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>
24 #include <asm/irq_regs.h>
26 #include "tick-internal.h"
29 * Per cpu nohz control structure
31 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
36 static ktime_t last_jiffies_update;
38 struct tick_sched *tick_get_tick_sched(int cpu)
40 return &per_cpu(tick_cpu_sched, cpu);
44 * Must be called with interrupts disabled !
46 static void tick_do_update_jiffies64(ktime_t now)
48 unsigned long ticks = 0;
52 * Do a quick check without holding xtime_lock:
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 < tick_period.tv64)
58 /* Reevalute with xtime_lock held */
59 write_seqlock(&xtime_lock);
61 delta = ktime_sub(now, last_jiffies_update);
62 if (delta.tv64 >= tick_period.tv64) {
64 delta = ktime_sub(delta, tick_period);
65 last_jiffies_update = ktime_add(last_jiffies_update,
68 /* Slow path for long timeouts */
69 if (unlikely(delta.tv64 >= tick_period.tv64)) {
70 s64 incr = ktime_to_ns(tick_period);
72 ticks = ktime_divns(delta, incr);
74 last_jiffies_update = ktime_add_ns(last_jiffies_update,
79 write_sequnlock(&xtime_lock);
83 * Initialize and return retrieve the jiffies update.
85 static ktime_t tick_init_jiffy_update(void)
89 write_seqlock(&xtime_lock);
90 /* Did we start the jiffies update yet ? */
91 if (last_jiffies_update.tv64 == 0)
92 last_jiffies_update = tick_next_period;
93 period = last_jiffies_update;
94 write_sequnlock(&xtime_lock);
99 * NOHZ - aka dynamic tick functionality
105 static int tick_nohz_enabled __read_mostly = 1;
108 * Enable / Disable tickless mode
110 static int __init setup_tick_nohz(char *str)
112 if (!strcmp(str, "off"))
113 tick_nohz_enabled = 0;
114 else if (!strcmp(str, "on"))
115 tick_nohz_enabled = 1;
121 __setup("nohz=", setup_tick_nohz);
124 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
126 * Called from interrupt entry when the CPU was idle
128 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
129 * must be updated. Otherwise an interrupt handler could use a stale jiffy
130 * value. We do this unconditionally on any cpu, as we don't know whether the
131 * cpu, which has the update task assigned is in a long sleep.
133 void tick_nohz_update_jiffies(void)
135 int cpu = smp_processor_id();
136 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
140 if (!ts->tick_stopped)
143 cpu_clear(cpu, nohz_cpu_mask);
145 ts->idle_waketime = now;
147 local_irq_save(flags);
148 tick_do_update_jiffies64(now);
149 local_irq_restore(flags);
151 touch_softlockup_watchdog();
154 void tick_nohz_stop_idle(int cpu)
156 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
158 if (ts->idle_active) {
161 delta = ktime_sub(now, ts->idle_entrytime);
162 ts->idle_lastupdate = now;
163 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
168 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
173 if (ts->idle_active) {
174 delta = ktime_sub(now, ts->idle_entrytime);
175 ts->idle_lastupdate = now;
176 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
178 ts->idle_entrytime = now;
183 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
185 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
187 *last_update_time = ktime_to_us(ts->idle_lastupdate);
188 return ktime_to_us(ts->idle_sleeptime);
192 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
194 * When the next event is more than a tick into the future, stop the idle tick
195 * Called either from the idle loop or from irq_exit() when an idle period was
196 * just interrupted by an interrupt which did not cause a reschedule.
