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 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 tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
);
203 * Re-evaluate the need for the tick on the current CPU
204 * and restart it if necessary.
206 void __tick_nohz_full_check(void)
208 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
210 if (tick_nohz_full_cpu(smp_processor_id())) {
211 if (ts
->tick_stopped
&& !is_idle_task(current
)) {
212 if (!can_stop_full_tick())
213 tick_nohz_restart_sched_tick(ts
, ktime_get());
218 static void nohz_full_kick_work_func(struct irq_work
*work
)
220 __tick_nohz_full_check();
223 static DEFINE_PER_CPU(struct irq_work
, nohz_full_kick_work
) = {
224 .func
= nohz_full_kick_work_func
,
228 * Kick the CPU if it's full dynticks in order to force it to
229 * re-evaluate its dependency on the tick and restart it if necessary.
231 void tick_nohz_full_kick_cpu(int cpu
)
233 if (!tick_nohz_full_cpu(cpu
))
236 irq_work_queue_on(&per_cpu(nohz_full_kick_work
, cpu
), cpu
);
239 static void nohz_full_kick_ipi(void *info
)
241 __tick_nohz_full_check();
245 * Kick all full dynticks CPUs in order to force these to re-evaluate
246 * their dependency on the tick and restart it if necessary.
248 void tick_nohz_full_kick_all(void)
250 if (!tick_nohz_full_running
)
254 smp_call_function_many(tick_nohz_full_mask
,
255 nohz_full_kick_ipi
, NULL
, false);
256 tick_nohz_full_kick();
261 * Re-evaluate the need for the tick as we switch the current task.
262 * It might need the tick due to per task/process properties:
263 * perf events, posix cpu timers, ...
265 void __tick_nohz_task_switch(struct task_struct
*tsk
)
269 local_irq_save(flags
);
271 if (!tick_nohz_full_cpu(smp_processor_id()))
274 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
275 tick_nohz_full_kick();
278 local_irq_restore(flags
);
281 /* Parse the boot-time nohz CPU list from the kernel parameters. */
282 static int __init
tick_nohz_full_setup(char *str
)
286 alloc_bootmem_cpumask_var(&tick_nohz_full_mask
);
287 alloc_bootmem_cpumask_var(&housekeeping_mask
);
288 if (cpulist_parse(str
, tick_nohz_full_mask
) < 0) {
289 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
293 cpu
= smp_processor_id();
294 if (cpumask_test_cpu(cpu
, tick_nohz_full_mask
)) {
295 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu
);
296 cpumask_clear_cpu(cpu
, tick_nohz_full_mask
);
298 cpumask_andnot(housekeeping_mask
,
299 cpu_possible_mask
, tick_nohz_full_mask
);
300 tick_nohz_full_running
= true;
304 __setup("nohz_full=", tick_nohz_full_setup
);
306 static int tick_nohz_cpu_down_callback(struct notifier_block
*nfb
,
307 unsigned long action
,
310 unsigned int cpu
= (unsigned long)hcpu
;
312 switch (action
& ~CPU_TASKS_FROZEN
) {
313 case CPU_DOWN_PREPARE
:
315 * If we handle the timekeeping duty for full dynticks CPUs,
316 * we can't safely shutdown that CPU.
318 if (tick_nohz_full_running
&& tick_do_timer_cpu
== cpu
)
326 * Worst case string length in chunks of CPU range seems 2 steps
327 * separations: 0,2,4,6,...
