projects/linux·blob v2.6.24-rc1 · refs

1474 lines · 35.7 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/irq.h>
36#include <linux/module.h>
37#include <linux/percpu.h>
38#include <linux/hrtimer.h>
39#include <linux/notifier.h>
40#include <linux/syscalls.h>
41#include <linux/kallsyms.h>
42#include <linux/interrupt.h>
43#include <linux/tick.h>
44#include <linux/seq_file.h>
45#include <linux/err.h>
46
47#include <asm/uaccess.h>
48
49/**
50 * ktime_get - get the monotonic time in ktime_t format
51 *
52 * returns the time in ktime_t format
53 */
54ktime_t ktime_get(void)
55{
56 struct timespec now;
57
58 ktime_get_ts(&now);
59
60 return timespec_to_ktime(now);
61}
62EXPORT_SYMBOL_GPL(ktime_get);
63
64/**
65 * ktime_get_real - get the real (wall-) time in ktime_t format
66 *
67 * returns the time in ktime_t format
68 */
69ktime_t ktime_get_real(void)
70{
71 struct timespec now;
72
73 getnstimeofday(&now);
74
75 return timespec_to_ktime(now);
76}
77
78EXPORT_SYMBOL_GPL(ktime_get_real);
79
80/*
81 * The timer bases:
82 *
83 * Note: If we want to add new timer bases, we have to skip the two
84 * clock ids captured by the cpu-timers. We do this by holding empty
85 * entries rather than doing math adjustment of the clock ids.
86 * This ensures that we capture erroneous accesses to these clock ids
87 * rather than moving them into the range of valid clock id's.
88 */
89DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
90{
91
92 .clock_base =
93 {
94 {
95 .index = CLOCK_REALTIME,
96 .get_time = &ktime_get_real,
97 .resolution = KTIME_LOW_RES,
98 },
99 {
100 .index = CLOCK_MONOTONIC,
101 .get_time = &ktime_get,
102 .resolution = KTIME_LOW_RES,
103 },
104 }
105};
106
107/**
108 * ktime_get_ts - get the monotonic clock in timespec format
109 * @ts: pointer to timespec variable
110 *
111 * The function calculates the monotonic clock from the realtime
112 * clock and the wall_to_monotonic offset and stores the result
113 * in normalized timespec format in the variable pointed to by @ts.
114 */
115void ktime_get_ts(struct timespec *ts)
116{
117 struct timespec tomono;
118 unsigned long seq;
119
120 do {
121 seq = read_seqbegin(&xtime_lock);
122 getnstimeofday(ts);
123 tomono = wall_to_monotonic;
124
125 } while (read_seqretry(&xtime_lock, seq));
126
127 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
128 ts->tv_nsec + tomono.tv_nsec);
129}
130EXPORT_SYMBOL_GPL(ktime_get_ts);
131
132/*
133 * Get the coarse grained time at the softirq based on xtime and
134 * wall_to_monotonic.
135 */
136static void hrtimer_get_softirq_time(struct hrtimer_cpu_base *base)
137{
138 ktime_t xtim, tomono;
139 struct timespec xts, tom;
140 unsigned long seq;
141
142 do {
143 seq = read_seqbegin(&xtime_lock);
144 xts = current_kernel_time();
145 tom = wall_to_monotonic;
146 } while (read_seqretry(&xtime_lock, seq));
147
148 xtim = timespec_to_ktime(xts);
149 tomono = timespec_to_ktime(tom);
150 base->clock_base[CLOCK_REALTIME].softirq_time = xtim;
151 base->clock_base[CLOCK_MONOTONIC].softirq_time =
152 ktime_add(xtim, tomono);
153}
154
155/*
156 * Helper function to check, whether the timer is running the callback
157 * function
158 */
159static inline int hrtimer_callback_running(struct hrtimer *timer)
160{
161 return timer->state & HRTIMER_STATE_CALLBACK;
162}
163
164/*
165 * Functions and macros which are different for UP/SMP systems are kept in a
166 * single place
167 */
168#ifdef CONFIG_SMP
169
170/*
171 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
172 * means that all timers which are tied to this base via timer->base are
173 * locked, and the base itself is locked too.
174 *
175 * So __run_timers/migrate_timers can safely modify all timers which could
176 * be found on the lists/queues.
177 *
178 * When the timer's base is locked, and the timer removed from list, it is
179 * possible to set timer->base = NULL and drop the lock: the timer remains
180 * locked.
181 */
182static
183struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
184 unsigned long *flags)
185{
186 struct hrtimer_clock_base *base;
187
188 for (;;) {
189 base = timer->base;
190 if (likely(base != NULL)) {
191 spin_lock_irqsave(&base->cpu_base->lock, *flags);
192 if (likely(base == timer->base))
193 return base;
194 /* The timer has migrated to another CPU: */
195 spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
196 }
197 cpu_relax();
198 }
199}
200
201/*
202 * Switch the timer base to the current CPU when possible.
203 */
204static inline struct hrtimer_clock_base *
205switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base)
206{
207 struct hrtimer_clock_base *new_base;
208 struct hrtimer_cpu_base *new_cpu_base;
209
210 new_cpu_base = &__get_cpu_var(hrtimer_bases);
211 new_base = &new_cpu_base->clock_base[base->index];
212
213 if (base != new_base) {
214 /*
215 * We are trying to schedule the timer on the local CPU.
216 * However we can't change timer's base while it is running,
217 * so we keep it on the same CPU. No hassle vs. reprogramming
218 * the event source in the high resolution case. The softirq
219 * code will take care of this when the timer function has
220 * completed. There is no conflict as we hold the lock until
221 * the timer is enqueued.
