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(int cpu)
163 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
167 if (ts->idle_active) {
168 delta = ktime_sub(now, ts->idle_entrytime);
169 ts->idle_lastupdate = now;
170 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
172 ts->idle_entrytime = now;
177 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
179 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
181 *last_update_time = ktime_to_us(ts->idle_lastupdate);
182 return ktime_to_us(ts->idle_sleeptime);
186 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
188 * When the next event is more than a tick into the future, stop the idle tick
189 * Called either from the idle loop or from irq_exit() when an idle period was
190 * just interrupted by an interrupt which did not cause a reschedule.
192 void tick_nohz_stop_sched_tick(void)
194 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
195 unsigned long rt_jiffies;
196 struct tick_sched *ts;
197 ktime_t last_update, expires, now;
198 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
201 local_irq_save(flags);
203 cpu = smp_processor_id();
204 now = tick_nohz_start_idle(cpu);
205 ts = &per_cpu(tick_cpu_sched, cpu);
208 * If this cpu is offline and it is the one which updates
209 * jiffies, then give up the assignment and let it be taken by
210 * the cpu which runs the tick timer next. If we don't drop
211 * this here the jiffies might be stale and do_timer() never
214 if (unlikely(!cpu_online(cpu))) {
215 if (cpu == tick_do_timer_cpu)
216 tick_do_timer_cpu = -1;
219 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
225 cpu = smp_processor_id();
226 if (unlikely(local_softirq_pending())) {
227 static int ratelimit;
229 if (ratelimit < 10) {
230 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
231 local_softirq_pending());
237 /* Read jiffies and the time when jiffies were updated last */
239 seq = read_seqbegin(&xtime_lock);
240 last_update = last_jiffies_update;
241 last_jiffies = jiffies;
242 } while (read_seqretry(&xtime_lock, seq));
244 /* Get the next timer wheel timer */
245 next_jiffies = get_next_timer_interrupt(last_jiffies);
246 delta_jiffies = next_jiffies - last_jiffies;
248 rt_jiffies = rt_needs_cpu(cpu);
249 if (rt_jiffies && rt_jiffies < delta_jiffies)
250 delta_jiffies = rt_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);
404 * Cancel the scheduled timer and restore the tick
406 ts->tick_stopped = 0;
407 ts->idle_exittime = now;
408 hrtimer_cancel(&ts->sched_timer);
409 ts->sched_timer.expires = ts->idle_tick;
412 /* Forward the time to expire in the future */
413 hrtimer_forward(&ts->sched_timer, now, tick_period);
415 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
416 hrtimer_start(&ts->sched_timer,
417 ts->sched_timer.expires,
419 /* Check, if the timer was already in the past */
420 if (hrtimer_active(&ts->sched_timer))
423 if (!tick_program_event(ts->sched_timer.expires, 0))
426 /* Update jiffies and reread time */
427 tick_do_update_jiffies64(now);
433 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
435 hrtimer_forward(&ts->sched_timer, now, tick_period);
436 return tick_program_event(ts->sched_timer.expires, 0);
440 * The nohz low res interrupt handler
442 static void tick_nohz_handler(struct clock_event_device *dev)
444 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
445 struct pt_regs *regs = get_irq_regs();
446 int cpu = smp_processor_id();
447 ktime_t now = ktime_get();
449 dev->next_event.tv64 = KTIME_MAX;
452 * Check if the do_timer duty was dropped. We don't care about
453 * concurrency: This happens only when the cpu in charge went
454 * into a long sleep. If two cpus happen to assign themself to
455 * this duty, then the jiffies update is still serialized by
458 if (unlikely(tick_do_timer_cpu == -1))
459 tick_do_timer_cpu = cpu;
461 /* Check, if the jiffies need an update */
462 if (tick_do_timer_cpu == cpu)
463 tick_do_update_jiffies64(now);
466 * When we are idle and the tick is stopped, we have to touch
467 * the watchdog as we might not schedule for a really long
468 * time. This happens on complete idle SMP systems while
469 * waiting on the login prompt. We also increment the "start
470 * of idle" jiffy stamp so the idle accounting adjustment we
471 * do when we go busy again does not account too much ticks.
