Merge tag 'regmap-fix-v5.11-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux/fpc-iii.git] / kernel / time / tick-sched.c
blobe10a4af887373bc8a693b750fd27ff1e2460df3f
1 // SPDX-License-Identifier: GPL-2.0
2 /*
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>
28 #include <linux/mm.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;
60 ktime_t delta, nextp;
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)))
73 return;
74 } else {
75 unsigned int seq;
78 * Avoid contention on jiffies_lock and protect the quick
79 * check with the sequence count.
81 do {
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))
87 return;
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
94 * update already.
96 if (ktime_before(now, tick_next_period)) {
97 raw_spin_unlock(&jiffies_lock);
98 return;
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,
111 incr * ticks);
112 } else {
113 last_jiffies_update = ktime_add_ns(last_jiffies_update,
114 TICK_NSEC);
117 /* Advance jiffies to complete the jiffies_seq protected job */
118 jiffies_64 += ticks;
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);
133 } else {
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);
148 calc_global_load();
150 raw_spin_unlock(&jiffies_lock);
151 update_wall_time();
155 * Initialize and return retrieve the jiffies update.
157 static ktime_t tick_init_jiffy_update(void)
159 ktime_t period;
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);
169 return period;
172 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
174 int cpu = smp_processor_id();
176 #ifdef CONFIG_NO_HZ_COMMON
178 * Check if the do_timer duty was dropped. We don't care about
179 * concurrency: This happens only when the CPU in charge went
180 * into a long sleep. If two CPUs happen to assign themselves to
181 * this duty, then the jiffies update is still serialized by
182 * jiffies_lock.
184 * If nohz_full is enabled, this should not happen because the
185 * tick_do_timer_cpu never relinquishes.
187 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
188 #ifdef CONFIG_NO_HZ_FULL
189 WARN_ON(tick_nohz_full_running);
190 #endif
191 tick_do_timer_cpu = cpu;
193 #endif
195 /* Check, if the jiffies need an update */
196 if (tick_do_timer_cpu == cpu)
197 tick_do_update_jiffies64(now);
199 if (ts->inidle)
200 ts->got_idle_tick = 1;
203 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
205 #ifdef CONFIG_NO_HZ_COMMON
207 * When we are idle and the tick is stopped, we have to touch
208 * the watchdog as we might not schedule for a really long
209 * time. This happens on complete idle SMP systems while
210 * waiting on the login prompt. We also increment the "start of
211 * idle" jiffy stamp so the idle accounting adjustment we do
212 * when we go busy again does not account too much ticks.
214 if (ts->tick_stopped) {
215 touch_softlockup_watchdog_sched();
216 if (is_idle_task(current))
217 ts->idle_jiffies++;
219 * In case the current tick fired too early past its expected
220 * expiration, make sure we don't bypass the next clock reprogramming
221 * to the same deadline.
223 ts->next_tick = 0;
225 #endif
226 update_process_times(user_mode(regs));
227 profile_tick(CPU_PROFILING);
229 #endif
231 #ifdef CONFIG_NO_HZ_FULL
232 cpumask_var_t tick_nohz_full_mask;
233 bool tick_nohz_full_running;
234 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
235 static atomic_t tick_dep_mask;
237 static bool check_tick_dependency(atomic_t *dep)
239 int val = atomic_read(dep);
241 if (val & TICK_DEP_MASK_POSIX_TIMER) {
242 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
243 return true;
246 if (val & TICK_DEP_MASK_PERF_EVENTS) {
247 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
248 return true;
251 if (val & TICK_DEP_MASK_SCHED) {
252 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
253 return true;
256 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
257 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
258 return true;
261 if (val & TICK_DEP_MASK_RCU) {
262 trace_tick_stop(0, TICK_DEP_MASK_RCU);
263 return true;
266 return false;
269 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
271 lockdep_assert_irqs_disabled();
273 if (unlikely(!cpu_online(cpu)))
274 return false;
276 if (check_tick_dependency(&tick_dep_mask))
277 return false;
279 if (check_tick_dependency(&ts->tick_dep_mask))
280 return false;
282 if (check_tick_dependency(&current->tick_dep_mask))
283 return false;
285 if (check_tick_dependency(&current->signal->tick_dep_mask))
286 return false;
288 return true;
291 static void nohz_full_kick_func(struct irq_work *work)
293 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
296 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
297 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
300 * Kick this CPU if it's full dynticks in order to force it to
301 * re-evaluate its dependency on the tick and restart it if necessary.
