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(void)
139 int cpu = smp_processor_id();
140 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
144 if (!ts->tick_stopped)
147 cpumask_clear_cpu(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);
227 * Call to tick_nohz_start_idle stops the last_update_time from being
228 * updated. Thus, it must not be called in the event we are called from
229 * irq_exit() with the prior state different than idle.
231 if (!inidle && !ts->inidle)
234 now = tick_nohz_start_idle(ts);
237 * If this cpu is offline and it is the one which updates
238 * jiffies, then give up the assignment and let it be taken by
239 * the cpu which runs the tick timer next. If we don't drop
240 * this here the jiffies might be stale and do_timer() never
243 if (unlikely(!cpu_online(cpu))) {
244 if (cpu == tick_do_timer_cpu)
245 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
248 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
256 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
257 static int ratelimit;
259 if (ratelimit < 10) {
260 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
261 local_softirq_pending());
268 /* Read jiffies and the time when jiffies were updated last */
270 seq = read_seqbegin(&xtime_lock);
271 last_update = last_jiffies_update;
272 last_jiffies = jiffies;
273 } while (read_seqretry(&xtime_lock, seq));
275 /* Get the next timer wheel timer */
276 next_jiffies = get_next_timer_interrupt(last_jiffies);
277 delta_jiffies = next_jiffies - last_jiffies;
279 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu))
282 * Do not stop the tick, if we are only one off
283 * or if the cpu is required for rcu
285 if (!ts->tick_stopped && delta_jiffies == 1)
288 /* Schedule the tick, if we are at least one jiffie off */
289 if ((long)delta_jiffies >= 1) {
292 * calculate the expiry time for the next timer wheel
295 expires = ktime_add_ns(last_update, tick_period.tv64 *
299 * If this cpu is the one which updates jiffies, then
300 * give up the assignment and let it be taken by the
301 * cpu which runs the tick timer next, which might be
302 * this cpu as well. If we don't drop this here the
303 * jiffies might be stale and do_timer() never
306 if (cpu == tick_do_timer_cpu)
307 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
309 if (delta_jiffies > 1)
310 cpumask_set_cpu(cpu, nohz_cpu_mask);
312 /* Skip reprogram of event if its not changed */
313 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
317 * nohz_stop_sched_tick can be called several times before
318 * the nohz_restart_sched_tick is called. This happens when
319 * interrupts arrive which do not cause a reschedule. In the
320 * first call we save the current tick time, so we can restart
321 * the scheduler tick in nohz_restart_sched_tick.
323 if (!ts->tick_stopped) {
324 if (select_nohz_load_balancer(1)) {
326 * sched tick not stopped!
328 cpumask_clear_cpu(cpu, nohz_cpu_mask);
332 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
333 ts->tick_stopped = 1;
334 ts->idle_jiffies = last_jiffies;
341 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
342 * there is no timer pending or at least extremly far
343 * into the future (12 days for HZ=1000). In this case
344 * we simply stop the tick timer:
346 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
347 ts->idle_expires.tv64 = KTIME_MAX;
348 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
349 hrtimer_cancel(&ts->sched_timer);
354 ts->idle_expires = expires;
356 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
357 hrtimer_start(&ts->sched_timer, expires,
358 HRTIMER_MODE_ABS_PINNED);
359 /* Check, if the timer was already in the past */
360 if (hrtimer_active(&ts->sched_timer))
362 } else if (!tick_program_event(expires, 0))
365 * We are past the event already. So we crossed a
366 * jiffie boundary. Update jiffies and raise the
369 tick_do_update_jiffies64(ktime_get());
370 cpumask_clear_cpu(cpu, nohz_cpu_mask);
372 raise_softirq_irqoff(TIMER_SOFTIRQ);
374 ts->next_jiffies = next_jiffies;
375 ts->last_jiffies = last_jiffies;
376 ts->sleep_length = ktime_sub(dev->next_event, now);
378 local_irq_restore(flags);
382 * tick_nohz_get_sleep_length - return the length of the current sleep
384 * Called from power state control code with interrupts disabled
386 ktime_t tick_nohz_get_sleep_length(void)
388 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
390 return ts->sleep_length;
393 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
395 hrtimer_cancel(&ts->sched_timer);
396 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
399 /* Forward the time to expire in the future */
400 hrtimer_forward(&ts->sched_timer, now, tick_period);
402 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
403 hrtimer_start_expires(&ts->sched_timer,
404 HRTIMER_MODE_ABS_PINNED);
405 /* Check, if the timer was already in the past */
406 if (hrtimer_active(&ts->sched_timer))
409 if (!tick_program_event(
410 hrtimer_get_expires(&ts->sched_timer), 0))
413 /* Update jiffies and reread time */
414 tick_do_update_jiffies64(now);
420 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
422 * Restart the idle tick when the CPU is woken up from idle
424 void tick_nohz_restart_sched_tick(void)
426 int cpu = smp_processor_id();
427 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
428 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
434 tick_nohz_stop_idle(cpu);
436 if (!ts->inidle || !ts->tick_stopped) {
446 /* Update jiffies first */
447 select_nohz_load_balancer(0);
449 tick_do_update_jiffies64(now);
450 cpumask_clear_cpu(cpu, nohz_cpu_mask);
452 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
454 * We stopped the tick in idle. Update process times would miss the
455 * time we slept as update_process_times does only a 1 tick
456 * accounting. Enforce that this is accounted to idle !
