xref: /openbmc/linux/kernel/time/tick-sched.c (revision 62c1256d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4  *  Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5  *  Copyright(C) 2006-2007  Timesys Corp., Thomas Gleixner
6  *
7  *  No idle tick implementation for low and high resolution timers
8  *
9  *  Started by: Thomas Gleixner and Ingo Molnar
10  */
11 #include <linux/cpu.h>
12 #include <linux/err.h>
13 #include <linux/hrtimer.h>
14 #include <linux/interrupt.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/percpu.h>
17 #include <linux/nmi.h>
18 #include <linux/profile.h>
19 #include <linux/sched/signal.h>
20 #include <linux/sched/clock.h>
21 #include <linux/sched/stat.h>
22 #include <linux/sched/nohz.h>
23 #include <linux/sched/loadavg.h>
24 #include <linux/module.h>
25 #include <linux/irq_work.h>
26 #include <linux/posix-timers.h>
27 #include <linux/context_tracking.h>
28 #include <linux/mm.h>
29 
30 #include <asm/irq_regs.h>
31 
32 #include "tick-internal.h"
33 
34 #include <trace/events/timer.h>
35 
36 /*
37  * Per-CPU nohz control structure
38  */
39 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40 
41 struct tick_sched *tick_get_tick_sched(int cpu)
42 {
43 	return &per_cpu(tick_cpu_sched, cpu);
44 }
45 
46 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47 /*
48  * The time, when the last jiffy update happened. Write access must hold
49  * jiffies_lock and jiffies_seq. tick_nohz_next_event() needs to get a
50  * consistent view of jiffies and last_jiffies_update.
51  */
52 static ktime_t last_jiffies_update;
53 
54 /*
55  * Must be called with interrupts disabled !
56  */
57 static void tick_do_update_jiffies64(ktime_t now)
58 {
59 	unsigned long ticks = 1;
60 	ktime_t delta, nextp;
61 
62 	/*
63 	 * 64bit can do a quick check without holding jiffies lock and
64 	 * without looking at the sequence count. The smp_load_acquire()
65 	 * pairs with the update done later in this function.
66 	 *
67 	 * 32bit cannot do that because the store of tick_next_period
68 	 * consists of two 32bit stores and the first store could move it
69 	 * to a random point in the future.
70 	 */
71 	if (IS_ENABLED(CONFIG_64BIT)) {
72 		if (ktime_before(now, smp_load_acquire(&tick_next_period)))
73 			return;
74 	} else {
75 		unsigned int seq;
76 
77 		/*
78 		 * Avoid contention on jiffies_lock and protect the quick
79 		 * check with the sequence count.
80 		 */
81 		do {
82 			seq = read_seqcount_begin(&jiffies_seq);
83 			nextp = tick_next_period;
84 		} while (read_seqcount_retry(&jiffies_seq, seq));
85 
86 		if (ktime_before(now, nextp))
87 			return;
88 	}
89 
90 	/* Quick check failed, i.e. update is required. */
91 	raw_spin_lock(&jiffies_lock);
92 	/*
93 	 * Reevaluate with the lock held. Another CPU might have done the
94 	 * update already.
95 	 */
96 	if (ktime_before(now, tick_next_period)) {
97 		raw_spin_unlock(&jiffies_lock);
98 		return;
99 	}
100 
101 	write_seqcount_begin(&jiffies_seq);
102 
103 	delta = ktime_sub(now, tick_next_period);
104 	if (unlikely(delta >= TICK_NSEC)) {
105 		/* Slow path for long idle sleep times */
106 		s64 incr = TICK_NSEC;
107 
108 		ticks += ktime_divns(delta, incr);
109 
110 		last_jiffies_update = ktime_add_ns(last_jiffies_update,
111 						   incr * ticks);
112 	} else {
113 		last_jiffies_update = ktime_add_ns(last_jiffies_update,
114 						   TICK_NSEC);
115 	}
116 
117 	/* Advance jiffies to complete the jiffies_seq protected job */
118 	jiffies_64 += ticks;
119 
120 	/*
121 	 * Keep the tick_next_period variable up to date.
122 	 */
123 	nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
124 
125 	if (IS_ENABLED(CONFIG_64BIT)) {
126 		/*
127 		 * Pairs with smp_load_acquire() in the lockless quick
128 		 * check above and ensures that the update to jiffies_64 is
129 		 * not reordered vs. the store to tick_next_period, neither
130 		 * by the compiler nor by the CPU.
131 		 */
132 		smp_store_release(&tick_next_period, nextp);
133 	} else {
134 		/*
135 		 * A plain store is good enough on 32bit as the quick check
136 		 * above is protected by the sequence count.
137 		 */
138 		tick_next_period = nextp;
139 	}
140 
141 	/*
142 	 * Release the sequence count. calc_global_load() below is not
143 	 * protected by it, but jiffies_lock needs to be held to prevent
144 	 * concurrent invocations.
145 	 */
146 	write_seqcount_end(&jiffies_seq);
147 
148 	calc_global_load();
149 
150 	raw_spin_unlock(&jiffies_lock);
151 	update_wall_time();
152 }
153 
154 /*
155  * Initialize and return retrieve the jiffies update.
156  */
157 static ktime_t tick_init_jiffy_update(void)
158 {
159 	ktime_t period;
160 
161 	raw_spin_lock(&jiffies_lock);
162 	write_seqcount_begin(&jiffies_seq);
163 	/* Did we start the jiffies update yet ? */
164 	if (last_jiffies_update == 0)
165 		last_jiffies_update = tick_next_period;
166 	period = last_jiffies_update;
167 	write_seqcount_end(&jiffies_seq);
168 	raw_spin_unlock(&jiffies_lock);
169 	return period;
170 }
171 
172 #define MAX_STALLED_JIFFIES 5
173 
174 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
175 {
176 	int cpu = smp_processor_id();
177 
178 #ifdef CONFIG_NO_HZ_COMMON
179 	/*
180 	 * Check if the do_timer duty was dropped. We don't care about
181 	 * concurrency: This happens only when the CPU in charge went
182 	 * into a long sleep. If two CPUs happen to assign themselves to
183 	 * this duty, then the jiffies update is still serialized by
184 	 * jiffies_lock.
