projects/linux·blob master · refs

1602 lines · 41.2 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2005-2006, Linutronix GmbH, Thomas Gleixner <[email protected]>
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/*
205 * Allow only one non-timekeeper CPU at a time update jiffies from
206 * the timer tick.
207 *
208 * Returns true if update was run.
209 */
210static bool tick_limited_update_jiffies64(struct tick_sched *ts, ktime_t now)
211{
212 static atomic_t in_progress;
213 int inp;
214
215 inp = atomic_read(&in_progress);
216 if (inp || !atomic_try_cmpxchg(&in_progress, &inp, 1))
217 return false;
218
219 if (ts->last_tick_jiffies == jiffies)
220 tick_do_update_jiffies64(now);
221 atomic_set(&in_progress, 0);
222 return true;
223}
224
225#define MAX_STALLED_JIFFIES 5
226
227static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
228{
229 int tick_cpu, cpu = smp_processor_id();
230
231 /*
232 * Check if the do_timer duty was dropped. We don't care about
233 * concurrency: This happens only when the CPU in charge went
234 * into a long sleep. If two CPUs happen to assign themselves to
235 * this duty, then the jiffies update is still serialized by
236 * 'jiffies_lock'.
237 *
238 * If nohz_full is enabled, this should not happen because the
239 * 'tick_do_timer_cpu' CPU never relinquishes.
240 */
241 tick_cpu = READ_ONCE(tick_do_timer_cpu);
242
243 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && unlikely(tick_cpu == TICK_DO_TIMER_NONE)) {
244#ifdef CONFIG_NO_HZ_FULL
245 WARN_ON_ONCE(tick_nohz_full_running);
246#endif
247 WRITE_ONCE(tick_do_timer_cpu, cpu);
248 tick_cpu = cpu;
249 }
250
251 /* Check if jiffies need an update */
252 if (tick_cpu == cpu)
253 tick_do_update_jiffies64(now);
254
255 /*
256 * If the jiffies update stalled for too long (timekeeper in stop_machine()
257 * or VMEXIT'ed for several msecs), force an update.
258 */
259 if (ts->last_tick_jiffies != jiffies) {
260 ts->stalled_jiffies = 0;
261 ts->last_tick_jiffies = READ_ONCE(jiffies);
262 } else {
263 if (++ts->stalled_jiffies >= MAX_STALLED_JIFFIES) {
264 if (tick_limited_update_jiffies64(ts, now)) {
265 ts->stalled_jiffies = 0;
266 ts->last_tick_jiffies = READ_ONCE(jiffies);
267 }
268 }
269 }
270
271 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE))
272 ts->got_idle_tick = 1;
273}
274
275static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
276{
277 /*
278 * When we are idle and the tick is stopped, we have to touch
279 * the watchdog as we might not schedule for a really long
280 * time. This happens on completely idle SMP systems while
281 * waiting on the login prompt. We also increment the "start of
282 * idle" jiffy stamp so the idle accounting adjustment we do
283 * when we go busy again does not account too many ticks.
284 */
285 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) &&
286 tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
287 touch_softlockup_watchdog_sched();
288 /*
289 * In case the current tick fired too early past its expected
290 * expiration, make sure we don't bypass the next clock reprogramming
291 * to the same deadline.
292 */
293 ts->next_tick = 0;
294 }
295
296 update_process_times(user_mode(regs));
297 profile_tick(CPU_PROFILING);
298}
299
300/*
301 * We rearm the timer until we get disabled by the idle code.
302 * Called with interrupts disabled.
303 */
304static enum hrtimer_restart tick_nohz_handler(struct hrtimer *timer)
305{
306 struct tick_sched *ts = container_of(timer, struct tick_sched, sched_timer);
307 struct pt_regs *regs = get_irq_regs();
308 ktime_t now = ktime_get();
309
310 tick_sched_do_timer(ts, now);
311
312 /*
313 * Do not call when we are not in IRQ context and have
314 * no valid 'regs' pointer
315 */
316 if (regs)
317 tick_sched_handle(ts, regs);
318 else
319 ts->next_tick = 0;
320
321 /*
322 * In dynticks mode, tick reprogram is deferred:
323 * - to the idle task if in dynticks-idle
324 * - to IRQ exit if in full-dynticks.
