projects/linux·blob v4.2-rc8 · refs

1803 lines · 46.0 KiB

1/*
2 * linux/kernel/hrtimer.c
3 *
4 * Copyright(C) 2005-2006, Thomas Gleixner <[email protected]>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 *
8 * High-resolution kernel timers
9 *
10 * In contrast to the low-resolution timeout API implemented in
11 * kernel/timer.c, hrtimers provide finer resolution and accuracy
12 * depending on system configuration and capabilities.
13 *
14 * These timers are currently used for:
15 * - itimers
16 * - POSIX timers
17 * - nanosleep
18 * - precise in-kernel timing
19 *
20 * Started by: Thomas Gleixner and Ingo Molnar
21 *
22 * Credits:
23 * based on kernel/timer.c
24 *
25 * Help, testing, suggestions, bugfixes, improvements were
26 * provided by:
27 *
28 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
29 * et. al.
30 *
31 * For licencing details see kernel-base/COPYING
32 */
33
34#include <linux/cpu.h>
35#include <linux/export.h>
36#include <linux/percpu.h>
37#include <linux/hrtimer.h>
38#include <linux/notifier.h>
39#include <linux/syscalls.h>
40#include <linux/kallsyms.h>
41#include <linux/interrupt.h>
42#include <linux/tick.h>
43#include <linux/seq_file.h>
44#include <linux/err.h>
45#include <linux/debugobjects.h>
46#include <linux/sched.h>
47#include <linux/sched/sysctl.h>
48#include <linux/sched/rt.h>
49#include <linux/sched/deadline.h>
50#include <linux/timer.h>
51#include <linux/freezer.h>
52
53#include <asm/uaccess.h>
54
55#include <trace/events/timer.h>
56
57#include "tick-internal.h"
58
59/*
60 * The timer bases:
61 *
62 * There are more clockids then hrtimer bases. Thus, we index
63 * into the timer bases by the hrtimer_base_type enum. When trying
64 * to reach a base using a clockid, hrtimer_clockid_to_base()
65 * is used to convert from clockid to the proper hrtimer_base_type.
66 */
67DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
68{
69 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
70 .seq = SEQCNT_ZERO(hrtimer_bases.seq),
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};
95
96static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
97 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
98 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
99 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
100 [CLOCK_TAI] = HRTIMER_BASE_TAI,
101};
102
103static inline int hrtimer_clockid_to_base(clockid_t clock_id)
104{
105 return hrtimer_clock_to_base_table[clock_id];
106}
107
108/*
109 * Functions and macros which are different for UP/SMP systems are kept in a
110 * single place
111 */
112#ifdef CONFIG_SMP
113
114/*
115 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
116 * such that hrtimer_callback_running() can unconditionally dereference
117 * timer->base->cpu_base
118 */
119static struct hrtimer_cpu_base migration_cpu_base = {
120 .seq = SEQCNT_ZERO(migration_cpu_base),
121 .clock_base = { { .cpu_base = &migration_cpu_base, }, },
122};
123
124#define migration_base migration_cpu_base.clock_base[0]
125
126/*
127 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
128 * means that all timers which are tied to this base via timer->base are
129 * locked, and the base itself is locked too.
130 *
131 * So __run_timers/migrate_timers can safely modify all timers which could
132 * be found on the lists/queues.
133 *
134 * When the timer's base is locked, and the timer removed from list, it is
135 * possible to set timer->base = &migration_base and drop the lock: the timer
136 * remains locked.
137 */
138static
139struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
140 unsigned long *flags)
141{
142 struct hrtimer_clock_base *base;
143
144 for (;;) {
145 base = timer->base;
146 if (likely(base != &migration_base)) {
147 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
148 if (likely(base == timer->base))
149 return base;
150 /* The timer has migrated to another CPU: */
151 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
152 }
153 cpu_relax();
154 }
155}
156
157/*
158 * With HIGHRES=y we do not migrate the timer when it is expiring
159 * before the next event on the target cpu because we cannot reprogram
160 * the target cpu hardware and we would cause it to fire late.
161 *
162 * Called with cpu_base->lock of target cpu held.
163 */
164static int
165hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
166{
167#ifdef CONFIG_HIGH_RES_TIMERS
168 ktime_t expires;
169
170 if (!new_base->cpu_base->hres_active)
171 return 0;
172
173 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
174 return expires.tv64 <= new_base->cpu_base->expires_next.tv64;
175#else
176 return 0;
177#endif
178}
179
180#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
181static inline
182struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
183 int pinned)
184{
185 if (pinned || !base->migration_enabled)
186 return this_cpu_ptr(&hrtimer_bases);
187 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
188}
189#else
190static inline
191struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
192 int pinned)
193{
194 return this_cpu_ptr(&hrtimer_bases);
195}
196#endif
197
198/*
199 * Switch the timer base to the current CPU when possible.
200 */
201static inline struct hrtimer_clock_base *
202switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
203 int pinned)
204{
205 struct hrtimer_cpu_base *new_cpu_base, *this_base;
206 struct hrtimer_clock_base *new_base;
207 int basenum = base->index;
208
209 this_base = this_cpu_ptr(&hrtimer_bases);
210 new_cpu_base = get_target_base(this_base, pinned);
211again:
212 new_base = &new_cpu_base->clock_base[basenum];
213
214 if (base != new_base) {
215 /*
216 * We are trying to move timer to new_base.
217 * However we can't change timer's base while it is running,
218 * so we keep it on the same CPU. No hassle vs. reprogramming
219 * the event source in the high resolution case. The softirq
220 * code will take care of this when the timer function has
221 * completed. There is no conflict as we hold the lock until
222 * the timer is enqueued.
