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/tick.h>
23 #include <linux/module.h>
25 #include <asm/irq_regs.h>
27 #include "tick-internal.h"
30 * Per cpu nohz control structure
32 static DEFINE_PER_CPU(struct tick_sched
, tick_cpu_sched
);
35 * The time, when the last jiffy update happened. Protected by xtime_lock.
37 static ktime_t last_jiffies_update
;
39 struct tick_sched
*tick_get_tick_sched(int cpu
)
41 return &per_cpu(tick_cpu_sched
, cpu
);
45 * Must be called with interrupts disabled !
47 static void tick_do_update_jiffies64(ktime_t now
)
49 unsigned long ticks
= 0;
53 * Do a quick check without holding xtime_lock:
55 delta
= ktime_sub(now
, last_jiffies_update
);
56 if (delta
.tv64
< tick_period
.tv64
)
59 /* Reevalute with xtime_lock held */
60 write_seqlock(&xtime_lock
);
62 delta
= ktime_sub(now
, last_jiffies_update
);
63 if (delta
.tv64
>= tick_period
.tv64
) {
65 delta
= ktime_sub(delta
, tick_period
);
66 last_jiffies_update
= ktime_add(last_jiffies_update
,
69 /* Slow path for long timeouts */
70 if (unlikely(delta
.tv64
>= tick_period
.tv64
)) {
71 s64 incr
= ktime_to_ns(tick_period
);
73 ticks
= ktime_divns(delta
, incr
);
75 last_jiffies_update
= ktime_add_ns(last_jiffies_update
,
80 /* Keep the tick_next_period variable up to date */
81 tick_next_period
= ktime_add(last_jiffies_update
, tick_period
);
83 write_sequnlock(&xtime_lock
);
87 * Initialize and return retrieve the jiffies update.
89 static ktime_t
tick_init_jiffy_update(void)
93 write_seqlock(&xtime_lock
);
94 /* Did we start the jiffies update yet ? */
95 if (last_jiffies_update
.tv64
== 0)
96 last_jiffies_update
= tick_next_period
;
97 period
= last_jiffies_update
;
98 write_sequnlock(&xtime_lock
);
103 * NOHZ - aka dynamic tick functionality
109 static int tick_nohz_enabled __read_mostly
= 1;
112 * Enable / Disable tickless mode
114 static int __init
setup_tick_nohz(char *str
)
116 if (!strcmp(str
, "off"))
117 tick_nohz_enabled
= 0;
118 else if (!strcmp(str
, "on"))
119 tick_nohz_enabled
= 1;
125 __setup("nohz=", setup_tick_nohz
);
128 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
130 * Called from interrupt entry when the CPU was idle
132 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
133 * must be updated. Otherwise an interrupt handler could use a stale jiffy
134 * value. We do this unconditionally on any cpu, as we don't know whether the
135 * cpu, which has the update task assigned is in a long sleep.
137 static void tick_nohz_update_jiffies(ktime_t now
)
139 int cpu
= smp_processor_id();
140 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
143 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
144 ts
->idle_waketime
= now
;
146 local_irq_save(flags
);
147 tick_do_update_jiffies64(now
);
148 local_irq_restore(flags
);
150 touch_softlockup_watchdog();
153 static void tick_nohz_stop_idle(int cpu
, ktime_t now
)
155 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
158 delta
= ktime_sub(now
, ts
->idle_entrytime
);
159 ts
->idle_lastupdate
= now
;
160 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
163 sched_clock_idle_wakeup_event(0);
166 static ktime_t
tick_nohz_start_idle(struct tick_sched
*ts
)
171 if (ts
->idle_active
) {
172 delta
= ktime_sub(now
, ts
->idle_entrytime
);
173 ts
->idle_lastupdate
= now
;
174 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
176 ts
->idle_entrytime
= now
;
178 sched_clock_idle_sleep_event();
182 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
184 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
186 if (!tick_nohz_enabled
)
190 *last_update_time
= ktime_to_us(ts
->idle_lastupdate
);
192 *last_update_time
= ktime_to_us(ktime_get());
194 return ktime_to_us(ts
->idle_sleeptime
);
196 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
199 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
201 * When the next event is more than a tick into the future, stop the idle tick
202 * Called either from the idle loop or from irq_exit() when an idle period was
203 * just interrupted by an interrupt which did not cause a reschedule.
