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/module.h>
24 #include <linux/irq_work.h>
25 #include <linux/posix-timers.h>
26 #include <linux/context_tracking.h>
29 #include <asm/irq_regs.h>
31 #include "tick-internal.h"
33 #include <trace/events/timer.h>
36 * Per-CPU nohz control structure
38 static DEFINE_PER_CPU(struct tick_sched
, tick_cpu_sched
);
40 struct tick_sched
*tick_get_tick_sched(int cpu
)
42 return &per_cpu(tick_cpu_sched
, cpu
);
45 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47 * The time, when the last jiffy update happened. Protected by jiffies_lock.
49 static ktime_t last_jiffies_update
;
52 * Must be called with interrupts disabled !
54 static void tick_do_update_jiffies64(ktime_t now
)
56 unsigned long ticks
= 0;
60 * Do a quick check without holding jiffies_lock:
62 delta
= ktime_sub(now
, last_jiffies_update
);
63 if (delta
< tick_period
)
66 /* Reevaluate with jiffies_lock held */
67 write_seqlock(&jiffies_lock
);
69 delta
= ktime_sub(now
, last_jiffies_update
);
70 if (delta
>= tick_period
) {
72 delta
= ktime_sub(delta
, tick_period
);
73 last_jiffies_update
= ktime_add(last_jiffies_update
,
76 /* Slow path for long timeouts */
77 if (unlikely(delta
>= tick_period
)) {
78 s64 incr
= ktime_to_ns(tick_period
);
80 ticks
= ktime_divns(delta
, incr
);
82 last_jiffies_update
= ktime_add_ns(last_jiffies_update
,
87 /* Keep the tick_next_period variable up to date */
88 tick_next_period
= ktime_add(last_jiffies_update
, tick_period
);
90 write_sequnlock(&jiffies_lock
);
93 write_sequnlock(&jiffies_lock
);
98 * Initialize and return retrieve the jiffies update.
100 static ktime_t
tick_init_jiffy_update(void)
104 write_seqlock(&jiffies_lock
);
105 /* Did we start the jiffies update yet ? */
106 if (last_jiffies_update
== 0)
107 last_jiffies_update
= tick_next_period
;
108 period
= last_jiffies_update
;
109 write_sequnlock(&jiffies_lock
);
113 static void tick_sched_do_timer(struct tick_sched
*ts
, ktime_t now
)
115 int cpu
= smp_processor_id();
117 #ifdef CONFIG_NO_HZ_COMMON
119 * Check if the do_timer duty was dropped. We don't care about
120 * concurrency: This happens only when the CPU in charge went
121 * into a long sleep. If two CPUs happen to assign themselves to
122 * this duty, then the jiffies update is still serialized by
125 * If nohz_full is enabled, this should not happen because the
126 * tick_do_timer_cpu never relinquishes.
128 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)) {
129 #ifdef CONFIG_NO_HZ_FULL
130 WARN_ON(tick_nohz_full_running
);
132 tick_do_timer_cpu
= cpu
;
136 /* Check, if the jiffies need an update */
137 if (tick_do_timer_cpu
== cpu
)
138 tick_do_update_jiffies64(now
);
141 ts
->got_idle_tick
= 1;
144 static void tick_sched_handle(struct tick_sched
*ts
, struct pt_regs
*regs
)
146 #ifdef CONFIG_NO_HZ_COMMON
148 * When we are idle and the tick is stopped, we have to touch
149 * the watchdog as we might not schedule for a really long
150 * time. This happens on complete idle SMP systems while
151 * waiting on the login prompt. We also increment the "start of
152 * idle" jiffy stamp so the idle accounting adjustment we do
153 * when we go busy again does not account too much ticks.
155 if (ts
->tick_stopped
) {
156 touch_softlockup_watchdog_sched();
157 if (is_idle_task(current
))
160 * In case the current tick fired too early past its expected
161 * expiration, make sure we don't bypass the next clock reprogramming
162 * to the same deadline.
