Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * NOHZ implementation for low and high resolution timers
8 *
9 * Started by: Thomas Gleixner and Ingo Molnar
10 */
11#include <linux/cpu.h>
12#include <linux/err.h>
13#include <linux/hrtimer.h>
14#include <linux/interrupt.h>
15#include <linux/kernel_stat.h>
16#include <linux/percpu.h>
17#include <linux/nmi.h>
18#include <linux/profile.h>
19#include <linux/sched/signal.h>
20#include <linux/sched/clock.h>
21#include <linux/sched/stat.h>
22#include <linux/sched/nohz.h>
23#include <linux/sched/loadavg.h>
24#include <linux/module.h>
25#include <linux/irq_work.h>
26#include <linux/posix-timers.h>
27#include <linux/context_tracking.h>
28#include <linux/mm.h>
29
30#include <asm/irq_regs.h>
31
32#include "tick-internal.h"
33
34#include <trace/events/timer.h>
35
36/*
37 * Per-CPU nohz control structure
38 */
39static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
40
41struct tick_sched *tick_get_tick_sched(int cpu)
42{
43 return &per_cpu(tick_cpu_sched, cpu);
44}
45
46#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
47/*
48 * The time when the last jiffy update happened. Write access must hold
49 * jiffies_lock and jiffies_seq. tick_nohz_next_event() needs to get a
50 * consistent view of jiffies and last_jiffies_update.
51 */
52static ktime_t last_jiffies_update;
53
54/*
55 * Must be called with interrupts disabled !
56 */
57static void tick_do_update_jiffies64(ktime_t now)
58{
59 unsigned long ticks = 1;
60 ktime_t delta, nextp;
61
62 /*
63 * 64-bit can do a quick check without holding the jiffies lock and
64 * without looking at the sequence count. The smp_load_acquire()
65 * pairs with the update done later in this function.
66 *
67 * 32-bit cannot do that because the store of 'tick_next_period'
68 * consists of two 32-bit stores, and the first store could be
69 * moved by the CPU to a random point in the future.
70 */
71 if (IS_ENABLED(CONFIG_64BIT)) {
72 if (ktime_before(now, smp_load_acquire(&tick_next_period)))
73 return;
74 } else {
75 unsigned int seq;
76
77 /*
78 * Avoid contention on 'jiffies_lock' and protect the quick
79 * check with the sequence count.
80 */
81 do {
82 seq = read_seqcount_begin(&jiffies_seq);
83 nextp = tick_next_period;
84 } while (read_seqcount_retry(&jiffies_seq, seq));
85
86 if (ktime_before(now, nextp))
87 return;
88 }
89
90 /* Quick check failed, i.e. update is required. */
91 raw_spin_lock(&jiffies_lock);
92 /*
93 * Re-evaluate with the lock held. Another CPU might have done the
94 * update already.
95 */
96 if (ktime_before(now, tick_next_period)) {
97 raw_spin_unlock(&jiffies_lock);
98 return;
99 }
100
101 write_seqcount_begin(&jiffies_seq);
102
103 delta = ktime_sub(now, tick_next_period);
104 if (unlikely(delta >= TICK_NSEC)) {
105 /* Slow path for long idle sleep times */
106 s64 incr = TICK_NSEC;
107
108 ticks += ktime_divns(delta, incr);
109
110 last_jiffies_update = ktime_add_ns(last_jiffies_update,
111 incr * ticks);
112 } else {
113 last_jiffies_update = ktime_add_ns(last_jiffies_update,
114 TICK_NSEC);
115 }
116
117 /* Advance jiffies to complete the 'jiffies_seq' protected job */
118 jiffies_64 += ticks;
119
120 /* Keep the tick_next_period variable up to date */
121 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
122
123 if (IS_ENABLED(CONFIG_64BIT)) {
124 /*
125 * Pairs with smp_load_acquire() in the lockless quick
126 * check above, and ensures that the update to 'jiffies_64' is
127 * not reordered vs. the store to 'tick_next_period', neither
128 * by the compiler nor by the CPU.
129 */
130 smp_store_release(&tick_next_period, nextp);
131 } else {
132 /*
133 * A plain store is good enough on 32-bit, as the quick check
134 * above is protected by the sequence count.
135 */
136 tick_next_period = nextp;
137 }
138
139 /*
140 * Release the sequence count. calc_global_load() below is not
141 * protected by it, but 'jiffies_lock' needs to be held to prevent
142 * concurrent invocations.
143 */
144 write_seqcount_end(&jiffies_seq);
145
146 calc_global_load();
147
148 raw_spin_unlock(&jiffies_lock);
149 update_wall_time();
150}
151
152/*
153 * Initialize and return retrieve the jiffies update.
154 */
155static ktime_t tick_init_jiffy_update(void)
156{
157 ktime_t period;
158
159 raw_spin_lock(&jiffies_lock);
160 write_seqcount_begin(&jiffies_seq);
161
162 /* Have we started the jiffies update yet ? */
163 if (last_jiffies_update == 0) {
164 u32 rem;
165
166 /*
167 * Ensure that the tick is aligned to a multiple of
168 * TICK_NSEC.
169 */
170 div_u64_rem(tick_next_period, TICK_NSEC, &rem);
171 if (rem)
172 tick_next_period += TICK_NSEC - rem;
173
174 last_jiffies_update = tick_next_period;
175 }
176 period = last_jiffies_update;
177
178 write_seqcount_end(&jiffies_seq);
179 raw_spin_unlock(&jiffies_lock);
180
181 return period;
182}
183
184#define MAX_STALLED_JIFFIES 5
185
186static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
187{
188 int cpu = smp_processor_id();
189
190#ifdef CONFIG_NO_HZ_COMMON
191 /*
192 * Check if the do_timer duty was dropped. We don't care about
193 * concurrency: This happens only when the CPU in charge went
194 * into a long sleep. If two CPUs happen to assign themselves to
195 * this duty, then the jiffies update is still serialized by
196 * 'jiffies_lock'.
