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>
27 #include <asm/irq_regs.h>
29 #include "tick-internal.h"
31 #include <trace/events/timer.h>
34 * Per cpu nohz control structure
36 DEFINE_PER_CPU(struct tick_sched
, tick_cpu_sched
);
39 * The time, when the last jiffy update happened. Protected by jiffies_lock.
41 static ktime_t last_jiffies_update
;
43 struct tick_sched
*tick_get_tick_sched(int cpu
)
45 return &per_cpu(tick_cpu_sched
, cpu
);
49 * Must be called with interrupts disabled !
51 static void tick_do_update_jiffies64(ktime_t now
)
53 unsigned long ticks
= 0;
57 * Do a quick check without holding jiffies_lock:
59 delta
= ktime_sub(now
, last_jiffies_update
);
60 if (delta
.tv64
< tick_period
.tv64
)
63 /* Reevalute with jiffies_lock held */
64 write_seqlock(&jiffies_lock
);
66 delta
= ktime_sub(now
, last_jiffies_update
);
67 if (delta
.tv64
>= tick_period
.tv64
) {
69 delta
= ktime_sub(delta
, tick_period
);
70 last_jiffies_update
= ktime_add(last_jiffies_update
,
73 /* Slow path for long timeouts */
74 if (unlikely(delta
.tv64
>= tick_period
.tv64
)) {
75 s64 incr
= ktime_to_ns(tick_period
);
77 ticks
= ktime_divns(delta
, incr
);
79 last_jiffies_update
= ktime_add_ns(last_jiffies_update
,
84 /* Keep the tick_next_period variable up to date */
85 tick_next_period
= ktime_add(last_jiffies_update
, tick_period
);
87 write_sequnlock(&jiffies_lock
);
91 * Initialize and return retrieve the jiffies update.
93 static ktime_t
tick_init_jiffy_update(void)
97 write_seqlock(&jiffies_lock
);
98 /* Did we start the jiffies update yet ? */
99 if (last_jiffies_update
.tv64
== 0)
100 last_jiffies_update
= tick_next_period
;
101 period
= last_jiffies_update
;
102 write_sequnlock(&jiffies_lock
);
107 static void tick_sched_do_timer(ktime_t now
)
109 int cpu
= smp_processor_id();
111 #ifdef CONFIG_NO_HZ_COMMON
113 * Check if the do_timer duty was dropped. We don't care about
114 * concurrency: This happens only when the cpu in charge went
115 * into a long sleep. If two cpus happen to assign themself to
116 * this duty, then the jiffies update is still serialized by
119 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)
120 && !tick_nohz_full_cpu(cpu
))
121 tick_do_timer_cpu
= cpu
;
124 /* Check, if the jiffies need an update */
125 if (tick_do_timer_cpu
== cpu
)
126 tick_do_update_jiffies64(now
);
129 static void tick_sched_handle(struct tick_sched
*ts
, struct pt_regs
*regs
)
131 #ifdef CONFIG_NO_HZ_COMMON
133 * When we are idle and the tick is stopped, we have to touch
134 * the watchdog as we might not schedule for a really long
135 * time. This happens on complete idle SMP systems while
136 * waiting on the login prompt. We also increment the "start of
137 * idle" jiffy stamp so the idle accounting adjustment we do
138 * when we go busy again does not account too much ticks.
140 if (ts
->tick_stopped
) {
141 touch_softlockup_watchdog();
142 if (is_idle_task(current
))
146 update_process_times(user_mode(regs
));
147 profile_tick(CPU_PROFILING
);
150 #ifdef CONFIG_NO_HZ_FULL
151 static cpumask_var_t nohz_full_mask
;
152 bool have_nohz_full_mask
;
154 static bool can_stop_full_tick(void)
156 WARN_ON_ONCE(!irqs_disabled());
158 if (!sched_can_stop_tick()) {
159 trace_tick_stop(0, "more than 1 task in runqueue\n");
163 if (!posix_cpu_timers_can_stop_tick(current
)) {
164 trace_tick_stop(0, "posix timers running\n");
168 if (!perf_event_can_stop_tick()) {
169 trace_tick_stop(0, "perf events running\n");
173 /* sched_clock_tick() needs us? */
174 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
176 * TODO: kick full dynticks CPUs when
177 * sched_clock_stable is set.
