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.tv64 < tick_period.tv64)
62 return;
63
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
66
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
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.tv64 >= tick_period.tv64)) {
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.tv64 == 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 themself 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(struct tick_sched *ts)
190{
191 WARN_ON_ONCE(!irqs_disabled());
192
193 if (check_tick_dependency(&tick_dep_mask))
194 return false;
195
196 if (check_tick_dependency(&ts->tick_dep_mask))
197 return false;
198
199 if (check_tick_dependency(¤t->tick_dep_mask))
200 return false;
201
202 if (check_tick_dependency(¤t->signal->tick_dep_mask))
203 return false;
204
205 return true;
206}
207
208static void nohz_full_kick_func(struct irq_work *work)
209{
210 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
211}
212
213static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
214 .func = nohz_full_kick_func,
215};
216
217/*
218 * Kick this CPU if it's full dynticks in order to force it to
219 * re-evaluate its dependency on the tick and restart it if necessary.
220 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
221 * is NMI safe.
222 */
223static void tick_nohz_full_kick(void)
224{
225 if (!tick_nohz_full_cpu(smp_processor_id()))
226 return;
227
228 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
229}
230
231/*
232 * Kick the CPU if it's full dynticks in order to force it to
233 * re-evaluate its dependency on the tick and restart it if necessary.
234 */
235void tick_nohz_full_kick_cpu(int cpu)
236{
237 if (!tick_nohz_full_cpu(cpu))
238 return;
239
240 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
241}
242
243/*
244 * Kick all full dynticks CPUs in order to force these to re-evaluate
245 * their dependency on the tick and restart it if necessary.
246 */
247static void tick_nohz_full_kick_all(void)
248{
249 int cpu;
250
251 if (!tick_nohz_full_running)
252 return;
253
254 preempt_disable();
255 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
256 tick_nohz_full_kick_cpu(cpu);
257 preempt_enable();
258}
259
260static void tick_nohz_dep_set_all(atomic_t *dep,
261 enum tick_dep_bits bit)
262{
263 int prev;
264
265 prev = atomic_fetch_or(dep, BIT(bit));
266 if (!prev)
267 tick_nohz_full_kick_all();
268}
269
270/*
271 * Set a global tick dependency. Used by perf events that rely on freq and
272 * by unstable clock.
273 */
274void tick_nohz_dep_set(enum tick_dep_bits bit)
275{
276 tick_nohz_dep_set_all(&tick_dep_mask, bit);
277}
278
279void tick_nohz_dep_clear(enum tick_dep_bits bit)
280{
281 atomic_andnot(BIT(bit), &tick_dep_mask);
282}
283
284/*
285 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
286 * manage events throttling.
287 */
288void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
289{
290 int prev;
291 struct tick_sched *ts;
292
293 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
294
295 prev = atomic_fetch_or(&ts->tick_dep_mask, BIT(bit));
296 if (!prev) {
297 preempt_disable();
298 /* Perf needs local kick that is NMI safe */
299 if (cpu == smp_processor_id()) {
300 tick_nohz_full_kick();
301 } else {
302 /* Remote irq work not NMI-safe */
303 if (!WARN_ON_ONCE(in_nmi()))
304 tick_nohz_full_kick_cpu(cpu);
305 }
306 preempt_enable();
307 }
308}
309
310void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
311{
312 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
313
314 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
315}
316
317/*
318 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
319 * per task timers.
320 */
321void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
322{
323 /*
324 * We could optimize this with just kicking the target running the task
325 * if that noise matters for nohz full users.
326 */
327 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
328}
329
330void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
331{
332 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
333}
334
335/*
336 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
337 * per process timers.
338 */
339void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
340{
341 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
342}
343
344void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
345{
346 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
347}
348
349/*
350 * Re-evaluate the need for the tick as we switch the current task.
351 * It might need the tick due to per task/process properties:
352 * perf events, posix cpu timers, ...