473 if (ts->tick_stopped) {
474 touch_softlockup_watchdog();
478 update_process_times(user_mode(regs));
479 profile_tick(CPU_PROFILING);
481 /* Do not restart, when we are in the idle loop */
482 if (ts->tick_stopped)
485 while (tick_nohz_reprogram(ts, now)) {
487 tick_do_update_jiffies64(now);
492 * tick_nohz_switch_to_nohz - switch to nohz mode
494 static void tick_nohz_switch_to_nohz(void)
496 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
499 if (!tick_nohz_enabled)
503 if (tick_switch_to_oneshot(tick_nohz_handler)) {
508 ts->nohz_mode = NOHZ_MODE_LOWRES;
511 * Recycle the hrtimer in ts, so we can share the
512 * hrtimer_forward with the highres code.
514 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
515 /* Get the next period */
516 next = tick_init_jiffy_update();
519 ts->sched_timer.expires = next;
520 if (!tick_program_event(next, 0))
522 next = ktime_add(next, tick_period);
526 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
532 static inline void tick_nohz_switch_to_nohz(void) { }
537 * High resolution timer specific code
539 #ifdef CONFIG_HIGH_RES_TIMERS
541 * We rearm the timer until we get disabled by the idle code.
542 * Called with interrupts disabled and timer->base->cpu_base->lock held.
544 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
546 struct tick_sched *ts =
547 container_of(timer, struct tick_sched, sched_timer);
548 struct pt_regs *regs = get_irq_regs();
549 ktime_t now = ktime_get();
550 int cpu = smp_processor_id();
554 * Check if the do_timer duty was dropped. We don't care about
555 * concurrency: This happens only when the cpu in charge went
556 * into a long sleep. If two cpus happen to assign themself to
557 * this duty, then the jiffies update is still serialized by
560 if (unlikely(tick_do_timer_cpu == -1))
561 tick_do_timer_cpu = cpu;
564 /* Check, if the jiffies need an update */
565 if (tick_do_timer_cpu == cpu)
566 tick_do_update_jiffies64(now);
569 * Do not call, when we are not in irq context and have
570 * no valid regs pointer
574 * When we are idle and the tick is stopped, we have to touch
575 * the watchdog as we might not schedule for a really long
576 * time. This happens on complete idle SMP systems while
577 * waiting on the login prompt. We also increment the "start of
578 * idle" jiffy stamp so the idle accounting adjustment we do
579 * when we go busy again does not account too much ticks.
581 if (ts->tick_stopped) {
582 touch_softlockup_watchdog();
585 update_process_times(user_mode(regs));
586 profile_tick(CPU_PROFILING);
589 /* Do not restart, when we are in the idle loop */
590 if (ts->tick_stopped)
591 return HRTIMER_NORESTART;
593 hrtimer_forward(timer, now, tick_period);
595 return HRTIMER_RESTART;
599 * tick_setup_sched_timer - setup the tick emulation timer
601 void tick_setup_sched_timer(void)
603 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
604 ktime_t now = ktime_get();
608 * Emulate tick processing via per-CPU hrtimers:
610 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
611 ts->sched_timer.function = tick_sched_timer;
612 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
614 /* Get the next period (per cpu) */
615 ts->sched_timer.expires = tick_init_jiffy_update();
616 offset = ktime_to_ns(tick_period) >> 1;
617 do_div(offset, num_possible_cpus());
618 offset *= smp_processor_id();
619 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
622 hrtimer_forward(&ts->sched_timer, now, tick_period);
623 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
625 /* Check, if the timer was already in the past */
626 if (hrtimer_active(&ts->sched_timer))
632 if (tick_nohz_enabled)
633 ts->nohz_mode = NOHZ_MODE_HIGHRES;
637 void tick_cancel_sched_timer(int cpu)
639 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
641 if (ts->sched_timer.base)
642 hrtimer_cancel(&ts->sched_timer);
644 ts->nohz_mode = NOHZ_MODE_INACTIVE;
646 #endif /* HIGH_RES_TIMERS */
649 * Async notification about clocksource changes
651 void tick_clock_notify(void)
655 for_each_possible_cpu(cpu)
656 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
660 * Async notification about clock event changes
662 void tick_oneshot_notify(void)
664 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
666 set_bit(0, &ts->check_clocks);
670 * Check, if a change happened, which makes oneshot possible.
672 * Called cyclic from the hrtimer softirq (driven by the timer
673 * softirq) allow_nohz signals, that we can switch into low-res nohz
674 * mode, because high resolution timers are disabled (either compile
677 int tick_check_oneshot_change(int allow_nohz)
679 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
681 if (!test_and_clear_bit(0, &ts->check_clocks))
684 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
687 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
693 tick_nohz_switch_to_nohz();