xref: /linux/kernel/time/tick-sched.c (revision 223981db9bafb80f558162c148f261e2ff043dbe)
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  *  NOHZ 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 	 * 64-bit can do a quick check without holding the 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 	 * 32-bit cannot do that because the store of 'tick_next_period'
68 	 * consists of two 32-bit stores, and the first store could be
69 	 * moved by the CPU 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 	 * Re-evaluate 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 	/* Keep the tick_next_period variable up to date */
121 	nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
122 
123 	if (IS_ENABLED(CONFIG_64BIT)) {
124 		/*
125 		 * Pairs with smp_load_acquire() in the lockless quick
126 		 * check above, and ensures that the update to 'jiffies_64' is
127 		 * not reordered vs. the store to 'tick_next_period', neither
128 		 * by the compiler nor by the CPU.
129 		 */
130 		smp_store_release(&tick_next_period, nextp);
131 	} else {
132 		/*
133 		 * A plain store is good enough on 32-bit, as the quick check
134 		 * above is protected by the sequence count.
135 		 */
136 		tick_next_period = nextp;
137 	}
138 
139 	/*
140 	 * Release the sequence count. calc_global_load() below is not
141 	 * protected by it, but 'jiffies_lock' needs to be held to prevent
142 	 * concurrent invocations.
143 	 */
144 	write_seqcount_end(&jiffies_seq);
145 
146 	calc_global_load();
147 
148 	raw_spin_unlock(&jiffies_lock);
149 	update_wall_time();
150 }
151 
152 /*
153  * Initialize and return retrieve the jiffies update.
154  */
155 static ktime_t tick_init_jiffy_update(void)
156 {
157 	ktime_t period;
158 
159 	raw_spin_lock(&jiffies_lock);
160 	write_seqcount_begin(&jiffies_seq);
161 
162 	/* Have we started the jiffies update yet ? */
163 	if (last_jiffies_update == 0) {
164 		u32 rem;
165 
166 		/*
167 		 * Ensure that the tick is aligned to a multiple of
168 		 * TICK_NSEC.
169 		 */
170 		div_u64_rem(tick_next_period, TICK_NSEC, &rem);
171 		if (rem)
172 			tick_next_period += TICK_NSEC - rem;
173 
174 		last_jiffies_update = tick_next_period;
175 	}
176 	period = last_jiffies_update;
177 
178 	write_seqcount_end(&jiffies_seq);
179 	raw_spin_unlock(&jiffies_lock);
180 
181 	return period;
182 }
183 
184 #define MAX_STALLED_JIFFIES 5
185 
186 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
187 {
188 	int cpu = smp_processor_id();
189 
190 #ifdef CONFIG_NO_HZ_COMMON
191 	/*
192 	 * Check if the do_timer duty was dropped. We don't care about
193 	 * concurrency: This happens only when the CPU in charge went
194 	 * into a long sleep. If two CPUs happen to assign themselves to
195 	 * this duty, then the jiffies update is still serialized by
196 	 * 'jiffies_lock'.
197 	 *
198 	 * If nohz_full is enabled, this should not happen because the
199 	 * 'tick_do_timer_cpu' CPU never relinquishes.
200 	 */
201 	if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
202 #ifdef CONFIG_NO_HZ_FULL
203 		WARN_ON_ONCE(tick_nohz_full_running);
204 #endif
205 		tick_do_timer_cpu = cpu;
206 	}
207 #endif
208 
209 	/* Check if jiffies need an update */
210 	if (tick_do_timer_cpu == cpu)
211 		tick_do_update_jiffies64(now);
212 
213 	/*
214 	 * If the jiffies update stalled for too long (timekeeper in stop_machine()
215 	 * or VMEXIT'ed for several msecs), force an update.
216 	 */
217 	if (ts->last_tick_jiffies != jiffies) {
218 		ts->stalled_jiffies = 0;
219 		ts->last_tick_jiffies = READ_ONCE(jiffies);
220 	} else {
221 		if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
222 			tick_do_update_jiffies64(now);
223 			ts->stalled_jiffies = 0;
224 			ts->last_tick_jiffies = READ_ONCE(jiffies);
225 		}
226 	}
227 
228 	if (ts->inidle)
229 		ts->got_idle_tick = 1;
230 }
231 
232 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
233 {
234 #ifdef CONFIG_NO_HZ_COMMON
235 	/*
236 	 * When we are idle and the tick is stopped, we have to touch
237 	 * the watchdog as we might not schedule for a really long
238 	 * time. This happens on completely idle SMP systems while
239 	 * waiting on the login prompt. We also increment the "start of
240 	 * idle" jiffy stamp so the idle accounting adjustment we do
241 	 * when we go busy again does not account too many ticks.
242 	 */
243 	if (ts->tick_stopped) {
244 		touch_softlockup_watchdog_sched();
245 		if (is_idle_task(current))
246 			ts->idle_jiffies++;
247 		/*
248 		 * In case the current tick fired too early past its expected
249 		 * expiration, make sure we don't bypass the next clock reprogramming
250 		 * to the same deadline.