197 *
198 * If nohz_full is enabled, this should not happen because the
199 * 'tick_do_timer_cpu' CPU never relinquishes.
200 */
201 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) {
202#ifdef CONFIG_NO_HZ_FULL
203 WARN_ON_ONCE(tick_nohz_full_running);
204#endif
205 tick_do_timer_cpu = cpu;
206 }
207#endif
208
209 /* Check if jiffies need an update */
210 if (tick_do_timer_cpu == cpu)
211 tick_do_update_jiffies64(now);
212
213 /*
214 * If the jiffies update stalled for too long (timekeeper in stop_machine()
215 * or VMEXIT'ed for several msecs), force an update.
216 */
217 if (ts->last_tick_jiffies != jiffies) {
218 ts->stalled_jiffies = 0;
219 ts->last_tick_jiffies = READ_ONCE(jiffies);
220 } else {
221 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) {
222 tick_do_update_jiffies64(now);
223 ts->stalled_jiffies = 0;
224 ts->last_tick_jiffies = READ_ONCE(jiffies);
225 }
226 }
227
228 if (ts->inidle)
229 ts->got_idle_tick = 1;
230}
231
232static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
233{
234#ifdef CONFIG_NO_HZ_COMMON
235 /*
236 * When we are idle and the tick is stopped, we have to touch
237 * the watchdog as we might not schedule for a really long
238 * time. This happens on completely idle SMP systems while
239 * waiting on the login prompt. We also increment the "start of
240 * idle" jiffy stamp so the idle accounting adjustment we do
241 * when we go busy again does not account too many ticks.
242 */
243 if (ts->tick_stopped) {
244 touch_softlockup_watchdog_sched();
245 if (is_idle_task(current))
246 ts->idle_jiffies++;
247 /*
248 * In case the current tick fired too early past its expected
249 * expiration, make sure we don't bypass the next clock reprogramming
250 * to the same deadline.
251 */
252 ts->next_tick = 0;
253 }
254#endif
255 update_process_times(user_mode(regs));
256 profile_tick(CPU_PROFILING);
257}
258#endif
259
260#ifdef CONFIG_NO_HZ_FULL
261cpumask_var_t tick_nohz_full_mask;
262EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
263bool tick_nohz_full_running;
264EXPORT_SYMBOL_GPL(tick_nohz_full_running);
265static atomic_t tick_dep_mask;
266
267static bool check_tick_dependency(atomic_t *dep)
268{
269 int val = atomic_read(dep);
270
271 if (val & TICK_DEP_MASK_POSIX_TIMER) {
272 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
273 return true;
274 }
275
276 if (val & TICK_DEP_MASK_PERF_EVENTS) {
277 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
278 return true;
279 }
280
281 if (val & TICK_DEP_MASK_SCHED) {
282 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
283 return true;
284 }
285
286 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
287 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
288 return true;
289 }
290
291 if (val & TICK_DEP_MASK_RCU) {
292 trace_tick_stop(0, TICK_DEP_MASK_RCU);
293 return true;
294 }
295
296 if (val & TICK_DEP_MASK_RCU_EXP) {
297 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
298 return true;
299 }
300
301 return false;
302}
303
304static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
305{
306 lockdep_assert_irqs_disabled();
307
308 if (unlikely(!cpu_online(cpu)))
309 return false;
310
311 if (check_tick_dependency(&tick_dep_mask))
312 return false;
313
314 if (check_tick_dependency(&ts->tick_dep_mask))
315 return false;
316
317 if (check_tick_dependency(¤t->tick_dep_mask))
318 return false;
319
320 if (check_tick_dependency(¤t->signal->tick_dep_mask))
321 return false;
322
323 return true;
324}
325
326static void nohz_full_kick_func(struct irq_work *work)
327{
328 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
329}
330
331static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
332 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
333
334/*
335 * Kick this CPU if it's full dynticks in order to force it to
336 * re-evaluate its dependency on the tick and restart it if necessary.
337 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
338 * is NMI safe.
339 */
340static void tick_nohz_full_kick(void)
341{
342 if (!tick_nohz_full_cpu(smp_processor_id()))
343 return;
344
345 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
346}
347
348/*
349 * Kick the CPU if it's full dynticks in order to force it to
350 * re-evaluate its dependency on the tick and restart it if necessary.
351 */
352void tick_nohz_full_kick_cpu(int cpu)
353{
354 if (!tick_nohz_full_cpu(cpu))
355 return;
356
357 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
358}
359
360static void tick_nohz_kick_task(struct task_struct *tsk)
361{
362 int cpu;
363
364 /*
365 * If the task is not running, run_posix_cpu_timers()
366 * has nothing to elapse, and an IPI can then be optimized out.
367 *
368 * activate_task() STORE p->tick_dep_mask
369 * STORE p->on_rq
370 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
371 * LOCK rq->lock LOAD p->on_rq
372 * smp_mb__after_spin_lock()
373 * tick_nohz_task_switch()
374 * LOAD p->tick_dep_mask
375 */
376 if (!sched_task_on_rq(tsk))
377 return;
378
379 /*
380 * If the task concurrently migrates to another CPU,
381 * we guarantee it sees the new tick dependency upon
382 * schedule.
383 *
384 * set_task_cpu(p, cpu);
385 * STORE p->cpu = @cpu
386 * __schedule() (switch to task 'p')
387 * LOCK rq->lock
388 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
389 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
390 * LOAD p->tick_dep_mask LOAD p->cpu
391 */
392 cpu = task_cpu(tsk);
393
394 preempt_disable();
395 if (cpu_online(cpu))
396 tick_nohz_full_kick_cpu(cpu);
397 preempt_enable();
398}
399
400/*
401 * Kick all full dynticks CPUs in order to force these to re-evaluate
402 * their dependency on the tick and restart it if necessary.
