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 = {
138 .clock_base = { { .cpu_base = &migration_cpu_base, }, },
141 #define migration_base migration_cpu_base.clock_base[0]
143 static inline bool is_migration_base(struct hrtimer_clock_base *base)
145 return base == &migration_base;
149 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
150 * means that all timers which are tied to this base via timer->base are
151 * locked, and the base itself is locked too.
153 * So __run_timers/migrate_timers can safely modify all timers which could
154 * be found on the lists/queues.
156 * When the timer's base is locked, and the timer removed from list, it is
157 * possible to set timer->base = &migration_base and drop the lock: the timer
161 struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
162 unsigned long *flags)
164 struct hrtimer_clock_base *base;
167 base = READ_ONCE(timer->base);
168 if (likely(base != &migration_base)) {
169 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
170 if (likely(base == timer->base))
172 /* The timer has migrated to another CPU: */
173 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
180 * We do not migrate the timer when it is expiring before the next
181 * event on the target cpu. When high resolution is enabled, we cannot
182 * reprogram the target cpu hardware and we would cause it to fire
183 * late. To keep it simple, we handle the high resolution enabled and
184 * disabled case similar.
186 * Called with cpu_base->lock of target cpu held.
189 hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
193 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
194 return expires < new_base->cpu_base->expires_next;
198 struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
201 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
202 if (static_branch_likely(&timers_migration_enabled) && !pinned)
203 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
209 * We switch the timer base to a power-optimized selected CPU target,
211 * - NO_HZ_COMMON is enabled
212 * - timer migration is enabled
213 * - the timer callback is not running
214 * - the timer is not the first expiring timer on the new target
216 * If one of the above requirements is not fulfilled we move the timer
217 * to the current CPU or leave it on the previously assigned CPU if
218 * the timer callback is currently running.
220 static inline struct hrtimer_clock_base *
221 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
224 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
225 struct hrtimer_clock_base *new_base;
226 int basenum = base->index;
228 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
229 new_cpu_base = get_target_base(this_cpu_base, pinned);
231 new_base = &new_cpu_base->clock_base[basenum];
233 if (base != new_base) {
235 * We are trying to move timer to new_base.
236 * However we can't change timer's base while it is running,
237 * so we keep it on the same CPU. No hassle vs. reprogramming
238 * the event source in the high resolution case. The softirq
239 * code will take care of this when the timer function has
240 * completed. There is no conflict as we hold the lock until
241 * the timer is enqueued.
243 if (unlikely(hrtimer_callback_running(timer)))
246 /* See the comment in lock_hrtimer_base() */
247 WRITE_ONCE(timer->base, &migration_base);
248 raw_spin_unlock(&base->cpu_base->lock);
249 raw_spin_lock(&new_base->cpu_base->lock);
251 if (new_cpu_base != this_cpu_base &&
252 hrtimer_check_target(timer, new_base)) {
253 raw_spin_unlock(&new_base->cpu_base->lock);
254 raw_spin_lock(&base->cpu_base->lock);
255 new_cpu_base = this_cpu_base;
256 WRITE_ONCE(timer->base, base);
259 WRITE_ONCE(timer->base, new_base);
261 if (new_cpu_base != this_cpu_base &&
262 hrtimer_check_target(timer, new_base)) {
263 new_cpu_base = this_cpu_base;
270 #else /* CONFIG_SMP */
272 static inline bool is_migration_base(struct hrtimer_clock_base *base)
277 static inline struct hrtimer_clock_base *
278 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
280 struct hrtimer_clock_base *base = timer->base;
282 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
287 # define switch_hrtimer_base(t, b, p) (b)
289 #endif /* !CONFIG_SMP */
292 * Functions for the union type storage format of ktime_t which are
293 * too large for inlining:
295 #if BITS_PER_LONG < 64
297 * Divide a ktime value by a nanosecond value
299 s64 __ktime_divns(const ktime_t kt, s64 div)
305 dclc = ktime_to_ns(kt);
306 tmp = dclc < 0 ? -dclc : dclc;
308 /* Make sure the divisor is less than 2^32: */
314 do_div(tmp, (unsigned long) div);
315 return dclc < 0 ? -tmp : tmp;
317 EXPORT_SYMBOL_GPL(__ktime_divns);
318 #endif /* BITS_PER_LONG >= 64 */
321 * Add two ktime values and do a safety check for overflow:
323 ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
325 ktime_t res = ktime_add_unsafe(lhs, rhs);
328 * We use KTIME_SEC_MAX here, the maximum timeout which we can
329 * return to user space in a timespec:
331 if (res < 0 || res < lhs || res < rhs)
332 res = ktime_set(KTIME_SEC_MAX, 0);
337 EXPORT_SYMBOL_GPL(ktime_add_safe);
339 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
341 static struct debug_obj_descr hrtimer_debug_descr;
343 static void *hrtimer_debug_hint(void *addr)
345 return ((struct hrtimer *) addr)->function;
349 * fixup_init is called when:
350 * - an active object is initialized
352 static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
354 struct hrtimer *timer = addr;
357 case ODEBUG_STATE_ACTIVE:
358 hrtimer_cancel(timer);
359 debug_object_init(timer, &hrtimer_debug_descr);
367 * fixup_activate is called when:
368 * - an active object is activated
369 * - an unknown non-static object is activated
371 static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
374 case ODEBUG_STATE_ACTIVE:
383 * fixup_free is called when:
384 * - an active object is freed
386 static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
388 struct hrtimer *timer = addr;
391 case ODEBUG_STATE_ACTIVE:
392 hrtimer_cancel(timer);
393 debug_object_free(timer, &hrtimer_debug_descr);
400 static struct debug_obj_descr hrtimer_debug_descr = {
402 .debug_hint = hrtimer_debug_hint,
403 .fixup_init = hrtimer_fixup_init,
404 .fixup_activate = hrtimer_fixup_activate,
405 .fixup_free = hrtimer_fixup_free,
408 static inline void debug_hrtimer_init(struct hrtimer *timer)
410 debug_object_init(timer, &hrtimer_debug_descr);
413 static inline void debug_hrtimer_activate(struct hrtimer *timer,
414 enum hrtimer_mode mode)
416 debug_object_activate(timer, &hrtimer_debug_descr);
419 static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
421 debug_object_deactivate(timer, &hrtimer_debug_descr);
424 static inline void debug_hrtimer_free(struct hrtimer *timer)
426 debug_object_free(timer, &hrtimer_debug_descr);
429 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
430 enum hrtimer_mode mode);