328 * This is NR_CPUS + sizeof('\0')
330 static char __initdata nohz_full_buf
[NR_CPUS
+ 1];
332 static int tick_nohz_init_all(void)
336 #ifdef CONFIG_NO_HZ_FULL_ALL
337 if (!alloc_cpumask_var(&tick_nohz_full_mask
, GFP_KERNEL
)) {
338 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
341 if (!alloc_cpumask_var(&housekeeping_mask
, GFP_KERNEL
)) {
342 pr_err("NO_HZ: Can't allocate not-full dynticks cpumask\n");
346 cpumask_setall(tick_nohz_full_mask
);
347 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask
);
348 cpumask_clear(housekeeping_mask
);
349 cpumask_set_cpu(smp_processor_id(), housekeeping_mask
);
350 tick_nohz_full_running
= true;
355 void __init
tick_nohz_init(void)
359 if (!tick_nohz_full_running
) {
360 if (tick_nohz_init_all() < 0)
364 for_each_cpu(cpu
, tick_nohz_full_mask
)
365 context_tracking_cpu_set(cpu
);
367 cpu_notifier(tick_nohz_cpu_down_callback
, 0);
368 cpulist_scnprintf(nohz_full_buf
, sizeof(nohz_full_buf
), tick_nohz_full_mask
);
369 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf
);
374 * NOHZ - aka dynamic tick functionality
376 #ifdef CONFIG_NO_HZ_COMMON
380 static int tick_nohz_enabled __read_mostly
= 1;
381 int tick_nohz_active __read_mostly
;
383 * Enable / Disable tickless mode
385 static int __init
setup_tick_nohz(char *str
)
387 if (!strcmp(str
, "off"))
388 tick_nohz_enabled
= 0;
389 else if (!strcmp(str
, "on"))
390 tick_nohz_enabled
= 1;
396 __setup("nohz=", setup_tick_nohz
);
399 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
401 * Called from interrupt entry when the CPU was idle
403 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
404 * must be updated. Otherwise an interrupt handler could use a stale jiffy
405 * value. We do this unconditionally on any cpu, as we don't know whether the
406 * cpu, which has the update task assigned is in a long sleep.
408 static void tick_nohz_update_jiffies(ktime_t now
)
412 __this_cpu_write(tick_cpu_sched
.idle_waketime
, now
);
414 local_irq_save(flags
);
415 tick_do_update_jiffies64(now
);
416 local_irq_restore(flags
);
418 touch_softlockup_watchdog();
422 * Updates the per cpu time idle statistics counters
425 update_ts_time_stats(int cpu
, struct tick_sched
*ts
, ktime_t now
, u64
*last_update_time
)
429 if (ts
->idle_active
) {
430 delta
= ktime_sub(now
, ts
->idle_entrytime
);
431 if (nr_iowait_cpu(cpu
) > 0)
432 ts
->iowait_sleeptime
= ktime_add(ts
->iowait_sleeptime
, delta
);
434 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
435 ts
->idle_entrytime
= now
;
438 if (last_update_time
)
439 *last_update_time
= ktime_to_us(now
);
443 static void tick_nohz_stop_idle(struct tick_sched
*ts
, ktime_t now
)
445 update_ts_time_stats(smp_processor_id(), ts
, now
, NULL
);
448 sched_clock_idle_wakeup_event(0);
451 static ktime_t
tick_nohz_start_idle(struct tick_sched
*ts
)
453 ktime_t now
= ktime_get();
455 ts
->idle_entrytime
= now
;
457 sched_clock_idle_sleep_event();
462 * get_cpu_idle_time_us - get the total idle time of a cpu
463 * @cpu: CPU number to query
464 * @last_update_time: variable to store update time in. Do not update
467 * Return the cummulative idle time (since boot) for a given
468 * CPU, in microseconds.
470 * This time is measured via accounting rather than sampling,
471 * and is as accurate as ktime_get() is.
473 * This function returns -1 if NOHZ is not enabled.
475 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
477 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
480 if (!tick_nohz_active
)
484 if (last_update_time
) {
485 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
486 idle
= ts
->idle_sleeptime
;
488 if (ts
->idle_active
&& !nr_iowait_cpu(cpu
)) {
489 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
491 idle
= ktime_add(ts
->idle_sleeptime
, delta
);
493 idle
= ts
->idle_sleeptime
;
497 return ktime_to_us(idle
);
500 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
503 * get_cpu_iowait_time_us - get the total iowait time of a cpu
504 * @cpu: CPU number to query
505 * @last_update_time: variable to store update time in. Do not update
508 * Return the cummulative iowait time (since boot) for a given
509 * CPU, in microseconds.
511 * This time is measured via accounting rather than sampling,
512 * and is as accurate as ktime_get() is.
514 * This function returns -1 if NOHZ is not enabled.
