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
24#include <asm/irq_regs.h>
25
26#include "tick-internal.h"
27
28/*
29 * Per cpu nohz control structure
30 */
31static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
32
33/*
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
35 */
36static ktime_t last_jiffies_update;
37
38struct tick_sched *tick_get_tick_sched(int cpu)
39{
40 return &per_cpu(tick_cpu_sched, cpu);
41}
42
43/*
44 * Must be called with interrupts disabled !
45 */
46static void tick_do_update_jiffies64(ktime_t now)
47{
48 unsigned long ticks = 0;
49 ktime_t delta;
50
51 /*
52 * Do a quick check without holding xtime_lock:
53 */
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 < tick_period.tv64)
56 return;
57
58 /* Reevalute with xtime_lock held */
59 write_seqlock(&xtime_lock);
60
61 delta = ktime_sub(now, last_jiffies_update);
62 if (delta.tv64 >= tick_period.tv64) {
63
64 delta = ktime_sub(delta, tick_period);
65 last_jiffies_update = ktime_add(last_jiffies_update,
66 tick_period);
67
68 /* Slow path for long timeouts */
69 if (unlikely(delta.tv64 >= tick_period.tv64)) {
70 s64 incr = ktime_to_ns(tick_period);
71
72 ticks = ktime_divns(delta, incr);
73
74 last_jiffies_update = ktime_add_ns(last_jiffies_update,
75 incr * ticks);
76 }
77 do_timer(++ticks);
78
79 /* Keep the tick_next_period variable up to date */
80 tick_next_period = ktime_add(last_jiffies_update, tick_period);
81 }
82 write_sequnlock(&xtime_lock);
83}
84
85/*
86 * Initialize and return retrieve the jiffies update.
87 */
88static ktime_t tick_init_jiffy_update(void)
89{
90 ktime_t period;
91
92 write_seqlock(&xtime_lock);
93 /* Did we start the jiffies update yet ? */
94 if (last_jiffies_update.tv64 == 0)
95 last_jiffies_update = tick_next_period;
96 period = last_jiffies_update;
97 write_sequnlock(&xtime_lock);
98 return period;
99}
100
101/*
102 * NOHZ - aka dynamic tick functionality
103 */
104#ifdef CONFIG_NO_HZ
105/*
106 * NO HZ enabled ?
107 */
108static int tick_nohz_enabled __read_mostly = 1;
109
110/*
111 * Enable / Disable tickless mode
112 */
113static int __init setup_tick_nohz(char *str)
114{
115 if (!strcmp(str, "off"))
116 tick_nohz_enabled = 0;
117 else if (!strcmp(str, "on"))
118 tick_nohz_enabled = 1;
119 else
120 return 0;
121 return 1;
122}
123
124__setup("nohz=", setup_tick_nohz);
125
126/**
127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
128 *
129 * Called from interrupt entry when the CPU was idle
130 *
131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
133 * value. We do this unconditionally on any cpu, as we don't know whether the
134 * cpu, which has the update task assigned is in a long sleep.
135 */
136static void tick_nohz_update_jiffies(ktime_t now)
137{
138 int cpu = smp_processor_id();
139 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
140 unsigned long flags;
141
142 ts->idle_waketime = now;
143
144 local_irq_save(flags);
145 tick_do_update_jiffies64(now);
146 local_irq_restore(flags);
147
148 calc_load_exit_idle();
149 touch_softlockup_watchdog();
150}
151
152/*
153 * Updates the per cpu time idle statistics counters
154 */
155static void
156update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
157{
158 ktime_t delta;
159
160 if (ts->idle_active) {
161 delta = ktime_sub(now, ts->idle_entrytime);
162 if (nr_iowait_cpu(cpu) > 0)
163 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
164 else
165 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
166 ts->idle_entrytime = now;
167 }
168
169 if (last_update_time)
170 *last_update_time = ktime_to_us(now);
171
172}
173
174static void tick_nohz_stop_idle(int cpu, ktime_t now)
175{
176 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
177
178 update_ts_time_stats(cpu, ts, now, NULL);
179 ts->idle_active = 0;
180
181 sched_clock_idle_wakeup_event(0);
182}
183
184static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
185{
186 ktime_t now = ktime_get();
187
188 ts->idle_entrytime = now;
189 ts->idle_active = 1;
190 sched_clock_idle_sleep_event();
191 return now;
192}
193
194/**
195 * get_cpu_idle_time_us - get the total idle time of a cpu
196 * @cpu: CPU number to query
197 * @last_update_time: variable to store update time in. Do not update
198 * counters if NULL.
199 *
200 * Return the cummulative idle time (since boot) for a given
201 * CPU, in microseconds.
202 *
203 * This time is measured via accounting rather than sampling,
204 * and is as accurate as ktime_get() is.
205 *
206 * This function returns -1 if NOHZ is not enabled.
207 */
208u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
209{
210 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
211 ktime_t now, idle;
212
213 if (!tick_nohz_enabled)
214 return -1;
215
216 now = ktime_get();
217 if (last_update_time) {
218 update_ts_time_stats(cpu, ts, now, last_update_time);
219 idle = ts->idle_sleeptime;
220 } else {
221 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
222 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
223
224 idle = ktime_add(ts->idle_sleeptime, delta);
225 } else {
226 idle = ts->idle_sleeptime;
227 }
228 }
229
230 return ktime_to_us(idle);
231
232}
233EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
234
235/**
236 * get_cpu_iowait_time_us - get the total iowait time of a cpu
237 * @cpu: CPU number to query
238 * @last_update_time: variable to store update time in. Do not update
239 * counters if NULL.
240 *
241 * Return the cummulative iowait time (since boot) for a given
242 * CPU, in microseconds.
243 *
244 * This time is measured via accounting rather than sampling,
245 * and is as accurate as ktime_get() is.
246 *
247 * This function returns -1 if NOHZ is not enabled.
248 */
249u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
250{
251 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
252 ktime_t now, iowait;
253
254 if (!tick_nohz_enabled)
255 return -1;
256
257 now = ktime_get();
258 if (last_update_time) {
259 update_ts_time_stats(cpu, ts, now, last_update_time);
260 iowait = ts->iowait_sleeptime;
261 } else {
262 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
263 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
264
265 iowait = ktime_add(ts->iowait_sleeptime, delta);
266 } else {
267 iowait = ts->iowait_sleeptime;
268 }
269 }
270
271 return ktime_to_us(iowait);
272}
273EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
274
275static void tick_nohz_stop_sched_tick(struct tick_sched *ts)
276{
277 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
278 unsigned long rcu_delta_jiffies;
279 ktime_t last_update, expires, now;
280 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
281 u64 time_delta;
282 int cpu;
283
284 cpu = smp_processor_id();
285 ts = &per_cpu(tick_cpu_sched, cpu);
286
287 now = tick_nohz_start_idle(cpu, ts);
288
289 /*
290 * If this cpu is offline and it is the one which updates
291 * jiffies, then give up the assignment and let it be taken by
292 * the cpu which runs the tick timer next. If we don't drop
293 * this here the jiffies might be stale and do_timer() never
294 * invoked.
295 */
296 if (unlikely(!cpu_online(cpu))) {
297 if (cpu == tick_do_timer_cpu)
298 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
299 }
300
301 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
302 return;
303
304 if (need_resched())
305 return;
306
307 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
308 static int ratelimit;
309
310 if (ratelimit < 10) {
311 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
312 (unsigned int) local_softirq_pending());
313 ratelimit++;
314 }
315 return;
316 }
317
318 ts->idle_calls++;
319 /* Read jiffies and the time when jiffies were updated last */
320 do {
321 seq = read_seqbegin(&xtime_lock);
322 last_update = last_jiffies_update;
323 last_jiffies = jiffies;
324 time_delta = timekeeping_max_deferment();
325 } while (read_seqretry(&xtime_lock, seq));
326
327 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) ||
328 arch_needs_cpu(cpu)) {
329 next_jiffies = last_jiffies + 1;
330 delta_jiffies = 1;
331 } else {
332 /* Get the next timer wheel timer */
333 next_jiffies = get_next_timer_interrupt(last_jiffies);
334 delta_jiffies = next_jiffies - last_jiffies;
335 if (rcu_delta_jiffies < delta_jiffies) {
336 next_jiffies = last_jiffies + rcu_delta_jiffies;
337 delta_jiffies = rcu_delta_jiffies;
338 }
339 }
340 /*
341 * Do not stop the tick, if we are only one off
342 * or if the cpu is required for rcu
343 */
344 if (!ts->tick_stopped && delta_jiffies == 1)
345 goto out;
346
347 /* Schedule the tick, if we are at least one jiffie off */
348 if ((long)delta_jiffies >= 1) {
349
350 /*
351 * If this cpu is the one which updates jiffies, then
352 * give up the assignment and let it be taken by the
353 * cpu which runs the tick timer next, which might be
354 * this cpu as well. If we don't drop this here the
355 * jiffies might be stale and do_timer() never
356 * invoked. Keep track of the fact that it was the one
357 * which had the do_timer() duty last. If this cpu is
358 * the one which had the do_timer() duty last, we
359 * limit the sleep time to the timekeeping
360 * max_deferement value which we retrieved
361 * above. Otherwise we can sleep as long as we want.
