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/perf_event.h>
26#include <linux/context_tracking.h>
27
28#include <asm/irq_regs.h>
29
30#include "tick-internal.h"
31
32#include <trace/events/timer.h>
33
34/*
35 * Per cpu nohz control structure
36 */
37DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
38
39/*
40 * The time, when the last jiffy update happened. Protected by jiffies_lock.
41 */
42static ktime_t last_jiffies_update;
43
44struct tick_sched *tick_get_tick_sched(int cpu)
45{
46 return &per_cpu(tick_cpu_sched, cpu);
47}
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();
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
155#ifdef CONFIG_NO_HZ_FULL
156cpumask_var_t tick_nohz_full_mask;
157bool tick_nohz_full_running;
158
159static bool can_stop_full_tick(void)
160{
161 WARN_ON_ONCE(!irqs_disabled());
162
163 if (!sched_can_stop_tick()) {
164 trace_tick_stop(0, "more than 1 task in runqueue\n");
165 return false;
166 }
167
168 if (!posix_cpu_timers_can_stop_tick(current)) {
169 trace_tick_stop(0, "posix timers running\n");
170 return false;
171 }
172
173 if (!perf_event_can_stop_tick()) {
174 trace_tick_stop(0, "perf events running\n");
175 return false;
176 }
177
178 /* sched_clock_tick() needs us? */
179#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
180 /*
181 * TODO: kick full dynticks CPUs when
182 * sched_clock_stable is set.
183 */
184 if (!sched_clock_stable()) {
185 trace_tick_stop(0, "unstable sched clock\n");
186 /*
187 * Don't allow the user to think they can get
188 * full NO_HZ with this machine.
189 */
190 WARN_ONCE(tick_nohz_full_running,
191 "NO_HZ FULL will not work with unstable sched clock");
192 return false;
193 }
194#endif
195
196 return true;
197}
198
199static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
200
201/*
202 * Re-evaluate the need for the tick on the current CPU
203 * and restart it if necessary.
204 */
205void __tick_nohz_full_check(void)
206{
207 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
208
209 if (tick_nohz_full_cpu(smp_processor_id())) {
210 if (ts->tick_stopped && !is_idle_task(current)) {
211 if (!can_stop_full_tick())
212 tick_nohz_restart_sched_tick(ts, ktime_get());
213 }
214 }
215}
216
217static void nohz_full_kick_work_func(struct irq_work *work)
218{
219 __tick_nohz_full_check();
220}
221
222static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
223 .func = nohz_full_kick_work_func,
224};
225
226/*
227 * Kick the current CPU if it's full dynticks in order to force it to
228 * re-evaluate its dependency on the tick and restart it if necessary.
229 */
230void tick_nohz_full_kick(void)
231{
232 if (tick_nohz_full_cpu(smp_processor_id()))
233 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
234}
235
236static void nohz_full_kick_ipi(void *info)
237{
238 __tick_nohz_full_check();
239}
240
241/*
242 * Kick all full dynticks CPUs in order to force these to re-evaluate
243 * their dependency on the tick and restart it if necessary.
244 */
245void tick_nohz_full_kick_all(void)
246{
247 if (!tick_nohz_full_running)
248 return;
249
250 preempt_disable();
251 smp_call_function_many(tick_nohz_full_mask,
252 nohz_full_kick_ipi, NULL, false);
253 tick_nohz_full_kick();
254 preempt_enable();
255}
256
257/*
258 * Re-evaluate the need for the tick as we switch the current task.
259 * It might need the tick due to per task/process properties:
260 * perf events, posix cpu timers, ...
261 */
262void __tick_nohz_task_switch(struct task_struct *tsk)
263{
264 unsigned long flags;
265
266 local_irq_save(flags);
267
268 if (!tick_nohz_full_cpu(smp_processor_id()))
269 goto out;
270
271 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
272 tick_nohz_full_kick();
273
274out:
275 local_irq_restore(flags);
276}
277
278/* Parse the boot-time nohz CPU list from the kernel parameters. */
279static int __init tick_nohz_full_setup(char *str)
280{
281 int cpu;
282
283 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
284 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
285 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
286 return 1;
287 }
288
289 cpu = smp_processor_id();
290 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
291 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
292 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
293 }
294 tick_nohz_full_running = true;
295
296 return 1;
297}
298__setup("nohz_full=", tick_nohz_full_setup);
299
300static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
301 unsigned long action,
302 void *hcpu)
303{
304 unsigned int cpu = (unsigned long)hcpu;
305
306 switch (action & ~CPU_TASKS_FROZEN) {
307 case CPU_DOWN_PREPARE:
308 /*
309 * If we handle the timekeeping duty for full dynticks CPUs,
310 * we can't safely shutdown that CPU.
311 */
312 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
313 return NOTIFY_BAD;
314 break;
315 }
316 return NOTIFY_OK;
317}
318
319/*
320 * Worst case string length in chunks of CPU range seems 2 steps
321 * separations: 0,2,4,6,...
322 * This is NR_CPUS + sizeof('\0')
323 */
324static char __initdata nohz_full_buf[NR_CPUS + 1];
325
326static int tick_nohz_init_all(void)
327{
328 int err = -1;
329
330#ifdef CONFIG_NO_HZ_FULL_ALL
331 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
332 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n");
333 return err;
334 }
335 err = 0;
336 cpumask_setall(tick_nohz_full_mask);
337 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask);
338 tick_nohz_full_running = true;
339#endif
340 return err;
341}
342
343void __init tick_nohz_init(void)
344{
345 int cpu;
346
347 if (!tick_nohz_full_running) {
348 if (tick_nohz_init_all() < 0)
349 return;
350 }
351
352 for_each_cpu(cpu, tick_nohz_full_mask)
353 context_tracking_cpu_set(cpu);
354
355 cpu_notifier(tick_nohz_cpu_down_callback, 0);
356 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
357 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
358}
359#endif
360
361/*
362 * NOHZ - aka dynamic tick functionality
363 */
364#ifdef CONFIG_NO_HZ_COMMON
365/*
366 * NO HZ enabled ?
367 */
368static int tick_nohz_enabled __read_mostly = 1;
369int tick_nohz_active __read_mostly;
370/*
371 * Enable / Disable tickless mode
372 */
373static int __init setup_tick_nohz(char *str)
374{
375 if (!strcmp(str, "off"))
376 tick_nohz_enabled = 0;
377 else if (!strcmp(str, "on"))
378 tick_nohz_enabled = 1;
379 else
380 return 0;
381 return 1;
382}
383
384__setup("nohz=", setup_tick_nohz);
385
386/**
387 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
388 *
389 * Called from interrupt entry when the CPU was idle
390 *
391 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
392 * must be updated. Otherwise an interrupt handler could use a stale jiffy
393 * value. We do this unconditionally on any cpu, as we don't know whether the
394 * cpu, which has the update task assigned is in a long sleep.
395 */
396static void tick_nohz_update_jiffies(ktime_t now)
397{
398 unsigned long flags;
399
400 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
401
402 local_irq_save(flags);
403 tick_do_update_jiffies64(now);
404 local_irq_restore(flags);
405
406 touch_softlockup_watchdog();
407}
408
409/*
410 * Updates the per cpu time idle statistics counters
411 */
412static void
413update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
414{
415 ktime_t delta;
416
417 if (ts->idle_active) {
418 delta = ktime_sub(now, ts->idle_entrytime);
419 if (nr_iowait_cpu(cpu) > 0)
420 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
421 else
422 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
423 ts->idle_entrytime = now;
424 }
425
426 if (last_update_time)
427 *last_update_time = ktime_to_us(now);
428
429}
430
431static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
432{
433 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
434 ts->idle_active = 0;
435
436 sched_clock_idle_wakeup_event(0);
437}
438
439static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
440{
441 ktime_t now = ktime_get();
442
443 ts->idle_entrytime = now;
444 ts->idle_active = 1;
445 sched_clock_idle_sleep_event();
446 return now;
447}
448
449/**
450 * get_cpu_idle_time_us - get the total idle time of a cpu
451 * @cpu: CPU number to query
452 * @last_update_time: variable to store update time in. Do not update
453 * counters if NULL.
454 *
455 * Return the cummulative idle time (since boot) for a given
456 * CPU, in microseconds.
457 *
458 * This time is measured via accounting rather than sampling,
459 * and is as accurate as ktime_get() is.
460 *
461 * This function returns -1 if NOHZ is not enabled.
462 */
463u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
464{
465 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
466 ktime_t now, idle;
467
468 if (!tick_nohz_active)
469 return -1;
470
471 now = ktime_get();
472 if (last_update_time) {
473 update_ts_time_stats(cpu, ts, now, last_update_time);
474 idle = ts->idle_sleeptime;
475 } else {
476 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
477 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
478
479 idle = ktime_add(ts->idle_sleeptime, delta);
480 } else {
481 idle = ts->idle_sleeptime;
482 }
483 }
484
485 return ktime_to_us(idle);
486
487}
488EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
489
490/**
491 * get_cpu_iowait_time_us - get the total iowait time of a cpu
492 * @cpu: CPU number to query
493 * @last_update_time: variable to store update time in. Do not update
494 * counters if NULL.