222 */
223 if (unlikely(hrtimer_callback_running(timer)))
224 return base;
225
226 /* See the comment in lock_timer_base() */
227 timer->base = NULL;
228 spin_unlock(&base->cpu_base->lock);
229 spin_lock(&new_base->cpu_base->lock);
230 timer->base = new_base;
231 }
232 return new_base;
233}
234
235#else /* CONFIG_SMP */
236
237static inline struct hrtimer_clock_base *
238lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
239{
240 struct hrtimer_clock_base *base = timer->base;
241
242 spin_lock_irqsave(&base->cpu_base->lock, *flags);
243
244 return base;
245}
246
247# define switch_hrtimer_base(t, b) (b)
248
249#endif /* !CONFIG_SMP */
250
251/*
252 * Functions for the union type storage format of ktime_t which are
253 * too large for inlining:
254 */
255#if BITS_PER_LONG < 64
256# ifndef CONFIG_KTIME_SCALAR
257/**
258 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
259 * @kt: addend
260 * @nsec: the scalar nsec value to add
261 *
262 * Returns the sum of kt and nsec in ktime_t format
263 */
264ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
265{
266 ktime_t tmp;
267
268 if (likely(nsec < NSEC_PER_SEC)) {
269 tmp.tv64 = nsec;
270 } else {
271 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
272
273 tmp = ktime_set((long)nsec, rem);
274 }
275
276 return ktime_add(kt, tmp);
277}
278
279EXPORT_SYMBOL_GPL(ktime_add_ns);
280
281/**
282 * ktime_sub_ns - Subtract a scalar nanoseconds value from a ktime_t variable
283 * @kt: minuend
284 * @nsec: the scalar nsec value to subtract
285 *
286 * Returns the subtraction of @nsec from @kt in ktime_t format
287 */
288ktime_t ktime_sub_ns(const ktime_t kt, u64 nsec)
289{
290 ktime_t tmp;
291
292 if (likely(nsec < NSEC_PER_SEC)) {
293 tmp.tv64 = nsec;
294 } else {
295 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
296
297 tmp = ktime_set((long)nsec, rem);
298 }
299
300 return ktime_sub(kt, tmp);
301}
302
303EXPORT_SYMBOL_GPL(ktime_sub_ns);
304# endif /* !CONFIG_KTIME_SCALAR */
305
306/*
307 * Divide a ktime value by a nanosecond value
308 */
309unsigned long ktime_divns(const ktime_t kt, s64 div)
310{
311 u64 dclc, inc, dns;
312 int sft = 0;
313
314 dclc = dns = ktime_to_ns(kt);
315 inc = div;
316 /* Make sure the divisor is less than 2^32: */
317 while (div >> 32) {
318 sft++;
319 div >>= 1;
320 }
321 dclc >>= sft;
322 do_div(dclc, (unsigned long) div);
323
324 return (unsigned long) dclc;
325}
326#endif /* BITS_PER_LONG >= 64 */
327
328/* High resolution timer related functions */
329#ifdef CONFIG_HIGH_RES_TIMERS
330
331/*
332 * High resolution timer enabled ?
333 */
334static int hrtimer_hres_enabled __read_mostly = 1;
335
336/*
337 * Enable / Disable high resolution mode
338 */
339static int __init setup_hrtimer_hres(char *str)
340{
341 if (!strcmp(str, "off"))
342 hrtimer_hres_enabled = 0;
343 else if (!strcmp(str, "on"))
344 hrtimer_hres_enabled = 1;
345 else
346 return 0;
347 return 1;
348}
349
350__setup("highres=", setup_hrtimer_hres);
351
352/*
353 * hrtimer_high_res_enabled - query, if the highres mode is enabled
354 */
355static inline int hrtimer_is_hres_enabled(void)
356{
357 return hrtimer_hres_enabled;
358}
359
360/*
361 * Is the high resolution mode active ?
362 */
363static inline int hrtimer_hres_active(void)
364{
365 return __get_cpu_var(hrtimer_bases).hres_active;
366}
367
368/*
369 * Reprogram the event source with checking both queues for the
370 * next event
371 * Called with interrupts disabled and base->lock held
372 */
373static void hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base)
374{
375 int i;
376 struct hrtimer_clock_base *base = cpu_base->clock_base;
377 ktime_t expires;
378
379 cpu_base->expires_next.tv64 = KTIME_MAX;
380
381 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
382 struct hrtimer *timer;
383
384 if (!base->first)
385 continue;
386 timer = rb_entry(base->first, struct hrtimer, node);
387 expires = ktime_sub(timer->expires, base->offset);
388 if (expires.tv64 < cpu_base->expires_next.tv64)
389 cpu_base->expires_next = expires;
390 }
391
392 if (cpu_base->expires_next.tv64 != KTIME_MAX)
393 tick_program_event(cpu_base->expires_next, 1);
394}
395
396/*
397 * Shared reprogramming for clock_realtime and clock_monotonic
398 *
399 * When a timer is enqueued and expires earlier than the already enqueued
400 * timers, we have to check, whether it expires earlier than the timer for
401 * which the clock event device was armed.
402 *
403 * Called with interrupts disabled and base->cpu_base.lock held
404 */
405static int hrtimer_reprogram(struct hrtimer *timer,
406 struct hrtimer_clock_base *base)
407{
408 ktime_t *expires_next = &__get_cpu_var(hrtimer_bases).expires_next;
409 ktime_t expires = ktime_sub(timer->expires, base->offset);
410 int res;
411
412 /*
413 * When the callback is running, we do not reprogram the clock event
414 * device. The timer callback is either running on a different CPU or
415 * the callback is executed in the hrtimer_interrupt context. The
416 * reprogramming is handled either by the softirq, which called the
417 * callback or at the end of the hrtimer_interrupt.