223 */
224 if (unlikely(hrtimer_callback_running(timer)))
225 return base;
226
227 /* See the comment in lock_hrtimer_base() */
228 timer->base = &migration_base;
229 raw_spin_unlock(&base->cpu_base->lock);
230 raw_spin_lock(&new_base->cpu_base->lock);
231
232 if (new_cpu_base != this_base &&
233 hrtimer_check_target(timer, new_base)) {
234 raw_spin_unlock(&new_base->cpu_base->lock);
235 raw_spin_lock(&base->cpu_base->lock);
236 new_cpu_base = this_base;
237 timer->base = base;
238 goto again;
239 }
240 timer->base = new_base;
241 } else {
242 if (new_cpu_base != this_base &&
243 hrtimer_check_target(timer, new_base)) {
244 new_cpu_base = this_base;
245 goto again;
246 }
247 }
248 return new_base;
249}
250
251#else /* CONFIG_SMP */
252
253static inline struct hrtimer_clock_base *
254lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
255{
256 struct hrtimer_clock_base *base = timer->base;
257
258 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
259
260 return base;
261}
262
263# define switch_hrtimer_base(t, b, p) (b)
264
265#endif /* !CONFIG_SMP */
266
267/*
268 * Functions for the union type storage format of ktime_t which are
269 * too large for inlining:
270 */
271#if BITS_PER_LONG < 64
272/*
273 * Divide a ktime value by a nanosecond value
274 */
275s64 __ktime_divns(const ktime_t kt, s64 div)
276{
277 int sft = 0;
278 s64 dclc;
279 u64 tmp;
280
281 dclc = ktime_to_ns(kt);
282 tmp = dclc < 0 ? -dclc : dclc;
283
284 /* Make sure the divisor is less than 2^32: */
285 while (div >> 32) {
286 sft++;
287 div >>= 1;
288 }
289 tmp >>= sft;
290 do_div(tmp, (unsigned long) div);
291 return dclc < 0 ? -tmp : tmp;
292}
293EXPORT_SYMBOL_GPL(__ktime_divns);
294#endif /* BITS_PER_LONG >= 64 */
295
296/*
297 * Add two ktime values and do a safety check for overflow:
298 */
299ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
300{
301 ktime_t res = ktime_add(lhs, rhs);
302
303 /*
304 * We use KTIME_SEC_MAX here, the maximum timeout which we can
305 * return to user space in a timespec:
306 */
307 if (res.tv64 < 0 || res.tv64 < lhs.tv64 || res.tv64 < rhs.tv64)
308 res = ktime_set(KTIME_SEC_MAX, 0);
309
310 return res;
311}
312
313EXPORT_SYMBOL_GPL(ktime_add_safe);
314
315#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
316
317static struct debug_obj_descr hrtimer_debug_descr;
318
319static void *hrtimer_debug_hint(void *addr)
320{
321 return ((struct hrtimer *) addr)->function;
322}
323
324/*
325 * fixup_init is called when:
326 * - an active object is initialized
327 */
328static int hrtimer_fixup_init(void *addr, enum debug_obj_state state)
329{
330 struct hrtimer *timer = addr;
331
332 switch (state) {
333 case ODEBUG_STATE_ACTIVE:
334 hrtimer_cancel(timer);
335 debug_object_init(timer, &hrtimer_debug_descr);
336 return 1;
337 default:
338 return 0;
339 }
340}
341
342/*
343 * fixup_activate is called when:
344 * - an active object is activated
345 * - an unknown object is activated (might be a statically initialized object)
346 */
347static int hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
348{
349 switch (state) {
350
351 case ODEBUG_STATE_NOTAVAILABLE:
352 WARN_ON_ONCE(1);
353 return 0;
354
355 case ODEBUG_STATE_ACTIVE:
356 WARN_ON(1);
357
358 default:
359 return 0;
360 }
361}
362
363/*
364 * fixup_free is called when:
365 * - an active object is freed
366 */
367static int hrtimer_fixup_free(void *addr, enum debug_obj_state state)
368{
369 struct hrtimer *timer = addr;
370
371 switch (state) {
372 case ODEBUG_STATE_ACTIVE:
373 hrtimer_cancel(timer);
374 debug_object_free(timer, &hrtimer_debug_descr);
375 return 1;
376 default:
377 return 0;
378 }
379}
380
381static struct debug_obj_descr hrtimer_debug_descr = {
382 .name = "hrtimer",
383 .debug_hint = hrtimer_debug_hint,
384 .fixup_init = hrtimer_fixup_init,
385 .fixup_activate = hrtimer_fixup_activate,
386 .fixup_free = hrtimer_fixup_free,
387};
388
389static inline void debug_hrtimer_init(struct hrtimer *timer)
390{
391 debug_object_init(timer, &hrtimer_debug_descr);
392}
393
394static inline void debug_hrtimer_activate(struct hrtimer *timer)
395{
396 debug_object_activate(timer, &hrtimer_debug_descr);
397}
398
399static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
400{
401 debug_object_deactivate(timer, &hrtimer_debug_descr);
402}
403
404static inline void debug_hrtimer_free(struct hrtimer *timer)
405{
406 debug_object_free(timer, &hrtimer_debug_descr);
407}
408
409static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
410 enum hrtimer_mode mode);
411
412void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
413 enum hrtimer_mode mode)
414{
415 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
416 __hrtimer_init(timer, clock_id, mode);
417}
418EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
419
420void destroy_hrtimer_on_stack(struct hrtimer *timer)
421{
422 debug_object_free(timer, &hrtimer_debug_descr);
423}
424
425#else
426static inline void debug_hrtimer_init(struct hrtimer *timer) { }
427static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
428static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
429#endif
430
431static inline void
432debug_init(struct hrtimer *timer, clockid_t clockid,
433 enum hrtimer_mode mode)
434{
435 debug_hrtimer_init(timer);
436 trace_hrtimer_init(timer, clockid, mode);
437}
438
439static inline void debug_activate(struct hrtimer *timer)
440{
441 debug_hrtimer_activate(timer);
442 trace_hrtimer_start(timer);
443}
444
445static inline void debug_deactivate(struct hrtimer *timer)
446{
447 debug_hrtimer_deactivate(timer);
448 trace_hrtimer_cancel(timer);
449}
450
451#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
452static inline void hrtimer_update_next_timer(struct hrtimer_cpu_base *cpu_base,
453 struct hrtimer *timer)
454{
455#ifdef CONFIG_HIGH_RES_TIMERS
456 cpu_base->next_timer = timer;
457#endif
458}
459
460static ktime_t __hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base)
461{
462 struct hrtimer_clock_base *base = cpu_base->clock_base;
463 ktime_t expires, expires_next = { .tv64 = KTIME_MAX };
464 unsigned int active = cpu_base->active_bases;
465
466 hrtimer_update_next_timer(cpu_base, NULL);
467 for (; active; base++, active >>= 1) {
468 struct timerqueue_node *next;
469 struct hrtimer *timer;
470
471 if (!(active & 0x01))
472 continue;
473
474 next = timerqueue_getnext(&base->active);
475 timer = container_of(next, struct hrtimer, node);
476 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
477 if (expires.tv64 < expires_next.tv64) {
478 expires_next = expires;
479 hrtimer_update_next_timer(cpu_base, timer);
480 }
481 }
482 /*
483 * clock_was_set() might have changed base->offset of any of
484 * the clock bases so the result might be negative. Fix it up
485 * to prevent a false positive in clockevents_program_event().
486 */
487 if (expires_next.tv64 < 0)
488 expires_next.tv64 = 0;
489 return expires_next;
490}
491#endif
492
493static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
494{
495 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
496 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
497 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
498
499 return ktime_get_update_offsets_now(&base->clock_was_set_seq,
500 offs_real, offs_boot, offs_tai);
501}
502
503/* High resolution timer related functions */
504#ifdef CONFIG_HIGH_RES_TIMERS
505
506/*
507 * High resolution timer enabled ?