362 */
363 if (cpu == tick_do_timer_cpu) {
364 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
365 ts->do_timer_last = 1;
366 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
367 time_delta = KTIME_MAX;
368 ts->do_timer_last = 0;
369 } else if (!ts->do_timer_last) {
370 time_delta = KTIME_MAX;
371 }
372
373 /*
374 * calculate the expiry time for the next timer wheel
375 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
376 * that there is no timer pending or at least extremely
377 * far into the future (12 days for HZ=1000). In this
378 * case we set the expiry to the end of time.
379 */
380 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
381 /*
382 * Calculate the time delta for the next timer event.
383 * If the time delta exceeds the maximum time delta
384 * permitted by the current clocksource then adjust
385 * the time delta accordingly to ensure the
386 * clocksource does not wrap.
387 */
388 time_delta = min_t(u64, time_delta,
389 tick_period.tv64 * delta_jiffies);
390 }
391
392 if (time_delta < KTIME_MAX)
393 expires = ktime_add_ns(last_update, time_delta);
394 else
395 expires.tv64 = KTIME_MAX;
396
397 /* Skip reprogram of event if its not changed */
398 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
399 goto out;
400
401 /*
402 * nohz_stop_sched_tick can be called several times before
403 * the nohz_restart_sched_tick is called. This happens when
404 * interrupts arrive which do not cause a reschedule. In the
405 * first call we save the current tick time, so we can restart
406 * the scheduler tick in nohz_restart_sched_tick.
407 */
408 if (!ts->tick_stopped) {
409 select_nohz_load_balancer(1);
410 calc_load_enter_idle();
411
412 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
413 ts->tick_stopped = 1;
414 ts->idle_jiffies = last_jiffies;
415 }
416
417 ts->idle_sleeps++;
418
419 /* Mark expires */
420 ts->idle_expires = expires;
421
422 /*
423 * If the expiration time == KTIME_MAX, then
424 * in this case we simply stop the tick timer.
425 */
426 if (unlikely(expires.tv64 == KTIME_MAX)) {
427 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
428 hrtimer_cancel(&ts->sched_timer);
429 goto out;
430 }
431
432 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
433 hrtimer_start(&ts->sched_timer, expires,
434 HRTIMER_MODE_ABS_PINNED);
435 /* Check, if the timer was already in the past */
436 if (hrtimer_active(&ts->sched_timer))
437 goto out;
438 } else if (!tick_program_event(expires, 0))
439 goto out;
440 /*
441 * We are past the event already. So we crossed a
442 * jiffie boundary. Update jiffies and raise the
443 * softirq.
444 */
445 tick_do_update_jiffies64(ktime_get());
446 }
447 raise_softirq_irqoff(TIMER_SOFTIRQ);
448out:
449 ts->next_jiffies = next_jiffies;
450 ts->last_jiffies = last_jiffies;
451 ts->sleep_length = ktime_sub(dev->next_event, now);
452}
453
454/**
455 * tick_nohz_idle_enter - stop the idle tick from the idle task
456 *
457 * When the next event is more than a tick into the future, stop the idle tick
458 * Called when we start the idle loop.
459 *
460 * The arch is responsible of calling:
461 *
462 * - rcu_idle_enter() after its last use of RCU before the CPU is put
463 * to sleep.
464 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
465 */
466void tick_nohz_idle_enter(void)
467{
468 struct tick_sched *ts;
469
470 WARN_ON_ONCE(irqs_disabled());
471
472 /*
473 * Update the idle state in the scheduler domain hierarchy
474 * when tick_nohz_stop_sched_tick() is called from the idle loop.
475 * State will be updated to busy during the first busy tick after
476 * exiting idle.
477 */
478 set_cpu_sd_state_idle();
479
480 local_irq_disable();
481
482 ts = &__get_cpu_var(tick_cpu_sched);
483 /*
484 * set ts->inidle unconditionally. even if the system did not
485 * switch to nohz mode the cpu frequency governers rely on the
486 * update of the idle time accounting in tick_nohz_start_idle().
487 */
488 ts->inidle = 1;
489 tick_nohz_stop_sched_tick(ts);
490
491 local_irq_enable();
492}
493
494/**
495 * tick_nohz_irq_exit - update next tick event from interrupt exit
496 *
497 * When an interrupt fires while we are idle and it doesn't cause
498 * a reschedule, it may still add, modify or delete a timer, enqueue
499 * an RCU callback, etc...
500 * So we need to re-calculate and reprogram the next tick event.
501 */
502void tick_nohz_irq_exit(void)
503{
504 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
505
506 if (!ts->inidle)
507 return;
508
509 tick_nohz_stop_sched_tick(ts);
510}
511
512/**
513 * tick_nohz_get_sleep_length - return the length of the current sleep
514 *
515 * Called from power state control code with interrupts disabled
516 */
517ktime_t tick_nohz_get_sleep_length(void)
518{
519 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
520
521 return ts->sleep_length;
522}
523
524static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
525{
526 hrtimer_cancel(&ts->sched_timer);
527 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
528
529 while (1) {
530 /* Forward the time to expire in the future */
531 hrtimer_forward(&ts->sched_timer, now, tick_period);
532
533 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
534 hrtimer_start_expires(&ts->sched_timer,
535 HRTIMER_MODE_ABS_PINNED);
536 /* Check, if the timer was already in the past */
537 if (hrtimer_active(&ts->sched_timer))
538 break;
539 } else {
540 if (!tick_program_event(
541 hrtimer_get_expires(&ts->sched_timer), 0))
542 break;
543 }
544 /* Reread time and update jiffies */
545 now = ktime_get();
546 tick_do_update_jiffies64(now);
547 }
548}
549
550/**
551 * tick_nohz_idle_exit - restart the idle tick from the idle task
552 *
553 * Restart the idle tick when the CPU is woken up from idle
554 * This also exit the RCU extended quiescent state. The CPU
555 * can use RCU again after this function is called.
556 */
557void tick_nohz_idle_exit(void)
558{
559 int cpu = smp_processor_id();
560 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
561#ifndef CONFIG_VIRT_CPU_ACCOUNTING
562 unsigned long ticks;
563#endif
564 ktime_t now;
565
566 local_irq_disable();
567
568 WARN_ON_ONCE(!ts->inidle);
569
570 ts->inidle = 0;
571
572 if (ts->idle_active || ts->tick_stopped)
573 now = ktime_get();
574
575 if (ts->idle_active)
576 tick_nohz_stop_idle(cpu, now);
577
578 if (!ts->tick_stopped) {
579 local_irq_enable();
580 return;
581 }
582
583 /* Update jiffies first */
584 select_nohz_load_balancer(0);
585 tick_do_update_jiffies64(now);
586 update_cpu_load_nohz();
587
588#ifndef CONFIG_VIRT_CPU_ACCOUNTING
589 /*
590 * We stopped the tick in idle. Update process times would miss the
591 * time we slept as update_process_times does only a 1 tick
592 * accounting. Enforce that this is accounted to idle !
593 */
594 ticks = jiffies - ts->idle_jiffies;
595 /*
596 * We might be one off. Do not randomly account a huge number of ticks!
597 */
598 if (ticks && ticks < LONG_MAX)
599 account_idle_ticks(ticks);
600#endif
601
602 calc_load_exit_idle();
603 touch_softlockup_watchdog();
604 /*
605 * Cancel the scheduled timer and restore the tick
606 */
607 ts->tick_stopped = 0;
608 ts->idle_exittime = now;
609
610 tick_nohz_restart(ts, now);
611
612 local_irq_enable();
613}
614
615static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
616{
617 hrtimer_forward(&ts->sched_timer, now, tick_period);
618 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
619}
620
621/*
622 * The nohz low res interrupt handler
623 */
624static void tick_nohz_handler(struct clock_event_device *dev)
625{
626 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
627 struct pt_regs *regs = get_irq_regs();
628 int cpu = smp_processor_id();
629 ktime_t now = ktime_get();
630
631 dev->next_event.tv64 = KTIME_MAX;
632
633 /*
634 * Check if the do_timer duty was dropped. We don't care about
635 * concurrency: This happens only when the cpu in charge went
636 * into a long sleep. If two cpus happen to assign themself to
637 * this duty, then the jiffies update is still serialized by
638 * xtime_lock.