418 */
419 if (hrtimer_callback_running(timer))
420 return 0;
421
422 if (expires.tv64 >= expires_next->tv64)
423 return 0;
424
425 /*
426 * Clockevents returns -ETIME, when the event was in the past.
427 */
428 res = tick_program_event(expires, 0);
429 if (!IS_ERR_VALUE(res))
430 *expires_next = expires;
431 return res;
432}
433
434
435/*
436 * Retrigger next event is called after clock was set
437 *
438 * Called with interrupts disabled via on_each_cpu()
439 */
440static void retrigger_next_event(void *arg)
441{
442 struct hrtimer_cpu_base *base;
443 struct timespec realtime_offset;
444 unsigned long seq;
445
446 if (!hrtimer_hres_active())
447 return;
448
449 do {
450 seq = read_seqbegin(&xtime_lock);
451 set_normalized_timespec(&realtime_offset,
452 -wall_to_monotonic.tv_sec,
453 -wall_to_monotonic.tv_nsec);
454 } while (read_seqretry(&xtime_lock, seq));
455
456 base = &__get_cpu_var(hrtimer_bases);
457
458 /* Adjust CLOCK_REALTIME offset */
459 spin_lock(&base->lock);
460 base->clock_base[CLOCK_REALTIME].offset =
461 timespec_to_ktime(realtime_offset);
462
463 hrtimer_force_reprogram(base);
464 spin_unlock(&base->lock);
465}
466
467/*
468 * Clock realtime was set
469 *
470 * Change the offset of the realtime clock vs. the monotonic
471 * clock.
472 *
473 * We might have to reprogram the high resolution timer interrupt. On
474 * SMP we call the architecture specific code to retrigger _all_ high
475 * resolution timer interrupts. On UP we just disable interrupts and
476 * call the high resolution interrupt code.
477 */
478void clock_was_set(void)
479{
480 /* Retrigger the CPU local events everywhere */
481 on_each_cpu(retrigger_next_event, NULL, 0, 1);
482}
483
484/*
485 * During resume we might have to reprogram the high resolution timer
486 * interrupt (on the local CPU):
487 */
488void hres_timers_resume(void)
489{
490 WARN_ON_ONCE(num_online_cpus() > 1);
491
492 /* Retrigger the CPU local events: */
493 retrigger_next_event(NULL);
494}
495
496/*
497 * Check, whether the timer is on the callback pending list
498 */
499static inline int hrtimer_cb_pending(const struct hrtimer *timer)
500{
501 return timer->state & HRTIMER_STATE_PENDING;
502}
503
504/*
505 * Remove a timer from the callback pending list
506 */
507static inline void hrtimer_remove_cb_pending(struct hrtimer *timer)
508{
509 list_del_init(&timer->cb_entry);
510}
511
512/*
513 * Initialize the high resolution related parts of cpu_base
514 */
515static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
516{
517 base->expires_next.tv64 = KTIME_MAX;
518 base->hres_active = 0;
519 INIT_LIST_HEAD(&base->cb_pending);
520}
521
522/*
523 * Initialize the high resolution related parts of a hrtimer
524 */
525static inline void hrtimer_init_timer_hres(struct hrtimer *timer)
526{
527 INIT_LIST_HEAD(&timer->cb_entry);
528}
529
530/*
531 * When High resolution timers are active, try to reprogram. Note, that in case
532 * the state has HRTIMER_STATE_CALLBACK set, no reprogramming and no expiry
533 * check happens. The timer gets enqueued into the rbtree. The reprogramming
534 * and expiry check is done in the hrtimer_interrupt or in the softirq.
535 */
536static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
537 struct hrtimer_clock_base *base)
538{
539 if (base->cpu_base->hres_active && hrtimer_reprogram(timer, base)) {
540
541 /* Timer is expired, act upon the callback mode */
542 switch(timer->cb_mode) {
543 case HRTIMER_CB_IRQSAFE_NO_RESTART:
544 /*
545 * We can call the callback from here. No restart
546 * happens, so no danger of recursion
547 */
548 BUG_ON(timer->function(timer) != HRTIMER_NORESTART);
549 return 1;
550 case HRTIMER_CB_IRQSAFE_NO_SOFTIRQ:
551 /*
552 * This is solely for the sched tick emulation with
553 * dynamic tick support to ensure that we do not
554 * restart the tick right on the edge and end up with
555 * the tick timer in the softirq ! The calling site
556 * takes care of this.
557 */
558 return 1;
559 case HRTIMER_CB_IRQSAFE:
560 case HRTIMER_CB_SOFTIRQ:
561 /*
562 * Move everything else into the softirq pending list !
