Loading...
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 */
39static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40
41struct 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 */
52static ktime_t last_jiffies_update;
53
54/*
55 * Must be called with interrupts disabled !
56 */
57static 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 */
157static 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
174static 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_ONCE(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
220static 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
249cpumask_var_t tick_nohz_full_mask;
250EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
251bool tick_nohz_full_running;
252EXPORT_SYMBOL_GPL(tick_nohz_full_running);
253static atomic_t tick_dep_mask;
254
255static 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
287static 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(¤t->tick_dep_mask))
301 return false;
302
303 if (check_tick_dependency(¤t->signal->tick_dep_mask))
304 return false;
305
306 return true;
307}
308
309static void nohz_full_kick_func(struct irq_work *work)
310{
311 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
312}
313
314static 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 */
323static 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 */
335void 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
343static 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 */
387static 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
400static 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 */
414void tick_nohz_dep_set(enum tick_dep_bits bit)
415{
416 tick_nohz_dep_set_all(&tick_dep_mask, bit);
417}
418
419void 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 */
428void 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}
449EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
450
451void 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}
457EXPORT_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 */
463void 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}
468EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
469
470void 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}
474EXPORT_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 */
480void 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
496void 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 */
506void __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(¤t->tick_dep_mask) ||
517 atomic_read(¤t->signal->tick_dep_mask))
518 tick_nohz_full_kick();
519 }
520}
521
522/* Get the boot-time nohz CPU list from the kernel parameters. */
523void __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
530static int tick_nohz_cpu_down(unsigned int cpu)
531{
532 /*
533 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
534 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
535 * CPUs. It must remain online when nohz full is enabled.
536 */
537 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
538 return -EBUSY;
539 return 0;
540}
541
542void __init tick_nohz_init(void)
543{
544 int cpu, ret;
545
546 if (!tick_nohz_full_running)
547 return;
548
549 /*
550 * Full dynticks uses irq work to drive the tick rescheduling on safe
551 * locking contexts. But then we need irq work to raise its own
552 * interrupts to avoid circular dependency on the tick
553 */
554 if (!arch_irq_work_has_interrupt()) {
555 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
556 cpumask_clear(tick_nohz_full_mask);
557 tick_nohz_full_running = false;
558 return;
559 }
560
561 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
562 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
563 cpu = smp_processor_id();
564
565 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
566 pr_warn("NO_HZ: Clearing %d from nohz_full range "
567 "for timekeeping\n", cpu);
568 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
569 }
570 }
571
572 for_each_cpu(cpu, tick_nohz_full_mask)
573 ct_cpu_track_user(cpu);
574
575 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
576 "kernel/nohz:predown", NULL,
577 tick_nohz_cpu_down);
578 WARN_ON(ret < 0);
579 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
580 cpumask_pr_args(tick_nohz_full_mask));
581}
582#endif
583
584/*
585 * NOHZ - aka dynamic tick functionality
586 */
587#ifdef CONFIG_NO_HZ_COMMON
588/*
589 * NO HZ enabled ?
590 */
591bool tick_nohz_enabled __read_mostly = true;
592unsigned long tick_nohz_active __read_mostly;
593/*
594 * Enable / Disable tickless mode
595 */
596static int __init setup_tick_nohz(char *str)
597{
598 return (kstrtobool(str, &tick_nohz_enabled) == 0);
599}
600
601__setup("nohz=", setup_tick_nohz);
602
603bool tick_nohz_tick_stopped(void)
604{
605 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
606
607 return ts->tick_stopped;
608}
609
610bool tick_nohz_tick_stopped_cpu(int cpu)
611{
612 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
613
614 return ts->tick_stopped;
615}
616
617/**
618 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
619 *
620 * Called from interrupt entry when the CPU was idle
621 *
622 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
623 * must be updated. Otherwise an interrupt handler could use a stale jiffy
624 * value. We do this unconditionally on any CPU, as we don't know whether the
625 * CPU, which has the update task assigned is in a long sleep.
626 */
627static void tick_nohz_update_jiffies(ktime_t now)
628{
629 unsigned long flags;
630
631 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
632
633 local_irq_save(flags);
634 tick_do_update_jiffies64(now);
635 local_irq_restore(flags);
636
637 touch_softlockup_watchdog_sched();
638}
639
640/*
641 * Updates the per-CPU time idle statistics counters
642 */
643static void
644update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
645{
646 ktime_t delta;
647
648 if (ts->idle_active) {
649 delta = ktime_sub(now, ts->idle_entrytime);
650 if (nr_iowait_cpu(cpu) > 0)
651 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
652 else
653 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
654 ts->idle_entrytime = now;
655 }
656
657 if (last_update_time)
658 *last_update_time = ktime_to_us(now);
659
660}
661
662static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
663{
664 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
665 ts->idle_active = 0;
666
667 sched_clock_idle_wakeup_event();
668}
669
670static void tick_nohz_start_idle(struct tick_sched *ts)
671{
672 ts->idle_entrytime = ktime_get();
673 ts->idle_active = 1;
674 sched_clock_idle_sleep_event();
675}
676
677/**
678 * get_cpu_idle_time_us - get the total idle time of a CPU
679 * @cpu: CPU number to query
680 * @last_update_time: variable to store update time in. Do not update
681 * counters if NULL.
682 *
683 * Return the cumulative idle time (since boot) for a given
684 * CPU, in microseconds.
685 *
686 * This time is measured via accounting rather than sampling,
687 * and is as accurate as ktime_get() is.
688 *
689 * This function returns -1 if NOHZ is not enabled.
690 */
691u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
692{
693 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
694 ktime_t now, idle;
695
696 if (!tick_nohz_active)
697 return -1;
698
699 now = ktime_get();
700 if (last_update_time) {
701 update_ts_time_stats(cpu, ts, now, last_update_time);
702 idle = ts->idle_sleeptime;
703 } else {
704 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
705 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
706
707 idle = ktime_add(ts->idle_sleeptime, delta);
708 } else {
709 idle = ts->idle_sleeptime;
710 }
711 }
712
713 return ktime_to_us(idle);
714
715}
716EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
717
718/**
719 * get_cpu_iowait_time_us - get the total iowait time of a CPU
720 * @cpu: CPU number to query
721 * @last_update_time: variable to store update time in. Do not update
722 * counters if NULL.
723 *
724 * Return the cumulative iowait time (since boot) for a given
725 * CPU, in microseconds.
726 *
727 * This time is measured via accounting rather than sampling,
728 * and is as accurate as ktime_get() is.
729 *
730 * This function returns -1 if NOHZ is not enabled.
