kernel/hrtimer.c

   1 /*
   2  *  linux/kernel/hrtimer.c
   3  *
   4  *  Copyright(C) 2005, Thomas Gleixner <tglx@linutronix.de>
   5  *  Copyright(C) 2005, Red Hat, Inc., Ingo Molnar
   6  *
   7  *  High-resolution kernel timers
   8  *
   9  *  In contrast to the low-resolution timeout API implemented in
  10  *  kernel/timer.c, hrtimers provide finer resolution and accuracy
  11  *  depending on system configuration and capabilities.
  12  *
  13  *  These timers are currently used for:
  14  *   - itimers
  15  *   - POSIX timers
  16  *   - nanosleep
  17  *   - precise in-kernel timing
  18  *
  19  *  Started by: Thomas Gleixner and Ingo Molnar
  20  *
  21  *  Credits:
  22  *      based on kernel/timer.c
  23  *
  24  *      Help, testing, suggestions, bugfixes, improvements were
  25  *      provided by:
  26  *
  27  *      George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
  28  *      et. al.
  29  *
  30  *  For licencing details see kernel-base/COPYING
  31  */
  32
  33 #include <linux/cpu.h>
  34 #include <linux/module.h>
  35 #include <linux/percpu.h>
  36 #include <linux/hrtimer.h>
  37 #include <linux/notifier.h>
  38 #include <linux/syscalls.h>
  39 #include <linux/interrupt.h>
  40
  41 #include <asm/uaccess.h>
  42
  43 /**
  44  * ktime_get - get the monotonic time in ktime_t format
  45  *
  46  * returns the time in ktime_t format
  47  */
  48 static ktime_t ktime_get(void)
  49 {
  50         struct timespec now;
  51
  52         ktime_get_ts(&now);
  53
  54         return timespec_to_ktime(now);
  55 }
  56
  57 /**
  58  * ktime_get_real - get the real (wall-) time in ktime_t format
  59  *
  60  * returns the time in ktime_t format
  61  */
  62 static ktime_t ktime_get_real(void)
  63 {
  64         struct timespec now;
  65
  66         getnstimeofday(&now);
  67
  68         return timespec_to_ktime(now);
  69 }
  70
  71 EXPORT_SYMBOL_GPL(ktime_get_real);
  72
  73 /*
  74  * The timer bases:
  75  *
  76  * Note: If we want to add new timer bases, we have to skip the two
  77  * clock ids captured by the cpu-timers. We do this by holding empty
  78  * entries rather than doing math adjustment of the clock ids.
  79  * This ensures that we capture erroneous accesses to these clock ids
  80  * rather than moving them into the range of valid clock id's.
  81  */
  82
  83 #define MAX_HRTIMER_BASES 2
  84
  85 static DEFINE_PER_CPU(struct hrtimer_base, hrtimer_bases[MAX_HRTIMER_BASES]) =
  86 {
  87         {
  88                 .index = CLOCK_REALTIME,
  89                 .get_time = &ktime_get_real,
  90                 .resolution = KTIME_REALTIME_RES,
  91         },
  92         {
  93                 .index = CLOCK_MONOTONIC,
  94                 .get_time = &ktime_get,
  95                 .resolution = KTIME_MONOTONIC_RES,
  96         },
  97 };
  98
  99 /**
 100  * ktime_get_ts - get the monotonic clock in timespec format
 101  *
 102  * @ts:         pointer to timespec variable
 103  *
 104  * The function calculates the monotonic clock from the realtime
 105  * clock and the wall_to_monotonic offset and stores the result
 106  * in normalized timespec format in the variable pointed to by ts.
 107  */
 108 void ktime_get_ts(struct timespec *ts)
 109 {
 110         struct timespec tomono;
 111         unsigned long seq;
 112
 113         do {
 114                 seq = read_seqbegin(&xtime_lock);
 115                 getnstimeofday(ts);
 116                 tomono = wall_to_monotonic;
 117
 118         } while (read_seqretry(&xtime_lock, seq));
 119
 120         set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
 121                                 ts->tv_nsec + tomono.tv_nsec);
 122 }
 123 EXPORT_SYMBOL_GPL(ktime_get_ts);
 124
 125 /*
 126  * Functions and macros which are different for UP/SMP systems are kept in a
 127  * single place
 128  */
 129 #ifdef CONFIG_SMP
 130
 131 #define set_curr_timer(b, t)            do { (b)->curr_timer = (t); } while (0)
 132
 133 /*
 134  * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
 135  * means that all timers which are tied to this base via timer->base are
 136  * locked, and the base itself is locked too.
