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 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
172 local_softirq_pending());
176 * When called from irq_exit we need to account the idle sleep time
179 if (ts->tick_stopped) {
180 delta = ktime_sub(now, ts->idle_entrytime);
181 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
184 ts->idle_entrytime = now;
187 /* Read jiffies and the time when jiffies were updated last */
189 seq = read_seqbegin(&xtime_lock);
190 last_update = last_jiffies_update;
191 last_jiffies = jiffies;
192 } while (read_seqretry(&xtime_lock, seq));
194 /* Get the next timer wheel timer */
195 next_jiffies = get_next_timer_interrupt(last_jiffies);
196 delta_jiffies = next_jiffies - last_jiffies;
198 if (rcu_needs_cpu(cpu))
201 * Do not stop the tick, if we are only one off
202 * or if the cpu is required for rcu
204 if (!ts->tick_stopped && delta_jiffies == 1)
207 /* Schedule the tick, if we are at least one jiffie off */
208 if ((long)delta_jiffies >= 1) {
210 if (delta_jiffies > 1)
211 cpu_set(cpu, nohz_cpu_mask);
213 * nohz_stop_sched_tick can be called several times before
214 * the nohz_restart_sched_tick is called. This happens when
215 * interrupts arrive which do not cause a reschedule. In the
216 * first call we save the current tick time, so we can restart
217 * the scheduler tick in nohz_restart_sched_tick.
219 if (!ts->tick_stopped) {
220 ts->idle_tick = ts->sched_timer.expires;
221 ts->tick_stopped = 1;
222 ts->idle_jiffies = last_jiffies;
226 * If this cpu is the one which updates jiffies, then
227 * give up the assignment and let it be taken by the
228 * cpu which runs the tick timer next, which might be
229 * this cpu as well. If we don't drop this here the
230 * jiffies might be stale and do_timer() never
233 if (cpu == tick_do_timer_cpu)
234 tick_do_timer_cpu = -1;
237 * calculate the expiry time for the next timer wheel
240 expires = ktime_add_ns(last_update, tick_period.tv64 *
242 ts->idle_expires = expires;
245 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
246 hrtimer_start(&ts->sched_timer, expires,
248 /* Check, if the timer was already in the past */
249 if (hrtimer_active(&ts->sched_timer))
251 } else if(!tick_program_event(expires, 0))
254 * We are past the event already. So we crossed a
255 * jiffie boundary. Update jiffies and raise the
258 tick_do_update_jiffies64(ktime_get());
259 cpu_clear(cpu, nohz_cpu_mask);
261 raise_softirq_irqoff(TIMER_SOFTIRQ);
263 ts->next_jiffies = next_jiffies;
264 ts->last_jiffies = last_jiffies;
266 local_irq_restore(flags);
270 * nohz_restart_sched_tick - restart the idle tick from the idle task
272 * Restart the idle tick when the CPU is woken up from idle
274 void tick_nohz_restart_sched_tick(void)
276 int cpu = smp_processor_id();
277 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
281 if (!ts->tick_stopped)
284 /* Update jiffies first */
288 tick_do_update_jiffies64(now);
289 cpu_clear(cpu, nohz_cpu_mask);
291 /* Account the idle time */
292 delta = ktime_sub(now, ts->idle_entrytime);
293 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
296 * We stopped the tick in idle. Update process times would miss the
297 * time we slept as update_process_times does only a 1 tick
298 * accounting. Enforce that this is accounted to idle !
300 ticks = jiffies - ts->idle_jiffies;
302 * We might be one off. Do not randomly account a huge number of ticks!
304 if (ticks && ticks < LONG_MAX) {
305 add_preempt_count(HARDIRQ_OFFSET);
306 account_system_time(current, HARDIRQ_OFFSET,
307 jiffies_to_cputime(ticks));
308 sub_preempt_count(HARDIRQ_OFFSET);
312 * Cancel the scheduled timer and restore the tick
314 ts->tick_stopped = 0;
315 hrtimer_cancel(&ts->sched_timer);
316 ts->sched_timer.expires = ts->idle_tick;
319 /* Forward the time to expire in the future */
320 hrtimer_forward(&ts->sched_timer, now, tick_period);
322 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
323 hrtimer_start(&ts->sched_timer,
324 ts->sched_timer.expires,
326 /* Check, if the timer was already in the past */
327 if (hrtimer_active(&ts->sched_timer))
330 if (!tick_program_event(ts->sched_timer.expires, 0))
333 /* Update jiffies and reread time */
334 tick_do_update_jiffies64(now);
340 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
342 hrtimer_forward(&ts->sched_timer, now, tick_period);
343 return tick_program_event(ts->sched_timer.expires, 0);
347 * The nohz low res interrupt handler
349 static void tick_nohz_handler(struct clock_event_device *dev)
351 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
352 struct pt_regs *regs = get_irq_regs();
353 int cpu = smp_processor_id();
354 ktime_t now = ktime_get();
356 dev->next_event.tv64 = KTIME_MAX;
359 * Check if the do_timer duty was dropped. We don't care about
360 * concurrency: This happens only when the cpu in charge went
361 * into a long sleep. If two cpus happen to assign themself to
362 * this duty, then the jiffies update is still serialized by
365 if (unlikely(tick_do_timer_cpu == -1))
366 tick_do_timer_cpu = cpu;
368 /* Check, if the jiffies need an update */
369 if (tick_do_timer_cpu == cpu)
370 tick_do_update_jiffies64(now);
373 * When we are idle and the tick is stopped, we have to touch
374 * the watchdog as we might not schedule for a really long
375 * time. This happens on complete idle SMP systems while
376 * waiting on the login prompt. We also increment the "start
377 * of idle" jiffy stamp so the idle accounting adjustment we
378 * do when we go busy again does not account too much ticks.
