Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * No idle tick implementation for low and high resolution timers
8 *
9 * Started by: Thomas Gleixner and Ingo Molnar
10 */
11#include <linux/cpu.h>
12#include <linux/err.h>
13#include <linux/hrtimer.h>
14#include <linux/interrupt.h>
15#include <linux/kernel_stat.h>
16#include <linux/percpu.h>
17#include <linux/nmi.h>
18#include <linux/profile.h>
19#include <linux/sched/signal.h>
20#include <linux/sched/clock.h>
21#include <linux/sched/stat.h>
22#include <linux/sched/nohz.h>
23#include <linux/sched/loadavg.h>
24#include <linux/module.h>
25#include <linux/irq_work.h>
26#include <linux/posix-timers.h>
27#include <linux/context_tracking.h>
28#include <linux/mm.h>
29
30#include <asm/irq_regs.h>
31
32#include "tick-internal.h"
33
34#include <trace/events/timer.h>
35
36/*
37 * Per-CPU nohz control structure
38 */
39static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40
41struct tick_sched *tick_get_tick_sched(int cpu)
42{
43 return &per_cpu(tick_cpu_sched, cpu);
44}
45
46#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47/*
48 * The time, when the last jiffy update happened. Write access must hold
49 * jiffies_lock and jiffies_seq. tick_nohz_next_event() needs to get a
50 * consistent view of jiffies and last_jiffies_update.
51 */
52static ktime_t last_jiffies_update;
53
54/*
55 * Must be called with interrupts disabled !
56 */
57static void tick_do_update_jiffies64(ktime_t now)
58{
59 unsigned long ticks = 1;
60 ktime_t delta, nextp;
61
62 /*
63 * 64bit can do a quick check without holding jiffies lock and
64 * without looking at the sequence count. The smp_load_acquire()
65 * pairs with the update done later in this function.
66 *
67 * 32bit cannot do that because the store of tick_next_period
68 * consists of two 32bit stores and the first store could move it
69 * to a random point in the future.
70 */
71 if (IS_ENABLED(CONFIG_64BIT)) {
72 if (ktime_before(now, smp_load_acquire(&tick_next_period)))
73 return;
74 } else {
75 unsigned int seq;
76
77 /*
78 * Avoid contention on jiffies_lock and protect the quick
79 * check with the sequence count.
80 */
81 do {
82 seq = read_seqcount_begin(&jiffies_seq);
83 nextp = tick_next_period;
84 } while (read_seqcount_retry(&jiffies_seq, seq));
85
86 if (ktime_before(now, nextp))
87 return;
88 }
89
90 /* Quick check failed, i.e. update is required. */
91 raw_spin_lock(&jiffies_lock);
92 /*
93 * Reevaluate with the lock held. Another CPU might have done the
94 * update already.
95 */
96 if (ktime_before(now, tick_next_period)) {
97 raw_spin_unlock(&jiffies_lock);
98 return;
99 }
100
101 write_seqcount_begin(&jiffies_seq);
102
103 delta = ktime_sub(now, tick_next_period);
104 if (unlikely(delta >= TICK_NSEC)) {
105 /* Slow path for long idle sleep times */
106 s64 incr = TICK_NSEC;
107
108 ticks += ktime_divns(delta, incr);
109
110 last_jiffies_update = ktime_add_ns(last_jiffies_update,
111 incr * ticks);
112 } else {
113 last_jiffies_update = ktime_add_ns(last_jiffies_update,
114 TICK_NSEC);
115 }
116
117 /* Advance jiffies to complete the jiffies_seq protected job */
118 jiffies_64 += ticks;
119
120 /*
121 * Keep the tick_next_period variable up to date.
122 */
123 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
124
125 if (IS_ENABLED(CONFIG_64BIT)) {
126 /*
127 * Pairs with smp_load_acquire() in the lockless quick
128 * check above and ensures that the update to jiffies_64 is
129 * not reordered vs. the store to tick_next_period, neither
130 * by the compiler nor by the CPU.
131 */
132 smp_store_release(&tick_next_period, nextp);
133 } else {
134 /*
135 * A plain store is good enough on 32bit as the quick check
136 * above is protected by the sequence count.
137 */
138 tick_next_period = nextp;
139 }
140
141 /*
142 * Release the sequence count. calc_global_load() below is not
143 * protected by it, but jiffies_lock needs to be held to prevent
144 * concurrent invocations.
145 */
146 write_seqcount_end(&jiffies_seq);
147
148 calc_global_load();
149
150 raw_spin_unlock(&jiffies_lock);
151 update_wall_time();
152}
153
154/*
155 * Initialize and return retrieve the jiffies update.
156 */
157static ktime_t tick_init_jiffy_update(void)
158{
159 ktime_t period;
160
161 raw_spin_lock(&jiffies_lock);
162 write_seqcount_begin(&jiffies_seq);
163 /* Did we start the jiffies update yet ? */
164 if (last_jiffies_update == 0)
165 last_jiffies_update = tick_next_period;
166 period = last_jiffies_update;
167 write_seqcount_end(&jiffies_seq);
168 raw_spin_unlock(&jiffies_lock);
169 return period;
170}
171
172#define MAX_STALLED_JIFFIES 5
173
174static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
175{
176 int cpu = smp_processor_id();
177
178#ifdef CONFIG_NO_HZ_COMMON
179 /*
180 * Check if the do_timer duty was dropped. We don't care about
181 * concurrency: This happens only when the CPU in charge went
182 * into a long sleep. If two CPUs happen to assign themselves to
183 * this duty, then the jiffies update is still serialized by
184 * jiffies_lock.
