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 touch_softlockup_watchdog();
145 cpu_clear(cpu, nohz_cpu_mask);
147 ts->idle_waketime = now;
149 local_irq_save(flags);
150 tick_do_update_jiffies64(now);
151 local_irq_restore(flags);
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(void)
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))
230 if (unlikely(local_softirq_pending())) {
231 static int ratelimit;
233 if (ratelimit < 10) {
234 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
235 local_softirq_pending());
241 /* Read jiffies and the time when jiffies were updated last */
243 seq = read_seqbegin(&xtime_lock);
244 last_update = last_jiffies_update;
245 last_jiffies = jiffies;
246 } while (read_seqretry(&xtime_lock, seq));
248 /* Get the next timer wheel timer */
249 next_jiffies = get_next_timer_interrupt(last_jiffies);
250 delta_jiffies = next_jiffies - last_jiffies;
252 if (rcu_needs_cpu(cpu))
255 * Do not stop the tick, if we are only one off
256 * or if the cpu is required for rcu
258 if (!ts->tick_stopped && delta_jiffies == 1)
261 /* Schedule the tick, if we are at least one jiffie off */
262 if ((long)delta_jiffies >= 1) {
264 if (delta_jiffies > 1)
265 cpu_set(cpu, nohz_cpu_mask);
267 * nohz_stop_sched_tick can be called several times before
268 * the nohz_restart_sched_tick is called. This happens when
269 * interrupts arrive which do not cause a reschedule. In the
270 * first call we save the current tick time, so we can restart
271 * the scheduler tick in nohz_restart_sched_tick.
273 if (!ts->tick_stopped) {
274 if (select_nohz_load_balancer(1)) {
276 * sched tick not stopped!
278 cpu_clear(cpu, nohz_cpu_mask);
282 ts->idle_tick = ts->sched_timer.expires;
283 ts->tick_stopped = 1;
284 ts->idle_jiffies = last_jiffies;
289 * If this cpu is the one which updates jiffies, then
290 * give up the assignment and let it be taken by the
291 * cpu which runs the tick timer next, which might be
292 * this cpu as well. If we don't drop this here the
293 * jiffies might be stale and do_timer() never
296 if (cpu == tick_do_timer_cpu)
297 tick_do_timer_cpu = -1;
302 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
303 * there is no timer pending or at least extremly far
304 * into the future (12 days for HZ=1000). In this case
305 * we simply stop the tick timer:
307 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
308 ts->idle_expires.tv64 = KTIME_MAX;
309 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
310 hrtimer_cancel(&ts->sched_timer);
315 * calculate the expiry time for the next timer wheel
318 expires = ktime_add_ns(last_update, tick_period.tv64 *
320 ts->idle_expires = expires;
322 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
323 hrtimer_start(&ts->sched_timer, expires,
325 /* Check, if the timer was already in the past */
326 if (hrtimer_active(&ts->sched_timer))
328 } else if (!tick_program_event(expires, 0))
331 * We are past the event already. So we crossed a
332 * jiffie boundary. Update jiffies and raise the
335 tick_do_update_jiffies64(ktime_get());
336 cpu_clear(cpu, nohz_cpu_mask);
338 raise_softirq_irqoff(TIMER_SOFTIRQ);
340 ts->next_jiffies = next_jiffies;
341 ts->last_jiffies = last_jiffies;
342 ts->sleep_length = ktime_sub(dev->next_event, now);
344 local_irq_restore(flags);
348 * tick_nohz_get_sleep_length - return the length of the current sleep
350 * Called from power state control code with interrupts disabled
352 ktime_t tick_nohz_get_sleep_length(void)
354 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
356 return ts->sleep_length;
360 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
362 * Restart the idle tick when the CPU is woken up from idle
364 void tick_nohz_restart_sched_tick(void)
366 int cpu = smp_processor_id();
367 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
372 tick_nohz_stop_idle(cpu);
374 if (!ts->tick_stopped) {
381 /* Update jiffies first */
382 select_nohz_load_balancer(0);
384 tick_do_update_jiffies64(now);
385 cpu_clear(cpu, nohz_cpu_mask);
388 * We stopped the tick in idle. Update process times would miss the
389 * time we slept as update_process_times does only a 1 tick
390 * accounting. Enforce that this is accounted to idle !
392 ticks = jiffies - ts->idle_jiffies;
394 * We might be one off. Do not randomly account a huge number of ticks!
396 if (ticks && ticks < LONG_MAX) {
397 add_preempt_count(HARDIRQ_OFFSET);
398 account_system_time(current, HARDIRQ_OFFSET,
399 jiffies_to_cputime(ticks));
400 sub_preempt_count(HARDIRQ_OFFSET);
403 touch_softlockup_watchdog();
405 * Cancel the scheduled timer and restore the tick
407 ts->tick_stopped = 0;
408 ts->idle_exittime = now;
409 hrtimer_cancel(&ts->sched_timer);
410 ts->sched_timer.expires = ts->idle_tick;
413 /* Forward the time to expire in the future */
414 hrtimer_forward(&ts->sched_timer, now, tick_period);
416 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
417 hrtimer_start(&ts->sched_timer,
418 ts->sched_timer.expires,
420 /* Check, if the timer was already in the past */
421 if (hrtimer_active(&ts->sched_timer))
424 if (!tick_program_event(ts->sched_timer.expires, 0))
427 /* Update jiffies and reread time */
428 tick_do_update_jiffies64(now);
434 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
436 hrtimer_forward(&ts->sched_timer, now, tick_period);
437 return tick_program_event(ts->sched_timer.expires, 0);
441 * The nohz low res interrupt handler
443 static void tick_nohz_handler(struct clock_event_device *dev)
445 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
446 struct pt_regs *regs = get_irq_regs();
447 int cpu = smp_processor_id();
448 ktime_t now = ktime_get();
450 dev->next_event.tv64 = KTIME_MAX;
453 * Check if the do_timer duty was dropped. We don't care about
454 * concurrency: This happens only when the cpu in charge went
455 * into a long sleep. If two cpus happen to assign themself to
456 * this duty, then the jiffies update is still serialized by
459 if (unlikely(tick_do_timer_cpu == -1))
460 tick_do_timer_cpu = cpu;
462 /* Check, if the jiffies need an update */
463 if (tick_do_timer_cpu == cpu)
464 tick_do_update_jiffies64(now);
467 * When we are idle and the tick is stopped, we have to touch
468 * the watchdog as we might not schedule for a really long
469 * time. This happens on complete idle SMP systems while
470 * waiting on the login prompt. We also increment the "start
471 * of idle" jiffy stamp so the idle accounting adjustment we
472 * do when we go busy again does not account too much ticks.