205 void tick_nohz_stop_sched_tick(int inidle
)
207 unsigned long seq
, last_jiffies
, next_jiffies
, delta_jiffies
, flags
;
208 struct tick_sched
*ts
;
209 ktime_t last_update
, expires
, now
;
210 struct clock_event_device
*dev
= __get_cpu_var(tick_cpu_device
).evtdev
;
213 local_irq_save(flags
);
215 cpu
= smp_processor_id();
216 ts
= &per_cpu(tick_cpu_sched
, cpu
);
219 * Call to tick_nohz_start_idle stops the last_update_time from being
220 * updated. Thus, it must not be called in the event we are called from
221 * irq_exit() with the prior state different than idle.
223 if (!inidle
&& !ts
->inidle
)
226 now
= tick_nohz_start_idle(ts
);
229 * If this cpu is offline and it is the one which updates
230 * jiffies, then give up the assignment and let it be taken by
231 * the cpu which runs the tick timer next. If we don't drop
232 * this here the jiffies might be stale and do_timer() never
235 if (unlikely(!cpu_online(cpu
))) {
236 if (cpu
== tick_do_timer_cpu
)
237 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
240 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
))
248 if (unlikely(local_softirq_pending() && cpu_online(cpu
))) {
249 static int ratelimit
;
251 if (ratelimit
< 10) {
252 printk(KERN_ERR
"NOHZ: local_softirq_pending %02x\n",
253 local_softirq_pending());
260 /* Read jiffies and the time when jiffies were updated last */
262 seq
= read_seqbegin(&xtime_lock
);
263 last_update
= last_jiffies_update
;
264 last_jiffies
= jiffies
;
265 } while (read_seqretry(&xtime_lock
, seq
));
267 /* Get the next timer wheel timer */
268 next_jiffies
= get_next_timer_interrupt(last_jiffies
);
269 delta_jiffies
= next_jiffies
- last_jiffies
;
271 if (rcu_needs_cpu(cpu
) || printk_needs_cpu(cpu
))
274 * Do not stop the tick, if we are only one off
275 * or if the cpu is required for rcu
277 if (!ts
->tick_stopped
&& delta_jiffies
== 1)
280 /* Schedule the tick, if we are at least one jiffie off */
281 if ((long)delta_jiffies
>= 1) {
284 * calculate the expiry time for the next timer wheel
287 expires
= ktime_add_ns(last_update
, tick_period
.tv64
*
291 * If this cpu is the one which updates jiffies, then
292 * give up the assignment and let it be taken by the
293 * cpu which runs the tick timer next, which might be
294 * this cpu as well. If we don't drop this here the
295 * jiffies might be stale and do_timer() never
298 if (cpu
== tick_do_timer_cpu
)
299 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
301 if (delta_jiffies
> 1)
302 cpumask_set_cpu(cpu
, nohz_cpu_mask
);
304 /* Skip reprogram of event if its not changed */
305 if (ts
->tick_stopped
&& ktime_equal(expires
, dev
->next_event
))
309 * nohz_stop_sched_tick can be called several times before
310 * the nohz_restart_sched_tick is called. This happens when
311 * interrupts arrive which do not cause a reschedule. In the
312 * first call we save the current tick time, so we can restart
313 * the scheduler tick in nohz_restart_sched_tick.
315 if (!ts
->tick_stopped
) {
316 if (select_nohz_load_balancer(1)) {
318 * sched tick not stopped!
320 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
324 ts
->idle_tick
= hrtimer_get_expires(&ts
->sched_timer
);
325 ts
->tick_stopped
= 1;
326 ts
->idle_jiffies
= last_jiffies
;
333 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
334 * there is no timer pending or at least extremly far
335 * into the future (12 days for HZ=1000). In this case
336 * we simply stop the tick timer:
338 if (unlikely(delta_jiffies
>= NEXT_TIMER_MAX_DELTA
)) {
339 ts
->idle_expires
.tv64
= KTIME_MAX
;
340 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
341 hrtimer_cancel(&ts
->sched_timer
);
346 ts
->idle_expires
= expires
;
348 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
349 hrtimer_start(&ts
->sched_timer
, expires
,
350 HRTIMER_MODE_ABS_PINNED
);
351 /* Check, if the timer was already in the past */
352 if (hrtimer_active(&ts
->sched_timer
))
354 } else if (!tick_program_event(expires
, 0))
357 * We are past the event already. So we crossed a
358 * jiffie boundary. Update jiffies and raise the
361 tick_do_update_jiffies64(ktime_get());
362 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
364 raise_softirq_irqoff(TIMER_SOFTIRQ
);
366 ts
->next_jiffies
= next_jiffies
;
367 ts
->last_jiffies
= last_jiffies
;
368 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
370 local_irq_restore(flags
);
374 * tick_nohz_get_sleep_length - return the length of the current sleep
376 * Called from power state control code with interrupts disabled
378 ktime_t
tick_nohz_get_sleep_length(void)
380 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
382 return ts
->sleep_length
;
385 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
387 hrtimer_cancel(&ts
->sched_timer
);
388 hrtimer_set_expires(&ts
->sched_timer
, ts
->idle_tick
);
391 /* Forward the time to expire in the future */
392 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
394 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
395 hrtimer_start_expires(&ts
->sched_timer
,
396 HRTIMER_MODE_ABS_PINNED
);
397 /* Check, if the timer was already in the past */
398 if (hrtimer_active(&ts
->sched_timer
))
401 if (!tick_program_event(
402 hrtimer_get_expires(&ts
->sched_timer
), 0))
405 /* Update jiffies and reread time */
406 tick_do_update_jiffies64(now
);
412 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
414 * Restart the idle tick when the CPU is woken up from idle
416 void tick_nohz_restart_sched_tick(void)
418 int cpu
= smp_processor_id();
419 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
420 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
426 if (ts
->idle_active
|| (ts
->inidle
&& ts
->tick_stopped
))
430 tick_nohz_stop_idle(cpu
, now
);
432 if (!ts
->inidle
|| !ts
->tick_stopped
) {
442 /* Update jiffies first */
443 select_nohz_load_balancer(0);
444 tick_do_update_jiffies64(now
);
445 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
447 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
449 * We stopped the tick in idle. Update process times would miss the
450 * time we slept as update_process_times does only a 1 tick
451 * accounting. Enforce that this is accounted to idle !