731 */
732u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
733{
734 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
735 ktime_t now, iowait;
736
737 if (!tick_nohz_active)
738 return -1;
739
740 now = ktime_get();
741 if (last_update_time) {
742 update_ts_time_stats(cpu, ts, now, last_update_time);
743 iowait = ts->iowait_sleeptime;
744 } else {
745 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
746 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
747
748 iowait = ktime_add(ts->iowait_sleeptime, delta);
749 } else {
750 iowait = ts->iowait_sleeptime;
751 }
752 }
753
754 return ktime_to_us(iowait);
755}
756EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
757
758static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
759{
760 hrtimer_cancel(&ts->sched_timer);
761 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
762
763 /* Forward the time to expire in the future */
764 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
765
766 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
767 hrtimer_start_expires(&ts->sched_timer,
768 HRTIMER_MODE_ABS_PINNED_HARD);
769 } else {
770 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
771 }
772
773 /*
774 * Reset to make sure next tick stop doesn't get fooled by past
775 * cached clock deadline.
776 */
777 ts->next_tick = 0;
778}
779
780static inline bool local_timer_softirq_pending(void)
781{
782 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
783}
784
785static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
786{
787 u64 basemono, next_tick, delta, expires;
788 unsigned long basejiff;
789 unsigned int seq;
790
791 /* Read jiffies and the time when jiffies were updated last */
792 do {
793 seq = read_seqcount_begin(&jiffies_seq);
794 basemono = last_jiffies_update;
795 basejiff = jiffies;
796 } while (read_seqcount_retry(&jiffies_seq, seq));
797 ts->last_jiffies = basejiff;
798 ts->timer_expires_base = basemono;
799
800 /*
801 * Keep the periodic tick, when RCU, architecture or irq_work
802 * requests it.
803 * Aside of that check whether the local timer softirq is
804 * pending. If so its a bad idea to call get_next_timer_interrupt()
805 * because there is an already expired timer, so it will request
806 * immediate expiry, which rearms the hardware timer with a
807 * minimal delta which brings us back to this place
808 * immediately. Lather, rinse and repeat...
809 */
810 if (rcu_needs_cpu() || arch_needs_cpu() ||
811 irq_work_needs_cpu() || local_timer_softirq_pending()) {
812 next_tick = basemono + TICK_NSEC;
813 } else {
814 /*
815 * Get the next pending timer. If high resolution
816 * timers are enabled this only takes the timer wheel
817 * timers into account. If high resolution timers are
818 * disabled this also looks at the next expiring
819 * hrtimer.
820 */
821 next_tick = get_next_timer_interrupt(basejiff, basemono);
822 ts->next_timer = next_tick;
823 }
824
825 /*
826 * If the tick is due in the next period, keep it ticking or
827 * force prod the timer.
828 */
829 delta = next_tick - basemono;
830 if (delta <= (u64)TICK_NSEC) {
831 /*
832 * Tell the timer code that the base is not idle, i.e. undo
833 * the effect of get_next_timer_interrupt():
834 */
835 timer_clear_idle();
836 /*
837 * We've not stopped the tick yet, and there's a timer in the
838 * next period, so no point in stopping it either, bail.
839 */
840 if (!ts->tick_stopped) {
841 ts->timer_expires = 0;
842 goto out;
843 }
844 }
845
846 /*
847 * If this CPU is the one which had the do_timer() duty last, we limit
848 * the sleep time to the timekeeping max_deferment value.
849 * Otherwise we can sleep as long as we want.
850 */
851 delta = timekeeping_max_deferment();
852 if (cpu != tick_do_timer_cpu &&
853 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
854 delta = KTIME_MAX;
855
856 /* Calculate the next expiry time */
857 if (delta < (KTIME_MAX - basemono))
858 expires = basemono + delta;
859 else
860 expires = KTIME_MAX;
861
862 ts->timer_expires = min_t(u64, expires, next_tick);
863
864out:
865 return ts->timer_expires;
866}
867
868static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
869{
870 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
871 u64 basemono = ts->timer_expires_base;
872 u64 expires = ts->timer_expires;
873 ktime_t tick = expires;
874
875 /* Make sure we won't be trying to stop it twice in a row. */
876 ts->timer_expires_base = 0;
877
878 /*
879 * If this CPU is the one which updates jiffies, then give up
880 * the assignment and let it be taken by the CPU which runs
881 * the tick timer next, which might be this CPU as well. If we
882 * don't drop this here the jiffies might be stale and
883 * do_timer() never invoked. Keep track of the fact that it
884 * was the one which had the do_timer() duty last.
885 */
886 if (cpu == tick_do_timer_cpu) {
887 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
888 ts->do_timer_last = 1;
889 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
890 ts->do_timer_last = 0;
891 }
892
893 /* Skip reprogram of event if its not changed */
894 if (ts->tick_stopped && (expires == ts->next_tick)) {
895 /* Sanity check: make sure clockevent is actually programmed */
896 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
897 return;
898
899 WARN_ON_ONCE(1);
900 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
901 basemono, ts->next_tick, dev->next_event,
902 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
903 }
904
905 /*
906 * nohz_stop_sched_tick can be called several times before
907 * the nohz_restart_sched_tick is called. This happens when
908 * interrupts arrive which do not cause a reschedule. In the
909 * first call we save the current tick time, so we can restart
910 * the scheduler tick in nohz_restart_sched_tick.
911 */
912 if (!ts->tick_stopped) {
913 calc_load_nohz_start();
914 quiet_vmstat();
915
916 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
917 ts->tick_stopped = 1;
918 trace_tick_stop(1, TICK_DEP_MASK_NONE);
919 }
920
921 ts->next_tick = tick;
922
923 /*
924 * If the expiration time == KTIME_MAX, then we simply stop
925 * the tick timer.
926 */
927 if (unlikely(expires == KTIME_MAX)) {
928 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
929 hrtimer_cancel(&ts->sched_timer);
930 else
931 tick_program_event(KTIME_MAX, 1);
932 return;
933 }
934
935 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
936 hrtimer_start(&ts->sched_timer, tick,
937 HRTIMER_MODE_ABS_PINNED_HARD);
938 } else {
939 hrtimer_set_expires(&ts->sched_timer, tick);
940 tick_program_event(tick, 1);
941 }
942}
943
944static void tick_nohz_retain_tick(struct tick_sched *ts)
945{
946 ts->timer_expires_base = 0;
947}
948
949#ifdef CONFIG_NO_HZ_FULL
950static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
951{
952 if (tick_nohz_next_event(ts, cpu))
953 tick_nohz_stop_tick(ts, cpu);
954 else
955 tick_nohz_retain_tick(ts);
956}
957#endif /* CONFIG_NO_HZ_FULL */
958
959static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
960{
961 /* Update jiffies first */
962 tick_do_update_jiffies64(now);
963
964 /*
965 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
966 * the clock forward checks in the enqueue path:
967 */
968 timer_clear_idle();
969
970 calc_load_nohz_stop();
971 touch_softlockup_watchdog_sched();
972 /*
973 * Cancel the scheduled timer and restore the tick
974 */
975 ts->tick_stopped = 0;
976 tick_nohz_restart(ts, now);
977}
978
979static void __tick_nohz_full_update_tick(struct tick_sched *ts,
980 ktime_t now)
981{
982#ifdef CONFIG_NO_HZ_FULL
983 int cpu = smp_processor_id();
984
985 if (can_stop_full_tick(cpu, ts))
986 tick_nohz_stop_sched_tick(ts, cpu);
987 else if (ts->tick_stopped)
988 tick_nohz_restart_sched_tick(ts, now);
989#endif
990}
991
992static void tick_nohz_full_update_tick(struct tick_sched *ts)
993{
994 if (!tick_nohz_full_cpu(smp_processor_id()))
995 return;
996
997 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
998 return;
999
1000 __tick_nohz_full_update_tick(ts, ktime_get());
1001}
1002
1003/*
1004 * A pending softirq outside an IRQ (or softirq disabled section) context
1005 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1006 * reach here due to the need_resched() early check in can_stop_idle_tick().
