1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API, aka timer wheel,
10 * hrtimers provide finer resolution and accuracy depending on system
11 * configuration and capabilities.
12 *
13 * Started by: Thomas Gleixner and Ingo Molnar
14 *
15 * Credits:
16 * Based on the original timer wheel code
17 *
18 * Help, testing, suggestions, bugfixes, improvements were
19 * provided by:
20 *
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
22 * et. al.
23 */
24
25 #include <linux/cpu.h>
26 #include <linux/export.h>
27 #include <linux/percpu.h>
28 #include <linux/hrtimer.h>
29 #include <linux/notifier.h>
30 #include <linux/syscalls.h>
31 #include <linux/interrupt.h>
32 #include <linux/tick.h>
33 #include <linux/err.h>
34 #include <linux/debugobjects.h>
35 #include <linux/sched/signal.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/sched/rt.h>
38 #include <linux/sched/deadline.h>
39 #include <linux/sched/nohz.h>
40 #include <linux/sched/debug.h>
41 #include <linux/timer.h>
42 #include <linux/freezer.h>
43 #include <linux/compat.h>
44
45 #include <linux/uaccess.h>
46
47 #include <trace/events/timer.h>
48
49 #include "tick-internal.h"
50
51 /*
52 * Masks for selecting the soft and hard context timers from
53 * cpu_base->active
54 */
55 #define MASK_SHIFT (HRTIMER_BASE_MONOTONIC_SOFT)
56 #define HRTIMER_ACTIVE_HARD ((1U << MASK_SHIFT) - 1)
57 #define HRTIMER_ACTIVE_SOFT (HRTIMER_ACTIVE_HARD << MASK_SHIFT)
58 #define HRTIMER_ACTIVE_ALL (HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
59
60 /*
61 * The timer bases:
62 *
63 * There are more clockids than hrtimer bases. Thus, we index
64 * into the timer bases by the hrtimer_base_type enum. When trying
65 * to reach a base using a clockid, hrtimer_clockid_to_base()
66 * is used to convert from clockid to the proper hrtimer_base_type.
67 */
68 DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
69 {
70 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
71 .clock_base =
72 {
73 {
74 .index = HRTIMER_BASE_MONOTONIC,
75 .clockid = CLOCK_MONOTONIC,
76 .get_time = &ktime_get,
77 },
78 {
79 .index = HRTIMER_BASE_REALTIME,
80 .clockid = CLOCK_REALTIME,
81 .get_time = &ktime_get_real,
82 },
83 {
84 .index = HRTIMER_BASE_BOOTTIME,
85 .clockid = CLOCK_BOOTTIME,
86 .get_time = &ktime_get_boottime,
87 },
88 {
89 .index = HRTIMER_BASE_TAI,
90 .clockid = CLOCK_TAI,
91 .get_time = &ktime_get_clocktai,
92 },
93 {
94 .index = HRTIMER_BASE_MONOTONIC_SOFT,
95 .clockid = CLOCK_MONOTONIC,
96 .get_time = &ktime_get,
97 },
98 {
99 .index = HRTIMER_BASE_REALTIME_SOFT,
100 .clockid = CLOCK_REALTIME,
101 .get_time = &ktime_get_real,
102 },
103 {
104 .index = HRTIMER_BASE_BOOTTIME_SOFT,
105 .clockid = CLOCK_BOOTTIME,
106 .get_time = &ktime_get_boottime,
107 },
108 {
109 .index = HRTIMER_BASE_TAI_SOFT,
110 .clockid = CLOCK_TAI,
111 .get_time = &ktime_get_clocktai,
112 },
113 }
114 };
115
116 static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
117 /* Make sure we catch unsupported clockids */
118 [0 ... MAX_CLOCKS - 1] = HRTIMER_MAX_CLOCK_BASES,
119
120 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
121 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
122 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
123 [CLOCK_TAI] = HRTIMER_BASE_TAI,
124 };
125
126 /*
127 * Functions and macros which are different for UP/SMP systems are kept in a
128 * single place
129 */
130 #ifdef CONFIG_SMP
131
132 /*
133 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
134 * such that hrtimer_callback_running() can unconditionally dereference
135 * timer->base->cpu_base
136 */
137 static struct hrtimer_cpu_base migration_cpu_base = {
138 .clock_base = { {
139 .cpu_base = &migration_cpu_base,
140 .seq = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
141 &migration_cpu_base.lock),
142 }, },
143 };
144
145 #define migration_base migration_cpu_base.clock_base[0]
146
is_migration_base(struct hrtimer_clock_base * base)147 static inline bool is_migration_base(struct hrtimer_clock_base *base)
148 {
149 return base == &migration_base;
150 }
151
152 /*
153 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
154 * means that all timers which are tied to this base via timer->base are
155 * locked, and the base itself is locked too.
156 *
157 * So __run_timers/migrate_timers can safely modify all timers which could
158 * be found on the lists/queues.
159 *
160 * When the timer's base is locked, and the timer removed from list, it is
161 * possible to set timer->base = &migration_base and drop the lock: the timer
162 * remains locked.
163 */
164 static
lock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)165 struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
166 unsigned long *flags)
167 __acquires(&timer->base->lock)
168 {
169 struct hrtimer_clock_base *base;
170
171 for (;;) {
172 base = READ_ONCE(timer->base);
173 if (likely(base != &migration_base)) {
174 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
175 if (likely(base == timer->base))
176 return base;
177 /* The timer has migrated to another CPU: */
178 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
179 }
180 cpu_relax();
181 }
182 }
183
184 /*
185 * We do not migrate the timer when it is expiring before the next
186 * event on the target cpu. When high resolution is enabled, we cannot
187 * reprogram the target cpu hardware and we would cause it to fire
188 * late. To keep it simple, we handle the high resolution enabled and
189 * disabled case similar.
190 *
191 * Called with cpu_base->lock of target cpu held.
192 */
193 static int
hrtimer_check_target(struct hrtimer * timer,struct hrtimer_clock_base * new_base)194 hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
195 {
196 ktime_t expires;
197
198 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
199 return expires < new_base->cpu_base->expires_next;
200 }
201
202 static inline
get_target_base(struct hrtimer_cpu_base * base,int pinned)203 struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
204 int pinned)
205 {
206 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
207 if (static_branch_likely(&timers_migration_enabled) && !pinned)
208 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
209 #endif
210 return base;
211 }
212
213 /*
214 * We switch the timer base to a power-optimized selected CPU target,
215 * if:
216 * - NO_HZ_COMMON is enabled
217 * - timer migration is enabled
218 * - the timer callback is not running
219 * - the timer is not the first expiring timer on the new target
220 *
221 * If one of the above requirements is not fulfilled we move the timer
222 * to the current CPU or leave it on the previously assigned CPU if
223 * the timer callback is currently running.
224 */
225 static inline struct hrtimer_clock_base *
switch_hrtimer_base(struct hrtimer * timer,struct hrtimer_clock_base * base,int pinned)226 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
227 int pinned)
228 {
229 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
230 struct hrtimer_clock_base *new_base;
231 int basenum = base->index;
232
233 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
234 new_cpu_base = get_target_base(this_cpu_base, pinned);
235 again:
236 new_base = &new_cpu_base->clock_base[basenum];
237
238 if (base != new_base) {
239 /*
240 * We are trying to move timer to new_base.
241 * However we can't change timer's base while it is running,
242 * so we keep it on the same CPU. No hassle vs. reprogramming
243 * the event source in the high resolution case. The softirq
244 * code will take care of this when the timer function has
245 * completed. There is no conflict as we hold the lock until
246 * the timer is enqueued.
