Loading...
1/*
2 * linux/kernel/time/tick-sched.c
3 *
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 *
8 * No idle tick implementation for low and high resolution timers
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
12 * Distribute under GPLv2.
13 */
14#include <linux/cpu.h>
15#include <linux/err.h>
16#include <linux/hrtimer.h>
17#include <linux/interrupt.h>
18#include <linux/kernel_stat.h>
19#include <linux/percpu.h>
20#include <linux/profile.h>
21#include <linux/sched.h>
22#include <linux/module.h>
23#include <linux/irq_work.h>
24#include <linux/posix-timers.h>
25#include <linux/context_tracking.h>
26
27#include <asm/irq_regs.h>
28
29#include "tick-internal.h"
30
31#include <trace/events/timer.h>
32
33/*
34 * Per cpu nohz control structure
35 */
36static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
37
38struct tick_sched *tick_get_tick_sched(int cpu)
39{
40 return &per_cpu(tick_cpu_sched, cpu);
41}
42
43#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
44/*
45 * The time, when the last jiffy update happened. Protected by jiffies_lock.
46 */
47static ktime_t last_jiffies_update;
48
49/*
50 * Must be called with interrupts disabled !
51 */
52static void tick_do_update_jiffies64(ktime_t now)
53{
54 unsigned long ticks = 0;
55 ktime_t delta;
56
57 /*
58 * Do a quick check without holding jiffies_lock:
59 */
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta.tv64 < tick_period.tv64)
62 return;
63
64 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
66
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
69
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
72 tick_period);
73
74 /* Slow path for long timeouts */
75 if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 s64 incr = ktime_to_ns(tick_period);
77
78 ticks = ktime_divns(delta, incr);
79
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
81 incr * ticks);
82 }
83 do_timer(++ticks);
84
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
87 } else {
88 write_sequnlock(&jiffies_lock);
89 return;
90 }
91 write_sequnlock(&jiffies_lock);
92 update_wall_time();
93}
94
95/*
96 * Initialize and return retrieve the jiffies update.
97 */
98static ktime_t tick_init_jiffy_update(void)
99{
100 ktime_t period;
101
102 write_seqlock(&jiffies_lock);
103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update.tv64 == 0)
105 last_jiffies_update = tick_next_period;
106 period = last_jiffies_update;
107 write_sequnlock(&jiffies_lock);
108 return period;
109}
110
111
112static void tick_sched_do_timer(ktime_t now)
113{
114 int cpu = smp_processor_id();
115
116#ifdef CONFIG_NO_HZ_COMMON
117 /*
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the cpu in charge went
120 * into a long sleep. If two cpus happen to assign themself to
121 * this duty, then the jiffies update is still serialized by
122 * jiffies_lock.
123 */
124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
125 && !tick_nohz_full_cpu(cpu))
126 tick_do_timer_cpu = cpu;
127#endif
128
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu == cpu)
131 tick_do_update_jiffies64(now);
132}
133
134static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
135{
136#ifdef CONFIG_NO_HZ_COMMON
137 /*
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
144 */
145 if (ts->tick_stopped) {
146 touch_softlockup_watchdog_sched();
147 if (is_idle_task(current))
148 ts->idle_jiffies++;
149 }
150#endif
151 update_process_times(user_mode(regs));
152 profile_tick(CPU_PROFILING);
153}
154#endif
155
156#ifdef CONFIG_NO_HZ_FULL
157cpumask_var_t tick_nohz_full_mask;
158cpumask_var_t housekeeping_mask;
159bool tick_nohz_full_running;
160static atomic_t tick_dep_mask;
161
162static bool check_tick_dependency(atomic_t *dep)
163{
164 int val = atomic_read(dep);
165
166 if (val & TICK_DEP_MASK_POSIX_TIMER) {
167 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
168 return true;
169 }
170
171 if (val & TICK_DEP_MASK_PERF_EVENTS) {
172 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
173 return true;
174 }
175
176 if (val & TICK_DEP_MASK_SCHED) {
177 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
178 return true;
179 }
180
181 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
182 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
183 return true;
184 }
185
186 return false;
187}
188
189static bool can_stop_full_tick(struct tick_sched *ts)
190{
191 WARN_ON_ONCE(!irqs_disabled());
192
193 if (check_tick_dependency(&tick_dep_mask))
194 return false;
195
196 if (check_tick_dependency(&ts->tick_dep_mask))
197 return false;
198
199 if (check_tick_dependency(¤t->tick_dep_mask))
200 return false;
201
202 if (check_tick_dependency(¤t->signal->tick_dep_mask))
203 return false;
204
205 return true;
206}
207
208static void nohz_full_kick_func(struct irq_work *work)
209{
210 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
211}
212
213static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
214 .func = nohz_full_kick_func,
215};
216
217/*
218 * Kick this CPU if it's full dynticks in order to force it to
219 * re-evaluate its dependency on the tick and restart it if necessary.
220 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
221 * is NMI safe.
222 */
223static void tick_nohz_full_kick(void)
224{
225 if (!tick_nohz_full_cpu(smp_processor_id()))
226 return;
227
228 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
229}
230
231/*
232 * Kick the CPU if it's full dynticks in order to force it to
233 * re-evaluate its dependency on the tick and restart it if necessary.
234 */
235void tick_nohz_full_kick_cpu(int cpu)
236{
237 if (!tick_nohz_full_cpu(cpu))
238 return;
239
240 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
241}
242
243/*
244 * Kick all full dynticks CPUs in order to force these to re-evaluate
245 * their dependency on the tick and restart it if necessary.
246 */
247static void tick_nohz_full_kick_all(void)
248{
249 int cpu;
250
251 if (!tick_nohz_full_running)
252 return;
253
254 preempt_disable();
255 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
256 tick_nohz_full_kick_cpu(cpu);
257 preempt_enable();
258}
259
260static void tick_nohz_dep_set_all(atomic_t *dep,
261 enum tick_dep_bits bit)
262{
263 int prev;
264
265 prev = atomic_fetch_or(dep, BIT(bit));
266 if (!prev)
267 tick_nohz_full_kick_all();
268}
269
270/*
271 * Set a global tick dependency. Used by perf events that rely on freq and
272 * by unstable clock.
273 */
274void tick_nohz_dep_set(enum tick_dep_bits bit)
275{
276 tick_nohz_dep_set_all(&tick_dep_mask, bit);
277}
278
279void tick_nohz_dep_clear(enum tick_dep_bits bit)
280{
281 atomic_andnot(BIT(bit), &tick_dep_mask);
282}
283
284/*
285 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
286 * manage events throttling.
287 */
288void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
289{
290 int prev;
291 struct tick_sched *ts;
292
293 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
294
295 prev = atomic_fetch_or(&ts->tick_dep_mask, BIT(bit));
296 if (!prev) {
297 preempt_disable();
298 /* Perf needs local kick that is NMI safe */
299 if (cpu == smp_processor_id()) {
300 tick_nohz_full_kick();
301 } else {
302 /* Remote irq work not NMI-safe */
303 if (!WARN_ON_ONCE(in_nmi()))
304 tick_nohz_full_kick_cpu(cpu);
305 }
306 preempt_enable();
307 }
308}
309
310void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
311{
312 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
313
314 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
315}
316
317/*
318 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
319 * per task timers.
