Loading...
1/*
2 * linux/kernel/hrtimer.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 * High-resolution kernel timers
9 *
10 * In contrast to the low-resolution timeout API implemented in
11 * kernel/timer.c, hrtimers provide finer resolution and accuracy
12 * depending on system configuration and capabilities.
13 *
14 * These timers are currently used for:
15 * - itimers
16 * - POSIX timers
17 * - nanosleep
18 * - precise in-kernel timing
19 *
20 * Started by: Thomas Gleixner and Ingo Molnar
21 *
22 * Credits:
23 * based on kernel/timer.c
24 *
25 * Help, testing, suggestions, bugfixes, improvements were
26 * provided by:
27 *
28 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
29 * et. al.
30 *
31 * For licencing details see kernel-base/COPYING
32 */
33
34#include <linux/cpu.h>
35#include <linux/export.h>
36#include <linux/percpu.h>
37#include <linux/hrtimer.h>
38#include <linux/notifier.h>
39#include <linux/syscalls.h>
40#include <linux/kallsyms.h>
41#include <linux/interrupt.h>
42#include <linux/tick.h>
43#include <linux/seq_file.h>
44#include <linux/err.h>
45#include <linux/debugobjects.h>
46#include <linux/sched.h>
47#include <linux/sched/sysctl.h>
48#include <linux/sched/rt.h>
49#include <linux/sched/deadline.h>
50#include <linux/timer.h>
51#include <linux/freezer.h>
52
53#include <asm/uaccess.h>
54
55#include <trace/events/timer.h>
56
57#include "tick-internal.h"
58
59/*
60 * The timer bases:
61 *
62 * There are more clockids than hrtimer bases. Thus, we index
63 * into the timer bases by the hrtimer_base_type enum. When trying
64 * to reach a base using a clockid, hrtimer_clockid_to_base()
65 * is used to convert from clockid to the proper hrtimer_base_type.
66 */
67DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
68{
69 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
70 .seq = SEQCNT_ZERO(hrtimer_bases.seq),
71 .clock_base =
72 {
73 {
74 .index = HRTIMER_BASE_MONOTONIC,
75 .clockid = CLOCK_MONOTONIC,
76 .get_time = &ktime_get,
77 },
78 {
79 .index = HRTIMER_BASE_REALTIME,
80 .clockid = CLOCK_REALTIME,
81 .get_time = &ktime_get_real,
82 },
83 {
84 .index = HRTIMER_BASE_BOOTTIME,
85 .clockid = CLOCK_BOOTTIME,
86 .get_time = &ktime_get_boottime,
87 },
88 {
89 .index = HRTIMER_BASE_TAI,
90 .clockid = CLOCK_TAI,
91 .get_time = &ktime_get_clocktai,
92 },
93 }
94};
95
96static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
97 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
98 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
99 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
100 [CLOCK_TAI] = HRTIMER_BASE_TAI,
101};
102
103static inline int hrtimer_clockid_to_base(clockid_t clock_id)
104{
105 return hrtimer_clock_to_base_table[clock_id];
106}
107
108/*
109 * Functions and macros which are different for UP/SMP systems are kept in a
110 * single place
111 */
112#ifdef CONFIG_SMP
113
114/*
115 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
116 * such that hrtimer_callback_running() can unconditionally dereference
117 * timer->base->cpu_base
118 */
119static struct hrtimer_cpu_base migration_cpu_base = {
120 .seq = SEQCNT_ZERO(migration_cpu_base),
121 .clock_base = { { .cpu_base = &migration_cpu_base, }, },
122};
123
124#define migration_base migration_cpu_base.clock_base[0]
125
126/*
127 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
128 * means that all timers which are tied to this base via timer->base are
129 * locked, and the base itself is locked too.
130 *
131 * So __run_timers/migrate_timers can safely modify all timers which could
132 * be found on the lists/queues.
133 *
134 * When the timer's base is locked, and the timer removed from list, it is
135 * possible to set timer->base = &migration_base and drop the lock: the timer
136 * remains locked.
137 */
138static
139struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
140 unsigned long *flags)
141{
142 struct hrtimer_clock_base *base;
143
144 for (;;) {
145 base = timer->base;
146 if (likely(base != &migration_base)) {
147 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
148 if (likely(base == timer->base))
149 return base;
150 /* The timer has migrated to another CPU: */
151 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
152 }
153 cpu_relax();
154 }
155}
156
157/*
158 * With HIGHRES=y we do not migrate the timer when it is expiring
159 * before the next event on the target cpu because we cannot reprogram
160 * the target cpu hardware and we would cause it to fire late.
161 *
162 * Called with cpu_base->lock of target cpu held.
163 */
164static int
165hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
166{
167#ifdef CONFIG_HIGH_RES_TIMERS
168 ktime_t expires;
169
170 if (!new_base->cpu_base->hres_active)
171 return 0;
172
173 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
174 return expires.tv64 <= new_base->cpu_base->expires_next.tv64;
175#else
176 return 0;
177#endif
178}
179
180#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
181static inline
182struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
183 int pinned)
184{
185 if (pinned || !base->migration_enabled)
186 return base;
187 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
188}
189#else
190static inline
191struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
192 int pinned)
193{
194 return base;
195}
196#endif
197
198/*
199 * We switch the timer base to a power-optimized selected CPU target,
200 * if:
201 * - NO_HZ_COMMON is enabled
202 * - timer migration is enabled
203 * - the timer callback is not running
204 * - the timer is not the first expiring timer on the new target
205 *
206 * If one of the above requirements is not fulfilled we move the timer
207 * to the current CPU or leave it on the previously assigned CPU if
208 * the timer callback is currently running.
209 */
210static inline struct hrtimer_clock_base *
211switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
212 int pinned)
213{
214 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
215 struct hrtimer_clock_base *new_base;
216 int basenum = base->index;
217
218 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
219 new_cpu_base = get_target_base(this_cpu_base, pinned);
220again:
221 new_base = &new_cpu_base->clock_base[basenum];
222
223 if (base != new_base) {
224 /*
225 * We are trying to move timer to new_base.
226 * However we can't change timer's base while it is running,
227 * so we keep it on the same CPU. No hassle vs. reprogramming
228 * the event source in the high resolution case. The softirq
229 * code will take care of this when the timer function has
230 * completed. There is no conflict as we hold the lock until
231 * the timer is enqueued.
232 */
233 if (unlikely(hrtimer_callback_running(timer)))
234 return base;
235
236 /* See the comment in lock_hrtimer_base() */
237 timer->base = &migration_base;
238 raw_spin_unlock(&base->cpu_base->lock);
239 raw_spin_lock(&new_base->cpu_base->lock);
240
241 if (new_cpu_base != this_cpu_base &&
242 hrtimer_check_target(timer, new_base)) {
243 raw_spin_unlock(&new_base->cpu_base->lock);
244 raw_spin_lock(&base->cpu_base->lock);
245 new_cpu_base = this_cpu_base;
246 timer->base = base;
247 goto again;
248 }
249 timer->base = new_base;
250 } else {
251 if (new_cpu_base != this_cpu_base &&
252 hrtimer_check_target(timer, new_base)) {
253 new_cpu_base = this_cpu_base;
254 goto again;
255 }
256 }
257 return new_base;
258}
259
260#else /* CONFIG_SMP */
261
262static inline struct hrtimer_clock_base *
263lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
264{
265 struct hrtimer_clock_base *base = timer->base;
266
267 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
268
269 return base;
270}
271
272# define switch_hrtimer_base(t, b, p) (b)
273
274#endif /* !CONFIG_SMP */
275
276/*
277 * Functions for the union type storage format of ktime_t which are
278 * too large for inlining:
279 */
280#if BITS_PER_LONG < 64
281/*
282 * Divide a ktime value by a nanosecond value
283 */
284s64 __ktime_divns(const ktime_t kt, s64 div)
285{
286 int sft = 0;
287 s64 dclc;
288 u64 tmp;
289
290 dclc = ktime_to_ns(kt);
291 tmp = dclc < 0 ? -dclc : dclc;
292
293 /* Make sure the divisor is less than 2^32: */
294 while (div >> 32) {
295 sft++;
296 div >>= 1;
297 }
298 tmp >>= sft;
299 do_div(tmp, (unsigned long) div);
300 return dclc < 0 ? -tmp : tmp;
301}
302EXPORT_SYMBOL_GPL(__ktime_divns);
303#endif /* BITS_PER_LONG >= 64 */
304
305/*
306 * Add two ktime values and do a safety check for overflow:
307 */
308ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
309{
310 ktime_t res = ktime_add(lhs, rhs);
311
312 /*
313 * We use KTIME_SEC_MAX here, the maximum timeout which we can
314 * return to user space in a timespec:
315 */
316 if (res.tv64 < 0 || res.tv64 < lhs.tv64 || res.tv64 < rhs.tv64)
317 res = ktime_set(KTIME_SEC_MAX, 0);
318
319 return res;
320}
321
322EXPORT_SYMBOL_GPL(ktime_add_safe);
323
324#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
325
326static struct debug_obj_descr hrtimer_debug_descr;
327
328static void *hrtimer_debug_hint(void *addr)
329{
330 return ((struct hrtimer *) addr)->function;
331}
332
333/*
334 * fixup_init is called when:
335 * - an active object is initialized
336 */
337static int hrtimer_fixup_init(void *addr, enum debug_obj_state state)
338{
339 struct hrtimer *timer = addr;
340
341 switch (state) {
342 case ODEBUG_STATE_ACTIVE:
343 hrtimer_cancel(timer);
344 debug_object_init(timer, &hrtimer_debug_descr);
345 return 1;
346 default:
347 return 0;
348 }
349}
350
351/*
352 * fixup_activate is called when:
353 * - an active object is activated
354 * - an unknown object is activated (might be a statically initialized object)
355 */
356static int hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
357{
358 switch (state) {
359
360 case ODEBUG_STATE_NOTAVAILABLE:
361 WARN_ON_ONCE(1);
362 return 0;
363
364 case ODEBUG_STATE_ACTIVE:
365 WARN_ON(1);
366
367 default:
368 return 0;
369 }
370}
371
372/*
373 * fixup_free is called when:
374 * - an active object is freed
375 */
376static int hrtimer_fixup_free(void *addr, enum debug_obj_state state)
377{
378 struct hrtimer *timer = addr;
379
380 switch (state) {
381 case ODEBUG_STATE_ACTIVE:
382 hrtimer_cancel(timer);
383 debug_object_free(timer, &hrtimer_debug_descr);
384 return 1;
385 default:
386 return 0;
387 }
388}
389
390static struct debug_obj_descr hrtimer_debug_descr = {
391 .name = "hrtimer",
392 .debug_hint = hrtimer_debug_hint,
393 .fixup_init = hrtimer_fixup_init,
394 .fixup_activate = hrtimer_fixup_activate,
395 .fixup_free = hrtimer_fixup_free,
396};
397
398static inline void debug_hrtimer_init(struct hrtimer *timer)
399{
400 debug_object_init(timer, &hrtimer_debug_descr);
401}
402
403static inline void debug_hrtimer_activate(struct hrtimer *timer)
404{
405 debug_object_activate(timer, &hrtimer_debug_descr);
406}
407
408static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
409{
410 debug_object_deactivate(timer, &hrtimer_debug_descr);
411}
412
413static inline void debug_hrtimer_free(struct hrtimer *timer)
414{
415 debug_object_free(timer, &hrtimer_debug_descr);
416}
417
418static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
419 enum hrtimer_mode mode);
420
421void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
422 enum hrtimer_mode mode)
423{
424 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
425 __hrtimer_init(timer, clock_id, mode);
426}
427EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
428
429void destroy_hrtimer_on_stack(struct hrtimer *timer)
430{
431 debug_object_free(timer, &hrtimer_debug_descr);
432}
433
434#else
435static inline void debug_hrtimer_init(struct hrtimer *timer) { }
436static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
437static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
438#endif
439
440static inline void
441debug_init(struct hrtimer *timer, clockid_t clockid,
442 enum hrtimer_mode mode)
443{
444 debug_hrtimer_init(timer);
445 trace_hrtimer_init(timer, clockid, mode);
446}
447
448static inline void debug_activate(struct hrtimer *timer)
449{
450 debug_hrtimer_activate(timer);
451 trace_hrtimer_start(timer);
452}
453
454static inline void debug_deactivate(struct hrtimer *timer)
455{
456 debug_hrtimer_deactivate(timer);
457 trace_hrtimer_cancel(timer);
458}
459
460#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
461static inline void hrtimer_update_next_timer(struct hrtimer_cpu_base *cpu_base,
462 struct hrtimer *timer)
463{
464#ifdef CONFIG_HIGH_RES_TIMERS
465 cpu_base->next_timer = timer;
466#endif
467}
468
469static ktime_t __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base)
470{
471 struct hrtimer_clock_base *base = cpu_base->clock_base;
472 ktime_t expires, expires_next = { .tv64 = KTIME_MAX };
473 unsigned int active = cpu_base->active_bases;
474
475 hrtimer_update_next_timer(cpu_base, NULL);
476 for (; active; base++, active >>= 1) {
477 struct timerqueue_node *next;
478 struct hrtimer *timer;
479
480 if (!(active & 0x01))
481 continue;
482
483 next = timerqueue_getnext(&base->active);
484 timer = container_of(next, struct hrtimer, node);
485 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
486 if (expires.tv64 < expires_next.tv64) {
487 expires_next = expires;
488 hrtimer_update_next_timer(cpu_base, timer);
489 }
490 }
491 /*
492 * clock_was_set() might have changed base->offset of any of
493 * the clock bases so the result might be negative. Fix it up
494 * to prevent a false positive in clockevents_program_event().
495 */
496 if (expires_next.tv64 < 0)
497 expires_next.tv64 = 0;
498 return expires_next;
499}
500#endif
501
502static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
503{
504 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
505 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
506 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
507
508 return ktime_get_update_offsets_now(&base->clock_was_set_seq,
509 offs_real, offs_boot, offs_tai);
510}
511
512/* High resolution timer related functions */
513#ifdef CONFIG_HIGH_RES_TIMERS
514
515/*
516 * High resolution timer enabled ?
517 */
518static bool hrtimer_hres_enabled __read_mostly = true;
519unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
520EXPORT_SYMBOL_GPL(hrtimer_resolution);
521
522/*
523 * Enable / Disable high resolution mode
524 */
525static int __init setup_hrtimer_hres(char *str)
526{
527 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
528}
529
530__setup("highres=", setup_hrtimer_hres);
531
532/*
533 * hrtimer_high_res_enabled - query, if the highres mode is enabled
534 */
535static inline int hrtimer_is_hres_enabled(void)
536{
537 return hrtimer_hres_enabled;
538}
539
540/*
541 * Is the high resolution mode active ?
542 */
543static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
544{
545 return cpu_base->hres_active;
546}
547
548static inline int hrtimer_hres_active(void)
549{
550 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
551}
552
553/*
554 * Reprogram the event source with checking both queues for the
555 * next event
556 * Called with interrupts disabled and base->lock held
557 */
558static void
559hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
560{
561 ktime_t expires_next;
562
563 if (!cpu_base->hres_active)
564 return;
565
566 expires_next = __hrtimer_get_next_event(cpu_base);
567
568 if (skip_equal && expires_next.tv64 == cpu_base->expires_next.tv64)
569 return;
570
571 cpu_base->expires_next.tv64 = expires_next.tv64;
572
573 /*
574 * If a hang was detected in the last timer interrupt then we
575 * leave the hang delay active in the hardware. We want the
576 * system to make progress. That also prevents the following
577 * scenario:
578 * T1 expires 50ms from now
579 * T2 expires 5s from now
580 *
581 * T1 is removed, so this code is called and would reprogram
582 * the hardware to 5s from now. Any hrtimer_start after that
583 * will not reprogram the hardware due to hang_detected being
584 * set. So we'd effectivly block all timers until the T2 event
585 * fires.
