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;
158 local_irq_save(flags);
160 cpu = smp_processor_id();
161 ts = &per_cpu(tick_cpu_sched, cpu);
163 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
169 cpu = smp_processor_id();
170 if (unlikely(local_softirq_pending())) {
171 static int ratelimit;
173 if (ratelimit < 10) {
174 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
175 local_softirq_pending());
182 * When called from irq_exit we need to account the idle sleep time
185 if (ts->tick_stopped) {
186 delta = ktime_sub(now, ts->idle_entrytime);
187 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
190 ts->idle_entrytime = now;
193 /* Read jiffies and the time when jiffies were updated last */
195 seq = read_seqbegin(&xtime_lock);
196 last_update = last_jiffies_update;
197 last_jiffies = jiffies;
198 } while (read_seqretry(&xtime_lock, seq));
200 /* Get the next timer wheel timer */
201 next_jiffies = get_next_timer_interrupt(last_jiffies);
202 delta_jiffies = next_jiffies - last_jiffies;
204 if (rcu_needs_cpu(cpu))
207 * Do not stop the tick, if we are only one off
208 * or if the cpu is required for rcu
210 if (!ts->tick_stopped && delta_jiffies == 1)
213 /* Schedule the tick, if we are at least one jiffie off */
214 if ((long)delta_jiffies >= 1) {
216 if (delta_jiffies > 1)
217 cpu_set(cpu, nohz_cpu_mask);
219 * nohz_stop_sched_tick can be called several times before
220 * the nohz_restart_sched_tick is called. This happens when
221 * interrupts arrive which do not cause a reschedule. In the
222 * first call we save the current tick time, so we can restart
223 * the scheduler tick in nohz_restart_sched_tick.
225 if (!ts->tick_stopped) {
226 if (select_nohz_load_balancer(1)) {
228 * sched tick not stopped!
230 cpu_clear(cpu, nohz_cpu_mask);
234 ts->idle_tick = ts->sched_timer.expires;
235 ts->tick_stopped = 1;
236 ts->idle_jiffies = last_jiffies;
240 * If this cpu is the one which updates jiffies, then
241 * give up the assignment and let it be taken by the
242 * cpu which runs the tick timer next, which might be
243 * this cpu as well. If we don't drop this here the
244 * jiffies might be stale and do_timer() never
247 if (cpu == tick_do_timer_cpu)
248 tick_do_timer_cpu = -1;
253 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
254 * there is no timer pending or at least extremly far
255 * into the future (12 days for HZ=1000). In this case
256 * we simply stop the tick timer:
258 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
259 ts->idle_expires.tv64 = KTIME_MAX;
260 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
261 hrtimer_cancel(&ts->sched_timer);
266 * calculate the expiry time for the next timer wheel
269 expires = ktime_add_ns(last_update, tick_period.tv64 *
271 ts->idle_expires = expires;
273 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
274 hrtimer_start(&ts->sched_timer, expires,
276 /* Check, if the timer was already in the past */
277 if (hrtimer_active(&ts->sched_timer))
279 } else if(!tick_program_event(expires, 0))
282 * We are past the event already. So we crossed a
283 * jiffie boundary. Update jiffies and raise the
286 tick_do_update_jiffies64(ktime_get());
287 cpu_clear(cpu, nohz_cpu_mask);
289 raise_softirq_irqoff(TIMER_SOFTIRQ);
291 ts->next_jiffies = next_jiffies;
292 ts->last_jiffies = last_jiffies;
294 local_irq_restore(flags);
298 * nohz_restart_sched_tick - restart the idle tick from the idle task
300 * Restart the idle tick when the CPU is woken up from idle
302 void tick_nohz_restart_sched_tick(void)
304 int cpu = smp_processor_id();
305 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
309 if (!ts->tick_stopped)
312 /* Update jiffies first */
316 select_nohz_load_balancer(0);
317 tick_do_update_jiffies64(now);
318 cpu_clear(cpu, nohz_cpu_mask);
320 /* Account the idle time */
321 delta = ktime_sub(now, ts->idle_entrytime);
322 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
325 * We stopped the tick in idle. Update process times would miss the
326 * time we slept as update_process_times does only a 1 tick
327 * accounting. Enforce that this is accounted to idle !