495 *
496 * Return the cummulative iowait time (since boot) for a given
497 * CPU, in microseconds.
498 *
499 * This time is measured via accounting rather than sampling,
500 * and is as accurate as ktime_get() is.
501 *
502 * This function returns -1 if NOHZ is not enabled.
503 */
504u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
505{
506 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
507 ktime_t now, iowait;
508
509 if (!tick_nohz_active)
510 return -1;
511
512 now = ktime_get();
513 if (last_update_time) {
514 update_ts_time_stats(cpu, ts, now, last_update_time);
515 iowait = ts->iowait_sleeptime;
516 } else {
517 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
518 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
519
520 iowait = ktime_add(ts->iowait_sleeptime, delta);
521 } else {
522 iowait = ts->iowait_sleeptime;
523 }
524 }
525
526 return ktime_to_us(iowait);
527}
528EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
529
530static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
531 ktime_t now, int cpu)
532{
533 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
534 ktime_t last_update, expires, ret = { .tv64 = 0 };
535 unsigned long rcu_delta_jiffies;
536 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
537 u64 time_delta;
538
539 time_delta = timekeeping_max_deferment();
540
541 /* Read jiffies and the time when jiffies were updated last */
542 do {
543 seq = read_seqbegin(&jiffies_lock);
544 last_update = last_jiffies_update;
545 last_jiffies = jiffies;
546 } while (read_seqretry(&jiffies_lock, seq));
547
548 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
549 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
550 next_jiffies = last_jiffies + 1;
551 delta_jiffies = 1;
552 } else {
553 /* Get the next timer wheel timer */
554 next_jiffies = get_next_timer_interrupt(last_jiffies);
555 delta_jiffies = next_jiffies - last_jiffies;
556 if (rcu_delta_jiffies < delta_jiffies) {
557 next_jiffies = last_jiffies + rcu_delta_jiffies;
558 delta_jiffies = rcu_delta_jiffies;
559 }
560 }
561
562 /*
563 * Do not stop the tick, if we are only one off (or less)
564 * or if the cpu is required for RCU:
565 */
566 if (!ts->tick_stopped && delta_jiffies <= 1)
567 goto out;
568
569 /* Schedule the tick, if we are at least one jiffie off */
570 if ((long)delta_jiffies >= 1) {
571
572 /*
573 * If this cpu is the one which updates jiffies, then
574 * give up the assignment and let it be taken by the
575 * cpu which runs the tick timer next, which might be
576 * this cpu as well. If we don't drop this here the
577 * jiffies might be stale and do_timer() never
578 * invoked. Keep track of the fact that it was the one
579 * which had the do_timer() duty last. If this cpu is
580 * the one which had the do_timer() duty last, we
581 * limit the sleep time to the timekeeping
582 * max_deferement value which we retrieved
583 * above. Otherwise we can sleep as long as we want.
584 */
585 if (cpu == tick_do_timer_cpu) {
586 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
587 ts->do_timer_last = 1;
588 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
589 time_delta = KTIME_MAX;
590 ts->do_timer_last = 0;
591 } else if (!ts->do_timer_last) {
592 time_delta = KTIME_MAX;
593 }
594
595#ifdef CONFIG_NO_HZ_FULL
596 if (!ts->inidle) {
597 time_delta = min(time_delta,
598 scheduler_tick_max_deferment());
599 }
600#endif
601
602 /*
603 * calculate the expiry time for the next timer wheel
604 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
605 * that there is no timer pending or at least extremely
606 * far into the future (12 days for HZ=1000). In this
607 * case we set the expiry to the end of time.
608 */
609 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
610 /*
611 * Calculate the time delta for the next timer event.
612 * If the time delta exceeds the maximum time delta
613 * permitted by the current clocksource then adjust
614 * the time delta accordingly to ensure the
615 * clocksource does not wrap.
616 */
617 time_delta = min_t(u64, time_delta,
618 tick_period.tv64 * delta_jiffies);
619 }
620
621 if (time_delta < KTIME_MAX)
622 expires = ktime_add_ns(last_update, time_delta);
623 else
624 expires.tv64 = KTIME_MAX;
625
626 /* Skip reprogram of event if its not changed */
627 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
628 goto out;
629
630 ret = expires;
631
632 /*
633 * nohz_stop_sched_tick can be called several times before
634 * the nohz_restart_sched_tick is called. This happens when
635 * interrupts arrive which do not cause a reschedule. In the
636 * first call we save the current tick time, so we can restart
637 * the scheduler tick in nohz_restart_sched_tick.
638 */
639 if (!ts->tick_stopped) {
640 nohz_balance_enter_idle(cpu);
641 calc_load_enter_idle();
642
643 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
644 ts->tick_stopped = 1;
645 trace_tick_stop(1, " ");
646 }
647
648 /*
649 * If the expiration time == KTIME_MAX, then
650 * in this case we simply stop the tick timer.
651 */
652 if (unlikely(expires.tv64 == KTIME_MAX)) {
653 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
654 hrtimer_cancel(&ts->sched_timer);
655 goto out;
656 }
657
658 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
659 hrtimer_start(&ts->sched_timer, expires,
660 HRTIMER_MODE_ABS_PINNED);
661 /* Check, if the timer was already in the past */
662 if (hrtimer_active(&ts->sched_timer))
663 goto out;
664 } else if (!tick_program_event(expires, 0))
665 goto out;
666 /*
667 * We are past the event already. So we crossed a
668 * jiffie boundary. Update jiffies and raise the
669 * softirq.
670 */
671 tick_do_update_jiffies64(ktime_get());
672 }
673 raise_softirq_irqoff(TIMER_SOFTIRQ);
674out:
675 ts->next_jiffies = next_jiffies;
676 ts->last_jiffies = last_jiffies;
677 ts->sleep_length = ktime_sub(dev->next_event, now);
678
679 return ret;
680}
681
682static void tick_nohz_full_stop_tick(struct tick_sched *ts)
683{
684#ifdef CONFIG_NO_HZ_FULL
685 int cpu = smp_processor_id();
686
687 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
688 return;
689
690 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
691 return;
692
693 if (!can_stop_full_tick())
694 return;
695
696 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
697#endif
698}
699
700static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
701{
702 /*
703 * If this cpu is offline and it is the one which updates
704 * jiffies, then give up the assignment and let it be taken by
705 * the cpu which runs the tick timer next. If we don't drop
706 * this here the jiffies might be stale and do_timer() never
707 * invoked.
708 */
709 if (unlikely(!cpu_online(cpu))) {
710 if (cpu == tick_do_timer_cpu)
711 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
712 return false;
713 }
714
715 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
716 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
717 return false;
718 }
719
720 if (need_resched())
721 return false;
722
723 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
724 static int ratelimit;
725
726 if (ratelimit < 10 &&
727 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
728 pr_warn("NOHZ: local_softirq_pending %02x\n",
729 (unsigned int) local_softirq_pending());
730 ratelimit++;
731 }
732 return false;
733 }
734
735 if (tick_nohz_full_enabled()) {
736 /*
737 * Keep the tick alive to guarantee timekeeping progression
738 * if there are full dynticks CPUs around
739 */
740 if (tick_do_timer_cpu == cpu)
741 return false;
742 /*
743 * Boot safety: make sure the timekeeping duty has been
744 * assigned before entering dyntick-idle mode,
745 */
746 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
747 return false;
748 }
749
750 return true;
751}
752
753static void __tick_nohz_idle_enter(struct tick_sched *ts)
754{
755 ktime_t now, expires;
756 int cpu = smp_processor_id();
757
758 now = tick_nohz_start_idle(ts);
759
760 if (can_stop_idle_tick(cpu, ts)) {
761 int was_stopped = ts->tick_stopped;
762
763 ts->idle_calls++;
764
765 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
766 if (expires.tv64 > 0LL) {
767 ts->idle_sleeps++;
768 ts->idle_expires = expires;
769 }
770
771 if (!was_stopped && ts->tick_stopped)
772 ts->idle_jiffies = ts->last_jiffies;
773 }
774}
775
776/**
777 * tick_nohz_idle_enter - stop the idle tick from the idle task
778 *
779 * When the next event is more than a tick into the future, stop the idle tick
780 * Called when we start the idle loop.
781 *
782 * The arch is responsible of calling:
783 *
784 * - rcu_idle_enter() after its last use of RCU before the CPU is put
785 * to sleep.
786 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
787 */
788void tick_nohz_idle_enter(void)
789{
790 struct tick_sched *ts;
791
792 WARN_ON_ONCE(irqs_disabled());
793
794 /*
795 * Update the idle state in the scheduler domain hierarchy
796 * when tick_nohz_stop_sched_tick() is called from the idle loop.
797 * State will be updated to busy during the first busy tick after
798 * exiting idle.
