xref: /linux/kernel/time/tick-sched.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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 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 /**
420  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
421  *
422  * Called from interrupt entry when the CPU was idle
423  *
424  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
425  * must be updated. Otherwise an interrupt handler could use a stale jiffy
426  * value. We do this unconditionally on any cpu, as we don't know whether the
427  * cpu, which has the update task assigned is in a long sleep.
428  */
429 static void tick_nohz_update_jiffies(ktime_t now)
430 {
431 	unsigned long flags;
432 
433 	__this_cpu_write(tick_cpu_sched.idle_waketime, now);
434 
435 	local_irq_save(flags);
436 	tick_do_update_jiffies64(now);
437 	local_irq_restore(flags);
438 
439 	touch_softlockup_watchdog();
440 }
441 
442 /*
443  * Updates the per cpu time idle statistics counters
444  */
445 static void
446 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
447 {
448 	ktime_t delta;
449 
450 	if (ts->idle_active) {
451 		delta = ktime_sub(now, ts->idle_entrytime);
452 		if (nr_iowait_cpu(cpu) > 0)
453 			ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
454 		else
455 			ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
456 		ts->idle_entrytime = now;
457 	}
458 
459 	if (last_update_time)
460 		*last_update_time = ktime_to_us(now);
461 
462 }
463 
464 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
465 {
466 	update_ts_time_stats(smp_processor_id(), ts, now, NULL);
467 	ts->idle_active = 0;
468 
469 	sched_clock_idle_wakeup_event(0);
470 }
471 
472 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
473 {
474 	ktime_t now = ktime_get();
475 
476 	ts->idle_entrytime = now;
477 	ts->idle_active = 1;
478 	sched_clock_idle_sleep_event();
479 	return now;
480 }
481 
482 /**
483  * get_cpu_idle_time_us - get the total idle time of a cpu
484  * @cpu: CPU number to query
485  * @last_update_time: variable to store update time in. Do not update
486  * counters if NULL.
487  *
488  * Return the cummulative idle time (since boot) for a given
489  * CPU, in microseconds.
490  *
491  * This time is measured via accounting rather than sampling,
492  * and is as accurate as ktime_get() is.
493  *
494  * This function returns -1 if NOHZ is not enabled.
495  */
496 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
497 {
498 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
499 	ktime_t now, idle;
500 
501 	if (!tick_nohz_active)
502 		return -1;
503 
504 	now = ktime_get();
505 	if (last_update_time) {
506 		update_ts_time_stats(cpu, ts, now, last_update_time);
507 		idle = ts->idle_sleeptime;
508 	} else {
509 		if (ts->idle_active && !nr_iowait_cpu(cpu)) {
510 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
511 
512 			idle = ktime_add(ts->idle_sleeptime, delta);
513 		} else {
514 			idle = ts->idle_sleeptime;
515 		}
516 	}
517 
518 	return ktime_to_us(idle);
519 
520 }
521 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
522 
523 /**
524  * get_cpu_iowait_time_us - get the total iowait time of a cpu
525  * @cpu: CPU number to query
526  * @last_update_time: variable to store update time in. Do not update
527  * counters if NULL.
528  *
529  * Return the cummulative iowait time (since boot) for a given
530  * CPU, in microseconds.
531  *
532  * This time is measured via accounting rather than sampling,
533  * and is as accurate as ktime_get() is.
534  *
535  * This function returns -1 if NOHZ is not enabled.