185 	 *
186 	 * If nohz_full is enabled, this should not happen because the
187 	 * tick_do_timer_cpu never relinquishes.
188 	 */
189 	if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
190 #ifdef CONFIG_NO_HZ_FULL
191 		WARN_ON(tick_nohz_full_running);
192 #endif
193 		tick_do_timer_cpu = cpu;
194 	}
195 #endif
196 
197 	/* Check, if the jiffies need an update */
198 	if (tick_do_timer_cpu == cpu)
199 		tick_do_update_jiffies64(now);
200 
201 	/*
202 	 * If jiffies update stalled for too long (timekeeper in stop_machine()
203 	 * or VMEXIT'ed for several msecs), force an update.
204 	 */
205 	if (ts->last_tick_jiffies != jiffies) {
206 		ts->stalled_jiffies = 0;
207 		ts->last_tick_jiffies = READ_ONCE(jiffies);
208 	} else {
209 		if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
210 			tick_do_update_jiffies64(now);
211 			ts->stalled_jiffies = 0;
212 			ts->last_tick_jiffies = READ_ONCE(jiffies);
213 		}
214 	}
215 
216 	if (ts->inidle)
217 		ts->got_idle_tick = 1;
218 }
219 
220 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
221 {
222 #ifdef CONFIG_NO_HZ_COMMON
223 	/*
224 	 * When we are idle and the tick is stopped, we have to touch
225 	 * the watchdog as we might not schedule for a really long
226 	 * time. This happens on complete idle SMP systems while
227 	 * waiting on the login prompt. We also increment the "start of
228 	 * idle" jiffy stamp so the idle accounting adjustment we do
229 	 * when we go busy again does not account too much ticks.
230 	 */
231 	if (ts->tick_stopped) {
232 		touch_softlockup_watchdog_sched();
233 		if (is_idle_task(current))
234 			ts->idle_jiffies++;
235 		/*
236 		 * In case the current tick fired too early past its expected
237 		 * expiration, make sure we don't bypass the next clock reprogramming
238 		 * to the same deadline.
239 		 */
240 		ts->next_tick = 0;
241 	}
242 #endif
243 	update_process_times(user_mode(regs));
244 	profile_tick(CPU_PROFILING);
245 }
246 #endif
247 
248 #ifdef CONFIG_NO_HZ_FULL
249 cpumask_var_t tick_nohz_full_mask;
250 EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
251 bool tick_nohz_full_running;
252 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
253 static atomic_t tick_dep_mask;
254 
255 static bool check_tick_dependency(atomic_t *dep)
256 {
257 	int val = atomic_read(dep);
258 
259 	if (val & TICK_DEP_MASK_POSIX_TIMER) {
260 		trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
261 		return true;
262 	}
263 
264 	if (val & TICK_DEP_MASK_PERF_EVENTS) {
265 		trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
266 		return true;
267 	}
268 
269 	if (val & TICK_DEP_MASK_SCHED) {
270 		trace_tick_stop(0, TICK_DEP_MASK_SCHED);
271 		return true;
272 	}
273 
274 	if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
275 		trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
276 		return true;
277 	}
278 
279 	if (val & TICK_DEP_MASK_RCU) {
280 		trace_tick_stop(0, TICK_DEP_MASK_RCU);
281 		return true;
282 	}
283 
284 	return false;
285 }
286 
287 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
288 {
289 	lockdep_assert_irqs_disabled();
290 
291 	if (unlikely(!cpu_online(cpu)))
292 		return false;
293 
294 	if (check_tick_dependency(&tick_dep_mask))
295 		return false;
296 
297 	if (check_tick_dependency(&ts->tick_dep_mask))
298 		return false;
299 
300 	if (check_tick_dependency(&current->tick_dep_mask))
301 		return false;
302 
303 	if (check_tick_dependency(&current->signal->tick_dep_mask))
304 		return false;
305 
306 	return true;
307 }
308 
309 static void nohz_full_kick_func(struct irq_work *work)
310 {
311 	/* Empty, the tick restart happens on tick_nohz_irq_exit() */
312 }
313 
314 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
315 	IRQ_WORK_INIT_HARD(nohz_full_kick_func);
316 
317 /*
318  * Kick this CPU if it's full dynticks in order to force it to
319  * re-evaluate its dependency on the tick and restart it if necessary.
320  * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
321  * is NMI safe.
322  */
323 static void tick_nohz_full_kick(void)
324 {
325 	if (!tick_nohz_full_cpu(smp_processor_id()))
326 		return;
327 
328 	irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
329 }
330 
331 /*
332  * Kick the CPU if it's full dynticks in order to force it to
333  * re-evaluate its dependency on the tick and restart it if necessary.
334  */
335 void tick_nohz_full_kick_cpu(int cpu)
336 {
337 	if (!tick_nohz_full_cpu(cpu))
338 		return;
339 
340 	irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
341 }
342 
343 static void tick_nohz_kick_task(struct task_struct *tsk)
344 {
345 	int cpu;
346 
347 	/*
348 	 * If the task is not running, run_posix_cpu_timers()
349 	 * has nothing to elapse, IPI can then be spared.
350 	 *
351 	 * activate_task()                      STORE p->tick_dep_mask
352 	 *   STORE p->on_rq
353 	 * __schedule() (switch to task 'p')    smp_mb() (atomic_fetch_or())
354 	 *   LOCK rq->lock                      LOAD p->on_rq
355 	 *   smp_mb__after_spin_lock()
356 	 *   tick_nohz_task_switch()
357 	 *     LOAD p->tick_dep_mask
358 	 */
359 	if (!sched_task_on_rq(tsk))
360 		return;
361 
362 	/*
363 	 * If the task concurrently migrates to another CPU,
364 	 * we guarantee it sees the new tick dependency upon
365 	 * schedule.