325 */
326 if (unlikely(tick_sched_flag_test(ts, TS_FLAG_STOPPED)))
327 return HRTIMER_NORESTART;
328
329 hrtimer_forward(timer, now, TICK_NSEC);
330
331 return HRTIMER_RESTART;
332}
333
334#ifdef CONFIG_NO_HZ_FULL
335cpumask_var_t tick_nohz_full_mask;
336EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
337bool tick_nohz_full_running;
338EXPORT_SYMBOL_GPL(tick_nohz_full_running);
339static atomic_t tick_dep_mask;
340
341static bool check_tick_dependency(atomic_t *dep)
342{
343 int val = atomic_read(dep);
344
345 if (likely(!tracepoint_enabled(tick_stop)))
346 return !!val;
347
348 if (val & TICK_DEP_MASK_POSIX_TIMER) {
349 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
350 return true;
351 }
352
353 if (val & TICK_DEP_MASK_PERF_EVENTS) {
354 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
355 return true;
356 }
357
358 if (val & TICK_DEP_MASK_SCHED) {
359 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
360 return true;
361 }
362
363 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
364 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
365 return true;
366 }
367
368 if (val & TICK_DEP_MASK_RCU) {
369 trace_tick_stop(0, TICK_DEP_MASK_RCU);
370 return true;
371 }
372
373 if (val & TICK_DEP_MASK_RCU_EXP) {
374 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
375 return true;
376 }
377
378 return false;
379}
380
381static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
382{
383 lockdep_assert_irqs_disabled();
384
385 if (unlikely(!cpu_online(cpu)))
386 return false;
387
388 if (check_tick_dependency(&tick_dep_mask))
389 return false;
390
391 if (check_tick_dependency(&ts->tick_dep_mask))
392 return false;
393
394 if (check_tick_dependency(&current->tick_dep_mask))
395 return false;
396
397 if (check_tick_dependency(&current->signal->tick_dep_mask))
398 return false;
399
400 return true;
401}
402
403static void nohz_full_kick_func(struct irq_work *work)
404{
405 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
406}
407
408static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
409 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
410
411/*
412 * Kick this CPU if it's full dynticks in order to force it to
413 * re-evaluate its dependency on the tick and restart it if necessary.
414 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
415 * is NMI safe.
416 */
417static void tick_nohz_full_kick(void)
418{
419 if (!tick_nohz_full_cpu(smp_processor_id()))
420 return;
421
422 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
423}
424
425/*
426 * Kick the CPU if it's full dynticks in order to force it to
427 * re-evaluate its dependency on the tick and restart it if necessary.
428 */
429void tick_nohz_full_kick_cpu(int cpu)
430{
431 if (!tick_nohz_full_cpu(cpu))
432 return;
433
434 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
435}
436
437static void tick_nohz_kick_task(struct task_struct *tsk)
438{
439 int cpu;
440
441 /*
442 * If the task is not running, run_posix_cpu_timers()
443 * has nothing to elapse, and an IPI can then be optimized out.
444 *
445 * activate_task() STORE p->tick_dep_mask
446 * STORE p->on_rq
447 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or())
448 * LOCK rq->lock LOAD p->on_rq
449 * smp_mb__after_spin_lock()
450 * tick_nohz_task_switch()
451 * LOAD p->tick_dep_mask
452 *
453 * XXX given a task picks up the dependency on schedule(), should we
454 * only care about tasks that are currently on the CPU instead of all
455 * that are on the runqueue?
456 *
457 * That is, does this want to be: task_on_cpu() / task_curr()?
458 */
459 if (!sched_task_on_rq(tsk))
460 return;
461
462 /*
463 * If the task concurrently migrates to another CPU,
464 * we guarantee it sees the new tick dependency upon
465 * schedule.
466 *
467 * set_task_cpu(p, cpu);
468 * STORE p->cpu = @cpu
469 * __schedule() (switch to task 'p')
470 * LOCK rq->lock
471 * smp_mb__after_spin_lock() STORE p->tick_dep_mask
472 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or())
473 * LOAD p->tick_dep_mask LOAD p->cpu
474 */
475 cpu = task_cpu(tsk);
476
477 preempt_disable();
478 if (cpu_online(cpu))
479 tick_nohz_full_kick_cpu(cpu);
480 preempt_enable();
481}
482
483/*
484 * Kick all full dynticks CPUs in order to force these to re-evaluate
485 * their dependency on the tick and restart it if necessary.
486 */
487static void tick_nohz_full_kick_all(void)
488{
489 int cpu;
490
491 if (!tick_nohz_full_running)
492 return;
493
494 preempt_disable();
495 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
496 tick_nohz_full_kick_cpu(cpu);
497 preempt_enable();
498}
499
500static void tick_nohz_dep_set_all(atomic_t *dep,
501 enum tick_dep_bits bit)
502{
503 int prev;
504
505 prev = atomic_fetch_or(BIT(bit), dep);
506 if (!prev)
507 tick_nohz_full_kick_all();
508}
509
510/*
511 * Set a global tick dependency. Used by perf events that rely on freq and
512 * unstable clocks.
513 */
514void tick_nohz_dep_set(enum tick_dep_bits bit)
515{
516 tick_nohz_dep_set_all(&tick_dep_mask, bit);
517}
518
519void tick_nohz_dep_clear(enum tick_dep_bits bit)
520{
521 atomic_andnot(BIT(bit), &tick_dep_mask);
522}
523
524/*
525 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
526 * manage event-throttling.
527 */
528void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
529{
530 int prev;
531 struct tick_sched *ts;
532
533 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
534
535 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
536 if (!prev) {
537 preempt_disable();
538 /* Perf needs local kick that is NMI safe */
539 if (cpu == smp_processor_id()) {
540 tick_nohz_full_kick();
541 } else {
542 /* Remote IRQ work not NMI-safe */
543 if (!WARN_ON_ONCE(in_nmi()))
544 tick_nohz_full_kick_cpu(cpu);
545 }
546 preempt_enable();
547 }
548}
549EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
550
551void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
552{
553 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
554
555 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
556}
557EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
558
559/*
560 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
561 * in order to elapse per task timers.
