2 * linux/kernel/time/tick-sched.c
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
8 * No idle tick implementation for low and high resolution timers
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * Distribute under GPLv2.
14 #include <linux/cpu.h>
15 #include <linux/err.h>
16 #include <linux/hrtimer.h>
17 #include <linux/interrupt.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/percpu.h>
20 #include <linux/profile.h>
21 #include <linux/sched.h>
22 #include <linux/module.h>
23 #include <linux/irq_work.h>
24 #include <linux/posix-timers.h>
25 #include <linux/perf_event.h>
26 #include <linux/context_tracking.h>
28 #include <asm/irq_regs.h>
30 #include "tick-internal.h"
32 #include <trace/events/timer.h>
35 * Per cpu nohz control structure
37 static DEFINE_PER_CPU(struct tick_sched
, tick_cpu_sched
);
40 * The time, when the last jiffy update happened. Protected by jiffies_lock.
42 static ktime_t last_jiffies_update
;
44 struct tick_sched
*tick_get_tick_sched(int cpu
)
46 return &per_cpu(tick_cpu_sched
, cpu
);
50 * Must be called with interrupts disabled !
52 static void tick_do_update_jiffies64(ktime_t now
)
54 unsigned long ticks
= 0;
58 * Do a quick check without holding jiffies_lock:
60 delta
= ktime_sub(now
, last_jiffies_update
);
61 if (delta
.tv64
< tick_period
.tv64
)
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock
);
67 delta
= ktime_sub(now
, last_jiffies_update
);
68 if (delta
.tv64
>= tick_period
.tv64
) {
70 delta
= ktime_sub(delta
, tick_period
);
71 last_jiffies_update
= ktime_add(last_jiffies_update
,
74 /* Slow path for long timeouts */
75 if (unlikely(delta
.tv64
>= tick_period
.tv64
)) {
76 s64 incr
= ktime_to_ns(tick_period
);
78 ticks
= ktime_divns(delta
, incr
);
80 last_jiffies_update
= ktime_add_ns(last_jiffies_update
,
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period
= ktime_add(last_jiffies_update
, tick_period
);
88 write_sequnlock(&jiffies_lock
);
91 write_sequnlock(&jiffies_lock
);
96 * Initialize and return retrieve the jiffies update.
98 static ktime_t
tick_init_jiffy_update(void)
102 write_seqlock(&jiffies_lock
);
103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update
.tv64
== 0)
105 last_jiffies_update
= tick_next_period
;
106 period
= last_jiffies_update
;
107 write_sequnlock(&jiffies_lock
);
112 static void tick_sched_do_timer(ktime_t now
)
114 int cpu
= smp_processor_id();
116 #ifdef CONFIG_NO_HZ_COMMON
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the cpu in charge went
120 * into a long sleep. If two cpus happen to assign themself to
121 * this duty, then the jiffies update is still serialized by
124 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)
125 && !tick_nohz_full_cpu(cpu
))
126 tick_do_timer_cpu
= cpu
;
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu
== cpu
)
131 tick_do_update_jiffies64(now
);
134 static void tick_sched_handle(struct tick_sched
*ts
, struct pt_regs
*regs
)
136 #ifdef CONFIG_NO_HZ_COMMON
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
145 if (ts
->tick_stopped
) {
146 touch_softlockup_watchdog();
147 if (is_idle_task(current
))
151 update_process_times(user_mode(regs
));
152 profile_tick(CPU_PROFILING
);
155 #ifdef CONFIG_NO_HZ_FULL
156 cpumask_var_t tick_nohz_full_mask
;
157 cpumask_var_t housekeeping_mask
;
158 bool tick_nohz_full_running
;
160 static bool can_stop_full_tick(void)
162 WARN_ON_ONCE(!irqs_disabled());
164 if (!sched_can_stop_tick()) {
165 trace_tick_stop(0, "more than 1 task in runqueue\n");
169 if (!posix_cpu_timers_can_stop_tick(current
)) {
170 trace_tick_stop(0, "posix timers running\n");
174 if (!perf_event_can_stop_tick()) {
175 trace_tick_stop(0, "perf events running\n");
179 /* sched_clock_tick() needs us? */
180 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
182 * TODO: kick full dynticks CPUs when
183 * sched_clock_stable is set.