586 */
587 if (cpu_base->hang_detected)
588 return;
589
590 tick_program_event(cpu_base->expires_next, 1);
591}
592
593/*
594 * When a timer is enqueued and expires earlier than the already enqueued
595 * timers, we have to check, whether it expires earlier than the timer for
596 * which the clock event device was armed.
597 *
598 * Called with interrupts disabled and base->cpu_base.lock held
599 */
600static void hrtimer_reprogram(struct hrtimer *timer,
601 struct hrtimer_clock_base *base)
602{
603 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
604 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
605
606 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
607
608 /*
609 * If the timer is not on the current cpu, we cannot reprogram
610 * the other cpus clock event device.
611 */
612 if (base->cpu_base != cpu_base)
613 return;
614
615 /*
616 * If the hrtimer interrupt is running, then it will
617 * reevaluate the clock bases and reprogram the clock event
618 * device. The callbacks are always executed in hard interrupt
619 * context so we don't need an extra check for a running
620 * callback.
621 */
622 if (cpu_base->in_hrtirq)
623 return;
624
625 /*
626 * CLOCK_REALTIME timer might be requested with an absolute
627 * expiry time which is less than base->offset. Set it to 0.
628 */
629 if (expires.tv64 < 0)
630 expires.tv64 = 0;
631
632 if (expires.tv64 >= cpu_base->expires_next.tv64)
633 return;
634
635 /* Update the pointer to the next expiring timer */
636 cpu_base->next_timer = timer;
637
638 /*
639 * If a hang was detected in the last timer interrupt then we
640 * do not schedule a timer which is earlier than the expiry
641 * which we enforced in the hang detection. We want the system
642 * to make progress.
643 */
644 if (cpu_base->hang_detected)
645 return;
646
647 /*
648 * Program the timer hardware. We enforce the expiry for
649 * events which are already in the past.
650 */
651 cpu_base->expires_next = expires;
652 tick_program_event(expires, 1);
653}
654
655/*
656 * Initialize the high resolution related parts of cpu_base
657 */
658static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
659{
660 base->expires_next.tv64 = KTIME_MAX;
661 base->hres_active = 0;
662}
663
664/*
665 * Retrigger next event is called after clock was set
666 *
667 * Called with interrupts disabled via on_each_cpu()
668 */
669static void retrigger_next_event(void *arg)
670{
671 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
672
673 if (!base->hres_active)
674 return;
675
676 raw_spin_lock(&base->lock);
677 hrtimer_update_base(base);
678 hrtimer_force_reprogram(base, 0);
679 raw_spin_unlock(&base->lock);
680}
681
682/*
683 * Switch to high resolution mode
684 */
685static void hrtimer_switch_to_hres(void)
686{
687 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
688
689 if (tick_init_highres()) {
690 printk(KERN_WARNING "Could not switch to high resolution "
691 "mode on CPU %d\n", base->cpu);
692 return;
693 }
694 base->hres_active = 1;
695 hrtimer_resolution = HIGH_RES_NSEC;
696
697 tick_setup_sched_timer();
698 /* "Retrigger" the interrupt to get things going */
699 retrigger_next_event(NULL);
700}
701
702static void clock_was_set_work(struct work_struct *work)
703{
704 clock_was_set();
705}
706
707static DECLARE_WORK(hrtimer_work, clock_was_set_work);
708
709/*
710 * Called from timekeeping and resume code to reprogramm the hrtimer
711 * interrupt device on all cpus.
712 */
713void clock_was_set_delayed(void)
714{
715 schedule_work(&hrtimer_work);
716}
717
718#else
719
720static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *b) { return 0; }
721static inline int hrtimer_hres_active(void) { return 0; }
722static inline int hrtimer_is_hres_enabled(void) { return 0; }
723static inline void hrtimer_switch_to_hres(void) { }
724static inline void
725hrtimer_force_reprogram(struct hrtimer_cpu_base *base, int skip_equal) { }
726static inline int hrtimer_reprogram(struct hrtimer *timer,
727 struct hrtimer_clock_base *base)
728{
729 return 0;
730}
731static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
732static inline void retrigger_next_event(void *arg) { }
733
734#endif /* CONFIG_HIGH_RES_TIMERS */
735
736/*
737 * Clock realtime was set
738 *
739 * Change the offset of the realtime clock vs. the monotonic
740 * clock.
741 *
742 * We might have to reprogram the high resolution timer interrupt. On
743 * SMP we call the architecture specific code to retrigger _all_ high
744 * resolution timer interrupts. On UP we just disable interrupts and
745 * call the high resolution interrupt code.
746 */
747void clock_was_set(void)
748{
749#ifdef CONFIG_HIGH_RES_TIMERS
750 /* Retrigger the CPU local events everywhere */
751 on_each_cpu(retrigger_next_event, NULL, 1);
752#endif
753 timerfd_clock_was_set();
754}
755
756/*
757 * During resume we might have to reprogram the high resolution timer
758 * interrupt on all online CPUs. However, all other CPUs will be
759 * stopped with IRQs interrupts disabled so the clock_was_set() call
760 * must be deferred.
761 */
762void hrtimers_resume(void)
763{
764 WARN_ONCE(!irqs_disabled(),
765 KERN_INFO "hrtimers_resume() called with IRQs enabled!");
766
767 /* Retrigger on the local CPU */
768 retrigger_next_event(NULL);
769 /* And schedule a retrigger for all others */
770 clock_was_set_delayed();
771}
772
773static inline void timer_stats_hrtimer_set_start_info(struct hrtimer *timer)
774{
775#ifdef CONFIG_TIMER_STATS
776 if (timer->start_site)
777 return;
778 timer->start_site = __builtin_return_address(0);
779 memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
780 timer->start_pid = current->pid;
781#endif
782}
783
784static inline void timer_stats_hrtimer_clear_start_info(struct hrtimer *timer)
785{
786#ifdef CONFIG_TIMER_STATS
787 timer->start_site = NULL;
788#endif
789}
790
791static inline void timer_stats_account_hrtimer(struct hrtimer *timer)
792{
793#ifdef CONFIG_TIMER_STATS
794 if (likely(!timer_stats_active))
795 return;
796 timer_stats_update_stats(timer, timer->start_pid, timer->start_site,
797 timer->function, timer->start_comm, 0);
798#endif
799}
800
801/*
802 * Counterpart to lock_hrtimer_base above:
803 */
804static inline
805void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
806{
807 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
808}
809
810/**
811 * hrtimer_forward - forward the timer expiry
812 * @timer: hrtimer to forward
813 * @now: forward past this time
814 * @interval: the interval to forward
815 *
816 * Forward the timer expiry so it will expire in the future.
817 * Returns the number of overruns.
818 *
819 * Can be safely called from the callback function of @timer. If
820 * called from other contexts @timer must neither be enqueued nor
821 * running the callback and the caller needs to take care of
822 * serialization.
823 *
824 * Note: This only updates the timer expiry value and does not requeue
825 * the timer.
826 */
827u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
828{
829 u64 orun = 1;
830 ktime_t delta;
831
832 delta = ktime_sub(now, hrtimer_get_expires(timer));
833
834 if (delta.tv64 < 0)
835 return 0;
836
837 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
838 return 0;
839
840 if (interval.tv64 < hrtimer_resolution)
841 interval.tv64 = hrtimer_resolution;
842
843 if (unlikely(delta.tv64 >= interval.tv64)) {
844 s64 incr = ktime_to_ns(interval);
845
846 orun = ktime_divns(delta, incr);
847 hrtimer_add_expires_ns(timer, incr * orun);
848 if (hrtimer_get_expires_tv64(timer) > now.tv64)
849 return orun;
850 /*
851 * This (and the ktime_add() below) is the
852 * correction for exact:
853 */
854 orun++;
855 }
856 hrtimer_add_expires(timer, interval);
857
858 return orun;
859}
860EXPORT_SYMBOL_GPL(hrtimer_forward);
861
862/*
863 * enqueue_hrtimer - internal function to (re)start a timer
864 *
865 * The timer is inserted in expiry order. Insertion into the
866 * red black tree is O(log(n)). Must hold the base lock.
867 *
868 * Returns 1 when the new timer is the leftmost timer in the tree.
869 */
870static int enqueue_hrtimer(struct hrtimer *timer,
871 struct hrtimer_clock_base *base)
872{
873 debug_activate(timer);
874
875 base->cpu_base->active_bases |= 1 << base->index;
876
877 timer->state = HRTIMER_STATE_ENQUEUED;
878
879 return timerqueue_add(&base->active, &timer->node);
880}
881
882/*
883 * __remove_hrtimer - internal function to remove a timer
884 *
885 * Caller must hold the base lock.
886 *
887 * High resolution timer mode reprograms the clock event device when the
888 * timer is the one which expires next. The caller can disable this by setting
889 * reprogram to zero. This is useful, when the context does a reprogramming
890 * anyway (e.g. timer interrupt)
891 */
892static void __remove_hrtimer(struct hrtimer *timer,
893 struct hrtimer_clock_base *base,
894 u8 newstate, int reprogram)
895{
896 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
897 u8 state = timer->state;
898
899 timer->state = newstate;
900 if (!(state & HRTIMER_STATE_ENQUEUED))
901 return;
902
903 if (!timerqueue_del(&base->active, &timer->node))
904 cpu_base->active_bases &= ~(1 << base->index);
905
906#ifdef CONFIG_HIGH_RES_TIMERS
907 /*
908 * Note: If reprogram is false we do not update
909 * cpu_base->next_timer. This happens when we remove the first
910 * timer on a remote cpu. No harm as we never dereference
911 * cpu_base->next_timer. So the worst thing what can happen is
912 * an superflous call to hrtimer_force_reprogram() on the
913 * remote cpu later on if the same timer gets enqueued again.
914 */
915 if (reprogram && timer == cpu_base->next_timer)
916 hrtimer_force_reprogram(cpu_base, 1);
917#endif
918}
919
920/*
921 * remove hrtimer, called with base lock held
922 */
923static inline int
924remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base, bool restart)
925{
926 if (hrtimer_is_queued(timer)) {
927 u8 state = timer->state;
928 int reprogram;
929
930 /*
931 * Remove the timer and force reprogramming when high
932 * resolution mode is active and the timer is on the current
933 * CPU. If we remove a timer on another CPU, reprogramming is
934 * skipped. The interrupt event on this CPU is fired and
935 * reprogramming happens in the interrupt handler. This is a
936 * rare case and less expensive than a smp call.
937 */
938 debug_deactivate(timer);
939 timer_stats_hrtimer_clear_start_info(timer);
940 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
941
942 if (!restart)
943 state = HRTIMER_STATE_INACTIVE;
944
945 __remove_hrtimer(timer, base, state, reprogram);
946 return 1;
947 }
948 return 0;
949}
950
951static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
952 const enum hrtimer_mode mode)
953{
954#ifdef CONFIG_TIME_LOW_RES
955 /*
956 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
957 * granular time values. For relative timers we add hrtimer_resolution
958 * (i.e. one jiffie) to prevent short timeouts.
959 */
960 timer->is_rel = mode & HRTIMER_MODE_REL;
961 if (timer->is_rel)
962 tim = ktime_add_safe(tim, ktime_set(0, hrtimer_resolution));
963#endif
964 return tim;
965}
966
967/**
968 * hrtimer_start_range_ns - (re)start an hrtimer on the current CPU
969 * @timer: the timer to be added
970 * @tim: expiry time
971 * @delta_ns: "slack" range for the timer
972 * @mode: expiry mode: absolute (HRTIMER_MODE_ABS) or
973 * relative (HRTIMER_MODE_REL)
974 */
975void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
976 u64 delta_ns, const enum hrtimer_mode mode)
977{
978 struct hrtimer_clock_base *base, *new_base;
979 unsigned long flags;
980 int leftmost;
981
982 base = lock_hrtimer_base(timer, &flags);
983
984 /* Remove an active timer from the queue: */
985 remove_hrtimer(timer, base, true);
986
987 if (mode & HRTIMER_MODE_REL)
988 tim = ktime_add_safe(tim, base->get_time());
989
990 tim = hrtimer_update_lowres(timer, tim, mode);
991
992 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
993
994 /* Switch the timer base, if necessary: */
995 new_base = switch_hrtimer_base(timer, base, mode & HRTIMER_MODE_PINNED);
996
997 timer_stats_hrtimer_set_start_info(timer);
998
999 leftmost = enqueue_hrtimer(timer, new_base);
1000 if (!leftmost)
1001 goto unlock;
1002
1003 if (!hrtimer_is_hres_active(timer)) {
1004 /*
1005 * Kick to reschedule the next tick to handle the new timer
1006 * on dynticks target.
1007 */
1008 if (new_base->cpu_base->nohz_active)
1009 wake_up_nohz_cpu(new_base->cpu_base->cpu);
1010 } else {
1011 hrtimer_reprogram(timer, new_base);
1012 }
1013unlock:
1014 unlock_hrtimer_base(timer, &flags);
1015}
1016EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1017
1018/**
1019 * hrtimer_try_to_cancel - try to deactivate a timer
1020 * @timer: hrtimer to stop
1021 *
1022 * Returns:
1023 * 0 when the timer was not active
1024 * 1 when the timer was active
1025 * -1 when the timer is currently excuting the callback function and
1026 * cannot be stopped
1027 */
1028int hrtimer_try_to_cancel(struct hrtimer *timer)
1029{
1030 struct hrtimer_clock_base *base;
1031 unsigned long flags;
1032 int ret = -1;
1033
1034 /*
1035 * Check lockless first. If the timer is not active (neither
1036 * enqueued nor running the callback, nothing to do here. The
1037 * base lock does not serialize against a concurrent enqueue,
1038 * so we can avoid taking it.
