projects/linux·blob v6.14 · refs

2361 lines · 66.0 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2005-2006, Thomas Gleixner <[email protected]>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API, aka timer wheel,
10 * hrtimers provide finer resolution and accuracy depending on system
11 * configuration and capabilities.
12 *
13 * Started by: Thomas Gleixner and Ingo Molnar
14 *
15 * Credits:
16 * Based on the original timer wheel code
17 *
18 * Help, testing, suggestions, bugfixes, improvements were
19 * provided by:
20 *
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
22 * et. al.
23 */
24
25#include <linux/cpu.h>
26#include <linux/export.h>
27#include <linux/percpu.h>
28#include <linux/hrtimer.h>
29#include <linux/notifier.h>
30#include <linux/syscalls.h>
31#include <linux/interrupt.h>
32#include <linux/tick.h>
33#include <linux/err.h>
34#include <linux/debugobjects.h>
35#include <linux/sched/signal.h>
36#include <linux/sched/sysctl.h>
37#include <linux/sched/rt.h>
38#include <linux/sched/deadline.h>
39#include <linux/sched/nohz.h>
40#include <linux/sched/debug.h>
41#include <linux/sched/isolation.h>
42#include <linux/timer.h>
43#include <linux/freezer.h>
44#include <linux/compat.h>
45
46#include <linux/uaccess.h>
47
48#include <trace/events/timer.h>
49
50#include "tick-internal.h"
51
52/*
53 * Masks for selecting the soft and hard context timers from
54 * cpu_base->active
55 */
56#define MASK_SHIFT (HRTIMER_BASE_MONOTONIC_SOFT)
57#define HRTIMER_ACTIVE_HARD ((1U << MASK_SHIFT) - 1)
58#define HRTIMER_ACTIVE_SOFT (HRTIMER_ACTIVE_HARD << MASK_SHIFT)
59#define HRTIMER_ACTIVE_ALL (HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
60
61static void retrigger_next_event(void *arg);
62
63/*
64 * The timer bases:
65 *
66 * There are more clockids than hrtimer bases. Thus, we index
67 * into the timer bases by the hrtimer_base_type enum. When trying
68 * to reach a base using a clockid, hrtimer_clockid_to_base()
69 * is used to convert from clockid to the proper hrtimer_base_type.
70 */
71DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
72{
73 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
74 .clock_base =
75 {
76 {
77 .index = HRTIMER_BASE_MONOTONIC,
78 .clockid = CLOCK_MONOTONIC,
79 .get_time = &ktime_get,
80 },
81 {
82 .index = HRTIMER_BASE_REALTIME,
83 .clockid = CLOCK_REALTIME,
84 .get_time = &ktime_get_real,
85 },
86 {
87 .index = HRTIMER_BASE_BOOTTIME,
88 .clockid = CLOCK_BOOTTIME,
89 .get_time = &ktime_get_boottime,
90 },
91 {
92 .index = HRTIMER_BASE_TAI,
93 .clockid = CLOCK_TAI,
94 .get_time = &ktime_get_clocktai,
95 },
96 {
97 .index = HRTIMER_BASE_MONOTONIC_SOFT,
98 .clockid = CLOCK_MONOTONIC,
99 .get_time = &ktime_get,
100 },
101 {
102 .index = HRTIMER_BASE_REALTIME_SOFT,
103 .clockid = CLOCK_REALTIME,
104 .get_time = &ktime_get_real,
105 },
106 {
107 .index = HRTIMER_BASE_BOOTTIME_SOFT,
108 .clockid = CLOCK_BOOTTIME,
109 .get_time = &ktime_get_boottime,
110 },
111 {
112 .index = HRTIMER_BASE_TAI_SOFT,
113 .clockid = CLOCK_TAI,
114 .get_time = &ktime_get_clocktai,
115 },
116 },
117 .csd = CSD_INIT(retrigger_next_event, NULL)
118};
119
120static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
121 /* Make sure we catch unsupported clockids */
122 [0 ... MAX_CLOCKS - 1] = HRTIMER_MAX_CLOCK_BASES,
123
124 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
125 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
126 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
127 [CLOCK_TAI] = HRTIMER_BASE_TAI,
128};
129
130static inline bool hrtimer_base_is_online(struct hrtimer_cpu_base *base)
131{
132 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
133 return true;
134 else
135 return likely(base->online);
136}
137
138/*
139 * Functions and macros which are different for UP/SMP systems are kept in a
140 * single place
141 */
142#ifdef CONFIG_SMP
143
144/*
145 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
146 * such that hrtimer_callback_running() can unconditionally dereference
147 * timer->base->cpu_base
148 */
149static struct hrtimer_cpu_base migration_cpu_base = {
150 .clock_base = { {
151 .cpu_base = &migration_cpu_base,
152 .seq = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
153 &migration_cpu_base.lock),
154 }, },
155};
156
157#define migration_base migration_cpu_base.clock_base[0]
158
159/*
160 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
161 * means that all timers which are tied to this base via timer->base are
162 * locked, and the base itself is locked too.
163 *
164 * So __run_timers/migrate_timers can safely modify all timers which could
165 * be found on the lists/queues.
166 *
167 * When the timer's base is locked, and the timer removed from list, it is
168 * possible to set timer->base = &migration_base and drop the lock: the timer
169 * remains locked.
170 */
171static
172struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
173 unsigned long *flags)
174 __acquires(&timer->base->lock)
175{
176 struct hrtimer_clock_base *base;
177
178 for (;;) {
179 base = READ_ONCE(timer->base);
180 if (likely(base != &migration_base)) {
181 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
182 if (likely(base == timer->base))
183 return base;
184 /* The timer has migrated to another CPU: */
185 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
186 }
187 cpu_relax();
188 }
189}
190
191/*
192 * Check if the elected target is suitable considering its next
193 * event and the hotplug state of the current CPU.
194 *
195 * If the elected target is remote and its next event is after the timer
196 * to queue, then a remote reprogram is necessary. However there is no
197 * guarantee the IPI handling the operation would arrive in time to meet
198 * the high resolution deadline. In this case the local CPU becomes a
199 * preferred target, unless it is offline.
200 *
201 * High and low resolution modes are handled the same way for simplicity.
202 *
203 * Called with cpu_base->lock of target cpu held.
204 */
205static bool hrtimer_suitable_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base,
206 struct hrtimer_cpu_base *new_cpu_base,
207 struct hrtimer_cpu_base *this_cpu_base)
208{
209 ktime_t expires;
210
211 /*
212 * The local CPU clockevent can be reprogrammed. Also get_target_base()
213 * guarantees it is online.
214 */
215 if (new_cpu_base == this_cpu_base)
216 return true;
217
218 /*
219 * The offline local CPU can't be the default target if the
220 * next remote target event is after this timer. Keep the
221 * elected new base. An IPI will we issued to reprogram
222 * it as a last resort.
223 */
224 if (!hrtimer_base_is_online(this_cpu_base))
225 return true;
226
227 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
228
229 return expires >= new_base->cpu_base->expires_next;
230}
231
232static inline struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base, int pinned)
233{
234 if (!hrtimer_base_is_online(base)) {
235 int cpu = cpumask_any_and(cpu_online_mask, housekeeping_cpumask(HK_TYPE_TIMER));
236
237 return &per_cpu(hrtimer_bases, cpu);
238 }
239
240#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
241 if (static_branch_likely(&timers_migration_enabled) && !pinned)
242 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
243#endif
244 return base;
245}
246
247/*
248 * We switch the timer base to a power-optimized selected CPU target,
249 * if:
250 * - NO_HZ_COMMON is enabled
251 * - timer migration is enabled
252 * - the timer callback is not running
253 * - the timer is not the first expiring timer on the new target
254 *
255 * If one of the above requirements is not fulfilled we move the timer
256 * to the current CPU or leave it on the previously assigned CPU if
257 * the timer callback is currently running.
258 */
259static inline struct hrtimer_clock_base *
260switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
261 int pinned)
262{
263 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
264 struct hrtimer_clock_base *new_base;
265 int basenum = base->index;
266
267 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
268 new_cpu_base = get_target_base(this_cpu_base, pinned);
269again:
270 new_base = &new_cpu_base->clock_base[basenum];
271
272 if (base != new_base) {
273 /*
274 * We are trying to move timer to new_base.
275 * However we can't change timer's base while it is running,
276 * so we keep it on the same CPU. No hassle vs. reprogramming
277 * the event source in the high resolution case. The softirq
278 * code will take care of this when the timer function has
279 * completed. There is no conflict as we hold the lock until
280 * the timer is enqueued.