185 *
186 * If nohz_full is enabled, this should not happen because the
187 * tick_do_timer_cpu never relinquishes.
188 */
189 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
190#ifdef CONFIG_NO_HZ_FULL
191 WARN_ON_ONCE(tick_nohz_full_running);
192#endif
193 tick_do_timer_cpu = cpu;
194 }
195#endif
196
197 /* Check, if the jiffies need an update */
198 if (tick_do_timer_cpu == cpu)
199 tick_do_update_jiffies64(now);
200
201 /*
202 * If jiffies update stalled for too long (timekeeper in stop_machine()
203 * or VMEXIT'ed for several msecs), force an update.
204 */
205 if (ts->last_tick_jiffies != jiffies) {
206 ts->stalled_jiffies = 0;
207 ts->last_tick_jiffies = READ_ONCE(jiffies);
208 } else {
209 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
210 tick_do_update_jiffies64(now);
211 ts->stalled_jiffies = 0;
212 ts->last_tick_jiffies = READ_ONCE(jiffies);
213 }
214 }
215
216 if (ts->inidle)
217 ts->got_idle_tick = 1;
218}
219
220static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
221{
222#ifdef CONFIG_NO_HZ_COMMON
223 /*
224 * When we are idle and the tick is stopped, we have to touch
225 * the watchdog as we might not schedule for a really long
226 * time. This happens on complete idle SMP systems while
227 * waiting on the login prompt. We also increment the "start of
228 * idle" jiffy stamp so the idle accounting adjustment we do
229 * when we go busy again does not account too much ticks.
230 */
231 if (ts->tick_stopped) {
232 touch_softlockup_watchdog_sched();
233 if (is_idle_task(current))
234 ts->idle_jiffies++;
235 /*
236 * In case the current tick fired too early past its expected
237 * expiration, make sure we don't bypass the next clock reprogramming
238 * to the same deadline.
239 */
240 ts->next_tick = 0;
241 }
242#endif
243 update_process_times(user_mode(regs));
244 profile_tick(CPU_PROFILING);
245}
246#endif
247
248#ifdef CONFIG_NO_HZ_FULL
249cpumask_var_t tick_nohz_full_mask;
250EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
251bool tick_nohz_full_running;
252EXPORT_SYMBOL_GPL(tick_nohz_full_running);
253static atomic_t tick_dep_mask;
254
255static bool check_tick_dependency(atomic_t *dep)
256{
257 int val = atomic_read(dep);
258
259 if (val & TICK_DEP_MASK_POSIX_TIMER) {
260 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
261 return true;
262 }
263
264 if (val & TICK_DEP_MASK_PERF_EVENTS) {
265 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
266 return true;
267 }
268
269 if (val & TICK_DEP_MASK_SCHED) {
270 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
271 return true;
272 }
273
274 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
275 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
276 return true;
277 }
278
279 if (val & TICK_DEP_MASK_RCU) {
280 trace_tick_stop(0, TICK_DEP_MASK_RCU);
281 return true;
282 }
283
284 return false;
285}
286
287static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
288{
289 lockdep_assert_irqs_disabled();
290
291 if (unlikely(!cpu_online(cpu)))
292 return false;
293
294 if (check_tick_dependency(&tick_dep_mask))
295 return false;
296
297 if (check_tick_dependency(&ts->tick_dep_mask))
298 return false;
299
300 if (check_tick_dependency(¤t->tick_dep_mask))
301 return false;
302
303 if (check_tick_dependency(¤t->signal->tick_dep_mask))
304 return false;
305
306 return true;
307}
308
309static void nohz_full_kick_func(struct irq_work *work)
310{
311 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
312}
313
314static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
315 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
316
317/*
318 * Kick this CPU if it's full dynticks in order to force it to
319 * re-evaluate its dependency on the tick and restart it if necessary.
320 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
321 * is NMI safe.
322 */
323static void tick_nohz_full_kick(void)
324{
325 if (!tick_nohz_full_cpu(smp_processor_id()))
326 return;
327
328 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
329}
330
331/*
332 * Kick the CPU if it's full dynticks in order to force it to
333 * re-evaluate its dependency on the tick and restart it if necessary.
334 */
335void tick_nohz_full_kick_cpu(int cpu)
336{
337 if (!tick_nohz_full_cpu(cpu))
338 return;
339
340 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
341}
342
343static void tick_nohz_kick_task(struct task_struct *tsk)
344{
345 int cpu;
346
347 /*
348 * If the task is not running, run_posix_cpu_timers()
349 * has nothing to elapse, IPI can then be spared.
350 *
351 * activate_task() STORE p->tick_dep_mask
352 * STORE p->on_rq
353 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
354 * LOCK rq->lock LOAD p->on_rq
355 * smp_mb__after_spin_lock()
356 * tick_nohz_task_switch()
357 * LOAD p->tick_dep_mask
358 */
359 if (!sched_task_on_rq(tsk))
360 return;
361
362 /*
363 * If the task concurrently migrates to another CPU,
364 * we guarantee it sees the new tick dependency upon
365 * schedule.
366 *
367 * set_task_cpu(p, cpu);
368 * STORE p->cpu = @cpu
369 * __schedule() (switch to task 'p')
370 * LOCK rq->lock
371 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
372 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
373 * LOAD p->tick_dep_mask LOAD p->cpu
374 */
375 cpu = task_cpu(tsk);
376
377 preempt_disable();
378 if (cpu_online(cpu))
379 tick_nohz_full_kick_cpu(cpu);
380 preempt_enable();
381}
382
383/*
384 * Kick all full dynticks CPUs in order to force these to re-evaluate
385 * their dependency on the tick and restart it if necessary.