167 update_process_times(user_mode(regs
));
168 profile_tick(CPU_PROFILING
);
172 #ifdef CONFIG_NO_HZ_FULL
173 cpumask_var_t tick_nohz_full_mask
;
174 bool tick_nohz_full_running
;
175 EXPORT_SYMBOL_GPL(tick_nohz_full_running
);
176 static atomic_t tick_dep_mask
;
178 static bool check_tick_dependency(atomic_t
*dep
)
180 int val
= atomic_read(dep
);
182 if (val
& TICK_DEP_MASK_POSIX_TIMER
) {
183 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER
);
187 if (val
& TICK_DEP_MASK_PERF_EVENTS
) {
188 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS
);
192 if (val
& TICK_DEP_MASK_SCHED
) {
193 trace_tick_stop(0, TICK_DEP_MASK_SCHED
);
197 if (val
& TICK_DEP_MASK_CLOCK_UNSTABLE
) {
198 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE
);
202 if (val
& TICK_DEP_MASK_RCU
) {
203 trace_tick_stop(0, TICK_DEP_MASK_RCU
);
210 static bool can_stop_full_tick(int cpu
, struct tick_sched
*ts
)
212 lockdep_assert_irqs_disabled();
214 if (unlikely(!cpu_online(cpu
)))
217 if (check_tick_dependency(&tick_dep_mask
))
220 if (check_tick_dependency(&ts
->tick_dep_mask
))
223 if (check_tick_dependency(¤t
->tick_dep_mask
))
226 if (check_tick_dependency(¤t
->signal
->tick_dep_mask
))
232 static void nohz_full_kick_func(struct irq_work
*work
)
234 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
237 static DEFINE_PER_CPU(struct irq_work
, nohz_full_kick_work
) = {
238 .func
= nohz_full_kick_func
,
242 * Kick this CPU if it's full dynticks in order to force it to
243 * re-evaluate its dependency on the tick and restart it if necessary.
244 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
247 static void tick_nohz_full_kick(void)
249 if (!tick_nohz_full_cpu(smp_processor_id()))
252 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work
));
256 * Kick the CPU if it's full dynticks in order to force it to
257 * re-evaluate its dependency on the tick and restart it if necessary.
259 void tick_nohz_full_kick_cpu(int cpu
)
261 if (!tick_nohz_full_cpu(cpu
))
264 irq_work_queue_on(&per_cpu(nohz_full_kick_work
, cpu
), cpu
);
268 * Kick all full dynticks CPUs in order to force these to re-evaluate
269 * their dependency on the tick and restart it if necessary.
271 static void tick_nohz_full_kick_all(void)
275 if (!tick_nohz_full_running
)
279 for_each_cpu_and(cpu
, tick_nohz_full_mask
, cpu_online_mask
)
280 tick_nohz_full_kick_cpu(cpu
);
284 static void tick_nohz_dep_set_all(atomic_t
*dep
,
285 enum tick_dep_bits bit
)
289 prev
= atomic_fetch_or(BIT(bit
), dep
);
291 tick_nohz_full_kick_all();
295 * Set a global tick dependency. Used by perf events that rely on freq and
298 void tick_nohz_dep_set(enum tick_dep_bits bit
)
300 tick_nohz_dep_set_all(&tick_dep_mask
, bit
);
303 void tick_nohz_dep_clear(enum tick_dep_bits bit
)
305 atomic_andnot(BIT(bit
), &tick_dep_mask
);
309 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
310 * manage events throttling.
312 void tick_nohz_dep_set_cpu(int cpu
, enum tick_dep_bits bit
)
315 struct tick_sched
*ts
;
317 ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
319 prev
= atomic_fetch_or(BIT(bit
), &ts
->tick_dep_mask
);
322 /* Perf needs local kick that is NMI safe */
323 if (cpu
== smp_processor_id()) {
324 tick_nohz_full_kick();
326 /* Remote irq work not NMI-safe */
327 if (!WARN_ON_ONCE(in_nmi()))
328 tick_nohz_full_kick_cpu(cpu
);
333 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu
);
335 void tick_nohz_dep_clear_cpu(int cpu
, enum tick_dep_bits bit
)
337 struct tick_sched
*ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
339 atomic_andnot(BIT(bit
), &ts
->tick_dep_mask
);
341 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu
);
344 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
347 void tick_nohz_dep_set_task(struct task_struct
*tsk
, enum tick_dep_bits bit
)
350 * We could optimize this with just kicking the target running the task
351 * if that noise matters for nohz full users.
353 tick_nohz_dep_set_all(&tsk
->tick_dep_mask
, bit
);
355 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task
);
357 void tick_nohz_dep_clear_task(struct task_struct
*tsk
, enum tick_dep_bits bit
)
359 atomic_andnot(BIT(bit
), &tsk
->tick_dep_mask
);
361 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task
);
364 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
365 * per process timers.