320 */
321void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
322{
323 /*
324 * We could optimize this with just kicking the target running the task
325 * if that noise matters for nohz full users.
326 */
327 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
328}
329
330void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
331{
332 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
333}
334
335/*
336 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
337 * per process timers.
338 */
339void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
340{
341 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
342}
343
344void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
345{
346 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
347}
348
349/*
350 * Re-evaluate the need for the tick as we switch the current task.
351 * It might need the tick due to per task/process properties:
352 * perf events, posix cpu timers, ...
353 */
354void __tick_nohz_task_switch(void)
355{
356 unsigned long flags;
357 struct tick_sched *ts;
358
359 local_irq_save(flags);
360
361 if (!tick_nohz_full_cpu(smp_processor_id()))
362 goto out;
363
364 ts = this_cpu_ptr(&tick_cpu_sched);
365
366 if (ts->tick_stopped) {
367 if (atomic_read(¤t->tick_dep_mask) ||
368 atomic_read(¤t->signal->tick_dep_mask))
369 tick_nohz_full_kick();
370 }
371out:
372 local_irq_restore(flags);
373}
374
375/* Parse the boot-time nohz CPU list from the kernel parameters. */
376static int __init tick_nohz_full_setup(char *str)
377{
378 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
379 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
380 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
381 free_bootmem_cpumask_var(tick_nohz_full_mask);
382 return 1;
383 }
384 tick_nohz_full_running = true;
385
386 return 1;
387}
388__setup("nohz_full=", tick_nohz_full_setup);
389
390static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
391 unsigned long action,
392 void *hcpu)
393{
394 unsigned int cpu = (unsigned long)hcpu;
395
396 switch (action & ~CPU_TASKS_FROZEN) {
397 case CPU_DOWN_PREPARE:
398 /*
399 * The boot CPU handles housekeeping duty (unbound timers,
400 * workqueues, timekeeping, ...) on behalf of full dynticks
401 * CPUs. It must remain online when nohz full is enabled.
402 */
403 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
404 return NOTIFY_BAD;
405 break;
406 }
407 return NOTIFY_OK;
408}
409
410static int tick_nohz_init_all(void)
411{
412 int err = -1;
413
414#ifdef CONFIG_NO_HZ_FULL_ALL
415 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
416 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
417 return err;
418 }
419 err = 0;
420 cpumask_setall(tick_nohz_full_mask);
421 tick_nohz_full_running = true;
422#endif
423 return err;
424}
425
426void __init tick_nohz_init(void)
427{
428 int cpu;
429
430 if (!tick_nohz_full_running) {
431 if (tick_nohz_init_all() < 0)
432 return;
433 }
434
435 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
436 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
437 cpumask_clear(tick_nohz_full_mask);
438 tick_nohz_full_running = false;
439 return;
440 }
441
442 /*
443 * Full dynticks uses irq work to drive the tick rescheduling on safe
444 * locking contexts. But then we need irq work to raise its own
445 * interrupts to avoid circular dependency on the tick
446 */
447 if (!arch_irq_work_has_interrupt()) {
448 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
449 cpumask_clear(tick_nohz_full_mask);
450 cpumask_copy(housekeeping_mask, cpu_possible_mask);
451 tick_nohz_full_running = false;
452 return;
453 }
454
455 cpu = smp_processor_id();
456
457 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
458 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
459 cpu);
460 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
461 }
462
463 cpumask_andnot(housekeeping_mask,
464 cpu_possible_mask, tick_nohz_full_mask);
465
466 for_each_cpu(cpu, tick_nohz_full_mask)
467 context_tracking_cpu_set(cpu);
468
469 cpu_notifier(tick_nohz_cpu_down_callback, 0);
470 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
471 cpumask_pr_args(tick_nohz_full_mask));
472
473 /*
474 * We need at least one CPU to handle housekeeping work such
475 * as timekeeping, unbound timers, workqueues, ...
476 */
477 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
478}
479#endif
480
481/*
482 * NOHZ - aka dynamic tick functionality
483 */
484#ifdef CONFIG_NO_HZ_COMMON
485/*
486 * NO HZ enabled ?
487 */
488bool tick_nohz_enabled __read_mostly = true;
489unsigned long tick_nohz_active __read_mostly;
490/*
491 * Enable / Disable tickless mode
492 */
493static int __init setup_tick_nohz(char *str)
494{
495 return (kstrtobool(str, &tick_nohz_enabled) == 0);
496}
497
498__setup("nohz=", setup_tick_nohz);
499
500int tick_nohz_tick_stopped(void)
501{
502 return __this_cpu_read(tick_cpu_sched.tick_stopped);
503}
504
505/**
506 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
507 *
508 * Called from interrupt entry when the CPU was idle
509 *
510 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
511 * must be updated. Otherwise an interrupt handler could use a stale jiffy
512 * value. We do this unconditionally on any cpu, as we don't know whether the
513 * cpu, which has the update task assigned is in a long sleep.
514 */
515static void tick_nohz_update_jiffies(ktime_t now)
516{
517 unsigned long flags;
518
519 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
520
521 local_irq_save(flags);
522 tick_do_update_jiffies64(now);
523 local_irq_restore(flags);
524
525 touch_softlockup_watchdog_sched();
526}
527
528/*
529 * Updates the per cpu time idle statistics counters
530 */
531static void
532update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
533{
534 ktime_t delta;
535
536 if (ts->idle_active) {
537 delta = ktime_sub(now, ts->idle_entrytime);
538 if (nr_iowait_cpu(cpu) > 0)
539 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
540 else
541 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
542 ts->idle_entrytime = now;
543 }
544
545 if (last_update_time)
546 *last_update_time = ktime_to_us(now);
547
548}
549
550static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
551{
552 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
553 ts->idle_active = 0;
554
555 sched_clock_idle_wakeup_event(0);
556}
557
558static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
559{
560 ktime_t now = ktime_get();
561
562 ts->idle_entrytime = now;
563 ts->idle_active = 1;
564 sched_clock_idle_sleep_event();
565 return now;
566}
567
568/**
569 * get_cpu_idle_time_us - get the total idle time of a cpu
570 * @cpu: CPU number to query
571 * @last_update_time: variable to store update time in. Do not update
572 * counters if NULL.
573 *
574 * Return the cummulative idle time (since boot) for a given
575 * CPU, in microseconds.
576 *
577 * This time is measured via accounting rather than sampling,
578 * and is as accurate as ktime_get() is.
579 *
580 * This function returns -1 if NOHZ is not enabled.
581 */
582u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
583{
584 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
585 ktime_t now, idle;
586
587 if (!tick_nohz_active)
588 return -1;
589
590 now = ktime_get();
591 if (last_update_time) {
592 update_ts_time_stats(cpu, ts, now, last_update_time);
593 idle = ts->idle_sleeptime;
594 } else {
595 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
596 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
597
598 idle = ktime_add(ts->idle_sleeptime, delta);
599 } else {
600 idle = ts->idle_sleeptime;
601 }
602 }
603
604 return ktime_to_us(idle);
605
606}
607EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
608
609/**
610 * get_cpu_iowait_time_us - get the total iowait time of a cpu
611 * @cpu: CPU number to query
612 * @last_update_time: variable to store update time in. Do not update
613 * counters if NULL.
614 *
615 * Return the cummulative iowait time (since boot) for a given
616 * CPU, in microseconds.
