sparseirq: remove some debug print out
[linux-2.6.git] / kernel / hrtimer.c
1 /*
2  *  linux/kernel/hrtimer.c
3  *
4  *  Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
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 #include <linux/debugobjects.h>
47
48 #include <asm/uaccess.h>
49
50 /**
51  * ktime_get - get the monotonic time in ktime_t format
52  *
53  * returns the time in ktime_t format
54  */
55 ktime_t ktime_get(void)
56 {
57         struct timespec now;
58
59         ktime_get_ts(&now);
60
61         return timespec_to_ktime(now);
62 }
63 EXPORT_SYMBOL_GPL(ktime_get);
64
65 /**
66  * ktime_get_real - get the real (wall-) time in ktime_t format
67  *
68  * returns the time in ktime_t format
69  */
70 ktime_t ktime_get_real(void)
71 {
72         struct timespec now;
73
74         getnstimeofday(&now);
75
76         return timespec_to_ktime(now);
77 }
78
79 EXPORT_SYMBOL_GPL(ktime_get_real);
80
81 /*
82  * The timer bases:
83  *
84  * Note: If we want to add new timer bases, we have to skip the two
85  * clock ids captured by the cpu-timers. We do this by holding empty
86  * entries rather than doing math adjustment of the clock ids.
87  * This ensures that we capture erroneous accesses to these clock ids
88  * rather than moving them into the range of valid clock id's.
89  */
90 DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
91 {
92
93         .clock_base =
94         {
95                 {
96                         .index = CLOCK_REALTIME,
97                         .get_time = &ktime_get_real,
98                         .resolution = KTIME_LOW_RES,
99                 },
100                 {
101                         .index = CLOCK_MONOTONIC,
102                         .get_time = &ktime_get,
103                         .resolution = KTIME_LOW_RES,
104                 },
105         }
106 };
107
108 /**
109  * ktime_get_ts - get the monotonic clock in timespec format
110  * @ts:         pointer to timespec variable
111  *
112  * The function calculates the monotonic clock from the realtime
113  * clock and the wall_to_monotonic offset and stores the result
114  * in normalized timespec format in the variable pointed to by @ts.
115  */
116 void ktime_get_ts(struct timespec *ts)
117 {
118         struct timespec tomono;
119         unsigned long seq;
120
121         do {
122                 seq = read_seqbegin(&xtime_lock);
123                 getnstimeofday(ts);
124                 tomono = wall_to_monotonic;
125
126         } while (read_seqretry(&xtime_lock, seq));
127
128         set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
129                                 ts->tv_nsec + tomono.tv_nsec);
130 }
131 EXPORT_SYMBOL_GPL(ktime_get_ts);
132
133 /*
134  * Get the coarse grained time at the softirq based on xtime and
135  * wall_to_monotonic.
136  */
137 static void hrtimer_get_softirq_time(struct hrtimer_cpu_base *base)
138 {
139         ktime_t xtim, tomono;
140         struct timespec xts, tom;
141         unsigned long seq;
142
143         do {
144                 seq = read_seqbegin(&xtime_lock);
145                 xts = current_kernel_time();
146                 tom = wall_to_monotonic;
147         } while (read_seqretry(&xtime_lock, seq));
148
149         xtim = timespec_to_ktime(xts);
150         tomono = timespec_to_ktime(tom);
151         base->clock_base[CLOCK_REALTIME].softirq_time = xtim;
152         base->clock_base[CLOCK_MONOTONIC].softirq_time =
153                 ktime_add(xtim, tomono);
154 }
155
156 /*
157  * Functions and macros which are different for UP/SMP systems are kept in a
158  * single place
159  */
160 #ifdef CONFIG_SMP
161
162 /*
163  * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
164  * means that all timers which are tied to this base via timer->base are
165  * locked, and the base itself is locked too.
166  *
167  * So __run_timers/migrate_timers can safely modify all timers which could
168  * be found on the lists/queues.
169  *
170  * When the timer's base is locked, and the timer removed from list, it is
171  * possible to set timer->base = NULL and drop the lock: the timer remains
172  * locked.
173  */
174 static
175 struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
176                                              unsigned long *flags)
177 {
178         struct hrtimer_clock_base *base;
179
180         for (;;) {
181                 base = timer->base;
182                 if (likely(base != NULL)) {
183                         spin_lock_irqsave(&base->cpu_base->lock, *flags);
184                         if (likely(base == timer->base))
185                                 return base;
186                         /* The timer has migrated to another CPU: */
187                         spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
188                 }
189                 cpu_relax();
190         }
191 }
192
193 /*
194  * Switch the timer base to the current CPU when possible.
195  */
196 static inline struct hrtimer_clock_base *
197 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base)
198 {
199         struct hrtimer_clock_base *new_base;
200         struct hrtimer_cpu_base *new_cpu_base;
201
202         new_cpu_base = &__get_cpu_var(hrtimer_bases);
203         new_base = &new_cpu_base->clock_base[base->index];
204
205         if (base != new_base) {
206                 /*
207                  * We are trying to schedule the timer on the local CPU.
208                  * However we can't change timer's base while it is running,
209                  * so we keep it on the same CPU. No hassle vs. reprogramming
210                  * the event source in the high resolution case. The softirq
211                  * code will take care of this when the timer function has
212                  * completed. There is no conflict as we hold the lock until
213                  * the timer is enqueued.