1039 */
1040 if (!hrtimer_active(timer))
1041 return 0;
1042
1043 base = lock_hrtimer_base(timer, &flags);
1044
1045 if (!hrtimer_callback_running(timer))
1046 ret = remove_hrtimer(timer, base, false);
1047
1048 unlock_hrtimer_base(timer, &flags);
1049
1050 return ret;
1051
1052}
1053EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1054
1055/**
1056 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1057 * @timer: the timer to be cancelled
1058 *
1059 * Returns:
1060 * 0 when the timer was not active
1061 * 1 when the timer was active
1062 */
1063int hrtimer_cancel(struct hrtimer *timer)
1064{
1065 for (;;) {
1066 int ret = hrtimer_try_to_cancel(timer);
1067
1068 if (ret >= 0)
1069 return ret;
1070 cpu_relax();
1071 }
1072}
1073EXPORT_SYMBOL_GPL(hrtimer_cancel);
1074
1075/**
1076 * hrtimer_get_remaining - get remaining time for the timer
1077 * @timer: the timer to read
1078 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1079 */
1080ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1081{
1082 unsigned long flags;
1083 ktime_t rem;
1084
1085 lock_hrtimer_base(timer, &flags);
1086 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1087 rem = hrtimer_expires_remaining_adjusted(timer);
1088 else
1089 rem = hrtimer_expires_remaining(timer);
1090 unlock_hrtimer_base(timer, &flags);
1091
1092 return rem;
1093}
1094EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1095
1096#ifdef CONFIG_NO_HZ_COMMON
1097/**
1098 * hrtimer_get_next_event - get the time until next expiry event
1099 *
1100 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1101 */
1102u64 hrtimer_get_next_event(void)
1103{
1104 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1105 u64 expires = KTIME_MAX;
1106 unsigned long flags;
1107
1108 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1109
1110 if (!__hrtimer_hres_active(cpu_base))
1111 expires = __hrtimer_get_next_event(cpu_base).tv64;
1112
1113 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1114
1115 return expires;
1116}
1117#endif
1118
1119static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1120 enum hrtimer_mode mode)
1121{
1122 struct hrtimer_cpu_base *cpu_base;
1123 int base;
1124
1125 memset(timer, 0, sizeof(struct hrtimer));
1126
1127 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1128
1129 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
1130 clock_id = CLOCK_MONOTONIC;
1131
1132 base = hrtimer_clockid_to_base(clock_id);
1133 timer->base = &cpu_base->clock_base[base];
1134 timerqueue_init(&timer->node);
1135
1136#ifdef CONFIG_TIMER_STATS
1137 timer->start_site = NULL;
1138 timer->start_pid = -1;
1139 memset(timer->start_comm, 0, TASK_COMM_LEN);
1140#endif
1141}
1142
1143/**
1144 * hrtimer_init - initialize a timer to the given clock
1145 * @timer: the timer to be initialized
1146 * @clock_id: the clock to be used
1147 * @mode: timer mode abs/rel
1148 */
1149void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1150 enum hrtimer_mode mode)
1151{
1152 debug_init(timer, clock_id, mode);
1153 __hrtimer_init(timer, clock_id, mode);
1154}
1155EXPORT_SYMBOL_GPL(hrtimer_init);
1156
1157/*
1158 * A timer is active, when it is enqueued into the rbtree or the
1159 * callback function is running or it's in the state of being migrated
1160 * to another cpu.
1161 *
1162 * It is important for this function to not return a false negative.
1163 */
1164bool hrtimer_active(const struct hrtimer *timer)
1165{
1166 struct hrtimer_cpu_base *cpu_base;
1167 unsigned int seq;
1168
1169 do {
1170 cpu_base = READ_ONCE(timer->base->cpu_base);
1171 seq = raw_read_seqcount_begin(&cpu_base->seq);
1172
1173 if (timer->state != HRTIMER_STATE_INACTIVE ||
1174 cpu_base->running == timer)
1175 return true;
1176
1177 } while (read_seqcount_retry(&cpu_base->seq, seq) ||
1178 cpu_base != READ_ONCE(timer->base->cpu_base));
1179
1180 return false;
1181}
1182EXPORT_SYMBOL_GPL(hrtimer_active);
1183
1184/*
1185 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1186 * distinct sections:
1187 *
1188 * - queued: the timer is queued
1189 * - callback: the timer is being ran
1190 * - post: the timer is inactive or (re)queued
1191 *
1192 * On the read side we ensure we observe timer->state and cpu_base->running
1193 * from the same section, if anything changed while we looked at it, we retry.
1194 * This includes timer->base changing because sequence numbers alone are
1195 * insufficient for that.
1196 *
1197 * The sequence numbers are required because otherwise we could still observe
1198 * a false negative if the read side got smeared over multiple consequtive
1199 * __run_hrtimer() invocations.
1200 */
1201
1202static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1203 struct hrtimer_clock_base *base,
1204 struct hrtimer *timer, ktime_t *now)
1205{
1206 enum hrtimer_restart (*fn)(struct hrtimer *);
1207 int restart;
1208
1209 lockdep_assert_held(&cpu_base->lock);
1210
1211 debug_deactivate(timer);
1212 cpu_base->running = timer;
1213
1214 /*
1215 * Separate the ->running assignment from the ->state assignment.
1216 *
1217 * As with a regular write barrier, this ensures the read side in
1218 * hrtimer_active() cannot observe cpu_base->running == NULL &&
1219 * timer->state == INACTIVE.
1220 */
1221 raw_write_seqcount_barrier(&cpu_base->seq);
1222
1223 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1224 timer_stats_account_hrtimer(timer);
1225 fn = timer->function;
1226
1227 /*
1228 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1229 * timer is restarted with a period then it becomes an absolute
1230 * timer. If its not restarted it does not matter.
1231 */
1232 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1233 timer->is_rel = false;
1234
1235 /*
1236 * Because we run timers from hardirq context, there is no chance
1237 * they get migrated to another cpu, therefore its safe to unlock
1238 * the timer base.
1239 */
1240 raw_spin_unlock(&cpu_base->lock);
1241 trace_hrtimer_expire_entry(timer, now);
1242 restart = fn(timer);
1243 trace_hrtimer_expire_exit(timer);
1244 raw_spin_lock(&cpu_base->lock);
1245
1246 /*
1247 * Note: We clear the running state after enqueue_hrtimer and
1248 * we do not reprogramm the event hardware. Happens either in
1249 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1250 *
1251 * Note: Because we dropped the cpu_base->lock above,
1252 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1253 * for us already.
1254 */
1255 if (restart != HRTIMER_NORESTART &&
1256 !(timer->state & HRTIMER_STATE_ENQUEUED))
1257 enqueue_hrtimer(timer, base);
1258
1259 /*
1260 * Separate the ->running assignment from the ->state assignment.
1261 *
1262 * As with a regular write barrier, this ensures the read side in
1263 * hrtimer_active() cannot observe cpu_base->running == NULL &&
1264 * timer->state == INACTIVE.
1265 */
1266 raw_write_seqcount_barrier(&cpu_base->seq);
1267
1268 WARN_ON_ONCE(cpu_base->running != timer);
1269 cpu_base->running = NULL;
1270}
1271
1272static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now)
1273{
1274 struct hrtimer_clock_base *base = cpu_base->clock_base;
1275 unsigned int active = cpu_base->active_bases;
1276
1277 for (; active; base++, active >>= 1) {
1278 struct timerqueue_node *node;
1279 ktime_t basenow;
1280
1281 if (!(active & 0x01))
1282 continue;
1283
1284 basenow = ktime_add(now, base->offset);
1285
1286 while ((node = timerqueue_getnext(&base->active))) {
1287 struct hrtimer *timer;
1288
1289 timer = container_of(node, struct hrtimer, node);
1290
1291 /*
1292 * The immediate goal for using the softexpires is
1293 * minimizing wakeups, not running timers at the
1294 * earliest interrupt after their soft expiration.
1295 * This allows us to avoid using a Priority Search
1296 * Tree, which can answer a stabbing querry for
1297 * overlapping intervals and instead use the simple
1298 * BST we already have.
1299 * We don't add extra wakeups by delaying timers that
1300 * are right-of a not yet expired timer, because that
1301 * timer will have to trigger a wakeup anyway.
1302 */
1303 if (basenow.tv64 < hrtimer_get_softexpires_tv64(timer))
1304 break;
1305
1306 __run_hrtimer(cpu_base, base, timer, &basenow);
1307 }
1308 }
1309}
1310
1311#ifdef CONFIG_HIGH_RES_TIMERS
1312
1313/*
1314 * High resolution timer interrupt
1315 * Called with interrupts disabled
1316 */
1317void hrtimer_interrupt(struct clock_event_device *dev)
1318{
1319 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1320 ktime_t expires_next, now, entry_time, delta;
1321 int retries = 0;
1322
1323 BUG_ON(!cpu_base->hres_active);
1324 cpu_base->nr_events++;
1325 dev->next_event.tv64 = KTIME_MAX;
1326
1327 raw_spin_lock(&cpu_base->lock);
1328 entry_time = now = hrtimer_update_base(cpu_base);
1329retry:
1330 cpu_base->in_hrtirq = 1;
1331 /*
1332 * We set expires_next to KTIME_MAX here with cpu_base->lock
1333 * held to prevent that a timer is enqueued in our queue via
1334 * the migration code. This does not affect enqueueing of
1335 * timers which run their callback and need to be requeued on
1336 * this CPU.
1337 */
1338 cpu_base->expires_next.tv64 = KTIME_MAX;
1339
1340 __hrtimer_run_queues(cpu_base, now);
1341
1342 /* Reevaluate the clock bases for the next expiry */
1343 expires_next = __hrtimer_get_next_event(cpu_base);
1344 /*
1345 * Store the new expiry value so the migration code can verify
1346 * against it.
1347 */
1348 cpu_base->expires_next = expires_next;
1349 cpu_base->in_hrtirq = 0;
1350 raw_spin_unlock(&cpu_base->lock);
1351
1352 /* Reprogramming necessary ? */
1353 if (!tick_program_event(expires_next, 0)) {
1354 cpu_base->hang_detected = 0;
1355 return;
1356 }
1357
1358 /*
1359 * The next timer was already expired due to:
1360 * - tracing
1361 * - long lasting callbacks
1362 * - being scheduled away when running in a VM
1363 *
1364 * We need to prevent that we loop forever in the hrtimer
1365 * interrupt routine. We give it 3 attempts to avoid
1366 * overreacting on some spurious event.
1367 *
1368 * Acquire base lock for updating the offsets and retrieving
1369 * the current time.
1370 */
1371 raw_spin_lock(&cpu_base->lock);
1372 now = hrtimer_update_base(cpu_base);
1373 cpu_base->nr_retries++;
1374 if (++retries < 3)
1375 goto retry;
1376 /*
1377 * Give the system a chance to do something else than looping
1378 * here. We stored the entry time, so we know exactly how long
1379 * we spent here. We schedule the next event this amount of
1380 * time away.
1381 */
1382 cpu_base->nr_hangs++;
1383 cpu_base->hang_detected = 1;
1384 raw_spin_unlock(&cpu_base->lock);
1385 delta = ktime_sub(now, entry_time);
1386 if ((unsigned int)delta.tv64 > cpu_base->max_hang_time)
1387 cpu_base->max_hang_time = (unsigned int) delta.tv64;
1388 /*
1389 * Limit it to a sensible value as we enforce a longer
1390 * delay. Give the CPU at least 100ms to catch up.
1391 */
1392 if (delta.tv64 > 100 * NSEC_PER_MSEC)
1393 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1394 else
1395 expires_next = ktime_add(now, delta);
1396 tick_program_event(expires_next, 1);
1397 printk_once(KERN_WARNING "hrtimer: interrupt took %llu ns\n",
1398 ktime_to_ns(delta));
1399}
1400
1401/*
1402 * local version of hrtimer_peek_ahead_timers() called with interrupts
1403 * disabled.
1404 */
1405static inline void __hrtimer_peek_ahead_timers(void)
1406{
1407 struct tick_device *td;
1408
1409 if (!hrtimer_hres_active())
1410 return;
1411
1412 td = this_cpu_ptr(&tick_cpu_device);
1413 if (td && td->evtdev)
1414 hrtimer_interrupt(td->evtdev);
1415}
1416
1417#else /* CONFIG_HIGH_RES_TIMERS */
1418
1419static inline void __hrtimer_peek_ahead_timers(void) { }
1420
1421#endif /* !CONFIG_HIGH_RES_TIMERS */
1422
1423/*
1424 * Called from run_local_timers in hardirq context every jiffy
1425 */
1426void hrtimer_run_queues(void)
1427{
1428 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1429 ktime_t now;
1430
1431 if (__hrtimer_hres_active(cpu_base))
1432 return;
1433
1434 /*
1435 * This _is_ ugly: We have to check periodically, whether we
1436 * can switch to highres and / or nohz mode. The clocksource
1437 * switch happens with xtime_lock held. Notification from
1438 * there only sets the check bit in the tick_oneshot code,
1439 * otherwise we might deadlock vs. xtime_lock.
1440 */
1441 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1442 hrtimer_switch_to_hres();
1443 return;
1444 }
1445
1446 raw_spin_lock(&cpu_base->lock);
1447 now = hrtimer_update_base(cpu_base);
1448 __hrtimer_run_queues(cpu_base, now);
1449 raw_spin_unlock(&cpu_base->lock);
1450}
1451
1452/*
1453 * Sleep related functions:
1454 */
1455static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1456{
1457 struct hrtimer_sleeper *t =
1458 container_of(timer, struct hrtimer_sleeper, timer);
1459 struct task_struct *task = t->task;
1460
1461 t->task = NULL;
1462 if (task)
1463 wake_up_process(task);
1464
1465 return HRTIMER_NORESTART;
1466}
1467
1468void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1469{
1470 sl->timer.function = hrtimer_wakeup;
1471 sl->task = task;
1472}
1473EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1474
1475static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1476{
1477 hrtimer_init_sleeper(t, current);
1478
1479 do {
1480 set_current_state(TASK_INTERRUPTIBLE);
1481 hrtimer_start_expires(&t->timer, mode);
1482
1483 if (likely(t->task))
1484 freezable_schedule();
1485
1486 hrtimer_cancel(&t->timer);
1487 mode = HRTIMER_MODE_ABS;
1488
1489 } while (t->task && !signal_pending(current));
1490
1491 __set_current_state(TASK_RUNNING);
1492
1493 return t->task == NULL;
1494}
1495
1496static int update_rmtp(struct hrtimer *timer, struct timespec __user *rmtp)
1497{
1498 struct timespec rmt;
1499 ktime_t rem;
1500
1501 rem = hrtimer_expires_remaining(timer);
1502 if (rem.tv64 <= 0)
1503 return 0;
1504 rmt = ktime_to_timespec(rem);
1505
1506 if (copy_to_user(rmtp, &rmt, sizeof(*rmtp)))
1507 return -EFAULT;
1508
1509 return 1;
1510}
1511
1512long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1513{
1514 struct hrtimer_sleeper t;
1515 struct timespec __user *rmtp;
1516 int ret = 0;
1517
1518 hrtimer_init_on_stack(&t.timer, restart->nanosleep.clockid,
1519 HRTIMER_MODE_ABS);
1520 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
1521
1522 if (do_nanosleep(&t, HRTIMER_MODE_ABS))
1523 goto out;
1524
1525 rmtp = restart->nanosleep.rmtp;
1526 if (rmtp) {
1527 ret = update_rmtp(&t.timer, rmtp);
1528 if (ret <= 0)
1529 goto out;
1530 }
1531
1532 /* The other values in restart are already filled in */
1533 ret = -ERESTART_RESTARTBLOCK;
1534out:
1535 destroy_hrtimer_on_stack(&t.timer);
1536 return ret;
1537}
1538
1539long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
1540 const enum hrtimer_mode mode, const clockid_t clockid)
1541{
1542 struct restart_block *restart;
1543 struct hrtimer_sleeper t;
1544 int ret = 0;
1545 u64 slack;
1546
1547 slack = current->timer_slack_ns;
1548 if (dl_task(current) || rt_task(current))
1549 slack = 0;
1550
1551 hrtimer_init_on_stack(&t.timer, clockid, mode);
1552 hrtimer_set_expires_range_ns(&t.timer, timespec_to_ktime(*rqtp), slack);
1553 if (do_nanosleep(&t, mode))
1554 goto out;
1555
1556 /* Absolute timers do not update the rmtp value and restart: */
1557 if (mode == HRTIMER_MODE_ABS) {
1558 ret = -ERESTARTNOHAND;
1559 goto out;
1560 }
1561
1562 if (rmtp) {
1563 ret = update_rmtp(&t.timer, rmtp);
1564 if (ret <= 0)
1565 goto out;
1566 }
1567
1568 restart = ¤t->restart_block;
1569 restart->fn = hrtimer_nanosleep_restart;
1570 restart->nanosleep.clockid = t.timer.base->clockid;
1571 restart->nanosleep.rmtp = rmtp;
1572 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
1573
1574 ret = -ERESTART_RESTARTBLOCK;
1575out:
1576 destroy_hrtimer_on_stack(&t.timer);
1577 return ret;
1578}
1579
1580SYSCALL_DEFINE2(nanosleep, struct timespec __user *, rqtp,
1581 struct timespec __user *, rmtp)
1582{
1583 struct timespec tu;
1584
1585 if (copy_from_user(&tu, rqtp, sizeof(tu)))
1586 return -EFAULT;
1587
1588 if (!timespec_valid(&tu))
1589 return -EINVAL;
1590
1591 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1592}
1593
1594/*
1595 * Functions related to boot-time initialization:
1596 */
1597static void init_hrtimers_cpu(int cpu)
1598{
1599 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1600 int i;
1601
1602 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1603 cpu_base->clock_base[i].cpu_base = cpu_base;
1604 timerqueue_init_head(&cpu_base->clock_base[i].active);
1605 }
1606
1607 cpu_base->cpu = cpu;
1608 hrtimer_init_hres(cpu_base);
1609}
1610
1611#ifdef CONFIG_HOTPLUG_CPU
1612
1613static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1614 struct hrtimer_clock_base *new_base)
1615{
1616 struct hrtimer *timer;
1617 struct timerqueue_node *node;
1618
1619 while ((node = timerqueue_getnext(&old_base->active))) {
1620 timer = container_of(node, struct hrtimer, node);
1621 BUG_ON(hrtimer_callback_running(timer));
1622 debug_deactivate(timer);
1623
1624 /*
1625 * Mark it as ENQUEUED not INACTIVE otherwise the
1626 * timer could be seen as !active and just vanish away
1627 * under us on another CPU
1628 */
1629 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
1630 timer->base = new_base;
1631 /*
1632 * Enqueue the timers on the new cpu. This does not
1633 * reprogram the event device in case the timer
1634 * expires before the earliest on this CPU, but we run
1635 * hrtimer_interrupt after we migrated everything to
1636 * sort out already expired timers and reprogram the
1637 * event device.