617 *
618 * This time is measured via accounting rather than sampling,
619 * and is as accurate as ktime_get() is.
620 *
621 * This function returns -1 if NOHZ is not enabled.
622 */
623u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
624{
625 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
626 ktime_t now, iowait;
627
628 if (!tick_nohz_active)
629 return -1;
630
631 now = ktime_get();
632 if (last_update_time) {
633 update_ts_time_stats(cpu, ts, now, last_update_time);
634 iowait = ts->iowait_sleeptime;
635 } else {
636 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
637 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
638
639 iowait = ktime_add(ts->iowait_sleeptime, delta);
640 } else {
641 iowait = ts->iowait_sleeptime;
642 }
643 }
644
645 return ktime_to_us(iowait);
646}
647EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
648
649static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
650{
651 hrtimer_cancel(&ts->sched_timer);
652 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
653
654 /* Forward the time to expire in the future */
655 hrtimer_forward(&ts->sched_timer, now, tick_period);
656
657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
658 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
659 else
660 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
661}
662
663static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
664 ktime_t now, int cpu)
665{
666 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
667 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
668 unsigned long seq, basejiff;
669 ktime_t tick;
670
671 /* Read jiffies and the time when jiffies were updated last */
672 do {
673 seq = read_seqbegin(&jiffies_lock);
674 basemono = last_jiffies_update.tv64;
675 basejiff = jiffies;
676 } while (read_seqretry(&jiffies_lock, seq));
677 ts->last_jiffies = basejiff;
678
679 if (rcu_needs_cpu(basemono, &next_rcu) ||
680 arch_needs_cpu() || irq_work_needs_cpu()) {
681 next_tick = basemono + TICK_NSEC;
682 } else {
683 /*
684 * Get the next pending timer. If high resolution
685 * timers are enabled this only takes the timer wheel
686 * timers into account. If high resolution timers are
687 * disabled this also looks at the next expiring
688 * hrtimer.
689 */
690 next_tmr = get_next_timer_interrupt(basejiff, basemono);
691 ts->next_timer = next_tmr;
692 /* Take the next rcu event into account */
693 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
694 }
695
696 /*
697 * If the tick is due in the next period, keep it ticking or
698 * force prod the timer.
699 */
700 delta = next_tick - basemono;
701 if (delta <= (u64)TICK_NSEC) {
702 tick.tv64 = 0;
703 /*
704 * We've not stopped the tick yet, and there's a timer in the
705 * next period, so no point in stopping it either, bail.
706 */
707 if (!ts->tick_stopped)
708 goto out;
709
710 /*
711 * If, OTOH, we did stop it, but there's a pending (expired)
712 * timer reprogram the timer hardware to fire now.
713 *
714 * We will not restart the tick proper, just prod the timer
715 * hardware into firing an interrupt to process the pending
716 * timers. Just like tick_irq_exit() will not restart the tick
717 * for 'normal' interrupts.
718 *
719 * Only once we exit the idle loop will we re-enable the tick,
720 * see tick_nohz_idle_exit().
721 */
722 if (delta == 0) {
723 tick_nohz_restart(ts, now);
724 goto out;
725 }
726 }
727
728 /*
729 * If this cpu is the one which updates jiffies, then give up
730 * the assignment and let it be taken by the cpu which runs
731 * the tick timer next, which might be this cpu as well. If we
732 * don't drop this here the jiffies might be stale and
733 * do_timer() never invoked. Keep track of the fact that it
734 * was the one which had the do_timer() duty last. If this cpu
735 * is the one which had the do_timer() duty last, we limit the
736 * sleep time to the timekeeping max_deferement value.
737 * Otherwise we can sleep as long as we want.
738 */
739 delta = timekeeping_max_deferment();
740 if (cpu == tick_do_timer_cpu) {
741 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
742 ts->do_timer_last = 1;
743 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
744 delta = KTIME_MAX;
745 ts->do_timer_last = 0;
746 } else if (!ts->do_timer_last) {
747 delta = KTIME_MAX;
748 }
749
750#ifdef CONFIG_NO_HZ_FULL
751 /* Limit the tick delta to the maximum scheduler deferment */
752 if (!ts->inidle)
753 delta = min(delta, scheduler_tick_max_deferment());
754#endif
755
756 /* Calculate the next expiry time */
757 if (delta < (KTIME_MAX - basemono))
758 expires = basemono + delta;
759 else
760 expires = KTIME_MAX;
761
762 expires = min_t(u64, expires, next_tick);
763 tick.tv64 = expires;
764
765 /* Skip reprogram of event if its not changed */
766 if (ts->tick_stopped && (expires == dev->next_event.tv64))
767 goto out;
768
769 /*
770 * nohz_stop_sched_tick can be called several times before
771 * the nohz_restart_sched_tick is called. This happens when
772 * interrupts arrive which do not cause a reschedule. In the
773 * first call we save the current tick time, so we can restart
774 * the scheduler tick in nohz_restart_sched_tick.
775 */
776 if (!ts->tick_stopped) {
777 nohz_balance_enter_idle(cpu);
778 calc_load_enter_idle();
779
780 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
781 ts->tick_stopped = 1;
782 trace_tick_stop(1, TICK_DEP_MASK_NONE);
783 }
784
785 /*
786 * If the expiration time == KTIME_MAX, then we simply stop
787 * the tick timer.
788 */
789 if (unlikely(expires == KTIME_MAX)) {
790 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
791 hrtimer_cancel(&ts->sched_timer);
792 goto out;
793 }
794
795 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
796 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
797 else
798 tick_program_event(tick, 1);
799out:
800 /* Update the estimated sleep length */
801 ts->sleep_length = ktime_sub(dev->next_event, now);
802 return tick;
803}
804
805static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active)
806{
807 /* Update jiffies first */
808 tick_do_update_jiffies64(now);
809 update_cpu_load_nohz(active);
810
811 calc_load_exit_idle();
812 touch_softlockup_watchdog_sched();
813 /*
814 * Cancel the scheduled timer and restore the tick
815 */
816 ts->tick_stopped = 0;
817 ts->idle_exittime = now;
818
819 tick_nohz_restart(ts, now);
820}
821
822static void tick_nohz_full_update_tick(struct tick_sched *ts)
823{
824#ifdef CONFIG_NO_HZ_FULL
825 int cpu = smp_processor_id();
826
827 if (!tick_nohz_full_cpu(cpu))
828 return;
829
830 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
831 return;
832
833 if (can_stop_full_tick(ts))
834 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
835 else if (ts->tick_stopped)
836 tick_nohz_restart_sched_tick(ts, ktime_get(), 1);
837#endif
838}
839
840static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
841{
842 /*
843 * If this cpu is offline and it is the one which updates
844 * jiffies, then give up the assignment and let it be taken by
845 * the cpu which runs the tick timer next. If we don't drop
846 * this here the jiffies might be stale and do_timer() never
847 * invoked.