353 */
354void __tick_nohz_task_switch(void)
355{
356 unsigned long flags;
357 struct tick_sched *ts;
358
359 local_irq_save(flags);
360
361 if (!tick_nohz_full_cpu(smp_processor_id()))
362 goto out;
363
364 ts = this_cpu_ptr(&tick_cpu_sched);
365
366 if (ts->tick_stopped) {
367 if (atomic_read(¤t->tick_dep_mask) ||
368 atomic_read(¤t->signal->tick_dep_mask))
369 tick_nohz_full_kick();
370 }
371out:
372 local_irq_restore(flags);
373}
374
375/* Parse the boot-time nohz CPU list from the kernel parameters. */
376static int __init tick_nohz_full_setup(char *str)
377{
378 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
379 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
380 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
381 free_bootmem_cpumask_var(tick_nohz_full_mask);
382 return 1;
383 }
384 tick_nohz_full_running = true;
385
386 return 1;
387}
388__setup("nohz_full=", tick_nohz_full_setup);
389
390static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
391 unsigned long action,
392 void *hcpu)
393{
394 unsigned int cpu = (unsigned long)hcpu;
395
396 switch (action & ~CPU_TASKS_FROZEN) {
397 case CPU_DOWN_PREPARE:
398 /*
399 * The boot CPU handles housekeeping duty (unbound timers,
400 * workqueues, timekeeping, ...) on behalf of full dynticks
401 * CPUs. It must remain online when nohz full is enabled.
402 */
403 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
404 return NOTIFY_BAD;
405 break;
406 }
407 return NOTIFY_OK;
408}
409
410static int tick_nohz_init_all(void)
411{
412 int err = -1;
413
414#ifdef CONFIG_NO_HZ_FULL_ALL
415 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
416 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
417 return err;
418 }
419 err = 0;
420 cpumask_setall(tick_nohz_full_mask);
421 tick_nohz_full_running = true;
422#endif
423 return err;
424}
425
426void __init tick_nohz_init(void)
427{
428 int cpu;
429
430 if (!tick_nohz_full_running) {
431 if (tick_nohz_init_all() < 0)
432 return;
433 }
434
435 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
436 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
437 cpumask_clear(tick_nohz_full_mask);
438 tick_nohz_full_running = false;
439 return;
440 }
441
442 /*
443 * Full dynticks uses irq work to drive the tick rescheduling on safe
444 * locking contexts. But then we need irq work to raise its own
445 * interrupts to avoid circular dependency on the tick
446 */
447 if (!arch_irq_work_has_interrupt()) {
448 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
449 cpumask_clear(tick_nohz_full_mask);
450 cpumask_copy(housekeeping_mask, cpu_possible_mask);
451 tick_nohz_full_running = false;
452 return;
453 }
454
455 cpu = smp_processor_id();
456
457 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
458 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
459 cpu);
460 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
461 }
462
463 cpumask_andnot(housekeeping_mask,
464 cpu_possible_mask, tick_nohz_full_mask);
465
466 for_each_cpu(cpu, tick_nohz_full_mask)
467 context_tracking_cpu_set(cpu);
468
469 cpu_notifier(tick_nohz_cpu_down_callback, 0);
470 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
471 cpumask_pr_args(tick_nohz_full_mask));
472
473 /*
474 * We need at least one CPU to handle housekeeping work such
475 * as timekeeping, unbound timers, workqueues, ...
476 */
477 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
478}
479#endif
480
481/*
482 * NOHZ - aka dynamic tick functionality
483 */
484#ifdef CONFIG_NO_HZ_COMMON
485/*
486 * NO HZ enabled ?
487 */
488bool tick_nohz_enabled __read_mostly = true;
489unsigned long tick_nohz_active __read_mostly;
490/*
491 * Enable / Disable tickless mode
492 */
493static int __init setup_tick_nohz(char *str)
494{
495 return (kstrtobool(str, &tick_nohz_enabled) == 0);
496}
497
498__setup("nohz=", setup_tick_nohz);
499
500int tick_nohz_tick_stopped(void)
501{
502 return __this_cpu_read(tick_cpu_sched.tick_stopped);
503}
504
505/**
506 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
507 *
508 * Called from interrupt entry when the CPU was idle
509 *
510 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
511 * must be updated. Otherwise an interrupt handler could use a stale jiffy
512 * value. We do this unconditionally on any cpu, as we don't know whether the
513 * cpu, which has the update task assigned is in a long sleep.