198 void tick_nohz_stop_sched_tick(int inidle)
200 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
201 struct tick_sched *ts;
202 ktime_t last_update, expires, now;
203 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
206 local_irq_save(flags);
208 cpu = smp_processor_id();
209 ts = &per_cpu(tick_cpu_sched, cpu);
210 now = tick_nohz_start_idle(ts);
213 * If this cpu is offline and it is the one which updates
214 * jiffies, then give up the assignment and let it be taken by
215 * the cpu which runs the tick timer next. If we don't drop
216 * this here the jiffies might be stale and do_timer() never
219 if (unlikely(!cpu_online(cpu))) {
220 if (cpu == tick_do_timer_cpu)
221 tick_do_timer_cpu = -1;
224 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
227 if (!inidle && !ts->inidle)
235 if (unlikely(local_softirq_pending())) {
236 static int ratelimit;
238 if (ratelimit < 10) {
239 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
240 local_softirq_pending());
247 /* Read jiffies and the time when jiffies were updated last */
249 seq = read_seqbegin(&xtime_lock);
250 last_update = last_jiffies_update;
251 last_jiffies = jiffies;
252 } while (read_seqretry(&xtime_lock, seq));
254 /* Get the next timer wheel timer */
255 next_jiffies = get_next_timer_interrupt(last_jiffies);
256 delta_jiffies = next_jiffies - last_jiffies;
258 if (rcu_needs_cpu(cpu))
261 * Do not stop the tick, if we are only one off
262 * or if the cpu is required for rcu
264 if (!ts->tick_stopped && delta_jiffies == 1)
267 /* Schedule the tick, if we are at least one jiffie off */
268 if ((long)delta_jiffies >= 1) {
270 if (delta_jiffies > 1)
271 cpu_set(cpu, nohz_cpu_mask);
273 * nohz_stop_sched_tick can be called several times before
274 * the nohz_restart_sched_tick is called. This happens when
275 * interrupts arrive which do not cause a reschedule. In the
276 * first call we save the current tick time, so we can restart
277 * the scheduler tick in nohz_restart_sched_tick.
279 if (!ts->tick_stopped) {
280 if (select_nohz_load_balancer(1)) {
282 * sched tick not stopped!
284 cpu_clear(cpu, nohz_cpu_mask);
288 ts->idle_tick = ts->sched_timer.expires;
289 ts->tick_stopped = 1;
290 ts->idle_jiffies = last_jiffies;
292 sched_clock_tick_stop(cpu);
296 * If this cpu is the one which updates jiffies, then
297 * give up the assignment and let it be taken by the
298 * cpu which runs the tick timer next, which might be
299 * this cpu as well. If we don't drop this here the
300 * jiffies might be stale and do_timer() never
303 if (cpu == tick_do_timer_cpu)
304 tick_do_timer_cpu = -1;
309 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
310 * there is no timer pending or at least extremly far
311 * into the future (12 days for HZ=1000). In this case
312 * we simply stop the tick timer:
314 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
315 ts->idle_expires.tv64 = KTIME_MAX;
316 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
317 hrtimer_cancel(&ts->sched_timer);
322 * calculate the expiry time for the next timer wheel
325 expires = ktime_add_ns(last_update, tick_period.tv64 *
327 ts->idle_expires = expires;
329 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
330 hrtimer_start(&ts->sched_timer, expires,
332 /* Check, if the timer was already in the past */
333 if (hrtimer_active(&ts->sched_timer))
335 } else if (!tick_program_event(expires, 0))
338 * We are past the event already. So we crossed a
339 * jiffie boundary. Update jiffies and raise the
342 tick_do_update_jiffies64(ktime_get());
343 cpu_clear(cpu, nohz_cpu_mask);
345 raise_softirq_irqoff(TIMER_SOFTIRQ);
347 ts->next_jiffies = next_jiffies;
348 ts->last_jiffies = last_jiffies;
349 ts->sleep_length = ktime_sub(dev->next_event, now);
351 local_irq_restore(flags);
355 * tick_nohz_get_sleep_length - return the length of the current sleep
357 * Called from power state control code with interrupts disabled
359 ktime_t tick_nohz_get_sleep_length(void)
361 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
363 return ts->sleep_length;
367 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
369 * Restart the idle tick when the CPU is woken up from idle
371 void tick_nohz_restart_sched_tick(void)
373 int cpu = smp_processor_id();
374 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
379 tick_nohz_stop_idle(cpu);
381 if (!ts->inidle || !ts->tick_stopped) {
391 /* Update jiffies first */
392 select_nohz_load_balancer(0);
394 tick_do_update_jiffies64(now);
395 sched_clock_tick_start(cpu);
396 cpu_clear(cpu, nohz_cpu_mask);
399 * We stopped the tick in idle. Update process times would miss the
400 * time we slept as update_process_times does only a 1 tick
401 * accounting. Enforce that this is accounted to idle !