848 */
849 if (unlikely(!cpu_online(cpu))) {
850 if (cpu == tick_do_timer_cpu)
851 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
852 return false;
853 }
854
855 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
856 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
857 return false;
858 }
859
860 if (need_resched())
861 return false;
862
863 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
864 static int ratelimit;
865
866 if (ratelimit < 10 &&
867 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
868 pr_warn("NOHZ: local_softirq_pending %02x\n",
869 (unsigned int) local_softirq_pending());
870 ratelimit++;
871 }
872 return false;
873 }
874
875 if (tick_nohz_full_enabled()) {
876 /*
877 * Keep the tick alive to guarantee timekeeping progression
878 * if there are full dynticks CPUs around
879 */
880 if (tick_do_timer_cpu == cpu)
881 return false;
882 /*
883 * Boot safety: make sure the timekeeping duty has been
884 * assigned before entering dyntick-idle mode,
885 */
886 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
887 return false;
888 }
889
890 return true;
891}
892
893static void __tick_nohz_idle_enter(struct tick_sched *ts)
894{
895 ktime_t now, expires;
896 int cpu = smp_processor_id();
897
898 now = tick_nohz_start_idle(ts);
899
900 if (can_stop_idle_tick(cpu, ts)) {
901 int was_stopped = ts->tick_stopped;
902
903 ts->idle_calls++;
904
905 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
906 if (expires.tv64 > 0LL) {
907 ts->idle_sleeps++;
908 ts->idle_expires = expires;
909 }
910
911 if (!was_stopped && ts->tick_stopped)
912 ts->idle_jiffies = ts->last_jiffies;
913 }
914}
915
916/**
917 * tick_nohz_idle_enter - stop the idle tick from the idle task
918 *
919 * When the next event is more than a tick into the future, stop the idle tick
920 * Called when we start the idle loop.
921 *
922 * The arch is responsible of calling:
923 *
924 * - rcu_idle_enter() after its last use of RCU before the CPU is put
925 * to sleep.
926 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
927 */
928void tick_nohz_idle_enter(void)
929{
930 struct tick_sched *ts;
931
932 WARN_ON_ONCE(irqs_disabled());
933
934 /*
935 * Update the idle state in the scheduler domain hierarchy
936 * when tick_nohz_stop_sched_tick() is called from the idle loop.
937 * State will be updated to busy during the first busy tick after
938 * exiting idle.
939 */
940 set_cpu_sd_state_idle();
941
942 local_irq_disable();
943
944 ts = this_cpu_ptr(&tick_cpu_sched);
945 ts->inidle = 1;
946 __tick_nohz_idle_enter(ts);
947
948 local_irq_enable();
949}
950
951/**
952 * tick_nohz_irq_exit - update next tick event from interrupt exit
953 *
954 * When an interrupt fires while we are idle and it doesn't cause
955 * a reschedule, it may still add, modify or delete a timer, enqueue
956 * an RCU callback, etc...
957 * So we need to re-calculate and reprogram the next tick event.
958 */
959void tick_nohz_irq_exit(void)
960{
961 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
962
963 if (ts->inidle)
964 __tick_nohz_idle_enter(ts);
965 else
966 tick_nohz_full_update_tick(ts);
967}
968
969/**
970 * tick_nohz_get_sleep_length - return the length of the current sleep
971 *
972 * Called from power state control code with interrupts disabled
973 */
974ktime_t tick_nohz_get_sleep_length(void)
975{
976 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
977
978 return ts->sleep_length;
979}
980
981static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
982{
983#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
984 unsigned long ticks;
985
986 if (vtime_accounting_cpu_enabled())
987 return;
988 /*
989 * We stopped the tick in idle. Update process times would miss the
990 * time we slept as update_process_times does only a 1 tick
991 * accounting. Enforce that this is accounted to idle !
992 */
993 ticks = jiffies - ts->idle_jiffies;
994 /*
995 * We might be one off. Do not randomly account a huge number of ticks!
996 */
997 if (ticks && ticks < LONG_MAX)
998 account_idle_ticks(ticks);
999#endif
1000}
1001
1002/**
1003 * tick_nohz_idle_exit - restart the idle tick from the idle task
1004 *
1005 * Restart the idle tick when the CPU is woken up from idle
1006 * This also exit the RCU extended quiescent state. The CPU
1007 * can use RCU again after this function is called.
1008 */
1009void tick_nohz_idle_exit(void)
1010{
1011 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1012 ktime_t now;
1013
1014 local_irq_disable();
1015
1016 WARN_ON_ONCE(!ts->inidle);
1017
1018 ts->inidle = 0;
1019
1020 if (ts->idle_active || ts->tick_stopped)
1021 now = ktime_get();
1022
1023 if (ts->idle_active)
1024 tick_nohz_stop_idle(ts, now);
1025
1026 if (ts->tick_stopped) {
1027 tick_nohz_restart_sched_tick(ts, now, 0);
1028 tick_nohz_account_idle_ticks(ts);
1029 }
1030
1031 local_irq_enable();
1032}
1033
1034/*
1035 * The nohz low res interrupt handler
1036 */
1037static void tick_nohz_handler(struct clock_event_device *dev)
1038{
1039 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1040 struct pt_regs *regs = get_irq_regs();
1041 ktime_t now = ktime_get();
1042
1043 dev->next_event.tv64 = KTIME_MAX;
1044
1045 tick_sched_do_timer(now);
1046 tick_sched_handle(ts, regs);
1047
1048 /* No need to reprogram if we are running tickless */
1049 if (unlikely(ts->tick_stopped))
1050 return;
1051
1052 hrtimer_forward(&ts->sched_timer, now, tick_period);
1053 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1054}
1055
1056static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1057{
1058 if (!tick_nohz_enabled)
1059 return;
1060 ts->nohz_mode = mode;
1061 /* One update is enough */
1062 if (!test_and_set_bit(0, &tick_nohz_active))
1063 timers_update_migration(true);
1064}
1065
1066/**
1067 * tick_nohz_switch_to_nohz - switch to nohz mode
1068 */
1069static void tick_nohz_switch_to_nohz(void)
1070{
1071 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1072 ktime_t next;
1073
1074 if (!tick_nohz_enabled)
1075 return;
1076
1077 if (tick_switch_to_oneshot(tick_nohz_handler))
1078 return;
1079
1080 /*
1081 * Recycle the hrtimer in ts, so we can share the
1082 * hrtimer_forward with the highres code.
1083 */
1084 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1085 /* Get the next period */
1086 next = tick_init_jiffy_update();
1087
1088 hrtimer_set_expires(&ts->sched_timer, next);
1089 hrtimer_forward_now(&ts->sched_timer, tick_period);
1090 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1091 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
1092}
1093
1094/*
1095 * When NOHZ is enabled and the tick is stopped, we need to kick the
1096 * tick timer from irq_enter() so that the jiffies update is kept
1097 * alive during long running softirqs. That's ugly as hell, but
1098 * correctness is key even if we need to fix the offending softirq in
1099 * the first place.
1100 *
1101 * Note, this is different to tick_nohz_restart. We just kick the
1102 * timer and do not touch the other magic bits which need to be done
1103 * when idle is left.
1104 */
1105static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
1106{
1107#if 0
1108 /* Switch back to 2.6.27 behaviour */
1109 ktime_t delta;
1110
1111 /*
1112 * Do not touch the tick device, when the next expiry is either
1113 * already reached or less/equal than the tick period.