 137  *
 138  * So __run_timers/migrate_timers can safely modify all timers which could
 139  * be found on the lists/queues.
 140  *
 141  * When the timer's base is locked, and the timer removed from list, it is
 142  * possible to set timer->base = NULL and drop the lock: the timer remains
 143  * locked.
 144  */
 145 static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
 146                                               unsigned long *flags)
 147 {
 148         struct hrtimer_base *base;
 149
 150         for (;;) {
 151                 base = timer->base;
 152                 if (likely(base != NULL)) {
 153                         spin_lock_irqsave(&base->lock, *flags);
 154                         if (likely(base == timer->base))
 155                                 return base;
 156                         /* The timer has migrated to another CPU: */
 157                         spin_unlock_irqrestore(&base->lock, *flags);
 158                 }
 159                 cpu_relax();
 160         }
 161 }
 162
 163 /*
 164  * Switch the timer base to the current CPU when possible.
 165  */
 166 static inline struct hrtimer_base *
 167 switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
 168 {
 169         struct hrtimer_base *new_base;
 170
 171         new_base = &__get_cpu_var(hrtimer_bases[base->index]);
 172
 173         if (base != new_base) {
 174                 /*
 175                  * We are trying to schedule the timer on the local CPU.
 176                  * However we can't change timer's base while it is running,
 177                  * so we keep it on the same CPU. No hassle vs. reprogramming
 178                  * the event source in the high resolution case. The softirq
 179                  * code will take care of this when the timer function has
 180                  * completed. There is no conflict as we hold the lock until
 181                  * the timer is enqueued.
 182                  */
 183                 if (unlikely(base->curr_timer == timer))
 184                         return base;
 185
 186                 /* See the comment in lock_timer_base() */
 187                 timer->base = NULL;
 188                 spin_unlock(&base->lock);
 189                 spin_lock(&new_base->lock);
 190                 timer->base = new_base;
 191         }
 192         return new_base;
 193 }
 194
 195 #else /* CONFIG_SMP */
 196
 197 #define set_curr_timer(b, t)            do { } while (0)
 198
 199 static inline struct hrtimer_base *
 200 lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 201 {
 202         struct hrtimer_base *base = timer->base;
 203
 204         spin_lock_irqsave(&base->lock, *flags);
 205
 206         return base;
 207 }
 208
 209 #define switch_hrtimer_base(t, b)       (b)
 210
 211 #endif  /* !CONFIG_SMP */
 212
 213 /*
 214  * Functions for the union type storage format of ktime_t which are
 215  * too large for inlining:
 216  */
 217 #if BITS_PER_LONG < 64
 218 # ifndef CONFIG_KTIME_SCALAR
 219 /**
 220  * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
 221  *
 222  * @kt:         addend
 223  * @nsec:       the scalar nsec value to add
 224  *
 225  * Returns the sum of kt and nsec in ktime_t format
 226  */
 227 ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
 228 {
 229         ktime_t tmp;
 230
 231         if (likely(nsec < NSEC_PER_SEC)) {
 232                 tmp.tv64 = nsec;
 233         } else {
 234                 unsigned long rem = do_div(nsec, NSEC_PER_SEC);
 235
 236                 tmp = ktime_set((long)nsec, rem);
 237         }
 238
 239         return ktime_add(kt, tmp);
 240 }
 241
 242 #else /* CONFIG_KTIME_SCALAR */
 243
 244 # endif /* !CONFIG_KTIME_SCALAR */
 245
 246 /*
 247  * Divide a ktime value by a nanosecond value
 248  */
 249 static unsigned long ktime_divns(const ktime_t kt, nsec_t div)
 250 {
 251         u64 dclc, inc, dns;
 252         int sft = 0;
 253
 254         dclc = dns = ktime_to_ns(kt);
 255         inc = div;
 256         /* Make sure the divisor is less than 2^32: */
 257         while (div >> 32) {
 258                 sft++;
 259                 div >>= 1;
 260         }
 261         dclc >>= sft;
 262         do_div(dclc, (unsigned long) div);
 263
 264         return (unsigned long) dclc;
 265 }
 266
 267 #else /* BITS_PER_LONG < 64 */
 268 # define ktime_divns(kt, div)           (unsigned long)((kt).tv64 / (div))
 269 #endif /* BITS_PER_LONG >= 64 */
 270
 271 /*
 272  * Counterpart to lock_timer_base above:
 273  */
 274 static inline
 275 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 276 {
 277         spin_unlock_irqrestore(&timer->base->lock, *flags);
 278 }
 279
 280 /**
 281  * hrtimer_forward - forward the timer expiry
 282  *
 283  * @timer:      hrtimer to forward
 284  * @interval:   the interval to forward
 285  *
 286  * Forward the timer expiry so it will expire in the future.