302 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
303 * is NMI safe.
305 static void tick_nohz_full_kick(void)
307 if (!tick_nohz_full_cpu(smp_processor_id()))
308 return;
310 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
314 * Kick the CPU if it's full dynticks in order to force it to
315 * re-evaluate its dependency on the tick and restart it if necessary.
317 void tick_nohz_full_kick_cpu(int cpu)
319 if (!tick_nohz_full_cpu(cpu))
320 return;
322 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
326 * Kick all full dynticks CPUs in order to force these to re-evaluate
327 * their dependency on the tick and restart it if necessary.
329 static void tick_nohz_full_kick_all(void)
331 int cpu;
333 if (!tick_nohz_full_running)
334 return;
336 preempt_disable();
337 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
338 tick_nohz_full_kick_cpu(cpu);
339 preempt_enable();
342 static void tick_nohz_dep_set_all(atomic_t *dep,
343 enum tick_dep_bits bit)
345 int prev;
347 prev = atomic_fetch_or(BIT(bit), dep);
348 if (!prev)
349 tick_nohz_full_kick_all();
353 * Set a global tick dependency. Used by perf events that rely on freq and
354 * by unstable clock.
356 void tick_nohz_dep_set(enum tick_dep_bits bit)
358 tick_nohz_dep_set_all(&tick_dep_mask, bit);
361 void tick_nohz_dep_clear(enum tick_dep_bits bit)
363 atomic_andnot(BIT(bit), &tick_dep_mask);
367 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
368 * manage events throttling.
370 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
372 int prev;
373 struct tick_sched *ts;
375 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
377 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
378 if (!prev) {
379 preempt_disable();
380 /* Perf needs local kick that is NMI safe */
381 if (cpu == smp_processor_id()) {
382 tick_nohz_full_kick();
383 } else {
384 /* Remote irq work not NMI-safe */
385 if (!WARN_ON_ONCE(in_nmi()))
386 tick_nohz_full_kick_cpu(cpu);
388 preempt_enable();
391 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
393 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
395 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
397 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
399 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
402 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
403 * in order to elapse per task timers.
405 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
407 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask)) {
408 if (tsk == current) {
409 preempt_disable();
410 tick_nohz_full_kick();
411 preempt_enable();
412 } else {
414 * Some future tick_nohz_full_kick_task()
415 * should optimize this.
417 tick_nohz_full_kick_all();
421 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
423 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
425 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
427 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
430 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
431 * per process timers.
433 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
435 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
438 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
440 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
444 * Re-evaluate the need for the tick as we switch the current task.
445 * It might need the tick due to per task/process properties:
446 * perf events, posix CPU timers, ...
448 void __tick_nohz_task_switch(void)
450 unsigned long flags;
451 struct tick_sched *ts;
453 local_irq_save(flags);
455 if (!tick_nohz_full_cpu(smp_processor_id()))
456 goto out;
458 ts = this_cpu_ptr(&tick_cpu_sched);
460 if (ts->tick_stopped) {
461 if (atomic_read(&current->tick_dep_mask) ||
462 atomic_read(&current->signal->tick_dep_mask))
463 tick_nohz_full_kick();
465 out:
466 local_irq_restore(flags);
469 /* Get the boot-time nohz CPU list from the kernel parameters. */
470 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
472 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
473 cpumask_copy(tick_nohz_full_mask, cpumask);
474 tick_nohz_full_running = true;
476 EXPORT_SYMBOL_GPL(tick_nohz_full_setup);
478 static int tick_nohz_cpu_down(unsigned int cpu)
481 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
482 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
483 * CPUs. It must remain online when nohz full is enabled.
