1 // SPDX-License-Identifier: GPL-2.0
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
7 * NOHZ implementation for low and high resolution timers
9 * Started by: Thomas Gleixner and Ingo Molnar
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>
30 #include <asm/irq_regs.h>
32 #include "tick-internal.h"
34 #include <trace/events/timer.h>
37 * Per-CPU nohz control structure
39 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
41 struct tick_sched *tick_get_tick_sched(int cpu)
43 return &per_cpu(tick_cpu_sched, cpu);
46 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
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.
52 static ktime_t last_jiffies_update;
55 * Must be called with interrupts disabled !
57 static void tick_do_update_jiffies64(ktime_t now)
59 unsigned long ticks = 1;
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.
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.
71 if (IS_ENABLED(CONFIG_64BIT)) {
72 if (ktime_before(now, smp_load_acquire(&tick_next_period)))
78 * Avoid contention on 'jiffies_lock' and protect the quick
79 * check with the sequence count.
82 seq = read_seqcount_begin(&jiffies_seq);
83 nextp = tick_next_period;
84 } while (read_seqcount_retry(&jiffies_seq, seq));
86 if (ktime_before(now, nextp))
90 /* Quick check failed, i.e. update is required. */
91 raw_spin_lock(&jiffies_lock);
93 * Re-evaluate with the lock held. Another CPU might have done the
96 if (ktime_before(now, tick_next_period)) {
97 raw_spin_unlock(&jiffies_lock);
101 write_seqcount_begin(&jiffies_seq);
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;
108 ticks += ktime_divns(delta, incr);
110 last_jiffies_update = ktime_add_ns(last_jiffies_update,
113 last_jiffies_update = ktime_add_ns(last_jiffies_update,
117 /* Advance jiffies to complete the 'jiffies_seq' protected job */
120 /* Keep the tick_next_period variable up to date */
121 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC);
123 if (IS_ENABLED(CONFIG_64BIT)) {
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.
130 smp_store_release(&tick_next_period, nextp);
133 * A plain store is good enough on 32-bit, as the quick check
134 * above is protected by the sequence count.
136 tick_next_period = nextp;
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.
144 write_seqcount_end(&jiffies_seq);
148 raw_spin_unlock(&jiffies_lock);
153 * Initialize and return retrieve the jiffies update.
155 static ktime_t tick_init_jiffy_update(void)
159 raw_spin_lock(&jiffies_lock);
160 write_seqcount_begin(&jiffies_seq);
162 /* Have we started the jiffies update yet ? */
163 if (last_jiffies_update == 0) {
167 * Ensure that the tick is aligned to a multiple of
170 div_u64_rem(tick_next_period, TICK_NSEC, &rem);
172 tick_next_period += TICK_NSEC - rem;
174 last_jiffies_update = tick_next_period;
176 period = last_jiffies_update;
178 write_seqcount_end(&jiffies_seq);
179 raw_spin_unlock(&jiffies_lock);
184 #define MAX_STALLED_JIFFIES 5
186 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
188 int cpu = smp_processor_id();
190 #ifdef CONFIG_NO_HZ_COMMON
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
198 * If nohz_full is enabled, this should not happen because the
199 * 'tick_do_timer_cpu' CPU never relinquishes.
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);
205 tick_do_timer_cpu = cpu;
209 /* Check if jiffies need an update */
210 if (tick_do_timer_cpu == cpu)
211 tick_do_update_jiffies64(now);
214 * If the jiffies update stalled for too long (timekeeper in stop_machine()
215 * or VMEXIT'ed for several msecs), force an update.
217 if (ts->last_tick_jiffies != jiffies) {
218 ts->stalled_jiffies = 0;
219 ts->last_tick_jiffies = READ_ONCE(jiffies);
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);
229 ts->got_idle_tick = 1;
232 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
234 #ifdef CONFIG_NO_HZ_COMMON
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.
243 if (ts->tick_stopped) {
244 touch_softlockup_watchdog_sched();
245 if (is_idle_task(current))
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.