 287  * Returns the number of overruns.
 288  */
 289 unsigned long
 290 hrtimer_forward(struct hrtimer *timer, ktime_t interval)
 291 {
 292         unsigned long orun = 1;
 293         ktime_t delta, now;
 294
 295         now = timer->base->get_time();
 296
 297         delta = ktime_sub(now, timer->expires);
 298
 299         if (delta.tv64 < 0)
 300                 return 0;
 301
 302         if (interval.tv64 < timer->base->resolution.tv64)
 303                 interval.tv64 = timer->base->resolution.tv64;
 304
 305         if (unlikely(delta.tv64 >= interval.tv64)) {
 306                 nsec_t incr = ktime_to_ns(interval);
 307
 308                 orun = ktime_divns(delta, incr);
 309                 timer->expires = ktime_add_ns(timer->expires, incr * orun);
 310                 if (timer->expires.tv64 > now.tv64)
 311                         return orun;
 312                 /*
 313                  * This (and the ktime_add() below) is the
 314                  * correction for exact:
 315                  */
 316                 orun++;
 317         }
 318         timer->expires = ktime_add(timer->expires, interval);
 319
 320         return orun;
 321 }
 322
 323 /*
 324  * enqueue_hrtimer - internal function to (re)start a timer
 325  *
 326  * The timer is inserted in expiry order. Insertion into the
 327  * red black tree is O(log(n)). Must hold the base lock.
 328  */
 329 static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 330 {
 331         struct rb_node **link = &base->active.rb_node;
 332         struct rb_node *parent = NULL;
 333         struct hrtimer *entry;
 334
 335         /*
 336          * Find the right place in the rbtree:
 337          */
 338         while (*link) {
 339                 parent = *link;
 340                 entry = rb_entry(parent, struct hrtimer, node);
 341                 /*
 342                  * We dont care about collisions. Nodes with
 343                  * the same expiry time stay together.
 344                  */
 345                 if (timer->expires.tv64 < entry->expires.tv64)
 346                         link = &(*link)->rb_left;
 347                 else
 348                         link = &(*link)->rb_right;
 349         }
 350
 351         /*
 352          * Insert the timer to the rbtree and check whether it
 353          * replaces the first pending timer
 354          */
 355         rb_link_node(&timer->node, parent, link);
 356         rb_insert_color(&timer->node, &base->active);
 357
 358         timer->state = HRTIMER_PENDING;
 359
 360         if (!base->first || timer->expires.tv64 <
 361             rb_entry(base->first, struct hrtimer, node)->expires.tv64)
 362                 base->first = &timer->node;
 363 }
 364
 365 /*
 366  * __remove_hrtimer - internal function to remove a timer
 367  *
 368  * Caller must hold the base lock.
 369  */
 370 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 371 {
 372         /*
 373          * Remove the timer from the rbtree and replace the
 374          * first entry pointer if necessary.
 375          */
 376         if (base->first == &timer->node)
 377                 base->first = rb_next(&timer->node);
 378         rb_erase(&timer->node, &base->active);
 379 }
 380
 381 /*
 382  * remove hrtimer, called with base lock held
 383  */
 384 static inline int
 385 remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 386 {
 387         if (hrtimer_active(timer)) {
 388                 __remove_hrtimer(timer, base);
 389                 timer->state = HRTIMER_INACTIVE;
 390                 return 1;
 391         }
 392         return 0;
 393 }
 394
 395 /**
 396  * hrtimer_start - (re)start an relative timer on the current CPU
 397  *
 398  * @timer:      the timer to be added
 399  * @tim:        expiry time
 400  * @mode:       expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
 401  *
 402  * Returns:
 403  *  0 on success
 404  *  1 when the timer was active
 405  */
 406 int
 407 hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
 408 {
 409         struct hrtimer_base *base, *new_base;
 410         unsigned long flags;
 411         int ret;
 412
 413         base = lock_hrtimer_base(timer, &flags);
 414
 415         /* Remove an active timer from the queue: */
 416         ret = remove_hrtimer(timer, base);
 417
 418         /* Switch the timer base, if necessary: */
 419         new_base = switch_hrtimer_base(timer, base);
 420
 421         if (mode == HRTIMER_REL) {
 422                 tim = ktime_add(tim, new_base->get_time());
 423                 /*
 424                  * CONFIG_TIME_LOW_RES is a temporary way for architectures
 425                  * to signal that they simply return xtime in
 426                  * do_gettimeoffset(). In this case we want to round up by
 427                  * resolution when starting a relative timer, to avoid short
 428                  * timeouts. This will go away with the GTOD framework.