247 */
248 if (unlikely(hrtimer_callback_running(timer)))
249 return base;
250
251 /* See the comment in lock_hrtimer_base() */
252 WRITE_ONCE(timer->base, &migration_base);
253 raw_spin_unlock(&base->cpu_base->lock);
254 raw_spin_lock(&new_base->cpu_base->lock);
255
256 if (new_cpu_base != this_cpu_base &&
257 hrtimer_check_target(timer, new_base)) {
258 raw_spin_unlock(&new_base->cpu_base->lock);
259 raw_spin_lock(&base->cpu_base->lock);
260 new_cpu_base = this_cpu_base;
261 WRITE_ONCE(timer->base, base);
262 goto again;
263 }
264 WRITE_ONCE(timer->base, new_base);
265 } else {
266 if (new_cpu_base != this_cpu_base &&
267 hrtimer_check_target(timer, new_base)) {
268 new_cpu_base = this_cpu_base;
269 goto again;
270 }
271 }
272 return new_base;
273 }
274
275 #else /* CONFIG_SMP */
276
is_migration_base(struct hrtimer_clock_base * base)277 static inline bool is_migration_base(struct hrtimer_clock_base *base)
278 {
279 return false;
280 }
281
282 static inline struct hrtimer_clock_base *
lock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)283 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
284 __acquires(&timer->base->cpu_base->lock)
285 {
286 struct hrtimer_clock_base *base = timer->base;
287
288 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
289
290 return base;
291 }
292
293 # define switch_hrtimer_base(t, b, p) (b)
294
295 #endif /* !CONFIG_SMP */
296
297 /*
298 * Functions for the union type storage format of ktime_t which are
299 * too large for inlining:
300 */
301 #if BITS_PER_LONG < 64
302 /*
303 * Divide a ktime value by a nanosecond value
304 */
__ktime_divns(const ktime_t kt,s64 div)305 s64 __ktime_divns(const ktime_t kt, s64 div)
306 {
307 int sft = 0;
308 s64 dclc;
309 u64 tmp;
310
311 dclc = ktime_to_ns(kt);
312 tmp = dclc < 0 ? -dclc : dclc;
313
314 /* Make sure the divisor is less than 2^32: */
315 while (div >> 32) {
316 sft++;
317 div >>= 1;
318 }
319 tmp >>= sft;
320 do_div(tmp, (u32) div);
321 return dclc < 0 ? -tmp : tmp;
322 }
323 EXPORT_SYMBOL_GPL(__ktime_divns);
324 #endif /* BITS_PER_LONG >= 64 */
325
326 /*
327 * Add two ktime values and do a safety check for overflow:
328 */
ktime_add_safe(const ktime_t lhs,const ktime_t rhs)329 ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
330 {
331 ktime_t res = ktime_add_unsafe(lhs, rhs);
332
333 /*
334 * We use KTIME_SEC_MAX here, the maximum timeout which we can
335 * return to user space in a timespec:
336 */
337 if (res < 0 || res < lhs || res < rhs)
338 res = ktime_set(KTIME_SEC_MAX, 0);
339
340 return res;
341 }
342
343 EXPORT_SYMBOL_GPL(ktime_add_safe);
344
345 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
346
347 static const struct debug_obj_descr hrtimer_debug_descr;
348
hrtimer_debug_hint(void * addr)349 static void *hrtimer_debug_hint(void *addr)
350 {
351 return ((struct hrtimer *) addr)->function;
352 }
353
354 /*
355 * fixup_init is called when:
356 * - an active object is initialized
357 */
hrtimer_fixup_init(void * addr,enum debug_obj_state state)358 static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
359 {
360 struct hrtimer *timer = addr;
361
362 switch (state) {
363 case ODEBUG_STATE_ACTIVE:
364 hrtimer_cancel(timer);
365 debug_object_init(timer, &hrtimer_debug_descr);
366 return true;
367 default:
368 return false;
369 }
370 }
371
372 /*
373 * fixup_activate is called when:
374 * - an active object is activated
375 * - an unknown non-static object is activated
376 */
hrtimer_fixup_activate(void * addr,enum debug_obj_state state)377 static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
378 {
379 switch (state) {
380 case ODEBUG_STATE_ACTIVE:
381 WARN_ON(1);
382 fallthrough;
383 default:
384 return false;
385 }
386 }
387
388 /*
389 * fixup_free is called when:
390 * - an active object is freed
391 */
hrtimer_fixup_free(void * addr,enum debug_obj_state state)392 static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
393 {
394 struct hrtimer *timer = addr;
395
396 switch (state) {
397 case ODEBUG_STATE_ACTIVE:
398 hrtimer_cancel(timer);
399 debug_object_free(timer, &hrtimer_debug_descr);
400 return true;
401 default:
402 return false;
403 }
404 }
405
406 static const struct debug_obj_descr hrtimer_debug_descr = {
407 .name = "hrtimer",
408 .debug_hint = hrtimer_debug_hint,
409 .fixup_init = hrtimer_fixup_init,
410 .fixup_activate = hrtimer_fixup_activate,
411 .fixup_free = hrtimer_fixup_free,
412 };
413
debug_hrtimer_init(struct hrtimer * timer)414 static inline void debug_hrtimer_init(struct hrtimer *timer)
415 {
416 debug_object_init(timer, &hrtimer_debug_descr);
417 }
418
debug_hrtimer_activate(struct hrtimer * timer,enum hrtimer_mode mode)419 static inline void debug_hrtimer_activate(struct hrtimer *timer,
420 enum hrtimer_mode mode)
421 {
422 debug_object_activate(timer, &hrtimer_debug_descr);
423 }
424
debug_hrtimer_deactivate(struct hrtimer * timer)425 static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
426 {
427 debug_object_deactivate(timer, &hrtimer_debug_descr);
428 }
429
430 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
431 enum hrtimer_mode mode);
432
hrtimer_init_on_stack(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)433 void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
434 enum hrtimer_mode mode)
435 {
436 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
437 __hrtimer_init(timer, clock_id, mode);
438 }
439 EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
440
441 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
442 clockid_t clock_id, enum hrtimer_mode mode);
443
hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)444 void hrtimer_init_sleeper_on_stack(struct hrtimer_sleeper *sl,
445 clockid_t clock_id, enum hrtimer_mode mode)
446 {
447 debug_object_init_on_stack(&sl->timer, &hrtimer_debug_descr);
448 __hrtimer_init_sleeper(sl, clock_id, mode);
449 }
450 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper_on_stack);
451
destroy_hrtimer_on_stack(struct hrtimer * timer)452 void destroy_hrtimer_on_stack(struct hrtimer *timer)
453 {
454 debug_object_free(timer, &hrtimer_debug_descr);
455 }
456 EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
457
458 #else
459
debug_hrtimer_init(struct hrtimer * timer)460 static inline void debug_hrtimer_init(struct hrtimer *timer) { }
debug_hrtimer_activate(struct hrtimer * timer,enum hrtimer_mode mode)461 static inline void debug_hrtimer_activate(struct hrtimer *timer,
462 enum hrtimer_mode mode) { }
debug_hrtimer_deactivate(struct hrtimer * timer)463 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
464 #endif
465
466 static inline void
debug_init(struct hrtimer * timer,clockid_t clockid,enum hrtimer_mode mode)467 debug_init(struct hrtimer *timer, clockid_t clockid,
468 enum hrtimer_mode mode)
469 {
470 debug_hrtimer_init(timer);
471 trace_hrtimer_init(timer, clockid, mode);
472 }
473
debug_activate(struct hrtimer * timer,enum hrtimer_mode mode)474 static inline void debug_activate(struct hrtimer *timer,
475 enum hrtimer_mode mode)
476 {
477 debug_hrtimer_activate(timer, mode);
478 trace_hrtimer_start(timer, mode);
479 }
480
debug_deactivate(struct hrtimer * timer)481 static inline void debug_deactivate(struct hrtimer *timer)
482 {
483 debug_hrtimer_deactivate(timer);
484 trace_hrtimer_cancel(timer);
485 }
486
487 static struct hrtimer_clock_base *
__next_base(struct hrtimer_cpu_base * cpu_base,unsigned int * active)488 __next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
489 {
490 unsigned int idx;
491
492 if (!*active)
493 return NULL;
494
495 idx = __ffs(*active);
496 *active &= ~(1U << idx);
497
498 return &cpu_base->clock_base[idx];
499 }
500
501 #define for_each_active_base(base, cpu_base, active) \
502 while ((base = __next_base((cpu_base), &(active))))
503
__hrtimer_next_event_base(struct hrtimer_cpu_base * cpu_base,const struct hrtimer * exclude,unsigned int active,ktime_t expires_next)504 static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
505 const struct hrtimer *exclude,
506 unsigned int active,
507 ktime_t expires_next)
508 {
509 struct hrtimer_clock_base *base;
510 ktime_t expires;
511
512 for_each_active_base(base, cpu_base, active) {
513 struct timerqueue_node *next;
514 struct hrtimer *timer;
515
516 next = timerqueue_getnext(&base->active);
517 timer = container_of(next, struct hrtimer, node);
518 if (timer == exclude) {
519 /* Get to the next timer in the queue. */
520 next = timerqueue_iterate_next(next);
521 if (!next)
522 continue;
523
524 timer = container_of(next, struct hrtimer, node);
525 }
526 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
527 if (expires < expires_next) {
528 expires_next = expires;
529
530 /* Skip cpu_base update if a timer is being excluded. */
531 if (exclude)
532 continue;
533
534 if (timer->is_soft)
535 cpu_base->softirq_next_timer = timer;
536 else
537 cpu_base->next_timer = timer;
538 }
539 }
540 /*
541 * clock_was_set() might have changed base->offset of any of
542 * the clock bases so the result might be negative. Fix it up
543 * to prevent a false positive in clockevents_program_event().
544 */
545 if (expires_next < 0)
546 expires_next = 0;
547 return expires_next;
548 }
549
550 /*
551 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
552 * but does not set cpu_base::*expires_next, that is done by
553 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
554 * cpu_base::*expires_next right away, reprogramming logic would no longer
555 * work.
556 *
557 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
558 * those timers will get run whenever the softirq gets handled, at the end of
559 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
560 *
561 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
562 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
563 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
564 *
565 * @active_mask must be one of:
566 * - HRTIMER_ACTIVE_ALL,
567 * - HRTIMER_ACTIVE_SOFT, or
568 * - HRTIMER_ACTIVE_HARD.
569 */
570 static ktime_t
__hrtimer_get_next_event(struct hrtimer_cpu_base * cpu_base,unsigned int active_mask)571 __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
572 {
573 unsigned int active;
574 struct hrtimer *next_timer = NULL;
575 ktime_t expires_next = KTIME_MAX;
576
577 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
578 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
579 cpu_base->softirq_next_timer = NULL;
580 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
581 active, KTIME_MAX);
582
583 next_timer = cpu_base->softirq_next_timer;
584 }
585
586 if (active_mask & HRTIMER_ACTIVE_HARD) {
587 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
588 cpu_base->next_timer = next_timer;
589 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
590 expires_next);
591 }
592
593 return expires_next;
594 }
595
hrtimer_update_next_event(struct hrtimer_cpu_base * cpu_base)596 static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
597 {
598 ktime_t expires_next, soft = KTIME_MAX;
599
600 /*
601 * If the soft interrupt has already been activated, ignore the
602 * soft bases. They will be handled in the already raised soft
603 * interrupt.
604 */
605 if (!cpu_base->softirq_activated) {
606 soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
607 /*
608 * Update the soft expiry time. clock_settime() might have
609 * affected it.
610 */
611 cpu_base->softirq_expires_next = soft;
612 }
613
614 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
615 /*
616 * If a softirq timer is expiring first, update cpu_base->next_timer
617 * and program the hardware with the soft expiry time.
618 */
619 if (expires_next > soft) {
620 cpu_base->next_timer = cpu_base->softirq_next_timer;
621 expires_next = soft;
622 }
623
624 return expires_next;
625 }
626
hrtimer_update_base(struct hrtimer_cpu_base * base)627 static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
628 {
629 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
630 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
631 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
632
633 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
634 offs_real, offs_boot, offs_tai);
635
636 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
637 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
638 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
639
640 return now;
641 }
642
643 /*
644 * Is the high resolution mode active ?
