this point.
1228 ktime_t now = ts->idle_entrytime;
1231 WARN_ON_ONCE(!ts->inidle);
1233 *delta_next = ktime_sub(dev->next_event, now);
1235 if (!can_stop_idle_tick(cpu, ts))
1238 next_event = tick_nohz_next_event(ts, cpu);
1243 * If the next highres timer to expire is earlier than next_event, the
1244 * idle governor needs to know that.
1246 next_event = min_t(u64, next_event,
1247 hrtimer_next_event_without(&ts->sched_timer));
1249 return ktime_sub(next_event, now);
1253 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1254 * for a particular CPU.
1256 * Called from the schedutil frequency scaling governor in scheduler context.
1258 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1260 struct tick_sched *ts = tick_get_tick_sched(cpu);
1262 return ts->idle_calls;
1266 * tick_nohz_get_idle_calls - return the current idle calls counter value
1268 * Called from the schedutil frequency scaling governor in scheduler context.
1270 unsigned long tick_nohz_get_idle_calls(void)
1272 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1274 return ts->idle_calls;
1277 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1280 unsigned long ticks;
1282 ts->idle_exittime = now;
1284 if (vtime_accounting_enabled_this_cpu())
1287 * We stopped the tick in idle. Update process times would miss the
1288 * time we slept as update_process_times does only a 1 tick
1289 * accounting. Enforce that this is accounted to idle !
1291 ticks = jiffies - ts->idle_jiffies;
1293 * We might be one off. Do not randomly account a huge number of ticks!
1295 if (ticks && ticks < LONG_MAX)
1296 account_idle_ticks(ticks);
1299 void tick_nohz_idle_restart_tick(void)
1301 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1303 if (ts->tick_stopped) {
1304 ktime_t now = ktime_get();
1305 tick_nohz_restart_sched_tick(ts, now);
1306 tick_nohz_account_idle_time(ts, now);
1310 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1312 if (tick_nohz_full_cpu(smp_processor_id()))
1313 __tick_nohz_full_update_tick(ts, now);
1315 tick_nohz_restart_sched_tick(ts, now);
1317 tick_nohz_account_idle_time(ts, now);
1321 * tick_nohz_idle_exit - restart the idle tick from the idle task
1323 * Restart the idle tick when the CPU is woken up from idle
1324 * This also exit the RCU extended quiescent state. The CPU
1325 * can use RCU again after this function is called.
1327 void tick_nohz_idle_exit(void)
1329 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1330 bool idle_active, tick_stopped;
1333 local_irq_disable();
1335 WARN_ON_ONCE(!ts->inidle);
1336 WARN_ON_ONCE(ts->timer_expires_base);
1339 idle_active = ts->idle_active;
1340 tick_stopped = ts->tick_stopped;
1342 if (idle_active || tick_stopped)
1346 tick_nohz_stop_idle(ts, now);
1349 tick_nohz_idle_update_tick(ts, now);
1355 * The nohz low res interrupt handler
1357 static void tick_nohz_handler(struct clock_event_device *dev)
1359 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1360 struct pt_regs *regs = get_irq_regs();
1361 ktime_t now = ktime_get();
1363 dev->next_event = KTIME_MAX;
1365 tick_sched_do_timer(ts, now);
1366 tick_sched_handle(ts, regs);
1368 if (unlikely(ts->tick_stopped)) {
1370 * The clockevent device is not reprogrammed, so change the
1371 * clock event device to ONESHOT_STOPPED to avoid spurious
1372 * interrupts on devices which might not be truly one shot.
1374 tick_program_event(KTIME_MAX, 1);
1378 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1379 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1382 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1384 if (!tick_nohz_enabled)
1386 ts->nohz_mode = mode;
1387 /* One update is enough */
1388 if (!test_and_set_bit(0, &tick_nohz_active))
1389 timers_update_nohz();
1393 * tick_nohz_switch_to_nohz - switch to nohz mode
1395 static void tick_nohz_switch_to_nohz(void)
1397 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1400 if (!tick_nohz_enabled)
1403 if (tick_switch_to_oneshot(tick_nohz_handler))
1407 * Recycle the hrtimer in ts, so we can share the
1408 * hrtimer_forward with the highres code.
