xref: /openbmc/linux/kernel/time/tick-sched.c (revision d7a3d85e)
1 /*
2  *  linux/kernel/time/tick-sched.c
3  *
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
7  *
8  *  No idle tick implementation for low and high resolution timers
9  *
10  *  Started by: Thomas Gleixner and Ingo Molnar
11  *
12  *  Distribute under GPLv2.
13  */
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/module.h>
23 #include <linux/irq_work.h>
24 #include <linux/posix-timers.h>
25 #include <linux/perf_event.h>
26 #include <linux/context_tracking.h>
27 
28 #include <asm/irq_regs.h>
29 
30 #include "tick-internal.h"
31 
32 #include <trace/events/timer.h>
33 
34 /*
35  * Per cpu nohz control structure
36  */
37 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
38 
39 /*
40  * The time, when the last jiffy update happened. Protected by jiffies_lock.
41  */
42 static ktime_t last_jiffies_update;
43 
44 struct tick_sched *tick_get_tick_sched(int cpu)
45 {
46 	return &per_cpu(tick_cpu_sched, cpu);
47 }
48 
49 /*
50  * Must be called with interrupts disabled !
51  */
52 static void tick_do_update_jiffies64(ktime_t now)
53 {
54 	unsigned long ticks = 0;
55 	ktime_t delta;
56 
57 	/*
58 	 * Do a quick check without holding jiffies_lock:
59 	 */
60 	delta = ktime_sub(now, last_jiffies_update);
61 	if (delta.tv64 < tick_period.tv64)
62 		return;
63 
64 	/* Reevalute with jiffies_lock held */
65 	write_seqlock(&jiffies_lock);
66 
67 	delta = ktime_sub(now, last_jiffies_update);
68 	if (delta.tv64 >= tick_period.tv64) {
69 
70 		delta = ktime_sub(delta, tick_period);
71 		last_jiffies_update = ktime_add(last_jiffies_update,
72 						tick_period);
73 
74 		/* Slow path for long timeouts */
75 		if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 			s64 incr = ktime_to_ns(tick_period);
77 
78 			ticks = ktime_divns(delta, incr);
79 
80 			last_jiffies_update = ktime_add_ns(last_jiffies_update,
81 							   incr * ticks);
82 		}
83 		do_timer(++ticks);
84 
85 		/* Keep the tick_next_period variable up to date */
86 		tick_next_period = ktime_add(last_jiffies_update, tick_period);
87 	} else {
88 		write_sequnlock(&jiffies_lock);
89 		return;
90 	}
91 	write_sequnlock(&jiffies_lock);
92 	update_wall_time();
93 }
94 
95 /*
96  * Initialize and return retrieve the jiffies update.
97  */
98 static ktime_t tick_init_jiffy_update(void)
99 {
100 	ktime_t period;
101 
102 	write_seqlock(&jiffies_lock);
103 	/* Did we start the jiffies update yet ? */
104 	if (last_jiffies_update.tv64 == 0)
105 		last_jiffies_update = tick_next_period;
106 	period = last_jiffies_update;
107 	write_sequnlock(&jiffies_lock);
108 	return period;
109 }
110 
111 
112 static void tick_sched_do_timer(ktime_t now)
113 {
114 	int cpu = smp_processor_id();
115 
116 #ifdef CONFIG_NO_HZ_COMMON
117 	/*
118 	 * Check if the do_timer duty was dropped. We don't care about
119 	 * concurrency: This happens only when the cpu in charge went
120 	 * into a long sleep. If two cpus happen to assign themself to
121 	 * this duty, then the jiffies update is still serialized by
122 	 * jiffies_lock.
123 	 */
124 	if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
125 	    && !tick_nohz_full_cpu(cpu))
126 		tick_do_timer_cpu = cpu;
127 #endif
128 
129 	/* Check, if the jiffies need an update */
130 	if (tick_do_timer_cpu == cpu)
131 		tick_do_update_jiffies64(now);
132 }
133 
134 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
135 {
136 #ifdef CONFIG_NO_HZ_COMMON
137 	/*
138 	 * When we are idle and the tick is stopped, we have to touch
139 	 * the watchdog as we might not schedule for a really long
140 	 * time. This happens on complete idle SMP systems while
141 	 * waiting on the login prompt. We also increment the "start of
142 	 * idle" jiffy stamp so the idle accounting adjustment we do
143 	 * when we go busy again does not account too much ticks.