251 		 */
252 		ts->next_tick = 0;
253 	}
254 #endif
255 	update_process_times(user_mode(regs));
256 	profile_tick(CPU_PROFILING);
257 }
258 #endif
259 
260 #ifdef CONFIG_NO_HZ_FULL
261 cpumask_var_t tick_nohz_full_mask;
262 EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
263 bool tick_nohz_full_running;
264 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
265 static atomic_t tick_dep_mask;
266 
267 static bool check_tick_dependency(atomic_t *dep)
268 {
269 	int val = atomic_read(dep);
270 
271 	if (val & TICK_DEP_MASK_POSIX_TIMER) {
272 		trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
273 		return true;
274 	}
275 
276 	if (val & TICK_DEP_MASK_PERF_EVENTS) {
277 		trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
278 		return true;
279 	}
280 
281 	if (val & TICK_DEP_MASK_SCHED) {
282 		trace_tick_stop(0, TICK_DEP_MASK_SCHED);
283 		return true;
284 	}
285 
286 	if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
287 		trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
288 		return true;
289 	}
290 
291 	if (val & TICK_DEP_MASK_RCU) {
292 		trace_tick_stop(0, TICK_DEP_MASK_RCU);
293 		return true;
294 	}
295 
296 	if (val & TICK_DEP_MASK_RCU_EXP) {
297 		trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
298 		return true;
299 	}
300 
301 	return false;
302 }
303 
304 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
305 {
306 	lockdep_assert_irqs_disabled();
307 
308 	if (unlikely(!cpu_online(cpu)))
309 		return false;
310 
311 	if (check_tick_dependency(&tick_dep_mask))
312 		return false;
313 
314 	if (check_tick_dependency(&ts->tick_dep_mask))
315 		return false;
316 
317 	if (check_tick_dependency(&current->tick_dep_mask))
318 		return false;
319 
320 	if (check_tick_dependency(&current->signal->tick_dep_mask))
321 		return false;
322 
323 	return true;
324 }
325 
326 static void nohz_full_kick_func(struct irq_work *work)
327 {
328 	/* Empty, the tick restart happens on tick_nohz_irq_exit() */
329 }
330 
331 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
332 	IRQ_WORK_INIT_HARD(nohz_full_kick_func);
333 
334 /*
335  * Kick this CPU if it's full dynticks in order to force it to
336  * re-evaluate its dependency on the tick and restart it if necessary.
337  * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
338  * is NMI safe.
339  */
340 static void tick_nohz_full_kick(void)
341 {
342 	if (!tick_nohz_full_cpu(smp_processor_id()))
343 		return;
344 
345 	irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
346 }
347 
348 /*
349  * Kick the CPU if it's full dynticks in order to force it to
350  * re-evaluate its dependency on the tick and restart it if necessary.
351  */
352 void tick_nohz_full_kick_cpu(int cpu)
353 {
354 	if (!tick_nohz_full_cpu(cpu))
355 		return;
356 
357 	irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
358 }
359 
360 static void tick_nohz_kick_task(struct task_struct *tsk)
361 {
362 	int cpu;
363 
364 	/*
365 	 * If the task is not running, run_posix_cpu_timers()
366 	 * has nothing to elapse, and an IPI can then be optimized out.
367 	 *
368 	 * activate_task()                      STORE p->tick_dep_mask
369 	 *   STORE p->on_rq
370 	 * __schedule() (switch to task 'p')    smp_mb() (atomic_fetch_or())
371 	 *   LOCK rq->lock                      LOAD p->on_rq
372 	 *   smp_mb__after_spin_lock()
373 	 *   tick_nohz_task_switch()
374 	 *     LOAD p->tick_dep_mask
375 	 */
376 	if (!sched_task_on_rq(tsk))
377 		return;
378 
379 	/*
380 	 * If the task concurrently migrates to another CPU,
381 	 * we guarantee it sees the new tick dependency upon
382 	 * schedule.
383 	 *
384 	 * set_task_cpu(p, cpu);
385 	 *   STORE p->cpu = @cpu
386 	 * __schedule() (switch to task 'p')
387 	 *   LOCK rq->lock
388 	 *   smp_mb__after_spin_lock()          STORE p->tick_dep_mask
389 	 *   tick_nohz_task_switch()            smp_mb() (atomic_fetch_or())
390 	 *      LOAD p->tick_dep_mask           LOAD p->cpu
391 	 */
392 	cpu = task_cpu(tsk);
393 
394 	preempt_disable();
395 	if (cpu_online(cpu))
396 		tick_nohz_full_kick_cpu(cpu);
397 	preempt_enable();
398 }
399 
400 /*
401  * Kick all full dynticks CPUs in order to force these to re-evaluate
402  * their dependency on the tick and restart it if necessary.
403  */
404 static void tick_nohz_full_kick_all(void)
405 {
406 	int cpu;
407 
408 	if (!tick_nohz_full_running)
409 		return;
410 
411 	preempt_disable();
412 	for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
413 		tick_nohz_full_kick_cpu(cpu);
414 	preempt_enable();
415 }
416 
417 static void tick_nohz_dep_set_all(atomic_t *dep,
418 				  enum tick_dep_bits bit)
419 {
420 	int prev;
421 
422 	prev = atomic_fetch_or(BIT(bit), dep);
423 	if (!prev)
424 		tick_nohz_full_kick_all();
425 }
426 
427 /*
428  * Set a global tick dependency. Used by perf events that rely on freq and
429  * unstable clocks.
430  */
431 void tick_nohz_dep_set(enum tick_dep_bits bit)
432 {
433 	tick_nohz_dep_set_all(&tick_dep_mask, bit);
434 }
435 
436 void tick_nohz_dep_clear(enum tick_dep_bits bit)
437 {
438 	atomic_andnot(BIT(bit), &tick_dep_mask);
439 }
440 
441 /*
442  * Set per-CPU tick dependency. Used by scheduler and perf events in order to
443  * manage event-throttling.
444  */
445 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
446 {
447 	int prev;
448 	struct tick_sched *ts;
449 
450 	ts = per_cpu_ptr(&tick_cpu_sched, cpu);
451 
452 	prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
453 	if (!prev) {
454 		preempt_disable();
455 		/* Perf needs local kick that is NMI safe */
456 		if (cpu == smp_processor_id()) {
457 			tick_nohz_full_kick();
458 		} else {
459 			/* Remote IRQ work not NMI-safe */
460 			if (!WARN_ON_ONCE(in_nmi()))
461 				tick_nohz_full_kick_cpu(cpu);
462 		}
463 		preempt_enable();
464 	}
465 }
466 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
467 
468 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
469 {
470 	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
471 
472 	atomic_andnot(BIT(bit), &ts->tick_dep_mask);
473 }
474 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
475 
476 /*
477  * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
478  * in order to elapse per task timers.