380 if (ts->tick_stopped) {
381 touch_softlockup_watchdog();
385 update_process_times(user_mode(regs));
386 profile_tick(CPU_PROFILING);
388 /* Do not restart, when we are in the idle loop */
389 if (ts->tick_stopped)
392 while (tick_nohz_reprogram(ts, now)) {
394 tick_do_update_jiffies64(now);
399 * tick_nohz_switch_to_nohz - switch to nohz mode
401 static void tick_nohz_switch_to_nohz(void)
403 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
406 if (!tick_nohz_enabled)
410 if (tick_switch_to_oneshot(tick_nohz_handler)) {
415 ts->nohz_mode = NOHZ_MODE_LOWRES;
418 * Recycle the hrtimer in ts, so we can share the
419 * hrtimer_forward with the highres code.
421 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
422 /* Get the next period */
423 next = tick_init_jiffy_update();
426 ts->sched_timer.expires = next;
427 if (!tick_program_event(next, 0))
429 next = ktime_add(next, tick_period);
433 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
439 static inline void tick_nohz_switch_to_nohz(void) { }
444 * High resolution timer specific code
446 #ifdef CONFIG_HIGH_RES_TIMERS
448 * We rearm the timer until we get disabled by the idle code
449 * Called with interrupts disabled and timer->base->cpu_base->lock held.
451 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
453 struct tick_sched *ts =
454 container_of(timer, struct tick_sched, sched_timer);
455 struct hrtimer_cpu_base *base = timer->base->cpu_base;
456 struct pt_regs *regs = get_irq_regs();
457 ktime_t now = ktime_get();
458 int cpu = smp_processor_id();
462 * Check if the do_timer duty was dropped. We don't care about
463 * concurrency: This happens only when the cpu in charge went
464 * into a long sleep. If two cpus happen to assign themself to
465 * this duty, then the jiffies update is still serialized by
468 if (unlikely(tick_do_timer_cpu == -1))
469 tick_do_timer_cpu = cpu;
472 /* Check, if the jiffies need an update */
473 if (tick_do_timer_cpu == cpu)
474 tick_do_update_jiffies64(now);
477 * Do not call, when we are not in irq context and have
478 * no valid regs pointer
482 * When we are idle and the tick is stopped, we have to touch
483 * the watchdog as we might not schedule for a really long
484 * time. This happens on complete idle SMP systems while
485 * waiting on the login prompt. We also increment the "start of
486 * idle" jiffy stamp so the idle accounting adjustment we do
487 * when we go busy again does not account too much ticks.
489 if (ts->tick_stopped) {
490 touch_softlockup_watchdog();
494 * update_process_times() might take tasklist_lock, hence
495 * drop the base lock. sched-tick hrtimers are per-CPU and
496 * never accessible by userspace APIs, so this is safe to do.
498 spin_unlock(&base->lock);
499 update_process_times(user_mode(regs));
500 profile_tick(CPU_PROFILING);
501 spin_lock(&base->lock);
504 /* Do not restart, when we are in the idle loop */
505 if (ts->tick_stopped)
506 return HRTIMER_NORESTART;
508 hrtimer_forward(timer, now, tick_period);
510 return HRTIMER_RESTART;
514 * tick_setup_sched_timer - setup the tick emulation timer
516 void tick_setup_sched_timer(void)
518 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
519 ktime_t now = ktime_get();
522 * Emulate tick processing via per-CPU hrtimers:
524 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
525 ts->sched_timer.function = tick_sched_timer;
526 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
528 /* Get the next period */
529 ts->sched_timer.expires = tick_init_jiffy_update();
532 hrtimer_forward(&ts->sched_timer, now, tick_period);
533 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires,
535 /* Check, if the timer was already in the past */
536 if (hrtimer_active(&ts->sched_timer))
542 if (tick_nohz_enabled)
543 ts->nohz_mode = NOHZ_MODE_HIGHRES;
547 void tick_cancel_sched_timer(int cpu)
549 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
551 if (ts->sched_timer.base)
552 hrtimer_cancel(&ts->sched_timer);
553 ts->tick_stopped = 0;
554 ts->nohz_mode = NOHZ_MODE_INACTIVE;
556 #endif /* HIGH_RES_TIMERS */
559 * Async notification about clocksource changes
561 void tick_clock_notify(void)
565 for_each_possible_cpu(cpu)
566 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
570 * Async notification about clock event changes
572 void tick_oneshot_notify(void)
574 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
576 set_bit(0, &ts->check_clocks);
580 * Check, if a change happened, which makes oneshot possible.
582 * Called cyclic from the hrtimer softirq (driven by the timer
583 * softirq) allow_nohz signals, that we can switch into low-res nohz
584 * mode, because high resolution timers are disabled (either compile
587 int tick_check_oneshot_change(int allow_nohz)
589 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
591 if (!test_and_clear_bit(0, &ts->check_clocks))
594 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
597 if (!timekeeping_is_continuous() || !tick_is_oneshot_available())
603 tick_nohz_switch_to_nohz();