386 */
387static void tick_nohz_full_kick_all(void)
388{
389 int cpu;
390
391 if (!tick_nohz_full_running)
392 return;
393
394 preempt_disable();
395 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
396 tick_nohz_full_kick_cpu(cpu);
397 preempt_enable();
398}
399
400static void tick_nohz_dep_set_all(atomic_t *dep,
401 enum tick_dep_bits bit)
402{
403 int prev;
404
405 prev = atomic_fetch_or(BIT(bit), dep);
406 if (!prev)
407 tick_nohz_full_kick_all();
408}
409
410/*
411 * Set a global tick dependency. Used by perf events that rely on freq and
412 * by unstable clock.
413 */
414void tick_nohz_dep_set(enum tick_dep_bits bit)
415{
416 tick_nohz_dep_set_all(&tick_dep_mask, bit);
417}
418
419void tick_nohz_dep_clear(enum tick_dep_bits bit)
420{
421 atomic_andnot(BIT(bit), &tick_dep_mask);
422}
423
424/*
425 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
426 * manage events throttling.
427 */
428void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
429{
430 int prev;
431 struct tick_sched *ts;
432
433 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
434
435 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
436 if (!prev) {
437 preempt_disable();
438 /* Perf needs local kick that is NMI safe */
439 if (cpu == smp_processor_id()) {
440 tick_nohz_full_kick();
441 } else {
442 /* Remote irq work not NMI-safe */
443 if (!WARN_ON_ONCE(in_nmi()))
444 tick_nohz_full_kick_cpu(cpu);
445 }
446 preempt_enable();
447 }
448}
449EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
450
451void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
452{
453 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
454
455 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
456}
457EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
458
459/*
460 * Set a per-task tick dependency. RCU need this. Also posix CPU timers
461 * in order to elapse per task timers.
462 */
463void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
464{
465 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
466 tick_nohz_kick_task(tsk);
467}
468EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
469
470void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
471{
472 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
473}
474EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
475
476/*
477 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
478 * per process timers.
479 */
480void tick_nohz_dep_set_signal(struct task_struct *tsk,
481 enum tick_dep_bits bit)
482{
483 int prev;
484 struct signal_struct *sig = tsk->signal;
485
486 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
487 if (!prev) {
488 struct task_struct *t;
489
490 lockdep_assert_held(&tsk->sighand->siglock);
491 __for_each_thread(sig, t)
492 tick_nohz_kick_task(t);
493 }
494}
495
496void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
497{
498 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
499}
500
501/*
502 * Re-evaluate the need for the tick as we switch the current task.
503 * It might need the tick due to per task/process properties:
504 * perf events, posix CPU timers, ...
505 */
506void __tick_nohz_task_switch(void)
507{
508 struct tick_sched *ts;
509
510 if (!tick_nohz_full_cpu(smp_processor_id()))
511 return;
512
513 ts = this_cpu_ptr(&tick_cpu_sched);
514
515 if (ts->tick_stopped) {
516 if (atomic_read(¤t->tick_dep_mask) ||
517 atomic_read(¤t->signal->tick_dep_mask))
518 tick_nohz_full_kick();
519 }
520}
521
522/* Get the boot-time nohz CPU list from the kernel parameters. */
523void __init tick_nohz_full_setup(cpumask_var_t cpumask)
524{
525 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
526 cpumask_copy(tick_nohz_full_mask, cpumask);
527 tick_nohz_full_running = true;
528}
529
530static int tick_nohz_cpu_down(unsigned int cpu)
531{
532 /*
533 * The tick_do_timer_cpu CPU handles housekeeping duty (unbound
534 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
535 * CPUs. It must remain online when nohz full is enabled.
536 */
537 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
538 return -EBUSY;
539 return 0;
540}
541
542void __init tick_nohz_init(void)
543{
544 int cpu, ret;
545
546 if (!tick_nohz_full_running)
547 return;
548
549 /*
550 * Full dynticks uses irq work to drive the tick rescheduling on safe
551 * locking contexts. But then we need irq work to raise its own
552 * interrupts to avoid circular dependency on the tick
553 */
554 if (!arch_irq_work_has_interrupt()) {
555 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
556 cpumask_clear(tick_nohz_full_mask);
557 tick_nohz_full_running = false;
558 return;
559 }
560
561 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
562 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
563 cpu = smp_processor_id();
564
565 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
566 pr_warn("NO_HZ: Clearing %d from nohz_full range "
567 "for timekeeping\n", cpu);
568 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
569 }
570 }
571
572 for_each_cpu(cpu, tick_nohz_full_mask)
573 ct_cpu_track_user(cpu);
574
575 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
576 "kernel/nohz:predown", NULL,
577 tick_nohz_cpu_down);
578 WARN_ON(ret < 0);
579 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
580 cpumask_pr_args(tick_nohz_full_mask));
581}
582#endif
583
584/*
585 * NOHZ - aka dynamic tick functionality
586 */
587#ifdef CONFIG_NO_HZ_COMMON
588/*
589 * NO HZ enabled ?