185 if (!sched_clock_stable()) {
186 trace_tick_stop(0, "unstable sched clock\n");
188 * Don't allow the user to think they can get
189 * full NO_HZ with this machine.
191 WARN_ONCE(tick_nohz_full_running
,
192 "NO_HZ FULL will not work with unstable sched clock");
200 static void nohz_full_kick_work_func(struct irq_work
*work
)
202 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
205 static DEFINE_PER_CPU(struct irq_work
, nohz_full_kick_work
) = {
206 .func
= nohz_full_kick_work_func
,
210 * Kick this CPU if it's full dynticks in order to force it to
211 * re-evaluate its dependency on the tick and restart it if necessary.
212 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
215 void tick_nohz_full_kick(void)
217 if (!tick_nohz_full_cpu(smp_processor_id()))
220 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work
));
224 * Kick the CPU if it's full dynticks in order to force it to
225 * re-evaluate its dependency on the tick and restart it if necessary.
227 void tick_nohz_full_kick_cpu(int cpu
)
229 if (!tick_nohz_full_cpu(cpu
))
232 irq_work_queue_on(&per_cpu(nohz_full_kick_work
, cpu
), cpu
);
235 static void nohz_full_kick_ipi(void *info
)
237 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
241 * Kick all full dynticks CPUs in order to force these to re-evaluate
242 * their dependency on the tick and restart it if necessary.
244 void tick_nohz_full_kick_all(void)
246 if (!tick_nohz_full_running
)
250 smp_call_function_many(tick_nohz_full_mask
,
251 nohz_full_kick_ipi
, NULL
, false);
252 tick_nohz_full_kick();
257 * Re-evaluate the need for the tick as we switch the current task.
258 * It might need the tick due to per task/process properties:
259 * perf events, posix cpu timers, ...
261 void __tick_nohz_task_switch(void)
265 local_irq_save(flags
);
267 if (!tick_nohz_full_cpu(smp_processor_id()))
270 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
271 tick_nohz_full_kick();
274 local_irq_restore(flags
);
277 /* Parse the boot-time nohz CPU list from the kernel parameters. */
278 static int __init
tick_nohz_full_setup(char *str
)
280 alloc_bootmem_cpumask_var(&tick_nohz_full_mask
);
281 if (cpulist_parse(str
, tick_nohz_full_mask
) < 0) {
282 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
283 free_bootmem_cpumask_var(tick_nohz_full_mask
);
286 tick_nohz_full_running
= true;
290 __setup("nohz_full=", tick_nohz_full_setup
);
292 static int tick_nohz_cpu_down_callback(struct notifier_block
*nfb
,
293 unsigned long action
,
296 unsigned int cpu
= (unsigned long)hcpu
;
298 switch (action
& ~CPU_TASKS_FROZEN
) {
299 case CPU_DOWN_PREPARE
:
301 * If we handle the timekeeping duty for full dynticks CPUs,
302 * we can't safely shutdown that CPU.
304 if (tick_nohz_full_running
&& tick_do_timer_cpu
== cpu
)
311 static int tick_nohz_init_all(void)
315 #ifdef CONFIG_NO_HZ_FULL_ALL
316 if (!alloc_cpumask_var(&tick_nohz_full_mask
, GFP_KERNEL
)) {
317 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
321 cpumask_setall(tick_nohz_full_mask
);
322 tick_nohz_full_running
= true;
327 void __init
tick_nohz_init(void)
331 if (!tick_nohz_full_running
) {
332 if (tick_nohz_init_all() < 0)
336 if (!alloc_cpumask_var(&housekeeping_mask
, GFP_KERNEL
)) {
337 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
338 cpumask_clear(tick_nohz_full_mask
);
339 tick_nohz_full_running
= false;
344 * Full dynticks uses irq work to drive the tick rescheduling on safe
345 * locking contexts. But then we need irq work to raise its own
346 * interrupts to avoid circular dependency on the tick
348 if (!arch_irq_work_has_interrupt()) {
349 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't "
350 "support irq work self-IPIs\n");
351 cpumask_clear(tick_nohz_full_mask
);
352 cpumask_copy(housekeeping_mask
, cpu_possible_mask
);
353 tick_nohz_full_running
= false;
357 cpu
= smp_processor_id();