1638 */
1639 enqueue_hrtimer(timer, new_base);
1640 }
1641}
1642
1643static void migrate_hrtimers(int scpu)
1644{
1645 struct hrtimer_cpu_base *old_base, *new_base;
1646 int i;
1647
1648 BUG_ON(cpu_online(scpu));
1649 tick_cancel_sched_timer(scpu);
1650
1651 local_irq_disable();
1652 old_base = &per_cpu(hrtimer_bases, scpu);
1653 new_base = this_cpu_ptr(&hrtimer_bases);
1654 /*
1655 * The caller is globally serialized and nobody else
1656 * takes two locks at once, deadlock is not possible.
1657 */
1658 raw_spin_lock(&new_base->lock);
1659 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1660
1661 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1662 migrate_hrtimer_list(&old_base->clock_base[i],
1663 &new_base->clock_base[i]);
1664 }
1665
1666 raw_spin_unlock(&old_base->lock);
1667 raw_spin_unlock(&new_base->lock);
1668
1669 /* Check, if we got expired work to do */
1670 __hrtimer_peek_ahead_timers();
1671 local_irq_enable();
1672}
1673
1674#endif /* CONFIG_HOTPLUG_CPU */
1675
1676static int hrtimer_cpu_notify(struct notifier_block *self,
1677 unsigned long action, void *hcpu)
1678{
1679 int scpu = (long)hcpu;
1680
1681 switch (action) {
1682
1683 case CPU_UP_PREPARE:
1684 case CPU_UP_PREPARE_FROZEN:
1685 init_hrtimers_cpu(scpu);
1686 break;
1687
1688#ifdef CONFIG_HOTPLUG_CPU
1689 case CPU_DEAD:
1690 case CPU_DEAD_FROZEN:
1691 migrate_hrtimers(scpu);
1692 break;
1693#endif
1694
1695 default:
1696 break;
1697 }
1698
1699 return NOTIFY_OK;
1700}
1701
1702static struct notifier_block hrtimers_nb = {
1703 .notifier_call = hrtimer_cpu_notify,
1704};
1705
1706void __init hrtimers_init(void)
1707{
1708 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
1709 (void *)(long)smp_processor_id());
1710 register_cpu_notifier(&hrtimers_nb);
1711}
1712
1713/**
1714 * schedule_hrtimeout_range_clock - sleep until timeout
1715 * @expires: timeout value (ktime_t)
1716 * @delta: slack in expires timeout (ktime_t)
1717 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1718 * @clock: timer clock, CLOCK_MONOTONIC or CLOCK_REALTIME
1719 */
1720int __sched
1721schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
1722 const enum hrtimer_mode mode, int clock)
1723{
1724 struct hrtimer_sleeper t;
1725
1726 /*
1727 * Optimize when a zero timeout value is given. It does not
1728 * matter whether this is an absolute or a relative time.
1729 */
1730 if (expires && !expires->tv64) {
1731 __set_current_state(TASK_RUNNING);
1732 return 0;
1733 }
1734
1735 /*
1736 * A NULL parameter means "infinite"
1737 */
1738 if (!expires) {
1739 schedule();
1740 return -EINTR;
1741 }
1742
1743 hrtimer_init_on_stack(&t.timer, clock, mode);
1744 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
1745
1746 hrtimer_init_sleeper(&t, current);
1747
1748 hrtimer_start_expires(&t.timer, mode);
1749
1750 if (likely(t.task))
1751 schedule();
1752
1753 hrtimer_cancel(&t.timer);
1754 destroy_hrtimer_on_stack(&t.timer);
1755
1756 __set_current_state(TASK_RUNNING);
1757
1758 return !t.task ? 0 : -EINTR;
1759}
1760
1761/**
1762 * schedule_hrtimeout_range - sleep until timeout
1763 * @expires: timeout value (ktime_t)
1764 * @delta: slack in expires timeout (ktime_t)
1765 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1766 *
1767 * Make the current task sleep until the given expiry time has
1768 * elapsed. The routine will return immediately unless
1769 * the current task state has been set (see set_current_state()).
1770 *
1771 * The @delta argument gives the kernel the freedom to schedule the
1772 * actual wakeup to a time that is both power and performance friendly.
1773 * The kernel give the normal best effort behavior for "@expires+@delta",
1774 * but may decide to fire the timer earlier, but no earlier than @expires.
1775 *
1776 * You can set the task state as follows -
1777 *
1778 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1779 * pass before the routine returns.
1780 *
1781 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1782 * delivered to the current task.
1783 *
1784 * The current task state is guaranteed to be TASK_RUNNING when this
1785 * routine returns.
1786 *
1787 * Returns 0 when the timer has expired otherwise -EINTR
1788 */
1789int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
1790 const enum hrtimer_mode mode)
1791{
1792 return schedule_hrtimeout_range_clock(expires, delta, mode,
1793 CLOCK_MONOTONIC);
1794}
1795EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
1796
1797/**
1798 * schedule_hrtimeout - sleep until timeout
1799 * @expires: timeout value (ktime_t)
1800 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1801 *
1802 * Make the current task sleep until the given expiry time has
1803 * elapsed. The routine will return immediately unless
1804 * the current task state has been set (see set_current_state()).
1805 *
1806 * You can set the task state as follows -
1807 *
1808 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1809 * pass before the routine returns.
1810 *
1811 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1812 * delivered to the current task.
1813 *
1814 * The current task state is guaranteed to be TASK_RUNNING when this
1815 * routine returns.
1816 *
1817 * Returns 0 when the timer has expired otherwise -EINTR
1818 */
1819int __sched schedule_hrtimeout(ktime_t *expires,
1820 const enum hrtimer_mode mode)
1821{
1822 return schedule_hrtimeout_range(expires, 0, mode);
1823}
1824EXPORT_SYMBOL_GPL(schedule_hrtimeout);
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 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API, aka timer wheel,
10 * hrtimers provide finer resolution and accuracy depending on system
11 * configuration and capabilities.
12 *
13 * Started by: Thomas Gleixner and Ingo Molnar
14 *
15 * Credits:
16 * Based on the original timer wheel code
17 *
18 * Help, testing, suggestions, bugfixes, improvements were
19 * provided by:
20 *
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
22 * et. al.
23 */
24
25#include <linux/cpu.h>
26#include <linux/export.h>
27#include <linux/percpu.h>
28#include <linux/hrtimer.h>
29#include <linux/notifier.h>
30#include <linux/syscalls.h>
31#include <linux/interrupt.h>
32#include <linux/tick.h>
33#include <linux/err.h>
34#include <linux/debugobjects.h>
35#include <linux/sched/signal.h>
36#include <linux/sched/sysctl.h>
37#include <linux/sched/rt.h>
38#include <linux/sched/deadline.h>
39#include <linux/sched/nohz.h>
40#include <linux/sched/debug.h>
41#include <linux/sched/isolation.h>
42#include <linux/timer.h>
43#include <linux/freezer.h>
44#include <linux/compat.h>
45
46#include <linux/uaccess.h>
47
48#include <trace/events/timer.h>
49
50#include "tick-internal.h"
51
52/*
53 * Masks for selecting the soft and hard context timers from
54 * cpu_base->active
55 */
56#define MASK_SHIFT (HRTIMER_BASE_MONOTONIC_SOFT)
57#define HRTIMER_ACTIVE_HARD ((1U << MASK_SHIFT) - 1)
58#define HRTIMER_ACTIVE_SOFT (HRTIMER_ACTIVE_HARD << MASK_SHIFT)
59#define HRTIMER_ACTIVE_ALL (HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
60
61static void retrigger_next_event(void *arg);
62
63/*
64 * The timer bases:
65 *
66 * There are more clockids than hrtimer bases. Thus, we index
67 * into the timer bases by the hrtimer_base_type enum. When trying
68 * to reach a base using a clockid, hrtimer_clockid_to_base()
69 * is used to convert from clockid to the proper hrtimer_base_type.
70 */
71DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
72{
73 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
74 .clock_base =
75 {
76 {
77 .index = HRTIMER_BASE_MONOTONIC,
78 .clockid = CLOCK_MONOTONIC,
79 .get_time = &ktime_get,
80 },
81 {
82 .index = HRTIMER_BASE_REALTIME,
83 .clockid = CLOCK_REALTIME,
84 .get_time = &ktime_get_real,
85 },
86 {
87 .index = HRTIMER_BASE_BOOTTIME,
88 .clockid = CLOCK_BOOTTIME,
89 .get_time = &ktime_get_boottime,
90 },
91 {
92 .index = HRTIMER_BASE_TAI,
93 .clockid = CLOCK_TAI,
94 .get_time = &ktime_get_clocktai,
95 },
96 {
97 .index = HRTIMER_BASE_MONOTONIC_SOFT,
98 .clockid = CLOCK_MONOTONIC,
99 .get_time = &ktime_get,
100 },
101 {
102 .index = HRTIMER_BASE_REALTIME_SOFT,
103 .clockid = CLOCK_REALTIME,
104 .get_time = &ktime_get_real,
105 },
106 {
107 .index = HRTIMER_BASE_BOOTTIME_SOFT,
108 .clockid = CLOCK_BOOTTIME,
109 .get_time = &ktime_get_boottime,
110 },
111 {
112 .index = HRTIMER_BASE_TAI_SOFT,
113 .clockid = CLOCK_TAI,
114 .get_time = &ktime_get_clocktai,
115 },
116 },
117 .csd = CSD_INIT(retrigger_next_event, NULL)
118};
119
120static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
121 /* Make sure we catch unsupported clockids */
122 [0 ... MAX_CLOCKS - 1] = HRTIMER_MAX_CLOCK_BASES,
123
124 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
125 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
126 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
127 [CLOCK_TAI] = HRTIMER_BASE_TAI,
128};
129
130static inline bool hrtimer_base_is_online(struct hrtimer_cpu_base *base)
131{
132 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
133 return true;
134 else
135 return likely(base->online);
136}
137
138/*
139 * Functions and macros which are different for UP/SMP systems are kept in a
140 * single place
141 */
142#ifdef CONFIG_SMP
143
144/*
145 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
146 * such that hrtimer_callback_running() can unconditionally dereference
147 * timer->base->cpu_base
148 */
149static struct hrtimer_cpu_base migration_cpu_base = {
150 .clock_base = { {
151 .cpu_base = &migration_cpu_base,
152 .seq = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
153 &migration_cpu_base.lock),
154 }, },
155};
156
157#define migration_base migration_cpu_base.clock_base[0]
158
159static inline bool is_migration_base(struct hrtimer_clock_base *base)
160{
161 return base == &migration_base;
162}
163
164/*
165 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
166 * means that all timers which are tied to this base via timer->base are
167 * locked, and the base itself is locked too.
168 *
169 * So __run_timers/migrate_timers can safely modify all timers which could
170 * be found on the lists/queues.
171 *
172 * When the timer's base is locked, and the timer removed from list, it is
173 * possible to set timer->base = &migration_base and drop the lock: the timer
174 * remains locked.
175 */
176static
177struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
178 unsigned long *flags)
179 __acquires(&timer->base->lock)
180{
181 struct hrtimer_clock_base *base;
182
183 for (;;) {
184 base = READ_ONCE(timer->base);
185 if (likely(base != &migration_base)) {
186 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
187 if (likely(base == timer->base))
188 return base;
189 /* The timer has migrated to another CPU: */
190 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
191 }
192 cpu_relax();
193 }
194}
195
196/*
197 * Check if the elected target is suitable considering its next
198 * event and the hotplug state of the current CPU.
199 *
200 * If the elected target is remote and its next event is after the timer
201 * to queue, then a remote reprogram is necessary. However there is no
202 * guarantee the IPI handling the operation would arrive in time to meet
203 * the high resolution deadline. In this case the local CPU becomes a
204 * preferred target, unless it is offline.
205 *
206 * High and low resolution modes are handled the same way for simplicity.
207 *
208 * Called with cpu_base->lock of target cpu held.
209 */
210static bool hrtimer_suitable_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base,
211 struct hrtimer_cpu_base *new_cpu_base,
212 struct hrtimer_cpu_base *this_cpu_base)
213{
214 ktime_t expires;
215
216 /*
217 * The local CPU clockevent can be reprogrammed. Also get_target_base()
218 * guarantees it is online.
219 */
220 if (new_cpu_base == this_cpu_base)
221 return true;
222
223 /*
224 * The offline local CPU can't be the default target if the
225 * next remote target event is after this timer. Keep the
226 * elected new base. An IPI will we issued to reprogram
227 * it as a last resort.
228 */
229 if (!hrtimer_base_is_online(this_cpu_base))
230 return true;
231
232 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
233
234 return expires >= new_base->cpu_base->expires_next;
235}
236
237static inline struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base, int pinned)
238{
239 if (!hrtimer_base_is_online(base)) {
240 int cpu = cpumask_any_and(cpu_online_mask, housekeeping_cpumask(HK_TYPE_TIMER));
241
242 return &per_cpu(hrtimer_bases, cpu);
243 }
244
245#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
246 if (static_branch_likely(&timers_migration_enabled) && !pinned)
247 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
248#endif
249 return base;
250}
251
252/*
253 * We switch the timer base to a power-optimized selected CPU target,
254 * if:
255 * - NO_HZ_COMMON is enabled
256 * - timer migration is enabled
257 * - the timer callback is not running
258 * - the timer is not the first expiring timer on the new target
259 *
260 * If one of the above requirements is not fulfilled we move the timer
261 * to the current CPU or leave it on the previously assigned CPU if
262 * the timer callback is currently running.
263 */
264static inline struct hrtimer_clock_base *
265switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
266 int pinned)
267{
268 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
269 struct hrtimer_clock_base *new_base;
270 int basenum = base->index;
271
272 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
273 new_cpu_base = get_target_base(this_cpu_base, pinned);
274again:
275 new_base = &new_cpu_base->clock_base[basenum];
276
277 if (base != new_base) {
278 /*
279 * We are trying to move timer to new_base.