366 	 *
367 	 * set_task_cpu(p, cpu);
368 	 *   STORE p->cpu = @cpu
369 	 * __schedule() (switch to task 'p')
370 	 *   LOCK rq->lock
371 	 *   smp_mb__after_spin_lock()          STORE p->tick_dep_mask
372 	 *   tick_nohz_task_switch()            smp_mb() (atomic_fetch_or())
373 	 *      LOAD p->tick_dep_mask           LOAD p->cpu
374 	 */
375 	cpu = task_cpu(tsk);
376 
377 	preempt_disable();
378 	if (cpu_online(cpu))
379 		tick_nohz_full_kick_cpu(cpu);
380 	preempt_enable();
381 }
382 
383 /*
384  * Kick all full dynticks CPUs in order to force these to re-evaluate
385  * their dependency on the tick and restart it if necessary.
386  */
387 static void tick_nohz_full_kick_all(void)
388 {
389 	int cpu;
390 
391 	if (!tick_nohz_full_running)
392 		return;
393 
394 	preempt_disable();
395 	for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
396 		tick_nohz_full_kick_cpu(cpu);
397 	preempt_enable();
398 }
399 
400 static void tick_nohz_dep_set_all(atomic_t *dep,
401 				  enum tick_dep_bits bit)
402 {
403 	int prev;
404 
405 	prev = atomic_fetch_or(BIT(bit), dep);
406 	if (!prev)
407 		tick_nohz_full_kick_all();
408 }
409 
410 /*
411  * Set a global tick dependency. Used by perf events that rely on freq and
412  * by unstable clock.
413  */
414 void tick_nohz_dep_set(enum tick_dep_bits bit)
415 {
416 	tick_nohz_dep_set_all(&tick_dep_mask, bit);
417 }
418 
419 void tick_nohz_dep_clear(enum tick_dep_bits bit)
420 {
421 	atomic_andnot(BIT(bit), &tick_dep_mask);
422 }
423 
424 /*
425  * Set per-CPU tick dependency. Used by scheduler and perf events in order to
426  * manage events throttling.
427  */
428 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
429 {
430 	int prev;
431 	struct tick_sched *ts;
432 
433 	ts = per_cpu_ptr(&tick_cpu_sched, cpu);
434 
435 	prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
436 	if (!prev) {
437 		preempt_disable();
438 		/* Perf needs local kick that is NMI safe */
439 		if (cpu == smp_processor_id()) {
440 			tick_nohz_full_kick();
441 		} else {
442 			/* Remote irq work not NMI-safe */
443 			if (!WARN_ON_ONCE(in_nmi()))
444 				tick_nohz_full_kick_cpu(cpu);
445 		}
446 		preempt_enable();
447 	}
448 }
449 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
450 
451 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
452 {
453 	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
454 
455 	atomic_andnot(BIT(bit), &ts->tick_dep_mask);
456 }
457 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
458 
459 /*
460  * Set a per-task tick dependency. RCU need this. Also posix CPU timers
461  * in order to elapse per task timers.
462  */
463 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
464 {
465 	if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
466 		tick_nohz_kick_task(tsk);
467 }
468 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
469 
470 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
471 {
472 	atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
473 }
474 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
475 
476 /*
477  * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
478  * per process timers.
479  */
480 void tick_nohz_dep_set_signal(struct task_struct *tsk,
481 			      enum tick_dep_bits bit)
482 {
483 	int prev;
484 	struct signal_struct *sig = tsk->signal;
485 
486 	prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
487 	if (!prev) {
488 		struct task_struct *t;
489 
490 		lockdep_assert_held(&tsk->sighand->siglock);
491 		__for_each_thread(sig, t)
492 			tick_nohz_kick_task(t);
493 	}
494 }
495 
496 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
497 {
498 	atomic_andnot(BIT(bit), &sig->tick_dep_mask);
499 }
500 
501 /*
502  * Re-evaluate the need for the tick as we switch the current task.
503  * It might need the tick due to per task/process properties:
504  * perf events, posix CPU timers, ...
505  */
506 void __tick_nohz_task_switch(void)
507 {
508 	struct tick_sched *ts;
509 
510 	if (!tick_nohz_full_cpu(smp_processor_id()))
511 		return;
512 
513 	ts = this_cpu_ptr(&tick_cpu_sched);
514 
515 	if (ts->tick_stopped) {
516 		if (atomic_read(&current->tick_dep_mask) ||
517 		    atomic_read(&current->signal->tick_dep_mask))
518 			tick_nohz_full_kick();
519 	}
520 }
521 
522 /* Get the boot-time nohz CPU list from the kernel parameters. */
523 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
524 {
525 	alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
526 	cpumask_copy(tick_nohz_full_mask, cpumask);
527 	tick_nohz_full_running = true;
528 }
529 EXPORT_SYMBOL_GPL(tick_nohz_full_setup);
530 
531 static int tick_nohz_cpu_down(unsigned int cpu)
532 {
533 	/*
534 	 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
535 	 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
536 	 * CPUs. It must remain online when nohz full is enabled.
537 	 */
538 	if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
539 		return -EBUSY;
540 	return 0;
541 }
542 
543 void __init tick_nohz_init(void)
544 {
545 	int cpu, ret;
546 
547 	if (!tick_nohz_full_running)
548 		return;
549 
550 	/*
551 	 * Full dynticks uses irq work to drive the tick rescheduling on safe
552 	 * locking contexts. But then we need irq work to raise its own
553 	 * interrupts to avoid circular dependency on the tick
554 	 */
555 	if (!arch_irq_work_has_interrupt()) {
556 		pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
557 		cpumask_clear(tick_nohz_full_mask);
558 		tick_nohz_full_running = false;
559 		return;
560 	}
561 
562 	if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
563 			!IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
564 		cpu = smp_processor_id();
565 
566 		if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
567 			pr_warn("NO_HZ: Clearing %d from nohz_full range "
568 				"for timekeeping\n", cpu);
569 			cpumask_clear_cpu(cpu, tick_nohz_full_mask);
570 		}
571 	}
572 
573 	for_each_cpu(cpu, tick_nohz_full_mask)
574 		context_tracking_cpu_set(cpu);
575 
576 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
577 					"kernel/nohz:predown", NULL,
578 					tick_nohz_cpu_down);
579 	WARN_ON(ret < 0);
580 	pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
581 		cpumask_pr_args(tick_nohz_full_mask));
582 }
583 #endif
584 
585 /*
586  * NOHZ - aka dynamic tick functionality
587  */
588 #ifdef CONFIG_NO_HZ_COMMON
589 /*
590  * NO HZ enabled ?
