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);
140 ts->idle_waketime = now;
142 local_irq_save(flags);
143 tick_do_update_jiffies64(now);
144 local_irq_restore(flags);
147 void tick_nohz_stop_idle(int cpu)
149 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
151 if (ts->idle_active) {
154 delta = ktime_sub(now, ts->idle_entrytime);
155 ts->idle_lastupdate = now;
156 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
161 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
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 ts = &per_cpu(tick_cpu_sched, cpu);
204 now = tick_nohz_start_idle(ts);
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 if (unlikely(local_softirq_pending())) {
225 static int ratelimit;
227 if (ratelimit < 10) {
228 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
229 local_softirq_pending());
235 /* Read jiffies and the time when jiffies were updated last */
237 seq = read_seqbegin(&xtime_lock);
238 last_update = last_jiffies_update;
239 last_jiffies = jiffies;
240 } while (read_seqretry(&xtime_lock, seq));
242 /* Get the next timer wheel timer */
243 next_jiffies = get_next_timer_interrupt(last_jiffies);
244 delta_jiffies = next_jiffies - last_jiffies;
246 rt_jiffies = rt_needs_cpu(cpu);
247 if (rt_jiffies && rt_jiffies < delta_jiffies)
248 delta_jiffies = rt_jiffies;
250 if (rcu_needs_cpu(cpu))
253 * Do not stop the tick, if we are only one off
254 * or if the cpu is required for rcu
256 if (!ts->tick_stopped && delta_jiffies == 1)
259 /* Schedule the tick, if we are at least one jiffie off */
260 if ((long)delta_jiffies >= 1) {
262 if (delta_jiffies > 1)
263 cpu_set(cpu, nohz_cpu_mask);
265 * nohz_stop_sched_tick can be called several times before
266 * the nohz_restart_sched_tick is called. This happens when
267 * interrupts arrive which do not cause a reschedule. In the
268 * first call we save the current tick time, so we can restart
269 * the scheduler tick in nohz_restart_sched_tick.
271 if (!ts->tick_stopped) {
272 if (select_nohz_load_balancer(1)) {
274 * sched tick not stopped!
276 cpu_clear(cpu, nohz_cpu_mask);
280 ts->idle_tick = ts->sched_timer.expires;
281 ts->tick_stopped = 1;
282 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) {
379 /* Update jiffies first */
380 select_nohz_load_balancer(0);
382 tick_do_update_jiffies64(now);
383 cpu_clear(cpu, nohz_cpu_mask);
386 * We stopped the tick in idle. Update process times would miss the
387 * time we slept as update_process_times does only a 1 tick
388 * accounting. Enforce that this is accounted to idle !
390 ticks = jiffies - ts->idle_jiffies;
392 * We might be one off. Do not randomly account a huge number of ticks!
394 if (ticks && ticks < LONG_MAX) {
395 add_preempt_count(HARDIRQ_OFFSET);
396 account_system_time(current, HARDIRQ_OFFSET,
397 jiffies_to_cputime(ticks));
398 sub_preempt_count(HARDIRQ_OFFSET);
402 * Cancel the scheduled timer and restore the tick
404 ts->tick_stopped = 0;
405 ts->idle_exittime = now;
406 hrtimer_cancel(&ts->sched_timer);
407 ts->sched_timer.expires = ts->idle_tick;
410 /* Forward the time to expire in the future */
411 hrtimer_forward(&ts->sched_timer, now, tick_period);
413 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
414 hrtimer_start(&ts->sched_timer,
415 ts->sched_timer.expires,
417 /* Check, if the timer was already in the past */
418 if (hrtimer_active(&ts->sched_timer))
421 if (!tick_program_event(ts->sched_timer.expires, 0))
424 /* Update jiffies and reread time */
425 tick_do_update_jiffies64(now);
431 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
433 hrtimer_forward(&ts->sched_timer, now, tick_period);
434 return tick_program_event(ts->sched_timer.expires, 0);
438 * The nohz low res interrupt handler
440 static void tick_nohz_handler(struct clock_event_device *dev)
442 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
443 struct pt_regs *regs = get_irq_regs();
444 int cpu = smp_processor_id();
445 ktime_t now = ktime_get();
447 dev->next_event.tv64 = KTIME_MAX;
450 * Check if the do_timer duty was dropped. We don't care about
451 * concurrency: This happens only when the cpu in charge went
452 * into a long sleep. If two cpus happen to assign themself to
453 * this duty, then the jiffies update is still serialized by
456 if (unlikely(tick_do_timer_cpu == -1))
457 tick_do_timer_cpu = cpu;
459 /* Check, if the jiffies need an update */
460 if (tick_do_timer_cpu == cpu)
461 tick_do_update_jiffies64(now);
464 * When we are idle and the tick is stopped, we have to touch
465 * the watchdog as we might not schedule for a really long
466 * time. This happens on complete idle SMP systems while
467 * waiting on the login prompt. We also increment the "start
468 * of idle" jiffy stamp so the idle accounting adjustment we
469 * do when we go busy again does not account too much ticks.