1410 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1411 /* Get the next period */
1412 next = tick_init_jiffy_update();
1414 hrtimer_set_expires(&ts->sched_timer, next);
1415 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1420 static inline void tick_nohz_irq_enter(void)
1422 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1425 if (!ts->idle_active && !ts->tick_stopped)
1428 if (ts->idle_active)
1429 tick_nohz_stop_idle(ts, now);
1431 * If all CPUs are idle. We may need to update a stale jiffies value.
1432 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1433 * alive but it might be busy looping with interrupts disabled in some
1434 * rare case (typically stop machine). So we must make sure we have a
1437 if (ts->tick_stopped)
1438 tick_nohz_update_jiffies(now);
1443 static inline void tick_nohz_switch_to_nohz(void) { }
1444 static inline void tick_nohz_irq_enter(void) { }
1445 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1447 #endif /* CONFIG_NO_HZ_COMMON */
1450 * Called from irq_enter to notify about the possible interruption of idle()
1452 void tick_irq_enter(void)
1454 tick_check_oneshot_broadcast_this_cpu();
1455 tick_nohz_irq_enter();
1459 * High resolution timer specific code
1461 #ifdef CONFIG_HIGH_RES_TIMERS
1463 * We rearm the timer until we get disabled by the idle code.
1464 * Called with interrupts disabled.
1466 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1468 struct tick_sched *ts =
1469 container_of(timer, struct tick_sched, sched_timer);
1470 struct pt_regs *regs = get_irq_regs();
1471 ktime_t now = ktime_get();
1473 tick_sched_do_timer(ts, now);
1476 * Do not call, when we are not in irq context and have
1477 * no valid regs pointer
1480 tick_sched_handle(ts, regs);
1484 /* No need to reprogram if we are in idle or full dynticks mode */
1485 if (unlikely(ts->tick_stopped))
1486 return HRTIMER_NORESTART;
1488 hrtimer_forward(timer, now, TICK_NSEC);
1490 return HRTIMER_RESTART;
1493 static int sched_skew_tick;
1495 static int __init skew_tick(char *str)
1497 get_option(&str, &sched_skew_tick);
1501 early_param("skew_tick", skew_tick);
1504 * tick_setup_sched_timer - setup the tick emulation timer
1506 void tick_setup_sched_timer(void)
1508 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1509 ktime_t now = ktime_get();
1512 * Emulate tick processing via per-CPU hrtimers:
1514 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1515 ts->sched_timer.function = tick_sched_timer;
1517 /* Get the next period (per-CPU) */
1518 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1520 /* Offset the tick to avert jiffies_lock contention. */
1521 if (sched_skew_tick) {
1522 u64 offset = TICK_NSEC >> 1;
1523 do_div(offset, num_possible_cpus());
1524 offset *= smp_processor_id();
1525 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1528 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1529 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1530 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1532 #endif /* HIGH_RES_TIMERS */
1534 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1535 void tick_cancel_sched_timer(int cpu)
1537 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1539 # ifdef CONFIG_HIGH_RES_TIMERS
1540 if (ts->sched_timer.base)
1541 hrtimer_cancel(&ts->sched_timer);
1544 memset(ts, 0, sizeof(*ts));
1549 * Async notification about clocksource changes
1551 void tick_clock_notify(void)
1555 for_each_possible_cpu(cpu)
1556 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1560 * Async notification about clock event changes
1562 void tick_oneshot_notify(void)
1564 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1566 set_bit(0, &ts->check_clocks);
1570 * Check, if a change happened, which makes oneshot possible.
1572 * Called cyclic from the hrtimer softirq (driven by the timer
1573 * softirq) allow_nohz signals, that we can switch into low-res nohz
1574 * mode, because high resolution timers are disabled (either compile
1575 * or runtime). Called with interrupts disabled.
1577 int tick_check_oneshot_change(int allow_nohz)
1579 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1581 if (!test_and_clear_bit(0, &ts->check_clocks))
1584 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1587 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1593 tick_nohz_switch_to_nohz();