563 */
564 list_add_tail(&timer->cb_entry,
565 &base->cpu_base->cb_pending);
566 timer->state = HRTIMER_STATE_PENDING;
567 raise_softirq(HRTIMER_SOFTIRQ);
568 return 1;
569 default:
570 BUG();
571 }
572 }
573 return 0;
574}
575
576/*
577 * Switch to high resolution mode
578 */
579static int hrtimer_switch_to_hres(void)
580{
581 int cpu = smp_processor_id();
582 struct hrtimer_cpu_base *base = &per_cpu(hrtimer_bases, cpu);
583 unsigned long flags;
584
585 if (base->hres_active)
586 return 1;
587
588 local_irq_save(flags);
589
590 if (tick_init_highres()) {
591 local_irq_restore(flags);
592 printk(KERN_WARNING "Could not switch to high resolution "
593 "mode on CPU %d\n", cpu);
594 return 0;
595 }
596 base->hres_active = 1;
597 base->clock_base[CLOCK_REALTIME].resolution = KTIME_HIGH_RES;
598 base->clock_base[CLOCK_MONOTONIC].resolution = KTIME_HIGH_RES;
599
600 tick_setup_sched_timer();
601
602 /* "Retrigger" the interrupt to get things going */
603 retrigger_next_event(NULL);
604 local_irq_restore(flags);
605 printk(KERN_INFO "Switched to high resolution mode on CPU %d\n",
606 smp_processor_id());
607 return 1;
608}
609
610#else
611
612static inline int hrtimer_hres_active(void) { return 0; }
613static inline int hrtimer_is_hres_enabled(void) { return 0; }
614static inline int hrtimer_switch_to_hres(void) { return 0; }
615static inline void hrtimer_force_reprogram(struct hrtimer_cpu_base *base) { }
616static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
617 struct hrtimer_clock_base *base)
618{
619 return 0;
620}
621static inline int hrtimer_cb_pending(struct hrtimer *timer) { return 0; }
622static inline void hrtimer_remove_cb_pending(struct hrtimer *timer) { }
623static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
624static inline void hrtimer_init_timer_hres(struct hrtimer *timer) { }
625
626#endif /* CONFIG_HIGH_RES_TIMERS */
627
628#ifdef CONFIG_TIMER_STATS
629void __timer_stats_hrtimer_set_start_info(struct hrtimer *timer, void *addr)
630{
631 if (timer->start_site)
632 return;
633
634 timer->start_site = addr;
635 memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
636 timer->start_pid = current->pid;
637}
638#endif
639
640/*
641 * Counterpart to lock_hrtimer_base above:
642 */
643static inline
644void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
645{
646 spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
647}
648
649/**
650 * hrtimer_forward - forward the timer expiry
651 * @timer: hrtimer to forward
652 * @now: forward past this time
653 * @interval: the interval to forward
654 *
655 * Forward the timer expiry so it will expire in the future.
656 * Returns the number of overruns.
657 */
658unsigned long
659hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
660{
661 unsigned long orun = 1;
662 ktime_t delta;
663
664 delta = ktime_sub(now, timer->expires);
665
666 if (delta.tv64 < 0)
667 return 0;
668
669 if (interval.tv64 < timer->base->resolution.tv64)
670 interval.tv64 = timer->base->resolution.tv64;
671
672 if (unlikely(delta.tv64 >= interval.tv64)) {
673 s64 incr = ktime_to_ns(interval);
674
675 orun = ktime_divns(delta, incr);
676 timer->expires = ktime_add_ns(timer->expires, incr * orun);
677 if (timer->expires.tv64 > now.tv64)
678 return orun;
679 /*
680 * This (and the ktime_add() below) is the
681 * correction for exact:
682 */
683 orun++;
684 }
685 timer->expires = ktime_add(timer->expires, interval);
686 /*
687 * Make sure, that the result did not wrap with a very large
688 * interval.
689 */
690 if (timer->expires.tv64 < 0)
691 timer->expires = ktime_set(KTIME_SEC_MAX, 0);
692
693 return orun;
694}
695EXPORT_SYMBOL_GPL(hrtimer_forward);
696
697/*
698 * enqueue_hrtimer - internal function to (re)start a timer
699 *
700 * The timer is inserted in expiry order. Insertion into the
701 * red black tree is O(log(n)). Must hold the base lock.
702 */
703static void enqueue_hrtimer(struct hrtimer *timer,
704 struct hrtimer_clock_base *base, int reprogram)
705{
706 struct rb_node **link = &base->active.rb_node;
707 struct rb_node *parent = NULL;
708 struct hrtimer *entry;
709 int leftmost = 1;
710
711 /*
712 * Find the right place in the rbtree:
713 */
714 while (*link) {
715 parent = *link;
716 entry = rb_entry(parent, struct hrtimer, node);
717 /*
718 * We dont care about collisions. Nodes with
719 * the same expiry time stay together.
720 */
721 if (timer->expires.tv64 < entry->expires.tv64) {
722 link = &(*link)->rb_left;
723 } else {
724 link = &(*link)->rb_right;
725 leftmost = 0;
726 }
727 }
728
729 /*
730 * Insert the timer to the rbtree and check whether it
731 * replaces the first pending timer
732 */
733 if (leftmost) {
734 /*
735 * Reprogram the clock event device. When the timer is already
736 * expired hrtimer_enqueue_reprogram has either called the
737 * callback or added it to the pending list and raised the
738 * softirq.
739 *
740 * This is a NOP for !HIGHRES
741 */
742 if (reprogram && hrtimer_enqueue_reprogram(timer, base))
743 return;
744
745 base->first = &timer->node;
746 }
747
748 rb_link_node(&timer->node, parent, link);
749 rb_insert_color(&timer->node, &base->active);
750 /*
751 * HRTIMER_STATE_ENQUEUED is or'ed to the current state to preserve the
752 * state of a possibly running callback.
753 */
754 timer->state |= HRTIMER_STATE_ENQUEUED;
755}
756
757/*
758 * __remove_hrtimer - internal function to remove a timer
759 *
760 * Caller must hold the base lock.
761 *
762 * High resolution timer mode reprograms the clock event device when the
763 * timer is the one which expires next. The caller can disable this by setting
764 * reprogram to zero. This is useful, when the context does a reprogramming
765 * anyway (e.g. timer interrupt)
766 */
767static void __remove_hrtimer(struct hrtimer *timer,
768 struct hrtimer_clock_base *base,
769 unsigned long newstate, int reprogram)
770{
771 /* High res. callback list. NOP for !HIGHRES */
772 if (hrtimer_cb_pending(timer))
773 hrtimer_remove_cb_pending(timer);
774 else {
775 /*
776 * Remove the timer from the rbtree and replace the
777 * first entry pointer if necessary.