255 update_process_times(user_mode(regs));
256 profile_tick(CPU_PROFILING);
260 #ifdef CONFIG_NO_HZ_FULL
261 cpumask_var_t tick_nohz_full_mask;
262 EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
263 bool tick_nohz_full_running;
264 EXPORT_SYMBOL_GPL(tick_nohz_full_running);
265 static atomic_t tick_dep_mask;
267 static bool check_tick_dependency(atomic_t *dep)
269 int val = atomic_read(dep);
271 if (val & TICK_DEP_MASK_POSIX_TIMER) {
272 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
276 if (val & TICK_DEP_MASK_PERF_EVENTS) {
277 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
281 if (val & TICK_DEP_MASK_SCHED) {
282 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
286 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
287 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
291 if (val & TICK_DEP_MASK_RCU) {
292 trace_tick_stop(0, TICK_DEP_MASK_RCU);
296 if (val & TICK_DEP_MASK_RCU_EXP) {
297 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
304 static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
306 lockdep_assert_irqs_disabled();
308 if (unlikely(!cpu_online(cpu)))
311 if (check_tick_dependency(&tick_dep_mask))
314 if (check_tick_dependency(&ts->tick_dep_mask))
317 if (check_tick_dependency(¤t->tick_dep_mask))
320 if (check_tick_dependency(¤t->signal->tick_dep_mask))
326 static void nohz_full_kick_func(struct irq_work *work)
328 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
331 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
332 IRQ_WORK_INIT_HARD(nohz_full_kick_func);
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(),
340 static void tick_nohz_full_kick(void)
342 if (!tick_nohz_full_cpu(smp_processor_id()))
345 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
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.
352 void tick_nohz_full_kick_cpu(int cpu)
354 if (!tick_nohz_full_cpu(cpu))
357 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
360 static void tick_nohz_kick_task(struct task_struct *tsk)
365 * If the task is not running, run_posix_cpu_timers()
366 * has nothing to elapse, and an IPI can then be optimized out.
368 * activate_task() STORE p->tick_dep_mask
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
376 if (!sched_task_on_rq(tsk))
380 * If the task concurrently migrates to another CPU,
381 * we guarantee it sees the new tick dependency upon
384 * set_task_cpu(p, cpu);
385 * STORE p->cpu = @cpu
386 * __schedule() (switch to task 'p')
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
396 tick_nohz_full_kick_cpu(cpu);
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.
404 static void tick_nohz_full_kick_all(void)
408 if (!tick_nohz_full_running)
412 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
413 tick_nohz_full_kick_cpu(cpu);
417 static void tick_nohz_dep_set_all(atomic_t *dep,
418 enum tick_dep_bits bit)
422 prev = atomic_fetch_or(BIT(bit), dep);
424 tick_nohz_full_kick_all();
428 * Set a global tick dependency. Used by perf events that rely on freq and
431 void tick_nohz_dep_set(enum tick_dep_bits bit)
433 tick_nohz_dep_set_all(&tick_dep_mask, bit);
436 void tick_nohz_dep_clear(enum tick_dep_bits bit)
438 atomic_andnot(BIT(bit), &tick_dep_mask);
442 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
443 * manage event-throttling.
445 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
448 struct tick_sched *ts;
450 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
452 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
455 /* Perf needs local kick that is NMI safe */
456 if (cpu == smp_processor_id()) {
457 tick_nohz_full_kick();
459 /* Remote IRQ work not NMI-safe */
460 if (!WARN_ON_ONCE(in_nmi()))
461 tick_nohz_full_kick_cpu(cpu);
466 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
468 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
470 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
472 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
474 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
477 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
478 * in order to elapse per task timers.
480 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
482 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
483 tick_nohz_kick_task(tsk);
485 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
487 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
489 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
491 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
494 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
495 * per process timers.
497 void tick_nohz_dep_set_signal(struct task_struct *tsk,
498 enum tick_dep_bits bit)
501 struct signal_struct *sig = tsk->signal;
503 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
505 struct task_struct *t;
507 lockdep_assert_held(&tsk->sighand->siglock);
508 __for_each_thread(sig, t)
509 tick_nohz_kick_task(t);
513 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
515 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
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, ...
523 void __tick_nohz_task_switch(void)
525 struct tick_sched *ts;
527 if (!tick_nohz_full_cpu(smp_processor_id()))
530 ts = this_cpu_ptr(&tick_cpu_sched);
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();
539 /* Get the boot-time nohz CPU list from the kernel parameters. */
540 void __init tick_nohz_full_setup(cpumask_var_t cpumask)
542 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
543 cpumask_copy(tick_nohz_full_mask, cpumask);
544 tick_nohz_full_running = true;
547 bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
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.