367 void tick_nohz_dep_set_signal(struct signal_struct
*sig
, enum tick_dep_bits bit
)
369 tick_nohz_dep_set_all(&sig
->tick_dep_mask
, bit
);
372 void tick_nohz_dep_clear_signal(struct signal_struct
*sig
, enum tick_dep_bits bit
)
374 atomic_andnot(BIT(bit
), &sig
->tick_dep_mask
);
378 * Re-evaluate the need for the tick as we switch the current task.
379 * It might need the tick due to per task/process properties:
380 * perf events, posix CPU timers, ...
382 void __tick_nohz_task_switch(void)
385 struct tick_sched
*ts
;
387 local_irq_save(flags
);
389 if (!tick_nohz_full_cpu(smp_processor_id()))
392 ts
= this_cpu_ptr(&tick_cpu_sched
);
394 if (ts
->tick_stopped
) {
395 if (atomic_read(¤t
->tick_dep_mask
) ||
396 atomic_read(¤t
->signal
->tick_dep_mask
))
397 tick_nohz_full_kick();
400 local_irq_restore(flags
);
403 /* Get the boot-time nohz CPU list from the kernel parameters. */
404 void __init
tick_nohz_full_setup(cpumask_var_t cpumask
)
406 alloc_bootmem_cpumask_var(&tick_nohz_full_mask
);
407 cpumask_copy(tick_nohz_full_mask
, cpumask
);
408 tick_nohz_full_running
= true;
410 EXPORT_SYMBOL_GPL(tick_nohz_full_setup
);
412 static int tick_nohz_cpu_down(unsigned int cpu
)
415 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
416 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
417 * CPUs. It must remain online when nohz full is enabled.
419 if (tick_nohz_full_running
&& tick_do_timer_cpu
== cpu
)
424 void __init
tick_nohz_init(void)
428 if (!tick_nohz_full_running
)
432 * Full dynticks uses irq work to drive the tick rescheduling on safe
433 * locking contexts. But then we need irq work to raise its own
434 * interrupts to avoid circular dependency on the tick
436 if (!arch_irq_work_has_interrupt()) {
437 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
438 cpumask_clear(tick_nohz_full_mask
);
439 tick_nohz_full_running
= false;
443 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP
) &&
444 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU
)) {
445 cpu
= smp_processor_id();
447 if (cpumask_test_cpu(cpu
, tick_nohz_full_mask
)) {
448 pr_warn("NO_HZ: Clearing %d from nohz_full range "
449 "for timekeeping\n", cpu
);
450 cpumask_clear_cpu(cpu
, tick_nohz_full_mask
);
454 for_each_cpu(cpu
, tick_nohz_full_mask
)
455 context_tracking_cpu_set(cpu
);
457 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
458 "kernel/nohz:predown", NULL
,
461 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
462 cpumask_pr_args(tick_nohz_full_mask
));
467 * NOHZ - aka dynamic tick functionality
469 #ifdef CONFIG_NO_HZ_COMMON
473 bool tick_nohz_enabled __read_mostly
= true;
474 unsigned long tick_nohz_active __read_mostly
;
476 * Enable / Disable tickless mode
478 static int __init
setup_tick_nohz(char *str
)
480 return (kstrtobool(str
, &tick_nohz_enabled
) == 0);
483 __setup("nohz=", setup_tick_nohz
);
485 bool tick_nohz_tick_stopped(void)
487 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
489 return ts
->tick_stopped
;
492 bool tick_nohz_tick_stopped_cpu(int cpu
)
494 struct tick_sched
*ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
496 return ts
->tick_stopped
;
500 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
502 * Called from interrupt entry when the CPU was idle
504 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
505 * must be updated. Otherwise an interrupt handler could use a stale jiffy
506 * value. We do this unconditionally on any CPU, as we don't know whether the
507 * CPU, which has the update task assigned is in a long sleep.
509 static void tick_nohz_update_jiffies(ktime_t now
)
513 __this_cpu_write(tick_cpu_sched
.idle_waketime
, now
);
515 local_irq_save(flags
);
516 tick_do_update_jiffies64(now
);
517 local_irq_restore(flags
);
519 touch_softlockup_watchdog_sched();
523 * Updates the per-CPU time idle statistics counters
526 update_ts_time_stats(int cpu
, struct tick_sched
*ts
, ktime_t now
, u64
*last_update_time
)
530 if (ts
->idle_active
) {
531 delta
= ktime_sub(now
, ts
->idle_entrytime
);
532 if (nr_iowait_cpu(cpu
) > 0)
533 ts
->iowait_sleeptime
= ktime_add(ts
->iowait_sleeptime
, delta
);
535 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
536 ts
->idle_entrytime
= now
;
539 if (last_update_time
)
540 *last_update_time
= ktime_to_us(now
);
544 static void tick_nohz_stop_idle(struct tick_sched
*ts
, ktime_t now
)
546 update_ts_time_stats(smp_processor_id(), ts
, now
, NULL
);
549 sched_clock_idle_wakeup_event();
552 static void tick_nohz_start_idle(struct tick_sched
*ts
)
554 ts
->idle_entrytime
= ktime_get();
556 sched_clock_idle_sleep_event();
560 * get_cpu_idle_time_us - get the total idle time of a CPU
561 * @cpu: CPU number to query
562 * @last_update_time: variable to store update time in. Do not update
565 * Return the cumulative idle time (since boot) for a given
566 * CPU, in microseconds.