214                  */
215                 if (unlikely(hrtimer_callback_running(timer)))
216                         return base;
217
218                 /* See the comment in lock_timer_base() */
219                 timer->base = NULL;
220                 spin_unlock(&base->cpu_base->lock);
221                 spin_lock(&new_base->cpu_base->lock);
222                 timer->base = new_base;
223         }
224         return new_base;
225 }
226
227 #else /* CONFIG_SMP */
228
229 static inline struct hrtimer_clock_base *
230 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
231 {
232         struct hrtimer_clock_base *base = timer->base;
233
234         spin_lock_irqsave(&base->cpu_base->lock, *flags);
235
236         return base;
237 }
238
239 # define switch_hrtimer_base(t, b)      (b)
240
241 #endif  /* !CONFIG_SMP */
242
243 /*
244  * Functions for the union type storage format of ktime_t which are
245  * too large for inlining:
246  */
247 #if BITS_PER_LONG < 64
248 # ifndef CONFIG_KTIME_SCALAR
249 /**
250  * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
251  * @kt:         addend
252  * @nsec:       the scalar nsec value to add
253  *
254  * Returns the sum of kt and nsec in ktime_t format
255  */
256 ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
257 {
258         ktime_t tmp;
259
260         if (likely(nsec < NSEC_PER_SEC)) {
261                 tmp.tv64 = nsec;
262         } else {
263                 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
264
265                 tmp = ktime_set((long)nsec, rem);
266         }
267
268         return ktime_add(kt, tmp);
269 }
270
271 EXPORT_SYMBOL_GPL(ktime_add_ns);
272
273 /**
274  * ktime_sub_ns - Subtract a scalar nanoseconds value from a ktime_t variable
275  * @kt:         minuend
276  * @nsec:       the scalar nsec value to subtract
277  *
278  * Returns the subtraction of @nsec from @kt in ktime_t format
279  */
280 ktime_t ktime_sub_ns(const ktime_t kt, u64 nsec)
281 {
282         ktime_t tmp;
283
284         if (likely(nsec < NSEC_PER_SEC)) {
285                 tmp.tv64 = nsec;
286         } else {
287                 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
288
289                 tmp = ktime_set((long)nsec, rem);
290         }
291
292         return ktime_sub(kt, tmp);
293 }
294
295 EXPORT_SYMBOL_GPL(ktime_sub_ns);
296 # endif /* !CONFIG_KTIME_SCALAR */
297
298 /*
299  * Divide a ktime value by a nanosecond value
300  */
301 u64 ktime_divns(const ktime_t kt, s64 div)
302 {
303         u64 dclc;
304         int sft = 0;
305
306         dclc = ktime_to_ns(kt);
307         /* Make sure the divisor is less than 2^32: */
308         while (div >> 32) {
309                 sft++;
310                 div >>= 1;
311         }
312         dclc >>= sft;
313         do_div(dclc, (unsigned long) div);
314
315         return dclc;
316 }
317 #endif /* BITS_PER_LONG >= 64 */
318
319 /*
320  * Add two ktime values and do a safety check for overflow:
321  */
322 ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
323 {
324         ktime_t res = ktime_add(lhs, rhs);
325
326         /*
327          * We use KTIME_SEC_MAX here, the maximum timeout which we can
328          * return to user space in a timespec:
329          */
330         if (res.tv64 < 0 || res.tv64 < lhs.tv64 || res.tv64 < rhs.tv64)
331                 res = ktime_set(KTIME_SEC_MAX, 0);
332
333         return res;
334 }
335
336 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
337
338 static struct debug_obj_descr hrtimer_debug_descr;
339
340 /*
341  * fixup_init is called when:
342  * - an active object is initialized
343  */
344 static int hrtimer_fixup_init(void *addr, enum debug_obj_state state)
345 {
346         struct hrtimer *timer = addr;
347
348         switch (state) {
349         case ODEBUG_STATE_ACTIVE:
350                 hrtimer_cancel(timer);
351                 debug_object_init(timer, &hrtimer_debug_descr);
352                 return 1;
353         default:
354                 return 0;
355         }
356 }
357
358 /*
359  * fixup_activate is called when:
360  * - an active object is activated
361  * - an unknown object is activated (might be a statically initialized object)
362  */
363 static int hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
364 {
365         switch (state) {
366
367         case ODEBUG_STATE_NOTAVAILABLE:
368                 WARN_ON_ONCE(1);
369                 return 0;
370
371         case ODEBUG_STATE_ACTIVE:
372                 WARN_ON(1);
373
374         default:
375                 return 0;
376         }
377 }
378
379 /*
380  * fixup_free is called when:
381  * - an active object is freed
382  */
383 static int hrtimer_fixup_free(void *addr, enum debug_obj_state state)
384 {
385         struct hrtimer *timer = addr;
386
387         switch (state) {
388         case ODEBUG_STATE_ACTIVE:
389                 hrtimer_cancel(timer);
390                 debug_object_free(timer, &hrtimer_debug_descr);
391                 return 1;
392         default:
393                 return 0;
394         }
395 }
396
397 static struct debug_obj_descr hrtimer_debug_descr = {
398         .name           = "hrtimer",
399         .fixup_init     = hrtimer_fixup_init,
400         .fixup_activate = hrtimer_fixup_activate,
401         .fixup_free     = hrtimer_fixup_free,
402 };
403
404 static inline void debug_hrtimer_init(struct hrtimer *timer)
405 {
406         debug_object_init(timer, &hrtimer_debug_descr);
407 }
408
409 static inline void debug_hrtimer_activate(struct hrtimer *timer)
410 {
411         debug_object_activate(timer, &hrtimer_debug_descr);
412 }
413
414 static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
415 {
416         debug_object_deactivate(timer, &hrtimer_debug_descr);
417 }
418
419 static inline void debug_hrtimer_free(struct hrtimer *timer)
420 {
421         debug_object_free(timer, &hrtimer_debug_descr);
422 }
423
424 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
425                            enum hrtimer_mode mode);
426
427 void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
428                            enum hrtimer_mode mode)
429 {
430         debug_object_init_on_stack(timer, &hrtimer_debug_descr);
431         __hrtimer_init(timer, clock_id, mode);
432 }
433
434 void destroy_hrtimer_on_stack(struct hrtimer *timer)
435 {
436         debug_object_free(timer, &hrtimer_debug_descr);
437 }
438
439 #else
440 static inline void debug_hrtimer_init(struct hrtimer *timer) { }
441 static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
442 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
443 #endif
444
445 /*
446  * Check, whether the timer is on the callback pending list
447  */
448 static inline int hrtimer_cb_pending(const struct hrtimer *timer)
449 {
450         return timer->state & HRTIMER_STATE_PENDING;
451 }
452
453 /*
454  * Remove a timer from the callback pending list
455  */
456 static inline void hrtimer_remove_cb_pending(struct hrtimer *timer)
457 {
458         list_del_init(&timer->cb_entry);
459 }
460
461 /* High resolution timer related functions */
462 #ifdef CONFIG_HIGH_RES_TIMERS
463
464 /*
465  * High resolution timer enabled ?
466  */
467 static int hrtimer_hres_enabled __read_mostly  = 1;
468
469 /*
470  * Enable / Disable high resolution mode
471  */
472 static int __init setup_hrtimer_hres(char *str)
473 {
474         if (!strcmp(str, "off"))
475                 hrtimer_hres_enabled = 0;
476         else if (!strcmp(str, "on"))
477                 hrtimer_hres_enabled = 1;
478         else
479                 return 0;
480         return 1;
481 }
482
483 __setup("highres=", setup_hrtimer_hres);
484
485 /*
486  * hrtimer_high_res_enabled - query, if the highres mode is enabled
487  */
488 static inline int hrtimer_is_hres_enabled(void)
489 {
490         return hrtimer_hres_enabled;
491 }
492
493 /*
494  * Is the high resolution mode active ?
495  */
496 static inline int hrtimer_hres_active(void)
497 {
498         return __get_cpu_var(hrtimer_bases).hres_active;
499 }
500
501 /*
502  * Reprogram the event source with checking both queues for the
503  * next event
504  * Called with interrupts disabled and base->lock held
505  */
506 static void hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base)
507 {
508         int i;
509         struct hrtimer_clock_base *base = cpu_base->clock_base;
510         ktime_t expires;
511
512         cpu_base->expires_next.tv64 = KTIME_MAX;
513
514         for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
515                 struct hrtimer *timer;
516
517                 if (!base->first)
518                         continue;
519                 timer = rb_entry(base->first, struct hrtimer, node);
520                 expires = ktime_sub(timer->expires, base->offset);
521                 if (expires.tv64 < cpu_base->expires_next.tv64)
522                         cpu_base->expires_next = expires;
523         }
524
525         if (cpu_base->expires_next.tv64 != KTIME_MAX)
526                 tick_program_event(cpu_base->expires_next, 1);
527 }
528
529 /*
530  * Shared reprogramming for clock_realtime and clock_monotonic
531  *
532  * When a timer is enqueued and expires earlier than the already enqueued
533  * timers, we have to check, whether it expires earlier than the timer for
534  * which the clock event device was armed.