144 	 */
145 	if (ts->tick_stopped) {
146 		touch_softlockup_watchdog();
147 		if (is_idle_task(current))
148 			ts->idle_jiffies++;
149 	}
150 #endif
151 	update_process_times(user_mode(regs));
152 	profile_tick(CPU_PROFILING);
153 }
154 
155 #ifdef CONFIG_NO_HZ_FULL
156 cpumask_var_t tick_nohz_full_mask;
157 cpumask_var_t housekeeping_mask;
158 bool tick_nohz_full_running;
159 
160 static bool can_stop_full_tick(void)
161 {
162 	WARN_ON_ONCE(!irqs_disabled());
163 
164 	if (!sched_can_stop_tick()) {
165 		trace_tick_stop(0, "more than 1 task in runqueue\n");
166 		return false;
167 	}
168 
169 	if (!posix_cpu_timers_can_stop_tick(current)) {
170 		trace_tick_stop(0, "posix timers running\n");
171 		return false;
172 	}
173 
174 	if (!perf_event_can_stop_tick()) {
175 		trace_tick_stop(0, "perf events running\n");
176 		return false;
177 	}
178 
179 	/* sched_clock_tick() needs us? */
180 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
181 	/*
182 	 * TODO: kick full dynticks CPUs when
183 	 * sched_clock_stable is set.
184 	 */
185 	if (!sched_clock_stable()) {
186 		trace_tick_stop(0, "unstable sched clock\n");
187 		/*
188 		 * Don't allow the user to think they can get
189 		 * full NO_HZ with this machine.
190 		 */
191 		WARN_ONCE(tick_nohz_full_running,
192 			  "NO_HZ FULL will not work with unstable sched clock");
193 		return false;
194 	}
195 #endif
196 
197 	return true;
198 }
199 
200 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
201 
202 /*
203  * Re-evaluate the need for the tick on the current CPU
204  * and restart it if necessary.
205  */
206 void __tick_nohz_full_check(void)
207 {
208 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
209 
210 	if (tick_nohz_full_cpu(smp_processor_id())) {
211 		if (ts->tick_stopped && !is_idle_task(current)) {
212 			if (!can_stop_full_tick())
213 				tick_nohz_restart_sched_tick(ts, ktime_get());
214 		}
215 	}
216 }
217 
218 static void nohz_full_kick_work_func(struct irq_work *work)
219 {
220 	__tick_nohz_full_check();
221 }
222 
223 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
224 	.func = nohz_full_kick_work_func,
225 };
226 
227 /*
228  * Kick this CPU if it's full dynticks in order to force it to
229  * re-evaluate its dependency on the tick and restart it if necessary.
230  * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
231  * is NMI safe.
232  */
233 void tick_nohz_full_kick(void)
234 {
235 	if (!tick_nohz_full_cpu(smp_processor_id()))
236 		return;
237 
238 	irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
239 }
240 
241 /*
242  * Kick the CPU if it's full dynticks in order to force it to
243  * re-evaluate its dependency on the tick and restart it if necessary.
244  */
245 void tick_nohz_full_kick_cpu(int cpu)
246 {
247 	if (!tick_nohz_full_cpu(cpu))
248 		return;
249 
250 	irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
251 }
252 
253 static void nohz_full_kick_ipi(void *info)
254 {
255 	__tick_nohz_full_check();
256 }
257 
258 /*
259  * Kick all full dynticks CPUs in order to force these to re-evaluate
260  * their dependency on the tick and restart it if necessary.
261  */
262 void tick_nohz_full_kick_all(void)
263 {
264 	if (!tick_nohz_full_running)
265 		return;
266 
267 	preempt_disable();
268 	smp_call_function_many(tick_nohz_full_mask,
269 			       nohz_full_kick_ipi, NULL, false);
270 	tick_nohz_full_kick();
271 	preempt_enable();
272 }
273 
274 /*
275  * Re-evaluate the need for the tick as we switch the current task.
276  * It might need the tick due to per task/process properties:
277  * perf events, posix cpu timers, ...
278  */
279 void __tick_nohz_task_switch(struct task_struct *tsk)
280 {
281 	unsigned long flags;
282 
283 	local_irq_save(flags);
284 
285 	if (!tick_nohz_full_cpu(smp_processor_id()))
286 		goto out;
287 
288 	if (tick_nohz_tick_stopped() && !can_stop_full_tick())
289 		tick_nohz_full_kick();
290 
291 out:
292 	local_irq_restore(flags);
293 }
294 
295 /* Parse the boot-time nohz CPU list from the kernel parameters. */
296 static int __init tick_nohz_full_setup(char *str)
297 {
298 	alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
299 	if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
300 		pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
301 		free_bootmem_cpumask_var(tick_nohz_full_mask);
302 		return 1;
303 	}
304 	tick_nohz_full_running = true;
305 
306 	return 1;
307 }
308 __setup("nohz_full=", tick_nohz_full_setup);
309 
310 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
311 						 unsigned long action,
312 						 void *hcpu)
313 {
314 	unsigned int cpu = (unsigned long)hcpu;
315 
316 	switch (action & ~CPU_TASKS_FROZEN) {
317 	case CPU_DOWN_PREPARE:
318 		/*
319 		 * If we handle the timekeeping duty for full dynticks CPUs,
320 		 * we can't safely shutdown that CPU.
