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/module.h>
24#include <linux/irq_work.h>
25#include <linux/posix-timers.h>
26#include <linux/context_tracking.h>
27#include <linux/mm.h>
28
29#include <asm/irq_regs.h>
30
31#include "tick-internal.h"
32
33#include <trace/events/timer.h>
34
35/*
36 * Per-CPU nohz control structure
37 */
38static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
39
40struct tick_sched *tick_get_tick_sched(int cpu)
41{
42 return &per_cpu(tick_cpu_sched, cpu);
43}
44
45#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
46/*
47 * The time, when the last jiffy update happened. Protected by jiffies_lock.
48 */
49static ktime_t last_jiffies_update;
50
51/*
52 * Must be called with interrupts disabled !
53 */
54static void tick_do_update_jiffies64(ktime_t now)
55{
56 unsigned long ticks = 0;
57 ktime_t delta;
58
59 /*
60 * Do a quick check without holding jiffies_lock:
61 */
62 delta = ktime_sub(now, last_jiffies_update);
63 if (delta < tick_period)
64 return;
65
66 /* Reevaluate with jiffies_lock held */
67 write_seqlock(&jiffies_lock);
68
69 delta = ktime_sub(now, last_jiffies_update);
70 if (delta >= tick_period) {
71
72 delta = ktime_sub(delta, tick_period);
73 last_jiffies_update = ktime_add(last_jiffies_update,
74 tick_period);
75
76 /* Slow path for long timeouts */
77 if (unlikely(delta >= tick_period)) {
78 s64 incr = ktime_to_ns(tick_period);
79
80 ticks = ktime_divns(delta, incr);
81
82 last_jiffies_update = ktime_add_ns(last_jiffies_update,
83 incr * ticks);
84 }
85 do_timer(++ticks);
86
87 /* Keep the tick_next_period variable up to date */
88 tick_next_period = ktime_add(last_jiffies_update, tick_period);
89 } else {
90 write_sequnlock(&jiffies_lock);
91 return;
92 }
93 write_sequnlock(&jiffies_lock);
94 update_wall_time();
95}
96
97/*
98 * Initialize and return retrieve the jiffies update.
99 */
100static ktime_t tick_init_jiffy_update(void)
101{
102 ktime_t period;
103
104 write_seqlock(&jiffies_lock);
105 /* Did we start the jiffies update yet ? */
106 if (last_jiffies_update == 0)
107 last_jiffies_update = tick_next_period;
108 period = last_jiffies_update;
109 write_sequnlock(&jiffies_lock);
110 return period;
111}
112
113static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
114{
115 int cpu = smp_processor_id();
116
117#ifdef CONFIG_NO_HZ_COMMON
118 /*
119 * Check if the do_timer duty was dropped. We don't care about
120 * concurrency: This happens only when the CPU in charge went
121 * into a long sleep. If two CPUs happen to assign themselves to
122 * this duty, then the jiffies update is still serialized by
123 * jiffies_lock.
124 *
125 * If nohz_full is enabled, this should not happen because the
126 * tick_do_timer_cpu never relinquishes.
127 */
128 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
129#ifdef CONFIG_NO_HZ_FULL
130 WARN_ON(tick_nohz_full_running);
131#endif
132 tick_do_timer_cpu = cpu;
133 }
134#endif
135
136 /* Check, if the jiffies need an update */
137 if (tick_do_timer_cpu == cpu)
138 tick_do_update_jiffies64(now);
139
140 if (ts->inidle)
141 ts->got_idle_tick = 1;
142}
143
144static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
145{
146#ifdef CONFIG_NO_HZ_COMMON
147 /*
148 * When we are idle and the tick is stopped, we have to touch
149 * the watchdog as we might not schedule for a really long
150 * time. This happens on complete idle SMP systems while
151 * waiting on the login prompt. We also increment the "start of
152 * idle" jiffy stamp so the idle accounting adjustment we do
153 * when we go busy again does not account too much ticks.
154 */
155 if (ts->tick_stopped) {
156 touch_softlockup_watchdog_sched();
157 if (is_idle_task(current))
158 ts->idle_jiffies++;
159 /*
160 * In case the current tick fired too early past its expected
161 * expiration, make sure we don't bypass the next clock reprogramming
162 * to the same deadline.
163 */
164 ts->next_tick = 0;
165 }
166#endif
167 update_process_times(user_mode(regs));
168 profile_tick(CPU_PROFILING);
169}
170#endif
171
172#ifdef CONFIG_NO_HZ_FULL
173cpumask_var_t tick_nohz_full_mask;
174bool tick_nohz_full_running;
175static atomic_t tick_dep_mask;
176
177static bool check_tick_dependency(atomic_t *dep)
178{
179 int val = atomic_read(dep);
180
181 if (val & TICK_DEP_MASK_POSIX_TIMER) {
182 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
183 return true;
184 }
185
186 if (val & TICK_DEP_MASK_PERF_EVENTS) {
187 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
188 return true;
189 }
190
191 if (val & TICK_DEP_MASK_SCHED) {
192 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
193 return true;
194 }
195
196 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
197 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
198 return true;
199 }
200
201 return false;
202}
203
204static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
205{
206 lockdep_assert_irqs_disabled();
207
208 if (unlikely(!cpu_online(cpu)))
209 return false;
210
211 if (check_tick_dependency(&tick_dep_mask))
212 return false;
213
214 if (check_tick_dependency(&ts->tick_dep_mask))
215 return false;
216
217 if (check_tick_dependency(¤t->tick_dep_mask))
218 return false;
219
220 if (check_tick_dependency(¤t->signal->tick_dep_mask))
221 return false;
222
223 return true;
224}
225
226static void nohz_full_kick_func(struct irq_work *work)
227{
228 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
229}
230
231static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
232 .func = nohz_full_kick_func,
233};
234
235/*
236 * Kick this CPU if it's full dynticks in order to force it to
237 * re-evaluate its dependency on the tick and restart it if necessary.
238 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
239 * is NMI safe.
240 */
241static void tick_nohz_full_kick(void)
242{
243 if (!tick_nohz_full_cpu(smp_processor_id()))
244 return;
245
246 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
247}
248
249/*
250 * Kick the CPU if it's full dynticks in order to force it to
251 * re-evaluate its dependency on the tick and restart it if necessary.
252 */
253void tick_nohz_full_kick_cpu(int cpu)
254{
255 if (!tick_nohz_full_cpu(cpu))
256 return;
257
258 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
259}
260
261/*
262 * Kick all full dynticks CPUs in order to force these to re-evaluate
263 * their dependency on the tick and restart it if necessary.
264 */
265static void tick_nohz_full_kick_all(void)
266{
267 int cpu;
268
269 if (!tick_nohz_full_running)
270 return;
271
272 preempt_disable();
273 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
274 tick_nohz_full_kick_cpu(cpu);
275 preempt_enable();
276}
277
278static void tick_nohz_dep_set_all(atomic_t *dep,
279 enum tick_dep_bits bit)
280{
281 int prev;
282
283 prev = atomic_fetch_or(BIT(bit), dep);
284 if (!prev)
285 tick_nohz_full_kick_all();
286}
287
288/*
289 * Set a global tick dependency. Used by perf events that rely on freq and
290 * by unstable clock.
291 */
292void tick_nohz_dep_set(enum tick_dep_bits bit)
293{
294 tick_nohz_dep_set_all(&tick_dep_mask, bit);
295}
296
297void tick_nohz_dep_clear(enum tick_dep_bits bit)
298{
299 atomic_andnot(BIT(bit), &tick_dep_mask);
300}
301
302/*
303 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
304 * manage events throttling.
305 */
306void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
307{
308 int prev;
309 struct tick_sched *ts;
310
311 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
312
313 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
314 if (!prev) {
315 preempt_disable();
316 /* Perf needs local kick that is NMI safe */
317 if (cpu == smp_processor_id()) {
318 tick_nohz_full_kick();
319 } else {
320 /* Remote irq work not NMI-safe */
321 if (!WARN_ON_ONCE(in_nmi()))
322 tick_nohz_full_kick_cpu(cpu);
323 }
324 preempt_enable();
325 }
326}
327
328void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
329{
330 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
331
332 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
333}
334
335/*
336 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
337 * per task timers.
338 */
339void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
340{
341 /*
342 * We could optimize this with just kicking the target running the task
343 * if that noise matters for nohz full users.
344 */
345 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
346}
347
348void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
349{
350 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
351}
352
353/*
354 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
355 * per process timers.
356 */
357void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
358{
359 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
360}
361
362void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
363{
364 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
365}
366
367/*
368 * Re-evaluate the need for the tick as we switch the current task.
369 * It might need the tick due to per task/process properties:
370 * perf events, posix CPU timers, ...