590 */
591bool tick_nohz_enabled __read_mostly = true;
592unsigned long tick_nohz_active __read_mostly;
593/*
594 * Enable / Disable tickless mode
595 */
596static int __init setup_tick_nohz(char *str)
597{
598 return (kstrtobool(str, &tick_nohz_enabled) == 0);
599}
600
601__setup("nohz=", setup_tick_nohz);
602
603bool tick_nohz_tick_stopped(void)
604{
605 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
606
607 return ts->tick_stopped;
608}
609
610bool tick_nohz_tick_stopped_cpu(int cpu)
611{
612 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
613
614 return ts->tick_stopped;
615}
616
617/**
618 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
619 *
620 * Called from interrupt entry when the CPU was idle
621 *
622 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
623 * must be updated. Otherwise an interrupt handler could use a stale jiffy
624 * value. We do this unconditionally on any CPU, as we don't know whether the
625 * CPU, which has the update task assigned is in a long sleep.
626 */
627static void tick_nohz_update_jiffies(ktime_t now)
628{
629 unsigned long flags;
630
631 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
632
633 local_irq_save(flags);
634 tick_do_update_jiffies64(now);
635 local_irq_restore(flags);
636
637 touch_softlockup_watchdog_sched();
638}
639
640/*
641 * Updates the per-CPU time idle statistics counters
642 */
643static void
644update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
645{
646 ktime_t delta;
647
648 if (ts->idle_active) {
649 delta = ktime_sub(now, ts->idle_entrytime);
650 if (nr_iowait_cpu(cpu) > 0)
651 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
652 else
653 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
654 ts->idle_entrytime = now;
655 }
656
657 if (last_update_time)
658 *last_update_time = ktime_to_us(now);
659
660}
661
662static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
663{
664 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
665 ts->idle_active = 0;
666
667 sched_clock_idle_wakeup_event();
668}
669
670static void tick_nohz_start_idle(struct tick_sched *ts)
671{
672 ts->idle_entrytime = ktime_get();
673 ts->idle_active = 1;
674 sched_clock_idle_sleep_event();
675}
676
677/**
678 * get_cpu_idle_time_us - get the total idle time of a CPU
679 * @cpu: CPU number to query
680 * @last_update_time: variable to store update time in. Do not update
681 * counters if NULL.
682 *
683 * Return the cumulative idle time (since boot) for a given
684 * CPU, in microseconds.
685 *
686 * This time is measured via accounting rather than sampling,
687 * and is as accurate as ktime_get() is.
688 *
689 * This function returns -1 if NOHZ is not enabled.
690 */
691u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
692{
693 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
694 ktime_t now, idle;
695
696 if (!tick_nohz_active)
697 return -1;
698
699 now = ktime_get();
700 if (last_update_time) {
701 update_ts_time_stats(cpu, ts, now, last_update_time);
702 idle = ts->idle_sleeptime;
703 } else {
704 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
705 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
706
707 idle = ktime_add(ts->idle_sleeptime, delta);
708 } else {
709 idle = ts->idle_sleeptime;
710 }
711 }
712
713 return ktime_to_us(idle);
714
715}
716EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
717
718/**
719 * get_cpu_iowait_time_us - get the total iowait time of a CPU
720 * @cpu: CPU number to query
721 * @last_update_time: variable to store update time in. Do not update
722 * counters if NULL.
723 *
724 * Return the cumulative iowait time (since boot) for a given
725 * CPU, in microseconds.
726 *
727 * This time is measured via accounting rather than sampling,
728 * and is as accurate as ktime_get() is.
729 *
730 * This function returns -1 if NOHZ is not enabled.
731 */
732u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
733{
734 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
735 ktime_t now, iowait;
736
737 if (!tick_nohz_active)
738 return -1;
739
740 now = ktime_get();
741 if (last_update_time) {
742 update_ts_time_stats(cpu, ts, now, last_update_time);
743 iowait = ts->iowait_sleeptime;
744 } else {
745 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
746 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
747
748 iowait = ktime_add(ts->iowait_sleeptime, delta);
749 } else {
750 iowait = ts->iowait_sleeptime;
751 }
752 }
753
754 return ktime_to_us(iowait);
755}
756EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
757
758static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
759{
760 hrtimer_cancel(&ts->sched_timer);
761 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
762
763 /* Forward the time to expire in the future */
764 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
765
766 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
767 hrtimer_start_expires(&ts->sched_timer,
768 HRTIMER_MODE_ABS_PINNED_HARD);
769 } else {
770 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
771 }
772
773 /*
774 * Reset to make sure next tick stop doesn't get fooled by past
775 * cached clock deadline.
776 */
777 ts->next_tick = 0;
778}
779
780static inline bool local_timer_softirq_pending(void)
781{
782 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
783}
784
785static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
786{
787 u64 basemono, next_tick, delta, expires;
788 unsigned long basejiff;
789 unsigned int seq;
790
791 /* Read jiffies and the time when jiffies were updated last */
792 do {
793 seq = read_seqcount_begin(&jiffies_seq);
794 basemono = last_jiffies_update;
795 basejiff = jiffies;
796 } while (read_seqcount_retry(&jiffies_seq, seq));
797 ts->last_jiffies = basejiff;
798 ts->timer_expires_base = basemono;
799
800 /*
801 * Keep the periodic tick, when RCU, architecture or irq_work
802 * requests it.
803 * Aside of that check whether the local timer softirq is
804 * pending. If so its a bad idea to call get_next_timer_interrupt()
805 * because there is an already expired timer, so it will request
806 * immediate expiry, which rearms the hardware timer with a
807 * minimal delta which brings us back to this place
808 * immediately. Lather, rinse and repeat...