321 		 */
322 		if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
323 			return NOTIFY_BAD;
324 		break;
325 	}
326 	return NOTIFY_OK;
327 }
328 
329 static int tick_nohz_init_all(void)
330 {
331 	int err = -1;
332 
333 #ifdef CONFIG_NO_HZ_FULL_ALL
334 	if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
335 		WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
336 		return err;
337 	}
338 	err = 0;
339 	cpumask_setall(tick_nohz_full_mask);
340 	tick_nohz_full_running = true;
341 #endif
342 	return err;
343 }
344 
345 void __init tick_nohz_init(void)
346 {
347 	int cpu;
348 
349 	if (!tick_nohz_full_running) {
350 		if (tick_nohz_init_all() < 0)
351 			return;
352 	}
353 
354 	if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
355 		WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
356 		cpumask_clear(tick_nohz_full_mask);
357 		tick_nohz_full_running = false;
358 		return;
359 	}
360 
361 	/*
362 	 * Full dynticks uses irq work to drive the tick rescheduling on safe
363 	 * locking contexts. But then we need irq work to raise its own
364 	 * interrupts to avoid circular dependency on the tick
365 	 */
366 	if (!arch_irq_work_has_interrupt()) {
367 		pr_warning("NO_HZ: Can't run full dynticks because arch doesn't "
368 			   "support irq work self-IPIs\n");
369 		cpumask_clear(tick_nohz_full_mask);
370 		cpumask_copy(housekeeping_mask, cpu_possible_mask);
371 		tick_nohz_full_running = false;
372 		return;
373 	}
374 
375 	cpu = smp_processor_id();
376 
377 	if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
378 		pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
379 		cpumask_clear_cpu(cpu, tick_nohz_full_mask);
380 	}
381 
382 	cpumask_andnot(housekeeping_mask,
383 		       cpu_possible_mask, tick_nohz_full_mask);
384 
385 	for_each_cpu(cpu, tick_nohz_full_mask)
386 		context_tracking_cpu_set(cpu);
387 
388 	cpu_notifier(tick_nohz_cpu_down_callback, 0);
389 	pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
390 		cpumask_pr_args(tick_nohz_full_mask));
391 }
392 #endif
393 
394 /*
395  * NOHZ - aka dynamic tick functionality
396  */
397 #ifdef CONFIG_NO_HZ_COMMON
398 /*
399  * NO HZ enabled ?
400  */
401 static int tick_nohz_enabled __read_mostly  = 1;
402 int tick_nohz_active  __read_mostly;
403 /*
404  * Enable / Disable tickless mode
405  */
406 static int __init setup_tick_nohz(char *str)
407 {
408 	if (!strcmp(str, "off"))
409 		tick_nohz_enabled = 0;
410 	else if (!strcmp(str, "on"))
411 		tick_nohz_enabled = 1;
412 	else
413 		return 0;
414 	return 1;
415 }
416 
417 __setup("nohz=", setup_tick_nohz);
418 
419 int tick_nohz_tick_stopped(void)
420 {
421 	return __this_cpu_read(tick_cpu_sched.tick_stopped);
422 }
423 
424 /**
425  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
426  *
427  * Called from interrupt entry when the CPU was idle
428  *
429  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
430  * must be updated. Otherwise an interrupt handler could use a stale jiffy
431  * value. We do this unconditionally on any cpu, as we don't know whether the
432  * cpu, which has the update task assigned is in a long sleep.
433  */
434 static void tick_nohz_update_jiffies(ktime_t now)
435 {
436 	unsigned long flags;
437 
438 	__this_cpu_write(tick_cpu_sched.idle_waketime, now);
439 
440 	local_irq_save(flags);
441 	tick_do_update_jiffies64(now);
442 	local_irq_restore(flags);
443 
444 	touch_softlockup_watchdog();
445 }
446 
447 /*
448  * Updates the per cpu time idle statistics counters
449  */
450 static void
451 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
452 {
453 	ktime_t delta;
454 
455 	if (ts->idle_active) {
456 		delta = ktime_sub(now, ts->idle_entrytime);
457 		if (nr_iowait_cpu(cpu) > 0)
458 			ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
459 		else
460 			ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
461 		ts->idle_entrytime = now;
462 	}
463 
464 	if (last_update_time)
465 		*last_update_time = ktime_to_us(now);
466 
467 }
468 
469 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
470 {
471 	update_ts_time_stats(smp_processor_id(), ts, now, NULL);
472 	ts->idle_active = 0;
473 
474 	sched_clock_idle_wakeup_event(0);
475 }
476 
477 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
478 {
479 	ktime_t now = ktime_get();
480 
481 	ts->idle_entrytime = now;
482 	ts->idle_active = 1;
483 	sched_clock_idle_sleep_event();
484 	return now;
485 }
486 
487 /**
488  * get_cpu_idle_time_us - get the total idle time of a cpu
489  * @cpu: CPU number to query
490  * @last_update_time: variable to store update time in. Do not update
491  * counters if NULL.