535  *
536  * Called with interrupts disabled and base->cpu_base.lock held
537  */
538 static int hrtimer_reprogram(struct hrtimer *timer,
539                              struct hrtimer_clock_base *base)
540 {
541         ktime_t *expires_next = &__get_cpu_var(hrtimer_bases).expires_next;
542         ktime_t expires = ktime_sub(timer->expires, base->offset);
543         int res;
544
545         WARN_ON_ONCE(timer->expires.tv64 < 0);
546
547         /*
548          * When the callback is running, we do not reprogram the clock event
549          * device. The timer callback is either running on a different CPU or
550          * the callback is executed in the hrtimer_interrupt context. The
551          * reprogramming is handled either by the softirq, which called the
552          * callback or at the end of the hrtimer_interrupt.
553          */
554         if (hrtimer_callback_running(timer))
555                 return 0;
556
557         /*
558          * CLOCK_REALTIME timer might be requested with an absolute
559          * expiry time which is less than base->offset. Nothing wrong
560          * about that, just avoid to call into the tick code, which
561          * has now objections against negative expiry values.
562          */
563         if (expires.tv64 < 0)
564                 return -ETIME;
565
566         if (expires.tv64 >= expires_next->tv64)
567                 return 0;
568
569         /*
570          * Clockevents returns -ETIME, when the event was in the past.
571          */
572         res = tick_program_event(expires, 0);
573         if (!IS_ERR_VALUE(res))
574                 *expires_next = expires;
575         return res;
576 }
577
578
579 /*
580  * Retrigger next event is called after clock was set
581  *
582  * Called with interrupts disabled via on_each_cpu()
583  */
584 static void retrigger_next_event(void *arg)
585 {
586         struct hrtimer_cpu_base *base;
587         struct timespec realtime_offset;
588         unsigned long seq;
589
590         if (!hrtimer_hres_active())
591                 return;
592
593         do {
594                 seq = read_seqbegin(&xtime_lock);
595                 set_normalized_timespec(&realtime_offset,
596                                         -wall_to_monotonic.tv_sec,
597                                         -wall_to_monotonic.tv_nsec);
598         } while (read_seqretry(&xtime_lock, seq));
599
600         base = &__get_cpu_var(hrtimer_bases);
601
602         /* Adjust CLOCK_REALTIME offset */
603         spin_lock(&base->lock);
604         base->clock_base[CLOCK_REALTIME].offset =
605                 timespec_to_ktime(realtime_offset);
606
607         hrtimer_force_reprogram(base);
608         spin_unlock(&base->lock);
609 }
610
611 /*
612  * Clock realtime was set
613  *
614  * Change the offset of the realtime clock vs. the monotonic
615  * clock.
616  *
617  * We might have to reprogram the high resolution timer interrupt. On
618  * SMP we call the architecture specific code to retrigger _all_ high
619  * resolution timer interrupts. On UP we just disable interrupts and
620  * call the high resolution interrupt code.
621  */
622 void clock_was_set(void)
623 {
624         /* Retrigger the CPU local events everywhere */
625         on_each_cpu(retrigger_next_event, NULL, 1);
626 }
627
628 /*
629  * During resume we might have to reprogram the high resolution timer
630  * interrupt (on the local CPU):
631  */
632 void hres_timers_resume(void)
633 {
634         /* Retrigger the CPU local events: */
635         retrigger_next_event(NULL);
636 }
637
638 /*
639  * Initialize the high resolution related parts of cpu_base
640  */
641 static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
642 {
643         base->expires_next.tv64 = KTIME_MAX;
644         base->hres_active = 0;
645 }
646
647 /*
648  * Initialize the high resolution related parts of a hrtimer
649  */
650 static inline void hrtimer_init_timer_hres(struct hrtimer *timer)
651 {
652 }
653
654 /*
655  * When High resolution timers are active, try to reprogram. Note, that in case
656  * the state has HRTIMER_STATE_CALLBACK set, no reprogramming and no expiry
657  * check happens. The timer gets enqueued into the rbtree. The reprogramming
658  * and expiry check is done in the hrtimer_interrupt or in the softirq.
659  */
660 static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
661                                             struct hrtimer_clock_base *base)
662 {
663         if (base->cpu_base->hres_active && hrtimer_reprogram(timer, base)) {
664
665                 /* Timer is expired, act upon the callback mode */
666                 switch(timer->cb_mode) {
667                 case HRTIMER_CB_IRQSAFE_NO_RESTART:
668                         debug_hrtimer_deactivate(timer);
669                         /*
670                          * We can call the callback from here. No restart
671                          * happens, so no danger of recursion
672                          */
673                         BUG_ON(timer->function(timer) != HRTIMER_NORESTART);
674                         return 1;
675                 case HRTIMER_CB_IRQSAFE_PERCPU:
676                 case HRTIMER_CB_IRQSAFE_UNLOCKED:
677                         /*
678                          * This is solely for the sched tick emulation with
679                          * dynamic tick support to ensure that we do not
680                          * restart the tick right on the edge and end up with
681                          * the tick timer in the softirq ! The calling site
682                          * takes care of this. Also used for hrtimer sleeper !
683                          */
684                         debug_hrtimer_deactivate(timer);
685                         return 1;
686                 case HRTIMER_CB_IRQSAFE:
687                 case HRTIMER_CB_SOFTIRQ:
688                         /*
689                          * Move everything else into the softirq pending list !