516 u64
get_cpu_iowait_time_us(int cpu
, u64
*last_update_time
)
518 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
521 if (!tick_nohz_active
)
525 if (last_update_time
) {
526 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
527 iowait
= ts
->iowait_sleeptime
;
529 if (ts
->idle_active
&& nr_iowait_cpu(cpu
) > 0) {
530 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
532 iowait
= ktime_add(ts
->iowait_sleeptime
, delta
);
534 iowait
= ts
->iowait_sleeptime
;
538 return ktime_to_us(iowait
);
540 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us
);
542 static ktime_t
tick_nohz_stop_sched_tick(struct tick_sched
*ts
,
543 ktime_t now
, int cpu
)
545 unsigned long seq
, last_jiffies
, next_jiffies
, delta_jiffies
;
546 ktime_t last_update
, expires
, ret
= { .tv64
= 0 };
547 unsigned long rcu_delta_jiffies
;
548 struct clock_event_device
*dev
= __get_cpu_var(tick_cpu_device
).evtdev
;
551 time_delta
= timekeeping_max_deferment();
553 /* Read jiffies and the time when jiffies were updated last */
555 seq
= read_seqbegin(&jiffies_lock
);
556 last_update
= last_jiffies_update
;
557 last_jiffies
= jiffies
;
558 } while (read_seqretry(&jiffies_lock
, seq
));
560 if (rcu_needs_cpu(cpu
, &rcu_delta_jiffies
) ||
561 arch_needs_cpu(cpu
) || irq_work_needs_cpu()) {
562 next_jiffies
= last_jiffies
+ 1;
565 /* Get the next timer wheel timer */
566 next_jiffies
= get_next_timer_interrupt(last_jiffies
);
567 delta_jiffies
= next_jiffies
- last_jiffies
;
568 if (rcu_delta_jiffies
< delta_jiffies
) {
569 next_jiffies
= last_jiffies
+ rcu_delta_jiffies
;
570 delta_jiffies
= rcu_delta_jiffies
;
575 * Do not stop the tick, if we are only one off (or less)
576 * or if the cpu is required for RCU:
578 if (!ts
->tick_stopped
&& delta_jiffies
<= 1)
581 /* Schedule the tick, if we are at least one jiffie off */
582 if ((long)delta_jiffies
>= 1) {
585 * If this cpu is the one which updates jiffies, then
586 * give up the assignment and let it be taken by the
587 * cpu which runs the tick timer next, which might be
588 * this cpu as well. If we don't drop this here the
589 * jiffies might be stale and do_timer() never
590 * invoked. Keep track of the fact that it was the one
591 * which had the do_timer() duty last. If this cpu is
592 * the one which had the do_timer() duty last, we
593 * limit the sleep time to the timekeeping
594 * max_deferement value which we retrieved
595 * above. Otherwise we can sleep as long as we want.
597 if (cpu
== tick_do_timer_cpu
) {
598 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
599 ts
->do_timer_last
= 1;
600 } else if (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
) {
601 time_delta
= KTIME_MAX
;
602 ts
->do_timer_last
= 0;
603 } else if (!ts
->do_timer_last
) {
604 time_delta
= KTIME_MAX
;
607 #ifdef CONFIG_NO_HZ_FULL
609 time_delta
= min(time_delta
,
610 scheduler_tick_max_deferment());
615 * calculate the expiry time for the next timer wheel
616 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
617 * that there is no timer pending or at least extremely
618 * far into the future (12 days for HZ=1000). In this
619 * case we set the expiry to the end of time.
621 if (likely(delta_jiffies
< NEXT_TIMER_MAX_DELTA
)) {
623 * Calculate the time delta for the next timer event.
624 * If the time delta exceeds the maximum time delta
625 * permitted by the current clocksource then adjust
626 * the time delta accordingly to ensure the
627 * clocksource does not wrap.
629 time_delta
= min_t(u64
, time_delta
,
630 tick_period
.tv64
* delta_jiffies
);
633 if (time_delta
< KTIME_MAX
)
634 expires
= ktime_add_ns(last_update
, time_delta
);
636 expires
.tv64
= KTIME_MAX
;
638 /* Skip reprogram of event if its not changed */
639 if (ts
->tick_stopped
&& ktime_equal(expires
, dev
->next_event
))
645 * nohz_stop_sched_tick can be called several times before
646 * the nohz_restart_sched_tick is called. This happens when
647 * interrupts arrive which do not cause a reschedule. In the
648 * first call we save the current tick time, so we can restart
649 * the scheduler tick in nohz_restart_sched_tick.
651 if (!ts
->tick_stopped
) {
652 nohz_balance_enter_idle(cpu
);
653 calc_load_enter_idle();
655 ts
->last_tick
= hrtimer_get_expires(&ts
->sched_timer
);
656 ts
->tick_stopped
= 1;
657 trace_tick_stop(1, " ");
661 * If the expiration time == KTIME_MAX, then
662 * in this case we simply stop the tick timer.