371 */
372void __tick_nohz_task_switch(void)
373{
374 unsigned long flags;
375 struct tick_sched *ts;
376
377 local_irq_save(flags);
378
379 if (!tick_nohz_full_cpu(smp_processor_id()))
380 goto out;
381
382 ts = this_cpu_ptr(&tick_cpu_sched);
383
384 if (ts->tick_stopped) {
385 if (atomic_read(¤t->tick_dep_mask) ||
386 atomic_read(¤t->signal->tick_dep_mask))
387 tick_nohz_full_kick();
388 }
389out:
390 local_irq_restore(flags);
391}
392
393/* Get the boot-time nohz CPU list from the kernel parameters. */
394void __init tick_nohz_full_setup(cpumask_var_t cpumask)
395{
396 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
397 cpumask_copy(tick_nohz_full_mask, cpumask);
398 tick_nohz_full_running = true;
399}
400
401static int tick_nohz_cpu_down(unsigned int cpu)
402{
403 /*
404 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
405 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
406 * CPUs. It must remain online when nohz full is enabled.
407 */
408 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
409 return -EBUSY;
410 return 0;
411}
412
413void __init tick_nohz_init(void)
414{
415 int cpu, ret;
416
417 if (!tick_nohz_full_running)
418 return;
419
420 /*
421 * Full dynticks uses irq work to drive the tick rescheduling on safe
422 * locking contexts. But then we need irq work to raise its own
423 * interrupts to avoid circular dependency on the tick
424 */
425 if (!arch_irq_work_has_interrupt()) {
426 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
427 cpumask_clear(tick_nohz_full_mask);
428 tick_nohz_full_running = false;
429 return;
430 }
431
432 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
433 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
434 cpu = smp_processor_id();
435
436 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
437 pr_warn("NO_HZ: Clearing %d from nohz_full range "
438 "for timekeeping\n", cpu);
439 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
440 }
441 }
442
443 for_each_cpu(cpu, tick_nohz_full_mask)
444 context_tracking_cpu_set(cpu);
445
446 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
447 "kernel/nohz:predown", NULL,
448 tick_nohz_cpu_down);
449 WARN_ON(ret < 0);
450 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
451 cpumask_pr_args(tick_nohz_full_mask));
452}
453#endif
454
455/*
456 * NOHZ - aka dynamic tick functionality
457 */
458#ifdef CONFIG_NO_HZ_COMMON
459/*
460 * NO HZ enabled ?
461 */
462bool tick_nohz_enabled __read_mostly = true;
463unsigned long tick_nohz_active __read_mostly;
464/*
465 * Enable / Disable tickless mode
466 */
467static int __init setup_tick_nohz(char *str)
468{
469 return (kstrtobool(str, &tick_nohz_enabled) == 0);
470}
471
472__setup("nohz=", setup_tick_nohz);
473
474bool tick_nohz_tick_stopped(void)
475{
476 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
477
478 return ts->tick_stopped;
479}
480
481bool tick_nohz_tick_stopped_cpu(int cpu)
482{
483 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
484
485 return ts->tick_stopped;
486}
487
488/**
489 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
490 *
491 * Called from interrupt entry when the CPU was idle
492 *
493 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
494 * must be updated. Otherwise an interrupt handler could use a stale jiffy
495 * value. We do this unconditionally on any CPU, as we don't know whether the
496 * CPU, which has the update task assigned is in a long sleep.
497 */
498static void tick_nohz_update_jiffies(ktime_t now)
499{
500 unsigned long flags;
501
502 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
503
504 local_irq_save(flags);
505 tick_do_update_jiffies64(now);
506 local_irq_restore(flags);
507
508 touch_softlockup_watchdog_sched();
509}
510
511/*
512 * Updates the per-CPU time idle statistics counters
513 */
514static void
515update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
516{
517 ktime_t delta;
518
519 if (ts->idle_active) {
520 delta = ktime_sub(now, ts->idle_entrytime);
521 if (nr_iowait_cpu(cpu) > 0)
522 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
523 else
524 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
525 ts->idle_entrytime = now;
526 }
527
528 if (last_update_time)
529 *last_update_time = ktime_to_us(now);
530
531}
532
533static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
534{
535 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
536 ts->idle_active = 0;
537
538 sched_clock_idle_wakeup_event();
539}
540
541static void tick_nohz_start_idle(struct tick_sched *ts)
542{
543 ts->idle_entrytime = ktime_get();
544 ts->idle_active = 1;
545 sched_clock_idle_sleep_event();
546}
547
548/**
549 * get_cpu_idle_time_us - get the total idle time of a CPU
550 * @cpu: CPU number to query
551 * @last_update_time: variable to store update time in. Do not update
552 * counters if NULL.
553 *
554 * Return the cumulative idle time (since boot) for a given
555 * CPU, in microseconds.
556 *
557 * This time is measured via accounting rather than sampling,
558 * and is as accurate as ktime_get() is.
559 *
560 * This function returns -1 if NOHZ is not enabled.
561 */
562u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
563{
564 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
565 ktime_t now, idle;
566
567 if (!tick_nohz_active)
568 return -1;
569
570 now = ktime_get();
571 if (last_update_time) {
572 update_ts_time_stats(cpu, ts, now, last_update_time);
573 idle = ts->idle_sleeptime;
574 } else {
575 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
576 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
577
578 idle = ktime_add(ts->idle_sleeptime, delta);
579 } else {
580 idle = ts->idle_sleeptime;
581 }
582 }
583
584 return ktime_to_us(idle);
585
586}
587EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
588
589/**
590 * get_cpu_iowait_time_us - get the total iowait time of a CPU
591 * @cpu: CPU number to query
592 * @last_update_time: variable to store update time in. Do not update
593 * counters if NULL.
594 *
595 * Return the cumulative iowait time (since boot) for a given
596 * CPU, in microseconds.
597 *
598 * This time is measured via accounting rather than sampling,
599 * and is as accurate as ktime_get() is.
600 *
601 * This function returns -1 if NOHZ is not enabled.
602 */
603u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
604{
605 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
606 ktime_t now, iowait;
607
608 if (!tick_nohz_active)
609 return -1;
610
611 now = ktime_get();
612 if (last_update_time) {
613 update_ts_time_stats(cpu, ts, now, last_update_time);
614 iowait = ts->iowait_sleeptime;
615 } else {
616 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
617 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
618
619 iowait = ktime_add(ts->iowait_sleeptime, delta);
620 } else {
621 iowait = ts->iowait_sleeptime;
622 }
623 }
624
625 return ktime_to_us(iowait);
626}
627EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
628
629static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
630{
631 hrtimer_cancel(&ts->sched_timer);
632 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
633
634 /* Forward the time to expire in the future */
635 hrtimer_forward(&ts->sched_timer, now, tick_period);
636
637 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
638 hrtimer_start_expires(&ts->sched_timer,
639 HRTIMER_MODE_ABS_PINNED_HARD);
640 } else {
641 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
642 }
643
644 /*
645 * Reset to make sure next tick stop doesn't get fooled by past
646 * cached clock deadline.
647 */
648 ts->next_tick = 0;
649}
650
651static inline bool local_timer_softirq_pending(void)
652{
653 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
654}
655
656static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
657{
658 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
659 unsigned long basejiff;
660 unsigned int seq;
661
662 /* Read jiffies and the time when jiffies were updated last */
663 do {
664 seq = read_seqbegin(&jiffies_lock);
665 basemono = last_jiffies_update;
666 basejiff = jiffies;
667 } while (read_seqretry(&jiffies_lock, seq));
668 ts->last_jiffies = basejiff;
669 ts->timer_expires_base = basemono;
670
671 /*
672 * Keep the periodic tick, when RCU, architecture or irq_work
673 * requests it.
674 * Aside of that check whether the local timer softirq is
675 * pending. If so its a bad idea to call get_next_timer_interrupt()
676 * because there is an already expired timer, so it will request
677 * immeditate expiry, which rearms the hardware timer with a
678 * minimal delta which brings us back to this place
679 * immediately. Lather, rinse and repeat...
680 */
681 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() ||
682 irq_work_needs_cpu() || local_timer_softirq_pending()) {
683 next_tick = basemono + TICK_NSEC;
684 } else {
685 /*
686 * Get the next pending timer. If high resolution
687 * timers are enabled this only takes the timer wheel
688 * timers into account. If high resolution timers are
689 * disabled this also looks at the next expiring
690 * hrtimer.
691 */
692 next_tmr = get_next_timer_interrupt(basejiff, basemono);
693 ts->next_timer = next_tmr;
694 /* Take the next rcu event into account */
695 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
696 }
697
698 /*
699 * If the tick is due in the next period, keep it ticking or
700 * force prod the timer.