479  */
480 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
481 {
482 	if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
483 		tick_nohz_kick_task(tsk);
484 }
485 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
486 
487 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
488 {
489 	atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
490 }
491 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
492 
493 /*
494  * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
495  * per process timers.
496  */
497 void tick_nohz_dep_set_signal(struct task_struct *tsk,
498 			      enum tick_dep_bits bit)
499 {
500 	int prev;
501 	struct signal_struct *sig = tsk->signal;
502 
503 	prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
504 	if (!prev) {
505 		struct task_struct *t;
506 
507 		lockdep_assert_held(&tsk->sighand->siglock);
508 		__for_each_thread(sig, t)
509 			tick_nohz_kick_task(t);
510 	}
511 }
512 
513 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
514 {
515 	atomic_andnot(BIT(bit), &sig->tick_dep_mask);
516 }
517 
518 /*
519  * Re-evaluate the need for the tick as we switch the current task.
520  * It might need the tick due to per task/process properties:
521  * perf events, posix CPU timers, ...
522  */
523 void __tick_nohz_task_switch(void)
524 {
525 	struct tick_sched *ts;
526 
527 	if (!tick_nohz_full_cpu(smp_processor_id()))
528 		return;
529 
530 	ts = this_cpu_ptr(&tick_cpu_sched);
531 
532 	if (ts->tick_stopped) {
533 		if (atomic_read(&current->tick_dep_mask) ||
534 		    atomic_read(&current->signal->tick_dep_mask))
535 			tick_nohz_full_kick();
536 	}
537 }
538 
539 /* Get the boot-time nohz CPU list from the kernel parameters. */
540 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
541 {
542 	alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
543 	cpumask_copy(tick_nohz_full_mask, cpumask);
544 	tick_nohz_full_running = true;
545 }
546 
547 bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
548 {
549 	/*
550 	 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
551 	 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
552 	 * CPUs. It must remain online when nohz full is enabled.
553 	 */
554 	if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
555 		return false;
556 	return true;
557 }
558 
559 static int tick_nohz_cpu_down(unsigned int cpu)
560 {
561 	return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
562 }
563 
564 void __init tick_nohz_init(void)
565 {
566 	int cpu, ret;
567 
568 	if (!tick_nohz_full_running)
569 		return;
570 
571 	/*
572 	 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
573 	 * locking contexts. But then we need IRQ work to raise its own
574 	 * interrupts to avoid circular dependency on the tick.
575 	 */
576 	if (!arch_irq_work_has_interrupt()) {
577 		pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
578 		cpumask_clear(tick_nohz_full_mask);
579 		tick_nohz_full_running = false;
580 		return;
581 	}
582 
583 	if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
584 			!IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
585 		cpu = smp_processor_id();
586 
587 		if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
588 			pr_warn("NO_HZ: Clearing %d from nohz_full range "
589 				"for timekeeping\n", cpu);
590 			cpumask_clear_cpu(cpu, tick_nohz_full_mask);
591 		}
592 	}
593 
594 	for_each_cpu(cpu, tick_nohz_full_mask)
595 		ct_cpu_track_user(cpu);
596 
597 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
598 					"kernel/nohz:predown", NULL,
599 					tick_nohz_cpu_down);
600 	WARN_ON(ret < 0);
601 	pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
602 		cpumask_pr_args(tick_nohz_full_mask));
603 }
604 #endif
605 
606 /*
607  * NOHZ - aka dynamic tick functionality
608  */
609 #ifdef CONFIG_NO_HZ_COMMON
610 /*
611  * NO HZ enabled ?
612  */
613 bool tick_nohz_enabled __read_mostly  = true;
614 unsigned long tick_nohz_active  __read_mostly;
615 /*
616  * Enable / Disable tickless mode
617  */
618 static int __init setup_tick_nohz(char *str)
619 {
620 	return (kstrtobool(str, &tick_nohz_enabled) == 0);
621 }
622 
623 __setup("nohz=", setup_tick_nohz);
624 
625 bool tick_nohz_tick_stopped(void)
626 {
627 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
628 
629 	return ts->tick_stopped;
630 }
631 
632 bool tick_nohz_tick_stopped_cpu(int cpu)
633 {
634 	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
635 
636 	return ts->tick_stopped;
637 }
638 
639 /**
640  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
641  *
642  * Called from interrupt entry when the CPU was idle
643  *
644  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
645  * must be updated. Otherwise an interrupt handler could use a stale jiffy
646  * value. We do this unconditionally on any CPU, as we don't know whether the
647  * CPU, which has the update task assigned, is in a long sleep.
648  */
649 static void tick_nohz_update_jiffies(ktime_t now)
650 {
651 	unsigned long flags;
652 
653 	__this_cpu_write(tick_cpu_sched.idle_waketime, now);
654 
655 	local_irq_save(flags);
656 	tick_do_update_jiffies64(now);
657 	local_irq_restore(flags);
658 
659 	touch_softlockup_watchdog_sched();
660 }
661 
662 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
663 {
664 	ktime_t delta;
665 
666 	if (WARN_ON_ONCE(!ts->idle_active))
667 		return;
668 
669 	delta = ktime_sub(now, ts->idle_entrytime);
670 
671 	write_seqcount_begin(&ts->idle_sleeptime_seq);
672 	if (nr_iowait_cpu(smp_processor_id()) > 0)
673 		ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
674 	else
675 		ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
676 
677 	ts->idle_entrytime = now;
678 	ts->idle_active = 0;
679 	write_seqcount_end(&ts->idle_sleeptime_seq);
680 
681 	sched_clock_idle_wakeup_event();
682 }
683 
684 static void tick_nohz_start_idle(struct tick_sched *ts)
685 {
686 	write_seqcount_begin(&ts->idle_sleeptime_seq);
687 	ts->idle_entrytime = ktime_get();
688 	ts->idle_active = 1;
689 	write_seqcount_end(&ts->idle_sleeptime_seq);
690 
691 	sched_clock_idle_sleep_event();
692 }
693 
694 static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
695 				 bool compute_delta, u64 *last_update_time)
696 {
697 	ktime_t now, idle;
698 	unsigned int seq;
699 
700 	if (!tick_nohz_active)
701 		return -1;
702 
703 	now = ktime_get();
704 	if (last_update_time)
705 		*last_update_time = ktime_to_us(now);
706 
707 	do {
708 		seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
709 
710 		if (ts->idle_active && compute_delta) {
711 			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
712 
713 			idle = ktime_add(*sleeptime, delta);
714 		} else {
715 			idle = *sleeptime;
716 		}
717 	} while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
718 
719 	return ktime_to_us(idle);
720 
721 }
722 
723 /**
724  * get_cpu_idle_time_us - get the total idle time of a CPU
725  * @cpu: CPU number to query
726  * @last_update_time: variable to store update time in. Do not update
727  * counters if NULL.