453 ticks
= jiffies
- ts
->idle_jiffies
;
455 * We might be one off. Do not randomly account a huge number of ticks!
457 if (ticks
&& ticks
< LONG_MAX
)
458 account_idle_ticks(ticks
);
461 touch_softlockup_watchdog();
463 * Cancel the scheduled timer and restore the tick
465 ts
->tick_stopped
= 0;
466 ts
->idle_exittime
= now
;
468 tick_nohz_restart(ts
, now
);
473 static int tick_nohz_reprogram(struct tick_sched
*ts
, ktime_t now
)
475 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
476 return tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 0);
480 * The nohz low res interrupt handler
482 static void tick_nohz_handler(struct clock_event_device
*dev
)
484 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
485 struct pt_regs
*regs
= get_irq_regs();
486 int cpu
= smp_processor_id();
487 ktime_t now
= ktime_get();
489 dev
->next_event
.tv64
= KTIME_MAX
;
492 * Check if the do_timer duty was dropped. We don't care about
493 * concurrency: This happens only when the cpu in charge went
494 * into a long sleep. If two cpus happen to assign themself to
495 * this duty, then the jiffies update is still serialized by
498 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
499 tick_do_timer_cpu
= cpu
;
501 /* Check, if the jiffies need an update */
502 if (tick_do_timer_cpu
== cpu
)
503 tick_do_update_jiffies64(now
);
506 * When we are idle and the tick is stopped, we have to touch
507 * the watchdog as we might not schedule for a really long
508 * time. This happens on complete idle SMP systems while
509 * waiting on the login prompt. We also increment the "start
510 * of idle" jiffy stamp so the idle accounting adjustment we
511 * do when we go busy again does not account too much ticks.
513 if (ts
->tick_stopped
) {
514 touch_softlockup_watchdog();
518 update_process_times(user_mode(regs
));
519 profile_tick(CPU_PROFILING
);
521 while (tick_nohz_reprogram(ts
, now
)) {
523 tick_do_update_jiffies64(now
);
528 * tick_nohz_switch_to_nohz - switch to nohz mode
530 static void tick_nohz_switch_to_nohz(void)
532 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
535 if (!tick_nohz_enabled
)
539 if (tick_switch_to_oneshot(tick_nohz_handler
)) {
544 ts
->nohz_mode
= NOHZ_MODE_LOWRES
;
547 * Recycle the hrtimer in ts, so we can share the
548 * hrtimer_forward with the highres code.
550 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
551 /* Get the next period */
552 next
= tick_init_jiffy_update();
555 hrtimer_set_expires(&ts
->sched_timer
, next
);
556 if (!tick_program_event(next
, 0))
558 next
= ktime_add(next
, tick_period
);
562 printk(KERN_INFO
"Switched to NOHz mode on CPU #%d\n",
567 * When NOHZ is enabled and the tick is stopped, we need to kick the
568 * tick timer from irq_enter() so that the jiffies update is kept
569 * alive during long running softirqs. That's ugly as hell, but
570 * correctness is key even if we need to fix the offending softirq in
573 * Note, this is different to tick_nohz_restart. We just kick the
574 * timer and do not touch the other magic bits which need to be done
577 static void tick_nohz_kick_tick(int cpu
, ktime_t now
)
580 /* Switch back to 2.6.27 behaviour */
582 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
586 * Do not touch the tick device, when the next expiry is either
587 * already reached or less/equal than the tick period.
