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 void tick_nohz_update_jiffies(void)
139 int cpu
= smp_processor_id();
140 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
144 if (!ts
->tick_stopped
)
147 cpu_clear(cpu
, nohz_cpu_mask
);
149 ts
->idle_waketime
= now
;
151 local_irq_save(flags
);
152 tick_do_update_jiffies64(now
);
153 local_irq_restore(flags
);
155 touch_softlockup_watchdog();
158 static void tick_nohz_stop_idle(int cpu
)
160 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
162 if (ts
->idle_active
) {
165 delta
= ktime_sub(now
, ts
->idle_entrytime
);
166 ts
->idle_lastupdate
= now
;
167 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
170 sched_clock_idle_wakeup_event(0);
174 static ktime_t
tick_nohz_start_idle(struct tick_sched
*ts
)
179 if (ts
->idle_active
) {
180 delta
= ktime_sub(now
, ts
->idle_entrytime
);
181 ts
->idle_lastupdate
= now
;
182 ts
->idle_sleeptime
= ktime_add(ts
->idle_sleeptime
, delta
);
184 ts
->idle_entrytime
= now
;
186 sched_clock_idle_sleep_event();
190 u64
get_cpu_idle_time_us(int cpu
, u64
*last_update_time
)
192 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
194 if (!tick_nohz_enabled
)
198 *last_update_time
= ktime_to_us(ts
->idle_lastupdate
);
200 *last_update_time
= ktime_to_us(ktime_get());
202 return ktime_to_us(ts
->idle_sleeptime
);
204 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us
);
207 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
209 * When the next event is more than a tick into the future, stop the idle tick
210 * Called either from the idle loop or from irq_exit() when an idle period was
211 * just interrupted by an interrupt which did not cause a reschedule.
213 void tick_nohz_stop_sched_tick(int inidle
)
215 unsigned long seq
, last_jiffies
, next_jiffies
, delta_jiffies
, flags
;
216 struct tick_sched
*ts
;
217 ktime_t last_update
, expires
, now
;
218 struct clock_event_device
*dev
= __get_cpu_var(tick_cpu_device
).evtdev
;
221 local_irq_save(flags
);
223 cpu
= smp_processor_id();
224 ts
= &per_cpu(tick_cpu_sched
, cpu
);
225 now
= tick_nohz_start_idle(ts
);
228 * If this cpu is offline and it is the one which updates
229 * jiffies, then give up the assignment and let it be taken by
230 * the cpu which runs the tick timer next. If we don't drop
231 * this here the jiffies might be stale and do_timer() never
234 if (unlikely(!cpu_online(cpu
))) {
235 if (cpu
== tick_do_timer_cpu
)
236 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
239 if (unlikely(ts
->nohz_mode
== NOHZ_MODE_INACTIVE
))
242 if (!inidle
&& !ts
->inidle
)
250 if (unlikely(local_softirq_pending())) {
251 static int ratelimit
;
253 if (ratelimit
< 10) {
254 printk(KERN_ERR
"NOHZ: local_softirq_pending %02x\n",
255 local_softirq_pending());
262 /* Read jiffies and the time when jiffies were updated last */
264 seq
= read_seqbegin(&xtime_lock
);
265 last_update
= last_jiffies_update
;
266 last_jiffies
= jiffies
;
267 } while (read_seqretry(&xtime_lock
, seq
));
269 /* Get the next timer wheel timer */
270 next_jiffies
= get_next_timer_interrupt(last_jiffies
);
271 delta_jiffies
= next_jiffies
- last_jiffies
;
273 if (rcu_needs_cpu(cpu
) || printk_needs_cpu(cpu
))
276 * Do not stop the tick, if we are only one off
277 * or if the cpu is required for rcu
279 if (!ts
->tick_stopped
&& delta_jiffies
== 1)
282 /* Schedule the tick, if we are at least one jiffie off */
283 if ((long)delta_jiffies
>= 1) {
285 if (delta_jiffies
> 1)
286 cpu_set(cpu
, nohz_cpu_mask
);
288 * nohz_stop_sched_tick can be called several times before
289 * the nohz_restart_sched_tick is called. This happens when
290 * interrupts arrive which do not cause a reschedule. In the
291 * first call we save the current tick time, so we can restart
292 * the scheduler tick in nohz_restart_sched_tick.