1007 *
1008 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1009 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1010 * triggering the below since wakep_softirqd() is ignored.
1011 *
1012 */
1013static bool report_idle_softirq(void)
1014{
1015 static int ratelimit;
1016 unsigned int pending = local_softirq_pending();
1017
1018 if (likely(!pending))
1019 return false;
1020
1021 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1022 if (!cpu_active(smp_processor_id())) {
1023 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1024 if (!pending)
1025 return false;
1026 }
1027
1028 if (ratelimit < 10)
1029 return false;
1030
1031 /* On RT, softirqs handling may be waiting on some lock */
1032 if (!local_bh_blocked())
1033 return false;
1034
1035 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1036 pending);
1037 ratelimit++;
1038
1039 return true;
1040}
1041
1042static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1043{
1044 /*
1045 * If this CPU is offline and it is the one which updates
1046 * jiffies, then give up the assignment and let it be taken by
1047 * the CPU which runs the tick timer next. If we don't drop
1048 * this here the jiffies might be stale and do_timer() never
1049 * invoked.
1050 */
1051 if (unlikely(!cpu_online(cpu))) {
1052 if (cpu == tick_do_timer_cpu)
1053 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1054 /*
1055 * Make sure the CPU doesn't get fooled by obsolete tick
1056 * deadline if it comes back online later.
1057 */
1058 ts->next_tick = 0;
1059 return false;
1060 }
1061
1062 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1063 return false;
1064
1065 if (need_resched())
1066 return false;
1067
1068 if (unlikely(report_idle_softirq()))
1069 return false;
1070
1071 if (tick_nohz_full_enabled()) {
1072 /*
1073 * Keep the tick alive to guarantee timekeeping progression
1074 * if there are full dynticks CPUs around
1075 */
1076 if (tick_do_timer_cpu == cpu)
1077 return false;
1078
1079 /* Should not happen for nohz-full */
1080 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1081 return false;
1082 }
1083
1084 return true;
1085}
1086
1087static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1088{
1089 ktime_t expires;
1090 int cpu = smp_processor_id();
1091
1092 /*
1093 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1094 * tick timer expiration time is known already.
1095 */
1096 if (ts->timer_expires_base)
1097 expires = ts->timer_expires;
1098 else if (can_stop_idle_tick(cpu, ts))
1099 expires = tick_nohz_next_event(ts, cpu);
1100 else
1101 return;
1102
1103 ts->idle_calls++;
1104
1105 if (expires > 0LL) {
1106 int was_stopped = ts->tick_stopped;
1107
1108 tick_nohz_stop_tick(ts, cpu);
1109
1110 ts->idle_sleeps++;
1111 ts->idle_expires = expires;
1112
1113 if (!was_stopped && ts->tick_stopped) {
1114 ts->idle_jiffies = ts->last_jiffies;
1115 nohz_balance_enter_idle(cpu);
1116 }
1117 } else {
1118 tick_nohz_retain_tick(ts);
1119 }
1120}
1121
1122/**
1123 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1124 *
1125 * When the next event is more than a tick into the future, stop the idle tick
1126 */
1127void tick_nohz_idle_stop_tick(void)
1128{
1129 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1130}
1131
1132void tick_nohz_idle_retain_tick(void)
1133{
1134 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1135 /*
1136 * Undo the effect of get_next_timer_interrupt() called from
1137 * tick_nohz_next_event().
1138 */
1139 timer_clear_idle();
1140}
1141
1142/**
1143 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1144 *
1145 * Called when we start the idle loop.
1146 */
1147void tick_nohz_idle_enter(void)
1148{
1149 struct tick_sched *ts;
1150
1151 lockdep_assert_irqs_enabled();
1152
1153 local_irq_disable();
1154
1155 ts = this_cpu_ptr(&tick_cpu_sched);
1156
1157 WARN_ON_ONCE(ts->timer_expires_base);
1158
1159 ts->inidle = 1;
1160 tick_nohz_start_idle(ts);
1161
1162 local_irq_enable();
1163}
1164
1165/**
1166 * tick_nohz_irq_exit - update next tick event from interrupt exit
1167 *
1168 * When an interrupt fires while we are idle and it doesn't cause
1169 * a reschedule, it may still add, modify or delete a timer, enqueue
1170 * an RCU callback, etc...
1171 * So we need to re-calculate and reprogram the next tick event.
1172 */
1173void tick_nohz_irq_exit(void)
1174{
1175 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1176
1177 if (ts->inidle)
1178 tick_nohz_start_idle(ts);
1179 else
1180 tick_nohz_full_update_tick(ts);
1181}
1182
1183/**
1184 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1185 */
1186bool tick_nohz_idle_got_tick(void)
1187{
1188 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1189
1190 if (ts->got_idle_tick) {
1191 ts->got_idle_tick = 0;
1192 return true;
1193 }
1194 return false;
1195}
1196
1197/**
1198 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1199 * or the tick, whatever that expires first. Note that, if the tick has been
1200 * stopped, it returns the next hrtimer.
1201 *
1202 * Called from power state control code with interrupts disabled
1203 */
1204ktime_t tick_nohz_get_next_hrtimer(void)
1205{
1206 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1207}
1208
1209/**
1210 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1211 * @delta_next: duration until the next event if the tick cannot be stopped
1212 *
1213 * Called from power state control code with interrupts disabled.
1214 *
1215 * The return value of this function and/or the value returned by it through the
1216 * @delta_next pointer can be negative which must be taken into account by its
1217 * callers.
1218 */
1219ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1220{
1221 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1222 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 int cpu = smp_processor_id();
1224 /*
1225 * The idle entry time is expected to be a sufficient approximation of
1226 * the current time at this point.
1227 */
1228 ktime_t now = ts->idle_entrytime;
1229 ktime_t next_event;
1230
1231 WARN_ON_ONCE(!ts->inidle);
1232
1233 *delta_next = ktime_sub(dev->next_event, now);
1234
1235 if (!can_stop_idle_tick(cpu, ts))
1236 return *delta_next;
1237
1238 next_event = tick_nohz_next_event(ts, cpu);
1239 if (!next_event)
1240 return *delta_next;
1241
1242 /*
1243 * If the next highres timer to expire is earlier than next_event, the
1244 * idle governor needs to know that.