568 * This time is measured via accounting rather than sampling,
569 * and is as accurate as ktime_get() is.
571 * This function returns -1 if NOHZ is not enabled.
573 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
575 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
578 if (!tick_nohz_active
)
582 if (last_update_time
) {
583 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
584 idle
= ts
->idle_sleeptime
;
586 if (ts
->idle_active
&& !nr_iowait_cpu(cpu
)) {
587 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
589 idle
= ktime_add(ts
->idle_sleeptime
, delta
);
591 idle
= ts
->idle_sleeptime
;
595 return ktime_to_us(idle
);
598 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
601 * get_cpu_iowait_time_us - get the total iowait time of a CPU
602 * @cpu: CPU number to query
603 * @last_update_time: variable to store update time in. Do not update
606 * Return the cumulative iowait time (since boot) for a given
607 * CPU, in microseconds.
609 * This time is measured via accounting rather than sampling,
610 * and is as accurate as ktime_get() is.
612 * This function returns -1 if NOHZ is not enabled.
614 u64
get_cpu_iowait_time_us(int cpu
, u64
*last_update_time
)
616 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
619 if (!tick_nohz_active
)
623 if (last_update_time
) {
624 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
625 iowait
= ts
->iowait_sleeptime
;
627 if (ts
->idle_active
&& nr_iowait_cpu(cpu
) > 0) {
628 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
630 iowait
= ktime_add(ts
->iowait_sleeptime
, delta
);
632 iowait
= ts
->iowait_sleeptime
;
636 return ktime_to_us(iowait
);
638 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us
);
640 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
642 hrtimer_cancel(&ts
->sched_timer
);
643 hrtimer_set_expires(&ts
->sched_timer
, ts
->last_tick
);
645 /* Forward the time to expire in the future */
646 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
648 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
649 hrtimer_start_expires(&ts
->sched_timer
,
650 HRTIMER_MODE_ABS_PINNED_HARD
);
652 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
656 * Reset to make sure next tick stop doesn't get fooled by past
657 * cached clock deadline.
662 static inline bool local_timer_softirq_pending(void)
664 return local_softirq_pending() & BIT(TIMER_SOFTIRQ
);
667 static ktime_t
tick_nohz_next_event(struct tick_sched
*ts
, int cpu
)
669 u64 basemono
, next_tick
, next_tmr
, next_rcu
, delta
, expires
;
670 unsigned long basejiff
;
673 /* Read jiffies and the time when jiffies were updated last */
675 seq
= read_seqbegin(&jiffies_lock
);
676 basemono
= last_jiffies_update
;
678 } while (read_seqretry(&jiffies_lock
, seq
));
679 ts
->last_jiffies
= basejiff
;
680 ts
->timer_expires_base
= basemono
;
683 * Keep the periodic tick, when RCU, architecture or irq_work
685 * Aside of that check whether the local timer softirq is
686 * pending. If so its a bad idea to call get_next_timer_interrupt()
687 * because there is an already expired timer, so it will request
688 * immeditate expiry, which rearms the hardware timer with a
689 * minimal delta which brings us back to this place
690 * immediately. Lather, rinse and repeat...
692 if (rcu_needs_cpu(basemono
, &next_rcu
) || arch_needs_cpu() ||
693 irq_work_needs_cpu() || local_timer_softirq_pending()) {
694 next_tick
= basemono
+ TICK_NSEC
;
697 * Get the next pending timer. If high resolution
698 * timers are enabled this only takes the timer wheel
699 * timers into account. If high resolution timers are
700 * disabled this also looks at the next expiring
703 next_tmr
= get_next_timer_interrupt(basejiff
, basemono
);
704 ts
->next_timer
= next_tmr
;
705 /* Take the next rcu event into account */
706 next_tick
= next_rcu
< next_tmr
? next_rcu
: next_tmr
;
710 * If the tick is due in the next period, keep it ticking or
711 * force prod the timer.