281 */
282 if (unlikely(hrtimer_callback_running(timer)))
283 return base;
284
285 /* See the comment in lock_hrtimer_base() */
286 WRITE_ONCE(timer->base, &migration_base);
287 raw_spin_unlock(&base->cpu_base->lock);
288 raw_spin_lock(&new_base->cpu_base->lock);
289
290 if (!hrtimer_suitable_target(timer, new_base, new_cpu_base,
291 this_cpu_base)) {
292 raw_spin_unlock(&new_base->cpu_base->lock);
293 raw_spin_lock(&base->cpu_base->lock);
294 new_cpu_base = this_cpu_base;
295 WRITE_ONCE(timer->base, base);
296 goto again;
297 }
298 WRITE_ONCE(timer->base, new_base);
299 } else {
300 if (!hrtimer_suitable_target(timer, new_base, new_cpu_base, this_cpu_base)) {
301 new_cpu_base = this_cpu_base;
302 goto again;
303 }
304 }
305 return new_base;
306}
307
308#else /* CONFIG_SMP */
309
310static inline struct hrtimer_clock_base *
311lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
312 __acquires(&timer->base->cpu_base->lock)
313{
314 struct hrtimer_clock_base *base = timer->base;
315
316 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
317
318 return base;
319}
320
321# define switch_hrtimer_base(t, b, p) (b)
322
323#endif /* !CONFIG_SMP */
324
325/*
326 * Functions for the union type storage format of ktime_t which are
327 * too large for inlining:
328 */
329#if BITS_PER_LONG < 64
330/*
331 * Divide a ktime value by a nanosecond value
332 */
333s64 __ktime_divns(const ktime_t kt, s64 div)
334{
335 int sft = 0;
336 s64 dclc;
337 u64 tmp;
338
339 dclc = ktime_to_ns(kt);
340 tmp = dclc < 0 ? -dclc : dclc;
341
342 /* Make sure the divisor is less than 2^32: */
343 while (div >> 32) {
344 sft++;
345 div >>= 1;
346 }
347 tmp >>= sft;
348 do_div(tmp, (u32) div);
349 return dclc < 0 ? -tmp : tmp;
350}
351EXPORT_SYMBOL_GPL(__ktime_divns);
352#endif /* BITS_PER_LONG >= 64 */
353
354/*
355 * Add two ktime values and do a safety check for overflow:
356 */
357ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
358{
359 ktime_t res = ktime_add_unsafe(lhs, rhs);
360
361 /*
362 * We use KTIME_SEC_MAX here, the maximum timeout which we can
363 * return to user space in a timespec:
364 */
365 if (res < 0 || res < lhs || res < rhs)
366 res = ktime_set(KTIME_SEC_MAX, 0);
367
368 return res;
369}
370
371EXPORT_SYMBOL_GPL(ktime_add_safe);
372
373#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
374
375static const struct debug_obj_descr hrtimer_debug_descr;
376
377static void *hrtimer_debug_hint(void *addr)
378{
379 return ((struct hrtimer *) addr)->function;
380}
381
382/*
383 * fixup_init is called when:
384 * - an active object is initialized
385 */
386static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
387{
388 struct hrtimer *timer = addr;
389
390 switch (state) {
391 case ODEBUG_STATE_ACTIVE:
392 hrtimer_cancel(timer);
393 debug_object_init(timer, &hrtimer_debug_descr);
394 return true;
395 default:
396 return false;
397 }
398}
399
400/*
401 * fixup_activate is called when:
402 * - an active object is activated
403 * - an unknown non-static object is activated
404 */
405static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
406{
407 switch (state) {
408 case ODEBUG_STATE_ACTIVE:
409 WARN_ON(1);
410 fallthrough;
411 default:
412 return false;
413 }
414}
415
416/*
417 * fixup_free is called when:
418 * - an active object is freed
419 */
420static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
421{
422 struct hrtimer *timer = addr;
423
424 switch (state) {
425 case ODEBUG_STATE_ACTIVE:
426 hrtimer_cancel(timer);
427 debug_object_free(timer, &hrtimer_debug_descr);
428 return true;
429 default:
430 return false;
431 }
432}
433
434static const struct debug_obj_descr hrtimer_debug_descr = {
435 .name = "hrtimer",
436 .debug_hint = hrtimer_debug_hint,
437 .fixup_init = hrtimer_fixup_init,
438 .fixup_activate = hrtimer_fixup_activate,
439 .fixup_free = hrtimer_fixup_free,
440};
441
442static inline void debug_hrtimer_init(struct hrtimer *timer)
443{
444 debug_object_init(timer, &hrtimer_debug_descr);
445}
446
447static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer)
448{
449 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
450}
451
452static inline void debug_hrtimer_activate(struct hrtimer *timer,
453 enum hrtimer_mode mode)
454{
455 debug_object_activate(timer, &hrtimer_debug_descr);
456}
457
458static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
459{
460 debug_object_deactivate(timer, &hrtimer_debug_descr);
461}
462
463void destroy_hrtimer_on_stack(struct hrtimer *timer)
464{
465 debug_object_free(timer, &hrtimer_debug_descr);
466}
467EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
468
469#else
470
471static inline void debug_hrtimer_init(struct hrtimer *timer) { }
472static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer) { }
473static inline void debug_hrtimer_activate(struct hrtimer *timer,
474 enum hrtimer_mode mode) { }
475static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
476#endif
477
478static inline void
479debug_init(struct hrtimer *timer, clockid_t clockid,
480 enum hrtimer_mode mode)
481{
482 debug_hrtimer_init(timer);
483 trace_hrtimer_init(timer, clockid, mode);
484}
485
486static inline void debug_init_on_stack(struct hrtimer *timer, clockid_t clockid,
487 enum hrtimer_mode mode)
488{
489 debug_hrtimer_init_on_stack(timer);
490 trace_hrtimer_init(timer, clockid, mode);
491}
492
493static inline void debug_activate(struct hrtimer *timer,
494 enum hrtimer_mode mode)
495{
496 debug_hrtimer_activate(timer, mode);
497 trace_hrtimer_start(timer, mode);
498}
499
500static inline void debug_deactivate(struct hrtimer *timer)
501{
502 debug_hrtimer_deactivate(timer);
503 trace_hrtimer_cancel(timer);
504}
505
506static struct hrtimer_clock_base *
507__next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
508{
509 unsigned int idx;
510
511 if (!*active)
512 return NULL;
513
514 idx = __ffs(*active);
515 *active &= ~(1U << idx);
516
517 return &cpu_base->clock_base[idx];
518}
519
520#define for_each_active_base(base, cpu_base, active) \
521 while ((base = __next_base((cpu_base), &(active))))
522
523static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
524 const struct hrtimer *exclude,
525 unsigned int active,
526 ktime_t expires_next)
527{
528 struct hrtimer_clock_base *base;
529 ktime_t expires;
530
531 for_each_active_base(base, cpu_base, active) {
532 struct timerqueue_node *next;
533 struct hrtimer *timer;
534
535 next = timerqueue_getnext(&base->active);
536 timer = container_of(next, struct hrtimer, node);
537 if (timer == exclude) {
538 /* Get to the next timer in the queue. */
539 next = timerqueue_iterate_next(next);
540 if (!next)
541 continue;
542
543 timer = container_of(next, struct hrtimer, node);
544 }
545 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
546 if (expires < expires_next) {
547 expires_next = expires;
548
549 /* Skip cpu_base update if a timer is being excluded. */
550 if (exclude)
551 continue;
552
553 if (timer->is_soft)
554 cpu_base->softirq_next_timer = timer;
555 else
556 cpu_base->next_timer = timer;
557 }
558 }
559 /*
560 * clock_was_set() might have changed base->offset of any of
561 * the clock bases so the result might be negative. Fix it up
562 * to prevent a false positive in clockevents_program_event().
563 */
564 if (expires_next < 0)
565 expires_next = 0;
566 return expires_next;
567}
568
569/*
570 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
571 * but does not set cpu_base::*expires_next, that is done by
572 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
573 * cpu_base::*expires_next right away, reprogramming logic would no longer
574 * work.
575 *
576 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
577 * those timers will get run whenever the softirq gets handled, at the end of
578 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
579 *
580 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
581 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
582 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
583 *
584 * @active_mask must be one of:
585 * - HRTIMER_ACTIVE_ALL,
586 * - HRTIMER_ACTIVE_SOFT, or
587 * - HRTIMER_ACTIVE_HARD.
588 */
589static ktime_t
590__hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
591{
592 unsigned int active;
593 struct hrtimer *next_timer = NULL;
594 ktime_t expires_next = KTIME_MAX;
595
596 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
597 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
598 cpu_base->softirq_next_timer = NULL;
599 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
600 active, KTIME_MAX);
601
602 next_timer = cpu_base->softirq_next_timer;
603 }
604
605 if (active_mask & HRTIMER_ACTIVE_HARD) {
606 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
607 cpu_base->next_timer = next_timer;
608 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
609 expires_next);
610 }
611
612 return expires_next;
613}
614
615static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
616{
617 ktime_t expires_next, soft = KTIME_MAX;
618
619 /*
620 * If the soft interrupt has already been activated, ignore the
621 * soft bases. They will be handled in the already raised soft
622 * interrupt.