432 void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
433 enum hrtimer_mode mode)
435 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
436 __hrtimer_init(timer, clock_id, mode);
438 EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
440 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
441 clockid_t clock_id, enum hrtimer_mode mode);
443 void hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper *sl,
444 clockid_t clock_id, enum hrtimer_mode mode)
446 debug_object_init_on_stack(&sl->timer, &hrtimer_debug_descr);
447 __hrtimer_init_sleeper(sl, clock_id, mode);
449 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper_on_stack);
451 void destroy_hrtimer_on_stack(struct hrtimer *timer)
453 debug_object_free(timer, &hrtimer_debug_descr);
455 EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
459 static inline void debug_hrtimer_init(struct hrtimer *timer) { }
460 static inline void debug_hrtimer_activate(struct hrtimer *timer,
461 enum hrtimer_mode mode) { }
462 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
466 debug_init(struct hrtimer *timer, clockid_t clockid,
467 enum hrtimer_mode mode)
469 debug_hrtimer_init(timer);
470 trace_hrtimer_init(timer, clockid, mode);
473 static inline void debug_activate(struct hrtimer *timer,
474 enum hrtimer_mode mode)
476 debug_hrtimer_activate(timer, mode);
477 trace_hrtimer_start(timer, mode);
480 static inline void debug_deactivate(struct hrtimer *timer)
482 debug_hrtimer_deactivate(timer);
483 trace_hrtimer_cancel(timer);
486 static struct hrtimer_clock_base *
487 __next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
494 idx = __ffs(*active);
495 *active &= ~(1U << idx);
497 return &cpu_base->clock_base[idx];
500 #define for_each_active_base(base, cpu_base, active) \
501 while ((base = __next_base((cpu_base), &(active))))
503 static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
504 const struct hrtimer *exclude,
506 ktime_t expires_next)
508 struct hrtimer_clock_base *base;
511 for_each_active_base(base, cpu_base, active) {
512 struct timerqueue_node *next;
513 struct hrtimer *timer;
515 next = timerqueue_getnext(&base->active);
516 timer = container_of(next, struct hrtimer, node);
517 if (timer == exclude) {
518 /* Get to the next timer in the queue. */
519 next = timerqueue_iterate_next(next);
523 timer = container_of(next, struct hrtimer, node);
525 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
526 if (expires < expires_next) {
527 expires_next = expires;
529 /* Skip cpu_base update if a timer is being excluded. */
534 cpu_base->softirq_next_timer = timer;
536 cpu_base->next_timer = timer;
540 * clock_was_set() might have changed base->offset of any of
541 * the clock bases so the result might be negative. Fix it up
542 * to prevent a false positive in clockevents_program_event().
544 if (expires_next < 0)
550 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
551 * but does not set cpu_base::*expires_next, that is done by
552 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
553 * cpu_base::*expires_next right away, reprogramming logic would no longer
556 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
557 * those timers will get run whenever the softirq gets handled, at the end of
558 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
560 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
561 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
562 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
564 * @active_mask must be one of:
565 * - HRTIMER_ACTIVE_ALL,
566 * - HRTIMER_ACTIVE_SOFT, or
567 * - HRTIMER_ACTIVE_HARD.
570 __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
573 struct hrtimer *next_timer = NULL;
574 ktime_t expires_next = KTIME_MAX;
576 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
577 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
578 cpu_base->softirq_next_timer = NULL;
579 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
582 next_timer = cpu_base->softirq_next_timer;
585 if (active_mask & HRTIMER_ACTIVE_HARD) {
586 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
587 cpu_base->next_timer = next_timer;
588 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
595 static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
597 ktime_t expires_next, soft = KTIME_MAX;
600 * If the soft interrupt has already been activated, ignore the
601 * soft bases. They will be handled in the already raised soft
604 if (!cpu_base->softirq_activated) {
605 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
607 * Update the soft expiry time. clock_settime() might have
610 cpu_base->softirq_expires_next = soft;
613 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
615 * If a softirq timer is expiring first, update cpu_base->next_timer
616 * and program the hardware with the soft expiry time.
618 if (expires_next > soft) {
619 cpu_base->next_timer = cpu_base->softirq_next_timer;
626 static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
628 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
629 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
630 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
632 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
633 offs_real, offs_boot, offs_tai);
635 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
636 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
637 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
643 * Is the high resolution mode active ?
645 static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
647 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
648 cpu_base->hres_active : 0;
651 static inline int hrtimer_hres_active(void)
653 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
657 * Reprogram the event source with checking both queues for the
659 * Called with interrupts disabled and base->lock held
662 hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
664 ktime_t expires_next;
666 expires_next = hrtimer_update_next_event(cpu_base);
668 if (skip_equal && expires_next == cpu_base->expires_next)
671 cpu_base->expires_next = expires_next;
674 * If hres is not active, hardware does not have to be
677 * If a hang was detected in the last timer interrupt then we
678 * leave the hang delay active in the hardware. We want the
679 * system to make progress. That also prevents the following
681 * T1 expires 50ms from now
682 * T2 expires 5s from now
684 * T1 is removed, so this code is called and would reprogram
685 * the hardware to 5s from now. Any hrtimer_start after that
686 * will not reprogram the hardware due to hang_detected being
687 * set. So we'd effectivly block all timers until the T2 event
690 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
693 tick_program_event(cpu_base->expires_next, 1);
696 /* High resolution timer related functions */
697 #ifdef CONFIG_HIGH_RES_TIMERS
700 * High resolution timer enabled ?