1245 */
1246 next_event = min_t(u64, next_event,
1247 hrtimer_next_event_without(&ts->sched_timer));
1248
1249 return ktime_sub(next_event, now);
1250}
1251
1252/**
1253 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1254 * for a particular CPU.
1255 *
1256 * Called from the schedutil frequency scaling governor in scheduler context.
1257 */
1258unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1259{
1260 struct tick_sched *ts = tick_get_tick_sched(cpu);
1261
1262 return ts->idle_calls;
1263}
1264
1265/**
1266 * tick_nohz_get_idle_calls - return the current idle calls counter value
1267 *
1268 * Called from the schedutil frequency scaling governor in scheduler context.
1269 */
1270unsigned long tick_nohz_get_idle_calls(void)
1271{
1272 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1273
1274 return ts->idle_calls;
1275}
1276
1277static void tick_nohz_account_idle_time(struct tick_sched *ts,
1278 ktime_t now)
1279{
1280 unsigned long ticks;
1281
1282 ts->idle_exittime = now;
1283
1284 if (vtime_accounting_enabled_this_cpu())
1285 return;
1286 /*
1287 * We stopped the tick in idle. Update process times would miss the
1288 * time we slept as update_process_times does only a 1 tick
1289 * accounting. Enforce that this is accounted to idle !
1290 */
1291 ticks = jiffies - ts->idle_jiffies;
1292 /*
1293 * We might be one off. Do not randomly account a huge number of ticks!
1294 */
1295 if (ticks && ticks < LONG_MAX)
1296 account_idle_ticks(ticks);
1297}
1298
1299void tick_nohz_idle_restart_tick(void)
1300{
1301 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1302
1303 if (ts->tick_stopped) {
1304 ktime_t now = ktime_get();
1305 tick_nohz_restart_sched_tick(ts, now);
1306 tick_nohz_account_idle_time(ts, now);
1307 }
1308}
1309
1310static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1311{
1312 if (tick_nohz_full_cpu(smp_processor_id()))
1313 __tick_nohz_full_update_tick(ts, now);
1314 else
1315 tick_nohz_restart_sched_tick(ts, now);
1316
1317 tick_nohz_account_idle_time(ts, now);
1318}
1319
1320/**
1321 * tick_nohz_idle_exit - restart the idle tick from the idle task
1322 *
1323 * Restart the idle tick when the CPU is woken up from idle
1324 * This also exit the RCU extended quiescent state. The CPU
1325 * can use RCU again after this function is called.
1326 */
1327void tick_nohz_idle_exit(void)
1328{
1329 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1330 bool idle_active, tick_stopped;
1331 ktime_t now;
1332
1333 local_irq_disable();
1334
1335 WARN_ON_ONCE(!ts->inidle);
1336 WARN_ON_ONCE(ts->timer_expires_base);
1337
1338 ts->inidle = 0;
1339 idle_active = ts->idle_active;
1340 tick_stopped = ts->tick_stopped;
1341
1342 if (idle_active || tick_stopped)
1343 now = ktime_get();
1344
1345 if (idle_active)
1346 tick_nohz_stop_idle(ts, now);
1347
1348 if (tick_stopped)
1349 tick_nohz_idle_update_tick(ts, now);
1350
1351 local_irq_enable();
1352}
1353
1354/*
1355 * The nohz low res interrupt handler
1356 */
1357static void tick_nohz_handler(struct clock_event_device *dev)
1358{
1359 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1360 struct pt_regs *regs = get_irq_regs();
1361 ktime_t now = ktime_get();
1362
1363 dev->next_event = KTIME_MAX;
1364
1365 tick_sched_do_timer(ts, now);
1366 tick_sched_handle(ts, regs);
1367
1368 if (unlikely(ts->tick_stopped)) {
1369 /*
1370 * The clockevent device is not reprogrammed, so change the
1371 * clock event device to ONESHOT_STOPPED to avoid spurious
1372 * interrupts on devices which might not be truly one shot.
1373 */
1374 tick_program_event(KTIME_MAX, 1);
1375 return;
1376 }
1377
1378 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1379 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1380}
1381
1382static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1383{
1384 if (!tick_nohz_enabled)
1385 return;
1386 ts->nohz_mode = mode;
1387 /* One update is enough */
1388 if (!test_and_set_bit(0, &tick_nohz_active))
1389 timers_update_nohz();
1390}
1391
1392/**
1393 * tick_nohz_switch_to_nohz - switch to nohz mode
1394 */
1395static void tick_nohz_switch_to_nohz(void)
1396{
1397 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1398 ktime_t next;
1399
1400 if (!tick_nohz_enabled)
1401 return;
1402
1403 if (tick_switch_to_oneshot(tick_nohz_handler))
1404 return;
1405
1406 /*
1407 * Recycle the hrtimer in ts, so we can share the
1408 * hrtimer_forward with the highres code.
1409 */
1410 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1411 /* Get the next period */
1412 next = tick_init_jiffy_update();
1413
1414 hrtimer_set_expires(&ts->sched_timer, next);
1415 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1418}
1419
1420static inline void tick_nohz_irq_enter(void)
1421{
1422 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1423 ktime_t now;
1424
1425 if (!ts->idle_active && !ts->tick_stopped)
1426 return;
1427 now = ktime_get();
1428 if (ts->idle_active)
1429 tick_nohz_stop_idle(ts, now);
1430 /*
1431 * If all CPUs are idle. We may need to update a stale jiffies value.
1432 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1433 * alive but it might be busy looping with interrupts disabled in some
1434 * rare case (typically stop machine). So we must make sure we have a
1435 * last resort.
1436 */
1437 if (ts->tick_stopped)
1438 tick_nohz_update_jiffies(now);
1439}
1440
1441#else
1442
1443static inline void tick_nohz_switch_to_nohz(void) { }
1444static inline void tick_nohz_irq_enter(void) { }
1445static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1446
1447#endif /* CONFIG_NO_HZ_COMMON */
1448
1449/*
1450 * Called from irq_enter to notify about the possible interruption of idle()
1451 */
1452void tick_irq_enter(void)
1453{
1454 tick_check_oneshot_broadcast_this_cpu();
1455 tick_nohz_irq_enter();
1456}
1457
1458/*
1459 * High resolution timer specific code
1460 */
1461#ifdef CONFIG_HIGH_RES_TIMERS
1462/*
1463 * We rearm the timer until we get disabled by the idle code.
1464 * Called with interrupts disabled.
