1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API, aka timer wheel,
10 * hrtimers provide finer resolution and accuracy depending on system
11 * configuration and capabilities.
12 *
13 * Started by: Thomas Gleixner and Ingo Molnar
14 *
15 * Credits:
16 * Based on the original timer wheel code
17 *
18 * Help, testing, suggestions, bugfixes, improvements were
19 * provided by:
20 *
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
22 * et. al.
23 */
24
25 #include <linux/cpu.h>
26 #include <linux/export.h>
27 #include <linux/percpu.h>
28 #include <linux/hrtimer.h>
29 #include <linux/notifier.h>
30 #include <linux/syscalls.h>
31 #include <linux/interrupt.h>
32 #include <linux/tick.h>
33 #include <linux/err.h>
34 #include <linux/debugobjects.h>
35 #include <linux/sched/signal.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/sched/rt.h>
38 #include <linux/sched/deadline.h>
39 #include <linux/sched/nohz.h>
40 #include <linux/sched/debug.h>
41 #include <linux/sched/isolation.h>
42 #include <linux/timer.h>
43 #include <linux/freezer.h>
44 #include <linux/compat.h>
45
46 #include <linux/uaccess.h>
47
48 #include <trace/events/timer.h>
49
50 #include "tick-internal.h"
51
52 /*
53 * Masks for selecting the soft and hard context timers from
54 * cpu_base->active
55 */
56 #define MASK_SHIFT (HRTIMER_BASE_MONOTONIC_SOFT)
57 #define HRTIMER_ACTIVE_HARD ((1U << MASK_SHIFT) - 1)
58 #define HRTIMER_ACTIVE_SOFT (HRTIMER_ACTIVE_HARD << MASK_SHIFT)
59 #define HRTIMER_ACTIVE_ALL (HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
60
61 /*
62 * The timer bases:
63 *
64 * There are more clockids than hrtimer bases. Thus, we index
65 * into the timer bases by the hrtimer_base_type enum. When trying
66 * to reach a base using a clockid, hrtimer_clockid_to_base()
67 * is used to convert from clockid to the proper hrtimer_base_type.
68 */
69 DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
70 {
71 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
72 .clock_base =
73 {
74 {
75 .index = HRTIMER_BASE_MONOTONIC,
76 .clockid = CLOCK_MONOTONIC,
77 .get_time = &ktime_get,
78 },
79 {
80 .index = HRTIMER_BASE_REALTIME,
81 .clockid = CLOCK_REALTIME,
82 .get_time = &ktime_get_real,
83 },
84 {
85 .index = HRTIMER_BASE_BOOTTIME,
86 .clockid = CLOCK_BOOTTIME,
87 .get_time = &ktime_get_boottime,
88 },
89 {
90 .index = HRTIMER_BASE_TAI,
91 .clockid = CLOCK_TAI,
92 .get_time = &ktime_get_clocktai,
93 },
94 {
95 .index = HRTIMER_BASE_MONOTONIC_SOFT,
96 .clockid = CLOCK_MONOTONIC,
97 .get_time = &ktime_get,
98 },
99 {
100 .index = HRTIMER_BASE_REALTIME_SOFT,
101 .clockid = CLOCK_REALTIME,
102 .get_time = &ktime_get_real,
103 },
104 {
105 .index = HRTIMER_BASE_BOOTTIME_SOFT,
106 .clockid = CLOCK_BOOTTIME,
107 .get_time = &ktime_get_boottime,
108 },
109 {
110 .index = HRTIMER_BASE_TAI_SOFT,
111 .clockid = CLOCK_TAI,
112 .get_time = &ktime_get_clocktai,
113 },
114 }
115 };
116
117 static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
118 /* Make sure we catch unsupported clockids */
119 [0 ... MAX_CLOCKS - 1] = HRTIMER_MAX_CLOCK_BASES,
120
121 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
122 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
123 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
124 [CLOCK_TAI] = HRTIMER_BASE_TAI,
125 };
126
127 /*
128 * Functions and macros which are different for UP/SMP systems are kept in a
129 * single place
130 */
131 #ifdef CONFIG_SMP
132
133 /*
134 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
135 * such that hrtimer_callback_running() can unconditionally dereference
136 * timer->base->cpu_base
137 */
138 static struct hrtimer_cpu_base migration_cpu_base = {
139 .clock_base = { {
140 .cpu_base = &migration_cpu_base,
141 .seq = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
142 &migration_cpu_base.lock),
143 }, },
144 };
145
146 #define migration_base migration_cpu_base.clock_base[0]
147
is_migration_base(struct hrtimer_clock_base * base)148 static inline bool is_migration_base(struct hrtimer_clock_base *base)
149 {
150 return base == &migration_base;
151 }
152
153 /*
154 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
155 * means that all timers which are tied to this base via timer->base are
156 * locked, and the base itself is locked too.
157 *
158 * So __run_timers/migrate_timers can safely modify all timers which could
159 * be found on the lists/queues.
160 *
161 * When the timer's base is locked, and the timer removed from list, it is
162 * possible to set timer->base = &migration_base and drop the lock: the timer
163 * remains locked.
164 */
165 static
lock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)166 struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
167 unsigned long *flags)
168 __acquires(&timer->base->lock)
169 {
170 struct hrtimer_clock_base *base;
171
172 for (;;) {
173 base = READ_ONCE(timer->base);
174 if (likely(base != &migration_base)) {
175 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
176 if (likely(base == timer->base))
177 return base;
178 /* The timer has migrated to another CPU: */
179 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
180 }
181 cpu_relax();
182 }
183 }
184
185 /*
186 * We do not migrate the timer when it is expiring before the next
187 * event on the target cpu. When high resolution is enabled, we cannot
188 * reprogram the target cpu hardware and we would cause it to fire
189 * late. To keep it simple, we handle the high resolution enabled and
190 * disabled case similar.
191 *
192 * Called with cpu_base->lock of target cpu held.
193 */
194 static int
hrtimer_check_target(struct hrtimer * timer,struct hrtimer_clock_base * new_base)195 hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
196 {
197 ktime_t expires;
198
199 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
200 return expires < new_base->cpu_base->expires_next;
201 }
202
203 static inline
get_target_base(struct hrtimer_cpu_base * base,int pinned)204 struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
205 int pinned)
206 {
207 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
208 if (static_branch_likely(&timers_migration_enabled) && !pinned)
209 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
210 #endif
211 return base;
212 }
213
214 /*
215 * We switch the timer base to a power-optimized selected CPU target,
216 * if:
217 * - NO_HZ_COMMON is enabled
218 * - timer migration is enabled
219 * - the timer callback is not running
220 * - the timer is not the first expiring timer on the new target
221 *
222 * If one of the above requirements is not fulfilled we move the timer
223 * to the current CPU or leave it on the previously assigned CPU if
224 * the timer callback is currently running.
225 */
226 static inline struct hrtimer_clock_base *
switch_hrtimer_base(struct hrtimer * timer,struct hrtimer_clock_base * base,int pinned)227 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
228 int pinned)
229 {
230 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
231 struct hrtimer_clock_base *new_base;
232 int basenum = base->index;
233
234 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
235 new_cpu_base = get_target_base(this_cpu_base, pinned);
236 again:
237 new_base = &new_cpu_base->clock_base[basenum];
238
239 if (base != new_base) {
240 /*
241 * We are trying to move timer to new_base.
242 * However we can't change timer's base while it is running,
243 * so we keep it on the same CPU. No hassle vs. reprogramming
244 * the event source in the high resolution case. The softirq
245 * code will take care of this when the timer function has
246 * completed. There is no conflict as we hold the lock until
247 * the timer is enqueued.