702 static bool hrtimer_hres_enabled __read_mostly = true;
703 unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
704 EXPORT_SYMBOL_GPL(hrtimer_resolution);
707 * Enable / Disable high resolution mode
709 static int __init setup_hrtimer_hres(char *str)
711 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
714 __setup("highres=", setup_hrtimer_hres);
717 * hrtimer_high_res_enabled - query, if the highres mode is enabled
719 static inline int hrtimer_is_hres_enabled(void)
721 return hrtimer_hres_enabled;
725 * Retrigger next event is called after clock was set
727 * Called with interrupts disabled via on_each_cpu()
729 static void retrigger_next_event(void *arg)
731 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
733 if (!__hrtimer_hres_active(base))
736 raw_spin_lock(&base->lock);
737 hrtimer_update_base(base);
738 hrtimer_force_reprogram(base, 0);
739 raw_spin_unlock(&base->lock);
743 * Switch to high resolution mode
745 static void hrtimer_switch_to_hres(void)
747 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
749 if (tick_init_highres()) {
750 pr_warn("Could not switch to high resolution mode on CPU %u\n",
754 base->hres_active = 1;
755 hrtimer_resolution = HIGH_RES_NSEC;
757 tick_setup_sched_timer();
758 /* "Retrigger" the interrupt to get things going */
759 retrigger_next_event(NULL);
764 static inline int hrtimer_is_hres_enabled(void) { return 0; }
765 static inline void hrtimer_switch_to_hres(void) { }
766 static inline void retrigger_next_event(void *arg) { }
768 #endif /* CONFIG_HIGH_RES_TIMERS */
771 * When a timer is enqueued and expires earlier than the already enqueued
772 * timers, we have to check, whether it expires earlier than the timer for
773 * which the clock event device was armed.
775 * Called with interrupts disabled and base->cpu_base.lock held
777 static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
779 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
780 struct hrtimer_clock_base *base = timer->base;
781 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
783 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
786 * CLOCK_REALTIME timer might be requested with an absolute
787 * expiry time which is less than base->offset. Set it to 0.
792 if (timer->is_soft) {
794 * soft hrtimer could be started on a remote CPU. In this
795 * case softirq_expires_next needs to be updated on the
796 * remote CPU. The soft hrtimer will not expire before the
797 * first hard hrtimer on the remote CPU -
798 * hrtimer_check_target() prevents this case.
800 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
802 if (timer_cpu_base->softirq_activated)
805 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
808 timer_cpu_base->softirq_next_timer = timer;
809 timer_cpu_base->softirq_expires_next = expires;
811 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
817 * If the timer is not on the current cpu, we cannot reprogram
818 * the other cpus clock event device.
820 if (base->cpu_base != cpu_base)
824 * If the hrtimer interrupt is running, then it will
825 * reevaluate the clock bases and reprogram the clock event
826 * device. The callbacks are always executed in hard interrupt
827 * context so we don't need an extra check for a running
830 if (cpu_base->in_hrtirq)
833 if (expires >= cpu_base->expires_next)
836 /* Update the pointer to the next expiring timer */
837 cpu_base->next_timer = timer;
838 cpu_base->expires_next = expires;
841 * If hres is not active, hardware does not have to be
844 * If a hang was detected in the last timer interrupt then we
845 * do not schedule a timer which is earlier than the expiry
846 * which we enforced in the hang detection. We want the system
849 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
853 * Program the timer hardware. We enforce the expiry for
854 * events which are already in the past.
856 tick_program_event(expires, 1);
860 * Clock realtime was set
862 * Change the offset of the realtime clock vs. the monotonic
865 * We might have to reprogram the high resolution timer interrupt. On
866 * SMP we call the architecture specific code to retrigger _all_ high
867 * resolution timer interrupts. On UP we just disable interrupts and
868 * call the high resolution interrupt code.
870 void clock_was_set(void)
872 #ifdef CONFIG_HIGH_RES_TIMERS
873 /* Retrigger the CPU local events everywhere */
874 on_each_cpu(retrigger_next_event, NULL, 1);
876 timerfd_clock_was_set();
879 static void clock_was_set_work(struct work_struct *work)
884 static DECLARE_WORK(hrtimer_work, clock_was_set_work);
887 * Called from timekeeping and resume code to reprogram the hrtimer
888 * interrupt device on all cpus and to notify timerfd.
890 void clock_was_set_delayed(void)
892 schedule_work(&hrtimer_work);
896 * During resume we might have to reprogram the high resolution timer
897 * interrupt on all online CPUs. However, all other CPUs will be
898 * stopped with IRQs interrupts disabled so the clock_was_set() call
901 void hrtimers_resume(void)
903 lockdep_assert_irqs_disabled();
904 /* Retrigger on the local CPU */
905 retrigger_next_event(NULL);
906 /* And schedule a retrigger for all others */
907 clock_was_set_delayed();
911 * Counterpart to lock_hrtimer_base above:
914 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
916 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
920 * hrtimer_forward - forward the timer expiry
921 * @timer: hrtimer to forward
922 * @now: forward past this time
923 * @interval: the interval to forward
925 * Forward the timer expiry so it will expire in the future.
926 * Returns the number of overruns.
928 * Can be safely called from the callback function of @timer. If
929 * called from other contexts @timer must neither be enqueued nor
930 * running the callback and the caller needs to take care of
933 * Note: This only updates the timer expiry value and does not requeue
936 u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
941 delta = ktime_sub(now, hrtimer_get_expires(timer));
946 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
949 if (interval < hrtimer_resolution)
950 interval = hrtimer_resolution;
952 if (unlikely(delta >= interval)) {
953 s64 incr = ktime_to_ns(interval);
955 orun = ktime_divns(delta, incr);
956 hrtimer_add_expires_ns(timer, incr * orun);
957 if (hrtimer_get_expires_tv64(timer) > now)
960 * This (and the ktime_add() below) is the
961 * correction for exact:
965 hrtimer_add_expires(timer, interval);
969 EXPORT_SYMBOL_GPL(hrtimer_forward);
972 * enqueue_hrtimer - internal function to (re)start a timer
974 * The timer is inserted in expiry order. Insertion into the
975 * red black tree is O(log(n)). Must hold the base lock.
977 * Returns 1 when the new timer is the leftmost timer in the tree.
979 static int enqueue_hrtimer(struct hrtimer *timer,
980 struct hrtimer_clock_base *base,
981 enum hrtimer_mode mode)
983 debug_activate(timer, mode);
984 WARN_ON_ONCE(!base->cpu_base->online);
986 base->cpu_base->active_bases |= 1 << base->index;
988 /* Pairs with the lockless read in hrtimer_is_queued() */
989 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
991 return timerqueue_add(&base->active, &timer->node);
995 * __remove_hrtimer - internal function to remove a timer
997 * Caller must hold the base lock.