179 if (!sched_clock_stable
) {
180 trace_tick_stop(0, "unstable sched clock\n");
182 * Don't allow the user to think they can get
183 * full NO_HZ with this machine.
185 WARN_ONCE(1, "NO_HZ FULL will not work with unstable sched clock");
193 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
);
196 * Re-evaluate the need for the tick on the current CPU
197 * and restart it if necessary.
199 void tick_nohz_full_check(void)
201 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
203 if (tick_nohz_full_cpu(smp_processor_id())) {
204 if (ts
->tick_stopped
&& !is_idle_task(current
)) {
205 if (!can_stop_full_tick())
206 tick_nohz_restart_sched_tick(ts
, ktime_get());
211 static void nohz_full_kick_work_func(struct irq_work
*work
)
213 tick_nohz_full_check();
216 static DEFINE_PER_CPU(struct irq_work
, nohz_full_kick_work
) = {
217 .func
= nohz_full_kick_work_func
,
221 * Kick the current CPU if it's full dynticks in order to force it to
222 * re-evaluate its dependency on the tick and restart it if necessary.
224 void tick_nohz_full_kick(void)
226 if (tick_nohz_full_cpu(smp_processor_id()))
227 irq_work_queue(&__get_cpu_var(nohz_full_kick_work
));
230 static void nohz_full_kick_ipi(void *info
)
232 tick_nohz_full_check();
236 * Kick all full dynticks CPUs in order to force these to re-evaluate
237 * their dependency on the tick and restart it if necessary.
239 void tick_nohz_full_kick_all(void)
241 if (!have_nohz_full_mask
)
245 smp_call_function_many(nohz_full_mask
,
246 nohz_full_kick_ipi
, NULL
, false);
251 * Re-evaluate the need for the tick as we switch the current task.
252 * It might need the tick due to per task/process properties:
253 * perf events, posix cpu timers, ...
255 void tick_nohz_task_switch(struct task_struct
*tsk
)
259 local_irq_save(flags
);
261 if (!tick_nohz_full_cpu(smp_processor_id()))
264 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
265 tick_nohz_full_kick();
268 local_irq_restore(flags
);
271 int tick_nohz_full_cpu(int cpu
)
273 if (!have_nohz_full_mask
)
276 return cpumask_test_cpu(cpu
, nohz_full_mask
);
279 /* Parse the boot-time nohz CPU list from the kernel parameters. */
280 static int __init
tick_nohz_full_setup(char *str
)
284 alloc_bootmem_cpumask_var(&nohz_full_mask
);
285 if (cpulist_parse(str
, nohz_full_mask
) < 0) {
286 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
290 cpu
= smp_processor_id();
291 if (cpumask_test_cpu(cpu
, nohz_full_mask
)) {
292 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu
);
293 cpumask_clear_cpu(cpu
, nohz_full_mask
);
295 have_nohz_full_mask
= true;
299 __setup("nohz_full=", tick_nohz_full_setup
);
301 static int tick_nohz_cpu_down_callback(struct notifier_block
*nfb
,
302 unsigned long action
,
305 unsigned int cpu
= (unsigned long)hcpu
;
307 switch (action
& ~CPU_TASKS_FROZEN
) {
308 case CPU_DOWN_PREPARE
:
310 * If we handle the timekeeping duty for full dynticks CPUs,
311 * we can't safely shutdown that CPU.
313 if (have_nohz_full_mask
&& tick_do_timer_cpu
== cpu
)
321 * Worst case string length in chunks of CPU range seems 2 steps
322 * separations: 0,2,4,6,...