359 if (cpumask_test_cpu(cpu
, tick_nohz_full_mask
)) {
360 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu
);
361 cpumask_clear_cpu(cpu
, tick_nohz_full_mask
);
364 cpumask_andnot(housekeeping_mask
,
365 cpu_possible_mask
, tick_nohz_full_mask
);
367 for_each_cpu(cpu
, tick_nohz_full_mask
)
368 context_tracking_cpu_set(cpu
);
370 cpu_notifier(tick_nohz_cpu_down_callback
, 0);
371 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
372 cpumask_pr_args(tick_nohz_full_mask
));
377 * NOHZ - aka dynamic tick functionality
379 #ifdef CONFIG_NO_HZ_COMMON
383 static int tick_nohz_enabled __read_mostly
= 1;
384 unsigned long tick_nohz_active __read_mostly
;
386 * Enable / Disable tickless mode
388 static int __init
setup_tick_nohz(char *str
)
390 if (!strcmp(str
, "off"))
391 tick_nohz_enabled
= 0;
392 else if (!strcmp(str
, "on"))
393 tick_nohz_enabled
= 1;
399 __setup("nohz=", setup_tick_nohz
);
401 int tick_nohz_tick_stopped(void)
403 return __this_cpu_read(tick_cpu_sched
.tick_stopped
);
407 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
409 * Called from interrupt entry when the CPU was idle
411 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
412 * must be updated. Otherwise an interrupt handler could use a stale jiffy
413 * value. We do this unconditionally on any cpu, as we don't know whether the
414 * cpu, which has the update task assigned is in a long sleep.
416 static void tick_nohz_update_jiffies(ktime_t now
)
420 __this_cpu_write(tick_cpu_sched
.idle_waketime
, now
);
422 local_irq_save(flags
);
423 tick_do_update_jiffies64(now
);
424 local_irq_restore(flags
);
426 touch_softlockup_watchdog();
430 * Updates the per cpu time idle statistics counters
433 update_ts_time_stats(int cpu
, struct tick_sched
*ts
, ktime_t now
, u64
*last_update_time
)
437 if (ts
->idle_active
) {
438 delta
= ktime_sub(now
, ts
->idle_entrytime
);
439 if (nr_iowait_cpu(cpu
) > 0)
440 ts
->iowait_sleeptime
= ktime_add(ts
->iowait_sleeptime
, delta
);
442 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
443 ts
->idle_entrytime
= now
;
446 if (last_update_time
)
447 *last_update_time
= ktime_to_us(now
);
451 static void tick_nohz_stop_idle(struct tick_sched
*ts
, ktime_t now
)
453 update_ts_time_stats(smp_processor_id(), ts
, now
, NULL
);
456 sched_clock_idle_wakeup_event(0);
459 static ktime_t
tick_nohz_start_idle(struct tick_sched
*ts
)
461 ktime_t now
= ktime_get();
463 ts
->idle_entrytime
= now
;
465 sched_clock_idle_sleep_event();
470 * get_cpu_idle_time_us - get the total idle time of a cpu
471 * @cpu: CPU number to query
472 * @last_update_time: variable to store update time in. Do not update
475 * Return the cummulative idle time (since boot) for a given
476 * CPU, in microseconds.
478 * This time is measured via accounting rather than sampling,
479 * and is as accurate as ktime_get() is.
481 * This function returns -1 if NOHZ is not enabled.
483 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
485 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
488 if (!tick_nohz_active
)
492 if (last_update_time
) {
493 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
494 idle
= ts
->idle_sleeptime
;
496 if (ts
->idle_active
&& !nr_iowait_cpu(cpu
)) {
497 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
499 idle
= ktime_add(ts
->idle_sleeptime
, delta
);
501 idle
= ts
->idle_sleeptime
;
505 return ktime_to_us(idle
);
508 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
511 * get_cpu_iowait_time_us - get the total iowait time of a cpu
512 * @cpu: CPU number to query
513 * @last_update_time: variable to store update time in. Do not update
516 * Return the cummulative iowait time (since boot) for a given
517 * CPU, in microseconds.
519 * This time is measured via accounting rather than sampling,
520 * and is as accurate as ktime_get() is.
522 * This function returns -1 if NOHZ is not enabled.