623 */
624 if (!cpu_base->softirq_activated) {
625 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
626 /*
627 * Update the soft expiry time. clock_settime() might have
628 * affected it.
629 */
630 cpu_base->softirq_expires_next = soft;
631 }
632
633 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
634 /*
635 * If a softirq timer is expiring first, update cpu_base->next_timer
636 * and program the hardware with the soft expiry time.
637 */
638 if (expires_next > soft) {
639 cpu_base->next_timer = cpu_base->softirq_next_timer;
640 expires_next = soft;
641 }
642
643 return expires_next;
644}
645
646static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
647{
648 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
649 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
650 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
651
652 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
653 offs_real, offs_boot, offs_tai);
654
655 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
656 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
657 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
658
659 return now;
660}
661
662/*
663 * Is the high resolution mode active ?
664 */
665static inline int hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
666{
667 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
668 cpu_base->hres_active : 0;
669}
670
671static void __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base,
672 struct hrtimer *next_timer,
673 ktime_t expires_next)
674{
675 cpu_base->expires_next = expires_next;
676
677 /*
678 * If hres is not active, hardware does not have to be
679 * reprogrammed yet.
680 *
681 * If a hang was detected in the last timer interrupt then we
682 * leave the hang delay active in the hardware. We want the
683 * system to make progress. That also prevents the following
684 * scenario:
685 * T1 expires 50ms from now
686 * T2 expires 5s from now
687 *
688 * T1 is removed, so this code is called and would reprogram
689 * the hardware to 5s from now. Any hrtimer_start after that
690 * will not reprogram the hardware due to hang_detected being
691 * set. So we'd effectively block all timers until the T2 event
692 * fires.
693 */
694 if (!hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
695 return;
696
697 tick_program_event(expires_next, 1);
698}
699
700/*
701 * Reprogram the event source with checking both queues for the
702 * next event
703 * Called with interrupts disabled and base->lock held
704 */
705static void
706hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
707{
708 ktime_t expires_next;
709
710 expires_next = hrtimer_update_next_event(cpu_base);
711
712 if (skip_equal && expires_next == cpu_base->expires_next)
713 return;
714
715 __hrtimer_reprogram(cpu_base, cpu_base->next_timer, expires_next);
716}
717
718/* High resolution timer related functions */
719#ifdef CONFIG_HIGH_RES_TIMERS
720
721/*
722 * High resolution timer enabled ?
723 */
724static bool hrtimer_hres_enabled __read_mostly = true;
725unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
726EXPORT_SYMBOL_GPL(hrtimer_resolution);
727
728/*
729 * Enable / Disable high resolution mode
730 */
731static int __init setup_hrtimer_hres(char *str)
732{
733 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
734}
735
736__setup("highres=", setup_hrtimer_hres);
737
738/*
739 * hrtimer_high_res_enabled - query, if the highres mode is enabled
740 */
741static inline int hrtimer_is_hres_enabled(void)
742{
743 return hrtimer_hres_enabled;
744}
745
746/*
747 * Switch to high resolution mode
748 */
749static void hrtimer_switch_to_hres(void)
750{
751 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
752
753 if (tick_init_highres()) {
754 pr_warn("Could not switch to high resolution mode on CPU %u\n",
755 base->cpu);
756 return;
757 }
758 base->hres_active = 1;
759 hrtimer_resolution = HIGH_RES_NSEC;
760
761 tick_setup_sched_timer(true);
762 /* "Retrigger" the interrupt to get things going */
763 retrigger_next_event(NULL);
764}
765
766#else
767
768static inline int hrtimer_is_hres_enabled(void) { return 0; }
769static inline void hrtimer_switch_to_hres(void) { }
770
771#endif /* CONFIG_HIGH_RES_TIMERS */
772/*
773 * Retrigger next event is called after clock was set with interrupts
774 * disabled through an SMP function call or directly from low level
775 * resume code.
776 *
777 * This is only invoked when:
778 * - CONFIG_HIGH_RES_TIMERS is enabled.
779 * - CONFIG_NOHZ_COMMON is enabled
780 *
781 * For the other cases this function is empty and because the call sites
782 * are optimized out it vanishes as well, i.e. no need for lots of
783 * #ifdeffery.
784 */
785static void retrigger_next_event(void *arg)
786{
787 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
788
789 /*
790 * When high resolution mode or nohz is active, then the offsets of
791 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
792 * next tick will take care of that.
793 *
794 * If high resolution mode is active then the next expiring timer
795 * must be reevaluated and the clock event device reprogrammed if
796 * necessary.
797 *
798 * In the NOHZ case the update of the offset and the reevaluation
799 * of the next expiring timer is enough. The return from the SMP
800 * function call will take care of the reprogramming in case the
801 * CPU was in a NOHZ idle sleep.
802 */
803 if (!hrtimer_hres_active(base) && !tick_nohz_active)
804 return;
805
806 raw_spin_lock(&base->lock);
807 hrtimer_update_base(base);
808 if (hrtimer_hres_active(base))
809 hrtimer_force_reprogram(base, 0);
810 else
811 hrtimer_update_next_event(base);
812 raw_spin_unlock(&base->lock);
813}
814
815/*
816 * When a timer is enqueued and expires earlier than the already enqueued
817 * timers, we have to check, whether it expires earlier than the timer for
818 * which the clock event device was armed.
819 *
820 * Called with interrupts disabled and base->cpu_base.lock held
821 */
822static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
823{
824 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
825 struct hrtimer_clock_base *base = timer->base;
826 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
827
828 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
829
830 /*
831 * CLOCK_REALTIME timer might be requested with an absolute
832 * expiry time which is less than base->offset. Set it to 0.
833 */
834 if (expires < 0)
835 expires = 0;
836
837 if (timer->is_soft) {
838 /*
839 * soft hrtimer could be started on a remote CPU. In this
840 * case softirq_expires_next needs to be updated on the
841 * remote CPU. The soft hrtimer will not expire before the
842 * first hard hrtimer on the remote CPU -
843 * hrtimer_check_target() prevents this case.
844 */
845 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
846
847 if (timer_cpu_base->softirq_activated)
848 return;
849
850 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
851 return;
852
853 timer_cpu_base->softirq_next_timer = timer;
854 timer_cpu_base->softirq_expires_next = expires;
855
856 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
857 !reprogram)
858 return;
859 }
860
861 /*
862 * If the timer is not on the current cpu, we cannot reprogram
863 * the other cpus clock event device.
864 */
865 if (base->cpu_base != cpu_base)
866 return;
867
868 if (expires >= cpu_base->expires_next)
869 return;
870
871 /*
872 * If the hrtimer interrupt is running, then it will reevaluate the
873 * clock bases and reprogram the clock event device.
874 */
875 if (cpu_base->in_hrtirq)
876 return;
877
878 cpu_base->next_timer = timer;
879
880 __hrtimer_reprogram(cpu_base, timer, expires);
881}
882
883static bool update_needs_ipi(struct hrtimer_cpu_base *cpu_base,
884 unsigned int active)
885{
886 struct hrtimer_clock_base *base;
887 unsigned int seq;
888 ktime_t expires;
889
890 /*
891 * Update the base offsets unconditionally so the following
892 * checks whether the SMP function call is required works.
893 *
894 * The update is safe even when the remote CPU is in the hrtimer
895 * interrupt or the hrtimer soft interrupt and expiring affected
896 * bases. Either it will see the update before handling a base or
897 * it will see it when it finishes the processing and reevaluates
898 * the next expiring timer.
899 */
900 seq = cpu_base->clock_was_set_seq;
901 hrtimer_update_base(cpu_base);
902
903 /*
904 * If the sequence did not change over the update then the
905 * remote CPU already handled it.
906 */
907 if (seq == cpu_base->clock_was_set_seq)
908 return false;
909
910 /*
911 * If the remote CPU is currently handling an hrtimer interrupt, it
912 * will reevaluate the first expiring timer of all clock bases
913 * before reprogramming. Nothing to do here.
914 */
915 if (cpu_base->in_hrtirq)
916 return false;
917
918 /*
919 * Walk the affected clock bases and check whether the first expiring
920 * timer in a clock base is moving ahead of the first expiring timer of
921 * @cpu_base. If so, the IPI must be invoked because per CPU clock
922 * event devices cannot be remotely reprogrammed.
923 */
924 active &= cpu_base->active_bases;
925
926 for_each_active_base(base, cpu_base, active) {
927 struct timerqueue_node *next;
928
929 next = timerqueue_getnext(&base->active);
930 expires = ktime_sub(next->expires, base->offset);
931 if (expires < cpu_base->expires_next)
932 return true;
933
934 /* Extra check for softirq clock bases */
935 if (base->clockid < HRTIMER_BASE_MONOTONIC_SOFT)
936 continue;
937 if (cpu_base->softirq_activated)
938 continue;
939 if (expires < cpu_base->softirq_expires_next)
940 return true;
941 }
942 return false;
943}
944
945/*
946 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
947 * CLOCK_BOOTTIME (for late sleep time injection).