514 */
515static void tick_nohz_update_jiffies(ktime_t now)
516{
517 unsigned long flags;
518
519 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
520
521 local_irq_save(flags);
522 tick_do_update_jiffies64(now);
523 local_irq_restore(flags);
524
525 touch_softlockup_watchdog_sched();
526}
527
528/*
529 * Updates the per cpu time idle statistics counters
530 */
531static void
532update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
533{
534 ktime_t delta;
535
536 if (ts->idle_active) {
537 delta = ktime_sub(now, ts->idle_entrytime);
538 if (nr_iowait_cpu(cpu) > 0)
539 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
540 else
541 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
542 ts->idle_entrytime = now;
543 }
544
545 if (last_update_time)
546 *last_update_time = ktime_to_us(now);
547
548}
549
550static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
551{
552 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
553 ts->idle_active = 0;
554
555 sched_clock_idle_wakeup_event(0);
556}
557
558static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
559{
560 ktime_t now = ktime_get();
561
562 ts->idle_entrytime = now;
563 ts->idle_active = 1;
564 sched_clock_idle_sleep_event();
565 return now;
566}
567
568/**
569 * get_cpu_idle_time_us - get the total idle time of a cpu
570 * @cpu: CPU number to query
571 * @last_update_time: variable to store update time in. Do not update
572 * counters if NULL.
573 *
574 * Return the cummulative idle time (since boot) for a given
575 * CPU, in microseconds.
576 *
577 * This time is measured via accounting rather than sampling,
578 * and is as accurate as ktime_get() is.
579 *
580 * This function returns -1 if NOHZ is not enabled.
581 */
582u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
583{
584 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
585 ktime_t now, idle;
586
587 if (!tick_nohz_active)
588 return -1;
589
590 now = ktime_get();
591 if (last_update_time) {
592 update_ts_time_stats(cpu, ts, now, last_update_time);
593 idle = ts->idle_sleeptime;
594 } else {
595 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
596 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
597
598 idle = ktime_add(ts->idle_sleeptime, delta);
599 } else {
600 idle = ts->idle_sleeptime;
601 }
602 }
603
604 return ktime_to_us(idle);
605
606}
607EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
608
609/**
610 * get_cpu_iowait_time_us - get the total iowait time of a cpu
611 * @cpu: CPU number to query
612 * @last_update_time: variable to store update time in. Do not update
613 * counters if NULL.
614 *
615 * Return the cummulative iowait time (since boot) for a given
616 * CPU, in microseconds.
617 *
618 * This time is measured via accounting rather than sampling,
619 * and is as accurate as ktime_get() is.
620 *
621 * This function returns -1 if NOHZ is not enabled.
622 */
623u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
624{
625 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
626 ktime_t now, iowait;
627
628 if (!tick_nohz_active)
629 return -1;
630
631 now = ktime_get();
632 if (last_update_time) {
633 update_ts_time_stats(cpu, ts, now, last_update_time);
634 iowait = ts->iowait_sleeptime;
635 } else {
636 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
637 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
638
639 iowait = ktime_add(ts->iowait_sleeptime, delta);
640 } else {
641 iowait = ts->iowait_sleeptime;
642 }
643 }
644
645 return ktime_to_us(iowait);
646}
647EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
648
649static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
650{
651 hrtimer_cancel(&ts->sched_timer);
652 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
653
654 /* Forward the time to expire in the future */
655 hrtimer_forward(&ts->sched_timer, now, tick_period);
656
657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
658 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
659 else
660 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
661}
662
663static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
664 ktime_t now, int cpu)
665{
666 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
667 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
668 unsigned long seq, basejiff;
669 ktime_t tick;
670
671 /* Read jiffies and the time when jiffies were updated last */
672 do {
673 seq = read_seqbegin(&jiffies_lock);
674 basemono = last_jiffies_update.tv64;
675 basejiff = jiffies;
676 } while (read_seqretry(&jiffies_lock, seq));
677 ts->last_jiffies = basejiff;
678
679 if (rcu_needs_cpu(basemono, &next_rcu) ||
680 arch_needs_cpu() || irq_work_needs_cpu()) {
681 next_tick = basemono + TICK_NSEC;
682 } else {
683 /*
684 * Get the next pending timer. If high resolution
685 * timers are enabled this only takes the timer wheel
686 * timers into account. If high resolution timers are
687 * disabled this also looks at the next expiring
688 * hrtimer.