492  *
493  * Return the cummulative idle time (since boot) for a given
494  * CPU, in microseconds.
495  *
496  * This time is measured via accounting rather than sampling,
497  * and is as accurate as ktime_get() is.
498  *
499  * This function returns -1 if NOHZ is not enabled.
500  */
501 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
502 {
503 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
504 	ktime_t now, idle;
505 
506 	if (!tick_nohz_active)
507 		return -1;
508 
509 	now = ktime_get();
510 	if (last_update_time) {
511 		update_ts_time_stats(cpu, ts, now, last_update_time);
512 		idle = ts->idle_sleeptime;
513 	} else {
514 		if (ts->idle_active && !nr_iowait_cpu(cpu)) {
515 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
516 
517 			idle = ktime_add(ts->idle_sleeptime, delta);
518 		} else {
519 			idle = ts->idle_sleeptime;
520 		}
521 	}
522 
523 	return ktime_to_us(idle);
524 
525 }
526 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
527 
528 /**
529  * get_cpu_iowait_time_us - get the total iowait time of a cpu
530  * @cpu: CPU number to query
531  * @last_update_time: variable to store update time in. Do not update
532  * counters if NULL.
533  *
534  * Return the cummulative iowait time (since boot) for a given
535  * CPU, in microseconds.
536  *
537  * This time is measured via accounting rather than sampling,
538  * and is as accurate as ktime_get() is.
539  *
540  * This function returns -1 if NOHZ is not enabled.
541  */
542 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
543 {
544 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
545 	ktime_t now, iowait;
546 
547 	if (!tick_nohz_active)
548 		return -1;
549 
550 	now = ktime_get();
551 	if (last_update_time) {
552 		update_ts_time_stats(cpu, ts, now, last_update_time);
553 		iowait = ts->iowait_sleeptime;
554 	} else {
555 		if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
556 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
557 
558 			iowait = ktime_add(ts->iowait_sleeptime, delta);
559 		} else {
560 			iowait = ts->iowait_sleeptime;
561 		}
562 	}
563 
564 	return ktime_to_us(iowait);
565 }
566 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
567 
568 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
569 					 ktime_t now, int cpu)
570 {
571 	unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
572 	ktime_t last_update, expires, ret = { .tv64 = 0 };
573 	unsigned long rcu_delta_jiffies;
574 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
575 	u64 time_delta;
576 
577 	time_delta = timekeeping_max_deferment();
578 
579 	/* Read jiffies and the time when jiffies were updated last */
580 	do {
581 		seq = read_seqbegin(&jiffies_lock);
582 		last_update = last_jiffies_update;
583 		last_jiffies = jiffies;
584 	} while (read_seqretry(&jiffies_lock, seq));
585 
586 	if (rcu_needs_cpu(&rcu_delta_jiffies) ||
587 	    arch_needs_cpu() || irq_work_needs_cpu()) {
588 		next_jiffies = last_jiffies + 1;
589 		delta_jiffies = 1;
590 	} else {
591 		/* Get the next timer wheel timer */
592 		next_jiffies = get_next_timer_interrupt(last_jiffies);
593 		delta_jiffies = next_jiffies - last_jiffies;
594 		if (rcu_delta_jiffies < delta_jiffies) {
595 			next_jiffies = last_jiffies + rcu_delta_jiffies;
596 			delta_jiffies = rcu_delta_jiffies;
597 		}
598 	}
599 
600 	/*
601 	 * Do not stop the tick, if we are only one off (or less)
602 	 * or if the cpu is required for RCU:
603 	 */
604 	if (!ts->tick_stopped && delta_jiffies <= 1)
605 		goto out;
606 
607 	/* Schedule the tick, if we are at least one jiffie off */
608 	if ((long)delta_jiffies >= 1) {
609 
610 		/*
611 		 * If this cpu is the one which updates jiffies, then
612 		 * give up the assignment and let it be taken by the
613 		 * cpu which runs the tick timer next, which might be
614 		 * this cpu as well. If we don't drop this here the
615 		 * jiffies might be stale and do_timer() never
616 		 * invoked. Keep track of the fact that it was the one
617 		 * which had the do_timer() duty last. If this cpu is
618 		 * the one which had the do_timer() duty last, we
619 		 * limit the sleep time to the timekeeping
620 		 * max_deferement value which we retrieved
621 		 * above. Otherwise we can sleep as long as we want.