809 */
810 if (rcu_needs_cpu() || arch_needs_cpu() ||
811 irq_work_needs_cpu() || local_timer_softirq_pending()) {
812 next_tick = basemono + TICK_NSEC;
813 } else {
814 /*
815 * Get the next pending timer. If high resolution
816 * timers are enabled this only takes the timer wheel
817 * timers into account. If high resolution timers are
818 * disabled this also looks at the next expiring
819 * hrtimer.
820 */
821 next_tick = get_next_timer_interrupt(basejiff, basemono);
822 ts->next_timer = next_tick;
823 }
824
825 /*
826 * If the tick is due in the next period, keep it ticking or
827 * force prod the timer.
828 */
829 delta = next_tick - basemono;
830 if (delta <= (u64)TICK_NSEC) {
831 /*
832 * Tell the timer code that the base is not idle, i.e. undo
833 * the effect of get_next_timer_interrupt():
834 */
835 timer_clear_idle();
836 /*
837 * We've not stopped the tick yet, and there's a timer in the
838 * next period, so no point in stopping it either, bail.
839 */
840 if (!ts->tick_stopped) {
841 ts->timer_expires = 0;
842 goto out;
843 }
844 }
845
846 /*
847 * If this CPU is the one which had the do_timer() duty last, we limit
848 * the sleep time to the timekeeping max_deferment value.
849 * Otherwise we can sleep as long as we want.
850 */
851 delta = timekeeping_max_deferment();
852 if (cpu != tick_do_timer_cpu &&
853 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
854 delta = KTIME_MAX;
855
856 /* Calculate the next expiry time */
857 if (delta < (KTIME_MAX - basemono))
858 expires = basemono + delta;
859 else
860 expires = KTIME_MAX;
861
862 ts->timer_expires = min_t(u64, expires, next_tick);
863
864out:
865 return ts->timer_expires;
866}
867
868static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
869{
870 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
871 u64 basemono = ts->timer_expires_base;
872 u64 expires = ts->timer_expires;
873 ktime_t tick = expires;
874
875 /* Make sure we won't be trying to stop it twice in a row. */
876 ts->timer_expires_base = 0;
877
878 /*
879 * If this CPU is the one which updates jiffies, then give up
880 * the assignment and let it be taken by the CPU which runs
881 * the tick timer next, which might be this CPU as well. If we
882 * don't drop this here the jiffies might be stale and
883 * do_timer() never invoked. Keep track of the fact that it
884 * was the one which had the do_timer() duty last.
885 */
886 if (cpu == tick_do_timer_cpu) {
887 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
888 ts->do_timer_last = 1;
889 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
890 ts->do_timer_last = 0;
891 }
892
893 /* Skip reprogram of event if its not changed */
894 if (ts->tick_stopped && (expires == ts->next_tick)) {
895 /* Sanity check: make sure clockevent is actually programmed */
896 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
897 return;
898
899 WARN_ON_ONCE(1);
900 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
901 basemono, ts->next_tick, dev->next_event,
902 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
903 }
904
905 /*
906 * nohz_stop_sched_tick can be called several times before
907 * the nohz_restart_sched_tick is called. This happens when
908 * interrupts arrive which do not cause a reschedule. In the
909 * first call we save the current tick time, so we can restart
910 * the scheduler tick in nohz_restart_sched_tick.
911 */
912 if (!ts->tick_stopped) {
913 calc_load_nohz_start();
914 quiet_vmstat();
915
916 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
917 ts->tick_stopped = 1;
918 trace_tick_stop(1, TICK_DEP_MASK_NONE);
919 }
920
921 ts->next_tick = tick;
922
923 /*
924 * If the expiration time == KTIME_MAX, then we simply stop
925 * the tick timer.
926 */
927 if (unlikely(expires == KTIME_MAX)) {
928 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
929 hrtimer_cancel(&ts->sched_timer);
930 else
931 tick_program_event(KTIME_MAX, 1);
932 return;
933 }
934
935 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
936 hrtimer_start(&ts->sched_timer, tick,
937 HRTIMER_MODE_ABS_PINNED_HARD);
938 } else {
939 hrtimer_set_expires(&ts->sched_timer, tick);
940 tick_program_event(tick, 1);
941 }
942}
943
944static void tick_nohz_retain_tick(struct tick_sched *ts)
945{
946 ts->timer_expires_base = 0;
947}
948
949#ifdef CONFIG_NO_HZ_FULL
950static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
951{
952 if (tick_nohz_next_event(ts, cpu))
953 tick_nohz_stop_tick(ts, cpu);
954 else
955 tick_nohz_retain_tick(ts);
956}
957#endif /* CONFIG_NO_HZ_FULL */
958
959static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
960{
961 /* Update jiffies first */
962 tick_do_update_jiffies64(now);
963
964 /*
965 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
966 * the clock forward checks in the enqueue path:
967 */
968 timer_clear_idle();
969
970 calc_load_nohz_stop();
971 touch_softlockup_watchdog_sched();
972 /*
973 * Cancel the scheduled timer and restore the tick
974 */
975 ts->tick_stopped = 0;
976 tick_nohz_restart(ts, now);
977}
978
979static void __tick_nohz_full_update_tick(struct tick_sched *ts,
980 ktime_t now)
981{
982#ifdef CONFIG_NO_HZ_FULL
983 int cpu = smp_processor_id();
984
985 if (can_stop_full_tick(cpu, ts))
986 tick_nohz_stop_sched_tick(ts, cpu);
987 else if (ts->tick_stopped)
988 tick_nohz_restart_sched_tick(ts, now);
989#endif
990}
991
992static void tick_nohz_full_update_tick(struct tick_sched *ts)
993{
994 if (!tick_nohz_full_cpu(smp_processor_id()))
995 return;
996
997 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
998 return;
999
1000 __tick_nohz_full_update_tick(ts, ktime_get());
1001}
1002
1003/*
1004 * A pending softirq outside an IRQ (or softirq disabled section) context
1005 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1006 * reach here due to the need_resched() early check in can_stop_idle_tick().
