1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 * High-resolution kernel timers
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.
13 * Started by: Thomas Gleixner and Ingo Molnar
16 * Based on the original timer wheel code
18 * Help, testing, suggestions, bugfixes, improvements were
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
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>
45 #include <linux/uaccess.h>
47 #include <trace/events/timer.h>
49 #include "tick-internal.h"
52 * Masks for selecting the soft and hard context timers from
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)
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.
68 DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
70 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
74 .index = HRTIMER_BASE_MONOTONIC,
75 .clockid = CLOCK_MONOTONIC,
76 .get_time = &ktime_get,
79 .index = HRTIMER_BASE_REALTIME,
80 .clockid = CLOCK_REALTIME,
81 .get_time = &ktime_get_real,
84 .index = HRTIMER_BASE_BOOTTIME,
85 .clockid = CLOCK_BOOTTIME,
86 .get_time = &ktime_get_boottime,
89 .index = HRTIMER_BASE_TAI,
91 .get_time = &ktime_get_clocktai,
94 .index = HRTIMER_BASE_MONOTONIC_SOFT,
95 .clockid = CLOCK_MONOTONIC,
96 .get_time = &ktime_get,
99 .index = HRTIMER_BASE_REALTIME_SOFT,
100 .clockid = CLOCK_REALTIME,
101 .get_time = &ktime_get_real,
104 .index = HRTIMER_BASE_BOOTTIME_SOFT,
105 .clockid = CLOCK_BOOTTIME,
106 .get_time = &ktime_get_boottime,
109 .index = HRTIMER_BASE_TAI_SOFT,
110 .clockid = CLOCK_TAI,
111 .get_time = &ktime_get_clocktai,
116 static 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,
120 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
121 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
122 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
123 [CLOCK_TAI] = HRTIMER_BASE_TAI,
127 * Functions and macros which are different for UP/SMP systems are kept in a
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
137 static struct hrtimer_cpu_base migration_cpu_base = {
139 .cpu_base = &migration_cpu_base,
140 .seq = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
141 &migration_cpu_base.lock),
145 #define migration_base migration_cpu_base.clock_base[0]
147 static inline bool is_migration_base(struct hrtimer_clock_base *base)
149 return base == &migration_base;
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.
157 * So __run_timers/migrate_timers can safely modify all timers which could
158 * be found on the lists/queues.
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
165 struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
166 unsigned long *flags)
168 struct hrtimer_clock_base *base;
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))
176 /* The timer has migrated to another CPU: */
177 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
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.
190 * Called with cpu_base->lock of target cpu held.
193 hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
197 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
198 return expires < new_base->cpu_base->expires_next;
202 struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
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());
213 * We switch the timer base to a power-optimized selected CPU target,
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
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.
224 static inline struct hrtimer_clock_base *
225 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
228 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
229 struct hrtimer_clock_base *new_base;
230 int basenum = base->index;
232 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
233 new_cpu_base = get_target_base(this_cpu_base, pinned);
235 new_base = &new_cpu_base->clock_base[basenum];
237 if (base != new_base) {
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.
247 if (unlikely(hrtimer_callback_running(timer)))
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);
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);
263 WRITE_ONCE(timer->base, new_base);
265 if (new_cpu_base != this_cpu_base &&
266 hrtimer_check_target(timer, new_base)) {
267 new_cpu_base = this_cpu_base;
274 #else /* CONFIG_SMP */
276 static inline bool is_migration_base(struct hrtimer_clock_base *base)
281 static inline struct hrtimer_clock_base *
282 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
284 struct hrtimer_clock_base *base = timer->base;
286 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
291 # define switch_hrtimer_base(t, b, p) (b)
293 #endif /* !CONFIG_SMP */
296 * Functions for the union type storage format of ktime_t which are
297 * too large for inlining:
299 #if BITS_PER_LONG < 64
301 * Divide a ktime value by a nanosecond value
303 s64 __ktime_divns(const ktime_t kt, s64 div)
309 dclc = ktime_to_ns(kt);
310 tmp = dclc < 0 ? -dclc : dclc;
312 /* Make sure the divisor is less than 2^32: */
318 do_div(tmp, (u32) div);
319 return dclc < 0 ? -tmp : tmp;
321 EXPORT_SYMBOL_GPL(__ktime_divns);
322 #endif /* BITS_PER_LONG >= 64 */
325 * Add two ktime values and do a safety check for overflow:
327 ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
329 ktime_t res = ktime_add_unsafe(lhs, rhs);
332 * We use KTIME_SEC_MAX here, the maximum timeout which we can
333 * return to user space in a timespec:
335 if (res < 0 || res < lhs || res < rhs)
336 res = ktime_set(KTIME_SEC_MAX, 0);
341 EXPORT_SYMBOL_GPL(ktime_add_safe);
343 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
345 static const struct debug_obj_descr hrtimer_debug_descr;
347 static void *hrtimer_debug_hint(void *addr)
349 return ((struct hrtimer *) addr)->function;
353 * fixup_init is called when:
354 * - an active object is initialized
356 static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
358 struct hrtimer *timer = addr;
361 case ODEBUG_STATE_ACTIVE:
362 hrtimer_cancel(timer);
363 debug_object_init(timer, &hrtimer_debug_descr);
371 * fixup_activate is called when:
372 * - an active object is activated
373 * - an unknown non-static object is activated
375 static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
378 case ODEBUG_STATE_ACTIVE:
387 * fixup_free is called when:
388 * - an active object is freed
390 static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
392 struct hrtimer *timer = addr;
395 case ODEBUG_STATE_ACTIVE:
396 hrtimer_cancel(timer);
397 debug_object_free(timer, &hrtimer_debug_descr);
404 static const struct debug_obj_descr hrtimer_debug_descr = {
406 .debug_hint = hrtimer_debug_hint,
407 .fixup_init = hrtimer_fixup_init,
408 .fixup_activate = hrtimer_fixup_activate,
409 .fixup_free = hrtimer_fixup_free,
412 static inline void debug_hrtimer_init(struct hrtimer *timer)
414 debug_object_init(timer, &hrtimer_debug_descr);
417 static inline void debug_hrtimer_activate(struct hrtimer *timer,
418 enum hrtimer_mode mode)
420 debug_object_activate(timer, &hrtimer_debug_descr);
423 static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
425 debug_object_deactivate(timer, &hrtimer_debug_descr);
428 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
429 enum hrtimer_mode mode);
431 void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
432 enum hrtimer_mode mode)
434 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
435 __hrtimer_init(timer, clock_id, mode);
437 EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
439 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
440 clockid_t clock_id, enum hrtimer_mode mode);
442 void hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper *sl,
443 clockid_t clock_id, enum hrtimer_mode mode)
445 debug_object_init_on_stack(&sl->timer, &hrtimer_debug_descr);
446 __hrtimer_init_sleeper(sl, clock_id, mode);
448 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper_on_stack);
450 void destroy_hrtimer_on_stack(struct hrtimer *timer)
452 debug_object_free(timer, &hrtimer_debug_descr);