248 */
249 if (unlikely(hrtimer_callback_running(timer)))
250 return base;
251
252 /* See the comment in lock_hrtimer_base() */
253 WRITE_ONCE(timer->base, &migration_base);
254 raw_spin_unlock(&base->cpu_base->lock);
255 raw_spin_lock(&new_base->cpu_base->lock);
256
257 if (new_cpu_base != this_cpu_base &&
258 hrtimer_check_target(timer, new_base)) {
259 raw_spin_unlock(&new_base->cpu_base->lock);
260 raw_spin_lock(&base->cpu_base->lock);
261 new_cpu_base = this_cpu_base;
262 WRITE_ONCE(timer->base, base);
263 goto again;
264 }
265 WRITE_ONCE(timer->base, new_base);
266 } else {
267 if (new_cpu_base != this_cpu_base &&
268 hrtimer_check_target(timer, new_base)) {
269 new_cpu_base = this_cpu_base;
270 goto again;
271 }
272 }
273 return new_base;
274 }
275
276 #else /* CONFIG_SMP */
277
is_migration_base(struct hrtimer_clock_base * base)278 static inline bool is_migration_base(struct hrtimer_clock_base *base)
279 {
280 return false;
281 }
282
283 static inline struct hrtimer_clock_base *
lock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)284 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
285 __acquires(&timer->base->cpu_base->lock)
286 {
287 struct hrtimer_clock_base *base = timer->base;
288
289 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
290
291 return base;
292 }
293
294 # define switch_hrtimer_base(t, b, p) (b)
295
296 #endif /* !CONFIG_SMP */
297
298 /*
299 * Functions for the union type storage format of ktime_t which are
300 * too large for inlining:
301 */
302 #if BITS_PER_LONG < 64
303 /*
304 * Divide a ktime value by a nanosecond value
305 */
__ktime_divns(const ktime_t kt,s64 div)306 s64 __ktime_divns(const ktime_t kt, s64 div)
307 {
308 int sft = 0;
309 s64 dclc;
310 u64 tmp;
311
312 dclc = ktime_to_ns(kt);
313 tmp = dclc < 0 ? -dclc : dclc;
314
315 /* Make sure the divisor is less than 2^32: */
316 while (div >> 32) {
317 sft++;
318 div >>= 1;
319 }
320 tmp >>= sft;
321 do_div(tmp, (u32) div);
322 return dclc < 0 ? -tmp : tmp;
323 }
324 EXPORT_SYMBOL_GPL(__ktime_divns);
325 #endif /* BITS_PER_LONG >= 64 */
326
327 /*
328 * Add two ktime values and do a safety check for overflow:
329 */
ktime_add_safe(const ktime_t lhs,const ktime_t rhs)330 ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
331 {
332 ktime_t res = ktime_add_unsafe(lhs, rhs);
333
334 /*
335 * We use KTIME_SEC_MAX here, the maximum timeout which we can
336 * return to user space in a timespec:
337 */
338 if (res < 0 || res < lhs || res < rhs)
339 res = ktime_set(KTIME_SEC_MAX, 0);
340
341 return res;
342 }
343
344 EXPORT_SYMBOL_GPL(ktime_add_safe);
345
346 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
347
348 static const struct debug_obj_descr hrtimer_debug_descr;
349
hrtimer_debug_hint(void * addr)350 static void *hrtimer_debug_hint(void *addr)
351 {
352 return ((struct hrtimer *) addr)->function;
353 }
354
355 /*
356 * fixup_init is called when:
357 * - an active object is initialized
358 */
hrtimer_fixup_init(void * addr,enum debug_obj_state state)359 static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
360 {
361 struct hrtimer *timer = addr;
362
363 switch (state) {
364 case ODEBUG_STATE_ACTIVE:
365 hrtimer_cancel(timer);
366 debug_object_init(timer, &hrtimer_debug_descr);
367 return true;
368 default:
369 return false;
370 }
371 }
372
373 /*
374 * fixup_activate is called when:
375 * - an active object is activated
376 * - an unknown non-static object is activated
377 */
hrtimer_fixup_activate(void * addr,enum debug_obj_state state)378 static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
379 {
380 switch (state) {
381 case ODEBUG_STATE_ACTIVE:
382 WARN_ON(1);
383 fallthrough;
384 default:
385 return false;
386 }
387 }
388
389 /*
390 * fixup_free is called when:
391 * - an active object is freed
392 */
hrtimer_fixup_free(void * addr,enum debug_obj_state state)393 static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
394 {
395 struct hrtimer *timer = addr;
396
397 switch (state) {
398 case ODEBUG_STATE_ACTIVE:
399 hrtimer_cancel(timer);
400 debug_object_free(timer, &hrtimer_debug_descr);
401 return true;
402 default:
403 return false;
404 }
405 }
406
407 static const struct debug_obj_descr hrtimer_debug_descr = {
408 .name = "hrtimer",
409 .debug_hint = hrtimer_debug_hint,
410 .fixup_init = hrtimer_fixup_init,
411 .fixup_activate = hrtimer_fixup_activate,
412 .fixup_free = hrtimer_fixup_free,
413 };
414
debug_hrtimer_init(struct hrtimer * timer)415 static inline void debug_hrtimer_init(struct hrtimer *timer)
416 {
417 debug_object_init(timer, &hrtimer_debug_descr);
418 }
419
debug_hrtimer_activate(struct hrtimer * timer,enum hrtimer_mode mode)420 static inline void debug_hrtimer_activate(struct hrtimer *timer,
421 enum hrtimer_mode mode)
422 {
423 debug_object_activate(timer, &hrtimer_debug_descr);
424 }
425
debug_hrtimer_deactivate(struct hrtimer * timer)426 static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
427 {
428 debug_object_deactivate(timer, &hrtimer_debug_descr);
429 }
430
431 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
432 enum hrtimer_mode mode);
433
hrtimer_init_on_stack(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)434 void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
435 enum hrtimer_mode mode)
436 {
437 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
438 __hrtimer_init(timer, clock_id, mode);
439 }
440 EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
441
442 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
443 clockid_t clock_id, enum hrtimer_mode mode);
444
hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)445 void hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper *sl,
446 clockid_t clock_id, enum hrtimer_mode mode)
447 {
448 debug_object_init_on_stack(&sl->timer, &hrtimer_debug_descr);
449 __hrtimer_init_sleeper(sl, clock_id, mode);
450 }
451 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper_on_stack);
452
destroy_hrtimer_on_stack(struct hrtimer * timer)453 void destroy_hrtimer_on_stack(struct hrtimer *timer)
454 {
455 debug_object_free(timer, &hrtimer_debug_descr);
456 }
457 EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
458
459 #else
460
debug_hrtimer_init(struct hrtimer * timer)461 static inline void debug_hrtimer_init(struct hrtimer *timer) { }
debug_hrtimer_activate(struct hrtimer * timer,enum hrtimer_mode mode)462 static inline void debug_hrtimer_activate(struct hrtimer *timer,
463 enum hrtimer_mode mode) { }
debug_hrtimer_deactivate(struct hrtimer * timer)464 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
465 #endif
466
467 static inline void
debug_init(struct hrtimer * timer,clockid_t clockid,enum hrtimer_mode mode)468 debug_init(struct hrtimer *timer, clockid_t clockid,
469 enum hrtimer_mode mode)
470 {
471 debug_hrtimer_init(timer);
472 trace_hrtimer_init(timer, clockid, mode);
473 }
474
debug_activate(struct hrtimer * timer,enum hrtimer_mode mode)475 static inline void debug_activate(struct hrtimer *timer,
476 enum hrtimer_mode mode)
477 {
478 debug_hrtimer_activate(timer, mode);
479 trace_hrtimer_start(timer, mode);
480 }
481
debug_deactivate(struct hrtimer * timer)482 static inline void debug_deactivate(struct hrtimer *timer)
483 {
484 debug_hrtimer_deactivate(timer);
485 trace_hrtimer_cancel(timer);
486 }
487
488 static struct hrtimer_clock_base *
__next_base(struct hrtimer_cpu_base * cpu_base,unsigned int * active)489 __next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
490 {
491 unsigned int idx;
492
493 if (!*active)
494 return NULL;
495
496 idx = __ffs(*active);
497 *active &= ~(1U << idx);
498
499 return &cpu_base->clock_base[idx];
500 }
501
502 #define for_each_active_base(base, cpu_base, active) \
503 while ((base = __next_base((cpu_base), &(active))))
504
__hrtimer_next_event_base(struct hrtimer_cpu_base * cpu_base,const struct hrtimer * exclude,unsigned int active,ktime_t expires_next)505 static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
506 const struct hrtimer *exclude,
507 unsigned int active,
508 ktime_t expires_next)
509 {
510 struct hrtimer_clock_base *base;
511 ktime_t expires;
512
513 for_each_active_base(base, cpu_base, active) {
514 struct timerqueue_node *next;
515 struct hrtimer *timer;
516
517 next = timerqueue_getnext(&base->active);
518 timer = container_of(next, struct hrtimer, node);
519 if (timer == exclude) {
520 /* Get to the next timer in the queue. */
521 next = timerqueue_iterate_next(next);
522 if (!next)
523 continue;
524
525 timer = container_of(next, struct hrtimer, node);
526 }
527 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
528 if (expires < expires_next) {
529 expires_next = expires;
530
531 /* Skip cpu_base update if a timer is being excluded. */
532 if (exclude)
533 continue;
534
535 if (timer->is_soft)
536 cpu_base->softirq_next_timer = timer;
537 else
538 cpu_base->next_timer = timer;
539 }
540 }
541 /*
542 * clock_was_set() might have changed base->offset of any of
543 * the clock bases so the result might be negative. Fix it up
544 * to prevent a false positive in clockevents_program_event().
545 */
546 if (expires_next < 0)
547 expires_next = 0;
548 return expires_next;
549 }
550
551 /*
552 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
553 * but does not set cpu_base::*expires_next, that is done by
554 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
555 * cpu_base::*expires_next right away, reprogramming logic would no longer
556 * work.
557 *
558 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
559 * those timers will get run whenever the softirq gets handled, at the end of
560 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
561 *
562 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
563 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
564 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
565 *
566 * @active_mask must be one of:
567 * - HRTIMER_ACTIVE_ALL,
568 * - HRTIMER_ACTIVE_SOFT, or
569 * - HRTIMER_ACTIVE_HARD.
570 */
571 static ktime_t
__hrtimer_get_next_event(struct hrtimer_cpu_base * cpu_base,unsigned int active_mask)572 __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
573 {
574 unsigned int active;
575 struct hrtimer *next_timer = NULL;
576 ktime_t expires_next = KTIME_MAX;
577
578 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
579 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
580 cpu_base->softirq_next_timer = NULL;
581 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
582 active, KTIME_MAX);
583
584 next_timer = cpu_base->softirq_next_timer;
585 }
586
587 if (active_mask & HRTIMER_ACTIVE_HARD) {
588 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
589 cpu_base->next_timer = next_timer;
590 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
591 expires_next);
592 }
593
594 return expires_next;
595 }
596
hrtimer_update_next_event(struct hrtimer_cpu_base * cpu_base)597 static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
598 {
599 ktime_t expires_next, soft = KTIME_MAX;
600
601 /*
602 * If the soft interrupt has already been activated, ignore the
603 * soft bases. They will be handled in the already raised soft
604 * interrupt.
605 */
606 if (!cpu_base->softirq_activated) {
607 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
608 /*
609 * Update the soft expiry time. clock_settime() might have
610 * affected it.
611 */
612 cpu_base->softirq_expires_next = soft;
613 }
614
615 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
616 /*
617 * If a softirq timer is expiring first, update cpu_base->next_timer
618 * and program the hardware with the soft expiry time.