508 */
509static int hrtimer_hres_enabled __read_mostly = 1;
510unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
511EXPORT_SYMBOL_GPL(hrtimer_resolution);
512
513/*
514 * Enable / Disable high resolution mode
515 */
516static int __init setup_hrtimer_hres(char *str)
517{
518 if (!strcmp(str, "off"))
519 hrtimer_hres_enabled = 0;
520 else if (!strcmp(str, "on"))
521 hrtimer_hres_enabled = 1;
522 else
523 return 0;
524 return 1;
525}
526
527__setup("highres=", setup_hrtimer_hres);
528
529/*
530 * hrtimer_high_res_enabled - query, if the highres mode is enabled
531 */
532static inline int hrtimer_is_hres_enabled(void)
533{
534 return hrtimer_hres_enabled;
535}
536
537/*
538 * Is the high resolution mode active ?
539 */
540static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
541{
542 return cpu_base->hres_active;
543}
544
545static inline int hrtimer_hres_active(void)
546{
547 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
548}
549
550/*
551 * Reprogram the event source with checking both queues for the
552 * next event
553 * Called with interrupts disabled and base->lock held
554 */
555static void
556hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
557{
558 ktime_t expires_next;
559
560 if (!cpu_base->hres_active)
561 return;
562
563 expires_next = __hrtimer_get_next_event(cpu_base);
564
565 if (skip_equal && expires_next.tv64 == cpu_base->expires_next.tv64)
566 return;
567
568 cpu_base->expires_next.tv64 = expires_next.tv64;
569
570 /*
571 * If a hang was detected in the last timer interrupt then we
572 * leave the hang delay active in the hardware. We want the
573 * system to make progress. That also prevents the following
574 * scenario:
575 * T1 expires 50ms from now
576 * T2 expires 5s from now
577 *
578 * T1 is removed, so this code is called and would reprogram
579 * the hardware to 5s from now. Any hrtimer_start after that
580 * will not reprogram the hardware due to hang_detected being
581 * set. So we'd effectivly block all timers until the T2 event
582 * fires.
583 */
584 if (cpu_base->hang_detected)
585 return;
586
587 tick_program_event(cpu_base->expires_next, 1);
588}
589
590/*
591 * When a timer is enqueued and expires earlier than the already enqueued
592 * timers, we have to check, whether it expires earlier than the timer for
593 * which the clock event device was armed.
594 *
595 * Called with interrupts disabled and base->cpu_base.lock held
596 */
597static void hrtimer_reprogram(struct hrtimer *timer,
598 struct hrtimer_clock_base *base)
599{
600 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
601 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
602
603 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
604
605 /*
606 * If the timer is not on the current cpu, we cannot reprogram
607 * the other cpus clock event device.
608 */
609 if (base->cpu_base != cpu_base)
610 return;
611
612 /*
613 * If the hrtimer interrupt is running, then it will
614 * reevaluate the clock bases and reprogram the clock event
615 * device. The callbacks are always executed in hard interrupt
616 * context so we don't need an extra check for a running
617 * callback.
618 */
619 if (cpu_base->in_hrtirq)
620 return;
621
622 /*
623 * CLOCK_REALTIME timer might be requested with an absolute
624 * expiry time which is less than base->offset. Set it to 0.
625 */
626 if (expires.tv64 < 0)
627 expires.tv64 = 0;
628
629 if (expires.tv64 >= cpu_base->expires_next.tv64)
630 return;
631
632 /* Update the pointer to the next expiring timer */
633 cpu_base->next_timer = timer;
634
635 /*
636 * If a hang was detected in the last timer interrupt then we
637 * do not schedule a timer which is earlier than the expiry
638 * which we enforced in the hang detection. We want the system
639 * to make progress.
640 */
641 if (cpu_base->hang_detected)
642 return;
643
644 /*
645 * Program the timer hardware. We enforce the expiry for
646 * events which are already in the past.
647 */
648 cpu_base->expires_next = expires;
649 tick_program_event(expires, 1);
650}
651
652/*
653 * Initialize the high resolution related parts of cpu_base
654 */
655static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
656{
657 base->expires_next.tv64 = KTIME_MAX;
658 base->hres_active = 0;
659}
660
661/*
662 * Retrigger next event is called after clock was set
663 *
664 * Called with interrupts disabled via on_each_cpu()
665 */
666static void retrigger_next_event(void *arg)
667{
668 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
669
670 if (!base->hres_active)
671 return;
672
673 raw_spin_lock(&base->lock);
674 hrtimer_update_base(base);
675 hrtimer_force_reprogram(base, 0);
676 raw_spin_unlock(&base->lock);
677}
678
679/*
680 * Switch to high resolution mode
681 */
682static int hrtimer_switch_to_hres(void)
683{
684 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
685
686 if (tick_init_highres()) {
687 printk(KERN_WARNING "Could not switch to high resolution "
688 "mode on CPU %d\n", base->cpu);
689 return 0;
690 }
691 base->hres_active = 1;
692 hrtimer_resolution = HIGH_RES_NSEC;
693
694 tick_setup_sched_timer();
695 /* "Retrigger" the interrupt to get things going */
696 retrigger_next_event(NULL);
697 return 1;
698}
699
700static void clock_was_set_work(struct work_struct *work)
701{
702 clock_was_set();
703}
704
705static DECLARE_WORK(hrtimer_work, clock_was_set_work);
706
707/*
708 * Called from timekeeping and resume code to reprogramm the hrtimer
709 * interrupt device on all cpus.
710 */
711void clock_was_set_delayed(void)
712{
713 schedule_work(&hrtimer_work);
714}
715
716#else
717
718static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *b) { return 0; }
719static inline int hrtimer_hres_active(void) { return 0; }
720static inline int hrtimer_is_hres_enabled(void) { return 0; }
721static inline int hrtimer_switch_to_hres(void) { return 0; }
722static inline void
723hrtimer_force_reprogram(struct hrtimer_cpu_base *base, int skip_equal) { }
724static inline int hrtimer_reprogram(struct hrtimer *timer,
725 struct hrtimer_clock_base *base)
726{
727 return 0;
728}
729static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
730static inline void retrigger_next_event(void *arg) { }
731
732#endif /* CONFIG_HIGH_RES_TIMERS */
733
734/*
735 * Clock realtime was set
736 *
737 * Change the offset of the realtime clock vs. the monotonic
738 * clock.
739 *
740 * We might have to reprogram the high resolution timer interrupt. On
741 * SMP we call the architecture specific code to retrigger _all_ high
742 * resolution timer interrupts. On UP we just disable interrupts and
743 * call the high resolution interrupt code.