329 ticks = jiffies - ts->idle_jiffies;
331 * We might be one off. Do not randomly account a huge number of ticks!
333 if (ticks && ticks < LONG_MAX) {
334 add_preempt_count(HARDIRQ_OFFSET);
335 account_system_time(current, HARDIRQ_OFFSET,
336 jiffies_to_cputime(ticks));
337 sub_preempt_count(HARDIRQ_OFFSET);
341 * Cancel the scheduled timer and restore the tick
343 ts->tick_stopped = 0;
344 hrtimer_cancel(&ts->sched_timer);
345 ts->sched_timer.expires = ts->idle_tick;
348 /* Forward the time to expire in the future */
349 hrtimer_forward(&ts->sched_timer, now, tick_period);
351 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
352 hrtimer_start(&ts->sched_timer,
353 ts->sched_timer.expires,
355 /* Check, if the timer was already in the past */
356 if (hrtimer_active(&ts->sched_timer))
359 if (!tick_program_event(ts->sched_timer.expires, 0))
362 /* Update jiffies and reread time */
363 tick_do_update_jiffies64(now);
369 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
371 hrtimer_forward(&ts->sched_timer, now, tick_period);
372 return tick_program_event(ts->sched_timer.expires, 0);
376 * The nohz low res interrupt handler
378 static void tick_nohz_handler(struct clock_event_device *dev)
380 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
381 struct pt_regs *regs = get_irq_regs();
382 int cpu = smp_processor_id();
383 ktime_t now = ktime_get();
385 dev->next_event.tv64 = KTIME_MAX;
388 * Check if the do_timer duty was dropped. We don't care about
389 * concurrency: This happens only when the cpu in charge went
390 * into a long sleep. If two cpus happen to assign themself to
391 * this duty, then the jiffies update is still serialized by
394 if (unlikely(tick_do_timer_cpu == -1))
395 tick_do_timer_cpu = cpu;
397 /* Check, if the jiffies need an update */
398 if (tick_do_timer_cpu == cpu)
399 tick_do_update_jiffies64(now);
402 * When we are idle and the tick is stopped, we have to touch
403 * the watchdog as we might not schedule for a really long
404 * time. This happens on complete idle SMP systems while
405 * waiting on the login prompt. We also increment the "start
406 * of idle" jiffy stamp so the idle accounting adjustment we
407 * do when we go busy again does not account too much ticks.
409 if (ts->tick_stopped) {
410 touch_softlockup_watchdog();
414 update_process_times(user_mode(regs));
415 profile_tick(CPU_PROFILING);
417 /* Do not restart, when we are in the idle loop */
418 if (ts->tick_stopped)
421 while (tick_nohz_reprogram(ts, now)) {
423 tick_do_update_jiffies64(now);
428 * tick_nohz_switch_to_nohz - switch to nohz mode
430 static void tick_nohz_switch_to_nohz(void)
432 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
435 if (!tick_nohz_enabled)
439 if (tick_switch_to_oneshot(tick_nohz_handler)) {
444 ts->nohz_mode = NOHZ_MODE_LOWRES;
447 * Recycle the hrtimer in ts, so we can share the
448 * hrtimer_forward with the highres code.
450 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
451 /* Get the next period */
452 next = tick_init_jiffy_update();
455 ts->sched_timer.expires = next;
456 if (!tick_program_event(next, 0))
458 next = ktime_add(next, tick_period);
462 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
468 static inline void tick_nohz_switch_to_nohz(void) { }
473 * High resolution timer specific code
475 #ifdef CONFIG_HIGH_RES_TIMERS
477 * We rearm the timer until we get disabled by the idle code
478 * Called with interrupts disabled and timer->base->cpu_base->lock held.