485 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
486 return -EBUSY;
487 return 0;
490 void __init tick_nohz_init(void)
492 int cpu, ret;
494 if (!tick_nohz_full_running)
495 return;
498 * Full dynticks uses irq work to drive the tick rescheduling on safe
499 * locking contexts. But then we need irq work to raise its own
500 * interrupts to avoid circular dependency on the tick
502 if (!arch_irq_work_has_interrupt()) {
503 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
504 cpumask_clear(tick_nohz_full_mask);
505 tick_nohz_full_running = false;
506 return;
509 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
510 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
511 cpu = smp_processor_id();
513 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
514 pr_warn("NO_HZ: Clearing %d from nohz_full range "
515 "for timekeeping\n", cpu);
516 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
520 for_each_cpu(cpu, tick_nohz_full_mask)
521 context_tracking_cpu_set(cpu);
523 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
524 "kernel/nohz:predown", NULL,
525 tick_nohz_cpu_down);
526 WARN_ON(ret < 0);
527 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
528 cpumask_pr_args(tick_nohz_full_mask));
530 #endif
533 * NOHZ - aka dynamic tick functionality
535 #ifdef CONFIG_NO_HZ_COMMON
537 * NO HZ enabled ?
539 bool tick_nohz_enabled __read_mostly = true;
540 unsigned long tick_nohz_active __read_mostly;
542 * Enable / Disable tickless mode
544 static int __init setup_tick_nohz(char *str)
546 return (kstrtobool(str, &tick_nohz_enabled) == 0);
549 __setup("nohz=", setup_tick_nohz);
551 bool tick_nohz_tick_stopped(void)
553 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
555 return ts->tick_stopped;
558 bool tick_nohz_tick_stopped_cpu(int cpu)
560 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
562 return ts->tick_stopped;
566 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
568 * Called from interrupt entry when the CPU was idle
570 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
571 * must be updated. Otherwise an interrupt handler could use a stale jiffy
572 * value. We do this unconditionally on any CPU, as we don't know whether the
573 * CPU, which has the update task assigned is in a long sleep.
575 static void tick_nohz_update_jiffies(ktime_t now)
577 unsigned long flags;
579 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
581 local_irq_save(flags);
582 tick_do_update_jiffies64(now);
583 local_irq_restore(flags);
585 touch_softlockup_watchdog_sched();
589 * Updates the per-CPU time idle statistics counters
591 static void
592 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
594 ktime_t delta;
596 if (ts->idle_active) {
597 delta = ktime_sub(now, ts->idle_entrytime);
598 if (nr_iowait_cpu(cpu) > 0)
599 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
600 else
601 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
602 ts->idle_entrytime = now;
605 if (last_update_time)
606 *last_update_time = ktime_to_us(now);
610 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
612 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
613 ts->idle_active = 0;
615 sched_clock_idle_wakeup_event();
618 static void tick_nohz_start_idle(struct tick_sched *ts)
620 ts->idle_entrytime = ktime_get();
621 ts->idle_active = 1;
622 sched_clock_idle_sleep_event();
626 * get_cpu_idle_time_us - get the total idle time of a CPU
627 * @cpu: CPU number to query
628 * @last_update_time: variable to store update time in. Do not update
629 * counters if NULL.
631 * Return the cumulative idle time (since boot) for a given
632 * CPU, in microseconds.
634 * This time is measured via accounting rather than sampling,
635 * and is as accurate as ktime_get() is.
637 * This function returns -1 if NOHZ is not enabled.
639 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
641 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
642 ktime_t now, idle;
644 if (!tick_nohz_active)
645 return -1;
647 now = ktime_get();
648 if (last_update_time) {
649 update_ts_time_stats(cpu, ts, now, last_update_time);
650 idle = ts->idle_sleeptime;
651 } else {
652 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
653 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
655 idle = ktime_add(ts->idle_sleeptime, delta);
656 } else {
657 idle = ts->idle_sleeptime;
661 return ktime_to_us(idle);
664 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
667 * get_cpu_iowait_time_us - get the total iowait time of a CPU
668 * @cpu: CPU number to query
669 * @last_update_time: variable to store update time in. Do not update
670 * counters if NULL.