799 */
800 set_cpu_sd_state_idle();
801
802 local_irq_disable();
803
804 ts = &__get_cpu_var(tick_cpu_sched);
805 ts->inidle = 1;
806 __tick_nohz_idle_enter(ts);
807
808 local_irq_enable();
809}
810EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
811
812/**
813 * tick_nohz_irq_exit - update next tick event from interrupt exit
814 *
815 * When an interrupt fires while we are idle and it doesn't cause
816 * a reschedule, it may still add, modify or delete a timer, enqueue
817 * an RCU callback, etc...
818 * So we need to re-calculate and reprogram the next tick event.
819 */
820void tick_nohz_irq_exit(void)
821{
822 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
823
824 if (ts->inidle)
825 __tick_nohz_idle_enter(ts);
826 else
827 tick_nohz_full_stop_tick(ts);
828}
829
830/**
831 * tick_nohz_get_sleep_length - return the length of the current sleep
832 *
833 * Called from power state control code with interrupts disabled
834 */
835ktime_t tick_nohz_get_sleep_length(void)
836{
837 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
838
839 return ts->sleep_length;
840}
841
842static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
843{
844 hrtimer_cancel(&ts->sched_timer);
845 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
846
847 while (1) {
848 /* Forward the time to expire in the future */
849 hrtimer_forward(&ts->sched_timer, now, tick_period);
850
851 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
852 hrtimer_start_expires(&ts->sched_timer,
853 HRTIMER_MODE_ABS_PINNED);
854 /* Check, if the timer was already in the past */
855 if (hrtimer_active(&ts->sched_timer))
856 break;
857 } else {
858 if (!tick_program_event(
859 hrtimer_get_expires(&ts->sched_timer), 0))
860 break;
861 }
862 /* Reread time and update jiffies */
863 now = ktime_get();
864 tick_do_update_jiffies64(now);
865 }
866}
867
868static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
869{
870 /* Update jiffies first */
871 tick_do_update_jiffies64(now);
872 update_cpu_load_nohz();
873
874 calc_load_exit_idle();
875 touch_softlockup_watchdog();
876 /*
877 * Cancel the scheduled timer and restore the tick
878 */
879 ts->tick_stopped = 0;
880 ts->idle_exittime = now;
881
882 tick_nohz_restart(ts, now);
883}
884
885static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
886{
887#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
888 unsigned long ticks;
889
890 if (vtime_accounting_enabled())
891 return;
892 /*
893 * We stopped the tick in idle. Update process times would miss the
894 * time we slept as update_process_times does only a 1 tick
895 * accounting. Enforce that this is accounted to idle !
896 */
897 ticks = jiffies - ts->idle_jiffies;
898 /*
899 * We might be one off. Do not randomly account a huge number of ticks!
900 */
901 if (ticks && ticks < LONG_MAX)
902 account_idle_ticks(ticks);
903#endif
904}
905
906/**
907 * tick_nohz_idle_exit - restart the idle tick from the idle task
908 *
909 * Restart the idle tick when the CPU is woken up from idle
910 * This also exit the RCU extended quiescent state. The CPU
911 * can use RCU again after this function is called.
912 */
913void tick_nohz_idle_exit(void)
914{
915 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
916 ktime_t now;
917
918 local_irq_disable();
919
920 WARN_ON_ONCE(!ts->inidle);
921
922 ts->inidle = 0;
923
924 if (ts->idle_active || ts->tick_stopped)
925 now = ktime_get();
926
927 if (ts->idle_active)
928 tick_nohz_stop_idle(ts, now);
929
930 if (ts->tick_stopped) {
931 tick_nohz_restart_sched_tick(ts, now);
932 tick_nohz_account_idle_ticks(ts);
933 }
934
935 local_irq_enable();
936}
937EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
938
939static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
940{
941 hrtimer_forward(&ts->sched_timer, now, tick_period);
942 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
943}
944
945/*
946 * The nohz low res interrupt handler
947 */
948static void tick_nohz_handler(struct clock_event_device *dev)
949{
950 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
951 struct pt_regs *regs = get_irq_regs();
952 ktime_t now = ktime_get();
953
954 dev->next_event.tv64 = KTIME_MAX;
955
956 tick_sched_do_timer(now);
957 tick_sched_handle(ts, regs);
958
959 while (tick_nohz_reprogram(ts, now)) {
960 now = ktime_get();
961 tick_do_update_jiffies64(now);
962 }
963}
964
965/**
966 * tick_nohz_switch_to_nohz - switch to nohz mode
967 */
968static void tick_nohz_switch_to_nohz(void)
969{
970 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
971 ktime_t next;
972
973 if (!tick_nohz_enabled)
974 return;
975
976 local_irq_disable();
977 if (tick_switch_to_oneshot(tick_nohz_handler)) {
978 local_irq_enable();
979 return;
980 }
981 tick_nohz_active = 1;
982 ts->nohz_mode = NOHZ_MODE_LOWRES;
983
984 /*
985 * Recycle the hrtimer in ts, so we can share the
986 * hrtimer_forward with the highres code.
987 */
988 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
989 /* Get the next period */
990 next = tick_init_jiffy_update();
991
992 for (;;) {
993 hrtimer_set_expires(&ts->sched_timer, next);
994 if (!tick_program_event(next, 0))
995 break;
996 next = ktime_add(next, tick_period);
997 }
998 local_irq_enable();
999}
1000
1001/*
1002 * When NOHZ is enabled and the tick is stopped, we need to kick the
1003 * tick timer from irq_enter() so that the jiffies update is kept
1004 * alive during long running softirqs. That's ugly as hell, but
1005 * correctness is key even if we need to fix the offending softirq in
1006 * the first place.
1007 *
1008 * Note, this is different to tick_nohz_restart. We just kick the
1009 * timer and do not touch the other magic bits which need to be done
1010 * when idle is left.
1011 */
1012static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1013{
1014#if 0
1015 /* Switch back to 2.6.27 behaviour */
1016 ktime_t delta;
1017
1018 /*
1019 * Do not touch the tick device, when the next expiry is either
1020 * already reached or less/equal than the tick period.
1021 */
1022 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1023 if (delta.tv64 <= tick_period.tv64)
1024 return;
1025
1026 tick_nohz_restart(ts, now);
1027#endif
1028}
1029
1030static inline void tick_nohz_irq_enter(void)
1031{
1032 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1033 ktime_t now;
1034
1035 if (!ts->idle_active && !ts->tick_stopped)
1036 return;
1037 now = ktime_get();
1038 if (ts->idle_active)
1039 tick_nohz_stop_idle(ts, now);
1040 if (ts->tick_stopped) {
1041 tick_nohz_update_jiffies(now);
1042 tick_nohz_kick_tick(ts, now);
1043 }
1044}
1045
1046#else
1047
1048static inline void tick_nohz_switch_to_nohz(void) { }
1049static inline void tick_nohz_irq_enter(void) { }
1050
1051#endif /* CONFIG_NO_HZ_COMMON */
1052
1053/*
1054 * Called from irq_enter to notify about the possible interruption of idle()
1055 */
1056void tick_irq_enter(void)
1057{
1058 tick_check_oneshot_broadcast_this_cpu();
1059 tick_nohz_irq_enter();
1060}
1061
1062/*
1063 * High resolution timer specific code
1064 */
1065#ifdef CONFIG_HIGH_RES_TIMERS
1066/*
1067 * We rearm the timer until we get disabled by the idle code.
1068 * Called with interrupts disabled.