524 u64
get_cpu_iowait_time_us(int cpu
, u64
*last_update_time
)
526 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
529 if (!tick_nohz_active
)
533 if (last_update_time
) {
534 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
535 iowait
= ts
->iowait_sleeptime
;
537 if (ts
->idle_active
&& nr_iowait_cpu(cpu
) > 0) {
538 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
540 iowait
= ktime_add(ts
->iowait_sleeptime
, delta
);
542 iowait
= ts
->iowait_sleeptime
;
546 return ktime_to_us(iowait
);
548 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us
);
550 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
552 hrtimer_cancel(&ts
->sched_timer
);
553 hrtimer_set_expires(&ts
->sched_timer
, ts
->last_tick
);
555 /* Forward the time to expire in the future */
556 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
558 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
559 hrtimer_start_expires(&ts
->sched_timer
, HRTIMER_MODE_ABS_PINNED
);
561 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
564 static ktime_t
tick_nohz_stop_sched_tick(struct tick_sched
*ts
,
565 ktime_t now
, int cpu
)
567 struct clock_event_device
*dev
= __this_cpu_read(tick_cpu_device
.evtdev
);
568 u64 basemono
, next_tick
, next_tmr
, next_rcu
, delta
, expires
;
569 unsigned long seq
, basejiff
;
572 /* Read jiffies and the time when jiffies were updated last */
574 seq
= read_seqbegin(&jiffies_lock
);
575 basemono
= last_jiffies_update
.tv64
;
577 } while (read_seqretry(&jiffies_lock
, seq
));
578 ts
->last_jiffies
= basejiff
;
580 if (rcu_needs_cpu(basemono
, &next_rcu
) ||
581 arch_needs_cpu() || irq_work_needs_cpu()) {
582 next_tick
= basemono
+ TICK_NSEC
;
585 * Get the next pending timer. If high resolution
586 * timers are enabled this only takes the timer wheel
587 * timers into account. If high resolution timers are
588 * disabled this also looks at the next expiring
591 next_tmr
= get_next_timer_interrupt(basejiff
, basemono
);
592 ts
->next_timer
= next_tmr
;
593 /* Take the next rcu event into account */
594 next_tick
= next_rcu
< next_tmr
? next_rcu
: next_tmr
;
598 * If the tick is due in the next period, keep it ticking or
601 delta
= next_tick
- basemono
;
602 if (delta
<= (u64
)TICK_NSEC
) {
604 if (!ts
->tick_stopped
)
607 /* Tick is stopped, but required now. Enforce it */
608 tick_nohz_restart(ts
, now
);
614 * If this cpu is the one which updates jiffies, then give up
615 * the assignment and let it be taken by the cpu which runs
616 * the tick timer next, which might be this cpu as well. If we
617 * don't drop this here the jiffies might be stale and
618 * do_timer() never invoked. Keep track of the fact that it
619 * was the one which had the do_timer() duty last. If this cpu
620 * is the one which had the do_timer() duty last, we limit the
621 * sleep time to the timekeeping max_deferement value.
622 * Otherwise we can sleep as long as we want.
624 delta
= timekeeping_max_deferment();
625 if (cpu
== tick_do_timer_cpu
) {
626 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
627 ts
->do_timer_last
= 1;
628 } else if (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
) {
630 ts
->do_timer_last
= 0;
631 } else if (!ts
->do_timer_last
) {
635 #ifdef CONFIG_NO_HZ_FULL
636 /* Limit the tick delta to the maximum scheduler deferment */
638 delta
= min(delta
, scheduler_tick_max_deferment());
641 /* Calculate the next expiry time */
642 if (delta
< (KTIME_MAX
- basemono
))
643 expires
= basemono
+ delta
;
647 expires
= min_t(u64
, expires
, next_tick
);
650 /* Skip reprogram of event if its not changed */
651 if (ts
->tick_stopped
&& (expires
== dev
->next_event
.tv64
))
655 * nohz_stop_sched_tick can be called several times before
656 * the nohz_restart_sched_tick is called. This happens when
657 * interrupts arrive which do not cause a reschedule. In the
658 * first call we save the current tick time, so we can restart
659 * the scheduler tick in nohz_restart_sched_tick.