728  *
729  * Return the cumulative idle time (since boot) for a given
730  * CPU, in microseconds. Note that this is partially broken due to
731  * the counter of iowait tasks that can be remotely updated without
732  * any synchronization. Therefore it is possible to observe backward
733  * values within two consecutive reads.
734  *
735  * This time is measured via accounting rather than sampling,
736  * and is as accurate as ktime_get() is.
737  *
738  * This function returns -1 if NOHZ is not enabled.
739  */
740 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
741 {
742 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
743 
744 	return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
745 				     !nr_iowait_cpu(cpu), last_update_time);
746 }
747 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
748 
749 /**
750  * get_cpu_iowait_time_us - get the total iowait time of a CPU
751  * @cpu: CPU number to query
752  * @last_update_time: variable to store update time in. Do not update
753  * counters if NULL.
754  *
755  * Return the cumulative iowait time (since boot) for a given
756  * CPU, in microseconds. Note this is partially broken due to
757  * the counter of iowait tasks that can be remotely updated without
758  * any synchronization. Therefore it is possible to observe backward
759  * values within two consecutive reads.
760  *
761  * This time is measured via accounting rather than sampling,
762  * and is as accurate as ktime_get() is.
763  *
764  * This function returns -1 if NOHZ is not enabled.
765  */
766 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
767 {
768 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
769 
770 	return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
771 				     nr_iowait_cpu(cpu), last_update_time);
772 }
773 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
774 
775 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
776 {
777 	hrtimer_cancel(&ts->sched_timer);
778 	hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
779 
780 	/* Forward the time to expire in the future */
781 	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
782 
783 	if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
784 		hrtimer_start_expires(&ts->sched_timer,
785 				      HRTIMER_MODE_ABS_PINNED_HARD);
786 	} else {
787 		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
788 	}
789 
790 	/*
791 	 * Reset to make sure the next tick stop doesn't get fooled by past
792 	 * cached clock deadline.
793 	 */
794 	ts->next_tick = 0;
795 }
796 
797 static inline bool local_timer_softirq_pending(void)
798 {
799 	return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
800 }
801 
802 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
803 {
804 	u64 basemono, next_tick, delta, expires;
805 	unsigned long basejiff;
806 	unsigned int seq;
807 
808 	/* Read jiffies and the time when jiffies were updated last */
809 	do {
810 		seq = read_seqcount_begin(&jiffies_seq);
811 		basemono = last_jiffies_update;
812 		basejiff = jiffies;
813 	} while (read_seqcount_retry(&jiffies_seq, seq));
814 	ts->last_jiffies = basejiff;
815 	ts->timer_expires_base = basemono;
816 
817 	/*
818 	 * Keep the periodic tick, when RCU, architecture or irq_work
819 	 * requests it.
820 	 * Aside of that, check whether the local timer softirq is
821 	 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
822 	 * because there is an already expired timer, so it will request
823 	 * immediate expiry, which rearms the hardware timer with a
824 	 * minimal delta, which brings us back to this place
825 	 * immediately. Lather, rinse and repeat...
826 	 */
827 	if (rcu_needs_cpu() || arch_needs_cpu() ||
828 	    irq_work_needs_cpu() || local_timer_softirq_pending()) {
829 		next_tick = basemono + TICK_NSEC;
830 	} else {
831 		/*
832 		 * Get the next pending timer. If high resolution
833 		 * timers are enabled this only takes the timer wheel
834 		 * timers into account. If high resolution timers are
835 		 * disabled this also looks at the next expiring
836 		 * hrtimer.
837 		 */
838 		next_tick = get_next_timer_interrupt(basejiff, basemono);
839 		ts->next_timer = next_tick;
840 	}
841 
842 	/*
843 	 * If the tick is due in the next period, keep it ticking or
844 	 * force prod the timer.
845 	 */
846 	delta = next_tick - basemono;
847 	if (delta <= (u64)TICK_NSEC) {
848 		/*
849 		 * Tell the timer code that the base is not idle, i.e. undo
850 		 * the effect of get_next_timer_interrupt():
851 		 */
852 		timer_clear_idle();
853 		/*
854 		 * We've not stopped the tick yet, and there's a timer in the
855 		 * next period, so no point in stopping it either, bail.
856 		 */
857 		if (!ts->tick_stopped) {
858 			ts->timer_expires = 0;
859 			goto out;
860 		}
861 	}
862 
863 	/*
864 	 * If this CPU is the one which had the do_timer() duty last, we limit
865 	 * the sleep time to the timekeeping 'max_deferment' value.
866 	 * Otherwise we can sleep as long as we want.