403 ticks = jiffies - ts->idle_jiffies;
405 * We might be one off. Do not randomly account a huge number of ticks!
407 if (ticks && ticks < LONG_MAX) {
408 add_preempt_count(HARDIRQ_OFFSET);
409 account_system_time(current, HARDIRQ_OFFSET,
410 jiffies_to_cputime(ticks));
411 sub_preempt_count(HARDIRQ_OFFSET);
414 touch_softlockup_watchdog();
416 * Cancel the scheduled timer and restore the tick
418 ts->tick_stopped = 0;
419 ts->idle_exittime = now;
420 hrtimer_cancel(&ts->sched_timer);
421 ts->sched_timer.expires = ts->idle_tick;
424 /* Forward the time to expire in the future */
425 hrtimer_forward(&ts->sched_timer, now, tick_period);
427 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
428 hrtimer_start(&ts->sched_timer,
429 ts->sched_timer.expires,
431 /* Check, if the timer was already in the past */
432 if (hrtimer_active(&ts->sched_timer))
435 if (!tick_program_event(ts->sched_timer.expires, 0))
438 /* Update jiffies and reread time */
439 tick_do_update_jiffies64(now);
445 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
447 hrtimer_forward(&ts->sched_timer, now, tick_period);
448 return tick_program_event(ts->sched_timer.expires, 0);
452 * The nohz low res interrupt handler
454 static void tick_nohz_handler(struct clock_event_device *dev)
456 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
457 struct pt_regs *regs = get_irq_regs();
458 int cpu = smp_processor_id();
459 ktime_t now = ktime_get();
461 dev->next_event.tv64 = KTIME_MAX;
464 * Check if the do_timer duty was dropped. We don't care about
465 * concurrency: This happens only when the cpu in charge went
466 * into a long sleep. If two cpus happen to assign themself to
467 * this duty, then the jiffies update is still serialized by
470 if (unlikely(tick_do_timer_cpu == -1))
471 tick_do_timer_cpu = cpu;
473 /* Check, if the jiffies need an update */
474 if (tick_do_timer_cpu == cpu)
475 tick_do_update_jiffies64(now);
478 * When we are idle and the tick is stopped, we have to touch
479 * the watchdog as we might not schedule for a really long
480 * time. This happens on complete idle SMP systems while
481 * waiting on the login prompt. We also increment the "start
482 * of idle" jiffy stamp so the idle accounting adjustment we
483 * do when we go busy again does not account too much ticks.
485 if (ts->tick_stopped) {
486 touch_softlockup_watchdog();
490 update_process_times(user_mode(regs));
491 profile_tick(CPU_PROFILING);
493 /* Do not restart, when we are in the idle loop */
494 if (ts->tick_stopped)
497 while (tick_nohz_reprogram(ts, now)) {
499 tick_do_update_jiffies64(now);
504 * tick_nohz_switch_to_nohz - switch to nohz mode
506 static void tick_nohz_switch_to_nohz(void)
508 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
511 if (!tick_nohz_enabled)
515 if (tick_switch_to_oneshot(tick_nohz_handler)) {
520 ts->nohz_mode = NOHZ_MODE_LOWRES;
523 * Recycle the hrtimer in ts, so we can share the
524 * hrtimer_forward with the highres code.