1069 */
1070static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1071{
1072 struct tick_sched *ts =
1073 container_of(timer, struct tick_sched, sched_timer);
1074 struct pt_regs *regs = get_irq_regs();
1075 ktime_t now = ktime_get();
1076
1077 tick_sched_do_timer(now);
1078
1079 /*
1080 * Do not call, when we are not in irq context and have
1081 * no valid regs pointer
1082 */
1083 if (regs)
1084 tick_sched_handle(ts, regs);
1085
1086 hrtimer_forward(timer, now, tick_period);
1087
1088 return HRTIMER_RESTART;
1089}
1090
1091static int sched_skew_tick;
1092
1093static int __init skew_tick(char *str)
1094{
1095 get_option(&str, &sched_skew_tick);
1096
1097 return 0;
1098}
1099early_param("skew_tick", skew_tick);
1100
1101/**
1102 * tick_setup_sched_timer - setup the tick emulation timer
1103 */
1104void tick_setup_sched_timer(void)
1105{
1106 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1107 ktime_t now = ktime_get();
1108
1109 /*
1110 * Emulate tick processing via per-CPU hrtimers:
1111 */
1112 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1113 ts->sched_timer.function = tick_sched_timer;
1114
1115 /* Get the next period (per cpu) */
1116 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1117
1118 /* Offset the tick to avert jiffies_lock contention. */
1119 if (sched_skew_tick) {
1120 u64 offset = ktime_to_ns(tick_period) >> 1;
1121 do_div(offset, num_possible_cpus());
1122 offset *= smp_processor_id();
1123 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1124 }
1125
1126 for (;;) {
1127 hrtimer_forward(&ts->sched_timer, now, tick_period);
1128 hrtimer_start_expires(&ts->sched_timer,
1129 HRTIMER_MODE_ABS_PINNED);
1130 /* Check, if the timer was already in the past */
1131 if (hrtimer_active(&ts->sched_timer))
1132 break;
1133 now = ktime_get();
1134 }
1135
1136#ifdef CONFIG_NO_HZ_COMMON
1137 if (tick_nohz_enabled) {
1138 ts->nohz_mode = NOHZ_MODE_HIGHRES;
1139 tick_nohz_active = 1;
1140 }
1141#endif
1142}
1143#endif /* HIGH_RES_TIMERS */
1144
1145#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1146void tick_cancel_sched_timer(int cpu)
1147{
1148 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1149
1150# ifdef CONFIG_HIGH_RES_TIMERS
1151 if (ts->sched_timer.base)
1152 hrtimer_cancel(&ts->sched_timer);
1153# endif
1154
1155 memset(ts, 0, sizeof(*ts));
1156}
1157#endif
1158
1159/**
1160 * Async notification about clocksource changes
1161 */
1162void tick_clock_notify(void)
1163{
1164 int cpu;
1165
1166 for_each_possible_cpu(cpu)
1167 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1168}
1169
1170/*
1171 * Async notification about clock event changes
1172 */
1173void tick_oneshot_notify(void)
1174{
1175 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1176
1177 set_bit(0, &ts->check_clocks);
1178}
1179
1180/**
1181 * Check, if a change happened, which makes oneshot possible.
1182 *
1183 * Called cyclic from the hrtimer softirq (driven by the timer
1184 * softirq) allow_nohz signals, that we can switch into low-res nohz
1185 * mode, because high resolution timers are disabled (either compile
1186 * or runtime).
1187 */
1188int tick_check_oneshot_change(int allow_nohz)
1189{
1190 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1191
1192 if (!test_and_clear_bit(0, &ts->check_clocks))
1193 return 0;
1194
1195 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1196 return 0;
1197
1198 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1199 return 0;
1200
1201 if (!allow_nohz)
1202 return 1;
1203
1204 tick_nohz_switch_to_nohz();
1205 return 0;
1206}
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 * NOHZ 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 * 64-bit can do a quick check without holding the 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 * 32-bit cannot do that because the store of 'tick_next_period'
68 * consists of two 32-bit stores, and the first store could be
69 * moved by the CPU 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 * Re-evaluate 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 /* Keep the tick_next_period variable up to date */
121 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
122
123 if (IS_ENABLED(CONFIG_64BIT)) {
124 /*
125 * Pairs with smp_load_acquire() in the lockless quick
126 * check above, and ensures that the update to 'jiffies_64' is
127 * not reordered vs. the store to 'tick_next_period', neither
128 * by the compiler nor by the CPU.
129 */
130 smp_store_release(&tick_next_period, nextp);
131 } else {
132 /*
133 * A plain store is good enough on 32-bit, as the quick check
134 * above is protected by the sequence count.
135 */
136 tick_next_period = nextp;
137 }
138
139 /*
140 * Release the sequence count. calc_global_load() below is not
141 * protected by it, but 'jiffies_lock' needs to be held to prevent
142 * concurrent invocations.
143 */
144 write_seqcount_end(&jiffies_seq);
145
146 calc_global_load();
147
148 raw_spin_unlock(&jiffies_lock);
149 update_wall_time();
150}
151
152/*
153 * Initialize and return retrieve the jiffies update.
154 */
155static ktime_t tick_init_jiffy_update(void)
156{
157 ktime_t period;
158
159 raw_spin_lock(&jiffies_lock);
160 write_seqcount_begin(&jiffies_seq);
161
162 /* Have we started the jiffies update yet ? */
163 if (last_jiffies_update == 0) {
164 u32 rem;
165
166 /*
167 * Ensure that the tick is aligned to a multiple of
168 * TICK_NSEC.
169 */
170 div_u64_rem(tick_next_period, TICK_NSEC, &rem);
171 if (rem)
172 tick_next_period += TICK_NSEC - rem;
173
174 last_jiffies_update = tick_next_period;
175 }
176 period = last_jiffies_update;
177
178 write_seqcount_end(&jiffies_seq);
179 raw_spin_unlock(&jiffies_lock);
180
181 return period;
182}
183
184#define MAX_STALLED_JIFFIES 5
185
186static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
187{
188 int cpu = smp_processor_id();
189
190#ifdef CONFIG_NO_HZ_COMMON
191 /*
192 * Check if the do_timer duty was dropped. We don't care about
193 * concurrency: This happens only when the CPU in charge went
194 * into a long sleep. If two CPUs happen to assign themselves to
195 * this duty, then the jiffies update is still serialized by
196 * 'jiffies_lock'.
197 *
198 * If nohz_full is enabled, this should not happen because the
199 * 'tick_do_timer_cpu' CPU never relinquishes.
200 */
201 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
202#ifdef CONFIG_NO_HZ_FULL
203 WARN_ON_ONCE(tick_nohz_full_running);
204#endif
205 tick_do_timer_cpu = cpu;
206 }
207#endif
208
209 /* Check if jiffies need an update */
210 if (tick_do_timer_cpu == cpu)
211 tick_do_update_jiffies64(now);
212
213 /*
214 * If the jiffies update stalled for too long (timekeeper in stop_machine()
215 * or VMEXIT'ed for several msecs), force an update.
216 */
217 if (ts->last_tick_jiffies != jiffies) {
218 ts->stalled_jiffies = 0;
219 ts->last_tick_jiffies = READ_ONCE(jiffies);
220 } else {
221 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
222 tick_do_update_jiffies64(now);
223 ts->stalled_jiffies = 0;
224 ts->last_tick_jiffies = READ_ONCE(jiffies);
225 }
226 }
227
228 if (ts->inidle)
229 ts->got_idle_tick = 1;
230}
231
232static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
233{
234#ifdef CONFIG_NO_HZ_COMMON
235 /*
236 * When we are idle and the tick is stopped, we have to touch
237 * the watchdog as we might not schedule for a really long
238 * time. This happens on completely idle SMP systems while
239 * waiting on the login prompt. We also increment the "start of
240 * idle" jiffy stamp so the idle accounting adjustment we do
241 * when we go busy again does not account too many ticks.
242 */
243 if (ts->tick_stopped) {
244 touch_softlockup_watchdog_sched();
245 if (is_idle_task(current))
246 ts->idle_jiffies++;
247 /*
248 * In case the current tick fired too early past its expected
249 * expiration, make sure we don't bypass the next clock reprogramming
250 * to the same deadline.
251 */
252 ts->next_tick = 0;
253 }
254#endif
255 update_process_times(user_mode(regs));
256 profile_tick(CPU_PROFILING);
257}
258#endif
259
260#ifdef CONFIG_NO_HZ_FULL
261cpumask_var_t tick_nohz_full_mask;
262EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
263bool tick_nohz_full_running;
264EXPORT_SYMBOL_GPL(tick_nohz_full_running);
265static atomic_t tick_dep_mask;
266
267static bool check_tick_dependency(atomic_t *dep)
268{
269 int val = atomic_read(dep);
270
271 if (val & TICK_DEP_MASK_POSIX_TIMER) {
272 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
273 return true;
274 }
275
276 if (val & TICK_DEP_MASK_PERF_EVENTS) {
277 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
278 return true;
279 }
280
281 if (val & TICK_DEP_MASK_SCHED) {
282 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
283 return true;
284 }
285
286 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
287 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
288 return true;
289 }
290
291 if (val & TICK_DEP_MASK_RCU) {
292 trace_tick_stop(0, TICK_DEP_MASK_RCU);
293 return true;
294 }
295
296 if (val & TICK_DEP_MASK_RCU_EXP) {
297 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
298 return true;
299 }
300
301 return false;
302}
303
304static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
305{
306 lockdep_assert_irqs_disabled();
307
308 if (unlikely(!cpu_online(cpu)))
309 return false;
310
311 if (check_tick_dependency(&tick_dep_mask))
312 return false;
313
314 if (check_tick_dependency(&ts->tick_dep_mask))
315 return false;
316
317 if (check_tick_dependency(¤t->tick_dep_mask))
318 return false;
319
320 if (check_tick_dependency(¤t->signal->tick_dep_mask))
321 return false;
322
323 return true;
324}
325
326static void nohz_full_kick_func(struct irq_work *work)
327{
328 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
329}
330
331static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
332 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
333
334/*
335 * Kick this CPU if it's full dynticks in order to force it to
336 * re-evaluate its dependency on the tick and restart it if necessary.
337 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
338 * is NMI safe.
339 */
340static void tick_nohz_full_kick(void)
341{
342 if (!tick_nohz_full_cpu(smp_processor_id()))
343 return;
344
345 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
346}
347
348/*
349 * Kick the CPU if it's full dynticks in order to force it to
350 * re-evaluate its dependency on the tick and restart it if necessary.