562 */
563void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
564{
565 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
566 tick_nohz_kick_task(tsk);
567}
568EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
569
570void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
571{
572 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
573}
574EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
575
576/*
577 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
578 * per process timers.
579 */
580void tick_nohz_dep_set_signal(struct task_struct *tsk,
581 enum tick_dep_bits bit)
582{
583 int prev;
584 struct signal_struct *sig = tsk->signal;
585
586 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
587 if (!prev) {
588 struct task_struct *t;
589
590 lockdep_assert_held(&tsk->sighand->siglock);
591 __for_each_thread(sig, t)
592 tick_nohz_kick_task(t);
593 }
594}
595
596void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
597{
598 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
599}
600
601/*
602 * Re-evaluate the need for the tick as we switch the current task.
603 * It might need the tick due to per task/process properties:
604 * perf events, posix CPU timers, ...
605 */
606void __tick_nohz_task_switch(void)
607{
608 struct tick_sched *ts;
609
610 if (!tick_nohz_full_cpu(smp_processor_id()))
611 return;
612
613 ts = this_cpu_ptr(&tick_cpu_sched);
614
615 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
616 if (atomic_read(&current->tick_dep_mask) ||
617 atomic_read(&current->signal->tick_dep_mask))
618 tick_nohz_full_kick();
619 }
620}
621
622/* Get the boot-time nohz CPU list from the kernel parameters. */
623void __init tick_nohz_full_setup(cpumask_var_t cpumask)
624{
625 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
626 cpumask_copy(tick_nohz_full_mask, cpumask);
627 tick_nohz_full_running = true;
628}
629
630bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
631{
632 /*
633 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
634 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
635 * CPUs. It must remain online when nohz full is enabled.
636 */
637 if (tick_nohz_full_running && READ_ONCE(tick_do_timer_cpu) == cpu)
638 return false;
639 return true;
640}
641
642static int tick_nohz_cpu_down(unsigned int cpu)
643{
644 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
645}
646
647void __init tick_nohz_init(void)
648{
649 int cpu, ret;
650
651 if (!tick_nohz_full_running)
652 return;
653
654 /*
655 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
656 * locking contexts. But then we need IRQ work to raise its own
657 * interrupts to avoid circular dependency on the tick.
658 */
659 if (!arch_irq_work_has_interrupt()) {
660 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
661 cpumask_clear(tick_nohz_full_mask);
662 tick_nohz_full_running = false;
663 return;
664 }
665
666 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
667 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
668 cpu = smp_processor_id();
669
670 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
671 pr_warn("NO_HZ: Clearing %d from nohz_full range "
672 "for timekeeping\n", cpu);
673 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
674 }
675 }
676
677 for_each_cpu(cpu, tick_nohz_full_mask)
678 ct_cpu_track_user(cpu);
679
680 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
681 "kernel/nohz:predown", NULL,
682 tick_nohz_cpu_down);
683 WARN_ON(ret < 0);
684 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
685 cpumask_pr_args(tick_nohz_full_mask));
686}
687#endif /* #ifdef CONFIG_NO_HZ_FULL */
688
689/*
690 * NOHZ - aka dynamic tick functionality
691 */
692#ifdef CONFIG_NO_HZ_COMMON
693/*
694 * NO HZ enabled ?
695 */
696bool tick_nohz_enabled __read_mostly = true;
697static unsigned long tick_nohz_active __read_mostly;
698/*
699 * Enable / Disable tickless mode
700 */
701static int __init setup_tick_nohz(char *str)
702{
703 return (kstrtobool(str, &tick_nohz_enabled) == 0);
704}
705
706__setup("nohz=", setup_tick_nohz);
707
708bool tick_nohz_is_active(void)
709{
710 return tick_nohz_active;
711}
712EXPORT_SYMBOL_GPL(tick_nohz_is_active);
713
714bool tick_nohz_tick_stopped(void)
715{
716 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
717
718 return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
719}
720
721bool tick_nohz_tick_stopped_cpu(int cpu)
722{
723 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
724
725 return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
726}
727
728/**
729 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
730 * @now: current ktime_t
731 *
732 * Called from interrupt entry when the CPU was idle
733 *
734 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
735 * must be updated. Otherwise an interrupt handler could use a stale jiffy
736 * value. We do this unconditionally on any CPU, as we don't know whether the
737 * CPU, which has the update task assigned, is in a long sleep.