454 EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
458 static inline void debug_hrtimer_init(struct hrtimer *timer) { }
459 static inline void debug_hrtimer_activate(struct hrtimer *timer,
460 enum hrtimer_mode mode) { }
461 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
465 debug_init(struct hrtimer *timer, clockid_t clockid,
466 enum hrtimer_mode mode)
468 debug_hrtimer_init(timer);
469 trace_hrtimer_init(timer, clockid, mode);
472 static inline void debug_activate(struct hrtimer *timer,
473 enum hrtimer_mode mode)
475 debug_hrtimer_activate(timer, mode);
476 trace_hrtimer_start(timer, mode);
479 static inline void debug_deactivate(struct hrtimer *timer)
481 debug_hrtimer_deactivate(timer);
482 trace_hrtimer_cancel(timer);
485 static struct hrtimer_clock_base *
486 __next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
493 idx = __ffs(*active);
494 *active &= ~(1U << idx);
496 return &cpu_base->clock_base[idx];
499 #define for_each_active_base(base, cpu_base, active) \
500 while ((base = __next_base((cpu_base), &(active))))
502 static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
503 const struct hrtimer *exclude,
505 ktime_t expires_next)
507 struct hrtimer_clock_base *base;
510 for_each_active_base(base, cpu_base, active) {
511 struct timerqueue_node *next;
512 struct hrtimer *timer;
514 next = timerqueue_getnext(&base->active);
515 timer = container_of(next, struct hrtimer, node);
516 if (timer == exclude) {
517 /* Get to the next timer in the queue. */
518 next = timerqueue_iterate_next(next);
522 timer = container_of(next, struct hrtimer, node);
524 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
525 if (expires < expires_next) {
526 expires_next = expires;
528 /* Skip cpu_base update if a timer is being excluded. */
533 cpu_base->softirq_next_timer = timer;
535 cpu_base->next_timer = timer;
539 * clock_was_set() might have changed base->offset of any of
540 * the clock bases so the result might be negative. Fix it up
541 * to prevent a false positive in clockevents_program_event().
543 if (expires_next < 0)
549 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
550 * but does not set cpu_base::*expires_next, that is done by
551 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
552 * cpu_base::*expires_next right away, reprogramming logic would no longer
555 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
556 * those timers will get run whenever the softirq gets handled, at the end of
557 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
559 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
560 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
561 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
563 * @active_mask must be one of:
564 * - HRTIMER_ACTIVE_ALL,
565 * - HRTIMER_ACTIVE_SOFT, or
566 * - HRTIMER_ACTIVE_HARD.
569 __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
572 struct hrtimer *next_timer = NULL;
573 ktime_t expires_next = KTIME_MAX;
575 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
576 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
577 cpu_base->softirq_next_timer = NULL;
578 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
581 next_timer = cpu_base->softirq_next_timer;
584 if (active_mask & HRTIMER_ACTIVE_HARD) {
585 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
586 cpu_base->next_timer = next_timer;
587 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
594 static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
596 ktime_t expires_next, soft = KTIME_MAX;
599 * If the soft interrupt has already been activated, ignore the
600 * soft bases. They will be handled in the already raised soft
603 if (!cpu_base->softirq_activated) {
604 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
606 * Update the soft expiry time. clock_settime() might have
609 cpu_base->softirq_expires_next = soft;
612 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
614 * If a softirq timer is expiring first, update cpu_base->next_timer
615 * and program the hardware with the soft expiry time.
617 if (expires_next > soft) {
618 cpu_base->next_timer = cpu_base->softirq_next_timer;
625 static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
627 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
628 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
629 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
631 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
632 offs_real, offs_boot, offs_tai);
634 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
635 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
636 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
642 * Is the high resolution mode active ?
644 static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
646 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
647 cpu_base->hres_active : 0;
650 static inline int hrtimer_hres_active(void)
652 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
656 __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal,
657 struct hrtimer *next_timer, ktime_t expires_next)
660 * If the hrtimer interrupt is running, then it will reevaluate the
661 * clock bases and reprogram the clock event device.
663 if (cpu_base->in_hrtirq)
666 if (expires_next > cpu_base->expires_next)
669 if (skip_equal && expires_next == cpu_base->expires_next)
672 cpu_base->next_timer = next_timer;
673 cpu_base->expires_next = expires_next;
676 * If hres is not active, hardware does not have to be
679 * If a hang was detected in the last timer interrupt then we
680 * leave the hang delay active in the hardware. We want the
681 * system to make progress. That also prevents the following
683 * T1 expires 50ms from now
684 * T2 expires 5s from now
686 * T1 is removed, so this code is called and would reprogram
687 * the hardware to 5s from now. Any hrtimer_start after that
688 * will not reprogram the hardware due to hang_detected being
689 * set. So we'd effectively block all timers until the T2 event
692 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
695 tick_program_event(expires_next, 1);
699 * Reprogram the event source with checking both queues for the
701 * Called with interrupts disabled and base->lock held
704 hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
706 ktime_t expires_next;
708 expires_next = hrtimer_update_next_event(cpu_base);
710 __hrtimer_reprogram(cpu_base, skip_equal, cpu_base->next_timer,
714 /* High resolution timer related functions */
715 #ifdef CONFIG_HIGH_RES_TIMERS
718 * High resolution timer enabled ?
720 static bool hrtimer_hres_enabled __read_mostly = true;
721 unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
722 EXPORT_SYMBOL_GPL(hrtimer_resolution);
725 * Enable / Disable high resolution mode
727 static int __init setup_hrtimer_hres(char *str)
729 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
732 __setup("highres=", setup_hrtimer_hres);
735 * hrtimer_high_res_enabled - query, if the highres mode is enabled
737 static inline int hrtimer_is_hres_enabled(void)
739 return hrtimer_hres_enabled;
742 static void retrigger_next_event(void *arg);
745 * Switch to high resolution mode
747 static void hrtimer_switch_to_hres(void)
749 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
751 if (tick_init_highres()) {
752 pr_warn("Could not switch to high resolution mode on CPU %u\n",
756 base->hres_active = 1;
757 hrtimer_resolution = HIGH_RES_NSEC;
759 tick_setup_sched_timer();
760 /* "Retrigger" the interrupt to get things going */
761 retrigger_next_event(NULL);
766 static inline int hrtimer_is_hres_enabled(void) { return 0; }
767 static inline void hrtimer_switch_to_hres(void) { }
769 #endif /* CONFIG_HIGH_RES_TIMERS */
771 * Retrigger next event is called after clock was set with interrupts
772 * disabled through an SMP function call or directly from low level
775 * This is only invoked when:
776 * - CONFIG_HIGH_RES_TIMERS is enabled.