458 ticks = jiffies - ts->idle_jiffies;
460 * We might be one off. Do not randomly account a huge number of ticks!
462 if (ticks && ticks < LONG_MAX)
463 account_idle_ticks(ticks);
466 touch_softlockup_watchdog();
468 * Cancel the scheduled timer and restore the tick
470 ts->tick_stopped = 0;
471 ts->idle_exittime = now;
473 tick_nohz_restart(ts, now);
478 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
480 hrtimer_forward(&ts->sched_timer, now, tick_period);
481 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
485 * The nohz low res interrupt handler
487 static void tick_nohz_handler(struct clock_event_device *dev)
489 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
490 struct pt_regs *regs = get_irq_regs();
491 int cpu = smp_processor_id();
492 ktime_t now = ktime_get();
494 dev->next_event.tv64 = KTIME_MAX;
497 * Check if the do_timer duty was dropped. We don't care about
498 * concurrency: This happens only when the cpu in charge went
499 * into a long sleep. If two cpus happen to assign themself to
500 * this duty, then the jiffies update is still serialized by
503 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
504 tick_do_timer_cpu = cpu;
506 /* Check, if the jiffies need an update */
507 if (tick_do_timer_cpu == cpu)
508 tick_do_update_jiffies64(now);
511 * When we are idle and the tick is stopped, we have to touch
512 * the watchdog as we might not schedule for a really long
513 * time. This happens on complete idle SMP systems while
514 * waiting on the login prompt. We also increment the "start
515 * of idle" jiffy stamp so the idle accounting adjustment we
516 * do when we go busy again does not account too much ticks.
518 if (ts->tick_stopped) {
519 touch_softlockup_watchdog();
523 update_process_times(user_mode(regs));
524 profile_tick(CPU_PROFILING);
526 while (tick_nohz_reprogram(ts, now)) {
528 tick_do_update_jiffies64(now);
533 * tick_nohz_switch_to_nohz - switch to nohz mode
535 static void tick_nohz_switch_to_nohz(void)
537 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
540 if (!tick_nohz_enabled)
544 if (tick_switch_to_oneshot(tick_nohz_handler)) {
549 ts->nohz_mode = NOHZ_MODE_LOWRES;
552 * Recycle the hrtimer in ts, so we can share the
553 * hrtimer_forward with the highres code.
555 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
556 /* Get the next period */
557 next = tick_init_jiffy_update();
560 hrtimer_set_expires(&ts->sched_timer, next);
561 if (!tick_program_event(next, 0))
563 next = ktime_add(next, tick_period);
567 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
572 * When NOHZ is enabled and the tick is stopped, we need to kick the
573 * tick timer from irq_enter() so that the jiffies update is kept
574 * alive during long running softirqs. That's ugly as hell, but
575 * correctness is key even if we need to fix the offending softirq in
578 * Note, this is different to tick_nohz_restart. We just kick the
579 * timer and do not touch the other magic bits which need to be done
582 static void tick_nohz_kick_tick(int cpu)
585 /* Switch back to 2.6.27 behaviour */
587 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
590 if (!ts->tick_stopped)
594 * Do not touch the tick device, when the next expiry is either
595 * already reached or less/equal than the tick period.