701 */
702 delta = next_tick - basemono;
703 if (delta <= (u64)TICK_NSEC) {
704 /*
705 * Tell the timer code that the base is not idle, i.e. undo
706 * the effect of get_next_timer_interrupt():
707 */
708 timer_clear_idle();
709 /*
710 * We've not stopped the tick yet, and there's a timer in the
711 * next period, so no point in stopping it either, bail.
712 */
713 if (!ts->tick_stopped) {
714 ts->timer_expires = 0;
715 goto out;
716 }
717 }
718
719 /*
720 * If this CPU is the one which had the do_timer() duty last, we limit
721 * the sleep time to the timekeeping max_deferment value.
722 * Otherwise we can sleep as long as we want.
723 */
724 delta = timekeeping_max_deferment();
725 if (cpu != tick_do_timer_cpu &&
726 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
727 delta = KTIME_MAX;
728
729 /* Calculate the next expiry time */
730 if (delta < (KTIME_MAX - basemono))
731 expires = basemono + delta;
732 else
733 expires = KTIME_MAX;
734
735 ts->timer_expires = min_t(u64, expires, next_tick);
736
737out:
738 return ts->timer_expires;
739}
740
741static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
742{
743 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
744 u64 basemono = ts->timer_expires_base;
745 u64 expires = ts->timer_expires;
746 ktime_t tick = expires;
747
748 /* Make sure we won't be trying to stop it twice in a row. */
749 ts->timer_expires_base = 0;
750
751 /*
752 * If this CPU is the one which updates jiffies, then give up
753 * the assignment and let it be taken by the CPU which runs
754 * the tick timer next, which might be this CPU as well. If we
755 * don't drop this here the jiffies might be stale and
756 * do_timer() never invoked. Keep track of the fact that it
757 * was the one which had the do_timer() duty last.
758 */
759 if (cpu == tick_do_timer_cpu) {
760 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
761 ts->do_timer_last = 1;
762 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
763 ts->do_timer_last = 0;
764 }
765
766 /* Skip reprogram of event if its not changed */
767 if (ts->tick_stopped && (expires == ts->next_tick)) {
768 /* Sanity check: make sure clockevent is actually programmed */
769 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
770 return;
771
772 WARN_ON_ONCE(1);
773 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
774 basemono, ts->next_tick, dev->next_event,
775 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
776 }
777
778 /*
779 * nohz_stop_sched_tick can be called several times before
780 * the nohz_restart_sched_tick is called. This happens when
781 * interrupts arrive which do not cause a reschedule. In the
782 * first call we save the current tick time, so we can restart
783 * the scheduler tick in nohz_restart_sched_tick.
784 */
785 if (!ts->tick_stopped) {
786 calc_load_nohz_start();
787 quiet_vmstat();
788
789 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
790 ts->tick_stopped = 1;
791 trace_tick_stop(1, TICK_DEP_MASK_NONE);
792 }
793
794 ts->next_tick = tick;
795
796 /*
797 * If the expiration time == KTIME_MAX, then we simply stop
798 * the tick timer.
799 */
800 if (unlikely(expires == KTIME_MAX)) {
801 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
802 hrtimer_cancel(&ts->sched_timer);
803 return;
804 }
805
806 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
807 hrtimer_start(&ts->sched_timer, tick,
808 HRTIMER_MODE_ABS_PINNED_HARD);
809 } else {
810 hrtimer_set_expires(&ts->sched_timer, tick);
811 tick_program_event(tick, 1);
812 }
813}
814
815static void tick_nohz_retain_tick(struct tick_sched *ts)
816{
817 ts->timer_expires_base = 0;
818}
819
820#ifdef CONFIG_NO_HZ_FULL
821static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
822{
823 if (tick_nohz_next_event(ts, cpu))
824 tick_nohz_stop_tick(ts, cpu);
825 else
826 tick_nohz_retain_tick(ts);
827}
828#endif /* CONFIG_NO_HZ_FULL */
829
830static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
831{
832 /* Update jiffies first */
833 tick_do_update_jiffies64(now);
834
835 /*
836 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
837 * the clock forward checks in the enqueue path:
838 */
839 timer_clear_idle();
840
841 calc_load_nohz_stop();
842 touch_softlockup_watchdog_sched();
843 /*
844 * Cancel the scheduled timer and restore the tick
845 */
846 ts->tick_stopped = 0;
847 ts->idle_exittime = now;
848
849 tick_nohz_restart(ts, now);
850}
851
852static void tick_nohz_full_update_tick(struct tick_sched *ts)
853{
854#ifdef CONFIG_NO_HZ_FULL
855 int cpu = smp_processor_id();
856
857 if (!tick_nohz_full_cpu(cpu))
858 return;
859
860 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
861 return;
862
863 if (can_stop_full_tick(cpu, ts))
864 tick_nohz_stop_sched_tick(ts, cpu);
865 else if (ts->tick_stopped)
866 tick_nohz_restart_sched_tick(ts, ktime_get());
867#endif
868}
869
870static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
871{
872 /*
873 * If this CPU is offline and it is the one which updates
874 * jiffies, then give up the assignment and let it be taken by
875 * the CPU which runs the tick timer next. If we don't drop
876 * this here the jiffies might be stale and do_timer() never
877 * invoked.
878 */
879 if (unlikely(!cpu_online(cpu))) {
880 if (cpu == tick_do_timer_cpu)
881 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
882 /*
883 * Make sure the CPU doesn't get fooled by obsolete tick
884 * deadline if it comes back online later.
885 */
886 ts->next_tick = 0;
887 return false;
888 }
889
890 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
891 return false;
892
893 if (need_resched())
894 return false;
895
896 if (unlikely(local_softirq_pending())) {
897 static int ratelimit;
898
899 if (ratelimit < 10 &&
900 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
901 pr_warn("NOHZ: local_softirq_pending %02x\n",
902 (unsigned int) local_softirq_pending());
903 ratelimit++;
904 }
905 return false;
906 }
907
908 if (tick_nohz_full_enabled()) {
909 /*
910 * Keep the tick alive to guarantee timekeeping progression
911 * if there are full dynticks CPUs around
912 */
913 if (tick_do_timer_cpu == cpu)
914 return false;
915 /*
916 * Boot safety: make sure the timekeeping duty has been
917 * assigned before entering dyntick-idle mode,
918 * tick_do_timer_cpu is TICK_DO_TIMER_BOOT
919 */
920 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_BOOT))
921 return false;
922
923 /* Should not happen for nohz-full */
924 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
925 return false;
926 }
927
928 return true;
929}
930
931static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
932{
933 ktime_t expires;
934 int cpu = smp_processor_id();
935
936 /*
937 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
938 * tick timer expiration time is known already.
939 */
940 if (ts->timer_expires_base)
941 expires = ts->timer_expires;
942 else if (can_stop_idle_tick(cpu, ts))
943 expires = tick_nohz_next_event(ts, cpu);
944 else
945 return;
946
947 ts->idle_calls++;
948
949 if (expires > 0LL) {
950 int was_stopped = ts->tick_stopped;
951
952 tick_nohz_stop_tick(ts, cpu);
953
954 ts->idle_sleeps++;
955 ts->idle_expires = expires;
956
957 if (!was_stopped && ts->tick_stopped) {
958 ts->idle_jiffies = ts->last_jiffies;
959 nohz_balance_enter_idle(cpu);
960 }
961 } else {
962 tick_nohz_retain_tick(ts);
963 }
964}
965
966/**
967 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
968 *
969 * When the next event is more than a tick into the future, stop the idle tick
970 */
971void tick_nohz_idle_stop_tick(void)
972{
973 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
974}
975
976void tick_nohz_idle_retain_tick(void)
977{
978 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
979 /*
980 * Undo the effect of get_next_timer_interrupt() called from
981 * tick_nohz_next_event().
982 */
983 timer_clear_idle();
984}
985
986/**
987 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
988 *
989 * Called when we start the idle loop.
990 */
991void tick_nohz_idle_enter(void)
992{
993 struct tick_sched *ts;
994
995 lockdep_assert_irqs_enabled();
996
997 local_irq_disable();
998
999 ts = this_cpu_ptr(&tick_cpu_sched);
1000
1001 WARN_ON_ONCE(ts->timer_expires_base);
1002
1003 ts->inidle = 1;
1004 tick_nohz_start_idle(ts);
1005
1006 local_irq_enable();
1007}
1008
1009/**
1010 * tick_nohz_irq_exit - update next tick event from interrupt exit
1011 *
1012 * When an interrupt fires while we are idle and it doesn't cause
1013 * a reschedule, it may still add, modify or delete a timer, enqueue
1014 * an RCU callback, etc...
1015 * So we need to re-calculate and reprogram the next tick event.