403 */
404static void tick_nohz_full_kick_all(void)
405{
406 int cpu;
407
408 if (!tick_nohz_full_running)
409 return;
410
411 preempt_disable();
412 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
413 tick_nohz_full_kick_cpu(cpu);
414 preempt_enable();
415}
416
417static void tick_nohz_dep_set_all(atomic_t *dep,
418 enum tick_dep_bits bit)
419{
420 int prev;
421
422 prev = atomic_fetch_or(BIT(bit), dep);
423 if (!prev)
424 tick_nohz_full_kick_all();
425}
426
427/*
428 * Set a global tick dependency. Used by perf events that rely on freq and
429 * unstable clocks.
430 */
431void tick_nohz_dep_set(enum tick_dep_bits bit)
432{
433 tick_nohz_dep_set_all(&tick_dep_mask, bit);
434}
435
436void tick_nohz_dep_clear(enum tick_dep_bits bit)
437{
438 atomic_andnot(BIT(bit), &tick_dep_mask);
439}
440
441/*
442 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
443 * manage event-throttling.
444 */
445void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
446{
447 int prev;
448 struct tick_sched *ts;
449
450 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
451
452 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
453 if (!prev) {
454 preempt_disable();
455 /* Perf needs local kick that is NMI safe */
456 if (cpu == smp_processor_id()) {
457 tick_nohz_full_kick();
458 } else {
459 /* Remote IRQ work not NMI-safe */
460 if (!WARN_ON_ONCE(in_nmi()))
461 tick_nohz_full_kick_cpu(cpu);
462 }
463 preempt_enable();
464 }
465}
466EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
467
468void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
469{
470 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
471
472 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
473}
474EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
475
476/*
477 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
478 * in order to elapse per task timers.
479 */
480void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
481{
482 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
483 tick_nohz_kick_task(tsk);
484}
485EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
486
487void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
488{
489 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
490}
491EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
492
493/*
494 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
495 * per process timers.
496 */
497void tick_nohz_dep_set_signal(struct task_struct *tsk,
498 enum tick_dep_bits bit)
499{
500 int prev;
501 struct signal_struct *sig = tsk->signal;
502
503 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
504 if (!prev) {
505 struct task_struct *t;
506
507 lockdep_assert_held(&tsk->sighand->siglock);
508 __for_each_thread(sig, t)
509 tick_nohz_kick_task(t);
510 }
511}
512
513void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
514{
515 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
516}
517
518/*
519 * Re-evaluate the need for the tick as we switch the current task.
520 * It might need the tick due to per task/process properties:
521 * perf events, posix CPU timers, ...
522 */
523void __tick_nohz_task_switch(void)
524{
525 struct tick_sched *ts;
526
527 if (!tick_nohz_full_cpu(smp_processor_id()))
528 return;
529
530 ts = this_cpu_ptr(&tick_cpu_sched);
531
532 if (ts->tick_stopped) {
533 if (atomic_read(¤t->tick_dep_mask) ||
534 atomic_read(¤t->signal->tick_dep_mask))
535 tick_nohz_full_kick();
536 }
537}
538
539/* Get the boot-time nohz CPU list from the kernel parameters. */
540void __init tick_nohz_full_setup(cpumask_var_t cpumask)
541{
542 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
543 cpumask_copy(tick_nohz_full_mask, cpumask);
544 tick_nohz_full_running = true;
545}
546
547bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
548{
549 /*
550 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
551 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
552 * CPUs. It must remain online when nohz full is enabled.
553 */
554 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
555 return false;
556 return true;
557}
558
559static int tick_nohz_cpu_down(unsigned int cpu)
560{
561 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
562}
563
564void __init tick_nohz_init(void)
565{
566 int cpu, ret;
567
568 if (!tick_nohz_full_running)
569 return;
570
571 /*
572 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
573 * locking contexts. But then we need IRQ work to raise its own
574 * interrupts to avoid circular dependency on the tick.
575 */
576 if (!arch_irq_work_has_interrupt()) {
577 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
578 cpumask_clear(tick_nohz_full_mask);
579 tick_nohz_full_running = false;
580 return;
581 }
582
583 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
584 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
585 cpu = smp_processor_id();
586
587 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
588 pr_warn("NO_HZ: Clearing %d from nohz_full range "
589 "for timekeeping\n", cpu);
590 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
591 }
592 }
593
594 for_each_cpu(cpu, tick_nohz_full_mask)
595 ct_cpu_track_user(cpu);
596
597 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
598 "kernel/nohz:predown", NULL,
599 tick_nohz_cpu_down);
600 WARN_ON(ret < 0);
601 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
602 cpumask_pr_args(tick_nohz_full_mask));
603}
604#endif
605
606/*
607 * NOHZ - aka dynamic tick functionality
608 */
609#ifdef CONFIG_NO_HZ_COMMON
610/*
611 * NO HZ enabled ?
612 */
613bool tick_nohz_enabled __read_mostly = true;
614unsigned long tick_nohz_active __read_mostly;
615/*
616 * Enable / Disable tickless mode
617 */
618static int __init setup_tick_nohz(char *str)
619{
620 return (kstrtobool(str, &tick_nohz_enabled) == 0);
621}
622
623__setup("nohz=", setup_tick_nohz);
624
625bool tick_nohz_tick_stopped(void)
626{
627 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
628
629 return ts->tick_stopped;
630}
631
632bool tick_nohz_tick_stopped_cpu(int cpu)
633{
634 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
635
636 return ts->tick_stopped;
637}
638
639/**
640 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
641 *
642 * Called from interrupt entry when the CPU was idle
643 *
644 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
645 * must be updated. Otherwise an interrupt handler could use a stale jiffy
646 * value. We do this unconditionally on any CPU, as we don't know whether the
647 * CPU, which has the update task assigned, is in a long sleep.