294 if (!ts
->tick_stopped
) {
295 if (select_nohz_load_balancer(1)) {
297 * sched tick not stopped!
299 cpu_clear(cpu
, nohz_cpu_mask
);
303 ts
->idle_tick
= ts
->sched_timer
.expires
;
304 ts
->tick_stopped
= 1;
305 ts
->idle_jiffies
= last_jiffies
;
310 * If this cpu is the one which updates jiffies, then
311 * give up the assignment and let it be taken by the
312 * cpu which runs the tick timer next, which might be
313 * this cpu as well. If we don't drop this here the
314 * jiffies might be stale and do_timer() never
317 if (cpu
== tick_do_timer_cpu
)
318 tick_do_timer_cpu
= TICK_DO_TIMER_NONE
;
323 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
324 * there is no timer pending or at least extremly far
325 * into the future (12 days for HZ=1000). In this case
326 * we simply stop the tick timer:
328 if (unlikely(delta_jiffies
>= NEXT_TIMER_MAX_DELTA
)) {
329 ts
->idle_expires
.tv64
= KTIME_MAX
;
330 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
)
331 hrtimer_cancel(&ts
->sched_timer
);
336 * calculate the expiry time for the next timer wheel
339 expires
= ktime_add_ns(last_update
, tick_period
.tv64
*
341 ts
->idle_expires
= expires
;
343 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
344 hrtimer_start(&ts
->sched_timer
, expires
,
346 /* Check, if the timer was already in the past */
347 if (hrtimer_active(&ts
->sched_timer
))
349 } else if (!tick_program_event(expires
, 0))
352 * We are past the event already. So we crossed a
353 * jiffie boundary. Update jiffies and raise the
356 tick_do_update_jiffies64(ktime_get());
357 cpu_clear(cpu
, nohz_cpu_mask
);
359 raise_softirq_irqoff(TIMER_SOFTIRQ
);
361 ts
->next_jiffies
= next_jiffies
;
362 ts
->last_jiffies
= last_jiffies
;
363 ts
->sleep_length
= ktime_sub(dev
->next_event
, now
);
365 local_irq_restore(flags
);
369 * tick_nohz_get_sleep_length - return the length of the current sleep
371 * Called from power state control code with interrupts disabled
373 ktime_t
tick_nohz_get_sleep_length(void)
375 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
377 return ts
->sleep_length
;
380 static void tick_nohz_restart(struct tick_sched
*ts
, ktime_t now
)
382 hrtimer_cancel(&ts
->sched_timer
);
383 ts
->sched_timer
.expires
= ts
->idle_tick
;
386 /* Forward the time to expire in the future */
387 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
389 if (ts
->nohz_mode
== NOHZ_MODE_HIGHRES
) {
390 hrtimer_start(&ts
->sched_timer
,
391 ts
->sched_timer
.expires
,
393 /* Check, if the timer was already in the past */
394 if (hrtimer_active(&ts
->sched_timer
))
397 if (!tick_program_event(ts
->sched_timer
.expires
, 0))
400 /* Update jiffies and reread time */
401 tick_do_update_jiffies64(now
);
407 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
409 * Restart the idle tick when the CPU is woken up from idle
411 void tick_nohz_restart_sched_tick(void)
413 int cpu
= smp_processor_id();
414 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
419 tick_nohz_stop_idle(cpu
);
421 if (!ts
->inidle
|| !ts
->tick_stopped
) {
431 /* Update jiffies first */
432 select_nohz_load_balancer(0);
434 tick_do_update_jiffies64(now
);
435 cpu_clear(cpu
, nohz_cpu_mask
);
438 * We stopped the tick in idle. Update process times would miss the
439 * time we slept as update_process_times does only a 1 tick
440 * accounting. Enforce that this is accounted to idle !
442 ticks
= jiffies
- ts
->idle_jiffies
;
444 * We might be one off. Do not randomly account a huge number of ticks!