323 * This is NR_CPUS + sizeof('\0')
325 static char __initdata nohz_full_buf
[NR_CPUS
+ 1];
327 static int tick_nohz_init_all(void)
331 #ifdef CONFIG_NO_HZ_FULL_ALL
332 if (!alloc_cpumask_var(&nohz_full_mask
, GFP_KERNEL
)) {
333 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
337 cpumask_setall(nohz_full_mask
);
338 cpumask_clear_cpu(smp_processor_id(), nohz_full_mask
);
339 have_nohz_full_mask
= true;
344 void __init
tick_nohz_init(void)
348 if (!have_nohz_full_mask
) {
349 if (tick_nohz_init_all() < 0)
353 cpu_notifier(tick_nohz_cpu_down_callback
, 0);
354 cpulist_scnprintf(nohz_full_buf
, sizeof(nohz_full_buf
), nohz_full_mask
);
355 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf
);
358 #define have_nohz_full_mask (0)
362 * NOHZ - aka dynamic tick functionality
364 #ifdef CONFIG_NO_HZ_COMMON
368 int tick_nohz_enabled __read_mostly
= 1;
371 * Enable / Disable tickless mode
373 static int __init
setup_tick_nohz(char *str
)
375 if (!strcmp(str
, "off"))
376 tick_nohz_enabled
= 0;
377 else if (!strcmp(str
, "on"))
378 tick_nohz_enabled
= 1;
384 __setup("nohz=", setup_tick_nohz
);
387 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
389 * Called from interrupt entry when the CPU was idle
391 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
392 * must be updated. Otherwise an interrupt handler could use a stale jiffy
393 * value. We do this unconditionally on any cpu, as we don't know whether the
394 * cpu, which has the update task assigned is in a long sleep.
396 static void tick_nohz_update_jiffies(ktime_t now
)
398 int cpu
= smp_processor_id();
399 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
402 ts
->idle_waketime
= now
;
404 local_irq_save(flags
);
405 tick_do_update_jiffies64(now
);
406 local_irq_restore(flags
);
408 touch_softlockup_watchdog();
412 * Updates the per cpu time idle statistics counters
415 update_ts_time_stats(int cpu
, struct tick_sched
*ts
, ktime_t now
, u64
*last_update_time
)
419 if (ts
->idle_active
) {
420 delta
= ktime_sub(now
, ts
->idle_entrytime
);
421 if (nr_iowait_cpu(cpu
) > 0)
422 ts
->iowait_sleeptime
= ktime_add(ts
->iowait_sleeptime
, delta
);
424 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
425 ts
->idle_entrytime
= now
;
428 if (last_update_time
)
429 *last_update_time
= ktime_to_us(now
);
433 static void tick_nohz_stop_idle(int cpu
, ktime_t now
)
435 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
437 update_ts_time_stats(cpu
, ts
, now
, NULL
);
440 sched_clock_idle_wakeup_event(0);
443 static ktime_t
tick_nohz_start_idle(int cpu
, struct tick_sched
*ts
)
445 ktime_t now
= ktime_get();
447 ts
->idle_entrytime
= now
;
449 sched_clock_idle_sleep_event();
454 * get_cpu_idle_time_us - get the total idle time of a cpu
455 * @cpu: CPU number to query
456 * @last_update_time: variable to store update time in. Do not update
459 * Return the cummulative idle time (since boot) for a given
460 * CPU, in microseconds.
462 * This time is measured via accounting rather than sampling,
463 * and is as accurate as ktime_get() is.
465 * This function returns -1 if NOHZ is not enabled.
467 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
469 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
472 if (!tick_nohz_enabled
)
476 if (last_update_time
) {
477 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
478 idle
= ts
->idle_sleeptime
;
480 if (ts
->idle_active
&& !nr_iowait_cpu(cpu
)) {
481 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
483 idle
= ktime_add(ts
->idle_sleeptime
, delta
);
485 idle
= ts
->idle_sleeptime
;
489 return ktime_to_us(idle
);
492 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
495 * get_cpu_iowait_time_us - get the total iowait time of a cpu
496 * @cpu: CPU number to query
497 * @last_update_time: variable to store update time in. Do not update
500 * Return the cummulative iowait time (since boot) for a given
501 * CPU, in microseconds.