672 * Return the cumulative iowait time (since boot) for a given
673 * CPU, in microseconds.
675 * This time is measured via accounting rather than sampling,
676 * and is as accurate as ktime_get() is.
678 * This function returns -1 if NOHZ is not enabled.
680 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
682 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
683 ktime_t now, iowait;
685 if (!tick_nohz_active)
686 return -1;
688 now = ktime_get();
689 if (last_update_time) {
690 update_ts_time_stats(cpu, ts, now, last_update_time);
691 iowait = ts->iowait_sleeptime;
692 } else {
693 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
694 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
696 iowait = ktime_add(ts->iowait_sleeptime, delta);
697 } else {
698 iowait = ts->iowait_sleeptime;
702 return ktime_to_us(iowait);
704 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
706 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
708 hrtimer_cancel(&ts->sched_timer);
709 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
711 /* Forward the time to expire in the future */
712 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
714 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
715 hrtimer_start_expires(&ts->sched_timer,
716 HRTIMER_MODE_ABS_PINNED_HARD);
717 } else {
718 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
722 * Reset to make sure next tick stop doesn't get fooled by past
723 * cached clock deadline.
725 ts->next_tick = 0;
728 static inline bool local_timer_softirq_pending(void)
730 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
733 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
735 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
736 unsigned long basejiff;
737 unsigned int seq;
739 /* Read jiffies and the time when jiffies were updated last */
740 do {
741 seq = read_seqcount_begin(&jiffies_seq);
742 basemono = last_jiffies_update;
743 basejiff = jiffies;
744 } while (read_seqcount_retry(&jiffies_seq, seq));
745 ts->last_jiffies = basejiff;
746 ts->timer_expires_base = basemono;
749 * Keep the periodic tick, when RCU, architecture or irq_work
750 * requests it.
751 * Aside of that check whether the local timer softirq is
752 * pending. If so its a bad idea to call get_next_timer_interrupt()
753 * because there is an already expired timer, so it will request
754 * immeditate expiry, which rearms the hardware timer with a
755 * minimal delta which brings us back to this place
756 * immediately. Lather, rinse and repeat...
758 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
759 irq_work_needs_cpu() || local_timer_softirq_pending()) {
760 next_tick = basemono + TICK_NSEC;
761 } else {
763 * Get the next pending timer. If high resolution
764 * timers are enabled this only takes the timer wheel
765 * timers into account. If high resolution timers are
766 * disabled this also looks at the next expiring
767 * hrtimer.
769 next_tmr = get_next_timer_interrupt(basejiff, basemono);
770 ts->next_timer = next_tmr;
771 /* Take the next rcu event into account */
772 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
776 * If the tick is due in the next period, keep it ticking or
777 * force prod the timer.
779 delta = next_tick - basemono;
780 if (delta <= (u64)TICK_NSEC) {
782 * Tell the timer code that the base is not idle, i.e. undo
783 * the effect of get_next_timer_interrupt():
785 timer_clear_idle();
787 * We've not stopped the tick yet, and there's a timer in the
788 * next period, so no point in stopping it either, bail.
790 if (!ts->tick_stopped) {
791 ts->timer_expires = 0;
792 goto out;
797 * If this CPU is the one which had the do_timer() duty last, we limit
798 * the sleep time to the timekeeping max_deferment value.
799 * Otherwise we can sleep as long as we want.
