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:
61 * The READ_ONCE() pairs with two updates done later in this function.
63 delta
= ktime_sub(now
, READ_ONCE(last_jiffies_update
));
64 if (delta
< tick_period
)
67 /* Reevaluate with jiffies_lock held */
68 write_seqlock(&jiffies_lock
);
70 delta
= ktime_sub(now
, last_jiffies_update
);
71 if (delta
>= tick_period
) {
73 delta
= ktime_sub(delta
, tick_period
);
74 /* Pairs with the lockless read in this function. */
75 WRITE_ONCE(last_jiffies_update
,
76 ktime_add(last_jiffies_update
, tick_period
));
78 /* Slow path for long timeouts */
79 if (unlikely(delta
>= tick_period
)) {
80 s64 incr
= ktime_to_ns(tick_period
);
82 ticks
= ktime_divns(delta
, incr
);
84 /* Pairs with the lockless read in this function. */
85 WRITE_ONCE(last_jiffies_update
,
86 ktime_add_ns(last_jiffies_update
,
91 /* Keep the tick_next_period variable up to date */
92 tick_next_period
= ktime_add(last_jiffies_update
, tick_period
);
94 write_sequnlock(&jiffies_lock
);
97 write_sequnlock(&jiffies_lock
);
102 * Initialize and return retrieve the jiffies update.
104 static ktime_t
tick_init_jiffy_update(void)
108 write_seqlock(&jiffies_lock
);
109 /* Did we start the jiffies update yet ? */
110 if (last_jiffies_update
== 0)
111 last_jiffies_update
= tick_next_period
;
112 period
= last_jiffies_update
;
113 write_sequnlock(&jiffies_lock
);
117 static void tick_sched_do_timer(struct tick_sched
*ts
, ktime_t now
)
119 int cpu
= smp_processor_id();
121 #ifdef CONFIG_NO_HZ_COMMON
123 * Check if the do_timer duty was dropped. We don't care about
124 * concurrency: This happens only when the CPU in charge went
125 * into a long sleep. If two CPUs happen to assign themselves to
126 * this duty, then the jiffies update is still serialized by
129 * If nohz_full is enabled, this should not happen because the
130 * tick_do_timer_cpu never relinquishes.
132 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)) {
133 #ifdef CONFIG_NO_HZ_FULL
134 WARN_ON(tick_nohz_full_running
);
136 tick_do_timer_cpu
= cpu
;
140 /* Check, if the jiffies need an update */
141 if (tick_do_timer_cpu
== cpu
)
142 tick_do_update_jiffies64(now
);
145 ts
->got_idle_tick
= 1;
148 static void tick_sched_handle(struct tick_sched
*ts
, struct pt_regs
*regs
)
150 #ifdef CONFIG_NO_HZ_COMMON
152 * When we are idle and the tick is stopped, we have to touch
153 * the watchdog as we might not schedule for a really long
154 * time. This happens on complete idle SMP systems while
155 * waiting on the login prompt. We also increment the "start of
156 * idle" jiffy stamp so the idle accounting adjustment we do
157 * when we go busy again does not account too much ticks.
159 if (ts
->tick_stopped
) {
160 touch_softlockup_watchdog_sched();
161 if (is_idle_task(current
))
164 * In case the current tick fired too early past its expected
165 * expiration, make sure we don't bypass the next clock reprogramming
166 * to the same deadline.
171 update_process_times(user_mode(regs
));
172 profile_tick(CPU_PROFILING
);
176 #ifdef CONFIG_NO_HZ_FULL
177 cpumask_var_t tick_nohz_full_mask
;
178 bool tick_nohz_full_running
;
179 EXPORT_SYMBOL_GPL(tick_nohz_full_running
);
180 static atomic_t tick_dep_mask
;
182 static bool check_tick_dependency(atomic_t
*dep
)
184 int val
= atomic_read(dep
);
186 if (val
& TICK_DEP_MASK_POSIX_TIMER
) {
187 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER
);
191 if (val
& TICK_DEP_MASK_PERF_EVENTS
) {
192 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS
);
196 if (val
& TICK_DEP_MASK_SCHED
) {
197 trace_tick_stop(0, TICK_DEP_MASK_SCHED
);
201 if (val
& TICK_DEP_MASK_CLOCK_UNSTABLE
) {
202 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE
);
206 if (val
& TICK_DEP_MASK_RCU
) {
207 trace_tick_stop(0, TICK_DEP_MASK_RCU
);
214 static bool can_stop_full_tick(int cpu
, struct tick_sched
*ts
)
216 lockdep_assert_irqs_disabled();
218 if (unlikely(!cpu_online(cpu
)))
221 if (check_tick_dependency(&tick_dep_mask
))
224 if (check_tick_dependency(&ts
->tick_dep_mask
))
227 if (check_tick_dependency(¤t
->tick_dep_mask
))
230 if (check_tick_dependency(¤t
->signal
->tick_dep_mask
))
236 static void nohz_full_kick_func(struct irq_work
*work
)
238 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
241 static DEFINE_PER_CPU(struct irq_work
, nohz_full_kick_work
) = {
242 .func
= nohz_full_kick_func
,
246 * Kick this CPU if it's full dynticks in order to force it to
247 * re-evaluate its dependency on the tick and restart it if necessary.