648 */
649static void tick_nohz_update_jiffies(ktime_t now)
650{
651 unsigned long flags;
652
653 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
654
655 local_irq_save(flags);
656 tick_do_update_jiffies64(now);
657 local_irq_restore(flags);
658
659 touch_softlockup_watchdog_sched();
660}
661
662static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
663{
664 ktime_t delta;
665
666 if (WARN_ON_ONCE(!ts->idle_active))
667 return;
668
669 delta = ktime_sub(now, ts->idle_entrytime);
670
671 write_seqcount_begin(&ts->idle_sleeptime_seq);
672 if (nr_iowait_cpu(smp_processor_id()) > 0)
673 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
674 else
675 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
676
677 ts->idle_entrytime = now;
678 ts->idle_active = 0;
679 write_seqcount_end(&ts->idle_sleeptime_seq);
680
681 sched_clock_idle_wakeup_event();
682}
683
684static void tick_nohz_start_idle(struct tick_sched *ts)
685{
686 write_seqcount_begin(&ts->idle_sleeptime_seq);
687 ts->idle_entrytime = ktime_get();
688 ts->idle_active = 1;
689 write_seqcount_end(&ts->idle_sleeptime_seq);
690
691 sched_clock_idle_sleep_event();
692}
693
694static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
695 bool compute_delta, u64 *last_update_time)
696{
697 ktime_t now, idle;
698 unsigned int seq;
699
700 if (!tick_nohz_active)
701 return -1;
702
703 now = ktime_get();
704 if (last_update_time)
705 *last_update_time = ktime_to_us(now);
706
707 do {
708 seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
709
710 if (ts->idle_active && compute_delta) {
711 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
712
713 idle = ktime_add(*sleeptime, delta);
714 } else {
715 idle = *sleeptime;
716 }
717 } while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
718
719 return ktime_to_us(idle);
720
721}
722
723/**
724 * get_cpu_idle_time_us - get the total idle time of a CPU
725 * @cpu: CPU number to query
726 * @last_update_time: variable to store update time in. Do not update
727 * counters if NULL.
728 *
729 * Return the cumulative idle time (since boot) for a given
730 * CPU, in microseconds. Note that this is partially broken due to
731 * the counter of iowait tasks that can be remotely updated without
732 * any synchronization. Therefore it is possible to observe backward
733 * values within two consecutive reads.
734 *
735 * This time is measured via accounting rather than sampling,
736 * and is as accurate as ktime_get() is.
737 *
738 * This function returns -1 if NOHZ is not enabled.
739 */
740u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
741{
742 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
743
744 return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
745 !nr_iowait_cpu(cpu), last_update_time);
746}
747EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
748
749/**
750 * get_cpu_iowait_time_us - get the total iowait time of a CPU
751 * @cpu: CPU number to query
752 * @last_update_time: variable to store update time in. Do not update
753 * counters if NULL.
754 *
755 * Return the cumulative iowait time (since boot) for a given
756 * CPU, in microseconds. Note this is partially broken due to
757 * the counter of iowait tasks that can be remotely updated without
758 * any synchronization. Therefore it is possible to observe backward
759 * values within two consecutive reads.
760 *
761 * This time is measured via accounting rather than sampling,
762 * and is as accurate as ktime_get() is.
763 *
764 * This function returns -1 if NOHZ is not enabled.
765 */
766u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
767{
768 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
769
770 return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
771 nr_iowait_cpu(cpu), last_update_time);
772}
773EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
774
775static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
776{
777 hrtimer_cancel(&ts->sched_timer);
778 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
779
780 /* Forward the time to expire in the future */
781 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
782
783 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
784 hrtimer_start_expires(&ts->sched_timer,
785 HRTIMER_MODE_ABS_PINNED_HARD);
786 } else {
787 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
788 }
789
790 /*
791 * Reset to make sure the next tick stop doesn't get fooled by past
792 * cached clock deadline.
793 */
794 ts->next_tick = 0;
795}
796
797static inline bool local_timer_softirq_pending(void)
798{
799 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
800}
801
802static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
803{
804 u64 basemono, next_tick, delta, expires;
805 unsigned long basejiff;
806 unsigned int seq;
807
808 /* Read jiffies and the time when jiffies were updated last */
809 do {
810 seq = read_seqcount_begin(&jiffies_seq);
811 basemono = last_jiffies_update;
812 basejiff = jiffies;
813 } while (read_seqcount_retry(&jiffies_seq, seq));
814 ts->last_jiffies = basejiff;
815 ts->timer_expires_base = basemono;
816
817 /*
818 * Keep the periodic tick, when RCU, architecture or irq_work
819 * requests it.
820 * Aside of that, check whether the local timer softirq is
821 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
822 * because there is an already expired timer, so it will request
823 * immediate expiry, which rearms the hardware timer with a
824 * minimal delta, which brings us back to this place
825 * immediately. Lather, rinse and repeat...
826 */
827 if (rcu_needs_cpu() || arch_needs_cpu() ||
828 irq_work_needs_cpu() || local_timer_softirq_pending()) {
829 next_tick = basemono + TICK_NSEC;
830 } else {
831 /*
832 * Get the next pending timer. If high resolution
833 * timers are enabled this only takes the timer wheel
834 * timers into account. If high resolution timers are
835 * disabled this also looks at the next expiring
836 * hrtimer.
837 */
838 next_tick = get_next_timer_interrupt(basejiff, basemono);
839 ts->next_timer = next_tick;
840 }
841
842 /* Make sure next_tick is never before basemono! */
843 if (WARN_ON_ONCE(basemono > next_tick))
844 next_tick = basemono;
845
846 /*
847 * If the tick is due in the next period, keep it ticking or
848 * force prod the timer.