474 if (ts->tick_stopped) {
475 touch_softlockup_watchdog();
479 update_process_times(user_mode(regs));
480 profile_tick(CPU_PROFILING);
482 /* Do not restart, when we are in the idle loop */
483 if (ts->tick_stopped)
486 while (tick_nohz_reprogram(ts, now)) {
488 tick_do_update_jiffies64(now);
493 * tick_nohz_switch_to_nohz - switch to nohz mode
495 static void tick_nohz_switch_to_nohz(void)
497 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
500 if (!tick_nohz_enabled)
504 if (tick_switch_to_oneshot(tick_nohz_handler)) {
509 ts->nohz_mode = NOHZ_MODE_LOWRES;
512 * Recycle the hrtimer in ts, so we can share the
513 * hrtimer_forward with the highres code.
515 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
516 /* Get the next period */
517 next = tick_init_jiffy_update();
520 ts->sched_timer.expires = next;
521 if (!tick_program_event(next, 0))
523 next = ktime_add(next, tick_period);
527 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
533 static inline void tick_nohz_switch_to_nohz(void) { }
538 * High resolution timer specific code
540 #ifdef CONFIG_HIGH_RES_TIMERS
542 * We rearm the timer until we get disabled by the idle code.
543 * Called with interrupts disabled and timer->base->cpu_base->lock held.
545 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
547 struct tick_sched *ts =
548 container_of(timer, struct tick_sched, sched_timer);
549 struct pt_regs *regs = get_irq_regs();
550 ktime_t now = ktime_get();
551 int cpu = smp_processor_id();
555 * Check if the do_timer duty was dropped. We don't care about
556 * concurrency: This happens only when the cpu in charge went
557 * into a long sleep. If two cpus happen to assign themself to
558 * this duty, then the jiffies update is still serialized by
561 if (unlikely(tick_do_timer_cpu == -1))
562 tick_do_timer_cpu = cpu;
565 /* Check, if the jiffies need an update */
566 if (tick_do_timer_cpu == cpu)
567 tick_do_update_jiffies64(now);
570 * Do not call, when we are not in irq context and have
571 * no valid regs pointer
575 * When we are idle and the tick is stopped, we have to touch
576 * the watchdog as we might not schedule for a really long
577 * time. This happens on complete idle SMP systems while
578 * waiting on the login prompt. We also increment the "start of
579 * idle" jiffy stamp so the idle accounting adjustment we do
580 * when we go busy again does not account too much ticks.
582 if (ts->tick_stopped) {
583 touch_softlockup_watchdog();
586 update_process_times(user_mode(regs));
587 profile_tick(CPU_PROFILING);
590 /* Do not restart, when we are in the idle loop */
591 if (ts->tick_stopped)
592 return HRTIMER_NORESTART;
594 hrtimer_forward(timer, now, tick_period);
596 return HRTIMER_RESTART;
600 * tick_setup_sched_timer - setup the tick emulation timer
602 void tick_setup_sched_timer(void)
604 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
605 ktime_t now = ktime_get();
609 * Emulate tick processing via per-CPU hrtimers:
611 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
612 ts->sched_timer.function = tick_sched_timer;
613 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
615 /* Get the next period (per cpu) */
616 ts->sched_timer.expires = tick_init_jiffy_update();
617 offset = ktime_to_ns(tick_period) >> 1;
618 do_div(offset, num_possible_cpus());
619 offset *= smp_processor_id();
620 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
623 hrtimer_forward(&ts->sched_timer, now, tick_period);
624 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
626 /* Check, if the timer was already in the past */
627 if (hrtimer_active(&ts->sched_timer))
633 if (tick_nohz_enabled)
634 ts->nohz_mode = NOHZ_MODE_HIGHRES;
638 void tick_cancel_sched_timer(int cpu)
640 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
642 if (ts->sched_timer.base)
643 hrtimer_cancel(&ts->sched_timer);
645 ts->nohz_mode = NOHZ_MODE_INACTIVE;
647 #endif /* HIGH_RES_TIMERS */
650 * Async notification about clocksource changes
652 void tick_clock_notify(void)
656 for_each_possible_cpu(cpu)
657 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
661 * Async notification about clock event changes
663 void tick_oneshot_notify(void)
665 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
667 set_bit(0, &ts->check_clocks);
671 * Check, if a change happened, which makes oneshot possible.
673 * Called cyclic from the hrtimer softirq (driven by the timer
674 * softirq) allow_nohz signals, that we can switch into low-res nohz
675 * mode, because high resolution timers are disabled (either compile
678 int tick_check_oneshot_change(int allow_nohz)
680 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
682 if (!test_and_clear_bit(0, &ts->check_clocks))
685 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
688 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
694 tick_nohz_switch_to_nohz();