778 */
779 if (base->first == &timer->node) {
780 base->first = rb_next(&timer->node);
781 /* Reprogram the clock event device. if enabled */
782 if (reprogram && hrtimer_hres_active())
783 hrtimer_force_reprogram(base->cpu_base);
784 }
785 rb_erase(&timer->node, &base->active);
786 }
787 timer->state = newstate;
788}
789
790/*
791 * remove hrtimer, called with base lock held
792 */
793static inline int
794remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base)
795{
796 if (hrtimer_is_queued(timer)) {
797 int reprogram;
798
799 /*
800 * Remove the timer and force reprogramming when high
801 * resolution mode is active and the timer is on the current
802 * CPU. If we remove a timer on another CPU, reprogramming is
803 * skipped. The interrupt event on this CPU is fired and
804 * reprogramming happens in the interrupt handler. This is a
805 * rare case and less expensive than a smp call.
806 */
807 timer_stats_hrtimer_clear_start_info(timer);
808 reprogram = base->cpu_base == &__get_cpu_var(hrtimer_bases);
809 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE,
810 reprogram);
811 return 1;
812 }
813 return 0;
814}
815
816/**
817 * hrtimer_start - (re)start an relative timer on the current CPU
818 * @timer: the timer to be added
819 * @tim: expiry time
820 * @mode: expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
821 *
822 * Returns:
823 * 0 on success
824 * 1 when the timer was active
825 */
826int
827hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
828{
829 struct hrtimer_clock_base *base, *new_base;
830 unsigned long flags;
831 int ret;
832
833 base = lock_hrtimer_base(timer, &flags);
834
835 /* Remove an active timer from the queue: */
836 ret = remove_hrtimer(timer, base);
837
838 /* Switch the timer base, if necessary: */
839 new_base = switch_hrtimer_base(timer, base);
840
841 if (mode == HRTIMER_MODE_REL) {
842 tim = ktime_add(tim, new_base->get_time());
843 /*
844 * CONFIG_TIME_LOW_RES is a temporary way for architectures
845 * to signal that they simply return xtime in
846 * do_gettimeoffset(). In this case we want to round up by
847 * resolution when starting a relative timer, to avoid short
848 * timeouts. This will go away with the GTOD framework.
849 */
850#ifdef CONFIG_TIME_LOW_RES
851 tim = ktime_add(tim, base->resolution);
852#endif
853 }
854 timer->expires = tim;
855
856 timer_stats_hrtimer_set_start_info(timer);
857
858 /*
859 * Only allow reprogramming if the new base is on this CPU.
860 * (it might still be on another CPU if the timer was pending)
861 */
862 enqueue_hrtimer(timer, new_base,
863 new_base->cpu_base == &__get_cpu_var(hrtimer_bases));
864
865 unlock_hrtimer_base(timer, &flags);
866
867 return ret;
868}
869EXPORT_SYMBOL_GPL(hrtimer_start);
870
871/**
872 * hrtimer_try_to_cancel - try to deactivate a timer
873 * @timer: hrtimer to stop
874 *
875 * Returns:
876 * 0 when the timer was not active
877 * 1 when the timer was active
878 * -1 when the timer is currently excuting the callback function and
879 * cannot be stopped
880 */
881int hrtimer_try_to_cancel(struct hrtimer *timer)
882{
883 struct hrtimer_clock_base *base;
884 unsigned long flags;
885 int ret = -1;
886
887 base = lock_hrtimer_base(timer, &flags);
888
889 if (!hrtimer_callback_running(timer))
890 ret = remove_hrtimer(timer, base);
891
892 unlock_hrtimer_base(timer, &flags);
893
894 return ret;
895
896}
897EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
898
899/**
900 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
901 * @timer: the timer to be cancelled
902 *
903 * Returns:
904 * 0 when the timer was not active
905 * 1 when the timer was active
906 */
907int hrtimer_cancel(struct hrtimer *timer)
908{
909 for (;;) {
910 int ret = hrtimer_try_to_cancel(timer);
911
912 if (ret >= 0)
913 return ret;
914 cpu_relax();
915 }
916}
917EXPORT_SYMBOL_GPL(hrtimer_cancel);
918
919/**
920 * hrtimer_get_remaining - get remaining time for the timer
921 * @timer: the timer to read
922 */
923ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
924{
925 struct hrtimer_clock_base *base;
926 unsigned long flags;
927 ktime_t rem;
928
929 base = lock_hrtimer_base(timer, &flags);
930 rem = ktime_sub(timer->expires, base->get_time());
931 unlock_hrtimer_base(timer, &flags);
932
933 return rem;
934}
935EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
936
937#if defined(CONFIG_NO_IDLE_HZ) || defined(CONFIG_NO_HZ)
938/**
939 * hrtimer_get_next_event - get the time until next expiry event
940 *
941 * Returns the delta to the next expiry event or KTIME_MAX if no timer
942 * is pending.