536  */
537 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
538 {
539 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
540 	ktime_t now, iowait;
541 
542 	if (!tick_nohz_active)
543 		return -1;
544 
545 	now = ktime_get();
546 	if (last_update_time) {
547 		update_ts_time_stats(cpu, ts, now, last_update_time);
548 		iowait = ts->iowait_sleeptime;
549 	} else {
550 		if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
551 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
552 
553 			iowait = ktime_add(ts->iowait_sleeptime, delta);
554 		} else {
555 			iowait = ts->iowait_sleeptime;
556 		}
557 	}
558 
559 	return ktime_to_us(iowait);
560 }
561 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
562 
563 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
564 					 ktime_t now, int cpu)
565 {
566 	unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
567 	ktime_t last_update, expires, ret = { .tv64 = 0 };
568 	unsigned long rcu_delta_jiffies;
569 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
570 	u64 time_delta;
571 
572 	time_delta = timekeeping_max_deferment();
573 
574 	/* Read jiffies and the time when jiffies were updated last */
575 	do {
576 		seq = read_seqbegin(&jiffies_lock);
577 		last_update = last_jiffies_update;
578 		last_jiffies = jiffies;
579 	} while (read_seqretry(&jiffies_lock, seq));
580 
581 	if (rcu_needs_cpu(&rcu_delta_jiffies) ||
582 	    arch_needs_cpu() || irq_work_needs_cpu()) {
583 		next_jiffies = last_jiffies + 1;
584 		delta_jiffies = 1;
585 	} else {
586 		/* Get the next timer wheel timer */
587 		next_jiffies = get_next_timer_interrupt(last_jiffies);
588 		delta_jiffies = next_jiffies - last_jiffies;
589 		if (rcu_delta_jiffies < delta_jiffies) {
590 			next_jiffies = last_jiffies + rcu_delta_jiffies;
591 			delta_jiffies = rcu_delta_jiffies;
592 		}
593 	}
594 
595 	/*
596 	 * Do not stop the tick, if we are only one off (or less)
597 	 * or if the cpu is required for RCU:
598 	 */
599 	if (!ts->tick_stopped && delta_jiffies <= 1)
600 		goto out;
601 
602 	/* Schedule the tick, if we are at least one jiffie off */
603 	if ((long)delta_jiffies >= 1) {
604 
605 		/*
606 		 * If this cpu is the one which updates jiffies, then
607 		 * give up the assignment and let it be taken by the
608 		 * cpu which runs the tick timer next, which might be
609 		 * this cpu as well. If we don't drop this here the
610 		 * jiffies might be stale and do_timer() never
611 		 * invoked. Keep track of the fact that it was the one
612 		 * which had the do_timer() duty last. If this cpu is
613 		 * the one which had the do_timer() duty last, we
614 		 * limit the sleep time to the timekeeping
615 		 * max_deferement value which we retrieved
616 		 * above. Otherwise we can sleep as long as we want.
617 		 */
618 		if (cpu == tick_do_timer_cpu) {
619 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
620 			ts->do_timer_last = 1;
621 		} else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
622 			time_delta = KTIME_MAX;
623 			ts->do_timer_last = 0;
624 		} else if (!ts->do_timer_last) {
625 			time_delta = KTIME_MAX;
626 		}
627 
628 #ifdef CONFIG_NO_HZ_FULL
629 		if (!ts->inidle) {
630 			time_delta = min(time_delta,
631 					 scheduler_tick_max_deferment());
632 		}
633 #endif
634 
635 		/*
636 		 * calculate the expiry time for the next timer wheel
637 		 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
638 		 * that there is no timer pending or at least extremely
639 		 * far into the future (12 days for HZ=1000). In this
640 		 * case we set the expiry to the end of time.
641 		 */
642 		if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
643 			/*
644 			 * Calculate the time delta for the next timer event.
645 			 * If the time delta exceeds the maximum time delta
646 			 * permitted by the current clocksource then adjust
647 			 * the time delta accordingly to ensure the
648 			 * clocksource does not wrap.
649 			 */
650 			time_delta = min_t(u64, time_delta,
651 					   tick_period.tv64 * delta_jiffies);
652 		}
653 
654 		if (time_delta < KTIME_MAX)
655 			expires = ktime_add_ns(last_update, time_delta);
656 		else
657 			expires.tv64 = KTIME_MAX;
658 
659 		/* Skip reprogram of event if its not changed */
660 		if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
661 			goto out;
662 
663 		ret = expires;
664 
665 		/*
666 		 * nohz_stop_sched_tick can be called several times before
667 		 * the nohz_restart_sched_tick is called. This happens when
668 		 * interrupts arrive which do not cause a reschedule. In the
669 		 * first call we save the current tick time, so we can restart
670 		 * the scheduler tick in nohz_restart_sched_tick.
671 		 */
672 		if (!ts->tick_stopped) {
673 			nohz_balance_enter_idle(cpu);
674 			calc_load_enter_idle();
675 
676 			ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
677 			ts->tick_stopped = 1;
678 			trace_tick_stop(1, " ");
679 		}
680 
681 		/*
682 		 * If the expiration time == KTIME_MAX, then
683 		 * in this case we simply stop the tick timer.