713 delta
= next_tick
- basemono
;
714 if (delta
<= (u64
)TICK_NSEC
) {
716 * Tell the timer code that the base is not idle, i.e. undo
717 * the effect of get_next_timer_interrupt():
721 * We've not stopped the tick yet, and there's a timer in the
722 * next period, so no point in stopping it either, bail.
724 if (!ts
->tick_stopped
) {
725 ts
->timer_expires
= 0;
731 * If this CPU is the one which had the do_timer() duty last, we limit
732 * the sleep time to the timekeeping max_deferment value.
733 * Otherwise we can sleep as long as we want.
735 delta
= timekeeping_max_deferment();
736 if (cpu
!= tick_do_timer_cpu
&&
737 (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
|| !ts
->do_timer_last
))
740 /* Calculate the next expiry time */
741 if (delta
< (KTIME_MAX
- basemono
))
742 expires
= basemono
+ delta
;
746 ts
->timer_expires
= min_t(u64
, expires
, next_tick
);
749 return ts
->timer_expires
;
752 static void tick_nohz_stop_tick(struct tick_sched
*ts
, int cpu
)
754 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
755 u64 basemono
= ts
->timer_expires_base
;
756 u64 expires
= ts
->timer_expires
;
757 ktime_t tick
= expires
;
759 /* Make sure we won't be trying to stop it twice in a row. */
760 ts
->timer_expires_base
= 0;
763 * If this CPU is the one which updates jiffies, then give up
764 * the assignment and let it be taken by the CPU which runs
765 * the tick timer next, which might be this CPU as well. If we
766 * don't drop this here the jiffies might be stale and
767 * do_timer() never invoked. Keep track of the fact that it
768 * was the one which had the do_timer() duty last.
770 if (cpu
== tick_do_timer_cpu
) {
771 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
772 ts
->do_timer_last
= 1;
773 } else if (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
) {
774 ts
->do_timer_last
= 0;
777 /* Skip reprogram of event if its not changed */
778 if (ts
->tick_stopped
&& (expires
== ts
->next_tick
)) {
779 /* Sanity check: make sure clockevent is actually programmed */
780 if (tick
== KTIME_MAX
|| ts
->next_tick
== hrtimer_get_expires(&ts
->sched_timer
))
784 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
785 basemono
, ts
->next_tick
, dev
->next_event
,
786 hrtimer_active(&ts
->sched_timer
), hrtimer_get_expires(&ts
->sched_timer
));
790 * nohz_stop_sched_tick can be called several times before
791 * the nohz_restart_sched_tick is called. This happens when
792 * interrupts arrive which do not cause a reschedule. In the
793 * first call we save the current tick time, so we can restart
794 * the scheduler tick in nohz_restart_sched_tick.
796 if (!ts
->tick_stopped
) {
797 calc_load_nohz_start();
800 ts
->last_tick
= hrtimer_get_expires(&ts
->sched_timer
);
801 ts
->tick_stopped
= 1;
802 trace_tick_stop(1, TICK_DEP_MASK_NONE
);
805 ts
->next_tick
= tick
;
808 * If the expiration time == KTIME_MAX, then we simply stop
811 if (unlikely(expires
== KTIME_MAX
)) {
812 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
813 hrtimer_cancel(&ts
->sched_timer
);
817 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
818 hrtimer_start(&ts
->sched_timer
, tick
,
819 HRTIMER_MODE_ABS_PINNED_HARD
);
821 hrtimer_set_expires(&ts
->sched_timer
, tick
);
822 tick_program_event(tick
, 1);
826 static void tick_nohz_retain_tick(struct tick_sched
*ts
)
828 ts
->timer_expires_base
= 0;
831 #ifdef CONFIG_NO_HZ_FULL
832 static void tick_nohz_stop_sched_tick(struct tick_sched
*ts
, int cpu
)
834 if (tick_nohz_next_event(ts
, cpu
))
835 tick_nohz_stop_tick(ts
, cpu
);
837 tick_nohz_retain_tick(ts
);
839 #endif /* CONFIG_NO_HZ_FULL */
841 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
)
843 /* Update jiffies first */
844 tick_do_update_jiffies64(now
);
847 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
848 * the clock forward checks in the enqueue path:
852 calc_load_nohz_stop();
853 touch_softlockup_watchdog_sched();
855 * Cancel the scheduled timer and restore the tick
857 ts
->tick_stopped
= 0;
858 ts
->idle_exittime
= now
;
860 tick_nohz_restart(ts
, now
);
863 static void tick_nohz_full_update_tick(struct tick_sched
*ts
)
865 #ifdef CONFIG_NO_HZ_FULL
866 int cpu
= smp_processor_id();
868 if (!tick_nohz_full_cpu(cpu
))
871 if (!ts
->tick_stopped
&& ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)
874 if (can_stop_full_tick(cpu
, ts
))
875 tick_nohz_stop_sched_tick(ts
, cpu
);
876 else if (ts
->tick_stopped
)
877 tick_nohz_restart_sched_tick(ts
, ktime_get());
881 static bool can_stop_idle_tick(int cpu
, struct tick_sched
*ts
)
884 * If this CPU is offline and it is the one which updates
885 * jiffies, then give up the assignment and let it be taken by
886 * the CPU which runs the tick timer next. If we don't drop
887 * this here the jiffies might be stale and do_timer() never
890 if (unlikely(!cpu_online(cpu
))) {
891 if (cpu
== tick_do_timer_cpu
)
892 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
894 * Make sure the CPU doesn't get fooled by obsolete tick
895 * deadline if it comes back online later.