849 */
850 delta = next_tick - basemono;
851 if (delta <= (u64)TICK_NSEC) {
852 /*
853 * Tell the timer code that the base is not idle, i.e. undo
854 * the effect of get_next_timer_interrupt():
855 */
856 timer_clear_idle();
857 /*
858 * We've not stopped the tick yet, and there's a timer in the
859 * next period, so no point in stopping it either, bail.
860 */
861 if (!ts->tick_stopped) {
862 ts->timer_expires = 0;
863 goto out;
864 }
865 }
866
867 /*
868 * If this CPU is the one which had the do_timer() duty last, we limit
869 * the sleep time to the timekeeping 'max_deferment' value.
870 * Otherwise we can sleep as long as we want.
871 */
872 delta = timekeeping_max_deferment();
873 if (cpu != tick_do_timer_cpu &&
874 (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last))
875 delta = KTIME_MAX;
876
877 /* Calculate the next expiry time */
878 if (delta < (KTIME_MAX - basemono))
879 expires = basemono + delta;
880 else
881 expires = KTIME_MAX;
882
883 ts->timer_expires = min_t(u64, expires, next_tick);
884
885out:
886 return ts->timer_expires;
887}
888
889static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
890{
891 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
892 u64 basemono = ts->timer_expires_base;
893 u64 expires = ts->timer_expires;
894
895 /* Make sure we won't be trying to stop it twice in a row. */
896 ts->timer_expires_base = 0;
897
898 /*
899 * If this CPU is the one which updates jiffies, then give up
900 * the assignment and let it be taken by the CPU which runs
901 * the tick timer next, which might be this CPU as well. If we
902 * don't drop this here, the jiffies might be stale and
903 * do_timer() never gets invoked. Keep track of the fact that it
904 * was the one which had the do_timer() duty last.
905 */
906 if (cpu == tick_do_timer_cpu) {
907 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
908 ts->do_timer_last = 1;
909 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
910 ts->do_timer_last = 0;
911 }
912
913 /* Skip reprogram of event if it's not changed */
914 if (ts->tick_stopped && (expires == ts->next_tick)) {
915 /* Sanity check: make sure clockevent is actually programmed */
916 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
917 return;
918
919 WARN_ON_ONCE(1);
920 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
921 basemono, ts->next_tick, dev->next_event,
922 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
923 }
924
925 /*
926 * tick_nohz_stop_tick() can be called several times before
927 * tick_nohz_restart_sched_tick() is called. This happens when
928 * interrupts arrive which do not cause a reschedule. In the first
929 * call we save the current tick time, so we can restart the
930 * scheduler tick in tick_nohz_restart_sched_tick().
931 */
932 if (!ts->tick_stopped) {
933 calc_load_nohz_start();
934 quiet_vmstat();
935
936 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
937 ts->tick_stopped = 1;
938 trace_tick_stop(1, TICK_DEP_MASK_NONE);
939 }
940
941 ts->next_tick = expires;
942
943 /*
944 * If the expiration time == KTIME_MAX, then we simply stop
945 * the tick timer.
946 */
947 if (unlikely(expires == KTIME_MAX)) {
948 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
949 hrtimer_cancel(&ts->sched_timer);
950 else
951 tick_program_event(KTIME_MAX, 1);
952 return;
953 }
954
955 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
956 hrtimer_start(&ts->sched_timer, expires,
957 HRTIMER_MODE_ABS_PINNED_HARD);
958 } else {
959 hrtimer_set_expires(&ts->sched_timer, expires);
960 tick_program_event(expires, 1);
961 }
962}
963
964static void tick_nohz_retain_tick(struct tick_sched *ts)
965{
966 ts->timer_expires_base = 0;
967}
968
969#ifdef CONFIG_NO_HZ_FULL
970static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
971{
972 if (tick_nohz_next_event(ts, cpu))
973 tick_nohz_stop_tick(ts, cpu);
974 else
975 tick_nohz_retain_tick(ts);
976}
977#endif /* CONFIG_NO_HZ_FULL */
978
979static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
980{
981 /* Update jiffies first */
982 tick_do_update_jiffies64(now);
983
984 /*
985 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
986 * the clock forward checks in the enqueue path:
987 */
988 timer_clear_idle();
989
990 calc_load_nohz_stop();
991 touch_softlockup_watchdog_sched();
992
993 /* Cancel the scheduled timer and restore the tick: */
994 ts->tick_stopped = 0;
995 tick_nohz_restart(ts, now);
996}
997
998static void __tick_nohz_full_update_tick(struct tick_sched *ts,
999 ktime_t now)
1000{
1001#ifdef CONFIG_NO_HZ_FULL
1002 int cpu = smp_processor_id();
1003
1004 if (can_stop_full_tick(cpu, ts))
1005 tick_nohz_stop_sched_tick(ts, cpu);
1006 else if (ts->tick_stopped)
1007 tick_nohz_restart_sched_tick(ts, now);
1008#endif
1009}
1010
1011static void tick_nohz_full_update_tick(struct tick_sched *ts)
1012{
1013 if (!tick_nohz_full_cpu(smp_processor_id()))
1014 return;
1015
1016 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1017 return;
1018
1019 __tick_nohz_full_update_tick(ts, ktime_get());
1020}
1021
1022/*
1023 * A pending softirq outside an IRQ (or softirq disabled section) context
1024 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1025 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1026 *
1027 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1028 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1029 * triggering the code below, since wakep_softirqd() is ignored.