999 * High resolution timer mode reprograms the clock event device when the
1000 * timer is the one which expires next. The caller can disable this by setting
1001 * reprogram to zero. This is useful, when the context does a reprogramming
1002 * anyway (e.g. timer interrupt)
1004 static void __remove_hrtimer(struct hrtimer *timer,
1005 struct hrtimer_clock_base *base,
1006 u8 newstate, int reprogram)
1008 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1009 u8 state = timer->state;
1011 /* Pairs with the lockless read in hrtimer_is_queued() */
1012 WRITE_ONCE(timer->state, newstate);
1013 if (!(state & HRTIMER_STATE_ENQUEUED))
1016 if (!timerqueue_del(&base->active, &timer->node))
1017 cpu_base->active_bases &= ~(1 << base->index);
1020 * Note: If reprogram is false we do not update
1021 * cpu_base->next_timer. This happens when we remove the first
1022 * timer on a remote cpu. No harm as we never dereference
1023 * cpu_base->next_timer. So the worst thing what can happen is
1024 * an superflous call to hrtimer_force_reprogram() on the
1025 * remote cpu later on if the same timer gets enqueued again.
1027 if (reprogram && timer == cpu_base->next_timer)
1028 hrtimer_force_reprogram(cpu_base, 1);
1032 * remove hrtimer, called with base lock held
1035 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1036 bool restart, bool keep_local)
1038 u8 state = timer->state;
1040 if (state & HRTIMER_STATE_ENQUEUED) {
1044 * Remove the timer and force reprogramming when high
1045 * resolution mode is active and the timer is on the current
1046 * CPU. If we remove a timer on another CPU, reprogramming is
1047 * skipped. The interrupt event on this CPU is fired and
1048 * reprogramming happens in the interrupt handler. This is a
1049 * rare case and less expensive than a smp call.
1051 debug_deactivate(timer);
1052 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1055 * If the timer is not restarted then reprogramming is
1056 * required if the timer is local. If it is local and about
1057 * to be restarted, avoid programming it twice (on removal
1058 * and a moment later when it's requeued).
1061 state = HRTIMER_STATE_INACTIVE;
1063 reprogram &= !keep_local;
1065 __remove_hrtimer(timer, base, state, reprogram);
1071 static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1072 const enum hrtimer_mode mode)
1074 #ifdef CONFIG_TIME_LOW_RES
1076 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1077 * granular time values. For relative timers we add hrtimer_resolution
1078 * (i.e. one jiffie) to prevent short timeouts.
1080 timer->is_rel = mode & HRTIMER_MODE_REL;
1082 tim = ktime_add_safe(tim, hrtimer_resolution);
1088 hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1093 * Find the next SOFT expiration.
1095 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1098 * reprogramming needs to be triggered, even if the next soft
1099 * hrtimer expires at the same time than the next hard
1100 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1102 if (expires == KTIME_MAX)
1106 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1107 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1109 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1112 static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1113 u64 delta_ns, const enum hrtimer_mode mode,
1114 struct hrtimer_clock_base *base)
1116 struct hrtimer_clock_base *new_base;
1117 bool force_local, first;
1120 * If the timer is on the local cpu base and is the first expiring
1121 * timer then this might end up reprogramming the hardware twice
1122 * (on removal and on enqueue). To avoid that by prevent the
1123 * reprogram on removal, keep the timer local to the current CPU
1124 * and enforce reprogramming after it is queued no matter whether
1125 * it is the new first expiring timer again or not.
1127 force_local = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1128 force_local &= base->cpu_base->next_timer == timer;
1131 * Remove an active timer from the queue. In case it is not queued
1132 * on the current CPU, make sure that remove_hrtimer() updates the
1133 * remote data correctly.
1135 * If it's on the current CPU and the first expiring timer, then
1136 * skip reprogramming, keep the timer local and enforce
1137 * reprogramming later if it was the first expiring timer. This
1138 * avoids programming the underlying clock event twice (once at
1139 * removal and once after enqueue).
1141 remove_hrtimer(timer, base, true, force_local);
1143 if (mode & HRTIMER_MODE_REL)
1144 tim = ktime_add_safe(tim, base->get_time());
1146 tim = hrtimer_update_lowres(timer, tim, mode);
1148 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1150 /* Switch the timer base, if necessary: */
1152 new_base = switch_hrtimer_base(timer, base,
1153 mode & HRTIMER_MODE_PINNED);
1158 first = enqueue_hrtimer(timer, new_base, mode);
1163 * Timer was forced to stay on the current CPU to avoid
1164 * reprogramming on removal and enqueue. Force reprogram the
1165 * hardware by evaluating the new first expiring timer.
1167 hrtimer_force_reprogram(new_base->cpu_base, 1);
1172 * hrtimer_start_range_ns - (re)start an hrtimer
1173 * @timer: the timer to be added
1175 * @delta_ns: "slack" range for the timer
1176 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1177 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1178 * softirq based mode is considered for debug purpose only!
1180 void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1181 u64 delta_ns, const enum hrtimer_mode mode)
1183 struct hrtimer_clock_base *base;
1184 unsigned long flags;
1187 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1188 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1189 * expiry mode because unmarked timers are moved to softirq expiry.
1191 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1192 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1194 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1196 base = lock_hrtimer_base(timer, &flags);
1198 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1199 hrtimer_reprogram(timer, true);
1201 unlock_hrtimer_base(timer, &flags);
1203 EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1206 * hrtimer_try_to_cancel - try to deactivate a timer
1207 * @timer: hrtimer to stop
1211 * * 0 when the timer was not active
1212 * * 1 when the timer was active
1213 * * -1 when the timer is currently executing the callback function and
1216 int hrtimer_try_to_cancel(struct hrtimer *timer)
1218 struct hrtimer_clock_base *base;
1219 unsigned long flags;
1223 * Check lockless first. If the timer is not active (neither
1224 * enqueued nor running the callback, nothing to do here. The
1225 * base lock does not serialize against a concurrent enqueue,
1226 * so we can avoid taking it.
1228 if (!hrtimer_active(timer))
1231 base = lock_hrtimer_base(timer, &flags);
1233 if (!hrtimer_callback_running(timer))
1234 ret = remove_hrtimer(timer, base, false, false);
1236 unlock_hrtimer_base(timer, &flags);
1241 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1243 #ifdef CONFIG_PREEMPT_RT
1244 static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1246 spin_lock_init(&base->softirq_expiry_lock);
1249 static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1251 spin_lock(&base->softirq_expiry_lock);
1254 static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1256 spin_unlock(&base->softirq_expiry_lock);
1260 * The counterpart to hrtimer_cancel_wait_running().
1262 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1263 * the timer callback to finish. Drop expiry_lock and reaquire it. That
1264 * allows the waiter to acquire the lock and make progress.