503 * This time is measured via accounting rather than sampling,
504 * and is as accurate as ktime_get() is.
506 * This function returns -1 if NOHZ is not enabled.
508 u64
get_cpu_iowait_time_us(int cpu
, u64
*last_update_time
)
510 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
513 if (!tick_nohz_enabled
)
517 if (last_update_time
) {
518 update_ts_time_stats(cpu
, ts
, now
, last_update_time
);
519 iowait
= ts
->iowait_sleeptime
;
521 if (ts
->idle_active
&& nr_iowait_cpu(cpu
) > 0) {
522 ktime_t delta
= ktime_sub(now
, ts
->idle_entrytime
);
524 iowait
= ktime_add(ts
->iowait_sleeptime
, delta
);
526 iowait
= ts
->iowait_sleeptime
;
530 return ktime_to_us(iowait
);
532 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us
);
534 static ktime_t
tick_nohz_stop_sched_tick(struct tick_sched
*ts
,
535 ktime_t now
, int cpu
)
537 unsigned long seq
, last_jiffies
, next_jiffies
, delta_jiffies
;
538 ktime_t last_update
, expires
, ret
= { .tv64
= 0 };
539 unsigned long rcu_delta_jiffies
;
540 struct clock_event_device
*dev
= __get_cpu_var(tick_cpu_device
).evtdev
;
543 /* Read jiffies and the time when jiffies were updated last */
545 seq
= read_seqbegin(&jiffies_lock
);
546 last_update
= last_jiffies_update
;
547 last_jiffies
= jiffies
;
548 time_delta
= timekeeping_max_deferment();
549 } while (read_seqretry(&jiffies_lock
, seq
));
551 if (rcu_needs_cpu(cpu
, &rcu_delta_jiffies
) ||
552 arch_needs_cpu(cpu
) || irq_work_needs_cpu()) {
553 next_jiffies
= last_jiffies
+ 1;
556 /* Get the next timer wheel timer */
557 next_jiffies
= get_next_timer_interrupt(last_jiffies
);
558 delta_jiffies
= next_jiffies
- last_jiffies
;
559 if (rcu_delta_jiffies
< delta_jiffies
) {
560 next_jiffies
= last_jiffies
+ rcu_delta_jiffies
;
561 delta_jiffies
= rcu_delta_jiffies
;
566 * Do not stop the tick, if we are only one off (or less)
567 * or if the cpu is required for RCU:
569 if (!ts
->tick_stopped
&& delta_jiffies
<= 1)
572 /* Schedule the tick, if we are at least one jiffie off */
573 if ((long)delta_jiffies
>= 1) {
576 * If this cpu is the one which updates jiffies, then
577 * give up the assignment and let it be taken by the
578 * cpu which runs the tick timer next, which might be
579 * this cpu as well. If we don't drop this here the
580 * jiffies might be stale and do_timer() never
581 * invoked. Keep track of the fact that it was the one
582 * which had the do_timer() duty last. If this cpu is
583 * the one which had the do_timer() duty last, we
584 * limit the sleep time to the timekeeping
585 * max_deferement value which we retrieved
586 * above. Otherwise we can sleep as long as we want.
588 if (cpu
== tick_do_timer_cpu
) {
589 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
590 ts
->do_timer_last
= 1;
591 } else if (tick_do_timer_cpu
!= TICK_DO_TIMER_NONE
) {
592 time_delta
= KTIME_MAX
;
593 ts
->do_timer_last
= 0;
594 } else if (!ts
->do_timer_last
) {
595 time_delta
= KTIME_MAX
;
598 #ifdef CONFIG_NO_HZ_FULL
600 time_delta
= min(time_delta
,
601 scheduler_tick_max_deferment());
606 * calculate the expiry time for the next timer wheel
607 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
608 * that there is no timer pending or at least extremely
609 * far into the future (12 days for HZ=1000). In this
610 * case we set the expiry to the end of time.