948 *
949 * This requires to update the offsets for these clocks
950 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
951 * also requires to eventually reprogram the per CPU clock event devices
952 * when the change moves an affected timer ahead of the first expiring
953 * timer on that CPU. Obviously remote per CPU clock event devices cannot
954 * be reprogrammed. The other reason why an IPI has to be sent is when the
955 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
956 * in the tick, which obviously might be stopped, so this has to bring out
957 * the remote CPU which might sleep in idle to get this sorted.
958 */
959void clock_was_set(unsigned int bases)
960{
961 struct hrtimer_cpu_base *cpu_base = raw_cpu_ptr(&hrtimer_bases);
962 cpumask_var_t mask;
963 int cpu;
964
965 if (!hrtimer_hres_active(cpu_base) && !tick_nohz_active)
966 goto out_timerfd;
967
968 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
969 on_each_cpu(retrigger_next_event, NULL, 1);
970 goto out_timerfd;
971 }
972
973 /* Avoid interrupting CPUs if possible */
974 cpus_read_lock();
975 for_each_online_cpu(cpu) {
976 unsigned long flags;
977
978 cpu_base = &per_cpu(hrtimer_bases, cpu);
979 raw_spin_lock_irqsave(&cpu_base->lock, flags);
980
981 if (update_needs_ipi(cpu_base, bases))
982 cpumask_set_cpu(cpu, mask);
983
984 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
985 }
986
987 preempt_disable();
988 smp_call_function_many(mask, retrigger_next_event, NULL, 1);
989 preempt_enable();
990 cpus_read_unlock();
991 free_cpumask_var(mask);
992
993out_timerfd:
994 timerfd_clock_was_set();
995}
996
997static void clock_was_set_work(struct work_struct *work)
998{
999 clock_was_set(CLOCK_SET_WALL);
1000}
1001
1002static DECLARE_WORK(hrtimer_work, clock_was_set_work);
1003
1004/*
1005 * Called from timekeeping code to reprogram the hrtimer interrupt device
1006 * on all cpus and to notify timerfd.
1007 */
1008void clock_was_set_delayed(void)
1009{
1010 schedule_work(&hrtimer_work);
1011}
1012
1013/*
1014 * Called during resume either directly from via timekeeping_resume()
1015 * or in the case of s2idle from tick_unfreeze() to ensure that the
1016 * hrtimers are up to date.
1017 */
1018void hrtimers_resume_local(void)
1019{
1020 lockdep_assert_irqs_disabled();
1021 /* Retrigger on the local CPU */
1022 retrigger_next_event(NULL);
1023}
1024
1025/*
1026 * Counterpart to lock_hrtimer_base above:
1027 */
1028static inline
1029void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
1030 __releases(&timer->base->cpu_base->lock)
1031{
1032 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
1033}
1034
1035/**
1036 * hrtimer_forward() - forward the timer expiry
1037 * @timer: hrtimer to forward
1038 * @now: forward past this time
1039 * @interval: the interval to forward
1040 *
1041 * Forward the timer expiry so it will expire in the future.
1042 *
1043 * .. note::
1044 * This only updates the timer expiry value and does not requeue the timer.
1045 *
1046 * There is also a variant of the function hrtimer_forward_now().
1047 *
1048 * Context: Can be safely called from the callback function of @timer. If called
1049 * from other contexts @timer must neither be enqueued nor running the
1050 * callback and the caller needs to take care of serialization.
1051 *
1052 * Return: The number of overruns are returned.
1053 */
1054u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
1055{
1056 u64 orun = 1;
1057 ktime_t delta;
1058
1059 delta = ktime_sub(now, hrtimer_get_expires(timer));
1060
1061 if (delta < 0)
1062 return 0;
1063
1064 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
1065 return 0;
1066
1067 if (interval < hrtimer_resolution)
1068 interval = hrtimer_resolution;
1069
1070 if (unlikely(delta >= interval)) {
1071 s64 incr = ktime_to_ns(interval);
1072
1073 orun = ktime_divns(delta, incr);
1074 hrtimer_add_expires_ns(timer, incr * orun);
1075 if (hrtimer_get_expires_tv64(timer) > now)
1076 return orun;
1077 /*
1078 * This (and the ktime_add() below) is the
1079 * correction for exact:
1080 */
1081 orun++;
1082 }
1083 hrtimer_add_expires(timer, interval);
1084
1085 return orun;
1086}
1087EXPORT_SYMBOL_GPL(hrtimer_forward);
1088
1089/*
1090 * enqueue_hrtimer - internal function to (re)start a timer
1091 *
1092 * The timer is inserted in expiry order. Insertion into the
1093 * red black tree is O(log(n)). Must hold the base lock.
1094 *
1095 * Returns true when the new timer is the leftmost timer in the tree.
1096 */
1097static bool enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1098 enum hrtimer_mode mode)
1099{
1100 debug_activate(timer, mode);
1101 WARN_ON_ONCE(!base->cpu_base->online);
1102
1103 base->cpu_base->active_bases |= 1 << base->index;
1104
1105 /* Pairs with the lockless read in hrtimer_is_queued() */
1106 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1107
1108 return timerqueue_add(&base->active, &timer->node);
1109}
1110
1111/*
1112 * __remove_hrtimer - internal function to remove a timer
1113 *
1114 * Caller must hold the base lock.
1115 *
1116 * High resolution timer mode reprograms the clock event device when the
1117 * timer is the one which expires next. The caller can disable this by setting
1118 * reprogram to zero. This is useful, when the context does a reprogramming
1119 * anyway (e.g. timer interrupt)
1120 */
1121static void __remove_hrtimer(struct hrtimer *timer,
1122 struct hrtimer_clock_base *base,
1123 u8 newstate, int reprogram)
1124{
1125 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1126 u8 state = timer->state;
1127
1128 /* Pairs with the lockless read in hrtimer_is_queued() */
1129 WRITE_ONCE(timer->state, newstate);
1130 if (!(state & HRTIMER_STATE_ENQUEUED))
1131 return;
1132
1133 if (!timerqueue_del(&base->active, &timer->node))
1134 cpu_base->active_bases &= ~(1 << base->index);
1135
1136 /*
1137 * Note: If reprogram is false we do not update
1138 * cpu_base->next_timer. This happens when we remove the first
1139 * timer on a remote cpu. No harm as we never dereference
1140 * cpu_base->next_timer. So the worst thing what can happen is
1141 * an superfluous call to hrtimer_force_reprogram() on the
1142 * remote cpu later on if the same timer gets enqueued again.
1143 */
1144 if (reprogram && timer == cpu_base->next_timer)
1145 hrtimer_force_reprogram(cpu_base, 1);
1146}
1147
1148/*
1149 * remove hrtimer, called with base lock held
1150 */
1151static inline int
1152remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1153 bool restart, bool keep_local)
1154{
1155 u8 state = timer->state;
1156
1157 if (state & HRTIMER_STATE_ENQUEUED) {
1158 bool reprogram;
1159
1160 /*
1161 * Remove the timer and force reprogramming when high
1162 * resolution mode is active and the timer is on the current
1163 * CPU. If we remove a timer on another CPU, reprogramming is
1164 * skipped. The interrupt event on this CPU is fired and
1165 * reprogramming happens in the interrupt handler. This is a
1166 * rare case and less expensive than a smp call.
1167 */
1168 debug_deactivate(timer);
1169 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1170
1171 /*
1172 * If the timer is not restarted then reprogramming is
1173 * required if the timer is local. If it is local and about
1174 * to be restarted, avoid programming it twice (on removal
1175 * and a moment later when it's requeued).
1176 */
1177 if (!restart)
1178 state = HRTIMER_STATE_INACTIVE;
1179 else
1180 reprogram &= !keep_local;
1181
1182 __remove_hrtimer(timer, base, state, reprogram);
1183 return 1;
1184 }
1185 return 0;
1186}
1187
1188static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1189 const enum hrtimer_mode mode)
1190{
1191#ifdef CONFIG_TIME_LOW_RES
1192 /*
1193 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1194 * granular time values. For relative timers we add hrtimer_resolution
1195 * (i.e. one jiffy) to prevent short timeouts.
1196 */
1197 timer->is_rel = mode & HRTIMER_MODE_REL;
1198 if (timer->is_rel)
1199 tim = ktime_add_safe(tim, hrtimer_resolution);
1200#endif
1201 return tim;
1202}
1203
1204static void
1205hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1206{
1207 ktime_t expires;
1208
1209 /*
1210 * Find the next SOFT expiration.