645 */
__hrtimer_hres_active(struct hrtimer_cpu_base * cpu_base)646 static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
647 {
648 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
649 cpu_base->hres_active : 0;
650 }
651
hrtimer_hres_active(void)652 static inline int hrtimer_hres_active(void)
653 {
654 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
655 }
656
__hrtimer_reprogram(struct hrtimer_cpu_base * cpu_base,struct hrtimer * next_timer,ktime_t expires_next)657 static void __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base,
658 struct hrtimer *next_timer,
659 ktime_t expires_next)
660 {
661 cpu_base->expires_next = expires_next;
662
663 /*
664 * If hres is not active, hardware does not have to be
665 * reprogrammed yet.
666 *
667 * If a hang was detected in the last timer interrupt then we
668 * leave the hang delay active in the hardware. We want the
669 * system to make progress. That also prevents the following
670 * scenario:
671 * T1 expires 50ms from now
672 * T2 expires 5s from now
673 *
674 * T1 is removed, so this code is called and would reprogram
675 * the hardware to 5s from now. Any hrtimer_start after that
676 * will not reprogram the hardware due to hang_detected being
677 * set. So we'd effectively block all timers until the T2 event
678 * fires.
679 */
680 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
681 return;
682
683 tick_program_event(expires_next, 1);
684 }
685
686 /*
687 * Reprogram the event source with checking both queues for the
688 * next event
689 * Called with interrupts disabled and base->lock held
690 */
691 static void
hrtimer_force_reprogram(struct hrtimer_cpu_base * cpu_base,int skip_equal)692 hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
693 {
694 ktime_t expires_next;
695
696 expires_next = hrtimer_update_next_event(cpu_base);
697
698 if (skip_equal && expires_next == cpu_base->expires_next)
699 return;
700
701 __hrtimer_reprogram(cpu_base, cpu_base->next_timer, expires_next);
702 }
703
704 /* High resolution timer related functions */
705 #ifdef CONFIG_HIGH_RES_TIMERS
706
707 /*
708 * High resolution timer enabled ?
709 */
710 static bool hrtimer_hres_enabled __read_mostly = true;
711 unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
712 EXPORT_SYMBOL_GPL(hrtimer_resolution);
713
714 /*
715 * Enable / Disable high resolution mode
716 */
setup_hrtimer_hres(char * str)717 static int __init setup_hrtimer_hres(char *str)
718 {
719 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
720 }
721
722 __setup("highres=", setup_hrtimer_hres);
723
724 /*
725 * hrtimer_high_res_enabled - query, if the highres mode is enabled
726 */
hrtimer_is_hres_enabled(void)727 static inline int hrtimer_is_hres_enabled(void)
728 {
729 return hrtimer_hres_enabled;
730 }
731
732 static void retrigger_next_event(void *arg);
733
734 /*
735 * Switch to high resolution mode
736 */
hrtimer_switch_to_hres(void)737 static void hrtimer_switch_to_hres(void)
738 {
739 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
740
741 if (tick_init_highres()) {
742 pr_warn("Could not switch to high resolution mode on CPU %u\n",
743 base->cpu);
744 return;
745 }
746 base->hres_active = 1;
747 hrtimer_resolution = HIGH_RES_NSEC;
748
749 tick_setup_sched_timer();
750 /* "Retrigger" the interrupt to get things going */
751 retrigger_next_event(NULL);
752 }
753
754 #else
755
hrtimer_is_hres_enabled(void)756 static inline int hrtimer_is_hres_enabled(void) { return 0; }
hrtimer_switch_to_hres(void)757 static inline void hrtimer_switch_to_hres(void) { }
758
759 #endif /* CONFIG_HIGH_RES_TIMERS */
760 /*
761 * Retrigger next event is called after clock was set with interrupts
762 * disabled through an SMP function call or directly from low level
763 * resume code.
764 *
765 * This is only invoked when:
766 * - CONFIG_HIGH_RES_TIMERS is enabled.
767 * - CONFIG_NOHZ_COMMON is enabled
768 *
769 * For the other cases this function is empty and because the call sites
770 * are optimized out it vanishes as well, i.e. no need for lots of
771 * #ifdeffery.
772 */
retrigger_next_event(void * arg)773 static void retrigger_next_event(void *arg)
774 {
775 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
776
777 /*
778 * When high resolution mode or nohz is active, then the offsets of
779 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
780 * next tick will take care of that.
781 *
782 * If high resolution mode is active then the next expiring timer
783 * must be reevaluated and the clock event device reprogrammed if
784 * necessary.
785 *
786 * In the NOHZ case the update of the offset and the reevaluation
787 * of the next expiring timer is enough. The return from the SMP
788 * function call will take care of the reprogramming in case the
789 * CPU was in a NOHZ idle sleep.
790 */
791 if (!__hrtimer_hres_active(base) && !tick_nohz_active)
792 return;
793
794 raw_spin_lock(&base->lock);
795 hrtimer_update_base(base);
796 if (__hrtimer_hres_active(base))
797 hrtimer_force_reprogram(base, 0);
798 else
799 hrtimer_update_next_event(base);
800 raw_spin_unlock(&base->lock);
801 }
802
803 /*
804 * When a timer is enqueued and expires earlier than the already enqueued
805 * timers, we have to check, whether it expires earlier than the timer for
806 * which the clock event device was armed.
807 *
808 * Called with interrupts disabled and base->cpu_base.lock held
809 */
hrtimer_reprogram(struct hrtimer * timer,bool reprogram)810 static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
811 {
812 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
813 struct hrtimer_clock_base *base = timer->base;
814 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
815
816 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
817
818 /*
819 * CLOCK_REALTIME timer might be requested with an absolute
820 * expiry time which is less than base->offset. Set it to 0.
821 */
822 if (expires < 0)
823 expires = 0;
824
825 if (timer->is_soft) {
826 /*
827 * soft hrtimer could be started on a remote CPU. In this
828 * case softirq_expires_next needs to be updated on the
829 * remote CPU. The soft hrtimer will not expire before the
830 * first hard hrtimer on the remote CPU -
831 * hrtimer_check_target() prevents this case.
832 */
833 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
834
835 if (timer_cpu_base->softirq_activated)
836 return;
837
838 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
839 return;
840
841 timer_cpu_base->softirq_next_timer = timer;
842 timer_cpu_base->softirq_expires_next = expires;
843
844 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
845 !reprogram)
846 return;
847 }
848
849 /*
850 * If the timer is not on the current cpu, we cannot reprogram
851 * the other cpus clock event device.
852 */
853 if (base->cpu_base != cpu_base)
854 return;
855
856 if (expires >= cpu_base->expires_next)
857 return;
858
859 /*
860 * If the hrtimer interrupt is running, then it will reevaluate the
861 * clock bases and reprogram the clock event device.
862 */
863 if (cpu_base->in_hrtirq)
864 return;
865
866 cpu_base->next_timer = timer;
867
868 __hrtimer_reprogram(cpu_base, timer, expires);
869 }
870
update_needs_ipi(struct hrtimer_cpu_base * cpu_base,unsigned int active)871 static bool update_needs_ipi(struct hrtimer_cpu_base *cpu_base,
872 unsigned int active)
873 {
874 struct hrtimer_clock_base *base;
875 unsigned int seq;
876 ktime_t expires;
877
878 /*
879 * Update the base offsets unconditionally so the following
880 * checks whether the SMP function call is required works.
881 *
882 * The update is safe even when the remote CPU is in the hrtimer
883 * interrupt or the hrtimer soft interrupt and expiring affected
884 * bases. Either it will see the update before handling a base or
885 * it will see it when it finishes the processing and reevaluates
886 * the next expiring timer.
887 */
888 seq = cpu_base->clock_was_set_seq;
889 hrtimer_update_base(cpu_base);
890
891 /*
892 * If the sequence did not change over the update then the
893 * remote CPU already handled it.
894 */
895 if (seq == cpu_base->clock_was_set_seq)
896 return false;
897
898 /*
899 * If the remote CPU is currently handling an hrtimer interrupt, it
900 * will reevaluate the first expiring timer of all clock bases
901 * before reprogramming. Nothing to do here.
902 */
903 if (cpu_base->in_hrtirq)
904 return false;
905
906 /*
907 * Walk the affected clock bases and check whether the first expiring
908 * timer in a clock base is moving ahead of the first expiring timer of
909 * @cpu_base. If so, the IPI must be invoked because per CPU clock
910 * event devices cannot be remotely reprogrammed.
911 */
912 active &= cpu_base->active_bases;
913
914 for_each_active_base(base, cpu_base, active) {
915 struct timerqueue_node *next;
916
917 next = timerqueue_getnext(&base->active);
918 expires = ktime_sub(next->expires, base->offset);
919 if (expires < cpu_base->expires_next)
920 return true;
921
922 /* Extra check for softirq clock bases */
923 if (base->clockid < HRTIMER_BASE_MONOTONIC_SOFT)
924 continue;
925 if (cpu_base->softirq_activated)
926 continue;
927 if (expires < cpu_base->softirq_expires_next)
928 return true;
929 }
930 return false;
931 }
932
933 /*
934 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
935 * CLOCK_BOOTTIME (for late sleep time injection).
936 *
937 * This requires to update the offsets for these clocks
938 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
939 * also requires to eventually reprogram the per CPU clock event devices
940 * when the change moves an affected timer ahead of the first expiring
941 * timer on that CPU. Obviously remote per CPU clock event devices cannot
942 * be reprogrammed. The other reason why an IPI has to be sent is when the
943 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
944 * in the tick, which obviously might be stopped, so this has to bring out
945 * the remote CPU which might sleep in idle to get this sorted.