591  */
592 bool tick_nohz_enabled __read_mostly  = true;
593 unsigned long tick_nohz_active  __read_mostly;
594 /*
595  * Enable / Disable tickless mode
596  */
597 static int __init setup_tick_nohz(char *str)
598 {
599 	return (kstrtobool(str, &tick_nohz_enabled) == 0);
600 }
601 
602 __setup("nohz=", setup_tick_nohz);
603 
604 bool tick_nohz_tick_stopped(void)
605 {
606 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
607 
608 	return ts->tick_stopped;
609 }
610 
611 bool tick_nohz_tick_stopped_cpu(int cpu)
612 {
613 	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
614 
615 	return ts->tick_stopped;
616 }
617 
618 /**
619  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
620  *
621  * Called from interrupt entry when the CPU was idle
622  *
623  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
624  * must be updated. Otherwise an interrupt handler could use a stale jiffy
625  * value. We do this unconditionally on any CPU, as we don't know whether the
626  * CPU, which has the update task assigned is in a long sleep.
627  */
628 static void tick_nohz_update_jiffies(ktime_t now)
629 {
630 	unsigned long flags;
631 
632 	__this_cpu_write(tick_cpu_sched.idle_waketime, now);
633 
634 	local_irq_save(flags);
635 	tick_do_update_jiffies64(now);
636 	local_irq_restore(flags);
637 
638 	touch_softlockup_watchdog_sched();
639 }
640 
641 /*
642  * Updates the per-CPU time idle statistics counters
643  */
644 static void
645 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
646 {
647 	ktime_t delta;
648 
649 	if (ts->idle_active) {
650 		delta = ktime_sub(now, ts->idle_entrytime);
651 		if (nr_iowait_cpu(cpu) > 0)
652 			ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
653 		else
654 			ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
655 		ts->idle_entrytime = now;
656 	}
657 
658 	if (last_update_time)
659 		*last_update_time = ktime_to_us(now);
660 
661 }
662 
663 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
664 {
665 	update_ts_time_stats(smp_processor_id(), ts, now, NULL);
666 	ts->idle_active = 0;
667 
668 	sched_clock_idle_wakeup_event();
669 }
670 
671 static void tick_nohz_start_idle(struct tick_sched *ts)
672 {
673 	ts->idle_entrytime = ktime_get();
674 	ts->idle_active = 1;
675 	sched_clock_idle_sleep_event();
676 }
677 
678 /**
679  * get_cpu_idle_time_us - get the total idle time of a CPU
680  * @cpu: CPU number to query
681  * @last_update_time: variable to store update time in. Do not update
682  * counters if NULL.
683  *
684  * Return the cumulative idle time (since boot) for a given
685  * CPU, in microseconds.
686  *
687  * This time is measured via accounting rather than sampling,
688  * and is as accurate as ktime_get() is.
689  *
690  * This function returns -1 if NOHZ is not enabled.
691  */
692 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
693 {
694 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
695 	ktime_t now, idle;
696 
697 	if (!tick_nohz_active)
698 		return -1;
699 
700 	now = ktime_get();
701 	if (last_update_time) {
702 		update_ts_time_stats(cpu, ts, now, last_update_time);
703 		idle = ts->idle_sleeptime;
704 	} else {
705 		if (ts->idle_active && !nr_iowait_cpu(cpu)) {
706 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
707 
708 			idle = ktime_add(ts->idle_sleeptime, delta);
709 		} else {
710 			idle = ts->idle_sleeptime;
711 		}
712 	}
713 
714 	return ktime_to_us(idle);
715 
716 }
717 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
718 
719 /**
720  * get_cpu_iowait_time_us - get the total iowait time of a CPU
721  * @cpu: CPU number to query
722  * @last_update_time: variable to store update time in. Do not update
723  * counters if NULL.
724  *
725  * Return the cumulative iowait time (since boot) for a given
726  * CPU, in microseconds.
727  *
728  * This time is measured via accounting rather than sampling,
729  * and is as accurate as ktime_get() is.
730  *
731  * This function returns -1 if NOHZ is not enabled.
732  */
733 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
734 {
735 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
736 	ktime_t now, iowait;
737 
738 	if (!tick_nohz_active)
739 		return -1;
740 
741 	now = ktime_get();
742 	if (last_update_time) {
743 		update_ts_time_stats(cpu, ts, now, last_update_time);
744 		iowait = ts->iowait_sleeptime;
745 	} else {
746 		if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
747 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
748 
749 			iowait = ktime_add(ts->iowait_sleeptime, delta);
750 		} else {
751 			iowait = ts->iowait_sleeptime;
752 		}
753 	}
754 
755 	return ktime_to_us(iowait);
756 }
757 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
758 
759 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
760 {
761 	hrtimer_cancel(&ts->sched_timer);
762 	hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
763 
764 	/* Forward the time to expire in the future */
765 	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
766 
767 	if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
768 		hrtimer_start_expires(&ts->sched_timer,
769 				      HRTIMER_MODE_ABS_PINNED_HARD);
770 	} else {
771 		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
772 	}
773 
774 	/*
775 	 * Reset to make sure next tick stop doesn't get fooled by past
776 	 * cached clock deadline.
777 	 */
778 	ts->next_tick = 0;
779 }
780 
781 static inline bool local_timer_softirq_pending(void)
782 {
783 	return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
784 }
785 
786 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
787 {
788 	u64 basemono, next_tick, delta, expires;
789 	unsigned long basejiff;
790 	unsigned int seq;
791 
792 	/* Read jiffies and the time when jiffies were updated last */
793 	do {
794 		seq = read_seqcount_begin(&jiffies_seq);
795 		basemono = last_jiffies_update;
796 		basejiff = jiffies;
797 	} while (read_seqcount_retry(&jiffies_seq, seq));
798 	ts->last_jiffies = basejiff;
799 	ts->timer_expires_base = basemono;
800 
801 	/*
802 	 * Keep the periodic tick, when RCU, architecture or irq_work
803 	 * requests it.
