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 "tick-internal.h"
27 * Per cpu nohz control structure
29 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
32 * The time, when the last jiffy update happened. Protected by xtime_lock.
34 static ktime_t last_jiffies_update;
37 * Must be called with interrupts disabled !
39 static void tick_do_update_jiffies64(ktime_t now)
41 unsigned long ticks = 0;
44 /* Reevalute with xtime_lock held */
45 write_seqlock(&xtime_lock);
47 delta = ktime_sub(now, last_jiffies_update);
48 if (delta.tv64 >= tick_period.tv64) {
50 delta = ktime_sub(delta, tick_period);
51 last_jiffies_update = ktime_add(last_jiffies_update,
54 /* Slow path for long timeouts */
55 if (unlikely(delta.tv64 >= tick_period.tv64)) {
56 s64 incr = ktime_to_ns(tick_period);
58 ticks = ktime_divns(delta, incr);
60 last_jiffies_update = ktime_add_ns(last_jiffies_update,
65 write_sequnlock(&xtime_lock);
69 * Initialize and return retrieve the jiffies update.
71 static ktime_t tick_init_jiffy_update(void)
75 write_seqlock(&xtime_lock);
76 /* Did we start the jiffies update yet ? */
77 if (last_jiffies_update.tv64 == 0)
78 last_jiffies_update = tick_next_period;
79 period = last_jiffies_update;
80 write_sequnlock(&xtime_lock);
85 * NOHZ - aka dynamic tick functionality
91 static int tick_nohz_enabled __read_mostly = 1;
94 * Enable / Disable tickless mode
96 static int __init setup_tick_nohz(char *str)
98 if (!strcmp(str, "off"))
99 tick_nohz_enabled = 0;
100 else if (!strcmp(str, "on"))
101 tick_nohz_enabled = 1;
107 __setup("nohz=", setup_tick_nohz);
110 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
112 * Called from interrupt entry when the CPU was idle
114 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
115 * must be updated. Otherwise an interrupt handler could use a stale jiffy
116 * value. We do this unconditionally on any cpu, as we don't know whether the
117 * cpu, which has the update task assigned is in a long sleep.
119 void tick_nohz_update_jiffies(void)
121 int cpu = smp_processor_id();
122 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
126 if (!ts->tick_stopped)
129 cpu_clear(cpu, nohz_cpu_mask);
132 local_irq_save(flags);
133 tick_do_update_jiffies64(now);
134 local_irq_restore(flags);
138 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
140 * When the next event is more than a tick into the future, stop the idle tick
141 * Called either from the idle loop or from irq_exit() when an idle period was
142 * just interrupted by an interrupt which did not cause a reschedule.
144 void tick_nohz_stop_sched_tick(void)
146 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
147 struct tick_sched *ts;
148 ktime_t last_update, expires, now, delta;
151 local_irq_save(flags);
153 cpu = smp_processor_id();
154 ts = &per_cpu(tick_cpu_sched, cpu);
156 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
162 cpu = smp_processor_id();
163 BUG_ON(local_softirq_pending());
167 * When called from irq_exit we need to account the idle sleep time
170 if (ts->tick_stopped) {
171 delta = ktime_sub(now, ts->idle_entrytime);
172 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
175 ts->idle_entrytime = now;
178 /* Read jiffies and the time when jiffies were updated last */
180 seq = read_seqbegin(&xtime_lock);
181 last_update = last_jiffies_update;
182 last_jiffies = jiffies;
183 } while (read_seqretry(&xtime_lock, seq));
185 /* Get the next timer wheel timer */
186 next_jiffies = get_next_timer_interrupt(last_jiffies);
187 delta_jiffies = next_jiffies - last_jiffies;
190 * Do not stop the tick, if we are only one off
191 * or if the cpu is required for rcu
193 if (!ts->tick_stopped && (delta_jiffies == 1 || rcu_needs_cpu(cpu)))
196 /* Schedule the tick, if we are at least one jiffie off */
197 if ((long)delta_jiffies >= 1) {
199 if (rcu_needs_cpu(cpu))
202 cpu_set(cpu, nohz_cpu_mask);
204 * nohz_stop_sched_tick can be called several times before
205 * the nohz_restart_sched_tick is called. This happens when
206 * interrupts arrive which do not cause a reschedule. In the
207 * first call we save the current tick time, so we can restart
208 * the scheduler tick in nohz_restart_sched_tick.