471 if (ts->tick_stopped) {
472 touch_softlockup_watchdog();
476 update_process_times(user_mode(regs));
477 profile_tick(CPU_PROFILING);
479 /* Do not restart, when we are in the idle loop */
480 if (ts->tick_stopped)
483 while (tick_nohz_reprogram(ts, now)) {
485 tick_do_update_jiffies64(now);
490 * tick_nohz_switch_to_nohz - switch to nohz mode
492 static void tick_nohz_switch_to_nohz(void)
494 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
497 if (!tick_nohz_enabled)
501 if (tick_switch_to_oneshot(tick_nohz_handler)) {
506 ts->nohz_mode = NOHZ_MODE_LOWRES;
509 * Recycle the hrtimer in ts, so we can share the
510 * hrtimer_forward with the highres code.
512 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
513 /* Get the next period */
514 next = tick_init_jiffy_update();
517 ts->sched_timer.expires = next;
518 if (!tick_program_event(next, 0))
520 next = ktime_add(next, tick_period);
524 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
530 static inline void tick_nohz_switch_to_nohz(void) { }
535 * High resolution timer specific code
537 #ifdef CONFIG_HIGH_RES_TIMERS
539 * We rearm the timer until we get disabled by the idle code.
540 * Called with interrupts disabled and timer->base->cpu_base->lock held.
542 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
544 struct tick_sched *ts =
545 container_of(timer, struct tick_sched, sched_timer);
546 struct pt_regs *regs = get_irq_regs();
547 ktime_t now = ktime_get();
548 int cpu = smp_processor_id();
552 * Check if the do_timer duty was dropped. We don't care about
553 * concurrency: This happens only when the cpu in charge went
554 * into a long sleep. If two cpus happen to assign themself to
555 * this duty, then the jiffies update is still serialized by
558 if (unlikely(tick_do_timer_cpu == -1))
559 tick_do_timer_cpu = cpu;
562 /* Check, if the jiffies need an update */
563 if (tick_do_timer_cpu == cpu)
564 tick_do_update_jiffies64(now);
567 * Do not call, when we are not in irq context and have
568 * no valid regs pointer
572 * When we are idle and the tick is stopped, we have to touch
573 * the watchdog as we might not schedule for a really long
574 * time. This happens on complete idle SMP systems while
575 * waiting on the login prompt. We also increment the "start of
576 * idle" jiffy stamp so the idle accounting adjustment we do
577 * when we go busy again does not account too much ticks.
579 if (ts->tick_stopped) {
580 touch_softlockup_watchdog();
583 update_process_times(user_mode(regs));
584 profile_tick(CPU_PROFILING);
587 /* Do not restart, when we are in the idle loop */
588 if (ts->tick_stopped)
589 return HRTIMER_NORESTART;
591 hrtimer_forward(timer, now, tick_period);
593 return HRTIMER_RESTART;
597 * tick_setup_sched_timer - setup the tick emulation timer
599 void tick_setup_sched_timer(void)
601 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
602 ktime_t now = ktime_get();
606 * Emulate tick processing via per-CPU hrtimers:
608 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
609 ts->sched_timer.function = tick_sched_timer;
610 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
612 /* Get the next period (per cpu) */
613 ts->sched_timer.expires = tick_init_jiffy_update();
614 offset = ktime_to_ns(tick_period) >> 1;
615 do_div(offset, num_possible_cpus());
616 offset *= smp_processor_id();
617 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
620 hrtimer_forward(&ts->sched_timer, now, tick_period);
621 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
623 /* Check, if the timer was already in the past */
624 if (hrtimer_active(&ts->sched_timer))
630 if (tick_nohz_enabled)
631 ts->nohz_mode = NOHZ_MODE_HIGHRES;
635 void tick_cancel_sched_timer(int cpu)
637 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
639 if (ts->sched_timer.base)
640 hrtimer_cancel(&ts->sched_timer);
642 ts->nohz_mode = NOHZ_MODE_INACTIVE;
644 #endif /* HIGH_RES_TIMERS */
647 * Async notification about clocksource changes
649 void tick_clock_notify(void)
653 for_each_possible_cpu(cpu)
654 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
658 * Async notification about clock event changes
660 void tick_oneshot_notify(void)
662 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
664 set_bit(0, &ts->check_clocks);
668 * Check, if a change happened, which makes oneshot possible.
670 * Called cyclic from the hrtimer softirq (driven by the timer
671 * softirq) allow_nohz signals, that we can switch into low-res nohz
672 * mode, because high resolution timers are disabled (either compile
675 int tick_check_oneshot_change(int allow_nohz)
677 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
679 if (!test_and_clear_bit(0, &ts->check_clocks))
682 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
685 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
691 tick_nohz_switch_to_nohz();