661 if (!ts
->tick_stopped
) {
662 nohz_balance_enter_idle(cpu
);
663 calc_load_enter_idle();
665 ts
->last_tick
= hrtimer_get_expires(&ts
->sched_timer
);
666 ts
->tick_stopped
= 1;
667 trace_tick_stop(1, " ");
671 * If the expiration time == KTIME_MAX, then we simply stop
674 if (unlikely(expires
== KTIME_MAX
)) {
675 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
676 hrtimer_cancel(&ts
->sched_timer
);
680 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
681 hrtimer_start(&ts
->sched_timer
, tick
, HRTIMER_MODE_ABS_PINNED
);
683 tick_program_event(tick
, 1);
685 /* Update the estimated sleep length */
686 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
690 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
)
692 /* Update jiffies first */
693 tick_do_update_jiffies64(now
);
694 update_cpu_load_nohz();
696 calc_load_exit_idle();
697 touch_softlockup_watchdog();
699 * Cancel the scheduled timer and restore the tick
701 ts
->tick_stopped
= 0;
702 ts
->idle_exittime
= now
;
704 tick_nohz_restart(ts
, now
);
707 static void tick_nohz_full_update_tick(struct tick_sched
*ts
)
709 #ifdef CONFIG_NO_HZ_FULL
710 int cpu
= smp_processor_id();
712 if (!tick_nohz_full_cpu(cpu
))
715 if (!ts
->tick_stopped
&& ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)
718 if (can_stop_full_tick())
719 tick_nohz_stop_sched_tick(ts
, ktime_get(), cpu
);
720 else if (ts
->tick_stopped
)
721 tick_nohz_restart_sched_tick(ts
, ktime_get());
725 static bool can_stop_idle_tick(int cpu
, struct tick_sched
*ts
)
728 * If this cpu is offline and it is the one which updates
729 * jiffies, then give up the assignment and let it be taken by
730 * the cpu which runs the tick timer next. If we don't drop
731 * this here the jiffies might be stale and do_timer() never
734 if (unlikely(!cpu_online(cpu
))) {
735 if (cpu
== tick_do_timer_cpu
)
736 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
740 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)) {
741 ts
->sleep_length
= (ktime_t
) { .tv64
= NSEC_PER_SEC
/HZ
};
748 if (unlikely(local_softirq_pending() && cpu_online(cpu
))) {
749 static int ratelimit
;
751 if (ratelimit
< 10 &&
752 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK
)) {
753 pr_warn("NOHZ: local_softirq_pending %02x\n",
754 (unsigned int) local_softirq_pending());
760 if (tick_nohz_full_enabled()) {
762 * Keep the tick alive to guarantee timekeeping progression
763 * if there are full dynticks CPUs around
765 if (tick_do_timer_cpu
== cpu
)
768 * Boot safety: make sure the timekeeping duty has been
769 * assigned before entering dyntick-idle mode,
771 if (tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)
778 static void __tick_nohz_idle_enter(struct tick_sched
*ts
)
780 ktime_t now
, expires
;
781 int cpu
= smp_processor_id();
783 now
= tick_nohz_start_idle(ts
);
785 if (can_stop_idle_tick(cpu
, ts
)) {
786 int was_stopped
= ts
->tick_stopped
;
790 expires
= tick_nohz_stop_sched_tick(ts
, now
, cpu
);
791 if (expires
.tv64
> 0LL) {
793 ts
->idle_expires
= expires
;
796 if (!was_stopped
&& ts
->tick_stopped
)
797 ts
->idle_jiffies
= ts
->last_jiffies
;
802 * tick_nohz_idle_enter - stop the idle tick from the idle task
804 * When the next event is more than a tick into the future, stop the idle tick
805 * Called when we start the idle loop.
807 * The arch is responsible of calling:
809 * - rcu_idle_enter() after its last use of RCU before the CPU is put
811 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
813 void tick_nohz_idle_enter(void)
815 struct tick_sched
*ts
;
817 WARN_ON_ONCE(irqs_disabled());
820 * Update the idle state in the scheduler domain hierarchy
821 * when tick_nohz_stop_sched_tick() is called from the idle loop.