554 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
559 static int tick_nohz_cpu_down(unsigned int cpu)
561 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
564 void __init tick_nohz_init(void)
568 if (!tick_nohz_full_running)
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.
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;
583 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
584 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
585 cpu = smp_processor_id();
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);
594 for_each_cpu(cpu, tick_nohz_full_mask)
595 ct_cpu_track_user(cpu);
597 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
598 "kernel/nohz:predown", NULL,
601 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
602 cpumask_pr_args(tick_nohz_full_mask));
607 * NOHZ - aka dynamic tick functionality
609 #ifdef CONFIG_NO_HZ_COMMON
613 bool tick_nohz_enabled __read_mostly = true;
614 unsigned long tick_nohz_active __read_mostly;
616 * Enable / Disable tickless mode
618 static int __init setup_tick_nohz(char *str)
620 return (kstrtobool(str, &tick_nohz_enabled) == 0);
623 __setup("nohz=", setup_tick_nohz);
625 bool tick_nohz_tick_stopped(void)
627 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
629 return ts->tick_stopped;
632 bool tick_nohz_tick_stopped_cpu(int cpu)
634 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
636 return ts->tick_stopped;
640 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
642 * Called from interrupt entry when the CPU was idle
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.
649 static void tick_nohz_update_jiffies(ktime_t now)
653 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
655 local_irq_save(flags);
656 tick_do_update_jiffies64(now);
657 local_irq_restore(flags);
659 touch_softlockup_watchdog_sched();
662 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
666 if (WARN_ON_ONCE(!ts->idle_active))
669 delta = ktime_sub(now, ts->idle_entrytime);
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);
675 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
677 ts->idle_entrytime = now;
679 write_seqcount_end(&ts->idle_sleeptime_seq);
681 sched_clock_idle_wakeup_event();
684 static void tick_nohz_start_idle(struct tick_sched *ts)
686 write_seqcount_begin(&ts->idle_sleeptime_seq);
687 ts->idle_entrytime = ktime_get();
689 write_seqcount_end(&ts->idle_sleeptime_seq);
691 sched_clock_idle_sleep_event();
694 static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
695 bool compute_delta, u64 *last_update_time)
700 if (!tick_nohz_active)
704 if (last_update_time)
705 *last_update_time = ktime_to_us(now);
708 seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
710 if (ts->idle_active && compute_delta) {
711 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
713 idle = ktime_add(*sleeptime, delta);
717 } while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
719 return ktime_to_us(idle);
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
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.
735 * This time is measured via accounting rather than sampling,
736 * and is as accurate as ktime_get() is.
738 * This function returns -1 if NOHZ is not enabled.
740 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
742 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
744 return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
745 !nr_iowait_cpu(cpu), last_update_time);
747 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
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
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.
761 * This time is measured via accounting rather than sampling,
762 * and is as accurate as ktime_get() is.
764 * This function returns -1 if NOHZ is not enabled.
766 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
768 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
770 return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
771 nr_iowait_cpu(cpu), last_update_time);
773 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
775 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
777 hrtimer_cancel(&ts->sched_timer);
778 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
780 /* Forward the time to expire in the future */
781 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
783 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
784 hrtimer_start_expires(&ts->sched_timer,
785 HRTIMER_MODE_ABS_PINNED_HARD);
787 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
791 * Reset to make sure the next tick stop doesn't get fooled by past
792 * cached clock deadline.
797 static inline bool local_timer_softirq_pending(void)
799 return local_softirq_pending() & BIT(TIMER_SOFTIRQ);
802 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
804 u64 basemono, next_tick, delta, expires;
805 unsigned long basejiff;
808 /* Read jiffies and the time when jiffies were updated last */
810 seq = read_seqcount_begin(&jiffies_seq);
811 basemono = last_jiffies_update;
813 } while (read_seqcount_retry(&jiffies_seq, seq));
814 ts->last_jiffies = basejiff;
815 ts->timer_expires_base = basemono;
818 * Keep the periodic tick, when RCU, architecture or irq_work
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...