248 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
251 static void tick_nohz_full_kick(void)
253 if (!tick_nohz_full_cpu(smp_processor_id()))
256 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work
));
260 * Kick the CPU if it's full dynticks in order to force it to
261 * re-evaluate its dependency on the tick and restart it if necessary.
263 void tick_nohz_full_kick_cpu(int cpu
)
265 if (!tick_nohz_full_cpu(cpu
))
268 irq_work_queue_on(&per_cpu(nohz_full_kick_work
, cpu
), cpu
);
272 * Kick all full dynticks CPUs in order to force these to re-evaluate
273 * their dependency on the tick and restart it if necessary.
275 static void tick_nohz_full_kick_all(void)
279 if (!tick_nohz_full_running
)
283 for_each_cpu_and(cpu
, tick_nohz_full_mask
, cpu_online_mask
)
284 tick_nohz_full_kick_cpu(cpu
);
288 static void tick_nohz_dep_set_all(atomic_t
*dep
,
289 enum tick_dep_bits bit
)
293 prev
= atomic_fetch_or(BIT(bit
), dep
);
295 tick_nohz_full_kick_all();
299 * Set a global tick dependency. Used by perf events that rely on freq and
302 void tick_nohz_dep_set(enum tick_dep_bits bit
)
304 tick_nohz_dep_set_all(&tick_dep_mask
, bit
);
307 void tick_nohz_dep_clear(enum tick_dep_bits bit
)
309 atomic_andnot(BIT(bit
), &tick_dep_mask
);
313 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
314 * manage events throttling.
316 void tick_nohz_dep_set_cpu(int cpu
, enum tick_dep_bits bit
)
319 struct tick_sched
*ts
;
321 ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
323 prev
= atomic_fetch_or(BIT(bit
), &ts
->tick_dep_mask
);
326 /* Perf needs local kick that is NMI safe */
327 if (cpu
== smp_processor_id()) {
328 tick_nohz_full_kick();
330 /* Remote irq work not NMI-safe */
331 if (!WARN_ON_ONCE(in_nmi()))
332 tick_nohz_full_kick_cpu(cpu
);
337 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu
);
339 void tick_nohz_dep_clear_cpu(int cpu
, enum tick_dep_bits bit
)
341 struct tick_sched
*ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
343 atomic_andnot(BIT(bit
), &ts
->tick_dep_mask
);
345 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu
);
348 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
351 void tick_nohz_dep_set_task(struct task_struct
*tsk
, enum tick_dep_bits bit
)
354 * We could optimize this with just kicking the target running the task
355 * if that noise matters for nohz full users.
357 tick_nohz_dep_set_all(&tsk
->tick_dep_mask
, bit
);
359 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task
);
361 void tick_nohz_dep_clear_task(struct task_struct
*tsk
, enum tick_dep_bits bit
)
363 atomic_andnot(BIT(bit
), &tsk
->tick_dep_mask
);
365 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task
);
368 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
369 * per process timers.
371 void tick_nohz_dep_set_signal(struct signal_struct
*sig
, enum tick_dep_bits bit
)
373 tick_nohz_dep_set_all(&sig
->tick_dep_mask
, bit
);
376 void tick_nohz_dep_clear_signal(struct signal_struct
*sig
, enum tick_dep_bits bit
)
378 atomic_andnot(BIT(bit
), &sig
->tick_dep_mask
);
382 * Re-evaluate the need for the tick as we switch the current task.
383 * It might need the tick due to per task/process properties:
384 * perf events, posix CPU timers, ...