690                          */
691                         list_add_tail(&timer->cb_entry,
692                                       &base->cpu_base->cb_pending);
693                         timer->state = HRTIMER_STATE_PENDING;
694                         return 1;
695                 default:
696                         BUG();
697                 }
698         }
699         return 0;
700 }
701
702 /*
703  * Switch to high resolution mode
704  */
705 static int hrtimer_switch_to_hres(void)
706 {
707         int cpu = smp_processor_id();
708         struct hrtimer_cpu_base *base = &per_cpu(hrtimer_bases, cpu);
709         unsigned long flags;
710
711         if (base->hres_active)
712                 return 1;
713
714         local_irq_save(flags);
715
716         if (tick_init_highres()) {
717                 local_irq_restore(flags);
718                 printk(KERN_WARNING "Could not switch to high resolution "
719                                     "mode on CPU %d\n", cpu);
720                 return 0;
721         }
722         base->hres_active = 1;
723         base->clock_base[CLOCK_REALTIME].resolution = KTIME_HIGH_RES;
724         base->clock_base[CLOCK_MONOTONIC].resolution = KTIME_HIGH_RES;
725
726         tick_setup_sched_timer();
727
728         /* "Retrigger" the interrupt to get things going */
729         retrigger_next_event(NULL);
730         local_irq_restore(flags);
731         printk(KERN_DEBUG "Switched to high resolution mode on CPU %d\n",
732                smp_processor_id());
733         return 1;
734 }
735
736 static inline void hrtimer_raise_softirq(void)
737 {
738         raise_softirq(HRTIMER_SOFTIRQ);
739 }
740
741 #else
742
743 static inline int hrtimer_hres_active(void) { return 0; }
744 static inline int hrtimer_is_hres_enabled(void) { return 0; }
745 static inline int hrtimer_switch_to_hres(void) { return 0; }
746 static inline void hrtimer_force_reprogram(struct hrtimer_cpu_base *base) { }
747 static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
748                                             struct hrtimer_clock_base *base)
749 {
750         return 0;
751 }
752 static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
753 static inline void hrtimer_init_timer_hres(struct hrtimer *timer) { }
754 static inline int hrtimer_reprogram(struct hrtimer *timer,
755                                     struct hrtimer_clock_base *base)
756 {
757         return 0;
758 }
759 static inline void hrtimer_raise_softirq(void) { }
760
761 #endif /* CONFIG_HIGH_RES_TIMERS */
762
763 #ifdef CONFIG_TIMER_STATS
764 void __timer_stats_hrtimer_set_start_info(struct hrtimer *timer, void *addr)
765 {
766         if (timer->start_site)
767                 return;
768
769         timer->start_site = addr;
770         memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
771         timer->start_pid = current->pid;
772 }
773 #endif
774
775 /*
776  * Counterpart to lock_hrtimer_base above:
777  */
778 static inline
779 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
780 {
781         spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
782 }
783
784 /**
785  * hrtimer_forward - forward the timer expiry
786  * @timer:      hrtimer to forward
787  * @now:        forward past this time
788  * @interval:   the interval to forward
789  *
790  * Forward the timer expiry so it will expire in the future.
791  * Returns the number of overruns.
792  */
793 u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
794 {
795         u64 orun = 1;
796         ktime_t delta;
797
798         delta = ktime_sub(now, timer->expires);
799
800         if (delta.tv64 < 0)
801                 return 0;
802
803         if (interval.tv64 < timer->base->resolution.tv64)
804                 interval.tv64 = timer->base->resolution.tv64;
805
806         if (unlikely(delta.tv64 >= interval.tv64)) {
807                 s64 incr = ktime_to_ns(interval);
808
809                 orun = ktime_divns(delta, incr);
810                 timer->expires = ktime_add_ns(timer->expires, incr * orun);
811                 if (timer->expires.tv64 > now.tv64)
812                         return orun;
813                 /*
814                  * This (and the ktime_add() below) is the
815                  * correction for exact:
816                  */
817                 orun++;
818         }
819         timer->expires = ktime_add_safe(timer->expires, interval);
820
821         return orun;
822 }
823 EXPORT_SYMBOL_GPL(hrtimer_forward);
824
825 /*
826  * enqueue_hrtimer - internal function to (re)start a timer
827  *
828  * The timer is inserted in expiry order. Insertion into the
829  * red black tree is O(log(n)). Must hold the base lock.
830  */
831 static void enqueue_hrtimer(struct hrtimer *timer,
832                             struct hrtimer_clock_base *base, int reprogram)
833 {
834         struct rb_node **link = &base->active.rb_node;
835         struct rb_node *parent = NULL;
836         struct hrtimer *entry;
837         int leftmost = 1;
838
839         debug_hrtimer_activate(timer);
840
841         /*
842          * Find the right place in the rbtree:
843          */
844         while (*link) {
845                 parent = *link;
846                 entry = rb_entry(parent, struct hrtimer, node);
847                 /*
848                  * We dont care about collisions. Nodes with
849                  * the same expiry time stay together.
850                  */
851                 if (timer->expires.tv64 < entry->expires.tv64) {
852                         link = &(*link)->rb_left;
853                 } else {
854                         link = &(*link)->rb_right;
855                         leftmost = 0;
856                 }
857         }
858
859         /*
860          * Insert the timer to the rbtree and check whether it
861          * replaces the first pending timer
862          */
863         if (leftmost) {
864                 /*
865                  * Reprogram the clock event device. When the timer is already
866                  * expired hrtimer_enqueue_reprogram has either called the
867                  * callback or added it to the pending list and raised the
868                  * softirq.
869                  *
870                  * This is a NOP for !HIGHRES
871                  */
872                 if (reprogram && hrtimer_enqueue_reprogram(timer, base))
873                         return;
874
875                 base->first = &timer->node;
876         }
877
878         rb_link_node(&timer->node, parent, link);
879         rb_insert_color(&timer->node, &base->active);
880         /*
881          * HRTIMER_STATE_ENQUEUED is or'ed to the current state to preserve the
882          * state of a possibly running callback.
883          */
884         timer->state |= HRTIMER_STATE_ENQUEUED;
885 }
886
887 /*
888  * __remove_hrtimer - internal function to remove a timer
889  *
890  * Caller must hold the base lock.
891  *
892  * High resolution timer mode reprograms the clock event device when the
893  * timer is the one which expires next. The caller can disable this by setting
894  * reprogram to zero. This is useful, when the context does a reprogramming
895  * anyway (e.g. timer interrupt)
896  */
897 static void __remove_hrtimer(struct hrtimer *timer,
898                              struct hrtimer_clock_base *base,
899                              unsigned long newstate, int reprogram)
900 {
901         /* High res. callback list. NOP for !HIGHRES */
902         if (hrtimer_cb_pending(timer))
903                 hrtimer_remove_cb_pending(timer);
904         else {
905                 /*
906                  * Remove the timer from the rbtree and replace the
907                  * first entry pointer if necessary.
908                  */
909                 if (base->first == &timer->node) {
910                         base->first = rb_next(&timer->node);
911                         /* Reprogram the clock event device. if enabled */
912                         if (reprogram && hrtimer_hres_active())
913                                 hrtimer_force_reprogram(base->cpu_base);
914                 }
915                 rb_erase(&timer->node, &base->active);
916         }
917         timer->state = newstate;
918 }
919
920 /*
921  * remove hrtimer, called with base lock held
922  */
923 static inline int
924 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base)
925 {
926         if (hrtimer_is_queued(timer)) {
927                 int reprogram;
928
929                 /*
930                  * Remove the timer and force reprogramming when high
931                  * resolution mode is active and the timer is on the current
932                  * CPU. If we remove a timer on another CPU, reprogramming is
933                  * skipped. The interrupt event on this CPU is fired and
934                  * reprogramming happens in the interrupt handler. This is a
935                  * rare case and less expensive than a smp call.