280 * However we can't change timer's base while it is running,
281 * so we keep it on the same CPU. No hassle vs. reprogramming
282 * the event source in the high resolution case. The softirq
283 * code will take care of this when the timer function has
284 * completed. There is no conflict as we hold the lock until
285 * the timer is enqueued.
286 */
287 if (unlikely(hrtimer_callback_running(timer)))
288 return base;
289
290 /* See the comment in lock_hrtimer_base() */
291 WRITE_ONCE(timer->base, &migration_base);
292 raw_spin_unlock(&base->cpu_base->lock);
293 raw_spin_lock(&new_base->cpu_base->lock);
294
295 if (!hrtimer_suitable_target(timer, new_base, new_cpu_base,
296 this_cpu_base)) {
297 raw_spin_unlock(&new_base->cpu_base->lock);
298 raw_spin_lock(&base->cpu_base->lock);
299 new_cpu_base = this_cpu_base;
300 WRITE_ONCE(timer->base, base);
301 goto again;
302 }
303 WRITE_ONCE(timer->base, new_base);
304 } else {
305 if (!hrtimer_suitable_target(timer, new_base, new_cpu_base, this_cpu_base)) {
306 new_cpu_base = this_cpu_base;
307 goto again;
308 }
309 }
310 return new_base;
311}
312
313#else /* CONFIG_SMP */
314
315static inline bool is_migration_base(struct hrtimer_clock_base *base)
316{
317 return false;
318}
319
320static inline struct hrtimer_clock_base *
321lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
322 __acquires(&timer->base->cpu_base->lock)
323{
324 struct hrtimer_clock_base *base = timer->base;
325
326 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
327
328 return base;
329}
330
331# define switch_hrtimer_base(t, b, p) (b)
332
333#endif /* !CONFIG_SMP */
334
335/*
336 * Functions for the union type storage format of ktime_t which are
337 * too large for inlining:
338 */
339#if BITS_PER_LONG < 64
340/*
341 * Divide a ktime value by a nanosecond value
342 */
343s64 __ktime_divns(const ktime_t kt, s64 div)
344{
345 int sft = 0;
346 s64 dclc;
347 u64 tmp;
348
349 dclc = ktime_to_ns(kt);
350 tmp = dclc < 0 ? -dclc : dclc;
351
352 /* Make sure the divisor is less than 2^32: */
353 while (div >> 32) {
354 sft++;
355 div >>= 1;
356 }
357 tmp >>= sft;
358 do_div(tmp, (u32) div);
359 return dclc < 0 ? -tmp : tmp;
360}
361EXPORT_SYMBOL_GPL(__ktime_divns);
362#endif /* BITS_PER_LONG >= 64 */
363
364/*
365 * Add two ktime values and do a safety check for overflow:
366 */
367ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
368{
369 ktime_t res = ktime_add_unsafe(lhs, rhs);
370
371 /*
372 * We use KTIME_SEC_MAX here, the maximum timeout which we can
373 * return to user space in a timespec:
374 */
375 if (res < 0 || res < lhs || res < rhs)
376 res = ktime_set(KTIME_SEC_MAX, 0);
377
378 return res;
379}
380
381EXPORT_SYMBOL_GPL(ktime_add_safe);
382
383#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
384
385static const struct debug_obj_descr hrtimer_debug_descr;
386
387static void *hrtimer_debug_hint(void *addr)
388{
389 return ((struct hrtimer *) addr)->function;
390}
391
392/*
393 * fixup_init is called when:
394 * - an active object is initialized
395 */
396static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
397{
398 struct hrtimer *timer = addr;
399
400 switch (state) {
401 case ODEBUG_STATE_ACTIVE:
402 hrtimer_cancel(timer);
403 debug_object_init(timer, &hrtimer_debug_descr);
404 return true;
405 default:
406 return false;
407 }
408}
409
410/*
411 * fixup_activate is called when:
412 * - an active object is activated
413 * - an unknown non-static object is activated
414 */
415static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
416{
417 switch (state) {
418 case ODEBUG_STATE_ACTIVE:
419 WARN_ON(1);
420 fallthrough;
421 default:
422 return false;
423 }
424}
425
426/*
427 * fixup_free is called when:
428 * - an active object is freed
429 */
430static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
431{
432 struct hrtimer *timer = addr;
433
434 switch (state) {
435 case ODEBUG_STATE_ACTIVE:
436 hrtimer_cancel(timer);
437 debug_object_free(timer, &hrtimer_debug_descr);
438 return true;
439 default:
440 return false;
441 }
442}
443
444static const struct debug_obj_descr hrtimer_debug_descr = {
445 .name = "hrtimer",
446 .debug_hint = hrtimer_debug_hint,
447 .fixup_init = hrtimer_fixup_init,
448 .fixup_activate = hrtimer_fixup_activate,
449 .fixup_free = hrtimer_fixup_free,
450};
451
452static inline void debug_hrtimer_init(struct hrtimer *timer)
453{
454 debug_object_init(timer, &hrtimer_debug_descr);
455}
456
457static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer)
458{
459 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
460}
461
462static inline void debug_hrtimer_activate(struct hrtimer *timer,
463 enum hrtimer_mode mode)
464{
465 debug_object_activate(timer, &hrtimer_debug_descr);
466}
467
468static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
469{
470 debug_object_deactivate(timer, &hrtimer_debug_descr);
471}
472
473void destroy_hrtimer_on_stack(struct hrtimer *timer)
474{
475 debug_object_free(timer, &hrtimer_debug_descr);
476}
477EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
478
479#else
480
481static inline void debug_hrtimer_init(struct hrtimer *timer) { }
482static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer) { }
483static inline void debug_hrtimer_activate(struct hrtimer *timer,
484 enum hrtimer_mode mode) { }
485static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
486#endif
487
488static inline void
489debug_init(struct hrtimer *timer, clockid_t clockid,
490 enum hrtimer_mode mode)
491{
492 debug_hrtimer_init(timer);
493 trace_hrtimer_init(timer, clockid, mode);
494}
495
496static inline void debug_init_on_stack(struct hrtimer *timer, clockid_t clockid,
497 enum hrtimer_mode mode)
498{
499 debug_hrtimer_init_on_stack(timer);
500 trace_hrtimer_init(timer, clockid, mode);
501}
502
503static inline void debug_activate(struct hrtimer *timer,
504 enum hrtimer_mode mode)
505{
506 debug_hrtimer_activate(timer, mode);
507 trace_hrtimer_start(timer, mode);
508}
509
510static inline void debug_deactivate(struct hrtimer *timer)
511{
512 debug_hrtimer_deactivate(timer);
513 trace_hrtimer_cancel(timer);
514}
515
516static struct hrtimer_clock_base *
517__next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
518{
519 unsigned int idx;
520
521 if (!*active)
522 return NULL;
523
524 idx = __ffs(*active);
525 *active &= ~(1U << idx);
526
527 return &cpu_base->clock_base[idx];
528}
529
530#define for_each_active_base(base, cpu_base, active) \
531 while ((base = __next_base((cpu_base), &(active))))
532
533static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
534 const struct hrtimer *exclude,
535 unsigned int active,
536 ktime_t expires_next)
537{
538 struct hrtimer_clock_base *base;
539 ktime_t expires;
540
541 for_each_active_base(base, cpu_base, active) {
542 struct timerqueue_node *next;
543 struct hrtimer *timer;
544
545 next = timerqueue_getnext(&base->active);
546 timer = container_of(next, struct hrtimer, node);
547 if (timer == exclude) {
548 /* Get to the next timer in the queue. */
549 next = timerqueue_iterate_next(next);
550 if (!next)
551 continue;
552
553 timer = container_of(next, struct hrtimer, node);
554 }
555 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
556 if (expires < expires_next) {
557 expires_next = expires;
558
559 /* Skip cpu_base update if a timer is being excluded. */
560 if (exclude)
561 continue;
562
563 if (timer->is_soft)
564 cpu_base->softirq_next_timer = timer;
565 else
566 cpu_base->next_timer = timer;
567 }
568 }
569 /*
570 * clock_was_set() might have changed base->offset of any of
571 * the clock bases so the result might be negative. Fix it up
572 * to prevent a false positive in clockevents_program_event().
573 */
574 if (expires_next < 0)
575 expires_next = 0;
576 return expires_next;
577}
578
579/*
580 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
581 * but does not set cpu_base::*expires_next, that is done by
582 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
583 * cpu_base::*expires_next right away, reprogramming logic would no longer
584 * work.
585 *
586 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
587 * those timers will get run whenever the softirq gets handled, at the end of
588 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
589 *
590 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
591 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
592 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
593 *
594 * @active_mask must be one of:
595 * - HRTIMER_ACTIVE_ALL,
596 * - HRTIMER_ACTIVE_SOFT, or
597 * - HRTIMER_ACTIVE_HARD.
598 */
599static ktime_t
600__hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
601{
602 unsigned int active;
603 struct hrtimer *next_timer = NULL;
604 ktime_t expires_next = KTIME_MAX;
605
606 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
607 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
608 cpu_base->softirq_next_timer = NULL;
609 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
610 active, KTIME_MAX);
611
612 next_timer = cpu_base->softirq_next_timer;
613 }
614
615 if (active_mask & HRTIMER_ACTIVE_HARD) {
616 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
617 cpu_base->next_timer = next_timer;
618 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
619 expires_next);
620 }
621
622 return expires_next;
623}
624
625static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
626{
627 ktime_t expires_next, soft = KTIME_MAX;
628
629 /*
630 * If the soft interrupt has already been activated, ignore the
631 * soft bases. They will be handled in the already raised soft
632 * interrupt.
633 */
634 if (!cpu_base->softirq_activated) {
635 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
636 /*
637 * Update the soft expiry time. clock_settime() might have
638 * affected it.
639 */
640 cpu_base->softirq_expires_next = soft;
641 }
642
643 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
644 /*
645 * If a softirq timer is expiring first, update cpu_base->next_timer
646 * and program the hardware with the soft expiry time.
647 */
648 if (expires_next > soft) {
649 cpu_base->next_timer = cpu_base->softirq_next_timer;
650 expires_next = soft;
651 }
652
653 return expires_next;
654}
655
656static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
657{
658 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
659 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
660 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
661
662 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
663 offs_real, offs_boot, offs_tai);
664
665 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
666 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
667 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
668
669 return now;
670}
671
672/*
673 * Is the high resolution mode active ?
674 */
675static inline int hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
676{
677 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
678 cpu_base->hres_active : 0;
679}
680
681static void __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base,
682 struct hrtimer *next_timer,
683 ktime_t expires_next)
684{
685 cpu_base->expires_next = expires_next;
686
687 /*
688 * If hres is not active, hardware does not have to be
689 * reprogrammed yet.
690 *
691 * If a hang was detected in the last timer interrupt then we
692 * leave the hang delay active in the hardware. We want the
693 * system to make progress. That also prevents the following
694 * scenario:
695 * T1 expires 50ms from now
696 * T2 expires 5s from now
697 *
698 * T1 is removed, so this code is called and would reprogram
699 * the hardware to 5s from now. Any hrtimer_start after that
700 * will not reprogram the hardware due to hang_detected being
701 * set. So we'd effectively block all timers until the T2 event
702 * fires.
703 */
704 if (!hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
705 return;
706
707 tick_program_event(expires_next, 1);
708}
709
710/*
711 * Reprogram the event source with checking both queues for the
712 * next event
713 * Called with interrupts disabled and base->lock held
714 */
715static void
716hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
717{
718 ktime_t expires_next;
719
720 expires_next = hrtimer_update_next_event(cpu_base);
721
722 if (skip_equal && expires_next == cpu_base->expires_next)
723 return;
724
725 __hrtimer_reprogram(cpu_base, cpu_base->next_timer, expires_next);
726}
727
728/* High resolution timer related functions */
729#ifdef CONFIG_HIGH_RES_TIMERS
730
731/*
732 * High resolution timer enabled ?
733 */
734static bool hrtimer_hres_enabled __read_mostly = true;
735unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
736EXPORT_SYMBOL_GPL(hrtimer_resolution);
737
738/*
739 * Enable / Disable high resolution mode
740 */
741static int __init setup_hrtimer_hres(char *str)
742{
743 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
744}
745
746__setup("highres=", setup_hrtimer_hres);
747
748/*
749 * hrtimer_high_res_enabled - query, if the highres mode is enabled
750 */
751static inline int hrtimer_is_hres_enabled(void)
752{
753 return hrtimer_hres_enabled;
754}
755
756/*
757 * Switch to high resolution mode
758 */
759static void hrtimer_switch_to_hres(void)
760{
761 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
762
763 if (tick_init_highres()) {
764 pr_warn("Could not switch to high resolution mode on CPU %u\n",
765 base->cpu);
766 return;
767 }
768 base->hres_active = 1;
769 hrtimer_resolution = HIGH_RES_NSEC;
770
771 tick_setup_sched_timer(true);
772 /* "Retrigger" the interrupt to get things going */
773 retrigger_next_event(NULL);
774}
775
776#else
777
778static inline int hrtimer_is_hres_enabled(void) { return 0; }
779static inline void hrtimer_switch_to_hres(void) { }
780
781#endif /* CONFIG_HIGH_RES_TIMERS */
782/*
783 * Retrigger next event is called after clock was set with interrupts
784 * disabled through an SMP function call or directly from low level
785 * resume code.
786 *
787 * This is only invoked when:
788 * - CONFIG_HIGH_RES_TIMERS is enabled.
789 * - CONFIG_NOHZ_COMMON is enabled
790 *
791 * For the other cases this function is empty and because the call sites
792 * are optimized out it vanishes as well, i.e. no need for lots of
793 * #ifdeffery.
794 */
795static void retrigger_next_event(void *arg)
796{
797 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
798
799 /*
800 * When high resolution mode or nohz is active, then the offsets of
801 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
802 * next tick will take care of that.
803 *
804 * If high resolution mode is active then the next expiring timer
805 * must be reevaluated and the clock event device reprogrammed if
806 * necessary.
807 *
808 * In the NOHZ case the update of the offset and the reevaluation
809 * of the next expiring timer is enough. The return from the SMP
810 * function call will take care of the reprogramming in case the
811 * CPU was in a NOHZ idle sleep.
812 */
813 if (!hrtimer_hres_active(base) && !tick_nohz_active)
814 return;
815
816 raw_spin_lock(&base->lock);
817 hrtimer_update_base(base);
818 if (hrtimer_hres_active(base))
819 hrtimer_force_reprogram(base, 0);
820 else
821 hrtimer_update_next_event(base);
822 raw_spin_unlock(&base->lock);
823}
824
825/*
826 * When a timer is enqueued and expires earlier than the already enqueued
827 * timers, we have to check, whether it expires earlier than the timer for
828 * which the clock event device was armed.
829 *
830 * Called with interrupts disabled and base->cpu_base.lock held
831 */
832static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
833{
834 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
835 struct hrtimer_clock_base *base = timer->base;
836 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
837
838 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
839
840 /*
841 * CLOCK_REALTIME timer might be requested with an absolute
842 * expiry time which is less than base->offset. Set it to 0.
843 */
844 if (expires < 0)
845 expires = 0;
846
847 if (timer->is_soft) {
848 /*
849 * soft hrtimer could be started on a remote CPU. In this
850 * case softirq_expires_next needs to be updated on the
851 * remote CPU. The soft hrtimer will not expire before the
852 * first hard hrtimer on the remote CPU -
853 * hrtimer_check_target() prevents this case.