943 */
944ktime_t hrtimer_get_next_event(void)
945{
946 struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
947 struct hrtimer_clock_base *base = cpu_base->clock_base;
948 ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
949 unsigned long flags;
950 int i;
951
952 spin_lock_irqsave(&cpu_base->lock, flags);
953
954 if (!hrtimer_hres_active()) {
955 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
956 struct hrtimer *timer;
957
958 if (!base->first)
959 continue;
960
961 timer = rb_entry(base->first, struct hrtimer, node);
962 delta.tv64 = timer->expires.tv64;
963 delta = ktime_sub(delta, base->get_time());
964 if (delta.tv64 < mindelta.tv64)
965 mindelta.tv64 = delta.tv64;
966 }
967 }
968
969 spin_unlock_irqrestore(&cpu_base->lock, flags);
970
971 if (mindelta.tv64 < 0)
972 mindelta.tv64 = 0;
973 return mindelta;
974}
975#endif
976
977/**
978 * hrtimer_init - initialize a timer to the given clock
979 * @timer: the timer to be initialized
980 * @clock_id: the clock to be used
981 * @mode: timer mode abs/rel
982 */
983void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
984 enum hrtimer_mode mode)
985{
986 struct hrtimer_cpu_base *cpu_base;
987
988 memset(timer, 0, sizeof(struct hrtimer));
989
990 cpu_base = &__raw_get_cpu_var(hrtimer_bases);
991
992 if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
993 clock_id = CLOCK_MONOTONIC;
994
995 timer->base = &cpu_base->clock_base[clock_id];
996 hrtimer_init_timer_hres(timer);
997
998#ifdef CONFIG_TIMER_STATS
999 timer->start_site = NULL;
1000 timer->start_pid = -1;
1001 memset(timer->start_comm, 0, TASK_COMM_LEN);
1002#endif
1003}
1004EXPORT_SYMBOL_GPL(hrtimer_init);
1005
1006/**
1007 * hrtimer_get_res - get the timer resolution for a clock
1008 * @which_clock: which clock to query
1009 * @tp: pointer to timespec variable to store the resolution
1010 *
1011 * Store the resolution of the clock selected by @which_clock in the
1012 * variable pointed to by @tp.
1013 */
1014int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
1015{
1016 struct hrtimer_cpu_base *cpu_base;
1017
1018 cpu_base = &__raw_get_cpu_var(hrtimer_bases);
1019 *tp = ktime_to_timespec(cpu_base->clock_base[which_clock].resolution);
1020
1021 return 0;
1022}
1023EXPORT_SYMBOL_GPL(hrtimer_get_res);
1024
1025#ifdef CONFIG_HIGH_RES_TIMERS
1026
1027/*
1028 * High resolution timer interrupt
1029 * Called with interrupts disabled
1030 */
1031void hrtimer_interrupt(struct clock_event_device *dev)
1032{
1033 struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1034 struct hrtimer_clock_base *base;
1035 ktime_t expires_next, now;
1036 int i, raise = 0;
1037
1038 BUG_ON(!cpu_base->hres_active);
1039 cpu_base->nr_events++;
1040 dev->next_event.tv64 = KTIME_MAX;
1041
1042 retry:
1043 now = ktime_get();
1044
1045 expires_next.tv64 = KTIME_MAX;
1046
1047 base = cpu_base->clock_base;
1048
1049 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1050 ktime_t basenow;
1051 struct rb_node *node;
1052
1053 spin_lock(&cpu_base->lock);
1054
1055 basenow = ktime_add(now, base->offset);
1056
1057 while ((node = base->first)) {
1058 struct hrtimer *timer;
1059
1060 timer = rb_entry(node, struct hrtimer, node);
1061
1062 if (basenow.tv64 < timer->expires.tv64) {
1063 ktime_t expires;
1064
1065 expires = ktime_sub(timer->expires,
1066 base->offset);
1067 if (expires.tv64 < expires_next.tv64)
1068 expires_next = expires;
1069 break;
1070 }
1071
1072 /* Move softirq callbacks to the pending list */
1073 if (timer->cb_mode == HRTIMER_CB_SOFTIRQ) {
1074 __remove_hrtimer(timer, base,
1075 HRTIMER_STATE_PENDING, 0);
1076 list_add_tail(&timer->cb_entry,
1077 &base->cpu_base->cb_pending);
1078 raise = 1;
1079 continue;
1080 }
1081
1082 __remove_hrtimer(timer, base,
1083 HRTIMER_STATE_CALLBACK, 0);
1084 timer_stats_account_hrtimer(timer);
1085
1086 /*
1087 * Note: We clear the CALLBACK bit after
1088 * enqueue_hrtimer to avoid reprogramming of
1089 * the event hardware. This happens at the end
1090 * of this function anyway.
1091 */
1092 if (timer->function(timer) != HRTIMER_NORESTART) {
1093 BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
1094 enqueue_hrtimer(timer, base, 0);
1095 }
1096 timer->state &= ~HRTIMER_STATE_CALLBACK;
1097 }
1098 spin_unlock(&cpu_base->lock);
1099 base++;
1100 }
1101
1102 cpu_base->expires_next = expires_next;
1103
1104 /* Reprogramming necessary ? */
1105 if (expires_next.tv64 != KTIME_MAX) {
1106 if (tick_program_event(expires_next, 0))
1107 goto retry;
1108 }
1109
1110 /* Raise softirq ? */
1111 if (raise)
1112 raise_softirq(HRTIMER_SOFTIRQ);
1113}
1114
1115static void run_hrtimer_softirq(struct softirq_action *h)
1116{
1117 struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1118
1119 spin_lock_irq(&cpu_base->lock);
1120
1121 while (!list_empty(&cpu_base->cb_pending)) {
1122 enum hrtimer_restart (*fn)(struct hrtimer *);
1123 struct hrtimer *timer;
1124 int restart;
1125
1126 timer = list_entry(cpu_base->cb_pending.next,
1127 struct hrtimer, cb_entry);
1128
1129 timer_stats_account_hrtimer(timer);
1130
1131 fn = timer->function;
1132 __remove_hrtimer(timer, timer->base, HRTIMER_STATE_CALLBACK, 0);
1133 spin_unlock_irq(&cpu_base->lock);
1134
1135 restart = fn(timer);
1136
1137 spin_lock_irq(&cpu_base->lock);
1138
1139 timer->state &= ~HRTIMER_STATE_CALLBACK;
1140 if (restart == HRTIMER_RESTART) {
1141 BUG_ON(hrtimer_active(timer));
1142 /*
1143 * Enqueue the timer, allow reprogramming of the event
1144 * device
1145 */
1146 enqueue_hrtimer(timer, timer->base, 1);
1147 } else if (hrtimer_active(timer)) {
1148 /*
1149 * If the timer was rearmed on another CPU, reprogram
1150 * the event device.