777 * - CONFIG_NOHZ_COMMON is enabled
779 * For the other cases this function is empty and because the call sites
780 * are optimized out it vanishes as well, i.e. no need for lots of
783 static void retrigger_next_event(void *arg)
785 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
788 * When high resolution mode or nohz is active, then the offsets of
789 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
790 * next tick will take care of that.
792 * If high resolution mode is active then the next expiring timer
793 * must be reevaluated and the clock event device reprogrammed if
796 * In the NOHZ case the update of the offset and the reevaluation
797 * of the next expiring timer is enough. The return from the SMP
798 * function call will take care of the reprogramming in case the
799 * CPU was in a NOHZ idle sleep.
801 if (!__hrtimer_hres_active(base) && !tick_nohz_active)
804 raw_spin_lock(&base->lock);
805 hrtimer_update_base(base);
806 if (__hrtimer_hres_active(base))
807 hrtimer_force_reprogram(base, 0);
809 hrtimer_update_next_event(base);
810 raw_spin_unlock(&base->lock);
814 * When a timer is enqueued and expires earlier than the already enqueued
815 * timers, we have to check, whether it expires earlier than the timer for
816 * which the clock event device was armed.
818 * Called with interrupts disabled and base->cpu_base.lock held
820 static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
822 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
823 struct hrtimer_clock_base *base = timer->base;
824 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
826 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
829 * CLOCK_REALTIME timer might be requested with an absolute
830 * expiry time which is less than base->offset. Set it to 0.
835 if (timer->is_soft) {
837 * soft hrtimer could be started on a remote CPU. In this
838 * case softirq_expires_next needs to be updated on the
839 * remote CPU. The soft hrtimer will not expire before the
840 * first hard hrtimer on the remote CPU -
841 * hrtimer_check_target() prevents this case.
843 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
845 if (timer_cpu_base->softirq_activated)
848 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
851 timer_cpu_base->softirq_next_timer = timer;
852 timer_cpu_base->softirq_expires_next = expires;
854 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
860 * If the timer is not on the current cpu, we cannot reprogram
861 * the other cpus clock event device.
863 if (base->cpu_base != cpu_base)
866 __hrtimer_reprogram(cpu_base, true, timer, expires);
870 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
871 * CLOCK_BOOTTIME (for late sleep time injection).
873 * This requires to update the offsets for these clocks
874 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
875 * also requires to eventually reprogram the per CPU clock event devices
876 * when the change moves an affected timer ahead of the first expiring
877 * timer on that CPU. Obviously remote per CPU clock event devices cannot
878 * be reprogrammed. The other reason why an IPI has to be sent is when the
879 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
880 * in the tick, which obviously might be stopped, so this has to bring out
881 * the remote CPU which might sleep in idle to get this sorted.
883 void clock_was_set(unsigned int bases)
888 if (!hrtimer_hres_active() && !tick_nohz_active)
891 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
892 on_each_cpu(retrigger_next_event, NULL, 1);
896 /* Avoid interrupting CPUs if possible */
898 for_each_online_cpu(cpu) {
899 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
902 raw_spin_lock_irqsave(&cpu_base->lock, flags);
904 * Only send the IPI when there are timers queued in one of
905 * the affected clock bases. Otherwise update the base
906 * remote to ensure that the next enqueue of a timer on
907 * such a clock base will see the correct offsets.
909 if (cpu_base->active_bases & bases)
910 cpumask_set_cpu(cpu, mask);
912 hrtimer_update_base(cpu_base);
913 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
917 smp_call_function_many(mask, retrigger_next_event, NULL, 1);
920 free_cpumask_var(mask);
923 timerfd_clock_was_set();
926 static void clock_was_set_work(struct work_struct *work)
928 clock_was_set(CLOCK_SET_WALL);
931 static DECLARE_WORK(hrtimer_work, clock_was_set_work);
934 * Called from timekeeping code to reprogram the hrtimer interrupt device
935 * on all cpus and to notify timerfd.
937 void clock_was_set_delayed(void)
939 schedule_work(&hrtimer_work);
943 * Called during resume either directly from via timekeeping_resume()
944 * or in the case of s2idle from tick_unfreeze() to ensure that the
945 * hrtimers are up to date.
947 void hrtimers_resume_local(void)
949 lockdep_assert_irqs_disabled();
950 /* Retrigger on the local CPU */
951 retrigger_next_event(NULL);
955 * Counterpart to lock_hrtimer_base above:
958 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
960 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
964 * hrtimer_forward - forward the timer expiry
965 * @timer: hrtimer to forward
966 * @now: forward past this time
967 * @interval: the interval to forward
969 * Forward the timer expiry so it will expire in the future.
970 * Returns the number of overruns.
972 * Can be safely called from the callback function of @timer. If
973 * called from other contexts @timer must neither be enqueued nor
974 * running the callback and the caller needs to take care of
977 * Note: This only updates the timer expiry value and does not requeue
980 u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
985 delta = ktime_sub(now, hrtimer_get_expires(timer));
990 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
993 if (interval < hrtimer_resolution)
994 interval = hrtimer_resolution;
996 if (unlikely(delta >= interval)) {
997 s64 incr = ktime_to_ns(interval);
999 orun = ktime_divns(delta, incr);
1000 hrtimer_add_expires_ns(timer, incr * orun);
1001 if (hrtimer_get_expires_tv64(timer) > now)
1004 * This (and the ktime_add() below) is the
1005 * correction for exact:
1009 hrtimer_add_expires(timer, interval);
1013 EXPORT_SYMBOL_GPL(hrtimer_forward);
1016 * enqueue_hrtimer - internal function to (re)start a timer
1018 * The timer is inserted in expiry order. Insertion into the
1019 * red black tree is O(log(n)). Must hold the base lock.
1021 * Returns 1 when the new timer is the leftmost timer in the tree.
