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 /* Keep the tick_next_period variable up to date */
80 tick_next_period = ktime_add(last_jiffies_update, tick_period);
82 write_sequnlock(&xtime_lock);
86 * Initialize and return retrieve the jiffies update.
88 static ktime_t tick_init_jiffy_update(void)
92 write_seqlock(&xtime_lock);
93 /* Did we start the jiffies update yet ? */
94 if (last_jiffies_update.tv64 == 0)
95 last_jiffies_update = tick_next_period;
96 period = last_jiffies_update;
97 write_sequnlock(&xtime_lock);
102 * NOHZ - aka dynamic tick functionality
108 static int tick_nohz_enabled __read_mostly = 1;
111 * Enable / Disable tickless mode
113 static int __init setup_tick_nohz(char *str)
115 if (!strcmp(str, "off"))
116 tick_nohz_enabled = 0;
117 else if (!strcmp(str, "on"))
118 tick_nohz_enabled = 1;
124 __setup("nohz=", setup_tick_nohz);
127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
129 * Called from interrupt entry when the CPU was idle
131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
133 * value. We do this unconditionally on any cpu, as we don't know whether the
134 * cpu, which has the update task assigned is in a long sleep.
136 void tick_nohz_update_jiffies(void)
138 int cpu = smp_processor_id();
139 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
143 if (!ts->tick_stopped)
146 cpu_clear(cpu, nohz_cpu_mask);
148 ts->idle_waketime = now;
150 local_irq_save(flags);
151 tick_do_update_jiffies64(now);
152 local_irq_restore(flags);
154 touch_softlockup_watchdog();
157 void tick_nohz_stop_idle(int cpu)
159 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
161 if (ts->idle_active) {
164 delta = ktime_sub(now, ts->idle_entrytime);
165 ts->idle_lastupdate = now;
166 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
169 sched_clock_idle_wakeup_event(0);
173 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
178 if (ts->idle_active) {
179 delta = ktime_sub(now, ts->idle_entrytime);
180 ts->idle_lastupdate = now;
181 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
183 ts->idle_entrytime = now;
185 sched_clock_idle_sleep_event();
189 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
191 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
193 *last_update_time = ktime_to_us(ts->idle_lastupdate);
194 return ktime_to_us(ts->idle_sleeptime);
198 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
200 * When the next event is more than a tick into the future, stop the idle tick
201 * Called either from the idle loop or from irq_exit() when an idle period was
202 * just interrupted by an interrupt which did not cause a reschedule.
204 void tick_nohz_stop_sched_tick(int inidle)
206 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
207 struct tick_sched *ts;
208 ktime_t last_update, expires, now;
209 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
212 local_irq_save(flags);
214 cpu = smp_processor_id();
215 ts = &per_cpu(tick_cpu_sched, cpu);
216 now = tick_nohz_start_idle(ts);
219 * If this cpu is offline and it is the one which updates
220 * jiffies, then give up the assignment and let it be taken by
221 * the cpu which runs the tick timer next. If we don't drop
222 * this here the jiffies might be stale and do_timer() never
225 if (unlikely(!cpu_online(cpu))) {
226 if (cpu == tick_do_timer_cpu)
227 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
230 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
233 if (!inidle && !ts->inidle)
241 if (unlikely(local_softirq_pending())) {
242 static int ratelimit;
244 if (ratelimit < 10) {
245 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
246 local_softirq_pending());
253 /* Read jiffies and the time when jiffies were updated last */
255 seq = read_seqbegin(&xtime_lock);
256 last_update = last_jiffies_update;
257 last_jiffies = jiffies;
258 } while (read_seqretry(&xtime_lock, seq));
260 /* Get the next timer wheel timer */
261 next_jiffies = get_next_timer_interrupt(last_jiffies);
262 delta_jiffies = next_jiffies - last_jiffies;
264 if (rcu_needs_cpu(cpu))
267 * Do not stop the tick, if we are only one off
268 * or if the cpu is required for rcu
270 if (!ts->tick_stopped && delta_jiffies == 1)
273 /* Schedule the tick, if we are at least one jiffie off */
274 if ((long)delta_jiffies >= 1) {
276 if (delta_jiffies > 1)
277 cpu_set(cpu, nohz_cpu_mask);
279 * nohz_stop_sched_tick can be called several times before
280 * the nohz_restart_sched_tick is called. This happens when
281 * interrupts arrive which do not cause a reschedule. In the
282 * first call we save the current tick time, so we can restart
283 * the scheduler tick in nohz_restart_sched_tick.
285 if (!ts->tick_stopped) {
286 if (select_nohz_load_balancer(1)) {
288 * sched tick not stopped!