936                  */
937                 debug_hrtimer_deactivate(timer);
938                 timer_stats_hrtimer_clear_start_info(timer);
939                 reprogram = base->cpu_base == &__get_cpu_var(hrtimer_bases);
940                 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE,
941                                  reprogram);
942                 return 1;
943         }
944         return 0;
945 }
946
947 /**
948  * hrtimer_start - (re)start an relative timer on the current CPU
949  * @timer:      the timer to be added
950  * @tim:        expiry time
951  * @mode:       expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
952  *
953  * Returns:
954  *  0 on success
955  *  1 when the timer was active
956  */
957 int
958 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
959 {
960         struct hrtimer_clock_base *base, *new_base;
961         unsigned long flags;
962         int ret, raise;
963
964         base = lock_hrtimer_base(timer, &flags);
965
966         /* Remove an active timer from the queue: */
967         ret = remove_hrtimer(timer, base);
968
969         /* Switch the timer base, if necessary: */
970         new_base = switch_hrtimer_base(timer, base);
971
972         if (mode == HRTIMER_MODE_REL) {
973                 tim = ktime_add_safe(tim, new_base->get_time());
974                 /*
975                  * CONFIG_TIME_LOW_RES is a temporary way for architectures
976                  * to signal that they simply return xtime in
977                  * do_gettimeoffset(). In this case we want to round up by
978                  * resolution when starting a relative timer, to avoid short
979                  * timeouts. This will go away with the GTOD framework.
980                  */
981 #ifdef CONFIG_TIME_LOW_RES
982                 tim = ktime_add_safe(tim, base->resolution);
983 #endif
984         }
985
986         timer->expires = tim;
987
988         timer_stats_hrtimer_set_start_info(timer);
989
990         /*
991          * Only allow reprogramming if the new base is on this CPU.
992          * (it might still be on another CPU if the timer was pending)
993          */
994         enqueue_hrtimer(timer, new_base,
995                         new_base->cpu_base == &__get_cpu_var(hrtimer_bases));
996
997         /*
998          * The timer may be expired and moved to the cb_pending
999          * list. We can not raise the softirq with base lock held due
1000          * to a possible deadlock with runqueue lock.
1001          */
1002         raise = timer->state == HRTIMER_STATE_PENDING;
1003
1004         /*
1005          * We use preempt_disable to prevent this task from migrating after
1006          * setting up the softirq and raising it. Otherwise, if me migrate
1007          * we will raise the softirq on the wrong CPU.
1008          */
1009         preempt_disable();
1010
1011         unlock_hrtimer_base(timer, &flags);
1012
1013         if (raise)
1014                 hrtimer_raise_softirq();
1015         preempt_enable();
1016
1017         return ret;
1018 }
1019 EXPORT_SYMBOL_GPL(hrtimer_start);
1020
1021 /**
1022  * hrtimer_try_to_cancel - try to deactivate a timer
1023  * @timer:      hrtimer to stop
1024  *
1025  * Returns:
1026  *  0 when the timer was not active
1027  *  1 when the timer was active
1028  * -1 when the timer is currently excuting the callback function and
1029  *    cannot be stopped
1030  */
1031 int hrtimer_try_to_cancel(struct hrtimer *timer)
1032 {
1033         struct hrtimer_clock_base *base;
1034         unsigned long flags;
1035         int ret = -1;
1036
1037         base = lock_hrtimer_base(timer, &flags);
1038
1039         if (!hrtimer_callback_running(timer))
1040                 ret = remove_hrtimer(timer, base);
1041
1042         unlock_hrtimer_base(timer, &flags);
1043
1044         return ret;
1045
1046 }
1047 EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1048
1049 /**
1050  * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1051  * @timer:      the timer to be cancelled
1052  *
1053  * Returns:
1054  *  0 when the timer was not active
1055  *  1 when the timer was active
1056  */
1057 int hrtimer_cancel(struct hrtimer *timer)
1058 {
1059         for (;;) {
1060                 int ret = hrtimer_try_to_cancel(timer);
1061
1062                 if (ret >= 0)
1063                         return ret;
1064                 cpu_relax();
1065         }
1066 }
1067 EXPORT_SYMBOL_GPL(hrtimer_cancel);
1068
1069 /**
1070  * hrtimer_get_remaining - get remaining time for the timer
1071  * @timer:      the timer to read
1072  */
1073 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
1074 {
1075         struct hrtimer_clock_base *base;
1076         unsigned long flags;
1077         ktime_t rem;
1078
1079         base = lock_hrtimer_base(timer, &flags);
1080         rem = ktime_sub(timer->expires, base->get_time());
1081         unlock_hrtimer_base(timer, &flags);
1082
1083         return rem;
1084 }
1085 EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
1086
1087 #ifdef CONFIG_NO_HZ
1088 /**
1089  * hrtimer_get_next_event - get the time until next expiry event
1090  *
1091  * Returns the delta to the next expiry event or KTIME_MAX if no timer
1092  * is pending.
1093  */
1094 ktime_t hrtimer_get_next_event(void)
1095 {
1096         struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1097         struct hrtimer_clock_base *base = cpu_base->clock_base;
1098         ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
1099         unsigned long flags;
1100         int i;
1101
1102         spin_lock_irqsave(&cpu_base->lock, flags);
1103
1104         if (!hrtimer_hres_active()) {
1105                 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
1106                         struct hrtimer *timer;
1107
1108                         if (!base->first)
1109                                 continue;
1110
1111                         timer = rb_entry(base->first, struct hrtimer, node);
1112                         delta.tv64 = timer->expires.tv64;
1113                         delta = ktime_sub(delta, base->get_time());
1114                         if (delta.tv64 < mindelta.tv64)
1115                                 mindelta.tv64 = delta.tv64;
1116                 }
1117         }
1118
1119         spin_unlock_irqrestore(&cpu_base->lock, flags);
1120
1121         if (mindelta.tv64 < 0)
1122                 mindelta.tv64 = 0;
1123         return mindelta;
1124 }
1125 #endif
1126
1127 static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1128                            enum hrtimer_mode mode)
1129 {
1130         struct hrtimer_cpu_base *cpu_base;
1131
1132         memset(timer, 0, sizeof(struct hrtimer));
1133
1134         cpu_base = &__raw_get_cpu_var(hrtimer_bases);
1135
1136         if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
1137                 clock_id = CLOCK_MONOTONIC;
1138
1139         timer->base = &cpu_base->clock_base[clock_id];
1140         INIT_LIST_HEAD(&timer->cb_entry);
1141         hrtimer_init_timer_hres(timer);
1142
1143 #ifdef CONFIG_TIMER_STATS
1144         timer->start_site = NULL;
1145         timer->start_pid = -1;
1146         memset(timer->start_comm, 0, TASK_COMM_LEN);
1147 #endif
1148 }
1149
1150 /**
1151  * hrtimer_init - initialize a timer to the given clock
1152  * @timer:      the timer to be initialized
1153  * @clock_id:   the clock to be used
1154  * @mode:       timer mode abs/rel
1155  */
1156 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1157                   enum hrtimer_mode mode)
1158 {
1159         debug_hrtimer_init(timer);
1160         __hrtimer_init(timer, clock_id, mode);
1161 }
1162 EXPORT_SYMBOL_GPL(hrtimer_init);
1163
1164 /**
1165  * hrtimer_get_res - get the timer resolution for a clock
1166  * @which_clock: which clock to query
1167  * @tp:          pointer to timespec variable to store the resolution
1168  *
1169  * Store the resolution of the clock selected by @which_clock in the
1170  * variable pointed to by @tp.