639 */
640 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
641 tick_do_timer_cpu = cpu;
642
643 /* Check, if the jiffies need an update */
644 if (tick_do_timer_cpu == cpu)
645 tick_do_update_jiffies64(now);
646
647 /*
648 * When we are idle and the tick is stopped, we have to touch
649 * the watchdog as we might not schedule for a really long
650 * time. This happens on complete idle SMP systems while
651 * waiting on the login prompt. We also increment the "start
652 * of idle" jiffy stamp so the idle accounting adjustment we
653 * do when we go busy again does not account too much ticks.
654 */
655 if (ts->tick_stopped) {
656 touch_softlockup_watchdog();
657 ts->idle_jiffies++;
658 }
659
660 update_process_times(user_mode(regs));
661 profile_tick(CPU_PROFILING);
662
663 while (tick_nohz_reprogram(ts, now)) {
664 now = ktime_get();
665 tick_do_update_jiffies64(now);
666 }
667}
668
669/**
670 * tick_nohz_switch_to_nohz - switch to nohz mode
671 */
672static void tick_nohz_switch_to_nohz(void)
673{
674 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
675 ktime_t next;
676
677 if (!tick_nohz_enabled)
678 return;
679
680 local_irq_disable();
681 if (tick_switch_to_oneshot(tick_nohz_handler)) {
682 local_irq_enable();
683 return;
684 }
685
686 ts->nohz_mode = NOHZ_MODE_LOWRES;
687
688 /*
689 * Recycle the hrtimer in ts, so we can share the
690 * hrtimer_forward with the highres code.
691 */
692 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
693 /* Get the next period */
694 next = tick_init_jiffy_update();
695
696 for (;;) {
697 hrtimer_set_expires(&ts->sched_timer, next);
698 if (!tick_program_event(next, 0))
699 break;
700 next = ktime_add(next, tick_period);
701 }
702 local_irq_enable();
703}
704
705/*
706 * When NOHZ is enabled and the tick is stopped, we need to kick the
707 * tick timer from irq_enter() so that the jiffies update is kept
708 * alive during long running softirqs. That's ugly as hell, but
709 * correctness is key even if we need to fix the offending softirq in
710 * the first place.
711 *
712 * Note, this is different to tick_nohz_restart. We just kick the
713 * timer and do not touch the other magic bits which need to be done
714 * when idle is left.
715 */
716static void tick_nohz_kick_tick(int cpu, ktime_t now)
717{
718#if 0
719 /* Switch back to 2.6.27 behaviour */
720
721 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
722 ktime_t delta;
723
724 /*
725 * Do not touch the tick device, when the next expiry is either
726 * already reached or less/equal than the tick period.
727 */
728 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
729 if (delta.tv64 <= tick_period.tv64)
730 return;
731
732 tick_nohz_restart(ts, now);
733#endif
734}
735
736static inline void tick_check_nohz(int cpu)
737{
738 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
739 ktime_t now;
740
741 if (!ts->idle_active && !ts->tick_stopped)
742 return;
743 now = ktime_get();
744 if (ts->idle_active)
745 tick_nohz_stop_idle(cpu, now);
746 if (ts->tick_stopped) {
747 tick_nohz_update_jiffies(now);
748 tick_nohz_kick_tick(cpu, now);
749 }
750}
751
752#else
753
754static inline void tick_nohz_switch_to_nohz(void) { }
755static inline void tick_check_nohz(int cpu) { }
756
757#endif /* NO_HZ */
758
759/*
760 * Called from irq_enter to notify about the possible interruption of idle()
761 */
762void tick_check_idle(int cpu)
763{
764 tick_check_oneshot_broadcast(cpu);
765 tick_check_nohz(cpu);
766}
767
768/*
769 * High resolution timer specific code
770 */
771#ifdef CONFIG_HIGH_RES_TIMERS
772/*
773 * We rearm the timer until we get disabled by the idle code.
774 * Called with interrupts disabled and timer->base->cpu_base->lock held.
775 */
776static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
777{
778 struct tick_sched *ts =
779 container_of(timer, struct tick_sched, sched_timer);
780 struct pt_regs *regs = get_irq_regs();
781 ktime_t now = ktime_get();
782 int cpu = smp_processor_id();
783
784#ifdef CONFIG_NO_HZ
785 /*
786 * Check if the do_timer duty was dropped. We don't care about
787 * concurrency: This happens only when the cpu in charge went
788 * into a long sleep. If two cpus happen to assign themself to
789 * this duty, then the jiffies update is still serialized by
790 * xtime_lock.
791 */
792 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
793 tick_do_timer_cpu = cpu;
794#endif
795
796 /* Check, if the jiffies need an update */
797 if (tick_do_timer_cpu == cpu)
798 tick_do_update_jiffies64(now);
799
800 /*
801 * Do not call, when we are not in irq context and have
802 * no valid regs pointer
803 */
804 if (regs) {
805 /*
806 * When we are idle and the tick is stopped, we have to touch
807 * the watchdog as we might not schedule for a really long
808 * time. This happens on complete idle SMP systems while
809 * waiting on the login prompt. We also increment the "start of
810 * idle" jiffy stamp so the idle accounting adjustment we do
811 * when we go busy again does not account too much ticks.
812 */
813 if (ts->tick_stopped) {
814 touch_softlockup_watchdog();
815 ts->idle_jiffies++;
816 }
817 update_process_times(user_mode(regs));
818 profile_tick(CPU_PROFILING);
819 }
820
821 hrtimer_forward(timer, now, tick_period);
822
823 return HRTIMER_RESTART;
824}
825
826static int sched_skew_tick;
827
828static int __init skew_tick(char *str)
829{
830 get_option(&str, &sched_skew_tick);
831
832 return 0;
833}
834early_param("skew_tick", skew_tick);
835
836/**
837 * tick_setup_sched_timer - setup the tick emulation timer
838 */
839void tick_setup_sched_timer(void)
840{
841 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
842 ktime_t now = ktime_get();
843
844 /*
845 * Emulate tick processing via per-CPU hrtimers:
846 */
847 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
848 ts->sched_timer.function = tick_sched_timer;
849
850 /* Get the next period (per cpu) */
851 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
852
853 /* Offset the tick to avert xtime_lock contention. */
854 if (sched_skew_tick) {
855 u64 offset = ktime_to_ns(tick_period) >> 1;
856 do_div(offset, num_possible_cpus());
857 offset *= smp_processor_id();
858 hrtimer_add_expires_ns(&ts->sched_timer, offset);
859 }
860
861 for (;;) {
862 hrtimer_forward(&ts->sched_timer, now, tick_period);
863 hrtimer_start_expires(&ts->sched_timer,
864 HRTIMER_MODE_ABS_PINNED);
865 /* Check, if the timer was already in the past */
866 if (hrtimer_active(&ts->sched_timer))
867 break;
868 now = ktime_get();
869 }
870
871#ifdef CONFIG_NO_HZ
872 if (tick_nohz_enabled)
873 ts->nohz_mode = NOHZ_MODE_HIGHRES;
874#endif
875}
876#endif /* HIGH_RES_TIMERS */
877
878#if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
879void tick_cancel_sched_timer(int cpu)
880{
881 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
882
883# ifdef CONFIG_HIGH_RES_TIMERS
884 if (ts->sched_timer.base)
885 hrtimer_cancel(&ts->sched_timer);
886# endif
887
888 ts->nohz_mode = NOHZ_MODE_INACTIVE;
889}
890#endif
891
892/**
893 * Async notification about clocksource changes
894 */
895void tick_clock_notify(void)
896{
897 int cpu;
898
899 for_each_possible_cpu(cpu)
900 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
901}
902
903/*
904 * Async notification about clock event changes
905 */
906void tick_oneshot_notify(void)
907{
908 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
909
910 set_bit(0, &ts->check_clocks);
911}
912
913/**
914 * Check, if a change happened, which makes oneshot possible.
915 *
916 * Called cyclic from the hrtimer softirq (driven by the timer
917 * softirq) allow_nohz signals, that we can switch into low-res nohz
918 * mode, because high resolution timers are disabled (either compile
919 * or runtime).
