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