598 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
599 if (delta.tv64 <= tick_period.tv64)
602 tick_nohz_restart(ts, now);
608 static inline void tick_nohz_switch_to_nohz(void) { }
613 * Called from irq_enter to notify about the possible interruption of idle()
615 void tick_check_idle(int cpu)
617 tick_check_oneshot_broadcast(cpu);
619 tick_nohz_stop_idle(cpu);
620 tick_nohz_update_jiffies();
621 tick_nohz_kick_tick(cpu);
626 * High resolution timer specific code
628 #ifdef CONFIG_HIGH_RES_TIMERS
630 * We rearm the timer until we get disabled by the idle code.
631 * Called with interrupts disabled and timer->base->cpu_base->lock held.
633 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
635 struct tick_sched *ts =
636 container_of(timer, struct tick_sched, sched_timer);
637 struct pt_regs *regs = get_irq_regs();
638 ktime_t now = ktime_get();
639 int cpu = smp_processor_id();
643 * Check if the do_timer duty was dropped. We don't care about
644 * concurrency: This happens only when the cpu in charge went
645 * into a long sleep. If two cpus happen to assign themself to
646 * this duty, then the jiffies update is still serialized by
649 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
650 tick_do_timer_cpu = cpu;
653 /* Check, if the jiffies need an update */
654 if (tick_do_timer_cpu == cpu)
655 tick_do_update_jiffies64(now);
658 * Do not call, when we are not in irq context and have
659 * no valid regs pointer
663 * When we are idle and the tick is stopped, we have to touch
664 * the watchdog as we might not schedule for a really long
665 * time. This happens on complete idle SMP systems while
666 * waiting on the login prompt. We also increment the "start of
667 * idle" jiffy stamp so the idle accounting adjustment we do
668 * when we go busy again does not account too much ticks.
670 if (ts->tick_stopped) {
671 touch_softlockup_watchdog();
674 update_process_times(user_mode(regs));
675 profile_tick(CPU_PROFILING);
678 hrtimer_forward(timer, now, tick_period);
680 return HRTIMER_RESTART;
684 * tick_setup_sched_timer - setup the tick emulation timer
686 void tick_setup_sched_timer(void)
688 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
689 ktime_t now = ktime_get();
693 * Emulate tick processing via per-CPU hrtimers:
695 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
696 ts->sched_timer.function = tick_sched_timer;
698 /* Get the next period (per cpu) */
699 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
700 offset = ktime_to_ns(tick_period) >> 1;
701 do_div(offset, num_possible_cpus());
702 offset *= smp_processor_id();
703 hrtimer_add_expires_ns(&ts->sched_timer, offset);
706 hrtimer_forward(&ts->sched_timer, now, tick_period);
707 hrtimer_start_expires(&ts->sched_timer,
708 HRTIMER_MODE_ABS_PINNED);
709 /* Check, if the timer was already in the past */
710 if (hrtimer_active(&ts->sched_timer))
716 if (tick_nohz_enabled)
717 ts->nohz_mode = NOHZ_MODE_HIGHRES;
720 #endif /* HIGH_RES_TIMERS */
722 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
723 void tick_cancel_sched_timer(int cpu)
725 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
727 # ifdef CONFIG_HIGH_RES_TIMERS
728 if (ts->sched_timer.base)
729 hrtimer_cancel(&ts->sched_timer);
732 ts->nohz_mode = NOHZ_MODE_INACTIVE;
737 * Async notification about clocksource changes
739 void tick_clock_notify(void)
743 for_each_possible_cpu(cpu)
744 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
748 * Async notification about clock event changes
750 void tick_oneshot_notify(void)
752 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
754 set_bit(0, &ts->check_clocks);
758 * Check, if a change happened, which makes oneshot possible.
760 * Called cyclic from the hrtimer softirq (driven by the timer
761 * softirq) allow_nohz signals, that we can switch into low-res nohz
762 * mode, because high resolution timers are disabled (either compile
765 int tick_check_oneshot_change(int allow_nohz)
767 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
769 if (!test_and_clear_bit(0, &ts->check_clocks))
772 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
775 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
781 tick_nohz_switch_to_nohz();