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);
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() && cpu_online(cpu))) {
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) {
286 * calculate the expiry time for the next timer wheel
289 expires = ktime_add_ns(last_update, tick_period.tv64 *
293 * If this cpu is the one which updates jiffies, then
294 * give up the assignment and let it be taken by the
295 * cpu which runs the tick timer next, which might be
296 * this cpu as well. If we don't drop this here the
297 * jiffies might be stale and do_timer() never
300 if (cpu == tick_do_timer_cpu)
301 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
303 if (delta_jiffies > 1)
304 cpumask_set_cpu(cpu, nohz_cpu_mask);
306 /* Skip reprogram of event if its not changed */
307 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
311 * nohz_stop_sched_tick can be called several times before
312 * the nohz_restart_sched_tick is called. This happens when
313 * interrupts arrive which do not cause a reschedule. In the
314 * first call we save the current tick time, so we can restart
315 * the scheduler tick in nohz_restart_sched_tick.
317 if (!ts->tick_stopped) {
318 if (select_nohz_load_balancer(1)) {
320 * sched tick not stopped!
322 cpumask_clear_cpu(cpu, nohz_cpu_mask);
326 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
327 ts->tick_stopped = 1;
328 ts->idle_jiffies = last_jiffies;
335 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
336 * there is no timer pending or at least extremly far
337 * into the future (12 days for HZ=1000). In this case
338 * we simply stop the tick timer:
340 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
341 ts->idle_expires.tv64 = KTIME_MAX;
342 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
343 hrtimer_cancel(&ts->sched_timer);
348 ts->idle_expires = expires;
350 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
351 hrtimer_start(&ts->sched_timer, expires,
352 HRTIMER_MODE_ABS_PINNED);
353 /* Check, if the timer was already in the past */
354 if (hrtimer_active(&ts->sched_timer))
356 } else if (!tick_program_event(expires, 0))
359 * We are past the event already. So we crossed a
360 * jiffie boundary. Update jiffies and raise the
363 tick_do_update_jiffies64(ktime_get());
364 cpumask_clear_cpu(cpu, nohz_cpu_mask);
366 raise_softirq_irqoff(TIMER_SOFTIRQ);
368 ts->next_jiffies = next_jiffies;
369 ts->last_jiffies = last_jiffies;
370 ts->sleep_length = ktime_sub(dev->next_event, now);
372 local_irq_restore(flags);
376 * tick_nohz_get_sleep_length - return the length of the current sleep
378 * Called from power state control code with interrupts disabled
380 ktime_t tick_nohz_get_sleep_length(void)
382 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
384 return ts->sleep_length;
387 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
389 hrtimer_cancel(&ts->sched_timer);
390 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
393 /* Forward the time to expire in the future */
394 hrtimer_forward(&ts->sched_timer, now, tick_period);
396 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
397 hrtimer_start_expires(&ts->sched_timer,
398 HRTIMER_MODE_ABS_PINNED);
399 /* Check, if the timer was already in the past */
400 if (hrtimer_active(&ts->sched_timer))
403 if (!tick_program_event(
404 hrtimer_get_expires(&ts->sched_timer), 0))
407 /* Update jiffies and reread time */
408 tick_do_update_jiffies64(now);
414 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
416 * Restart the idle tick when the CPU is woken up from idle
418 void tick_nohz_restart_sched_tick(void)
420 int cpu = smp_processor_id();
421 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
422 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
428 tick_nohz_stop_idle(cpu);
430 if (!ts->inidle || !ts->tick_stopped) {
440 /* Update jiffies first */
441 select_nohz_load_balancer(0);
443 tick_do_update_jiffies64(now);
444 cpumask_clear_cpu(cpu, nohz_cpu_mask);
446 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
448 * We stopped the tick in idle. Update process times would miss the
449 * time we slept as update_process_times does only a 1 tick
450 * accounting. Enforce that this is accounted to idle !
452 ticks = jiffies - ts->idle_jiffies;
454 * We might be one off. Do not randomly account a huge number of ticks!
456 if (ticks && ticks < LONG_MAX)
457 account_idle_ticks(ticks);
460 touch_softlockup_watchdog();
462 * Cancel the scheduled timer and restore the tick
464 ts->tick_stopped = 0;
465 ts->idle_exittime = now;
467 tick_nohz_restart(ts, now);
472 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
474 hrtimer_forward(&ts->sched_timer, now, tick_period);
475 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
479 * The nohz low res interrupt handler
481 static void tick_nohz_handler(struct clock_event_device *dev)
483 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
484 struct pt_regs *regs = get_irq_regs();
485 int cpu = smp_processor_id();
486 ktime_t now = ktime_get();
488 dev->next_event.tv64 = KTIME_MAX;
491 * Check if the do_timer duty was dropped. We don't care about
492 * concurrency: This happens only when the cpu in charge went
493 * into a long sleep. If two cpus happen to assign themself to
494 * this duty, then the jiffies update is still serialized by
497 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
498 tick_do_timer_cpu = cpu;
500 /* Check, if the jiffies need an update */
501 if (tick_do_timer_cpu == cpu)
502 tick_do_update_jiffies64(now);
505 * When we are idle and the tick is stopped, we have to touch
506 * the watchdog as we might not schedule for a really long
507 * time. This happens on complete idle SMP systems while
508 * waiting on the login prompt. We also increment the "start
509 * of idle" jiffy stamp so the idle accounting adjustment we
510 * do when we go busy again does not account too much ticks.