664 if (unlikely(expires
.tv64
== KTIME_MAX
)) {
665 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
666 hrtimer_cancel(&ts
->sched_timer
);
670 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
671 hrtimer_start(&ts
->sched_timer
, expires
,
672 HRTIMER_MODE_ABS_PINNED
);
673 /* Check, if the timer was already in the past */
674 if (hrtimer_active(&ts
->sched_timer
))
676 } else if (!tick_program_event(expires
, 0))
679 * We are past the event already. So we crossed a
680 * jiffie boundary. Update jiffies and raise the
683 tick_do_update_jiffies64(ktime_get());
685 raise_softirq_irqoff(TIMER_SOFTIRQ
);
687 ts
->next_jiffies
= next_jiffies
;
688 ts
->last_jiffies
= last_jiffies
;
689 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
694 static void tick_nohz_full_stop_tick(struct tick_sched
*ts
)
696 #ifdef CONFIG_NO_HZ_FULL
697 int cpu
= smp_processor_id();
699 if (!tick_nohz_full_cpu(cpu
) || is_idle_task(current
))
702 if (!ts
->tick_stopped
&& ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)
705 if (!can_stop_full_tick())
708 tick_nohz_stop_sched_tick(ts
, ktime_get(), cpu
);
712 static bool can_stop_idle_tick(int cpu
, struct tick_sched
*ts
)
715 * If this cpu is offline and it is the one which updates
716 * jiffies, then give up the assignment and let it be taken by
717 * the cpu which runs the tick timer next. If we don't drop
718 * this here the jiffies might be stale and do_timer() never
721 if (unlikely(!cpu_online(cpu
))) {
722 if (cpu
== tick_do_timer_cpu
)
723 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
727 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)) {
728 ts
->sleep_length
= (ktime_t
) { .tv64
= NSEC_PER_SEC
/HZ
};
735 if (unlikely(local_softirq_pending() && cpu_online(cpu
))) {
736 static int ratelimit
;
738 if (ratelimit
< 10 &&
739 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK
)) {
740 pr_warn("NOHZ: local_softirq_pending %02x\n",
741 (unsigned int) local_softirq_pending());
747 if (tick_nohz_full_enabled()) {
749 * Keep the tick alive to guarantee timekeeping progression
750 * if there are full dynticks CPUs around
752 if (tick_do_timer_cpu
== cpu
)
755 * Boot safety: make sure the timekeeping duty has been
756 * assigned before entering dyntick-idle mode,
758 if (tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)
765 static void __tick_nohz_idle_enter(struct tick_sched
*ts
)
767 ktime_t now
, expires
;
768 int cpu
= smp_processor_id();
770 now
= tick_nohz_start_idle(ts
);
772 if (can_stop_idle_tick(cpu
, ts
)) {
773 int was_stopped
= ts
->tick_stopped
;
777 expires
= tick_nohz_stop_sched_tick(ts
, now
, cpu
);
778 if (expires
.tv64
> 0LL) {
780 ts
->idle_expires
= expires
;
783 if (!was_stopped
&& ts
->tick_stopped
)
784 ts
->idle_jiffies
= ts
->last_jiffies
;
789 * tick_nohz_idle_enter - stop the idle tick from the idle task
791 * When the next event is more than a tick into the future, stop the idle tick
792 * Called when we start the idle loop.
794 * The arch is responsible of calling:
796 * - rcu_idle_enter() after its last use of RCU before the CPU is put
798 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
800 void tick_nohz_idle_enter(void)
802 struct tick_sched
*ts
;
804 WARN_ON_ONCE(irqs_disabled());
807 * Update the idle state in the scheduler domain hierarchy
808 * when tick_nohz_stop_sched_tick() is called from the idle loop.
809 * State will be updated to busy during the first busy tick after
812 set_cpu_sd_state_idle();
816 ts
= &__get_cpu_var(tick_cpu_sched
);
818 __tick_nohz_idle_enter(ts
);
822 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter
);
825 * tick_nohz_irq_exit - update next tick event from interrupt exit
827 * When an interrupt fires while we are idle and it doesn't cause
828 * a reschedule, it may still add, modify or delete a timer, enqueue
829 * an RCU callback, etc...
830 * So we need to re-calculate and reprogram the next tick event.