1023 static int enqueue_hrtimer(struct hrtimer *timer,
1024 struct hrtimer_clock_base *base,
1025 enum hrtimer_mode mode)
1027 debug_activate(timer, mode);
1029 base->cpu_base->active_bases |= 1 << base->index;
1031 /* Pairs with the lockless read in hrtimer_is_queued() */
1032 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1034 return timerqueue_add(&base->active, &timer->node);
1038 * __remove_hrtimer - internal function to remove a timer
1040 * Caller must hold the base lock.
1042 * High resolution timer mode reprograms the clock event device when the
1043 * timer is the one which expires next. The caller can disable this by setting
1044 * reprogram to zero. This is useful, when the context does a reprogramming
1045 * anyway (e.g. timer interrupt)
1047 static void __remove_hrtimer(struct hrtimer *timer,
1048 struct hrtimer_clock_base *base,
1049 u8 newstate, int reprogram)
1051 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1052 u8 state = timer->state;
1054 /* Pairs with the lockless read in hrtimer_is_queued() */
1055 WRITE_ONCE(timer->state, newstate);
1056 if (!(state & HRTIMER_STATE_ENQUEUED))
1059 if (!timerqueue_del(&base->active, &timer->node))
1060 cpu_base->active_bases &= ~(1 << base->index);
1063 * Note: If reprogram is false we do not update
1064 * cpu_base->next_timer. This happens when we remove the first
1065 * timer on a remote cpu. No harm as we never dereference
1066 * cpu_base->next_timer. So the worst thing what can happen is
1067 * an superfluous call to hrtimer_force_reprogram() on the
1068 * remote cpu later on if the same timer gets enqueued again.
1070 if (reprogram && timer == cpu_base->next_timer)
1071 hrtimer_force_reprogram(cpu_base, 1);
1075 * remove hrtimer, called with base lock held
1078 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1079 bool restart, bool keep_local)
1081 u8 state = timer->state;
1083 if (state & HRTIMER_STATE_ENQUEUED) {
1087 * Remove the timer and force reprogramming when high
1088 * resolution mode is active and the timer is on the current
1089 * CPU. If we remove a timer on another CPU, reprogramming is
1090 * skipped. The interrupt event on this CPU is fired and
1091 * reprogramming happens in the interrupt handler. This is a
1092 * rare case and less expensive than a smp call.
1094 debug_deactivate(timer);
1095 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1098 * If the timer is not restarted then reprogramming is
1099 * required if the timer is local. If it is local and about
1100 * to be restarted, avoid programming it twice (on removal
1101 * and a moment later when it's requeued).
1104 state = HRTIMER_STATE_INACTIVE;
1106 reprogram &= !keep_local;
1108 __remove_hrtimer(timer, base, state, reprogram);
1114 static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1115 const enum hrtimer_mode mode)
1117 #ifdef CONFIG_TIME_LOW_RES
1119 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1120 * granular time values. For relative timers we add hrtimer_resolution
1121 * (i.e. one jiffie) to prevent short timeouts.
1123 timer->is_rel = mode & HRTIMER_MODE_REL;
1125 tim = ktime_add_safe(tim, hrtimer_resolution);
1131 hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1136 * Find the next SOFT expiration.
1138 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1141 * reprogramming needs to be triggered, even if the next soft
1142 * hrtimer expires at the same time than the next hard
1143 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1145 if (expires == KTIME_MAX)
1149 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1150 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1152 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1155 static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1156 u64 delta_ns, const enum hrtimer_mode mode,
1157 struct hrtimer_clock_base *base)
1159 struct hrtimer_clock_base *new_base;
1160 bool force_local, first;
1163 * If the timer is on the local cpu base and is the first expiring
1164 * timer then this might end up reprogramming the hardware twice
1165 * (on removal and on enqueue). To avoid that by prevent the
1166 * reprogram on removal, keep the timer local to the current CPU
1167 * and enforce reprogramming after it is queued no matter whether
1168 * it is the new first expiring timer again or not.
1170 force_local = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1171 force_local &= base->cpu_base->next_timer == timer;
1174 * Remove an active timer from the queue. In case it is not queued
1175 * on the current CPU, make sure that remove_hrtimer() updates the
1176 * remote data correctly.
1178 * If it's on the current CPU and the first expiring timer, then
1179 * skip reprogramming, keep the timer local and enforce
1180 * reprogramming later if it was the first expiring timer. This
1181 * avoids programming the underlying clock event twice (once at
1182 * removal and once after enqueue).
1184 remove_hrtimer(timer, base, true, force_local);
1186 if (mode & HRTIMER_MODE_REL)
1187 tim = ktime_add_safe(tim, base->get_time());
1189 tim = hrtimer_update_lowres(timer, tim, mode);
1191 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1193 /* Switch the timer base, if necessary: */
1195 new_base = switch_hrtimer_base(timer, base,
1196 mode & HRTIMER_MODE_PINNED);
1201 first = enqueue_hrtimer(timer, new_base, mode);
1206 * Timer was forced to stay on the current CPU to avoid
1207 * reprogramming on removal and enqueue. Force reprogram the
1208 * hardware by evaluating the new first expiring timer.
1210 hrtimer_force_reprogram(new_base->cpu_base, 1);
1215 * hrtimer_start_range_ns - (re)start an hrtimer
1216 * @timer: the timer to be added
1218 * @delta_ns: "slack" range for the timer
1219 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1220 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1221 * softirq based mode is considered for debug purpose only!
1223 void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1224 u64 delta_ns, const enum hrtimer_mode mode)
1226 struct hrtimer_clock_base *base;
1227 unsigned long flags;
1230 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1231 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1232 * expiry mode because unmarked timers are moved to softirq expiry.
1234 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1235 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1237 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1239 base = lock_hrtimer_base(timer, &flags);
1241 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1242 hrtimer_reprogram(timer, true);
1244 unlock_hrtimer_base(timer, &flags);
1246 EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1249 * hrtimer_try_to_cancel - try to deactivate a timer
1250 * @timer: hrtimer to stop
1254 * * 0 when the timer was not active
1255 * * 1 when the timer was active
1256 * * -1 when the timer is currently executing the callback function and
1259 int hrtimer_try_to_cancel(struct hrtimer *timer)
1261 struct hrtimer_clock_base *base;
1262 unsigned long flags;
1266 * Check lockless first. If the timer is not active (neither
1267 * enqueued nor running the callback, nothing to do here. The
1268 * base lock does not serialize against a concurrent enqueue,
1269 * so we can avoid taking it.