386 void __tick_nohz_task_switch(void)
389 struct tick_sched
*ts
;
391 local_irq_save(flags
);
393 if (!tick_nohz_full_cpu(smp_processor_id()))
396 ts
= this_cpu_ptr(&tick_cpu_sched
);
398 if (ts
->tick_stopped
) {
399 if (atomic_read(¤t
->tick_dep_mask
) ||
400 atomic_read(¤t
->signal
->tick_dep_mask
))
401 tick_nohz_full_kick();
404 local_irq_restore(flags
);
407 /* Get the boot-time nohz CPU list from the kernel parameters. */
408 void __init
tick_nohz_full_setup(cpumask_var_t cpumask
)
410 alloc_bootmem_cpumask_var(&tick_nohz_full_mask
);
411 cpumask_copy(tick_nohz_full_mask
, cpumask
);
412 tick_nohz_full_running
= true;
414 EXPORT_SYMBOL_GPL(tick_nohz_full_setup
);
416 static int tick_nohz_cpu_down(unsigned int cpu
)
419 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
420 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
421 * CPUs. It must remain online when nohz full is enabled.
423 if (tick_nohz_full_running
&& tick_do_timer_cpu
== cpu
)
428 void __init
tick_nohz_init(void)
432 if (!tick_nohz_full_running
)
436 * Full dynticks uses irq work to drive the tick rescheduling on safe
437 * locking contexts. But then we need irq work to raise its own
438 * interrupts to avoid circular dependency on the tick
440 if (!arch_irq_work_has_interrupt()) {
441 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
442 cpumask_clear(tick_nohz_full_mask
);
443 tick_nohz_full_running
= false;
447 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP
) &&
448 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU
)) {
449 cpu
= smp_processor_id();
451 if (cpumask_test_cpu(cpu
, tick_nohz_full_mask
)) {
452 pr_warn("NO_HZ: Clearing %d from nohz_full range "
453 "for timekeeping\n", cpu
);
454 cpumask_clear_cpu(cpu
, tick_nohz_full_mask
);
458 for_each_cpu(cpu
, tick_nohz_full_mask
)
459 context_tracking_cpu_set(cpu
);
461 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
462 "kernel/nohz:predown", NULL
,
465 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
466 cpumask_pr_args(tick_nohz_full_mask
));
471 * NOHZ - aka dynamic tick functionality
473 #ifdef CONFIG_NO_HZ_COMMON
477 bool tick_nohz_enabled __read_mostly
= true;
478 unsigned long tick_nohz_active __read_mostly
;
480 * Enable / Disable tickless mode
482 static int __init
setup_tick_nohz(char *str
)
484 return (kstrtobool(str
, &tick_nohz_enabled
) == 0);
487 __setup("nohz=", setup_tick_nohz
);
489 bool tick_nohz_tick_stopped(void)
491 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
493 return ts
->tick_stopped
;
496 bool tick_nohz_tick_stopped_cpu(int cpu
)
498 struct tick_sched
*ts
= per_cpu_ptr(&tick_cpu_sched
, cpu
);
500 return ts
->tick_stopped
;
504 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
506 * Called from interrupt entry when the CPU was idle
508 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
509 * must be updated. Otherwise an interrupt handler could use a stale jiffy
510 * value. We do this unconditionally on any CPU, as we don't know whether the
511 * CPU, which has the update task assigned is in a long sleep.
513 static void tick_nohz_update_jiffies(ktime_t now
)
517 __this_cpu_write(tick_cpu_sched
.idle_waketime
, now
);
519 local_irq_save(flags
);
520 tick_do_update_jiffies64(now
);
521 local_irq_restore(flags
);
523 touch_softlockup_watchdog_sched();
527 * Updates the per-CPU time idle statistics counters
530 update_ts_time_stats(int cpu
, struct tick_sched
*ts
, ktime_t now
, u64
*last_update_time
)
534 if (ts
->idle_active
) {
535 delta
= ktime_sub(now
, ts
->idle_entrytime
);
536 if (nr_iowait_cpu(cpu
) > 0)
537 ts
->iowait_sleeptime
= ktime_add(ts
->iowait_sleeptime
, delta
);
539 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
540 ts
->idle_entrytime
= now
;
543 if (last_update_time
)
544 *last_update_time
= ktime_to_us(now
);
548 static void tick_nohz_stop_idle(struct tick_sched
*ts
, ktime_t now
)
550 update_ts_time_stats(smp_processor_id(), ts
, now
, NULL
);
553 sched_clock_idle_wakeup_event();
556 static void tick_nohz_start_idle(struct tick_sched
*ts
)
558 ts
->idle_entrytime
= ktime_get();
560 sched_clock_idle_sleep_event();
564 * get_cpu_idle_time_us - get the total idle time of a CPU
565 * @cpu: CPU number to query
566 * @last_update_time: variable to store update time in. Do not update
569 * Return the cumulative idle time (since boot) for a given
570 * CPU, in microseconds.
