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;
51 /* Reevalute with xtime_lock held */
52 write_seqlock(&xtime_lock);
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 >= tick_period.tv64) {
57 delta = ktime_sub(delta, tick_period);
58 last_jiffies_update = ktime_add(last_jiffies_update,
61 /* Slow path for long timeouts */
62 if (unlikely(delta.tv64 >= tick_period.tv64)) {
63 s64 incr = ktime_to_ns(tick_period);
65 ticks = ktime_divns(delta, incr);
67 last_jiffies_update = ktime_add_ns(last_jiffies_update,
72 write_sequnlock(&xtime_lock);
76 * Initialize and return retrieve the jiffies update.
78 static ktime_t tick_init_jiffy_update(void)
82 write_seqlock(&xtime_lock);
83 /* Did we start the jiffies update yet ? */
84 if (last_jiffies_update.tv64 == 0)
85 last_jiffies_update = tick_next_period;
86 period = last_jiffies_update;
87 write_sequnlock(&xtime_lock);
92 * NOHZ - aka dynamic tick functionality
98 static int tick_nohz_enabled __read_mostly = 1;
101 * Enable / Disable tickless mode
103 static int __init setup_tick_nohz(char *str)
105 if (!strcmp(str, "off"))
106 tick_nohz_enabled = 0;
107 else if (!strcmp(str, "on"))
108 tick_nohz_enabled = 1;
114 __setup("nohz=", setup_tick_nohz);
117 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
119 * Called from interrupt entry when the CPU was idle
121 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
122 * must be updated. Otherwise an interrupt handler could use a stale jiffy
123 * value. We do this unconditionally on any cpu, as we don't know whether the
124 * cpu, which has the update task assigned is in a long sleep.
126 void tick_nohz_update_jiffies(void)
128 int cpu = smp_processor_id();
129 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
133 if (!ts->tick_stopped)
136 touch_softlockup_watchdog();
138 cpu_clear(cpu, nohz_cpu_mask);
141 local_irq_save(flags);
142 tick_do_update_jiffies64(now);
143 local_irq_restore(flags);
146 void tick_nohz_stop_idle(int cpu)
148 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
150 if (ts->idle_active) {
153 delta = ktime_sub(now, ts->idle_entrytime);
154 ts->idle_lastupdate = now;
155 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
160 static ktime_t tick_nohz_start_idle(int cpu)
162 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
166 if (ts->idle_active) {
167 delta = ktime_sub(now, ts->idle_entrytime);
168 ts->idle_lastupdate = now;
169 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
171 ts->idle_entrytime = now;
176 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
178 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
180 *last_update_time = ktime_to_us(ts->idle_lastupdate);
181 return ktime_to_us(ts->idle_sleeptime);
185 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
187 * When the next event is more than a tick into the future, stop the idle tick
188 * Called either from the idle loop or from irq_exit() when an idle period was
189 * just interrupted by an interrupt which did not cause a reschedule.
191 void tick_nohz_stop_sched_tick(void)
193 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
194 unsigned long rt_jiffies;
195 struct tick_sched *ts;
196 ktime_t last_update, expires, now;
197 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
200 local_irq_save(flags);
202 cpu = smp_processor_id();
203 now = tick_nohz_start_idle(cpu);
204 ts = &per_cpu(tick_cpu_sched, cpu);
207 * If this cpu is offline and it is the one which updates
208 * jiffies, then give up the assignment and let it be taken by
209 * the cpu which runs the tick timer next. If we don't drop
210 * this here the jiffies might be stale and do_timer() never
213 if (unlikely(!cpu_online(cpu))) {
214 if (cpu == tick_do_timer_cpu)
215 tick_do_timer_cpu = -1;
218 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
224 cpu = smp_processor_id();
225 if (unlikely(local_softirq_pending())) {
226 static int ratelimit;
228 if (ratelimit < 10) {
229 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
230 local_softirq_pending());
236 /* Read jiffies and the time when jiffies were updated last */
238 seq = read_seqbegin(&xtime_lock);
239 last_update = last_jiffies_update;
240 last_jiffies = jiffies;
241 } while (read_seqretry(&xtime_lock, seq));
243 /* Get the next timer wheel timer */
244 next_jiffies = get_next_timer_interrupt(last_jiffies);
245 delta_jiffies = next_jiffies - last_jiffies;
247 rt_jiffies = rt_needs_cpu(cpu);
248 if (rt_jiffies && rt_jiffies < delta_jiffies)
249 delta_jiffies = rt_jiffies;
251 if (rcu_needs_cpu(cpu))
254 * Do not stop the tick, if we are only one off
255 * or if the cpu is required for rcu
257 if (!ts->tick_stopped && delta_jiffies == 1)
260 /* Schedule the tick, if we are at least one jiffie off */
261 if ((long)delta_jiffies >= 1) {
263 if (delta_jiffies > 1)
264 cpu_set(cpu, nohz_cpu_mask);
266 * nohz_stop_sched_tick can be called several times before
267 * the nohz_restart_sched_tick is called. This happens when
268 * interrupts arrive which do not cause a reschedule. In the
269 * first call we save the current tick time, so we can restart
270 * the scheduler tick in nohz_restart_sched_tick.