 429                  */
 430 #ifdef CONFIG_TIME_LOW_RES
 431                 tim = ktime_add(tim, base->resolution);
 432 #endif
 433         }
 434         timer->expires = tim;
 435
 436         enqueue_hrtimer(timer, new_base);
 437
 438         unlock_hrtimer_base(timer, &flags);
 439
 440         return ret;
 441 }
 442
 443 /**
 444  * hrtimer_try_to_cancel - try to deactivate a timer
 445  *
 446  * @timer:      hrtimer to stop
 447  *
 448  * Returns:
 449  *  0 when the timer was not active
 450  *  1 when the timer was active
 451  * -1 when the timer is currently excuting the callback function and
 452  *    can not be stopped
 453  */
 454 int hrtimer_try_to_cancel(struct hrtimer *timer)
 455 {
 456         struct hrtimer_base *base;
 457         unsigned long flags;
 458         int ret = -1;
 459
 460         base = lock_hrtimer_base(timer, &flags);
 461
 462         if (base->curr_timer != timer)
 463                 ret = remove_hrtimer(timer, base);
 464
 465         unlock_hrtimer_base(timer, &flags);
 466
 467         return ret;
 468
 469 }
 470
 471 /**
 472  * hrtimer_cancel - cancel a timer and wait for the handler to finish.
 473  *
 474  * @timer:      the timer to be cancelled
 475  *
 476  * Returns:
 477  *  0 when the timer was not active
 478  *  1 when the timer was active
 479  */
 480 int hrtimer_cancel(struct hrtimer *timer)
 481 {
 482         for (;;) {
 483                 int ret = hrtimer_try_to_cancel(timer);
 484
 485                 if (ret >= 0)
 486                         return ret;
 487         }
 488 }
 489
 490 /**
 491  * hrtimer_get_remaining - get remaining time for the timer
 492  *
 493  * @timer:      the timer to read
 494  */
 495 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
 496 {
 497         struct hrtimer_base *base;
 498         unsigned long flags;
 499         ktime_t rem;
 500
 501         base = lock_hrtimer_base(timer, &flags);
 502         rem = ktime_sub(timer->expires, timer->base->get_time());
 503         unlock_hrtimer_base(timer, &flags);
 504
 505         return rem;
 506 }
 507
 508 /**
 509  * hrtimer_init - initialize a timer to the given clock
 510  *
 511  * @timer:      the timer to be initialized
 512  * @clock_id:   the clock to be used
 513  * @mode:       timer mode abs/rel
 514  */
 515 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
 516                   enum hrtimer_mode mode)
 517 {
 518         struct hrtimer_base *bases;
 519
 520         memset(timer, 0, sizeof(struct hrtimer));
 521
 522         bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
 523
 524         if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
 525                 clock_id = CLOCK_MONOTONIC;
 526
 527         timer->base = &bases[clock_id];
 528 }
 529
 530 /**
 531  * hrtimer_get_res - get the timer resolution for a clock
 532  *
 533  * @which_clock: which clock to query
 534  * @tp:          pointer to timespec variable to store the resolution
 535  *
 536  * Store the resolution of the clock selected by which_clock in the
 537  * variable pointed to by tp.