1016 */
1017void tick_nohz_irq_exit(void)
1018{
1019 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1020
1021 if (ts->inidle)
1022 tick_nohz_start_idle(ts);
1023 else
1024 tick_nohz_full_update_tick(ts);
1025}
1026
1027/**
1028 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1029 */
1030bool tick_nohz_idle_got_tick(void)
1031{
1032 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1033
1034 if (ts->got_idle_tick) {
1035 ts->got_idle_tick = 0;
1036 return true;
1037 }
1038 return false;
1039}
1040
1041/**
1042 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1043 * or the tick, whatever that expires first. Note that, if the tick has been
1044 * stopped, it returns the next hrtimer.
1045 *
1046 * Called from power state control code with interrupts disabled
1047 */
1048ktime_t tick_nohz_get_next_hrtimer(void)
1049{
1050 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1051}
1052
1053/**
1054 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1055 * @delta_next: duration until the next event if the tick cannot be stopped
1056 *
1057 * Called from power state control code with interrupts disabled
1058 */
1059ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1060{
1061 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1062 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1063 int cpu = smp_processor_id();
1064 /*
1065 * The idle entry time is expected to be a sufficient approximation of
1066 * the current time at this point.
1067 */
1068 ktime_t now = ts->idle_entrytime;
1069 ktime_t next_event;
1070
1071 WARN_ON_ONCE(!ts->inidle);
1072
1073 *delta_next = ktime_sub(dev->next_event, now);
1074
1075 if (!can_stop_idle_tick(cpu, ts))
1076 return *delta_next;
1077
1078 next_event = tick_nohz_next_event(ts, cpu);
1079 if (!next_event)
1080 return *delta_next;
1081
1082 /*
1083 * If the next highres timer to expire is earlier than next_event, the
1084 * idle governor needs to know that.
1085 */
1086 next_event = min_t(u64, next_event,
1087 hrtimer_next_event_without(&ts->sched_timer));
1088
1089 return ktime_sub(next_event, now);
1090}
1091
1092/**
1093 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1094 * for a particular CPU.
1095 *
1096 * Called from the schedutil frequency scaling governor in scheduler context.
1097 */
1098unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1099{
1100 struct tick_sched *ts = tick_get_tick_sched(cpu);
1101
1102 return ts->idle_calls;
1103}
1104
1105/**
1106 * tick_nohz_get_idle_calls - return the current idle calls counter value
1107 *
1108 * Called from the schedutil frequency scaling governor in scheduler context.
1109 */
1110unsigned long tick_nohz_get_idle_calls(void)
1111{
1112 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1113
1114 return ts->idle_calls;
1115}
1116
1117static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1118{
1119#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
1120 unsigned long ticks;
1121
1122 if (vtime_accounting_cpu_enabled())
1123 return;
1124 /*
1125 * We stopped the tick in idle. Update process times would miss the
1126 * time we slept as update_process_times does only a 1 tick
1127 * accounting. Enforce that this is accounted to idle !
1128 */
1129 ticks = jiffies - ts->idle_jiffies;
1130 /*
1131 * We might be one off. Do not randomly account a huge number of ticks!
1132 */
1133 if (ticks && ticks < LONG_MAX)
1134 account_idle_ticks(ticks);
1135#endif
1136}
1137
1138static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now)
1139{
1140 tick_nohz_restart_sched_tick(ts, now);
1141 tick_nohz_account_idle_ticks(ts);
1142}
1143
1144void tick_nohz_idle_restart_tick(void)
1145{
1146 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1147
1148 if (ts->tick_stopped)
1149 __tick_nohz_idle_restart_tick(ts, ktime_get());
1150}
1151
1152/**
1153 * tick_nohz_idle_exit - restart the idle tick from the idle task
1154 *
1155 * Restart the idle tick when the CPU is woken up from idle
1156 * This also exit the RCU extended quiescent state. The CPU
1157 * can use RCU again after this function is called.
1158 */
1159void tick_nohz_idle_exit(void)
1160{
1161 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1162 bool idle_active, tick_stopped;
1163 ktime_t now;
1164
1165 local_irq_disable();
1166
1167 WARN_ON_ONCE(!ts->inidle);
1168 WARN_ON_ONCE(ts->timer_expires_base);
1169
1170 ts->inidle = 0;
1171 idle_active = ts->idle_active;
1172 tick_stopped = ts->tick_stopped;
1173
1174 if (idle_active || tick_stopped)
1175 now = ktime_get();
1176
1177 if (idle_active)
1178 tick_nohz_stop_idle(ts, now);
1179
1180 if (tick_stopped)
1181 __tick_nohz_idle_restart_tick(ts, now);
1182
1183 local_irq_enable();
1184}
1185
1186/*
1187 * The nohz low res interrupt handler
1188 */
1189static void tick_nohz_handler(struct clock_event_device *dev)
1190{
1191 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1192 struct pt_regs *regs = get_irq_regs();
1193 ktime_t now = ktime_get();
1194
1195 dev->next_event = KTIME_MAX;
1196
1197 tick_sched_do_timer(ts, now);
1198 tick_sched_handle(ts, regs);
1199
1200 /* No need to reprogram if we are running tickless */
1201 if (unlikely(ts->tick_stopped))
1202 return;
1203
1204 hrtimer_forward(&ts->sched_timer, now, tick_period);
1205 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1206}
1207
1208static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1209{
1210 if (!tick_nohz_enabled)
1211 return;
1212 ts->nohz_mode = mode;
1213 /* One update is enough */
1214 if (!test_and_set_bit(0, &tick_nohz_active))
1215 timers_update_nohz();
1216}
1217
1218/**
1219 * tick_nohz_switch_to_nohz - switch to nohz mode
1220 */
1221static void tick_nohz_switch_to_nohz(void)
1222{
1223 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1224 ktime_t next;
1225
1226 if (!tick_nohz_enabled)
1227 return;
1228
1229 if (tick_switch_to_oneshot(tick_nohz_handler))
1230 return;
1231
1232 /*
1233 * Recycle the hrtimer in ts, so we can share the
1234 * hrtimer_forward with the highres code.
1235 */
1236 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1237 /* Get the next period */
1238 next = tick_init_jiffy_update();
1239
1240 hrtimer_set_expires(&ts->sched_timer, next);
1241 hrtimer_forward_now(&ts->sched_timer, tick_period);
1242 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1243 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1244}
1245
1246static inline void tick_nohz_irq_enter(void)
1247{
1248 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1249 ktime_t now;
1250
1251 if (!ts->idle_active && !ts->tick_stopped)
1252 return;
1253 now = ktime_get();
1254 if (ts->idle_active)
1255 tick_nohz_stop_idle(ts, now);
1256 if (ts->tick_stopped)
1257 tick_nohz_update_jiffies(now);
1258}
1259
1260#else
1261
1262static inline void tick_nohz_switch_to_nohz(void) { }
1263static inline void tick_nohz_irq_enter(void) { }
1264static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1265
1266#endif /* CONFIG_NO_HZ_COMMON */
1267
1268/*
1269 * Called from irq_enter to notify about the possible interruption of idle()
1270 */
1271void tick_irq_enter(void)
1272{
1273 tick_check_oneshot_broadcast_this_cpu();
1274 tick_nohz_irq_enter();
1275}
1276
1277/*
1278 * High resolution timer specific code
1279 */
1280#ifdef CONFIG_HIGH_RES_TIMERS
1281/*
1282 * We rearm the timer until we get disabled by the idle code.
1283 * Called with interrupts disabled.