622 		 */
623 		if (cpu == tick_do_timer_cpu) {
624 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
625 			ts->do_timer_last = 1;
626 		} else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
627 			time_delta = KTIME_MAX;
628 			ts->do_timer_last = 0;
629 		} else if (!ts->do_timer_last) {
630 			time_delta = KTIME_MAX;
631 		}
632 
633 #ifdef CONFIG_NO_HZ_FULL
634 		if (!ts->inidle) {
635 			time_delta = min(time_delta,
636 					 scheduler_tick_max_deferment());
637 		}
638 #endif
639 
640 		/*
641 		 * calculate the expiry time for the next timer wheel
642 		 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
643 		 * that there is no timer pending or at least extremely
644 		 * far into the future (12 days for HZ=1000). In this
645 		 * case we set the expiry to the end of time.
646 		 */
647 		if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
648 			/*
649 			 * Calculate the time delta for the next timer event.
650 			 * If the time delta exceeds the maximum time delta
651 			 * permitted by the current clocksource then adjust
652 			 * the time delta accordingly to ensure the
653 			 * clocksource does not wrap.
654 			 */
655 			time_delta = min_t(u64, time_delta,
656 					   tick_period.tv64 * delta_jiffies);
657 		}
658 
659 		if (time_delta < KTIME_MAX)
660 			expires = ktime_add_ns(last_update, time_delta);
661 		else
662 			expires.tv64 = KTIME_MAX;
663 
664 		/* Skip reprogram of event if its not changed */
665 		if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
666 			goto out;
667 
668 		ret = expires;
669 
670 		/*
671 		 * nohz_stop_sched_tick can be called several times before
672 		 * the nohz_restart_sched_tick is called. This happens when
673 		 * interrupts arrive which do not cause a reschedule. In the
674 		 * first call we save the current tick time, so we can restart
675 		 * the scheduler tick in nohz_restart_sched_tick.
676 		 */
677 		if (!ts->tick_stopped) {
678 			nohz_balance_enter_idle(cpu);
679 			calc_load_enter_idle();
680 
681 			ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
682 			ts->tick_stopped = 1;
683 			trace_tick_stop(1, " ");
684 		}
685 
686 		/*
687 		 * If the expiration time == KTIME_MAX, then
688 		 * in this case we simply stop the tick timer.
689 		 */
690 		 if (unlikely(expires.tv64 == KTIME_MAX)) {
691 			if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
692 				hrtimer_cancel(&ts->sched_timer);
693 			goto out;
694 		}
695 
696 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
697 			hrtimer_start(&ts->sched_timer, expires,
698 				      HRTIMER_MODE_ABS_PINNED);
699 			/* Check, if the timer was already in the past */
700 			if (hrtimer_active(&ts->sched_timer))
701 				goto out;
702 		} else if (!tick_program_event(expires, 0))
703 				goto out;
704 		/*
705 		 * We are past the event already. So we crossed a
706 		 * jiffie boundary. Update jiffies and raise the
707 		 * softirq.
708 		 */
709 		tick_do_update_jiffies64(ktime_get());
710 	}
711 	raise_softirq_irqoff(TIMER_SOFTIRQ);
712 out:
713 	ts->next_jiffies = next_jiffies;
714 	ts->last_jiffies = last_jiffies;
715 	ts->sleep_length = ktime_sub(dev->next_event, now);
716 
717 	return ret;
718 }
719 
720 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
721 {
722 #ifdef CONFIG_NO_HZ_FULL
723 	int cpu = smp_processor_id();
724 
725 	if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
726 		return;
727 
728 	if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
729 		return;
730 
731 	if (!can_stop_full_tick())
732 		return;
733 
734 	tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
735 #endif
736 }
737 
738 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
739 {
740 	/*
741 	 * If this cpu is offline and it is the one which updates
742 	 * jiffies, then give up the assignment and let it be taken by
743 	 * the cpu which runs the tick timer next. If we don't drop
744 	 * this here the jiffies might be stale and do_timer() never
745 	 * invoked.