 538  */
 539 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
 540 {
 541         struct hrtimer_base *bases;
 542
 543         bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
 544         *tp = ktime_to_timespec(bases[which_clock].resolution);
 545
 546         return 0;
 547 }
 548
 549 /*
 550  * Expire the per base hrtimer-queue:
 551  */
 552 static inline void run_hrtimer_queue(struct hrtimer_base *base)
 553 {
 554         ktime_t now = base->get_time();
 555         struct rb_node *node;
 556
 557         spin_lock_irq(&base->lock);
 558
 559         while ((node = base->first)) {
 560                 struct hrtimer *timer;
 561                 int (*fn)(void *);
 562                 int restart;
 563                 void *data;
 564
 565                 timer = rb_entry(node, struct hrtimer, node);
 566                 if (now.tv64 <= timer->expires.tv64)
 567                         break;
 568
 569                 fn = timer->function;
 570                 data = timer->data;
 571                 set_curr_timer(base, timer);
 572                 timer->state = HRTIMER_RUNNING;
 573                 __remove_hrtimer(timer, base);
 574                 spin_unlock_irq(&base->lock);
 575
 576                 /*
 577                  * fn == NULL is special case for the simplest timer
 578                  * variant - wake up process and do not restart:
 579                  */
 580                 if (!fn) {
 581                         wake_up_process(data);
 582                         restart = HRTIMER_NORESTART;
 583                 } else
 584                         restart = fn(data);
 585
 586                 spin_lock_irq(&base->lock);
 587
 588                 /* Another CPU has added back the timer */
 589                 if (timer->state != HRTIMER_RUNNING)
 590                         continue;
 591
 592                 if (restart == HRTIMER_RESTART)
 593                         enqueue_hrtimer(timer, base);
 594                 else
 595                         timer->state = HRTIMER_EXPIRED;
 596         }
 597         set_curr_timer(base, NULL);
 598         spin_unlock_irq(&base->lock);
 599 }
 600
 601 /*
 602  * Called from timer softirq every jiffy, expire hrtimers:
 603  */
 604 void hrtimer_run_queues(void)
 605 {
 606         struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
 607         int i;
 608
 609         for (i = 0; i < MAX_HRTIMER_BASES; i++)
 610                 run_hrtimer_queue(&base[i]);
 611 }
 612
 613 /*
 614  * Sleep related functions:
 615  */
 616
 617 /**
 618  * schedule_hrtimer - sleep until timeout
 619  *
 620  * @timer:      hrtimer variable initialized with the correct clock base
 621  * @mode:       timeout value is abs/rel
 622  *
 623  * Make the current task sleep until @timeout is
 624  * elapsed.
 625  *
 626  * You can set the task state as follows -
 627  *
 628  * %TASK_UNINTERRUPTIBLE - at least @timeout is guaranteed to
 629  * pass before the routine returns. The routine will return 0
 630  *
 631  * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
 632  * delivered to the current task. In this case the remaining time
 633  * will be returned
 634  *
 635  * The current task state is guaranteed to be TASK_RUNNING when this
 636  * routine returns.
 637  */
 638 static ktime_t __sched
 639 schedule_hrtimer(struct hrtimer *timer, const enum hrtimer_mode mode)
 640 {
 641         /* fn stays NULL, meaning single-shot wakeup: */
 642         timer->data = current;
 643
 644         hrtimer_start(timer, timer->expires, mode);
 645
 646         schedule();
 647         hrtimer_cancel(timer);
 648
 649         /* Return the remaining time: */
 650         if (timer->state != HRTIMER_EXPIRED)
 651                 return ktime_sub(timer->expires, timer->base->get_time());
 652         else
 653                 return (ktime_t) {.tv64 = 0 };
 654 }
 655
 656 static inline ktime_t __sched
 657 schedule_hrtimer_interruptible(struct hrtimer *timer,
 658                                const enum hrtimer_mode mode)
 659 {
 660         set_current_state(TASK_INTERRUPTIBLE);
 661
 662         return schedule_hrtimer(timer, mode);
 663 }
 664
 665 static long __sched nanosleep_restart(struct restart_block *restart)
 666 {
 667         struct timespec __user *rmtp;
 668         struct timespec tu;
 669         void *rfn_save = restart->fn;
 670         struct hrtimer timer;
 671         ktime_t rem;
 672
 673         restart->fn = do_no_restart_syscall;
 674
 675         hrtimer_init(&timer, (clockid_t) restart->arg3, HRTIMER_ABS);
 676
 677         timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
 678
 679         rem = schedule_hrtimer_interruptible(&timer, HRTIMER_ABS);
 680
 681         if (rem.tv64 <= 0)
 682                 return 0;
 683