446 if (ticks
&& ticks
< LONG_MAX
) {
447 add_preempt_count(HARDIRQ_OFFSET
);
448 account_system_time(current
, HARDIRQ_OFFSET
,
449 jiffies_to_cputime(ticks
));
450 sub_preempt_count(HARDIRQ_OFFSET
);
453 touch_softlockup_watchdog();
455 * Cancel the scheduled timer and restore the tick
457 ts
->tick_stopped
= 0;
458 ts
->idle_exittime
= now
;
459 tick_nohz_restart(ts
, now
);
463 static int tick_nohz_reprogram(struct tick_sched
*ts
, ktime_t now
)
465 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
466 return tick_program_event(ts
->sched_timer
.expires
, 0);
470 * The nohz low res interrupt handler
472 static void tick_nohz_handler(struct clock_event_device
*dev
)
474 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
475 struct pt_regs
*regs
= get_irq_regs();
476 int cpu
= smp_processor_id();
477 ktime_t now
= ktime_get();
479 dev
->next_event
.tv64
= KTIME_MAX
;
482 * Check if the do_timer duty was dropped. We don't care about
483 * concurrency: This happens only when the cpu in charge went
484 * into a long sleep. If two cpus happen to assign themself to
485 * this duty, then the jiffies update is still serialized by
488 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
489 tick_do_timer_cpu
= cpu
;
491 /* Check, if the jiffies need an update */
492 if (tick_do_timer_cpu
== cpu
)
493 tick_do_update_jiffies64(now
);
496 * When we are idle and the tick is stopped, we have to touch
497 * the watchdog as we might not schedule for a really long
498 * time. This happens on complete idle SMP systems while
499 * waiting on the login prompt. We also increment the "start
500 * of idle" jiffy stamp so the idle accounting adjustment we
501 * do when we go busy again does not account too much ticks.
503 if (ts
->tick_stopped
) {
504 touch_softlockup_watchdog();
508 update_process_times(user_mode(regs
));
509 profile_tick(CPU_PROFILING
);
511 while (tick_nohz_reprogram(ts
, now
)) {
513 tick_do_update_jiffies64(now
);
518 * tick_nohz_switch_to_nohz - switch to nohz mode
520 static void tick_nohz_switch_to_nohz(void)
522 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
525 if (!tick_nohz_enabled
)
529 if (tick_switch_to_oneshot(tick_nohz_handler
)) {
534 ts
->nohz_mode
= NOHZ_MODE_LOWRES
;
537 * Recycle the hrtimer in ts, so we can share the
538 * hrtimer_forward with the highres code.
540 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
541 /* Get the next period */
542 next
= tick_init_jiffy_update();
545 ts
->sched_timer
.expires
= next
;
546 if (!tick_program_event(next
, 0))
548 next
= ktime_add(next
, tick_period
);
552 printk(KERN_INFO
"Switched to NOHz mode on CPU #%d\n",
557 * When NOHZ is enabled and the tick is stopped, we need to kick the
558 * tick timer from irq_enter() so that the jiffies update is kept
559 * alive during long running softirqs. That's ugly as hell, but
560 * correctness is key even if we need to fix the offending softirq in
563 * Note, this is different to tick_nohz_restart. We just kick the
564 * timer and do not touch the other magic bits which need to be done
567 static void tick_nohz_kick_tick(int cpu
)
569 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
571 if (!ts
->tick_stopped
)
574 tick_nohz_restart(ts
, ktime_get());
579 static inline void tick_nohz_switch_to_nohz(void) { }
584 * Called from irq_enter to notify about the possible interruption of idle()
586 void tick_check_idle(int cpu
)
588 tick_check_oneshot_broadcast(cpu
);
590 tick_nohz_stop_idle(cpu
);
591 tick_nohz_update_jiffies();
592 tick_nohz_kick_tick(cpu
);
597 * High resolution timer specific code
599 #ifdef CONFIG_HIGH_RES_TIMERS
601 * We rearm the timer until we get disabled by the idle code.
602 * Called with interrupts disabled and timer->base->cpu_base->lock held.
