kernel/time/timekeeping.c

   1 /*
   2  *  linux/kernel/time/timekeeping.c
   3  *
   4  *  Kernel timekeeping code and accessor functions
   5  *
   6  *  This code was moved from linux/kernel/timer.c.
   7  *  Please see that file for copyright and history logs.
   8  *
   9  */
  10
  11 #include <linux/module.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/percpu.h>
  14 #include <linux/init.h>
  15 #include <linux/mm.h>
  16 #include <linux/sysdev.h>
  17 #include <linux/clocksource.h>
  18 #include <linux/jiffies.h>
  19 #include <linux/time.h>
  20 #include <linux/tick.h>
  21
  22
  23 /*
  24  * This read-write spinlock protects us from races in SMP while
  25  * playing with xtime and avenrun.
  26  */
  27 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
  28
  29
  30 /*
  31  * The current time
  32  * wall_to_monotonic is what we need to add to xtime (or xtime corrected
  33  * for sub jiffie times) to get to monotonic time.  Monotonic is pegged
  34  * at zero at system boot time, so wall_to_monotonic will be negative,
  35  * however, we will ALWAYS keep the tv_nsec part positive so we can use
  36  * the usual normalization.
  37  *
  38  * wall_to_monotonic is moved after resume from suspend for the monotonic
  39  * time not to jump. We need to add total_sleep_time to wall_to_monotonic
  40  * to get the real boot based time offset.
  41  *
  42  * - wall_to_monotonic is no longer the boot time, getboottime must be
  43  * used instead.
  44  */
  45 struct timespec xtime __attribute__ ((aligned (16)));
  46 struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
  47 static unsigned long total_sleep_time;          /* seconds */
  48
  49 static struct timespec xtime_cache __attribute__ ((aligned (16)));
  50 static inline void update_xtime_cache(u64 nsec)
  51 {
  52         xtime_cache = xtime;
  53         timespec_add_ns(&xtime_cache, nsec);
  54 }
  55
  56 static struct clocksource *clock; /* pointer to current clocksource */
  57
  58
  59 #ifdef CONFIG_GENERIC_TIME
  60 /**
  61  * __get_nsec_offset - Returns nanoseconds since last call to periodic_hook
  62  *
  63  * private function, must hold xtime_lock lock when being
  64  * called. Returns the number of nanoseconds since the
  65  * last call to update_wall_time() (adjusted by NTP scaling)
  66  */
  67 static inline s64 __get_nsec_offset(void)
  68 {
  69         cycle_t cycle_now, cycle_delta;
  70         s64 ns_offset;
  71
  72         /* read clocksource: */
  73         cycle_now = clocksource_read(clock);
  74
  75         /* calculate the delta since the last update_wall_time: */
  76         cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
  77
  78         /* convert to nanoseconds: */
  79         ns_offset = cyc2ns(clock, cycle_delta);
  80
  81         return ns_offset;
  82 }
  83
  84 /**
  85  * __get_realtime_clock_ts - Returns the time of day in a timespec
  86  * @ts:         pointer to the timespec to be set
  87  *
  88  * Returns the time of day in a timespec. Used by
  89  * do_gettimeofday() and get_realtime_clock_ts().
  90  */
  91 static inline void __get_realtime_clock_ts(struct timespec *ts)
  92 {
  93         unsigned long seq;
  94         s64 nsecs;
  95
  96         do {
  97                 seq = read_seqbegin(&xtime_lock);
  98
  99                 *ts = xtime;
 100                 nsecs = __get_nsec_offset();
 101
 102         } while (read_seqretry(&xtime_lock, seq));
 103
 104         timespec_add_ns(ts, nsecs);
 105 }
 106
 107 /**
 108  * getnstimeofday - Returns the time of day in a timespec
 109  * @ts:         pointer to the timespec to be set
 110  *
 111  * Returns the time of day in a timespec.