867 	 */
868 	delta = timekeeping_max_deferment();
869 	if (cpu != tick_do_timer_cpu &&
870 	    (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
871 		delta = KTIME_MAX;
872 
873 	/* Calculate the next expiry time */
874 	if (delta < (KTIME_MAX - basemono))
875 		expires = basemono + delta;
876 	else
877 		expires = KTIME_MAX;
878 
879 	ts->timer_expires = min_t(u64, expires, next_tick);
880 
881 out:
882 	return ts->timer_expires;
883 }
884 
885 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
886 {
887 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
888 	u64 basemono = ts->timer_expires_base;
889 	u64 expires = ts->timer_expires;
890 	ktime_t tick = expires;
891 
892 	/* Make sure we won't be trying to stop it twice in a row. */
893 	ts->timer_expires_base = 0;
894 
895 	/*
896 	 * If this CPU is the one which updates jiffies, then give up
897 	 * the assignment and let it be taken by the CPU which runs
898 	 * the tick timer next, which might be this CPU as well. If we
899 	 * don't drop this here, the jiffies might be stale and
900 	 * do_timer() never gets invoked. Keep track of the fact that it
901 	 * was the one which had the do_timer() duty last.
902 	 */
903 	if (cpu == tick_do_timer_cpu) {
904 		tick_do_timer_cpu = TICK_DO_TIMER_NONE;
905 		ts->do_timer_last = 1;
906 	} else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
907 		ts->do_timer_last = 0;
908 	}
909 
910 	/* Skip reprogram of event if it's not changed */
911 	if (ts->tick_stopped && (expires == ts->next_tick)) {
912 		/* Sanity check: make sure clockevent is actually programmed */
913 		if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
914 			return;
915 
916 		WARN_ON_ONCE(1);
917 		printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
918 			    basemono, ts->next_tick, dev->next_event,
919 			    hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
920 	}
921 
922 	/*
923 	 * nohz_stop_sched_tick() can be called several times before
924 	 * nohz_restart_sched_tick() is called. This happens when
925 	 * interrupts arrive which do not cause a reschedule. In the
926 	 * first call we save the current tick time, so we can restart
927 	 * the scheduler tick in nohz_restart_sched_tick().
928 	 */
929 	if (!ts->tick_stopped) {
930 		calc_load_nohz_start();
931 		quiet_vmstat();
932 
933 		ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
934 		ts->tick_stopped = 1;
935 		trace_tick_stop(1, TICK_DEP_MASK_NONE);
936 	}
937 
938 	ts->next_tick = tick;
939 
940 	/*
941 	 * If the expiration time == KTIME_MAX, then we simply stop
942 	 * the tick timer.
943 	 */
944 	if (unlikely(expires == KTIME_MAX)) {
945 		if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
946 			hrtimer_cancel(&ts->sched_timer);
947 		else
948 			tick_program_event(KTIME_MAX, 1);
949 		return;
950 	}
951 
952 	if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
953 		hrtimer_start(&ts->sched_timer, tick,
954 			      HRTIMER_MODE_ABS_PINNED_HARD);
955 	} else {
956 		hrtimer_set_expires(&ts->sched_timer, tick);
957 		tick_program_event(tick, 1);
958 	}
959 }
960 
961 static void tick_nohz_retain_tick(struct tick_sched *ts)
962 {
963 	ts->timer_expires_base = 0;
964 }
965 
966 #ifdef CONFIG_NO_HZ_FULL
967 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
968 {
969 	if (tick_nohz_next_event(ts, cpu))
970 		tick_nohz_stop_tick(ts, cpu);
971 	else
972 		tick_nohz_retain_tick(ts);
973 }
974 #endif /* CONFIG_NO_HZ_FULL */
975 
976 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
977 {
978 	/* Update jiffies first */
979 	tick_do_update_jiffies64(now);
980 
981 	/*
982 	 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
983 	 * the clock forward checks in the enqueue path:
984 	 */
985 	timer_clear_idle();
986 
987 	calc_load_nohz_stop();
988 	touch_softlockup_watchdog_sched();
989 
990 	/* Cancel the scheduled timer and restore the tick: */
991 	ts->tick_stopped  = 0;
992 	tick_nohz_restart(ts, now);
993 }
994 
995 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
996 					 ktime_t now)
997 {
998 #ifdef CONFIG_NO_HZ_FULL
999 	int cpu = smp_processor_id();
1000 
1001 	if (can_stop_full_tick(cpu, ts))
1002 		tick_nohz_stop_sched_tick(ts, cpu);
1003 	else if (ts->tick_stopped)
1004 		tick_nohz_restart_sched_tick(ts, now);
1005 #endif
1006 }
1007 
1008 static void tick_nohz_full_update_tick(struct tick_sched *ts)
1009 {
1010 	if (!tick_nohz_full_cpu(smp_processor_id()))
1011 		return;
1012 
1013 	if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1014 		return;
1015 
1016 	__tick_nohz_full_update_tick(ts, ktime_get());
1017 }
1018 
1019 /*
1020  * A pending softirq outside an IRQ (or softirq disabled section) context
1021  * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1022  * reach this code due to the need_resched() early check in can_stop_idle_tick().
1023  *
1024  * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1025  * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1026  * triggering the code below, since wakep_softirqd() is ignored.
1027  *
1028  */
1029 static bool report_idle_softirq(void)
1030 {
1031 	static int ratelimit;
1032 	unsigned int pending = local_softirq_pending();
1033 
1034 	if (likely(!pending))
1035 		return false;
1036 
1037 	/* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1038 	if (!cpu_active(smp_processor_id())) {
1039 		pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1040 		if (!pending)
1041 			return false;
1042 	}
1043 
1044 	if (ratelimit >= 10)
1045 		return false;
1046 
1047 	/* On RT, softirq handling may be waiting on some lock */
1048 	if (local_bh_blocked())
1049 		return false;
1050 
1051 	pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1052 		pending);
1053 	ratelimit++;
1054 
1055 	return true;
1056 }
1057 
1058 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1059 {
1060 	/*
1061 	 * If this CPU is offline and it is the one which updates
1062 	 * jiffies, then give up the assignment and let it be taken by
1063 	 * the CPU which runs the tick timer next. If we don't drop
1064 	 * this here, the jiffies might be stale and do_timer() never
1065 	 * gets invoked.