946 */
clock_was_set(unsigned int bases)947 void clock_was_set(unsigned int bases)
948 {
949 struct hrtimer_cpu_base *cpu_base = raw_cpu_ptr(&hrtimer_bases);
950 cpumask_var_t mask;
951 int cpu;
952
953 if (!__hrtimer_hres_active(cpu_base) && !tick_nohz_active)
954 goto out_timerfd;
955
956 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
957 on_each_cpu(retrigger_next_event, NULL, 1);
958 goto out_timerfd;
959 }
960
961 /* Avoid interrupting CPUs if possible */
962 cpus_read_lock();
963 for_each_online_cpu(cpu) {
964 unsigned long flags;
965
966 cpu_base = &per_cpu(hrtimer_bases, cpu);
967 raw_spin_lock_irqsave(&cpu_base->lock, flags);
968
969 if (update_needs_ipi(cpu_base, bases))
970 cpumask_set_cpu(cpu, mask);
971
972 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
973 }
974
975 preempt_disable();
976 smp_call_function_many(mask, retrigger_next_event, NULL, 1);
977 preempt_enable();
978 cpus_read_unlock();
979 free_cpumask_var(mask);
980
981 out_timerfd:
982 timerfd_clock_was_set();
983 }
984
clock_was_set_work(struct work_struct * work)985 static void clock_was_set_work(struct work_struct *work)
986 {
987 clock_was_set(CLOCK_SET_WALL);
988 }
989
990 static DECLARE_WORK(hrtimer_work, clock_was_set_work);
991
992 /*
993 * Called from timekeeping code to reprogram the hrtimer interrupt device
994 * on all cpus and to notify timerfd.
995 */
clock_was_set_delayed(void)996 void clock_was_set_delayed(void)
997 {
998 schedule_work(&hrtimer_work);
999 }
1000
1001 /*
1002 * Called during resume either directly from via timekeeping_resume()
1003 * or in the case of s2idle from tick_unfreeze() to ensure that the
1004 * hrtimers are up to date.
1005 */
hrtimers_resume_local(void)1006 void hrtimers_resume_local(void)
1007 {
1008 lockdep_assert_irqs_disabled();
1009 /* Retrigger on the local CPU */
1010 retrigger_next_event(NULL);
1011 }
1012
1013 /*
1014 * Counterpart to lock_hrtimer_base above:
1015 */
1016 static inline
unlock_hrtimer_base(const struct hrtimer * timer,unsigned long * flags)1017 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
1018 __releases(&timer->base->cpu_base->lock)
1019 {
1020 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
1021 }
1022
1023 /**
1024 * hrtimer_forward - forward the timer expiry
1025 * @timer: hrtimer to forward
1026 * @now: forward past this time
1027 * @interval: the interval to forward
1028 *
1029 * Forward the timer expiry so it will expire in the future.
1030 * Returns the number of overruns.
1031 *
1032 * Can be safely called from the callback function of @timer. If
1033 * called from other contexts @timer must neither be enqueued nor
1034 * running the callback and the caller needs to take care of
1035 * serialization.
1036 *
1037 * Note: This only updates the timer expiry value and does not requeue
1038 * the timer.
1039 */
hrtimer_forward(struct hrtimer * timer,ktime_t now,ktime_t interval)1040 u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
1041 {
1042 u64 orun = 1;
1043 ktime_t delta;
1044
1045 delta = ktime_sub(now, hrtimer_get_expires(timer));
1046
1047 if (delta < 0)
1048 return 0;
1049
1050 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
1051 return 0;
1052
1053 if (interval < hrtimer_resolution)
1054 interval = hrtimer_resolution;
1055
1056 if (unlikely(delta >= interval)) {
1057 s64 incr = ktime_to_ns(interval);
1058
1059 orun = ktime_divns(delta, incr);
1060 hrtimer_add_expires_ns(timer, incr * orun);
1061 if (hrtimer_get_expires_tv64(timer) > now)
1062 return orun;
1063 /*
1064 * This (and the ktime_add() below) is the
1065 * correction for exact:
1066 */
1067 orun++;
1068 }
1069 hrtimer_add_expires(timer, interval);
1070
1071 return orun;
1072 }
1073 EXPORT_SYMBOL_GPL(hrtimer_forward);
1074
1075 /*
1076 * enqueue_hrtimer - internal function to (re)start a timer
1077 *
1078 * The timer is inserted in expiry order. Insertion into the
1079 * red black tree is O(log(n)). Must hold the base lock.
1080 *
1081 * Returns 1 when the new timer is the leftmost timer in the tree.
1082 */
enqueue_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,enum hrtimer_mode mode)1083 static int enqueue_hrtimer(struct hrtimer *timer,
1084 struct hrtimer_clock_base *base,
1085 enum hrtimer_mode mode)
1086 {
1087 debug_activate(timer, mode);
1088 WARN_ON_ONCE(!base->cpu_base->online);
1089
1090 base->cpu_base->active_bases |= 1 << base->index;
1091
1092 /* Pairs with the lockless read in hrtimer_is_queued() */
1093 WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1094
1095 return timerqueue_add(&base->active, &timer->node);
1096 }
1097
1098 /*
1099 * __remove_hrtimer - internal function to remove a timer
1100 *
1101 * Caller must hold the base lock.
1102 *
1103 * High resolution timer mode reprograms the clock event device when the
1104 * timer is the one which expires next. The caller can disable this by setting
1105 * reprogram to zero. This is useful, when the context does a reprogramming
1106 * anyway (e.g. timer interrupt)
1107 */
__remove_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,u8 newstate,int reprogram)1108 static void __remove_hrtimer(struct hrtimer *timer,
1109 struct hrtimer_clock_base *base,
1110 u8 newstate, int reprogram)
1111 {
1112 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1113 u8 state = timer->state;
1114
1115 /* Pairs with the lockless read in hrtimer_is_queued() */
1116 WRITE_ONCE(timer->state, newstate);
1117 if (!(state & HRTIMER_STATE_ENQUEUED))
1118 return;
1119
1120 if (!timerqueue_del(&base->active, &timer->node))
1121 cpu_base->active_bases &= ~(1 << base->index);
1122
1123 /*
1124 * Note: If reprogram is false we do not update
1125 * cpu_base->next_timer. This happens when we remove the first
1126 * timer on a remote cpu. No harm as we never dereference
1127 * cpu_base->next_timer. So the worst thing what can happen is
1128 * an superfluous call to hrtimer_force_reprogram() on the
1129 * remote cpu later on if the same timer gets enqueued again.
1130 */
1131 if (reprogram && timer == cpu_base->next_timer)
1132 hrtimer_force_reprogram(cpu_base, 1);
1133 }
1134
1135 /*
1136 * remove hrtimer, called with base lock held
1137 */
1138 static inline int
remove_hrtimer(struct hrtimer * timer,struct hrtimer_clock_base * base,bool restart,bool keep_local)1139 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1140 bool restart, bool keep_local)
1141 {
1142 u8 state = timer->state;
1143
1144 if (state & HRTIMER_STATE_ENQUEUED) {
1145 bool reprogram;
1146
1147 /*
1148 * Remove the timer and force reprogramming when high
1149 * resolution mode is active and the timer is on the current
1150 * CPU. If we remove a timer on another CPU, reprogramming is
1151 * skipped. The interrupt event on this CPU is fired and
1152 * reprogramming happens in the interrupt handler. This is a
1153 * rare case and less expensive than a smp call.
1154 */
1155 debug_deactivate(timer);
1156 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1157
1158 /*
1159 * If the timer is not restarted then reprogramming is
1160 * required if the timer is local. If it is local and about
1161 * to be restarted, avoid programming it twice (on removal
1162 * and a moment later when it's requeued).
1163 */
1164 if (!restart)
1165 state = HRTIMER_STATE_INACTIVE;
1166 else
1167 reprogram &= !keep_local;
1168
1169 __remove_hrtimer(timer, base, state, reprogram);
1170 return 1;
1171 }
1172 return 0;
1173 }
1174
hrtimer_update_lowres(struct hrtimer * timer,ktime_t tim,const enum hrtimer_mode mode)1175 static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1176 const enum hrtimer_mode mode)
1177 {
1178 #ifdef CONFIG_TIME_LOW_RES
1179 /*
1180 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1181 * granular time values. For relative timers we add hrtimer_resolution
1182 * (i.e. one jiffie) to prevent short timeouts.
1183 */
1184 timer->is_rel = mode & HRTIMER_MODE_REL;
1185 if (timer->is_rel)
1186 tim = ktime_add_safe(tim, hrtimer_resolution);
1187 #endif
1188 return tim;
1189 }
1190
1191 static void
hrtimer_update_softirq_timer(struct hrtimer_cpu_base * cpu_base,bool reprogram)1192 hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1193 {
1194 ktime_t expires;
1195
1196 /*
1197 * Find the next SOFT expiration.
1198 */
1199 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1200
1201 /*
1202 * reprogramming needs to be triggered, even if the next soft
1203 * hrtimer expires at the same time than the next hard
1204 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1205 */
1206 if (expires == KTIME_MAX)
1207 return;
1208
1209 /*
1210 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1211 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1212 */
1213 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1214 }
1215
__hrtimer_start_range_ns(struct hrtimer * timer,ktime_t tim,u64 delta_ns,const enum hrtimer_mode mode,struct hrtimer_clock_base * base)1216 static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1217 u64 delta_ns, const enum hrtimer_mode mode,
1218 struct hrtimer_clock_base *base)
1219 {
1220 struct hrtimer_clock_base *new_base;
1221 bool force_local, first;
1222
1223 /*
1224 * If the timer is on the local cpu base and is the first expiring
1225 * timer then this might end up reprogramming the hardware twice
1226 * (on removal and on enqueue). To avoid that by prevent the
1227 * reprogram on removal, keep the timer local to the current CPU
1228 * and enforce reprogramming after it is queued no matter whether
1229 * it is the new first expiring timer again or not.