1171  */
1172 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
1173 {
1174         struct hrtimer_cpu_base *cpu_base;
1175
1176         cpu_base = &__raw_get_cpu_var(hrtimer_bases);
1177         *tp = ktime_to_timespec(cpu_base->clock_base[which_clock].resolution);
1178
1179         return 0;
1180 }
1181 EXPORT_SYMBOL_GPL(hrtimer_get_res);
1182
1183 static void run_hrtimer_pending(struct hrtimer_cpu_base *cpu_base)
1184 {
1185         spin_lock_irq(&cpu_base->lock);
1186
1187         while (!list_empty(&cpu_base->cb_pending)) {
1188                 enum hrtimer_restart (*fn)(struct hrtimer *);
1189                 struct hrtimer *timer;
1190                 int restart;
1191
1192                 timer = list_entry(cpu_base->cb_pending.next,
1193                                    struct hrtimer, cb_entry);
1194
1195                 debug_hrtimer_deactivate(timer);
1196                 timer_stats_account_hrtimer(timer);
1197
1198                 fn = timer->function;
1199                 __remove_hrtimer(timer, timer->base, HRTIMER_STATE_CALLBACK, 0);
1200                 spin_unlock_irq(&cpu_base->lock);
1201
1202                 restart = fn(timer);
1203
1204                 spin_lock_irq(&cpu_base->lock);
1205
1206                 timer->state &= ~HRTIMER_STATE_CALLBACK;
1207                 if (restart == HRTIMER_RESTART) {
1208                         BUG_ON(hrtimer_active(timer));
1209                         /*
1210                          * Enqueue the timer, allow reprogramming of the event
1211                          * device
1212                          */
1213                         enqueue_hrtimer(timer, timer->base, 1);
1214                 } else if (hrtimer_active(timer)) {
1215                         /*
1216                          * If the timer was rearmed on another CPU, reprogram
1217                          * the event device.
1218                          */
1219                         struct hrtimer_clock_base *base = timer->base;
1220
1221                         if (base->first == &timer->node &&
1222                             hrtimer_reprogram(timer, base)) {
1223                                 /*
1224                                  * Timer is expired. Thus move it from tree to
1225                                  * pending list again.
1226                                  */
1227                                 __remove_hrtimer(timer, base,
1228                                                  HRTIMER_STATE_PENDING, 0);
1229                                 list_add_tail(&timer->cb_entry,
1230                                               &base->cpu_base->cb_pending);
1231                         }
1232                 }
1233         }
1234         spin_unlock_irq(&cpu_base->lock);
1235 }
1236
1237 static void __run_hrtimer(struct hrtimer *timer)
1238 {
1239         struct hrtimer_clock_base *base = timer->base;
1240         struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1241         enum hrtimer_restart (*fn)(struct hrtimer *);
1242         int restart;
1243
1244         debug_hrtimer_deactivate(timer);
1245         __remove_hrtimer(timer, base, HRTIMER_STATE_CALLBACK, 0);
1246         timer_stats_account_hrtimer(timer);
1247
1248         fn = timer->function;
1249         if (timer->cb_mode == HRTIMER_CB_IRQSAFE_PERCPU ||
1250             timer->cb_mode == HRTIMER_CB_IRQSAFE_UNLOCKED) {
1251                 /*
1252                  * Used for scheduler timers, avoid lock inversion with
1253                  * rq->lock and tasklist_lock.
1254                  *
1255                  * These timers are required to deal with enqueue expiry
1256                  * themselves and are not allowed to migrate.
1257                  */
1258                 spin_unlock(&cpu_base->lock);
1259                 restart = fn(timer);
1260                 spin_lock(&cpu_base->lock);
1261         } else
1262                 restart = fn(timer);
1263
1264         /*
1265          * Note: We clear the CALLBACK bit after enqueue_hrtimer to avoid
1266          * reprogramming of the event hardware. This happens at the end of this
1267          * function anyway.
1268          */
1269         if (restart != HRTIMER_NORESTART) {
1270                 BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
1271                 enqueue_hrtimer(timer, base, 0);
1272         }
1273         timer->state &= ~HRTIMER_STATE_CALLBACK;
1274 }
1275
1276 #ifdef CONFIG_HIGH_RES_TIMERS
1277
1278 /*
1279  * High resolution timer interrupt
1280  * Called with interrupts disabled
1281  */
1282 void hrtimer_interrupt(struct clock_event_device *dev)
1283 {
1284         struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1285         struct hrtimer_clock_base *base;
1286         ktime_t expires_next, now;
1287         int i, raise = 0;
1288
1289         BUG_ON(!cpu_base->hres_active);
1290         cpu_base->nr_events++;
1291         dev->next_event.tv64 = KTIME_MAX;
1292
1293  retry:
1294         now = ktime_get();
1295
1296         expires_next.tv64 = KTIME_MAX;
1297
1298         base = cpu_base->clock_base;
1299
1300         for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1301                 ktime_t basenow;
1302                 struct rb_node *node;
1303
1304                 spin_lock(&cpu_base->lock);
1305
1306                 basenow = ktime_add(now, base->offset);
1307
1308                 while ((node = base->first)) {
1309                         struct hrtimer *timer;
1310
1311                         timer = rb_entry(node, struct hrtimer, node);
1312
1313                         if (basenow.tv64 < timer->expires.tv64) {
1314                                 ktime_t expires;
1315
1316                                 expires = ktime_sub(timer->expires,
1317                                                     base->offset);
1318                                 if (expires.tv64 < expires_next.tv64)
1319                                         expires_next = expires;
1320                                 break;
1321                         }
1322
1323                         /* Move softirq callbacks to the pending list */
1324                         if (timer->cb_mode == HRTIMER_CB_SOFTIRQ) {
1325                                 __remove_hrtimer(timer, base,
1326                                                  HRTIMER_STATE_PENDING, 0);
1327                                 list_add_tail(&timer->cb_entry,
1328                                               &base->cpu_base->cb_pending);
1329                                 raise = 1;
1330                                 continue;
1331                         }
1332
1333                         __run_hrtimer(timer);
1334                 }
1335                 spin_unlock(&cpu_base->lock);
1336                 base++;
1337         }
1338
1339         cpu_base->expires_next = expires_next;
1340
1341         /* Reprogramming necessary ? */
1342         if (expires_next.tv64 != KTIME_MAX) {
1343                 if (tick_program_event(expires_next, 0))
1344                         goto retry;
1345         }
1346
1347         /* Raise softirq ? */
1348         if (raise)
1349                 raise_softirq(HRTIMER_SOFTIRQ);
1350 }
1351
1352 static void run_hrtimer_softirq(struct softirq_action *h)
1353 {
1354         run_hrtimer_pending(&__get_cpu_var(hrtimer_bases));
1355 }
1356
1357 #endif  /* CONFIG_HIGH_RES_TIMERS */
1358
1359 /*
1360  * Called from timer softirq every jiffy, expire hrtimers:
1361  *
1362  * For HRT its the fall back code to run the softirq in the timer
1363  * softirq context in case the hrtimer initialization failed or has
1364  * not been done yet.