1211 */
1212 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1213
1214 /*
1215 * reprogramming needs to be triggered, even if the next soft
1216 * hrtimer expires at the same time than the next hard
1217 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1218 */
1219 if (expires == KTIME_MAX)
1220 return;
1221
1222 /*
1223 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1224 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1225 */
1226 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1227}
1228
1229static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1230 u64 delta_ns, const enum hrtimer_mode mode,
1231 struct hrtimer_clock_base *base)
1232{
1233 struct hrtimer_cpu_base *this_cpu_base = this_cpu_ptr(&hrtimer_bases);
1234 struct hrtimer_clock_base *new_base;
1235 bool force_local, first;
1236
1237 /*
1238 * If the timer is on the local cpu base and is the first expiring
1239 * timer then this might end up reprogramming the hardware twice
1240 * (on removal and on enqueue). To avoid that by prevent the
1241 * reprogram on removal, keep the timer local to the current CPU
1242 * and enforce reprogramming after it is queued no matter whether
1243 * it is the new first expiring timer again or not.
1244 */
1245 force_local = base->cpu_base == this_cpu_base;
1246 force_local &= base->cpu_base->next_timer == timer;
1247
1248 /*
1249 * Don't force local queuing if this enqueue happens on a unplugged
1250 * CPU after hrtimer_cpu_dying() has been invoked.
1251 */
1252 force_local &= this_cpu_base->online;
1253
1254 /*
1255 * Remove an active timer from the queue. In case it is not queued
1256 * on the current CPU, make sure that remove_hrtimer() updates the
1257 * remote data correctly.
1258 *
1259 * If it's on the current CPU and the first expiring timer, then
1260 * skip reprogramming, keep the timer local and enforce
1261 * reprogramming later if it was the first expiring timer. This
1262 * avoids programming the underlying clock event twice (once at
1263 * removal and once after enqueue).
1264 */
1265 remove_hrtimer(timer, base, true, force_local);
1266
1267 if (mode & HRTIMER_MODE_REL)
1268 tim = ktime_add_safe(tim, base->get_time());
1269
1270 tim = hrtimer_update_lowres(timer, tim, mode);
1271
1272 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1273
1274 /* Switch the timer base, if necessary: */
1275 if (!force_local) {
1276 new_base = switch_hrtimer_base(timer, base,
1277 mode & HRTIMER_MODE_PINNED);
1278 } else {
1279 new_base = base;
1280 }
1281
1282 first = enqueue_hrtimer(timer, new_base, mode);
1283 if (!force_local) {
1284 /*
1285 * If the current CPU base is online, then the timer is
1286 * never queued on a remote CPU if it would be the first
1287 * expiring timer there.
1288 */
1289 if (hrtimer_base_is_online(this_cpu_base))
1290 return first;
1291
1292 /*
1293 * Timer was enqueued remote because the current base is
1294 * already offline. If the timer is the first to expire,
1295 * kick the remote CPU to reprogram the clock event.
1296 */
1297 if (first) {
1298 struct hrtimer_cpu_base *new_cpu_base = new_base->cpu_base;
1299
1300 smp_call_function_single_async(new_cpu_base->cpu, &new_cpu_base->csd);
1301 }
1302 return 0;
1303 }
1304
1305 /*
1306 * Timer was forced to stay on the current CPU to avoid
1307 * reprogramming on removal and enqueue. Force reprogram the
1308 * hardware by evaluating the new first expiring timer.
1309 */
1310 hrtimer_force_reprogram(new_base->cpu_base, 1);
1311 return 0;
1312}
1313
1314/**
1315 * hrtimer_start_range_ns - (re)start an hrtimer
1316 * @timer: the timer to be added
1317 * @tim: expiry time
1318 * @delta_ns: "slack" range for the timer
1319 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1320 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1321 * softirq based mode is considered for debug purpose only!
1322 */
1323void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1324 u64 delta_ns, const enum hrtimer_mode mode)
1325{
1326 struct hrtimer_clock_base *base;
1327 unsigned long flags;
1328
1329 if (WARN_ON_ONCE(!timer->function))
1330 return;
1331 /*
1332 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1333 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1334 * expiry mode because unmarked timers are moved to softirq expiry.
1335 */
1336 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1337 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1338 else
1339 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1340
1341 base = lock_hrtimer_base(timer, &flags);
1342
1343 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1344 hrtimer_reprogram(timer, true);
1345
1346 unlock_hrtimer_base(timer, &flags);
1347}
1348EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1349
1350/**
1351 * hrtimer_try_to_cancel - try to deactivate a timer
1352 * @timer: hrtimer to stop
1353 *
1354 * Returns:
1355 *
1356 * * 0 when the timer was not active
1357 * * 1 when the timer was active
1358 * * -1 when the timer is currently executing the callback function and
1359 * cannot be stopped
1360 */
1361int hrtimer_try_to_cancel(struct hrtimer *timer)
1362{
1363 struct hrtimer_clock_base *base;
1364 unsigned long flags;
1365 int ret = -1;
1366
1367 /*
1368 * Check lockless first. If the timer is not active (neither
1369 * enqueued nor running the callback, nothing to do here. The
1370 * base lock does not serialize against a concurrent enqueue,
1371 * so we can avoid taking it.
1372 */
1373 if (!hrtimer_active(timer))
1374 return 0;
1375
1376 base = lock_hrtimer_base(timer, &flags);
1377
1378 if (!hrtimer_callback_running(timer))
1379 ret = remove_hrtimer(timer, base, false, false);
1380
1381 unlock_hrtimer_base(timer, &flags);
1382
1383 return ret;
1384
1385}
1386EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1387
1388#ifdef CONFIG_PREEMPT_RT
1389static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1390{
1391 spin_lock_init(&base->softirq_expiry_lock);
1392}
1393
1394static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1395 __acquires(&base->softirq_expiry_lock)
1396{
1397 spin_lock(&base->softirq_expiry_lock);
1398}
1399
1400static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1401 __releases(&base->softirq_expiry_lock)
1402{
1403 spin_unlock(&base->softirq_expiry_lock);
1404}
1405
1406/*
1407 * The counterpart to hrtimer_cancel_wait_running().
1408 *
1409 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1410 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1411 * allows the waiter to acquire the lock and make progress.
1412 */
1413static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1414 unsigned long flags)
1415{
1416 if (atomic_read(&cpu_base->timer_waiters)) {
1417 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1418 spin_unlock(&cpu_base->softirq_expiry_lock);
1419 spin_lock(&cpu_base->softirq_expiry_lock);
1420 raw_spin_lock_irq(&cpu_base->lock);
1421 }
1422}
1423
1424#ifdef CONFIG_SMP
1425static __always_inline bool is_migration_base(struct hrtimer_clock_base *base)
1426{
1427 return base == &migration_base;
1428}
1429#else
1430static __always_inline bool is_migration_base(struct hrtimer_clock_base *base)
1431{
1432 return false;
1433}
1434#endif
1435
1436/*
1437 * This function is called on PREEMPT_RT kernels when the fast path
1438 * deletion of a timer failed because the timer callback function was
1439 * running.
1440 *
1441 * This prevents priority inversion: if the soft irq thread is preempted
1442 * in the middle of a timer callback, then calling del_timer_sync() can
1443 * lead to two issues:
1444 *
1445 * - If the caller is on a remote CPU then it has to spin wait for the timer
1446 * handler to complete. This can result in unbound priority inversion.
1447 *
1448 * - If the caller originates from the task which preempted the timer
1449 * handler on the same CPU, then spin waiting for the timer handler to
1450 * complete is never going to end.
1451 */
1452void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1453{
1454 /* Lockless read. Prevent the compiler from reloading it below */
1455 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1456
1457 /*
1458 * Just relax if the timer expires in hard interrupt context or if
1459 * it is currently on the migration base.
1460 */
1461 if (!timer->is_soft || is_migration_base(base)) {
1462 cpu_relax();
1463 return;
1464 }
1465
1466 /*
1467 * Mark the base as contended and grab the expiry lock, which is
1468 * held by the softirq across the timer callback. Drop the lock
1469 * immediately so the softirq can expire the next timer. In theory
1470 * the timer could already be running again, but that's more than
1471 * unlikely and just causes another wait loop.
1472 */
1473 atomic_inc(&base->cpu_base->timer_waiters);
1474 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1475 atomic_dec(&base->cpu_base->timer_waiters);
1476 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1477}
1478#else
1479static inline void
1480hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1481static inline void
1482hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1483static inline void
1484hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
1485static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1486 unsigned long flags) { }
1487#endif
1488
1489/**
1490 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1491 * @timer: the timer to be cancelled
1492 *
1493 * Returns:
1494 * 0 when the timer was not active
1495 * 1 when the timer was active
1496 */
1497int hrtimer_cancel(struct hrtimer *timer)
1498{
1499 int ret;
1500
1501 do {
1502 ret = hrtimer_try_to_cancel(timer);
1503
1504 if (ret < 0)
1505 hrtimer_cancel_wait_running(timer);
1506 } while (ret < 0);
1507 return ret;
1508}
1509EXPORT_SYMBOL_GPL(hrtimer_cancel);
1510
1511/**
1512 * __hrtimer_get_remaining - get remaining time for the timer
1513 * @timer: the timer to read
1514 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1515 */
1516ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1517{
1518 unsigned long flags;
1519 ktime_t rem;
1520
1521 lock_hrtimer_base(timer, &flags);
1522 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1523 rem = hrtimer_expires_remaining_adjusted(timer);
1524 else
1525 rem = hrtimer_expires_remaining(timer);
1526 unlock_hrtimer_base(timer, &flags);
1527
1528 return rem;
1529}
1530EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1531
1532#ifdef CONFIG_NO_HZ_COMMON
1533/**
1534 * hrtimer_get_next_event - get the time until next expiry event
1535 *
1536 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1537 */
1538u64 hrtimer_get_next_event(void)
1539{
1540 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1541 u64 expires = KTIME_MAX;
1542 unsigned long flags;
1543
1544 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1545
1546 if (!hrtimer_hres_active(cpu_base))
1547 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1548
1549 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1550
1551 return expires;
1552}
1553
1554/**
1555 * hrtimer_next_event_without - time until next expiry event w/o one timer
1556 * @exclude: timer to exclude
1557 *
1558 * Returns the next expiry time over all timers except for the @exclude one or
1559 * KTIME_MAX if none of them is pending.