1465 */
1466static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1467{
1468 struct tick_sched *ts =
1469 container_of(timer, struct tick_sched, sched_timer);
1470 struct pt_regs *regs = get_irq_regs();
1471 ktime_t now = ktime_get();
1472
1473 tick_sched_do_timer(ts, now);
1474
1475 /*
1476 * Do not call, when we are not in irq context and have
1477 * no valid regs pointer
1478 */
1479 if (regs)
1480 tick_sched_handle(ts, regs);
1481 else
1482 ts->next_tick = 0;
1483
1484 /* No need to reprogram if we are in idle or full dynticks mode */
1485 if (unlikely(ts->tick_stopped))
1486 return HRTIMER_NORESTART;
1487
1488 hrtimer_forward(timer, now, TICK_NSEC);
1489
1490 return HRTIMER_RESTART;
1491}
1492
1493static int sched_skew_tick;
1494
1495static int __init skew_tick(char *str)
1496{
1497 get_option(&str, &sched_skew_tick);
1498
1499 return 0;
1500}
1501early_param("skew_tick", skew_tick);
1502
1503/**
1504 * tick_setup_sched_timer - setup the tick emulation timer
1505 */
1506void tick_setup_sched_timer(void)
1507{
1508 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1509 ktime_t now = ktime_get();
1510
1511 /*
1512 * Emulate tick processing via per-CPU hrtimers:
1513 */
1514 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1515 ts->sched_timer.function = tick_sched_timer;
1516
1517 /* Get the next period (per-CPU) */
1518 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1519
1520 /* Offset the tick to avert jiffies_lock contention. */
1521 if (sched_skew_tick) {
1522 u64 offset = TICK_NSEC >> 1;
1523 do_div(offset, num_possible_cpus());
1524 offset *= smp_processor_id();
1525 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1526 }
1527
1528 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1529 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1530 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1531}
1532#endif /* HIGH_RES_TIMERS */
1533
1534#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1535void tick_cancel_sched_timer(int cpu)
1536{
1537 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1538
1539# ifdef CONFIG_HIGH_RES_TIMERS
1540 if (ts->sched_timer.base)
1541 hrtimer_cancel(&ts->sched_timer);
1542# endif
1543
1544 memset(ts, 0, sizeof(*ts));
1545}
1546#endif
1547
1548/*
1549 * Async notification about clocksource changes
1550 */
1551void tick_clock_notify(void)
1552{
1553 int cpu;
1554
1555 for_each_possible_cpu(cpu)
1556 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1557}
1558
1559/*
1560 * Async notification about clock event changes
1561 */
1562void tick_oneshot_notify(void)
1563{
1564 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1565
1566 set_bit(0, &ts->check_clocks);
1567}
1568
1569/*
1570 * Check, if a change happened, which makes oneshot possible.
1571 *
1572 * Called cyclic from the hrtimer softirq (driven by the timer
1573 * softirq) allow_nohz signals, that we can switch into low-res nohz
1574 * mode, because high resolution timers are disabled (either compile
1575 * or runtime). Called with interrupts disabled.
1576 */
1577int tick_check_oneshot_change(int allow_nohz)
1578{
1579 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1580
1581 if (!test_and_clear_bit(0, &ts->check_clocks))
1582 return 0;
1583
1584 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1585 return 0;
1586
1587 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1588 return 0;
1589
1590 if (!allow_nohz)
1591 return 1;
1592
1593 tick_nohz_switch_to_nohz();
1594 return 0;
1595}
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 */
39static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40
41struct 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 */
52static ktime_t last_jiffies_update;
53
54/*
55 * Must be called with interrupts disabled !
56 */
57static 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 */
155static 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
186static 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
232static 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
261cpumask_var_t tick_nohz_full_mask;
262EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
263bool tick_nohz_full_running;
264EXPORT_SYMBOL_GPL(tick_nohz_full_running);
265static atomic_t tick_dep_mask;
266
267static 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
304static 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(¤t->tick_dep_mask))
318 return false;
319
320 if (check_tick_dependency(¤t->signal->tick_dep_mask))
321 return false;
322
323 return true;
324}
325
326static void nohz_full_kick_func(struct irq_work *work)
327{
328 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
329}
330
331static 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 */
340static 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 */
352void 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
360static 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 */
404static 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
417static 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 */
431void tick_nohz_dep_set(enum tick_dep_bits bit)
432{
433 tick_nohz_dep_set_all(&tick_dep_mask, bit);
434}
435
436void 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 */
445void 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}
466EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
467
468void 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}
474EXPORT_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 */
480void 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}
485EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
486
487void 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}
491EXPORT_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 */
497void 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
513void 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 */
523void __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(¤t->tick_dep_mask) ||
534 atomic_read(¤t->signal->tick_dep_mask))
535 tick_nohz_full_kick();
536 }
537}
538
539/* Get the boot-time nohz CPU list from the kernel parameters. */
540void __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
547bool 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
559static int tick_nohz_cpu_down(unsigned int cpu)
560{
561 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
562}
563
564void __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 */
613bool tick_nohz_enabled __read_mostly = true;
614unsigned long tick_nohz_active __read_mostly;
615/*
616 * Enable / Disable tickless mode
617 */
618static 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
625bool 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
632bool 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 */
649static 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
662static 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
684static 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
694static 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 */
740u64 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}
747EXPORT_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 */
766u64 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}
773EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
774
775static 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
797static inline bool local_timer_softirq_pending(void)
798{
799 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
800}
801
802static 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 /* Make sure next_tick is never before basemono! */
843 if (WARN_ON_ONCE(basemono > next_tick))
844 next_tick = basemono;
845
846 /*
847 * If the tick is due in the next period, keep it ticking or
848 * force prod the timer.
849 */
850 delta = next_tick - basemono;
851 if (delta <= (u64)TICK_NSEC) {
852 /*
853 * Tell the timer code that the base is not idle, i.e. undo
854 * the effect of get_next_timer_interrupt():
855 */
856 timer_clear_idle();
857 /*
858 * We've not stopped the tick yet, and there's a timer in the
859 * next period, so no point in stopping it either, bail.
860 */
861 if (!ts->tick_stopped) {
862 ts->timer_expires = 0;
863 goto out;
864 }
865 }
866
867 /*
868 * If this CPU is the one which had the do_timer() duty last, we limit
869 * the sleep time to the timekeeping 'max_deferment' value.
870 * Otherwise we can sleep as long as we want.
871 */
872 delta = timekeeping_max_deferment();
873 if (cpu != tick_do_timer_cpu &&
874 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
875 delta = KTIME_MAX;
876
877 /* Calculate the next expiry time */
878 if (delta < (KTIME_MAX - basemono))
879 expires = basemono + delta;
880 else
881 expires = KTIME_MAX;
882
883 ts->timer_expires = min_t(u64, expires, next_tick);
884
885out:
886 return ts->timer_expires;
887}
888
889static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
890{
891 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
892 u64 basemono = ts->timer_expires_base;
893 u64 expires = ts->timer_expires;
894
895 /* Make sure we won't be trying to stop it twice in a row. */
896 ts->timer_expires_base = 0;
897
898 /*
899 * If this CPU is the one which updates jiffies, then give up
900 * the assignment and let it be taken by the CPU which runs
901 * the tick timer next, which might be this CPU as well. If we
902 * don't drop this here, the jiffies might be stale and
903 * do_timer() never gets invoked. Keep track of the fact that it
904 * was the one which had the do_timer() duty last.