1066 	 */
1067 	if (unlikely(!cpu_online(cpu))) {
1068 		if (cpu == tick_do_timer_cpu)
1069 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1070 		/*
1071 		 * Make sure the CPU doesn't get fooled by obsolete tick
1072 		 * deadline if it comes back online later.
1073 		 */
1074 		ts->next_tick = 0;
1075 		return false;
1076 	}
1077 
1078 	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1079 		return false;
1080 
1081 	if (need_resched())
1082 		return false;
1083 
1084 	if (unlikely(report_idle_softirq()))
1085 		return false;
1086 
1087 	if (tick_nohz_full_enabled()) {
1088 		/*
1089 		 * Keep the tick alive to guarantee timekeeping progression
1090 		 * if there are full dynticks CPUs around
1091 		 */
1092 		if (tick_do_timer_cpu == cpu)
1093 			return false;
1094 
1095 		/* Should not happen for nohz-full */
1096 		if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1097 			return false;
1098 	}
1099 
1100 	return true;
1101 }
1102 
1103 /**
1104  * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1105  *
1106  * When the next event is more than a tick into the future, stop the idle tick
1107  */
1108 void tick_nohz_idle_stop_tick(void)
1109 {
1110 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1111 	int cpu = smp_processor_id();
1112 	ktime_t expires;
1113 
1114 	/*
1115 	 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1116 	 * tick timer expiration time is known already.
1117 	 */
1118 	if (ts->timer_expires_base)
1119 		expires = ts->timer_expires;
1120 	else if (can_stop_idle_tick(cpu, ts))
1121 		expires = tick_nohz_next_event(ts, cpu);
1122 	else
1123 		return;
1124 
1125 	ts->idle_calls++;
1126 
1127 	if (expires > 0LL) {
1128 		int was_stopped = ts->tick_stopped;
1129 
1130 		tick_nohz_stop_tick(ts, cpu);
1131 
1132 		ts->idle_sleeps++;
1133 		ts->idle_expires = expires;
1134 
1135 		if (!was_stopped && ts->tick_stopped) {
1136 			ts->idle_jiffies = ts->last_jiffies;
1137 			nohz_balance_enter_idle(cpu);
1138 		}
1139 	} else {
1140 		tick_nohz_retain_tick(ts);
1141 	}
1142 }
1143 
1144 void tick_nohz_idle_retain_tick(void)
1145 {
1146 	tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1147 	/*
1148 	 * Undo the effect of get_next_timer_interrupt() called from
1149 	 * tick_nohz_next_event().
1150 	 */
1151 	timer_clear_idle();
1152 }
1153 
1154 /**
1155  * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1156  *
1157  * Called when we start the idle loop.
1158  */
1159 void tick_nohz_idle_enter(void)
1160 {
1161 	struct tick_sched *ts;
1162 
1163 	lockdep_assert_irqs_enabled();
1164 
1165 	local_irq_disable();
1166 
1167 	ts = this_cpu_ptr(&tick_cpu_sched);
1168 
1169 	WARN_ON_ONCE(ts->timer_expires_base);
1170 
1171 	ts->inidle = 1;
1172 	tick_nohz_start_idle(ts);
1173 
1174 	local_irq_enable();
1175 }
1176 
1177 /**
1178  * tick_nohz_irq_exit - Notify the tick about IRQ exit
1179  *
1180  * A timer may have been added/modified/deleted either by the current IRQ,
1181  * or by another place using this IRQ as a notification. This IRQ may have
1182  * also updated the RCU callback list. These events may require a
1183  * re-evaluation of the next tick. Depending on the context:
1184  *
1185  * 1) If the CPU is idle and no resched is pending, just proceed with idle
1186  *    time accounting. The next tick will be re-evaluated on the next idle
1187  *    loop iteration.
1188  *
1189  * 2) If the CPU is nohz_full:
1190  *
1191  *    2.1) If there is any tick dependency, restart the tick if stopped.
1192  *
1193  *    2.2) If there is no tick dependency, (re-)evaluate the next tick and
1194  *         stop/update it accordingly.
1195  */
1196 void tick_nohz_irq_exit(void)
1197 {
1198 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1199 
1200 	if (ts->inidle)
1201 		tick_nohz_start_idle(ts);
1202 	else
1203 		tick_nohz_full_update_tick(ts);
1204 }
1205 
1206 /**
1207  * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1208  */
1209 bool tick_nohz_idle_got_tick(void)
1210 {
1211 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1212 
1213 	if (ts->got_idle_tick) {
1214 		ts->got_idle_tick = 0;
1215 		return true;
1216 	}
1217 	return false;
1218 }
1219 
1220 /**
1221  * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1222  * or the tick, whichever expires first. Note that, if the tick has been
1223  * stopped, it returns the next hrtimer.
1224  *
1225  * Called from power state control code with interrupts disabled
1226  */
1227 ktime_t tick_nohz_get_next_hrtimer(void)
1228 {
1229 	return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1230 }
1231 
1232 /**
1233  * tick_nohz_get_sleep_length - return the expected length of the current sleep
1234  * @delta_next: duration until the next event if the tick cannot be stopped
1235  *
1236  * Called from power state control code with interrupts disabled.
1237  *
1238  * The return value of this function and/or the value returned by it through the
1239  * @delta_next pointer can be negative which must be taken into account by its
1240  * callers.
1241  */
1242 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1243 {
1244 	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1245 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1246 	int cpu = smp_processor_id();
1247 	/*
1248 	 * The idle entry time is expected to be a sufficient approximation of
1249 	 * the current time at this point.