1230 */
1231 force_local = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1232 force_local &= base->cpu_base->next_timer == timer;
1233
1234 /*
1235 * Remove an active timer from the queue. In case it is not queued
1236 * on the current CPU, make sure that remove_hrtimer() updates the
1237 * remote data correctly.
1238 *
1239 * If it's on the current CPU and the first expiring timer, then
1240 * skip reprogramming, keep the timer local and enforce
1241 * reprogramming later if it was the first expiring timer. This
1242 * avoids programming the underlying clock event twice (once at
1243 * removal and once after enqueue).
1244 */
1245 remove_hrtimer(timer, base, true, force_local);
1246
1247 if (mode & HRTIMER_MODE_REL)
1248 tim = ktime_add_safe(tim, base->get_time());
1249
1250 tim = hrtimer_update_lowres(timer, tim, mode);
1251
1252 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1253
1254 /* Switch the timer base, if necessary: */
1255 if (!force_local) {
1256 new_base = switch_hrtimer_base(timer, base,
1257 mode & HRTIMER_MODE_PINNED);
1258 } else {
1259 new_base = base;
1260 }
1261
1262 first = enqueue_hrtimer(timer, new_base, mode);
1263 if (!force_local)
1264 return first;
1265
1266 /*
1267 * Timer was forced to stay on the current CPU to avoid
1268 * reprogramming on removal and enqueue. Force reprogram the
1269 * hardware by evaluating the new first expiring timer.
1270 */
1271 hrtimer_force_reprogram(new_base->cpu_base, 1);
1272 return 0;
1273 }
1274
1275 /**
1276 * hrtimer_start_range_ns - (re)start an hrtimer
1277 * @timer: the timer to be added
1278 * @tim: expiry time
1279 * @delta_ns: "slack" range for the timer
1280 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1281 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1282 * softirq based mode is considered for debug purpose only!
1283 */
hrtimer_start_range_ns(struct hrtimer * timer,ktime_t tim,u64 delta_ns,const enum hrtimer_mode mode)1284 void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1285 u64 delta_ns, const enum hrtimer_mode mode)
1286 {
1287 struct hrtimer_clock_base *base;
1288 unsigned long flags;
1289
1290 /*
1291 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1292 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1293 * expiry mode because unmarked timers are moved to softirq expiry.
1294 */
1295 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1296 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1297 else
1298 WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1299
1300 base = lock_hrtimer_base(timer, &flags);
1301
1302 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1303 hrtimer_reprogram(timer, true);
1304
1305 unlock_hrtimer_base(timer, &flags);
1306 }
1307 EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1308
1309 /**
1310 * hrtimer_try_to_cancel - try to deactivate a timer
1311 * @timer: hrtimer to stop
1312 *
1313 * Returns:
1314 *
1315 * * 0 when the timer was not active
1316 * * 1 when the timer was active
1317 * * -1 when the timer is currently executing the callback function and
1318 * cannot be stopped
1319 */
hrtimer_try_to_cancel(struct hrtimer * timer)1320 int hrtimer_try_to_cancel(struct hrtimer *timer)
1321 {
1322 struct hrtimer_clock_base *base;
1323 unsigned long flags;
1324 int ret = -1;
1325
1326 /*
1327 * Check lockless first. If the timer is not active (neither
1328 * enqueued nor running the callback, nothing to do here. The
1329 * base lock does not serialize against a concurrent enqueue,
1330 * so we can avoid taking it.
1331 */
1332 if (!hrtimer_active(timer))
1333 return 0;
1334
1335 base = lock_hrtimer_base(timer, &flags);
1336
1337 if (!hrtimer_callback_running(timer))
1338 ret = remove_hrtimer(timer, base, false, false);
1339
1340 unlock_hrtimer_base(timer, &flags);
1341
1342 return ret;
1343
1344 }
1345 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1346
1347 #ifdef CONFIG_PREEMPT_RT
hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base * base)1348 static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1349 {
1350 spin_lock_init(&base->softirq_expiry_lock);
1351 }
1352
hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base * base)1353 static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1354 {
1355 spin_lock(&base->softirq_expiry_lock);
1356 }
1357
hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base * base)1358 static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1359 {
1360 spin_unlock(&base->softirq_expiry_lock);
1361 }
1362
1363 /*
1364 * The counterpart to hrtimer_cancel_wait_running().
1365 *
1366 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1367 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1368 * allows the waiter to acquire the lock and make progress.
1369 */
hrtimer_sync_wait_running(struct hrtimer_cpu_base * cpu_base,unsigned long flags)1370 static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1371 unsigned long flags)
1372 {
1373 if (atomic_read(&cpu_base->timer_waiters)) {
1374 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1375 spin_unlock(&cpu_base->softirq_expiry_lock);
1376 spin_lock(&cpu_base->softirq_expiry_lock);
1377 raw_spin_lock_irq(&cpu_base->lock);
1378 }
1379 }
1380
1381 /*
1382 * This function is called on PREEMPT_RT kernels when the fast path
1383 * deletion of a timer failed because the timer callback function was
1384 * running.
1385 *
1386 * This prevents priority inversion: if the soft irq thread is preempted
1387 * in the middle of a timer callback, then calling del_timer_sync() can
1388 * lead to two issues:
1389 *
1390 * - If the caller is on a remote CPU then it has to spin wait for the timer
1391 * handler to complete. This can result in unbound priority inversion.
1392 *
1393 * - If the caller originates from the task which preempted the timer
1394 * handler on the same CPU, then spin waiting for the timer handler to
1395 * complete is never going to end.
1396 */
hrtimer_cancel_wait_running(const struct hrtimer * timer)1397 void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1398 {
1399 /* Lockless read. Prevent the compiler from reloading it below */
1400 struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1401
1402 /*
1403 * Just relax if the timer expires in hard interrupt context or if
1404 * it is currently on the migration base.
1405 */
1406 if (!timer->is_soft || is_migration_base(base)) {
1407 cpu_relax();
1408 return;
1409 }
1410
1411 /*
1412 * Mark the base as contended and grab the expiry lock, which is
1413 * held by the softirq across the timer callback. Drop the lock
1414 * immediately so the softirq can expire the next timer. In theory
1415 * the timer could already be running again, but that's more than
1416 * unlikely and just causes another wait loop.
1417 */
1418 atomic_inc(&base->cpu_base->timer_waiters);
1419 spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1420 atomic_dec(&base->cpu_base->timer_waiters);
1421 spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1422 }
1423 #else
1424 static inline void
hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base * base)1425 hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1426 static inline void
hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base * base)1427 hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1428 static inline void
hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base * base)1429 hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
hrtimer_sync_wait_running(struct hrtimer_cpu_base * base,unsigned long flags)1430 static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1431 unsigned long flags) { }
1432 #endif
1433
1434 /**
1435 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1436 * @timer: the timer to be cancelled
1437 *
1438 * Returns:
1439 * 0 when the timer was not active
1440 * 1 when the timer was active
1441 */
hrtimer_cancel(struct hrtimer * timer)1442 int hrtimer_cancel(struct hrtimer *timer)
1443 {
1444 int ret;
1445
1446 do {
1447 ret = hrtimer_try_to_cancel(timer);
1448
1449 if (ret < 0)
1450 hrtimer_cancel_wait_running(timer);
1451 } while (ret < 0);
1452 return ret;
1453 }
1454 EXPORT_SYMBOL_GPL(hrtimer_cancel);
1455
1456 /**
1457 * __hrtimer_get_remaining - get remaining time for the timer
1458 * @timer: the timer to read
1459 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1460 */
__hrtimer_get_remaining(const struct hrtimer * timer,bool adjust)1461 ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1462 {
1463 unsigned long flags;
1464 ktime_t rem;
1465
1466 lock_hrtimer_base(timer, &flags);
1467 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1468 rem = hrtimer_expires_remaining_adjusted(timer);
1469 else
1470 rem = hrtimer_expires_remaining(timer);
1471 unlock_hrtimer_base(timer, &flags);
1472
1473 return rem;
1474 }
1475 EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1476
1477 #ifdef CONFIG_NO_HZ_COMMON
1478 /**
1479 * hrtimer_get_next_event - get the time until next expiry event
1480 *
1481 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1482 */
hrtimer_get_next_event(void)1483 u64 hrtimer_get_next_event(void)
1484 {
1485 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1486 u64 expires = KTIME_MAX;
1487 unsigned long flags;
1488
1489 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1490
1491 if (!__hrtimer_hres_active(cpu_base))
1492 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1493
1494 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1495
1496 return expires;
1497 }
1498
1499 /**
1500 * hrtimer_next_event_without - time until next expiry event w/o one timer
1501 * @exclude: timer to exclude
1502 *
1503 * Returns the next expiry time over all timers except for the @exclude one or
1504 * KTIME_MAX if none of them is pending.
1505 */
hrtimer_next_event_without(const struct hrtimer * exclude)1506 u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1507 {
1508 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1509 u64 expires = KTIME_MAX;
1510 unsigned long flags;
1511
1512 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1513
1514 if (__hrtimer_hres_active(cpu_base)) {
1515 unsigned int active;
1516
1517 if (!cpu_base->softirq_activated) {
1518 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1519 expires = __hrtimer_next_event_base(cpu_base, exclude,
1520 active, KTIME_MAX);
1521 }
1522 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1523 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1524 expires);
1525 }
1526
1527 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1528
1529 return expires;
1530 }
1531 #endif
1532
hrtimer_clockid_to_base(clockid_t clock_id)1533 static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1534 {
1535 if (likely(clock_id < MAX_CLOCKS)) {
1536 int base = hrtimer_clock_to_base_table[clock_id];
1537
1538 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1539 return base;
1540 }
1541 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1542 return HRTIMER_BASE_MONOTONIC;
1543 }
1544
__hrtimer_init(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)1545 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1546 enum hrtimer_mode mode)
1547 {
1548 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1549 struct hrtimer_cpu_base *cpu_base;
1550 int base;
1551
1552 /*
1553 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1554 * marked for hard interrupt expiry mode are moved into soft
1555 * interrupt context for latency reasons and because the callbacks
1556 * can invoke functions which might sleep on RT, e.g. spin_lock().