920 */
921int tick_check_oneshot_change(int allow_nohz)
922{
923 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
924
925 if (!test_and_clear_bit(0, &ts->check_clocks))
926 return 0;
927
928 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
929 return 0;
930
931 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
932 return 0;
933
934 if (!allow_nohz)
935 return 1;
936
937 tick_nohz_switch_to_nohz();
938 return 0;
939}
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/compiler.h>
12#include <linux/cpu.h>
13#include <linux/err.h>
14#include <linux/hrtimer.h>
15#include <linux/interrupt.h>
16#include <linux/kernel_stat.h>
17#include <linux/percpu.h>
18#include <linux/nmi.h>
19#include <linux/profile.h>
20#include <linux/sched/signal.h>
21#include <linux/sched/clock.h>
22#include <linux/sched/stat.h>
23#include <linux/sched/nohz.h>
24#include <linux/sched/loadavg.h>
25#include <linux/module.h>
26#include <linux/irq_work.h>
27#include <linux/posix-timers.h>
28#include <linux/context_tracking.h>
29#include <linux/mm.h>
30
31#include <asm/irq_regs.h>
32
33#include "tick-internal.h"
34
35#include <trace/events/timer.h>
36
37/*
38 * Per-CPU nohz control structure
39 */
40static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
41
42struct tick_sched *tick_get_tick_sched(int cpu)
43{
44 return &per_cpu(tick_cpu_sched, cpu);
45}
46
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
184static inline int tick_sched_flag_test(struct tick_sched *ts,
185 unsigned long flag)
186{
187 return !!(ts->flags & flag);
188}
189
190static inline void tick_sched_flag_set(struct tick_sched *ts,
191 unsigned long flag)
192{
193 lockdep_assert_irqs_disabled();
194 ts->flags |= flag;
195}
196
197static inline void tick_sched_flag_clear(struct tick_sched *ts,
198 unsigned long flag)
199{
200 lockdep_assert_irqs_disabled();
201 ts->flags &= ~flag;
202}
203
204#define MAX_STALLED_JIFFIES 5
205
206static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
207{
208 int tick_cpu, cpu = smp_processor_id();
209
210 /*
211 * Check if the do_timer duty was dropped. We don't care about
212 * concurrency: This happens only when the CPU in charge went
213 * into a long sleep. If two CPUs happen to assign themselves to
214 * this duty, then the jiffies update is still serialized by
215 * 'jiffies_lock'.
216 *
217 * If nohz_full is enabled, this should not happen because the
218 * 'tick_do_timer_cpu' CPU never relinquishes.
219 */
220 tick_cpu = READ_ONCE(tick_do_timer_cpu);
221
222 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && unlikely(tick_cpu == TICK_DO_TIMER_NONE)) {
223#ifdef CONFIG_NO_HZ_FULL
224 WARN_ON_ONCE(tick_nohz_full_running);
225#endif
226 WRITE_ONCE(tick_do_timer_cpu, cpu);
227 tick_cpu = cpu;
228 }
229
230 /* Check if jiffies need an update */
231 if (tick_cpu == cpu)
232 tick_do_update_jiffies64(now);
233
234 /*
235 * If the jiffies update stalled for too long (timekeeper in stop_machine()
236 * or VMEXIT'ed for several msecs), force an update.
237 */
238 if (ts->last_tick_jiffies != jiffies) {
239 ts->stalled_jiffies = 0;
240 ts->last_tick_jiffies = READ_ONCE(jiffies);
241 } else {
242 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
243 tick_do_update_jiffies64(now);
244 ts->stalled_jiffies = 0;
245 ts->last_tick_jiffies = READ_ONCE(jiffies);
246 }
247 }
248
249 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE))
250 ts->got_idle_tick = 1;
251}
252
253static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
254{
255 /*
256 * When we are idle and the tick is stopped, we have to touch
257 * the watchdog as we might not schedule for a really long
258 * time. This happens on completely idle SMP systems while
259 * waiting on the login prompt. We also increment the "start of
260 * idle" jiffy stamp so the idle accounting adjustment we do
261 * when we go busy again does not account too many ticks.
262 */
263 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) &&
264 tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
265 touch_softlockup_watchdog_sched();
266 if (is_idle_task(current))
267 ts->idle_jiffies++;
268 /*
269 * In case the current tick fired too early past its expected
270 * expiration, make sure we don't bypass the next clock reprogramming
271 * to the same deadline.
272 */
273 ts->next_tick = 0;
274 }
275
276 update_process_times(user_mode(regs));
277 profile_tick(CPU_PROFILING);
278}
279
280/*
281 * We rearm the timer until we get disabled by the idle code.
282 * Called with interrupts disabled.
283 */
284static enum hrtimer_restart tick_nohz_handler(struct hrtimer *timer)
285{
286 struct tick_sched *ts = container_of(timer, struct tick_sched, sched_timer);
287 struct pt_regs *regs = get_irq_regs();
288 ktime_t now = ktime_get();
289
290 tick_sched_do_timer(ts, now);
291
292 /*
293 * Do not call when we are not in IRQ context and have
294 * no valid 'regs' pointer
295 */
296 if (regs)
297 tick_sched_handle(ts, regs);
298 else
299 ts->next_tick = 0;
300
301 /*
302 * In dynticks mode, tick reprogram is deferred:
303 * - to the idle task if in dynticks-idle
304 * - to IRQ exit if in full-dynticks.
305 */
306 if (unlikely(tick_sched_flag_test(ts, TS_FLAG_STOPPED)))
307 return HRTIMER_NORESTART;
308
309 hrtimer_forward(timer, now, TICK_NSEC);
310
311 return HRTIMER_RESTART;
312}
313
314static void tick_sched_timer_cancel(struct tick_sched *ts)
315{
316 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES))
317 hrtimer_cancel(&ts->sched_timer);
318 else if (tick_sched_flag_test(ts, TS_FLAG_NOHZ))
319 tick_program_event(KTIME_MAX, 1);
320}
321
322#ifdef CONFIG_NO_HZ_FULL
323cpumask_var_t tick_nohz_full_mask;
324EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
325bool tick_nohz_full_running;
326EXPORT_SYMBOL_GPL(tick_nohz_full_running);
327static atomic_t tick_dep_mask;
328
329static bool check_tick_dependency(atomic_t *dep)
330{
331 int val = atomic_read(dep);
332
333 if (val & TICK_DEP_MASK_POSIX_TIMER) {
334 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
335 return true;
336 }
337
338 if (val & TICK_DEP_MASK_PERF_EVENTS) {
339 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
340 return true;
341 }
342
343 if (val & TICK_DEP_MASK_SCHED) {
344 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
345 return true;
346 }
347
348 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
349 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
350 return true;
351 }
352
353 if (val & TICK_DEP_MASK_RCU) {
354 trace_tick_stop(0, TICK_DEP_MASK_RCU);
355 return true;
356 }
357
358 if (val & TICK_DEP_MASK_RCU_EXP) {
359 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
360 return true;
361 }
362
363 return false;
364}
365
366static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
367{
368 lockdep_assert_irqs_disabled();
369
370 if (unlikely(!cpu_online(cpu)))
371 return false;
372
373 if (check_tick_dependency(&tick_dep_mask))
374 return false;
375
376 if (check_tick_dependency(&ts->tick_dep_mask))
377 return false;
378
379 if (check_tick_dependency(¤t->tick_dep_mask))
380 return false;
381
382 if (check_tick_dependency(¤t->signal->tick_dep_mask))
383 return false;
384
385 return true;
386}
387
388static void nohz_full_kick_func(struct irq_work *work)
389{
390 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
391}
392
393static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
394 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
395
396/*
397 * Kick this CPU if it's full dynticks in order to force it to
398 * re-evaluate its dependency on the tick and restart it if necessary.
399 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
400 * is NMI safe.
401 */
402static void tick_nohz_full_kick(void)
403{
404 if (!tick_nohz_full_cpu(smp_processor_id()))
405 return;
406
407 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
408}
409
410/*
411 * Kick the CPU if it's full dynticks in order to force it to
412 * re-evaluate its dependency on the tick and restart it if necessary.
413 */
414void tick_nohz_full_kick_cpu(int cpu)
415{
416 if (!tick_nohz_full_cpu(cpu))
417 return;
418
419 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
420}
421
422static void tick_nohz_kick_task(struct task_struct *tsk)
423{
424 int cpu;
425
426 /*
427 * If the task is not running, run_posix_cpu_timers()
428 * has nothing to elapse, and an IPI can then be optimized out.
429 *
430 * activate_task() STORE p->tick_dep_mask
431 * STORE p->on_rq
432 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
433 * LOCK rq->lock LOAD p->on_rq
434 * smp_mb__after_spin_lock()
435 * tick_nohz_task_switch()
436 * LOAD p->tick_dep_mask
437 */
438 if (!sched_task_on_rq(tsk))
439 return;
440
441 /*
442 * If the task concurrently migrates to another CPU,
443 * we guarantee it sees the new tick dependency upon
444 * schedule.
445 *
446 * set_task_cpu(p, cpu);
447 * STORE p->cpu = @cpu
448 * __schedule() (switch to task 'p')
449 * LOCK rq->lock
450 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
451 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
452 * LOAD p->tick_dep_mask LOAD p->cpu
453 */
454 cpu = task_cpu(tsk);
455
456 preempt_disable();
457 if (cpu_online(cpu))
458 tick_nohz_full_kick_cpu(cpu);
459 preempt_enable();
460}
461
462/*
463 * Kick all full dynticks CPUs in order to force these to re-evaluate
464 * their dependency on the tick and restart it if necessary.