272 if (!ts->tick_stopped) {
273 if (select_nohz_load_balancer(1)) {
275 * sched tick not stopped!
277 cpu_clear(cpu, nohz_cpu_mask);
281 ts->idle_tick = ts->sched_timer.expires;
282 ts->tick_stopped = 1;
283 ts->idle_jiffies = last_jiffies;
287 * If this cpu is the one which updates jiffies, then
288 * give up the assignment and let it be taken by the
289 * cpu which runs the tick timer next, which might be
290 * this cpu as well. If we don't drop this here the
291 * jiffies might be stale and do_timer() never
294 if (cpu == tick_do_timer_cpu)
295 tick_do_timer_cpu = -1;
300 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
301 * there is no timer pending or at least extremly far
302 * into the future (12 days for HZ=1000). In this case
303 * we simply stop the tick timer:
305 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
306 ts->idle_expires.tv64 = KTIME_MAX;
307 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
308 hrtimer_cancel(&ts->sched_timer);
313 * calculate the expiry time for the next timer wheel
316 expires = ktime_add_ns(last_update, tick_period.tv64 *
318 ts->idle_expires = expires;
320 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
321 hrtimer_start(&ts->sched_timer, expires,
323 /* Check, if the timer was already in the past */
324 if (hrtimer_active(&ts->sched_timer))
326 } else if (!tick_program_event(expires, 0))
329 * We are past the event already. So we crossed a
330 * jiffie boundary. Update jiffies and raise the
333 tick_do_update_jiffies64(ktime_get());
334 cpu_clear(cpu, nohz_cpu_mask);
336 raise_softirq_irqoff(TIMER_SOFTIRQ);
338 ts->next_jiffies = next_jiffies;
339 ts->last_jiffies = last_jiffies;
340 ts->sleep_length = ktime_sub(dev->next_event, now);
342 local_irq_restore(flags);
346 * tick_nohz_get_sleep_length - return the length of the current sleep
348 * Called from power state control code with interrupts disabled
350 ktime_t tick_nohz_get_sleep_length(void)
352 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
354 return ts->sleep_length;
358 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
360 * Restart the idle tick when the CPU is woken up from idle
362 void tick_nohz_restart_sched_tick(void)
364 int cpu = smp_processor_id();
365 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
370 tick_nohz_stop_idle(cpu);
372 if (!ts->tick_stopped) {
377 /* Update jiffies first */
378 select_nohz_load_balancer(0);
380 tick_do_update_jiffies64(now);
381 cpu_clear(cpu, nohz_cpu_mask);
384 * We stopped the tick in idle. Update process times would miss the
385 * time we slept as update_process_times does only a 1 tick
386 * accounting. Enforce that this is accounted to idle !
388 ticks = jiffies - ts->idle_jiffies;
390 * We might be one off. Do not randomly account a huge number of ticks!
392 if (ticks && ticks < LONG_MAX) {
393 add_preempt_count(HARDIRQ_OFFSET);
394 account_system_time(current, HARDIRQ_OFFSET,
395 jiffies_to_cputime(ticks));
396 sub_preempt_count(HARDIRQ_OFFSET);
400 * Cancel the scheduled timer and restore the tick
402 ts->tick_stopped = 0;
403 hrtimer_cancel(&ts->sched_timer);
404 ts->sched_timer.expires = ts->idle_tick;
407 /* Forward the time to expire in the future */
408 hrtimer_forward(&ts->sched_timer, now, tick_period);
410 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
411 hrtimer_start(&ts->sched_timer,
412 ts->sched_timer.expires,
414 /* Check, if the timer was already in the past */
415 if (hrtimer_active(&ts->sched_timer))
418 if (!tick_program_event(ts->sched_timer.expires, 0))
421 /* Update jiffies and reread time */
422 tick_do_update_jiffies64(now);
428 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
430 hrtimer_forward(&ts->sched_timer, now, tick_period);
431 return tick_program_event(ts->sched_timer.expires, 0);
435 * The nohz low res interrupt handler
437 static void tick_nohz_handler(struct clock_event_device *dev)
439 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
440 struct pt_regs *regs = get_irq_regs();
441 int cpu = smp_processor_id();
442 ktime_t now = ktime_get();
444 dev->next_event.tv64 = KTIME_MAX;
447 * Check if the do_timer duty was dropped. We don't care about
448 * concurrency: This happens only when the cpu in charge went
449 * into a long sleep. If two cpus happen to assign themself to
450 * this duty, then the jiffies update is still serialized by
453 if (unlikely(tick_do_timer_cpu == -1))
454 tick_do_timer_cpu = cpu;
456 /* Check, if the jiffies need an update */
457 if (tick_do_timer_cpu == cpu)
458 tick_do_update_jiffies64(now);
461 * When we are idle and the tick is stopped, we have to touch
462 * the watchdog as we might not schedule for a really long
463 * time. This happens on complete idle SMP systems while
464 * waiting on the login prompt. We also increment the "start
465 * of idle" jiffy stamp so the idle accounting adjustment we
466 * do when we go busy again does not account too much ticks.