1266 static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1267 unsigned long flags)
1269 if (atomic_read(&cpu_base->timer_waiters)) {
1270 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1271 spin_unlock(&cpu_base->softirq_expiry_lock);
1272 spin_lock(&cpu_base->softirq_expiry_lock);
1273 raw_spin_lock_irq(&cpu_base->lock);
1278 * This function is called on PREEMPT_RT kernels when the fast path
1279 * deletion of a timer failed because the timer callback function was
1282 * This prevents priority inversion: if the soft irq thread is preempted
1283 * in the middle of a timer callback, then calling del_timer_sync() can
1284 * lead to two issues:
1286 * - If the caller is on a remote CPU then it has to spin wait for the timer
1287 * handler to complete. This can result in unbound priority inversion.
1289 * - If the caller originates from the task which preempted the timer
1290 * handler on the same CPU, then spin waiting for the timer handler to
1291 * complete is never going to end.
1293 void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1295 /* Lockless read. Prevent the compiler from reloading it below */
1296 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1299 * Just relax if the timer expires in hard interrupt context or if
1300 * it is currently on the migration base.
1302 if (!timer->is_soft || is_migration_base(base)) {
1308 * Mark the base as contended and grab the expiry lock, which is
1309 * held by the softirq across the timer callback. Drop the lock
1310 * immediately so the softirq can expire the next timer. In theory
1311 * the timer could already be running again, but that's more than
1312 * unlikely and just causes another wait loop.
1314 atomic_inc(&base->cpu_base->timer_waiters);
1315 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1316 atomic_dec(&base->cpu_base->timer_waiters);
1317 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1321 hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1323 hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1325 hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
1326 static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1327 unsigned long flags) { }
1331 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1332 * @timer: the timer to be cancelled
1335 * 0 when the timer was not active
1336 * 1 when the timer was active
1338 int hrtimer_cancel(struct hrtimer *timer)
1343 ret = hrtimer_try_to_cancel(timer);
1346 hrtimer_cancel_wait_running(timer);
1350 EXPORT_SYMBOL_GPL(hrtimer_cancel);
1353 * hrtimer_get_remaining - get remaining time for the timer
1354 * @timer: the timer to read
1355 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1357 ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1359 unsigned long flags;
1362 lock_hrtimer_base(timer, &flags);
1363 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1364 rem = hrtimer_expires_remaining_adjusted(timer);
1366 rem = hrtimer_expires_remaining(timer);
1367 unlock_hrtimer_base(timer, &flags);
1371 EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1373 #ifdef CONFIG_NO_HZ_COMMON
1375 * hrtimer_get_next_event - get the time until next expiry event
1377 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1379 u64 hrtimer_get_next_event(void)
1381 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1382 u64 expires = KTIME_MAX;
1383 unsigned long flags;
1385 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1387 if (!__hrtimer_hres_active(cpu_base))
1388 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1390 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1396 * hrtimer_next_event_without - time until next expiry event w/o one timer
1397 * @exclude: timer to exclude
1399 * Returns the next expiry time over all timers except for the @exclude one or
1400 * KTIME_MAX if none of them is pending.
1402 u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1404 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1405 u64 expires = KTIME_MAX;
1406 unsigned long flags;
1408 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1410 if (__hrtimer_hres_active(cpu_base)) {
1411 unsigned int active;
1413 if (!cpu_base->softirq_activated) {
1414 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1415 expires = __hrtimer_next_event_base(cpu_base, exclude,
1418 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1419 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1423 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1429 static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1431 if (likely(clock_id < MAX_CLOCKS)) {
1432 int base = hrtimer_clock_to_base_table[clock_id];
1434 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1437 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1438 return HRTIMER_BASE_MONOTONIC;
1441 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1442 enum hrtimer_mode mode)
1444 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1445 struct hrtimer_cpu_base *cpu_base;
1449 * On PREEMPT_RT enabled kernels hrtimers which are not explicitely
1450 * marked for hard interrupt expiry mode are moved into soft
1451 * interrupt context for latency reasons and because the callbacks
1452 * can invoke functions which might sleep on RT, e.g. spin_lock().
1454 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1457 memset(timer, 0, sizeof(struct hrtimer));
1459 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1462 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1463 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1464 * ensure POSIX compliance.
1466 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1467 clock_id = CLOCK_MONOTONIC;
1469 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1470 base += hrtimer_clockid_to_base(clock_id);
1471 timer->is_soft = softtimer;
1472 timer->is_hard = !softtimer;
1473 timer->base = &cpu_base->clock_base[base];
1474 timerqueue_init(&timer->node);
1478 * hrtimer_init - initialize a timer to the given clock
1479 * @timer: the timer to be initialized
1480 * @clock_id: the clock to be used
1481 * @mode: The modes which are relevant for intitialization:
1482 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1483 * HRTIMER_MODE_REL_SOFT
1485 * The PINNED variants of the above can be handed in,
1486 * but the PINNED bit is ignored as pinning happens
1487 * when the hrtimer is started
1489 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1490 enum hrtimer_mode mode)
1492 debug_init(timer, clock_id, mode);
1493 __hrtimer_init(timer, clock_id, mode);
1495 EXPORT_SYMBOL_GPL(hrtimer_init);
1498 * A timer is active, when it is enqueued into the rbtree or the
1499 * callback function is running or it's in the state of being migrated
1502 * It is important for this function to not return a false negative.
1504 bool hrtimer_active(const struct hrtimer *timer)
1506 struct hrtimer_clock_base *base;
1510 base = READ_ONCE(timer->base);
1511 seq = raw_read_seqcount_begin(&base->seq);
1513 if (timer->state != HRTIMER_STATE_INACTIVE ||
1514 base->running == timer)
1517 } while (read_seqcount_retry(&base->seq, seq) ||
1518 base != READ_ONCE(timer->base));
1522 EXPORT_SYMBOL_GPL(hrtimer_active);
1525 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1526 * distinct sections:
1528 * - queued: the timer is queued
1529 * - callback: the timer is being ran
1530 * - post: the timer is inactive or (re)queued
1532 * On the read side we ensure we observe timer->state and cpu_base->running
1533 * from the same section, if anything changed while we looked at it, we retry.
1534 * This includes timer->base changing because sequence numbers alone are
1535 * insufficient for that.
1537 * The sequence numbers are required because otherwise we could still observe
1538 * a false negative if the read side got smeared over multiple consequtive
1539 * __run_hrtimer() invocations.
