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 if (rcu_needs_cpu(cpu
) || printk_needs_cpu(cpu
) ||
268 arch_needs_cpu(cpu
)) {
269 next_jiffies
= last_jiffies
+ 1;
272 /* Get the next timer wheel timer */
273 next_jiffies
= get_next_timer_interrupt(last_jiffies
);
274 delta_jiffies
= next_jiffies
- last_jiffies
;
277 * Do not stop the tick, if we are only one off
278 * or if the cpu is required for rcu
280 if (!ts
->tick_stopped
&& delta_jiffies
== 1)
283 /* Schedule the tick, if we are at least one jiffie off */
284 if ((long)delta_jiffies
>= 1) {
287 * calculate the expiry time for the next timer wheel
290 expires
= ktime_add_ns(last_update
, tick_period
.tv64
*
294 * If this cpu is the one which updates jiffies, then
295 * give up the assignment and let it be taken by the
296 * cpu which runs the tick timer next, which might be
297 * this cpu as well. If we don't drop this here the
298 * jiffies might be stale and do_timer() never
301 if (cpu
== tick_do_timer_cpu
)
302 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
304 if (delta_jiffies
> 1)
305 cpumask_set_cpu(cpu
, nohz_cpu_mask
);
307 /* Skip reprogram of event if its not changed */
308 if (ts
->tick_stopped
&& ktime_equal(expires
, dev
->next_event
))
312 * nohz_stop_sched_tick can be called several times before
313 * the nohz_restart_sched_tick is called. This happens when
314 * interrupts arrive which do not cause a reschedule. In the
315 * first call we save the current tick time, so we can restart
316 * the scheduler tick in nohz_restart_sched_tick.
318 if (!ts
->tick_stopped
) {
319 if (select_nohz_load_balancer(1)) {
321 * sched tick not stopped!
323 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
327 ts
->idle_tick
= hrtimer_get_expires(&ts
->sched_timer
);
328 ts
->tick_stopped
= 1;
329 ts
->idle_jiffies
= last_jiffies
;
336 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
337 * there is no timer pending or at least extremly far
338 * into the future (12 days for HZ=1000). In this case
339 * we simply stop the tick timer:
341 if (unlikely(delta_jiffies
>= NEXT_TIMER_MAX_DELTA
)) {
342 ts
->idle_expires
.tv64
= KTIME_MAX
;
343 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
344 hrtimer_cancel(&ts
->sched_timer
);
349 ts
->idle_expires
= expires
;
351 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
352 hrtimer_start(&ts
->sched_timer
, expires
,
353 HRTIMER_MODE_ABS_PINNED
);
354 /* Check, if the timer was already in the past */
355 if (hrtimer_active(&ts
->sched_timer
))
357 } else if (!tick_program_event(expires
, 0))
360 * We are past the event already. So we crossed a
361 * jiffie boundary. Update jiffies and raise the
364 tick_do_update_jiffies64(ktime_get());
365 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
367 raise_softirq_irqoff(TIMER_SOFTIRQ
);
369 ts
->next_jiffies
= next_jiffies
;
370 ts
->last_jiffies
= last_jiffies
;
371 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
373 local_irq_restore(flags
);
377 * tick_nohz_get_sleep_length - return the length of the current sleep
379 * Called from power state control code with interrupts disabled
381 ktime_t
tick_nohz_get_sleep_length(void)
383 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
385 return ts
->sleep_length
;
388 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
390 hrtimer_cancel(&ts
->sched_timer
);
391 hrtimer_set_expires(&ts
->sched_timer
, ts
->idle_tick
);
394 /* Forward the time to expire in the future */
395 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
397 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
398 hrtimer_start_expires(&ts
->sched_timer
,
399 HRTIMER_MODE_ABS_PINNED
);
400 /* Check, if the timer was already in the past */
401 if (hrtimer_active(&ts
->sched_timer
))
404 if (!tick_program_event(
405 hrtimer_get_expires(&ts
->sched_timer
), 0))
408 /* Update jiffies and reread time */
409 tick_do_update_jiffies64(now
);
415 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
417 * Restart the idle tick when the CPU is woken up from idle
419 void tick_nohz_restart_sched_tick(void)
421 int cpu
= smp_processor_id();
422 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
423 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
429 if (ts
->idle_active
|| (ts
->inidle
&& ts
->tick_stopped
))
433 tick_nohz_stop_idle(cpu
, now
);
435 if (!ts
->inidle
|| !ts
->tick_stopped
) {
445 /* Update jiffies first */
446 select_nohz_load_balancer(0);
447 tick_do_update_jiffies64(now
);
448 cpumask_clear_cpu(cpu
, nohz_cpu_mask
);
450 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
452 * We stopped the tick in idle. Update process times would miss the
453 * time we slept as update_process_times does only a 1 tick
454 * accounting. Enforce that this is accounted to idle !