905 */
906 if (cpu == tick_do_timer_cpu) {
907 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
908 ts->do_timer_last = 1;
909 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
910 ts->do_timer_last = 0;
911 }
912
913 /* Skip reprogram of event if it's not changed */
914 if (ts->tick_stopped && (expires == ts->next_tick)) {
915 /* Sanity check: make sure clockevent is actually programmed */
916 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
917 return;
918
919 WARN_ON_ONCE(1);
920 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
921 basemono, ts->next_tick, dev->next_event,
922 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
923 }
924
925 /*
926 * tick_nohz_stop_tick() can be called several times before
927 * tick_nohz_restart_sched_tick() is called. This happens when
928 * interrupts arrive which do not cause a reschedule. In the first
929 * call we save the current tick time, so we can restart the
930 * scheduler tick in tick_nohz_restart_sched_tick().
931 */
932 if (!ts->tick_stopped) {
933 calc_load_nohz_start();
934 quiet_vmstat();
935
936 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
937 ts->tick_stopped = 1;
938 trace_tick_stop(1, TICK_DEP_MASK_NONE);
939 }
940
941 ts->next_tick = expires;
942
943 /*
944 * If the expiration time == KTIME_MAX, then we simply stop
945 * the tick timer.
946 */
947 if (unlikely(expires == KTIME_MAX)) {
948 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
949 hrtimer_cancel(&ts->sched_timer);
950 else
951 tick_program_event(KTIME_MAX, 1);
952 return;
953 }
954
955 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
956 hrtimer_start(&ts->sched_timer, expires,
957 HRTIMER_MODE_ABS_PINNED_HARD);
958 } else {
959 hrtimer_set_expires(&ts->sched_timer, expires);
960 tick_program_event(expires, 1);
961 }
962}
963
964static void tick_nohz_retain_tick(struct tick_sched *ts)
965{
966 ts->timer_expires_base = 0;
967}
968
969#ifdef CONFIG_NO_HZ_FULL
970static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
971{
972 if (tick_nohz_next_event(ts, cpu))
973 tick_nohz_stop_tick(ts, cpu);
974 else
975 tick_nohz_retain_tick(ts);
976}
977#endif /* CONFIG_NO_HZ_FULL */
978
979static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
980{
981 /* Update jiffies first */
982 tick_do_update_jiffies64(now);
983
984 /*
985 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
986 * the clock forward checks in the enqueue path:
987 */
988 timer_clear_idle();
989
990 calc_load_nohz_stop();
991 touch_softlockup_watchdog_sched();
992
993 /* Cancel the scheduled timer and restore the tick: */
994 ts->tick_stopped = 0;
995 tick_nohz_restart(ts, now);
996}
997
998static void __tick_nohz_full_update_tick(struct tick_sched *ts,
999 ktime_t now)
1000{
1001#ifdef CONFIG_NO_HZ_FULL
1002 int cpu = smp_processor_id();
1003
1004 if (can_stop_full_tick(cpu, ts))
1005 tick_nohz_stop_sched_tick(ts, cpu);
1006 else if (ts->tick_stopped)
1007 tick_nohz_restart_sched_tick(ts, now);
1008#endif
1009}
1010
1011static void tick_nohz_full_update_tick(struct tick_sched *ts)
1012{
1013 if (!tick_nohz_full_cpu(smp_processor_id()))
1014 return;
1015
1016 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1017 return;
1018
1019 __tick_nohz_full_update_tick(ts, ktime_get());
1020}
1021
1022/*
1023 * A pending softirq outside an IRQ (or softirq disabled section) context
1024 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1025 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1026 *
1027 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1028 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1029 * triggering the code below, since wakep_softirqd() is ignored.
1030 *
1031 */
1032static bool report_idle_softirq(void)
1033{
1034 static int ratelimit;
1035 unsigned int pending = local_softirq_pending();
1036
1037 if (likely(!pending))
1038 return false;
1039
1040 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1041 if (!cpu_active(smp_processor_id())) {
1042 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1043 if (!pending)
1044 return false;
1045 }
1046
1047 if (ratelimit >= 10)
1048 return false;
1049
1050 /* On RT, softirq handling may be waiting on some lock */
1051 if (local_bh_blocked())
1052 return false;
1053
1054 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1055 pending);
1056 ratelimit++;
1057
1058 return true;
1059}
1060
1061static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1062{
1063 /*
1064 * If this CPU is offline and it is the one which updates
1065 * jiffies, then give up the assignment and let it be taken by
1066 * the CPU which runs the tick timer next. If we don't drop
1067 * this here, the jiffies might be stale and do_timer() never
1068 * gets invoked.
1069 */
1070 if (unlikely(!cpu_online(cpu))) {
1071 if (cpu == tick_do_timer_cpu)
1072 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1073 /*
1074 * Make sure the CPU doesn't get fooled by obsolete tick
1075 * deadline if it comes back online later.
1076 */
1077 ts->next_tick = 0;
1078 return false;
1079 }
1080
1081 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1082 return false;
1083
1084 if (need_resched())
1085 return false;
1086
1087 if (unlikely(report_idle_softirq()))
1088 return false;
1089
1090 if (tick_nohz_full_enabled()) {
1091 /*
1092 * Keep the tick alive to guarantee timekeeping progression
1093 * if there are full dynticks CPUs around
1094 */
1095 if (tick_do_timer_cpu == cpu)
1096 return false;
1097
1098 /* Should not happen for nohz-full */
1099 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1100 return false;
1101 }
1102
1103 return true;
1104}
1105
1106/**
1107 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1108 *
1109 * When the next event is more than a tick into the future, stop the idle tick
1110 */
1111void tick_nohz_idle_stop_tick(void)
1112{
1113 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1114 int cpu = smp_processor_id();
1115 ktime_t expires;
1116
1117 /*
1118 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1119 * tick timer expiration time is known already.
1120 */
1121 if (ts->timer_expires_base)
1122 expires = ts->timer_expires;
1123 else if (can_stop_idle_tick(cpu, ts))
1124 expires = tick_nohz_next_event(ts, cpu);
1125 else
1126 return;
1127
1128 ts->idle_calls++;
1129
1130 if (expires > 0LL) {
1131 int was_stopped = ts->tick_stopped;
1132
1133 tick_nohz_stop_tick(ts, cpu);
1134
1135 ts->idle_sleeps++;
1136 ts->idle_expires = expires;
1137
1138 if (!was_stopped && ts->tick_stopped) {
1139 ts->idle_jiffies = ts->last_jiffies;
1140 nohz_balance_enter_idle(cpu);
1141 }
1142 } else {
1143 tick_nohz_retain_tick(ts);
1144 }
1145}
1146
1147void tick_nohz_idle_retain_tick(void)
1148{
1149 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1150 /*
1151 * Undo the effect of get_next_timer_interrupt() called from
1152 * tick_nohz_next_event().
1153 */
1154 timer_clear_idle();
1155}
1156
1157/**
1158 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1159 *
1160 * Called when we start the idle loop.