744 */
745void clock_was_set(void)
746{
747#ifdef CONFIG_HIGH_RES_TIMERS
748 /* Retrigger the CPU local events everywhere */
749 on_each_cpu(retrigger_next_event, NULL, 1);
750#endif
751 timerfd_clock_was_set();
752}
753
754/*
755 * During resume we might have to reprogram the high resolution timer
756 * interrupt on all online CPUs. However, all other CPUs will be
757 * stopped with IRQs interrupts disabled so the clock_was_set() call
758 * must be deferred.
759 */
760void hrtimers_resume(void)
761{
762 WARN_ONCE(!irqs_disabled(),
763 KERN_INFO "hrtimers_resume() called with IRQs enabled!");
764
765 /* Retrigger on the local CPU */
766 retrigger_next_event(NULL);
767 /* And schedule a retrigger for all others */
768 clock_was_set_delayed();
769}
770
771static inline void timer_stats_hrtimer_set_start_info(struct hrtimer *timer)
772{
773#ifdef CONFIG_TIMER_STATS
774 if (timer->start_site)
775 return;
776 timer->start_site = __builtin_return_address(0);
777 memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
778 timer->start_pid = current->pid;
779#endif
780}
781
782static inline void timer_stats_hrtimer_clear_start_info(struct hrtimer *timer)
783{
784#ifdef CONFIG_TIMER_STATS
785 timer->start_site = NULL;
786#endif
787}
788
789static inline void timer_stats_account_hrtimer(struct hrtimer *timer)
790{
791#ifdef CONFIG_TIMER_STATS
792 if (likely(!timer_stats_active))
793 return;
794 timer_stats_update_stats(timer, timer->start_pid, timer->start_site,
795 timer->function, timer->start_comm, 0);
796#endif
797}
798
799/*
800 * Counterpart to lock_hrtimer_base above:
801 */
802static inline
803void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
804{
805 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
806}
807
808/**
809 * hrtimer_forward - forward the timer expiry
810 * @timer: hrtimer to forward
811 * @now: forward past this time
812 * @interval: the interval to forward
813 *
814 * Forward the timer expiry so it will expire in the future.
815 * Returns the number of overruns.
816 *
817 * Can be safely called from the callback function of @timer. If
818 * called from other contexts @timer must neither be enqueued nor
819 * running the callback and the caller needs to take care of
820 * serialization.
821 *
822 * Note: This only updates the timer expiry value and does not requeue
823 * the timer.
824 */
825u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
826{
827 u64 orun = 1;
828 ktime_t delta;
829
830 delta = ktime_sub(now, hrtimer_get_expires(timer));
831
832 if (delta.tv64 < 0)
833 return 0;
834
835 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
836 return 0;
837
838 if (interval.tv64 < hrtimer_resolution)
839 interval.tv64 = hrtimer_resolution;
840
841 if (unlikely(delta.tv64 >= interval.tv64)) {
842 s64 incr = ktime_to_ns(interval);
843
844 orun = ktime_divns(delta, incr);
845 hrtimer_add_expires_ns(timer, incr * orun);
846 if (hrtimer_get_expires_tv64(timer) > now.tv64)
847 return orun;
848 /*
849 * This (and the ktime_add() below) is the
850 * correction for exact:
851 */
852 orun++;
853 }
854 hrtimer_add_expires(timer, interval);
855
856 return orun;
857}
858EXPORT_SYMBOL_GPL(hrtimer_forward);
859
860/*
861 * enqueue_hrtimer - internal function to (re)start a timer
862 *
863 * The timer is inserted in expiry order. Insertion into the
864 * red black tree is O(log(n)). Must hold the base lock.
865 *
866 * Returns 1 when the new timer is the leftmost timer in the tree.
867 */
868static int enqueue_hrtimer(struct hrtimer *timer,
869 struct hrtimer_clock_base *base)
870{
871 debug_activate(timer);
872
873 base->cpu_base->active_bases |= 1 << base->index;
874
875 timer->state = HRTIMER_STATE_ENQUEUED;
876
877 return timerqueue_add(&base->active, &timer->node);
878}
879
880/*
881 * __remove_hrtimer - internal function to remove a timer
882 *
883 * Caller must hold the base lock.
884 *
885 * High resolution timer mode reprograms the clock event device when the
886 * timer is the one which expires next. The caller can disable this by setting
887 * reprogram to zero. This is useful, when the context does a reprogramming
888 * anyway (e.g. timer interrupt)
889 */
890static void __remove_hrtimer(struct hrtimer *timer,
891 struct hrtimer_clock_base *base,
892 unsigned long newstate, int reprogram)
893{
894 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
895 unsigned int state = timer->state;
896
897 timer->state = newstate;
898 if (!(state & HRTIMER_STATE_ENQUEUED))
899 return;
900
901 if (!timerqueue_del(&base->active, &timer->node))
902 cpu_base->active_bases &= ~(1 << base->index);
903
904#ifdef CONFIG_HIGH_RES_TIMERS
905 /*
906 * Note: If reprogram is false we do not update
907 * cpu_base->next_timer. This happens when we remove the first
908 * timer on a remote cpu. No harm as we never dereference
909 * cpu_base->next_timer. So the worst thing what can happen is
910 * an superflous call to hrtimer_force_reprogram() on the
911 * remote cpu later on if the same timer gets enqueued again.
912 */
913 if (reprogram && timer == cpu_base->next_timer)
914 hrtimer_force_reprogram(cpu_base, 1);
915#endif
916}
917
918/*
919 * remove hrtimer, called with base lock held
920 */
921static inline int
922remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base, bool restart)
923{
924 if (hrtimer_is_queued(timer)) {
925 unsigned long state = timer->state;
926 int reprogram;
927
928 /*
929 * Remove the timer and force reprogramming when high
930 * resolution mode is active and the timer is on the current
931 * CPU. If we remove a timer on another CPU, reprogramming is
932 * skipped. The interrupt event on this CPU is fired and
933 * reprogramming happens in the interrupt handler. This is a
934 * rare case and less expensive than a smp call.
935 */
936 debug_deactivate(timer);
937 timer_stats_hrtimer_clear_start_info(timer);
938 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
939
940 if (!restart)
941 state = HRTIMER_STATE_INACTIVE;
942
943 __remove_hrtimer(timer, base, state, reprogram);
944 return 1;
945 }
946 return 0;
947}
948
949/**
950 * hrtimer_start_range_ns - (re)start an hrtimer on the current CPU
951 * @timer: the timer to be added
952 * @tim: expiry time
953 * @delta_ns: "slack" range for the timer
954 * @mode: expiry mode: absolute (HRTIMER_MODE_ABS) or
955 * relative (HRTIMER_MODE_REL)
956 */
957void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
958 unsigned long delta_ns, const enum hrtimer_mode mode)
959{
960 struct hrtimer_clock_base *base, *new_base;
961 unsigned long flags;
962 int leftmost;
963
964 base = lock_hrtimer_base(timer, &flags);
965
966 /* Remove an active timer from the queue: */
967 remove_hrtimer(timer, base, true);
968
969 if (mode & HRTIMER_MODE_REL) {
970 tim = ktime_add_safe(tim, base->get_time());
971 /*
972 * CONFIG_TIME_LOW_RES is a temporary way for architectures
973 * to signal that they simply return xtime in
974 * do_gettimeoffset(). In this case we want to round up by
975 * resolution when starting a relative timer, to avoid short
976 * timeouts. This will go away with the GTOD framework.