1151 */
1152 if (timer->base->first == &timer->node)
1153 hrtimer_reprogram(timer, timer->base);
1154 }
1155 }
1156 spin_unlock_irq(&cpu_base->lock);
1157}
1158
1159#endif /* CONFIG_HIGH_RES_TIMERS */
1160
1161/*
1162 * Expire the per base hrtimer-queue:
1163 */
1164static inline void run_hrtimer_queue(struct hrtimer_cpu_base *cpu_base,
1165 int index)
1166{
1167 struct rb_node *node;
1168 struct hrtimer_clock_base *base = &cpu_base->clock_base[index];
1169
1170 if (!base->first)
1171 return;
1172
1173 if (base->get_softirq_time)
1174 base->softirq_time = base->get_softirq_time();
1175
1176 spin_lock_irq(&cpu_base->lock);
1177
1178 while ((node = base->first)) {
1179 struct hrtimer *timer;
1180 enum hrtimer_restart (*fn)(struct hrtimer *);
1181 int restart;
1182
1183 timer = rb_entry(node, struct hrtimer, node);
1184 if (base->softirq_time.tv64 <= timer->expires.tv64)
1185 break;
1186
1187#ifdef CONFIG_HIGH_RES_TIMERS
1188 WARN_ON_ONCE(timer->cb_mode == HRTIMER_CB_IRQSAFE_NO_SOFTIRQ);
1189#endif
1190 timer_stats_account_hrtimer(timer);
1191
1192 fn = timer->function;
1193 __remove_hrtimer(timer, base, HRTIMER_STATE_CALLBACK, 0);
1194 spin_unlock_irq(&cpu_base->lock);
1195
1196 restart = fn(timer);
1197
1198 spin_lock_irq(&cpu_base->lock);
1199
1200 timer->state &= ~HRTIMER_STATE_CALLBACK;
1201 if (restart != HRTIMER_NORESTART) {
1202 BUG_ON(hrtimer_active(timer));
1203 enqueue_hrtimer(timer, base, 0);
1204 }
1205 }
1206 spin_unlock_irq(&cpu_base->lock);
1207}
1208
1209/*
1210 * Called from timer softirq every jiffy, expire hrtimers:
1211 *
1212 * For HRT its the fall back code to run the softirq in the timer
1213 * softirq context in case the hrtimer initialization failed or has
1214 * not been done yet.
1215 */
1216void hrtimer_run_queues(void)
1217{
1218 struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1219 int i;
1220
1221 if (hrtimer_hres_active())
1222 return;
1223
1224 /*
1225 * This _is_ ugly: We have to check in the softirq context,
1226 * whether we can switch to highres and / or nohz mode. The
1227 * clocksource switch happens in the timer interrupt with
1228 * xtime_lock held. Notification from there only sets the
1229 * check bit in the tick_oneshot code, otherwise we might
1230 * deadlock vs. xtime_lock.
1231 */
1232 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled()))
1233 if (hrtimer_switch_to_hres())
1234 return;
1235
1236 hrtimer_get_softirq_time(cpu_base);
1237
1238 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
1239 run_hrtimer_queue(cpu_base, i);
1240}
1241
1242/*
1243 * Sleep related functions:
1244 */
1245static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1246{
1247 struct hrtimer_sleeper *t =
1248 container_of(timer, struct hrtimer_sleeper, timer);
1249 struct task_struct *task = t->task;
1250
1251 t->task = NULL;
1252 if (task)
1253 wake_up_process(task);
1254
1255 return HRTIMER_NORESTART;
1256}
1257
1258void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1259{
1260 sl->timer.function = hrtimer_wakeup;
1261 sl->task = task;
1262#ifdef CONFIG_HIGH_RES_TIMERS
1263 sl->timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_RESTART;
1264#endif
1265}
1266
1267static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1268{
1269 hrtimer_init_sleeper(t, current);
1270
1271 do {
1272 set_current_state(TASK_INTERRUPTIBLE);
1273 hrtimer_start(&t->timer, t->timer.expires, mode);
1274
1275 if (likely(t->task))
1276 schedule();
1277
1278 hrtimer_cancel(&t->timer);
1279 mode = HRTIMER_MODE_ABS;
1280
1281 } while (t->task && !signal_pending(current));
1282
1283 return t->task == NULL;
1284}
1285
1286long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1287{
1288 struct hrtimer_sleeper t;
1289 struct timespec *rmtp;
1290 ktime_t time;
1291
1292 restart->fn = do_no_restart_syscall;
1293
1294 hrtimer_init(&t.timer, restart->arg0, HRTIMER_MODE_ABS);
1295 t.timer.expires.tv64 = ((u64)restart->arg3 << 32) | (u64) restart->arg2;
1296
1297 if (do_nanosleep(&t, HRTIMER_MODE_ABS))
1298 return 0;
1299
1300 rmtp = (struct timespec *)restart->arg1;
1301 if (rmtp) {
1302 time = ktime_sub(t.timer.expires, t.timer.base->get_time());
1303 if (time.tv64 <= 0)
1304 return 0;
1305 *rmtp = ktime_to_timespec(time);
1306 }
1307
1308 restart->fn = hrtimer_nanosleep_restart;
1309
1310 /* The other values in restart are already filled in */