832 void tick_nohz_irq_exit(void)
834 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
837 __tick_nohz_idle_enter(ts
);
839 tick_nohz_full_stop_tick(ts
);
843 * tick_nohz_get_sleep_length - return the length of the current sleep
845 * Called from power state control code with interrupts disabled
847 ktime_t
tick_nohz_get_sleep_length(void)
849 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
851 return ts
->sleep_length
;
854 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
856 hrtimer_cancel(&ts
->sched_timer
);
857 hrtimer_set_expires(&ts
->sched_timer
, ts
->last_tick
);
860 /* Forward the time to expire in the future */
861 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
863 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
864 hrtimer_start_expires(&ts
->sched_timer
,
865 HRTIMER_MODE_ABS_PINNED
);
866 /* Check, if the timer was already in the past */
867 if (hrtimer_active(&ts
->sched_timer
))
870 if (!tick_program_event(
871 hrtimer_get_expires(&ts
->sched_timer
), 0))
874 /* Reread time and update jiffies */
876 tick_do_update_jiffies64(now
);
880 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
)
882 /* Update jiffies first */
883 tick_do_update_jiffies64(now
);
884 update_cpu_load_nohz();
886 calc_load_exit_idle();
887 touch_softlockup_watchdog();
889 * Cancel the scheduled timer and restore the tick
891 ts
->tick_stopped
= 0;
892 ts
->idle_exittime
= now
;
894 tick_nohz_restart(ts
, now
);
897 static void tick_nohz_account_idle_ticks(struct tick_sched
*ts
)
899 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
902 if (vtime_accounting_enabled())
905 * We stopped the tick in idle. Update process times would miss the
906 * time we slept as update_process_times does only a 1 tick
907 * accounting. Enforce that this is accounted to idle !
909 ticks
= jiffies
- ts
->idle_jiffies
;
911 * We might be one off. Do not randomly account a huge number of ticks!
913 if (ticks
&& ticks
< LONG_MAX
)
914 account_idle_ticks(ticks
);
919 * tick_nohz_idle_exit - restart the idle tick from the idle task
921 * Restart the idle tick when the CPU is woken up from idle
922 * This also exit the RCU extended quiescent state. The CPU
923 * can use RCU again after this function is called.
925 void tick_nohz_idle_exit(void)
927 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
932 WARN_ON_ONCE(!ts
->inidle
);
936 if (ts
->idle_active
|| ts
->tick_stopped
)
940 tick_nohz_stop_idle(ts
, now
);
942 if (ts
->tick_stopped
) {
943 tick_nohz_restart_sched_tick(ts
, now
);
944 tick_nohz_account_idle_ticks(ts
);
949 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit
);
951 static int tick_nohz_reprogram(struct tick_sched
*ts
, ktime_t now
)
953 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
954 return tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 0);
958 * The nohz low res interrupt handler
960 static void tick_nohz_handler(struct clock_event_device
*dev
)
962 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
963 struct pt_regs
*regs
= get_irq_regs();
964 ktime_t now
= ktime_get();
966 dev
->next_event
.tv64
= KTIME_MAX
;
968 tick_sched_do_timer(now
);
969 tick_sched_handle(ts
, regs
);
971 while (tick_nohz_reprogram(ts
, now
)) {
973 tick_do_update_jiffies64(now
);
978 * tick_nohz_switch_to_nohz - switch to nohz mode
980 static void tick_nohz_switch_to_nohz(void)
982 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
985 if (!tick_nohz_enabled
)
989 if (tick_switch_to_oneshot(tick_nohz_handler
)) {
993 tick_nohz_active
= 1;
994 ts
->nohz_mode
= NOHZ_MODE_LOWRES
;
997 * Recycle the hrtimer in ts, so we can share the
998 * hrtimer_forward with the highres code.
1000 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
1001 /* Get the next period */
1002 next
= tick_init_jiffy_update();
1005 hrtimer_set_expires(&ts
->sched_timer
, next
);
1006 if (!tick_program_event(next
, 0))
1008 next
= ktime_add(next
, tick_period
);
1014 * When NOHZ is enabled and the tick is stopped, we need to kick the
1015 * tick timer from irq_enter() so that the jiffies update is kept
1016 * alive during long running softirqs. That's ugly as hell, but
1017 * correctness is key even if we need to fix the offending softirq in
1020 * Note, this is different to tick_nohz_restart. We just kick the
1021 * timer and do not touch the other magic bits which need to be done
1022 * when idle is left.
1024 static void tick_nohz_kick_tick(struct tick_sched
*ts
, ktime_t now
)
1027 /* Switch back to 2.6.27 behaviour */
1031 * Do not touch the tick device, when the next expiry is either
1032 * already reached or less/equal than the tick period.