1542 static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1543 struct hrtimer_clock_base *base,
1544 struct hrtimer *timer, ktime_t *now,
1545 unsigned long flags)
1547 enum hrtimer_restart (*fn)(struct hrtimer *);
1550 lockdep_assert_held(&cpu_base->lock);
1552 debug_deactivate(timer);
1553 base->running = timer;
1556 * Separate the ->running assignment from the ->state assignment.
1558 * As with a regular write barrier, this ensures the read side in
1559 * hrtimer_active() cannot observe base->running == NULL &&
1560 * timer->state == INACTIVE.
1562 raw_write_seqcount_barrier(&base->seq);
1564 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1565 fn = timer->function;
1568 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1569 * timer is restarted with a period then it becomes an absolute
1570 * timer. If its not restarted it does not matter.
1572 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1573 timer->is_rel = false;
1576 * The timer is marked as running in the CPU base, so it is
1577 * protected against migration to a different CPU even if the lock
1580 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1581 trace_hrtimer_expire_entry(timer, now);
1582 restart = fn(timer);
1583 trace_hrtimer_expire_exit(timer);
1584 raw_spin_lock_irq(&cpu_base->lock);
1587 * Note: We clear the running state after enqueue_hrtimer and
1588 * we do not reprogram the event hardware. Happens either in
1589 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1591 * Note: Because we dropped the cpu_base->lock above,
1592 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1595 if (restart != HRTIMER_NORESTART &&
1596 !(timer->state & HRTIMER_STATE_ENQUEUED))
1597 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
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.timer == NULL &&
1604 * timer->state == INACTIVE.
1606 raw_write_seqcount_barrier(&base->seq);
1608 WARN_ON_ONCE(base->running != timer);
1609 base->running = NULL;
1612 static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1613 unsigned long flags, unsigned int active_mask)
1615 struct hrtimer_clock_base *base;
1616 unsigned int active = cpu_base->active_bases & active_mask;
1618 for_each_active_base(base, cpu_base, active) {
1619 struct timerqueue_node *node;
1622 basenow = ktime_add(now, base->offset);
1624 while ((node = timerqueue_getnext(&base->active))) {
1625 struct hrtimer *timer;
1627 timer = container_of(node, struct hrtimer, node);
1630 * The immediate goal for using the softexpires is
1631 * minimizing wakeups, not running timers at the
1632 * earliest interrupt after their soft expiration.
1633 * This allows us to avoid using a Priority Search
1634 * Tree, which can answer a stabbing querry for
1635 * overlapping intervals and instead use the simple
1636 * BST we already have.
1637 * We don't add extra wakeups by delaying timers that
1638 * are right-of a not yet expired timer, because that
1639 * timer will have to trigger a wakeup anyway.
1641 if (basenow < hrtimer_get_softexpires_tv64(timer))
1644 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1645 if (active_mask == HRTIMER_ACTIVE_SOFT)
1646 hrtimer_sync_wait_running(cpu_base, flags);
1651 static __latent_entropy void hrtimer_run_softirq(struct softirq_action *h)
1653 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1654 unsigned long flags;
1657 hrtimer_cpu_base_lock_expiry(cpu_base);
1658 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1660 now = hrtimer_update_base(cpu_base);
1661 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1663 cpu_base->softirq_activated = 0;
1664 hrtimer_update_softirq_timer(cpu_base, true);
1666 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1667 hrtimer_cpu_base_unlock_expiry(cpu_base);
1670 #ifdef CONFIG_HIGH_RES_TIMERS
1673 * High resolution timer interrupt
1674 * Called with interrupts disabled
1676 void hrtimer_interrupt(struct clock_event_device *dev)
1678 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1679 ktime_t expires_next, now, entry_time, delta;
1680 unsigned long flags;
1683 BUG_ON(!cpu_base->hres_active);
1684 cpu_base->nr_events++;
1685 dev->next_event = KTIME_MAX;
1687 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1688 entry_time = now = hrtimer_update_base(cpu_base);
1690 cpu_base->in_hrtirq = 1;
1692 * We set expires_next to KTIME_MAX here with cpu_base->lock
1693 * held to prevent that a timer is enqueued in our queue via
1694 * the migration code. This does not affect enqueueing of
1695 * timers which run their callback and need to be requeued on
1698 cpu_base->expires_next = KTIME_MAX;
1700 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1701 cpu_base->softirq_expires_next = KTIME_MAX;
1702 cpu_base->softirq_activated = 1;
1703 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1706 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1708 /* Reevaluate the clock bases for the [soft] next expiry */
1709 expires_next = hrtimer_update_next_event(cpu_base);
1711 * Store the new expiry value so the migration code can verify
1714 cpu_base->expires_next = expires_next;
1715 cpu_base->in_hrtirq = 0;
1716 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1718 /* Reprogramming necessary ? */
1719 if (!tick_program_event(expires_next, 0)) {
1720 cpu_base->hang_detected = 0;
1725 * The next timer was already expired due to:
1727 * - long lasting callbacks
1728 * - being scheduled away when running in a VM
1730 * We need to prevent that we loop forever in the hrtimer
1731 * interrupt routine. We give it 3 attempts to avoid
1732 * overreacting on some spurious event.
1734 * Acquire base lock for updating the offsets and retrieving
1737 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1738 now = hrtimer_update_base(cpu_base);
1739 cpu_base->nr_retries++;
1743 * Give the system a chance to do something else than looping
1744 * here. We stored the entry time, so we know exactly how long
1745 * we spent here. We schedule the next event this amount of
1748 cpu_base->nr_hangs++;
1749 cpu_base->hang_detected = 1;
1750 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1752 delta = ktime_sub(now, entry_time);
1753 if ((unsigned int)delta > cpu_base->max_hang_time)
1754 cpu_base->max_hang_time = (unsigned int) delta;
1756 * Limit it to a sensible value as we enforce a longer
1757 * delay. Give the CPU at least 100ms to catch up.