619 */
620 if (expires_next > soft) {
621 cpu_base->next_timer = cpu_base->softirq_next_timer;
622 expires_next = soft;
623 }
624
625 return expires_next;
626 }
627
hrtimer_update_base(struct hrtimer_cpu_base * base)628 static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
629 {
630 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
631 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
632 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
633
634 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
635 offs_real, offs_boot, offs_tai);
636
637 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
638 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
639 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
640
641 return now;
642 }
643
644 /*
645 * Is the high resolution mode active ?
646 */
hrtimer_hres_active(struct hrtimer_cpu_base * cpu_base)647 static inline int hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
648 {
649 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
650 cpu_base->hres_active : 0;
651 }
652
__hrtimer_reprogram(struct hrtimer_cpu_base * cpu_base,struct hrtimer * next_timer,ktime_t expires_next)653 static void __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base,
654 struct hrtimer *next_timer,
655 ktime_t expires_next)
656 {
657 cpu_base->expires_next = expires_next;
658
659 /*
660 * If hres is not active, hardware does not have to be
661 * reprogrammed yet.
662 *
663 * If a hang was detected in the last timer interrupt then we
664 * leave the hang delay active in the hardware. We want the
665 * system to make progress. That also prevents the following
666 * scenario:
667 * T1 expires 50ms from now
668 * T2 expires 5s from now
669 *
670 * T1 is removed, so this code is called and would reprogram
671 * the hardware to 5s from now. Any hrtimer_start after that
672 * will not reprogram the hardware due to hang_detected being
673 * set. So we'd effectively block all timers until the T2 event
674 * fires.
675 */
676 if (!hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
677 return;
678
679 tick_program_event(expires_next, 1);
680 }
681
682 /*
683 * Reprogram the event source with checking both queues for the
684 * next event
685 * Called with interrupts disabled and base->lock held
686 */
687 static void
hrtimer_force_reprogram(struct hrtimer_cpu_base * cpu_base,int skip_equal)688 hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
689 {
690 ktime_t expires_next;
691
692 expires_next = hrtimer_update_next_event(cpu_base);
693
694 if (skip_equal && expires_next == cpu_base->expires_next)
695 return;
696
697 __hrtimer_reprogram(cpu_base, cpu_base->next_timer, expires_next);
698 }
699
700 /* High resolution timer related functions */
701 #ifdef CONFIG_HIGH_RES_TIMERS
702
703 /*
704 * High resolution timer enabled ?
705 */
706 static bool hrtimer_hres_enabled __read_mostly = true;
707 unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
708 EXPORT_SYMBOL_GPL(hrtimer_resolution);
709
710 /*
711 * Enable / Disable high resolution mode
712 */
setup_hrtimer_hres(char * str)713 static int __init setup_hrtimer_hres(char *str)
714 {
715 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
716 }
717
718 __setup("highres=", setup_hrtimer_hres);
719
720 /*
721 * hrtimer_high_res_enabled - query, if the highres mode is enabled
722 */
hrtimer_is_hres_enabled(void)723 static inline int hrtimer_is_hres_enabled(void)
724 {
725 return hrtimer_hres_enabled;
726 }
727
728 static void retrigger_next_event(void *arg);
729
730 /*
731 * Switch to high resolution mode
732 */
hrtimer_switch_to_hres(void)733 static void hrtimer_switch_to_hres(void)
734 {
735 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
736
737 if (tick_init_highres()) {
738 pr_warn("Could not switch to high resolution mode on CPU %u\n",
739 base->cpu);
740 return;
741 }
742 base->hres_active = 1;
743 hrtimer_resolution = HIGH_RES_NSEC;
744
745 tick_setup_sched_timer(true);
746 /* "Retrigger" the interrupt to get things going */
747 retrigger_next_event(NULL);
748 }
749
750 #else
751
hrtimer_is_hres_enabled(void)752 static inline int hrtimer_is_hres_enabled(void) { return 0; }
hrtimer_switch_to_hres(void)753 static inline void hrtimer_switch_to_hres(void) { }
754
755 #endif /* CONFIG_HIGH_RES_TIMERS */
756 /*
757 * Retrigger next event is called after clock was set with interrupts
758 * disabled through an SMP function call or directly from low level
759 * resume code.
760 *
761 * This is only invoked when:
762 * - CONFIG_HIGH_RES_TIMERS is enabled.
763 * - CONFIG_NOHZ_COMMON is enabled
764 *
765 * For the other cases this function is empty and because the call sites
766 * are optimized out it vanishes as well, i.e. no need for lots of
767 * #ifdeffery.
768 */
retrigger_next_event(void * arg)769 static void retrigger_next_event(void *arg)
770 {
771 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
772
773 /*
774 * When high resolution mode or nohz is active, then the offsets of
775 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
776 * next tick will take care of that.
777 *
778 * If high resolution mode is active then the next expiring timer
779 * must be reevaluated and the clock event device reprogrammed if
780 * necessary.
781 *
782 * In the NOHZ case the update of the offset and the reevaluation
783 * of the next expiring timer is enough. The return from the SMP
784 * function call will take care of the reprogramming in case the
785 * CPU was in a NOHZ idle sleep.
786 */
787 if (!hrtimer_hres_active(base) && !tick_nohz_active)
788 return;
789
790 raw_spin_lock(&base->lock);
791 hrtimer_update_base(base);
792 if (hrtimer_hres_active(base))
793 hrtimer_force_reprogram(base, 0);
794 else
795 hrtimer_update_next_event(base);
796 raw_spin_unlock(&base->lock);
797 }
798
799 /*
800 * When a timer is enqueued and expires earlier than the already enqueued
801 * timers, we have to check, whether it expires earlier than the timer for
802 * which the clock event device was armed.
803 *
804 * Called with interrupts disabled and base->cpu_base.lock held
805 */
hrtimer_reprogram(struct hrtimer * timer,bool reprogram)806 static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
807 {
808 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
809 struct hrtimer_clock_base *base = timer->base;
810 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
811
812 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
813
814 /*
815 * CLOCK_REALTIME timer might be requested with an absolute
816 * expiry time which is less than base->offset. Set it to 0.
817 */
818 if (expires < 0)
819 expires = 0;
820
821 if (timer->is_soft) {
822 /*
823 * soft hrtimer could be started on a remote CPU. In this
824 * case softirq_expires_next needs to be updated on the
825 * remote CPU. The soft hrtimer will not expire before the
826 * first hard hrtimer on the remote CPU -
827 * hrtimer_check_target() prevents this case.
828 */
829 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
830
831 if (timer_cpu_base->softirq_activated)
832 return;
833
834 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
835 return;
836
837 timer_cpu_base->softirq_next_timer = timer;
838 timer_cpu_base->softirq_expires_next = expires;
839
840 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
841 !reprogram)
842 return;
843 }
844
845 /*
846 * If the timer is not on the current cpu, we cannot reprogram
847 * the other cpus clock event device.
848 */
849 if (base->cpu_base != cpu_base)
850 return;
851
852 if (expires >= cpu_base->expires_next)
853 return;
854
855 /*
856 * If the hrtimer interrupt is running, then it will reevaluate the
857 * clock bases and reprogram the clock event device.
858 */
859 if (cpu_base->in_hrtirq)
860 return;
861
862 cpu_base->next_timer = timer;
863
864 __hrtimer_reprogram(cpu_base, timer, expires);
865 }
866
update_needs_ipi(struct hrtimer_cpu_base * cpu_base,unsigned int active)867 static bool update_needs_ipi(struct hrtimer_cpu_base *cpu_base,
868 unsigned int active)
869 {
870 struct hrtimer_clock_base *base;
871 unsigned int seq;
872 ktime_t expires;
873
874 /*
875 * Update the base offsets unconditionally so the following
876 * checks whether the SMP function call is required works.
877 *
878 * The update is safe even when the remote CPU is in the hrtimer
879 * interrupt or the hrtimer soft interrupt and expiring affected
880 * bases. Either it will see the update before handling a base or
881 * it will see it when it finishes the processing and reevaluates
882 * the next expiring timer.
883 */
884 seq = cpu_base->clock_was_set_seq;
885 hrtimer_update_base(cpu_base);
886
887 /*
888 * If the sequence did not change over the update then the
889 * remote CPU already handled it.
890 */
891 if (seq == cpu_base->clock_was_set_seq)
892 return false;
893
894 /*
895 * If the remote CPU is currently handling an hrtimer interrupt, it
896 * will reevaluate the first expiring timer of all clock bases
897 * before reprogramming. Nothing to do here.
898 */
899 if (cpu_base->in_hrtirq)
900 return false;
901
902 /*
903 * Walk the affected clock bases and check whether the first expiring
904 * timer in a clock base is moving ahead of the first expiring timer of
905 * @cpu_base. If so, the IPI must be invoked because per CPU clock
906 * event devices cannot be remotely reprogrammed.
907 */
908 active &= cpu_base->active_bases;
909
910 for_each_active_base(base, cpu_base, active) {
911 struct timerqueue_node *next;
912
913 next = timerqueue_getnext(&base->active);
914 expires = ktime_sub(next->expires, base->offset);
915 if (expires < cpu_base->expires_next)
916 return true;
917
918 /* Extra check for softirq clock bases */
919 if (base->clockid < HRTIMER_BASE_MONOTONIC_SOFT)
920 continue;
921 if (cpu_base->softirq_activated)
922 continue;
923 if (expires < cpu_base->softirq_expires_next)
924 return true;
925 }
926 return false;
927 }
928
929 /*
930 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
931 * CLOCK_BOOTTIME (for late sleep time injection).