746 	 */
747 	if (unlikely(!cpu_online(cpu))) {
748 		if (cpu == tick_do_timer_cpu)
749 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
750 		return false;
751 	}
752 
753 	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
754 		ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
755 		return false;
756 	}
757 
758 	if (need_resched())
759 		return false;
760 
761 	if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
762 		static int ratelimit;
763 
764 		if (ratelimit < 10 &&
765 		    (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
766 			pr_warn("NOHZ: local_softirq_pending %02x\n",
767 				(unsigned int) local_softirq_pending());
768 			ratelimit++;
769 		}
770 		return false;
771 	}
772 
773 	if (tick_nohz_full_enabled()) {
774 		/*
775 		 * Keep the tick alive to guarantee timekeeping progression
776 		 * if there are full dynticks CPUs around
777 		 */
778 		if (tick_do_timer_cpu == cpu)
779 			return false;
780 		/*
781 		 * Boot safety: make sure the timekeeping duty has been
782 		 * assigned before entering dyntick-idle mode,
783 		 */
784 		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
785 			return false;
786 	}
787 
788 	return true;
789 }
790 
791 static void __tick_nohz_idle_enter(struct tick_sched *ts)
792 {
793 	ktime_t now, expires;
794 	int cpu = smp_processor_id();
795 
796 	now = tick_nohz_start_idle(ts);
797 
798 	if (can_stop_idle_tick(cpu, ts)) {
799 		int was_stopped = ts->tick_stopped;
800 
801 		ts->idle_calls++;
802 
803 		expires = tick_nohz_stop_sched_tick(ts, now, cpu);
804 		if (expires.tv64 > 0LL) {
805 			ts->idle_sleeps++;
806 			ts->idle_expires = expires;
807 		}
808 
809 		if (!was_stopped && ts->tick_stopped)
810 			ts->idle_jiffies = ts->last_jiffies;
811 	}
812 }
813 
814 /**
815  * tick_nohz_idle_enter - stop the idle tick from the idle task
816  *
817  * When the next event is more than a tick into the future, stop the idle tick
818  * Called when we start the idle loop.
819  *
820  * The arch is responsible of calling:
821  *
822  * - rcu_idle_enter() after its last use of RCU before the CPU is put
823  *  to sleep.
824  * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
825  */
826 void tick_nohz_idle_enter(void)
827 {
828 	struct tick_sched *ts;
829 
830 	WARN_ON_ONCE(irqs_disabled());
831 
832 	/*
833  	 * Update the idle state in the scheduler domain hierarchy
834  	 * when tick_nohz_stop_sched_tick() is called from the idle loop.
835  	 * State will be updated to busy during the first busy tick after
836  	 * exiting idle.
837  	 */
838 	set_cpu_sd_state_idle();
839 
840 	local_irq_disable();
841 
842 	ts = this_cpu_ptr(&tick_cpu_sched);
843 	ts->inidle = 1;
844 	__tick_nohz_idle_enter(ts);
845 
846 	local_irq_enable();
847 }
848 
849 /**
850  * tick_nohz_irq_exit - update next tick event from interrupt exit
851  *
852  * When an interrupt fires while we are idle and it doesn't cause
853  * a reschedule, it may still add, modify or delete a timer, enqueue
854  * an RCU callback, etc...
855  * So we need to re-calculate and reprogram the next tick event.
856  */
857 void tick_nohz_irq_exit(void)
858 {
859 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
860 
861 	if (ts->inidle)
862 		__tick_nohz_idle_enter(ts);
863 	else
864 		tick_nohz_full_stop_tick(ts);
865 }
866 
867 /**
868  * tick_nohz_get_sleep_length - return the length of the current sleep
869  *
870  * Called from power state control code with interrupts disabled
871  */
872 ktime_t tick_nohz_get_sleep_length(void)
873 {
874 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
875 
876 	return ts->sleep_length;
877 }
878 
879 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
880 {
881 	hrtimer_cancel(&ts->sched_timer);
882 	hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
883 
884 	while (1) {
885 		/* Forward the time to expire in the future */
886 		hrtimer_forward(&ts->sched_timer, now, tick_period);
887 
888 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
889 			hrtimer_start_expires(&ts->sched_timer,
890 					      HRTIMER_MODE_ABS_PINNED);
891 			/* Check, if the timer was already in the past */
892 			if (hrtimer_active(&ts->sched_timer))
893 				break;
894 		} else {
895 			if (!tick_program_event(
896 				hrtimer_get_expires(&ts->sched_timer), 0))
897 				break;
898 		}
899 		/* Reread time and update jiffies */
900 		now = ktime_get();
901 		tick_do_update_jiffies64(now);
902 	}
903 }
904 
905 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
906 {
907 	/* Update jiffies first */
908 	tick_do_update_jiffies64(now);
909 	update_cpu_load_nohz();
910 
911 	calc_load_exit_idle();
912 	touch_softlockup_watchdog();
913 	/*
914 	 * Cancel the scheduled timer and restore the tick
915 	 */
916 	ts->tick_stopped  = 0;
917 	ts->idle_exittime = now;
918 
919 	tick_nohz_restart(ts, now);
920 }
921 
922 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
923 {
924 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
925 	unsigned long ticks;
926 
927 	if (vtime_accounting_enabled())
928 		return;
929 	/*
930 	 * We stopped the tick in idle. Update process times would miss the
931 	 * time we slept as update_process_times does only a 1 tick
932 	 * accounting. Enforce that this is accounted to idle !
933 	 */
934 	ticks = jiffies - ts->idle_jiffies;
935 	/*
936 	 * We might be one off. Do not randomly account a huge number of ticks!