901 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
))
907 if (unlikely(local_softirq_pending())) {
908 static int ratelimit
;
910 if (ratelimit
< 10 &&
911 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK
)) {
912 pr_warn("NOHZ: local_softirq_pending %02x\n",
913 (unsigned int) local_softirq_pending());
919 if (tick_nohz_full_enabled()) {
921 * Keep the tick alive to guarantee timekeeping progression
922 * if there are full dynticks CPUs around
924 if (tick_do_timer_cpu
== cpu
)
927 * Boot safety: make sure the timekeeping duty has been
928 * assigned before entering dyntick-idle mode,
929 * tick_do_timer_cpu is TICK_DO_TIMER_BOOT
931 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_BOOT
))
934 /* Should not happen for nohz-full */
935 if (WARN_ON_ONCE(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
942 static void __tick_nohz_idle_stop_tick(struct tick_sched
*ts
)
945 int cpu
= smp_processor_id();
948 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
949 * tick timer expiration time is known already.
951 if (ts
->timer_expires_base
)
952 expires
= ts
->timer_expires
;
953 else if (can_stop_idle_tick(cpu
, ts
))
954 expires
= tick_nohz_next_event(ts
, cpu
);
961 int was_stopped
= ts
->tick_stopped
;
963 tick_nohz_stop_tick(ts
, cpu
);
966 ts
->idle_expires
= expires
;
968 if (!was_stopped
&& ts
->tick_stopped
) {
969 ts
->idle_jiffies
= ts
->last_jiffies
;
970 nohz_balance_enter_idle(cpu
);
973 tick_nohz_retain_tick(ts
);
978 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
980 * When the next event is more than a tick into the future, stop the idle tick
982 void tick_nohz_idle_stop_tick(void)
984 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched
));
987 void tick_nohz_idle_retain_tick(void)
989 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched
));
991 * Undo the effect of get_next_timer_interrupt() called from
992 * tick_nohz_next_event().
998 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1000 * Called when we start the idle loop.
1002 void tick_nohz_idle_enter(void)
1004 struct tick_sched
*ts
;
1006 lockdep_assert_irqs_enabled();
1008 local_irq_disable();
1010 ts
= this_cpu_ptr(&tick_cpu_sched
);
1012 WARN_ON_ONCE(ts
->timer_expires_base
);
1015 tick_nohz_start_idle(ts
);
1021 * tick_nohz_irq_exit - update next tick event from interrupt exit
1023 * When an interrupt fires while we are idle and it doesn't cause
1024 * a reschedule, it may still add, modify or delete a timer, enqueue
1025 * an RCU callback, etc...
1026 * So we need to re-calculate and reprogram the next tick event.
1028 void tick_nohz_irq_exit(void)
1030 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1033 tick_nohz_start_idle(ts
);
1035 tick_nohz_full_update_tick(ts
);
1039 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1041 bool tick_nohz_idle_got_tick(void)
1043 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1045 if (ts
->got_idle_tick
) {
1046 ts
->got_idle_tick
= 0;
1053 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1054 * or the tick, whatever that expires first. Note that, if the tick has been
1055 * stopped, it returns the next hrtimer.