1007 *
1008 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1009 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1010 * triggering the below since wakep_softirqd() is ignored.
1011 *
1012 */
1013static bool report_idle_softirq(void)
1014{
1015 static int ratelimit;
1016 unsigned int pending = local_softirq_pending();
1017
1018 if (likely(!pending))
1019 return false;
1020
1021 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1022 if (!cpu_active(smp_processor_id())) {
1023 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1024 if (!pending)
1025 return false;
1026 }
1027
1028 if (ratelimit < 10)
1029 return false;
1030
1031 /* On RT, softirqs handling may be waiting on some lock */
1032 if (!local_bh_blocked())
1033 return false;
1034
1035 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1036 pending);
1037 ratelimit++;
1038
1039 return true;
1040}
1041
1042static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1043{
1044 /*
1045 * If this CPU is offline and it is the one which updates
1046 * jiffies, then give up the assignment and let it be taken by
1047 * the CPU which runs the tick timer next. If we don't drop
1048 * this here the jiffies might be stale and do_timer() never
1049 * invoked.
1050 */
1051 if (unlikely(!cpu_online(cpu))) {
1052 if (cpu == tick_do_timer_cpu)
1053 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1054 /*
1055 * Make sure the CPU doesn't get fooled by obsolete tick
1056 * deadline if it comes back online later.
1057 */
1058 ts->next_tick = 0;
1059 return false;
1060 }
1061
1062 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1063 return false;
1064
1065 if (need_resched())
1066 return false;
1067
1068 if (unlikely(report_idle_softirq()))
1069 return false;
1070
1071 if (tick_nohz_full_enabled()) {
1072 /*
1073 * Keep the tick alive to guarantee timekeeping progression
1074 * if there are full dynticks CPUs around
1075 */
1076 if (tick_do_timer_cpu == cpu)
1077 return false;
1078
1079 /* Should not happen for nohz-full */
1080 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1081 return false;
1082 }
1083
1084 return true;
1085}
1086
1087static void __tick_nohz_idle_stop_tick(struct tick_sched *ts)
1088{
1089 ktime_t expires;
1090 int cpu = smp_processor_id();
1091
1092 /*
1093 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1094 * tick timer expiration time is known already.
1095 */
1096 if (ts->timer_expires_base)
1097 expires = ts->timer_expires;
1098 else if (can_stop_idle_tick(cpu, ts))
1099 expires = tick_nohz_next_event(ts, cpu);
1100 else
1101 return;
1102
1103 ts->idle_calls++;
1104
1105 if (expires > 0LL) {
1106 int was_stopped = ts->tick_stopped;
1107
1108 tick_nohz_stop_tick(ts, cpu);
1109
1110 ts->idle_sleeps++;
1111 ts->idle_expires = expires;
1112
1113 if (!was_stopped && ts->tick_stopped) {
1114 ts->idle_jiffies = ts->last_jiffies;
1115 nohz_balance_enter_idle(cpu);
1116 }
1117 } else {
1118 tick_nohz_retain_tick(ts);
1119 }
1120}
1121
1122/**
1123 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1124 *
1125 * When the next event is more than a tick into the future, stop the idle tick
1126 */
1127void tick_nohz_idle_stop_tick(void)
1128{
1129 __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched));
1130}
1131
1132void tick_nohz_idle_retain_tick(void)
1133{
1134 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1135 /*
1136 * Undo the effect of get_next_timer_interrupt() called from
1137 * tick_nohz_next_event().
1138 */
1139 timer_clear_idle();
1140}
1141
1142/**
1143 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1144 *
1145 * Called when we start the idle loop.
1146 */
1147void tick_nohz_idle_enter(void)
1148{
1149 struct tick_sched *ts;
1150
1151 lockdep_assert_irqs_enabled();
1152
1153 local_irq_disable();
1154
1155 ts = this_cpu_ptr(&tick_cpu_sched);
1156
1157 WARN_ON_ONCE(ts->timer_expires_base);
1158
1159 ts->inidle = 1;
1160 tick_nohz_start_idle(ts);
1161
1162 local_irq_enable();
1163}
1164
1165/**
1166 * tick_nohz_irq_exit - update next tick event from interrupt exit
1167 *
1168 * When an interrupt fires while we are idle and it doesn't cause
1169 * a reschedule, it may still add, modify or delete a timer, enqueue
1170 * an RCU callback, etc...
1171 * So we need to re-calculate and reprogram the next tick event.
1172 */
1173void tick_nohz_irq_exit(void)
1174{
1175 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1176
1177 if (ts->inidle)
1178 tick_nohz_start_idle(ts);
1179 else
1180 tick_nohz_full_update_tick(ts);
1181}
1182
1183/**
1184 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1185 */
1186bool tick_nohz_idle_got_tick(void)
1187{
1188 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1189
1190 if (ts->got_idle_tick) {
1191 ts->got_idle_tick = 0;
1192 return true;
1193 }
1194 return false;
1195}
1196
1197/**
1198 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1199 * or the tick, whatever that expires first. Note that, if the tick has been
1200 * stopped, it returns the next hrtimer.