822 * State will be updated to busy during the first busy tick after
825 set_cpu_sd_state_idle();
829 ts
= this_cpu_ptr(&tick_cpu_sched
);
831 __tick_nohz_idle_enter(ts
);
837 * tick_nohz_irq_exit - update next tick event from interrupt exit
839 * When an interrupt fires while we are idle and it doesn't cause
840 * a reschedule, it may still add, modify or delete a timer, enqueue
841 * an RCU callback, etc...
842 * So we need to re-calculate and reprogram the next tick event.
844 void tick_nohz_irq_exit(void)
846 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
849 __tick_nohz_idle_enter(ts
);
851 tick_nohz_full_update_tick(ts
);
855 * tick_nohz_get_sleep_length - return the length of the current sleep
857 * Called from power state control code with interrupts disabled
859 ktime_t
tick_nohz_get_sleep_length(void)
861 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
863 return ts
->sleep_length
;
866 static void tick_nohz_account_idle_ticks(struct tick_sched
*ts
)
868 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
871 if (vtime_accounting_enabled())
874 * We stopped the tick in idle. Update process times would miss the
875 * time we slept as update_process_times does only a 1 tick
876 * accounting. Enforce that this is accounted to idle !
878 ticks
= jiffies
- ts
->idle_jiffies
;
880 * We might be one off. Do not randomly account a huge number of ticks!
882 if (ticks
&& ticks
< LONG_MAX
)
883 account_idle_ticks(ticks
);
888 * tick_nohz_idle_exit - restart the idle tick from the idle task
890 * Restart the idle tick when the CPU is woken up from idle
891 * This also exit the RCU extended quiescent state. The CPU
892 * can use RCU again after this function is called.
894 void tick_nohz_idle_exit(void)
896 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
901 WARN_ON_ONCE(!ts
->inidle
);
905 if (ts
->idle_active
|| ts
->tick_stopped
)
909 tick_nohz_stop_idle(ts
, now
);
911 if (ts
->tick_stopped
) {
912 tick_nohz_restart_sched_tick(ts
, now
);
913 tick_nohz_account_idle_ticks(ts
);
920 * The nohz low res interrupt handler
922 static void tick_nohz_handler(struct clock_event_device
*dev
)
924 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
925 struct pt_regs
*regs
= get_irq_regs();
926 ktime_t now
= ktime_get();
928 dev
->next_event
.tv64
= KTIME_MAX
;
930 tick_sched_do_timer(now
);
931 tick_sched_handle(ts
, regs
);
933 /* No need to reprogram if we are running tickless */
934 if (unlikely(ts
->tick_stopped
))
937 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
938 tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 1);
941 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
)
943 if (!tick_nohz_enabled
)
945 ts
->nohz_mode
= mode
;
946 /* One update is enough */
947 if (!test_and_set_bit(0, &tick_nohz_active
))
948 timers_update_migration(true);
952 * tick_nohz_switch_to_nohz - switch to nohz mode
954 static void tick_nohz_switch_to_nohz(void)
956 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
959 if (!tick_nohz_enabled
)
962 if (tick_switch_to_oneshot(tick_nohz_handler
))
966 * Recycle the hrtimer in ts, so we can share the
967 * hrtimer_forward with the highres code.
969 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
970 /* Get the next period */
971 next
= tick_init_jiffy_update();
973 hrtimer_forward_now(&ts
->sched_timer
, tick_period
);
974 hrtimer_set_expires(&ts
->sched_timer
, next
);
975 tick_program_event(next
, 1);
976 tick_nohz_activate(ts
, NOHZ_MODE_LOWRES
);
980 * When NOHZ is enabled and the tick is stopped, we need to kick the
981 * tick timer from irq_enter() so that the jiffies update is kept
982 * alive during long running softirqs. That's ugly as hell, but
983 * correctness is key even if we need to fix the offending softirq in
986 * Note, this is different to tick_nohz_restart. We just kick the
987 * timer and do not touch the other magic bits which need to be done
990 static void tick_nohz_kick_tick(struct tick_sched
*ts
, ktime_t now
)
993 /* Switch back to 2.6.27 behaviour */
997 * Do not touch the tick device, when the next expiry is either
998 * already reached or less/equal than the tick period.