526 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
527 /* Get the next period */
528 next = tick_init_jiffy_update();
531 ts->sched_timer.expires = next;
532 if (!tick_program_event(next, 0))
534 next = ktime_add(next, tick_period);
538 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
544 static inline void tick_nohz_switch_to_nohz(void) { }
549 * High resolution timer specific code
551 #ifdef CONFIG_HIGH_RES_TIMERS
553 * We rearm the timer until we get disabled by the idle code.
554 * Called with interrupts disabled and timer->base->cpu_base->lock held.
556 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
558 struct tick_sched *ts =
559 container_of(timer, struct tick_sched, sched_timer);
560 struct pt_regs *regs = get_irq_regs();
561 ktime_t now = ktime_get();
562 int cpu = smp_processor_id();
566 * Check if the do_timer duty was dropped. We don't care about
567 * concurrency: This happens only when the cpu in charge went
568 * into a long sleep. If two cpus happen to assign themself to
569 * this duty, then the jiffies update is still serialized by
572 if (unlikely(tick_do_timer_cpu == -1))
573 tick_do_timer_cpu = cpu;
576 /* Check, if the jiffies need an update */
577 if (tick_do_timer_cpu == cpu)
578 tick_do_update_jiffies64(now);
581 * Do not call, when we are not in irq context and have
582 * no valid regs pointer
586 * When we are idle and the tick is stopped, we have to touch
587 * the watchdog as we might not schedule for a really long
588 * time. This happens on complete idle SMP systems while
589 * waiting on the login prompt. We also increment the "start of
590 * idle" jiffy stamp so the idle accounting adjustment we do
591 * when we go busy again does not account too much ticks.
593 if (ts->tick_stopped) {
594 touch_softlockup_watchdog();
597 update_process_times(user_mode(regs));
598 profile_tick(CPU_PROFILING);
601 /* Do not restart, when we are in the idle loop */
602 if (ts->tick_stopped)
603 return HRTIMER_NORESTART;
605 hrtimer_forward(timer, now, tick_period);
607 return HRTIMER_RESTART;
611 * tick_setup_sched_timer - setup the tick emulation timer
613 void tick_setup_sched_timer(void)
615 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
616 ktime_t now = ktime_get();
620 * Emulate tick processing via per-CPU hrtimers:
622 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
623 ts->sched_timer.function = tick_sched_timer;
624 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
626 /* Get the next period (per cpu) */
627 ts->sched_timer.expires = tick_init_jiffy_update();
628 offset = ktime_to_ns(tick_period) >> 1;
629 do_div(offset, num_possible_cpus());
630 offset *= smp_processor_id();
631 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
634 hrtimer_forward(&ts->sched_timer, now, tick_period);
635 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
637 /* Check, if the timer was already in the past */
638 if (hrtimer_active(&ts->sched_timer))
644 if (tick_nohz_enabled)
645 ts->nohz_mode = NOHZ_MODE_HIGHRES;
649 void tick_cancel_sched_timer(int cpu)
651 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
653 if (ts->sched_timer.base)
654 hrtimer_cancel(&ts->sched_timer);
656 ts->nohz_mode = NOHZ_MODE_INACTIVE;
658 #endif /* HIGH_RES_TIMERS */
661 * Async notification about clocksource changes
663 void tick_clock_notify(void)
667 for_each_possible_cpu(cpu)
668 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
672 * Async notification about clock event changes
674 void tick_oneshot_notify(void)
676 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
678 set_bit(0, &ts->check_clocks);
682 * Check, if a change happened, which makes oneshot possible.
684 * Called cyclic from the hrtimer softirq (driven by the timer
685 * softirq) allow_nohz signals, that we can switch into low-res nohz
686 * mode, because high resolution timers are disabled (either compile
689 int tick_check_oneshot_change(int allow_nohz)
691 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
693 if (!test_and_clear_bit(0, &ts->check_clocks))
696 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
699 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
705 tick_nohz_switch_to_nohz();