1271 if (!hrtimer_active(timer))
1274 base = lock_hrtimer_base(timer, &flags);
1276 if (!hrtimer_callback_running(timer))
1277 ret = remove_hrtimer(timer, base, false, false);
1279 unlock_hrtimer_base(timer, &flags);
1284 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1286 #ifdef CONFIG_PREEMPT_RT
1287 static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1289 spin_lock_init(&base->softirq_expiry_lock);
1292 static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1294 spin_lock(&base->softirq_expiry_lock);
1297 static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1299 spin_unlock(&base->softirq_expiry_lock);
1303 * The counterpart to hrtimer_cancel_wait_running().
1305 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1306 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1307 * allows the waiter to acquire the lock and make progress.
1309 static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1310 unsigned long flags)
1312 if (atomic_read(&cpu_base->timer_waiters)) {
1313 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1314 spin_unlock(&cpu_base->softirq_expiry_lock);
1315 spin_lock(&cpu_base->softirq_expiry_lock);
1316 raw_spin_lock_irq(&cpu_base->lock);
1321 * This function is called on PREEMPT_RT kernels when the fast path
1322 * deletion of a timer failed because the timer callback function was
1325 * This prevents priority inversion: if the soft irq thread is preempted
1326 * in the middle of a timer callback, then calling del_timer_sync() can
1327 * lead to two issues:
1329 * - If the caller is on a remote CPU then it has to spin wait for the timer
1330 * handler to complete. This can result in unbound priority inversion.
1332 * - If the caller originates from the task which preempted the timer
1333 * handler on the same CPU, then spin waiting for the timer handler to
1334 * complete is never going to end.
1336 void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1338 /* Lockless read. Prevent the compiler from reloading it below */
1339 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1342 * Just relax if the timer expires in hard interrupt context or if
1343 * it is currently on the migration base.
1345 if (!timer->is_soft || is_migration_base(base)) {
1351 * Mark the base as contended and grab the expiry lock, which is
1352 * held by the softirq across the timer callback. Drop the lock
1353 * immediately so the softirq can expire the next timer. In theory
1354 * the timer could already be running again, but that's more than
1355 * unlikely and just causes another wait loop.
1357 atomic_inc(&base->cpu_base->timer_waiters);
1358 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1359 atomic_dec(&base->cpu_base->timer_waiters);
1360 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1364 hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1366 hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1368 hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
1369 static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1370 unsigned long flags) { }
1374 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1375 * @timer: the timer to be cancelled
1378 * 0 when the timer was not active
1379 * 1 when the timer was active
1381 int hrtimer_cancel(struct hrtimer *timer)
1386 ret = hrtimer_try_to_cancel(timer);
1389 hrtimer_cancel_wait_running(timer);
1393 EXPORT_SYMBOL_GPL(hrtimer_cancel);
1396 * __hrtimer_get_remaining - get remaining time for the timer
1397 * @timer: the timer to read
1398 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1400 ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1402 unsigned long flags;
1405 lock_hrtimer_base(timer, &flags);
1406 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1407 rem = hrtimer_expires_remaining_adjusted(timer);
1409 rem = hrtimer_expires_remaining(timer);
1410 unlock_hrtimer_base(timer, &flags);
1414 EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1416 #ifdef CONFIG_NO_HZ_COMMON
1418 * hrtimer_get_next_event - get the time until next expiry event
1420 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1422 u64 hrtimer_get_next_event(void)
1424 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1425 u64 expires = KTIME_MAX;
1426 unsigned long flags;
1428 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1430 if (!__hrtimer_hres_active(cpu_base))
1431 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1433 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1439 * hrtimer_next_event_without - time until next expiry event w/o one timer
1440 * @exclude: timer to exclude
1442 * Returns the next expiry time over all timers except for the @exclude one or
1443 * KTIME_MAX if none of them is pending.
1445 u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1447 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1448 u64 expires = KTIME_MAX;
1449 unsigned long flags;
1451 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1453 if (__hrtimer_hres_active(cpu_base)) {
1454 unsigned int active;
1456 if (!cpu_base->softirq_activated) {
1457 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1458 expires = __hrtimer_next_event_base(cpu_base, exclude,
1461 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1462 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1466 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1472 static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1474 if (likely(clock_id < MAX_CLOCKS)) {
1475 int base = hrtimer_clock_to_base_table[clock_id];
1477 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1480 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1481 return HRTIMER_BASE_MONOTONIC;
1484 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1485 enum hrtimer_mode mode)
1487 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1488 struct hrtimer_cpu_base *cpu_base;
1492 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1493 * marked for hard interrupt expiry mode are moved into soft
1494 * interrupt context for latency reasons and because the callbacks
1495 * can invoke functions which might sleep on RT, e.g. spin_lock().
1497 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1500 memset(timer, 0, sizeof(struct hrtimer));
1502 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1505 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1506 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1507 * ensure POSIX compliance.
1509 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1510 clock_id = CLOCK_MONOTONIC;
1512 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1513 base += hrtimer_clockid_to_base(clock_id);
1514 timer->is_soft = softtimer;
1515 timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1516 timer->base = &cpu_base->clock_base[base];
1517 timerqueue_init(&timer->node);
1521 * hrtimer_init - initialize a timer to the given clock
1522 * @timer: the timer to be initialized
1523 * @clock_id: the clock to be used
1524 * @mode: The modes which are relevant for initialization:
1525 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1526 * HRTIMER_MODE_REL_SOFT
1528 * The PINNED variants of the above can be handed in,
1529 * but the PINNED bit is ignored as pinning happens
1530 * when the hrtimer is started
1532 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1533 enum hrtimer_mode mode)
1535 debug_init(timer, clock_id, mode);
1536 __hrtimer_init(timer, clock_id, mode);
1538 EXPORT_SYMBOL_GPL(hrtimer_init);
1541 * A timer is active, when it is enqueued into the rbtree or the
1542 * callback function is running or it's in the state of being migrated
1545 * It is important for this function to not return a false negative.
1547 bool hrtimer_active(const struct hrtimer *timer)
1549 struct hrtimer_clock_base *base;
1553 base = READ_ONCE(timer->base);
1554 seq = raw_read_seqcount_begin(&base->seq);
1556 if (timer->state != HRTIMER_STATE_INACTIVE ||
1557 base->running == timer)
1560 } while (read_seqcount_retry(&base->seq, seq) ||
1561 base != READ_ONCE(timer->base));
1565 EXPORT_SYMBOL_GPL(hrtimer_active);
1568 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1569 * distinct sections:
1571 * - queued: the timer is queued
1572 * - callback: the timer is being ran
1573 * - post: the timer is inactive or (re)queued
1575 * On the read side we ensure we observe timer->state and cpu_base->running
1576 * from the same section, if anything changed while we looked at it, we retry.