689 */
690 next_tmr = get_next_timer_interrupt(basejiff, basemono);
691 ts->next_timer = next_tmr;
692 /* Take the next rcu event into account */
693 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
694 }
695
696 /*
697 * If the tick is due in the next period, keep it ticking or
698 * force prod the timer.
699 */
700 delta = next_tick - basemono;
701 if (delta <= (u64)TICK_NSEC) {
702 tick.tv64 = 0;
703 /*
704 * We've not stopped the tick yet, and there's a timer in the
705 * next period, so no point in stopping it either, bail.
706 */
707 if (!ts->tick_stopped)
708 goto out;
709
710 /*
711 * If, OTOH, we did stop it, but there's a pending (expired)
712 * timer reprogram the timer hardware to fire now.
713 *
714 * We will not restart the tick proper, just prod the timer
715 * hardware into firing an interrupt to process the pending
716 * timers. Just like tick_irq_exit() will not restart the tick
717 * for 'normal' interrupts.
718 *
719 * Only once we exit the idle loop will we re-enable the tick,
720 * see tick_nohz_idle_exit().
721 */
722 if (delta == 0) {
723 tick_nohz_restart(ts, now);
724 goto out;
725 }
726 }
727
728 /*
729 * If this cpu is the one which updates jiffies, then give up
730 * the assignment and let it be taken by the cpu which runs
731 * the tick timer next, which might be this cpu as well. If we
732 * don't drop this here the jiffies might be stale and
733 * do_timer() never invoked. Keep track of the fact that it
734 * was the one which had the do_timer() duty last. If this cpu
735 * is the one which had the do_timer() duty last, we limit the
736 * sleep time to the timekeeping max_deferement value.
737 * Otherwise we can sleep as long as we want.
738 */
739 delta = timekeeping_max_deferment();
740 if (cpu == tick_do_timer_cpu) {
741 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
742 ts->do_timer_last = 1;
743 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
744 delta = KTIME_MAX;
745 ts->do_timer_last = 0;
746 } else if (!ts->do_timer_last) {
747 delta = KTIME_MAX;
748 }
749
750#ifdef CONFIG_NO_HZ_FULL
751 /* Limit the tick delta to the maximum scheduler deferment */
752 if (!ts->inidle)
753 delta = min(delta, scheduler_tick_max_deferment());
754#endif
755
756 /* Calculate the next expiry time */
757 if (delta < (KTIME_MAX - basemono))
758 expires = basemono + delta;
759 else
760 expires = KTIME_MAX;
761
762 expires = min_t(u64, expires, next_tick);
763 tick.tv64 = expires;
764
765 /* Skip reprogram of event if its not changed */
766 if (ts->tick_stopped && (expires == dev->next_event.tv64))
767 goto out;
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 nohz_balance_enter_idle(cpu);
778 calc_load_enter_idle();
779
780 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
781 ts->tick_stopped = 1;
782 trace_tick_stop(1, TICK_DEP_MASK_NONE);
783 }
784
785 /*
786 * If the expiration time == KTIME_MAX, then we simply stop
787 * the tick timer.
788 */
789 if (unlikely(expires == KTIME_MAX)) {
790 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
791 hrtimer_cancel(&ts->sched_timer);
792 goto out;
793 }
794
795 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
796 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
797 else
798 tick_program_event(tick, 1);
799out:
800 /* Update the estimated sleep length */
801 ts->sleep_length = ktime_sub(dev->next_event, now);
802 return tick;
803}
804
805static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active)
806{
807 /* Update jiffies first */
808 tick_do_update_jiffies64(now);
809 update_cpu_load_nohz(active);
810
811 calc_load_exit_idle();
812 touch_softlockup_watchdog_sched();
813 /*
814 * Cancel the scheduled timer and restore the tick
815 */
816 ts->tick_stopped = 0;
817 ts->idle_exittime = now;
818
819 tick_nohz_restart(ts, now);
820}
821
822static void tick_nohz_full_update_tick(struct tick_sched *ts)
823{
824#ifdef CONFIG_NO_HZ_FULL
825 int cpu = smp_processor_id();
826
827 if (!tick_nohz_full_cpu(cpu))
828 return;
829
830 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
831 return;
832
833 if (can_stop_full_tick(ts))
834 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
835 else if (ts->tick_stopped)
836 tick_nohz_restart_sched_tick(ts, ktime_get(), 1);
837#endif
838}
839
840static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
841{
842 /*
843 * If this cpu is offline and it is the one which updates
844 * jiffies, then give up the assignment and let it be taken by
845 * the cpu which runs the tick timer next. If we don't drop
846 * this here the jiffies might be stale and do_timer() never
847 * invoked.
