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);
164 * If this cpu is offline and it is the one which updates
165 * jiffies, then give up the assignment and let it be taken by
166 * the cpu which runs the tick timer next. If we don't drop
167 * this here the jiffies might be stale and do_timer() never
170 if (unlikely(!cpu_online(cpu))) {
171 if (cpu == tick_do_timer_cpu)
172 tick_do_timer_cpu = -1;
175 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
181 cpu = smp_processor_id();
182 if (unlikely(local_softirq_pending())) {
183 static int ratelimit;
185 if (ratelimit < 10) {
186 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
187 local_softirq_pending());
194 * When called from irq_exit we need to account the idle sleep time
197 if (ts->tick_stopped) {
198 delta = ktime_sub(now, ts->idle_entrytime);
199 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
202 ts->idle_entrytime = now;
205 /* Read jiffies and the time when jiffies were updated last */
207 seq = read_seqbegin(&xtime_lock);
208 last_update = last_jiffies_update;
209 last_jiffies = jiffies;
210 } while (read_seqretry(&xtime_lock, seq));
212 /* Get the next timer wheel timer */
213 next_jiffies = get_next_timer_interrupt(last_jiffies);
214 delta_jiffies = next_jiffies - last_jiffies;
216 if (rcu_needs_cpu(cpu))
219 * Do not stop the tick, if we are only one off
220 * or if the cpu is required for rcu
222 if (!ts->tick_stopped && delta_jiffies == 1)
225 /* Schedule the tick, if we are at least one jiffie off */
226 if ((long)delta_jiffies >= 1) {
228 if (delta_jiffies > 1)
229 cpu_set(cpu, nohz_cpu_mask);
231 * nohz_stop_sched_tick can be called several times before
232 * the nohz_restart_sched_tick is called. This happens when
233 * interrupts arrive which do not cause a reschedule. In the
234 * first call we save the current tick time, so we can restart
235 * the scheduler tick in nohz_restart_sched_tick.
237 if (!ts->tick_stopped) {
238 if (select_nohz_load_balancer(1)) {
240 * sched tick not stopped!
242 cpu_clear(cpu, nohz_cpu_mask);
246 ts->idle_tick = ts->sched_timer.expires;
247 ts->tick_stopped = 1;
248 ts->idle_jiffies = last_jiffies;
252 * If this cpu is the one which updates jiffies, then
253 * give up the assignment and let it be taken by the
254 * cpu which runs the tick timer next, which might be
255 * this cpu as well. If we don't drop this here the
256 * jiffies might be stale and do_timer() never
259 if (cpu == tick_do_timer_cpu)
260 tick_do_timer_cpu = -1;
265 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
266 * there is no timer pending or at least extremly far
267 * into the future (12 days for HZ=1000). In this case
268 * we simply stop the tick timer:
270 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
271 ts->idle_expires.tv64 = KTIME_MAX;
272 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
273 hrtimer_cancel(&ts->sched_timer);
278 * calculate the expiry time for the next timer wheel
281 expires = ktime_add_ns(last_update, tick_period.tv64 *
283 ts->idle_expires = expires;
285 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
286 hrtimer_start(&ts->sched_timer, expires,
288 /* Check, if the timer was already in the past */
289 if (hrtimer_active(&ts->sched_timer))
291 } else if(!tick_program_event(expires, 0))
294 * We are past the event already. So we crossed a
295 * jiffie boundary. Update jiffies and raise the
298 tick_do_update_jiffies64(ktime_get());
299 cpu_clear(cpu, nohz_cpu_mask);
301 raise_softirq_irqoff(TIMER_SOFTIRQ);
303 ts->next_jiffies = next_jiffies;
304 ts->last_jiffies = last_jiffies;
306 local_irq_restore(flags);
310 * nohz_restart_sched_tick - restart the idle tick from the idle task
312 * Restart the idle tick when the CPU is woken up from idle
314 void tick_nohz_restart_sched_tick(void)
316 int cpu = smp_processor_id();
317 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
321 if (!ts->tick_stopped)
324 /* Update jiffies first */
328 select_nohz_load_balancer(0);
329 tick_do_update_jiffies64(now);
330 cpu_clear(cpu, nohz_cpu_mask);
332 /* Account the idle time */
333 delta = ktime_sub(now, ts->idle_entrytime);
334 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
337 * We stopped the tick in idle. Update process times would miss the
338 * time we slept as update_process_times does only a 1 tick
339 * accounting. Enforce that this is accounted to idle !