1284 */
1285static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1286{
1287 struct tick_sched *ts =
1288 container_of(timer, struct tick_sched, sched_timer);
1289 struct pt_regs *regs = get_irq_regs();
1290 ktime_t now = ktime_get();
1291
1292 tick_sched_do_timer(ts, now);
1293
1294 /*
1295 * Do not call, when we are not in irq context and have
1296 * no valid regs pointer
1297 */
1298 if (regs)
1299 tick_sched_handle(ts, regs);
1300 else
1301 ts->next_tick = 0;
1302
1303 /* No need to reprogram if we are in idle or full dynticks mode */
1304 if (unlikely(ts->tick_stopped))
1305 return HRTIMER_NORESTART;
1306
1307 hrtimer_forward(timer, now, tick_period);
1308
1309 return HRTIMER_RESTART;
1310}
1311
1312static int sched_skew_tick;
1313
1314static int __init skew_tick(char *str)
1315{
1316 get_option(&str, &sched_skew_tick);
1317
1318 return 0;
1319}
1320early_param("skew_tick", skew_tick);
1321
1322/**
1323 * tick_setup_sched_timer - setup the tick emulation timer
1324 */
1325void tick_setup_sched_timer(void)
1326{
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1328 ktime_t now = ktime_get();
1329
1330 /*
1331 * Emulate tick processing via per-CPU hrtimers:
1332 */
1333 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1334 ts->sched_timer.function = tick_sched_timer;
1335
1336 /* Get the next period (per-CPU) */
1337 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1338
1339 /* Offset the tick to avert jiffies_lock contention. */
1340 if (sched_skew_tick) {
1341 u64 offset = ktime_to_ns(tick_period) >> 1;
1342 do_div(offset, num_possible_cpus());
1343 offset *= smp_processor_id();
1344 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1345 }
1346
1347 hrtimer_forward(&ts->sched_timer, now, tick_period);
1348 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1349 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1350}
1351#endif /* HIGH_RES_TIMERS */
1352
1353#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1354void tick_cancel_sched_timer(int cpu)
1355{
1356 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1357
1358# ifdef CONFIG_HIGH_RES_TIMERS
1359 if (ts->sched_timer.base)
1360 hrtimer_cancel(&ts->sched_timer);
1361# endif
1362
1363 memset(ts, 0, sizeof(*ts));
1364}
1365#endif
1366
1367/**
1368 * Async notification about clocksource changes
1369 */
1370void tick_clock_notify(void)
1371{
1372 int cpu;
1373
1374 for_each_possible_cpu(cpu)
1375 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1376}
1377
1378/*
1379 * Async notification about clock event changes
1380 */
1381void tick_oneshot_notify(void)
1382{
1383 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1384
1385 set_bit(0, &ts->check_clocks);
1386}
1387
1388/**
1389 * Check, if a change happened, which makes oneshot possible.
1390 *
1391 * Called cyclic from the hrtimer softirq (driven by the timer
1392 * softirq) allow_nohz signals, that we can switch into low-res nohz
1393 * mode, because high resolution timers are disabled (either compile
1394 * or runtime). Called with interrupts disabled.
1395 */
1396int tick_check_oneshot_change(int allow_nohz)
1397{
1398 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1399
1400 if (!test_and_clear_bit(0, &ts->check_clocks))
1401 return 0;
1402
1403 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1404 return 0;
1405
1406 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1407 return 0;
1408
1409 if (!allow_nohz)
1410 return 1;
1411
1412 tick_nohz_switch_to_nohz();
1413 return 0;
1414}
1/*
2 * linux/kernel/time/tick-sched.c
3 *
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 *
8 * No idle tick implementation for low and high resolution timers
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
12 * Distribute under GPLv2.
13 */
14#include <linux/cpu.h>
15#include <linux/err.h>
16#include <linux/hrtimer.h>
17#include <linux/interrupt.h>
18#include <linux/kernel_stat.h>
19#include <linux/percpu.h>
20#include <linux/profile.h>
21#include <linux/sched.h>
22#include <linux/module.h>
23#include <linux/irq_work.h>
24#include <linux/posix-timers.h>
25#include <linux/context_tracking.h>
26
27#include <asm/irq_regs.h>
28
29#include "tick-internal.h"
30
31#include <trace/events/timer.h>
32
33/*
34 * Per-CPU nohz control structure
35 */
36static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
37
38struct tick_sched *tick_get_tick_sched(int cpu)
39{
40 return &per_cpu(tick_cpu_sched, cpu);
41}
42
43#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
44/*
45 * The time, when the last jiffy update happened. Protected by jiffies_lock.
46 */
47static ktime_t last_jiffies_update;
48
49/*
50 * Must be called with interrupts disabled !
51 */
52static void tick_do_update_jiffies64(ktime_t now)
53{
54 unsigned long ticks = 0;
55 ktime_t delta;
56
57 /*
58 * Do a quick check without holding jiffies_lock:
59 */
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta < tick_period)
62 return;
63
64 /* Reevaluate with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
66
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta >= tick_period) {
69
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
72 tick_period);
73
74 /* Slow path for long timeouts */
75 if (unlikely(delta >= tick_period)) {
76 s64 incr = ktime_to_ns(tick_period);
77
78 ticks = ktime_divns(delta, incr);
79
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
81 incr * ticks);
82 }
83 do_timer(++ticks);
84
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
87 } else {
88 write_sequnlock(&jiffies_lock);
89 return;
90 }
91 write_sequnlock(&jiffies_lock);
92 update_wall_time();
93}
94
95/*
96 * Initialize and return retrieve the jiffies update.
97 */
98static ktime_t tick_init_jiffy_update(void)
99{
100 ktime_t period;
101
102 write_seqlock(&jiffies_lock);
103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update == 0)
105 last_jiffies_update = tick_next_period;
106 period = last_jiffies_update;
107 write_sequnlock(&jiffies_lock);
108 return period;
109}
110
111
112static void tick_sched_do_timer(ktime_t now)
113{
114 int cpu = smp_processor_id();
115
116#ifdef CONFIG_NO_HZ_COMMON
117 /*
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the CPU in charge went
120 * into a long sleep. If two CPUs happen to assign themselves to
121 * this duty, then the jiffies update is still serialized by
122 * jiffies_lock.
123 */
124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
125 && !tick_nohz_full_cpu(cpu))
126 tick_do_timer_cpu = cpu;
127#endif
128
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu == cpu)
131 tick_do_update_jiffies64(now);
132}
133
134static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
135{
136#ifdef CONFIG_NO_HZ_COMMON
137 /*
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
144 */
145 if (ts->tick_stopped) {
146 touch_softlockup_watchdog_sched();
147 if (is_idle_task(current))
148 ts->idle_jiffies++;
149 }
150#endif
151 update_process_times(user_mode(regs));
152 profile_tick(CPU_PROFILING);
153}
154#endif
155
156#ifdef CONFIG_NO_HZ_FULL
157cpumask_var_t tick_nohz_full_mask;
158cpumask_var_t housekeeping_mask;
159bool tick_nohz_full_running;
160static atomic_t tick_dep_mask;
161
162static bool check_tick_dependency(atomic_t *dep)
163{
164 int val = atomic_read(dep);
165
166 if (val & TICK_DEP_MASK_POSIX_TIMER) {
167 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
168 return true;
169 }
170
171 if (val & TICK_DEP_MASK_PERF_EVENTS) {
172 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
173 return true;
174 }
175
176 if (val & TICK_DEP_MASK_SCHED) {
177 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
178 return true;
179 }
180
181 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
182 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
183 return true;
184 }
185
186 return false;
187}
188
189static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
190{
191 WARN_ON_ONCE(!irqs_disabled());
192
193 if (unlikely(!cpu_online(cpu)))
194 return false;
195
196 if (check_tick_dependency(&tick_dep_mask))
197 return false;
198
199 if (check_tick_dependency(&ts->tick_dep_mask))
200 return false;
201
202 if (check_tick_dependency(¤t->tick_dep_mask))
203 return false;
204
205 if (check_tick_dependency(¤t->signal->tick_dep_mask))
206 return false;
207
208 return true;
209}
210
211static void nohz_full_kick_func(struct irq_work *work)
212{
213 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
214}
215
216static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
217 .func = nohz_full_kick_func,
218};
219
220/*
221 * Kick this CPU if it's full dynticks in order to force it to
222 * re-evaluate its dependency on the tick and restart it if necessary.
223 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
224 * is NMI safe.
225 */
226static void tick_nohz_full_kick(void)
227{
228 if (!tick_nohz_full_cpu(smp_processor_id()))
229 return;
230
231 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
232}
233
234/*
235 * Kick the CPU if it's full dynticks in order to force it to
236 * re-evaluate its dependency on the tick and restart it if necessary.
237 */
238void tick_nohz_full_kick_cpu(int cpu)
239{
240 if (!tick_nohz_full_cpu(cpu))
241 return;
242
243 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
244}
245
246/*
247 * Kick all full dynticks CPUs in order to force these to re-evaluate
248 * their dependency on the tick and restart it if necessary.
249 */
250static void tick_nohz_full_kick_all(void)
251{
252 int cpu;
253
254 if (!tick_nohz_full_running)
255 return;
256
257 preempt_disable();
258 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
259 tick_nohz_full_kick_cpu(cpu);
260 preempt_enable();
261}
262
263static void tick_nohz_dep_set_all(atomic_t *dep,
264 enum tick_dep_bits bit)
265{
266 int prev;
267
268 prev = atomic_fetch_or(BIT(bit), dep);
269 if (!prev)
270 tick_nohz_full_kick_all();
271}
272
273/*
274 * Set a global tick dependency. Used by perf events that rely on freq and
275 * by unstable clock.
276 */
277void tick_nohz_dep_set(enum tick_dep_bits bit)
278{
279 tick_nohz_dep_set_all(&tick_dep_mask, bit);
280}
281
282void tick_nohz_dep_clear(enum tick_dep_bits bit)
283{
284 atomic_andnot(BIT(bit), &tick_dep_mask);
285}
286
287/*
288 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
289 * manage events throttling.