512 if (ts->tick_stopped) {
513 touch_softlockup_watchdog();
517 update_process_times(user_mode(regs));
518 profile_tick(CPU_PROFILING);
520 while (tick_nohz_reprogram(ts, now)) {
522 tick_do_update_jiffies64(now);
527 * tick_nohz_switch_to_nohz - switch to nohz mode
529 static void tick_nohz_switch_to_nohz(void)
531 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
534 if (!tick_nohz_enabled)
538 if (tick_switch_to_oneshot(tick_nohz_handler)) {
543 ts->nohz_mode = NOHZ_MODE_LOWRES;
546 * Recycle the hrtimer in ts, so we can share the
547 * hrtimer_forward with the highres code.
549 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
550 /* Get the next period */
551 next = tick_init_jiffy_update();
554 hrtimer_set_expires(&ts->sched_timer, next);
555 if (!tick_program_event(next, 0))
557 next = ktime_add(next, tick_period);
561 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
566 * When NOHZ is enabled and the tick is stopped, we need to kick the
567 * tick timer from irq_enter() so that the jiffies update is kept
568 * alive during long running softirqs. That's ugly as hell, but
569 * correctness is key even if we need to fix the offending softirq in
572 * Note, this is different to tick_nohz_restart. We just kick the
573 * timer and do not touch the other magic bits which need to be done
576 static void tick_nohz_kick_tick(int cpu)
579 /* Switch back to 2.6.27 behaviour */
581 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
584 if (!ts->tick_stopped)
588 * Do not touch the tick device, when the next expiry is either
589 * 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);
602 static inline void tick_nohz_switch_to_nohz(void) { }
607 * Called from irq_enter to notify about the possible interruption of idle()
609 void tick_check_idle(int cpu)
611 tick_check_oneshot_broadcast(cpu);
613 tick_nohz_stop_idle(cpu);
614 tick_nohz_update_jiffies();
615 tick_nohz_kick_tick(cpu);
620 * High resolution timer specific code
622 #ifdef CONFIG_HIGH_RES_TIMERS
624 * We rearm the timer until we get disabled by the idle code.
625 * Called with interrupts disabled and timer->base->cpu_base->lock held.
627 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
629 struct tick_sched *ts =
630 container_of(timer, struct tick_sched, sched_timer);
631 struct pt_regs *regs = get_irq_regs();
632 ktime_t now = ktime_get();
633 int cpu = smp_processor_id();
637 * Check if the do_timer duty was dropped. We don't care about
638 * concurrency: This happens only when the cpu in charge went
639 * into a long sleep. If two cpus happen to assign themself to
640 * this duty, then the jiffies update is still serialized by
643 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
644 tick_do_timer_cpu = cpu;
647 /* Check, if the jiffies need an update */
648 if (tick_do_timer_cpu == cpu)
649 tick_do_update_jiffies64(now);
652 * Do not call, when we are not in irq context and have
653 * no valid regs pointer
657 * When we are idle and the tick is stopped, we have to touch
658 * the watchdog as we might not schedule for a really long
659 * time. This happens on complete idle SMP systems while
660 * waiting on the login prompt. We also increment the "start of
661 * idle" jiffy stamp so the idle accounting adjustment we do
662 * when we go busy again does not account too much ticks.
664 if (ts->tick_stopped) {
665 touch_softlockup_watchdog();
668 update_process_times(user_mode(regs));
669 profile_tick(CPU_PROFILING);
672 hrtimer_forward(timer, now, tick_period);
674 return HRTIMER_RESTART;
678 * tick_setup_sched_timer - setup the tick emulation timer
680 void tick_setup_sched_timer(void)
682 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
683 ktime_t now = ktime_get();
687 * Emulate tick processing via per-CPU hrtimers:
689 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
690 ts->sched_timer.function = tick_sched_timer;
692 /* Get the next period (per cpu) */
693 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
694 offset = ktime_to_ns(tick_period) >> 1;
695 do_div(offset, num_possible_cpus());
696 offset *= smp_processor_id();
697 hrtimer_add_expires_ns(&ts->sched_timer, offset);
700 hrtimer_forward(&ts->sched_timer, now, tick_period);
701 hrtimer_start_expires(&ts->sched_timer,
702 HRTIMER_MODE_ABS_PINNED);
703 /* Check, if the timer was already in the past */
704 if (hrtimer_active(&ts->sched_timer))
710 if (tick_nohz_enabled)
711 ts->nohz_mode = NOHZ_MODE_HIGHRES;
714 #endif /* HIGH_RES_TIMERS */
716 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
717 void tick_cancel_sched_timer(int cpu)
719 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
721 # ifdef CONFIG_HIGH_RES_TIMERS
722 if (ts->sched_timer.base)
723 hrtimer_cancel(&ts->sched_timer);
726 ts->nohz_mode = NOHZ_MODE_INACTIVE;
731 * Async notification about clocksource changes
733 void tick_clock_notify(void)
737 for_each_possible_cpu(cpu)
738 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
742 * Async notification about clock event changes
744 void tick_oneshot_notify(void)
746 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
748 set_bit(0, &ts->check_clocks);
752 * Check, if a change happened, which makes oneshot possible.
754 * Called cyclic from the hrtimer softirq (driven by the timer
755 * softirq) allow_nohz signals, that we can switch into low-res nohz
756 * mode, because high resolution timers are disabled (either compile
759 int tick_check_oneshot_change(int allow_nohz)
761 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
763 if (!test_and_clear_bit(0, &ts->check_clocks))
766 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
769 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
775 tick_nohz_switch_to_nohz();