1365  */
1366 void hrtimer_run_pending(void)
1367 {
1368         struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1369
1370         if (hrtimer_hres_active())
1371                 return;
1372
1373         /*
1374          * This _is_ ugly: We have to check in the softirq context,
1375          * whether we can switch to highres and / or nohz mode. The
1376          * clocksource switch happens in the timer interrupt with
1377          * xtime_lock held. Notification from there only sets the
1378          * check bit in the tick_oneshot code, otherwise we might
1379          * deadlock vs. xtime_lock.
1380          */
1381         if (tick_check_oneshot_change(!hrtimer_is_hres_enabled()))
1382                 hrtimer_switch_to_hres();
1383
1384         run_hrtimer_pending(cpu_base);
1385 }
1386
1387 /*
1388  * Called from hardirq context every jiffy
1389  */
1390 void hrtimer_run_queues(void)
1391 {
1392         struct rb_node *node;
1393         struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1394         struct hrtimer_clock_base *base;
1395         int index, gettime = 1;
1396
1397         if (hrtimer_hres_active())
1398                 return;
1399
1400         for (index = 0; index < HRTIMER_MAX_CLOCK_BASES; index++) {
1401                 base = &cpu_base->clock_base[index];
1402
1403                 if (!base->first)
1404                         continue;
1405
1406                 if (base->get_softirq_time)
1407                         base->softirq_time = base->get_softirq_time();
1408                 else if (gettime) {
1409                         hrtimer_get_softirq_time(cpu_base);
1410                         gettime = 0;
1411                 }
1412
1413                 spin_lock(&cpu_base->lock);
1414
1415                 while ((node = base->first)) {
1416                         struct hrtimer *timer;
1417
1418                         timer = rb_entry(node, struct hrtimer, node);
1419                         if (base->softirq_time.tv64 <= timer->expires.tv64)
1420                                 break;
1421
1422                         if (timer->cb_mode == HRTIMER_CB_SOFTIRQ) {
1423                                 __remove_hrtimer(timer, base,
1424                                         HRTIMER_STATE_PENDING, 0);
1425                                 list_add_tail(&timer->cb_entry,
1426                                         &base->cpu_base->cb_pending);
1427                                 continue;
1428                         }
1429
1430                         __run_hrtimer(timer);
1431                 }
1432                 spin_unlock(&cpu_base->lock);
1433         }
1434 }
1435
1436 /*
1437  * Sleep related functions:
1438  */
1439 static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1440 {
1441         struct hrtimer_sleeper *t =
1442                 container_of(timer, struct hrtimer_sleeper, timer);
1443         struct task_struct *task = t->task;
1444
1445         t->task = NULL;
1446         if (task)
1447                 wake_up_process(task);
1448
1449         return HRTIMER_NORESTART;
1450 }
1451
1452 void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1453 {
1454         sl->timer.function = hrtimer_wakeup;
1455         sl->task = task;
1456 #ifdef CONFIG_HIGH_RES_TIMERS
1457         sl->timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
1458 #endif
1459 }
1460
1461 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1462 {
1463         hrtimer_init_sleeper(t, current);
1464
1465         do {
1466                 set_current_state(TASK_INTERRUPTIBLE);
1467                 hrtimer_start(&t->timer, t->timer.expires, mode);
1468                 if (!hrtimer_active(&t->timer))
1469                         t->task = NULL;
1470
1471                 if (likely(t->task))
1472                         schedule();
1473
1474                 hrtimer_cancel(&t->timer);
1475                 mode = HRTIMER_MODE_ABS;
1476
1477         } while (t->task && !signal_pending(current));
1478
1479         __set_current_state(TASK_RUNNING);
1480
1481         return t->task == NULL;
1482 }
1483
1484 static int update_rmtp(struct hrtimer *timer, struct timespec __user *rmtp)
1485 {
1486         struct timespec rmt;
1487         ktime_t rem;
1488
1489         rem = ktime_sub(timer->expires, timer->base->get_time());
1490         if (rem.tv64 <= 0)
1491                 return 0;
1492         rmt = ktime_to_timespec(rem);
1493
1494         if (copy_to_user(rmtp, &rmt, sizeof(*rmtp)))
1495                 return -EFAULT;
1496
1497         return 1;
1498 }
1499
1500 long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1501 {
1502         struct hrtimer_sleeper t;
1503         struct timespec __user  *rmtp;
1504         int ret = 0;
1505
1506         hrtimer_init_on_stack(&t.timer, restart->nanosleep.index,
1507                                 HRTIMER_MODE_ABS);
1508         t.timer.expires.tv64 = restart->nanosleep.expires;
1509
1510         if (do_nanosleep(&t, HRTIMER_MODE_ABS))
1511                 goto out;
1512
1513         rmtp = restart->nanosleep.rmtp;
1514         if (rmtp) {
1515                 ret = update_rmtp(&t.timer, rmtp);
1516                 if (ret <= 0)
1517                         goto out;
1518         }
1519
1520         /* The other values in restart are already filled in */
1521         ret = -ERESTART_RESTARTBLOCK;
1522 out:
1523         destroy_hrtimer_on_stack(&t.timer);
1524         return ret;
1525 }
1526
1527 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
1528                        const enum hrtimer_mode mode, const clockid_t clockid)
1529 {
1530         struct restart_block *restart;
1531         struct hrtimer_sleeper t;
1532         int ret = 0;
1533
1534         hrtimer_init_on_stack(&t.timer, clockid, mode);
1535         t.timer.expires = timespec_to_ktime(*rqtp);
1536         if (do_nanosleep(&t, mode))
1537                 goto out;
1538
1539         /* Absolute timers do not update the rmtp value and restart: */
1540         if (mode == HRTIMER_MODE_ABS) {
1541                 ret = -ERESTARTNOHAND;
1542                 goto out;
1543         }
1544
1545         if (rmtp) {
1546                 ret = update_rmtp(&t.timer, rmtp);
1547                 if (ret <= 0)
1548                         goto out;
1549         }
1550
1551         restart = &current_thread_info()->restart_block;
1552         restart->fn = hrtimer_nanosleep_restart;
1553         restart->nanosleep.index = t.timer.base->index;
1554         restart->nanosleep.rmtp = rmtp;
1555         restart->nanosleep.expires = t.timer.expires.tv64;
1556
1557         ret = -ERESTART_RESTARTBLOCK;
1558 out:
1559         destroy_hrtimer_on_stack(&t.timer);
1560         return ret;
1561 }
1562
1563 asmlinkage long
1564 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
1565 {
1566         struct timespec tu;
1567
1568         if (copy_from_user(&tu, rqtp, sizeof(tu)))
1569                 return -EFAULT;
