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