738 */
739static void tick_nohz_update_jiffies(ktime_t now)
740{
741 unsigned long flags;
742
743 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
744
745 local_irq_save(flags);
746 tick_do_update_jiffies64(now);
747 local_irq_restore(flags);
748
749 touch_softlockup_watchdog_sched();
750}
751
752/* Simplified variant of hrtimer_forward_now() */
753static ktime_t tick_forward_now(ktime_t expires, ktime_t now)
754{
755 ktime_t delta = now - expires;
756
757 if (likely(delta < TICK_NSEC))
758 return expires + TICK_NSEC;
759
760 expires += TICK_NSEC * ktime_divns(delta, TICK_NSEC);
761 if (expires > now)
762 return expires;
763 return expires + TICK_NSEC;
764}
765
766static void tick_nohz_forward_last_tick(struct tick_sched *ts, ktime_t now)
767{
768 if (now >= ts->last_tick)
769 ts->last_tick = tick_forward_now(ts->last_tick, now);
770}
771
772static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
773{
774 ktime_t expires;
775
776 tick_nohz_forward_last_tick(ts, now);
777 expires = ts->last_tick;
778
779 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
780 hrtimer_start(&ts->sched_timer, expires, HRTIMER_MODE_ABS_PINNED_HARD);
781 } else {
782 hrtimer_set_expires(&ts->sched_timer, expires);
783 tick_program_event(expires, 1);
784 }
785
786 /*
787 * Reset to make sure the next tick stop doesn't get fooled by past
788 * cached clock deadline.
789 */
790 ts->next_tick = 0;
791}
792
793static inline bool local_timer_softirq_pending(void)
794{
795 return local_timers_pending() & BIT(TIMER_SOFTIRQ);
796}
797
798/*
799 * Read jiffies and the time when jiffies were updated last
800 */
801u64 get_jiffies_update(unsigned long *basej)
802{
803 unsigned long basejiff;
804 unsigned int seq;
805 u64 basemono;
806
807 do {
808 seq = read_seqcount_begin(&jiffies_seq);
809 basemono = last_jiffies_update;
810 basejiff = jiffies;
811 } while (read_seqcount_retry(&jiffies_seq, seq));
812 *basej = basejiff;
813 return basemono;
814}
815
816/**
817 * tick_nohz_next_event() - return the clock monotonic based next event
818 * @ts: pointer to tick_sched struct
819 * @cpu: CPU number
820 *
821 * Return:
822 * *%0 - When the next event is a maximum of TICK_NSEC in the future
823 * and the tick is not stopped yet
824 * *%next_event - Next event based on clock monotonic
825 */
826static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
827{
828 u64 basemono, next_tick, delta, expires;
829 unsigned long basejiff;
830 int tick_cpu;
831
832 basemono = get_jiffies_update(&basejiff);
833 ts->last_jiffies = basejiff;
834 ts->timer_expires_base = basemono;
835
836 /*
837 * Keep the periodic tick, when RCU or irq_work requests it.
838 * Aside of that, check whether the local timer softirq is
839 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
840 * because there is an already expired timer, so it will request
841 * immediate expiry, which rearms the hardware timer with a
842 * minimal delta, which brings us back to this place
843 * immediately. Lather, rinse and repeat...
844 */
845 if (rcu_needs_cpu() || irq_work_needs_cpu() ||
846 local_timer_softirq_pending()) {
847 next_tick = basemono + TICK_NSEC;
848 } else {
849 /*
850 * Get the next pending timer. If high resolution
851 * timers are enabled this only takes the timer wheel
852 * timers into account. If high resolution timers are
853 * disabled this also looks at the next expiring
854 * hrtimer.
855 */
856 next_tick = get_next_timer_interrupt(basejiff, basemono);
857 ts->next_timer = next_tick;
858 }
859
860 /* Make sure next_tick is never before basemono! */
861 if (WARN_ON_ONCE(basemono > next_tick))
862 next_tick = basemono;
863
864 /*
865 * If the tick is due in the next period, keep it ticking or
866 * force prod the timer.
867 */
868 delta = next_tick - basemono;
869 if (delta <= (u64)TICK_NSEC) {
870 /*
871 * We've not stopped the tick yet, and there's a timer in the
872 * next period, so no point in stopping it either, bail.
873 */
874 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
875 ts->timer_expires = 0;
876 goto out;
877 }
878 }
879
880 /*
881 * If this CPU is the one which had the do_timer() duty last, we limit
882 * the sleep time to the timekeeping 'max_deferment' value.
883 * Otherwise we can sleep as long as we want.
884 */
885 delta = timekeeping_max_deferment();
886 tick_cpu = READ_ONCE(tick_do_timer_cpu);
887 if (tick_cpu != cpu &&
888 (tick_cpu != TICK_DO_TIMER_NONE || !tick_sched_flag_test(ts, TS_FLAG_DO_TIMER_LAST)))
889 delta = KTIME_MAX;
890
891 /* Calculate the next expiry time */
892 if (delta < (KTIME_MAX - basemono))
893 expires = basemono + delta;
894 else
895 expires = KTIME_MAX;
896
897 ts->timer_expires = min_t(u64, expires, next_tick);
898
899out:
900 return ts->timer_expires;
901}
902
903static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
904{
905 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
906 unsigned long basejiff = ts->last_jiffies;
907 u64 basemono = ts->timer_expires_base;
908 bool timer_idle = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
909 int tick_cpu;
910 u64 expires;
911
912 /* Make sure we won't be trying to stop it twice in a row. */
913 ts->timer_expires_base = 0;
914
915 /*
916 * Now the tick should be stopped definitely - so the timer base needs
917 * to be marked idle as well to not miss a newly queued timer.