977 */
978#ifdef CONFIG_TIME_LOW_RES
979 tim = ktime_add_safe(tim, ktime_set(0, hrtimer_resolution));
980#endif
981 }
982
983 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
984
985 /* Switch the timer base, if necessary: */
986 new_base = switch_hrtimer_base(timer, base, mode & HRTIMER_MODE_PINNED);
987
988 timer_stats_hrtimer_set_start_info(timer);
989
990 leftmost = enqueue_hrtimer(timer, new_base);
991 if (!leftmost)
992 goto unlock;
993
994 if (!hrtimer_is_hres_active(timer)) {
995 /*
996 * Kick to reschedule the next tick to handle the new timer
997 * on dynticks target.
998 */
999 if (new_base->cpu_base->nohz_active)
1000 wake_up_nohz_cpu(new_base->cpu_base->cpu);
1001 } else {
1002 hrtimer_reprogram(timer, new_base);
1003 }
1004unlock:
1005 unlock_hrtimer_base(timer, &flags);
1006}
1007EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1008
1009/**
1010 * hrtimer_try_to_cancel - try to deactivate a timer
1011 * @timer: hrtimer to stop
1012 *
1013 * Returns:
1014 * 0 when the timer was not active
1015 * 1 when the timer was active
1016 * -1 when the timer is currently excuting the callback function and
1017 * cannot be stopped
1018 */
1019int hrtimer_try_to_cancel(struct hrtimer *timer)
1020{
1021 struct hrtimer_clock_base *base;
1022 unsigned long flags;
1023 int ret = -1;
1024
1025 /*
1026 * Check lockless first. If the timer is not active (neither
1027 * enqueued nor running the callback, nothing to do here. The
1028 * base lock does not serialize against a concurrent enqueue,
1029 * so we can avoid taking it.
1030 */
1031 if (!hrtimer_active(timer))
1032 return 0;
1033
1034 base = lock_hrtimer_base(timer, &flags);
1035
1036 if (!hrtimer_callback_running(timer))
1037 ret = remove_hrtimer(timer, base, false);
1038
1039 unlock_hrtimer_base(timer, &flags);
1040
1041 return ret;
1042
1043}
1044EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1045
1046/**
1047 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1048 * @timer: the timer to be cancelled
1049 *
1050 * Returns:
1051 * 0 when the timer was not active
1052 * 1 when the timer was active
1053 */
1054int hrtimer_cancel(struct hrtimer *timer)
1055{
1056 for (;;) {
1057 int ret = hrtimer_try_to_cancel(timer);
1058
1059 if (ret >= 0)
1060 return ret;
1061 cpu_relax();
1062 }
1063}
1064EXPORT_SYMBOL_GPL(hrtimer_cancel);
1065
1066/**
1067 * hrtimer_get_remaining - get remaining time for the timer
1068 * @timer: the timer to read
1069 */
1070ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
1071{
1072 unsigned long flags;
1073 ktime_t rem;
1074
1075 lock_hrtimer_base(timer, &flags);
1076 rem = hrtimer_expires_remaining(timer);
1077 unlock_hrtimer_base(timer, &flags);
1078
1079 return rem;
1080}
1081EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
1082
1083#ifdef CONFIG_NO_HZ_COMMON
1084/**
1085 * hrtimer_get_next_event - get the time until next expiry event
1086 *
1087 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1088 */
1089u64 hrtimer_get_next_event(void)
1090{
1091 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1092 u64 expires = KTIME_MAX;
1093 unsigned long flags;
1094
1095 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1096
1097 if (!__hrtimer_hres_active(cpu_base))
1098 expires = __hrtimer_get_next_event(cpu_base).tv64;
1099
1100 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1101
1102 return expires;
1103}
1104#endif
1105
1106static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1107 enum hrtimer_mode mode)
1108{
1109 struct hrtimer_cpu_base *cpu_base;
1110 int base;
1111
1112 memset(timer, 0, sizeof(struct hrtimer));
1113
1114 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1115
1116 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
1117 clock_id = CLOCK_MONOTONIC;
1118
1119 base = hrtimer_clockid_to_base(clock_id);
1120 timer->base = &cpu_base->clock_base[base];
1121 timerqueue_init(&timer->node);
1122
1123#ifdef CONFIG_TIMER_STATS
1124 timer->start_site = NULL;
1125 timer->start_pid = -1;
1126 memset(timer->start_comm, 0, TASK_COMM_LEN);
1127#endif
1128}
1129
1130/**
1131 * hrtimer_init - initialize a timer to the given clock
1132 * @timer: the timer to be initialized
1133 * @clock_id: the clock to be used
1134 * @mode: timer mode abs/rel
1135 */
1136void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1137 enum hrtimer_mode mode)
1138{
1139 debug_init(timer, clock_id, mode);
1140 __hrtimer_init(timer, clock_id, mode);
1141}
1142EXPORT_SYMBOL_GPL(hrtimer_init);
1143
1144/*
1145 * A timer is active, when it is enqueued into the rbtree or the
1146 * callback function is running or it's in the state of being migrated
1147 * to another cpu.
1148 *
1149 * It is important for this function to not return a false negative.
1150 */
1151bool hrtimer_active(const struct hrtimer *timer)
1152{
1153 struct hrtimer_cpu_base *cpu_base;
1154 unsigned int seq;
1155
1156 do {
1157 cpu_base = READ_ONCE(timer->base->cpu_base);
1158 seq = raw_read_seqcount_begin(&cpu_base->seq);
1159
1160 if (timer->state != HRTIMER_STATE_INACTIVE ||
1161 cpu_base->running == timer)
1162 return true;
1163
1164 } while (read_seqcount_retry(&cpu_base->seq, seq) ||
1165 cpu_base != READ_ONCE(timer->base->cpu_base));
1166
1167 return false;
1168}
1169EXPORT_SYMBOL_GPL(hrtimer_active);
1170
1171/*
1172 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1173 * distinct sections:
1174 *
1175 * - queued: the timer is queued
1176 * - callback: the timer is being ran
1177 * - post: the timer is inactive or (re)queued
1178 *
1179 * On the read side we ensure we observe timer->state and cpu_base->running
1180 * from the same section, if anything changed while we looked at it, we retry.
1181 * This includes timer->base changing because sequence numbers alone are
1182 * insufficient for that.