210 if (!ts->tick_stopped) {
211 ts->idle_tick = ts->sched_timer.expires;
212 ts->tick_stopped = 1;
213 ts->idle_jiffies = last_jiffies;
216 * calculate the expiry time for the next timer wheel
219 expires = ktime_add_ns(last_update, tick_period.tv64 *
221 ts->idle_expires = expires;
224 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
225 hrtimer_start(&ts->sched_timer, expires,
227 /* Check, if the timer was already in the past */
228 if (hrtimer_active(&ts->sched_timer))
230 } else if(!tick_program_event(expires, 0))
233 * We are past the event already. So we crossed a
234 * jiffie boundary. Update jiffies and raise the
237 tick_do_update_jiffies64(ktime_get());
238 cpu_clear(cpu, nohz_cpu_mask);
240 raise_softirq_irqoff(TIMER_SOFTIRQ);
242 ts->next_jiffies = next_jiffies;
243 ts->last_jiffies = last_jiffies;
245 local_irq_restore(flags);
249 * nohz_restart_sched_tick - restart the idle tick from the idle task
251 * Restart the idle tick when the CPU is woken up from idle
253 void tick_nohz_restart_sched_tick(void)
255 int cpu = smp_processor_id();
256 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
260 if (!ts->tick_stopped)
263 /* Update jiffies first */
267 tick_do_update_jiffies64(now);
268 cpu_clear(cpu, nohz_cpu_mask);
270 /* Account the idle time */
271 delta = ktime_sub(now, ts->idle_entrytime);
272 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
275 * We stopped the tick in idle. Update process times would miss the
276 * time we slept as update_process_times does only a 1 tick
277 * accounting. Enforce that this is accounted to idle !
279 ticks = jiffies - ts->idle_jiffies;
281 * We might be one off. Do not randomly account a huge number of ticks!
283 if (ticks && ticks < LONG_MAX) {
284 add_preempt_count(HARDIRQ_OFFSET);
285 account_system_time(current, HARDIRQ_OFFSET,
286 jiffies_to_cputime(ticks));
287 sub_preempt_count(HARDIRQ_OFFSET);
291 * Cancel the scheduled timer and restore the tick
293 ts->tick_stopped = 0;
294 hrtimer_cancel(&ts->sched_timer);
295 ts->sched_timer.expires = ts->idle_tick;
298 /* Forward the time to expire in the future */
299 hrtimer_forward(&ts->sched_timer, now, tick_period);
301 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
302 hrtimer_start(&ts->sched_timer,
303 ts->sched_timer.expires,
305 /* Check, if the timer was already in the past */
306 if (hrtimer_active(&ts->sched_timer))
309 if (!tick_program_event(ts->sched_timer.expires, 0))
312 /* Update jiffies and reread time */
313 tick_do_update_jiffies64(now);
319 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
321 hrtimer_forward(&ts->sched_timer, now, tick_period);
322 return tick_program_event(ts->sched_timer.expires, 0);
326 * The nohz low res interrupt handler
328 static void tick_nohz_handler(struct clock_event_device *dev)
330 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
331 struct pt_regs *regs = get_irq_regs();
332 ktime_t now = ktime_get();
334 dev->next_event.tv64 = KTIME_MAX;
336 /* Check, if the jiffies need an update */
337 tick_do_update_jiffies64(now);
340 * When we are idle and the tick is stopped, we have to touch
341 * the watchdog as we might not schedule for a really long
342 * time. This happens on complete idle SMP systems while
343 * waiting on the login prompt. We also increment the "start
344 * of idle" jiffy stamp so the idle accounting adjustment we
345 * do when we go busy again does not account too much ticks.