589 delta
= ktime_sub(hrtimer_get_expires(&ts
->sched_timer
), now
);
590 if (delta
.tv64
<= tick_period
.tv64
)
593 tick_nohz_restart(ts
, now
);
597 static inline void tick_check_nohz(int cpu
)
599 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
602 if (!ts
->idle_active
&& !ts
->tick_stopped
)
606 tick_nohz_stop_idle(cpu
, now
);
607 if (ts
->tick_stopped
) {
608 tick_nohz_update_jiffies(now
);
609 tick_nohz_kick_tick(cpu
, now
);
615 static inline void tick_nohz_switch_to_nohz(void) { }
616 static inline void tick_check_nohz(int cpu
) { }
621 * Called from irq_enter to notify about the possible interruption of idle()
623 void tick_check_idle(int cpu
)
625 tick_check_oneshot_broadcast(cpu
);
626 tick_check_nohz(cpu
);
630 * High resolution timer specific code
632 #ifdef CONFIG_HIGH_RES_TIMERS
634 * We rearm the timer until we get disabled by the idle code.
635 * Called with interrupts disabled and timer->base->cpu_base->lock held.
637 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
639 struct tick_sched
*ts
=
640 container_of(timer
, struct tick_sched
, sched_timer
);
641 struct pt_regs
*regs
= get_irq_regs();
642 ktime_t now
= ktime_get();
643 int cpu
= smp_processor_id();
647 * Check if the do_timer duty was dropped. We don't care about
648 * concurrency: This happens only when the cpu in charge went
649 * into a long sleep. If two cpus happen to assign themself to
650 * this duty, then the jiffies update is still serialized by
653 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
654 tick_do_timer_cpu
= cpu
;
657 /* Check, if the jiffies need an update */
658 if (tick_do_timer_cpu
== cpu
)
659 tick_do_update_jiffies64(now
);
662 * Do not call, when we are not in irq context and have
663 * no valid regs pointer
667 * When we are idle and the tick is stopped, we have to touch
668 * the watchdog as we might not schedule for a really long
669 * time. This happens on complete idle SMP systems while
670 * waiting on the login prompt. We also increment the "start of
671 * idle" jiffy stamp so the idle accounting adjustment we do
672 * when we go busy again does not account too much ticks.
674 if (ts
->tick_stopped
) {
675 touch_softlockup_watchdog();
678 update_process_times(user_mode(regs
));
679 profile_tick(CPU_PROFILING
);
682 hrtimer_forward(timer
, now
, tick_period
);
684 return HRTIMER_RESTART
;
688 * tick_setup_sched_timer - setup the tick emulation timer
690 void tick_setup_sched_timer(void)
692 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
693 ktime_t now
= ktime_get();
697 * Emulate tick processing via per-CPU hrtimers:
699 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
700 ts
->sched_timer
.function
= tick_sched_timer
;
702 /* Get the next period (per cpu) */
703 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
704 offset
= ktime_to_ns(tick_period
) >> 1;
705 do_div(offset
, num_possible_cpus());
706 offset
*= smp_processor_id();
707 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
710 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
711 hrtimer_start_expires(&ts
->sched_timer
,
712 HRTIMER_MODE_ABS_PINNED
);
713 /* Check, if the timer was already in the past */
714 if (hrtimer_active(&ts
->sched_timer
))
720 if (tick_nohz_enabled
)
721 ts
->nohz_mode
= NOHZ_MODE_HIGHRES
;
724 #endif /* HIGH_RES_TIMERS */
726 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
727 void tick_cancel_sched_timer(int cpu
)
729 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
731 # ifdef CONFIG_HIGH_RES_TIMERS
732 if (ts
->sched_timer
.base
)
733 hrtimer_cancel(&ts
->sched_timer
);
736 ts
->nohz_mode
= NOHZ_MODE_INACTIVE
;
741 * Async notification about clocksource changes
743 void tick_clock_notify(void)
747 for_each_possible_cpu(cpu
)
748 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
752 * Async notification about clock event changes
754 void tick_oneshot_notify(void)
756 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
758 set_bit(0, &ts
->check_clocks
);
762 * Check, if a change happened, which makes oneshot possible.
764 * Called cyclic from the hrtimer softirq (driven by the timer
765 * softirq) allow_nohz signals, that we can switch into low-res nohz
766 * mode, because high resolution timers are disabled (either compile
769 int tick_check_oneshot_change(int allow_nohz
)
771 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
773 if (!test_and_clear_bit(0, &ts
->check_clocks
))
776 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
779 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
785 tick_nohz_switch_to_nohz();