1183 *
1184 * The sequence numbers are required because otherwise we could still observe
1185 * a false negative if the read side got smeared over multiple consequtive
1186 * __run_hrtimer() invocations.
1187 */
1188
1189static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1190 struct hrtimer_clock_base *base,
1191 struct hrtimer *timer, ktime_t *now)
1192{
1193 enum hrtimer_restart (*fn)(struct hrtimer *);
1194 int restart;
1195
1196 lockdep_assert_held(&cpu_base->lock);
1197
1198 debug_deactivate(timer);
1199 cpu_base->running = timer;
1200
1201 /*
1202 * Separate the ->running assignment from the ->state assignment.
1203 *
1204 * As with a regular write barrier, this ensures the read side in
1205 * hrtimer_active() cannot observe cpu_base->running == NULL &&
1206 * timer->state == INACTIVE.
1207 */
1208 raw_write_seqcount_barrier(&cpu_base->seq);
1209
1210 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1211 timer_stats_account_hrtimer(timer);
1212 fn = timer->function;
1213
1214 /*
1215 * Because we run timers from hardirq context, there is no chance
1216 * they get migrated to another cpu, therefore its safe to unlock
1217 * the timer base.
1218 */
1219 raw_spin_unlock(&cpu_base->lock);
1220 trace_hrtimer_expire_entry(timer, now);
1221 restart = fn(timer);
1222 trace_hrtimer_expire_exit(timer);
1223 raw_spin_lock(&cpu_base->lock);
1224
1225 /*
1226 * Note: We clear the running state after enqueue_hrtimer and
1227 * we do not reprogramm the event hardware. Happens either in
1228 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1229 *
1230 * Note: Because we dropped the cpu_base->lock above,
1231 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1232 * for us already.
1233 */
1234 if (restart != HRTIMER_NORESTART &&
1235 !(timer->state & HRTIMER_STATE_ENQUEUED))
1236 enqueue_hrtimer(timer, base);
1237
1238 /*
1239 * Separate the ->running assignment from the ->state assignment.
1240 *
1241 * As with a regular write barrier, this ensures the read side in
1242 * hrtimer_active() cannot observe cpu_base->running == NULL &&
1243 * timer->state == INACTIVE.
1244 */
1245 raw_write_seqcount_barrier(&cpu_base->seq);
1246
1247 WARN_ON_ONCE(cpu_base->running != timer);
1248 cpu_base->running = NULL;
1249}
1250
1251static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now)
1252{
1253 struct hrtimer_clock_base *base = cpu_base->clock_base;
1254 unsigned int active = cpu_base->active_bases;
1255
1256 for (; active; base++, active >>= 1) {
1257 struct timerqueue_node *node;
1258 ktime_t basenow;
1259
1260 if (!(active & 0x01))
1261 continue;
1262
1263 basenow = ktime_add(now, base->offset);
1264
1265 while ((node = timerqueue_getnext(&base->active))) {
1266 struct hrtimer *timer;
1267
1268 timer = container_of(node, struct hrtimer, node);
1269
1270 /*
1271 * The immediate goal for using the softexpires is
1272 * minimizing wakeups, not running timers at the
1273 * earliest interrupt after their soft expiration.
1274 * This allows us to avoid using a Priority Search
1275 * Tree, which can answer a stabbing querry for
1276 * overlapping intervals and instead use the simple
1277 * BST we already have.
1278 * We don't add extra wakeups by delaying timers that
1279 * are right-of a not yet expired timer, because that
1280 * timer will have to trigger a wakeup anyway.
1281 */
1282 if (basenow.tv64 < hrtimer_get_softexpires_tv64(timer))
1283 break;
1284
1285 __run_hrtimer(cpu_base, base, timer, &basenow);
1286 }
1287 }
1288}
1289
1290#ifdef CONFIG_HIGH_RES_TIMERS
1291
1292/*
1293 * High resolution timer interrupt
1294 * Called with interrupts disabled
1295 */
1296void hrtimer_interrupt(struct clock_event_device *dev)
1297{
1298 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1299 ktime_t expires_next, now, entry_time, delta;
1300 int retries = 0;
1301
1302 BUG_ON(!cpu_base->hres_active);
1303 cpu_base->nr_events++;
1304 dev->next_event.tv64 = KTIME_MAX;
1305
1306 raw_spin_lock(&cpu_base->lock);
1307 entry_time = now = hrtimer_update_base(cpu_base);
1308retry:
1309 cpu_base->in_hrtirq = 1;
1310 /*
1311 * We set expires_next to KTIME_MAX here with cpu_base->lock
1312 * held to prevent that a timer is enqueued in our queue via
1313 * the migration code. This does not affect enqueueing of
1314 * timers which run their callback and need to be requeued on
1315 * this CPU.
1316 */
1317 cpu_base->expires_next.tv64 = KTIME_MAX;
1318
1319 __hrtimer_run_queues(cpu_base, now);
1320
1321 /* Reevaluate the clock bases for the next expiry */
1322 expires_next = __hrtimer_get_next_event(cpu_base);
1323 /*
1324 * Store the new expiry value so the migration code can verify
1325 * against it.
1326 */
1327 cpu_base->expires_next = expires_next;
1328 cpu_base->in_hrtirq = 0;
1329 raw_spin_unlock(&cpu_base->lock);
1330
1331 /* Reprogramming necessary ? */
1332 if (!tick_program_event(expires_next, 0)) {
1333 cpu_base->hang_detected = 0;
1334 return;
1335 }
1336
1337 /*
1338 * The next timer was already expired due to:
1339 * - tracing
1340 * - long lasting callbacks
1341 * - being scheduled away when running in a VM
1342 *
1343 * We need to prevent that we loop forever in the hrtimer
1344 * interrupt routine. We give it 3 attempts to avoid
1345 * overreacting on some spurious event.
1346 *
1347 * Acquire base lock for updating the offsets and retrieving
1348 * the current time.
1349 */
1350 raw_spin_lock(&cpu_base->lock);
1351 now = hrtimer_update_base(cpu_base);
1352 cpu_base->nr_retries++;
1353 if (++retries < 3)
1354 goto retry;
1355 /*
1356 * Give the system a chance to do something else than looping
1357 * here. We stored the entry time, so we know exactly how long
1358 * we spent here. We schedule the next event this amount of
1359 * time away.
1360 */
1361 cpu_base->nr_hangs++;
1362 cpu_base->hang_detected = 1;
1363 raw_spin_unlock(&cpu_base->lock);
1364 delta = ktime_sub(now, entry_time);
1365 if ((unsigned int)delta.tv64 > cpu_base->max_hang_time)
1366 cpu_base->max_hang_time = (unsigned int) delta.tv64;
1367 /*
1368 * Limit it to a sensible value as we enforce a longer
1369 * delay. Give the CPU at least 100ms to catch up.