932 *
933 * This requires to update the offsets for these clocks
934 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
935 * also requires to eventually reprogram the per CPU clock event devices
936 * when the change moves an affected timer ahead of the first expiring
937 * timer on that CPU. Obviously remote per CPU clock event devices cannot
938 * be reprogrammed. The other reason why an IPI has to be sent is when the
939 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
940 * in the tick, which obviously might be stopped, so this has to bring out
941 * the remote CPU which might sleep in idle to get this sorted.
942 */
clock_was_set(unsigned int bases)943 void clock_was_set(unsigned int bases)
944 {
945 struct hrtimer_cpu_base *cpu_base = raw_cpu_ptr(&hrtimer_bases);
946 cpumask_var_t mask;
947 int cpu;
948
949 if (!hrtimer_hres_active(cpu_base) && !tick_nohz_active)
950 goto out_timerfd;
951
952 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
953 on_each_cpu(retrigger_next_event, NULL, 1);
954 goto out_timerfd;
955 }
956
957 /* Avoid interrupting CPUs if possible */
958 cpus_read_lock();
959 for_each_online_cpu(cpu) {
960 unsigned long flags;
961
962 cpu_base = &per_cpu(hrtimer_bases, cpu);
963 raw_spin_lock_irqsave(&cpu_base->lock, flags);
964
965 if (update_needs_ipi(cpu_base, bases))
966 cpumask_set_cpu(cpu, mask);
967
968 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
969 }
970
971 preempt_disable();
972 smp_call_function_many(mask, retrigger_next_event, NULL, 1);
973 preempt_enable();
974 cpus_read_unlock();
975 free_cpumask_var(mask);
976
977 out_timerfd:
978 timerfd_clock_was_set();
979 }
980
clock_was_set_work(struct work_struct * work)981 static void clock_was_set_work(struct work_struct *work)
982 {
983 clock_was_set(CLOCK_SET_WALL);
984 }
985
986 static DECLARE_WORK(hrtimer_work, clock_was_set_work);
987
988 /*
989 * Called from timekeeping code to reprogram the hrtimer interrupt device
990 * on all cpus and to notify timerfd.
991 */
clock_was_set_delayed(void)992 void clock_was_set_delayed(void)
993 {
994 schedule_work(&hrtimer_work);
995 }
996
997 /*
998 * Called during resume either directly from via timekeeping_resume()
999 * or in the case of s2idle from tick_unfreeze() to ensure that the
1000 * hrtimers are up to date.
1001 */
hrtimers_resume_local(void)1002 void hrtimers_resume_local(void)
1003 {
1004 lockdep_assert_irqs_disabled();
1005 /* Retrigger on the local CPU */
1006 retrigger_next_event(NULL);
1007 }
1008
1009 /*
1010 * Counterpart to lock_hrtimer_base above:
1011 */
1012 static inline
unlock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)1013 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
1014 __releases(&timer->base->cpu_base->lock)
1015 {
1016 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
1017 }
1018
1019 /**
1020 * hrtimer_forward() - forward the timer expiry
1021 * @timer: hrtimer to forward
1022 * @now: forward past this time
1023 * @interval: the interval to forward
1024 *
1025 * Forward the timer expiry so it will expire in the future.
1026 *
1027 * .. note::
1028 * This only updates the timer expiry value and does not requeue the timer.
1029 *
1030 * There is also a variant of the function hrtimer_forward_now().
1031 *
1032 * Context: Can be safely called from the callback function of @timer. If called
1033 * from other contexts @timer must neither be enqueued nor running the
1034 * callback and the caller needs to take care of serialization.
1035 *
1036 * Return: The number of overruns are returned.
1037 */
hrtimer_forward(struct hrtimer * timer,ktime_t now,ktime_t interval)1038 u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
1039 {
1040 u64 orun = 1;
1041 ktime_t delta;
1042
1043 delta = ktime_sub(now, hrtimer_get_expires(timer));
1044
1045 if (delta < 0)
1046 return 0;
1047
1048 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
1049 return 0;
1050
1051 if (interval < hrtimer_resolution)
1052 interval = hrtimer_resolution;
1053
1054 if (unlikely(delta >= interval)) {
1055 s64 incr = ktime_to_ns(interval);
1056
1057 orun = ktime_divns(delta, incr);
1058 hrtimer_add_expires_ns(timer, incr * orun);
1059 if (hrtimer_get_expires_tv64(timer) > now)
1060 return orun;
1061 /*
1062 * This (and the ktime_add() below) is the
1063 * correction for exact:
1064 */
1065 orun++;
1066 }
1067 hrtimer_add_expires(timer, interval);
1068
1069 return orun;
1070 }
1071 EXPORT_SYMBOL_GPL(hrtimer_forward);
1072
1073 /*
1074 * enqueue_hrtimer - internal function to (re)start a timer
1075 *
1076 * The timer is inserted in expiry order. Insertion into the
1077 * red black tree is O(log(n)). Must hold the base lock.
1078 *
1079 * Returns 1 when the new timer is the leftmost timer in the tree.
1080 */
enqueue_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,enum hrtimer_mode mode)1081 static int enqueue_hrtimer(struct hrtimer *timer,
1082 struct hrtimer_clock_base *base,
1083 enum hrtimer_mode mode)
1084 {
1085 debug_activate(timer, mode);
1086 WARN_ON_ONCE(!base->cpu_base->online);
1087
1088 base->cpu_base->active_bases |= 1 << base->index;
1089
1090 /* Pairs with the lockless read in hrtimer_is_queued() */
1091 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1092
1093 return timerqueue_add(&base->active, &timer->node);
1094 }
1095
1096 /*
1097 * __remove_hrtimer - internal function to remove a timer
1098 *
1099 * Caller must hold the base lock.
1100 *
1101 * High resolution timer mode reprograms the clock event device when the
1102 * timer is the one which expires next. The caller can disable this by setting
1103 * reprogram to zero. This is useful, when the context does a reprogramming
1104 * anyway (e.g. timer interrupt)
1105 */
__remove_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,u8 newstate,int reprogram)1106 static void __remove_hrtimer(struct hrtimer *timer,
1107 struct hrtimer_clock_base *base,
1108 u8 newstate, int reprogram)
1109 {
1110 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1111 u8 state = timer->state;
1112
1113 /* Pairs with the lockless read in hrtimer_is_queued() */
1114 WRITE_ONCE(timer->state, newstate);
1115 if (!(state & HRTIMER_STATE_ENQUEUED))
1116 return;
1117
1118 if (!timerqueue_del(&base->active, &timer->node))
1119 cpu_base->active_bases &= ~(1 << base->index);
1120
1121 /*
1122 * Note: If reprogram is false we do not update
1123 * cpu_base->next_timer. This happens when we remove the first
1124 * timer on a remote cpu. No harm as we never dereference
1125 * cpu_base->next_timer. So the worst thing what can happen is
1126 * an superfluous call to hrtimer_force_reprogram() on the
1127 * remote cpu later on if the same timer gets enqueued again.
1128 */
1129 if (reprogram && timer == cpu_base->next_timer)
1130 hrtimer_force_reprogram(cpu_base, 1);
1131 }
1132
1133 /*
1134 * remove hrtimer, called with base lock held
1135 */
1136 static inline int
remove_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,bool restart,bool keep_local)1137 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1138 bool restart, bool keep_local)
1139 {
1140 u8 state = timer->state;
1141
1142 if (state & HRTIMER_STATE_ENQUEUED) {
1143 bool reprogram;
1144
1145 /*
1146 * Remove the timer and force reprogramming when high
1147 * resolution mode is active and the timer is on the current
1148 * CPU. If we remove a timer on another CPU, reprogramming is
1149 * skipped. The interrupt event on this CPU is fired and
1150 * reprogramming happens in the interrupt handler. This is a
1151 * rare case and less expensive than a smp call.
1152 */
1153 debug_deactivate(timer);
1154 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1155
1156 /*
1157 * If the timer is not restarted then reprogramming is
1158 * required if the timer is local. If it is local and about
1159 * to be restarted, avoid programming it twice (on removal
1160 * and a moment later when it's requeued).
1161 */
1162 if (!restart)
1163 state = HRTIMER_STATE_INACTIVE;
1164 else
1165 reprogram &= !keep_local;
1166
1167 __remove_hrtimer(timer, base, state, reprogram);
1168 return 1;
1169 }
1170 return 0;
1171 }
1172
hrtimer_update_lowres(struct hrtimer * timer,ktime_t tim,const enum hrtimer_mode mode)1173 static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1174 const enum hrtimer_mode mode)
1175 {
1176 #ifdef CONFIG_TIME_LOW_RES
1177 /*
1178 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1179 * granular time values. For relative timers we add hrtimer_resolution
1180 * (i.e. one jiffie) to prevent short timeouts.
1181 */
1182 timer->is_rel = mode & HRTIMER_MODE_REL;
1183 if (timer->is_rel)
1184 tim = ktime_add_safe(tim, hrtimer_resolution);
1185 #endif
1186 return tim;
1187 }
1188
1189 static void
hrtimer_update_softirq_timer(struct hrtimer_cpu_base * cpu_base,bool reprogram)1190 hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1191 {
1192 ktime_t expires;
1193
1194 /*
1195 * Find the next SOFT expiration.
1196 */
1197 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1198
1199 /*
1200 * reprogramming needs to be triggered, even if the next soft
1201 * hrtimer expires at the same time than the next hard
1202 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1203 */
1204 if (expires == KTIME_MAX)
1205 return;
1206
1207 /*
1208 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1209 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1210 */
1211 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1212 }
1213
__hrtimer_start_range_ns(struct hrtimer * timer,ktime_t tim,u64 delta_ns,const enum hrtimer_mode mode,struct hrtimer_clock_base * base)1214 static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1215 u64 delta_ns, const enum hrtimer_mode mode,
1216 struct hrtimer_clock_base *base)
1217 {
1218 struct hrtimer_clock_base *new_base;
1219 bool force_local, first;
1220
1221 /*
1222 * If the timer is on the local cpu base and is the first expiring
1223 * timer then this might end up reprogramming the hardware twice
1224 * (on removal and on enqueue). To avoid that by prevent the
1225 * reprogram on removal, keep the timer local to the current CPU
1226 * and enforce reprogramming after it is queued no matter whether
1227 * it is the new first expiring timer again or not.