1114 */
1115 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
1116 if (delta.tv64 <= tick_period.tv64)
1117 return;
1118
1119 tick_nohz_restart(ts, now);
1120#endif
1121}
1122
1123static inline void tick_nohz_irq_enter(void)
1124{
1125 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1126 ktime_t now;
1127
1128 if (!ts->idle_active && !ts->tick_stopped)
1129 return;
1130 now = ktime_get();
1131 if (ts->idle_active)
1132 tick_nohz_stop_idle(ts, now);
1133 if (ts->tick_stopped) {
1134 tick_nohz_update_jiffies(now);
1135 tick_nohz_kick_tick(ts, now);
1136 }
1137}
1138
1139#else
1140
1141static inline void tick_nohz_switch_to_nohz(void) { }
1142static inline void tick_nohz_irq_enter(void) { }
1143static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
1144
1145#endif /* CONFIG_NO_HZ_COMMON */
1146
1147/*
1148 * Called from irq_enter to notify about the possible interruption of idle()
1149 */
1150void tick_irq_enter(void)
1151{
1152 tick_check_oneshot_broadcast_this_cpu();
1153 tick_nohz_irq_enter();
1154}
1155
1156/*
1157 * High resolution timer specific code
1158 */
1159#ifdef CONFIG_HIGH_RES_TIMERS
1160/*
1161 * We rearm the timer until we get disabled by the idle code.
1162 * Called with interrupts disabled.
1163 */
1164static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1165{
1166 struct tick_sched *ts =
1167 container_of(timer, struct tick_sched, sched_timer);
1168 struct pt_regs *regs = get_irq_regs();
1169 ktime_t now = ktime_get();
1170
1171 tick_sched_do_timer(now);
1172
1173 /*
1174 * Do not call, when we are not in irq context and have
1175 * no valid regs pointer
1176 */
1177 if (regs)
1178 tick_sched_handle(ts, regs);
1179
1180 /* No need to reprogram if we are in idle or full dynticks mode */
1181 if (unlikely(ts->tick_stopped))
1182 return HRTIMER_NORESTART;
1183
1184 hrtimer_forward(timer, now, tick_period);
1185
1186 return HRTIMER_RESTART;
1187}
1188
1189static int sched_skew_tick;
1190
1191static int __init skew_tick(char *str)
1192{
1193 get_option(&str, &sched_skew_tick);
1194
1195 return 0;
1196}
1197early_param("skew_tick", skew_tick);
1198
1199/**
1200 * tick_setup_sched_timer - setup the tick emulation timer
1201 */
1202void tick_setup_sched_timer(void)
1203{
1204 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1205 ktime_t now = ktime_get();
1206
1207 /*
1208 * Emulate tick processing via per-CPU hrtimers:
1209 */
1210 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1211 ts->sched_timer.function = tick_sched_timer;
1212
1213 /* Get the next period (per cpu) */
1214 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1215
1216 /* Offset the tick to avert jiffies_lock contention. */
1217 if (sched_skew_tick) {
1218 u64 offset = ktime_to_ns(tick_period) >> 1;
1219 do_div(offset, num_possible_cpus());
1220 offset *= smp_processor_id();
1221 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1222 }
1223
1224 hrtimer_forward(&ts->sched_timer, now, tick_period);
1225 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
1226 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
1227}
1228#endif /* HIGH_RES_TIMERS */
1229
1230#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
1231void tick_cancel_sched_timer(int cpu)
1232{
1233 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1234
1235# ifdef CONFIG_HIGH_RES_TIMERS
1236 if (ts->sched_timer.base)
1237 hrtimer_cancel(&ts->sched_timer);
1238# endif
1239
1240 memset(ts, 0, sizeof(*ts));
1241}
1242#endif
1243
1244/**
1245 * Async notification about clocksource changes
1246 */
1247void tick_clock_notify(void)
1248{
1249 int cpu;
1250
1251 for_each_possible_cpu(cpu)
1252 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1253}
1254
1255/*
1256 * Async notification about clock event changes
1257 */
1258void tick_oneshot_notify(void)
1259{
1260 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1261
1262 set_bit(0, &ts->check_clocks);
1263}
1264
1265/**
1266 * Check, if a change happened, which makes oneshot possible.
1267 *
1268 * Called cyclic from the hrtimer softirq (driven by the timer
1269 * softirq) allow_nohz signals, that we can switch into low-res nohz
1270 * mode, because high resolution timers are disabled (either compile
1271 * or runtime). Called with interrupts disabled.
1272 */
1273int tick_check_oneshot_change(int allow_nohz)
1274{
1275 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1276
1277 if (!test_and_clear_bit(0, &ts->check_clocks))
1278 return 0;
1279
1280 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1281 return 0;
1282
1283 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1284 return 0;
1285
1286 if (!allow_nohz)
1287 return 1;
1288
1289 tick_nohz_switch_to_nohz();
1290 return 0;
1291}
1/*
2 * linux/kernel/time/tick-sched.c
3 *
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 *
8 * No idle tick implementation for low and high resolution timers
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
12 * Distribute under GPLv2.
13 */
14#include <linux/cpu.h>
15#include <linux/err.h>
16#include <linux/hrtimer.h>
17#include <linux/interrupt.h>
18#include <linux/kernel_stat.h>
19#include <linux/percpu.h>
20#include <linux/profile.h>
21#include <linux/sched.h>
22#include <linux/module.h>
23
24#include <asm/irq_regs.h>
25
26#include "tick-internal.h"
27
28/*
29 * Per cpu nohz control structure
30 */
31static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
32
33/*
34 * The time, when the last jiffy update happened. Protected by xtime_lock.
35 */
36static ktime_t last_jiffies_update;
37
38struct tick_sched *tick_get_tick_sched(int cpu)
39{
40 return &per_cpu(tick_cpu_sched, cpu);
41}
42
43/*
44 * Must be called with interrupts disabled !
45 */
46static void tick_do_update_jiffies64(ktime_t now)
47{
48 unsigned long ticks = 0;
49 ktime_t delta;
50
51 /*
52 * Do a quick check without holding xtime_lock:
53 */
54 delta = ktime_sub(now, last_jiffies_update);
55 if (delta.tv64 < tick_period.tv64)
56 return;
57
58 /* Reevalute with xtime_lock held */
59 write_seqlock(&xtime_lock);
60
61 delta = ktime_sub(now, last_jiffies_update);
62 if (delta.tv64 >= tick_period.tv64) {
63
64 delta = ktime_sub(delta, tick_period);
65 last_jiffies_update = ktime_add(last_jiffies_update,
66 tick_period);
67
68 /* Slow path for long timeouts */
69 if (unlikely(delta.tv64 >= tick_period.tv64)) {
70 s64 incr = ktime_to_ns(tick_period);
71
72 ticks = ktime_divns(delta, incr);
73
74 last_jiffies_update = ktime_add_ns(last_jiffies_update,
75 incr * ticks);
76 }
77 do_timer(++ticks);
78
79 /* Keep the tick_next_period variable up to date */
80 tick_next_period = ktime_add(last_jiffies_update, tick_period);
81 }
82 write_sequnlock(&xtime_lock);
83}
84
85/*
86 * Initialize and return retrieve the jiffies update.
87 */
88static ktime_t tick_init_jiffy_update(void)
89{
90 ktime_t period;
91
92 write_seqlock(&xtime_lock);
93 /* Did we start the jiffies update yet ? */
94 if (last_jiffies_update.tv64 == 0)
95 last_jiffies_update = tick_next_period;
96 period = last_jiffies_update;
97 write_sequnlock(&xtime_lock);
98 return period;
99}
100
101/*
102 * NOHZ - aka dynamic tick functionality
103 */
104#ifdef CONFIG_NO_HZ
105/*
106 * NO HZ enabled ?
