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 * For licencing details see kernel-base/COPYING
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 cpu_clear(cpu, nohz_cpu_mask);
139 local_irq_save(flags);
140 tick_do_update_jiffies64(now);
141 local_irq_restore(flags);
145 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
147 * When the next event is more than a tick into the future, stop the idle tick
148 * Called either from the idle loop or from irq_exit() when an idle period was
149 * just interrupted by an interrupt which did not cause a reschedule.
151 void tick_nohz_stop_sched_tick(void)
153 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
154 struct tick_sched *ts;
155 ktime_t last_update, expires, now, delta;
156 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
159 local_irq_save(flags);
161 cpu = smp_processor_id();
162 ts = &per_cpu(tick_cpu_sched, cpu);
165 * If this cpu is offline and it is the one which updates
166 * jiffies, then give up the assignment and let it be taken by
167 * the cpu which runs the tick timer next. If we don't drop
168 * this here the jiffies might be stale and do_timer() never
171 if (unlikely(!cpu_online(cpu))) {
172 if (cpu == tick_do_timer_cpu)
173 tick_do_timer_cpu = -1;
176 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
182 cpu = smp_processor_id();
183 if (unlikely(local_softirq_pending())) {
184 static int ratelimit;
186 if (ratelimit < 10) {
187 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
188 local_softirq_pending());
195 * When called from irq_exit we need to account the idle sleep time
198 if (ts->tick_stopped) {
199 delta = ktime_sub(now, ts->idle_entrytime);
200 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
203 ts->idle_entrytime = now;
206 /* Read jiffies and the time when jiffies were updated last */
208 seq = read_seqbegin(&xtime_lock);
209 last_update = last_jiffies_update;
210 last_jiffies = jiffies;
211 } while (read_seqretry(&xtime_lock, seq));
213 /* Get the next timer wheel timer */
214 next_jiffies = get_next_timer_interrupt(last_jiffies);
215 delta_jiffies = next_jiffies - last_jiffies;
217 if (rcu_needs_cpu(cpu))
220 * Do not stop the tick, if we are only one off
221 * or if the cpu is required for rcu
223 if (!ts->tick_stopped && delta_jiffies == 1)
226 /* Schedule the tick, if we are at least one jiffie off */
227 if ((long)delta_jiffies >= 1) {
229 if (delta_jiffies > 1)
230 cpu_set(cpu, nohz_cpu_mask);
232 * nohz_stop_sched_tick can be called several times before
233 * the nohz_restart_sched_tick is called. This happens when
234 * interrupts arrive which do not cause a reschedule. In the
235 * first call we save the current tick time, so we can restart
236 * the scheduler tick in nohz_restart_sched_tick.
238 if (!ts->tick_stopped) {
239 if (select_nohz_load_balancer(1)) {
241 * sched tick not stopped!
243 cpu_clear(cpu, nohz_cpu_mask);
247 ts->idle_tick = ts->sched_timer.expires;
248 ts->tick_stopped = 1;
249 ts->idle_jiffies = last_jiffies;
253 * If this cpu is the one which updates jiffies, then
254 * give up the assignment and let it be taken by the
255 * cpu which runs the tick timer next, which might be
256 * this cpu as well. If we don't drop this here the
257 * jiffies might be stale and do_timer() never
260 if (cpu == tick_do_timer_cpu)
261 tick_do_timer_cpu = -1;
266 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
267 * there is no timer pending or at least extremly far
268 * into the future (12 days for HZ=1000). In this case
269 * we simply stop the tick timer:
271 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
272 ts->idle_expires.tv64 = KTIME_MAX;
273 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
274 hrtimer_cancel(&ts->sched_timer);
279 * calculate the expiry time for the next timer wheel
282 expires = ktime_add_ns(last_update, tick_period.tv64 *
284 ts->idle_expires = expires;
286 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
287 hrtimer_start(&ts->sched_timer, expires,
289 /* Check, if the timer was already in the past */
290 if (hrtimer_active(&ts->sched_timer))
292 } else if(!tick_program_event(expires, 0))
295 * We are past the event already. So we crossed a
296 * jiffie boundary. Update jiffies and raise the
299 tick_do_update_jiffies64(ktime_get());
300 cpu_clear(cpu, nohz_cpu_mask);
302 raise_softirq_irqoff(TIMER_SOFTIRQ);
304 ts->next_jiffies = next_jiffies;
305 ts->last_jiffies = last_jiffies;
306 ts->sleep_length = ktime_sub(dev->next_event, now);
308 local_irq_restore(flags);
312 * tick_nohz_get_sleep_length - return the length of the current sleep
314 * Called from power state control code with interrupts disabled
316 ktime_t tick_nohz_get_sleep_length(void)
318 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
320 return ts->sleep_length;
323 EXPORT_SYMBOL_GPL(tick_nohz_get_sleep_length);
326 * nohz_restart_sched_tick - restart the idle tick from the idle task
328 * Restart the idle tick when the CPU is woken up from idle
330 void tick_nohz_restart_sched_tick(void)
332 int cpu = smp_processor_id();
333 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
337 if (!ts->tick_stopped)
340 /* Update jiffies first */
344 select_nohz_load_balancer(0);
345 tick_do_update_jiffies64(now);
346 cpu_clear(cpu, nohz_cpu_mask);
348 /* Account the idle time */
349 delta = ktime_sub(now, ts->idle_entrytime);
350 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
353 * We stopped the tick in idle. Update process times would miss the
354 * time we slept as update_process_times does only a 1 tick
355 * accounting. Enforce that this is accounted to idle !