1577 * This includes timer->base changing because sequence numbers alone are
1578 * insufficient for that.
1580 * The sequence numbers are required because otherwise we could still observe
1581 * a false negative if the read side got smeared over multiple consecutive
1582 * __run_hrtimer() invocations.
1585 static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1586 struct hrtimer_clock_base *base,
1587 struct hrtimer *timer, ktime_t *now,
1588 unsigned long flags) __must_hold(&cpu_base->lock)
1590 enum hrtimer_restart (*fn)(struct hrtimer *);
1591 bool expires_in_hardirq;
1594 lockdep_assert_held(&cpu_base->lock);
1596 debug_deactivate(timer);
1597 base->running = timer;
1600 * Separate the ->running assignment from the ->state assignment.
1602 * As with a regular write barrier, this ensures the read side in
1603 * hrtimer_active() cannot observe base->running == NULL &&
1604 * timer->state == INACTIVE.
1606 raw_write_seqcount_barrier(&base->seq);
1608 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1609 fn = timer->function;
1612 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1613 * timer is restarted with a period then it becomes an absolute
1614 * timer. If its not restarted it does not matter.
1616 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1617 timer->is_rel = false;
1620 * The timer is marked as running in the CPU base, so it is
1621 * protected against migration to a different CPU even if the lock
1624 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1625 trace_hrtimer_expire_entry(timer, now);
1626 expires_in_hardirq = lockdep_hrtimer_enter(timer);
1628 restart = fn(timer);
1630 lockdep_hrtimer_exit(expires_in_hardirq);
1631 trace_hrtimer_expire_exit(timer);
1632 raw_spin_lock_irq(&cpu_base->lock);
1635 * Note: We clear the running state after enqueue_hrtimer and
1636 * we do not reprogram the event hardware. Happens either in
1637 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1639 * Note: Because we dropped the cpu_base->lock above,
1640 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1643 if (restart != HRTIMER_NORESTART &&
1644 !(timer->state & HRTIMER_STATE_ENQUEUED))
1645 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1648 * Separate the ->running assignment from the ->state assignment.
1650 * As with a regular write barrier, this ensures the read side in
1651 * hrtimer_active() cannot observe base->running.timer == NULL &&
1652 * timer->state == INACTIVE.
1654 raw_write_seqcount_barrier(&base->seq);
1656 WARN_ON_ONCE(base->running != timer);
1657 base->running = NULL;
1660 static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1661 unsigned long flags, unsigned int active_mask)
1663 struct hrtimer_clock_base *base;
1664 unsigned int active = cpu_base->active_bases & active_mask;
1666 for_each_active_base(base, cpu_base, active) {
1667 struct timerqueue_node *node;
1670 basenow = ktime_add(now, base->offset);
1672 while ((node = timerqueue_getnext(&base->active))) {
1673 struct hrtimer *timer;
1675 timer = container_of(node, struct hrtimer, node);
1678 * The immediate goal for using the softexpires is
1679 * minimizing wakeups, not running timers at the
1680 * earliest interrupt after their soft expiration.
1681 * This allows us to avoid using a Priority Search
1682 * Tree, which can answer a stabbing query for
1683 * overlapping intervals and instead use the simple
1684 * BST we already have.
1685 * We don't add extra wakeups by delaying timers that
1686 * are right-of a not yet expired timer, because that
1687 * timer will have to trigger a wakeup anyway.
1689 if (basenow < hrtimer_get_softexpires_tv64(timer))
1692 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1693 if (active_mask == HRTIMER_ACTIVE_SOFT)
1694 hrtimer_sync_wait_running(cpu_base, flags);
1699 static __latent_entropy void hrtimer_run_softirq(struct softirq_action *h)
1701 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1702 unsigned long flags;
1705 hrtimer_cpu_base_lock_expiry(cpu_base);
1706 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1708 now = hrtimer_update_base(cpu_base);
1709 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1711 cpu_base->softirq_activated = 0;
1712 hrtimer_update_softirq_timer(cpu_base, true);
1714 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1715 hrtimer_cpu_base_unlock_expiry(cpu_base);
1718 #ifdef CONFIG_HIGH_RES_TIMERS
1721 * High resolution timer interrupt
1722 * Called with interrupts disabled
1724 void hrtimer_interrupt(struct clock_event_device *dev)
1726 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1727 ktime_t expires_next, now, entry_time, delta;
1728 unsigned long flags;
1731 BUG_ON(!cpu_base->hres_active);
1732 cpu_base->nr_events++;
1733 dev->next_event = KTIME_MAX;
1735 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1736 entry_time = now = hrtimer_update_base(cpu_base);
1738 cpu_base->in_hrtirq = 1;
1740 * We set expires_next to KTIME_MAX here with cpu_base->lock
1741 * held to prevent that a timer is enqueued in our queue via
1742 * the migration code. This does not affect enqueueing of
1743 * timers which run their callback and need to be requeued on
1746 cpu_base->expires_next = KTIME_MAX;
1748 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1749 cpu_base->softirq_expires_next = KTIME_MAX;
1750 cpu_base->softirq_activated = 1;
1751 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1754 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1756 /* Reevaluate the clock bases for the [soft] next expiry */
1757 expires_next = hrtimer_update_next_event(cpu_base);
1759 * Store the new expiry value so the migration code can verify
1762 cpu_base->expires_next = expires_next;
1763 cpu_base->in_hrtirq = 0;
1764 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1766 /* Reprogramming necessary ? */
1767 if (!tick_program_event(expires_next, 0)) {
1768 cpu_base->hang_detected = 0;
1773 * The next timer was already expired due to:
1775 * - long lasting callbacks
1776 * - being scheduled away when running in a VM
1778 * We need to prevent that we loop forever in the hrtimer
1779 * interrupt routine. We give it 3 attempts to avoid
1780 * overreacting on some spurious event.
1782 * Acquire base lock for updating the offsets and retrieving
1785 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1786 now = hrtimer_update_base(cpu_base);
1787 cpu_base->nr_retries++;
1791 * Give the system a chance to do something else than looping
1792 * here. We stored the entry time, so we know exactly how long
1793 * we spent here. We schedule the next event this amount of
1796 cpu_base->nr_hangs++;
1797 cpu_base->hang_detected = 1;
1798 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1800 delta = ktime_sub(now, entry_time);
1801 if ((unsigned int)delta > cpu_base->max_hang_time)
1802 cpu_base->max_hang_time = (unsigned int) delta;
1804 * Limit it to a sensible value as we enforce a longer
1805 * delay. Give the CPU at least 100ms to catch up.