456 ticks
= jiffies
- ts
->idle_jiffies
;
458 * We might be one off. Do not randomly account a huge number of ticks!
460 if (ticks
&& ticks
< LONG_MAX
)
461 account_idle_ticks(ticks
);
464 touch_softlockup_watchdog();
466 * Cancel the scheduled timer and restore the tick
468 ts
->tick_stopped
= 0;
469 ts
->idle_exittime
= now
;
471 tick_nohz_restart(ts
, now
);
476 static int tick_nohz_reprogram(struct tick_sched
*ts
, ktime_t now
)
478 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
479 return tick_program_event(hrtimer_get_expires(&ts
->sched_timer
), 0);
483 * The nohz low res interrupt handler
485 static void tick_nohz_handler(struct clock_event_device
*dev
)
487 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
488 struct pt_regs
*regs
= get_irq_regs();
489 int cpu
= smp_processor_id();
490 ktime_t now
= ktime_get();
492 dev
->next_event
.tv64
= KTIME_MAX
;
495 * Check if the do_timer duty was dropped. We don't care about
496 * concurrency: This happens only when the cpu in charge went
497 * into a long sleep. If two cpus happen to assign themself to
498 * this duty, then the jiffies update is still serialized by
501 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
502 tick_do_timer_cpu
= cpu
;
504 /* Check, if the jiffies need an update */
505 if (tick_do_timer_cpu
== cpu
)
506 tick_do_update_jiffies64(now
);
509 * When we are idle and the tick is stopped, we have to touch
510 * the watchdog as we might not schedule for a really long
511 * time. This happens on complete idle SMP systems while
512 * waiting on the login prompt. We also increment the "start
513 * of idle" jiffy stamp so the idle accounting adjustment we
514 * do when we go busy again does not account too much ticks.
516 if (ts
->tick_stopped
) {
517 touch_softlockup_watchdog();
521 update_process_times(user_mode(regs
));
522 profile_tick(CPU_PROFILING
);
524 while (tick_nohz_reprogram(ts
, now
)) {
526 tick_do_update_jiffies64(now
);
531 * tick_nohz_switch_to_nohz - switch to nohz mode
533 static void tick_nohz_switch_to_nohz(void)
535 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
538 if (!tick_nohz_enabled
)
542 if (tick_switch_to_oneshot(tick_nohz_handler
)) {
547 ts
->nohz_mode
= NOHZ_MODE_LOWRES
;
550 * Recycle the hrtimer in ts, so we can share the
551 * hrtimer_forward with the highres code.
553 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
554 /* Get the next period */
555 next
= tick_init_jiffy_update();
558 hrtimer_set_expires(&ts
->sched_timer
, next
);
559 if (!tick_program_event(next
, 0))
561 next
= ktime_add(next
, tick_period
);
565 printk(KERN_INFO
"Switched to NOHz mode on CPU #%d\n",
570 * When NOHZ is enabled and the tick is stopped, we need to kick the
571 * tick timer from irq_enter() so that the jiffies update is kept
572 * alive during long running softirqs. That's ugly as hell, but
573 * correctness is key even if we need to fix the offending softirq in
576 * Note, this is different to tick_nohz_restart. We just kick the
577 * timer and do not touch the other magic bits which need to be done
580 static void tick_nohz_kick_tick(int cpu
, ktime_t now
)
583 /* Switch back to 2.6.27 behaviour */
585 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
589 * Do not touch the tick device, when the next expiry is either
590 * already reached or less/equal than the tick period.