827 if (rcu_needs_cpu() || arch_needs_cpu() ||
828 irq_work_needs_cpu() || local_timer_softirq_pending()) {
829 next_tick = basemono + TICK_NSEC;
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
838 next_tick = get_next_timer_interrupt(basejiff, basemono);
839 ts->next_timer = next_tick;
842 /* Make sure next_tick is never before basemono! */
843 if (WARN_ON_ONCE(basemono > next_tick))
844 next_tick = basemono;
847 * If the tick is due in the next period, keep it ticking or
848 * force prod the timer.
850 delta = next_tick - basemono;
851 if (delta <= (u64)TICK_NSEC) {
853 * Tell the timer code that the base is not idle, i.e. undo
854 * the effect of get_next_timer_interrupt():
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.
861 if (!ts->tick_stopped) {
862 ts->timer_expires = 0;
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.
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))
877 /* Calculate the next expiry time */
878 if (delta < (KTIME_MAX - basemono))
879 expires = basemono + delta;
883 ts->timer_expires = min_t(u64, expires, next_tick);
886 return ts->timer_expires;
889 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
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;
895 /* Make sure we won't be trying to stop it twice in a row. */
896 ts->timer_expires_base = 0;
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.
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;
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))
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));
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().
932 if (!ts->tick_stopped) {
933 calc_load_nohz_start();
936 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
937 ts->tick_stopped = 1;
938 trace_tick_stop(1, TICK_DEP_MASK_NONE);
941 ts->next_tick = expires;
944 * If the expiration time == KTIME_MAX, then we simply stop
947 if (unlikely(expires == KTIME_MAX)) {
948 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
949 hrtimer_cancel(&ts->sched_timer);
951 tick_program_event(KTIME_MAX, 1);
955 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
956 hrtimer_start(&ts->sched_timer, expires,
957 HRTIMER_MODE_ABS_PINNED_HARD);
959 hrtimer_set_expires(&ts->sched_timer, expires);
960 tick_program_event(expires, 1);
964 static void tick_nohz_retain_tick(struct tick_sched *ts)
966 ts->timer_expires_base = 0;
969 #ifdef CONFIG_NO_HZ_FULL
970 static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu)
972 if (tick_nohz_next_event(ts, cpu))
973 tick_nohz_stop_tick(ts, cpu);
975 tick_nohz_retain_tick(ts);
977 #endif /* CONFIG_NO_HZ_FULL */
979 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
981 /* Update jiffies first */
982 tick_do_update_jiffies64(now);
985 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
986 * the clock forward checks in the enqueue path:
990 calc_load_nohz_stop();
991 touch_softlockup_watchdog_sched();
993 /* Cancel the scheduled timer and restore the tick: */
994 ts->tick_stopped = 0;
995 tick_nohz_restart(ts, now);
998 static void __tick_nohz_full_update_tick(struct tick_sched *ts,
1001 #ifdef CONFIG_NO_HZ_FULL
1002 int cpu = smp_processor_id();
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);
1011 static void tick_nohz_full_update_tick(struct tick_sched *ts)
1013 if (!tick_nohz_full_cpu(smp_processor_id()))
1016 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
1019 __tick_nohz_full_update_tick(ts, ktime_get());
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().
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.
1032 static bool report_idle_softirq(void)
1034 static int ratelimit;
1035 unsigned int pending = local_softirq_pending();
1037 if (likely(!pending))
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;
1047 if (ratelimit >= 10)
1050 /* On RT, softirq handling may be waiting on some lock */
1051 if (local_bh_blocked())
1054 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1061 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
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
1070 if (unlikely(!cpu_online(cpu))) {
1071 if (cpu == tick_do_timer_cpu)
1072 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
1074 * Make sure the CPU doesn't get fooled by obsolete tick
1075 * deadline if it comes back online later.
1081 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
1087 if (unlikely(report_idle_softirq()))
1090 if (tick_nohz_full_enabled()) {
1092 * Keep the tick alive to guarantee timekeeping progression
1093 * if there are full dynticks CPUs around
1095 if (tick_do_timer_cpu == cpu)
1098 /* Should not happen for nohz-full */
1099 if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
1107 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1109 * When the next event is more than a tick into the future, stop the idle tick
1111 void tick_nohz_idle_stop_tick(void)
1113 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1114 int cpu = smp_processor_id();
1118 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1119 * tick timer expiration time is known already.
