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