1560 */
1561u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1562{
1563 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1564 u64 expires = KTIME_MAX;
1565 unsigned long flags;
1566
1567 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1568
1569 if (hrtimer_hres_active(cpu_base)) {
1570 unsigned int active;
1571
1572 if (!cpu_base->softirq_activated) {
1573 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1574 expires = __hrtimer_next_event_base(cpu_base, exclude,
1575 active, KTIME_MAX);
1576 }
1577 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1578 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1579 expires);
1580 }
1581
1582 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1583
1584 return expires;
1585}
1586#endif
1587
1588static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1589{
1590 if (likely(clock_id < MAX_CLOCKS)) {
1591 int base = hrtimer_clock_to_base_table[clock_id];
1592
1593 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1594 return base;
1595 }
1596 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1597 return HRTIMER_BASE_MONOTONIC;
1598}
1599
1600static enum hrtimer_restart hrtimer_dummy_timeout(struct hrtimer *unused)
1601{
1602 return HRTIMER_NORESTART;
1603}
1604
1605static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1606 enum hrtimer_mode mode)
1607{
1608 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1609 struct hrtimer_cpu_base *cpu_base;
1610 int base;
1611
1612 /*
1613 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1614 * marked for hard interrupt expiry mode are moved into soft
1615 * interrupt context for latency reasons and because the callbacks
1616 * can invoke functions which might sleep on RT, e.g. spin_lock().
1617 */
1618 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1619 softtimer = true;
1620
1621 memset(timer, 0, sizeof(struct hrtimer));
1622
1623 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1624
1625 /*
1626 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1627 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1628 * ensure POSIX compliance.
1629 */
1630 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1631 clock_id = CLOCK_MONOTONIC;
1632
1633 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1634 base += hrtimer_clockid_to_base(clock_id);
1635 timer->is_soft = softtimer;
1636 timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1637 timer->base = &cpu_base->clock_base[base];
1638 timerqueue_init(&timer->node);
1639}
1640
1641static void __hrtimer_setup(struct hrtimer *timer,
1642 enum hrtimer_restart (*function)(struct hrtimer *),
1643 clockid_t clock_id, enum hrtimer_mode mode)
1644{
1645 __hrtimer_init(timer, clock_id, mode);
1646
1647 if (WARN_ON_ONCE(!function))
1648 timer->function = hrtimer_dummy_timeout;
1649 else
1650 timer->function = function;
1651}
1652
1653/**
1654 * hrtimer_init - initialize a timer to the given clock
1655 * @timer: the timer to be initialized
1656 * @clock_id: the clock to be used
1657 * @mode: The modes which are relevant for initialization:
1658 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1659 * HRTIMER_MODE_REL_SOFT
1660 *
1661 * The PINNED variants of the above can be handed in,
1662 * but the PINNED bit is ignored as pinning happens
1663 * when the hrtimer is started
1664 */
1665void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1666 enum hrtimer_mode mode)
1667{
1668 debug_init(timer, clock_id, mode);
1669 __hrtimer_init(timer, clock_id, mode);
1670}
1671EXPORT_SYMBOL_GPL(hrtimer_init);
1672
1673/**
1674 * hrtimer_setup - initialize a timer to the given clock
1675 * @timer: the timer to be initialized
1676 * @function: the callback function
1677 * @clock_id: the clock to be used
1678 * @mode: The modes which are relevant for initialization:
1679 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1680 * HRTIMER_MODE_REL_SOFT
1681 *
1682 * The PINNED variants of the above can be handed in,
1683 * but the PINNED bit is ignored as pinning happens
1684 * when the hrtimer is started
1685 */
1686void hrtimer_setup(struct hrtimer *timer, enum hrtimer_restart (*function)(struct hrtimer *),
1687 clockid_t clock_id, enum hrtimer_mode mode)
1688{
1689 debug_init(timer, clock_id, mode);
1690 __hrtimer_setup(timer, function, clock_id, mode);
1691}
1692EXPORT_SYMBOL_GPL(hrtimer_setup);
1693
1694/**
1695 * hrtimer_setup_on_stack - initialize a timer on stack memory
1696 * @timer: The timer to be initialized
1697 * @function: the callback function
1698 * @clock_id: The clock to be used
1699 * @mode: The timer mode
1700 *
1701 * Similar to hrtimer_setup(), except that this one must be used if struct hrtimer is in stack
1702 * memory.
1703 */
1704void hrtimer_setup_on_stack(struct hrtimer *timer,
1705 enum hrtimer_restart (*function)(struct hrtimer *),
1706 clockid_t clock_id, enum hrtimer_mode mode)
1707{
1708 debug_init_on_stack(timer, clock_id, mode);
1709 __hrtimer_setup(timer, function, clock_id, mode);
1710}
1711EXPORT_SYMBOL_GPL(hrtimer_setup_on_stack);
1712
1713/*
1714 * A timer is active, when it is enqueued into the rbtree or the
1715 * callback function is running or it's in the state of being migrated
1716 * to another cpu.
1717 *
1718 * It is important for this function to not return a false negative.
1719 */
1720bool hrtimer_active(const struct hrtimer *timer)
1721{
1722 struct hrtimer_clock_base *base;
1723 unsigned int seq;
1724
1725 do {
1726 base = READ_ONCE(timer->base);
1727 seq = raw_read_seqcount_begin(&base->seq);
1728
1729 if (timer->state != HRTIMER_STATE_INACTIVE ||
1730 base->running == timer)
1731 return true;
1732
1733 } while (read_seqcount_retry(&base->seq, seq) ||
1734 base != READ_ONCE(timer->base));
1735
1736 return false;
1737}
1738EXPORT_SYMBOL_GPL(hrtimer_active);
1739
1740/*
1741 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1742 * distinct sections:
1743 *
1744 * - queued: the timer is queued
1745 * - callback: the timer is being ran
1746 * - post: the timer is inactive or (re)queued
1747 *
1748 * On the read side we ensure we observe timer->state and cpu_base->running
1749 * from the same section, if anything changed while we looked at it, we retry.
1750 * This includes timer->base changing because sequence numbers alone are
1751 * insufficient for that.
1752 *
1753 * The sequence numbers are required because otherwise we could still observe
1754 * a false negative if the read side got smeared over multiple consecutive
1755 * __run_hrtimer() invocations.
1756 */
1757
1758static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1759 struct hrtimer_clock_base *base,
1760 struct hrtimer *timer, ktime_t *now,
1761 unsigned long flags) __must_hold(&cpu_base->lock)
1762{
1763 enum hrtimer_restart (*fn)(struct hrtimer *);
1764 bool expires_in_hardirq;
1765 int restart;
1766
1767 lockdep_assert_held(&cpu_base->lock);
1768
1769 debug_deactivate(timer);
1770 base->running = timer;
1771
1772 /*
1773 * Separate the ->running assignment from the ->state assignment.
1774 *
1775 * As with a regular write barrier, this ensures the read side in
1776 * hrtimer_active() cannot observe base->running == NULL &&
1777 * timer->state == INACTIVE.
1778 */
1779 raw_write_seqcount_barrier(&base->seq);
1780
1781 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1782 fn = timer->function;
1783
1784 /*
1785 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1786 * timer is restarted with a period then it becomes an absolute
1787 * timer. If its not restarted it does not matter.
1788 */
1789 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1790 timer->is_rel = false;
1791
1792 /*
1793 * The timer is marked as running in the CPU base, so it is
1794 * protected against migration to a different CPU even if the lock
1795 * is dropped.
1796 */
1797 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1798 trace_hrtimer_expire_entry(timer, now);
1799 expires_in_hardirq = lockdep_hrtimer_enter(timer);
1800
1801 restart = fn(timer);
1802
1803 lockdep_hrtimer_exit(expires_in_hardirq);
1804 trace_hrtimer_expire_exit(timer);
1805 raw_spin_lock_irq(&cpu_base->lock);
1806
1807 /*
1808 * Note: We clear the running state after enqueue_hrtimer and
1809 * we do not reprogram the event hardware. Happens either in
1810 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1811 *
1812 * Note: Because we dropped the cpu_base->lock above,
1813 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1814 * for us already.