1807 if (delta > 100 * NSEC_PER_MSEC)
1808 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1810 expires_next = ktime_add(now, delta);
1811 tick_program_event(expires_next, 1);
1812 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1815 /* called with interrupts disabled */
1816 static inline void __hrtimer_peek_ahead_timers(void)
1818 struct tick_device *td;
1820 if (!hrtimer_hres_active())
1823 td = this_cpu_ptr(&tick_cpu_device);
1824 if (td && td->evtdev)
1825 hrtimer_interrupt(td->evtdev);
1828 #else /* CONFIG_HIGH_RES_TIMERS */
1830 static inline void __hrtimer_peek_ahead_timers(void) { }
1832 #endif /* !CONFIG_HIGH_RES_TIMERS */
1835 * Called from run_local_timers in hardirq context every jiffy
1837 void hrtimer_run_queues(void)
1839 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1840 unsigned long flags;
1843 if (__hrtimer_hres_active(cpu_base))
1847 * This _is_ ugly: We have to check periodically, whether we
1848 * can switch to highres and / or nohz mode. The clocksource
1849 * switch happens with xtime_lock held. Notification from
1850 * there only sets the check bit in the tick_oneshot code,
1851 * otherwise we might deadlock vs. xtime_lock.
1853 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1854 hrtimer_switch_to_hres();
1858 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1859 now = hrtimer_update_base(cpu_base);
1861 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1862 cpu_base->softirq_expires_next = KTIME_MAX;
1863 cpu_base->softirq_activated = 1;
1864 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1867 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1868 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1872 * Sleep related functions:
1874 static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1876 struct hrtimer_sleeper *t =
1877 container_of(timer, struct hrtimer_sleeper, timer);
1878 struct task_struct *task = t->task;
1882 wake_up_process(task);
1884 return HRTIMER_NORESTART;
1888 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
1889 * @sl: sleeper to be started
1890 * @mode: timer mode abs/rel
1892 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
1893 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
1895 void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
1896 enum hrtimer_mode mode)
1899 * Make the enqueue delivery mode check work on RT. If the sleeper
1900 * was initialized for hard interrupt delivery, force the mode bit.
1901 * This is a special case for hrtimer_sleepers because
1902 * hrtimer_init_sleeper() determines the delivery mode on RT so the
1903 * fiddling with this decision is avoided at the call sites.
1905 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
1906 mode |= HRTIMER_MODE_HARD;
1908 hrtimer_start_expires(&sl->timer, mode);
1910 EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
1912 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
1913 clockid_t clock_id, enum hrtimer_mode mode)
1916 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1917 * marked for hard interrupt expiry mode are moved into soft
1918 * interrupt context either for latency reasons or because the
1919 * hrtimer callback takes regular spinlocks or invokes other
1920 * functions which are not suitable for hard interrupt context on
1923 * The hrtimer_sleeper callback is RT compatible in hard interrupt
1924 * context, but there is a latency concern: Untrusted userspace can
1925 * spawn many threads which arm timers for the same expiry time on
1926 * the same CPU. That causes a latency spike due to the wakeup of
1927 * a gazillion threads.
1929 * OTOH, privileged real-time user space applications rely on the
1930 * low latency of hard interrupt wakeups. If the current task is in
1931 * a real-time scheduling class, mark the mode for hard interrupt
1934 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1935 if (task_is_realtime(current) && !(mode & HRTIMER_MODE_SOFT))
1936 mode |= HRTIMER_MODE_HARD;
1939 __hrtimer_init(&sl->timer, clock_id, mode);
1940 sl->timer.function = hrtimer_wakeup;
1945 * hrtimer_init_sleeper - initialize sleeper to the given clock
1946 * @sl: sleeper to be initialized
1947 * @clock_id: the clock to be used
1948 * @mode: timer mode abs/rel
1950 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, clockid_t clock_id,
1951 enum hrtimer_mode mode)
1953 debug_init(&sl->timer, clock_id, mode);
1954 __hrtimer_init_sleeper(sl, clock_id, mode);
1957 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1959 int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
1961 switch(restart->nanosleep.type) {
1962 #ifdef CONFIG_COMPAT_32BIT_TIME
1964 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
1969 if (put_timespec64(ts, restart->nanosleep.rmtp))
1975 return -ERESTART_RESTARTBLOCK;
1978 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1980 struct restart_block *restart;
1983 set_current_state(TASK_INTERRUPTIBLE);
1984 hrtimer_sleeper_start_expires(t, mode);
1986 if (likely(t->task))
1987 freezable_schedule();
1989 hrtimer_cancel(&t->timer);
1990 mode = HRTIMER_MODE_ABS;
1992 } while (t->task && !signal_pending(current));
1994 __set_current_state(TASK_RUNNING);
1999 restart = ¤t->restart_block;
2000 if (restart->nanosleep.type != TT_NONE) {
2001 ktime_t rem = hrtimer_expires_remaining(&t->timer);
2002 struct timespec64 rmt;
2006 rmt = ktime_to_timespec64(rem);
2008 return nanosleep_copyout(restart, &rmt);
2010 return -ERESTART_RESTARTBLOCK;
2013 static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2015 struct hrtimer_sleeper t;
2018 hrtimer_init_sleeper_on_stack(&t, restart->nanosleep.clockid,
2020 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
2021 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2022 destroy_hrtimer_on_stack(&t.timer);
2026 long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2027 const clockid_t clockid)
2029 struct restart_block *restart;
2030 struct hrtimer_sleeper t;
2034 slack = current->timer_slack_ns;
2035 if (dl_task(current) || rt_task(current))
2038 hrtimer_init_sleeper_on_stack(&t, clockid, mode);
2039 hrtimer_set_expires_range_ns(&t.timer, rqtp, slack);
2040 ret = do_nanosleep(&t, mode);
2041 if (ret != -ERESTART_RESTARTBLOCK)
2044 /* Absolute timers do not update the rmtp value and restart: */
2045 if (mode == HRTIMER_MODE_ABS) {
2046 ret = -ERESTARTNOHAND;
2050 restart = ¤t->restart_block;
2051 restart->nanosleep.clockid = t.timer.base->clockid;
2052 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2053 set_restart_fn(restart, hrtimer_nanosleep_restart);
2055 destroy_hrtimer_on_stack(&t.timer);
2061 SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2062 struct __kernel_timespec __user *, rmtp)
2064 struct timespec64 tu;
2066 if (get_timespec64(&tu, rqtp))
2069 if (!timespec64_valid(&tu))
2072 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2073 current->restart_block.nanosleep.rmtp = rmtp;