290 cpu_clear(cpu, nohz_cpu_mask);
294 ts->idle_tick = ts->sched_timer.expires;
295 ts->tick_stopped = 1;
296 ts->idle_jiffies = last_jiffies;
301 * If this cpu is the one which updates jiffies, then
302 * give up the assignment and let it be taken by the
303 * cpu which runs the tick timer next, which might be
304 * this cpu as well. If we don't drop this here the
305 * jiffies might be stale and do_timer() never
308 if (cpu == tick_do_timer_cpu)
309 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
314 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
315 * there is no timer pending or at least extremly far
316 * into the future (12 days for HZ=1000). In this case
317 * we simply stop the tick timer:
319 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
320 ts->idle_expires.tv64 = KTIME_MAX;
321 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
322 hrtimer_cancel(&ts->sched_timer);
327 * calculate the expiry time for the next timer wheel
330 expires = ktime_add_ns(last_update, tick_period.tv64 *
332 ts->idle_expires = expires;
334 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
335 hrtimer_start(&ts->sched_timer, expires,
337 /* Check, if the timer was already in the past */
338 if (hrtimer_active(&ts->sched_timer))
340 } else if (!tick_program_event(expires, 0))
343 * We are past the event already. So we crossed a
344 * jiffie boundary. Update jiffies and raise the
347 tick_do_update_jiffies64(ktime_get());
348 cpu_clear(cpu, nohz_cpu_mask);
350 raise_softirq_irqoff(TIMER_SOFTIRQ);
352 ts->next_jiffies = next_jiffies;
353 ts->last_jiffies = last_jiffies;
354 ts->sleep_length = ktime_sub(dev->next_event, now);
356 local_irq_restore(flags);
360 * tick_nohz_get_sleep_length - return the length of the current sleep
362 * Called from power state control code with interrupts disabled
364 ktime_t tick_nohz_get_sleep_length(void)
366 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
368 return ts->sleep_length;
372 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
374 * Restart the idle tick when the CPU is woken up from idle
376 void tick_nohz_restart_sched_tick(void)
378 int cpu = smp_processor_id();
379 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
384 tick_nohz_stop_idle(cpu);
386 if (!ts->inidle || !ts->tick_stopped) {
396 /* Update jiffies first */
397 select_nohz_load_balancer(0);
399 tick_do_update_jiffies64(now);
400 cpu_clear(cpu, nohz_cpu_mask);
403 * We stopped the tick in idle. Update process times would miss the
404 * time we slept as update_process_times does only a 1 tick
405 * accounting. Enforce that this is accounted to idle !
407 ticks = jiffies - ts->idle_jiffies;
409 * We might be one off. Do not randomly account a huge number of ticks!
411 if (ticks && ticks < LONG_MAX) {
412 add_preempt_count(HARDIRQ_OFFSET);
413 account_system_time(current, HARDIRQ_OFFSET,
414 jiffies_to_cputime(ticks));
415 sub_preempt_count(HARDIRQ_OFFSET);
418 touch_softlockup_watchdog();
420 * Cancel the scheduled timer and restore the tick
422 ts->tick_stopped = 0;
423 ts->idle_exittime = now;
424 hrtimer_cancel(&ts->sched_timer);
425 ts->sched_timer.expires = ts->idle_tick;
428 /* Forward the time to expire in the future */
429 hrtimer_forward(&ts->sched_timer, now, tick_period);
431 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
432 hrtimer_start(&ts->sched_timer,
433 ts->sched_timer.expires,
435 /* Check, if the timer was already in the past */
436 if (hrtimer_active(&ts->sched_timer))
439 if (!tick_program_event(ts->sched_timer.expires, 0))
442 /* Update jiffies and reread time */
443 tick_do_update_jiffies64(now);
449 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
451 hrtimer_forward(&ts->sched_timer, now, tick_period);
452 return tick_program_event(ts->sched_timer.expires, 0);
456 * The nohz low res interrupt handler
458 static void tick_nohz_handler(struct clock_event_device *dev)
460 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
461 struct pt_regs *regs = get_irq_regs();
462 int cpu = smp_processor_id();
463 ktime_t now = ktime_get();
465 dev->next_event.tv64 = KTIME_MAX;
468 * Check if the do_timer duty was dropped. We don't care about
469 * concurrency: This happens only when the cpu in charge went
470 * into a long sleep. If two cpus happen to assign themself to
471 * this duty, then the jiffies update is still serialized by
474 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
475 tick_do_timer_cpu = cpu;
477 /* Check, if the jiffies need an update */
478 if (tick_do_timer_cpu == cpu)
479 tick_do_update_jiffies64(now);
482 * When we are idle and the tick is stopped, we have to touch
483 * the watchdog as we might not schedule for a really long
484 * time. This happens on complete idle SMP systems while
485 * waiting on the login prompt. We also increment the "start
486 * of idle" jiffy stamp so the idle accounting adjustment we
487 * do when we go busy again does not account too much ticks.