1034 delta
= ktime_sub(hrtimer_get_expires(&ts
->sched_timer
), now
);
1035 if (delta
.tv64
<= tick_period
.tv64
)
1038 tick_nohz_restart(ts
, now
);
1042 static inline void tick_nohz_irq_enter(void)
1044 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
1047 if (!ts
->idle_active
&& !ts
->tick_stopped
)
1050 if (ts
->idle_active
)
1051 tick_nohz_stop_idle(ts
, now
);
1052 if (ts
->tick_stopped
) {
1053 tick_nohz_update_jiffies(now
);
1054 tick_nohz_kick_tick(ts
, now
);
1060 static inline void tick_nohz_switch_to_nohz(void) { }
1061 static inline void tick_nohz_irq_enter(void) { }
1063 #endif /* CONFIG_NO_HZ_COMMON */
1066 * Called from irq_enter to notify about the possible interruption of idle()
1068 void tick_irq_enter(void)
1070 tick_check_oneshot_broadcast_this_cpu();
1071 tick_nohz_irq_enter();
1075 * High resolution timer specific code
1077 #ifdef CONFIG_HIGH_RES_TIMERS
1079 * We rearm the timer until we get disabled by the idle code.
1080 * Called with interrupts disabled.
1082 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
1084 struct tick_sched
*ts
=
1085 container_of(timer
, struct tick_sched
, sched_timer
);
1086 struct pt_regs
*regs
= get_irq_regs();
1087 ktime_t now
= ktime_get();
1089 tick_sched_do_timer(now
);
1092 * Do not call, when we are not in irq context and have
1093 * no valid regs pointer
1096 tick_sched_handle(ts
, regs
);
1098 hrtimer_forward(timer
, now
, tick_period
);
1100 return HRTIMER_RESTART
;
1103 static int sched_skew_tick
;
1105 static int __init
skew_tick(char *str
)
1107 get_option(&str
, &sched_skew_tick
);
1111 early_param("skew_tick", skew_tick
);
1114 * tick_setup_sched_timer - setup the tick emulation timer
1116 void tick_setup_sched_timer(void)
1118 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
1119 ktime_t now
= ktime_get();
1122 * Emulate tick processing via per-CPU hrtimers:
1124 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
1125 ts
->sched_timer
.function
= tick_sched_timer
;
1127 /* Get the next period (per cpu) */
1128 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
1130 /* Offset the tick to avert jiffies_lock contention. */
1131 if (sched_skew_tick
) {
1132 u64 offset
= ktime_to_ns(tick_period
) >> 1;
1133 do_div(offset
, num_possible_cpus());
1134 offset
*= smp_processor_id();
1135 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
1139 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1140 hrtimer_start_expires(&ts
->sched_timer
,
1141 HRTIMER_MODE_ABS_PINNED
);
1142 /* Check, if the timer was already in the past */
1143 if (hrtimer_active(&ts
->sched_timer
))
1148 #ifdef CONFIG_NO_HZ_COMMON
1149 if (tick_nohz_enabled
) {
1150 ts
->nohz_mode
= NOHZ_MODE_HIGHRES
;
1151 tick_nohz_active
= 1;
1155 #endif /* HIGH_RES_TIMERS */
1157 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1158 void tick_cancel_sched_timer(int cpu
)
1160 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1162 # ifdef CONFIG_HIGH_RES_TIMERS
1163 if (ts
->sched_timer
.base
)
1164 hrtimer_cancel(&ts
->sched_timer
);
1167 memset(ts
, 0, sizeof(*ts
));
1172 * Async notification about clocksource changes
1174 void tick_clock_notify(void)
1178 for_each_possible_cpu(cpu
)
1179 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
1183 * Async notification about clock event changes
1185 void tick_oneshot_notify(void)
1187 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
1189 set_bit(0, &ts
->check_clocks
);
1193 * Check, if a change happened, which makes oneshot possible.
1195 * Called cyclic from the hrtimer softirq (driven by the timer
1196 * softirq) allow_nohz signals, that we can switch into low-res nohz
1197 * mode, because high resolution timers are disabled (either compile
1200 int tick_check_oneshot_change(int allow_nohz
)
1202 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
1204 if (!test_and_clear_bit(0, &ts
->check_clocks
))
1207 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
1210 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1216 tick_nohz_switch_to_nohz();