1250 	 */
1251 	ktime_t now = ts->idle_entrytime;
1252 	ktime_t next_event;
1253 
1254 	WARN_ON_ONCE(!ts->inidle);
1255 
1256 	*delta_next = ktime_sub(dev->next_event, now);
1257 
1258 	if (!can_stop_idle_tick(cpu, ts))
1259 		return *delta_next;
1260 
1261 	next_event = tick_nohz_next_event(ts, cpu);
1262 	if (!next_event)
1263 		return *delta_next;
1264 
1265 	/*
1266 	 * If the next highres timer to expire is earlier than 'next_event', the
1267 	 * idle governor needs to know that.
1268 	 */
1269 	next_event = min_t(u64, next_event,
1270 			   hrtimer_next_event_without(&ts->sched_timer));
1271 
1272 	return ktime_sub(next_event, now);
1273 }
1274 
1275 /**
1276  * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1277  * for a particular CPU.
1278  *
1279  * Called from the schedutil frequency scaling governor in scheduler context.
1280  */
1281 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1282 {
1283 	struct tick_sched *ts = tick_get_tick_sched(cpu);
1284 
1285 	return ts->idle_calls;
1286 }
1287 
1288 /**
1289  * tick_nohz_get_idle_calls - return the current idle calls counter value
1290  *
1291  * Called from the schedutil frequency scaling governor in scheduler context.
1292  */
1293 unsigned long tick_nohz_get_idle_calls(void)
1294 {
1295 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1296 
1297 	return ts->idle_calls;
1298 }
1299 
1300 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1301 					ktime_t now)
1302 {
1303 	unsigned long ticks;
1304 
1305 	ts->idle_exittime = now;
1306 
1307 	if (vtime_accounting_enabled_this_cpu())
1308 		return;
1309 	/*
1310 	 * We stopped the tick in idle. update_process_times() would miss the
1311 	 * time we slept, as it does only a 1 tick accounting.
1312 	 * Enforce that this is accounted to idle !
1313 	 */
1314 	ticks = jiffies - ts->idle_jiffies;
1315 	/*
1316 	 * We might be one off. Do not randomly account a huge number of ticks!
1317 	 */
1318 	if (ticks && ticks < LONG_MAX)
1319 		account_idle_ticks(ticks);
1320 }
1321 
1322 void tick_nohz_idle_restart_tick(void)
1323 {
1324 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1325 
1326 	if (ts->tick_stopped) {
1327 		ktime_t now = ktime_get();
1328 		tick_nohz_restart_sched_tick(ts, now);
1329 		tick_nohz_account_idle_time(ts, now);
1330 	}
1331 }
1332 
1333 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1334 {
1335 	if (tick_nohz_full_cpu(smp_processor_id()))
1336 		__tick_nohz_full_update_tick(ts, now);
1337 	else
1338 		tick_nohz_restart_sched_tick(ts, now);
1339 
1340 	tick_nohz_account_idle_time(ts, now);
1341 }
1342 
1343 /**
1344  * tick_nohz_idle_exit - Update the tick upon idle task exit
1345  *
1346  * When the idle task exits, update the tick depending on the
1347  * following situations:
1348  *
1349  * 1) If the CPU is not in nohz_full mode (most cases), then
1350  *    restart the tick.
1351  *
1352  * 2) If the CPU is in nohz_full mode (corner case):
1353  *   2.1) If the tick can be kept stopped (no tick dependencies)
1354  *        then re-evaluate the next tick and try to keep it stopped
1355  *        as long as possible.
1356  *   2.2) If the tick has dependencies, restart the tick.
1357  *
1358  */
1359 void tick_nohz_idle_exit(void)
1360 {
1361 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1362 	bool idle_active, tick_stopped;
1363 	ktime_t now;
1364 
1365 	local_irq_disable();
1366 
1367 	WARN_ON_ONCE(!ts->inidle);
1368 	WARN_ON_ONCE(ts->timer_expires_base);
1369 
1370 	ts->inidle = 0;
1371 	idle_active = ts->idle_active;
1372 	tick_stopped = ts->tick_stopped;
1373 
1374 	if (idle_active || tick_stopped)
1375 		now = ktime_get();
1376 
1377 	if (idle_active)
1378 		tick_nohz_stop_idle(ts, now);
1379 
1380 	if (tick_stopped)
1381 		tick_nohz_idle_update_tick(ts, now);
1382 
1383 	local_irq_enable();
1384 }
1385 
1386 /*
1387  * In low-resolution mode, the tick handler must be implemented directly
1388  * at the clockevent level. hrtimer can't be used instead, because its
1389  * infrastructure actually relies on the tick itself as a backend in
1390  * low-resolution mode (see hrtimer_run_queues()).
1391  *
1392  * This low-resolution handler still makes use of some hrtimer APIs meanwhile
1393  * for convenience with expiration calculation and forwarding.
1394  */
1395 static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1396 {
1397 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1398 	struct pt_regs *regs = get_irq_regs();
1399 	ktime_t now = ktime_get();
1400 
1401 	dev->next_event = KTIME_MAX;
1402 
1403 	tick_sched_do_timer(ts, now);
1404 	tick_sched_handle(ts, regs);
1405 
1406 	/*
1407 	 * In dynticks mode, tick reprogram is deferred:
1408 	 * - to the idle task if in dynticks-idle
1409 	 * - to IRQ exit if in full-dynticks.
1410 	 */
1411 	if (likely(!ts->tick_stopped)) {
1412 		hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1413 		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1414 	}
1415 
1416 }
1417 
1418 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1419 {
1420 	if (!tick_nohz_enabled)
1421 		return;
1422 	ts->nohz_mode = mode;
1423 	/* One update is enough */
1424 	if (!test_and_set_bit(0, &tick_nohz_active))
1425 		timers_update_nohz();
1426 }
1427 
1428 /**
1429  * tick_nohz_switch_to_nohz - switch to NOHZ mode
1430  */
1431 static void tick_nohz_switch_to_nohz(void)
1432 {
1433 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1434 	ktime_t next;
1435 
1436 	if (!tick_nohz_enabled)
1437 		return;
1438 
1439 	if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1440 		return;
1441 
1442 	/*
1443 	 * Recycle the hrtimer in 'ts', so we can share the
1444 	 * hrtimer_forward_now() function with the highres code.
