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