1030 *
1031 */
1032static bool report_idle_softirq(void)
1033{
1034 static int ratelimit;
1035 unsigned int pending = local_softirq_pending();
1036
1037 if (likely(!pending))
1038 return false;
1039
1040 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1041 if (!cpu_active(smp_processor_id())) {
1042 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1043 if (!pending)
1044 return false;
1045 }
1046
1047 if (ratelimit >= 10)
1048 return false;
1049
1050 /* On RT, softirq handling may be waiting on some lock */
1051 if (local_bh_blocked())
1052 return false;
1053
1054 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1055 pending);
1056 ratelimit++;
1057
1058 return true;
1059}
1060
1061static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1062{
1063 /*
1064 * If this CPU is offline and it is the one which updates
1065 * jiffies, then give up the assignment and let it be taken by
1066 * the CPU which runs the tick timer next. If we don't drop
1067 * this here, the jiffies might be stale and do_timer() never
1068 * gets invoked.
1069 */
1070 if (unlikely(!cpu_online(cpu))) {
1071 if (cpu == tick_do_timer_cpu)
1072 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1073 /*
1074 * Make sure the CPU doesn't get fooled by obsolete tick
1075 * deadline if it comes back online later.
1076 */
1077 ts->next_tick = 0;
1078 return false;
1079 }
1080
1081 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1082 return false;
1083
1084 if (need_resched())
1085 return false;
1086
1087 if (unlikely(report_idle_softirq()))
1088 return false;
1089
1090 if (tick_nohz_full_enabled()) {
1091 /*
1092 * Keep the tick alive to guarantee timekeeping progression
1093 * if there are full dynticks CPUs around
1094 */
1095 if (tick_do_timer_cpu == cpu)
1096 return false;
1097
1098 /* Should not happen for nohz-full */
1099 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1100 return false;
1101 }
1102
1103 return true;
1104}
1105
1106/**
1107 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1108 *
1109 * When the next event is more than a tick into the future, stop the idle tick
1110 */
1111void tick_nohz_idle_stop_tick(void)
1112{
1113 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1114 int cpu = smp_processor_id();
1115 ktime_t expires;
1116
1117 /*
1118 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1119 * tick timer expiration time is known already.
1120 */
1121 if (ts->timer_expires_base)
1122 expires = ts->timer_expires;
1123 else if (can_stop_idle_tick(cpu, ts))
1124 expires = tick_nohz_next_event(ts, cpu);
1125 else
1126 return;
1127
1128 ts->idle_calls++;
1129
1130 if (expires > 0LL) {
1131 int was_stopped = ts->tick_stopped;
1132
1133 tick_nohz_stop_tick(ts, cpu);
1134
1135 ts->idle_sleeps++;
1136 ts->idle_expires = expires;
1137
1138 if (!was_stopped && ts->tick_stopped) {
1139 ts->idle_jiffies = ts->last_jiffies;
1140 nohz_balance_enter_idle(cpu);
1141 }
1142 } else {
1143 tick_nohz_retain_tick(ts);
1144 }
1145}
1146
1147void tick_nohz_idle_retain_tick(void)
1148{
1149 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1150 /*
1151 * Undo the effect of get_next_timer_interrupt() called from
1152 * tick_nohz_next_event().
1153 */
1154 timer_clear_idle();
1155}
1156
1157/**
1158 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1159 *
1160 * Called when we start the idle loop.
1161 */
1162void tick_nohz_idle_enter(void)
1163{
1164 struct tick_sched *ts;
1165
1166 lockdep_assert_irqs_enabled();
1167
1168 local_irq_disable();
1169
1170 ts = this_cpu_ptr(&tick_cpu_sched);
1171
1172 WARN_ON_ONCE(ts->timer_expires_base);
1173
1174 ts->inidle = 1;
1175 tick_nohz_start_idle(ts);
1176
1177 local_irq_enable();
1178}
1179
1180/**
1181 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1182 *
1183 * A timer may have been added/modified/deleted either by the current IRQ,
1184 * or by another place using this IRQ as a notification. This IRQ may have
1185 * also updated the RCU callback list. These events may require a
1186 * re-evaluation of the next tick. Depending on the context:
1187 *
1188 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1189 * time accounting. The next tick will be re-evaluated on the next idle
1190 * loop iteration.
1191 *
1192 * 2) If the CPU is nohz_full:
1193 *
1194 * 2.1) If there is any tick dependency, restart the tick if stopped.
1195 *
1196 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1197 * stop/update it accordingly.
1198 */
1199void tick_nohz_irq_exit(void)
1200{
1201 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1202
1203 if (ts->inidle)
1204 tick_nohz_start_idle(ts);
1205 else
1206 tick_nohz_full_update_tick(ts);
1207}
1208
1209/**
1210 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1211 */
1212bool tick_nohz_idle_got_tick(void)
1213{
1214 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1215
1216 if (ts->got_idle_tick) {
1217 ts->got_idle_tick = 0;
1218 return true;
1219 }
1220 return false;
1221}
1222
1223/**
1224 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1225 * or the tick, whichever expires first. Note that, if the tick has been
1226 * stopped, it returns the next hrtimer.
1227 *
1228 * Called from power state control code with interrupts disabled
1229 */
1230ktime_t tick_nohz_get_next_hrtimer(void)
1231{
1232 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1233}
1234
1235/**
1236 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1237 * @delta_next: duration until the next event if the tick cannot be stopped
1238 *
1239 * Called from power state control code with interrupts disabled.
1240 *
1241 * The return value of this function and/or the value returned by it through the
1242 * @delta_next pointer can be negative which must be taken into account by its
1243 * callers.
1244 */
1245ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1246{
1247 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1248 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1249 int cpu = smp_processor_id();
1250 /*
1251 * The idle entry time is expected to be a sufficient approximation of
1252 * the current time at this point.