1201 *
1202 * Called from power state control code with interrupts disabled
1203 */
1204ktime_t tick_nohz_get_next_hrtimer(void)
1205{
1206 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1207}
1208
1209/**
1210 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1211 * @delta_next: duration until the next event if the tick cannot be stopped
1212 *
1213 * Called from power state control code with interrupts disabled.
1214 *
1215 * The return value of this function and/or the value returned by it through the
1216 * @delta_next pointer can be negative which must be taken into account by its
1217 * callers.
1218 */
1219ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1220{
1221 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1222 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1223 int cpu = smp_processor_id();
1224 /*
1225 * The idle entry time is expected to be a sufficient approximation of
1226 * the current time at this point.
1227 */
1228 ktime_t now = ts->idle_entrytime;
1229 ktime_t next_event;
1230
1231 WARN_ON_ONCE(!ts->inidle);
1232
1233 *delta_next = ktime_sub(dev->next_event, now);
1234
1235 if (!can_stop_idle_tick(cpu, ts))
1236 return *delta_next;
1237
1238 next_event = tick_nohz_next_event(ts, cpu);
1239 if (!next_event)
1240 return *delta_next;
1241
1242 /*
1243 * If the next highres timer to expire is earlier than next_event, the
1244 * idle governor needs to know that.
1245 */
1246 next_event = min_t(u64, next_event,
1247 hrtimer_next_event_without(&ts->sched_timer));
1248
1249 return ktime_sub(next_event, now);
1250}
1251
1252/**
1253 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1254 * for a particular CPU.
1255 *
1256 * Called from the schedutil frequency scaling governor in scheduler context.
1257 */
1258unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1259{
1260 struct tick_sched *ts = tick_get_tick_sched(cpu);
1261
1262 return ts->idle_calls;
1263}
1264
1265/**
1266 * tick_nohz_get_idle_calls - return the current idle calls counter value
1267 *
1268 * Called from the schedutil frequency scaling governor in scheduler context.
1269 */
1270unsigned long tick_nohz_get_idle_calls(void)
1271{
1272 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1273
1274 return ts->idle_calls;
1275}
1276
1277static void tick_nohz_account_idle_time(struct tick_sched *ts,
1278 ktime_t now)
1279{
1280 unsigned long ticks;
1281
1282 ts->idle_exittime = now;
1283
1284 if (vtime_accounting_enabled_this_cpu())
1285 return;
1286 /*
1287 * We stopped the tick in idle. Update process times would miss the
1288 * time we slept as update_process_times does only a 1 tick
1289 * accounting. Enforce that this is accounted to idle !
1290 */
1291 ticks = jiffies - ts->idle_jiffies;
1292 /*
1293 * We might be one off. Do not randomly account a huge number of ticks!
1294 */
1295 if (ticks && ticks < LONG_MAX)
1296 account_idle_ticks(ticks);
1297}
1298
1299void tick_nohz_idle_restart_tick(void)
1300{
1301 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1302
1303 if (ts->tick_stopped) {
1304 ktime_t now = ktime_get();
1305 tick_nohz_restart_sched_tick(ts, now);
1306 tick_nohz_account_idle_time(ts, now);
1307 }
1308}
1309
1310static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1311{
1312 if (tick_nohz_full_cpu(smp_processor_id()))
1313 __tick_nohz_full_update_tick(ts, now);
1314 else
1315 tick_nohz_restart_sched_tick(ts, now);
1316
1317 tick_nohz_account_idle_time(ts, now);
1318}
1319
1320/**
1321 * tick_nohz_idle_exit - restart the idle tick from the idle task
1322 *
1323 * Restart the idle tick when the CPU is woken up from idle
1324 * This also exit the RCU extended quiescent state. The CPU
1325 * can use RCU again after this function is called.
1326 */
1327void tick_nohz_idle_exit(void)
1328{
1329 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1330 bool idle_active, tick_stopped;
1331 ktime_t now;
1332
1333 local_irq_disable();
1334
1335 WARN_ON_ONCE(!ts->inidle);
1336 WARN_ON_ONCE(ts->timer_expires_base);
1337
1338 ts->inidle = 0;
1339 idle_active = ts->idle_active;
1340 tick_stopped = ts->tick_stopped;
1341
1342 if (idle_active || tick_stopped)
1343 now = ktime_get();
1344
1345 if (idle_active)
1346 tick_nohz_stop_idle(ts, now);
1347
1348 if (tick_stopped)
1349 tick_nohz_idle_update_tick(ts, now);
1350
1351 local_irq_enable();
1352}
1353
1354/*
1355 * The nohz low res interrupt handler
1356 */
1357static void tick_nohz_handler(struct clock_event_device *dev)
1358{
1359 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1360 struct pt_regs *regs = get_irq_regs();
1361 ktime_t now = ktime_get();
1362
1363 dev->next_event = KTIME_MAX;
1364
1365 tick_sched_do_timer(ts, now);
1366 tick_sched_handle(ts, regs);
1367
1368 if (unlikely(ts->tick_stopped)) {
1369 /*
1370 * The clockevent device is not reprogrammed, so change the
1371 * clock event device to ONESHOT_STOPPED to avoid spurious
1372 * interrupts on devices which might not be truly one shot.
1373 */
1374 tick_program_event(KTIME_MAX, 1);
1375 return;
1376 }
1377
1378 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1379 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1380}
1381
1382static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1383{
1384 if (!tick_nohz_enabled)
1385 return;
1386 ts->nohz_mode = mode;
1387 /* One update is enough */
1388 if (!test_and_set_bit(0, &tick_nohz_active))
1389 timers_update_nohz();
1390}
1391
1392/**
1393 * tick_nohz_switch_to_nohz - switch to nohz mode
1394 */
1395static void tick_nohz_switch_to_nohz(void)
1396{
1397 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1398 ktime_t next;
1399
1400 if (!tick_nohz_enabled)
1401 return;
1402
1403 if (tick_switch_to_oneshot(tick_nohz_handler))
1404 return;
1405
1406 /*
1407 * Recycle the hrtimer in ts, so we can share the
1408 * hrtimer_forward with the highres code.