572 * This time is measured via accounting rather than sampling,
573 * and is as accurate as ktime_get() is.
575 * This function returns -1 if NOHZ is not enabled.
577 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
579 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
582 if (!tick_nohz_active
)
586 if (last_update_time
) {
587 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
588 idle
= ts
->idle_sleeptime
;
590 if (ts
->idle_active
&& !nr_iowait_cpu(cpu
)) {
591 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
593 idle
= ktime_add(ts
->idle_sleeptime
, delta
);
595 idle
= ts
->idle_sleeptime
;
599 return ktime_to_us(idle
);
602 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
605 * get_cpu_iowait_time_us - get the total iowait time of a CPU
606 * @cpu: CPU number to query
607 * @last_update_time: variable to store update time in. Do not update
610 * Return the cumulative iowait time (since boot) for a given
611 * CPU, in microseconds.
613 * This time is measured via accounting rather than sampling,
614 * and is as accurate as ktime_get() is.
616 * This function returns -1 if NOHZ is not enabled.
618 u64
get_cpu_iowait_time_us(int cpu
, u64
*last_update_time
)
620 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
623 if (!tick_nohz_active
)
627 if (last_update_time
) {
628 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
629 iowait
= ts
->iowait_sleeptime
;
631 if (ts
->idle_active
&& nr_iowait_cpu(cpu
) > 0) {
632 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
634 iowait
= ktime_add(ts
->iowait_sleeptime
, delta
);
636 iowait
= ts
->iowait_sleeptime
;
640 return ktime_to_us(iowait
);
642 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us
);
644 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
646 hrtimer_cancel(&ts
->sched_timer
);
647 hrtimer_set_expires(&ts
->sched_timer
, ts
->last_tick
);
649 /* Forward the time to expire in the future */
650 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
652 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
653 hrtimer_start_expires(&ts
->sched_timer
,
654 HRTIMER_MODE_ABS_PINNED_HARD
);
656 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
660 * Reset to make sure next tick stop doesn't get fooled by past
661 * cached clock deadline.
666 static inline bool local_timer_softirq_pending(void)
668 return local_softirq_pending() & BIT(TIMER_SOFTIRQ
);
671 static ktime_t
tick_nohz_next_event(struct tick_sched
*ts
, int cpu
)
673 u64 basemono
, next_tick
, next_tmr
, next_rcu
, delta
, expires
;
674 unsigned long basejiff
;
677 /* Read jiffies and the time when jiffies were updated last */
679 seq
= read_seqbegin(&jiffies_lock
);
680 basemono
= last_jiffies_update
;
682 } while (read_seqretry(&jiffies_lock
, seq
));
683 ts
->last_jiffies
= basejiff
;
684 ts
->timer_expires_base
= basemono
;
687 * Keep the periodic tick, when RCU, architecture or irq_work
689 * Aside of that check whether the local timer softirq is
690 * pending. If so its a bad idea to call get_next_timer_interrupt()
691 * because there is an already expired timer, so it will request
692 * immeditate expiry, which rearms the hardware timer with a
693 * minimal delta which brings us back to this place
694 * immediately. Lather, rinse and repeat...
696 if (rcu_needs_cpu(basemono
, &next_rcu
) || arch_needs_cpu() ||
697 irq_work_needs_cpu() || local_timer_softirq_pending()) {
698 next_tick
= basemono
+ TICK_NSEC
;
701 * Get the next pending timer. If high resolution
702 * timers are enabled this only takes the timer wheel
703 * timers into account. If high resolution timers are
704 * disabled this also looks at the next expiring
707 next_tmr
= get_next_timer_interrupt(basejiff
, basemono
);
708 ts
->next_timer
= next_tmr
;
709 /* Take the next rcu event into account */
710 next_tick
= next_rcu
< next_tmr
? next_rcu
: next_tmr
;
714 * If the tick is due in the next period, keep it ticking or
715 * force prod the timer.