1253 */
1254 ktime_t now = ts->idle_entrytime;
1255 ktime_t next_event;
1256
1257 WARN_ON_ONCE(!ts->inidle);
1258
1259 *delta_next = ktime_sub(dev->next_event, now);
1260
1261 if (!can_stop_idle_tick(cpu, ts))
1262 return *delta_next;
1263
1264 next_event = tick_nohz_next_event(ts, cpu);
1265 if (!next_event)
1266 return *delta_next;
1267
1268 /*
1269 * If the next highres timer to expire is earlier than 'next_event', the
1270 * idle governor needs to know that.
1271 */
1272 next_event = min_t(u64, next_event,
1273 hrtimer_next_event_without(&ts->sched_timer));
1274
1275 return ktime_sub(next_event, now);
1276}
1277
1278/**
1279 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1280 * for a particular CPU.
1281 *
1282 * Called from the schedutil frequency scaling governor in scheduler context.
1283 */
1284unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1285{
1286 struct tick_sched *ts = tick_get_tick_sched(cpu);
1287
1288 return ts->idle_calls;
1289}
1290
1291/**
1292 * tick_nohz_get_idle_calls - return the current idle calls counter value
1293 *
1294 * Called from the schedutil frequency scaling governor in scheduler context.
1295 */
1296unsigned long tick_nohz_get_idle_calls(void)
1297{
1298 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1299
1300 return ts->idle_calls;
1301}
1302
1303static void tick_nohz_account_idle_time(struct tick_sched *ts,
1304 ktime_t now)
1305{
1306 unsigned long ticks;
1307
1308 ts->idle_exittime = now;
1309
1310 if (vtime_accounting_enabled_this_cpu())
1311 return;
1312 /*
1313 * We stopped the tick in idle. update_process_times() would miss the
1314 * time we slept, as it does only a 1 tick accounting.
1315 * Enforce that this is accounted to idle !
1316 */
1317 ticks = jiffies - ts->idle_jiffies;
1318 /*
1319 * We might be one off. Do not randomly account a huge number of ticks!
1320 */
1321 if (ticks && ticks < LONG_MAX)
1322 account_idle_ticks(ticks);
1323}
1324
1325void tick_nohz_idle_restart_tick(void)
1326{
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1328
1329 if (ts->tick_stopped) {
1330 ktime_t now = ktime_get();
1331 tick_nohz_restart_sched_tick(ts, now);
1332 tick_nohz_account_idle_time(ts, now);
1333 }
1334}
1335
1336static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1337{
1338 if (tick_nohz_full_cpu(smp_processor_id()))
1339 __tick_nohz_full_update_tick(ts, now);
1340 else
1341 tick_nohz_restart_sched_tick(ts, now);
1342
1343 tick_nohz_account_idle_time(ts, now);
1344}
1345
1346/**
1347 * tick_nohz_idle_exit - Update the tick upon idle task exit
1348 *
1349 * When the idle task exits, update the tick depending on the
1350 * following situations:
1351 *
1352 * 1) If the CPU is not in nohz_full mode (most cases), then
1353 * restart the tick.
1354 *
1355 * 2) If the CPU is in nohz_full mode (corner case):
1356 * 2.1) If the tick can be kept stopped (no tick dependencies)
1357 * then re-evaluate the next tick and try to keep it stopped
1358 * as long as possible.
1359 * 2.2) If the tick has dependencies, restart the tick.
1360 *
1361 */
1362void tick_nohz_idle_exit(void)
1363{
1364 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1365 bool idle_active, tick_stopped;
1366 ktime_t now;
1367
1368 local_irq_disable();
1369
1370 WARN_ON_ONCE(!ts->inidle);
1371 WARN_ON_ONCE(ts->timer_expires_base);
1372
1373 ts->inidle = 0;
1374 idle_active = ts->idle_active;
1375 tick_stopped = ts->tick_stopped;
1376
1377 if (idle_active || tick_stopped)
1378 now = ktime_get();
1379
1380 if (idle_active)
1381 tick_nohz_stop_idle(ts, now);
1382
1383 if (tick_stopped)
1384 tick_nohz_idle_update_tick(ts, now);
1385
1386 local_irq_enable();
1387}
1388
1389/*
1390 * In low-resolution mode, the tick handler must be implemented directly
1391 * at the clockevent level. hrtimer can't be used instead, because its
1392 * infrastructure actually relies on the tick itself as a backend in
1393 * low-resolution mode (see hrtimer_run_queues()).
1394 *
1395 * This low-resolution handler still makes use of some hrtimer APIs meanwhile
1396 * for convenience with expiration calculation and forwarding.
1397 */
1398static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1399{
1400 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1401 struct pt_regs *regs = get_irq_regs();
1402 ktime_t now = ktime_get();
1403
1404 dev->next_event = KTIME_MAX;
1405
1406 tick_sched_do_timer(ts, now);
1407 tick_sched_handle(ts, regs);
1408
1409 /*
1410 * In dynticks mode, tick reprogram is deferred:
1411 * - to the idle task if in dynticks-idle
1412 * - to IRQ exit if in full-dynticks.
1413 */
1414 if (likely(!ts->tick_stopped)) {
1415 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1416 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1417 }
1418
1419}
1420
1421static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1422{
1423 if (!tick_nohz_enabled)
1424 return;
1425 ts->nohz_mode = mode;
1426 /* One update is enough */
1427 if (!test_and_set_bit(0, &tick_nohz_active))
1428 timers_update_nohz();
1429}
1430
1431/**
1432 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1433 */
1434static void tick_nohz_switch_to_nohz(void)
1435{
1436 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1437 ktime_t next;
1438
1439 if (!tick_nohz_enabled)
1440 return;
1441
1442 if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1443 return;
1444
1445 /*
1446 * Recycle the hrtimer in 'ts', so we can share the
1447 * hrtimer_forward_now() function with the highres code.