1161 */
1162void tick_nohz_idle_enter(void)
1163{
1164 struct tick_sched *ts;
1165
1166 lockdep_assert_irqs_enabled();
1167
1168 local_irq_disable();
1169
1170 ts = this_cpu_ptr(&tick_cpu_sched);
1171
1172 WARN_ON_ONCE(ts->timer_expires_base);
1173
1174 ts->inidle = 1;
1175 tick_nohz_start_idle(ts);
1176
1177 local_irq_enable();
1178}
1179
1180/**
1181 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1182 *
1183 * A timer may have been added/modified/deleted either by the current IRQ,
1184 * or by another place using this IRQ as a notification. This IRQ may have
1185 * also updated the RCU callback list. These events may require a
1186 * re-evaluation of the next tick. Depending on the context:
1187 *
1188 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1189 * time accounting. The next tick will be re-evaluated on the next idle
1190 * loop iteration.
1191 *
1192 * 2) If the CPU is nohz_full:
1193 *
1194 * 2.1) If there is any tick dependency, restart the tick if stopped.
1195 *
1196 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1197 * stop/update it accordingly.
1198 */
1199void tick_nohz_irq_exit(void)
1200{
1201 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1202
1203 if (ts->inidle)
1204 tick_nohz_start_idle(ts);
1205 else
1206 tick_nohz_full_update_tick(ts);
1207}
1208
1209/**
1210 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1211 */
1212bool tick_nohz_idle_got_tick(void)
1213{
1214 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1215
1216 if (ts->got_idle_tick) {
1217 ts->got_idle_tick = 0;
1218 return true;
1219 }
1220 return false;
1221}
1222
1223/**
1224 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1225 * or the tick, whichever expires first. Note that, if the tick has been
1226 * stopped, it returns the next hrtimer.
1227 *
1228 * Called from power state control code with interrupts disabled
1229 */
1230ktime_t tick_nohz_get_next_hrtimer(void)
1231{
1232 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1233}
1234
1235/**
1236 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1237 * @delta_next: duration until the next event if the tick cannot be stopped
1238 *
1239 * Called from power state control code with interrupts disabled.
1240 *
1241 * The return value of this function and/or the value returned by it through the
1242 * @delta_next pointer can be negative which must be taken into account by its
1243 * callers.
1244 */
1245ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1246{
1247 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1248 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1249 int cpu = smp_processor_id();
1250 /*
1251 * The idle entry time is expected to be a sufficient approximation of
1252 * the current time at this point.
1253 */
1254 ktime_t now = ts->idle_entrytime;
1255 ktime_t next_event;
1256
1257 WARN_ON_ONCE(!ts->inidle);
1258
1259 *delta_next = ktime_sub(dev->next_event, now);
1260
1261 if (!can_stop_idle_tick(cpu, ts))
1262 return *delta_next;
1263
1264 next_event = tick_nohz_next_event(ts, cpu);
1265 if (!next_event)
1266 return *delta_next;
1267
1268 /*
1269 * If the next highres timer to expire is earlier than 'next_event', the
1270 * idle governor needs to know that.
1271 */
1272 next_event = min_t(u64, next_event,
1273 hrtimer_next_event_without(&ts->sched_timer));
1274
1275 return ktime_sub(next_event, now);
1276}
1277
1278/**
1279 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1280 * for a particular CPU.
1281 *
1282 * Called from the schedutil frequency scaling governor in scheduler context.
1283 */
1284unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1285{
1286 struct tick_sched *ts = tick_get_tick_sched(cpu);
1287
1288 return ts->idle_calls;
1289}
1290
1291/**
1292 * tick_nohz_get_idle_calls - return the current idle calls counter value
1293 *
1294 * Called from the schedutil frequency scaling governor in scheduler context.
1295 */
1296unsigned long tick_nohz_get_idle_calls(void)
1297{
1298 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1299
1300 return ts->idle_calls;
1301}
1302
1303static void tick_nohz_account_idle_time(struct tick_sched *ts,
1304 ktime_t now)
1305{
1306 unsigned long ticks;
1307
1308 ts->idle_exittime = now;
1309
1310 if (vtime_accounting_enabled_this_cpu())
1311 return;
1312 /*
1313 * We stopped the tick in idle. update_process_times() would miss the
1314 * time we slept, as it does only a 1 tick accounting.
1315 * Enforce that this is accounted to idle !
1316 */
1317 ticks = jiffies - ts->idle_jiffies;
1318 /*
1319 * We might be one off. Do not randomly account a huge number of ticks!
1320 */
1321 if (ticks && ticks < LONG_MAX)
1322 account_idle_ticks(ticks);
1323}
1324
1325void tick_nohz_idle_restart_tick(void)
1326{
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1328
1329 if (ts->tick_stopped) {
1330 ktime_t now = ktime_get();
1331 tick_nohz_restart_sched_tick(ts, now);
1332 tick_nohz_account_idle_time(ts, now);
1333 }
1334}
1335
1336static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1337{
1338 if (tick_nohz_full_cpu(smp_processor_id()))
1339 __tick_nohz_full_update_tick(ts, now);
1340 else
1341 tick_nohz_restart_sched_tick(ts, now);
1342
1343 tick_nohz_account_idle_time(ts, now);
1344}
1345
1346/**
1347 * tick_nohz_idle_exit - Update the tick upon idle task exit
1348 *
1349 * When the idle task exits, update the tick depending on the
1350 * following situations:
1351 *
1352 * 1) If the CPU is not in nohz_full mode (most cases), then
1353 * restart the tick.
1354 *
1355 * 2) If the CPU is in nohz_full mode (corner case):
1356 * 2.1) If the tick can be kept stopped (no tick dependencies)
1357 * then re-evaluate the next tick and try to keep it stopped
1358 * as long as possible.
1359 * 2.2) If the tick has dependencies, restart the tick.
1360 *
1361 */
1362void tick_nohz_idle_exit(void)
1363{
1364 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1365 bool idle_active, tick_stopped;
1366 ktime_t now;
1367
1368 local_irq_disable();
1369
1370 WARN_ON_ONCE(!ts->inidle);
1371 WARN_ON_ONCE(ts->timer_expires_base);
1372
1373 ts->inidle = 0;
1374 idle_active = ts->idle_active;
1375 tick_stopped = ts->tick_stopped;
1376
1377 if (idle_active || tick_stopped)
1378 now = ktime_get();
1379
1380 if (idle_active)
1381 tick_nohz_stop_idle(ts, now);
1382
1383 if (tick_stopped)
1384 tick_nohz_idle_update_tick(ts, now);
1385
1386 local_irq_enable();
1387}
1388
1389/*
1390 * In low-resolution mode, the tick handler must be implemented directly
1391 * at the clockevent level. hrtimer can't be used instead, because its
1392 * infrastructure actually relies on the tick itself as a backend in
1393 * low-resolution mode (see hrtimer_run_queues()).
1394 *
1395 * This low-resolution handler still makes use of some hrtimer APIs meanwhile
1396 * for convenience with expiration calculation and forwarding.