1057 * Called from power state control code with interrupts disabled
1059 ktime_t
tick_nohz_get_next_hrtimer(void)
1061 return __this_cpu_read(tick_cpu_device
.evtdev
)->next_event
;
1065 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1066 * @delta_next: duration until the next event if the tick cannot be stopped
1068 * Called from power state control code with interrupts disabled
1070 ktime_t
tick_nohz_get_sleep_length(ktime_t
*delta_next
)
1072 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
1073 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1074 int cpu
= smp_processor_id();
1076 * The idle entry time is expected to be a sufficient approximation of
1077 * the current time at this point.
1079 ktime_t now
= ts
->idle_entrytime
;
1082 WARN_ON_ONCE(!ts
->inidle
);
1084 *delta_next
= ktime_sub(dev
->next_event
, now
);
1086 if (!can_stop_idle_tick(cpu
, ts
))
1089 next_event
= tick_nohz_next_event(ts
, cpu
);
1094 * If the next highres timer to expire is earlier than next_event, the
1095 * idle governor needs to know that.
1097 next_event
= min_t(u64
, next_event
,
1098 hrtimer_next_event_without(&ts
->sched_timer
));
1100 return ktime_sub(next_event
, now
);
1104 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1105 * for a particular CPU.
1107 * Called from the schedutil frequency scaling governor in scheduler context.
1109 unsigned long tick_nohz_get_idle_calls_cpu(int cpu
)
1111 struct tick_sched
*ts
= tick_get_tick_sched(cpu
);
1113 return ts
->idle_calls
;
1117 * tick_nohz_get_idle_calls - return the current idle calls counter value
1119 * Called from the schedutil frequency scaling governor in scheduler context.
1121 unsigned long tick_nohz_get_idle_calls(void)
1123 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1125 return ts
->idle_calls
;
1128 static void tick_nohz_account_idle_ticks(struct tick_sched
*ts
)
1130 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1131 unsigned long ticks
;
1133 if (vtime_accounting_enabled_this_cpu())
1136 * We stopped the tick in idle. Update process times would miss the
1137 * time we slept as update_process_times does only a 1 tick
1138 * accounting. Enforce that this is accounted to idle !
1140 ticks
= jiffies
- ts
->idle_jiffies
;
1142 * We might be one off. Do not randomly account a huge number of ticks!
1144 if (ticks
&& ticks
< LONG_MAX
)
1145 account_idle_ticks(ticks
);
1149 static void __tick_nohz_idle_restart_tick(struct tick_sched
*ts
, ktime_t now
)
1151 tick_nohz_restart_sched_tick(ts
, now
);
1152 tick_nohz_account_idle_ticks(ts
);
1155 void tick_nohz_idle_restart_tick(void)
1157 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1159 if (ts
->tick_stopped
)
1160 __tick_nohz_idle_restart_tick(ts
, ktime_get());
1164 * tick_nohz_idle_exit - restart the idle tick from the idle task
1166 * Restart the idle tick when the CPU is woken up from idle
1167 * This also exit the RCU extended quiescent state. The CPU
1168 * can use RCU again after this function is called.
1170 void tick_nohz_idle_exit(void)
1172 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1173 bool idle_active
, tick_stopped
;
1176 local_irq_disable();
1178 WARN_ON_ONCE(!ts
->inidle
);
1179 WARN_ON_ONCE(ts
->timer_expires_base
);
1182 idle_active
= ts
->idle_active
;
1183 tick_stopped
= ts
->tick_stopped
;
1185 if (idle_active
|| tick_stopped
)
1189 tick_nohz_stop_idle(ts
, now
);
1192 __tick_nohz_idle_restart_tick(ts
, now
);
1198 * The nohz low res interrupt handler
1200 static void tick_nohz_handler(struct clock_event_device
*dev
)
1202 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1203 struct pt_regs
*regs
= get_irq_regs();
1204 ktime_t now
= ktime_get();
1206 dev
->next_event
= KTIME_MAX
;
1208 tick_sched_do_timer(ts
, now
);
1209 tick_sched_handle(ts
, regs
);
1211 /* No need to reprogram if we are running tickless */
1212 if (unlikely(ts
->tick_stopped
))
1215 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1216 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
1219 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
)
1221 if (!tick_nohz_enabled
)
1223 ts
->nohz_mode
= mode
;
1224 /* One update is enough */
1225 if (!test_and_set_bit(0, &tick_nohz_active
))
1226 timers_update_nohz();
1230 * tick_nohz_switch_to_nohz - switch to nohz mode
1232 static void tick_nohz_switch_to_nohz(void)
1234 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1237 if (!tick_nohz_enabled
)
1240 if (tick_switch_to_oneshot(tick_nohz_handler
))
1244 * Recycle the hrtimer in ts, so we can share the
1245 * hrtimer_forward with the highres code.