801 delta = timekeeping_max_deferment();
802 if (cpu != tick_do_timer_cpu &&
803 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
804 delta = KTIME_MAX;
806 /* Calculate the next expiry time */
807 if (delta < (KTIME_MAX - basemono))
808 expires = basemono + delta;
809 else
810 expires = KTIME_MAX;
812 ts->timer_expires = min_t(u64, expires, next_tick);
814 out:
815 return ts->timer_expires;
818 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
820 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
821 u64 basemono = ts->timer_expires_base;
822 u64 expires = ts->timer_expires;
823 ktime_t tick = expires;
825 /* Make sure we won't be trying to stop it twice in a row. */
826 ts->timer_expires_base = 0;
829 * If this CPU is the one which updates jiffies, then give up
830 * the assignment and let it be taken by the CPU which runs
831 * the tick timer next, which might be this CPU as well. If we
832 * don't drop this here the jiffies might be stale and
833 * do_timer() never invoked. Keep track of the fact that it
834 * was the one which had the do_timer() duty last.
836 if (cpu == tick_do_timer_cpu) {
837 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
838 ts->do_timer_last = 1;
839 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
840 ts->do_timer_last = 0;
843 /* Skip reprogram of event if its not changed */
844 if (ts->tick_stopped && (expires == ts->next_tick)) {
845 /* Sanity check: make sure clockevent is actually programmed */
846 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
847 return;
849 WARN_ON_ONCE(1);
850 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
851 basemono, ts->next_tick, dev->next_event,
852 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
856 * nohz_stop_sched_tick can be called several times before
857 * the nohz_restart_sched_tick is called. This happens when
858 * interrupts arrive which do not cause a reschedule. In the
859 * first call we save the current tick time, so we can restart
860 * the scheduler tick in nohz_restart_sched_tick.
862 if (!ts->tick_stopped) {
863 calc_load_nohz_start();
864 quiet_vmstat();
866 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
867 ts->tick_stopped = 1;
868 trace_tick_stop(1, TICK_DEP_MASK_NONE);
871 ts->next_tick = tick;
874 * If the expiration time == KTIME_MAX, then we simply stop
875 * the tick timer.
877 if (unlikely(expires == KTIME_MAX)) {
878 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
879 hrtimer_cancel(&ts->sched_timer);
880 return;
883 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
884 hrtimer_start(&ts->sched_timer, tick,
885 HRTIMER_MODE_ABS_PINNED_HARD);
886 } else {
887 hrtimer_set_expires(&ts->sched_timer, tick);
888 tick_program_event(tick, 1);
892 static void tick_nohz_retain_tick(struct tick_sched *ts)
894 ts->timer_expires_base = 0;
897 #ifdef CONFIG_NO_HZ_FULL
898 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
900 if (tick_nohz_next_event(ts, cpu))
901 tick_nohz_stop_tick(ts, cpu);
902 else
903 tick_nohz_retain_tick(ts);
905 #endif /* CONFIG_NO_HZ_FULL */
907 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
909 /* Update jiffies first */
910 tick_do_update_jiffies64(now);
913 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
914 * the clock forward checks in the enqueue path:
916 timer_clear_idle();
918 calc_load_nohz_stop();
919 touch_softlockup_watchdog_sched();
921 * Cancel the scheduled timer and restore the tick
923 ts->tick_stopped = 0;
924 ts->idle_exittime = now;
926 tick_nohz_restart(ts, now);
929 static void tick_nohz_full_update_tick(struct tick_sched *ts)
931 #ifdef CONFIG_NO_HZ_FULL
932 int cpu = smp_processor_id();
934 if (!tick_nohz_full_cpu(cpu))
935 return;
937 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
938 return;
940 if (can_stop_full_tick(cpu, ts))
941 tick_nohz_stop_sched_tick(ts, cpu);
942 else if (ts->tick_stopped)
943 tick_nohz_restart_sched_tick(ts, ktime_get());
944 #endif
947 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
950 * If this CPU is offline and it is the one which updates
951 * jiffies, then give up the assignment and let it be taken by
952 * the CPU which runs the tick timer next. If we don't drop
953 * this here the jiffies might be stale and do_timer() never
954 * invoked.
956 if (unlikely(!cpu_online(cpu))) {
957 if (cpu == tick_do_timer_cpu)
958 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
960 * Make sure the CPU doesn't get fooled by obsolete tick
961 * deadline if it comes back online later.