684 		 */
685 		 if (unlikely(expires.tv64 == KTIME_MAX)) {
686 			if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
687 				hrtimer_cancel(&ts->sched_timer);
688 			goto out;
689 		}
690 
691 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
692 			hrtimer_start(&ts->sched_timer, expires,
693 				      HRTIMER_MODE_ABS_PINNED);
694 			/* Check, if the timer was already in the past */
695 			if (hrtimer_active(&ts->sched_timer))
696 				goto out;
697 		} else if (!tick_program_event(expires, 0))
698 				goto out;
699 		/*
700 		 * We are past the event already. So we crossed a
701 		 * jiffie boundary. Update jiffies and raise the
702 		 * softirq.
703 		 */
704 		tick_do_update_jiffies64(ktime_get());
705 	}
706 	raise_softirq_irqoff(TIMER_SOFTIRQ);
707 out:
708 	ts->next_jiffies = next_jiffies;
709 	ts->last_jiffies = last_jiffies;
710 	ts->sleep_length = ktime_sub(dev->next_event, now);
711 
712 	return ret;
713 }
714 
715 static void tick_nohz_full_stop_tick(struct tick_sched *ts)
716 {
717 #ifdef CONFIG_NO_HZ_FULL
718 	int cpu = smp_processor_id();
719 
720 	if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
721 		return;
722 
723 	if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
724 		return;
725 
726 	if (!can_stop_full_tick())
727 		return;
728 
729 	tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
730 #endif
731 }
732 
733 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
734 {
735 	/*
736 	 * If this cpu is offline and it is the one which updates
737 	 * jiffies, then give up the assignment and let it be taken by
738 	 * the cpu which runs the tick timer next. If we don't drop
739 	 * this here the jiffies might be stale and do_timer() never
740 	 * invoked.
741 	 */
742 	if (unlikely(!cpu_online(cpu))) {
743 		if (cpu == tick_do_timer_cpu)
744 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
745 		return false;
746 	}
747 
748 	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
749 		ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
750 		return false;
751 	}
752 
753 	if (need_resched())
754 		return false;
755 
756 	if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
757 		static int ratelimit;
758 
759 		if (ratelimit < 10 &&
760 		    (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
761 			pr_warn("NOHZ: local_softirq_pending %02x\n",
762 				(unsigned int) local_softirq_pending());
763 			ratelimit++;
764 		}
765 		return false;
766 	}
767 
768 	if (tick_nohz_full_enabled()) {
769 		/*
770 		 * Keep the tick alive to guarantee timekeeping progression
771 		 * if there are full dynticks CPUs around
772 		 */
773 		if (tick_do_timer_cpu == cpu)
774 			return false;
775 		/*
776 		 * Boot safety: make sure the timekeeping duty has been
777 		 * assigned before entering dyntick-idle mode,
778 		 */
779 		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
780 			return false;
781 	}
782 
783 	return true;
784 }
785 
786 static void __tick_nohz_idle_enter(struct tick_sched *ts)
787 {
788 	ktime_t now, expires;
789 	int cpu = smp_processor_id();
790 
791 	now = tick_nohz_start_idle(ts);
792 
793 	if (can_stop_idle_tick(cpu, ts)) {
794 		int was_stopped = ts->tick_stopped;
795 
796 		ts->idle_calls++;
797 
798 		expires = tick_nohz_stop_sched_tick(ts, now, cpu);
799 		if (expires.tv64 > 0LL) {
800 			ts->idle_sleeps++;
801 			ts->idle_expires = expires;
802 		}
803 
804 		if (!was_stopped && ts->tick_stopped)
805 			ts->idle_jiffies = ts->last_jiffies;
806 	}
807 }
808 
809 /**
810  * tick_nohz_idle_enter - stop the idle tick from the idle task
811  *
812  * When the next event is more than a tick into the future, stop the idle tick
813  * Called when we start the idle loop.
814  *
815  * The arch is responsible of calling:
816  *
817  * - rcu_idle_enter() after its last use of RCU before the CPU is put
818  *  to sleep.
819  * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
820  */
821 void tick_nohz_idle_enter(void)
822 {
823 	struct tick_sched *ts;
824 
825 	WARN_ON_ONCE(irqs_disabled());
826 
827 	/*
828  	 * Update the idle state in the scheduler domain hierarchy
829  	 * when tick_nohz_stop_sched_tick() is called from the idle loop.