804 	 * Aside of that check whether the local timer softirq is
805 	 * pending. If so its a bad idea to call get_next_timer_interrupt()
806 	 * because there is an already expired timer, so it will request
807 	 * immediate expiry, which rearms the hardware timer with a
808 	 * minimal delta which brings us back to this place
809 	 * immediately. Lather, rinse and repeat...
810 	 */
811 	if (rcu_needs_cpu() || arch_needs_cpu() ||
812 	    irq_work_needs_cpu() || local_timer_softirq_pending()) {
813 		next_tick = basemono + TICK_NSEC;
814 	} else {
815 		/*
816 		 * Get the next pending timer. If high resolution
817 		 * timers are enabled this only takes the timer wheel
818 		 * timers into account. If high resolution timers are
819 		 * disabled this also looks at the next expiring
820 		 * hrtimer.
821 		 */
822 		next_tick = get_next_timer_interrupt(basejiff, basemono);
823 		ts->next_timer = next_tick;
824 	}
825 
826 	/*
827 	 * If the tick is due in the next period, keep it ticking or
828 	 * force prod the timer.
829 	 */
830 	delta = next_tick - basemono;
831 	if (delta <= (u64)TICK_NSEC) {
832 		/*
833 		 * Tell the timer code that the base is not idle, i.e. undo
834 		 * the effect of get_next_timer_interrupt():
835 		 */
836 		timer_clear_idle();
837 		/*
838 		 * We've not stopped the tick yet, and there's a timer in the
839 		 * next period, so no point in stopping it either, bail.
840 		 */
841 		if (!ts->tick_stopped) {
842 			ts->timer_expires = 0;
843 			goto out;
844 		}
845 	}
846 
847 	/*
848 	 * If this CPU is the one which had the do_timer() duty last, we limit
849 	 * the sleep time to the timekeeping max_deferment value.
850 	 * Otherwise we can sleep as long as we want.
851 	 */
852 	delta = timekeeping_max_deferment();
853 	if (cpu != tick_do_timer_cpu &&
854 	    (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
855 		delta = KTIME_MAX;
856 
857 	/* Calculate the next expiry time */
858 	if (delta < (KTIME_MAX - basemono))
859 		expires = basemono + delta;
860 	else
861 		expires = KTIME_MAX;
862 
863 	ts->timer_expires = min_t(u64, expires, next_tick);
864 
865 out:
866 	return ts->timer_expires;
867 }
868 
869 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
870 {
871 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
872 	u64 basemono = ts->timer_expires_base;
873 	u64 expires = ts->timer_expires;
874 	ktime_t tick = expires;
875 
876 	/* Make sure we won't be trying to stop it twice in a row. */
877 	ts->timer_expires_base = 0;
878 
879 	/*
880 	 * If this CPU is the one which updates jiffies, then give up
881 	 * the assignment and let it be taken by the CPU which runs
882 	 * the tick timer next, which might be this CPU as well. If we
883 	 * don't drop this here the jiffies might be stale and
884 	 * do_timer() never invoked. Keep track of the fact that it
885 	 * was the one which had the do_timer() duty last.
886 	 */
887 	if (cpu == tick_do_timer_cpu) {
888 		tick_do_timer_cpu = TICK_DO_TIMER_NONE;
889 		ts->do_timer_last = 1;
890 	} else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
891 		ts->do_timer_last = 0;
892 	}
893 
894 	/* Skip reprogram of event if its not changed */
895 	if (ts->tick_stopped && (expires == ts->next_tick)) {
896 		/* Sanity check: make sure clockevent is actually programmed */
897 		if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
898 			return;
899 
900 		WARN_ON_ONCE(1);
901 		printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
902 			    basemono, ts->next_tick, dev->next_event,
903 			    hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
904 	}
905 
906 	/*
907 	 * nohz_stop_sched_tick can be called several times before
908 	 * the nohz_restart_sched_tick is called. This happens when
909 	 * interrupts arrive which do not cause a reschedule. In the
910 	 * first call we save the current tick time, so we can restart
911 	 * the scheduler tick in nohz_restart_sched_tick.
912 	 */
913 	if (!ts->tick_stopped) {
914 		calc_load_nohz_start();
915 		quiet_vmstat();
916 
917 		ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
918 		ts->tick_stopped = 1;
919 		trace_tick_stop(1, TICK_DEP_MASK_NONE);
920 	}
921 
922 	ts->next_tick = tick;
923 
924 	/*
925 	 * If the expiration time == KTIME_MAX, then we simply stop
926 	 * the tick timer.
927 	 */
928 	if (unlikely(expires == KTIME_MAX)) {
929 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
930 			hrtimer_cancel(&ts->sched_timer);
931 		else
932 			tick_program_event(KTIME_MAX, 1);
933 		return;
934 	}
935 
936 	if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
937 		hrtimer_start(&ts->sched_timer, tick,
938 			      HRTIMER_MODE_ABS_PINNED_HARD);
939 	} else {
940 		hrtimer_set_expires(&ts->sched_timer, tick);
941 		tick_program_event(tick, 1);
942 	}
943 }
944 
945 static void tick_nohz_retain_tick(struct tick_sched *ts)
946 {
947 	ts->timer_expires_base = 0;
948 }
949 
950 #ifdef CONFIG_NO_HZ_FULL
951 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
952 {
953 	if (tick_nohz_next_event(ts, cpu))
954 		tick_nohz_stop_tick(ts, cpu);
955 	else
956 		tick_nohz_retain_tick(ts);
957 }
958 #endif /* CONFIG_NO_HZ_FULL */
959 
960 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
961 {
962 	/* Update jiffies first */
963 	tick_do_update_jiffies64(now);
964 
965 	/*
966 	 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
967 	 * the clock forward checks in the enqueue path:
968 	 */
969 	timer_clear_idle();
970 
971 	calc_load_nohz_stop();
972 	touch_softlockup_watchdog_sched();
973 	/*
974 	 * Cancel the scheduled timer and restore the tick
975 	 */
976 	ts->tick_stopped  = 0;
977 	tick_nohz_restart(ts, now);
978 }
979 
980 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
981 					 ktime_t now)
982 {
983 #ifdef CONFIG_NO_HZ_FULL
984 	int cpu = smp_processor_id();
985 
986 	if (can_stop_full_tick(cpu, ts))
987 		tick_nohz_stop_sched_tick(ts, cpu);
988 	else if (ts->tick_stopped)
989 		tick_nohz_restart_sched_tick(ts, now);
990 #endif
991 }
992 
993 static void tick_nohz_full_update_tick(struct tick_sched *ts)
994 {
995 	if (!tick_nohz_full_cpu(smp_processor_id()))
996 		return;
997 
998 	if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
999 		return;
1000 
1001 	__tick_nohz_full_update_tick(ts, ktime_get());
1002 }
1003 
1004 /*
1005  * A pending softirq outside an IRQ (or softirq disabled section) context
1006  * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1007  * reach here due to the need_resched() early check in can_stop_idle_tick().