963 ts->next_tick = 0;
964 return false;
967 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
968 return false;
970 if (need_resched())
971 return false;
973 if (unlikely(local_softirq_pending())) {
974 static int ratelimit;
976 if (ratelimit < 10 &&
977 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
978 pr_warn("NOHZ tick-stop error: Non-RCU local softirq work is pending, handler #%02x!!!\n",
979 (unsigned int) local_softirq_pending());
980 ratelimit++;
982 return false;
985 if (tick_nohz_full_enabled()) {
987 * Keep the tick alive to guarantee timekeeping progression
988 * if there are full dynticks CPUs around
990 if (tick_do_timer_cpu == cpu)
991 return false;
993 /* Should not happen for nohz-full */
994 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
995 return false;
998 return true;
1001 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1003 ktime_t expires;
1004 int cpu = smp_processor_id();
1007 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1008 * tick timer expiration time is known already.
1010 if (ts->timer_expires_base)
1011 expires = ts->timer_expires;
1012 else if (can_stop_idle_tick(cpu, ts))
1013 expires = tick_nohz_next_event(ts, cpu);
1014 else
1015 return;
1017 ts->idle_calls++;
1019 if (expires > 0LL) {
1020 int was_stopped = ts->tick_stopped;
1022 tick_nohz_stop_tick(ts, cpu);
1024 ts->idle_sleeps++;
1025 ts->idle_expires = expires;
1027 if (!was_stopped && ts->tick_stopped) {
1028 ts->idle_jiffies = ts->last_jiffies;
1029 nohz_balance_enter_idle(cpu);
1031 } else {
1032 tick_nohz_retain_tick(ts);
1037 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1039 * When the next event is more than a tick into the future, stop the idle tick
1041 void tick_nohz_idle_stop_tick(void)
1043 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1046 void tick_nohz_idle_retain_tick(void)
1048 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1050 * Undo the effect of get_next_timer_interrupt() called from
1051 * tick_nohz_next_event().
1053 timer_clear_idle();
1057 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1059 * Called when we start the idle loop.
1061 void tick_nohz_idle_enter(void)
1063 struct tick_sched *ts;
1065 lockdep_assert_irqs_enabled();
1067 local_irq_disable();
1069 ts = this_cpu_ptr(&tick_cpu_sched);
1071 WARN_ON_ONCE(ts->timer_expires_base);
1073 ts->inidle = 1;
1074 tick_nohz_start_idle(ts);
1076 local_irq_enable();
1080 * tick_nohz_irq_exit - update next tick event from interrupt exit
1082 * When an interrupt fires while we are idle and it doesn't cause
1083 * a reschedule, it may still add, modify or delete a timer, enqueue
1084 * an RCU callback, etc...
1085 * So we need to re-calculate and reprogram the next tick event.
1087 void tick_nohz_irq_exit(void)
1089 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1091 if (ts->inidle)
1092 tick_nohz_start_idle(ts);
1093 else
1094 tick_nohz_full_update_tick(ts);
1098 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1100 bool tick_nohz_idle_got_tick(void)
1102 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1104 if (ts->got_idle_tick) {
1105 ts->got_idle_tick = 0;
1106 return true;
1108 return false;
1112 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1113 * or the tick, whatever that expires first. Note that, if the tick has been
1114 * stopped, it returns the next hrtimer.
1116 * Called from power state control code with interrupts disabled
1118 ktime_t tick_nohz_get_next_hrtimer(void)
1120 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1124 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1125 * @delta_next: duration until the next event if the tick cannot be stopped
1127 * Called from power state control code with interrupts disabled
1129 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1131 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1132 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1133 int cpu = smp_processor_id();
1135 * The idle entry time is expected to be a sufficient approximation of
1136 * the current time at this point.
1138 ktime_t now = ts->idle_entrytime;
1139 ktime_t next_event;
1141 WARN_ON_ONCE(!ts->inidle);
1143 *delta_next = ktime_sub(dev->next_event, now);
1145 if (!can_stop_idle_tick(cpu, ts))
1146 return *delta_next;
1148 next_event = tick_nohz_next_event(ts, cpu);
1149 if (!next_event)
1150 return *delta_next;
1153 * If the next highres timer to expire is earlier than next_event, the
1154 * idle governor needs to know that.