1557 */
1558 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1559 softtimer = true;
1560
1561 memset(timer, 0, sizeof(struct hrtimer));
1562
1563 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1564
1565 /*
1566 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1567 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1568 * ensure POSIX compliance.
1569 */
1570 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1571 clock_id = CLOCK_MONOTONIC;
1572
1573 base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1574 base += hrtimer_clockid_to_base(clock_id);
1575 timer->is_soft = softtimer;
1576 timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1577 timer->base = &cpu_base->clock_base[base];
1578 timerqueue_init(&timer->node);
1579 }
1580
1581 /**
1582 * hrtimer_init - initialize a timer to the given clock
1583 * @timer: the timer to be initialized
1584 * @clock_id: the clock to be used
1585 * @mode: The modes which are relevant for initialization:
1586 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1587 * HRTIMER_MODE_REL_SOFT
1588 *
1589 * The PINNED variants of the above can be handed in,
1590 * but the PINNED bit is ignored as pinning happens
1591 * when the hrtimer is started
1592 */
hrtimer_init(struct hrtimer * timer,clockid_t clock_id,enum hrtimer_mode mode)1593 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1594 enum hrtimer_mode mode)
1595 {
1596 debug_init(timer, clock_id, mode);
1597 __hrtimer_init(timer, clock_id, mode);
1598 }
1599 EXPORT_SYMBOL_GPL(hrtimer_init);
1600
1601 /*
1602 * A timer is active, when it is enqueued into the rbtree or the
1603 * callback function is running or it's in the state of being migrated
1604 * to another cpu.
1605 *
1606 * It is important for this function to not return a false negative.
1607 */
hrtimer_active(const struct hrtimer * timer)1608 bool hrtimer_active(const struct hrtimer *timer)
1609 {
1610 struct hrtimer_clock_base *base;
1611 unsigned int seq;
1612
1613 do {
1614 base = READ_ONCE(timer->base);
1615 seq = raw_read_seqcount_begin(&base->seq);
1616
1617 if (timer->state != HRTIMER_STATE_INACTIVE ||
1618 base->running == timer)
1619 return true;
1620
1621 } while (read_seqcount_retry(&base->seq, seq) ||
1622 base != READ_ONCE(timer->base));
1623
1624 return false;
1625 }
1626 EXPORT_SYMBOL_GPL(hrtimer_active);
1627
1628 /*
1629 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1630 * distinct sections:
1631 *
1632 * - queued: the timer is queued
1633 * - callback: the timer is being ran
1634 * - post: the timer is inactive or (re)queued
1635 *
1636 * On the read side we ensure we observe timer->state and cpu_base->running
1637 * from the same section, if anything changed while we looked at it, we retry.
1638 * This includes timer->base changing because sequence numbers alone are
1639 * insufficient for that.
1640 *
1641 * The sequence numbers are required because otherwise we could still observe
1642 * a false negative if the read side got smeared over multiple consecutive
1643 * __run_hrtimer() invocations.
1644 */
1645
__run_hrtimer(struct hrtimer_cpu_base * cpu_base,struct hrtimer_clock_base * base,struct hrtimer * timer,ktime_t * now,unsigned long flags)1646 static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1647 struct hrtimer_clock_base *base,
1648 struct hrtimer *timer, ktime_t *now,
1649 unsigned long flags) __must_hold(&cpu_base->lock)
1650 {
1651 enum hrtimer_restart (*fn)(struct hrtimer *);
1652 bool expires_in_hardirq;
1653 int restart;
1654
1655 lockdep_assert_held(&cpu_base->lock);
1656
1657 debug_deactivate(timer);
1658 base->running = timer;
1659
1660 /*
1661 * Separate the ->running assignment from the ->state assignment.
1662 *
1663 * As with a regular write barrier, this ensures the read side in
1664 * hrtimer_active() cannot observe base->running == NULL &&
1665 * timer->state == INACTIVE.
1666 */
1667 raw_write_seqcount_barrier(&base->seq);
1668
1669 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1670 fn = timer->function;
1671
1672 /*
1673 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1674 * timer is restarted with a period then it becomes an absolute
1675 * timer. If its not restarted it does not matter.
1676 */
1677 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1678 timer->is_rel = false;
1679
1680 /*
1681 * The timer is marked as running in the CPU base, so it is
1682 * protected against migration to a different CPU even if the lock
1683 * is dropped.
1684 */
1685 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1686 trace_hrtimer_expire_entry(timer, now);
1687 expires_in_hardirq = lockdep_hrtimer_enter(timer);
1688
1689 restart = fn(timer);
1690
1691 lockdep_hrtimer_exit(expires_in_hardirq);
1692 trace_hrtimer_expire_exit(timer);
1693 raw_spin_lock_irq(&cpu_base->lock);
1694
1695 /*
1696 * Note: We clear the running state after enqueue_hrtimer and
1697 * we do not reprogram the event hardware. Happens either in
1698 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1699 *
1700 * Note: Because we dropped the cpu_base->lock above,
1701 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1702 * for us already.
1703 */
1704 if (restart != HRTIMER_NORESTART &&
1705 !(timer->state & HRTIMER_STATE_ENQUEUED))
1706 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1707
1708 /*
1709 * Separate the ->running assignment from the ->state assignment.
1710 *
1711 * As with a regular write barrier, this ensures the read side in
1712 * hrtimer_active() cannot observe base->running.timer == NULL &&
1713 * timer->state == INACTIVE.
1714 */
1715 raw_write_seqcount_barrier(&base->seq);
1716
1717 WARN_ON_ONCE(base->running != timer);
1718 base->running = NULL;
1719 }
1720
__hrtimer_run_queues(struct hrtimer_cpu_base * cpu_base,ktime_t now,unsigned long flags,unsigned int active_mask)1721 static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1722 unsigned long flags, unsigned int active_mask)
1723 {
1724 struct hrtimer_clock_base *base;
1725 unsigned int active = cpu_base->active_bases & active_mask;
1726
1727 for_each_active_base(base, cpu_base, active) {
1728 struct timerqueue_node *node;
1729 ktime_t basenow;
1730
1731 basenow = ktime_add(now, base->offset);
1732
1733 while ((node = timerqueue_getnext(&base->active))) {
1734 struct hrtimer *timer;
1735
1736 timer = container_of(node, struct hrtimer, node);
1737
1738 /*
1739 * The immediate goal for using the softexpires is
1740 * minimizing wakeups, not running timers at the
1741 * earliest interrupt after their soft expiration.
1742 * This allows us to avoid using a Priority Search
1743 * Tree, which can answer a stabbing query for
1744 * overlapping intervals and instead use the simple
1745 * BST we already have.
1746 * We don't add extra wakeups by delaying timers that
1747 * are right-of a not yet expired timer, because that
1748 * timer will have to trigger a wakeup anyway.
1749 */
1750 if (basenow < hrtimer_get_softexpires_tv64(timer))
1751 break;
1752
1753 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1754 if (active_mask == HRTIMER_ACTIVE_SOFT)
1755 hrtimer_sync_wait_running(cpu_base, flags);
1756 }
1757 }
1758 }
1759
hrtimer_run_softirq(struct softirq_action * h)1760 static __latent_entropy void hrtimer_run_softirq(struct softirq_action *h)
1761 {
1762 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1763 unsigned long flags;
1764 ktime_t now;
1765
1766 hrtimer_cpu_base_lock_expiry(cpu_base);
1767 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1768
1769 now = hrtimer_update_base(cpu_base);
1770 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1771
1772 cpu_base->softirq_activated = 0;
1773 hrtimer_update_softirq_timer(cpu_base, true);
1774
1775 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1776 hrtimer_cpu_base_unlock_expiry(cpu_base);
1777 }
1778
1779 #ifdef CONFIG_HIGH_RES_TIMERS
1780
1781 /*
1782 * High resolution timer interrupt
1783 * Called with interrupts disabled
1784 */
hrtimer_interrupt(struct clock_event_device * dev)1785 void hrtimer_interrupt(struct clock_event_device *dev)
1786 {
1787 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1788 ktime_t expires_next, now, entry_time, delta;
1789 unsigned long flags;
1790 int retries = 0;
1791
1792 BUG_ON(!cpu_base->hres_active);
1793 cpu_base->nr_events++;
1794 dev->next_event = KTIME_MAX;
1795
1796 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1797 entry_time = now = hrtimer_update_base(cpu_base);
1798 retry:
1799 cpu_base->in_hrtirq = 1;
1800 /*
1801 * We set expires_next to KTIME_MAX here with cpu_base->lock
1802 * held to prevent that a timer is enqueued in our queue via
1803 * the migration code. This does not affect enqueueing of
1804 * timers which run their callback and need to be requeued on
1805 * this CPU.
1806 */
1807 cpu_base->expires_next = KTIME_MAX;
1808
1809 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1810 cpu_base->softirq_expires_next = KTIME_MAX;
1811 cpu_base->softirq_activated = 1;
1812 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1813 }
1814
1815 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1816
1817 /* Reevaluate the clock bases for the [soft] next expiry */
1818 expires_next = hrtimer_update_next_event(cpu_base);
1819 /*
1820 * Store the new expiry value so the migration code can verify
1821 * against it.