107 */
108static int tick_nohz_enabled __read_mostly = 1;
109
110/*
111 * Enable / Disable tickless mode
112 */
113static int __init setup_tick_nohz(char *str)
114{
115 if (!strcmp(str, "off"))
116 tick_nohz_enabled = 0;
117 else if (!strcmp(str, "on"))
118 tick_nohz_enabled = 1;
119 else
120 return 0;
121 return 1;
122}
123
124__setup("nohz=", setup_tick_nohz);
125
126/**
127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
128 *
129 * Called from interrupt entry when the CPU was idle
130 *
131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
133 * value. We do this unconditionally on any cpu, as we don't know whether the
134 * cpu, which has the update task assigned is in a long sleep.
135 */
136static void tick_nohz_update_jiffies(ktime_t now)
137{
138 int cpu = smp_processor_id();
139 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
140 unsigned long flags;
141
142 cpumask_clear_cpu(cpu, nohz_cpu_mask);
143 ts->idle_waketime = now;
144
145 local_irq_save(flags);
146 tick_do_update_jiffies64(now);
147 local_irq_restore(flags);
148
149 touch_softlockup_watchdog();
150}
151
152/*
153 * Updates the per cpu time idle statistics counters
154 */
155static void
156update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
157{
158 ktime_t delta;
159
160 if (ts->idle_active) {
161 delta = ktime_sub(now, ts->idle_entrytime);
162 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
163 if (nr_iowait_cpu(cpu) > 0)
164 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
165 ts->idle_entrytime = now;
166 }
167
168 if (last_update_time)
169 *last_update_time = ktime_to_us(now);
170
171}
172
173static void tick_nohz_stop_idle(int cpu, ktime_t now)
174{
175 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
176
177 update_ts_time_stats(cpu, ts, now, NULL);
178 ts->idle_active = 0;
179
180 sched_clock_idle_wakeup_event(0);
181}
182
183static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
184{
185 ktime_t now;
186
187 now = ktime_get();
188
189 update_ts_time_stats(cpu, ts, now, NULL);
190
191 ts->idle_entrytime = now;
192 ts->idle_active = 1;
193 sched_clock_idle_sleep_event();
194 return now;
195}
196
197/**
198 * get_cpu_idle_time_us - get the total idle time of a cpu
199 * @cpu: CPU number to query
200 * @last_update_time: variable to store update time in
201 *
202 * Return the cummulative idle time (since boot) for a given
203 * CPU, in microseconds. The idle time returned includes
204 * the iowait time (unlike what "top" and co report).
205 *
206 * This time is measured via accounting rather than sampling,
207 * and is as accurate as ktime_get() is.
208 *
209 * This function returns -1 if NOHZ is not enabled.
210 */
211u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
212{
213 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
214
215 if (!tick_nohz_enabled)
216 return -1;
217
218 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
219
220 return ktime_to_us(ts->idle_sleeptime);
221}
222EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
223
224/*
225 * get_cpu_iowait_time_us - get the total iowait time of a cpu
226 * @cpu: CPU number to query
227 * @last_update_time: variable to store update time in
228 *
229 * Return the cummulative iowait time (since boot) for a given
230 * CPU, in microseconds.
231 *
232 * This time is measured via accounting rather than sampling,
233 * and is as accurate as ktime_get() is.
234 *
235 * This function returns -1 if NOHZ is not enabled.
236 */
237u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
238{
239 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
240
241 if (!tick_nohz_enabled)
242 return -1;
243
244 update_ts_time_stats(cpu, ts, ktime_get(), last_update_time);
245
246 return ktime_to_us(ts->iowait_sleeptime);
247}
248EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
249
250/**
251 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
252 *
253 * When the next event is more than a tick into the future, stop the idle tick
254 * Called either from the idle loop or from irq_exit() when an idle period was
255 * just interrupted by an interrupt which did not cause a reschedule.
256 */
257void tick_nohz_stop_sched_tick(int inidle)
258{
259 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
260 struct tick_sched *ts;
261 ktime_t last_update, expires, now;
262 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
263 u64 time_delta;
264 int cpu;
265
266 local_irq_save(flags);
267
268 cpu = smp_processor_id();
269 ts = &per_cpu(tick_cpu_sched, cpu);
270
271 /*
272 * Call to tick_nohz_start_idle stops the last_update_time from being
273 * updated. Thus, it must not be called in the event we are called from
274 * irq_exit() with the prior state different than idle.
275 */
276 if (!inidle && !ts->inidle)
277 goto end;
278
279 /*
280 * Set ts->inidle unconditionally. Even if the system did not
281 * switch to NOHZ mode the cpu frequency governers rely on the
282 * update of the idle time accounting in tick_nohz_start_idle().
283 */
284 ts->inidle = 1;
285
286 now = tick_nohz_start_idle(cpu, ts);
287
288 /*
289 * If this cpu is offline and it is the one which updates
290 * jiffies, then give up the assignment and let it be taken by
291 * the cpu which runs the tick timer next. If we don't drop
292 * this here the jiffies might be stale and do_timer() never
293 * invoked.
294 */
295 if (unlikely(!cpu_online(cpu))) {
296 if (cpu == tick_do_timer_cpu)
297 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
298 }
299
300 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
301 goto end;
302
303 if (need_resched())
304 goto end;
305
306 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
307 static int ratelimit;
308
309 if (ratelimit < 10) {
310 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
311 (unsigned int) local_softirq_pending());
312 ratelimit++;
313 }
314 goto end;
315 }
316
317 ts->idle_calls++;
318 /* Read jiffies and the time when jiffies were updated last */
319 do {
320 seq = read_seqbegin(&xtime_lock);
321 last_update = last_jiffies_update;
322 last_jiffies = jiffies;
323 time_delta = timekeeping_max_deferment();
324 } while (read_seqretry(&xtime_lock, seq));
325
326 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
327 arch_needs_cpu(cpu)) {
328 next_jiffies = last_jiffies + 1;
329 delta_jiffies = 1;
330 } else {
331 /* Get the next timer wheel timer */
332 next_jiffies = get_next_timer_interrupt(last_jiffies);
333 delta_jiffies = next_jiffies - last_jiffies;
334 }
335 /*
336 * Do not stop the tick, if we are only one off
337 * or if the cpu is required for rcu
338 */
339 if (!ts->tick_stopped && delta_jiffies == 1)
340 goto out;
341
342 /* Schedule the tick, if we are at least one jiffie off */
343 if ((long)delta_jiffies >= 1) {
344
345 /*
346 * If this cpu is the one which updates jiffies, then
347 * give up the assignment and let it be taken by the
348 * cpu which runs the tick timer next, which might be
349 * this cpu as well. If we don't drop this here the
350 * jiffies might be stale and do_timer() never
351 * invoked. Keep track of the fact that it was the one
352 * which had the do_timer() duty last. If this cpu is
353 * the one which had the do_timer() duty last, we
354 * limit the sleep time to the timekeeping
355 * max_deferement value which we retrieved
356 * above. Otherwise we can sleep as long as we want.