937 	 */
938 	if (ticks && ticks < LONG_MAX)
939 		account_idle_ticks(ticks);
940 #endif
941 }
942 
943 /**
944  * tick_nohz_idle_exit - restart the idle tick from the idle task
945  *
946  * Restart the idle tick when the CPU is woken up from idle
947  * This also exit the RCU extended quiescent state. The CPU
948  * can use RCU again after this function is called.
949  */
950 void tick_nohz_idle_exit(void)
951 {
952 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
953 	ktime_t now;
954 
955 	local_irq_disable();
956 
957 	WARN_ON_ONCE(!ts->inidle);
958 
959 	ts->inidle = 0;
960 
961 	if (ts->idle_active || ts->tick_stopped)
962 		now = ktime_get();
963 
964 	if (ts->idle_active)
965 		tick_nohz_stop_idle(ts, now);
966 
967 	if (ts->tick_stopped) {
968 		tick_nohz_restart_sched_tick(ts, now);
969 		tick_nohz_account_idle_ticks(ts);
970 	}
971 
972 	local_irq_enable();
973 }
974 
975 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
976 {
977 	hrtimer_forward(&ts->sched_timer, now, tick_period);
978 	return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
979 }
980 
981 /*
982  * The nohz low res interrupt handler
983  */
984 static void tick_nohz_handler(struct clock_event_device *dev)
985 {
986 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
987 	struct pt_regs *regs = get_irq_regs();
988 	ktime_t now = ktime_get();
989 
990 	dev->next_event.tv64 = KTIME_MAX;
991 
992 	tick_sched_do_timer(now);
993 	tick_sched_handle(ts, regs);
994 
995 	/* No need to reprogram if we are running tickless  */
996 	if (unlikely(ts->tick_stopped))
997 		return;
998 
999 	while (tick_nohz_reprogram(ts, now)) {
1000 		now = ktime_get();
1001 		tick_do_update_jiffies64(now);
1002 	}
1003 }
1004 
1005 /**
1006  * tick_nohz_switch_to_nohz - switch to nohz mode
1007  */
1008 static void tick_nohz_switch_to_nohz(void)
1009 {
1010 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1011 	ktime_t next;
1012 
1013 	if (!tick_nohz_enabled)
1014 		return;
1015 
1016 	local_irq_disable();
1017 	if (tick_switch_to_oneshot(tick_nohz_handler)) {
1018 		local_irq_enable();
1019 		return;
1020 	}
1021 	tick_nohz_active = 1;
1022 	ts->nohz_mode = NOHZ_MODE_LOWRES;
1023 
1024 	/*
1025 	 * Recycle the hrtimer in ts, so we can share the
1026 	 * hrtimer_forward with the highres code.
1027 	 */
1028 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1029 	/* Get the next period */
1030 	next = tick_init_jiffy_update();
1031 
1032 	for (;;) {
1033 		hrtimer_set_expires(&ts->sched_timer, next);
1034 		if (!tick_program_event(next, 0))
1035 			break;
1036 		next = ktime_add(next, tick_period);
1037 	}
1038 	local_irq_enable();
1039 }
1040 
1041 /*
1042  * When NOHZ is enabled and the tick is stopped, we need to kick the
1043  * tick timer from irq_enter() so that the jiffies update is kept
1044  * alive during long running softirqs. That's ugly as hell, but
1045  * correctness is key even if we need to fix the offending softirq in
1046  * the first place.
1047  *
1048  * Note, this is different to tick_nohz_restart. We just kick the
1049  * timer and do not touch the other magic bits which need to be done
1050  * when idle is left.
1051  */
1052 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1053 {
1054 #if 0
1055 	/* Switch back to 2.6.27 behaviour */
1056 	ktime_t delta;
1057 
1058 	/*
1059 	 * Do not touch the tick device, when the next expiry is either
1060 	 * already reached or less/equal than the tick period.
1061 	 */
1062 	delta =	ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1063 	if (delta.tv64 <= tick_period.tv64)
1064 		return;
1065 
1066 	tick_nohz_restart(ts, now);
1067 #endif
1068 }
1069 
1070 static inline void tick_nohz_irq_enter(void)
1071 {
1072 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1073 	ktime_t now;
1074 
1075 	if (!ts->idle_active && !ts->tick_stopped)
1076 		return;
1077 	now = ktime_get();
1078 	if (ts->idle_active)
1079 		tick_nohz_stop_idle(ts, now);
1080 	if (ts->tick_stopped) {
1081 		tick_nohz_update_jiffies(now);
1082 		tick_nohz_kick_tick(ts, now);
1083 	}
1084 }
1085 
1086 #else
1087 
1088 static inline void tick_nohz_switch_to_nohz(void) { }
1089 static inline void tick_nohz_irq_enter(void) { }
1090 
1091 #endif /* CONFIG_NO_HZ_COMMON */
1092 
1093 /*
1094  * Called from irq_enter to notify about the possible interruption of idle()
1095  */
1096 void tick_irq_enter(void)
1097 {
1098 	tick_check_oneshot_broadcast_this_cpu();
1099 	tick_nohz_irq_enter();
1100 }
1101 
1102 /*
1103  * High resolution timer specific code
1104  */
1105 #ifdef CONFIG_HIGH_RES_TIMERS
1106 /*
1107  * We rearm the timer until we get disabled by the idle code.