1409 */
1410 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1411 /* Get the next period */
1412 next = tick_init_jiffy_update();
1413
1414 hrtimer_set_expires(&ts->sched_timer, next);
1415 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1418}
1419
1420static inline void tick_nohz_irq_enter(void)
1421{
1422 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1423 ktime_t now;
1424
1425 if (!ts->idle_active && !ts->tick_stopped)
1426 return;
1427 now = ktime_get();
1428 if (ts->idle_active)
1429 tick_nohz_stop_idle(ts, now);
1430 /*
1431 * If all CPUs are idle. We may need to update a stale jiffies value.
1432 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1433 * alive but it might be busy looping with interrupts disabled in some
1434 * rare case (typically stop machine). So we must make sure we have a
1435 * last resort.
1436 */
1437 if (ts->tick_stopped)
1438 tick_nohz_update_jiffies(now);
1439}
1440
1441#else
1442
1443static inline void tick_nohz_switch_to_nohz(void) { }
1444static inline void tick_nohz_irq_enter(void) { }
1445static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1446
1447#endif /* CONFIG_NO_HZ_COMMON */
1448
1449/*
1450 * Called from irq_enter to notify about the possible interruption of idle()
1451 */
1452void tick_irq_enter(void)
1453{
1454 tick_check_oneshot_broadcast_this_cpu();
1455 tick_nohz_irq_enter();
1456}
1457
1458/*
1459 * High resolution timer specific code
1460 */
1461#ifdef CONFIG_HIGH_RES_TIMERS
1462/*
1463 * We rearm the timer until we get disabled by the idle code.
1464 * Called with interrupts disabled.
1465 */
1466static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1467{
1468 struct tick_sched *ts =
1469 container_of(timer, struct tick_sched, sched_timer);
1470 struct pt_regs *regs = get_irq_regs();
1471 ktime_t now = ktime_get();
1472
1473 tick_sched_do_timer(ts, now);
1474
1475 /*
1476 * Do not call, when we are not in irq context and have
1477 * no valid regs pointer
1478 */
1479 if (regs)
1480 tick_sched_handle(ts, regs);
1481 else
1482 ts->next_tick = 0;
1483
1484 /* No need to reprogram if we are in idle or full dynticks mode */
1485 if (unlikely(ts->tick_stopped))
1486 return HRTIMER_NORESTART;
1487
1488 hrtimer_forward(timer, now, TICK_NSEC);
1489
1490 return HRTIMER_RESTART;
1491}
1492
1493static int sched_skew_tick;
1494
1495static int __init skew_tick(char *str)
1496{
1497 get_option(&str, &sched_skew_tick);
1498
1499 return 0;
1500}
1501early_param("skew_tick", skew_tick);
1502
1503/**
1504 * tick_setup_sched_timer - setup the tick emulation timer
1505 */
1506void tick_setup_sched_timer(void)
1507{
1508 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1509 ktime_t now = ktime_get();
1510
1511 /*
1512 * Emulate tick processing via per-CPU hrtimers:
1513 */
1514 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1515 ts->sched_timer.function = tick_sched_timer;
1516
1517 /* Get the next period (per-CPU) */
1518 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1519
1520 /* Offset the tick to avert jiffies_lock contention. */
1521 if (sched_skew_tick) {
1522 u64 offset = TICK_NSEC >> 1;
1523 do_div(offset, num_possible_cpus());
1524 offset *= smp_processor_id();
1525 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1526 }
1527
1528 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1529 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1530 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1531}
1532#endif /* HIGH_RES_TIMERS */
1533
1534#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1535void tick_cancel_sched_timer(int cpu)
1536{
1537 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1538
1539# ifdef CONFIG_HIGH_RES_TIMERS
1540 if (ts->sched_timer.base)
1541 hrtimer_cancel(&ts->sched_timer);
1542# endif
1543
1544 memset(ts, 0, sizeof(*ts));
1545}
1546#endif
1547
1548/*
1549 * Async notification about clocksource changes
1550 */
1551void tick_clock_notify(void)
1552{
1553 int cpu;
1554
1555 for_each_possible_cpu(cpu)
1556 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1557}
1558
1559/*
1560 * Async notification about clock event changes
1561 */
1562void tick_oneshot_notify(void)
1563{
1564 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1565
1566 set_bit(0, &ts->check_clocks);
1567}
1568
1569/*
1570 * Check, if a change happened, which makes oneshot possible.
1571 *
1572 * Called cyclic from the hrtimer softirq (driven by the timer
1573 * softirq) allow_nohz signals, that we can switch into low-res nohz
1574 * mode, because high resolution timers are disabled (either compile
1575 * or runtime). Called with interrupts disabled.
1576 */
1577int tick_check_oneshot_change(int allow_nohz)
1578{
1579 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1580
1581 if (!test_and_clear_bit(0, &ts->check_clocks))
1582 return 0;
1583
1584 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1585 return 0;
1586
1587 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1588 return 0;
1589
1590 if (!allow_nohz)
1591 return 1;
1592
1593 tick_nohz_switch_to_nohz();
1594 return 0;
1595}
1/*
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}