1247 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_HARD
);
1248 /* Get the next period */
1249 next
= tick_init_jiffy_update();
1251 hrtimer_set_expires(&ts
->sched_timer
, next
);
1252 hrtimer_forward_now(&ts
->sched_timer
, tick_period
);
1253 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
1254 tick_nohz_activate(ts
, NOHZ_MODE_LOWRES
);
1257 static inline void tick_nohz_irq_enter(void)
1259 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1262 if (!ts
->idle_active
&& !ts
->tick_stopped
)
1265 if (ts
->idle_active
)
1266 tick_nohz_stop_idle(ts
, now
);
1267 if (ts
->tick_stopped
)
1268 tick_nohz_update_jiffies(now
);
1273 static inline void tick_nohz_switch_to_nohz(void) { }
1274 static inline void tick_nohz_irq_enter(void) { }
1275 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
) { }
1277 #endif /* CONFIG_NO_HZ_COMMON */
1280 * Called from irq_enter to notify about the possible interruption of idle()
1282 void tick_irq_enter(void)
1284 tick_check_oneshot_broadcast_this_cpu();
1285 tick_nohz_irq_enter();
1289 * High resolution timer specific code
1291 #ifdef CONFIG_HIGH_RES_TIMERS
1293 * We rearm the timer until we get disabled by the idle code.
1294 * Called with interrupts disabled.
1296 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
1298 struct tick_sched
*ts
=
1299 container_of(timer
, struct tick_sched
, sched_timer
);
1300 struct pt_regs
*regs
= get_irq_regs();
1301 ktime_t now
= ktime_get();
1303 tick_sched_do_timer(ts
, now
);
1306 * Do not call, when we are not in irq context and have
1307 * no valid regs pointer
1310 tick_sched_handle(ts
, regs
);
1314 /* No need to reprogram if we are in idle or full dynticks mode */
1315 if (unlikely(ts
->tick_stopped
))
1316 return HRTIMER_NORESTART
;
1318 hrtimer_forward(timer
, now
, tick_period
);
1320 return HRTIMER_RESTART
;
1323 static int sched_skew_tick
;
1325 static int __init
skew_tick(char *str
)
1327 get_option(&str
, &sched_skew_tick
);
1331 early_param("skew_tick", skew_tick
);
1334 * tick_setup_sched_timer - setup the tick emulation timer
1336 void tick_setup_sched_timer(void)
1338 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1339 ktime_t now
= ktime_get();
1342 * Emulate tick processing via per-CPU hrtimers:
1344 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_HARD
);
1345 ts
->sched_timer
.function
= tick_sched_timer
;
1347 /* Get the next period (per-CPU) */
1348 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
1350 /* Offset the tick to avert jiffies_lock contention. */
1351 if (sched_skew_tick
) {
1352 u64 offset
= ktime_to_ns(tick_period
) >> 1;
1353 do_div(offset
, num_possible_cpus());
1354 offset
*= smp_processor_id();
1355 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
1358 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1359 hrtimer_start_expires(&ts
->sched_timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
1360 tick_nohz_activate(ts
, NOHZ_MODE_HIGHRES
);
1362 #endif /* HIGH_RES_TIMERS */
1364 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1365 void tick_cancel_sched_timer(int cpu
)
1367 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1369 # ifdef CONFIG_HIGH_RES_TIMERS
1370 if (ts
->sched_timer
.base
)
1371 hrtimer_cancel(&ts
->sched_timer
);
1374 memset(ts
, 0, sizeof(*ts
));
1379 * Async notification about clocksource changes
1381 void tick_clock_notify(void)
1385 for_each_possible_cpu(cpu
)
1386 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
1390 * Async notification about clock event changes
1392 void tick_oneshot_notify(void)
1394 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1396 set_bit(0, &ts
->check_clocks
);
1400 * Check, if a change happened, which makes oneshot possible.
1402 * Called cyclic from the hrtimer softirq (driven by the timer
1403 * softirq) allow_nohz signals, that we can switch into low-res nohz
1404 * mode, because high resolution timers are disabled (either compile
1405 * or runtime). Called with interrupts disabled.
1407 int tick_check_oneshot_change(int allow_nohz
)
1409 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1411 if (!test_and_clear_bit(0, &ts
->check_clocks
))
1414 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
1417 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1423 tick_nohz_switch_to_nohz();