357 */
358 if (cpu == tick_do_timer_cpu) {
359 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
360 ts->do_timer_last = 1;
361 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
362 time_delta = KTIME_MAX;
363 ts->do_timer_last = 0;
364 } else if (!ts->do_timer_last) {
365 time_delta = KTIME_MAX;
366 }
367
368 /*
369 * calculate the expiry time for the next timer wheel
370 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
371 * that there is no timer pending or at least extremely
372 * far into the future (12 days for HZ=1000). In this
373 * case we set the expiry to the end of time.
374 */
375 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
376 /*
377 * Calculate the time delta for the next timer event.
378 * If the time delta exceeds the maximum time delta
379 * permitted by the current clocksource then adjust
380 * the time delta accordingly to ensure the
381 * clocksource does not wrap.
382 */
383 time_delta = min_t(u64, time_delta,
384 tick_period.tv64 * delta_jiffies);
385 }
386
387 if (time_delta < KTIME_MAX)
388 expires = ktime_add_ns(last_update, time_delta);
389 else
390 expires.tv64 = KTIME_MAX;
391
392 if (delta_jiffies > 1)
393 cpumask_set_cpu(cpu, nohz_cpu_mask);
394
395 /* Skip reprogram of event if its not changed */
396 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
397 goto out;
398
399 /*
400 * nohz_stop_sched_tick can be called several times before
401 * the nohz_restart_sched_tick is called. This happens when
402 * interrupts arrive which do not cause a reschedule. In the
403 * first call we save the current tick time, so we can restart
404 * the scheduler tick in nohz_restart_sched_tick.
405 */
406 if (!ts->tick_stopped) {
407 select_nohz_load_balancer(1);
408
409 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
410 ts->tick_stopped = 1;
411 ts->idle_jiffies = last_jiffies;
412 rcu_enter_nohz();
413 }
414
415 ts->idle_sleeps++;
416
417 /* Mark expires */
418 ts->idle_expires = expires;
419
420 /*
421 * If the expiration time == KTIME_MAX, then
422 * in this case we simply stop the tick timer.
423 */
424 if (unlikely(expires.tv64 == KTIME_MAX)) {
425 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
426 hrtimer_cancel(&ts->sched_timer);
427 goto out;
428 }
429
430 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
431 hrtimer_start(&ts->sched_timer, expires,
432 HRTIMER_MODE_ABS_PINNED);
433 /* Check, if the timer was already in the past */
434 if (hrtimer_active(&ts->sched_timer))
435 goto out;
436 } else if (!tick_program_event(expires, 0))
437 goto out;
438 /*
439 * We are past the event already. So we crossed a
440 * jiffie boundary. Update jiffies and raise the
441 * softirq.
442 */
443 tick_do_update_jiffies64(ktime_get());
444 cpumask_clear_cpu(cpu, nohz_cpu_mask);
445 }
446 raise_softirq_irqoff(TIMER_SOFTIRQ);
447out:
448 ts->next_jiffies = next_jiffies;
449 ts->last_jiffies = last_jiffies;
450 ts->sleep_length = ktime_sub(dev->next_event, now);
451end:
452 local_irq_restore(flags);
453}
454
455/**
456 * tick_nohz_get_sleep_length - return the length of the current sleep
457 *
458 * Called from power state control code with interrupts disabled
459 */
460ktime_t tick_nohz_get_sleep_length(void)
461{
462 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
463
464 return ts->sleep_length;
465}
466
467static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
468{
469 hrtimer_cancel(&ts->sched_timer);
470 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
471
472 while (1) {
473 /* Forward the time to expire in the future */
474 hrtimer_forward(&ts->sched_timer, now, tick_period);
475
476 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
477 hrtimer_start_expires(&ts->sched_timer,
478 HRTIMER_MODE_ABS_PINNED);
479 /* Check, if the timer was already in the past */
480 if (hrtimer_active(&ts->sched_timer))
481 break;
482 } else {
483 if (!tick_program_event(
484 hrtimer_get_expires(&ts->sched_timer), 0))
485 break;
486 }
487 /* Update jiffies and reread time */
488 tick_do_update_jiffies64(now);
489 now = ktime_get();
490 }
491}
492
493/**
494 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
495 *
496 * Restart the idle tick when the CPU is woken up from idle
497 */
498void tick_nohz_restart_sched_tick(void)
499{
500 int cpu = smp_processor_id();
501 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
502#ifndef CONFIG_VIRT_CPU_ACCOUNTING
503 unsigned long ticks;
504#endif
505 ktime_t now;
506
507 local_irq_disable();
508 if (ts->idle_active || (ts->inidle && ts->tick_stopped))
509 now = ktime_get();
510
511 if (ts->idle_active)
512 tick_nohz_stop_idle(cpu, now);
513
514 if (!ts->inidle || !ts->tick_stopped) {
515 ts->inidle = 0;
516 local_irq_enable();
517 return;
518 }
519
520 ts->inidle = 0;
521
522 rcu_exit_nohz();
523
524 /* Update jiffies first */
525 select_nohz_load_balancer(0);
526 tick_do_update_jiffies64(now);
527 cpumask_clear_cpu(cpu, nohz_cpu_mask);
528
529#ifndef CONFIG_VIRT_CPU_ACCOUNTING
530 /*
531 * We stopped the tick in idle. Update process times would miss the
532 * time we slept as update_process_times does only a 1 tick
533 * accounting. Enforce that this is accounted to idle !
534 */
535 ticks = jiffies - ts->idle_jiffies;
536 /*
537 * We might be one off. Do not randomly account a huge number of ticks!
538 */
539 if (ticks && ticks < LONG_MAX)
540 account_idle_ticks(ticks);
541#endif
542
543 touch_softlockup_watchdog();
544 /*
545 * Cancel the scheduled timer and restore the tick
546 */
547 ts->tick_stopped = 0;
548 ts->idle_exittime = now;
549
550 tick_nohz_restart(ts, now);
551
552 local_irq_enable();
553}
554
555static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
556{
557 hrtimer_forward(&ts->sched_timer, now, tick_period);
558 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
559}
560
561/*
562 * The nohz low res interrupt handler
563 */
564static void tick_nohz_handler(struct clock_event_device *dev)
565{
566 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
567 struct pt_regs *regs = get_irq_regs();
568 int cpu = smp_processor_id();
569 ktime_t now = ktime_get();
570
571 dev->next_event.tv64 = KTIME_MAX;
572
573 /*
574 * Check if the do_timer duty was dropped. We don't care about
575 * concurrency: This happens only when the cpu in charge went
576 * into a long sleep. If two cpus happen to assign themself to
577 * this duty, then the jiffies update is still serialized by
578 * xtime_lock.
579 */
580 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
581 tick_do_timer_cpu = cpu;
582
583 /* Check, if the jiffies need an update */
584 if (tick_do_timer_cpu == cpu)
585 tick_do_update_jiffies64(now);
586
587 /*
588 * When we are idle and the tick is stopped, we have to touch
589 * the watchdog as we might not schedule for a really long
590 * time. This happens on complete idle SMP systems while
591 * waiting on the login prompt. We also increment the "start
592 * of idle" jiffy stamp so the idle accounting adjustment we
593 * do when we go busy again does not account too much ticks.