848 */
849 if (unlikely(!cpu_online(cpu))) {
850 if (cpu == tick_do_timer_cpu)
851 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
852 return false;
853 }
854
855 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
856 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
857 return false;
858 }
859
860 if (need_resched())
861 return false;
862
863 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
864 static int ratelimit;
865
866 if (ratelimit < 10 &&
867 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
868 pr_warn("NOHZ: local_softirq_pending %02x\n",
869 (unsigned int) local_softirq_pending());
870 ratelimit++;
871 }
872 return false;
873 }
874
875 if (tick_nohz_full_enabled()) {
876 /*
877 * Keep the tick alive to guarantee timekeeping progression
878 * if there are full dynticks CPUs around
879 */
880 if (tick_do_timer_cpu == cpu)
881 return false;
882 /*
883 * Boot safety: make sure the timekeeping duty has been
884 * assigned before entering dyntick-idle mode,
885 */
886 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
887 return false;
888 }
889
890 return true;
891}
892
893static void __tick_nohz_idle_enter(struct tick_sched *ts)
894{
895 ktime_t now, expires;
896 int cpu = smp_processor_id();
897
898 now = tick_nohz_start_idle(ts);
899
900 if (can_stop_idle_tick(cpu, ts)) {
901 int was_stopped = ts->tick_stopped;
902
903 ts->idle_calls++;
904
905 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
906 if (expires.tv64 > 0LL) {
907 ts->idle_sleeps++;
908 ts->idle_expires = expires;
909 }
910
911 if (!was_stopped && ts->tick_stopped)
912 ts->idle_jiffies = ts->last_jiffies;
913 }
914}
915
916/**
917 * tick_nohz_idle_enter - stop the idle tick from the idle task
918 *
919 * When the next event is more than a tick into the future, stop the idle tick
920 * Called when we start the idle loop.
921 *
922 * The arch is responsible of calling:
923 *
924 * - rcu_idle_enter() after its last use of RCU before the CPU is put
925 * to sleep.
926 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
927 */
928void tick_nohz_idle_enter(void)
929{
930 struct tick_sched *ts;
931
932 WARN_ON_ONCE(irqs_disabled());
933
934 /*
935 * Update the idle state in the scheduler domain hierarchy
936 * when tick_nohz_stop_sched_tick() is called from the idle loop.
937 * State will be updated to busy during the first busy tick after
938 * exiting idle.
939 */
940 set_cpu_sd_state_idle();
941
942 local_irq_disable();
943
944 ts = this_cpu_ptr(&tick_cpu_sched);
945 ts->inidle = 1;
946 __tick_nohz_idle_enter(ts);
947
948 local_irq_enable();
949}
950
951/**
952 * tick_nohz_irq_exit - update next tick event from interrupt exit
953 *
954 * When an interrupt fires while we are idle and it doesn't cause
955 * a reschedule, it may still add, modify or delete a timer, enqueue
956 * an RCU callback, etc...
957 * So we need to re-calculate and reprogram the next tick event.
958 */
959void tick_nohz_irq_exit(void)
960{
961 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
962
963 if (ts->inidle)
964 __tick_nohz_idle_enter(ts);
965 else
966 tick_nohz_full_update_tick(ts);
967}
968
969/**
970 * tick_nohz_get_sleep_length - return the length of the current sleep
971 *
972 * Called from power state control code with interrupts disabled
973 */
974ktime_t tick_nohz_get_sleep_length(void)
975{
976 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
977
978 return ts->sleep_length;
979}
980
981static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
982{
983#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
984 unsigned long ticks;
985
986 if (vtime_accounting_cpu_enabled())
987 return;
988 /*
989 * We stopped the tick in idle. Update process times would miss the
990 * time we slept as update_process_times does only a 1 tick
991 * accounting. Enforce that this is accounted to idle !
992 */
993 ticks = jiffies - ts->idle_jiffies;
994 /*
995 * We might be one off. Do not randomly account a huge number of ticks!