465 */
466static void tick_nohz_full_kick_all(void)
467{
468 int cpu;
469
470 if (!tick_nohz_full_running)
471 return;
472
473 preempt_disable();
474 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
475 tick_nohz_full_kick_cpu(cpu);
476 preempt_enable();
477}
478
479static void tick_nohz_dep_set_all(atomic_t *dep,
480 enum tick_dep_bits bit)
481{
482 int prev;
483
484 prev = atomic_fetch_or(BIT(bit), dep);
485 if (!prev)
486 tick_nohz_full_kick_all();
487}
488
489/*
490 * Set a global tick dependency. Used by perf events that rely on freq and
491 * unstable clocks.
492 */
493void tick_nohz_dep_set(enum tick_dep_bits bit)
494{
495 tick_nohz_dep_set_all(&tick_dep_mask, bit);
496}
497
498void tick_nohz_dep_clear(enum tick_dep_bits bit)
499{
500 atomic_andnot(BIT(bit), &tick_dep_mask);
501}
502
503/*
504 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
505 * manage event-throttling.
506 */
507void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
508{
509 int prev;
510 struct tick_sched *ts;
511
512 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
513
514 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
515 if (!prev) {
516 preempt_disable();
517 /* Perf needs local kick that is NMI safe */
518 if (cpu == smp_processor_id()) {
519 tick_nohz_full_kick();
520 } else {
521 /* Remote IRQ work not NMI-safe */
522 if (!WARN_ON_ONCE(in_nmi()))
523 tick_nohz_full_kick_cpu(cpu);
524 }
525 preempt_enable();
526 }
527}
528EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
529
530void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
531{
532 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
533
534 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
535}
536EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
537
538/*
539 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
540 * in order to elapse per task timers.
541 */
542void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
543{
544 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
545 tick_nohz_kick_task(tsk);
546}
547EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
548
549void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
550{
551 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
552}
553EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
554
555/*
556 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
557 * per process timers.
558 */
559void tick_nohz_dep_set_signal(struct task_struct *tsk,
560 enum tick_dep_bits bit)
561{
562 int prev;
563 struct signal_struct *sig = tsk->signal;
564
565 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
566 if (!prev) {
567 struct task_struct *t;
568
569 lockdep_assert_held(&tsk->sighand->siglock);
570 __for_each_thread(sig, t)
571 tick_nohz_kick_task(t);
572 }
573}
574
575void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
576{
577 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
578}
579
580/*
581 * Re-evaluate the need for the tick as we switch the current task.
582 * It might need the tick due to per task/process properties:
583 * perf events, posix CPU timers, ...
584 */
585void __tick_nohz_task_switch(void)
586{
587 struct tick_sched *ts;
588
589 if (!tick_nohz_full_cpu(smp_processor_id()))
590 return;
591
592 ts = this_cpu_ptr(&tick_cpu_sched);
593
594 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
595 if (atomic_read(¤t->tick_dep_mask) ||
596 atomic_read(¤t->signal->tick_dep_mask))
597 tick_nohz_full_kick();
598 }
599}
600
601/* Get the boot-time nohz CPU list from the kernel parameters. */
602void __init tick_nohz_full_setup(cpumask_var_t cpumask)
603{
604 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
605 cpumask_copy(tick_nohz_full_mask, cpumask);
606 tick_nohz_full_running = true;
607}
608
609bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
610{
611 /*
612 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
613 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
614 * CPUs. It must remain online when nohz full is enabled.
615 */
616 if (tick_nohz_full_running && READ_ONCE(tick_do_timer_cpu) == cpu)
617 return false;
618 return true;
619}
620
621static int tick_nohz_cpu_down(unsigned int cpu)
622{
623 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
624}
625
626void __init tick_nohz_init(void)
627{
628 int cpu, ret;
629
630 if (!tick_nohz_full_running)
631 return;
632
633 /*
634 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
635 * locking contexts. But then we need IRQ work to raise its own
636 * interrupts to avoid circular dependency on the tick.
637 */
638 if (!arch_irq_work_has_interrupt()) {
639 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
640 cpumask_clear(tick_nohz_full_mask);
641 tick_nohz_full_running = false;
642 return;
643 }
644
645 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
646 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
647 cpu = smp_processor_id();
648
649 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
650 pr_warn("NO_HZ: Clearing %d from nohz_full range "
651 "for timekeeping\n", cpu);
652 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
653 }
654 }
655
656 for_each_cpu(cpu, tick_nohz_full_mask)
657 ct_cpu_track_user(cpu);
658
659 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
660 "kernel/nohz:predown", NULL,
661 tick_nohz_cpu_down);
662 WARN_ON(ret < 0);
663 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
664 cpumask_pr_args(tick_nohz_full_mask));
665}
666#endif /* #ifdef CONFIG_NO_HZ_FULL */
667
668/*
669 * NOHZ - aka dynamic tick functionality
670 */
671#ifdef CONFIG_NO_HZ_COMMON
672/*
673 * NO HZ enabled ?
674 */
675bool tick_nohz_enabled __read_mostly = true;
676unsigned long tick_nohz_active __read_mostly;
677/*
678 * Enable / Disable tickless mode
679 */
680static int __init setup_tick_nohz(char *str)
681{
682 return (kstrtobool(str, &tick_nohz_enabled) == 0);
683}
684
685__setup("nohz=", setup_tick_nohz);
686
687bool tick_nohz_tick_stopped(void)
688{
689 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
690
691 return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
692}
693
694bool tick_nohz_tick_stopped_cpu(int cpu)
695{
696 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
697
698 return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
699}
700
701/**
702 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
703 * @now: current ktime_t
704 *
705 * Called from interrupt entry when the CPU was idle
706 *
707 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
708 * must be updated. Otherwise an interrupt handler could use a stale jiffy
709 * value. We do this unconditionally on any CPU, as we don't know whether the
710 * CPU, which has the update task assigned, is in a long sleep.
711 */
712static void tick_nohz_update_jiffies(ktime_t now)
713{
714 unsigned long flags;
715
716 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
717
718 local_irq_save(flags);
719 tick_do_update_jiffies64(now);
720 local_irq_restore(flags);
721
722 touch_softlockup_watchdog_sched();
723}
724
725static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
726{
727 ktime_t delta;
728
729 if (WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE)))
730 return;
731
732 delta = ktime_sub(now, ts->idle_entrytime);
733
734 write_seqcount_begin(&ts->idle_sleeptime_seq);
735 if (nr_iowait_cpu(smp_processor_id()) > 0)
736 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
737 else
738 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
739
740 ts->idle_entrytime = now;
741 tick_sched_flag_clear(ts, TS_FLAG_IDLE_ACTIVE);
742 write_seqcount_end(&ts->idle_sleeptime_seq);
743
744 sched_clock_idle_wakeup_event();
745}
746
747static void tick_nohz_start_idle(struct tick_sched *ts)
748{
749 write_seqcount_begin(&ts->idle_sleeptime_seq);
750 ts->idle_entrytime = ktime_get();
751 tick_sched_flag_set(ts, TS_FLAG_IDLE_ACTIVE);
752 write_seqcount_end(&ts->idle_sleeptime_seq);
753
754 sched_clock_idle_sleep_event();
755}
756
757static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
758 bool compute_delta, u64 *last_update_time)
759{
760 ktime_t now, idle;
761 unsigned int seq;
762
763 if (!tick_nohz_active)
764 return -1;
765
766 now = ktime_get();
767 if (last_update_time)
768 *last_update_time = ktime_to_us(now);
769
770 do {
771 seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
772
773 if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE) && compute_delta) {
774 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
775
776 idle = ktime_add(*sleeptime, delta);
777 } else {
778 idle = *sleeptime;
779 }
780 } while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
781
782 return ktime_to_us(idle);
783
784}
785
786/**
787 * get_cpu_idle_time_us - get the total idle time of a CPU
788 * @cpu: CPU number to query
789 * @last_update_time: variable to store update time in. Do not update
790 * counters if NULL.
791 *
792 * Return the cumulative idle time (since boot) for a given
793 * CPU, in microseconds. Note that this is partially broken due to
794 * the counter of iowait tasks that can be remotely updated without
795 * any synchronization. Therefore it is possible to observe backward
796 * values within two consecutive reads.
797 *
798 * This time is measured via accounting rather than sampling,
799 * and is as accurate as ktime_get() is.
800 *
801 * Return: -1 if NOHZ is not enabled, else total idle time of the @cpu
802 */
803u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
804{
805 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
806
807 return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
808 !nr_iowait_cpu(cpu), last_update_time);
809}
810EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
811
812/**
813 * get_cpu_iowait_time_us - get the total iowait time of a CPU
814 * @cpu: CPU number to query
815 * @last_update_time: variable to store update time in. Do not update
816 * counters if NULL.