717 delta
= next_tick
- basemono
;
718 if (delta
<= (u64
)TICK_NSEC
) {
720 * Tell the timer code that the base is not idle, i.e. undo
721 * the effect of get_next_timer_interrupt():
725 * We've not stopped the tick yet, and there's a timer in the
726 * next period, so no point in stopping it either, bail.
728 if (!ts
->tick_stopped
) {
729 ts
->timer_expires
= 0;
735 * If this CPU is the one which had the do_timer() duty last, we limit
736 * the sleep time to the timekeeping max_deferment value.
737 * Otherwise we can sleep as long as we want.
739 delta
= timekeeping_max_deferment();
740 if (cpu
!= tick_do_timer_cpu
&&
741 (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
|| !ts
->do_timer_last
))
744 /* Calculate the next expiry time */
745 if (delta
< (KTIME_MAX
- basemono
))
746 expires
= basemono
+ delta
;
750 ts
->timer_expires
= min_t(u64
, expires
, next_tick
);
753 return ts
->timer_expires
;
756 static void tick_nohz_stop_tick(struct tick_sched
*ts
, int cpu
)
758 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
759 u64 basemono
= ts
->timer_expires_base
;
760 u64 expires
= ts
->timer_expires
;
761 ktime_t tick
= expires
;
763 /* Make sure we won't be trying to stop it twice in a row. */
764 ts
->timer_expires_base
= 0;
767 * If this CPU is the one which updates jiffies, then give up
768 * the assignment and let it be taken by the CPU which runs
769 * the tick timer next, which might be this CPU as well. If we
770 * don't drop this here the jiffies might be stale and
771 * do_timer() never invoked. Keep track of the fact that it
772 * was the one which had the do_timer() duty last.
774 if (cpu
== tick_do_timer_cpu
) {
775 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
776 ts
->do_timer_last
= 1;
777 } else if (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
) {
778 ts
->do_timer_last
= 0;
781 /* Skip reprogram of event if its not changed */
782 if (ts
->tick_stopped
&& (expires
== ts
->next_tick
)) {
783 /* Sanity check: make sure clockevent is actually programmed */
784 if (tick
== KTIME_MAX
|| ts
->next_tick
== hrtimer_get_expires(&ts
->sched_timer
))
788 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
789 basemono
, ts
->next_tick
, dev
->next_event
,
790 hrtimer_active(&ts
->sched_timer
), hrtimer_get_expires(&ts
->sched_timer
));
794 * nohz_stop_sched_tick can be called several times before
795 * the nohz_restart_sched_tick is called. This happens when
796 * interrupts arrive which do not cause a reschedule. In the
797 * first call we save the current tick time, so we can restart
798 * the scheduler tick in nohz_restart_sched_tick.
800 if (!ts
->tick_stopped
) {
801 calc_load_nohz_start();
804 ts
->last_tick
= hrtimer_get_expires(&ts
->sched_timer
);
805 ts
->tick_stopped
= 1;
806 trace_tick_stop(1, TICK_DEP_MASK_NONE
);
809 ts
->next_tick
= tick
;
812 * If the expiration time == KTIME_MAX, then we simply stop
815 if (unlikely(expires
== KTIME_MAX
)) {
816 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
817 hrtimer_cancel(&ts
->sched_timer
);
821 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
822 hrtimer_start(&ts
->sched_timer
, tick
,
823 HRTIMER_MODE_ABS_PINNED_HARD
);
825 hrtimer_set_expires(&ts
->sched_timer
, tick
);
826 tick_program_event(tick
, 1);
830 static void tick_nohz_retain_tick(struct tick_sched
*ts
)
832 ts
->timer_expires_base
= 0;
835 #ifdef CONFIG_NO_HZ_FULL
836 static void tick_nohz_stop_sched_tick(struct tick_sched
*ts
, int cpu
)
838 if (tick_nohz_next_event(ts
, cpu
))
839 tick_nohz_stop_tick(ts
, cpu
);
841 tick_nohz_retain_tick(ts
);
843 #endif /* CONFIG_NO_HZ_FULL */
845 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
)
847 /* Update jiffies first */
848 tick_do_update_jiffies64(now
);
851 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
852 * the clock forward checks in the enqueue path:
856 calc_load_nohz_stop();
857 touch_softlockup_watchdog_sched();
859 * Cancel the scheduled timer and restore the tick
861 ts
->tick_stopped
= 0;
862 ts
->idle_exittime
= now
;
864 tick_nohz_restart(ts
, now
);
867 static void tick_nohz_full_update_tick(struct tick_sched
*ts
)
869 #ifdef CONFIG_NO_HZ_FULL
870 int cpu
= smp_processor_id();
872 if (!tick_nohz_full_cpu(cpu
))
875 if (!ts
->tick_stopped
&& ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)
878 if (can_stop_full_tick(cpu
, ts
))
879 tick_nohz_stop_sched_tick(ts
, cpu
);
880 else if (ts
->tick_stopped
)
881 tick_nohz_restart_sched_tick(ts
, ktime_get());
885 static bool can_stop_idle_tick(int cpu
, struct tick_sched
*ts
)
888 * If this CPU is offline and it is the one which updates
889 * jiffies, then give up the assignment and let it be taken by
890 * the CPU which runs the tick timer next. If we don't drop
891 * this here the jiffies might be stale and do_timer() never
894 if (unlikely(!cpu_online(cpu
))) {
895 if (cpu
== tick_do_timer_cpu
)
896 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
898 * Make sure the CPU doesn't get fooled by obsolete tick
899 * deadline if it comes back online later.