612 if (likely(delta_jiffies
< NEXT_TIMER_MAX_DELTA
)) {
614 * Calculate the time delta for the next timer event.
615 * If the time delta exceeds the maximum time delta
616 * permitted by the current clocksource then adjust
617 * the time delta accordingly to ensure the
618 * clocksource does not wrap.
620 time_delta
= min_t(u64
, time_delta
,
621 tick_period
.tv64
* delta_jiffies
);
624 if (time_delta
< KTIME_MAX
)
625 expires
= ktime_add_ns(last_update
, time_delta
);
627 expires
.tv64
= KTIME_MAX
;
629 /* Skip reprogram of event if its not changed */
630 if (ts
->tick_stopped
&& ktime_equal(expires
, dev
->next_event
))
636 * nohz_stop_sched_tick can be called several times before
637 * the nohz_restart_sched_tick is called. This happens when
638 * interrupts arrive which do not cause a reschedule. In the
639 * first call we save the current tick time, so we can restart
640 * the scheduler tick in nohz_restart_sched_tick.
642 if (!ts
->tick_stopped
) {
643 nohz_balance_enter_idle(cpu
);
644 calc_load_enter_idle();
646 ts
->last_tick
= hrtimer_get_expires(&ts
->sched_timer
);
647 ts
->tick_stopped
= 1;
648 trace_tick_stop(1, " ");
652 * If the expiration time == KTIME_MAX, then
653 * in this case we simply stop the tick timer.
655 if (unlikely(expires
.tv64
== KTIME_MAX
)) {
656 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
657 hrtimer_cancel(&ts
->sched_timer
);
661 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
662 hrtimer_start(&ts
->sched_timer
, expires
,
663 HRTIMER_MODE_ABS_PINNED
);
664 /* Check, if the timer was already in the past */
665 if (hrtimer_active(&ts
->sched_timer
))
667 } else if (!tick_program_event(expires
, 0))
670 * We are past the event already. So we crossed a
671 * jiffie boundary. Update jiffies and raise the
674 tick_do_update_jiffies64(ktime_get());
676 raise_softirq_irqoff(TIMER_SOFTIRQ
);
678 ts
->next_jiffies
= next_jiffies
;
679 ts
->last_jiffies
= last_jiffies
;
680 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
685 static void tick_nohz_full_stop_tick(struct tick_sched
*ts
)
687 #ifdef CONFIG_NO_HZ_FULL
688 int cpu
= smp_processor_id();
690 if (!tick_nohz_full_cpu(cpu
) || is_idle_task(current
))
693 if (!ts
->tick_stopped
&& ts
->nohz_mode
== NOHZ_MODE_INACTIVE
)
696 if (!can_stop_full_tick())
699 tick_nohz_stop_sched_tick(ts
, ktime_get(), cpu
);
703 static bool can_stop_idle_tick(int cpu
, struct tick_sched
*ts
)
706 * If this cpu is offline and it is the one which updates
707 * jiffies, then give up the assignment and let it be taken by
708 * the cpu which runs the tick timer next. If we don't drop
709 * this here the jiffies might be stale and do_timer() never
712 if (unlikely(!cpu_online(cpu
))) {
713 if (cpu
== tick_do_timer_cpu
)
714 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
718 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
))
724 if (unlikely(local_softirq_pending() && cpu_online(cpu
))) {
725 static int ratelimit
;
727 if (ratelimit
< 10 &&
728 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK
)) {
729 pr_warn("NOHZ: local_softirq_pending %02x\n",
730 (unsigned int) local_softirq_pending());
736 if (have_nohz_full_mask
) {
738 * Keep the tick alive to guarantee timekeeping progression
739 * if there are full dynticks CPUs around
741 if (tick_do_timer_cpu
== cpu
)
744 * Boot safety: make sure the timekeeping duty has been
745 * assigned before entering dyntick-idle mode,
747 if (tick_do_timer_cpu
== TICK_DO_TIMER_NONE
)
754 static void __tick_nohz_idle_enter(struct tick_sched
*ts
)
756 ktime_t now
, expires
;
757 int cpu
= smp_processor_id();
759 now
= tick_nohz_start_idle(cpu
, ts
);
761 if (can_stop_idle_tick(cpu
, ts
)) {
762 int was_stopped
= ts
->tick_stopped
;
766 expires
= tick_nohz_stop_sched_tick(ts
, now
, cpu
);
767 if (expires
.tv64
> 0LL) {
769 ts
->idle_expires
= expires
;
772 if (!was_stopped
&& ts
->tick_stopped
)
773 ts
->idle_jiffies
= ts
->last_jiffies
;
778 * tick_nohz_idle_enter - stop the idle tick from the idle task
780 * When the next event is more than a tick into the future, stop the idle tick
781 * Called when we start the idle loop.