341 ticks = jiffies - ts->idle_jiffies;
343 * We might be one off. Do not randomly account a huge number of ticks!
345 if (ticks && ticks < LONG_MAX) {
346 add_preempt_count(HARDIRQ_OFFSET);
347 account_system_time(current, HARDIRQ_OFFSET,
348 jiffies_to_cputime(ticks));
349 sub_preempt_count(HARDIRQ_OFFSET);
353 * Cancel the scheduled timer and restore the tick
355 ts->tick_stopped = 0;
356 hrtimer_cancel(&ts->sched_timer);
357 ts->sched_timer.expires = ts->idle_tick;
360 /* Forward the time to expire in the future */
361 hrtimer_forward(&ts->sched_timer, now, tick_period);
363 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
364 hrtimer_start(&ts->sched_timer,
365 ts->sched_timer.expires,
367 /* Check, if the timer was already in the past */
368 if (hrtimer_active(&ts->sched_timer))
371 if (!tick_program_event(ts->sched_timer.expires, 0))
374 /* Update jiffies and reread time */
375 tick_do_update_jiffies64(now);
381 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
383 hrtimer_forward(&ts->sched_timer, now, tick_period);
384 return tick_program_event(ts->sched_timer.expires, 0);
388 * The nohz low res interrupt handler
390 static void tick_nohz_handler(struct clock_event_device *dev)
392 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
393 struct pt_regs *regs = get_irq_regs();
394 int cpu = smp_processor_id();
395 ktime_t now = ktime_get();
397 dev->next_event.tv64 = KTIME_MAX;
400 * Check if the do_timer duty was dropped. We don't care about
401 * concurrency: This happens only when the cpu in charge went
402 * into a long sleep. If two cpus happen to assign themself to
403 * this duty, then the jiffies update is still serialized by
406 if (unlikely(tick_do_timer_cpu == -1))
407 tick_do_timer_cpu = cpu;
409 /* Check, if the jiffies need an update */
410 if (tick_do_timer_cpu == cpu)
411 tick_do_update_jiffies64(now);
414 * When we are idle and the tick is stopped, we have to touch
415 * the watchdog as we might not schedule for a really long
416 * time. This happens on complete idle SMP systems while
417 * waiting on the login prompt. We also increment the "start
418 * of idle" jiffy stamp so the idle accounting adjustment we
419 * do when we go busy again does not account too much ticks.
421 if (ts->tick_stopped) {
422 touch_softlockup_watchdog();
426 update_process_times(user_mode(regs));
427 profile_tick(CPU_PROFILING);
429 /* Do not restart, when we are in the idle loop */
430 if (ts->tick_stopped)
433 while (tick_nohz_reprogram(ts, now)) {
435 tick_do_update_jiffies64(now);
440 * tick_nohz_switch_to_nohz - switch to nohz mode
442 static void tick_nohz_switch_to_nohz(void)
444 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
447 if (!tick_nohz_enabled)
451 if (tick_switch_to_oneshot(tick_nohz_handler)) {
456 ts->nohz_mode = NOHZ_MODE_LOWRES;
459 * Recycle the hrtimer in ts, so we can share the
460 * hrtimer_forward with the highres code.
462 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
463 /* Get the next period */
464 next = tick_init_jiffy_update();
467 ts->sched_timer.expires = next;
468 if (!tick_program_event(next, 0))
470 next = ktime_add(next, tick_period);
474 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
480 static inline void tick_nohz_switch_to_nohz(void) { }
485 * High resolution timer specific code
487 #ifdef CONFIG_HIGH_RES_TIMERS
489 * We rearm the timer until we get disabled by the idle code
490 * Called with interrupts disabled and timer->base->cpu_base->lock held.