1759 if (delta > 100 * NSEC_PER_MSEC)
1760 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1762 expires_next = ktime_add(now, delta);
1763 tick_program_event(expires_next, 1);
1764 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1767 /* called with interrupts disabled */
1768 static inline void __hrtimer_peek_ahead_timers(void)
1770 struct tick_device *td;
1772 if (!hrtimer_hres_active())
1775 td = this_cpu_ptr(&tick_cpu_device);
1776 if (td && td->evtdev)
1777 hrtimer_interrupt(td->evtdev);
1780 #else /* CONFIG_HIGH_RES_TIMERS */
1782 static inline void __hrtimer_peek_ahead_timers(void) { }
1784 #endif /* !CONFIG_HIGH_RES_TIMERS */
1787 * Called from run_local_timers in hardirq context every jiffy
1789 void hrtimer_run_queues(void)
1791 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1792 unsigned long flags;
1795 if (__hrtimer_hres_active(cpu_base))
1799 * This _is_ ugly: We have to check periodically, whether we
1800 * can switch to highres and / or nohz mode. The clocksource
1801 * switch happens with xtime_lock held. Notification from
1802 * there only sets the check bit in the tick_oneshot code,
1803 * otherwise we might deadlock vs. xtime_lock.
1805 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1806 hrtimer_switch_to_hres();
1810 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1811 now = hrtimer_update_base(cpu_base);
1813 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1814 cpu_base->softirq_expires_next = KTIME_MAX;
1815 cpu_base->softirq_activated = 1;
1816 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1819 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1820 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1824 * Sleep related functions:
1826 static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1828 struct hrtimer_sleeper *t =
1829 container_of(timer, struct hrtimer_sleeper, timer);
1830 struct task_struct *task = t->task;
1834 wake_up_process(task);
1836 return HRTIMER_NORESTART;
1840 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
1841 * @sl: sleeper to be started
1842 * @mode: timer mode abs/rel
1844 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
1845 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
1847 void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
1848 enum hrtimer_mode mode)
1851 * Make the enqueue delivery mode check work on RT. If the sleeper
1852 * was initialized for hard interrupt delivery, force the mode bit.
1853 * This is a special case for hrtimer_sleepers because
1854 * hrtimer_init_sleeper() determines the delivery mode on RT so the
1855 * fiddling with this decision is avoided at the call sites.
1857 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
1858 mode |= HRTIMER_MODE_HARD;
1860 hrtimer_start_expires(&sl->timer, mode);
1862 EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
1864 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
1865 clockid_t clock_id, enum hrtimer_mode mode)
1868 * On PREEMPT_RT enabled kernels hrtimers which are not explicitely
1869 * marked for hard interrupt expiry mode are moved into soft
1870 * interrupt context either for latency reasons or because the
1871 * hrtimer callback takes regular spinlocks or invokes other
1872 * functions which are not suitable for hard interrupt context on
1875 * The hrtimer_sleeper callback is RT compatible in hard interrupt
1876 * context, but there is a latency concern: Untrusted userspace can
1877 * spawn many threads which arm timers for the same expiry time on
1878 * the same CPU. That causes a latency spike due to the wakeup of
1879 * a gazillion threads.
1881 * OTOH, priviledged real-time user space applications rely on the
1882 * low latency of hard interrupt wakeups. If the current task is in
1883 * a real-time scheduling class, mark the mode for hard interrupt
1886 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1887 if (task_is_realtime(current) && !(mode & HRTIMER_MODE_SOFT))
1888 mode |= HRTIMER_MODE_HARD;
1891 __hrtimer_init(&sl->timer, clock_id, mode);
1892 sl->timer.function = hrtimer_wakeup;
1897 * hrtimer_init_sleeper - initialize sleeper to the given clock
1898 * @sl: sleeper to be initialized
1899 * @clock_id: the clock to be used
1900 * @mode: timer mode abs/rel
1902 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, clockid_t clock_id,
1903 enum hrtimer_mode mode)
1905 debug_init(&sl->timer, clock_id, mode);
1906 __hrtimer_init_sleeper(sl, clock_id, mode);
1909 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1911 int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
1913 switch(restart->nanosleep.type) {
1914 #ifdef CONFIG_COMPAT_32BIT_TIME
1916 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
1921 if (put_timespec64(ts, restart->nanosleep.rmtp))
1927 return -ERESTART_RESTARTBLOCK;
1930 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1932 struct restart_block *restart;
1935 set_current_state(TASK_INTERRUPTIBLE);
1936 hrtimer_sleeper_start_expires(t, mode);
1938 if (likely(t->task))
1939 freezable_schedule();
1941 hrtimer_cancel(&t->timer);
1942 mode = HRTIMER_MODE_ABS;
1944 } while (t->task && !signal_pending(current));
1946 __set_current_state(TASK_RUNNING);
1951 restart = ¤t->restart_block;
1952 if (restart->nanosleep.type != TT_NONE) {
1953 ktime_t rem = hrtimer_expires_remaining(&t->timer);
1954 struct timespec64 rmt;
1958 rmt = ktime_to_timespec64(rem);
1960 return nanosleep_copyout(restart, &rmt);
1962 return -ERESTART_RESTARTBLOCK;
1965 static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1967 struct hrtimer_sleeper t;
1970 hrtimer_init_sleeper_on_stack(&t, restart->nanosleep.clockid,
1972 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
1973 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
1974 destroy_hrtimer_on_stack(&t.timer);
1978 long hrtimer_nanosleep(const struct timespec64 *rqtp,
1979 const enum hrtimer_mode mode, const clockid_t clockid)
1981 struct restart_block *restart;
1982 struct hrtimer_sleeper t;
1986 slack = current->timer_slack_ns;
1987 if (dl_task(current) || rt_task(current))
1990 hrtimer_init_sleeper_on_stack(&t, clockid, mode);
1991 hrtimer_set_expires_range_ns(&t.timer, timespec64_to_ktime(*rqtp), slack);
1992 ret = do_nanosleep(&t, mode);
1993 if (ret != -ERESTART_RESTARTBLOCK)
1996 /* Absolute timers do not update the rmtp value and restart: */
1997 if (mode == HRTIMER_MODE_ABS) {
1998 ret = -ERESTARTNOHAND;
2002 restart = ¤t->restart_block;
2003 restart->nanosleep.clockid = t.timer.base->clockid;
2004 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2005 set_restart_fn(restart, hrtimer_nanosleep_restart);
2007 destroy_hrtimer_on_stack(&t.timer);
2011 #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2013 SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2014 struct __kernel_timespec __user *, rmtp)
2016 struct timespec64 tu;
2018 if (get_timespec64(&tu, rqtp))
2021 if (!timespec64_valid(&tu))
2024 current->restart_block.fn = do_no_restart_syscall;