351 */
352void tick_nohz_full_kick_cpu(int cpu)
353{
354 if (!tick_nohz_full_cpu(cpu))
355 return;
356
357 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
358}
359
360static void tick_nohz_kick_task(struct task_struct *tsk)
361{
362 int cpu;
363
364 /*
365 * If the task is not running, run_posix_cpu_timers()
366 * has nothing to elapse, and an IPI can then be optimized out.
367 *
368 * activate_task() STORE p->tick_dep_mask
369 * STORE p->on_rq
370 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
371 * LOCK rq->lock LOAD p->on_rq
372 * smp_mb__after_spin_lock()
373 * tick_nohz_task_switch()
374 * LOAD p->tick_dep_mask
375 */
376 if (!sched_task_on_rq(tsk))
377 return;
378
379 /*
380 * If the task concurrently migrates to another CPU,
381 * we guarantee it sees the new tick dependency upon
382 * schedule.
383 *
384 * set_task_cpu(p, cpu);
385 * STORE p->cpu = @cpu
386 * __schedule() (switch to task 'p')
387 * LOCK rq->lock
388 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
389 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
390 * LOAD p->tick_dep_mask LOAD p->cpu
391 */
392 cpu = task_cpu(tsk);
393
394 preempt_disable();
395 if (cpu_online(cpu))
396 tick_nohz_full_kick_cpu(cpu);
397 preempt_enable();
398}
399
400/*
401 * Kick all full dynticks CPUs in order to force these to re-evaluate
402 * their dependency on the tick and restart it if necessary.
403 */
404static void tick_nohz_full_kick_all(void)
405{
406 int cpu;
407
408 if (!tick_nohz_full_running)
409 return;
410
411 preempt_disable();
412 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
413 tick_nohz_full_kick_cpu(cpu);
414 preempt_enable();
415}
416
417static void tick_nohz_dep_set_all(atomic_t *dep,
418 enum tick_dep_bits bit)
419{
420 int prev;
421
422 prev = atomic_fetch_or(BIT(bit), dep);
423 if (!prev)
424 tick_nohz_full_kick_all();
425}
426
427/*
428 * Set a global tick dependency. Used by perf events that rely on freq and
429 * unstable clocks.
430 */
431void tick_nohz_dep_set(enum tick_dep_bits bit)
432{
433 tick_nohz_dep_set_all(&tick_dep_mask, bit);
434}
435
436void tick_nohz_dep_clear(enum tick_dep_bits bit)
437{
438 atomic_andnot(BIT(bit), &tick_dep_mask);
439}
440
441/*
442 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
443 * manage event-throttling.
444 */
445void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
446{
447 int prev;
448 struct tick_sched *ts;
449
450 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
451
452 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
453 if (!prev) {
454 preempt_disable();
455 /* Perf needs local kick that is NMI safe */
456 if (cpu == smp_processor_id()) {
457 tick_nohz_full_kick();
458 } else {
459 /* Remote IRQ work not NMI-safe */
460 if (!WARN_ON_ONCE(in_nmi()))
461 tick_nohz_full_kick_cpu(cpu);
462 }
463 preempt_enable();
464 }
465}
466EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
467
468void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
469{
470 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
471
472 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
473}
474EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
475
476/*
477 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
478 * in order to elapse per task timers.
479 */
480void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
481{
482 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
483 tick_nohz_kick_task(tsk);
484}
485EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
486
487void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
488{
489 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
490}
491EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
492
493/*
494 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
495 * per process timers.
496 */
497void tick_nohz_dep_set_signal(struct task_struct *tsk,
498 enum tick_dep_bits bit)
499{
500 int prev;
501 struct signal_struct *sig = tsk->signal;
502
503 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
504 if (!prev) {
505 struct task_struct *t;
506
507 lockdep_assert_held(&tsk->sighand->siglock);
508 __for_each_thread(sig, t)
509 tick_nohz_kick_task(t);
510 }
511}
512
513void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
514{
515 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
516}
517
518/*
519 * Re-evaluate the need for the tick as we switch the current task.
520 * It might need the tick due to per task/process properties:
521 * perf events, posix CPU timers, ...
522 */
523void __tick_nohz_task_switch(void)
524{
525 struct tick_sched *ts;
526
527 if (!tick_nohz_full_cpu(smp_processor_id()))
528 return;
529
530 ts = this_cpu_ptr(&tick_cpu_sched);
531
532 if (ts->tick_stopped) {
533 if (atomic_read(¤t->tick_dep_mask) ||
534 atomic_read(¤t->signal->tick_dep_mask))
535 tick_nohz_full_kick();
536 }
537}
538
539/* Get the boot-time nohz CPU list from the kernel parameters. */
540void __init tick_nohz_full_setup(cpumask_var_t cpumask)
541{
542 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
543 cpumask_copy(tick_nohz_full_mask, cpumask);
544 tick_nohz_full_running = true;
545}
546
547bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
548{
549 /*
550 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
551 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
552 * CPUs. It must remain online when nohz full is enabled.
553 */
554 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
555 return false;
556 return true;
557}
558
559static int tick_nohz_cpu_down(unsigned int cpu)
560{
561 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
562}
563
564void __init tick_nohz_init(void)
565{
566 int cpu, ret;
567
568 if (!tick_nohz_full_running)
569 return;
570
571 /*
572 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
573 * locking contexts. But then we need IRQ work to raise its own
574 * interrupts to avoid circular dependency on the tick.
575 */
576 if (!arch_irq_work_has_interrupt()) {
577 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
578 cpumask_clear(tick_nohz_full_mask);
579 tick_nohz_full_running = false;
580 return;
581 }
582
583 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
584 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
585 cpu = smp_processor_id();
586
587 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
588 pr_warn("NO_HZ: Clearing %d from nohz_full range "
589 "for timekeeping\n", cpu);
590 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
591 }
592 }
593
594 for_each_cpu(cpu, tick_nohz_full_mask)
595 ct_cpu_track_user(cpu);
596
597 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
598 "kernel/nohz:predown", NULL,
599 tick_nohz_cpu_down);
600 WARN_ON(ret < 0);
601 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
602 cpumask_pr_args(tick_nohz_full_mask));
603}
604#endif
605
606/*
607 * NOHZ - aka dynamic tick functionality
608 */
609#ifdef CONFIG_NO_HZ_COMMON
610/*
611 * NO HZ enabled ?
612 */
613bool tick_nohz_enabled __read_mostly = true;
614unsigned long tick_nohz_active __read_mostly;
615/*
616 * Enable / Disable tickless mode
617 */
618static int __init setup_tick_nohz(char *str)
619{
620 return (kstrtobool(str, &tick_nohz_enabled) == 0);
621}
622
623__setup("nohz=", setup_tick_nohz);
624
625bool tick_nohz_tick_stopped(void)
626{
627 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
628
629 return ts->tick_stopped;
630}
631
632bool tick_nohz_tick_stopped_cpu(int cpu)
633{
634 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
635
636 return ts->tick_stopped;
637}
638
639/**
640 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
641 *
642 * Called from interrupt entry when the CPU was idle
643 *
644 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
645 * must be updated. Otherwise an interrupt handler could use a stale jiffy
646 * value. We do this unconditionally on any CPU, as we don't know whether the
647 * CPU, which has the update task assigned, is in a long sleep.
648 */
649static void tick_nohz_update_jiffies(ktime_t now)
650{
651 unsigned long flags;
652
653 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
654
655 local_irq_save(flags);
656 tick_do_update_jiffies64(now);
657 local_irq_restore(flags);
658
659 touch_softlockup_watchdog_sched();
660}
661
662static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
663{
664 ktime_t delta;
665
666 if (WARN_ON_ONCE(!ts->idle_active))
667 return;
668
669 delta = ktime_sub(now, ts->idle_entrytime);
670
671 write_seqcount_begin(&ts->idle_sleeptime_seq);
672 if (nr_iowait_cpu(smp_processor_id()) > 0)
673 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
674 else
675 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
676
677 ts->idle_entrytime = now;
678 ts->idle_active = 0;
679 write_seqcount_end(&ts->idle_sleeptime_seq);
680
681 sched_clock_idle_wakeup_event();
682}
683
684static void tick_nohz_start_idle(struct tick_sched *ts)
685{
686 write_seqcount_begin(&ts->idle_sleeptime_seq);
687 ts->idle_entrytime = ktime_get();
688 ts->idle_active = 1;
689 write_seqcount_end(&ts->idle_sleeptime_seq);
690
691 sched_clock_idle_sleep_event();
692}
693
694static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
695 bool compute_delta, u64 *last_update_time)
696{
697 ktime_t now, idle;
698 unsigned int seq;
699
700 if (!tick_nohz_active)
701 return -1;
702
703 now = ktime_get();
704 if (last_update_time)
705 *last_update_time = ktime_to_us(now);
706
707 do {
708 seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
709
710 if (ts->idle_active && compute_delta) {
711 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
712
713 idle = ktime_add(*sleeptime, delta);
714 } else {
715 idle = *sleeptime;
716 }
717 } while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
718
719 return ktime_to_us(idle);
720
721}
722
723/**
724 * get_cpu_idle_time_us - get the total idle time of a CPU
725 * @cpu: CPU number to query
726 * @last_update_time: variable to store update time in. Do not update
727 * counters if NULL.
