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