2025 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2026 current->restart_block.nanosleep.rmtp = rmtp;
2027 return hrtimer_nanosleep(&tu, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
2032 #ifdef CONFIG_COMPAT_32BIT_TIME
2034 SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2035 struct old_timespec32 __user *, rmtp)
2037 struct timespec64 tu;
2039 if (get_old_timespec32(&tu, rqtp))
2042 if (!timespec64_valid(&tu))
2045 current->restart_block.fn = do_no_restart_syscall;
2046 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2047 current->restart_block.nanosleep.compat_rmtp = rmtp;
2048 return hrtimer_nanosleep(&tu, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
2053 * Functions related to boot-time initialization:
2055 int hrtimers_prepare_cpu(unsigned int cpu)
2057 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2060 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2061 cpu_base->clock_base[i].cpu_base = cpu_base;
2062 timerqueue_init_head(&cpu_base->clock_base[i].active);
2065 cpu_base->cpu = cpu;
2066 cpu_base->active_bases = 0;
2067 cpu_base->hres_active = 0;
2068 cpu_base->hang_detected = 0;
2069 cpu_base->next_timer = NULL;
2070 cpu_base->softirq_next_timer = NULL;
2071 cpu_base->expires_next = KTIME_MAX;
2072 cpu_base->softirq_expires_next = KTIME_MAX;
2073 cpu_base->online = 1;
2074 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2078 #ifdef CONFIG_HOTPLUG_CPU
2080 static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2081 struct hrtimer_clock_base *new_base)
2083 struct hrtimer *timer;
2084 struct timerqueue_node *node;
2086 while ((node = timerqueue_getnext(&old_base->active))) {
2087 timer = container_of(node, struct hrtimer, node);
2088 BUG_ON(hrtimer_callback_running(timer));
2089 debug_deactivate(timer);
2092 * Mark it as ENQUEUED not INACTIVE otherwise the
2093 * timer could be seen as !active and just vanish away
2094 * under us on another CPU
2096 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2097 timer->base = new_base;
2099 * Enqueue the timers on the new cpu. This does not
2100 * reprogram the event device in case the timer
2101 * expires before the earliest on this CPU, but we run
2102 * hrtimer_interrupt after we migrated everything to
2103 * sort out already expired timers and reprogram the
2106 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2110 int hrtimers_cpu_dying(unsigned int dying_cpu)
2112 struct hrtimer_cpu_base *old_base, *new_base;
2113 int i, ncpu = cpumask_first(cpu_active_mask);
2115 tick_cancel_sched_timer(dying_cpu);
2117 old_base = this_cpu_ptr(&hrtimer_bases);
2118 new_base = &per_cpu(hrtimer_bases, ncpu);
2121 * The caller is globally serialized and nobody else
2122 * takes two locks at once, deadlock is not possible.
2124 raw_spin_lock(&old_base->lock);
2125 raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2127 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2128 migrate_hrtimer_list(&old_base->clock_base[i],
2129 &new_base->clock_base[i]);
2133 * The migration might have changed the first expiring softirq
2134 * timer on this CPU. Update it.
2136 __hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
2137 /* Tell the other CPU to retrigger the next event */
2138 smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2140 raw_spin_unlock(&new_base->lock);
2141 old_base->online = 0;
2142 raw_spin_unlock(&old_base->lock);
2147 #endif /* CONFIG_HOTPLUG_CPU */
2149 void __init hrtimers_init(void)
2151 hrtimers_prepare_cpu(smp_processor_id());
2152 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2156 * schedule_hrtimeout_range_clock - sleep until timeout
2157 * @expires: timeout value (ktime_t)
2158 * @delta: slack in expires timeout (ktime_t)
2160 * @clock_id: timer clock to be used
2163 schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
2164 const enum hrtimer_mode mode, clockid_t clock_id)
2166 struct hrtimer_sleeper t;
2169 * Optimize when a zero timeout value is given. It does not
2170 * matter whether this is an absolute or a relative time.
2172 if (expires && *expires == 0) {
2173 __set_current_state(TASK_RUNNING);
2178 * A NULL parameter means "infinite"
2185 hrtimer_init_sleeper_on_stack(&t, clock_id, mode);
2186 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
2187 hrtimer_sleeper_start_expires(&t, mode);
2192 hrtimer_cancel(&t.timer);
2193 destroy_hrtimer_on_stack(&t.timer);
2195 __set_current_state(TASK_RUNNING);
2197 return !t.task ? 0 : -EINTR;
2201 * schedule_hrtimeout_range - sleep until timeout
2202 * @expires: timeout value (ktime_t)
2203 * @delta: slack in expires timeout (ktime_t)
2206 * Make the current task sleep until the given expiry time has
2207 * elapsed. The routine will return immediately unless
2208 * the current task state has been set (see set_current_state()).
2210 * The @delta argument gives the kernel the freedom to schedule the
2211 * actual wakeup to a time that is both power and performance friendly.
2212 * The kernel give the normal best effort behavior for "@expires+@delta",
2213 * but may decide to fire the timer earlier, but no earlier than @expires.
2215 * You can set the task state as follows -
2217 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2218 * pass before the routine returns unless the current task is explicitly
2219 * woken up, (e.g. by wake_up_process()).
2221 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2222 * delivered to the current task or the current task is explicitly woken
2225 * The current task state is guaranteed to be TASK_RUNNING when this
2228 * Returns 0 when the timer has expired. If the task was woken before the
2229 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2230 * by an explicit wakeup, it returns -EINTR.
2232 int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
2233 const enum hrtimer_mode mode)
2235 return schedule_hrtimeout_range_clock(expires, delta, mode,
2238 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
2241 * schedule_hrtimeout - sleep until timeout
2242 * @expires: timeout value (ktime_t)
2245 * Make the current task sleep until the given expiry time has
2246 * elapsed. The routine will return immediately unless
2247 * the current task state has been set (see set_current_state()).
2249 * You can set the task state as follows -
2251 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2252 * pass before the routine returns unless the current task is explicitly
2253 * woken up, (e.g. by wake_up_process()).
2255 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2256 * delivered to the current task or the current task is explicitly woken
2259 * The current task state is guaranteed to be TASK_RUNNING when this
2262 * Returns 0 when the timer has expired. If the task was woken before the
2263 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2264 * by an explicit wakeup, it returns -EINTR.
2266 int __sched schedule_hrtimeout(ktime_t *expires,
2267 const enum hrtimer_mode mode)
2269 return schedule_hrtimeout_range(expires, 0, mode);
2271 EXPORT_SYMBOL_GPL(schedule_hrtimeout);