1008  *
1009  * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1010  * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1011  * triggering the below since wakep_softirqd() is ignored.
1012  *
1013  */
1014 static bool report_idle_softirq(void)
1015 {
1016 	static int ratelimit;
1017 	unsigned int pending = local_softirq_pending();
1018 
1019 	if (likely(!pending))
1020 		return false;
1021 
1022 	/* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1023 	if (!cpu_active(smp_processor_id())) {
1024 		pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1025 		if (!pending)
1026 			return false;
1027 	}
1028 
1029 	if (ratelimit < 10)
1030 		return false;
1031 
1032 	/* On RT, softirqs handling may be waiting on some lock */
1033 	if (!local_bh_blocked())
1034 		return false;
1035 
1036 	pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1037 		pending);
1038 	ratelimit++;
1039 
1040 	return true;
1041 }
1042 
1043 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1044 {
1045 	/*
1046 	 * If this CPU is offline and it is the one which updates
1047 	 * jiffies, then give up the assignment and let it be taken by
1048 	 * the CPU which runs the tick timer next. If we don't drop
1049 	 * this here the jiffies might be stale and do_timer() never
1050 	 * invoked.
1051 	 */
1052 	if (unlikely(!cpu_online(cpu))) {
1053 		if (cpu == tick_do_timer_cpu)
1054 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1055 		/*
1056 		 * Make sure the CPU doesn't get fooled by obsolete tick
1057 		 * deadline if it comes back online later.
1058 		 */
1059 		ts->next_tick = 0;
1060 		return false;
1061 	}
1062 
1063 	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1064 		return false;
1065 
1066 	if (need_resched())
1067 		return false;
1068 
1069 	if (unlikely(report_idle_softirq()))
1070 		return false;
1071 
1072 	if (tick_nohz_full_enabled()) {
1073 		/*
1074 		 * Keep the tick alive to guarantee timekeeping progression
1075 		 * if there are full dynticks CPUs around
1076 		 */
1077 		if (tick_do_timer_cpu == cpu)
1078 			return false;
1079 
1080 		/* Should not happen for nohz-full */
1081 		if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1082 			return false;
1083 	}
1084 
1085 	return true;
1086 }
1087 
1088 static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1089 {
1090 	ktime_t expires;
1091 	int cpu = smp_processor_id();
1092 
1093 	/*
1094 	 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1095 	 * tick timer expiration time is known already.
1096 	 */
1097 	if (ts->timer_expires_base)
1098 		expires = ts->timer_expires;
1099 	else if (can_stop_idle_tick(cpu, ts))
1100 		expires = tick_nohz_next_event(ts, cpu);
1101 	else
1102 		return;
1103 
1104 	ts->idle_calls++;
1105 
1106 	if (expires > 0LL) {
1107 		int was_stopped = ts->tick_stopped;
1108 
1109 		tick_nohz_stop_tick(ts, cpu);
1110 
1111 		ts->idle_sleeps++;
1112 		ts->idle_expires = expires;
1113 
1114 		if (!was_stopped && ts->tick_stopped) {
1115 			ts->idle_jiffies = ts->last_jiffies;
1116 			nohz_balance_enter_idle(cpu);
1117 		}
1118 	} else {
1119 		tick_nohz_retain_tick(ts);
1120 	}
1121 }
1122 
1123 /**
1124  * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1125  *
1126  * When the next event is more than a tick into the future, stop the idle tick
1127  */
1128 void tick_nohz_idle_stop_tick(void)
1129 {
1130 	__tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1131 }
1132 
1133 void tick_nohz_idle_retain_tick(void)
1134 {
1135 	tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1136 	/*
1137 	 * Undo the effect of get_next_timer_interrupt() called from
1138 	 * tick_nohz_next_event().
1139 	 */
1140 	timer_clear_idle();
1141 }
1142 
1143 /**
1144  * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1145  *
1146  * Called when we start the idle loop.
1147  */
1148 void tick_nohz_idle_enter(void)
1149 {
1150 	struct tick_sched *ts;
1151 
1152 	lockdep_assert_irqs_enabled();
1153 
1154 	local_irq_disable();
1155 
1156 	ts = this_cpu_ptr(&tick_cpu_sched);
1157 
1158 	WARN_ON_ONCE(ts->timer_expires_base);
1159 
1160 	ts->inidle = 1;
1161 	tick_nohz_start_idle(ts);
1162 
1163 	local_irq_enable();
1164 }
1165 
1166 /**
1167  * tick_nohz_irq_exit - update next tick event from interrupt exit
1168  *
1169  * When an interrupt fires while we are idle and it doesn't cause
1170  * a reschedule, it may still add, modify or delete a timer, enqueue
1171  * an RCU callback, etc...
1172  * So we need to re-calculate and reprogram the next tick event.
1173  */
1174 void tick_nohz_irq_exit(void)
1175 {
1176 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1177 
1178 	if (ts->inidle)
1179 		tick_nohz_start_idle(ts);
1180 	else
1181 		tick_nohz_full_update_tick(ts);
1182 }
1183 
1184 /**
1185  * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1186  */
1187 bool tick_nohz_idle_got_tick(void)
1188 {
1189 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1190 
1191 	if (ts->got_idle_tick) {
1192 		ts->got_idle_tick = 0;
1193 		return true;
1194 	}
1195 	return false;
1196 }
1197 
1198 /**
1199  * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1200  * or the tick, whatever that expires first. Note that, if the tick has been
1201  * stopped, it returns the next hrtimer.