817 *
818 * Return the cumulative iowait time (since boot) for a given
819 * CPU, in microseconds. Note this is partially broken due to
820 * the counter of iowait tasks that can be remotely updated without
821 * any synchronization. Therefore it is possible to observe backward
822 * values within two consecutive reads.
823 *
824 * This time is measured via accounting rather than sampling,
825 * and is as accurate as ktime_get() is.
826 *
827 * Return: -1 if NOHZ is not enabled, else total iowait time of @cpu
828 */
829u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
830{
831 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
832
833 return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
834 nr_iowait_cpu(cpu), last_update_time);
835}
836EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
837
838static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
839{
840 hrtimer_cancel(&ts->sched_timer);
841 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
842
843 /* Forward the time to expire in the future */
844 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
845
846 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
847 hrtimer_start_expires(&ts->sched_timer,
848 HRTIMER_MODE_ABS_PINNED_HARD);
849 } else {
850 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
851 }
852
853 /*
854 * Reset to make sure the next tick stop doesn't get fooled by past
855 * cached clock deadline.
856 */
857 ts->next_tick = 0;
858}
859
860static inline bool local_timer_softirq_pending(void)
861{
862 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
863}
864
865/*
866 * Read jiffies and the time when jiffies were updated last
867 */
868u64 get_jiffies_update(unsigned long *basej)
869{
870 unsigned long basejiff;
871 unsigned int seq;
872 u64 basemono;
873
874 do {
875 seq = read_seqcount_begin(&jiffies_seq);
876 basemono = last_jiffies_update;
877 basejiff = jiffies;
878 } while (read_seqcount_retry(&jiffies_seq, seq));
879 *basej = basejiff;
880 return basemono;
881}
882
883/**
884 * tick_nohz_next_event() - return the clock monotonic based next event
885 * @ts: pointer to tick_sched struct
886 * @cpu: CPU number
887 *
888 * Return:
889 * *%0 - When the next event is a maximum of TICK_NSEC in the future
890 * and the tick is not stopped yet
891 * *%next_event - Next event based on clock monotonic
892 */
893static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
894{
895 u64 basemono, next_tick, delta, expires;
896 unsigned long basejiff;
897 int tick_cpu;
898
899 basemono = get_jiffies_update(&basejiff);
900 ts->last_jiffies = basejiff;
901 ts->timer_expires_base = basemono;
902
903 /*
904 * Keep the periodic tick, when RCU, architecture or irq_work
905 * requests it.
906 * Aside of that, check whether the local timer softirq is
907 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
908 * because there is an already expired timer, so it will request
909 * immediate expiry, which rearms the hardware timer with a
910 * minimal delta, which brings us back to this place
911 * immediately. Lather, rinse and repeat...
912 */
913 if (rcu_needs_cpu() || arch_needs_cpu() ||
914 irq_work_needs_cpu() || local_timer_softirq_pending()) {
915 next_tick = basemono + TICK_NSEC;
916 } else {
917 /*
918 * Get the next pending timer. If high resolution
919 * timers are enabled this only takes the timer wheel
920 * timers into account. If high resolution timers are
921 * disabled this also looks at the next expiring
922 * hrtimer.
923 */
924 next_tick = get_next_timer_interrupt(basejiff, basemono);
925 ts->next_timer = next_tick;
926 }
927
928 /* Make sure next_tick is never before basemono! */
929 if (WARN_ON_ONCE(basemono > next_tick))
930 next_tick = basemono;
931
932 /*
933 * If the tick is due in the next period, keep it ticking or
934 * force prod the timer.
935 */
936 delta = next_tick - basemono;
937 if (delta <= (u64)TICK_NSEC) {
938 /*
939 * We've not stopped the tick yet, and there's a timer in the
940 * next period, so no point in stopping it either, bail.
941 */
942 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
943 ts->timer_expires = 0;
944 goto out;
945 }
946 }
947
948 /*
949 * If this CPU is the one which had the do_timer() duty last, we limit
950 * the sleep time to the timekeeping 'max_deferment' value.
951 * Otherwise we can sleep as long as we want.
952 */
953 delta = timekeeping_max_deferment();
954 tick_cpu = READ_ONCE(tick_do_timer_cpu);
955 if (tick_cpu != cpu &&
956 (tick_cpu != TICK_DO_TIMER_NONE || !tick_sched_flag_test(ts, TS_FLAG_DO_TIMER_LAST)))
957 delta = KTIME_MAX;
958
959 /* Calculate the next expiry time */
960 if (delta < (KTIME_MAX - basemono))
961 expires = basemono + delta;
962 else
963 expires = KTIME_MAX;
964
965 ts->timer_expires = min_t(u64, expires, next_tick);
966
967out:
968 return ts->timer_expires;
969}
970
971static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
972{
973 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
974 unsigned long basejiff = ts->last_jiffies;
975 u64 basemono = ts->timer_expires_base;
976 bool timer_idle = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
977 int tick_cpu;
978 u64 expires;
979
980 /* Make sure we won't be trying to stop it twice in a row. */
981 ts->timer_expires_base = 0;
982
983 /*
984 * Now the tick should be stopped definitely - so the timer base needs
985 * to be marked idle as well to not miss a newly queued timer.
986 */
987 expires = timer_base_try_to_set_idle(basejiff, basemono, &timer_idle);
988 if (expires > ts->timer_expires) {
989 /*
990 * This path could only happen when the first timer was removed
991 * between calculating the possible sleep length and now (when
992 * high resolution mode is not active, timer could also be a
993 * hrtimer).
994 *
995 * We have to stick to the original calculated expiry value to
996 * not stop the tick for too long with a shallow C-state (which
997 * was programmed by cpuidle because of an early next expiration
998 * value).
999 */
1000 expires = ts->timer_expires;
1001 }
1002
1003 /* If the timer base is not idle, retain the not yet stopped tick. */
1004 if (!timer_idle)
1005 return;
1006
1007 /*
1008 * If this CPU is the one which updates jiffies, then give up
1009 * the assignment and let it be taken by the CPU which runs
1010 * the tick timer next, which might be this CPU as well. If we
1011 * don't drop this here, the jiffies might be stale and
1012 * do_timer() never gets invoked. Keep track of the fact that it
1013 * was the one which had the do_timer() duty last.
1014 */
1015 tick_cpu = READ_ONCE(tick_do_timer_cpu);
1016 if (tick_cpu == cpu) {
1017 WRITE_ONCE(tick_do_timer_cpu, TICK_DO_TIMER_NONE);
1018 tick_sched_flag_set(ts, TS_FLAG_DO_TIMER_LAST);
1019 } else if (tick_cpu != TICK_DO_TIMER_NONE) {
1020 tick_sched_flag_clear(ts, TS_FLAG_DO_TIMER_LAST);
1021 }
1022
1023 /* Skip reprogram of event if it's not changed */
1024 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED) && (expires == ts->next_tick)) {
1025 /* Sanity check: make sure clockevent is actually programmed */
1026 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
1027 return;
1028
1029 WARN_ON_ONCE(1);
1030 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
1031 basemono, ts->next_tick, dev->next_event,
1032 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
1033 }
1034
1035 /*
1036 * tick_nohz_stop_tick() can be called several times before
1037 * tick_nohz_restart_sched_tick() is called. This happens when
1038 * interrupts arrive which do not cause a reschedule. In the first
1039 * call we save the current tick time, so we can restart the
1040 * scheduler tick in tick_nohz_restart_sched_tick().
1041 */
1042 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1043 calc_load_nohz_start();
1044 quiet_vmstat();
1045
1046 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
1047 tick_sched_flag_set(ts, TS_FLAG_STOPPED);
1048 trace_tick_stop(1, TICK_DEP_MASK_NONE);
1049 }
1050
1051 ts->next_tick = expires;
1052
1053 /*
1054 * If the expiration time == KTIME_MAX, then we simply stop
1055 * the tick timer.