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);
1130 if (expires > 0LL) {
1131 int was_stopped = ts->tick_stopped;
1133 tick_nohz_stop_tick(ts, cpu);
1136 ts->idle_expires = expires;
1138 if (!was_stopped && ts->tick_stopped) {
1139 ts->idle_jiffies = ts->last_jiffies;
1140 nohz_balance_enter_idle(cpu);
1143 tick_nohz_retain_tick(ts);
1147 void tick_nohz_idle_retain_tick(void)
1149 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1151 * Undo the effect of get_next_timer_interrupt() called from
1152 * tick_nohz_next_event().
1158 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1160 * Called when we start the idle loop.
1162 void tick_nohz_idle_enter(void)
1164 struct tick_sched *ts;
1166 lockdep_assert_irqs_enabled();
1168 local_irq_disable();
1170 ts = this_cpu_ptr(&tick_cpu_sched);
1172 WARN_ON_ONCE(ts->timer_expires_base);
1175 tick_nohz_start_idle(ts);
1181 * tick_nohz_irq_exit - Notify the tick about IRQ exit
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:
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
1192 * 2) If the CPU is nohz_full:
1194 * 2.1) If there is any tick dependency, restart the tick if stopped.
1196 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and
1197 * stop/update it accordingly.
1199 void tick_nohz_irq_exit(void)
1201 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1204 tick_nohz_start_idle(ts);
1206 tick_nohz_full_update_tick(ts);
1210 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1212 bool tick_nohz_idle_got_tick(void)
1214 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1216 if (ts->got_idle_tick) {
1217 ts->got_idle_tick = 0;
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.
1228 * Called from power state control code with interrupts disabled
1230 ktime_t tick_nohz_get_next_hrtimer(void)
1232 return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
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
1239 * Called from power state control code with interrupts disabled.
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
1245 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
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();
1251 * The idle entry time is expected to be a sufficient approximation of
1252 * the current time at this point.
1254 ktime_t now = ts->idle_entrytime;
1257 WARN_ON_ONCE(!ts->inidle);
1259 *delta_next = ktime_sub(dev->next_event, now);
1261 if (!can_stop_idle_tick(cpu, ts))
1264 next_event = tick_nohz_next_event(ts, cpu);
1269 * If the next highres timer to expire is earlier than 'next_event', the
1270 * idle governor needs to know that.
1272 next_event = min_t(u64, next_event,
1273 hrtimer_next_event_without(&ts->sched_timer));
1275 return ktime_sub(next_event, now);
1279 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1280 * for a particular CPU.
1282 * Called from the schedutil frequency scaling governor in scheduler context.
1284 unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1286 struct tick_sched *ts = tick_get_tick_sched(cpu);
1288 return ts->idle_calls;
1292 * tick_nohz_get_idle_calls - return the current idle calls counter value
1294 * Called from the schedutil frequency scaling governor in scheduler context.
1296 unsigned long tick_nohz_get_idle_calls(void)
1298 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1300 return ts->idle_calls;
1303 static void tick_nohz_account_idle_time(struct tick_sched *ts,
1306 unsigned long ticks;
1308 ts->idle_exittime = now;
1310 if (vtime_accounting_enabled_this_cpu())
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 !
1317 ticks = jiffies - ts->idle_jiffies;
1319 * We might be one off. Do not randomly account a huge number of ticks!
1321 if (ticks && ticks < LONG_MAX)
1322 account_idle_ticks(ticks);
1325 void tick_nohz_idle_restart_tick(void)
1327 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
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);
1336 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1338 if (tick_nohz_full_cpu(smp_processor_id()))
1339 __tick_nohz_full_update_tick(ts, now);
1341 tick_nohz_restart_sched_tick(ts, now);
1343 tick_nohz_account_idle_time(ts, now);
1347 * tick_nohz_idle_exit - Update the tick upon idle task exit
1349 * When the idle task exits, update the tick depending on the
1350 * following situations:
1352 * 1) If the CPU is not in nohz_full mode (most cases), then
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.