830  	 * State will be updated to busy during the first busy tick after
831  	 * exiting idle.
832  	 */
833 	set_cpu_sd_state_idle();
834 
835 	local_irq_disable();
836 
837 	ts = this_cpu_ptr(&tick_cpu_sched);
838 	ts->inidle = 1;
839 	__tick_nohz_idle_enter(ts);
840 
841 	local_irq_enable();
842 }
843 
844 /**
845  * tick_nohz_irq_exit - update next tick event from interrupt exit
846  *
847  * When an interrupt fires while we are idle and it doesn't cause
848  * a reschedule, it may still add, modify or delete a timer, enqueue
849  * an RCU callback, etc...
850  * So we need to re-calculate and reprogram the next tick event.
851  */
852 void tick_nohz_irq_exit(void)
853 {
854 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
855 
856 	if (ts->inidle)
857 		__tick_nohz_idle_enter(ts);
858 	else
859 		tick_nohz_full_stop_tick(ts);
860 }
861 
862 /**
863  * tick_nohz_get_sleep_length - return the length of the current sleep
864  *
865  * Called from power state control code with interrupts disabled
866  */
867 ktime_t tick_nohz_get_sleep_length(void)
868 {
869 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
870 
871 	return ts->sleep_length;
872 }
873 
874 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
875 {
876 	hrtimer_cancel(&ts->sched_timer);
877 	hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
878 
879 	while (1) {
880 		/* Forward the time to expire in the future */
881 		hrtimer_forward(&ts->sched_timer, now, tick_period);
882 
883 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
884 			hrtimer_start_expires(&ts->sched_timer,
885 					      HRTIMER_MODE_ABS_PINNED);
886 			/* Check, if the timer was already in the past */
887 			if (hrtimer_active(&ts->sched_timer))
888 				break;
889 		} else {
890 			if (!tick_program_event(
891 				hrtimer_get_expires(&ts->sched_timer), 0))
892 				break;
893 		}
894 		/* Reread time and update jiffies */
895 		now = ktime_get();
896 		tick_do_update_jiffies64(now);
897 	}
898 }
899 
900 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
901 {
902 	/* Update jiffies first */
903 	tick_do_update_jiffies64(now);
904 	update_cpu_load_nohz();
905 
906 	calc_load_exit_idle();
907 	touch_softlockup_watchdog();
908 	/*
909 	 * Cancel the scheduled timer and restore the tick
910 	 */
911 	ts->tick_stopped  = 0;
912 	ts->idle_exittime = now;
913 
914 	tick_nohz_restart(ts, now);
915 }
916 
917 static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
918 {
919 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
920 	unsigned long ticks;
921 
922 	if (vtime_accounting_enabled())
923 		return;
924 	/*
925 	 * We stopped the tick in idle. Update process times would miss the
926 	 * time we slept as update_process_times does only a 1 tick
927 	 * accounting. Enforce that this is accounted to idle !
928 	 */
929 	ticks = jiffies - ts->idle_jiffies;
930 	/*
931 	 * We might be one off. Do not randomly account a huge number of ticks!
932 	 */
933 	if (ticks && ticks < LONG_MAX)
934 		account_idle_ticks(ticks);
935 #endif
936 }
937 
938 /**
939  * tick_nohz_idle_exit - restart the idle tick from the idle task
940  *
941  * Restart the idle tick when the CPU is woken up from idle
942  * This also exit the RCU extended quiescent state. The CPU
943  * can use RCU again after this function is called.