728 *
729 * Return the cumulative idle time (since boot) for a given
730 * CPU, in microseconds. Note that this is partially broken due to
731 * the counter of iowait tasks that can be remotely updated without
732 * any synchronization. Therefore it is possible to observe backward
733 * values within two consecutive reads.
734 *
735 * This time is measured via accounting rather than sampling,
736 * and is as accurate as ktime_get() is.
737 *
738 * This function returns -1 if NOHZ is not enabled.
739 */
740u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
741{
742 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
743
744 return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
745 !nr_iowait_cpu(cpu), last_update_time);
746}
747EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
748
749/**
750 * get_cpu_iowait_time_us - get the total iowait time of a CPU
751 * @cpu: CPU number to query
752 * @last_update_time: variable to store update time in. Do not update
753 * counters if NULL.
754 *
755 * Return the cumulative iowait time (since boot) for a given
756 * CPU, in microseconds. Note this is partially broken due to
757 * the counter of iowait tasks that can be remotely updated without
758 * any synchronization. Therefore it is possible to observe backward
759 * values within two consecutive reads.
760 *
761 * This time is measured via accounting rather than sampling,
762 * and is as accurate as ktime_get() is.
763 *
764 * This function returns -1 if NOHZ is not enabled.
765 */
766u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
767{
768 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
769
770 return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
771 nr_iowait_cpu(cpu), last_update_time);
772}
773EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
774
775static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
776{
777 hrtimer_cancel(&ts->sched_timer);
778 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
779
780 /* Forward the time to expire in the future */
781 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
782
783 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
784 hrtimer_start_expires(&ts->sched_timer,
785 HRTIMER_MODE_ABS_PINNED_HARD);
786 } else {
787 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
788 }
789
790 /*
791 * Reset to make sure the next tick stop doesn't get fooled by past
792 * cached clock deadline.
793 */
794 ts->next_tick = 0;
795}
796
797static inline bool local_timer_softirq_pending(void)
798{
799 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
800}
801
802static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
803{
804 u64 basemono, next_tick, delta, expires;
805 unsigned long basejiff;
806 unsigned int seq;
807
808 /* Read jiffies and the time when jiffies were updated last */
809 do {
810 seq = read_seqcount_begin(&jiffies_seq);
811 basemono = last_jiffies_update;
812 basejiff = jiffies;
813 } while (read_seqcount_retry(&jiffies_seq, seq));
814 ts->last_jiffies = basejiff;
815 ts->timer_expires_base = basemono;
816
817 /*
818 * Keep the periodic tick, when RCU, architecture or irq_work
819 * requests it.
820 * Aside of that, check whether the local timer softirq is
821 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
822 * because there is an already expired timer, so it will request
823 * immediate expiry, which rearms the hardware timer with a
824 * minimal delta, which brings us back to this place
825 * immediately. Lather, rinse and repeat...
826 */
827 if (rcu_needs_cpu() || arch_needs_cpu() ||
828 irq_work_needs_cpu() || local_timer_softirq_pending()) {
829 next_tick = basemono + TICK_NSEC;
830 } else {
831 /*
832 * Get the next pending timer. If high resolution
833 * timers are enabled this only takes the timer wheel
834 * timers into account. If high resolution timers are
835 * disabled this also looks at the next expiring
836 * hrtimer.
837 */
838 next_tick = get_next_timer_interrupt(basejiff, basemono);
839 ts->next_timer = next_tick;
840 }
841
842 /* Make sure next_tick is never before basemono! */
843 if (WARN_ON_ONCE(basemono > next_tick))
844 next_tick = basemono;
845
846 /*
847 * If the tick is due in the next period, keep it ticking or
848 * force prod the timer.
849 */
850 delta = next_tick - basemono;
851 if (delta <= (u64)TICK_NSEC) {
852 /*
853 * Tell the timer code that the base is not idle, i.e. undo
854 * the effect of get_next_timer_interrupt():
855 */
856 timer_clear_idle();
857 /*
858 * We've not stopped the tick yet, and there's a timer in the
859 * next period, so no point in stopping it either, bail.
860 */
861 if (!ts->tick_stopped) {
862 ts->timer_expires = 0;
863 goto out;
864 }
865 }
866
867 /*
868 * If this CPU is the one which had the do_timer() duty last, we limit
869 * the sleep time to the timekeeping 'max_deferment' value.
870 * Otherwise we can sleep as long as we want.
871 */
872 delta = timekeeping_max_deferment();
873 if (cpu != tick_do_timer_cpu &&
874 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
875 delta = KTIME_MAX;
876
877 /* Calculate the next expiry time */
878 if (delta < (KTIME_MAX - basemono))
879 expires = basemono + delta;
880 else
881 expires = KTIME_MAX;
882
883 ts->timer_expires = min_t(u64, expires, next_tick);
884
885out:
886 return ts->timer_expires;
887}
888
889static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
890{
891 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
892 u64 basemono = ts->timer_expires_base;
893 u64 expires = ts->timer_expires;
894
895 /* Make sure we won't be trying to stop it twice in a row. */
896 ts->timer_expires_base = 0;
897
898 /*
899 * If this CPU is the one which updates jiffies, then give up
900 * the assignment and let it be taken by the CPU which runs
901 * the tick timer next, which might be this CPU as well. If we
902 * don't drop this here, the jiffies might be stale and
903 * do_timer() never gets invoked. Keep track of the fact that it
904 * was the one which had the do_timer() duty last.
905 */
906 if (cpu == tick_do_timer_cpu) {
907 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
908 ts->do_timer_last = 1;
909 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
910 ts->do_timer_last = 0;
911 }
912
913 /* Skip reprogram of event if it's not changed */
914 if (ts->tick_stopped && (expires == ts->next_tick)) {
915 /* Sanity check: make sure clockevent is actually programmed */
916 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
917 return;
918
919 WARN_ON_ONCE(1);
920 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
921 basemono, ts->next_tick, dev->next_event,
922 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
923 }
924
925 /*
926 * tick_nohz_stop_tick() can be called several times before
927 * tick_nohz_restart_sched_tick() is called. This happens when
928 * interrupts arrive which do not cause a reschedule. In the first
929 * call we save the current tick time, so we can restart the
930 * scheduler tick in tick_nohz_restart_sched_tick().
931 */
932 if (!ts->tick_stopped) {
933 calc_load_nohz_start();
934 quiet_vmstat();
935
936 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
937 ts->tick_stopped = 1;
938 trace_tick_stop(1, TICK_DEP_MASK_NONE);
939 }
940
941 ts->next_tick = expires;
942
943 /*
944 * If the expiration time == KTIME_MAX, then we simply stop
945 * the tick timer.
946 */
947 if (unlikely(expires == KTIME_MAX)) {
948 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
949 hrtimer_cancel(&ts->sched_timer);
950 else
951 tick_program_event(KTIME_MAX, 1);
952 return;
953 }
954
955 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
956 hrtimer_start(&ts->sched_timer, expires,
957 HRTIMER_MODE_ABS_PINNED_HARD);
958 } else {
959 hrtimer_set_expires(&ts->sched_timer, expires);
960 tick_program_event(expires, 1);
961 }
962}
963
964static void tick_nohz_retain_tick(struct tick_sched *ts)
965{
966 ts->timer_expires_base = 0;
967}
968
969#ifdef CONFIG_NO_HZ_FULL
970static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
971{
972 if (tick_nohz_next_event(ts, cpu))
973 tick_nohz_stop_tick(ts, cpu);
974 else
975 tick_nohz_retain_tick(ts);
976}
977#endif /* CONFIG_NO_HZ_FULL */
978
979static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
980{
981 /* Update jiffies first */
982 tick_do_update_jiffies64(now);
983
984 /*
985 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
986 * the clock forward checks in the enqueue path:
987 */
988 timer_clear_idle();
989
990 calc_load_nohz_stop();
991 touch_softlockup_watchdog_sched();
992
993 /* Cancel the scheduled timer and restore the tick: */
994 ts->tick_stopped = 0;
995 tick_nohz_restart(ts, now);
996}
997
998static void __tick_nohz_full_update_tick(struct tick_sched *ts,
999 ktime_t now)
1000{
1001#ifdef CONFIG_NO_HZ_FULL
1002 int cpu = smp_processor_id();
1003
1004 if (can_stop_full_tick(cpu, ts))
1005 tick_nohz_stop_sched_tick(ts, cpu);
1006 else if (ts->tick_stopped)
1007 tick_nohz_restart_sched_tick(ts, now);
1008#endif
1009}
1010
1011static void tick_nohz_full_update_tick(struct tick_sched *ts)
1012{
1013 if (!tick_nohz_full_cpu(smp_processor_id()))
1014 return;
1015
1016 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1017 return;
1018
1019 __tick_nohz_full_update_tick(ts, ktime_get());
1020}
1021
1022/*
1023 * A pending softirq outside an IRQ (or softirq disabled section) context
1024 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1025 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1026 *
1027 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1028 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1029 * triggering the code below, since wakep_softirqd() is ignored.