1202  *
1203  * Called from power state control code with interrupts disabled
1204  */
1205 ktime_t tick_nohz_get_next_hrtimer(void)
1206 {
1207 	return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1208 }
1209 
1210 /**
1211  * tick_nohz_get_sleep_length - return the expected length of the current sleep
1212  * @delta_next: duration until the next event if the tick cannot be stopped
1213  *
1214  * Called from power state control code with interrupts disabled.
1215  *
1216  * The return value of this function and/or the value returned by it through the
1217  * @delta_next pointer can be negative which must be taken into account by its
1218  * callers.
1219  */
1220 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1221 {
1222 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1223 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1224 	int cpu = smp_processor_id();
1225 	/*
1226 	 * The idle entry time is expected to be a sufficient approximation of
1227 	 * the current time at this point.
1228 	 */
1229 	ktime_t now = ts->idle_entrytime;
1230 	ktime_t next_event;
1231 
1232 	WARN_ON_ONCE(!ts->inidle);
1233 
1234 	*delta_next = ktime_sub(dev->next_event, now);
1235 
1236 	if (!can_stop_idle_tick(cpu, ts))
1237 		return *delta_next;
1238 
1239 	next_event = tick_nohz_next_event(ts, cpu);
1240 	if (!next_event)
1241 		return *delta_next;
1242 
1243 	/*
1244 	 * If the next highres timer to expire is earlier than next_event, the
1245 	 * idle governor needs to know that.
1246 	 */
1247 	next_event = min_t(u64, next_event,
1248 			   hrtimer_next_event_without(&ts->sched_timer));
1249 
1250 	return ktime_sub(next_event, now);
1251 }
1252 
1253 /**
1254  * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1255  * for a particular CPU.
1256  *
1257  * Called from the schedutil frequency scaling governor in scheduler context.
1258  */
1259 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1260 {
1261 	struct tick_sched *ts = tick_get_tick_sched(cpu);
1262 
1263 	return ts->idle_calls;
1264 }
1265 
1266 /**
1267  * tick_nohz_get_idle_calls - return the current idle calls counter value
1268  *
1269  * Called from the schedutil frequency scaling governor in scheduler context.
1270  */
1271 unsigned long tick_nohz_get_idle_calls(void)
1272 {
1273 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1274 
1275 	return ts->idle_calls;
1276 }
1277 
1278 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1279 					ktime_t now)
1280 {
1281 	unsigned long ticks;
1282 
1283 	ts->idle_exittime = now;
1284 
1285 	if (vtime_accounting_enabled_this_cpu())
1286 		return;
1287 	/*
1288 	 * We stopped the tick in idle. Update process times would miss the
1289 	 * time we slept as update_process_times does only a 1 tick
1290 	 * accounting. Enforce that this is accounted to idle !
1291 	 */
1292 	ticks = jiffies - ts->idle_jiffies;
1293 	/*
1294 	 * We might be one off. Do not randomly account a huge number of ticks!
1295 	 */
1296 	if (ticks && ticks < LONG_MAX)
1297 		account_idle_ticks(ticks);
1298 }
1299 
1300 void tick_nohz_idle_restart_tick(void)
1301 {
1302 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1303 
1304 	if (ts->tick_stopped) {
1305 		ktime_t now = ktime_get();
1306 		tick_nohz_restart_sched_tick(ts, now);
1307 		tick_nohz_account_idle_time(ts, now);
1308 	}
1309 }
1310 
1311 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1312 {
1313 	if (tick_nohz_full_cpu(smp_processor_id()))
1314 		__tick_nohz_full_update_tick(ts, now);
1315 	else
1316 		tick_nohz_restart_sched_tick(ts, now);
1317 
1318 	tick_nohz_account_idle_time(ts, now);
1319 }
1320 
1321 /**
1322  * tick_nohz_idle_exit - restart the idle tick from the idle task
1323  *
1324  * Restart the idle tick when the CPU is woken up from idle
1325  * This also exit the RCU extended quiescent state. The CPU
1326  * can use RCU again after this function is called.
1327  */
1328 void tick_nohz_idle_exit(void)
1329 {
1330 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1331 	bool idle_active, tick_stopped;
1332 	ktime_t now;
1333 
1334 	local_irq_disable();
1335 
1336 	WARN_ON_ONCE(!ts->inidle);
1337 	WARN_ON_ONCE(ts->timer_expires_base);
1338 
1339 	ts->inidle = 0;
1340 	idle_active = ts->idle_active;
1341 	tick_stopped = ts->tick_stopped;
1342 
1343 	if (idle_active || tick_stopped)
1344 		now = ktime_get();
1345 
1346 	if (idle_active)
1347 		tick_nohz_stop_idle(ts, now);
1348 
1349 	if (tick_stopped)
1350 		tick_nohz_idle_update_tick(ts, now);
1351 
1352 	local_irq_enable();
1353 }
1354 
1355 /*
1356  * The nohz low res interrupt handler
1357  */
1358 static void tick_nohz_handler(struct clock_event_device *dev)
1359 {
1360 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1361 	struct pt_regs *regs = get_irq_regs();
1362 	ktime_t now = ktime_get();
1363 
1364 	dev->next_event = KTIME_MAX;
1365 
1366 	tick_sched_do_timer(ts, now);
1367 	tick_sched_handle(ts, regs);
1368 
1369 	if (unlikely(ts->tick_stopped)) {
1370 		/*
1371 		 * The clockevent device is not reprogrammed, so change the
1372 		 * clock event device to ONESHOT_STOPPED to avoid spurious
1373 		 * interrupts on devices which might not be truly one shot.