604 static enum hrtimer_restart
tick_sched_timer(struct hrtimer
*timer
)
606 struct tick_sched
*ts
=
607 container_of(timer
, struct tick_sched
, sched_timer
);
608 struct pt_regs
*regs
= get_irq_regs();
609 ktime_t now
= ktime_get();
610 int cpu
= smp_processor_id();
614 * Check if the do_timer duty was dropped. We don't care about
615 * concurrency: This happens only when the cpu in charge went
616 * into a long sleep. If two cpus happen to assign themself to
617 * this duty, then the jiffies update is still serialized by
620 if (unlikely(tick_do_timer_cpu
== TICK_DO_TIMER_NONE
))
621 tick_do_timer_cpu
= cpu
;
624 /* Check, if the jiffies need an update */
625 if (tick_do_timer_cpu
== cpu
)
626 tick_do_update_jiffies64(now
);
629 * Do not call, when we are not in irq context and have
630 * no valid regs pointer
634 * When we are idle and the tick is stopped, we have to touch
635 * the watchdog as we might not schedule for a really long
636 * time. This happens on complete idle SMP systems while
637 * waiting on the login prompt. We also increment the "start of
638 * idle" jiffy stamp so the idle accounting adjustment we do
639 * when we go busy again does not account too much ticks.
641 if (ts
->tick_stopped
) {
642 touch_softlockup_watchdog();
645 update_process_times(user_mode(regs
));
646 profile_tick(CPU_PROFILING
);
649 hrtimer_forward(timer
, now
, tick_period
);
651 return HRTIMER_RESTART
;
655 * tick_setup_sched_timer - setup the tick emulation timer
657 void tick_setup_sched_timer(void)
659 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
660 ktime_t now
= ktime_get();
664 * Emulate tick processing via per-CPU hrtimers:
666 hrtimer_init(&ts
->sched_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS
);
667 ts
->sched_timer
.function
= tick_sched_timer
;
668 ts
->sched_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
670 /* Get the next period (per cpu) */
671 ts
->sched_timer
.expires
= tick_init_jiffy_update();
672 offset
= ktime_to_ns(tick_period
) >> 1;
673 do_div(offset
, num_possible_cpus());
674 offset
*= smp_processor_id();
675 ts
->sched_timer
.expires
= ktime_add_ns(ts
->sched_timer
.expires
, offset
);
678 hrtimer_forward(&ts
->sched_timer
, now
, tick_period
);
679 hrtimer_start(&ts
->sched_timer
, ts
->sched_timer
.expires
,
681 /* Check, if the timer was already in the past */
682 if (hrtimer_active(&ts
->sched_timer
))
688 if (tick_nohz_enabled
)
689 ts
->nohz_mode
= NOHZ_MODE_HIGHRES
;
692 #endif /* HIGH_RES_TIMERS */
694 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
695 void tick_cancel_sched_timer(int cpu
)
697 struct tick_sched
*ts
= &per_cpu(tick_cpu_sched
, cpu
);
699 # ifdef CONFIG_HIGH_RES_TIMERS
700 if (ts
->sched_timer
.base
)
701 hrtimer_cancel(&ts
->sched_timer
);
704 ts
->nohz_mode
= NOHZ_MODE_INACTIVE
;
709 * Async notification about clocksource changes
711 void tick_clock_notify(void)
715 for_each_possible_cpu(cpu
)
716 set_bit(0, &per_cpu(tick_cpu_sched
, cpu
).check_clocks
);
720 * Async notification about clock event changes
722 void tick_oneshot_notify(void)
724 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
726 set_bit(0, &ts
->check_clocks
);
730 * Check, if a change happened, which makes oneshot possible.
732 * Called cyclic from the hrtimer softirq (driven by the timer
733 * softirq) allow_nohz signals, that we can switch into low-res nohz
734 * mode, because high resolution timers are disabled (either compile
737 int tick_check_oneshot_change(int allow_nohz
)
739 struct tick_sched
*ts
= &__get_cpu_var(tick_cpu_sched
);
741 if (!test_and_clear_bit(0, &ts
->check_clocks
))
744 if (ts
->nohz_mode
!= NOHZ_MODE_INACTIVE
)
747 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
753 tick_nohz_switch_to_nohz();