1030 *
1031 */
1032static bool report_idle_softirq(void)
1033{
1034 static int ratelimit;
1035 unsigned int pending = local_softirq_pending();
1036
1037 if (likely(!pending))
1038 return false;
1039
1040 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1041 if (!cpu_active(smp_processor_id())) {
1042 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1043 if (!pending)
1044 return false;
1045 }
1046
1047 if (ratelimit >= 10)
1048 return false;
1049
1050 /* On RT, softirq handling may be waiting on some lock */
1051 if (local_bh_blocked())
1052 return false;
1053
1054 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1055 pending);
1056 ratelimit++;
1057
1058 return true;
1059}
1060
1061static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1062{
1063 /*
1064 * If this CPU is offline and it is the one which updates
1065 * jiffies, then give up the assignment and let it be taken by
1066 * the CPU which runs the tick timer next. If we don't drop
1067 * this here, the jiffies might be stale and do_timer() never
1068 * gets invoked.
1069 */
1070 if (unlikely(!cpu_online(cpu))) {
1071 if (cpu == tick_do_timer_cpu)
1072 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1073 /*
1074 * Make sure the CPU doesn't get fooled by obsolete tick
1075 * deadline if it comes back online later.
1076 */
1077 ts->next_tick = 0;
1078 return false;
1079 }
1080
1081 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1082 return false;
1083
1084 if (need_resched())
1085 return false;
1086
1087 if (unlikely(report_idle_softirq()))
1088 return false;
1089
1090 if (tick_nohz_full_enabled()) {
1091 /*
1092 * Keep the tick alive to guarantee timekeeping progression
1093 * if there are full dynticks CPUs around
1094 */
1095 if (tick_do_timer_cpu == cpu)
1096 return false;
1097
1098 /* Should not happen for nohz-full */
1099 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1100 return false;
1101 }
1102
1103 return true;
1104}
1105
1106/**
1107 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1108 *
1109 * When the next event is more than a tick into the future, stop the idle tick
1110 */
1111void tick_nohz_idle_stop_tick(void)
1112{
1113 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1114 int cpu = smp_processor_id();
1115 ktime_t expires;
1116
1117 /*
1118 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1119 * tick timer expiration time is known already.
1120 */
1121 if (ts->timer_expires_base)
1122 expires = ts->timer_expires;
1123 else if (can_stop_idle_tick(cpu, ts))
1124 expires = tick_nohz_next_event(ts, cpu);
1125 else
1126 return;
1127
1128 ts->idle_calls++;
1129
1130 if (expires > 0LL) {
1131 int was_stopped = ts->tick_stopped;
1132
1133 tick_nohz_stop_tick(ts, cpu);
1134
1135 ts->idle_sleeps++;
1136 ts->idle_expires = expires;
1137
1138 if (!was_stopped && ts->tick_stopped) {
1139 ts->idle_jiffies = ts->last_jiffies;
1140 nohz_balance_enter_idle(cpu);
1141 }
1142 } else {
1143 tick_nohz_retain_tick(ts);
1144 }
1145}
1146
1147void tick_nohz_idle_retain_tick(void)
1148{
1149 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1150 /*
1151 * Undo the effect of get_next_timer_interrupt() called from
1152 * tick_nohz_next_event().
1153 */
1154 timer_clear_idle();
1155}
1156
1157/**
1158 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1159 *
1160 * Called when we start the idle loop.
1161 */
1162void tick_nohz_idle_enter(void)
1163{
1164 struct tick_sched *ts;
1165
1166 lockdep_assert_irqs_enabled();
1167
1168 local_irq_disable();
1169
1170 ts = this_cpu_ptr(&tick_cpu_sched);
1171
1172 WARN_ON_ONCE(ts->timer_expires_base);
1173
1174 ts->inidle = 1;
1175 tick_nohz_start_idle(ts);
1176
1177 local_irq_enable();
1178}
1179
1180/**
1181 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1182 *
1183 * A timer may have been added/modified/deleted either by the current IRQ,
1184 * or by another place using this IRQ as a notification. This IRQ may have
1185 * also updated the RCU callback list. These events may require a
1186 * re-evaluation of the next tick. Depending on the context:
1187 *
1188 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1189 * time accounting. The next tick will be re-evaluated on the next idle
1190 * loop iteration.
1191 *
1192 * 2) If the CPU is nohz_full:
1193 *
1194 * 2.1) If there is any tick dependency, restart the tick if stopped.
1195 *
1196 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1197 * stop/update it accordingly.
1198 */
1199void tick_nohz_irq_exit(void)
1200{
1201 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1202
1203 if (ts->inidle)
1204 tick_nohz_start_idle(ts);
1205 else
1206 tick_nohz_full_update_tick(ts);
1207}
1208
1209/**
1210 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1211 */
1212bool tick_nohz_idle_got_tick(void)
1213{
1214 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1215
1216 if (ts->got_idle_tick) {
1217 ts->got_idle_tick = 0;
1218 return true;
1219 }
1220 return false;
1221}
1222
1223/**
1224 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1225 * or the tick, whichever expires first. Note that, if the tick has been
1226 * stopped, it returns the next hrtimer.
1227 *
1228 * Called from power state control code with interrupts disabled
1229 */
1230ktime_t tick_nohz_get_next_hrtimer(void)
1231{
1232 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1233}
1234
1235/**
1236 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1237 * @delta_next: duration until the next event if the tick cannot be stopped
1238 *
1239 * Called from power state control code with interrupts disabled.
1240 *
1241 * The return value of this function and/or the value returned by it through the
1242 * @delta_next pointer can be negative which must be taken into account by its
1243 * callers.
1244 */
1245ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1246{
1247 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1248 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1249 int cpu = smp_processor_id();
1250 /*
1251 * The idle entry time is expected to be a sufficient approximation of
1252 * the current time at this point.
1253 */
1254 ktime_t now = ts->idle_entrytime;
1255 ktime_t next_event;
1256
1257 WARN_ON_ONCE(!ts->inidle);
1258
1259 *delta_next = ktime_sub(dev->next_event, now);
1260
1261 if (!can_stop_idle_tick(cpu, ts))
1262 return *delta_next;
1263
1264 next_event = tick_nohz_next_event(ts, cpu);
1265 if (!next_event)
1266 return *delta_next;
1267
1268 /*
1269 * If the next highres timer to expire is earlier than 'next_event', the
1270 * idle governor needs to know that.
1271 */
1272 next_event = min_t(u64, next_event,
1273 hrtimer_next_event_without(&ts->sched_timer));
1274
1275 return ktime_sub(next_event, now);
1276}
1277
1278/**
1279 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1280 * for a particular CPU.
1281 *
1282 * Called from the schedutil frequency scaling governor in scheduler context.
1283 */
1284unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1285{
1286 struct tick_sched *ts = tick_get_tick_sched(cpu);
1287
1288 return ts->idle_calls;
1289}
1290
1291/**
1292 * tick_nohz_get_idle_calls - return the current idle calls counter value
1293 *
1294 * Called from the schedutil frequency scaling governor in scheduler context.
1295 */
1296unsigned long tick_nohz_get_idle_calls(void)
1297{
1298 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1299
1300 return ts->idle_calls;
1301}
1302
1303static void tick_nohz_account_idle_time(struct tick_sched *ts,
1304 ktime_t now)
1305{
1306 unsigned long ticks;
1307
1308 ts->idle_exittime = now;
1309
1310 if (vtime_accounting_enabled_this_cpu())
1311 return;
1312 /*
1313 * We stopped the tick in idle. update_process_times() would miss the
1314 * time we slept, as it does only a 1 tick accounting.
1315 * Enforce that this is accounted to idle !
1316 */
1317 ticks = jiffies - ts->idle_jiffies;
1318 /*
1319 * We might be one off. Do not randomly account a huge number of ticks!
1320 */
1321 if (ticks && ticks < LONG_MAX)
1322 account_idle_ticks(ticks);
1323}
1324
1325void tick_nohz_idle_restart_tick(void)
1326{
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1328
1329 if (ts->tick_stopped) {
1330 ktime_t now = ktime_get();
1331 tick_nohz_restart_sched_tick(ts, now);
1332 tick_nohz_account_idle_time(ts, now);
1333 }
1334}
1335
1336static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1337{
1338 if (tick_nohz_full_cpu(smp_processor_id()))
1339 __tick_nohz_full_update_tick(ts, now);
1340 else
1341 tick_nohz_restart_sched_tick(ts, now);
1342
1343 tick_nohz_account_idle_time(ts, now);
1344}
1345
1346/**
1347 * tick_nohz_idle_exit - Update the tick upon idle task exit
1348 *
1349 * When the idle task exits, update the tick depending on the
1350 * following situations:
1351 *
1352 * 1) If the CPU is not in nohz_full mode (most cases), then
1353 * restart the tick.