944  */
945 void tick_nohz_idle_exit(void)
946 {
947 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
948 	ktime_t now;
949 
950 	local_irq_disable();
951 
952 	WARN_ON_ONCE(!ts->inidle);
953 
954 	ts->inidle = 0;
955 
956 	if (ts->idle_active || ts->tick_stopped)
957 		now = ktime_get();
958 
959 	if (ts->idle_active)
960 		tick_nohz_stop_idle(ts, now);
961 
962 	if (ts->tick_stopped) {
963 		tick_nohz_restart_sched_tick(ts, now);
964 		tick_nohz_account_idle_ticks(ts);
965 	}
966 
967 	local_irq_enable();
968 }
969 
970 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
971 {
972 	hrtimer_forward(&ts->sched_timer, now, tick_period);
973 	return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
974 }
975 
976 /*
977  * The nohz low res interrupt handler
978  */
979 static void tick_nohz_handler(struct clock_event_device *dev)
980 {
981 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
982 	struct pt_regs *regs = get_irq_regs();
983 	ktime_t now = ktime_get();
984 
985 	dev->next_event.tv64 = KTIME_MAX;
986 
987 	tick_sched_do_timer(now);
988 	tick_sched_handle(ts, regs);
989 
990 	/* No need to reprogram if we are running tickless  */
991 	if (unlikely(ts->tick_stopped))
992 		return;
993 
994 	while (tick_nohz_reprogram(ts, now)) {
995 		now = ktime_get();
996 		tick_do_update_jiffies64(now);
997 	}
998 }
999 
1000 /**
1001  * tick_nohz_switch_to_nohz - switch to nohz mode
1002  */
1003 static void tick_nohz_switch_to_nohz(void)
1004 {
1005 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1006 	ktime_t next;
1007 
1008 	if (!tick_nohz_enabled)
1009 		return;
1010 
1011 	local_irq_disable();
1012 	if (tick_switch_to_oneshot(tick_nohz_handler)) {
1013 		local_irq_enable();
1014 		return;
1015 	}
1016 	tick_nohz_active = 1;
1017 	ts->nohz_mode = NOHZ_MODE_LOWRES;
1018 
1019 	/*
1020 	 * Recycle the hrtimer in ts, so we can share the
1021 	 * hrtimer_forward with the highres code.
1022 	 */
1023 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1024 	/* Get the next period */
1025 	next = tick_init_jiffy_update();
1026 
1027 	for (;;) {
1028 		hrtimer_set_expires(&ts->sched_timer, next);
1029 		if (!tick_program_event(next, 0))
1030 			break;
1031 		next = ktime_add(next, tick_period);
1032 	}
1033 	local_irq_enable();
1034 }
1035 
1036 /*
1037  * When NOHZ is enabled and the tick is stopped, we need to kick the
1038  * tick timer from irq_enter() so that the jiffies update is kept
1039  * alive during long running softirqs. That's ugly as hell, but
1040  * correctness is key even if we need to fix the offending softirq in
1041  * the first place.
1042  *
1043  * Note, this is different to tick_nohz_restart. We just kick the
1044  * timer and do not touch the other magic bits which need to be done
1045  * when idle is left.
1046  */
1047 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1048 {
1049 #if 0
1050 	/* Switch back to 2.6.27 behaviour */
1051 	ktime_t delta;
1052 
1053 	/*
1054 	 * Do not touch the tick device, when the next expiry is either
1055 	 * already reached or less/equal than the tick period.
1056 	 */
1057 	delta =	ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1058 	if (delta.tv64 <= tick_period.tv64)
1059 		return;
1060 
1061 	tick_nohz_restart(ts, now);
1062 #endif
1063 }
1064 
1065 static inline void tick_nohz_irq_enter(void)
1066 {
1067 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1068 	ktime_t now;
1069 
1070 	if (!ts->idle_active && !ts->tick_stopped)
1071 		return;
1072 	now = ktime_get();
1073 	if (ts->idle_active)
1074 		tick_nohz_stop_idle(ts, now);
1075 	if (ts->tick_stopped) {
1076 		tick_nohz_update_jiffies(now);
1077 		tick_nohz_kick_tick(ts, now);
1078 	}
1079 }
1080 
1081 #else
1082 
1083 static inline void tick_nohz_switch_to_nohz(void) { }
1084 static inline void tick_nohz_irq_enter(void) { }
1085 
1086 #endif /* CONFIG_NO_HZ_COMMON */
1087 
1088 /*
1089  * Called from irq_enter to notify about the possible interruption of idle()
1090  */
1091 void tick_irq_enter(void)
1092 {
1093 	tick_check_oneshot_broadcast_this_cpu();
1094 	tick_nohz_irq_enter();
1095 }
1096 
1097 /*
1098  * High resolution timer specific code
1099  */
1100 #ifdef CONFIG_HIGH_RES_TIMERS
1101 /*
1102  * We rearm the timer until we get disabled by the idle code.
1103  * Called with interrupts disabled.