905 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
))
911 if (unlikely(local_softirq_pending())) {
912 static int ratelimit
;
914 if (ratelimit
< 10 &&
915 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK
)) {
916 pr_warn("NOHZ: local_softirq_pending %02x\n",
917 (unsigned int) local_softirq_pending());
923 if (tick_nohz_full_enabled()) {
925 * Keep the tick alive to guarantee timekeeping progression
926 * if there are full dynticks CPUs around
928 if (tick_do_timer_cpu
== cpu
)
931 * Boot safety: make sure the timekeeping duty has been
932 * assigned before entering dyntick-idle mode,
933 * tick_do_timer_cpu is TICK_DO_TIMER_BOOT
935 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_BOOT
))
938 /* Should not happen for nohz-full */
939 if (WARN_ON_ONCE(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
946 static void __tick_nohz_idle_stop_tick(struct tick_sched
*ts
)
949 int cpu
= smp_processor_id();
952 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
953 * tick timer expiration time is known already.
955 if (ts
->timer_expires_base
)
956 expires
= ts
->timer_expires
;
957 else if (can_stop_idle_tick(cpu
, ts
))
958 expires
= tick_nohz_next_event(ts
, cpu
);
965 int was_stopped
= ts
->tick_stopped
;
967 tick_nohz_stop_tick(ts
, cpu
);
970 ts
->idle_expires
= expires
;
972 if (!was_stopped
&& ts
->tick_stopped
) {
973 ts
->idle_jiffies
= ts
->last_jiffies
;
974 nohz_balance_enter_idle(cpu
);
977 tick_nohz_retain_tick(ts
);
982 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
984 * When the next event is more than a tick into the future, stop the idle tick
986 void tick_nohz_idle_stop_tick(void)
988 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched
));
991 void tick_nohz_idle_retain_tick(void)
993 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched
));
995 * Undo the effect of get_next_timer_interrupt() called from
996 * tick_nohz_next_event().
1002 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1004 * Called when we start the idle loop.
1006 void tick_nohz_idle_enter(void)
1008 struct tick_sched
*ts
;
1010 lockdep_assert_irqs_enabled();
1012 local_irq_disable();
1014 ts
= this_cpu_ptr(&tick_cpu_sched
);
1016 WARN_ON_ONCE(ts
->timer_expires_base
);
1019 tick_nohz_start_idle(ts
);
1025 * tick_nohz_irq_exit - update next tick event from interrupt exit
1027 * When an interrupt fires while we are idle and it doesn't cause
1028 * a reschedule, it may still add, modify or delete a timer, enqueue
1029 * an RCU callback, etc...
1030 * So we need to re-calculate and reprogram the next tick event.
1032 void tick_nohz_irq_exit(void)
1034 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1037 tick_nohz_start_idle(ts
);
1039 tick_nohz_full_update_tick(ts
);
1043 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1045 bool tick_nohz_idle_got_tick(void)
1047 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1049 if (ts
->got_idle_tick
) {
1050 ts
->got_idle_tick
= 0;
1057 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1058 * or the tick, whatever that expires first. Note that, if the tick has been
1059 * stopped, it returns the next hrtimer.