468 if (ts->tick_stopped) {
469 touch_softlockup_watchdog();
473 update_process_times(user_mode(regs));
474 profile_tick(CPU_PROFILING);
476 /* Do not restart, when we are in the idle loop */
477 if (ts->tick_stopped)
480 while (tick_nohz_reprogram(ts, now)) {
482 tick_do_update_jiffies64(now);
487 * tick_nohz_switch_to_nohz - switch to nohz mode
489 static void tick_nohz_switch_to_nohz(void)
491 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
494 if (!tick_nohz_enabled)
498 if (tick_switch_to_oneshot(tick_nohz_handler)) {
503 ts->nohz_mode = NOHZ_MODE_LOWRES;
506 * Recycle the hrtimer in ts, so we can share the
507 * hrtimer_forward with the highres code.
509 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
510 /* Get the next period */
511 next = tick_init_jiffy_update();
514 ts->sched_timer.expires = next;
515 if (!tick_program_event(next, 0))
517 next = ktime_add(next, tick_period);
521 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
527 static inline void tick_nohz_switch_to_nohz(void) { }
532 * High resolution timer specific code
534 #ifdef CONFIG_HIGH_RES_TIMERS
536 * We rearm the timer until we get disabled by the idle code.
537 * Called with interrupts disabled and timer->base->cpu_base->lock held.
539 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
541 struct tick_sched *ts =
542 container_of(timer, struct tick_sched, sched_timer);
543 struct pt_regs *regs = get_irq_regs();
544 ktime_t now = ktime_get();
545 int cpu = smp_processor_id();
549 * Check if the do_timer duty was dropped. We don't care about
550 * concurrency: This happens only when the cpu in charge went
551 * into a long sleep. If two cpus happen to assign themself to
552 * this duty, then the jiffies update is still serialized by
555 if (unlikely(tick_do_timer_cpu == -1))
556 tick_do_timer_cpu = cpu;
559 /* Check, if the jiffies need an update */
560 if (tick_do_timer_cpu == cpu)
561 tick_do_update_jiffies64(now);
564 * Do not call, when we are not in irq context and have
565 * no valid regs pointer
569 * When we are idle and the tick is stopped, we have to touch
570 * the watchdog as we might not schedule for a really long
571 * time. This happens on complete idle SMP systems while
572 * waiting on the login prompt. We also increment the "start of
573 * idle" jiffy stamp so the idle accounting adjustment we do
574 * when we go busy again does not account too much ticks.
576 if (ts->tick_stopped) {
577 touch_softlockup_watchdog();
580 update_process_times(user_mode(regs));
581 profile_tick(CPU_PROFILING);
584 /* Do not restart, when we are in the idle loop */
585 if (ts->tick_stopped)
586 return HRTIMER_NORESTART;
588 hrtimer_forward(timer, now, tick_period);
590 return HRTIMER_RESTART;
594 * tick_setup_sched_timer - setup the tick emulation timer
596 void tick_setup_sched_timer(void)
598 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
599 ktime_t now = ktime_get();
603 * Emulate tick processing via per-CPU hrtimers:
605 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
606 ts->sched_timer.function = tick_sched_timer;
607 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
609 /* Get the next period (per cpu) */
610 ts->sched_timer.expires = tick_init_jiffy_update();
611 offset = ktime_to_ns(tick_period) >> 1;
612 do_div(offset, num_possible_cpus());
613 offset *= smp_processor_id();
614 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
617 hrtimer_forward(&ts->sched_timer, now, tick_period);
618 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
620 /* Check, if the timer was already in the past */
621 if (hrtimer_active(&ts->sched_timer))
627 if (tick_nohz_enabled)
628 ts->nohz_mode = NOHZ_MODE_HIGHRES;
632 void tick_cancel_sched_timer(int cpu)
634 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
636 if (ts->sched_timer.base)
637 hrtimer_cancel(&ts->sched_timer);
638 ts->tick_stopped = 0;
639 ts->nohz_mode = NOHZ_MODE_INACTIVE;
641 #endif /* HIGH_RES_TIMERS */
644 * Async notification about clocksource changes
646 void tick_clock_notify(void)
650 for_each_possible_cpu(cpu)
651 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
655 * Async notification about clock event changes
657 void tick_oneshot_notify(void)
659 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
661 set_bit(0, &ts->check_clocks);
665 * Check, if a change happened, which makes oneshot possible.
667 * Called cyclic from the hrtimer softirq (driven by the timer
668 * softirq) allow_nohz signals, that we can switch into low-res nohz
669 * mode, because high resolution timers are disabled (either compile
672 int tick_check_oneshot_change(int allow_nohz)
674 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
676 if (!test_and_clear_bit(0, &ts->check_clocks))
679 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
682 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
688 tick_nohz_switch_to_nohz();