357 ticks = jiffies - ts->idle_jiffies;
359 * We might be one off. Do not randomly account a huge number of ticks!
361 if (ticks && ticks < LONG_MAX) {
362 add_preempt_count(HARDIRQ_OFFSET);
363 account_system_time(current, HARDIRQ_OFFSET,
364 jiffies_to_cputime(ticks));
365 sub_preempt_count(HARDIRQ_OFFSET);
369 * Cancel the scheduled timer and restore the tick
371 ts->tick_stopped = 0;
372 hrtimer_cancel(&ts->sched_timer);
373 ts->sched_timer.expires = ts->idle_tick;
376 /* Forward the time to expire in the future */
377 hrtimer_forward(&ts->sched_timer, now, tick_period);
379 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
380 hrtimer_start(&ts->sched_timer,
381 ts->sched_timer.expires,
383 /* Check, if the timer was already in the past */
384 if (hrtimer_active(&ts->sched_timer))
387 if (!tick_program_event(ts->sched_timer.expires, 0))
390 /* Update jiffies and reread time */
391 tick_do_update_jiffies64(now);
397 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
399 hrtimer_forward(&ts->sched_timer, now, tick_period);
400 return tick_program_event(ts->sched_timer.expires, 0);
404 * The nohz low res interrupt handler
406 static void tick_nohz_handler(struct clock_event_device *dev)
408 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
409 struct pt_regs *regs = get_irq_regs();
410 int cpu = smp_processor_id();
411 ktime_t now = ktime_get();
413 dev->next_event.tv64 = KTIME_MAX;
416 * Check if the do_timer duty was dropped. We don't care about
417 * concurrency: This happens only when the cpu in charge went
418 * into a long sleep. If two cpus happen to assign themself to
419 * this duty, then the jiffies update is still serialized by
422 if (unlikely(tick_do_timer_cpu == -1))
423 tick_do_timer_cpu = cpu;
425 /* Check, if the jiffies need an update */
426 if (tick_do_timer_cpu == cpu)
427 tick_do_update_jiffies64(now);
430 * When we are idle and the tick is stopped, we have to touch
431 * the watchdog as we might not schedule for a really long
432 * time. This happens on complete idle SMP systems while
433 * waiting on the login prompt. We also increment the "start
434 * of idle" jiffy stamp so the idle accounting adjustment we
435 * do when we go busy again does not account too much ticks.
437 if (ts->tick_stopped) {
438 touch_softlockup_watchdog();
442 update_process_times(user_mode(regs));
443 profile_tick(CPU_PROFILING);
445 /* Do not restart, when we are in the idle loop */
446 if (ts->tick_stopped)
449 while (tick_nohz_reprogram(ts, now)) {
451 tick_do_update_jiffies64(now);
456 * tick_nohz_switch_to_nohz - switch to nohz mode
458 static void tick_nohz_switch_to_nohz(void)
460 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
463 if (!tick_nohz_enabled)
467 if (tick_switch_to_oneshot(tick_nohz_handler)) {
472 ts->nohz_mode = NOHZ_MODE_LOWRES;
475 * Recycle the hrtimer in ts, so we can share the
476 * hrtimer_forward with the highres code.