1311 return -ERESTART_RESTARTBLOCK;
1312}
1313
1314long hrtimer_nanosleep(struct timespec *rqtp, struct timespec *rmtp,
1315 const enum hrtimer_mode mode, const clockid_t clockid)
1316{
1317 struct restart_block *restart;
1318 struct hrtimer_sleeper t;
1319 ktime_t rem;
1320
1321 hrtimer_init(&t.timer, clockid, mode);
1322 t.timer.expires = timespec_to_ktime(*rqtp);
1323 if (do_nanosleep(&t, mode))
1324 return 0;
1325
1326 /* Absolute timers do not update the rmtp value and restart: */
1327 if (mode == HRTIMER_MODE_ABS)
1328 return -ERESTARTNOHAND;
1329
1330 if (rmtp) {
1331 rem = ktime_sub(t.timer.expires, t.timer.base->get_time());
1332 if (rem.tv64 <= 0)
1333 return 0;
1334 *rmtp = ktime_to_timespec(rem);
1335 }
1336
1337 restart = &current_thread_info()->restart_block;
1338 restart->fn = hrtimer_nanosleep_restart;
1339 restart->arg0 = (unsigned long) t.timer.base->index;
1340 restart->arg1 = (unsigned long) rmtp;
1341 restart->arg2 = t.timer.expires.tv64 & 0xFFFFFFFF;
1342 restart->arg3 = t.timer.expires.tv64 >> 32;
1343
1344 return -ERESTART_RESTARTBLOCK;
1345}
1346
1347asmlinkage long
1348sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
1349{
1350 struct timespec tu, rmt;
1351 int ret;
1352
1353 if (copy_from_user(&tu, rqtp, sizeof(tu)))
1354 return -EFAULT;
1355
1356 if (!timespec_valid(&tu))
1357 return -EINVAL;
1358
1359 ret = hrtimer_nanosleep(&tu, rmtp ? &rmt : NULL, HRTIMER_MODE_REL,
1360 CLOCK_MONOTONIC);
1361
1362 if (ret && rmtp) {
1363 if (copy_to_user(rmtp, &rmt, sizeof(*rmtp)))
1364 return -EFAULT;
1365 }
1366
1367 return ret;
1368}
1369
1370/*
1371 * Functions related to boot-time initialization:
1372 */
1373static void __devinit init_hrtimers_cpu(int cpu)
1374{
1375 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1376 int i;
1377
1378 spin_lock_init(&cpu_base->lock);
1379 lockdep_set_class(&cpu_base->lock, &cpu_base->lock_key);
1380
1381 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
1382 cpu_base->clock_base[i].cpu_base = cpu_base;
1383
1384 hrtimer_init_hres(cpu_base);
1385}
1386
1387#ifdef CONFIG_HOTPLUG_CPU
1388
1389static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1390 struct hrtimer_clock_base *new_base)
1391{
1392 struct hrtimer *timer;
1393 struct rb_node *node;
1394
1395 while ((node = rb_first(&old_base->active))) {
1396 timer = rb_entry(node, struct hrtimer, node);
1397 BUG_ON(hrtimer_callback_running(timer));
1398 __remove_hrtimer(timer, old_base, HRTIMER_STATE_INACTIVE, 0);
1399 timer->base = new_base;
1400 /*
1401 * Enqueue the timer. Allow reprogramming of the event device
1402 */
1403 enqueue_hrtimer(timer, new_base, 1);
1404 }
1405}
1406
1407static void migrate_hrtimers(int cpu)
1408{
1409 struct hrtimer_cpu_base *old_base, *new_base;
1410 int i;
1411
1412 BUG_ON(cpu_online(cpu));
1413 old_base = &per_cpu(hrtimer_bases, cpu);
1414 new_base = &get_cpu_var(hrtimer_bases);
1415
1416 tick_cancel_sched_timer(cpu);
1417
1418 local_irq_disable();
1419 double_spin_lock(&new_base->lock, &old_base->lock,
1420 smp_processor_id() < cpu);
1421
1422 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1423 migrate_hrtimer_list(&old_base->clock_base[i],
1424 &new_base->clock_base[i]);
1425 }
1426
1427 double_spin_unlock(&new_base->lock, &old_base->lock,
1428 smp_processor_id() < cpu);
1429 local_irq_enable();
1430 put_cpu_var(hrtimer_bases);
1431}
1432#endif /* CONFIG_HOTPLUG_CPU */
1433
1434static int __cpuinit hrtimer_cpu_notify(struct notifier_block *self,
1435 unsigned long action, void *hcpu)
1436{
1437 unsigned int cpu = (long)hcpu;
1438
1439 switch (action) {
1440
1441 case CPU_UP_PREPARE:
1442 case CPU_UP_PREPARE_FROZEN:
1443 init_hrtimers_cpu(cpu);
1444 break;
1445
1446#ifdef CONFIG_HOTPLUG_CPU
1447 case CPU_DEAD:
1448 case CPU_DEAD_FROZEN:
1449 clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &cpu);
1450 migrate_hrtimers(cpu);
1451 break;
1452#endif
1453
1454 default:
1455 break;
1456 }
1457
1458 return NOTIFY_OK;
1459}
1460
1461static struct notifier_block __cpuinitdata hrtimers_nb = {
1462 .notifier_call = hrtimer_cpu_notify,
1463};
1464
1465void __init hrtimers_init(void)
1466{
1467 hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
1468 (void *)(long)smp_processor_id());
1469 register_cpu_notifier(&hrtimers_nb);
1470#ifdef CONFIG_HIGH_RES_TIMERS
1471 open_softirq(HRTIMER_SOFTIRQ, run_hrtimer_softirq, NULL);
1472#endif
1473}
1474