1061 * Called from power state control code with interrupts disabled
1063 ktime_t
tick_nohz_get_next_hrtimer(void)
1065 return __this_cpu_read(tick_cpu_device
.evtdev
)->next_event
;
1069 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1070 * @delta_next: duration until the next event if the tick cannot be stopped
1072 * Called from power state control code with interrupts disabled
1074 ktime_t
tick_nohz_get_sleep_length(ktime_t
*delta_next
)
1076 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
1077 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1078 int cpu
= smp_processor_id();
1080 * The idle entry time is expected to be a sufficient approximation of
1081 * the current time at this point.
1083 ktime_t now
= ts
->idle_entrytime
;
1086 WARN_ON_ONCE(!ts
->inidle
);
1088 *delta_next
= ktime_sub(dev
->next_event
, now
);
1090 if (!can_stop_idle_tick(cpu
, ts
))
1093 next_event
= tick_nohz_next_event(ts
, cpu
);
1098 * If the next highres timer to expire is earlier than next_event, the
1099 * idle governor needs to know that.
1101 next_event
= min_t(u64
, next_event
,
1102 hrtimer_next_event_without(&ts
->sched_timer
));
1104 return ktime_sub(next_event
, now
);
1108 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1109 * for a particular CPU.
1111 * Called from the schedutil frequency scaling governor in scheduler context.
1113 unsigned long tick_nohz_get_idle_calls_cpu(int cpu
)
1115 struct tick_sched
*ts
= tick_get_tick_sched(cpu
);
1117 return ts
->idle_calls
;
1121 * tick_nohz_get_idle_calls - return the current idle calls counter value
1123 * Called from the schedutil frequency scaling governor in scheduler context.
1125 unsigned long tick_nohz_get_idle_calls(void)
1127 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1129 return ts
->idle_calls
;
1132 static void tick_nohz_account_idle_ticks(struct tick_sched
*ts
)
1134 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1135 unsigned long ticks
;
1137 if (vtime_accounting_enabled_this_cpu())
1140 * We stopped the tick in idle. Update process times would miss the
1141 * time we slept as update_process_times does only a 1 tick
1142 * accounting. Enforce that this is accounted to idle !
1144 ticks
= jiffies
- ts
->idle_jiffies
;
1146 * We might be one off. Do not randomly account a huge number of ticks!
1148 if (ticks
&& ticks
< LONG_MAX
)
1149 account_idle_ticks(ticks
);
1153 static void __tick_nohz_idle_restart_tick(struct tick_sched
*ts
, ktime_t now
)
1155 tick_nohz_restart_sched_tick(ts
, now
);
1156 tick_nohz_account_idle_ticks(ts
);
1159 void tick_nohz_idle_restart_tick(void)
1161 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1163 if (ts
->tick_stopped
)
1164 __tick_nohz_idle_restart_tick(ts
, ktime_get());
1168 * tick_nohz_idle_exit - restart the idle tick from the idle task
1170 * Restart the idle tick when the CPU is woken up from idle
1171 * This also exit the RCU extended quiescent state. The CPU
1172 * can use RCU again after this function is called.
1174 void tick_nohz_idle_exit(void)
1176 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1177 bool idle_active
, tick_stopped
;
1180 local_irq_disable();
1182 WARN_ON_ONCE(!ts
->inidle
);
1183 WARN_ON_ONCE(ts
->timer_expires_base
);
1186 idle_active
= ts
->idle_active
;
1187 tick_stopped
= ts
->tick_stopped
;
1189 if (idle_active
|| tick_stopped
)
1193 tick_nohz_stop_idle(ts
, now
);
1196 __tick_nohz_idle_restart_tick(ts
, now
);
1202 * The nohz low res interrupt handler
1204 static void tick_nohz_handler(struct clock_event_device
*dev
)
1206 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1207 struct pt_regs
*regs
= get_irq_regs();
1208 ktime_t now
= ktime_get();
1210 dev
->next_event
= KTIME_MAX
;
1212 tick_sched_do_timer(ts
, now
);
1213 tick_sched_handle(ts
, regs
);
1215 /* No need to reprogram if we are running tickless */
1216 if (unlikely(ts
->tick_stopped
))
1219 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1220 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
1223 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
)
1225 if (!tick_nohz_enabled
)
1227 ts
->nohz_mode
= mode
;
1228 /* One update is enough */
1229 if (!test_and_set_bit(0, &tick_nohz_active
))
1230 timers_update_nohz();
1234 * tick_nohz_switch_to_nohz - switch to nohz mode
1236 static void tick_nohz_switch_to_nohz(void)
1238 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1241 if (!tick_nohz_enabled
)
1244 if (tick_switch_to_oneshot(tick_nohz_handler
))
1248 * Recycle the hrtimer in ts, so we can share the
1249 * hrtimer_forward with the highres code.