1822 */
1823 cpu_base->expires_next = expires_next;
1824 cpu_base->in_hrtirq = 0;
1825 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1826
1827 /* Reprogramming necessary ? */
1828 if (!tick_program_event(expires_next, 0)) {
1829 cpu_base->hang_detected = 0;
1830 return;
1831 }
1832
1833 /*
1834 * The next timer was already expired due to:
1835 * - tracing
1836 * - long lasting callbacks
1837 * - being scheduled away when running in a VM
1838 *
1839 * We need to prevent that we loop forever in the hrtimer
1840 * interrupt routine. We give it 3 attempts to avoid
1841 * overreacting on some spurious event.
1842 *
1843 * Acquire base lock for updating the offsets and retrieving
1844 * the current time.
1845 */
1846 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1847 now = hrtimer_update_base(cpu_base);
1848 cpu_base->nr_retries++;
1849 if (++retries < 3)
1850 goto retry;
1851 /*
1852 * Give the system a chance to do something else than looping
1853 * here. We stored the entry time, so we know exactly how long
1854 * we spent here. We schedule the next event this amount of
1855 * time away.
1856 */
1857 cpu_base->nr_hangs++;
1858 cpu_base->hang_detected = 1;
1859 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1860
1861 delta = ktime_sub(now, entry_time);
1862 if ((unsigned int)delta > cpu_base->max_hang_time)
1863 cpu_base->max_hang_time = (unsigned int) delta;
1864 /*
1865 * Limit it to a sensible value as we enforce a longer
1866 * delay. Give the CPU at least 100ms to catch up.
1867 */
1868 if (delta > 100 * NSEC_PER_MSEC)
1869 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1870 else
1871 expires_next = ktime_add(now, delta);
1872 tick_program_event(expires_next, 1);
1873 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1874 }
1875
1876 /* called with interrupts disabled */
__hrtimer_peek_ahead_timers(void)1877 static inline void __hrtimer_peek_ahead_timers(void)
1878 {
1879 struct tick_device *td;
1880
1881 if (!hrtimer_hres_active())
1882 return;
1883
1884 td = this_cpu_ptr(&tick_cpu_device);
1885 if (td && td->evtdev)
1886 hrtimer_interrupt(td->evtdev);
1887 }
1888
1889 #else /* CONFIG_HIGH_RES_TIMERS */
1890
__hrtimer_peek_ahead_timers(void)1891 static inline void __hrtimer_peek_ahead_timers(void) { }
1892
1893 #endif /* !CONFIG_HIGH_RES_TIMERS */
1894
1895 /*
1896 * Called from run_local_timers in hardirq context every jiffy
1897 */
hrtimer_run_queues(void)1898 void hrtimer_run_queues(void)
1899 {
1900 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1901 unsigned long flags;
1902 ktime_t now;
1903
1904 if (__hrtimer_hres_active(cpu_base))
1905 return;
1906
1907 /*
1908 * This _is_ ugly: We have to check periodically, whether we
1909 * can switch to highres and / or nohz mode. The clocksource
1910 * switch happens with xtime_lock held. Notification from
1911 * there only sets the check bit in the tick_oneshot code,
1912 * otherwise we might deadlock vs. xtime_lock.
1913 */
1914 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1915 hrtimer_switch_to_hres();
1916 return;
1917 }
1918
1919 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1920 now = hrtimer_update_base(cpu_base);
1921
1922 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1923 cpu_base->softirq_expires_next = KTIME_MAX;
1924 cpu_base->softirq_activated = 1;
1925 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1926 }
1927
1928 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1929 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1930 }
1931
1932 /*
1933 * Sleep related functions:
1934 */
hrtimer_wakeup(struct hrtimer * timer)1935 static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1936 {
1937 struct hrtimer_sleeper *t =
1938 container_of(timer, struct hrtimer_sleeper, timer);
1939 struct task_struct *task = t->task;
1940
1941 t->task = NULL;
1942 if (task)
1943 wake_up_process(task);
1944
1945 return HRTIMER_NORESTART;
1946 }
1947
1948 /**
1949 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
1950 * @sl: sleeper to be started
1951 * @mode: timer mode abs/rel
1952 *
1953 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
1954 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
1955 */
hrtimer_sleeper_start_expires(struct hrtimer_sleeper * sl,enum hrtimer_mode mode)1956 void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
1957 enum hrtimer_mode mode)
1958 {
1959 /*
1960 * Make the enqueue delivery mode check work on RT. If the sleeper
1961 * was initialized for hard interrupt delivery, force the mode bit.
1962 * This is a special case for hrtimer_sleepers because
1963 * hrtimer_init_sleeper() determines the delivery mode on RT so the
1964 * fiddling with this decision is avoided at the call sites.
1965 */
1966 if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
1967 mode |= HRTIMER_MODE_HARD;
1968
1969 hrtimer_start_expires(&sl->timer, mode);
1970 }
1971 EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
1972
__hrtimer_init_sleeper(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)1973 static void __hrtimer_init_sleeper(struct hrtimer_sleeper *sl,
1974 clockid_t clock_id, enum hrtimer_mode mode)
1975 {
1976 /*
1977 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1978 * marked for hard interrupt expiry mode are moved into soft
1979 * interrupt context either for latency reasons or because the
1980 * hrtimer callback takes regular spinlocks or invokes other
1981 * functions which are not suitable for hard interrupt context on
1982 * PREEMPT_RT.
1983 *
1984 * The hrtimer_sleeper callback is RT compatible in hard interrupt
1985 * context, but there is a latency concern: Untrusted userspace can
1986 * spawn many threads which arm timers for the same expiry time on
1987 * the same CPU. That causes a latency spike due to the wakeup of
1988 * a gazillion threads.
1989 *
1990 * OTOH, privileged real-time user space applications rely on the
1991 * low latency of hard interrupt wakeups. If the current task is in
1992 * a real-time scheduling class, mark the mode for hard interrupt
1993 * expiry.
1994 */
1995 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1996 if (task_is_realtime(current) && !(mode & HRTIMER_MODE_SOFT))
1997 mode |= HRTIMER_MODE_HARD;
1998 }
1999
2000 __hrtimer_init(&sl->timer, clock_id, mode);
2001 sl->timer.function = hrtimer_wakeup;
2002 sl->task = current;
2003 }
2004
2005 /**
2006 * hrtimer_init_sleeper - initialize sleeper to the given clock
2007 * @sl: sleeper to be initialized
2008 * @clock_id: the clock to be used
2009 * @mode: timer mode abs/rel
2010 */
hrtimer_init_sleeper(struct hrtimer_sleeper * sl,clockid_t clock_id,enum hrtimer_mode mode)2011 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, clockid_t clock_id,
2012 enum hrtimer_mode mode)
2013 {
2014 debug_init(&sl->timer, clock_id, mode);
2015 __hrtimer_init_sleeper(sl, clock_id, mode);
2016
2017 }
2018 EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
2019
nanosleep_copyout(struct restart_block * restart,struct timespec64 * ts)2020 int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
2021 {
2022 switch(restart->nanosleep.type) {
2023 #ifdef CONFIG_COMPAT_32BIT_TIME
2024 case TT_COMPAT:
2025 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
2026 return -EFAULT;
2027 break;
2028 #endif
2029 case TT_NATIVE:
2030 if (put_timespec64(ts, restart->nanosleep.rmtp))
2031 return -EFAULT;
2032 break;
2033 default:
2034 BUG();
2035 }
2036 return -ERESTART_RESTARTBLOCK;
2037 }
2038
do_nanosleep(struct hrtimer_sleeper * t,enum hrtimer_mode mode)2039 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
2040 {
2041 struct restart_block *restart;
2042
2043 do {
2044 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
2045 hrtimer_sleeper_start_expires(t, mode);
2046
2047 if (likely(t->task))
2048 schedule();
2049
2050 hrtimer_cancel(&t->timer);
2051 mode = HRTIMER_MODE_ABS;
2052
2053 } while (t->task && !signal_pending(current));
2054
2055 __set_current_state(TASK_RUNNING);
2056
2057 if (!t->task)
2058 return 0;
2059
2060 restart = ¤t->restart_block;
2061 if (restart->nanosleep.type != TT_NONE) {
2062 ktime_t rem = hrtimer_expires_remaining(&t->timer);
2063 struct timespec64 rmt;
2064
2065 if (rem <= 0)
2066 return 0;
2067 rmt = ktime_to_timespec64(rem);
2068
2069 return nanosleep_copyout(restart, &rmt);
2070 }
2071 return -ERESTART_RESTARTBLOCK;
2072 }
2073
hrtimer_nanosleep_restart(struct restart_block * restart)2074 static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2075 {
2076 struct hrtimer_sleeper t;
2077 int ret;
2078
2079 hrtimer_init_sleeper_on_stack(&t, restart->nanosleep.clockid,
2080 HRTIMER_MODE_ABS);
2081 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
2082 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2083 destroy_hrtimer_on_stack(&t.timer);
2084 return ret;
2085 }
2086
hrtimer_nanosleep(ktime_t rqtp,const enum hrtimer_mode mode,const clockid_t clockid)2087 long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2088 const clockid_t clockid)
2089 {
2090 struct restart_block *restart;
2091 struct hrtimer_sleeper t;
2092 int ret = 0;
2093 u64 slack;
2094
2095 slack = current->timer_slack_ns;
2096 if (rt_task(current))
2097 slack = 0;
2098
2099 hrtimer_init_sleeper_on_stack(&t, clockid, mode);
2100 hrtimer_set_expires_range_ns(&t.timer, rqtp, slack);
2101 ret = do_nanosleep(&t, mode);
2102 if (ret != -ERESTART_RESTARTBLOCK)
2103 goto out;
2104
2105 /* Absolute timers do not update the rmtp value and restart: */
2106 if (mode == HRTIMER_MODE_ABS) {
2107 ret = -ERESTARTNOHAND;
2108 goto out;
2109 }
2110
2111 restart = ¤t->restart_block;
2112 restart->nanosleep.clockid = t.timer.base->clockid;