996 */
997 if (ticks && ticks < LONG_MAX)
998 account_idle_ticks(ticks);
999#endif
1000}
1001
1002/**
1003 * tick_nohz_idle_exit - restart the idle tick from the idle task
1004 *
1005 * Restart the idle tick when the CPU is woken up from idle
1006 * This also exit the RCU extended quiescent state. The CPU
1007 * can use RCU again after this function is called.
1008 */
1009void tick_nohz_idle_exit(void)
1010{
1011 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1012 ktime_t now;
1013
1014 local_irq_disable();
1015
1016 WARN_ON_ONCE(!ts->inidle);
1017
1018 ts->inidle = 0;
1019
1020 if (ts->idle_active || ts->tick_stopped)
1021 now = ktime_get();
1022
1023 if (ts->idle_active)
1024 tick_nohz_stop_idle(ts, now);
1025
1026 if (ts->tick_stopped) {
1027 tick_nohz_restart_sched_tick(ts, now, 0);
1028 tick_nohz_account_idle_ticks(ts);
1029 }
1030
1031 local_irq_enable();
1032}
1033
1034/*
1035 * The nohz low res interrupt handler
1036 */
1037static void tick_nohz_handler(struct clock_event_device *dev)
1038{
1039 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1040 struct pt_regs *regs = get_irq_regs();
1041 ktime_t now = ktime_get();
1042
1043 dev->next_event.tv64 = KTIME_MAX;
1044
1045 tick_sched_do_timer(now);
1046 tick_sched_handle(ts, regs);
1047
1048 /* No need to reprogram if we are running tickless */
1049 if (unlikely(ts->tick_stopped))
1050 return;
1051
1052 hrtimer_forward(&ts->sched_timer, now, tick_period);
1053 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1054}
1055
1056static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1057{
1058 if (!tick_nohz_enabled)
1059 return;
1060 ts->nohz_mode = mode;
1061 /* One update is enough */
1062 if (!test_and_set_bit(0, &tick_nohz_active))
1063 timers_update_migration(true);
1064}
1065
1066/**
1067 * tick_nohz_switch_to_nohz - switch to nohz mode
1068 */
1069static void tick_nohz_switch_to_nohz(void)
1070{
1071 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1072 ktime_t next;
1073
1074 if (!tick_nohz_enabled)
1075 return;
1076
1077 if (tick_switch_to_oneshot(tick_nohz_handler))
1078 return;
1079
1080 /*
1081 * Recycle the hrtimer in ts, so we can share the
1082 * hrtimer_forward with the highres code.
1083 */
1084 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1085 /* Get the next period */
1086 next = tick_init_jiffy_update();
1087
1088 hrtimer_set_expires(&ts->sched_timer, next);
1089 hrtimer_forward_now(&ts->sched_timer, tick_period);
1090 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1091 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1092}
1093
1094/*
1095 * When NOHZ is enabled and the tick is stopped, we need to kick the
1096 * tick timer from irq_enter() so that the jiffies update is kept
1097 * alive during long running softirqs. That's ugly as hell, but
1098 * correctness is key even if we need to fix the offending softirq in
1099 * the first place.
1100 *
1101 * Note, this is different to tick_nohz_restart. We just kick the
1102 * timer and do not touch the other magic bits which need to be done
1103 * when idle is left.
1104 */
1105static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1106{
1107#if 0
1108 /* Switch back to 2.6.27 behaviour */
1109 ktime_t delta;
1110
1111 /*
1112 * Do not touch the tick device, when the next expiry is either
1113 * already reached or less/equal than the tick period.