1156 next_event = min_t(u64, next_event,
1157 hrtimer_next_event_without(&ts->sched_timer));
1159 return ktime_sub(next_event, now);
1163 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1164 * for a particular CPU.
1166 * Called from the schedutil frequency scaling governor in scheduler context.
1168 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1170 struct tick_sched *ts = tick_get_tick_sched(cpu);
1172 return ts->idle_calls;
1176 * tick_nohz_get_idle_calls - return the current idle calls counter value
1178 * Called from the schedutil frequency scaling governor in scheduler context.
1180 unsigned long tick_nohz_get_idle_calls(void)
1182 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1184 return ts->idle_calls;
1187 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1189 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1190 unsigned long ticks;
1192 if (vtime_accounting_enabled_this_cpu())
1193 return;
1195 * We stopped the tick in idle. Update process times would miss the
1196 * time we slept as update_process_times does only a 1 tick
1197 * accounting. Enforce that this is accounted to idle !
1199 ticks = jiffies - ts->idle_jiffies;
1201 * We might be one off. Do not randomly account a huge number of ticks!
1203 if (ticks && ticks < LONG_MAX)
1204 account_idle_ticks(ticks);
1205 #endif
1208 static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1210 tick_nohz_restart_sched_tick(ts, now);
1211 tick_nohz_account_idle_ticks(ts);
1214 void tick_nohz_idle_restart_tick(void)
1216 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1218 if (ts->tick_stopped)
1219 __tick_nohz_idle_restart_tick(ts, ktime_get());
1223 * tick_nohz_idle_exit - restart the idle tick from the idle task
1225 * Restart the idle tick when the CPU is woken up from idle
1226 * This also exit the RCU extended quiescent state. The CPU
1227 * can use RCU again after this function is called.
1229 void tick_nohz_idle_exit(void)
1231 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1232 bool idle_active, tick_stopped;
1233 ktime_t now;
1235 local_irq_disable();
1237 WARN_ON_ONCE(!ts->inidle);
1238 WARN_ON_ONCE(ts->timer_expires_base);
1240 ts->inidle = 0;
1241 idle_active = ts->idle_active;
1242 tick_stopped = ts->tick_stopped;
1244 if (idle_active || tick_stopped)
1245 now = ktime_get();
1247 if (idle_active)
1248 tick_nohz_stop_idle(ts, now);
1250 if (tick_stopped)
1251 __tick_nohz_idle_restart_tick(ts, now);
1253 local_irq_enable();
1257 * The nohz low res interrupt handler
1259 static void tick_nohz_handler(struct clock_event_device *dev)
1261 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1262 struct pt_regs *regs = get_irq_regs();
1263 ktime_t now = ktime_get();
1265 dev->next_event = KTIME_MAX;
1267 tick_sched_do_timer(ts, now);
1268 tick_sched_handle(ts, regs);
1270 /* No need to reprogram if we are running tickless */
1271 if (unlikely(ts->tick_stopped))
1272 return;
1274 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1275 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1278 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1280 if (!tick_nohz_enabled)
1281 return;
1282 ts->nohz_mode = mode;
1283 /* One update is enough */
1284 if (!test_and_set_bit(0, &tick_nohz_active))
1285 timers_update_nohz();
1289 * tick_nohz_switch_to_nohz - switch to nohz mode
1291 static void tick_nohz_switch_to_nohz(void)
1293 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1294 ktime_t next;
1296 if (!tick_nohz_enabled)
1297 return;
1299 if (tick_switch_to_oneshot(tick_nohz_handler))
1300 return;
1303 * Recycle the hrtimer in ts, so we can share the
1304 * hrtimer_forward with the highres code.