1228 */
1229 force_local = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1230 force_local &= base->cpu_base->next_timer == timer;
1231
1232 /*
1233 * Remove an active timer from the queue. In case it is not queued
1234 * on the current CPU, make sure that remove_hrtimer() updates the
1235 * remote data correctly.
1236 *
1237 * If it's on the current CPU and the first expiring timer, then
1238 * skip reprogramming, keep the timer local and enforce
1239 * reprogramming later if it was the first expiring timer. This
1240 * avoids programming the underlying clock event twice (once at
1241 * removal and once after enqueue).
1242 */
1243 remove_hrtimer(timer, base, true, force_local);
1244
1245 if (mode & HRTIMER_MODE_REL)
1246 tim = ktime_add_safe(tim, base->get_time());
1247
1248 tim = hrtimer_update_lowres(timer, tim, mode);
1249
1250 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1251
1252 /* Switch the timer base, if necessary: */
1253 if (!force_local) {
1254 new_base = switch_hrtimer_base(timer, base,
1255 mode & HRTIMER_MODE_PINNED);
1256 } else {
1257 new_base = base;
1258 }
1259
1260 first = enqueue_hrtimer(timer, new_base, mode);
1261 if (!force_local)
1262 return first;
1263
1264 /*
1265 * Timer was forced to stay on the current CPU to avoid
1266 * reprogramming on removal and enqueue. Force reprogram the
1267 * hardware by evaluating the new first expiring timer.
1268 */
1269 hrtimer_force_reprogram(new_base->cpu_base, 1);
1270 return 0;
1271 }
1272
1273 /**
1274 * hrtimer_start_range_ns - (re)start an hrtimer
1275 * @timer: the timer to be added
1276 * @tim: expiry time
1277 * @delta_ns: "slack" range for the timer
1278 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1279 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1280 * softirq based mode is considered for debug purpose only!
1281 */
hrtimer_start_range_ns(struct hrtimer * timer,ktime_t tim,u64 delta_ns,const enum hrtimer_mode mode)1282 void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1283 u64 delta_ns, const enum hrtimer_mode mode)
1284 {
1285 struct hrtimer_clock_base *base;
1286 unsigned long flags;
1287
1288 /*
1289 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1290 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1291 * expiry mode because unmarked timers are moved to softirq expiry.
1292 */
1293 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1294 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1295 else
1296 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1297
1298 base = lock_hrtimer_base(timer, &flags);
1299
1300 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1301 hrtimer_reprogram(timer, true);
1302
1303 unlock_hrtimer_base(timer, &flags);
1304 }
1305 EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1306
1307 /**
1308 * hrtimer_try_to_cancel - try to deactivate a timer
1309 * @timer: hrtimer to stop
1310 *
1311 * Returns:
1312 *
1313 * * 0 when the timer was not active
1314 * * 1 when the timer was active
1315 * * -1 when the timer is currently executing the callback function and
1316 * cannot be stopped
1317 */
hrtimer_try_to_cancel(struct hrtimer * timer)1318 int hrtimer_try_to_cancel(struct hrtimer *timer)
1319 {
1320 struct hrtimer_clock_base *base;
1321 unsigned long flags;
1322 int ret = -1;
1323
1324 /*
1325 * Check lockless first. If the timer is not active (neither
1326 * enqueued nor running the callback, nothing to do here. The
1327 * base lock does not serialize against a concurrent enqueue,
1328 * so we can avoid taking it.
1329 */
1330 if (!hrtimer_active(timer))
1331 return 0;
1332
1333 base = lock_hrtimer_base(timer, &flags);
1334
1335 if (!hrtimer_callback_running(timer))
1336 ret = remove_hrtimer(timer, base, false, false);
1337
1338 unlock_hrtimer_base(timer, &flags);
1339
1340 return ret;
1341
1342 }
1343 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1344
1345 #ifdef CONFIG_PREEMPT_RT
hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base * base)1346 static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1347 {
1348 spin_lock_init(&base->softirq_expiry_lock);
1349 }
1350
hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base * base)1351 static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1352 {
1353 spin_lock(&base->softirq_expiry_lock);
1354 }
1355
hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base * base)1356 static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1357 {
1358 spin_unlock(&base->softirq_expiry_lock);
1359 }
1360
1361 /*
1362 * The counterpart to hrtimer_cancel_wait_running().
1363 *
1364 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1365 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1366 * allows the waiter to acquire the lock and make progress.
1367 */
hrtimer_sync_wait_running(struct hrtimer_cpu_base * cpu_base,unsigned long flags)1368 static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1369 unsigned long flags)
1370 {
1371 if (atomic_read(&cpu_base->timer_waiters)) {
1372 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1373 spin_unlock(&cpu_base->softirq_expiry_lock);
1374 spin_lock(&cpu_base->softirq_expiry_lock);
1375 raw_spin_lock_irq(&cpu_base->lock);
1376 }
1377 }
1378
1379 /*
1380 * This function is called on PREEMPT_RT kernels when the fast path
1381 * deletion of a timer failed because the timer callback function was
1382 * running.
1383 *
1384 * This prevents priority inversion: if the soft irq thread is preempted
1385 * in the middle of a timer callback, then calling del_timer_sync() can
1386 * lead to two issues:
1387 *
1388 * - If the caller is on a remote CPU then it has to spin wait for the timer
1389 * handler to complete. This can result in unbound priority inversion.
1390 *
1391 * - If the caller originates from the task which preempted the timer
1392 * handler on the same CPU, then spin waiting for the timer handler to
1393 * complete is never going to end.
1394 */
hrtimer_cancel_wait_running(const struct hrtimer * timer)1395 void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1396 {
1397 /* Lockless read. Prevent the compiler from reloading it below */
1398 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1399
1400 /*
1401 * Just relax if the timer expires in hard interrupt context or if
1402 * it is currently on the migration base.
1403 */
1404 if (!timer->is_soft || is_migration_base(base)) {
1405 cpu_relax();
1406 return;
1407 }
1408
1409 /*
1410 * Mark the base as contended and grab the expiry lock, which is
1411 * held by the softirq across the timer callback. Drop the lock
1412 * immediately so the softirq can expire the next timer. In theory
1413 * the timer could already be running again, but that's more than
1414 * unlikely and just causes another wait loop.
1415 */
1416 atomic_inc(&base->cpu_base->timer_waiters);
1417 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1418 atomic_dec(&base->cpu_base->timer_waiters);
1419 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1420 }
1421 #else
1422 static inline void
hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base * base)1423 hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1424 static inline void
hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base * base)1425 hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1426 static inline void
hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base * base)1427 hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
hrtimer_sync_wait_running(struct hrtimer_cpu_base * base,unsigned long flags)1428 static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1429 unsigned long flags) { }
1430 #endif
1431
1432 /**
1433 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1434 * @timer: the timer to be cancelled
1435 *
1436 * Returns:
1437 * 0 when the timer was not active
1438 * 1 when the timer was active
1439 */
hrtimer_cancel(struct hrtimer * timer)1440 int hrtimer_cancel(struct hrtimer *timer)
1441 {
1442 int ret;
1443
1444 do {
1445 ret = hrtimer_try_to_cancel(timer);
1446
1447 if (ret < 0)
1448 hrtimer_cancel_wait_running(timer);
1449 } while (ret < 0);
1450 return ret;
1451 }
1452 EXPORT_SYMBOL_GPL(hrtimer_cancel);
1453
1454 /**
1455 * __hrtimer_get_remaining - get remaining time for the timer
1456 * @timer: the timer to read
1457 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1458 */
__hrtimer_get_remaining(const struct hrtimer * timer,bool adjust)1459 ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1460 {
1461 unsigned long flags;
1462 ktime_t rem;
1463
1464 lock_hrtimer_base(timer, &flags);
1465 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1466 rem = hrtimer_expires_remaining_adjusted(timer);
1467 else
1468 rem = hrtimer_expires_remaining(timer);
1469 unlock_hrtimer_base(timer, &flags);
1470
1471 return rem;
1472 }
1473 EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1474
1475 #ifdef CONFIG_NO_HZ_COMMON
1476 /**
1477 * hrtimer_get_next_event - get the time until next expiry event
1478 *
1479 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1480 */
hrtimer_get_next_event(void)1481 u64 hrtimer_get_next_event(void)
1482 {
1483 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1484 u64 expires = KTIME_MAX;
1485 unsigned long flags;
1486
1487 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1488
1489 if (!hrtimer_hres_active(cpu_base))
1490 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1491
1492 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1493
1494 return expires;
1495 }
1496
1497 /**
1498 * hrtimer_next_event_without - time until next expiry event w/o one timer
1499 * @exclude: timer to exclude
1500 *
1501 * Returns the next expiry time over all timers except for the @exclude one or
1502 * KTIME_MAX if none of them is pending.