594 */
595 if (ts->tick_stopped) {
596 touch_softlockup_watchdog();
597 ts->idle_jiffies++;
598 }
599
600 update_process_times(user_mode(regs));
601 profile_tick(CPU_PROFILING);
602
603 while (tick_nohz_reprogram(ts, now)) {
604 now = ktime_get();
605 tick_do_update_jiffies64(now);
606 }
607}
608
609/**
610 * tick_nohz_switch_to_nohz - switch to nohz mode
611 */
612static void tick_nohz_switch_to_nohz(void)
613{
614 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
615 ktime_t next;
616
617 if (!tick_nohz_enabled)
618 return;
619
620 local_irq_disable();
621 if (tick_switch_to_oneshot(tick_nohz_handler)) {
622 local_irq_enable();
623 return;
624 }
625
626 ts->nohz_mode = NOHZ_MODE_LOWRES;
627
628 /*
629 * Recycle the hrtimer in ts, so we can share the
630 * hrtimer_forward with the highres code.
631 */
632 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
633 /* Get the next period */
634 next = tick_init_jiffy_update();
635
636 for (;;) {
637 hrtimer_set_expires(&ts->sched_timer, next);
638 if (!tick_program_event(next, 0))
639 break;
640 next = ktime_add(next, tick_period);
641 }
642 local_irq_enable();
643
644 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", smp_processor_id());
645}
646
647/*
648 * When NOHZ is enabled and the tick is stopped, we need to kick the
649 * tick timer from irq_enter() so that the jiffies update is kept
650 * alive during long running softirqs. That's ugly as hell, but
651 * correctness is key even if we need to fix the offending softirq in
652 * the first place.
653 *
654 * Note, this is different to tick_nohz_restart. We just kick the
655 * timer and do not touch the other magic bits which need to be done
656 * when idle is left.
657 */
658static void tick_nohz_kick_tick(int cpu, ktime_t now)
659{
660#if 0
661 /* Switch back to 2.6.27 behaviour */
662
663 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
664 ktime_t delta;
665
666 /*
667 * Do not touch the tick device, when the next expiry is either
668 * already reached or less/equal than the tick period.
669 */
670 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
671 if (delta.tv64 <= tick_period.tv64)
672 return;
673
674 tick_nohz_restart(ts, now);
675#endif
676}
677
678static inline void tick_check_nohz(int cpu)
679{
680 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
681 ktime_t now;
682
683 if (!ts->idle_active && !ts->tick_stopped)
684 return;
685 now = ktime_get();
686 if (ts->idle_active)
687 tick_nohz_stop_idle(cpu, now);
688 if (ts->tick_stopped) {
689 tick_nohz_update_jiffies(now);
690 tick_nohz_kick_tick(cpu, now);
691 }
692}
693
694#else
695
696static inline void tick_nohz_switch_to_nohz(void) { }
697static inline void tick_check_nohz(int cpu) { }
698
699#endif /* NO_HZ */
700
701/*
702 * Called from irq_enter to notify about the possible interruption of idle()
703 */
704void tick_check_idle(int cpu)
705{
706 tick_check_oneshot_broadcast(cpu);
707 tick_check_nohz(cpu);
708}
709
710/*
711 * High resolution timer specific code
712 */
713#ifdef CONFIG_HIGH_RES_TIMERS
714/*
715 * We rearm the timer until we get disabled by the idle code.
716 * Called with interrupts disabled and timer->base->cpu_base->lock held.
717 */
718static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
719{
720 struct tick_sched *ts =
721 container_of(timer, struct tick_sched, sched_timer);
722 struct pt_regs *regs = get_irq_regs();
723 ktime_t now = ktime_get();
724 int cpu = smp_processor_id();
725
726#ifdef CONFIG_NO_HZ
727 /*
728 * Check if the do_timer duty was dropped. We don't care about
729 * concurrency: This happens only when the cpu in charge went
730 * into a long sleep. If two cpus happen to assign themself to
731 * this duty, then the jiffies update is still serialized by
732 * xtime_lock.
733 */
734 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
735 tick_do_timer_cpu = cpu;
736#endif
737
738 /* Check, if the jiffies need an update */
739 if (tick_do_timer_cpu == cpu)
740 tick_do_update_jiffies64(now);
741
742 /*
743 * Do not call, when we are not in irq context and have
744 * no valid regs pointer
745 */
746 if (regs) {
747 /*
748 * When we are idle and the tick is stopped, we have to touch
749 * the watchdog as we might not schedule for a really long
750 * time. This happens on complete idle SMP systems while
751 * waiting on the login prompt. We also increment the "start of
752 * idle" jiffy stamp so the idle accounting adjustment we do
753 * when we go busy again does not account too much ticks.
754 */
755 if (ts->tick_stopped) {
756 touch_softlockup_watchdog();
757 ts->idle_jiffies++;
758 }
759 update_process_times(user_mode(regs));
760 profile_tick(CPU_PROFILING);
761 }
762
763 hrtimer_forward(timer, now, tick_period);
764
765 return HRTIMER_RESTART;
766}
767
768/**
769 * tick_setup_sched_timer - setup the tick emulation timer
770 */
771void tick_setup_sched_timer(void)
772{
773 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
774 ktime_t now = ktime_get();
775
776 /*
777 * Emulate tick processing via per-CPU hrtimers:
778 */
779 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
780 ts->sched_timer.function = tick_sched_timer;
781
782 /* Get the next period (per cpu) */
783 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
784
785 for (;;) {
786 hrtimer_forward(&ts->sched_timer, now, tick_period);
787 hrtimer_start_expires(&ts->sched_timer,
788 HRTIMER_MODE_ABS_PINNED);
789 /* Check, if the timer was already in the past */
790 if (hrtimer_active(&ts->sched_timer))
791 break;
792 now = ktime_get();
793 }
794
795#ifdef CONFIG_NO_HZ
796 if (tick_nohz_enabled) {
797 ts->nohz_mode = NOHZ_MODE_HIGHRES;
798 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", smp_processor_id());
799 }
800#endif
801}
802#endif /* HIGH_RES_TIMERS */
803
804#if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
805void tick_cancel_sched_timer(int cpu)
806{
807 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
808
809# ifdef CONFIG_HIGH_RES_TIMERS
810 if (ts->sched_timer.base)
811 hrtimer_cancel(&ts->sched_timer);
812# endif
813
814 ts->nohz_mode = NOHZ_MODE_INACTIVE;
815}
816#endif
817
818/**
819 * Async notification about clocksource changes
820 */
821void tick_clock_notify(void)
822{
823 int cpu;
824
825 for_each_possible_cpu(cpu)
826 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
827}
828
829/*
830 * Async notification about clock event changes
831 */
832void tick_oneshot_notify(void)
833{
834 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
835
836 set_bit(0, &ts->check_clocks);
837}
838
839/**
840 * Check, if a change happened, which makes oneshot possible.
841 *
842 * Called cyclic from the hrtimer softirq (driven by the timer
843 * softirq) allow_nohz signals, that we can switch into low-res nohz
844 * mode, because high resolution timers are disabled (either compile
845 * or runtime).
846 */
847int tick_check_oneshot_change(int allow_nohz)
848{
849 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
850
851 if (!test_and_clear_bit(0, &ts->check_clocks))
852 return 0;
853
854 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
855 return 0;
856
857 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
858 return 0;
859
860 if (!allow_nohz)
861 return 1;
862
863 tick_nohz_switch_to_nohz();
864 return 0;
865}