2113 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
2114 set_restart_fn(restart, hrtimer_nanosleep_restart);
2115 out:
2116 destroy_hrtimer_on_stack(&t.timer);
2117 return ret;
2118 }
2119
2120 #ifdef CONFIG_64BIT
2121
SYSCALL_DEFINE2(nanosleep,struct __kernel_timespec __user *,rqtp,struct __kernel_timespec __user *,rmtp)2122 SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2123 struct __kernel_timespec __user *, rmtp)
2124 {
2125 struct timespec64 tu;
2126
2127 if (get_timespec64(&tu, rqtp))
2128 return -EFAULT;
2129
2130 if (!timespec64_valid(&tu))
2131 return -EINVAL;
2132
2133 current->restart_block.fn = do_no_restart_syscall;
2134 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2135 current->restart_block.nanosleep.rmtp = rmtp;
2136 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2137 CLOCK_MONOTONIC);
2138 }
2139
2140 #endif
2141
2142 #ifdef CONFIG_COMPAT_32BIT_TIME
2143
SYSCALL_DEFINE2(nanosleep_time32,struct old_timespec32 __user *,rqtp,struct old_timespec32 __user *,rmtp)2144 SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2145 struct old_timespec32 __user *, rmtp)
2146 {
2147 struct timespec64 tu;
2148
2149 if (get_old_timespec32(&tu, rqtp))
2150 return -EFAULT;
2151
2152 if (!timespec64_valid(&tu))
2153 return -EINVAL;
2154
2155 current->restart_block.fn = do_no_restart_syscall;
2156 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2157 current->restart_block.nanosleep.compat_rmtp = rmtp;
2158 return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2159 CLOCK_MONOTONIC);
2160 }
2161 #endif
2162
2163 /*
2164 * Functions related to boot-time initialization:
2165 */
hrtimers_prepare_cpu(unsigned int cpu)2166 int hrtimers_prepare_cpu(unsigned int cpu)
2167 {
2168 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2169 int i;
2170
2171 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2172 struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2173
2174 clock_b->cpu_base = cpu_base;
2175 seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2176 timerqueue_init_head(&clock_b->active);
2177 }
2178
2179 cpu_base->cpu = cpu;
2180 cpu_base->active_bases = 0;
2181 cpu_base->hres_active = 0;
2182 cpu_base->hang_detected = 0;
2183 cpu_base->next_timer = NULL;
2184 cpu_base->softirq_next_timer = NULL;
2185 cpu_base->expires_next = KTIME_MAX;
2186 cpu_base->softirq_expires_next = KTIME_MAX;
2187 cpu_base->online = 1;
2188 hrtimer_cpu_base_init_expiry_lock(cpu_base);
2189 return 0;
2190 }
2191
2192 #ifdef CONFIG_HOTPLUG_CPU
2193
migrate_hrtimer_list(struct hrtimer_clock_base * old_base,struct hrtimer_clock_base * new_base)2194 static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2195 struct hrtimer_clock_base *new_base)
2196 {
2197 struct hrtimer *timer;
2198 struct timerqueue_node *node;
2199
2200 while ((node = timerqueue_getnext(&old_base->active))) {
2201 timer = container_of(node, struct hrtimer, node);
2202 BUG_ON(hrtimer_callback_running(timer));
2203 debug_deactivate(timer);
2204
2205 /*
2206 * Mark it as ENQUEUED not INACTIVE otherwise the
2207 * timer could be seen as !active and just vanish away
2208 * under us on another CPU
2209 */
2210 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2211 timer->base = new_base;
2212 /*
2213 * Enqueue the timers on the new cpu. This does not
2214 * reprogram the event device in case the timer
2215 * expires before the earliest on this CPU, but we run
2216 * hrtimer_interrupt after we migrated everything to
2217 * sort out already expired timers and reprogram the
2218 * event device.
2219 */
2220 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2221 }
2222 }
2223
hrtimers_cpu_dying(unsigned int dying_cpu)2224 int hrtimers_cpu_dying(unsigned int dying_cpu)
2225 {
2226 struct hrtimer_cpu_base *old_base, *new_base;
2227 int i, ncpu = cpumask_first(cpu_active_mask);
2228
2229 tick_cancel_sched_timer(dying_cpu);
2230
2231 old_base = this_cpu_ptr(&hrtimer_bases);
2232 new_base = &per_cpu(hrtimer_bases, ncpu);
2233
2234 /*
2235 * The caller is globally serialized and nobody else
2236 * takes two locks at once, deadlock is not possible.
2237 */
2238 raw_spin_lock(&old_base->lock);
2239 raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2240
2241 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2242 migrate_hrtimer_list(&old_base->clock_base[i],
2243 &new_base->clock_base[i]);
2244 }
2245
2246 /*
2247 * The migration might have changed the first expiring softirq
2248 * timer on this CPU. Update it.
2249 */
2250 __hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
2251 /* Tell the other CPU to retrigger the next event */
2252 smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2253
2254 raw_spin_unlock(&new_base->lock);
2255 old_base->online = 0;
2256 raw_spin_unlock(&old_base->lock);
2257
2258 return 0;
2259 }
2260
2261 #endif /* CONFIG_HOTPLUG_CPU */
2262
hrtimers_init(void)2263 void __init hrtimers_init(void)
2264 {
2265 hrtimers_prepare_cpu(smp_processor_id());
2266 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2267 }
2268
2269 /**
2270 * schedule_hrtimeout_range_clock - sleep until timeout
2271 * @expires: timeout value (ktime_t)
2272 * @delta: slack in expires timeout (ktime_t) for SCHED_OTHER tasks
2273 * @mode: timer mode
2274 * @clock_id: timer clock to be used
2275 */
2276 int __sched
schedule_hrtimeout_range_clock(ktime_t * expires,u64 delta,const enum hrtimer_mode mode,clockid_t clock_id)2277 schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
2278 const enum hrtimer_mode mode, clockid_t clock_id)
2279 {
2280 struct hrtimer_sleeper t;
2281
2282 /*
2283 * Optimize when a zero timeout value is given. It does not
2284 * matter whether this is an absolute or a relative time.
2285 */
2286 if (expires && *expires == 0) {
2287 __set_current_state(TASK_RUNNING);
2288 return 0;
2289 }
2290
2291 /*
2292 * A NULL parameter means "infinite"
2293 */
2294 if (!expires) {
2295 schedule();
2296 return -EINTR;
2297 }
2298
2299 /*
2300 * Override any slack passed by the user if under
2301 * rt contraints.
2302 */
2303 if (rt_task(current))
2304 delta = 0;
2305
2306 hrtimer_init_sleeper_on_stack(&t, clock_id, mode);
2307 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
2308 hrtimer_sleeper_start_expires(&t, mode);
2309
2310 if (likely(t.task))
2311 schedule();
2312
2313 hrtimer_cancel(&t.timer);
2314 destroy_hrtimer_on_stack(&t.timer);
2315
2316 __set_current_state(TASK_RUNNING);
2317
2318 return !t.task ? 0 : -EINTR;
2319 }
2320 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range_clock);
2321
2322 /**
2323 * schedule_hrtimeout_range - sleep until timeout
2324 * @expires: timeout value (ktime_t)
2325 * @delta: slack in expires timeout (ktime_t) for SCHED_OTHER tasks
2326 * @mode: timer mode
2327 *
2328 * Make the current task sleep until the given expiry time has
2329 * elapsed. The routine will return immediately unless
2330 * the current task state has been set (see set_current_state()).
2331 *
2332 * The @delta argument gives the kernel the freedom to schedule the
2333 * actual wakeup to a time that is both power and performance friendly
2334 * for regular (non RT/DL) tasks.
2335 * The kernel give the normal best effort behavior for "@expires+@delta",
2336 * but may decide to fire the timer earlier, but no earlier than @expires.
2337 *
2338 * You can set the task state as follows -
2339 *
2340 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2341 * pass before the routine returns unless the current task is explicitly
2342 * woken up, (e.g. by wake_up_process()).
2343 *
2344 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2345 * delivered to the current task or the current task is explicitly woken
2346 * up.
2347 *
2348 * The current task state is guaranteed to be TASK_RUNNING when this
2349 * routine returns.
2350 *
2351 * Returns 0 when the timer has expired. If the task was woken before the
2352 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2353 * by an explicit wakeup, it returns -EINTR.
2354 */
schedule_hrtimeout_range(ktime_t * expires,u64 delta,const enum hrtimer_mode mode)2355 int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
2356 const enum hrtimer_mode mode)
2357 {
2358 return schedule_hrtimeout_range_clock(expires, delta, mode,
2359 CLOCK_MONOTONIC);
2360 }
2361 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
2362
2363 /**
2364 * schedule_hrtimeout - sleep until timeout
2365 * @expires: timeout value (ktime_t)
2366 * @mode: timer mode
2367 *
2368 * Make the current task sleep until the given expiry time has
2369 * elapsed. The routine will return immediately unless
2370 * the current task state has been set (see set_current_state()).
2371 *
2372 * You can set the task state as follows -
2373 *
2374 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2375 * pass before the routine returns unless the current task is explicitly
2376 * woken up, (e.g. by wake_up_process()).
2377 *
2378 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2379 * delivered to the current task or the current task is explicitly woken
2380 * up.
2381 *
2382 * The current task state is guaranteed to be TASK_RUNNING when this
2383 * routine returns.
2384 *
2385 * Returns 0 when the timer has expired. If the task was woken before the
2386 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2387 * by an explicit wakeup, it returns -EINTR.
2388 */
schedule_hrtimeout(ktime_t * expires,const enum hrtimer_mode mode)2389 int __sched schedule_hrtimeout(ktime_t *expires,
2390 const enum hrtimer_mode mode)
2391 {
2392 return schedule_hrtimeout_range(expires, 0, mode);
2393 }
2394 EXPORT_SYMBOL_GPL(schedule_hrtimeout);
2395