1251 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_HARD
);
1252 /* Get the next period */
1253 next
= tick_init_jiffy_update();
1255 hrtimer_set_expires(&ts
->sched_timer
, next
);
1256 hrtimer_forward_now(&ts
->sched_timer
, tick_period
);
1257 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
1258 tick_nohz_activate(ts
, NOHZ_MODE_LOWRES
);
1261 static inline void tick_nohz_irq_enter(void)
1263 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1266 if (!ts
->idle_active
&& !ts
->tick_stopped
)
1269 if (ts
->idle_active
)
1270 tick_nohz_stop_idle(ts
, now
);
1271 if (ts
->tick_stopped
)
1272 tick_nohz_update_jiffies(now
);
1277 static inline void tick_nohz_switch_to_nohz(void) { }
1278 static inline void tick_nohz_irq_enter(void) { }
1279 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
) { }
1281 #endif /* CONFIG_NO_HZ_COMMON */
1284 * Called from irq_enter to notify about the possible interruption of idle()
1286 void tick_irq_enter(void)
1288 tick_check_oneshot_broadcast_this_cpu();
1289 tick_nohz_irq_enter();
1293 * High resolution timer specific code
1295 #ifdef CONFIG_HIGH_RES_TIMERS
1297 * We rearm the timer until we get disabled by the idle code.
1298 * Called with interrupts disabled.
1300 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
1302 struct tick_sched
*ts
=
1303 container_of(timer
, struct tick_sched
, sched_timer
);
1304 struct pt_regs
*regs
= get_irq_regs();
1305 ktime_t now
= ktime_get();
1307 tick_sched_do_timer(ts
, now
);
1310 * Do not call, when we are not in irq context and have
1311 * no valid regs pointer
1314 tick_sched_handle(ts
, regs
);
1318 /* No need to reprogram if we are in idle or full dynticks mode */
1319 if (unlikely(ts
->tick_stopped
))
1320 return HRTIMER_NORESTART
;
1322 hrtimer_forward(timer
, now
, tick_period
);
1324 return HRTIMER_RESTART
;
1327 static int sched_skew_tick
;
1329 static int __init
skew_tick(char *str
)
1331 get_option(&str
, &sched_skew_tick
);
1335 early_param("skew_tick", skew_tick
);
1338 * tick_setup_sched_timer - setup the tick emulation timer
1340 void tick_setup_sched_timer(void)
1342 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1343 ktime_t now
= ktime_get();
1346 * Emulate tick processing via per-CPU hrtimers:
1348 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_HARD
);
1349 ts
->sched_timer
.function
= tick_sched_timer
;
1351 /* Get the next period (per-CPU) */
1352 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
1354 /* Offset the tick to avert jiffies_lock contention. */
1355 if (sched_skew_tick
) {
1356 u64 offset
= ktime_to_ns(tick_period
) >> 1;
1357 do_div(offset
, num_possible_cpus());
1358 offset
*= smp_processor_id();
1359 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
1362 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1363 hrtimer_start_expires(&ts
->sched_timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
1364 tick_nohz_activate(ts
, NOHZ_MODE_HIGHRES
);
1366 #endif /* HIGH_RES_TIMERS */
1368 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1369 void tick_cancel_sched_timer(int cpu
)
1371 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1373 # ifdef CONFIG_HIGH_RES_TIMERS
1374 if (ts
->sched_timer
.base
)
1375 hrtimer_cancel(&ts
->sched_timer
);
1378 memset(ts
, 0, sizeof(*ts
));
1383 * Async notification about clocksource changes
1385 void tick_clock_notify(void)
1389 for_each_possible_cpu(cpu
)
1390 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
1394 * Async notification about clock event changes
1396 void tick_oneshot_notify(void)
1398 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1400 set_bit(0, &ts
->check_clocks
);
1404 * Check, if a change happened, which makes oneshot possible.
1406 * Called cyclic from the hrtimer softirq (driven by the timer
1407 * softirq) allow_nohz signals, that we can switch into low-res nohz
1408 * mode, because high resolution timers are disabled (either compile
1409 * or runtime). Called with interrupts disabled.
1411 int tick_check_oneshot_change(int allow_nohz
)
1413 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1415 if (!test_and_clear_bit(0, &ts
->check_clocks
))
1418 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
1421 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1427 tick_nohz_switch_to_nohz();