1374 		 */
1375 		tick_program_event(KTIME_MAX, 1);
1376 		return;
1377 	}
1378 
1379 	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1380 	tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1381 }
1382 
1383 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1384 {
1385 	if (!tick_nohz_enabled)
1386 		return;
1387 	ts->nohz_mode = mode;
1388 	/* One update is enough */
1389 	if (!test_and_set_bit(0, &tick_nohz_active))
1390 		timers_update_nohz();
1391 }
1392 
1393 /**
1394  * tick_nohz_switch_to_nohz - switch to nohz mode
1395  */
1396 static void tick_nohz_switch_to_nohz(void)
1397 {
1398 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1399 	ktime_t next;
1400 
1401 	if (!tick_nohz_enabled)
1402 		return;
1403 
1404 	if (tick_switch_to_oneshot(tick_nohz_handler))
1405 		return;
1406 
1407 	/*
1408 	 * Recycle the hrtimer in ts, so we can share the
1409 	 * hrtimer_forward with the highres code.
1410 	 */
1411 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1412 	/* Get the next period */
1413 	next = tick_init_jiffy_update();
1414 
1415 	hrtimer_set_expires(&ts->sched_timer, next);
1416 	hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1417 	tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1418 	tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1419 }
1420 
1421 static inline void tick_nohz_irq_enter(void)
1422 {
1423 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1424 	ktime_t now;
1425 
1426 	if (!ts->idle_active && !ts->tick_stopped)
1427 		return;
1428 	now = ktime_get();
1429 	if (ts->idle_active)
1430 		tick_nohz_stop_idle(ts, now);
1431 	/*
1432 	 * If all CPUs are idle. We may need to update a stale jiffies value.
1433 	 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1434 	 * alive but it might be busy looping with interrupts disabled in some
1435 	 * rare case (typically stop machine). So we must make sure we have a
1436 	 * last resort.
1437 	 */
1438 	if (ts->tick_stopped)
1439 		tick_nohz_update_jiffies(now);
1440 }
1441 
1442 #else
1443 
1444 static inline void tick_nohz_switch_to_nohz(void) { }
1445 static inline void tick_nohz_irq_enter(void) { }
1446 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1447 
1448 #endif /* CONFIG_NO_HZ_COMMON */
1449 
1450 /*
1451  * Called from irq_enter to notify about the possible interruption of idle()
1452  */
1453 void tick_irq_enter(void)
1454 {
1455 	tick_check_oneshot_broadcast_this_cpu();
1456 	tick_nohz_irq_enter();
1457 }
1458 
1459 /*
1460  * High resolution timer specific code
1461  */
1462 #ifdef CONFIG_HIGH_RES_TIMERS
1463 /*
1464  * We rearm the timer until we get disabled by the idle code.
1465  * Called with interrupts disabled.
1466  */
1467 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1468 {
1469 	struct tick_sched *ts =
1470 		container_of(timer, struct tick_sched, sched_timer);
1471 	struct pt_regs *regs = get_irq_regs();
1472 	ktime_t now = ktime_get();
1473 
1474 	tick_sched_do_timer(ts, now);
1475 
1476 	/*
1477 	 * Do not call, when we are not in irq context and have
1478 	 * no valid regs pointer
1479 	 */
1480 	if (regs)
1481 		tick_sched_handle(ts, regs);
1482 	else
1483 		ts->next_tick = 0;
1484 
1485 	/* No need to reprogram if we are in idle or full dynticks mode */
1486 	if (unlikely(ts->tick_stopped))
1487 		return HRTIMER_NORESTART;
1488 
1489 	hrtimer_forward(timer, now, TICK_NSEC);
1490 
1491 	return HRTIMER_RESTART;
1492 }
1493 
1494 static int sched_skew_tick;
1495 
1496 static int __init skew_tick(char *str)
1497 {
1498 	get_option(&str, &sched_skew_tick);
1499 
1500 	return 0;
1501 }
1502 early_param("skew_tick", skew_tick);
1503 
1504 /**
1505  * tick_setup_sched_timer - setup the tick emulation timer
1506  */
1507 void tick_setup_sched_timer(void)
1508 {
1509 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1510 	ktime_t now = ktime_get();
1511 
1512 	/*
1513 	 * Emulate tick processing via per-CPU hrtimers:
1514 	 */
1515 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1516 	ts->sched_timer.function = tick_sched_timer;
1517 
1518 	/* Get the next period (per-CPU) */
1519 	hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1520 
1521 	/* Offset the tick to avert jiffies_lock contention. */
1522 	if (sched_skew_tick) {
1523 		u64 offset = TICK_NSEC >> 1;
1524 		do_div(offset, num_possible_cpus());
1525 		offset *= smp_processor_id();
1526 		hrtimer_add_expires_ns(&ts->sched_timer, offset);
1527 	}
1528 
1529 	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1530 	hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1531 	tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1532 }
1533 #endif /* HIGH_RES_TIMERS */
1534 
1535 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1536 void tick_cancel_sched_timer(int cpu)
1537 {
1538 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1539 
1540 # ifdef CONFIG_HIGH_RES_TIMERS
1541 	if (ts->sched_timer.base)
1542 		hrtimer_cancel(&ts->sched_timer);
1543 # endif
1544 
1545 	memset(ts, 0, sizeof(*ts));
1546 }
1547 #endif
1548 
1549 /**
1550  * Async notification about clocksource changes
1551  */
1552 void tick_clock_notify(void)
1553 {
1554 	int cpu;
1555 
1556 	for_each_possible_cpu(cpu)
1557 		set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1558 }
1559 
1560 /*
1561  * Async notification about clock event changes
1562  */
1563 void tick_oneshot_notify(void)
1564 {
1565 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1566 
1567 	set_bit(0, &ts->check_clocks);
1568 }
1569 
1570 /**
1571  * Check, if a change happened, which makes oneshot possible.
1572  *
1573  * Called cyclic from the hrtimer softirq (driven by the timer
1574  * softirq) allow_nohz signals, that we can switch into low-res nohz
1575  * mode, because high resolution timers are disabled (either compile
1576  * or runtime). Called with interrupts disabled.
1577  */
1578 int tick_check_oneshot_change(int allow_nohz)
1579 {
1580 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1581 
1582 	if (!test_and_clear_bit(0, &ts->check_clocks))
1583 		return 0;
1584 
1585 	if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1586 		return 0;
1587 
1588 	if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1589 		return 0;
1590 
1591 	if (!allow_nohz)
1592 		return 1;
1593 
1594 	tick_nohz_switch_to_nohz();
1595 	return 0;
1596 }
1597