854 */
855 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
856
857 if (timer_cpu_base->softirq_activated)
858 return;
859
860 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
861 return;
862
863 timer_cpu_base->softirq_next_timer = timer;
864 timer_cpu_base->softirq_expires_next = expires;
865
866 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
867 !reprogram)
868 return;
869 }
870
871 /*
872 * If the timer is not on the current cpu, we cannot reprogram
873 * the other cpus clock event device.
874 */
875 if (base->cpu_base != cpu_base)
876 return;
877
878 if (expires >= cpu_base->expires_next)
879 return;
880
881 /*
882 * If the hrtimer interrupt is running, then it will reevaluate the
883 * clock bases and reprogram the clock event device.
884 */
885 if (cpu_base->in_hrtirq)
886 return;
887
888 cpu_base->next_timer = timer;
889
890 __hrtimer_reprogram(cpu_base, timer, expires);
891}
892
893static bool update_needs_ipi(struct hrtimer_cpu_base *cpu_base,
894 unsigned int active)
895{
896 struct hrtimer_clock_base *base;
897 unsigned int seq;
898 ktime_t expires;
899
900 /*
901 * Update the base offsets unconditionally so the following
902 * checks whether the SMP function call is required works.
903 *
904 * The update is safe even when the remote CPU is in the hrtimer
905 * interrupt or the hrtimer soft interrupt and expiring affected
906 * bases. Either it will see the update before handling a base or
907 * it will see it when it finishes the processing and reevaluates
908 * the next expiring timer.
909 */
910 seq = cpu_base->clock_was_set_seq;
911 hrtimer_update_base(cpu_base);
912
913 /*
914 * If the sequence did not change over the update then the
915 * remote CPU already handled it.
916 */
917 if (seq == cpu_base->clock_was_set_seq)
918 return false;
919
920 /*
921 * If the remote CPU is currently handling an hrtimer interrupt, it
922 * will reevaluate the first expiring timer of all clock bases
923 * before reprogramming. Nothing to do here.
924 */
925 if (cpu_base->in_hrtirq)
926 return false;
927
928 /*
929 * Walk the affected clock bases and check whether the first expiring
930 * timer in a clock base is moving ahead of the first expiring timer of
931 * @cpu_base. If so, the IPI must be invoked because per CPU clock
932 * event devices cannot be remotely reprogrammed.
933 */
934 active &= cpu_base->active_bases;
935
936 for_each_active_base(base, cpu_base, active) {
937 struct timerqueue_node *next;
938
939 next = timerqueue_getnext(&base->active);
940 expires = ktime_sub(next->expires, base->offset);
941 if (expires < cpu_base->expires_next)
942 return true;
943
944 /* Extra check for softirq clock bases */
945 if (base->clockid < HRTIMER_BASE_MONOTONIC_SOFT)
946 continue;
947 if (cpu_base->softirq_activated)
948 continue;
949 if (expires < cpu_base->softirq_expires_next)
950 return true;
951 }
952 return false;
953}
954
955/*
956 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
957 * CLOCK_BOOTTIME (for late sleep time injection).
958 *
959 * This requires to update the offsets for these clocks
960 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
961 * also requires to eventually reprogram the per CPU clock event devices
962 * when the change moves an affected timer ahead of the first expiring
963 * timer on that CPU. Obviously remote per CPU clock event devices cannot
964 * be reprogrammed. The other reason why an IPI has to be sent is when the
965 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
966 * in the tick, which obviously might be stopped, so this has to bring out
967 * the remote CPU which might sleep in idle to get this sorted.
968 */
969void clock_was_set(unsigned int bases)
970{
971 struct hrtimer_cpu_base *cpu_base = raw_cpu_ptr(&hrtimer_bases);
972 cpumask_var_t mask;
973 int cpu;
974
975 if (!hrtimer_hres_active(cpu_base) && !tick_nohz_active)
976 goto out_timerfd;
977
978 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
979 on_each_cpu(retrigger_next_event, NULL, 1);
980 goto out_timerfd;
981 }
982
983 /* Avoid interrupting CPUs if possible */
984 cpus_read_lock();
985 for_each_online_cpu(cpu) {
986 unsigned long flags;
987
988 cpu_base = &per_cpu(hrtimer_bases, cpu);
989 raw_spin_lock_irqsave(&cpu_base->lock, flags);
990
991 if (update_needs_ipi(cpu_base, bases))
992 cpumask_set_cpu(cpu, mask);
993
994 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
995 }
996
997 preempt_disable();
998 smp_call_function_many(mask, retrigger_next_event, NULL, 1);
999 preempt_enable();
1000 cpus_read_unlock();
1001 free_cpumask_var(mask);
1002
1003out_timerfd:
1004 timerfd_clock_was_set();
1005}
1006
1007static void clock_was_set_work(struct work_struct *work)
1008{
1009 clock_was_set(CLOCK_SET_WALL);
1010}
1011
1012static DECLARE_WORK(hrtimer_work, clock_was_set_work);
1013
1014/*
1015 * Called from timekeeping code to reprogram the hrtimer interrupt device
1016 * on all cpus and to notify timerfd.
1017 */
1018void clock_was_set_delayed(void)
1019{
1020 schedule_work(&hrtimer_work);
1021}
1022
1023/*
1024 * Called during resume either directly from via timekeeping_resume()
1025 * or in the case of s2idle from tick_unfreeze() to ensure that the
1026 * hrtimers are up to date.
1027 */
1028void hrtimers_resume_local(void)
1029{
1030 lockdep_assert_irqs_disabled();
1031 /* Retrigger on the local CPU */
1032 retrigger_next_event(NULL);
1033}
1034
1035/*
1036 * Counterpart to lock_hrtimer_base above:
1037 */
1038static inline
1039void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
1040 __releases(&timer->base->cpu_base->lock)
1041{
1042 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
1043}
1044
1045/**
1046 * hrtimer_forward() - forward the timer expiry
1047 * @timer: hrtimer to forward
1048 * @now: forward past this time
1049 * @interval: the interval to forward
1050 *
1051 * Forward the timer expiry so it will expire in the future.
1052 *
1053 * .. note::
1054 * This only updates the timer expiry value and does not requeue the timer.
1055 *
1056 * There is also a variant of the function hrtimer_forward_now().
1057 *
1058 * Context: Can be safely called from the callback function of @timer. If called
1059 * from other contexts @timer must neither be enqueued nor running the
1060 * callback and the caller needs to take care of serialization.
1061 *
1062 * Return: The number of overruns are returned.
1063 */
1064u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
1065{
1066 u64 orun = 1;
1067 ktime_t delta;
1068
1069 delta = ktime_sub(now, hrtimer_get_expires(timer));
1070
1071 if (delta < 0)
1072 return 0;
1073
1074 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
1075 return 0;
1076
1077 if (interval < hrtimer_resolution)
1078 interval = hrtimer_resolution;
1079
1080 if (unlikely(delta >= interval)) {
1081 s64 incr = ktime_to_ns(interval);
1082
1083 orun = ktime_divns(delta, incr);
1084 hrtimer_add_expires_ns(timer, incr * orun);
1085 if (hrtimer_get_expires_tv64(timer) > now)
1086 return orun;
1087 /*
1088 * This (and the ktime_add() below) is the
1089 * correction for exact:
1090 */
1091 orun++;
1092 }
1093 hrtimer_add_expires(timer, interval);
1094
1095 return orun;
1096}
1097EXPORT_SYMBOL_GPL(hrtimer_forward);
1098
1099/*
1100 * enqueue_hrtimer - internal function to (re)start a timer
1101 *
1102 * The timer is inserted in expiry order. Insertion into the
1103 * red black tree is O(log(n)). Must hold the base lock.
1104 *
1105 * Returns 1 when the new timer is the leftmost timer in the tree.
1106 */
1107static int enqueue_hrtimer(struct hrtimer *timer,
1108 struct hrtimer_clock_base *base,
1109 enum hrtimer_mode mode)
1110{
1111 debug_activate(timer, mode);
1112 WARN_ON_ONCE(!base->cpu_base->online);
1113
1114 base->cpu_base->active_bases |= 1 << base->index;
1115
1116 /* Pairs with the lockless read in hrtimer_is_queued() */
1117 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1118
1119 return timerqueue_add(&base->active, &timer->node);
1120}
1121
1122/*
1123 * __remove_hrtimer - internal function to remove a timer
1124 *
1125 * Caller must hold the base lock.
1126 *
1127 * High resolution timer mode reprograms the clock event device when the
1128 * timer is the one which expires next. The caller can disable this by setting
1129 * reprogram to zero. This is useful, when the context does a reprogramming
1130 * anyway (e.g. timer interrupt)
1131 */
1132static void __remove_hrtimer(struct hrtimer *timer,
1133 struct hrtimer_clock_base *base,
1134 u8 newstate, int reprogram)
1135{
1136 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1137 u8 state = timer->state;
1138
1139 /* Pairs with the lockless read in hrtimer_is_queued() */
1140 WRITE_ONCE(timer->state, newstate);
1141 if (!(state & HRTIMER_STATE_ENQUEUED))
1142 return;
1143
1144 if (!timerqueue_del(&base->active, &timer->node))
1145 cpu_base->active_bases &= ~(1 << base->index);
1146
1147 /*
1148 * Note: If reprogram is false we do not update
1149 * cpu_base->next_timer. This happens when we remove the first
1150 * timer on a remote cpu. No harm as we never dereference
1151 * cpu_base->next_timer. So the worst thing what can happen is
1152 * an superfluous call to hrtimer_force_reprogram() on the
1153 * remote cpu later on if the same timer gets enqueued again.
1154 */
1155 if (reprogram && timer == cpu_base->next_timer)
1156 hrtimer_force_reprogram(cpu_base, 1);
1157}
1158
1159/*
1160 * remove hrtimer, called with base lock held
1161 */
1162static inline int
1163remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1164 bool restart, bool keep_local)
1165{
1166 u8 state = timer->state;
1167
1168 if (state & HRTIMER_STATE_ENQUEUED) {
1169 bool reprogram;
1170
1171 /*
1172 * Remove the timer and force reprogramming when high
1173 * resolution mode is active and the timer is on the current
1174 * CPU. If we remove a timer on another CPU, reprogramming is
1175 * skipped. The interrupt event on this CPU is fired and
1176 * reprogramming happens in the interrupt handler. This is a
1177 * rare case and less expensive than a smp call.
1178 */
1179 debug_deactivate(timer);
1180 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1181
1182 /*
1183 * If the timer is not restarted then reprogramming is
1184 * required if the timer is local. If it is local and about
1185 * to be restarted, avoid programming it twice (on removal
1186 * and a moment later when it's requeued).
1187 */
1188 if (!restart)
1189 state = HRTIMER_STATE_INACTIVE;
1190 else
1191 reprogram &= !keep_local;
1192
1193 __remove_hrtimer(timer, base, state, reprogram);
1194 return 1;
1195 }
1196 return 0;
1197}
1198
1199static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1200 const enum hrtimer_mode mode)
1201{
1202#ifdef CONFIG_TIME_LOW_RES
1203 /*
1204 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1205 * granular time values. For relative timers we add hrtimer_resolution
1206 * (i.e. one jiffy) to prevent short timeouts.
1207 */
1208 timer->is_rel = mode & HRTIMER_MODE_REL;
1209 if (timer->is_rel)
1210 tim = ktime_add_safe(tim, hrtimer_resolution);
1211#endif
1212 return tim;
1213}
1214
1215static void
1216hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1217{
1218 ktime_t expires;
1219
1220 /*
1221 * Find the next SOFT expiration.
1222 */
1223 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1224
1225 /*
1226 * reprogramming needs to be triggered, even if the next soft
1227 * hrtimer expires at the same time than the next hard
1228 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1229 */
1230 if (expires == KTIME_MAX)
1231 return;
1232
1233 /*
1234 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1235 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1236 */
1237 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1238}
1239
1240static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1241 u64 delta_ns, const enum hrtimer_mode mode,
1242 struct hrtimer_clock_base *base)
1243{
1244 struct hrtimer_cpu_base *this_cpu_base = this_cpu_ptr(&hrtimer_bases);
1245 struct hrtimer_clock_base *new_base;
1246 bool force_local, first;
1247
1248 /*
1249 * If the timer is on the local cpu base and is the first expiring
1250 * timer then this might end up reprogramming the hardware twice
1251 * (on removal and on enqueue). To avoid that by prevent the
1252 * reprogram on removal, keep the timer local to the current CPU
1253 * and enforce reprogramming after it is queued no matter whether
1254 * it is the new first expiring timer again or not.
1255 */
1256 force_local = base->cpu_base == this_cpu_base;
1257 force_local &= base->cpu_base->next_timer == timer;
1258
1259 /*
1260 * Don't force local queuing if this enqueue happens on a unplugged
1261 * CPU after hrtimer_cpu_dying() has been invoked.
1262 */
1263 force_local &= this_cpu_base->online;
1264
1265 /*
1266 * Remove an active timer from the queue. In case it is not queued
1267 * on the current CPU, make sure that remove_hrtimer() updates the
1268 * remote data correctly.
1269 *
1270 * If it's on the current CPU and the first expiring timer, then
1271 * skip reprogramming, keep the timer local and enforce
1272 * reprogramming later if it was the first expiring timer. This
1273 * avoids programming the underlying clock event twice (once at
1274 * removal and once after enqueue).
1275 */
1276 remove_hrtimer(timer, base, true, force_local);
1277
1278 if (mode & HRTIMER_MODE_REL)
1279 tim = ktime_add_safe(tim, base->get_time());
1280
1281 tim = hrtimer_update_lowres(timer, tim, mode);
1282
1283 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1284
1285 /* Switch the timer base, if necessary: */
1286 if (!force_local) {
1287 new_base = switch_hrtimer_base(timer, base,
1288 mode & HRTIMER_MODE_PINNED);
1289 } else {
1290 new_base = base;
1291 }
1292
1293 first = enqueue_hrtimer(timer, new_base, mode);
1294 if (!force_local) {
1295 /*
1296 * If the current CPU base is online, then the timer is
1297 * never queued on a remote CPU if it would be the first
1298 * expiring timer there.
1299 */
1300 if (hrtimer_base_is_online(this_cpu_base))
1301 return first;
1302
1303 /*
1304 * Timer was enqueued remote because the current base is
1305 * already offline. If the timer is the first to expire,
1306 * kick the remote CPU to reprogram the clock event.
1307 */
1308 if (first) {
1309 struct hrtimer_cpu_base *new_cpu_base = new_base->cpu_base;
1310
1311 smp_call_function_single_async(new_cpu_base->cpu, &new_cpu_base->csd);
1312 }
1313 return 0;
1314 }
1315
1316 /*
1317 * Timer was forced to stay on the current CPU to avoid
1318 * reprogramming on removal and enqueue. Force reprogram the
1319 * hardware by evaluating the new first expiring timer.
1320 */
1321 hrtimer_force_reprogram(new_base->cpu_base, 1);
1322 return 0;
1323}
1324
1325/**
1326 * hrtimer_start_range_ns - (re)start an hrtimer
1327 * @timer: the timer to be added
1328 * @tim: expiry time
1329 * @delta_ns: "slack" range for the timer
1330 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1331 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1332 * softirq based mode is considered for debug purpose only!
1333 */
1334void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1335 u64 delta_ns, const enum hrtimer_mode mode)
1336{
1337 struct hrtimer_clock_base *base;
1338 unsigned long flags;
1339
1340 if (WARN_ON_ONCE(!timer->function))
1341 return;
1342 /*
1343 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1344 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1345 * expiry mode because unmarked timers are moved to softirq expiry.