347 if (ts->tick_stopped) {
348 touch_softlockup_watchdog();
352 update_process_times(user_mode(regs));
353 profile_tick(CPU_PROFILING);
355 /* Do not restart, when we are in the idle loop */
356 if (ts->tick_stopped)
359 while (tick_nohz_reprogram(ts, now)) {
361 tick_do_update_jiffies64(now);
366 * tick_nohz_switch_to_nohz - switch to nohz mode
368 static void tick_nohz_switch_to_nohz(void)
370 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
373 if (!tick_nohz_enabled)
377 if (tick_switch_to_oneshot(tick_nohz_handler)) {
382 ts->nohz_mode = NOHZ_MODE_LOWRES;
385 * Recycle the hrtimer in ts, so we can share the
386 * hrtimer_forward with the highres code.
388 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
389 /* Get the next period */
390 next = tick_init_jiffy_update();
393 ts->sched_timer.expires = next;
394 if (!tick_program_event(next, 0))
396 next = ktime_add(next, tick_period);
400 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
406 static inline void tick_nohz_switch_to_nohz(void) { }
411 * High resolution timer specific code
413 #ifdef CONFIG_HIGH_RES_TIMERS
415 * We rearm the timer until we get disabled by the idle code
416 * Called with interrupts disabled and timer->base->cpu_base->lock held.
418 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
420 struct tick_sched *ts =
421 container_of(timer, struct tick_sched, sched_timer);
422 struct hrtimer_cpu_base *base = timer->base->cpu_base;
423 struct pt_regs *regs = get_irq_regs();
424 ktime_t now = ktime_get();
426 /* Check, if the jiffies need an update */
427 tick_do_update_jiffies64(now);
430 * Do not call, when we are not in irq context and have
431 * no valid regs pointer
435 * When we are idle and the tick is stopped, we have to touch
436 * the watchdog as we might not schedule for a really long
437 * time. This happens on complete idle SMP systems while
438 * waiting on the login prompt. We also increment the "start of
439 * idle" jiffy stamp so the idle accounting adjustment we do
440 * when we go busy again does not account too much ticks.
442 if (ts->tick_stopped) {
443 touch_softlockup_watchdog();
447 * update_process_times() might take tasklist_lock, hence
448 * drop the base lock. sched-tick hrtimers are per-CPU and
449 * never accessible by userspace APIs, so this is safe to do.
451 spin_unlock(&base->lock);
452 update_process_times(user_mode(regs));
453 profile_tick(CPU_PROFILING);
454 spin_lock(&base->lock);
457 /* Do not restart, when we are in the idle loop */
458 if (ts->tick_stopped)
459 return HRTIMER_NORESTART;
461 hrtimer_forward(timer, now, tick_period);
463 return HRTIMER_RESTART;
467 * tick_setup_sched_timer - setup the tick emulation timer
469 void tick_setup_sched_timer(void)
471 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
472 ktime_t now = ktime_get();
475 * Emulate tick processing via per-CPU hrtimers:
477 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
478 ts->sched_timer.function = tick_sched_timer;
479 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
481 /* Get the next period */
482 ts->sched_timer.expires = tick_init_jiffy_update();
485 hrtimer_forward(&ts->sched_timer, now, tick_period);
486 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
488 /* Check, if the timer was already in the past */
489 if (hrtimer_active(&ts->sched_timer))
495 if (tick_nohz_enabled)
496 ts->nohz_mode = NOHZ_MODE_HIGHRES;
500 void tick_cancel_sched_timer(int cpu)
502 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
504 if (ts->sched_timer.base)
505 hrtimer_cancel(&ts->sched_timer);
506 ts->tick_stopped = 0;
507 ts->nohz_mode = NOHZ_MODE_INACTIVE;
509 #endif /* HIGH_RES_TIMERS */
512 * Async notification about clocksource changes
514 void tick_clock_notify(void)
518 for_each_possible_cpu(cpu)
519 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
523 * Async notification about clock event changes
525 void tick_oneshot_notify(void)
527 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
529 set_bit(0, &ts->check_clocks);
533 * Check, if a change happened, which makes oneshot possible.
535 * Called cyclic from the hrtimer softirq (driven by the timer
536 * softirq) allow_nohz signals, that we can switch into low-res nohz
537 * mode, because high resolution timers are disabled (either compile
540 int tick_check_oneshot_change(int allow_nohz)
542 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
544 if (!test_and_clear_bit(0, &ts->check_clocks))
547 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
550 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
556 tick_nohz_switch_to_nohz();