290 */
291void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
292{
293 int prev;
294 struct tick_sched *ts;
295
296 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
297
298 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
299 if (!prev) {
300 preempt_disable();
301 /* Perf needs local kick that is NMI safe */
302 if (cpu == smp_processor_id()) {
303 tick_nohz_full_kick();
304 } else {
305 /* Remote irq work not NMI-safe */
306 if (!WARN_ON_ONCE(in_nmi()))
307 tick_nohz_full_kick_cpu(cpu);
308 }
309 preempt_enable();
310 }
311}
312
313void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
314{
315 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
316
317 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
318}
319
320/*
321 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
322 * per task timers.
323 */
324void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
325{
326 /*
327 * We could optimize this with just kicking the target running the task
328 * if that noise matters for nohz full users.
329 */
330 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
331}
332
333void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
334{
335 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
336}
337
338/*
339 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
340 * per process timers.
341 */
342void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
343{
344 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
345}
346
347void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
348{
349 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
350}
351
352/*
353 * Re-evaluate the need for the tick as we switch the current task.
354 * It might need the tick due to per task/process properties:
355 * perf events, posix CPU timers, ...
356 */
357void __tick_nohz_task_switch(void)
358{
359 unsigned long flags;
360 struct tick_sched *ts;
361
362 local_irq_save(flags);
363
364 if (!tick_nohz_full_cpu(smp_processor_id()))
365 goto out;
366
367 ts = this_cpu_ptr(&tick_cpu_sched);
368
369 if (ts->tick_stopped) {
370 if (atomic_read(¤t->tick_dep_mask) ||
371 atomic_read(¤t->signal->tick_dep_mask))
372 tick_nohz_full_kick();
373 }
374out:
375 local_irq_restore(flags);
376}
377
378/* Parse the boot-time nohz CPU list from the kernel parameters. */
379static int __init tick_nohz_full_setup(char *str)
380{
381 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
382 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
383 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
384 free_bootmem_cpumask_var(tick_nohz_full_mask);
385 return 1;
386 }
387 tick_nohz_full_running = true;
388
389 return 1;
390}
391__setup("nohz_full=", tick_nohz_full_setup);
392
393static int tick_nohz_cpu_down(unsigned int cpu)
394{
395 /*
396 * The boot CPU handles housekeeping duty (unbound timers,
397 * workqueues, timekeeping, ...) on behalf of full dynticks
398 * CPUs. It must remain online when nohz full is enabled.
399 */
400 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
401 return -EBUSY;
402 return 0;
403}
404
405static int tick_nohz_init_all(void)
406{
407 int err = -1;
408
409#ifdef CONFIG_NO_HZ_FULL_ALL
410 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
411 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
412 return err;
413 }
414 err = 0;
415 cpumask_setall(tick_nohz_full_mask);
416 tick_nohz_full_running = true;
417#endif
418 return err;
419}
420
421void __init tick_nohz_init(void)
422{
423 int cpu, ret;
424
425 if (!tick_nohz_full_running) {
426 if (tick_nohz_init_all() < 0)
427 return;
428 }
429
430 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
431 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
432 cpumask_clear(tick_nohz_full_mask);
433 tick_nohz_full_running = false;
434 return;
435 }
436
437 /*
438 * Full dynticks uses irq work to drive the tick rescheduling on safe
439 * locking contexts. But then we need irq work to raise its own
440 * interrupts to avoid circular dependency on the tick
441 */
442 if (!arch_irq_work_has_interrupt()) {
443 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
444 cpumask_clear(tick_nohz_full_mask);
445 cpumask_copy(housekeeping_mask, cpu_possible_mask);
446 tick_nohz_full_running = false;
447 return;
448 }
449
450 cpu = smp_processor_id();
451
452 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
453 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
454 cpu);
455 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
456 }
457
458 cpumask_andnot(housekeeping_mask,
459 cpu_possible_mask, tick_nohz_full_mask);
460
461 for_each_cpu(cpu, tick_nohz_full_mask)
462 context_tracking_cpu_set(cpu);
463
464 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
465 "kernel/nohz:predown", NULL,
466 tick_nohz_cpu_down);
467 WARN_ON(ret < 0);
468 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
469 cpumask_pr_args(tick_nohz_full_mask));
470
471 /*
472 * We need at least one CPU to handle housekeeping work such
473 * as timekeeping, unbound timers, workqueues, ...
474 */
475 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
476}
477#endif
478
479/*
480 * NOHZ - aka dynamic tick functionality
481 */
482#ifdef CONFIG_NO_HZ_COMMON
483/*
484 * NO HZ enabled ?
485 */
486bool tick_nohz_enabled __read_mostly = true;
487unsigned long tick_nohz_active __read_mostly;
488/*
489 * Enable / Disable tickless mode
490 */
491static int __init setup_tick_nohz(char *str)
492{
493 return (kstrtobool(str, &tick_nohz_enabled) == 0);
494}
495
496__setup("nohz=", setup_tick_nohz);
497
498int tick_nohz_tick_stopped(void)
499{
500 return __this_cpu_read(tick_cpu_sched.tick_stopped);
501}
502
503/**
504 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
505 *
506 * Called from interrupt entry when the CPU was idle
507 *
508 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
509 * must be updated. Otherwise an interrupt handler could use a stale jiffy
510 * value. We do this unconditionally on any CPU, as we don't know whether the
511 * CPU, which has the update task assigned is in a long sleep.
512 */
513static void tick_nohz_update_jiffies(ktime_t now)
514{
515 unsigned long flags;
516
517 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
518
519 local_irq_save(flags);
520 tick_do_update_jiffies64(now);
521 local_irq_restore(flags);
522
523 touch_softlockup_watchdog_sched();
524}
525
526/*
527 * Updates the per-CPU time idle statistics counters
528 */
529static void
530update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
531{
532 ktime_t delta;
533
534 if (ts->idle_active) {
535 delta = ktime_sub(now, ts->idle_entrytime);
536 if (nr_iowait_cpu(cpu) > 0)
537 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
538 else
539 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
540 ts->idle_entrytime = now;
541 }
542
543 if (last_update_time)
544 *last_update_time = ktime_to_us(now);
545
546}
547
548static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
549{
550 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
551 ts->idle_active = 0;
552
553 sched_clock_idle_wakeup_event(0);
554}
555
556static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
557{
558 ktime_t now = ktime_get();
559
560 ts->idle_entrytime = now;
561 ts->idle_active = 1;
562 sched_clock_idle_sleep_event();
563 return now;
564}
565
566/**
567 * get_cpu_idle_time_us - get the total idle time of a CPU
568 * @cpu: CPU number to query
569 * @last_update_time: variable to store update time in. Do not update
570 * counters if NULL.
571 *
572 * Return the cumulative idle time (since boot) for a given
573 * CPU, in microseconds.
574 *
575 * This time is measured via accounting rather than sampling,
576 * and is as accurate as ktime_get() is.
577 *
578 * This function returns -1 if NOHZ is not enabled.
579 */
580u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
581{
582 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
583 ktime_t now, idle;
584
585 if (!tick_nohz_active)
586 return -1;
587
588 now = ktime_get();
589 if (last_update_time) {
590 update_ts_time_stats(cpu, ts, now, last_update_time);
591 idle = ts->idle_sleeptime;
592 } else {
593 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
594 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
595
596 idle = ktime_add(ts->idle_sleeptime, delta);
597 } else {
598 idle = ts->idle_sleeptime;
599 }
600 }
601
602 return ktime_to_us(idle);
603
604}
605EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
606
607/**
608 * get_cpu_iowait_time_us - get the total iowait time of a CPU
609 * @cpu: CPU number to query
610 * @last_update_time: variable to store update time in. Do not update
611 * counters if NULL.
612 *
613 * Return the cumulative iowait time (since boot) for a given
614 * CPU, in microseconds.
615 *
616 * This time is measured via accounting rather than sampling,
617 * and is as accurate as ktime_get() is.
618 *
619 * This function returns -1 if NOHZ is not enabled.