478 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
479 /* Get the next period */
480 next = tick_init_jiffy_update();
483 ts->sched_timer.expires = next;
484 if (!tick_program_event(next, 0))
486 next = ktime_add(next, tick_period);
490 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
496 static inline void tick_nohz_switch_to_nohz(void) { }
501 * High resolution timer specific code
503 #ifdef CONFIG_HIGH_RES_TIMERS
505 * We rearm the timer until we get disabled by the idle code
506 * Called with interrupts disabled and timer->base->cpu_base->lock held.
508 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
510 struct tick_sched *ts =
511 container_of(timer, struct tick_sched, sched_timer);
512 struct hrtimer_cpu_base *base = timer->base->cpu_base;
513 struct pt_regs *regs = get_irq_regs();
514 ktime_t now = ktime_get();
515 int cpu = smp_processor_id();
519 * Check if the do_timer duty was dropped. We don't care about
520 * concurrency: This happens only when the cpu in charge went
521 * into a long sleep. If two cpus happen to assign themself to
522 * this duty, then the jiffies update is still serialized by
525 if (unlikely(tick_do_timer_cpu == -1))
526 tick_do_timer_cpu = cpu;
529 /* Check, if the jiffies need an update */
530 if (tick_do_timer_cpu == cpu)
531 tick_do_update_jiffies64(now);
534 * Do not call, when we are not in irq context and have
535 * no valid regs pointer
539 * When we are idle and the tick is stopped, we have to touch
540 * the watchdog as we might not schedule for a really long
541 * time. This happens on complete idle SMP systems while
542 * waiting on the login prompt. We also increment the "start of
543 * idle" jiffy stamp so the idle accounting adjustment we do
544 * when we go busy again does not account too much ticks.
546 if (ts->tick_stopped) {
547 touch_softlockup_watchdog();
551 * update_process_times() might take tasklist_lock, hence
552 * drop the base lock. sched-tick hrtimers are per-CPU and
553 * never accessible by userspace APIs, so this is safe to do.
555 spin_unlock(&base->lock);
556 update_process_times(user_mode(regs));
557 profile_tick(CPU_PROFILING);
558 spin_lock(&base->lock);
561 /* Do not restart, when we are in the idle loop */
562 if (ts->tick_stopped)
563 return HRTIMER_NORESTART;
565 hrtimer_forward(timer, now, tick_period);
567 return HRTIMER_RESTART;
571 * tick_setup_sched_timer - setup the tick emulation timer
573 void tick_setup_sched_timer(void)
575 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
576 ktime_t now = ktime_get();
580 * Emulate tick processing via per-CPU hrtimers:
582 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
583 ts->sched_timer.function = tick_sched_timer;
584 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
586 /* Get the next period (per cpu) */
587 ts->sched_timer.expires = tick_init_jiffy_update();
588 offset = ktime_to_ns(tick_period) >> 1;
589 do_div(offset, num_possible_cpus());
590 offset *= smp_processor_id();
591 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
594 hrtimer_forward(&ts->sched_timer, now, tick_period);
595 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
597 /* Check, if the timer was already in the past */
598 if (hrtimer_active(&ts->sched_timer))
604 if (tick_nohz_enabled)
605 ts->nohz_mode = NOHZ_MODE_HIGHRES;
609 void tick_cancel_sched_timer(int cpu)
611 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
613 if (ts->sched_timer.base)
614 hrtimer_cancel(&ts->sched_timer);
615 ts->tick_stopped = 0;
616 ts->nohz_mode = NOHZ_MODE_INACTIVE;
618 #endif /* HIGH_RES_TIMERS */
621 * Async notification about clocksource changes
623 void tick_clock_notify(void)
627 for_each_possible_cpu(cpu)
628 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
632 * Async notification about clock event changes
634 void tick_oneshot_notify(void)
636 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
638 set_bit(0, &ts->check_clocks);
642 * Check, if a change happened, which makes oneshot possible.
644 * Called cyclic from the hrtimer softirq (driven by the timer
645 * softirq) allow_nohz signals, that we can switch into low-res nohz
646 * mode, because high resolution timers are disabled (either compile
649 int tick_check_oneshot_change(int allow_nohz)
651 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
653 if (!test_and_clear_bit(0, &ts->check_clocks))
656 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
659 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
665 tick_nohz_switch_to_nohz();