783 * The arch is responsible of calling:
785 * - rcu_idle_enter() after its last use of RCU before the CPU is put
787 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
789 void tick_nohz_idle_enter(void)
791 struct tick_sched
*ts
;
793 WARN_ON_ONCE(irqs_disabled());
796 * Update the idle state in the scheduler domain hierarchy
797 * when tick_nohz_stop_sched_tick() is called from the idle loop.
798 * State will be updated to busy during the first busy tick after
801 set_cpu_sd_state_idle();
805 ts
= &__get_cpu_var(tick_cpu_sched
);
807 * set ts->inidle unconditionally. even if the system did not
808 * switch to nohz mode the cpu frequency governers rely on the
809 * update of the idle time accounting in tick_nohz_start_idle().
812 __tick_nohz_idle_enter(ts
);
816 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter
);
819 * tick_nohz_irq_exit - update next tick event from interrupt exit
821 * When an interrupt fires while we are idle and it doesn't cause
822 * a reschedule, it may still add, modify or delete a timer, enqueue
823 * an RCU callback, etc...
824 * So we need to re-calculate and reprogram the next tick event.
826 void tick_nohz_irq_exit(void)
828 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
831 /* Cancel the timer because CPU already waken up from the C-states*/
832 menu_hrtimer_cancel();
833 __tick_nohz_idle_enter(ts
);
835 tick_nohz_full_stop_tick(ts
);
840 * tick_nohz_get_sleep_length - return the length of the current sleep
842 * Called from power state control code with interrupts disabled
844 ktime_t
tick_nohz_get_sleep_length(void)
846 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
848 return ts
->sleep_length
;
851 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
853 hrtimer_cancel(&ts
->sched_timer
);
854 hrtimer_set_expires(&ts
->sched_timer
, ts
->last_tick
);
857 /* Forward the time to expire in the future */
858 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
860 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
861 hrtimer_start_expires(&ts
->sched_timer
,
862 HRTIMER_MODE_ABS_PINNED
);
863 /* Check, if the timer was already in the past */
864 if (hrtimer_active(&ts
->sched_timer
))
867 if (!tick_program_event(
868 hrtimer_get_expires(&ts
->sched_timer
), 0))
871 /* Reread time and update jiffies */
873 tick_do_update_jiffies64(now
);
877 static void tick_nohz_restart_sched_tick(struct tick_sched
*ts
, ktime_t now
)
879 /* Update jiffies first */
880 tick_do_update_jiffies64(now
);
881 update_cpu_load_nohz();
883 calc_load_exit_idle();
884 touch_softlockup_watchdog();
886 * Cancel the scheduled timer and restore the tick
888 ts
->tick_stopped
= 0;
889 ts
->idle_exittime
= now
;
891 tick_nohz_restart(ts
, now
);
894 static void tick_nohz_account_idle_ticks(struct tick_sched
*ts
)
896 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
899 if (vtime_accounting_enabled())
902 * We stopped the tick in idle. Update process times would miss the
903 * time we slept as update_process_times does only a 1 tick
904 * accounting. Enforce that this is accounted to idle !
906 ticks
= jiffies
- ts
->idle_jiffies
;
908 * We might be one off. Do not randomly account a huge number of ticks!