 684         rmtp = (struct timespec __user *) restart->arg2;
 685         tu = ktime_to_timespec(rem);
 686         if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
 687                 return -EFAULT;
 688
 689         restart->fn = rfn_save;
 690
 691         /* The other values in restart are already filled in */
 692         return -ERESTART_RESTARTBLOCK;
 693 }
 694
 695 long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
 696                        const enum hrtimer_mode mode, const clockid_t clockid)
 697 {
 698         struct restart_block *restart;
 699         struct hrtimer timer;
 700         struct timespec tu;
 701         ktime_t rem;
 702
 703         hrtimer_init(&timer, clockid, mode);
 704
 705         timer.expires = timespec_to_ktime(*rqtp);
 706
 707         rem = schedule_hrtimer_interruptible(&timer, mode);
 708         if (rem.tv64 <= 0)
 709                 return 0;
 710
 711         /* Absolute timers do not update the rmtp value and restart: */
 712         if (mode == HRTIMER_ABS)
 713                 return -ERESTARTNOHAND;
 714
 715         tu = ktime_to_timespec(rem);
 716
 717         if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
 718                 return -EFAULT;
 719
 720         restart = &current_thread_info()->restart_block;
 721         restart->fn = nanosleep_restart;
 722         restart->arg0 = timer.expires.tv64 & 0xFFFFFFFF;
 723         restart->arg1 = timer.expires.tv64 >> 32;
 724         restart->arg2 = (unsigned long) rmtp;
 725         restart->arg3 = (unsigned long) timer.base->index;
 726
 727         return -ERESTART_RESTARTBLOCK;
 728 }
 729
 730 asmlinkage long
 731 sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
 732 {
 733         struct timespec tu;
 734
 735         if (copy_from_user(&tu, rqtp, sizeof(tu)))
 736                 return -EFAULT;
 737
 738         if (!timespec_valid(&tu))
 739                 return -EINVAL;
 740
 741         return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
 742 }
 743
 744 /*
 745  * Functions related to boot-time initialization:
 746  */
 747 static void __devinit init_hrtimers_cpu(int cpu)
 748 {
 749         struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
 750         int i;
 751
 752         for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
 753                 spin_lock_init(&base->lock);
 754 }
 755
 756 #ifdef CONFIG_HOTPLUG_CPU
 757
 758 static void migrate_hrtimer_list(struct hrtimer_base *old_base,
 759                                 struct hrtimer_base *new_base)
 760 {
 761         struct hrtimer *timer;
 762         struct rb_node *node;
 763
 764         while ((node = rb_first(&old_base->active))) {
 765                 timer = rb_entry(node, struct hrtimer, node);
 766                 __remove_hrtimer(timer, old_base);
 767                 timer->base = new_base;
 768                 enqueue_hrtimer(timer, new_base);
 769         }
 770 }
 771
 772 static void migrate_hrtimers(int cpu)
 773 {
 774         struct hrtimer_base *old_base, *new_base;
 775         int i;
 776
 777         BUG_ON(cpu_online(cpu));
 778         old_base = per_cpu(hrtimer_bases, cpu);
 779         new_base = get_cpu_var(hrtimer_bases);
 780
 781         local_irq_disable();
 782
 783         for (i = 0; i < MAX_HRTIMER_BASES; i++) {
 784
 785                 spin_lock(&new_base->lock);
 786                 spin_lock(&old_base->lock);
 787
 788                 BUG_ON(old_base->curr_timer);
 789
 790                 migrate_hrtimer_list(old_base, new_base);
 791
 792                 spin_unlock(&old_base->lock);
 793                 spin_unlock(&new_base->lock);
 794                 old_base++;
 795                 new_base++;
 796         }
 797
 798         local_irq_enable();
 799         put_cpu_var(hrtimer_bases);
 800 }
 801 #endif /* CONFIG_HOTPLUG_CPU */
 802
 803 static int __devinit hrtimer_cpu_notify(struct notifier_block *self,
 804                                         unsigned long action, void *hcpu)
 805 {
 806         long cpu = (long)hcpu;
 807
 808         switch (action) {
 809
 810         case CPU_UP_PREPARE:
 811                 init_hrtimers_cpu(cpu);
 812                 break;
 813
 814 #ifdef CONFIG_HOTPLUG_CPU
 815         case CPU_DEAD:
 816                 migrate_hrtimers(cpu);
 817                 break;
 818 #endif
 819
 820         default:
 821                 break;
 822         }
 823
 824         return NOTIFY_OK;
 825 }
 826
 827 static struct notifier_block __devinitdata hrtimers_nb = {
 828         .notifier_call = hrtimer_cpu_notify,
 829 };
 830
 831 void __init hrtimers_init(void)
 832 {
 833         hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
 834                           (void *)(long)smp_processor_id());
 835         register_cpu_notifier(&hrtimers_nb);
 836 }
 837