1306 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1307 /* Get the next period */
1308 next = tick_init_jiffy_update();
1310 hrtimer_set_expires(&ts->sched_timer, next);
1311 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1312 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1313 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1316 static inline void tick_nohz_irq_enter(void)
1318 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1319 ktime_t now;
1321 if (!ts->idle_active && !ts->tick_stopped)
1322 return;
1323 now = ktime_get();
1324 if (ts->idle_active)
1325 tick_nohz_stop_idle(ts, now);
1326 if (ts->tick_stopped)
1327 tick_nohz_update_jiffies(now);
1330 #else
1332 static inline void tick_nohz_switch_to_nohz(void) { }
1333 static inline void tick_nohz_irq_enter(void) { }
1334 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1336 #endif /* CONFIG_NO_HZ_COMMON */
1339 * Called from irq_enter to notify about the possible interruption of idle()
1341 void tick_irq_enter(void)
1343 tick_check_oneshot_broadcast_this_cpu();
1344 tick_nohz_irq_enter();
1348 * High resolution timer specific code
1350 #ifdef CONFIG_HIGH_RES_TIMERS
1352 * We rearm the timer until we get disabled by the idle code.
1353 * Called with interrupts disabled.
1355 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1357 struct tick_sched *ts =
1358 container_of(timer, struct tick_sched, sched_timer);
1359 struct pt_regs *regs = get_irq_regs();
1360 ktime_t now = ktime_get();
1362 tick_sched_do_timer(ts, now);
1365 * Do not call, when we are not in irq context and have
1366 * no valid regs pointer
1368 if (regs)
1369 tick_sched_handle(ts, regs);
1370 else
1371 ts->next_tick = 0;
1373 /* No need to reprogram if we are in idle or full dynticks mode */
1374 if (unlikely(ts->tick_stopped))
1375 return HRTIMER_NORESTART;
1377 hrtimer_forward(timer, now, TICK_NSEC);
1379 return HRTIMER_RESTART;
1382 static int sched_skew_tick;
1384 static int __init skew_tick(char *str)
1386 get_option(&str, &sched_skew_tick);
1388 return 0;
1390 early_param("skew_tick", skew_tick);
1393 * tick_setup_sched_timer - setup the tick emulation timer
1395 void tick_setup_sched_timer(void)
1397 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1398 ktime_t now = ktime_get();
1401 * Emulate tick processing via per-CPU hrtimers:
1403 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1404 ts->sched_timer.function = tick_sched_timer;
1406 /* Get the next period (per-CPU) */
1407 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1409 /* Offset the tick to avert jiffies_lock contention. */
1410 if (sched_skew_tick) {
1411 u64 offset = TICK_NSEC >> 1;
1412 do_div(offset, num_possible_cpus());
1413 offset *= smp_processor_id();
1414 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1417 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1418 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1419 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1421 #endif /* HIGH_RES_TIMERS */
1423 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1424 void tick_cancel_sched_timer(int cpu)
1426 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1428 # ifdef CONFIG_HIGH_RES_TIMERS
1429 if (ts->sched_timer.base)
1430 hrtimer_cancel(&ts->sched_timer);
1431 # endif
1433 memset(ts, 0, sizeof(*ts));
1435 #endif
1438 * Async notification about clocksource changes
1440 void tick_clock_notify(void)
1442 int cpu;
1444 for_each_possible_cpu(cpu)
1445 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1449 * Async notification about clock event changes
1451 void tick_oneshot_notify(void)
1453 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1455 set_bit(0, &ts->check_clocks);
1459 * Check, if a change happened, which makes oneshot possible.
1461 * Called cyclic from the hrtimer softirq (driven by the timer
1462 * softirq) allow_nohz signals, that we can switch into low-res nohz
1463 * mode, because high resolution timers are disabled (either compile
1464 * or runtime). Called with interrupts disabled.
1466 int tick_check_oneshot_change(int allow_nohz)
1468 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1470 if (!test_and_clear_bit(0, &ts->check_clocks))
1471 return 0;
1473 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1474 return 0;
1476 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1477 return 0;
1479 if (!allow_nohz)
1480 return 1;
1482 tick_nohz_switch_to_nohz();
1483 return 0;