1354 *
1355 * 2) If the CPU is in nohz_full mode (corner case):
1356 * 2.1) If the tick can be kept stopped (no tick dependencies)
1357 * then re-evaluate the next tick and try to keep it stopped
1358 * as long as possible.
1359 * 2.2) If the tick has dependencies, restart the tick.
1360 *
1361 */
1362void tick_nohz_idle_exit(void)
1363{
1364 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1365 bool idle_active, tick_stopped;
1366 ktime_t now;
1367
1368 local_irq_disable();
1369
1370 WARN_ON_ONCE(!ts->inidle);
1371 WARN_ON_ONCE(ts->timer_expires_base);
1372
1373 ts->inidle = 0;
1374 idle_active = ts->idle_active;
1375 tick_stopped = ts->tick_stopped;
1376
1377 if (idle_active || tick_stopped)
1378 now = ktime_get();
1379
1380 if (idle_active)
1381 tick_nohz_stop_idle(ts, now);
1382
1383 if (tick_stopped)
1384 tick_nohz_idle_update_tick(ts, now);
1385
1386 local_irq_enable();
1387}
1388
1389/*
1390 * In low-resolution mode, the tick handler must be implemented directly
1391 * at the clockevent level. hrtimer can't be used instead, because its
1392 * infrastructure actually relies on the tick itself as a backend in
1393 * low-resolution mode (see hrtimer_run_queues()).
1394 *
1395 * This low-resolution handler still makes use of some hrtimer APIs meanwhile
1396 * for convenience with expiration calculation and forwarding.
1397 */
1398static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1399{
1400 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1401 struct pt_regs *regs = get_irq_regs();
1402 ktime_t now = ktime_get();
1403
1404 dev->next_event = KTIME_MAX;
1405
1406 tick_sched_do_timer(ts, now);
1407 tick_sched_handle(ts, regs);
1408
1409 /*
1410 * In dynticks mode, tick reprogram is deferred:
1411 * - to the idle task if in dynticks-idle
1412 * - to IRQ exit if in full-dynticks.
1413 */
1414 if (likely(!ts->tick_stopped)) {
1415 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 }
1418
1419}
1420
1421static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1422{
1423 if (!tick_nohz_enabled)
1424 return;
1425 ts->nohz_mode = mode;
1426 /* One update is enough */
1427 if (!test_and_set_bit(0, &tick_nohz_active))
1428 timers_update_nohz();
1429}
1430
1431/**
1432 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1433 */
1434static void tick_nohz_switch_to_nohz(void)
1435{
1436 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1437 ktime_t next;
1438
1439 if (!tick_nohz_enabled)
1440 return;
1441
1442 if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1443 return;
1444
1445 /*
1446 * Recycle the hrtimer in 'ts', so we can share the
1447 * hrtimer_forward_now() function with the highres code.
1448 */
1449 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1450 /* Get the next period */
1451 next = tick_init_jiffy_update();
1452
1453 hrtimer_set_expires(&ts->sched_timer, next);
1454 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1455 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1456 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1457}
1458
1459static inline void tick_nohz_irq_enter(void)
1460{
1461 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1462 ktime_t now;
1463
1464 if (!ts->idle_active && !ts->tick_stopped)
1465 return;
1466 now = ktime_get();
1467 if (ts->idle_active)
1468 tick_nohz_stop_idle(ts, now);
1469 /*
1470 * If all CPUs are idle we may need to update a stale jiffies value.
1471 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1472 * alive but it might be busy looping with interrupts disabled in some
1473 * rare case (typically stop machine). So we must make sure we have a
1474 * last resort.
1475 */
1476 if (ts->tick_stopped)
1477 tick_nohz_update_jiffies(now);
1478}
1479
1480#else
1481
1482static inline void tick_nohz_switch_to_nohz(void) { }
1483static inline void tick_nohz_irq_enter(void) { }
1484static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1485
1486#endif /* CONFIG_NO_HZ_COMMON */
1487
1488/*
1489 * Called from irq_enter() to notify about the possible interruption of idle()
1490 */
1491void tick_irq_enter(void)
1492{
1493 tick_check_oneshot_broadcast_this_cpu();
1494 tick_nohz_irq_enter();
1495}
1496
1497/*
1498 * High resolution timer specific code
1499 */
1500#ifdef CONFIG_HIGH_RES_TIMERS
1501/*
1502 * We rearm the timer until we get disabled by the idle code.
1503 * Called with interrupts disabled.
1504 */
1505static enum hrtimer_restart tick_nohz_highres_handler(struct hrtimer *timer)
1506{
1507 struct tick_sched *ts =
1508 container_of(timer, struct tick_sched, sched_timer);
1509 struct pt_regs *regs = get_irq_regs();
1510 ktime_t now = ktime_get();
1511
1512 tick_sched_do_timer(ts, now);
1513
1514 /*
1515 * Do not call when we are not in IRQ context and have
1516 * no valid 'regs' pointer
1517 */
1518 if (regs)
1519 tick_sched_handle(ts, regs);
1520 else
1521 ts->next_tick = 0;
1522
1523 /*
1524 * In dynticks mode, tick reprogram is deferred:
1525 * - to the idle task if in dynticks-idle
1526 * - to IRQ exit if in full-dynticks.
1527 */
1528 if (unlikely(ts->tick_stopped))
1529 return HRTIMER_NORESTART;
1530
1531 hrtimer_forward(timer, now, TICK_NSEC);
1532
1533 return HRTIMER_RESTART;
1534}
1535
1536static int sched_skew_tick;
1537
1538static int __init skew_tick(char *str)
1539{
1540 get_option(&str, &sched_skew_tick);
1541
1542 return 0;
1543}
1544early_param("skew_tick", skew_tick);
1545
1546/**
1547 * tick_setup_sched_timer - setup the tick emulation timer
1548 */
1549void tick_setup_sched_timer(void)
1550{
1551 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1552 ktime_t now = ktime_get();
1553
1554 /* Emulate tick processing via per-CPU hrtimers: */
1555 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1556 ts->sched_timer.function = tick_nohz_highres_handler;
1557
1558 /* Get the next period (per-CPU) */
1559 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1560
1561 /* Offset the tick to avert 'jiffies_lock' contention. */
1562 if (sched_skew_tick) {
1563 u64 offset = TICK_NSEC >> 1;
1564 do_div(offset, num_possible_cpus());
1565 offset *= smp_processor_id();
1566 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1567 }
1568
1569 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1570 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1571 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1572}
1573#endif /* HIGH_RES_TIMERS */
1574
1575#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1576void tick_cancel_sched_timer(int cpu)
1577{
1578 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1579 ktime_t idle_sleeptime, iowait_sleeptime;
1580 unsigned long idle_calls, idle_sleeps;
1581
1582# ifdef CONFIG_HIGH_RES_TIMERS
1583 if (ts->sched_timer.base)
1584 hrtimer_cancel(&ts->sched_timer);
1585# endif
1586
1587 idle_sleeptime = ts->idle_sleeptime;
1588 iowait_sleeptime = ts->iowait_sleeptime;
1589 idle_calls = ts->idle_calls;
1590 idle_sleeps = ts->idle_sleeps;
1591 memset(ts, 0, sizeof(*ts));
1592 ts->idle_sleeptime = idle_sleeptime;
1593 ts->iowait_sleeptime = iowait_sleeptime;
1594 ts->idle_calls = idle_calls;
1595 ts->idle_sleeps = idle_sleeps;
1596}
1597#endif
1598
1599/*
1600 * Async notification about clocksource changes
1601 */
1602void tick_clock_notify(void)
1603{
1604 int cpu;
1605
1606 for_each_possible_cpu(cpu)
1607 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1608}
1609
1610/*
1611 * Async notification about clock event changes
1612 */
1613void tick_oneshot_notify(void)
1614{
1615 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1616
1617 set_bit(0, &ts->check_clocks);
1618}
1619
1620/*
1621 * Check if a change happened, which makes oneshot possible.
1622 *
1623 * Called cyclically from the hrtimer softirq (driven by the timer
1624 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1625 * mode, because high resolution timers are disabled (either compile
1626 * or runtime). Called with interrupts disabled.
1627 */
1628int tick_check_oneshot_change(int allow_nohz)
1629{
1630 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1631
1632 if (!test_and_clear_bit(0, &ts->check_clocks))
1633 return 0;
1634
1635 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1636 return 0;
1637
1638 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1639 return 0;
1640
1641 if (!allow_nohz)
1642 return 1;
1643
1644 tick_nohz_switch_to_nohz();
1645 return 0;
1646}