489 if (ts->tick_stopped) {
490 touch_softlockup_watchdog();
494 update_process_times(user_mode(regs));
495 profile_tick(CPU_PROFILING);
497 /* Do not restart, when we are in the idle loop */
498 if (ts->tick_stopped)
501 while (tick_nohz_reprogram(ts, now)) {
503 tick_do_update_jiffies64(now);
508 * tick_nohz_switch_to_nohz - switch to nohz mode
510 static void tick_nohz_switch_to_nohz(void)
512 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
515 if (!tick_nohz_enabled)
519 if (tick_switch_to_oneshot(tick_nohz_handler)) {
524 ts->nohz_mode = NOHZ_MODE_LOWRES;
527 * Recycle the hrtimer in ts, so we can share the
528 * hrtimer_forward with the highres code.
530 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
531 /* Get the next period */
532 next = tick_init_jiffy_update();
535 ts->sched_timer.expires = next;
536 if (!tick_program_event(next, 0))
538 next = ktime_add(next, tick_period);
542 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
548 static inline void tick_nohz_switch_to_nohz(void) { }
553 * High resolution timer specific code
555 #ifdef CONFIG_HIGH_RES_TIMERS
557 * We rearm the timer until we get disabled by the idle code.
558 * Called with interrupts disabled and timer->base->cpu_base->lock held.
560 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
562 struct tick_sched *ts =
563 container_of(timer, struct tick_sched, sched_timer);
564 struct pt_regs *regs = get_irq_regs();
565 ktime_t now = ktime_get();
566 int cpu = smp_processor_id();
570 * Check if the do_timer duty was dropped. We don't care about
571 * concurrency: This happens only when the cpu in charge went
572 * into a long sleep. If two cpus happen to assign themself to
573 * this duty, then the jiffies update is still serialized by
576 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
577 tick_do_timer_cpu = cpu;
580 /* Check, if the jiffies need an update */
581 if (tick_do_timer_cpu == cpu)
582 tick_do_update_jiffies64(now);
585 * Do not call, when we are not in irq context and have
586 * no valid regs pointer
590 * When we are idle and the tick is stopped, we have to touch
591 * the watchdog as we might not schedule for a really long
592 * time. This happens on complete idle SMP systems while
593 * waiting on the login prompt. We also increment the "start of
594 * idle" jiffy stamp so the idle accounting adjustment we do
595 * when we go busy again does not account too much ticks.
597 if (ts->tick_stopped) {
598 touch_softlockup_watchdog();
601 update_process_times(user_mode(regs));
602 profile_tick(CPU_PROFILING);
605 /* Do not restart, when we are in the idle loop */
606 if (ts->tick_stopped)
607 return HRTIMER_NORESTART;
609 hrtimer_forward(timer, now, tick_period);
611 return HRTIMER_RESTART;
615 * tick_setup_sched_timer - setup the tick emulation timer
617 void tick_setup_sched_timer(void)
619 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
620 ktime_t now = ktime_get();
624 * Emulate tick processing via per-CPU hrtimers:
626 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
627 ts->sched_timer.function = tick_sched_timer;
628 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
630 /* Get the next period (per cpu) */
631 ts->sched_timer.expires = tick_init_jiffy_update();
632 offset = ktime_to_ns(tick_period) >> 1;
633 do_div(offset, num_possible_cpus());
634 offset *= smp_processor_id();
635 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
638 hrtimer_forward(&ts->sched_timer, now, tick_period);
639 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
641 /* Check, if the timer was already in the past */
642 if (hrtimer_active(&ts->sched_timer))
648 if (tick_nohz_enabled)
649 ts->nohz_mode = NOHZ_MODE_HIGHRES;
652 #endif /* HIGH_RES_TIMERS */
654 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
655 void tick_cancel_sched_timer(int cpu)
657 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
659 # ifdef CONFIG_HIGH_RES_TIMERS
660 if (ts->sched_timer.base)
661 hrtimer_cancel(&ts->sched_timer);
664 ts->nohz_mode = NOHZ_MODE_INACTIVE;
669 * Async notification about clocksource changes
671 void tick_clock_notify(void)
675 for_each_possible_cpu(cpu)
676 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
680 * Async notification about clock event changes
682 void tick_oneshot_notify(void)
684 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
686 set_bit(0, &ts->check_clocks);
690 * Check, if a change happened, which makes oneshot possible.
692 * Called cyclic from the hrtimer softirq (driven by the timer
693 * softirq) allow_nohz signals, that we can switch into low-res nohz
694 * mode, because high resolution timers are disabled (either compile
697 int tick_check_oneshot_change(int allow_nohz)
699 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
701 if (!test_and_clear_bit(0, &ts->check_clocks))
704 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
707 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
713 tick_nohz_switch_to_nohz();