1056 */
1057 if (unlikely(expires == KTIME_MAX)) {
1058 tick_sched_timer_cancel(ts);
1059 return;
1060 }
1061
1062 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
1063 hrtimer_start(&ts->sched_timer, expires,
1064 HRTIMER_MODE_ABS_PINNED_HARD);
1065 } else {
1066 hrtimer_set_expires(&ts->sched_timer, expires);
1067 tick_program_event(expires, 1);
1068 }
1069}
1070
1071static void tick_nohz_retain_tick(struct tick_sched *ts)
1072{
1073 ts->timer_expires_base = 0;
1074}
1075
1076#ifdef CONFIG_NO_HZ_FULL
1077static void tick_nohz_full_stop_tick(struct tick_sched *ts, int cpu)
1078{
1079 if (tick_nohz_next_event(ts, cpu))
1080 tick_nohz_stop_tick(ts, cpu);
1081 else
1082 tick_nohz_retain_tick(ts);
1083}
1084#endif /* CONFIG_NO_HZ_FULL */
1085
1086static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
1087{
1088 /* Update jiffies first */
1089 tick_do_update_jiffies64(now);
1090
1091 /*
1092 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
1093 * the clock forward checks in the enqueue path:
1094 */
1095 timer_clear_idle();
1096
1097 calc_load_nohz_stop();
1098 touch_softlockup_watchdog_sched();
1099
1100 /* Cancel the scheduled timer and restore the tick: */
1101 tick_sched_flag_clear(ts, TS_FLAG_STOPPED);
1102 tick_nohz_restart(ts, now);
1103}
1104
1105static void __tick_nohz_full_update_tick(struct tick_sched *ts,
1106 ktime_t now)
1107{
1108#ifdef CONFIG_NO_HZ_FULL
1109 int cpu = smp_processor_id();
1110
1111 if (can_stop_full_tick(cpu, ts))
1112 tick_nohz_full_stop_tick(ts, cpu);
1113 else if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1114 tick_nohz_restart_sched_tick(ts, now);
1115#endif
1116}
1117
1118static void tick_nohz_full_update_tick(struct tick_sched *ts)
1119{
1120 if (!tick_nohz_full_cpu(smp_processor_id()))
1121 return;
1122
1123 if (!tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1124 return;
1125
1126 __tick_nohz_full_update_tick(ts, ktime_get());
1127}
1128
1129/*
1130 * A pending softirq outside an IRQ (or softirq disabled section) context
1131 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1132 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1133 *
1134 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1135 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1136 * triggering the code below, since wakep_softirqd() is ignored.
1137 *
1138 */
1139static bool report_idle_softirq(void)
1140{
1141 static int ratelimit;
1142 unsigned int pending = local_softirq_pending();
1143
1144 if (likely(!pending))
1145 return false;
1146
1147 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1148 if (!cpu_active(smp_processor_id())) {
1149 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1150 if (!pending)
1151 return false;
1152 }
1153
1154 if (ratelimit >= 10)
1155 return false;
1156
1157 /* On RT, softirq handling may be waiting on some lock */
1158 if (local_bh_blocked())
1159 return false;
1160
1161 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1162 pending);
1163 ratelimit++;
1164
1165 return true;
1166}
1167
1168static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1169{
1170 WARN_ON_ONCE(cpu_is_offline(cpu));
1171
1172 if (unlikely(!tick_sched_flag_test(ts, TS_FLAG_NOHZ)))
1173 return false;
1174
1175 if (need_resched())
1176 return false;
1177
1178 if (unlikely(report_idle_softirq()))
1179 return false;
1180
1181 if (tick_nohz_full_enabled()) {
1182 int tick_cpu = READ_ONCE(tick_do_timer_cpu);
1183
1184 /*
1185 * Keep the tick alive to guarantee timekeeping progression
1186 * if there are full dynticks CPUs around
1187 */
1188 if (tick_cpu == cpu)
1189 return false;
1190
1191 /* Should not happen for nohz-full */
1192 if (WARN_ON_ONCE(tick_cpu == TICK_DO_TIMER_NONE))
1193 return false;
1194 }
1195
1196 return true;
1197}
1198
1199/**
1200 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1201 *
1202 * When the next event is more than a tick into the future, stop the idle tick
1203 */
1204void tick_nohz_idle_stop_tick(void)
1205{
1206 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1207 int cpu = smp_processor_id();
1208 ktime_t expires;
1209
1210 /*
1211 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1212 * tick timer expiration time is known already.
1213 */
1214 if (ts->timer_expires_base)
1215 expires = ts->timer_expires;
1216 else if (can_stop_idle_tick(cpu, ts))
1217 expires = tick_nohz_next_event(ts, cpu);
1218 else
1219 return;
1220
1221 ts->idle_calls++;
1222
1223 if (expires > 0LL) {
1224 int was_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
1225
1226 tick_nohz_stop_tick(ts, cpu);
1227
1228 ts->idle_sleeps++;
1229 ts->idle_expires = expires;
1230
1231 if (!was_stopped && tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1232 ts->idle_jiffies = ts->last_jiffies;
1233 nohz_balance_enter_idle(cpu);
1234 }
1235 } else {
1236 tick_nohz_retain_tick(ts);
1237 }
1238}
1239
1240void tick_nohz_idle_retain_tick(void)
1241{
1242 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1243}
1244
1245/**
1246 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1247 *
1248 * Called when we start the idle loop.
1249 */
1250void tick_nohz_idle_enter(void)
1251{
1252 struct tick_sched *ts;
1253
1254 lockdep_assert_irqs_enabled();
1255
1256 local_irq_disable();
1257
1258 ts = this_cpu_ptr(&tick_cpu_sched);
1259
1260 WARN_ON_ONCE(ts->timer_expires_base);
1261
1262 tick_sched_flag_set(ts, TS_FLAG_INIDLE);
1263 tick_nohz_start_idle(ts);
1264
1265 local_irq_enable();
1266}
1267
1268/**
1269 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1270 *
1271 * A timer may have been added/modified/deleted either by the current IRQ,
1272 * or by another place using this IRQ as a notification. This IRQ may have
1273 * also updated the RCU callback list. These events may require a
1274 * re-evaluation of the next tick. Depending on the context:
1275 *
1276 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1277 * time accounting. The next tick will be re-evaluated on the next idle
1278 * loop iteration.
1279 *
1280 * 2) If the CPU is nohz_full:
1281 *
1282 * 2.1) If there is any tick dependency, restart the tick if stopped.
1283 *
1284 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1285 * stop/update it accordingly.
1286 */
1287void tick_nohz_irq_exit(void)
1288{
1289 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1290
1291 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE))
1292 tick_nohz_start_idle(ts);
1293 else
1294 tick_nohz_full_update_tick(ts);
1295}
1296
1297/**
1298 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1299 *
1300 * Return: %true if the tick handler has run, otherwise %false
1301 */
1302bool tick_nohz_idle_got_tick(void)
1303{
1304 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1305
1306 if (ts->got_idle_tick) {
1307 ts->got_idle_tick = 0;
1308 return true;
1309 }
1310 return false;
1311}
1312
1313/**
1314 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1315 * or the tick, whichever expires first. Note that, if the tick has been
1316 * stopped, it returns the next hrtimer.
1317 *
1318 * Called from power state control code with interrupts disabled
1319 *
1320 * Return: the next expiration time
1321 */
1322ktime_t tick_nohz_get_next_hrtimer(void)
1323{
1324 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1325}
1326
1327/**
1328 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1329 * @delta_next: duration until the next event if the tick cannot be stopped
1330 *
1331 * Called from power state control code with interrupts disabled.
1332 *
1333 * The return value of this function and/or the value returned by it through the
1334 * @delta_next pointer can be negative which must be taken into account by its
1335 * callers.
1336 *
1337 * Return: the expected length of the current sleep
1338 */
1339ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1340{
1341 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1342 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1343 int cpu = smp_processor_id();
1344 /*
1345 * The idle entry time is expected to be a sufficient approximation of
1346 * the current time at this point.
1347 */
1348 ktime_t now = ts->idle_entrytime;
1349 ktime_t next_event;
1350
1351 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1352
1353 *delta_next = ktime_sub(dev->next_event, now);
1354
1355 if (!can_stop_idle_tick(cpu, ts))
1356 return *delta_next;
1357
1358 next_event = tick_nohz_next_event(ts, cpu);
1359 if (!next_event)
1360 return *delta_next;
1361
1362 /*
1363 * If the next highres timer to expire is earlier than 'next_event', the
1364 * idle governor needs to know that.
1365 */
1366 next_event = min_t(u64, next_event,
1367 hrtimer_next_event_without(&ts->sched_timer));
1368
1369 return ktime_sub(next_event, now);
1370}
1371
1372/**
1373 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1374 * for a particular CPU.
1375 * @cpu: target CPU number
1376 *
1377 * Called from the schedutil frequency scaling governor in scheduler context.
1378 *
1379 * Return: the current idle calls counter value for @cpu
1380 */
1381unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1382{
1383 struct tick_sched *ts = tick_get_tick_sched(cpu);
1384
1385 return ts->idle_calls;
1386}
1387
1388/**
1389 * tick_nohz_get_idle_calls - return the current idle calls counter value
1390 *
1391 * Called from the schedutil frequency scaling governor in scheduler context.