1362 void tick_nohz_idle_exit(void)
1364 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1365 bool idle_active, tick_stopped;
1368 local_irq_disable();
1370 WARN_ON_ONCE(!ts->inidle);
1371 WARN_ON_ONCE(ts->timer_expires_base);
1374 idle_active = ts->idle_active;
1375 tick_stopped = ts->tick_stopped;
1377 if (idle_active || tick_stopped)
1381 tick_nohz_stop_idle(ts, now);
1384 tick_nohz_idle_update_tick(ts, now);
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()).
1395 * This low-resolution handler still makes use of some hrtimer APIs meanwhile
1396 * for convenience with expiration calculation and forwarding.
1398 static void tick_nohz_lowres_handler(struct clock_event_device *dev)
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();
1404 dev->next_event = KTIME_MAX;
1406 tick_sched_do_timer(ts, now);
1407 tick_sched_handle(ts, regs);
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.
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);
1421 static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1423 if (!tick_nohz_enabled)
1425 ts->nohz_mode = mode;
1426 /* One update is enough */
1427 if (!test_and_set_bit(0, &tick_nohz_active))
1428 timers_update_nohz();
1432 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1434 static void tick_nohz_switch_to_nohz(void)
1436 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1439 if (!tick_nohz_enabled)
1442 if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1446 * Recycle the hrtimer in 'ts', so we can share the
1447 * hrtimer_forward_now() function with the highres code.
1449 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1450 /* Get the next period */
1451 next = tick_init_jiffy_update();
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);
1459 static inline void tick_nohz_irq_enter(void)
1461 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1464 if (!ts->idle_active && !ts->tick_stopped)
1467 if (ts->idle_active)
1468 tick_nohz_stop_idle(ts, now);
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
1476 if (ts->tick_stopped)
1477 tick_nohz_update_jiffies(now);
1482 static inline void tick_nohz_switch_to_nohz(void) { }
1483 static inline void tick_nohz_irq_enter(void) { }
1484 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1486 #endif /* CONFIG_NO_HZ_COMMON */
1489 * Called from irq_enter() to notify about the possible interruption of idle()
1491 void tick_irq_enter(void)
1493 tick_check_oneshot_broadcast_this_cpu();
1494 tick_nohz_irq_enter();
1498 * High resolution timer specific code
1500 #ifdef CONFIG_HIGH_RES_TIMERS
1502 * We rearm the timer until we get disabled by the idle code.
1503 * Called with interrupts disabled.
1505 static enum hrtimer_restart tick_nohz_highres_handler(struct hrtimer *timer)
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();
1512 tick_sched_do_timer(ts, now);
1515 * Do not call when we are not in IRQ context and have
1516 * no valid 'regs' pointer
1519 tick_sched_handle(ts, regs);
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.
1528 if (unlikely(ts->tick_stopped))
1529 return HRTIMER_NORESTART;
1531 hrtimer_forward(timer, now, TICK_NSEC);
1533 return HRTIMER_RESTART;
1536 static int sched_skew_tick;
1538 static int __init skew_tick(char *str)
1540 get_option(&str, &sched_skew_tick);
1544 early_param("skew_tick", skew_tick);
1547 * tick_setup_sched_timer - setup the tick emulation timer
1549 void tick_setup_sched_timer(void)
1551 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1552 ktime_t now = ktime_get();
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;
1558 /* Get the next period (per-CPU) */
1559 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
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);
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);
1573 #endif /* HIGH_RES_TIMERS */
1575 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1576 void tick_cancel_sched_timer(int cpu)
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;
1582 # ifdef CONFIG_HIGH_RES_TIMERS
1583 if (ts->sched_timer.base)
1584 hrtimer_cancel(&ts->sched_timer);
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;
1600 * Async notification about clocksource changes
1602 void tick_clock_notify(void)
1606 for_each_possible_cpu(cpu)
1607 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1611 * Async notification about clock event changes
1613 void tick_oneshot_notify(void)
1615 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1617 set_bit(0, &ts->check_clocks);
1621 * Check if a change happened, which makes oneshot possible.
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.
1628 int tick_check_oneshot_change(int allow_nohz)
1630 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1632 if (!test_and_clear_bit(0, &ts->check_clocks))
1635 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1638 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1644 tick_nohz_switch_to_nohz();