620 */
621u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
622{
623 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
624 ktime_t now, iowait;
625
626 if (!tick_nohz_active)
627 return -1;
628
629 now = ktime_get();
630 if (last_update_time) {
631 update_ts_time_stats(cpu, ts, now, last_update_time);
632 iowait = ts->iowait_sleeptime;
633 } else {
634 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
635 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
636
637 iowait = ktime_add(ts->iowait_sleeptime, delta);
638 } else {
639 iowait = ts->iowait_sleeptime;
640 }
641 }
642
643 return ktime_to_us(iowait);
644}
645EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
646
647static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
648{
649 hrtimer_cancel(&ts->sched_timer);
650 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
651
652 /* Forward the time to expire in the future */
653 hrtimer_forward(&ts->sched_timer, now, tick_period);
654
655 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
656 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
657 else
658 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
659}
660
661static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
662 ktime_t now, int cpu)
663{
664 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
665 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
666 unsigned long seq, basejiff;
667 ktime_t tick;
668
669 /* Read jiffies and the time when jiffies were updated last */
670 do {
671 seq = read_seqbegin(&jiffies_lock);
672 basemono = last_jiffies_update;
673 basejiff = jiffies;
674 } while (read_seqretry(&jiffies_lock, seq));
675 ts->last_jiffies = basejiff;
676
677 if (rcu_needs_cpu(basemono, &next_rcu) ||
678 arch_needs_cpu() || irq_work_needs_cpu()) {
679 next_tick = basemono + TICK_NSEC;
680 } else {
681 /*
682 * Get the next pending timer. If high resolution
683 * timers are enabled this only takes the timer wheel
684 * timers into account. If high resolution timers are
685 * disabled this also looks at the next expiring
686 * hrtimer.
687 */
688 next_tmr = get_next_timer_interrupt(basejiff, basemono);
689 ts->next_timer = next_tmr;
690 /* Take the next rcu event into account */
691 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
692 }
693
694 /*
695 * If the tick is due in the next period, keep it ticking or
696 * force prod the timer.
697 */
698 delta = next_tick - basemono;
699 if (delta <= (u64)TICK_NSEC) {
700 tick = 0;
701
702 /*
703 * Tell the timer code that the base is not idle, i.e. undo
704 * the effect of get_next_timer_interrupt():
705 */
706 timer_clear_idle();
707 /*
708 * We've not stopped the tick yet, and there's a timer in the
709 * next period, so no point in stopping it either, bail.
710 */
711 if (!ts->tick_stopped)
712 goto out;
713
714 /*
715 * If, OTOH, we did stop it, but there's a pending (expired)
716 * timer reprogram the timer hardware to fire now.
717 *
718 * We will not restart the tick proper, just prod the timer
719 * hardware into firing an interrupt to process the pending
720 * timers. Just like tick_irq_exit() will not restart the tick
721 * for 'normal' interrupts.
722 *
723 * Only once we exit the idle loop will we re-enable the tick,
724 * see tick_nohz_idle_exit().
725 */
726 if (delta == 0) {
727 tick_nohz_restart(ts, now);
728 goto out;
729 }
730 }
731
732 /*
733 * If this CPU is the one which updates jiffies, then give up
734 * the assignment and let it be taken by the CPU which runs
735 * the tick timer next, which might be this CPU as well. If we
736 * don't drop this here the jiffies might be stale and
737 * do_timer() never invoked. Keep track of the fact that it
738 * was the one which had the do_timer() duty last. If this CPU
739 * is the one which had the do_timer() duty last, we limit the
740 * sleep time to the timekeeping max_deferment value.
741 * Otherwise we can sleep as long as we want.
742 */
743 delta = timekeeping_max_deferment();
744 if (cpu == tick_do_timer_cpu) {
745 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
746 ts->do_timer_last = 1;
747 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
748 delta = KTIME_MAX;
749 ts->do_timer_last = 0;
750 } else if (!ts->do_timer_last) {
751 delta = KTIME_MAX;
752 }
753
754#ifdef CONFIG_NO_HZ_FULL
755 /* Limit the tick delta to the maximum scheduler deferment */
756 if (!ts->inidle)
757 delta = min(delta, scheduler_tick_max_deferment());
758#endif
759
760 /* Calculate the next expiry time */
761 if (delta < (KTIME_MAX - basemono))
762 expires = basemono + delta;
763 else
764 expires = KTIME_MAX;
765
766 expires = min_t(u64, expires, next_tick);
767 tick = expires;
768
769 /* Skip reprogram of event if its not changed */
770 if (ts->tick_stopped && (expires == dev->next_event))
771 goto out;
772
773 /*
774 * nohz_stop_sched_tick can be called several times before
775 * the nohz_restart_sched_tick is called. This happens when
776 * interrupts arrive which do not cause a reschedule. In the
777 * first call we save the current tick time, so we can restart
778 * the scheduler tick in nohz_restart_sched_tick.
779 */
780 if (!ts->tick_stopped) {
781 nohz_balance_enter_idle(cpu);
782 calc_load_enter_idle();
783 cpu_load_update_nohz_start();
784
785 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
786 ts->tick_stopped = 1;
787 trace_tick_stop(1, TICK_DEP_MASK_NONE);
788 }
789
790 /*
791 * If the expiration time == KTIME_MAX, then we simply stop
792 * the tick timer.
793 */
794 if (unlikely(expires == KTIME_MAX)) {
795 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
796 hrtimer_cancel(&ts->sched_timer);
797 goto out;
798 }
799
800 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
801 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
802 else
803 tick_program_event(tick, 1);
804out:
805 /* Update the estimated sleep length */
806 ts->sleep_length = ktime_sub(dev->next_event, now);
807 return tick;
808}
809
810static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
811{
812 /* Update jiffies first */
813 tick_do_update_jiffies64(now);
814 cpu_load_update_nohz_stop();
815
816 /*
817 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
818 * the clock forward checks in the enqueue path:
819 */
820 timer_clear_idle();
821
822 calc_load_exit_idle();
823 touch_softlockup_watchdog_sched();
824 /*
825 * Cancel the scheduled timer and restore the tick
826 */
827 ts->tick_stopped = 0;
828 ts->idle_exittime = now;
829
830 tick_nohz_restart(ts, now);
831}
832
833static void tick_nohz_full_update_tick(struct tick_sched *ts)
834{
835#ifdef CONFIG_NO_HZ_FULL
836 int cpu = smp_processor_id();
837
838 if (!tick_nohz_full_cpu(cpu))
839 return;
840
841 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
842 return;
843
844 if (can_stop_full_tick(cpu, ts))
845 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
846 else if (ts->tick_stopped)
847 tick_nohz_restart_sched_tick(ts, ktime_get());
848#endif
849}
850
851static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
852{
853 /*
854 * If this CPU is offline and it is the one which updates
855 * jiffies, then give up the assignment and let it be taken by
856 * the CPU which runs the tick timer next. If we don't drop
857 * this here the jiffies might be stale and do_timer() never
858 * invoked.
859 */
860 if (unlikely(!cpu_online(cpu))) {
861 if (cpu == tick_do_timer_cpu)
862 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
863 return false;
864 }
865
866 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
867 ts->sleep_length = NSEC_PER_SEC / HZ;
868 return false;
869 }
870
871 if (need_resched())
872 return false;
873
874 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
875 static int ratelimit;
876
877 if (ratelimit < 10 &&
878 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
879 pr_warn("NOHZ: local_softirq_pending %02x\n",
880 (unsigned int) local_softirq_pending());
881 ratelimit++;
882 }
883 return false;
884 }
885
886 if (tick_nohz_full_enabled()) {
887 /*
888 * Keep the tick alive to guarantee timekeeping progression
889 * if there are full dynticks CPUs around
890 */
891 if (tick_do_timer_cpu == cpu)
892 return false;
893 /*
894 * Boot safety: make sure the timekeeping duty has been
895 * assigned before entering dyntick-idle mode,
896 */
897 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
898 return false;
899 }
900
901 return true;
902}
903
904static void __tick_nohz_idle_enter(struct tick_sched *ts)
905{
906 ktime_t now, expires;
907 int cpu = smp_processor_id();
908
909 now = tick_nohz_start_idle(ts);
910
911 if (can_stop_idle_tick(cpu, ts)) {
912 int was_stopped = ts->tick_stopped;
913
914 ts->idle_calls++;
915
916 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
917 if (expires > 0LL) {
918 ts->idle_sleeps++;
919 ts->idle_expires = expires;
920 }
921
922 if (!was_stopped && ts->tick_stopped)
923 ts->idle_jiffies = ts->last_jiffies;
924 }
925}
926
927/**
928 * tick_nohz_idle_enter - stop the idle tick from the idle task
929 *
930 * When the next event is more than a tick into the future, stop the idle tick
931 * Called when we start the idle loop.
932 *
933 * The arch is responsible of calling:
934 *
935 * - rcu_idle_enter() after its last use of RCU before the CPU is put
936 * to sleep.
937 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
938 */
939void tick_nohz_idle_enter(void)
940{
941 struct tick_sched *ts;
942
943 WARN_ON_ONCE(irqs_disabled());
944
945 /*
946 * Update the idle state in the scheduler domain hierarchy
947 * when tick_nohz_stop_sched_tick() is called from the idle loop.
948 * State will be updated to busy during the first busy tick after
949 * exiting idle.