1815 */
1816 if (restart != HRTIMER_NORESTART &&
1817 !(timer->state & HRTIMER_STATE_ENQUEUED))
1818 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1819
1820 /*
1821 * Separate the ->running assignment from the ->state assignment.
1822 *
1823 * As with a regular write barrier, this ensures the read side in
1824 * hrtimer_active() cannot observe base->running.timer == NULL &&
1825 * timer->state == INACTIVE.
1826 */
1827 raw_write_seqcount_barrier(&base->seq);
1828
1829 WARN_ON_ONCE(base->running != timer);
1830 base->running = NULL;
1831}
1832
1833static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1834 unsigned long flags, unsigned int active_mask)
1835{
1836 struct hrtimer_clock_base *base;
1837 unsigned int active = cpu_base->active_bases & active_mask;
1838
1839 for_each_active_base(base, cpu_base, active) {
1840 struct timerqueue_node *node;
1841 ktime_t basenow;
1842
1843 basenow = ktime_add(now, base->offset);
1844
1845 while ((node = timerqueue_getnext(&base->active))) {
1846 struct hrtimer *timer;
1847
1848 timer = container_of(node, struct hrtimer, node);
1849
1850 /*
1851 * The immediate goal for using the softexpires is
1852 * minimizing wakeups, not running timers at the
1853 * earliest interrupt after their soft expiration.
1854 * This allows us to avoid using a Priority Search
1855 * Tree, which can answer a stabbing query for
1856 * overlapping intervals and instead use the simple
1857 * BST we already have.
1858 * We don't add extra wakeups by delaying timers that
1859 * are right-of a not yet expired timer, because that
1860 * timer will have to trigger a wakeup anyway.
1861 */
1862 if (basenow < hrtimer_get_softexpires_tv64(timer))
1863 break;
1864
1865 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1866 if (active_mask == HRTIMER_ACTIVE_SOFT)
1867 hrtimer_sync_wait_running(cpu_base, flags);
1868 }
1869 }
1870}
1871
1872static __latent_entropy void hrtimer_run_softirq(void)
1873{
1874 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1875 unsigned long flags;
1876 ktime_t now;
1877
1878 hrtimer_cpu_base_lock_expiry(cpu_base);
1879 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1880
1881 now = hrtimer_update_base(cpu_base);
1882 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1883
1884 cpu_base->softirq_activated = 0;
1885 hrtimer_update_softirq_timer(cpu_base, true);
1886
1887 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1888 hrtimer_cpu_base_unlock_expiry(cpu_base);
1889}
1890
1891#ifdef CONFIG_HIGH_RES_TIMERS
1892
1893/*
1894 * High resolution timer interrupt
1895 * Called with interrupts disabled
1896 */
1897void hrtimer_interrupt(struct clock_event_device *dev)
1898{
1899 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1900 ktime_t expires_next, now, entry_time, delta;
1901 unsigned long flags;
1902 int retries = 0;
1903
1904 BUG_ON(!cpu_base->hres_active);
1905 cpu_base->nr_events++;
1906 dev->next_event = KTIME_MAX;
1907
1908 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1909 entry_time = now = hrtimer_update_base(cpu_base);
1910retry:
1911 cpu_base->in_hrtirq = 1;
1912 /*
1913 * We set expires_next to KTIME_MAX here with cpu_base->lock
1914 * held to prevent that a timer is enqueued in our queue via
1915 * the migration code. This does not affect enqueueing of
1916 * timers which run their callback and need to be requeued on
1917 * this CPU.
1918 */
1919 cpu_base->expires_next = KTIME_MAX;
1920
1921 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1922 cpu_base->softirq_expires_next = KTIME_MAX;
1923 cpu_base->softirq_activated = 1;
1924 raise_timer_softirq(HRTIMER_SOFTIRQ);
1925 }
1926
1927 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1928
1929 /* Reevaluate the clock bases for the [soft] next expiry */
1930 expires_next = hrtimer_update_next_event(cpu_base);
1931 /*
1932 * Store the new expiry value so the migration code can verify
1933 * against it.
1934 */
1935 cpu_base->expires_next = expires_next;
1936 cpu_base->in_hrtirq = 0;
1937 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1938
1939 /* Reprogramming necessary ? */
1940 if (!tick_program_event(expires_next, 0)) {
1941 cpu_base->hang_detected = 0;
1942 return;
1943 }
1944
1945 /*
1946 * The next timer was already expired due to:
1947 * - tracing
1948 * - long lasting callbacks
1949 * - being scheduled away when running in a VM
1950 *
1951 * We need to prevent that we loop forever in the hrtimer
1952 * interrupt routine. We give it 3 attempts to avoid
1953 * overreacting on some spurious event.
1954 *
1955 * Acquire base lock for updating the offsets and retrieving
1956 * the current time.
1957 */
1958 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1959 now = hrtimer_update_base(cpu_base);
1960 cpu_base->nr_retries++;
1961 if (++retries < 3)
1962 goto retry;
1963 /*
1964 * Give the system a chance to do something else than looping
1965 * here. We stored the entry time, so we know exactly how long
1966 * we spent here. We schedule the next event this amount of
1967 * time away.
1968 */
1969 cpu_base->nr_hangs++;
1970 cpu_base->hang_detected = 1;
1971 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1972
1973 delta = ktime_sub(now, entry_time);
1974 if ((unsigned int)delta > cpu_base->max_hang_time)
1975 cpu_base->max_hang_time = (unsigned int) delta;
1976 /*
1977 * Limit it to a sensible value as we enforce a longer
1978 * delay. Give the CPU at least 100ms to catch up.
1979 */
1980 if (delta > 100 * NSEC_PER_MSEC)
1981 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1982 else
1983 expires_next = ktime_add(now, delta);
1984 tick_program_event(expires_next, 1);
1985 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1986}
1987#endif /* !CONFIG_HIGH_RES_TIMERS */
1988
1989/*
1990 * Called from run_local_timers in hardirq context every jiffy
1991 */
1992void hrtimer_run_queues(void)
1993{
1994 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1995 unsigned long flags;
1996 ktime_t now;
1997
1998 if (hrtimer_hres_active(cpu_base))
1999 return;
2000
2001 /*
2002 * This _is_ ugly: We have to check periodically, whether we
2003 * can switch to highres and / or nohz mode. The clocksource
2004 * switch happens with xtime_lock held. Notification from
2005 * there only sets the check bit in the tick_oneshot code,
2006 * otherwise we might deadlock vs. xtime_lock.
2007 */
2008 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
2009 hrtimer_switch_to_hres();
2010 return;
2011 }
2012
2013 raw_spin_lock_irqsave(&cpu_base->lock, flags);
2014 now = hrtimer_update_base(cpu_base);
2015
2016 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
2017 cpu_base->softirq_expires_next = KTIME_MAX;
2018 cpu_base->softirq_activated = 1;
2019 raise_timer_softirq(HRTIMER_SOFTIRQ);
2020 }
2021
2022 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
2023 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
2024}
2025
2026/*
2027 * Sleep related functions:
2028 */
2029static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
2030{
2031 struct hrtimer_sleeper *t =
2032 container_of(timer, struct hrtimer_sleeper, timer);
2033 struct task_struct *task = t->task;
2034
2035 t->task = NULL;
2036 if (task)
2037 wake_up_process(task);
2038
2039 return HRTIMER_NORESTART;
2040}
2041
2042/**
2043 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
2044 * @sl: sleeper to be started
2045 * @mode: timer mode abs/rel
2046 *
2047 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
2048 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
2049 */
2050void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
2051 enum hrtimer_mode mode)
2052{
2053 /*
2054 * Make the enqueue delivery mode check work on RT. If the sleeper
2055 * was initialized for hard interrupt delivery, force the mode bit.
2056 * This is a special case for hrtimer_sleepers because
2057 * __hrtimer_init_sleeper() determines the delivery mode on RT so the
2058 * fiddling with this decision is avoided at the call sites.
2059 */
2060 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
2061 mode |= HRTIMER_MODE_HARD;
2062
2063 hrtimer_start_expires(&sl->timer, mode);
2064}
2065EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
2066
2067static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
2068 clockid_t clock_id, enum hrtimer_mode mode)
2069{
2070 /*
2071 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
2072 * marked for hard interrupt expiry mode are moved into soft
2073 * interrupt context either for latency reasons or because the
2074 * hrtimer callback takes regular spinlocks or invokes other
2075 * functions which are not suitable for hard interrupt context on
2076 * PREEMPT_RT.
2077 *
2078 * The hrtimer_sleeper callback is RT compatible in hard interrupt
2079 * context, but there is a latency concern: Untrusted userspace can
2080 * spawn many threads which arm timers for the same expiry time on
2081 * the same CPU. That causes a latency spike due to the wakeup of
2082 * a gazillion threads.