1570
1571         if (!timespec_valid(&tu))
1572                 return -EINVAL;
1573
1574         return hrtimer_nanosleep(&tu, rmtp, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1575 }
1576
1577 /*
1578  * Functions related to boot-time initialization:
1579  */
1580 static void __cpuinit init_hrtimers_cpu(int cpu)
1581 {
1582         struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1583         int i;
1584
1585         spin_lock_init(&cpu_base->lock);
1586
1587         for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
1588                 cpu_base->clock_base[i].cpu_base = cpu_base;
1589
1590         INIT_LIST_HEAD(&cpu_base->cb_pending);
1591         hrtimer_init_hres(cpu_base);
1592 }
1593
1594 #ifdef CONFIG_HOTPLUG_CPU
1595
1596 static int migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1597                                 struct hrtimer_clock_base *new_base, int dcpu)
1598 {
1599         struct hrtimer *timer;
1600         struct rb_node *node;
1601         int raise = 0;
1602
1603         while ((node = rb_first(&old_base->active))) {
1604                 timer = rb_entry(node, struct hrtimer, node);
1605                 BUG_ON(hrtimer_callback_running(timer));
1606                 debug_hrtimer_deactivate(timer);
1607
1608                 /*
1609                  * Should not happen. Per CPU timers should be
1610                  * canceled _before_ the migration code is called
1611                  */
1612                 if (timer->cb_mode == HRTIMER_CB_IRQSAFE_PERCPU) {
1613                         __remove_hrtimer(timer, old_base,
1614                                          HRTIMER_STATE_INACTIVE, 0);
1615                         WARN(1, "hrtimer (%p %p)active but cpu %d dead\n",
1616                              timer, timer->function, dcpu);
1617                         continue;
1618                 }
1619
1620                 /*
1621                  * Mark it as STATE_MIGRATE not INACTIVE otherwise the
1622                  * timer could be seen as !active and just vanish away
1623                  * under us on another CPU
1624                  */
1625                 __remove_hrtimer(timer, old_base, HRTIMER_STATE_MIGRATE, 0);
1626                 timer->base = new_base;
1627                 /*
1628                  * Enqueue the timer. Allow reprogramming of the event device
1629                  */
1630                 enqueue_hrtimer(timer, new_base, 1);
1631
1632 #ifdef CONFIG_HIGH_RES_TIMERS
1633                 /*
1634                  * Happens with high res enabled when the timer was
1635                  * already expired and the callback mode is
1636                  * HRTIMER_CB_IRQSAFE_UNLOCKED (hrtimer_sleeper). The
1637                  * enqueue code does not move them to the soft irq
1638                  * pending list for performance/latency reasons, but
1639                  * in the migration state, we need to do that
1640                  * otherwise we end up with a stale timer.
1641                  */
1642                 if (timer->state == HRTIMER_STATE_MIGRATE) {
1643                         timer->state = HRTIMER_STATE_PENDING;
1644                         list_add_tail(&timer->cb_entry,
1645                                       &new_base->cpu_base->cb_pending);
1646                         raise = 1;
1647                 }
1648 #endif
1649                 /* Clear the migration state bit */
1650                 timer->state &= ~HRTIMER_STATE_MIGRATE;
1651         }
1652         return raise;
1653 }
1654
1655 #ifdef CONFIG_HIGH_RES_TIMERS
1656 static int migrate_hrtimer_pending(struct hrtimer_cpu_base *old_base,
1657                                    struct hrtimer_cpu_base *new_base)
1658 {
1659         struct hrtimer *timer;
1660         int raise = 0;
1661
1662         while (!list_empty(&old_base->cb_pending)) {
1663                 timer = list_entry(old_base->cb_pending.next,
1664                                    struct hrtimer, cb_entry);
1665
1666                 __remove_hrtimer(timer, timer->base, HRTIMER_STATE_PENDING, 0);
1667                 timer->base = &new_base->clock_base[timer->base->index];
1668                 list_add_tail(&timer->cb_entry, &new_base->cb_pending);
1669                 raise = 1;
1670         }
1671         return raise;
1672 }
1673 #else
1674 static int migrate_hrtimer_pending(struct hrtimer_cpu_base *old_base,
1675                                    struct hrtimer_cpu_base *new_base)
1676 {
1677         return 0;
1678 }
1679 #endif
1680
1681 static void migrate_hrtimers(int cpu)
1682 {
1683         struct hrtimer_cpu_base *old_base, *new_base;
1684         int i, raise = 0;
1685
1686         BUG_ON(cpu_online(cpu));
1687         old_base = &per_cpu(hrtimer_bases, cpu);
1688         new_base = &get_cpu_var(hrtimer_bases);
1689
1690         tick_cancel_sched_timer(cpu);
1691
1692         local_irq_disable();
1693         spin_lock(&new_base->lock);
1694         spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1695
1696         for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1697                 if (migrate_hrtimer_list(&old_base->clock_base[i],
1698                                          &new_base->clock_base[i], cpu))
1699                         raise = 1;
1700         }
1701
1702         if (migrate_hrtimer_pending(old_base, new_base))
1703                 raise = 1;
1704
1705         spin_unlock(&old_base->lock);
1706         spin_unlock(&new_base->lock);
1707         local_irq_enable();
1708         put_cpu_var(hrtimer_bases);
1709
1710         if (raise)
1711                 hrtimer_raise_softirq();
1712 }
1713 #endif /* CONFIG_HOTPLUG_CPU */
1714
1715 static int __cpuinit hrtimer_cpu_notify(struct notifier_block *self,
1716                                         unsigned long action, void *hcpu)
1717 {
1718         unsigned int cpu = (long)hcpu;
1719
1720         switch (action) {
1721
1722         case CPU_UP_PREPARE:
1723         case CPU_UP_PREPARE_FROZEN:
1724                 init_hrtimers_cpu(cpu);
1725                 break;
1726
1727 #ifdef CONFIG_HOTPLUG_CPU
1728         case CPU_DEAD:
1729         case CPU_DEAD_FROZEN:
1730                 clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &cpu);
1731                 migrate_hrtimers(cpu);
1732                 break;
1733 #endif
1734
1735         default:
1736                 break;
1737         }
1738
1739         return NOTIFY_OK;
1740 }
1741
1742 static struct notifier_block __cpuinitdata hrtimers_nb = {
1743         .notifier_call = hrtimer_cpu_notify,
1744 };
1745
1746 void __init hrtimers_init(void)
1747 {
1748         hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
1749                           (void *)(long)smp_processor_id());
1750         register_cpu_notifier(&hrtimers_nb);
1751 #ifdef CONFIG_HIGH_RES_TIMERS
1752         open_softirq(HRTIMER_SOFTIRQ, run_hrtimer_softirq);
1753 #endif
1754 }
1755