1503 */
hrtimer_next_event_without(const struct hrtimer * exclude)1504 u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1505 {
1506 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1507 u64 expires = KTIME_MAX;
1508 unsigned long flags;
1509
1510 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1511
1512 if (hrtimer_hres_active(cpu_base)) {
1513 unsigned int active;
1514
1515 if (!cpu_base->softirq_activated) {
1516 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1517 expires = __hrtimer_next_event_base(cpu_base, exclude,
1518 active, KTIME_MAX);
1519 }
1520 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1521 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1522 expires);
1523 }
1524
1525 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1526
1527 return expires;
1528 }
1529 #endif
1530
hrtimer_clockid_to_base(clockid_t clock_id)1531 static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1532 {
1533 if (likely(clock_id < MAX_CLOCKS)) {
1534 int base = hrtimer_clock_to_base_table[clock_id];
1535
1536 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1537 return base;
1538 }
1539 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1540 return HRTIMER_BASE_MONOTONIC;
1541 }
1542
__hrtimer_init(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)1543 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1544 enum hrtimer_mode mode)
1545 {
1546 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1547 struct hrtimer_cpu_base *cpu_base;
1548 int base;
1549
1550 /*
1551 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1552 * marked for hard interrupt expiry mode are moved into soft
1553 * interrupt context for latency reasons and because the callbacks
1554 * can invoke functions which might sleep on RT, e.g. spin_lock().
1555 */
1556 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1557 softtimer = true;
1558
1559 memset(timer, 0, sizeof(struct hrtimer));
1560
1561 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1562
1563 /*
1564 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1565 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1566 * ensure POSIX compliance.
1567 */
1568 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1569 clock_id = CLOCK_MONOTONIC;
1570
1571 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1572 base += hrtimer_clockid_to_base(clock_id);
1573 timer->is_soft = softtimer;
1574 timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1575 timer->base = &cpu_base->clock_base[base];
1576 timerqueue_init(&timer->node);
1577 }
1578
1579 /**
1580 * hrtimer_init - initialize a timer to the given clock
1581 * @timer: the timer to be initialized
1582 * @clock_id: the clock to be used
1583 * @mode: The modes which are relevant for initialization:
1584 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1585 * HRTIMER_MODE_REL_SOFT
1586 *
1587 * The PINNED variants of the above can be handed in,
1588 * but the PINNED bit is ignored as pinning happens
1589 * when the hrtimer is started
1590 */
hrtimer_init(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)1591 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1592 enum hrtimer_mode mode)
1593 {
1594 debug_init(timer, clock_id, mode);
1595 __hrtimer_init(timer, clock_id, mode);
1596 }
1597 EXPORT_SYMBOL_GPL(hrtimer_init);
1598
1599 /*
1600 * A timer is active, when it is enqueued into the rbtree or the
1601 * callback function is running or it's in the state of being migrated
1602 * to another cpu.
1603 *
1604 * It is important for this function to not return a false negative.
1605 */
hrtimer_active(const struct hrtimer * timer)1606 bool hrtimer_active(const struct hrtimer *timer)
1607 {
1608 struct hrtimer_clock_base *base;
1609 unsigned int seq;
1610
1611 do {
1612 base = READ_ONCE(timer->base);
1613 seq = raw_read_seqcount_begin(&base->seq);
1614
1615 if (timer->state != HRTIMER_STATE_INACTIVE ||
1616 base->running == timer)
1617 return true;
1618
1619 } while (read_seqcount_retry(&base->seq, seq) ||
1620 base != READ_ONCE(timer->base));
1621
1622 return false;
1623 }
1624 EXPORT_SYMBOL_GPL(hrtimer_active);
1625
1626 /*
1627 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1628 * distinct sections:
1629 *
1630 * - queued: the timer is queued
1631 * - callback: the timer is being ran
1632 * - post: the timer is inactive or (re)queued
1633 *
1634 * On the read side we ensure we observe timer->state and cpu_base->running
1635 * from the same section, if anything changed while we looked at it, we retry.
1636 * This includes timer->base changing because sequence numbers alone are
1637 * insufficient for that.
1638 *
1639 * The sequence numbers are required because otherwise we could still observe
1640 * a false negative if the read side got smeared over multiple consecutive
1641 * __run_hrtimer() invocations.
1642 */
1643
__run_hrtimer(struct hrtimer_cpu_base * cpu_base,struct hrtimer_clock_base * base,struct hrtimer * timer,ktime_t * now,unsigned long flags)1644 static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1645 struct hrtimer_clock_base *base,
1646 struct hrtimer *timer, ktime_t *now,
1647 unsigned long flags) __must_hold(&cpu_base->lock)
1648 {
1649 enum hrtimer_restart (*fn)(struct hrtimer *);
1650 bool expires_in_hardirq;
1651 int restart;
1652
1653 lockdep_assert_held(&cpu_base->lock);
1654
1655 debug_deactivate(timer);
1656 base->running = timer;
1657
1658 /*
1659 * Separate the ->running assignment from the ->state assignment.
1660 *
1661 * As with a regular write barrier, this ensures the read side in
1662 * hrtimer_active() cannot observe base->running == NULL &&
1663 * timer->state == INACTIVE.
1664 */
1665 raw_write_seqcount_barrier(&base->seq);
1666
1667 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1668 fn = timer->function;
1669
1670 /*
1671 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1672 * timer is restarted with a period then it becomes an absolute
1673 * timer. If its not restarted it does not matter.
1674 */
1675 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1676 timer->is_rel = false;
1677
1678 /*
1679 * The timer is marked as running in the CPU base, so it is
1680 * protected against migration to a different CPU even if the lock
1681 * is dropped.
1682 */
1683 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1684 trace_hrtimer_expire_entry(timer, now);
1685 expires_in_hardirq = lockdep_hrtimer_enter(timer);
1686
1687 restart = fn(timer);
1688
1689 lockdep_hrtimer_exit(expires_in_hardirq);
1690 trace_hrtimer_expire_exit(timer);
1691 raw_spin_lock_irq(&cpu_base->lock);
1692
1693 /*
1694 * Note: We clear the running state after enqueue_hrtimer and
1695 * we do not reprogram the event hardware. Happens either in
1696 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1697 *
1698 * Note: Because we dropped the cpu_base->lock above,
1699 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1700 * for us already.
1701 */
1702 if (restart != HRTIMER_NORESTART &&
1703 !(timer->state & HRTIMER_STATE_ENQUEUED))
1704 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1705
1706 /*
1707 * Separate the ->running assignment from the ->state assignment.
1708 *
1709 * As with a regular write barrier, this ensures the read side in
1710 * hrtimer_active() cannot observe base->running.timer == NULL &&
1711 * timer->state == INACTIVE.
1712 */
1713 raw_write_seqcount_barrier(&base->seq);
1714
1715 WARN_ON_ONCE(base->running != timer);
1716 base->running = NULL;
1717 }
1718
__hrtimer_run_queues(struct hrtimer_cpu_base * cpu_base,ktime_t now,unsigned long flags,unsigned int active_mask)1719 static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1720 unsigned long flags, unsigned int active_mask)
1721 {
1722 struct hrtimer_clock_base *base;
1723 unsigned int active = cpu_base->active_bases & active_mask;
1724
1725 for_each_active_base(base, cpu_base, active) {
1726 struct timerqueue_node *node;
1727 ktime_t basenow;
1728
1729 basenow = ktime_add(now, base->offset);
1730
1731 while ((node = timerqueue_getnext(&base->active))) {
1732 struct hrtimer *timer;
1733
1734 timer = container_of(node, struct hrtimer, node);
1735
1736 /*
1737 * The immediate goal for using the softexpires is
1738 * minimizing wakeups, not running timers at the
1739 * earliest interrupt after their soft expiration.
1740 * This allows us to avoid using a Priority Search
1741 * Tree, which can answer a stabbing query for
1742 * overlapping intervals and instead use the simple
1743 * BST we already have.
1744 * We don't add extra wakeups by delaying timers that
1745 * are right-of a not yet expired timer, because that
1746 * timer will have to trigger a wakeup anyway.
1747 */
1748 if (basenow < hrtimer_get_softexpires_tv64(timer))
1749 break;
1750
1751 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1752 if (active_mask == HRTIMER_ACTIVE_SOFT)
1753 hrtimer_sync_wait_running(cpu_base, flags);
1754 }
1755 }
1756 }
1757
hrtimer_run_softirq(struct softirq_action * h)1758 static __latent_entropy void hrtimer_run_softirq(struct softirq_action *h)
1759 {
1760 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1761 unsigned long flags;
1762 ktime_t now;
1763
1764 hrtimer_cpu_base_lock_expiry(cpu_base);
1765 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1766
1767 now = hrtimer_update_base(cpu_base);
1768 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1769
1770 cpu_base->softirq_activated = 0;
1771 hrtimer_update_softirq_timer(cpu_base, true);
1772
1773 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1774 hrtimer_cpu_base_unlock_expiry(cpu_base);
1775 }
1776
1777 #ifdef CONFIG_HIGH_RES_TIMERS
1778
1779 /*
1780 * High resolution timer interrupt
1781 * Called with interrupts disabled
1782 */
hrtimer_interrupt(struct clock_event_device * dev)1783 void hrtimer_interrupt(struct clock_event_device *dev)
1784 {
1785 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1786 ktime_t expires_next, now, entry_time, delta;
1787 unsigned long flags;
1788 int retries = 0;
1789
1790 BUG_ON(!cpu_base->hres_active);
1791 cpu_base->nr_events++;
1792 dev->next_event = KTIME_MAX;
1793
1794 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1795 entry_time = now = hrtimer_update_base(cpu_base);
1796 retry:
1797 cpu_base->in_hrtirq = 1;
1798 /*
1799 * We set expires_next to KTIME_MAX here with cpu_base->lock
1800 * held to prevent that a timer is enqueued in our queue via
1801 * the migration code. This does not affect enqueueing of
1802 * timers which run their callback and need to be requeued on
1803 * this CPU.