1370 */
1371 if (delta.tv64 > 100 * NSEC_PER_MSEC)
1372 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1373 else
1374 expires_next = ktime_add(now, delta);
1375 tick_program_event(expires_next, 1);
1376 printk_once(KERN_WARNING "hrtimer: interrupt took %llu ns\n",
1377 ktime_to_ns(delta));
1378}
1379
1380/*
1381 * local version of hrtimer_peek_ahead_timers() called with interrupts
1382 * disabled.
1383 */
1384static inline void __hrtimer_peek_ahead_timers(void)
1385{
1386 struct tick_device *td;
1387
1388 if (!hrtimer_hres_active())
1389 return;
1390
1391 td = this_cpu_ptr(&tick_cpu_device);
1392 if (td && td->evtdev)
1393 hrtimer_interrupt(td->evtdev);
1394}
1395
1396#else /* CONFIG_HIGH_RES_TIMERS */
1397
1398static inline void __hrtimer_peek_ahead_timers(void) { }
1399
1400#endif /* !CONFIG_HIGH_RES_TIMERS */
1401
1402/*
1403 * Called from run_local_timers in hardirq context every jiffy
1404 */
1405void hrtimer_run_queues(void)
1406{
1407 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1408 ktime_t now;
1409
1410 if (__hrtimer_hres_active(cpu_base))
1411 return;
1412
1413 /*
1414 * This _is_ ugly: We have to check periodically, whether we
1415 * can switch to highres and / or nohz mode. The clocksource
1416 * switch happens with xtime_lock held. Notification from
1417 * there only sets the check bit in the tick_oneshot code,
1418 * otherwise we might deadlock vs. xtime_lock.
1419 */
1420 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1421 hrtimer_switch_to_hres();
1422 return;
1423 }
1424
1425 raw_spin_lock(&cpu_base->lock);
1426 now = hrtimer_update_base(cpu_base);
1427 __hrtimer_run_queues(cpu_base, now);
1428 raw_spin_unlock(&cpu_base->lock);
1429}
1430
1431/*
1432 * Sleep related functions:
1433 */
1434static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1435{
1436 struct hrtimer_sleeper *t =
1437 container_of(timer, struct hrtimer_sleeper, timer);
1438 struct task_struct *task = t->task;
1439
1440 t->task = NULL;
1441 if (task)
1442 wake_up_process(task);
1443
1444 return HRTIMER_NORESTART;
1445}
1446
1447void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1448{
1449 sl->timer.function = hrtimer_wakeup;
1450 sl->task = task;
1451}
1452EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1453
1454static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1455{
1456 hrtimer_init_sleeper(t, current);
1457
1458 do {
1459 set_current_state(TASK_INTERRUPTIBLE);
1460 hrtimer_start_expires(&t->timer, mode);
1461
1462 if (likely(t->task))
1463 freezable_schedule();
1464
1465 hrtimer_cancel(&t->timer);
1466 mode = HRTIMER_MODE_ABS;
1467
1468 } while (t->task && !signal_pending(current));
1469
1470 __set_current_state(TASK_RUNNING);
1471
1472 return t->task == NULL;
1473}
1474
1475static int update_rmtp(struct hrtimer *timer, struct timespec __user *rmtp)
1476{
1477 struct timespec rmt;
1478 ktime_t rem;
1479
1480 rem = hrtimer_expires_remaining(timer);
1481 if (rem.tv64 <= 0)
1482 return 0;
1483 rmt = ktime_to_timespec(rem);
1484
1485 if (copy_to_user(rmtp, &rmt, sizeof(*rmtp)))
1486 return -EFAULT;
1487
1488 return 1;
1489}
1490
1491long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1492{
1493 struct hrtimer_sleeper t;
1494 struct timespec __user *rmtp;
1495 int ret = 0;
1496
1497 hrtimer_init_on_stack(&t.timer, restart->nanosleep.clockid,
1498 HRTIMER_MODE_ABS);
1499 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
1500
1501 if (do_nanosleep(&t, HRTIMER_MODE_ABS))
1502 goto out;
1503
1504 rmtp = restart->nanosleep.rmtp;
1505 if (rmtp) {
1506 ret = update_rmtp(&t.timer, rmtp);
1507 if (ret <= 0)
1508 goto out;
1509 }
1510
1511 /* The other values in restart are already filled in */
1512 ret = -ERESTART_RESTARTBLOCK;
1513out:
1514 destroy_hrtimer_on_stack(&t.timer);
1515 return ret;
1516}
1517
1518long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
1519 const enum hrtimer_mode mode, const clockid_t clockid)
1520{
1521 struct restart_block *restart;
1522 struct hrtimer_sleeper t;
1523 int ret = 0;
1524 unsigned long slack;
1525
1526 slack = current->timer_slack_ns;
1527 if (dl_task(current) || rt_task(current))
1528 slack = 0;
1529
1530 hrtimer_init_on_stack(&t.timer, clockid, mode);
1531 hrtimer_set_expires_range_ns(&t.timer, timespec_to_ktime(*rqtp), slack);
1532 if (do_nanosleep(&t, mode))
1533 goto out;
1534
1535 /* Absolute timers do not update the rmtp value and restart: */
1536 if (mode == HRTIMER_MODE_ABS) {
1537 ret = -ERESTARTNOHAND;
1538 goto out;
1539 }
1540
1541 if (rmtp) {
1542 ret = update_rmtp(&t.timer, rmtp);
1543 if (ret <= 0)
1544 goto out;
1545 }
1546
1547 restart = &current->restart_block;
1548 restart->fn = hrtimer_nanosleep_restart;
1549 restart->nanosleep.clockid = t.timer.base->clockid;
1550 restart->nanosleep.rmtp = rmtp;
1551 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
1552
1553 ret = -ERESTART_RESTARTBLOCK;
1554out:
1555 destroy_hrtimer_on_stack(&t.timer);
1556 return ret;
1557}
1558
1559SYSCALL_DEFINE2(nanosleep, struct timespec __user *, rqtp,
1560 struct timespec __user *, rmtp)
1561{
1562 struct timespec tu;
1563
1564 if (copy_from_user(&tu, rqtp, sizeof(tu)))
1565 return -EFAULT;
1566
1567 if (!timespec_valid(&tu))
1568 return -EINVAL;
1569
1570 return hrtimer_nanosleep(&tu, rmtp, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1571}
1572
1573/*
1574 * Functions related to boot-time initialization:
1575 */
1576static void init_hrtimers_cpu(int cpu)
1577{
1578 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1579 int i;
1580
1581 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1582 cpu_base->clock_base[i].cpu_base = cpu_base;
1583 timerqueue_init_head(&cpu_base->clock_base[i].active);
1584 }
1585
1586 cpu_base->cpu = cpu;
1587 hrtimer_init_hres(cpu_base);
1588}
1589
1590#ifdef CONFIG_HOTPLUG_CPU
1591
1592static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1593 struct hrtimer_clock_base *new_base)
1594{
1595 struct hrtimer *timer;
1596 struct timerqueue_node *node;
1597
1598 while ((node = timerqueue_getnext(&old_base->active))) {
1599 timer = container_of(node, struct hrtimer, node);
1600 BUG_ON(hrtimer_callback_running(timer));
1601 debug_deactivate(timer);
1602
1603 /*
1604 * Mark it as ENQUEUED not INACTIVE otherwise the
1605 * timer could be seen as !active and just vanish away
1606 * under us on another CPU
1607 */
1608 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
1609 timer->base = new_base;
1610 /*
1611 * Enqueue the timers on the new cpu. This does not
1612 * reprogram the event device in case the timer
1613 * expires before the earliest on this CPU, but we run
1614 * hrtimer_interrupt after we migrated everything to
1615 * sort out already expired timers and reprogram the
1616 * event device.