1448 */
1449 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1450 /* Get the next period */
1451 next = tick_init_jiffy_update();
1452
1453 hrtimer_set_expires(&ts->sched_timer, next);
1454 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1455 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1456 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1457}
1458
1459static inline void tick_nohz_irq_enter(void)
1460{
1461 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1462 ktime_t now;
1463
1464 if (!ts->idle_active && !ts->tick_stopped)
1465 return;
1466 now = ktime_get();
1467 if (ts->idle_active)
1468 tick_nohz_stop_idle(ts, now);
1469 /*
1470 * If all CPUs are idle we may need to update a stale jiffies value.
1471 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1472 * alive but it might be busy looping with interrupts disabled in some
1473 * rare case (typically stop machine). So we must make sure we have a
1474 * last resort.
1475 */
1476 if (ts->tick_stopped)
1477 tick_nohz_update_jiffies(now);
1478}
1479
1480#else
1481
1482static inline void tick_nohz_switch_to_nohz(void) { }
1483static inline void tick_nohz_irq_enter(void) { }
1484static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1485
1486#endif /* CONFIG_NO_HZ_COMMON */
1487
1488/*
1489 * Called from irq_enter() to notify about the possible interruption of idle()
1490 */
1491void tick_irq_enter(void)
1492{
1493 tick_check_oneshot_broadcast_this_cpu();
1494 tick_nohz_irq_enter();
1495}
1496
1497/*
1498 * High resolution timer specific code
1499 */
1500#ifdef CONFIG_HIGH_RES_TIMERS
1501/*
1502 * We rearm the timer until we get disabled by the idle code.
1503 * Called with interrupts disabled.
1504 */
1505static enum hrtimer_restart tick_nohz_highres_handler(struct hrtimer *timer)
1506{
1507 struct tick_sched *ts =
1508 container_of(timer, struct tick_sched, sched_timer);
1509 struct pt_regs *regs = get_irq_regs();
1510 ktime_t now = ktime_get();
1511
1512 tick_sched_do_timer(ts, now);
1513
1514 /*
1515 * Do not call when we are not in IRQ context and have
1516 * no valid 'regs' pointer
1517 */
1518 if (regs)
1519 tick_sched_handle(ts, regs);
1520 else
1521 ts->next_tick = 0;
1522
1523 /*
1524 * In dynticks mode, tick reprogram is deferred:
1525 * - to the idle task if in dynticks-idle
1526 * - to IRQ exit if in full-dynticks.
1527 */
1528 if (unlikely(ts->tick_stopped))
1529 return HRTIMER_NORESTART;
1530
1531 hrtimer_forward(timer, now, TICK_NSEC);
1532
1533 return HRTIMER_RESTART;
1534}
1535
1536static int sched_skew_tick;
1537
1538static int __init skew_tick(char *str)
1539{
1540 get_option(&str, &sched_skew_tick);
1541
1542 return 0;
1543}
1544early_param("skew_tick", skew_tick);
1545
1546/**
1547 * tick_setup_sched_timer - setup the tick emulation timer
1548 */
1549void tick_setup_sched_timer(void)
1550{
1551 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1552 ktime_t now = ktime_get();
1553
1554 /* Emulate tick processing via per-CPU hrtimers: */
1555 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1556 ts->sched_timer.function = tick_nohz_highres_handler;
1557
1558 /* Get the next period (per-CPU) */
1559 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1560
1561 /* Offset the tick to avert 'jiffies_lock' contention. */
1562 if (sched_skew_tick) {
1563 u64 offset = TICK_NSEC >> 1;
1564 do_div(offset, num_possible_cpus());
1565 offset *= smp_processor_id();
1566 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1567 }
1568
1569 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
1570 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1571 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1572}
1573#endif /* HIGH_RES_TIMERS */
1574
1575#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1576void tick_cancel_sched_timer(int cpu)
1577{
1578 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1579 ktime_t idle_sleeptime, iowait_sleeptime;
1580 unsigned long idle_calls, idle_sleeps;
1581
1582# ifdef CONFIG_HIGH_RES_TIMERS
1583 if (ts->sched_timer.base)
1584 hrtimer_cancel(&ts->sched_timer);
1585# endif
1586
1587 idle_sleeptime = ts->idle_sleeptime;
1588 iowait_sleeptime = ts->iowait_sleeptime;
1589 idle_calls = ts->idle_calls;
1590 idle_sleeps = ts->idle_sleeps;
1591 memset(ts, 0, sizeof(*ts));
1592 ts->idle_sleeptime = idle_sleeptime;
1593 ts->iowait_sleeptime = iowait_sleeptime;
1594 ts->idle_calls = idle_calls;
1595 ts->idle_sleeps = idle_sleeps;
1596}
1597#endif
1598
1599/*
1600 * Async notification about clocksource changes
1601 */
1602void tick_clock_notify(void)
1603{
1604 int cpu;
1605
1606 for_each_possible_cpu(cpu)
1607 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1608}
1609
1610/*
1611 * Async notification about clock event changes
1612 */
1613void tick_oneshot_notify(void)
1614{
1615 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1616
1617 set_bit(0, &ts->check_clocks);
1618}
1619
1620/*
1621 * Check if a change happened, which makes oneshot possible.
1622 *
1623 * Called cyclically from the hrtimer softirq (driven by the timer
1624 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1625 * mode, because high resolution timers are disabled (either compile
1626 * or runtime). Called with interrupts disabled.
1627 */
1628int tick_check_oneshot_change(int allow_nohz)
1629{
1630 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1631
1632 if (!test_and_clear_bit(0, &ts->check_clocks))
1633 return 0;
1634
1635 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1636 return 0;
1637
1638 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1639 return 0;
1640
1641 if (!allow_nohz)
1642 return 1;
1643
1644 tick_nohz_switch_to_nohz();
1645 return 0;
1646}
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}