1804 */
1805 cpu_base->expires_next = KTIME_MAX;
1806
1807 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1808 cpu_base->softirq_expires_next = KTIME_MAX;
1809 cpu_base->softirq_activated = 1;
1810 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1811 }
1812
1813 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1814
1815 /* Reevaluate the clock bases for the [soft] next expiry */
1816 expires_next = hrtimer_update_next_event(cpu_base);
1817 /*
1818 * Store the new expiry value so the migration code can verify
1819 * against it.
1820 */
1821 cpu_base->expires_next = expires_next;
1822 cpu_base->in_hrtirq = 0;
1823 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1824
1825 /* Reprogramming necessary ? */
1826 if (!tick_program_event(expires_next, 0)) {
1827 cpu_base->hang_detected = 0;
1828 return;
1829 }
1830
1831 /*
1832 * The next timer was already expired due to:
1833 * - tracing
1834 * - long lasting callbacks
1835 * - being scheduled away when running in a VM
1836 *
1837 * We need to prevent that we loop forever in the hrtimer
1838 * interrupt routine. We give it 3 attempts to avoid
1839 * overreacting on some spurious event.
1840 *
1841 * Acquire base lock for updating the offsets and retrieving
1842 * the current time.
1843 */
1844 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1845 now = hrtimer_update_base(cpu_base);
1846 cpu_base->nr_retries++;
1847 if (++retries < 3)
1848 goto retry;
1849 /*
1850 * Give the system a chance to do something else than looping
1851 * here. We stored the entry time, so we know exactly how long
1852 * we spent here. We schedule the next event this amount of
1853 * time away.
1854 */
1855 cpu_base->nr_hangs++;
1856 cpu_base->hang_detected = 1;
1857 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1858
1859 delta = ktime_sub(now, entry_time);
1860 if ((unsigned int)delta > cpu_base->max_hang_time)
1861 cpu_base->max_hang_time = (unsigned int) delta;
1862 /*
1863 * Limit it to a sensible value as we enforce a longer
1864 * delay. Give the CPU at least 100ms to catch up.
1865 */
1866 if (delta > 100 * NSEC_PER_MSEC)
1867 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1868 else
1869 expires_next = ktime_add(now, delta);
1870 tick_program_event(expires_next, 1);
1871 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1872 }
1873 #endif /* !CONFIG_HIGH_RES_TIMERS */
1874
1875 /*
1876 * Called from run_local_timers in hardirq context every jiffy
1877 */
hrtimer_run_queues(void)1878 void hrtimer_run_queues(void)
1879 {
1880 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1881 unsigned long flags;
1882 ktime_t now;
1883
1884 if (hrtimer_hres_active(cpu_base))
1885 return;
1886
1887 /*
1888 * This _is_ ugly: We have to check periodically, whether we
1889 * can switch to highres and / or nohz mode. The clocksource
1890 * switch happens with xtime_lock held. Notification from
1891 * there only sets the check bit in the tick_oneshot code,
1892 * otherwise we might deadlock vs. xtime_lock.
1893 */
1894 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1895 hrtimer_switch_to_hres();
1896 return;
1897 }
1898
1899 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1900 now = hrtimer_update_base(cpu_base);
1901
1902 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1903 cpu_base->softirq_expires_next = KTIME_MAX;
1904 cpu_base->softirq_activated = 1;
1905 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1906 }
1907
1908 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1909 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1910 }
1911
1912 /*
1913 * Sleep related functions:
1914 */
hrtimer_wakeup(struct hrtimer * timer)1915 static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1916 {
1917 struct hrtimer_sleeper *t =
1918 container_of(timer, struct hrtimer_sleeper, timer);
1919 struct task_struct *task = t->task;
1920
1921 t->task = NULL;
1922 if (task)
1923 wake_up_process(task);
1924
1925 return HRTIMER_NORESTART;
1926 }
1927
1928 /**
1929 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
1930 * @sl: sleeper to be started
1931 * @mode: timer mode abs/rel
1932 *
1933 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
1934 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
1935 */
hrtimer_sleeper_start_expires(struct hrtimer_sleeper * sl,enum hrtimer_mode mode)1936 void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
1937 enum hrtimer_mode mode)
1938 {
1939 /*
1940 * Make the enqueue delivery mode check work on RT. If the sleeper
1941 * was initialized for hard interrupt delivery, force the mode bit.
1942 * This is a special case for hrtimer_sleepers because
1943 * hrtimer_init_sleeper() determines the delivery mode on RT so the
1944 * fiddling with this decision is avoided at the call sites.
1945 */
1946 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
1947 mode |= HRTIMER_MODE_HARD;
1948
1949 hrtimer_start_expires(&sl->timer, mode);
1950 }
1951 EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
1952
__hrtimer_init_sleeper(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)1953 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
1954 clockid_t clock_id, enum hrtimer_mode mode)
1955 {
1956 /*
1957 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1958 * marked for hard interrupt expiry mode are moved into soft
1959 * interrupt context either for latency reasons or because the
1960 * hrtimer callback takes regular spinlocks or invokes other
1961 * functions which are not suitable for hard interrupt context on
1962 * PREEMPT_RT.
1963 *
1964 * The hrtimer_sleeper callback is RT compatible in hard interrupt
1965 * context, but there is a latency concern: Untrusted userspace can
1966 * spawn many threads which arm timers for the same expiry time on
1967 * the same CPU. That causes a latency spike due to the wakeup of
1968 * a gazillion threads.
1969 *
1970 * OTOH, privileged real-time user space applications rely on the
1971 * low latency of hard interrupt wakeups. If the current task is in
1972 * a real-time scheduling class, mark the mode for hard interrupt
1973 * expiry.
1974 */
1975 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1976 if (task_is_realtime(current) && !(mode & HRTIMER_MODE_SOFT))
1977 mode |= HRTIMER_MODE_HARD;
1978 }
1979
1980 __hrtimer_init(&sl->timer, clock_id, mode);
1981 sl->timer.function = hrtimer_wakeup;
1982 sl->task = current;
1983 }
1984
1985 /**
1986 * hrtimer_init_sleeper - initialize sleeper to the given clock
1987 * @sl: sleeper to be initialized
1988 * @clock_id: the clock to be used
1989 * @mode: timer mode abs/rel
1990 */
hrtimer_init_sleeper(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)1991 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, clockid_t clock_id,
1992 enum hrtimer_mode mode)
1993 {
1994 debug_init(&sl->timer, clock_id, mode);
1995 __hrtimer_init_sleeper(sl, clock_id, mode);
1996
1997 }
1998 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1999
nanosleep_copyout(struct restart_block * restart,struct timespec64 * ts)2000 int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
2001 {
2002 switch(restart->nanosleep.type) {
2003 #ifdef CONFIG_COMPAT_32BIT_TIME
2004 case TT_COMPAT:
2005 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
2006 return -EFAULT;
2007 break;
2008 #endif
2009 case TT_NATIVE:
2010 if (put_timespec64(ts, restart->nanosleep.rmtp))
2011 return -EFAULT;
2012 break;
2013 default:
2014 BUG();
2015 }
2016 return -ERESTART_RESTARTBLOCK;
2017 }
2018
do_nanosleep(struct hrtimer_sleeper * t,enum hrtimer_mode mode)2019 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
2020 {
2021 struct restart_block *restart;
2022
2023 do {
2024 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
2025 hrtimer_sleeper_start_expires(t, mode);
2026
2027 if (likely(t->task))
2028 schedule();
2029
2030 hrtimer_cancel(&t->timer);
2031 mode = HRTIMER_MODE_ABS;
2032
2033 } while (t->task && !signal_pending(current));
2034
2035 __set_current_state(TASK_RUNNING);
2036
2037 if (!t->task)
2038 return 0;
2039
2040 restart = ¤t->restart_block;
2041 if (restart->nanosleep.type != TT_NONE) {
2042 ktime_t rem = hrtimer_expires_remaining(&t->timer);
2043 struct timespec64 rmt;
2044
2045 if (rem <= 0)
2046 return 0;
2047 rmt = ktime_to_timespec64(rem);
2048
2049 return nanosleep_copyout(restart, &rmt);
2050 }
2051 return -ERESTART_RESTARTBLOCK;
2052 }
2053
hrtimer_nanosleep_restart(struct restart_block * restart)2054 static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2055 {
2056 struct hrtimer_sleeper t;
2057 int ret;
2058
2059 hrtimer_init_sleeper_on_stack(&t, restart->nanosleep.clockid,
2060 HRTIMER_MODE_ABS);
2061 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
2062 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2063 destroy_hrtimer_on_stack(&t.timer);
2064 return ret;
2065 }
2066
hrtimer_nanosleep(ktime_t rqtp,const enum hrtimer_mode mode,const clockid_t clockid)2067 long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2068 const clockid_t clockid)
2069 {
2070 struct restart_block *restart;
2071 struct hrtimer_sleeper t;
2072 int ret = 0;
2073 u64 slack;
2074
2075 slack = current->timer_slack_ns;
2076 if (rt_task(current))
2077 slack = 0;
2078
2079 hrtimer_init_sleeper_on_stack(&t, clockid, mode);
2080 hrtimer_set_expires_range_ns(&t.timer, rqtp, slack);
2081 ret = do_nanosleep(&t, mode);
2082 if (ret != -ERESTART_RESTARTBLOCK)
2083 goto out;
2084
2085 /* Absolute timers do not update the rmtp value and restart: */
2086 if (mode == HRTIMER_MODE_ABS) {
2087 ret = -ERESTARTNOHAND;
2088 goto out;
2089 }
2090
2091 restart = ¤t->restart_block;
2092 restart->nanosleep.clockid = t.timer.base->clockid;