1617 */
1618 enqueue_hrtimer(timer, new_base);
1619 }
1620}
1621
1622static void migrate_hrtimers(int scpu)
1623{
1624 struct hrtimer_cpu_base *old_base, *new_base;
1625 int i;
1626
1627 BUG_ON(cpu_online(scpu));
1628 tick_cancel_sched_timer(scpu);
1629
1630 local_irq_disable();
1631 old_base = &per_cpu(hrtimer_bases, scpu);
1632 new_base = this_cpu_ptr(&hrtimer_bases);
1633 /*
1634 * The caller is globally serialized and nobody else
1635 * takes two locks at once, deadlock is not possible.
1636 */
1637 raw_spin_lock(&new_base->lock);
1638 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1639
1640 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1641 migrate_hrtimer_list(&old_base->clock_base[i],
1642 &new_base->clock_base[i]);
1643 }
1644
1645 raw_spin_unlock(&old_base->lock);
1646 raw_spin_unlock(&new_base->lock);
1647
1648 /* Check, if we got expired work to do */
1649 __hrtimer_peek_ahead_timers();
1650 local_irq_enable();
1651}
1652
1653#endif /* CONFIG_HOTPLUG_CPU */
1654
1655static int hrtimer_cpu_notify(struct notifier_block *self,
1656 unsigned long action, void *hcpu)
1657{
1658 int scpu = (long)hcpu;
1659
1660 switch (action) {
1661
1662 case CPU_UP_PREPARE:
1663 case CPU_UP_PREPARE_FROZEN:
1664 init_hrtimers_cpu(scpu);
1665 break;
1666
1667#ifdef CONFIG_HOTPLUG_CPU
1668 case CPU_DEAD:
1669 case CPU_DEAD_FROZEN:
1670 migrate_hrtimers(scpu);
1671 break;
1672#endif
1673
1674 default:
1675 break;
1676 }
1677
1678 return NOTIFY_OK;
1679}
1680
1681static struct notifier_block hrtimers_nb = {
1682 .notifier_call = hrtimer_cpu_notify,
1683};
1684
1685void __init hrtimers_init(void)
1686{
1687 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
1688 (void *)(long)smp_processor_id());
1689 register_cpu_notifier(&hrtimers_nb);
1690}
1691
1692/**
1693 * schedule_hrtimeout_range_clock - sleep until timeout
1694 * @expires: timeout value (ktime_t)
1695 * @delta: slack in expires timeout (ktime_t)
1696 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1697 * @clock: timer clock, CLOCK_MONOTONIC or CLOCK_REALTIME
1698 */
1699int __sched
1700schedule_hrtimeout_range_clock(ktime_t *expires, unsigned long delta,
1701 const enum hrtimer_mode mode, int clock)
1702{
1703 struct hrtimer_sleeper t;
1704
1705 /*
1706 * Optimize when a zero timeout value is given. It does not
1707 * matter whether this is an absolute or a relative time.
1708 */
1709 if (expires && !expires->tv64) {
1710 __set_current_state(TASK_RUNNING);
1711 return 0;
1712 }
1713
1714 /*
1715 * A NULL parameter means "infinite"
1716 */
1717 if (!expires) {
1718 schedule();
1719 return -EINTR;
1720 }
1721
1722 hrtimer_init_on_stack(&t.timer, clock, mode);
1723 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
1724
1725 hrtimer_init_sleeper(&t, current);
1726
1727 hrtimer_start_expires(&t.timer, mode);
1728
1729 if (likely(t.task))
1730 schedule();
1731
1732 hrtimer_cancel(&t.timer);
1733 destroy_hrtimer_on_stack(&t.timer);
1734
1735 __set_current_state(TASK_RUNNING);
1736
1737 return !t.task ? 0 : -EINTR;
1738}
1739
1740/**
1741 * schedule_hrtimeout_range - sleep until timeout
1742 * @expires: timeout value (ktime_t)
1743 * @delta: slack in expires timeout (ktime_t)
1744 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1745 *
1746 * Make the current task sleep until the given expiry time has
1747 * elapsed. The routine will return immediately unless
1748 * the current task state has been set (see set_current_state()).
1749 *
1750 * The @delta argument gives the kernel the freedom to schedule the
1751 * actual wakeup to a time that is both power and performance friendly.
1752 * The kernel give the normal best effort behavior for "@expires+@delta",
1753 * but may decide to fire the timer earlier, but no earlier than @expires.
1754 *
1755 * You can set the task state as follows -
1756 *
1757 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1758 * pass before the routine returns.
1759 *
1760 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1761 * delivered to the current task.
1762 *
1763 * The current task state is guaranteed to be TASK_RUNNING when this
1764 * routine returns.
1765 *
1766 * Returns 0 when the timer has expired otherwise -EINTR
1767 */
1768int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
1769 const enum hrtimer_mode mode)
1770{
1771 return schedule_hrtimeout_range_clock(expires, delta, mode,
1772 CLOCK_MONOTONIC);
1773}
1774EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
1775
1776/**
1777 * schedule_hrtimeout - sleep until timeout
1778 * @expires: timeout value (ktime_t)
1779 * @mode: timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1780 *
1781 * Make the current task sleep until the given expiry time has
1782 * elapsed. The routine will return immediately unless
1783 * the current task state has been set (see set_current_state()).
1784 *
1785 * You can set the task state as follows -
1786 *
1787 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1788 * pass before the routine returns.
1789 *
1790 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1791 * delivered to the current task.
1792 *
1793 * The current task state is guaranteed to be TASK_RUNNING when this
1794 * routine returns.
1795 *
1796 * Returns 0 when the timer has expired otherwise -EINTR
1797 */
1798int __sched schedule_hrtimeout(ktime_t *expires,
1799 const enum hrtimer_mode mode)
1800{
1801 return schedule_hrtimeout_range(expires, 0, mode);
1802}
1803EXPORT_SYMBOL_GPL(schedule_hrtimeout);