950 */
951 set_cpu_sd_state_idle();
952
953 local_irq_disable();
954
955 ts = this_cpu_ptr(&tick_cpu_sched);
956 ts->inidle = 1;
957 __tick_nohz_idle_enter(ts);
958
959 local_irq_enable();
960}
961
962/**
963 * tick_nohz_irq_exit - update next tick event from interrupt exit
964 *
965 * When an interrupt fires while we are idle and it doesn't cause
966 * a reschedule, it may still add, modify or delete a timer, enqueue
967 * an RCU callback, etc...
968 * So we need to re-calculate and reprogram the next tick event.
969 */
970void tick_nohz_irq_exit(void)
971{
972 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
973
974 if (ts->inidle)
975 __tick_nohz_idle_enter(ts);
976 else
977 tick_nohz_full_update_tick(ts);
978}
979
980/**
981 * tick_nohz_get_sleep_length - return the length of the current sleep
982 *
983 * Called from power state control code with interrupts disabled
984 */
985ktime_t tick_nohz_get_sleep_length(void)
986{
987 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
988
989 return ts->sleep_length;
990}
991
992static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
993{
994#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
995 unsigned long ticks;
996
997 if (vtime_accounting_cpu_enabled())
998 return;
999 /*
1000 * We stopped the tick in idle. Update process times would miss the
1001 * time we slept as update_process_times does only a 1 tick
1002 * accounting. Enforce that this is accounted to idle !
1003 */
1004 ticks = jiffies - ts->idle_jiffies;
1005 /*
1006 * We might be one off. Do not randomly account a huge number of ticks!
1007 */
1008 if (ticks && ticks < LONG_MAX)
1009 account_idle_ticks(ticks);
1010#endif
1011}
1012
1013/**
1014 * tick_nohz_idle_exit - restart the idle tick from the idle task
1015 *
1016 * Restart the idle tick when the CPU is woken up from idle
1017 * This also exit the RCU extended quiescent state. The CPU
1018 * can use RCU again after this function is called.
1019 */
1020void tick_nohz_idle_exit(void)
1021{
1022 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1023 ktime_t now;
1024
1025 local_irq_disable();
1026
1027 WARN_ON_ONCE(!ts->inidle);
1028
1029 ts->inidle = 0;
1030
1031 if (ts->idle_active || ts->tick_stopped)
1032 now = ktime_get();
1033
1034 if (ts->idle_active)
1035 tick_nohz_stop_idle(ts, now);
1036
1037 if (ts->tick_stopped) {
1038 tick_nohz_restart_sched_tick(ts, now);
1039 tick_nohz_account_idle_ticks(ts);
1040 }
1041
1042 local_irq_enable();
1043}
1044
1045/*
1046 * The nohz low res interrupt handler
1047 */
1048static void tick_nohz_handler(struct clock_event_device *dev)
1049{
1050 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1051 struct pt_regs *regs = get_irq_regs();
1052 ktime_t now = ktime_get();
1053
1054 dev->next_event = KTIME_MAX;
1055
1056 tick_sched_do_timer(now);
1057 tick_sched_handle(ts, regs);
1058
1059 /* No need to reprogram if we are running tickless */
1060 if (unlikely(ts->tick_stopped))
1061 return;
1062
1063 hrtimer_forward(&ts->sched_timer, now, tick_period);
1064 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1065}
1066
1067static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1068{
1069 if (!tick_nohz_enabled)
1070 return;
1071 ts->nohz_mode = mode;
1072 /* One update is enough */
1073 if (!test_and_set_bit(0, &tick_nohz_active))
1074 timers_update_migration(true);
1075}
1076
1077/**
1078 * tick_nohz_switch_to_nohz - switch to nohz mode
1079 */
1080static void tick_nohz_switch_to_nohz(void)
1081{
1082 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1083 ktime_t next;
1084
1085 if (!tick_nohz_enabled)
1086 return;
1087
1088 if (tick_switch_to_oneshot(tick_nohz_handler))
1089 return;
1090
1091 /*
1092 * Recycle the hrtimer in ts, so we can share the
1093 * hrtimer_forward with the highres code.
1094 */
1095 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1096 /* Get the next period */
1097 next = tick_init_jiffy_update();
1098
1099 hrtimer_set_expires(&ts->sched_timer, next);
1100 hrtimer_forward_now(&ts->sched_timer, tick_period);
1101 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1102 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1103}
1104
1105static inline void tick_nohz_irq_enter(void)
1106{
1107 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1108 ktime_t now;
1109
1110 if (!ts->idle_active && !ts->tick_stopped)
1111 return;
1112 now = ktime_get();
1113 if (ts->idle_active)
1114 tick_nohz_stop_idle(ts, now);
1115 if (ts->tick_stopped)
1116 tick_nohz_update_jiffies(now);
1117}
1118
1119#else
1120
1121static inline void tick_nohz_switch_to_nohz(void) { }
1122static inline void tick_nohz_irq_enter(void) { }
1123static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1124
1125#endif /* CONFIG_NO_HZ_COMMON */
1126
1127/*
1128 * Called from irq_enter to notify about the possible interruption of idle()
1129 */
1130void tick_irq_enter(void)
1131{
1132 tick_check_oneshot_broadcast_this_cpu();
1133 tick_nohz_irq_enter();
1134}
1135
1136/*
1137 * High resolution timer specific code
1138 */
1139#ifdef CONFIG_HIGH_RES_TIMERS
1140/*
1141 * We rearm the timer until we get disabled by the idle code.
1142 * Called with interrupts disabled.
1143 */
1144static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1145{
1146 struct tick_sched *ts =
1147 container_of(timer, struct tick_sched, sched_timer);
1148 struct pt_regs *regs = get_irq_regs();
1149 ktime_t now = ktime_get();
1150
1151 tick_sched_do_timer(now);
1152
1153 /*
1154 * Do not call, when we are not in irq context and have
1155 * no valid regs pointer
1156 */
1157 if (regs)
1158 tick_sched_handle(ts, regs);
1159
1160 /* No need to reprogram if we are in idle or full dynticks mode */
1161 if (unlikely(ts->tick_stopped))
1162 return HRTIMER_NORESTART;
1163
1164 hrtimer_forward(timer, now, tick_period);
1165
1166 return HRTIMER_RESTART;
1167}
1168
1169static int sched_skew_tick;
1170
1171static int __init skew_tick(char *str)
1172{
1173 get_option(&str, &sched_skew_tick);
1174
1175 return 0;
1176}
1177early_param("skew_tick", skew_tick);
1178
1179/**
1180 * tick_setup_sched_timer - setup the tick emulation timer
1181 */
1182void tick_setup_sched_timer(void)
1183{
1184 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1185 ktime_t now = ktime_get();
1186
1187 /*
1188 * Emulate tick processing via per-CPU hrtimers:
1189 */
1190 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1191 ts->sched_timer.function = tick_sched_timer;
1192
1193 /* Get the next period (per-CPU) */
1194 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1195
1196 /* Offset the tick to avert jiffies_lock contention. */
1197 if (sched_skew_tick) {
1198 u64 offset = ktime_to_ns(tick_period) >> 1;
1199 do_div(offset, num_possible_cpus());
1200 offset *= smp_processor_id();
1201 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1202 }
1203
1204 hrtimer_forward(&ts->sched_timer, now, tick_period);
1205 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1206 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1207}
1208#endif /* HIGH_RES_TIMERS */
1209
1210#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1211void tick_cancel_sched_timer(int cpu)
1212{
1213 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1214
1215# ifdef CONFIG_HIGH_RES_TIMERS
1216 if (ts->sched_timer.base)
1217 hrtimer_cancel(&ts->sched_timer);
1218# endif
1219
1220 memset(ts, 0, sizeof(*ts));
1221}
1222#endif
1223
1224/**
1225 * Async notification about clocksource changes
1226 */
1227void tick_clock_notify(void)
1228{
1229 int cpu;
1230
1231 for_each_possible_cpu(cpu)
1232 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1233}
1234
1235/*
1236 * Async notification about clock event changes
1237 */
1238void tick_oneshot_notify(void)
1239{
1240 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1241
1242 set_bit(0, &ts->check_clocks);
1243}
1244
1245/**
1246 * Check, if a change happened, which makes oneshot possible.
1247 *
1248 * Called cyclic from the hrtimer softirq (driven by the timer
1249 * softirq) allow_nohz signals, that we can switch into low-res nohz
1250 * mode, because high resolution timers are disabled (either compile
1251 * or runtime). Called with interrupts disabled.
1252 */
1253int tick_check_oneshot_change(int allow_nohz)
1254{
1255 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1256
1257 if (!test_and_clear_bit(0, &ts->check_clocks))
1258 return 0;
1259
1260 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1261 return 0;
1262
1263 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1264 return 0;
1265
1266 if (!allow_nohz)
1267 return 1;
1268
1269 tick_nohz_switch_to_nohz();
1270 return 0;
1271}