1104  */
1105 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1106 {
1107 	struct tick_sched *ts =
1108 		container_of(timer, struct tick_sched, sched_timer);
1109 	struct pt_regs *regs = get_irq_regs();
1110 	ktime_t now = ktime_get();
1111 
1112 	tick_sched_do_timer(now);
1113 
1114 	/*
1115 	 * Do not call, when we are not in irq context and have
1116 	 * no valid regs pointer
1117 	 */
1118 	if (regs)
1119 		tick_sched_handle(ts, regs);
1120 
1121 	/* No need to reprogram if we are in idle or full dynticks mode */
1122 	if (unlikely(ts->tick_stopped))
1123 		return HRTIMER_NORESTART;
1124 
1125 	hrtimer_forward(timer, now, tick_period);
1126 
1127 	return HRTIMER_RESTART;
1128 }
1129 
1130 static int sched_skew_tick;
1131 
1132 static int __init skew_tick(char *str)
1133 {
1134 	get_option(&str, &sched_skew_tick);
1135 
1136 	return 0;
1137 }
1138 early_param("skew_tick", skew_tick);
1139 
1140 /**
1141  * tick_setup_sched_timer - setup the tick emulation timer
1142  */
1143 void tick_setup_sched_timer(void)
1144 {
1145 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1146 	ktime_t now = ktime_get();
1147 
1148 	/*
1149 	 * Emulate tick processing via per-CPU hrtimers:
1150 	 */
1151 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1152 	ts->sched_timer.function = tick_sched_timer;
1153 
1154 	/* Get the next period (per cpu) */
1155 	hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1156 
1157 	/* Offset the tick to avert jiffies_lock contention. */
1158 	if (sched_skew_tick) {
1159 		u64 offset = ktime_to_ns(tick_period) >> 1;
1160 		do_div(offset, num_possible_cpus());
1161 		offset *= smp_processor_id();
1162 		hrtimer_add_expires_ns(&ts->sched_timer, offset);
1163 	}
1164 
1165 	for (;;) {
1166 		hrtimer_forward(&ts->sched_timer, now, tick_period);
1167 		hrtimer_start_expires(&ts->sched_timer,
1168 				      HRTIMER_MODE_ABS_PINNED);
1169 		/* Check, if the timer was already in the past */
1170 		if (hrtimer_active(&ts->sched_timer))
1171 			break;
1172 		now = ktime_get();
1173 	}
1174 
1175 #ifdef CONFIG_NO_HZ_COMMON
1176 	if (tick_nohz_enabled) {
1177 		ts->nohz_mode = NOHZ_MODE_HIGHRES;
1178 		tick_nohz_active = 1;
1179 	}
1180 #endif
1181 }
1182 #endif /* HIGH_RES_TIMERS */
1183 
1184 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1185 void tick_cancel_sched_timer(int cpu)
1186 {
1187 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1188 
1189 # ifdef CONFIG_HIGH_RES_TIMERS
1190 	if (ts->sched_timer.base)
1191 		hrtimer_cancel(&ts->sched_timer);
1192 # endif
1193 
1194 	memset(ts, 0, sizeof(*ts));
1195 }
1196 #endif
1197 
1198 /**
1199  * Async notification about clocksource changes
1200  */
1201 void tick_clock_notify(void)
1202 {
1203 	int cpu;
1204 
1205 	for_each_possible_cpu(cpu)
1206 		set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1207 }
1208 
1209 /*
1210  * Async notification about clock event changes
1211  */
1212 void tick_oneshot_notify(void)
1213 {
1214 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1215 
1216 	set_bit(0, &ts->check_clocks);
1217 }
1218 
1219 /**
1220  * Check, if a change happened, which makes oneshot possible.
1221  *
1222  * Called cyclic from the hrtimer softirq (driven by the timer
1223  * softirq) allow_nohz signals, that we can switch into low-res nohz
1224  * mode, because high resolution timers are disabled (either compile
1225  * or runtime).
1226  */
1227 int tick_check_oneshot_change(int allow_nohz)
1228 {
1229 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1230 
1231 	if (!test_and_clear_bit(0, &ts->check_clocks))
1232 		return 0;
1233 
1234 	if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1235 		return 0;
1236 
1237 	if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1238 		return 0;
1239 
1240 	if (!allow_nohz)
1241 		return 1;
1242 
1243 	tick_nohz_switch_to_nohz();
1244 	return 0;
1245 }
1246