1108  * Called with interrupts disabled.
1109  */
1110 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1111 {
1112 	struct tick_sched *ts =
1113 		container_of(timer, struct tick_sched, sched_timer);
1114 	struct pt_regs *regs = get_irq_regs();
1115 	ktime_t now = ktime_get();
1116 
1117 	tick_sched_do_timer(now);
1118 
1119 	/*
1120 	 * Do not call, when we are not in irq context and have
1121 	 * no valid regs pointer
1122 	 */
1123 	if (regs)
1124 		tick_sched_handle(ts, regs);
1125 
1126 	/* No need to reprogram if we are in idle or full dynticks mode */
1127 	if (unlikely(ts->tick_stopped))
1128 		return HRTIMER_NORESTART;
1129 
1130 	hrtimer_forward(timer, now, tick_period);
1131 
1132 	return HRTIMER_RESTART;
1133 }
1134 
1135 static int sched_skew_tick;
1136 
1137 static int __init skew_tick(char *str)
1138 {
1139 	get_option(&str, &sched_skew_tick);
1140 
1141 	return 0;
1142 }
1143 early_param("skew_tick", skew_tick);
1144 
1145 /**
1146  * tick_setup_sched_timer - setup the tick emulation timer
1147  */
1148 void tick_setup_sched_timer(void)
1149 {
1150 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1151 	ktime_t now = ktime_get();
1152 
1153 	/*
1154 	 * Emulate tick processing via per-CPU hrtimers:
1155 	 */
1156 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1157 	ts->sched_timer.function = tick_sched_timer;
1158 
1159 	/* Get the next period (per cpu) */
1160 	hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1161 
1162 	/* Offset the tick to avert jiffies_lock contention. */
1163 	if (sched_skew_tick) {
1164 		u64 offset = ktime_to_ns(tick_period) >> 1;
1165 		do_div(offset, num_possible_cpus());
1166 		offset *= smp_processor_id();
1167 		hrtimer_add_expires_ns(&ts->sched_timer, offset);
1168 	}
1169 
1170 	for (;;) {
1171 		hrtimer_forward(&ts->sched_timer, now, tick_period);
1172 		hrtimer_start_expires(&ts->sched_timer,
1173 				      HRTIMER_MODE_ABS_PINNED);
1174 		/* Check, if the timer was already in the past */
1175 		if (hrtimer_active(&ts->sched_timer))
1176 			break;
1177 		now = ktime_get();
1178 	}
1179 
1180 #ifdef CONFIG_NO_HZ_COMMON
1181 	if (tick_nohz_enabled) {
1182 		ts->nohz_mode = NOHZ_MODE_HIGHRES;
1183 		tick_nohz_active = 1;
1184 	}
1185 #endif
1186 }
1187 #endif /* HIGH_RES_TIMERS */
1188 
1189 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1190 void tick_cancel_sched_timer(int cpu)
1191 {
1192 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1193 
1194 # ifdef CONFIG_HIGH_RES_TIMERS
1195 	if (ts->sched_timer.base)
1196 		hrtimer_cancel(&ts->sched_timer);
1197 # endif
1198 
1199 	memset(ts, 0, sizeof(*ts));
1200 }
1201 #endif
1202 
1203 /**
1204  * Async notification about clocksource changes
1205  */
1206 void tick_clock_notify(void)
1207 {
1208 	int cpu;
1209 
1210 	for_each_possible_cpu(cpu)
1211 		set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1212 }
1213 
1214 /*
1215  * Async notification about clock event changes
1216  */
1217 void tick_oneshot_notify(void)
1218 {
1219 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1220 
1221 	set_bit(0, &ts->check_clocks);
1222 }
1223 
1224 /**
1225  * Check, if a change happened, which makes oneshot possible.
1226  *
1227  * Called cyclic from the hrtimer softirq (driven by the timer
1228  * softirq) allow_nohz signals, that we can switch into low-res nohz
1229  * mode, because high resolution timers are disabled (either compile
1230  * or runtime).
1231  */
1232 int tick_check_oneshot_change(int allow_nohz)
1233 {
1234 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1235 
1236 	if (!test_and_clear_bit(0, &ts->check_clocks))
1237 		return 0;
1238 
1239 	if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1240 		return 0;
1241 
1242 	if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1243 		return 0;
1244 
1245 	if (!allow_nohz)
1246 		return 1;
1247 
1248 	tick_nohz_switch_to_nohz();
1249 	return 0;
1250 }
1251