1000 delta
= ktime_sub(hrtimer_get_expires(&ts
->sched_timer
), now
);
1001 if (delta
.tv64
<= tick_period
.tv64
)
1004 tick_nohz_restart(ts
, now
);
1008 static inline void tick_nohz_irq_enter(void)
1010 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1013 if (!ts
->idle_active
&& !ts
->tick_stopped
)
1016 if (ts
->idle_active
)
1017 tick_nohz_stop_idle(ts
, now
);
1018 if (ts
->tick_stopped
) {
1019 tick_nohz_update_jiffies(now
);
1020 tick_nohz_kick_tick(ts
, now
);
1026 static inline void tick_nohz_switch_to_nohz(void) { }
1027 static inline void tick_nohz_irq_enter(void) { }
1028 static inline void tick_nohz_activate(struct tick_sched
*ts
, int mode
) { }
1030 #endif /* CONFIG_NO_HZ_COMMON */
1033 * Called from irq_enter to notify about the possible interruption of idle()
1035 void tick_irq_enter(void)
1037 tick_check_oneshot_broadcast_this_cpu();
1038 tick_nohz_irq_enter();
1042 * High resolution timer specific code
1044 #ifdef CONFIG_HIGH_RES_TIMERS
1046 * We rearm the timer until we get disabled by the idle code.
1047 * Called with interrupts disabled.
1049 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
1051 struct tick_sched
*ts
=
1052 container_of(timer
, struct tick_sched
, sched_timer
);
1053 struct pt_regs
*regs
= get_irq_regs();
1054 ktime_t now
= ktime_get();
1056 tick_sched_do_timer(now
);
1059 * Do not call, when we are not in irq context and have
1060 * no valid regs pointer
1063 tick_sched_handle(ts
, regs
);
1065 /* No need to reprogram if we are in idle or full dynticks mode */
1066 if (unlikely(ts
->tick_stopped
))
1067 return HRTIMER_NORESTART
;
1069 hrtimer_forward(timer
, now
, tick_period
);
1071 return HRTIMER_RESTART
;
1074 static int sched_skew_tick
;
1076 static int __init
skew_tick(char *str
)
1078 get_option(&str
, &sched_skew_tick
);
1082 early_param("skew_tick", skew_tick
);
1085 * tick_setup_sched_timer - setup the tick emulation timer
1087 void tick_setup_sched_timer(void)
1089 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1090 ktime_t now
= ktime_get();
1093 * Emulate tick processing via per-CPU hrtimers:
1095 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
1096 ts
->sched_timer
.function
= tick_sched_timer
;
1098 /* Get the next period (per cpu) */
1099 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
1101 /* Offset the tick to avert jiffies_lock contention. */
1102 if (sched_skew_tick
) {
1103 u64 offset
= ktime_to_ns(tick_period
) >> 1;
1104 do_div(offset
, num_possible_cpus());
1105 offset
*= smp_processor_id();
1106 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
1109 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1110 hrtimer_start_expires(&ts
->sched_timer
, HRTIMER_MODE_ABS_PINNED
);
1111 tick_nohz_activate(ts
, NOHZ_MODE_HIGHRES
);
1113 #endif /* HIGH_RES_TIMERS */
1115 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1116 void tick_cancel_sched_timer(int cpu
)
1118 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1120 # ifdef CONFIG_HIGH_RES_TIMERS
1121 if (ts
->sched_timer
.base
)
1122 hrtimer_cancel(&ts
->sched_timer
);
1125 memset(ts
, 0, sizeof(*ts
));
1130 * Async notification about clocksource changes
1132 void tick_clock_notify(void)
1136 for_each_possible_cpu(cpu
)
1137 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
1141 * Async notification about clock event changes
1143 void tick_oneshot_notify(void)
1145 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1147 set_bit(0, &ts
->check_clocks
);
1151 * Check, if a change happened, which makes oneshot possible.
1153 * Called cyclic from the hrtimer softirq (driven by the timer
1154 * softirq) allow_nohz signals, that we can switch into low-res nohz
1155 * mode, because high resolution timers are disabled (either compile
1156 * or runtime). Called with interrupts disabled.
1158 int tick_check_oneshot_change(int allow_nohz
)
1160 struct tick_sched
*ts
= this_cpu_ptr(&tick_cpu_sched
);
1162 if (!test_and_clear_bit(0, &ts
->check_clocks
))
1165 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
1168 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1174 tick_nohz_switch_to_nohz();