2083 *
2084 * OTOH, privileged real-time user space applications rely on the
2085 * low latency of hard interrupt wakeups. If the current task is in
2086 * a real-time scheduling class, mark the mode for hard interrupt
2087 * expiry.
2088 */
2089 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
2090 if (rt_or_dl_task_policy(current) && !(mode & HRTIMER_MODE_SOFT))
2091 mode |= HRTIMER_MODE_HARD;
2092 }
2093
2094 __hrtimer_init(&sl->timer, clock_id, mode);
2095 sl->timer.function = hrtimer_wakeup;
2096 sl->task = current;
2097}
2098
2099/**
2100 * hrtimer_setup_sleeper_on_stack - initialize a sleeper in stack memory
2101 * @sl: sleeper to be initialized
2102 * @clock_id: the clock to be used
2103 * @mode: timer mode abs/rel
2104 */
2105void hrtimer_setup_sleeper_on_stack(struct hrtimer_sleeper *sl,
2106 clockid_t clock_id, enum hrtimer_mode mode)
2107{
2108 debug_init_on_stack(&sl->timer, clock_id, mode);
2109 __hrtimer_init_sleeper(sl, clock_id, mode);
2110}
2111EXPORT_SYMBOL_GPL(hrtimer_setup_sleeper_on_stack);
2112
2113int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
2114{
2115 switch(restart->nanosleep.type) {
2116#ifdef CONFIG_COMPAT_32BIT_TIME
2117 case TT_COMPAT:
2118 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
2119 return -EFAULT;
2120 break;
2121#endif
2122 case TT_NATIVE:
2123 if (put_timespec64(ts, restart->nanosleep.rmtp))
2124 return -EFAULT;
2125 break;
2126 default:
2127 BUG();
2128 }
2129 return -ERESTART_RESTARTBLOCK;
2130}
2131
2132static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
2133{
2134 struct restart_block *restart;
2135
2136 do {
2137 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
2138 hrtimer_sleeper_start_expires(t, mode);
2139
2140 if (likely(t->task))
2141 schedule();
2142
2143 hrtimer_cancel(&t->timer);
2144 mode = HRTIMER_MODE_ABS;
2145
2146 } while (t->task && !signal_pending(current));
2147
2148 __set_current_state(TASK_RUNNING);
2149
2150 if (!t->task)
2151 return 0;
2152
2153 restart = &current->restart_block;
2154 if (restart->nanosleep.type != TT_NONE) {
2155 ktime_t rem = hrtimer_expires_remaining(&t->timer);
2156 struct timespec64 rmt;
2157
2158 if (rem <= 0)
2159 return 0;
2160 rmt = ktime_to_timespec64(rem);
2161
2162 return nanosleep_copyout(restart, &rmt);
2163 }
2164 return -ERESTART_RESTARTBLOCK;
2165}
2166
2167static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2168{
2169 struct hrtimer_sleeper t;
2170 int ret;
2171
2172 hrtimer_setup_sleeper_on_stack(&t, restart->nanosleep.clockid, HRTIMER_MODE_ABS);
2173 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
2174 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2175 destroy_hrtimer_on_stack(&t.timer);
2176 return ret;
2177}
2178
2179long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2180 const clockid_t clockid)
2181{
2182 struct restart_block *restart;
2183 struct hrtimer_sleeper t;
2184 int ret = 0;
2185
2186 hrtimer_setup_sleeper_on_stack(&t, clockid, mode);
2187 hrtimer_set_expires_range_ns(&t.timer, rqtp, current->timer_slack_ns);
2188 ret = do_nanosleep(&t, mode);
2189 if (ret != -ERESTART_RESTARTBLOCK)
2190 goto out;
2191
2192 /* Absolute timers do not update the rmtp value and restart: */
2193 if (mode == HRTIMER_MODE_ABS) {
2194 ret = -ERESTARTNOHAND;
2195 goto out;
2196 }
2197
2198 restart = &current->restart_block;
2199 restart->nanosleep.clockid = t.timer.base->clockid;
2200 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2201 set_restart_fn(restart, hrtimer_nanosleep_restart);
2202out:
2203 destroy_hrtimer_on_stack(&t.timer);
2204 return ret;
2205}
2206
2207#ifdef CONFIG_64BIT
2208
2209SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2210 struct __kernel_timespec __user *, rmtp)
2211{
2212 struct timespec64 tu;
2213
2214 if (get_timespec64(&tu, rqtp))
2215 return -EFAULT;
2216
2217 if (!timespec64_valid(&tu))
2218 return -EINVAL;
2219
2220 current->restart_block.fn = do_no_restart_syscall;
2221 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2222 current->restart_block.nanosleep.rmtp = rmtp;
2223 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2224 CLOCK_MONOTONIC);
2225}
2226
2227#endif
2228
2229#ifdef CONFIG_COMPAT_32BIT_TIME
2230
2231SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2232 struct old_timespec32 __user *, rmtp)
2233{
2234 struct timespec64 tu;
2235
2236 if (get_old_timespec32(&tu, rqtp))
2237 return -EFAULT;
2238
2239 if (!timespec64_valid(&tu))
2240 return -EINVAL;
2241
2242 current->restart_block.fn = do_no_restart_syscall;
2243 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2244 current->restart_block.nanosleep.compat_rmtp = rmtp;
2245 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2246 CLOCK_MONOTONIC);
2247}
2248#endif
2249
2250/*
2251 * Functions related to boot-time initialization:
2252 */
2253int hrtimers_prepare_cpu(unsigned int cpu)
2254{
2255 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2256 int i;
2257
2258 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2259 struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2260
2261 clock_b->cpu_base = cpu_base;
2262 seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2263 timerqueue_init_head(&clock_b->active);
2264 }
2265
2266 cpu_base->cpu = cpu;
2267 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2268 return 0;
2269}
2270
2271int hrtimers_cpu_starting(unsigned int cpu)
2272{
2273 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
2274
2275 /* Clear out any left over state from a CPU down operation */
2276 cpu_base->active_bases = 0;
2277 cpu_base->hres_active = 0;
2278 cpu_base->hang_detected = 0;
2279 cpu_base->next_timer = NULL;
2280 cpu_base->softirq_next_timer = NULL;
2281 cpu_base->expires_next = KTIME_MAX;
2282 cpu_base->softirq_expires_next = KTIME_MAX;
2283 cpu_base->online = 1;
2284 return 0;
2285}
2286
2287#ifdef CONFIG_HOTPLUG_CPU
2288
2289static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2290 struct hrtimer_clock_base *new_base)
2291{
2292 struct hrtimer *timer;
2293 struct timerqueue_node *node;
2294
2295 while ((node = timerqueue_getnext(&old_base->active))) {
2296 timer = container_of(node, struct hrtimer, node);
2297 BUG_ON(hrtimer_callback_running(timer));
2298 debug_deactivate(timer);
2299
2300 /*
2301 * Mark it as ENQUEUED not INACTIVE otherwise the
2302 * timer could be seen as !active and just vanish away
2303 * under us on another CPU
2304 */
2305 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2306 timer->base = new_base;
2307 /*
2308 * Enqueue the timers on the new cpu. This does not
2309 * reprogram the event device in case the timer
2310 * expires before the earliest on this CPU, but we run
2311 * hrtimer_interrupt after we migrated everything to
2312 * sort out already expired timers and reprogram the
2313 * event device.
2314 */
2315 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2316 }
2317}
2318
2319int hrtimers_cpu_dying(unsigned int dying_cpu)
2320{
2321 int i, ncpu = cpumask_any_and(cpu_active_mask, housekeeping_cpumask(HK_TYPE_TIMER));
2322 struct hrtimer_cpu_base *old_base, *new_base;
2323
2324 old_base = this_cpu_ptr(&hrtimer_bases);
2325 new_base = &per_cpu(hrtimer_bases, ncpu);
2326
2327 /*
2328 * The caller is globally serialized and nobody else
2329 * takes two locks at once, deadlock is not possible.
2330 */
2331 raw_spin_lock(&old_base->lock);
2332 raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2333
2334 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2335 migrate_hrtimer_list(&old_base->clock_base[i],
2336 &new_base->clock_base[i]);
2337 }
2338
2339 /*
2340 * The migration might have changed the first expiring softirq
2341 * timer on this CPU. Update it.
2342 */
2343 __hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
2344 /* Tell the other CPU to retrigger the next event */
2345 smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2346
2347 raw_spin_unlock(&new_base->lock);
2348 old_base->online = 0;
2349 raw_spin_unlock(&old_base->lock);
2350
2351 return 0;
2352}
2353
2354#endif /* CONFIG_HOTPLUG_CPU */
2355
2356void __init hrtimers_init(void)
2357{
2358 hrtimers_prepare_cpu(smp_processor_id());
2359 hrtimers_cpu_starting(smp_processor_id());
2360 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2361}