910 if (ticks
&& ticks
< LONG_MAX
)
911 account_idle_ticks(ticks
);
916 * tick_nohz_idle_exit - restart the idle tick from the idle task
918 * Restart the idle tick when the CPU is woken up from idle
919 * This also exit the RCU extended quiescent state. The CPU
920 * can use RCU again after this function is called.
922 void tick_nohz_idle_exit(void)
924 int cpu
= smp_processor_id();
925 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
930 WARN_ON_ONCE(!ts
->inidle
);
934 /* Cancel the timer because CPU already waken up from the C-states*/
935 menu_hrtimer_cancel();
936 if (ts
->idle_active
|| ts
->tick_stopped
)
940 tick_nohz_stop_idle(cpu
, 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
)) {
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(int cpu
, ktime_t now
)
1027 /* Switch back to 2.6.27 behaviour */
1029 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1033 * Do not touch the tick device, when the next expiry is either
1034 * already reached or less/equal than the tick period.
1036 delta
= ktime_sub(hrtimer_get_expires(&ts
->sched_timer
), now
);
1037 if (delta
.tv64
<= tick_period
.tv64
)
1040 tick_nohz_restart(ts
, now
);
1044 static inline void tick_check_nohz(int cpu
)
1046 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
1049 if (!ts
->idle_active
&& !ts
->tick_stopped
)
1052 if (ts
->idle_active
)
1053 tick_nohz_stop_idle(cpu
, now
);
1054 if (ts
->tick_stopped
) {
1055 tick_nohz_update_jiffies(now
);
1056 tick_nohz_kick_tick(cpu
, now
);
1062 static inline void tick_nohz_switch_to_nohz(void) { }
1063 static inline void tick_check_nohz(int cpu
) { }
1065 #endif /* CONFIG_NO_HZ_COMMON */
1068 * Called from irq_enter to notify about the possible interruption of idle()
1070 void tick_check_idle(int cpu
)
1072 tick_check_oneshot_broadcast(cpu
);
1073 tick_check_nohz(cpu
);
1077 * High resolution timer specific code
1079 #ifdef CONFIG_HIGH_RES_TIMERS
1081 * We rearm the timer until we get disabled by the idle code.
1082 * Called with interrupts disabled.
1084 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
1086 struct tick_sched
*ts
=
1087 container_of(timer
, struct tick_sched
, sched_timer
);
1088 struct pt_regs
*regs
= get_irq_regs();
1089 ktime_t now
= ktime_get();
1091 tick_sched_do_timer(now
);
1094 * Do not call, when we are not in irq context and have
1095 * no valid regs pointer
1098 tick_sched_handle(ts
, regs
);
1100 hrtimer_forward(timer
, now
, tick_period
);
1102 return HRTIMER_RESTART
;
1105 static int sched_skew_tick
;
1107 static int __init
skew_tick(char *str
)
1109 get_option(&str
, &sched_skew_tick
);
1113 early_param("skew_tick", skew_tick
);
1116 * tick_setup_sched_timer - setup the tick emulation timer
1118 void tick_setup_sched_timer(void)
1120 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
1121 ktime_t now
= ktime_get();
1124 * Emulate tick processing via per-CPU hrtimers:
1126 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
1127 ts
->sched_timer
.function
= tick_sched_timer
;
1129 /* Get the next period (per cpu) */
1130 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
1132 /* Offset the tick to avert jiffies_lock contention. */
1133 if (sched_skew_tick
) {
1134 u64 offset
= ktime_to_ns(tick_period
) >> 1;
1135 do_div(offset
, num_possible_cpus());
1136 offset
*= smp_processor_id();
1137 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
1141 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
1142 hrtimer_start_expires(&ts
->sched_timer
,
1143 HRTIMER_MODE_ABS_PINNED
);
1144 /* Check, if the timer was already in the past */
1145 if (hrtimer_active(&ts
->sched_timer
))
1150 #ifdef CONFIG_NO_HZ_COMMON
1151 if (tick_nohz_enabled
)
1152 ts
->nohz_mode
= NOHZ_MODE_HIGHRES
;
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();