1397 */
1398static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1399{
1400 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1401 struct pt_regs *regs = get_irq_regs();
1402 ktime_t now = ktime_get();
1403
1404 dev->next_event = KTIME_MAX;
1405
1406 tick_sched_do_timer(ts, now);
1407 tick_sched_handle(ts, regs);
1408
1409 /*
1410 * In dynticks mode, tick reprogram is deferred:
1411 * - to the idle task if in dynticks-idle
1412 * - to IRQ exit if in full-dynticks.
1413 */
1414 if (likely(!ts->tick_stopped)) {
1415 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 }
1418
1419}
1420
1421static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1422{
1423 if (!tick_nohz_enabled)
1424 return;
1425 ts->nohz_mode = mode;
1426 /* One update is enough */
1427 if (!test_and_set_bit(0, &tick_nohz_active))
1428 timers_update_nohz();
1429}
1430
1431/**
1432 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1433 */
1434static void tick_nohz_switch_to_nohz(void)
1435{
1436 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1437 ktime_t next;
1438
1439 if (!tick_nohz_enabled)
1440 return;
1441
1442 if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1443 return;
1444
1445 /*
1446 * Recycle the hrtimer in 'ts', so we can share the
1447 * hrtimer_forward_now() function with the highres code.
1448 */
1449 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1450 /* Get the next period */
1451 next = tick_init_jiffy_update();
1452
1453 hrtimer_set_expires(&ts->sched_timer, next);
1454 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1455 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1456 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1457}
1458
1459static inline void tick_nohz_irq_enter(void)
1460{
1461 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1462 ktime_t now;
1463
1464 if (!ts->idle_active && !ts->tick_stopped)
1465 return;
1466 now = ktime_get();
1467 if (ts->idle_active)
1468 tick_nohz_stop_idle(ts, now);
1469 /*
1470 * If all CPUs are idle we may need to update a stale jiffies value.
1471 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1472 * alive but it might be busy looping with interrupts disabled in some
1473 * rare case (typically stop machine). So we must make sure we have a
1474 * last resort.
1475 */
1476 if (ts->tick_stopped)
1477 tick_nohz_update_jiffies(now);
1478}
1479
1480#else
1481
1482static inline void tick_nohz_switch_to_nohz(void) { }
1483static inline void tick_nohz_irq_enter(void) { }
1484static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1485
1486#endif /* CONFIG_NO_HZ_COMMON */
1487
1488/*
1489 * Called from irq_enter() to notify about the possible interruption of idle()
1490 */
1491void tick_irq_enter(void)
1492{
1493 tick_check_oneshot_broadcast_this_cpu();
1494 tick_nohz_irq_enter();
1495}
1496
1497/*
1498 * High resolution timer specific code
1499 */
1500#ifdef CONFIG_HIGH_RES_TIMERS
1501/*
1502 * We rearm the timer until we get disabled by the idle code.
1503 * Called with interrupts disabled.
1504 */
1505static enum hrtimer_restart tick_nohz_highres_handler(struct hrtimer *timer)
1506{
1507 struct tick_sched *ts =
1508 container_of(timer, struct tick_sched, sched_timer);
1509 struct pt_regs *regs = get_irq_regs();
1510 ktime_t now = ktime_get();
1511
1512 tick_sched_do_timer(ts, now);
1513
1514 /*
1515 * Do not call when we are not in IRQ context and have
1516 * no valid 'regs' pointer
1517 */
1518 if (regs)
1519 tick_sched_handle(ts, regs);
1520 else
1521 ts->next_tick = 0;
1522
1523 /*
1524 * In dynticks mode, tick reprogram is deferred:
1525 * - to the idle task if in dynticks-idle
1526 * - to IRQ exit if in full-dynticks.
1527 */
1528 if (unlikely(ts->tick_stopped))
1529 return HRTIMER_NORESTART;
1530
1531 hrtimer_forward(timer, now, TICK_NSEC);
1532
1533 return HRTIMER_RESTART;
1534}
1535
1536static int sched_skew_tick;
1537
1538static int __init skew_tick(char *str)
1539{
1540 get_option(&str, &sched_skew_tick);
1541
1542 return 0;
1543}
1544early_param("skew_tick", skew_tick);
1545
1546/**
1547 * tick_setup_sched_timer - setup the tick emulation timer
1548 */
1549void tick_setup_sched_timer(void)
1550{
1551 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1552 ktime_t now = ktime_get();
1553
1554 /* Emulate tick processing via per-CPU hrtimers: */
1555 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1556 ts->sched_timer.function = tick_nohz_highres_handler;
1557
1558 /* Get the next period (per-CPU) */
1559 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1560
1561 /* Offset the tick to avert 'jiffies_lock' contention. */
1562 if (sched_skew_tick) {
1563 u64 offset = TICK_NSEC >> 1;
1564 do_div(offset, num_possible_cpus());
1565 offset *= smp_processor_id();
1566 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1567 }
1568
1569 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1570 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1571 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1572}
1573#endif /* HIGH_RES_TIMERS */
1574
1575#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1576void tick_cancel_sched_timer(int cpu)
1577{
1578 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1579 ktime_t idle_sleeptime, iowait_sleeptime;
1580 unsigned long idle_calls, idle_sleeps;
1581
1582# ifdef CONFIG_HIGH_RES_TIMERS
1583 if (ts->sched_timer.base)
1584 hrtimer_cancel(&ts->sched_timer);
1585# endif
1586
1587 idle_sleeptime = ts->idle_sleeptime;
1588 iowait_sleeptime = ts->iowait_sleeptime;
1589 idle_calls = ts->idle_calls;
1590 idle_sleeps = ts->idle_sleeps;
1591 memset(ts, 0, sizeof(*ts));
1592 ts->idle_sleeptime = idle_sleeptime;
1593 ts->iowait_sleeptime = iowait_sleeptime;
1594 ts->idle_calls = idle_calls;
1595 ts->idle_sleeps = idle_sleeps;
1596}
1597#endif
1598
1599/*
1600 * Async notification about clocksource changes
1601 */
1602void tick_clock_notify(void)
1603{
1604 int cpu;
1605
1606 for_each_possible_cpu(cpu)
1607 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1608}
1609
1610/*
1611 * Async notification about clock event changes
1612 */
1613void tick_oneshot_notify(void)
1614{
1615 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1616
1617 set_bit(0, &ts->check_clocks);
1618}
1619
1620/*
1621 * Check if a change happened, which makes oneshot possible.
1622 *
1623 * Called cyclically from the hrtimer softirq (driven by the timer
1624 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1625 * mode, because high resolution timers are disabled (either compile
1626 * or runtime). Called with interrupts disabled.
1627 */
1628int tick_check_oneshot_change(int allow_nohz)
1629{
1630 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1631
1632 if (!test_and_clear_bit(0, &ts->check_clocks))
1633 return 0;
1634
1635 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1636 return 0;
1637
1638 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1639 return 0;
1640
1641 if (!allow_nohz)
1642 return 1;
1643
1644 tick_nohz_switch_to_nohz();
1645 return 0;
1646}