2074 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2080 #ifdef CONFIG_COMPAT_32BIT_TIME
2082 SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2083 struct old_timespec32 __user *, rmtp)
2085 struct timespec64 tu;
2087 if (get_old_timespec32(&tu, rqtp))
2090 if (!timespec64_valid(&tu))
2093 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2094 current->restart_block.nanosleep.compat_rmtp = rmtp;
2095 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2101 * Functions related to boot-time initialization:
2103 int hrtimers_prepare_cpu(unsigned int cpu)
2105 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2108 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2109 struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2111 clock_b->cpu_base = cpu_base;
2112 seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2113 timerqueue_init_head(&clock_b->active);
2116 cpu_base->cpu = cpu;
2117 cpu_base->active_bases = 0;
2118 cpu_base->hres_active = 0;
2119 cpu_base->hang_detected = 0;
2120 cpu_base->next_timer = NULL;
2121 cpu_base->softirq_next_timer = NULL;
2122 cpu_base->expires_next = KTIME_MAX;
2123 cpu_base->softirq_expires_next = KTIME_MAX;
2124 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2128 #ifdef CONFIG_HOTPLUG_CPU
2130 static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2131 struct hrtimer_clock_base *new_base)
2133 struct hrtimer *timer;
2134 struct timerqueue_node *node;
2136 while ((node = timerqueue_getnext(&old_base->active))) {
2137 timer = container_of(node, struct hrtimer, node);
2138 BUG_ON(hrtimer_callback_running(timer));
2139 debug_deactivate(timer);
2142 * Mark it as ENQUEUED not INACTIVE otherwise the
2143 * timer could be seen as !active and just vanish away
2144 * under us on another CPU
2146 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2147 timer->base = new_base;
2149 * Enqueue the timers on the new cpu. This does not
2150 * reprogram the event device in case the timer
2151 * expires before the earliest on this CPU, but we run
2152 * hrtimer_interrupt after we migrated everything to
2153 * sort out already expired timers and reprogram the
2156 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2160 int hrtimers_dead_cpu(unsigned int scpu)
2162 struct hrtimer_cpu_base *old_base, *new_base;
2165 BUG_ON(cpu_online(scpu));
2166 tick_cancel_sched_timer(scpu);
2169 * this BH disable ensures that raise_softirq_irqoff() does
2170 * not wakeup ksoftirqd (and acquire the pi-lock) while
2171 * holding the cpu_base lock
2174 local_irq_disable();
2175 old_base = &per_cpu(hrtimer_bases, scpu);
2176 new_base = this_cpu_ptr(&hrtimer_bases);
2178 * The caller is globally serialized and nobody else
2179 * takes two locks at once, deadlock is not possible.
2181 raw_spin_lock(&new_base->lock);
2182 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
2184 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2185 migrate_hrtimer_list(&old_base->clock_base[i],
2186 &new_base->clock_base[i]);
2190 * The migration might have changed the first expiring softirq
2191 * timer on this CPU. Update it.
2193 hrtimer_update_softirq_timer(new_base, false);
2195 raw_spin_unlock(&old_base->lock);
2196 raw_spin_unlock(&new_base->lock);
2198 /* Check, if we got expired work to do */
2199 __hrtimer_peek_ahead_timers();
2205 #endif /* CONFIG_HOTPLUG_CPU */
2207 void __init hrtimers_init(void)
2209 hrtimers_prepare_cpu(smp_processor_id());
2210 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2214 * schedule_hrtimeout_range_clock - sleep until timeout
2215 * @expires: timeout value (ktime_t)
2216 * @delta: slack in expires timeout (ktime_t)
2218 * @clock_id: timer clock to be used
2221 schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
2222 const enum hrtimer_mode mode, clockid_t clock_id)
2224 struct hrtimer_sleeper t;
2227 * Optimize when a zero timeout value is given. It does not
2228 * matter whether this is an absolute or a relative time.
2230 if (expires && *expires == 0) {
2231 __set_current_state(TASK_RUNNING);
2236 * A NULL parameter means "infinite"
2243 hrtimer_init_sleeper_on_stack(&t, clock_id, mode);
2244 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
2245 hrtimer_sleeper_start_expires(&t, mode);
2250 hrtimer_cancel(&t.timer);
2251 destroy_hrtimer_on_stack(&t.timer);
2253 __set_current_state(TASK_RUNNING);
2255 return !t.task ? 0 : -EINTR;
2259 * schedule_hrtimeout_range - sleep until timeout
2260 * @expires: timeout value (ktime_t)
2261 * @delta: slack in expires timeout (ktime_t)
2264 * Make the current task sleep until the given expiry time has
2265 * elapsed. The routine will return immediately unless
2266 * the current task state has been set (see set_current_state()).
2268 * The @delta argument gives the kernel the freedom to schedule the
2269 * actual wakeup to a time that is both power and performance friendly.
2270 * The kernel give the normal best effort behavior for "@expires+@delta",
2271 * but may decide to fire the timer earlier, but no earlier than @expires.
2273 * You can set the task state as follows -
2275 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2276 * pass before the routine returns unless the current task is explicitly
2277 * woken up, (e.g. by wake_up_process()).
2279 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2280 * delivered to the current task or the current task is explicitly woken
2283 * The current task state is guaranteed to be TASK_RUNNING when this
2286 * Returns 0 when the timer has expired. If the task was woken before the
2287 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2288 * by an explicit wakeup, it returns -EINTR.
2290 int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
2291 const enum hrtimer_mode mode)
2293 return schedule_hrtimeout_range_clock(expires, delta, mode,
2296 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
2299 * schedule_hrtimeout - sleep until timeout
2300 * @expires: timeout value (ktime_t)
2303 * Make the current task sleep until the given expiry time has
2304 * elapsed. The routine will return immediately unless
2305 * the current task state has been set (see set_current_state()).
2307 * You can set the task state as follows -
2309 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2310 * pass before the routine returns unless the current task is explicitly
2311 * woken up, (e.g. by wake_up_process()).
2313 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2314 * delivered to the current task or the current task is explicitly woken
2317 * The current task state is guaranteed to be TASK_RUNNING when this
2320 * Returns 0 when the timer has expired. If the task was woken before the
2321 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2322 * by an explicit wakeup, it returns -EINTR.
2324 int __sched schedule_hrtimeout(ktime_t *expires,
2325 const enum hrtimer_mode mode)
2327 return schedule_hrtimeout_range(expires, 0, mode);
2329 EXPORT_SYMBOL_GPL(schedule_hrtimeout);