2093 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2094 set_restart_fn(restart, hrtimer_nanosleep_restart);
2095 out:
2096 destroy_hrtimer_on_stack(&t.timer);
2097 return ret;
2098 }
2099
2100 #ifdef CONFIG_64BIT
2101
SYSCALL_DEFINE2(nanosleep,struct __kernel_timespec __user *,rqtp,struct __kernel_timespec __user *,rmtp)2102 SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2103 struct __kernel_timespec __user *, rmtp)
2104 {
2105 struct timespec64 tu;
2106
2107 if (get_timespec64(&tu, rqtp))
2108 return -EFAULT;
2109
2110 if (!timespec64_valid(&tu))
2111 return -EINVAL;
2112
2113 current->restart_block.fn = do_no_restart_syscall;
2114 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2115 current->restart_block.nanosleep.rmtp = rmtp;
2116 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2117 CLOCK_MONOTONIC);
2118 }
2119
2120 #endif
2121
2122 #ifdef CONFIG_COMPAT_32BIT_TIME
2123
SYSCALL_DEFINE2(nanosleep_time32,struct old_timespec32 __user *,rqtp,struct old_timespec32 __user *,rmtp)2124 SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2125 struct old_timespec32 __user *, rmtp)
2126 {
2127 struct timespec64 tu;
2128
2129 if (get_old_timespec32(&tu, rqtp))
2130 return -EFAULT;
2131
2132 if (!timespec64_valid(&tu))
2133 return -EINVAL;
2134
2135 current->restart_block.fn = do_no_restart_syscall;
2136 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2137 current->restart_block.nanosleep.compat_rmtp = rmtp;
2138 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2139 CLOCK_MONOTONIC);
2140 }
2141 #endif
2142
2143 /*
2144 * Functions related to boot-time initialization:
2145 */
hrtimers_prepare_cpu(unsigned int cpu)2146 int hrtimers_prepare_cpu(unsigned int cpu)
2147 {
2148 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2149 int i;
2150
2151 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2152 struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2153
2154 clock_b->cpu_base = cpu_base;
2155 seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2156 timerqueue_init_head(&clock_b->active);
2157 }
2158
2159 cpu_base->cpu = cpu;
2160 cpu_base->active_bases = 0;
2161 cpu_base->hres_active = 0;
2162 cpu_base->hang_detected = 0;
2163 cpu_base->next_timer = NULL;
2164 cpu_base->softirq_next_timer = NULL;
2165 cpu_base->expires_next = KTIME_MAX;
2166 cpu_base->softirq_expires_next = KTIME_MAX;
2167 cpu_base->online = 1;
2168 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2169 return 0;
2170 }
2171
2172 #ifdef CONFIG_HOTPLUG_CPU
2173
migrate_hrtimer_list(struct hrtimer_clock_base * old_base,struct hrtimer_clock_base * new_base)2174 static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2175 struct hrtimer_clock_base *new_base)
2176 {
2177 struct hrtimer *timer;
2178 struct timerqueue_node *node;
2179
2180 while ((node = timerqueue_getnext(&old_base->active))) {
2181 timer = container_of(node, struct hrtimer, node);
2182 BUG_ON(hrtimer_callback_running(timer));
2183 debug_deactivate(timer);
2184
2185 /*
2186 * Mark it as ENQUEUED not INACTIVE otherwise the
2187 * timer could be seen as !active and just vanish away
2188 * under us on another CPU
2189 */
2190 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2191 timer->base = new_base;
2192 /*
2193 * Enqueue the timers on the new cpu. This does not
2194 * reprogram the event device in case the timer
2195 * expires before the earliest on this CPU, but we run
2196 * hrtimer_interrupt after we migrated everything to
2197 * sort out already expired timers and reprogram the
2198 * event device.
2199 */
2200 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2201 }
2202 }
2203
hrtimers_cpu_dying(unsigned int dying_cpu)2204 int hrtimers_cpu_dying(unsigned int dying_cpu)
2205 {
2206 int i, ncpu = cpumask_any_and(cpu_active_mask, housekeeping_cpumask(HK_TYPE_TIMER));
2207 struct hrtimer_cpu_base *old_base, *new_base;
2208
2209 old_base = this_cpu_ptr(&hrtimer_bases);
2210 new_base = &per_cpu(hrtimer_bases, ncpu);
2211
2212 /*
2213 * The caller is globally serialized and nobody else
2214 * takes two locks at once, deadlock is not possible.
2215 */
2216 raw_spin_lock(&old_base->lock);
2217 raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2218
2219 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2220 migrate_hrtimer_list(&old_base->clock_base[i],
2221 &new_base->clock_base[i]);
2222 }
2223
2224 /*
2225 * The migration might have changed the first expiring softirq
2226 * timer on this CPU. Update it.
2227 */
2228 __hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
2229 /* Tell the other CPU to retrigger the next event */
2230 smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2231
2232 raw_spin_unlock(&new_base->lock);
2233 old_base->online = 0;
2234 raw_spin_unlock(&old_base->lock);
2235
2236 return 0;
2237 }
2238
2239 #endif /* CONFIG_HOTPLUG_CPU */
2240
hrtimers_init(void)2241 void __init hrtimers_init(void)
2242 {
2243 hrtimers_prepare_cpu(smp_processor_id());
2244 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2245 }
2246
2247 /**
2248 * schedule_hrtimeout_range_clock - sleep until timeout
2249 * @expires: timeout value (ktime_t)
2250 * @delta: slack in expires timeout (ktime_t) for SCHED_OTHER tasks
2251 * @mode: timer mode
2252 * @clock_id: timer clock to be used
2253 */
2254 int __sched
schedule_hrtimeout_range_clock(ktime_t * expires,u64 delta,const enum hrtimer_mode mode,clockid_t clock_id)2255 schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
2256 const enum hrtimer_mode mode, clockid_t clock_id)
2257 {
2258 struct hrtimer_sleeper t;
2259
2260 /*
2261 * Optimize when a zero timeout value is given. It does not
2262 * matter whether this is an absolute or a relative time.
2263 */
2264 if (expires && *expires == 0) {
2265 __set_current_state(TASK_RUNNING);
2266 return 0;
2267 }
2268
2269 /*
2270 * A NULL parameter means "infinite"
2271 */
2272 if (!expires) {
2273 schedule();
2274 return -EINTR;
2275 }
2276
2277 /*
2278 * Override any slack passed by the user if under
2279 * rt contraints.
2280 */
2281 if (rt_task(current))
2282 delta = 0;
2283
2284 hrtimer_init_sleeper_on_stack(&t, clock_id, mode);
2285 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
2286 hrtimer_sleeper_start_expires(&t, mode);
2287
2288 if (likely(t.task))
2289 schedule();
2290
2291 hrtimer_cancel(&t.timer);
2292 destroy_hrtimer_on_stack(&t.timer);
2293
2294 __set_current_state(TASK_RUNNING);
2295
2296 return !t.task ? 0 : -EINTR;
2297 }
2298 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range_clock);
2299
2300 /**
2301 * schedule_hrtimeout_range - sleep until timeout
2302 * @expires: timeout value (ktime_t)
2303 * @delta: slack in expires timeout (ktime_t) for SCHED_OTHER tasks
2304 * @mode: timer mode
2305 *
2306 * Make the current task sleep until the given expiry time has
2307 * elapsed. The routine will return immediately unless
2308 * the current task state has been set (see set_current_state()).
2309 *
2310 * The @delta argument gives the kernel the freedom to schedule the
2311 * actual wakeup to a time that is both power and performance friendly
2312 * for regular (non RT/DL) tasks.
2313 * The kernel give the normal best effort behavior for "@expires+@delta",
2314 * but may decide to fire the timer earlier, but no earlier than @expires.
2315 *
2316 * You can set the task state as follows -
2317 *
2318 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2319 * pass before the routine returns unless the current task is explicitly
2320 * woken up, (e.g. by wake_up_process()).
2321 *
2322 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2323 * delivered to the current task or the current task is explicitly woken
2324 * up.
2325 *
2326 * The current task state is guaranteed to be TASK_RUNNING when this
2327 * routine returns.
2328 *
2329 * Returns 0 when the timer has expired. If the task was woken before the
2330 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2331 * by an explicit wakeup, it returns -EINTR.
2332 */
schedule_hrtimeout_range(ktime_t * expires,u64 delta,const enum hrtimer_mode mode)2333 int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
2334 const enum hrtimer_mode mode)
2335 {
2336 return schedule_hrtimeout_range_clock(expires, delta, mode,
2337 CLOCK_MONOTONIC);
2338 }
2339 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
2340
2341 /**
2342 * schedule_hrtimeout - sleep until timeout
2343 * @expires: timeout value (ktime_t)
2344 * @mode: timer mode
2345 *
2346 * Make the current task sleep until the given expiry time has
2347 * elapsed. The routine will return immediately unless
2348 * the current task state has been set (see set_current_state()).
2349 *
2350 * You can set the task state as follows -
2351 *
2352 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2353 * pass before the routine returns unless the current task is explicitly
2354 * woken up, (e.g. by wake_up_process()).
2355 *
2356 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2357 * delivered to the current task or the current task is explicitly woken
2358 * up.
2359 *
2360 * The current task state is guaranteed to be TASK_RUNNING when this
2361 * routine returns.
2362 *
2363 * Returns 0 when the timer has expired. If the task was woken before the
2364 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2365 * by an explicit wakeup, it returns -EINTR.
2366 */
schedule_hrtimeout(ktime_t * expires,const enum hrtimer_mode mode)2367 int __sched schedule_hrtimeout(ktime_t *expires,
2368 const enum hrtimer_mode mode)
2369 {
2370 return schedule_hrtimeout_range(expires, 0, mode);
2371 }
2372 EXPORT_SYMBOL_GPL(schedule_hrtimeout);
2373