918 */
919 expires = timer_base_try_to_set_idle(basejiff, basemono, &timer_idle);
920 if (expires > ts->timer_expires) {
921 /*
922 * This path could only happen when the first timer was removed
923 * between calculating the possible sleep length and now (when
924 * high resolution mode is not active, timer could also be a
925 * hrtimer).
926 *
927 * We have to stick to the original calculated expiry value to
928 * not stop the tick for too long with a shallow C-state (which
929 * was programmed by cpuidle because of an early next expiration
930 * value).
931 */
932 expires = ts->timer_expires;
933 }
934
935 /* If the timer base is not idle, retain the not yet stopped tick. */
936 if (!timer_idle)
937 return;
938
939 /*
940 * If this CPU is the one which updates jiffies, then give up
941 * the assignment and let it be taken by the CPU which runs
942 * the tick timer next, which might be this CPU as well. If we
943 * don't drop this here, the jiffies might be stale and
944 * do_timer() never gets invoked. Keep track of the fact that it
945 * was the one which had the do_timer() duty last.
946 */
947 tick_cpu = READ_ONCE(tick_do_timer_cpu);
948 if (tick_cpu == cpu) {
949 WRITE_ONCE(tick_do_timer_cpu, TICK_DO_TIMER_NONE);
950 tick_sched_flag_set(ts, TS_FLAG_DO_TIMER_LAST);
951 } else if (tick_cpu != TICK_DO_TIMER_NONE) {
952 tick_sched_flag_clear(ts, TS_FLAG_DO_TIMER_LAST);
953 }
954
955 /* Skip reprogram of event if it's not changed */
956 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED) && (expires == ts->next_tick)) {
957 /* Sanity check: make sure clockevent is actually programmed */
958 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
959 return;
960
961 WARN_ONCE(1, "basemono: %llu ts->next_tick: %llu dev->next_event: %llu "
962 "timer->active: %d timer->expires: %llu\n", basemono, ts->next_tick,
963 dev->next_event, hrtimer_active(&ts->sched_timer),
964 hrtimer_get_expires(&ts->sched_timer));
965 }
966
967 /*
968 * tick_nohz_stop_tick() can be called several times before
969 * tick_nohz_restart_sched_tick() is called. This happens when
970 * interrupts arrive which do not cause a reschedule. In the first
971 * call we save the current tick time, so we can restart the
972 * scheduler tick in tick_nohz_restart_sched_tick().
973 */
974 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
975 calc_load_nohz_start();
976 quiet_vmstat();
977
978 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
979 tick_sched_flag_set(ts, TS_FLAG_STOPPED);
980 trace_tick_stop(1, TICK_DEP_MASK_NONE);
981 }
982
983 ts->next_tick = expires;
984
985 /*
986 * If the expiration time == KTIME_MAX, then we simply stop
987 * the tick timer.
988 */
989 if (unlikely(expires == KTIME_MAX)) {
990 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES))
991 hrtimer_cancel(&ts->sched_timer);
992 else
993 tick_program_event(KTIME_MAX, 1);
994 return;
995 }
996
997 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
998 hrtimer_start(&ts->sched_timer, expires,
999 HRTIMER_MODE_ABS_PINNED_HARD);
1000 } else {
1001 hrtimer_set_expires(&ts->sched_timer, expires);
1002 tick_program_event(expires, 1);
1003 }
1004}
1005
1006static void tick_nohz_retain_tick(struct tick_sched *ts)
1007{
1008 ts->timer_expires_base = 0;
1009}
1010
1011#ifdef CONFIG_NO_HZ_FULL
1012static void tick_nohz_full_stop_tick(struct tick_sched *ts, int cpu)
1013{
1014 if (tick_nohz_next_event(ts, cpu))
1015 tick_nohz_stop_tick(ts, cpu);
1016 else
1017 tick_nohz_retain_tick(ts);
1018}
1019#endif /* CONFIG_NO_HZ_FULL */
1020
1021static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
1022{
1023 /* Update jiffies first */
1024 tick_do_update_jiffies64(now);
1025
1026 /*
1027 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
1028 * the clock forward checks in the enqueue path:
1029 */
1030 timer_clear_idle();
1031
1032 calc_load_nohz_stop();
1033 touch_softlockup_watchdog_sched();
1034
1035 /* Cancel the scheduled timer and restore the tick: */
1036 tick_sched_flag_clear(ts, TS_FLAG_STOPPED);
1037 tick_nohz_restart(ts, now);
1038}
1039
1040static void __tick_nohz_full_update_tick(struct tick_sched *ts,
1041 ktime_t now)
1042{
1043#ifdef CONFIG_NO_HZ_FULL
1044 int cpu = smp_processor_id();
1045
1046 if (can_stop_full_tick(cpu, ts))
1047 tick_nohz_full_stop_tick(ts, cpu);
1048 else if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1049 tick_nohz_restart_sched_tick(ts, now);
1050#endif
1051}
1052
1053static void tick_nohz_full_update_tick(struct tick_sched *ts)
1054{
1055 if (!tick_nohz_full_cpu(smp_processor_id()))
1056 return;
1057
1058 if (!tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1059 return;
1060
1061 __tick_nohz_full_update_tick(ts, ktime_get());
1062}
1063
1064/*
1065 * A pending softirq outside an IRQ (or softirq disabled section) context
1066 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1067 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1068 *
1069 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1070 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1071 * triggering the code below, since wakep_softirqd() is ignored.
