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