1346 */
1347 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1348 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1349 else
1350 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1351
1352 base = lock_hrtimer_base(timer, &flags);
1353
1354 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1355 hrtimer_reprogram(timer, true);
1356
1357 unlock_hrtimer_base(timer, &flags);
1358}
1359EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1360
1361/**
1362 * hrtimer_try_to_cancel - try to deactivate a timer
1363 * @timer: hrtimer to stop
1364 *
1365 * Returns:
1366 *
1367 * * 0 when the timer was not active
1368 * * 1 when the timer was active
1369 * * -1 when the timer is currently executing the callback function and
1370 * cannot be stopped
1371 */
1372int hrtimer_try_to_cancel(struct hrtimer *timer)
1373{
1374 struct hrtimer_clock_base *base;
1375 unsigned long flags;
1376 int ret = -1;
1377
1378 /*
1379 * Check lockless first. If the timer is not active (neither
1380 * enqueued nor running the callback, nothing to do here. The
1381 * base lock does not serialize against a concurrent enqueue,
1382 * so we can avoid taking it.
1383 */
1384 if (!hrtimer_active(timer))
1385 return 0;
1386
1387 base = lock_hrtimer_base(timer, &flags);
1388
1389 if (!hrtimer_callback_running(timer))
1390 ret = remove_hrtimer(timer, base, false, false);
1391
1392 unlock_hrtimer_base(timer, &flags);
1393
1394 return ret;
1395
1396}
1397EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1398
1399#ifdef CONFIG_PREEMPT_RT
1400static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1401{
1402 spin_lock_init(&base->softirq_expiry_lock);
1403}
1404
1405static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1406 __acquires(&base->softirq_expiry_lock)
1407{
1408 spin_lock(&base->softirq_expiry_lock);
1409}
1410
1411static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1412 __releases(&base->softirq_expiry_lock)
1413{
1414 spin_unlock(&base->softirq_expiry_lock);
1415}
1416
1417/*
1418 * The counterpart to hrtimer_cancel_wait_running().
1419 *
1420 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1421 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1422 * allows the waiter to acquire the lock and make progress.
1423 */
1424static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1425 unsigned long flags)
1426{
1427 if (atomic_read(&cpu_base->timer_waiters)) {
1428 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1429 spin_unlock(&cpu_base->softirq_expiry_lock);
1430 spin_lock(&cpu_base->softirq_expiry_lock);
1431 raw_spin_lock_irq(&cpu_base->lock);
1432 }
1433}
1434
1435/*
1436 * This function is called on PREEMPT_RT kernels when the fast path
1437 * deletion of a timer failed because the timer callback function was
1438 * running.
1439 *
1440 * This prevents priority inversion: if the soft irq thread is preempted
1441 * in the middle of a timer callback, then calling del_timer_sync() can
1442 * lead to two issues:
1443 *
1444 * - If the caller is on a remote CPU then it has to spin wait for the timer
1445 * handler to complete. This can result in unbound priority inversion.
1446 *
1447 * - If the caller originates from the task which preempted the timer
1448 * handler on the same CPU, then spin waiting for the timer handler to
1449 * complete is never going to end.
1450 */
1451void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1452{
1453 /* Lockless read. Prevent the compiler from reloading it below */
1454 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1455
1456 /*
1457 * Just relax if the timer expires in hard interrupt context or if
1458 * it is currently on the migration base.
1459 */
1460 if (!timer->is_soft || is_migration_base(base)) {
1461 cpu_relax();
1462 return;
1463 }
1464
1465 /*
1466 * Mark the base as contended and grab the expiry lock, which is
1467 * held by the softirq across the timer callback. Drop the lock
1468 * immediately so the softirq can expire the next timer. In theory
1469 * the timer could already be running again, but that's more than
1470 * unlikely and just causes another wait loop.
1471 */
1472 atomic_inc(&base->cpu_base->timer_waiters);
1473 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1474 atomic_dec(&base->cpu_base->timer_waiters);
1475 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1476}
1477#else
1478static inline void
1479hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1480static inline void
1481hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1482static inline void
1483hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
1484static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1485 unsigned long flags) { }
1486#endif
1487
1488/**
1489 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1490 * @timer: the timer to be cancelled
1491 *
1492 * Returns:
1493 * 0 when the timer was not active
1494 * 1 when the timer was active
1495 */
1496int hrtimer_cancel(struct hrtimer *timer)
1497{
1498 int ret;
1499
1500 do {
1501 ret = hrtimer_try_to_cancel(timer);
1502
1503 if (ret < 0)
1504 hrtimer_cancel_wait_running(timer);
1505 } while (ret < 0);
1506 return ret;
1507}
1508EXPORT_SYMBOL_GPL(hrtimer_cancel);
1509
1510/**
1511 * __hrtimer_get_remaining - get remaining time for the timer
1512 * @timer: the timer to read
1513 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1514 */
1515ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1516{
1517 unsigned long flags;
1518 ktime_t rem;
1519
1520 lock_hrtimer_base(timer, &flags);
1521 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1522 rem = hrtimer_expires_remaining_adjusted(timer);
1523 else
1524 rem = hrtimer_expires_remaining(timer);
1525 unlock_hrtimer_base(timer, &flags);
1526
1527 return rem;
1528}
1529EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1530
1531#ifdef CONFIG_NO_HZ_COMMON
1532/**
1533 * hrtimer_get_next_event - get the time until next expiry event
1534 *
1535 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1536 */
1537u64 hrtimer_get_next_event(void)
1538{
1539 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1540 u64 expires = KTIME_MAX;
1541 unsigned long flags;
1542
1543 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1544
1545 if (!hrtimer_hres_active(cpu_base))
1546 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1547
1548 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1549
1550 return expires;
1551}
1552
1553/**
1554 * hrtimer_next_event_without - time until next expiry event w/o one timer
1555 * @exclude: timer to exclude
1556 *
1557 * Returns the next expiry time over all timers except for the @exclude one or
1558 * KTIME_MAX if none of them is pending.
1559 */
1560u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1561{
1562 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1563 u64 expires = KTIME_MAX;
1564 unsigned long flags;
1565
1566 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1567
1568 if (hrtimer_hres_active(cpu_base)) {
1569 unsigned int active;
1570
1571 if (!cpu_base->softirq_activated) {
1572 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1573 expires = __hrtimer_next_event_base(cpu_base, exclude,
1574 active, KTIME_MAX);
1575 }
1576 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1577 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1578 expires);
1579 }
1580
1581 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1582
1583 return expires;
1584}
1585#endif
1586
1587static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1588{
1589 if (likely(clock_id < MAX_CLOCKS)) {
1590 int base = hrtimer_clock_to_base_table[clock_id];
1591
1592 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1593 return base;
1594 }
1595 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1596 return HRTIMER_BASE_MONOTONIC;
1597}
1598
1599static enum hrtimer_restart hrtimer_dummy_timeout(struct hrtimer *unused)
1600{
1601 return HRTIMER_NORESTART;
1602}
1603
1604static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1605 enum hrtimer_mode mode)
1606{
1607 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1608 struct hrtimer_cpu_base *cpu_base;
1609 int base;
1610
1611 /*
1612 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1613 * marked for hard interrupt expiry mode are moved into soft
1614 * interrupt context for latency reasons and because the callbacks
1615 * can invoke functions which might sleep on RT, e.g. spin_lock().
1616 */
1617 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1618 softtimer = true;
1619
1620 memset(timer, 0, sizeof(struct hrtimer));
1621
1622 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1623
1624 /*
1625 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1626 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1627 * ensure POSIX compliance.
1628 */
1629 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1630 clock_id = CLOCK_MONOTONIC;
1631
1632 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1633 base += hrtimer_clockid_to_base(clock_id);
1634 timer->is_soft = softtimer;
1635 timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1636 timer->base = &cpu_base->clock_base[base];
1637 timerqueue_init(&timer->node);
1638}
1639
1640static void __hrtimer_setup(struct hrtimer *timer,
1641 enum hrtimer_restart (*function)(struct hrtimer *),
1642 clockid_t clock_id, enum hrtimer_mode mode)
1643{
1644 __hrtimer_init(timer, clock_id, mode);
1645
1646 if (WARN_ON_ONCE(!function))
1647 timer->function = hrtimer_dummy_timeout;
1648 else
1649 timer->function = function;
1650}
1651
1652/**
1653 * hrtimer_init - initialize a timer to the given clock
1654 * @timer: the timer to be initialized
1655 * @clock_id: the clock to be used
1656 * @mode: The modes which are relevant for initialization:
1657 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1658 * HRTIMER_MODE_REL_SOFT
1659 *
1660 * The PINNED variants of the above can be handed in,
1661 * but the PINNED bit is ignored as pinning happens
1662 * when the hrtimer is started
1663 */
1664void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1665 enum hrtimer_mode mode)
1666{
1667 debug_init(timer, clock_id, mode);
1668 __hrtimer_init(timer, clock_id, mode);
1669}
1670EXPORT_SYMBOL_GPL(hrtimer_init);
1671
1672/**
1673 * hrtimer_setup - initialize a timer to the given clock
1674 * @timer: the timer to be initialized
1675 * @function: the callback function
1676 * @clock_id: the clock to be used
1677 * @mode: The modes which are relevant for initialization:
1678 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1679 * HRTIMER_MODE_REL_SOFT
1680 *
1681 * The PINNED variants of the above can be handed in,
1682 * but the PINNED bit is ignored as pinning happens
1683 * when the hrtimer is started
1684 */
1685void hrtimer_setup(struct hrtimer *timer, enum hrtimer_restart (*function)(struct hrtimer *),
1686 clockid_t clock_id, enum hrtimer_mode mode)
1687{
1688 debug_init(timer, clock_id, mode);
1689 __hrtimer_setup(timer, function, clock_id, mode);
1690}
1691EXPORT_SYMBOL_GPL(hrtimer_setup);
1692
1693/**
1694 * hrtimer_setup_on_stack - initialize a timer on stack memory
1695 * @timer: The timer to be initialized
1696 * @function: the callback function
1697 * @clock_id: The clock to be used
1698 * @mode: The timer mode
1699 *
1700 * Similar to hrtimer_setup(), except that this one must be used if struct hrtimer is in stack
1701 * memory.
1702 */
1703void hrtimer_setup_on_stack(struct hrtimer *timer,
1704 enum hrtimer_restart (*function)(struct hrtimer *),
1705 clockid_t clock_id, enum hrtimer_mode mode)
1706{
1707 debug_init_on_stack(timer, clock_id, mode);
1708 __hrtimer_setup(timer, function, clock_id, mode);
1709}
1710EXPORT_SYMBOL_GPL(hrtimer_setup_on_stack);
1711
1712/*
1713 * A timer is active, when it is enqueued into the rbtree or the
1714 * callback function is running or it's in the state of being migrated
1715 * to another cpu.
1716 *
1717 * It is important for this function to not return a false negative.
1718 */
1719bool hrtimer_active(const struct hrtimer *timer)
1720{
1721 struct hrtimer_clock_base *base;
1722 unsigned int seq;
1723
1724 do {
1725 base = READ_ONCE(timer->base);
1726 seq = raw_read_seqcount_begin(&base->seq);
1727
1728 if (timer->state != HRTIMER_STATE_INACTIVE ||
1729 base->running == timer)
1730 return true;
1731
1732 } while (read_seqcount_retry(&base->seq, seq) ||
1733 base != READ_ONCE(timer->base));
1734
1735 return false;
1736}
1737EXPORT_SYMBOL_GPL(hrtimer_active);
1738
1739/*
1740 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1741 * distinct sections:
1742 *
1743 * - queued: the timer is queued
1744 * - callback: the timer is being ran
1745 * - post: the timer is inactive or (re)queued
1746 *
1747 * On the read side we ensure we observe timer->state and cpu_base->running
1748 * from the same section, if anything changed while we looked at it, we retry.
1749 * This includes timer->base changing because sequence numbers alone are
1750 * insufficient for that.
1751 *
1752 * The sequence numbers are required because otherwise we could still observe
1753 * a false negative if the read side got smeared over multiple consecutive
1754 * __run_hrtimer() invocations.
1755 */
1756
1757static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1758 struct hrtimer_clock_base *base,
1759 struct hrtimer *timer, ktime_t *now,
1760 unsigned long flags) __must_hold(&cpu_base->lock)
1761{
1762 enum hrtimer_restart (*fn)(struct hrtimer *);
1763 bool expires_in_hardirq;
1764 int restart;
1765
1766 lockdep_assert_held(&cpu_base->lock);
1767
1768 debug_deactivate(timer);
1769 base->running = timer;
1770
1771 /*
1772 * Separate the ->running assignment from the ->state assignment.
1773 *
1774 * As with a regular write barrier, this ensures the read side in
1775 * hrtimer_active() cannot observe base->running == NULL &&
1776 * timer->state == INACTIVE.
1777 */
1778 raw_write_seqcount_barrier(&base->seq);
1779
1780 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1781 fn = timer->function;
1782
1783 /*
1784 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1785 * timer is restarted with a period then it becomes an absolute
1786 * timer. If its not restarted it does not matter.
1787 */
1788 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1789 timer->is_rel = false;
1790
1791 /*
1792 * The timer is marked as running in the CPU base, so it is
1793 * protected against migration to a different CPU even if the lock
1794 * is dropped.
1795 */
1796 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1797 trace_hrtimer_expire_entry(timer, now);
1798 expires_in_hardirq = lockdep_hrtimer_enter(timer);
1799
1800 restart = fn(timer);
1801
1802 lockdep_hrtimer_exit(expires_in_hardirq);
1803 trace_hrtimer_expire_exit(timer);
1804 raw_spin_lock_irq(&cpu_base->lock);
1805
1806 /*
1807 * Note: We clear the running state after enqueue_hrtimer and
1808 * we do not reprogram the event hardware. Happens either in
1809 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1810 *
1811 * Note: Because we dropped the cpu_base->lock above,
1812 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1813 * for us already.
1814 */
1815 if (restart != HRTIMER_NORESTART &&
1816 !(timer->state & HRTIMER_STATE_ENQUEUED))
1817 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1818
1819 /*
1820 * Separate the ->running assignment from the ->state assignment.
1821 *
1822 * As with a regular write barrier, this ensures the read side in
1823 * hrtimer_active() cannot observe base->running.timer == NULL &&
1824 * timer->state == INACTIVE.
1825 */
1826 raw_write_seqcount_barrier(&base->seq);
1827
1828 WARN_ON_ONCE(base->running != timer);
1829 base->running = NULL;
1830}
1831
1832static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1833 unsigned long flags, unsigned int active_mask)
1834{
1835 struct hrtimer_clock_base *base;
1836 unsigned int active = cpu_base->active_bases & active_mask;
1837
1838 for_each_active_base(base, cpu_base, active) {
1839 struct timerqueue_node *node;
1840 ktime_t basenow;
1841
1842 basenow = ktime_add(now, base->offset);
1843
1844 while ((node = timerqueue_getnext(&base->active))) {
1845 struct hrtimer *timer;
1846
1847 timer = container_of(node, struct hrtimer, node);
1848
1849 /*
1850 * The immediate goal for using the softexpires is
1851 * minimizing wakeups, not running timers at the
1852 * earliest interrupt after their soft expiration.
1853 * This allows us to avoid using a Priority Search
1854 * Tree, which can answer a stabbing query for
1855 * overlapping intervals and instead use the simple
1856 * BST we already have.
1857 * We don't add extra wakeups by delaying timers that
1858 * are right-of a not yet expired timer, because that
1859 * timer will have to trigger a wakeup anyway.