1392 *
1393 * Return: the current idle calls counter value for the current CPU
1394 */
1395unsigned long tick_nohz_get_idle_calls(void)
1396{
1397 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1398
1399 return ts->idle_calls;
1400}
1401
1402static void tick_nohz_account_idle_time(struct tick_sched *ts,
1403 ktime_t now)
1404{
1405 unsigned long ticks;
1406
1407 ts->idle_exittime = now;
1408
1409 if (vtime_accounting_enabled_this_cpu())
1410 return;
1411 /*
1412 * We stopped the tick in idle. update_process_times() would miss the
1413 * time we slept, as it does only a 1 tick accounting.
1414 * Enforce that this is accounted to idle !
1415 */
1416 ticks = jiffies - ts->idle_jiffies;
1417 /*
1418 * We might be one off. Do not randomly account a huge number of ticks!
1419 */
1420 if (ticks && ticks < LONG_MAX)
1421 account_idle_ticks(ticks);
1422}
1423
1424void tick_nohz_idle_restart_tick(void)
1425{
1426 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1427
1428 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1429 ktime_t now = ktime_get();
1430 tick_nohz_restart_sched_tick(ts, now);
1431 tick_nohz_account_idle_time(ts, now);
1432 }
1433}
1434
1435static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1436{
1437 if (tick_nohz_full_cpu(smp_processor_id()))
1438 __tick_nohz_full_update_tick(ts, now);
1439 else
1440 tick_nohz_restart_sched_tick(ts, now);
1441
1442 tick_nohz_account_idle_time(ts, now);
1443}
1444
1445/**
1446 * tick_nohz_idle_exit - Update the tick upon idle task exit
1447 *
1448 * When the idle task exits, update the tick depending on the
1449 * following situations:
1450 *
1451 * 1) If the CPU is not in nohz_full mode (most cases), then
1452 * restart the tick.
1453 *
1454 * 2) If the CPU is in nohz_full mode (corner case):
1455 * 2.1) If the tick can be kept stopped (no tick dependencies)
1456 * then re-evaluate the next tick and try to keep it stopped
1457 * as long as possible.
1458 * 2.2) If the tick has dependencies, restart the tick.
1459 *
1460 */
1461void tick_nohz_idle_exit(void)
1462{
1463 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1464 bool idle_active, tick_stopped;
1465 ktime_t now;
1466
1467 local_irq_disable();
1468
1469 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1470 WARN_ON_ONCE(ts->timer_expires_base);
1471
1472 tick_sched_flag_clear(ts, TS_FLAG_INIDLE);
1473 idle_active = tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE);
1474 tick_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
1475
1476 if (idle_active || tick_stopped)
1477 now = ktime_get();
1478
1479 if (idle_active)
1480 tick_nohz_stop_idle(ts, now);
1481
1482 if (tick_stopped)
1483 tick_nohz_idle_update_tick(ts, now);
1484
1485 local_irq_enable();
1486}
1487
1488/*
1489 * In low-resolution mode, the tick handler must be implemented directly
1490 * at the clockevent level. hrtimer can't be used instead, because its
1491 * infrastructure actually relies on the tick itself as a backend in
1492 * low-resolution mode (see hrtimer_run_queues()).
1493 */
1494static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1495{
1496 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1497
1498 dev->next_event = KTIME_MAX;
1499
1500 if (likely(tick_nohz_handler(&ts->sched_timer) == HRTIMER_RESTART))
1501 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1502}
1503
1504static inline void tick_nohz_activate(struct tick_sched *ts)
1505{
1506 if (!tick_nohz_enabled)
1507 return;
1508 tick_sched_flag_set(ts, TS_FLAG_NOHZ);
1509 /* One update is enough */
1510 if (!test_and_set_bit(0, &tick_nohz_active))
1511 timers_update_nohz();
1512}
1513
1514/**
1515 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1516 */
1517static void tick_nohz_switch_to_nohz(void)
1518{
1519 if (!tick_nohz_enabled)
1520 return;
1521
1522 if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1523 return;
1524
1525 /*
1526 * Recycle the hrtimer in 'ts', so we can share the
1527 * highres code.
1528 */
1529 tick_setup_sched_timer(false);
1530}
1531
1532static inline void tick_nohz_irq_enter(void)
1533{
1534 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1535 ktime_t now;
1536
1537 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED | TS_FLAG_IDLE_ACTIVE))
1538 return;
1539 now = ktime_get();
1540 if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE))
1541 tick_nohz_stop_idle(ts, now);
1542 /*
1543 * If all CPUs are idle we may need to update a stale jiffies value.
1544 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1545 * alive but it might be busy looping with interrupts disabled in some
1546 * rare case (typically stop machine). So we must make sure we have a
1547 * last resort.
1548 */
1549 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1550 tick_nohz_update_jiffies(now);
1551}
1552
1553#else
1554
1555static inline void tick_nohz_switch_to_nohz(void) { }
1556static inline void tick_nohz_irq_enter(void) { }
1557static inline void tick_nohz_activate(struct tick_sched *ts) { }
1558
1559#endif /* CONFIG_NO_HZ_COMMON */
1560
1561/*
1562 * Called from irq_enter() to notify about the possible interruption of idle()
1563 */
1564void tick_irq_enter(void)
1565{
1566 tick_check_oneshot_broadcast_this_cpu();
1567 tick_nohz_irq_enter();
1568}
1569
1570static int sched_skew_tick;
1571
1572static int __init skew_tick(char *str)
1573{
1574 get_option(&str, &sched_skew_tick);
1575
1576 return 0;
1577}
1578early_param("skew_tick", skew_tick);
1579
1580/**
1581 * tick_setup_sched_timer - setup the tick emulation timer
1582 * @hrtimer: whether to use the hrtimer or not
1583 */
1584void tick_setup_sched_timer(bool hrtimer)
1585{
1586 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1587
1588 /* Emulate tick processing via per-CPU hrtimers: */
1589 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1590
1591 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer) {
1592 tick_sched_flag_set(ts, TS_FLAG_HIGHRES);
1593 ts->sched_timer.function = tick_nohz_handler;
1594 }
1595
1596 /* Get the next period (per-CPU) */
1597 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1598
1599 /* Offset the tick to avert 'jiffies_lock' contention. */
1600 if (sched_skew_tick) {
1601 u64 offset = TICK_NSEC >> 1;
1602 do_div(offset, num_possible_cpus());
1603 offset *= smp_processor_id();
1604 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1605 }
1606
1607 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1608 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer)
1609 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1610 else
1611 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1612 tick_nohz_activate(ts);
1613}
1614
1615/*
1616 * Shut down the tick and make sure the CPU won't try to retake the timekeeping
1617 * duty before disabling IRQs in idle for the last time.
1618 */
1619void tick_sched_timer_dying(int cpu)
1620{
1621 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
1622 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1623 struct clock_event_device *dev = td->evtdev;
1624 ktime_t idle_sleeptime, iowait_sleeptime;
1625 unsigned long idle_calls, idle_sleeps;
1626
1627 /* This must happen before hrtimers are migrated! */
1628 tick_sched_timer_cancel(ts);
1629
1630 /*
1631 * If the clockevents doesn't support CLOCK_EVT_STATE_ONESHOT_STOPPED,
1632 * make sure not to call low-res tick handler.
1633 */
1634 if (tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1635 dev->event_handler = clockevents_handle_noop;
1636
1637 idle_sleeptime = ts->idle_sleeptime;
1638 iowait_sleeptime = ts->iowait_sleeptime;
1639 idle_calls = ts->idle_calls;
1640 idle_sleeps = ts->idle_sleeps;
1641 memset(ts, 0, sizeof(*ts));
1642 ts->idle_sleeptime = idle_sleeptime;
1643 ts->iowait_sleeptime = iowait_sleeptime;
1644 ts->idle_calls = idle_calls;
1645 ts->idle_sleeps = idle_sleeps;
1646}
1647
1648/*
1649 * Async notification about clocksource changes
1650 */
1651void tick_clock_notify(void)
1652{
1653 int cpu;
1654
1655 for_each_possible_cpu(cpu)
1656 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1657}
1658
1659/*
1660 * Async notification about clock event changes
1661 */
1662void tick_oneshot_notify(void)
1663{
1664 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1665
1666 set_bit(0, &ts->check_clocks);
1667}
1668
1669/*
1670 * Check if a change happened, which makes oneshot possible.
1671 *
1672 * Called cyclically from the hrtimer softirq (driven by the timer
1673 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1674 * mode, because high resolution timers are disabled (either compile
1675 * or runtime). Called with interrupts disabled.
1676 */
1677int tick_check_oneshot_change(int allow_nohz)
1678{
1679 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1680
1681 if (!test_and_clear_bit(0, &ts->check_clocks))
1682 return 0;
1683
1684 if (tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1685 return 0;
1686
1687 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1688 return 0;
1689
1690 if (!allow_nohz)
1691 return 1;
1692
1693 tick_nohz_switch_to_nohz();
1694 return 0;
1695}