1072 *
1073 */
1074static bool report_idle_softirq(void)
1075{
1076 static int ratelimit;
1077 unsigned int pending = local_softirq_pending();
1078
1079 if (likely(!pending))
1080 return false;
1081
1082 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1083 if (!cpu_active(smp_processor_id())) {
1084 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1085 if (!pending)
1086 return false;
1087 }
1088
1089 /* On RT, softirq handling may be waiting on some lock */
1090 if (local_bh_blocked())
1091 return false;
1092
1093 if (ratelimit < 10) {
1094 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1095 pending);
1096 ratelimit++;
1097 }
1098
1099 return true;
1100}
1101
1102static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1103{
1104 WARN_ON_ONCE(cpu_is_offline(cpu));
1105
1106 if (unlikely(!tick_sched_flag_test(ts, TS_FLAG_NOHZ)))
1107 return false;
1108
1109 if (need_resched())
1110 return false;
1111
1112 if (unlikely(report_idle_softirq()))
1113 return false;
1114
1115 if (tick_nohz_full_enabled()) {
1116 int tick_cpu = READ_ONCE(tick_do_timer_cpu);
1117
1118 /*
1119 * Keep the tick alive to guarantee timekeeping progression
1120 * if there are full dynticks CPUs around
1121 */
1122 if (tick_cpu == cpu)
1123 return false;
1124
1125 /* Should not happen for nohz-full */
1126 if (WARN_ON_ONCE(tick_cpu == TICK_DO_TIMER_NONE))
1127 return false;
1128 }
1129
1130 return true;
1131}
1132
1133/**
1134 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1135 *
1136 * When the next event is more than a tick into the future, stop the idle tick
1137 */
1138void tick_nohz_idle_stop_tick(void)
1139{
1140 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1141 int cpu = smp_processor_id();
1142 ktime_t expires;
1143
1144 /*
1145 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1146 * tick timer expiration time is known already.
1147 */
1148 if (ts->timer_expires_base)
1149 expires = ts->timer_expires;
1150 else if (can_stop_idle_tick(cpu, ts))
1151 expires = tick_nohz_next_event(ts, cpu);
1152 else
1153 return;
1154
1155 ts->idle_calls++;
1156
1157 if (expires > 0LL) {
1158 int was_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
1159
1160 tick_nohz_stop_tick(ts, cpu);
1161
1162 ts->idle_sleeps++;
1163 ts->idle_expires = expires;
1164
1165 if (!was_stopped && tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1166 kcpustat_dyntick_start(ts->idle_entrytime);
1167 nohz_balance_enter_idle(cpu);
1168 }
1169 } else {
1170 tick_nohz_retain_tick(ts);
1171 }
1172}
1173
1174void tick_nohz_idle_retain_tick(void)
1175{
1176 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1177}
1178
1179static void tick_nohz_clock_sleep(struct tick_sched *ts)
1180{
1181 tick_sched_flag_set(ts, TS_FLAG_IDLE_ACTIVE);
1182 sched_clock_idle_sleep_event();
1183}
1184
1185static void tick_nohz_clock_wakeup(struct tick_sched *ts)
1186{
1187 if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE)) {
1188 tick_sched_flag_clear(ts, TS_FLAG_IDLE_ACTIVE);
1189 sched_clock_idle_wakeup_event();
1190 }
1191}
1192
1193/**
1194 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1195 *
1196 * Called when we start the idle loop.
1197 */
1198void tick_nohz_idle_enter(void)
1199{
1200 struct tick_sched *ts;
1201
1202 lockdep_assert_irqs_enabled();
1203
1204 local_irq_disable();
1205
1206 ts = this_cpu_ptr(&tick_cpu_sched);
1207 WARN_ON_ONCE(ts->timer_expires_base);
1208 tick_sched_flag_set(ts, TS_FLAG_INIDLE);
1209 ts->idle_entrytime = ktime_get();
1210 tick_nohz_clock_sleep(ts);
1211
1212 local_irq_enable();
1213}
1214
1215/**
1216 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1217 *
1218 * A timer may have been added/modified/deleted either by the current IRQ,
1219 * or by another place using this IRQ as a notification. This IRQ may have
1220 * also updated the RCU callback list. These events may require a
1221 * re-evaluation of the next tick. Depending on the context:
1222 *
1223 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1224 * time accounting. The next tick will be re-evaluated on the next idle
1225 * loop iteration.
1226 *
1227 * 2) If the CPU is nohz_full:
1228 *
1229 * 2.1) If there is any tick dependency, restart the tick if stopped.
1230 *
1231 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1232 * stop/update it accordingly.