480 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
482 struct tick_sched *ts =
483 container_of(timer, struct tick_sched, sched_timer);
484 struct hrtimer_cpu_base *base = timer->base->cpu_base;
485 struct pt_regs *regs = get_irq_regs();
486 ktime_t now = ktime_get();
487 int cpu = smp_processor_id();
491 * Check if the do_timer duty was dropped. We don't care about
492 * concurrency: This happens only when the cpu in charge went
493 * into a long sleep. If two cpus happen to assign themself to
494 * this duty, then the jiffies update is still serialized by
497 if (unlikely(tick_do_timer_cpu == -1))
498 tick_do_timer_cpu = cpu;
501 /* Check, if the jiffies need an update */
502 if (tick_do_timer_cpu == cpu)
503 tick_do_update_jiffies64(now);
506 * Do not call, when we are not in irq context and have
507 * no valid regs pointer
511 * When we are idle and the tick is stopped, we have to touch
512 * the watchdog as we might not schedule for a really long
513 * time. This happens on complete idle SMP systems while
514 * waiting on the login prompt. We also increment the "start of
515 * idle" jiffy stamp so the idle accounting adjustment we do
516 * when we go busy again does not account too much ticks.
518 if (ts->tick_stopped) {
519 touch_softlockup_watchdog();
523 * update_process_times() might take tasklist_lock, hence
524 * drop the base lock. sched-tick hrtimers are per-CPU and
525 * never accessible by userspace APIs, so this is safe to do.
527 spin_unlock(&base->lock);
528 update_process_times(user_mode(regs));
529 profile_tick(CPU_PROFILING);
530 spin_lock(&base->lock);
533 /* Do not restart, when we are in the idle loop */
534 if (ts->tick_stopped)
535 return HRTIMER_NORESTART;
537 hrtimer_forward(timer, now, tick_period);
539 return HRTIMER_RESTART;
543 * tick_setup_sched_timer - setup the tick emulation timer
545 void tick_setup_sched_timer(void)
547 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
548 ktime_t now = ktime_get();
552 * Emulate tick processing via per-CPU hrtimers:
554 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
555 ts->sched_timer.function = tick_sched_timer;
556 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
558 /* Get the next period (per cpu) */
559 ts->sched_timer.expires = tick_init_jiffy_update();
560 offset = ktime_to_ns(tick_period) >> 1;
561 do_div(offset, NR_CPUS);
562 offset *= smp_processor_id();
563 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
566 hrtimer_forward(&ts->sched_timer, now, tick_period);
567 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
569 /* Check, if the timer was already in the past */
570 if (hrtimer_active(&ts->sched_timer))
576 if (tick_nohz_enabled)
577 ts->nohz_mode = NOHZ_MODE_HIGHRES;
581 void tick_cancel_sched_timer(int cpu)
583 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
585 if (ts->sched_timer.base)
586 hrtimer_cancel(&ts->sched_timer);
587 ts->tick_stopped = 0;
588 ts->nohz_mode = NOHZ_MODE_INACTIVE;
590 #endif /* HIGH_RES_TIMERS */
593 * Async notification about clocksource changes
595 void tick_clock_notify(void)
599 for_each_possible_cpu(cpu)
600 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
604 * Async notification about clock event changes
606 void tick_oneshot_notify(void)
608 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
610 set_bit(0, &ts->check_clocks);
614 * Check, if a change happened, which makes oneshot possible.
616 * Called cyclic from the hrtimer softirq (driven by the timer
617 * softirq) allow_nohz signals, that we can switch into low-res nohz
618 * mode, because high resolution timers are disabled (either compile
621 int tick_check_oneshot_change(int allow_nohz)
623 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
625 if (!test_and_clear_bit(0, &ts->check_clocks))
628 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
631 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
637 tick_nohz_switch_to_nohz();