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