1233 */
1234void tick_nohz_irq_exit(void)
1235{
1236 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1237
1238 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE)) {
1239 tick_nohz_clock_sleep(ts);
1240 ts->idle_entrytime = ktime_get();
1241 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1242 kcpustat_irq_exit(ts->idle_entrytime);
1243 } else {
1244 tick_nohz_full_update_tick(ts);
1245 }
1246}
1247
1248/**
1249 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1250 *
1251 * Return: %true if the tick handler has run, otherwise %false
1252 */
1253bool tick_nohz_idle_got_tick(void)
1254{
1255 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1256
1257 if (ts->got_idle_tick) {
1258 ts->got_idle_tick = 0;
1259 return true;
1260 }
1261 return false;
1262}
1263
1264/**
1265 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1266 * or the tick, whichever expires first. Note that, if the tick has been
1267 * stopped, it returns the next hrtimer.
1268 *
1269 * Called from power state control code with interrupts disabled
1270 *
1271 * Return: the next expiration time
1272 */
1273ktime_t tick_nohz_get_next_hrtimer(void)
1274{
1275 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1276}
1277
1278/**
1279 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1280 * @delta_next: duration until the next event if the tick cannot be stopped
1281 *
1282 * Called from power state control code with interrupts disabled.
1283 *
1284 * The return value of this function and/or the value returned by it through the
1285 * @delta_next pointer can be negative which must be taken into account by its
1286 * callers.
1287 *
1288 * Return: the expected length of the current sleep
1289 */
1290ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1291{
1292 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1293 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1294 int cpu = smp_processor_id();
1295 /*
1296 * The idle entry time is expected to be a sufficient approximation of
1297 * the current time at this point.
1298 */
1299 ktime_t now = ts->idle_entrytime;
1300 ktime_t next_event;
1301
1302 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1303
1304 *delta_next = ktime_sub(dev->next_event, now);
1305
1306 if (!can_stop_idle_tick(cpu, ts))
1307 return *delta_next;
1308
1309 next_event = tick_nohz_next_event(ts, cpu);
1310 if (!next_event)
1311 return *delta_next;
1312
1313 /*
1314 * If the next highres timer to expire is earlier than 'next_event', the
1315 * idle governor needs to know that.
1316 */
1317 next_event = min(next_event, hrtimer_next_event_without(&ts->sched_timer));
1318
1319 return ktime_sub(next_event, now);
1320}
1321
1322/**
1323 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1324 * for a particular CPU.
1325 * @cpu: target CPU number
1326 *
1327 * Called from the schedutil frequency scaling governor in scheduler context.
1328 *
1329 * Return: the current idle calls counter value for @cpu
1330 */
1331unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1332{
1333 struct tick_sched *ts = tick_get_tick_sched(cpu);
1334
1335 return ts->idle_calls;
1336}
1337
1338static void tick_nohz_dyntick_stop(struct tick_sched *ts, ktime_t now)
1339{
1340 tick_nohz_forward_last_tick(ts, now);
1341 kcpustat_dyntick_stop(now, ts->last_tick - TICK_NSEC);
1342}
1343
1344void tick_nohz_idle_restart_tick(void)
1345{
1346 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1347
1348 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1349 /*
1350 * Update entrytime here in case the tick restart is due to temporary
1351 * polling on forced broadcast. The tick may be stopped again later within
1352 * the same idle trip. The idle_entrytime was updated recently but make sure
1353 * no tiny amount of idle time is accounted twice.
1354 */
1355 ts->idle_entrytime = ktime_get();
1356 tick_nohz_dyntick_stop(ts, ts->idle_entrytime);
1357 tick_nohz_restart_sched_tick(ts, ts->idle_entrytime);
1358 }
1359}
1360
1361static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1362{
1363 if (tick_nohz_full_cpu(smp_processor_id()))
1364 __tick_nohz_full_update_tick(ts, now);
1365 else
1366 tick_nohz_restart_sched_tick(ts, now);
1367}
1368
1369/**
1370 * tick_nohz_idle_exit - Update the tick upon idle task exit
1371 *
1372 * When the idle task exits, update the tick depending on the
1373 * following situations:
1374 *
1375 * 1) If the CPU is not in nohz_full mode (most cases), then
1376 * restart the tick.
1377 *
1378 * 2) If the CPU is in nohz_full mode (corner case):
1379 * 2.1) If the tick can be kept stopped (no tick dependencies)
1380 * then re-evaluate the next tick and try to keep it stopped
1381 * as long as possible.
1382 * 2.2) If the tick has dependencies, restart the tick.
1383 *
1384 */
1385void tick_nohz_idle_exit(void)
1386{
1387 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1388 ktime_t now;
1389
1390 local_irq_disable();
1391
1392 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1393 WARN_ON_ONCE(ts->timer_expires_base);
1394
1395 tick_sched_flag_clear(ts, TS_FLAG_INIDLE);
1396 tick_nohz_clock_wakeup(ts);
1397
1398 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1399 now = ktime_get();
1400 tick_nohz_dyntick_stop(ts, now);
1401 tick_nohz_idle_update_tick(ts, now);
1402 }
1403
1404 local_irq_enable();
1405}
1406
1407/*
1408 * In low-resolution mode, the tick handler must be implemented directly
1409 * at the clockevent level. hrtimer can't be used instead, because its
1410 * infrastructure actually relies on the tick itself as a backend in
1411 * low-resolution mode (see hrtimer_run_queues()).