492 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
494 struct tick_sched *ts =
495 container_of(timer, struct tick_sched, sched_timer);
496 struct hrtimer_cpu_base *base = timer->base->cpu_base;
497 struct pt_regs *regs = get_irq_regs();
498 ktime_t now = ktime_get();
499 int cpu = smp_processor_id();
503 * Check if the do_timer duty was dropped. We don't care about
504 * concurrency: This happens only when the cpu in charge went
505 * into a long sleep. If two cpus happen to assign themself to
506 * this duty, then the jiffies update is still serialized by
509 if (unlikely(tick_do_timer_cpu == -1))
510 tick_do_timer_cpu = cpu;
513 /* Check, if the jiffies need an update */
514 if (tick_do_timer_cpu == cpu)
515 tick_do_update_jiffies64(now);
518 * Do not call, when we are not in irq context and have
519 * no valid regs pointer
523 * When we are idle and the tick is stopped, we have to touch
524 * the watchdog as we might not schedule for a really long
525 * time. This happens on complete idle SMP systems while
526 * waiting on the login prompt. We also increment the "start of
527 * idle" jiffy stamp so the idle accounting adjustment we do
528 * when we go busy again does not account too much ticks.
530 if (ts->tick_stopped) {
531 touch_softlockup_watchdog();
535 * update_process_times() might take tasklist_lock, hence
536 * drop the base lock. sched-tick hrtimers are per-CPU and
537 * never accessible by userspace APIs, so this is safe to do.
539 spin_unlock(&base->lock);
540 update_process_times(user_mode(regs));
541 profile_tick(CPU_PROFILING);
542 spin_lock(&base->lock);
545 /* Do not restart, when we are in the idle loop */
546 if (ts->tick_stopped)
547 return HRTIMER_NORESTART;
549 hrtimer_forward(timer, now, tick_period);
551 return HRTIMER_RESTART;
555 * tick_setup_sched_timer - setup the tick emulation timer
557 void tick_setup_sched_timer(void)
559 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
560 ktime_t now = ktime_get();
564 * Emulate tick processing via per-CPU hrtimers:
566 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
567 ts->sched_timer.function = tick_sched_timer;
568 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
570 /* Get the next period (per cpu) */
571 ts->sched_timer.expires = tick_init_jiffy_update();
572 offset = ktime_to_ns(tick_period) >> 1;
573 do_div(offset, NR_CPUS);
574 offset *= smp_processor_id();
575 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset);
578 hrtimer_forward(&ts->sched_timer, now, tick_period);
579 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
581 /* Check, if the timer was already in the past */
582 if (hrtimer_active(&ts->sched_timer))
588 if (tick_nohz_enabled)
589 ts->nohz_mode = NOHZ_MODE_HIGHRES;
593 void tick_cancel_sched_timer(int cpu)
595 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
597 if (ts->sched_timer.base)
598 hrtimer_cancel(&ts->sched_timer);
599 ts->tick_stopped = 0;
600 ts->nohz_mode = NOHZ_MODE_INACTIVE;
602 #endif /* HIGH_RES_TIMERS */
605 * Async notification about clocksource changes
607 void tick_clock_notify(void)
611 for_each_possible_cpu(cpu)
612 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
616 * Async notification about clock event changes
618 void tick_oneshot_notify(void)
620 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
622 set_bit(0, &ts->check_clocks);
626 * Check, if a change happened, which makes oneshot possible.
628 * Called cyclic from the hrtimer softirq (driven by the timer
629 * softirq) allow_nohz signals, that we can switch into low-res nohz
630 * mode, because high resolution timers are disabled (either compile
633 int tick_check_oneshot_change(int allow_nohz)
635 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
637 if (!test_and_clear_bit(0, &ts->check_clocks))
640 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
643 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
649 tick_nohz_switch_to_nohz();