1860 */
1861 if (basenow < hrtimer_get_softexpires_tv64(timer))
1862 break;
1863
1864 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1865 if (active_mask == HRTIMER_ACTIVE_SOFT)
1866 hrtimer_sync_wait_running(cpu_base, flags);
1867 }
1868 }
1869}
1870
1871static __latent_entropy void hrtimer_run_softirq(void)
1872{
1873 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1874 unsigned long flags;
1875 ktime_t now;
1876
1877 hrtimer_cpu_base_lock_expiry(cpu_base);
1878 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1879
1880 now = hrtimer_update_base(cpu_base);
1881 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1882
1883 cpu_base->softirq_activated = 0;
1884 hrtimer_update_softirq_timer(cpu_base, true);
1885
1886 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1887 hrtimer_cpu_base_unlock_expiry(cpu_base);
1888}
1889
1890#ifdef CONFIG_HIGH_RES_TIMERS
1891
1892/*
1893 * High resolution timer interrupt
1894 * Called with interrupts disabled
1895 */
1896void hrtimer_interrupt(struct clock_event_device *dev)
1897{
1898 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1899 ktime_t expires_next, now, entry_time, delta;
1900 unsigned long flags;
1901 int retries = 0;
1902
1903 BUG_ON(!cpu_base->hres_active);
1904 cpu_base->nr_events++;
1905 dev->next_event = KTIME_MAX;
1906
1907 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1908 entry_time = now = hrtimer_update_base(cpu_base);
1909retry:
1910 cpu_base->in_hrtirq = 1;
1911 /*
1912 * We set expires_next to KTIME_MAX here with cpu_base->lock
1913 * held to prevent that a timer is enqueued in our queue via
1914 * the migration code. This does not affect enqueueing of
1915 * timers which run their callback and need to be requeued on
1916 * this CPU.
1917 */
1918 cpu_base->expires_next = KTIME_MAX;
1919
1920 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1921 cpu_base->softirq_expires_next = KTIME_MAX;
1922 cpu_base->softirq_activated = 1;
1923 raise_timer_softirq(HRTIMER_SOFTIRQ);
1924 }
1925
1926 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1927
1928 /* Reevaluate the clock bases for the [soft] next expiry */
1929 expires_next = hrtimer_update_next_event(cpu_base);
1930 /*
1931 * Store the new expiry value so the migration code can verify
1932 * against it.
1933 */
1934 cpu_base->expires_next = expires_next;
1935 cpu_base->in_hrtirq = 0;
1936 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1937
1938 /* Reprogramming necessary ? */
1939 if (!tick_program_event(expires_next, 0)) {
1940 cpu_base->hang_detected = 0;
1941 return;
1942 }
1943
1944 /*
1945 * The next timer was already expired due to:
1946 * - tracing
1947 * - long lasting callbacks
1948 * - being scheduled away when running in a VM
1949 *
1950 * We need to prevent that we loop forever in the hrtimer
1951 * interrupt routine. We give it 3 attempts to avoid
1952 * overreacting on some spurious event.
1953 *
1954 * Acquire base lock for updating the offsets and retrieving
1955 * the current time.
1956 */
1957 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1958 now = hrtimer_update_base(cpu_base);
1959 cpu_base->nr_retries++;
1960 if (++retries < 3)
1961 goto retry;
1962 /*
1963 * Give the system a chance to do something else than looping
1964 * here. We stored the entry time, so we know exactly how long
1965 * we spent here. We schedule the next event this amount of
1966 * time away.
1967 */
1968 cpu_base->nr_hangs++;
1969 cpu_base->hang_detected = 1;
1970 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1971
1972 delta = ktime_sub(now, entry_time);
1973 if ((unsigned int)delta > cpu_base->max_hang_time)
1974 cpu_base->max_hang_time = (unsigned int) delta;
1975 /*
1976 * Limit it to a sensible value as we enforce a longer
1977 * delay. Give the CPU at least 100ms to catch up.
1978 */
1979 if (delta > 100 * NSEC_PER_MSEC)
1980 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1981 else
1982 expires_next = ktime_add(now, delta);
1983 tick_program_event(expires_next, 1);
1984 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1985}
1986#endif /* !CONFIG_HIGH_RES_TIMERS */
1987
1988/*
1989 * Called from run_local_timers in hardirq context every jiffy
1990 */
1991void hrtimer_run_queues(void)
1992{
1993 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1994 unsigned long flags;
1995 ktime_t now;
1996
1997 if (hrtimer_hres_active(cpu_base))
1998 return;
1999
2000 /*
2001 * This _is_ ugly: We have to check periodically, whether we
2002 * can switch to highres and / or nohz mode. The clocksource
2003 * switch happens with xtime_lock held. Notification from
2004 * there only sets the check bit in the tick_oneshot code,
2005 * otherwise we might deadlock vs. xtime_lock.
2006 */
2007 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
2008 hrtimer_switch_to_hres();
2009 return;
2010 }
2011
2012 raw_spin_lock_irqsave(&cpu_base->lock, flags);
2013 now = hrtimer_update_base(cpu_base);
2014
2015 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
2016 cpu_base->softirq_expires_next = KTIME_MAX;
2017 cpu_base->softirq_activated = 1;
2018 raise_timer_softirq(HRTIMER_SOFTIRQ);
2019 }
2020
2021 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
2022 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
2023}
2024
2025/*
2026 * Sleep related functions:
2027 */
2028static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
2029{
2030 struct hrtimer_sleeper *t =
2031 container_of(timer, struct hrtimer_sleeper, timer);
2032 struct task_struct *task = t->task;
2033
2034 t->task = NULL;
2035 if (task)
2036 wake_up_process(task);
2037
2038 return HRTIMER_NORESTART;
2039}
2040
2041/**
2042 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
2043 * @sl: sleeper to be started
2044 * @mode: timer mode abs/rel
2045 *
2046 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
2047 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
2048 */
2049void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
2050 enum hrtimer_mode mode)
2051{
2052 /*
2053 * Make the enqueue delivery mode check work on RT. If the sleeper
2054 * was initialized for hard interrupt delivery, force the mode bit.
2055 * This is a special case for hrtimer_sleepers because
2056 * __hrtimer_init_sleeper() determines the delivery mode on RT so the
2057 * fiddling with this decision is avoided at the call sites.
2058 */
2059 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
2060 mode |= HRTIMER_MODE_HARD;
2061
2062 hrtimer_start_expires(&sl->timer, mode);
2063}
2064EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
2065
2066static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
2067 clockid_t clock_id, enum hrtimer_mode mode)
2068{
2069 /*
2070 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
2071 * marked for hard interrupt expiry mode are moved into soft
2072 * interrupt context either for latency reasons or because the
2073 * hrtimer callback takes regular spinlocks or invokes other
2074 * functions which are not suitable for hard interrupt context on
2075 * PREEMPT_RT.
2076 *
2077 * The hrtimer_sleeper callback is RT compatible in hard interrupt
2078 * context, but there is a latency concern: Untrusted userspace can
2079 * spawn many threads which arm timers for the same expiry time on
2080 * the same CPU. That causes a latency spike due to the wakeup of
2081 * a gazillion threads.
2082 *
2083 * OTOH, privileged real-time user space applications rely on the
2084 * low latency of hard interrupt wakeups. If the current task is in
2085 * a real-time scheduling class, mark the mode for hard interrupt
2086 * expiry.
2087 */
2088 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
2089 if (rt_or_dl_task_policy(current) && !(mode & HRTIMER_MODE_SOFT))
2090 mode |= HRTIMER_MODE_HARD;
2091 }
2092
2093 __hrtimer_init(&sl->timer, clock_id, mode);
2094 sl->timer.function = hrtimer_wakeup;
2095 sl->task = current;
2096}
2097
2098/**
2099 * hrtimer_setup_sleeper_on_stack - initialize a sleeper in stack memory
2100 * @sl: sleeper to be initialized
2101 * @clock_id: the clock to be used
2102 * @mode: timer mode abs/rel
2103 */
2104void hrtimer_setup_sleeper_on_stack(struct hrtimer_sleeper *sl,
2105 clockid_t clock_id, enum hrtimer_mode mode)
2106{
2107 debug_init_on_stack(&sl->timer, clock_id, mode);
2108 __hrtimer_init_sleeper(sl, clock_id, mode);
2109}
2110EXPORT_SYMBOL_GPL(hrtimer_setup_sleeper_on_stack);
2111
2112int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
2113{
2114 switch(restart->nanosleep.type) {
2115#ifdef CONFIG_COMPAT_32BIT_TIME
2116 case TT_COMPAT:
2117 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
2118 return -EFAULT;
2119 break;
2120#endif
2121 case TT_NATIVE:
2122 if (put_timespec64(ts, restart->nanosleep.rmtp))
2123 return -EFAULT;
2124 break;
2125 default:
2126 BUG();
2127 }
2128 return -ERESTART_RESTARTBLOCK;
2129}
2130
2131static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
2132{
2133 struct restart_block *restart;
2134
2135 do {
2136 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
2137 hrtimer_sleeper_start_expires(t, mode);
2138
2139 if (likely(t->task))
2140 schedule();
2141
2142 hrtimer_cancel(&t->timer);
2143 mode = HRTIMER_MODE_ABS;
2144
2145 } while (t->task && !signal_pending(current));
2146
2147 __set_current_state(TASK_RUNNING);
2148
2149 if (!t->task)
2150 return 0;
2151
2152 restart = ¤t->restart_block;
2153 if (restart->nanosleep.type != TT_NONE) {
2154 ktime_t rem = hrtimer_expires_remaining(&t->timer);
2155 struct timespec64 rmt;
2156
2157 if (rem <= 0)
2158 return 0;
2159 rmt = ktime_to_timespec64(rem);
2160
2161 return nanosleep_copyout(restart, &rmt);
2162 }
2163 return -ERESTART_RESTARTBLOCK;
2164}
2165
2166static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2167{
2168 struct hrtimer_sleeper t;
2169 int ret;
2170
2171 hrtimer_setup_sleeper_on_stack(&t, restart->nanosleep.clockid, HRTIMER_MODE_ABS);
2172 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
2173 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2174 destroy_hrtimer_on_stack(&t.timer);
2175 return ret;
2176}
2177
2178long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2179 const clockid_t clockid)
2180{
2181 struct restart_block *restart;
2182 struct hrtimer_sleeper t;
2183 int ret = 0;
2184
2185 hrtimer_setup_sleeper_on_stack(&t, clockid, mode);
2186 hrtimer_set_expires_range_ns(&t.timer, rqtp, current->timer_slack_ns);
2187 ret = do_nanosleep(&t, mode);
2188 if (ret != -ERESTART_RESTARTBLOCK)
2189 goto out;
2190
2191 /* Absolute timers do not update the rmtp value and restart: */
2192 if (mode == HRTIMER_MODE_ABS) {
2193 ret = -ERESTARTNOHAND;
2194 goto out;
2195 }
2196
2197 restart = ¤t->restart_block;
2198 restart->nanosleep.clockid = t.timer.base->clockid;
2199 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2200 set_restart_fn(restart, hrtimer_nanosleep_restart);
2201out:
2202 destroy_hrtimer_on_stack(&t.timer);
2203 return ret;
2204}
2205
2206#ifdef CONFIG_64BIT
2207
2208SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2209 struct __kernel_timespec __user *, rmtp)
2210{
2211 struct timespec64 tu;
2212
2213 if (get_timespec64(&tu, rqtp))
2214 return -EFAULT;
2215
2216 if (!timespec64_valid(&tu))
2217 return -EINVAL;
2218
2219 current->restart_block.fn = do_no_restart_syscall;
2220 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2221 current->restart_block.nanosleep.rmtp = rmtp;
2222 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2223 CLOCK_MONOTONIC);
2224}
2225
2226#endif
2227
2228#ifdef CONFIG_COMPAT_32BIT_TIME
2229
2230SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2231 struct old_timespec32 __user *, rmtp)
2232{
2233 struct timespec64 tu;
2234
2235 if (get_old_timespec32(&tu, rqtp))
2236 return -EFAULT;
2237
2238 if (!timespec64_valid(&tu))
2239 return -EINVAL;
2240
2241 current->restart_block.fn = do_no_restart_syscall;
2242 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2243 current->restart_block.nanosleep.compat_rmtp = rmtp;
2244 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2245 CLOCK_MONOTONIC);
2246}
2247#endif
2248
2249/*
2250 * Functions related to boot-time initialization:
2251 */
2252int hrtimers_prepare_cpu(unsigned int cpu)
2253{
2254 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2255 int i;
2256
2257 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2258 struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2259
2260 clock_b->cpu_base = cpu_base;
2261 seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2262 timerqueue_init_head(&clock_b->active);
2263 }
2264
2265 cpu_base->cpu = cpu;
2266 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2267 return 0;
2268}
2269
2270int hrtimers_cpu_starting(unsigned int cpu)
2271{
2272 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
2273
2274 /* Clear out any left over state from a CPU down operation */
2275 cpu_base->active_bases = 0;
2276 cpu_base->hres_active = 0;
2277 cpu_base->hang_detected = 0;
2278 cpu_base->next_timer = NULL;
2279 cpu_base->softirq_next_timer = NULL;
2280 cpu_base->expires_next = KTIME_MAX;
2281 cpu_base->softirq_expires_next = KTIME_MAX;
2282 cpu_base->online = 1;
2283 return 0;
2284}
2285
2286#ifdef CONFIG_HOTPLUG_CPU
2287
2288static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2289 struct hrtimer_clock_base *new_base)
2290{
2291 struct hrtimer *timer;
2292 struct timerqueue_node *node;
2293
2294 while ((node = timerqueue_getnext(&old_base->active))) {
2295 timer = container_of(node, struct hrtimer, node);
2296 BUG_ON(hrtimer_callback_running(timer));
2297 debug_deactivate(timer);
2298
2299 /*
2300 * Mark it as ENQUEUED not INACTIVE otherwise the
2301 * timer could be seen as !active and just vanish away
2302 * under us on another CPU
2303 */
2304 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2305 timer->base = new_base;
2306 /*
2307 * Enqueue the timers on the new cpu. This does not
2308 * reprogram the event device in case the timer
2309 * expires before the earliest on this CPU, but we run
2310 * hrtimer_interrupt after we migrated everything to
2311 * sort out already expired timers and reprogram the
2312 * event device.
2313 */
2314 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2315 }
2316}
2317
2318int hrtimers_cpu_dying(unsigned int dying_cpu)
2319{
2320 int i, ncpu = cpumask_any_and(cpu_active_mask, housekeeping_cpumask(HK_TYPE_TIMER));
2321 struct hrtimer_cpu_base *old_base, *new_base;
2322
2323 old_base = this_cpu_ptr(&hrtimer_bases);
2324 new_base = &per_cpu(hrtimer_bases, ncpu);
2325
2326 /*
2327 * The caller is globally serialized and nobody else
2328 * takes two locks at once, deadlock is not possible.
2329 */
2330 raw_spin_lock(&old_base->lock);
2331 raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2332
2333 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2334 migrate_hrtimer_list(&old_base->clock_base[i],
2335 &new_base->clock_base[i]);
2336 }
2337
2338 /*
2339 * The migration might have changed the first expiring softirq
2340 * timer on this CPU. Update it.
2341 */
2342 __hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
2343 /* Tell the other CPU to retrigger the next event */
2344 smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2345
2346 raw_spin_unlock(&new_base->lock);
2347 old_base->online = 0;
2348 raw_spin_unlock(&old_base->lock);
2349
2350 return 0;
2351}
2352
2353#endif /* CONFIG_HOTPLUG_CPU */
2354
2355void __init hrtimers_init(void)
2356{
2357 hrtimers_prepare_cpu(smp_processor_id());
2358 hrtimers_cpu_starting(smp_processor_id());
2359 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2360}