1412 */
1413static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1414{
1415 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1416
1417 dev->next_event = KTIME_MAX;
1418 dev->next_event_forced = 0;
1419
1420 if (likely(tick_nohz_handler(&ts->sched_timer) == HRTIMER_RESTART))
1421 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1422}
1423
1424static inline void tick_nohz_activate(struct tick_sched *ts)
1425{
1426 if (!tick_nohz_enabled)
1427 return;
1428 tick_sched_flag_set(ts, TS_FLAG_NOHZ);
1429 /* One update is enough */
1430 if (!test_and_set_bit(0, &tick_nohz_active))
1431 timers_update_nohz();
1432}
1433
1434/**
1435 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1436 */
1437static void tick_nohz_switch_to_nohz(void)
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 * highres code.
1448 */
1449 tick_setup_sched_timer(false);
1450}
1451
1452static inline void tick_nohz_irq_enter(void)
1453{
1454 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1455 ktime_t now;
1456
1457 tick_nohz_clock_wakeup(ts);
1458
1459 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1460 return;
1461
1462 now = ktime_get();
1463 kcpustat_irq_enter(now);
1464
1465 /*
1466 * If all CPUs are idle we may need to update a stale jiffies value.
1467 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1468 * alive but it might be busy looping with interrupts disabled in some
1469 * rare case (typically stop machine). So we must make sure we have a
1470 * last resort.
1471 */
1472 tick_nohz_update_jiffies(now);
1473}
1474
1475#else
1476
1477static inline void tick_nohz_switch_to_nohz(void) { }
1478static inline void tick_nohz_irq_enter(void) { }
1479static inline void tick_nohz_activate(struct tick_sched *ts) { }
1480
1481#endif /* CONFIG_NO_HZ_COMMON */
1482
1483/*
1484 * Called from irq_enter() to notify about the possible interruption of idle()
1485 */
1486void tick_irq_enter(void)
1487{
1488 tick_check_oneshot_broadcast_this_cpu();
1489 tick_nohz_irq_enter();
1490}
1491
1492static int sched_skew_tick;
1493
1494static int __init skew_tick(char *str)
1495{
1496 get_option(&str, &sched_skew_tick);
1497
1498 return 0;
1499}
1500early_param("skew_tick", skew_tick);
1501
1502/**
1503 * tick_setup_sched_timer - setup the tick emulation timer
1504 * @hrtimer: whether to use the hrtimer or not
1505 */
1506void tick_setup_sched_timer(bool hrtimer)
1507{
1508 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1509
1510 /* Emulate tick processing via per-CPU hrtimers: */
1511 hrtimer_setup(&ts->sched_timer, tick_nohz_handler, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1512
1513 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer)
1514 tick_sched_flag_set(ts, TS_FLAG_HIGHRES);
1515
1516 /* Get the next period (per-CPU) */
1517 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1518
1519 /* Offset the tick to avert 'jiffies_lock' contention. */
1520 if (sched_skew_tick) {
1521 u64 offset = TICK_NSEC >> 1;
1522 do_div(offset, num_possible_cpus());
1523 offset *= smp_processor_id();
1524 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1525 }
1526
1527 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1528 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer)
1529 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1530 else
1531 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1532 tick_nohz_activate(ts);
1533}
1534
1535/*
1536 * Shut down the tick and make sure the CPU won't try to retake the timekeeping
1537 * duty before disabling IRQs in idle for the last time.
1538 */
1539void tick_sched_timer_dying(int cpu)
1540{
1541 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1542 unsigned long idle_calls, idle_sleeps;
1543
1544 /* This must happen before hrtimers are migrated! */
1545 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES))
1546 hrtimer_cancel(&ts->sched_timer);
1547
1548 idle_calls = ts->idle_calls;
1549 idle_sleeps = ts->idle_sleeps;
1550 memset(ts, 0, sizeof(*ts));
1551 ts->idle_calls = idle_calls;
1552 ts->idle_sleeps = idle_sleeps;
1553}
1554
1555/*
1556 * Async notification about clocksource changes
1557 */
1558void tick_clock_notify(void)
1559{
1560 int cpu;
1561
1562 for_each_possible_cpu(cpu)
1563 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1564}
1565
1566/*
1567 * Async notification about clock event changes
1568 */
1569void tick_oneshot_notify(void)
1570{
1571 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1572
1573 set_bit(0, &ts->check_clocks);
1574}
1575
1576/*
1577 * Check if a change happened, which makes oneshot possible.
1578 *
1579 * Called cyclically from the hrtimer softirq (driven by the timer
1580 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1581 * mode, because high resolution timers are disabled (either compile
1582 * or runtime). Called with interrupts disabled.
1583 */
1584int tick_check_oneshot_change(int allow_nohz)
1585{
1586 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1587
1588 if (!test_and_clear_bit(0, &ts->check_clocks))
1589 return 0;
1590
1591 if (tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1592 return 0;
1593
1594 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1595 return 0;
1596
1597 if (!allow_nohz)
1598 return 1;
1599
1600 tick_nohz_switch_to_nohz();
1601 return 0;
1602}