1445 	 */
1446 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1447 	/* Get the next period */
1448 	next = tick_init_jiffy_update();
1449 
1450 	hrtimer_set_expires(&ts->sched_timer, next);
1451 	hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1452 	tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1453 	tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1454 }
1455 
1456 static inline void tick_nohz_irq_enter(void)
1457 {
1458 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1459 	ktime_t now;
1460 
1461 	if (!ts->idle_active && !ts->tick_stopped)
1462 		return;
1463 	now = ktime_get();
1464 	if (ts->idle_active)
1465 		tick_nohz_stop_idle(ts, now);
1466 	/*
1467 	 * If all CPUs are idle we may need to update a stale jiffies value.
1468 	 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1469 	 * alive but it might be busy looping with interrupts disabled in some
1470 	 * rare case (typically stop machine). So we must make sure we have a
1471 	 * last resort.
1472 	 */
1473 	if (ts->tick_stopped)
1474 		tick_nohz_update_jiffies(now);
1475 }
1476 
1477 #else
1478 
1479 static inline void tick_nohz_switch_to_nohz(void) { }
1480 static inline void tick_nohz_irq_enter(void) { }
1481 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1482 
1483 #endif /* CONFIG_NO_HZ_COMMON */
1484 
1485 /*
1486  * Called from irq_enter() to notify about the possible interruption of idle()
1487  */
1488 void tick_irq_enter(void)
1489 {
1490 	tick_check_oneshot_broadcast_this_cpu();
1491 	tick_nohz_irq_enter();
1492 }
1493 
1494 /*
1495  * High resolution timer specific code
1496  */
1497 #ifdef CONFIG_HIGH_RES_TIMERS
1498 /*
1499  * We rearm the timer until we get disabled by the idle code.
1500  * Called with interrupts disabled.
1501  */
1502 static enum hrtimer_restart tick_nohz_highres_handler(struct hrtimer *timer)
1503 {
1504 	struct tick_sched *ts =
1505 		container_of(timer, struct tick_sched, sched_timer);
1506 	struct pt_regs *regs = get_irq_regs();
1507 	ktime_t now = ktime_get();
1508 
1509 	tick_sched_do_timer(ts, now);
1510 
1511 	/*
1512 	 * Do not call when we are not in IRQ context and have
1513 	 * no valid 'regs' pointer
1514 	 */
1515 	if (regs)
1516 		tick_sched_handle(ts, regs);
1517 	else
1518 		ts->next_tick = 0;
1519 
1520 	/*
1521 	 * In dynticks mode, tick reprogram is deferred:
1522 	 * - to the idle task if in dynticks-idle
1523 	 * - to IRQ exit if in full-dynticks.
1524 	 */
1525 	if (unlikely(ts->tick_stopped))
1526 		return HRTIMER_NORESTART;
1527 
1528 	hrtimer_forward(timer, now, TICK_NSEC);
1529 
1530 	return HRTIMER_RESTART;
1531 }
1532 
1533 static int sched_skew_tick;
1534 
1535 static int __init skew_tick(char *str)
1536 {
1537 	get_option(&str, &sched_skew_tick);
1538 
1539 	return 0;
1540 }
1541 early_param("skew_tick", skew_tick);
1542 
1543 /**
1544  * tick_setup_sched_timer - setup the tick emulation timer
1545  */
1546 void tick_setup_sched_timer(void)
1547 {
1548 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1549 	ktime_t now = ktime_get();
1550 
1551 	/* Emulate tick processing via per-CPU hrtimers: */
1552 	hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1553 	ts->sched_timer.function = tick_nohz_highres_handler;
1554 
1555 	/* Get the next period (per-CPU) */
1556 	hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1557 
1558 	/* Offset the tick to avert 'jiffies_lock' contention. */
1559 	if (sched_skew_tick) {
1560 		u64 offset = TICK_NSEC >> 1;
1561 		do_div(offset, num_possible_cpus());
1562 		offset *= smp_processor_id();
1563 		hrtimer_add_expires_ns(&ts->sched_timer, offset);
1564 	}
1565 
1566 	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1567 	hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1568 	tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1569 }
1570 #endif /* HIGH_RES_TIMERS */
1571 
1572 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1573 void tick_cancel_sched_timer(int cpu)
1574 {
1575 	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1576 
1577 # ifdef CONFIG_HIGH_RES_TIMERS
1578 	if (ts->sched_timer.base)
1579 		hrtimer_cancel(&ts->sched_timer);
1580 # endif
1581 
1582 	memset(ts, 0, sizeof(*ts));
1583 }
1584 #endif
1585 
1586 /*
1587  * Async notification about clocksource changes
1588  */
1589 void tick_clock_notify(void)
1590 {
1591 	int cpu;
1592 
1593 	for_each_possible_cpu(cpu)
1594 		set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1595 }
1596 
1597 /*
1598  * Async notification about clock event changes
1599  */
1600 void tick_oneshot_notify(void)
1601 {
1602 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1603 
1604 	set_bit(0, &ts->check_clocks);
1605 }
1606 
1607 /*
1608  * Check if a change happened, which makes oneshot possible.
1609  *
1610  * Called cyclically from the hrtimer softirq (driven by the timer
1611  * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1612  * mode, because high resolution timers are disabled (either compile
1613  * or runtime). Called with interrupts disabled.
1614  */
1615 int tick_check_oneshot_change(int allow_nohz)
1616 {
1617 	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1618 
1619 	if (!test_and_clear_bit(0, &ts->check_clocks))
1620 		return 0;
1621 
1622 	if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1623 		return 0;
1624 
1625 	if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1626 		return 0;
1627 
1628 	if (!allow_nohz)
1629 		return 1;
1630 
1631 	tick_nohz_switch_to_nohz();
1632 	return 0;
1633 }
1634