592 delta
= ktime_sub(hrtimer_get_expires(&ts
->sched_timer
), now
);
593 if (delta
.tv64
<= tick_period
.tv64
)
596 tick_nohz_restart(ts
, now
);
600 static inline void tick_check_nohz(int cpu
)
602 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
605 if (!ts
->idle_active
&& !ts
->tick_stopped
)
609 tick_nohz_stop_idle(cpu
, now
);
610 if (ts
->tick_stopped
) {
611 tick_nohz_update_jiffies(now
);
612 tick_nohz_kick_tick(cpu
, now
);
618 static inline void tick_nohz_switch_to_nohz(void) { }
619 static inline void tick_check_nohz(int cpu
) { }
624 * Called from irq_enter to notify about the possible interruption of idle()
626 void tick_check_idle(int cpu
)
628 tick_check_oneshot_broadcast(cpu
);
629 tick_check_nohz(cpu
);
633 * High resolution timer specific code
635 #ifdef CONFIG_HIGH_RES_TIMERS
637 * We rearm the timer until we get disabled by the idle code.
638 * Called with interrupts disabled and timer->base->cpu_base->lock held.
640 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
642 struct tick_sched
*ts
=
643 container_of(timer
, struct tick_sched
, sched_timer
);
644 struct pt_regs
*regs
= get_irq_regs();
645 ktime_t now
= ktime_get();
646 int cpu
= smp_processor_id();
650 * Check if the do_timer duty was dropped. We don't care about
651 * concurrency: This happens only when the cpu in charge went
652 * into a long sleep. If two cpus happen to assign themself to
653 * this duty, then the jiffies update is still serialized by
656 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
657 tick_do_timer_cpu
= cpu
;
660 /* Check, if the jiffies need an update */
661 if (tick_do_timer_cpu
== cpu
)
662 tick_do_update_jiffies64(now
);
665 * Do not call, when we are not in irq context and have
666 * no valid regs pointer
670 * When we are idle and the tick is stopped, we have to touch
671 * the watchdog as we might not schedule for a really long
672 * time. This happens on complete idle SMP systems while
673 * waiting on the login prompt. We also increment the "start of
674 * idle" jiffy stamp so the idle accounting adjustment we do
675 * when we go busy again does not account too much ticks.
677 if (ts
->tick_stopped
) {
678 touch_softlockup_watchdog();
681 update_process_times(user_mode(regs
));
682 profile_tick(CPU_PROFILING
);
685 hrtimer_forward(timer
, now
, tick_period
);
687 return HRTIMER_RESTART
;
691 * tick_setup_sched_timer - setup the tick emulation timer
693 void tick_setup_sched_timer(void)
695 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
696 ktime_t now
= ktime_get();
700 * Emulate tick processing via per-CPU hrtimers:
702 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
703 ts
->sched_timer
.function
= tick_sched_timer
;
705 /* Get the next period (per cpu) */
706 hrtimer_set_expires(&ts
->sched_timer
, tick_init_jiffy_update());
707 offset
= ktime_to_ns(tick_period
) >> 1;
708 do_div(offset
, num_possible_cpus());
709 offset
*= smp_processor_id();
710 hrtimer_add_expires_ns(&ts
->sched_timer
, offset
);
713 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
714 hrtimer_start_expires(&ts
->sched_timer
,
715 HRTIMER_MODE_ABS_PINNED
);
716 /* Check, if the timer was already in the past */
717 if (hrtimer_active(&ts
->sched_timer
))
723 if (tick_nohz_enabled
)
724 ts
->nohz_mode
= NOHZ_MODE_HIGHRES
;
727 #endif /* HIGH_RES_TIMERS */
729 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
730 void tick_cancel_sched_timer(int cpu
)
732 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
734 # ifdef CONFIG_HIGH_RES_TIMERS
735 if (ts
->sched_timer
.base
)
736 hrtimer_cancel(&ts
->sched_timer
);
739 ts
->nohz_mode
= NOHZ_MODE_INACTIVE
;
744 * Async notification about clocksource changes
746 void tick_clock_notify(void)
750 for_each_possible_cpu(cpu
)
751 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
755 * Async notification about clock event changes
757 void tick_oneshot_notify(void)
759 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
761 set_bit(0, &ts
->check_clocks
);
765 * Check, if a change happened, which makes oneshot possible.
767 * Called cyclic from the hrtimer softirq (driven by the timer
768 * softirq) allow_nohz signals, that we can switch into low-res nohz
769 * mode, because high resolution timers are disabled (either compile
772 int tick_check_oneshot_change(int allow_nohz
)
774 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
776 if (!test_and_clear_bit(0, &ts
->check_clocks
))
779 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
782 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
788 tick_nohz_switch_to_nohz();