1114 */
1115 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1116 if (delta.tv64 <= tick_period.tv64)
1117 return;
1118
1119 tick_nohz_restart(ts, now);
1120#endif
1121}
1122
1123static inline void tick_nohz_irq_enter(void)
1124{
1125 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1126 ktime_t now;
1127
1128 if (!ts->idle_active && !ts->tick_stopped)
1129 return;
1130 now = ktime_get();
1131 if (ts->idle_active)
1132 tick_nohz_stop_idle(ts, now);
1133 if (ts->tick_stopped) {
1134 tick_nohz_update_jiffies(now);
1135 tick_nohz_kick_tick(ts, now);
1136 }
1137}
1138
1139#else
1140
1141static inline void tick_nohz_switch_to_nohz(void) { }
1142static inline void tick_nohz_irq_enter(void) { }
1143static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1144
1145#endif /* CONFIG_NO_HZ_COMMON */
1146
1147/*
1148 * Called from irq_enter to notify about the possible interruption of idle()
1149 */
1150void tick_irq_enter(void)
1151{
1152 tick_check_oneshot_broadcast_this_cpu();
1153 tick_nohz_irq_enter();
1154}
1155
1156/*
1157 * High resolution timer specific code
1158 */
1159#ifdef CONFIG_HIGH_RES_TIMERS
1160/*
1161 * We rearm the timer until we get disabled by the idle code.
1162 * Called with interrupts disabled.
1163 */
1164static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1165{
1166 struct tick_sched *ts =
1167 container_of(timer, struct tick_sched, sched_timer);
1168 struct pt_regs *regs = get_irq_regs();
1169 ktime_t now = ktime_get();
1170
1171 tick_sched_do_timer(now);
1172
1173 /*
1174 * Do not call, when we are not in irq context and have
1175 * no valid regs pointer
1176 */
1177 if (regs)
1178 tick_sched_handle(ts, regs);
1179
1180 /* No need to reprogram if we are in idle or full dynticks mode */
1181 if (unlikely(ts->tick_stopped))
1182 return HRTIMER_NORESTART;
1183
1184 hrtimer_forward(timer, now, tick_period);
1185
1186 return HRTIMER_RESTART;
1187}
1188
1189static int sched_skew_tick;
1190
1191static int __init skew_tick(char *str)
1192{
1193 get_option(&str, &sched_skew_tick);
1194
1195 return 0;
1196}
1197early_param("skew_tick", skew_tick);
1198
1199/**
1200 * tick_setup_sched_timer - setup the tick emulation timer
1201 */
1202void tick_setup_sched_timer(void)
1203{
1204 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1205 ktime_t now = ktime_get();
1206
1207 /*
1208 * Emulate tick processing via per-CPU hrtimers:
1209 */
1210 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1211 ts->sched_timer.function = tick_sched_timer;
1212
1213 /* Get the next period (per cpu) */
1214 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1215
1216 /* Offset the tick to avert jiffies_lock contention. */
1217 if (sched_skew_tick) {
1218 u64 offset = ktime_to_ns(tick_period) >> 1;
1219 do_div(offset, num_possible_cpus());
1220 offset *= smp_processor_id();
1221 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1222 }
1223
1224 hrtimer_forward(&ts->sched_timer, now, tick_period);
1225 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1226 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1227}
1228#endif /* HIGH_RES_TIMERS */
1229
1230#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1231void tick_cancel_sched_timer(int cpu)
1232{
1233 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1234
1235# ifdef CONFIG_HIGH_RES_TIMERS
1236 if (ts->sched_timer.base)
1237 hrtimer_cancel(&ts->sched_timer);
1238# endif
1239
1240 memset(ts, 0, sizeof(*ts));
1241}
1242#endif
1243
1244/**
1245 * Async notification about clocksource changes
1246 */
1247void tick_clock_notify(void)
1248{
1249 int cpu;
1250
1251 for_each_possible_cpu(cpu)
1252 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1253}
1254
1255/*
1256 * Async notification about clock event changes
1257 */
1258void tick_oneshot_notify(void)
1259{
1260 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1261
1262 set_bit(0, &ts->check_clocks);
1263}
1264
1265/**
1266 * Check, if a change happened, which makes oneshot possible.
1267 *
1268 * Called cyclic from the hrtimer softirq (driven by the timer
1269 * softirq) allow_nohz signals, that we can switch into low-res nohz
1270 * mode, because high resolution timers are disabled (either compile
1271 * or runtime). Called with interrupts disabled.
1272 */
1273int tick_check_oneshot_change(int allow_nohz)
1274{
1275 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1276
1277 if (!test_and_clear_bit(0, &ts->check_clocks))
1278 return 0;
1279
1280 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1281 return 0;
1282
1283 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1284 return 0;
1285
1286 if (!allow_nohz)
1287 return 1;
1288
1289 tick_nohz_switch_to_nohz();
1290 return 0;
1291}
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}