 112  */
 113 void getnstimeofday(struct timespec *ts)
 114 {
 115         __get_realtime_clock_ts(ts);
 116 }
 117
 118 EXPORT_SYMBOL(getnstimeofday);
 119
 120 /**
 121  * do_gettimeofday - Returns the time of day in a timeval
 122  * @tv:         pointer to the timeval to be set
 123  *
 124  * NOTE: Users should be converted to using get_realtime_clock_ts()
 125  */
 126 void do_gettimeofday(struct timeval *tv)
 127 {
 128         struct timespec now;
 129
 130         __get_realtime_clock_ts(&now);
 131         tv->tv_sec = now.tv_sec;
 132         tv->tv_usec = now.tv_nsec/1000;
 133 }
 134
 135 EXPORT_SYMBOL(do_gettimeofday);
 136 /**
 137  * do_settimeofday - Sets the time of day
 138  * @tv:         pointer to the timespec variable containing the new time
 139  *
 140  * Sets the time of day to the new time and update NTP and notify hrtimers
 141  */
 142 int do_settimeofday(struct timespec *tv)
 143 {
 144         unsigned long flags;
 145         time_t wtm_sec, sec = tv->tv_sec;
 146         long wtm_nsec, nsec = tv->tv_nsec;
 147
 148         if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 149                 return -EINVAL;
 150
 151         write_seqlock_irqsave(&xtime_lock, flags);
 152
 153         nsec -= __get_nsec_offset();
 154
 155         wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
 156         wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
 157
 158         set_normalized_timespec(&xtime, sec, nsec);
 159         set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
 160
 161         clock->error = 0;
 162         ntp_clear();
 163
 164         update_vsyscall(&xtime, clock);
 165
 166         write_sequnlock_irqrestore(&xtime_lock, flags);
 167
 168         /* signal hrtimers about time change */
 169         clock_was_set();
 170
 171         return 0;
 172 }
 173
 174 EXPORT_SYMBOL(do_settimeofday);
 175
 176 /**
 177  * change_clocksource - Swaps clocksources if a new one is available
 178  *
 179  * Accumulates current time interval and initializes new clocksource
 180  */
 181 static void change_clocksource(void)
 182 {
 183         struct clocksource *new;
 184         cycle_t now;
 185         u64 nsec;
 186
 187         new = clocksource_get_next();
 188
 189         if (clock == new)
 190                 return;
 191
 192         now = clocksource_read(new);
 193         nsec =  __get_nsec_offset();
 194         timespec_add_ns(&xtime, nsec);
 195
 196         clock = new;
 197         clock->cycle_last = now;
 198
 199         clock->error = 0;
 200         clock->xtime_nsec = 0;
 201         clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
 202
 203         tick_clock_notify();
 204
 205         printk(KERN_INFO "Time: %s clocksource has been installed.\n",
 206                clock->name);
 207 }
 208 #else
 209 static inline void change_clocksource(void) { }
 210 static inline s64 __get_nsec_offset(void) { return 0; }
 211 #endif
 212
 213 /**
 214  * timekeeping_is_continuous - check to see if timekeeping is free running
 215  */
 216 int timekeeping_is_continuous(void)
 217 {
 218         unsigned long seq;
 219         int ret;
 220
 221         do {
 222                 seq = read_seqbegin(&xtime_lock);
 223
 224                 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
 225
 226         } while (read_seqretry(&xtime_lock, seq));
 227
 228         return ret;
 229 }
 230
 231 /**
 232  * read_persistent_clock -  Return time in seconds from the persistent clock.
 233  *
 234  * Weak dummy function for arches that do not yet support it.
 235  * Returns seconds from epoch using the battery backed persistent clock.
 236  * Returns zero if unsupported.
 237  *
 238  *  XXX - Do be sure to remove it once all arches implement it.
 239  */
 240 unsigned long __attribute__((weak)) read_persistent_clock(void)
 241 {
 242         return 0;
 243 }
 244
 245 /*
 246  * timekeeping_init - Initializes the clocksource and common timekeeping values
 247  */
 248 void __init timekeeping_init(void)
 249 {
 250         unsigned long flags;
 251         unsigned long sec = read_persistent_clock();
 252
 253         write_seqlock_irqsave(&xtime_lock, flags);
 254
 255         ntp_clear();
 256
 257         clock = clocksource_get_next();
 258         clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
 259         clock->cycle_last = clocksource_read(clock);
 260
 261         xtime.tv_sec = sec;
 262         xtime.tv_nsec = 0;
 263         set_normalized_timespec(&wall_to_monotonic,
 264                 -xtime.tv_sec, -xtime.tv_nsec);
 265         total_sleep_time = 0;
 266
 267         write_sequnlock_irqrestore(&xtime_lock, flags);
 268 }
 269
 270 /* flag for if timekeeping is suspended */
 271 static int timekeeping_suspended;
 272 /* time in seconds when suspend began */
 273 static unsigned long timekeeping_suspend_time;
 274 /* xtime offset when we went into suspend */
 275 static s64 timekeeping_suspend_nsecs;
 276
 277 /**
 278  * timekeeping_resume - Resumes the generic timekeeping subsystem.
 279  * @dev:        unused
 280  *
 281  * This is for the generic clocksource timekeeping.
 282  * xtime/wall_to_monotonic/jiffies/etc are
 283  * still managed by arch specific suspend/resume code.
 284  */
 285 static int timekeeping_resume(struct sys_device *dev)
 286 {
 287         unsigned long flags;
 288         unsigned long now = read_persistent_clock();
 289
 290         clocksource_resume();
 291
 292         write_seqlock_irqsave(&xtime_lock, flags);
 293
 294         if (now && (now > timekeeping_suspend_time)) {
 295                 unsigned long sleep_length = now - timekeeping_suspend_time;
 296
 297                 xtime.tv_sec += sleep_length;
 298                 wall_to_monotonic.tv_sec -= sleep_length;
 299                 total_sleep_time += sleep_length;
 300         }
 301         /* Make sure that we have the correct xtime reference */
 302         timespec_add_ns(&xtime, timekeeping_suspend_nsecs);
 303         /* re-base the last cycle value */
 304         clock->cycle_last = clocksource_read(clock);
 305         clock->error = 0;
 306         timekeeping_suspended = 0;
 307         write_sequnlock_irqrestore(&xtime_lock, flags);
 308
 309         touch_softlockup_watchdog();
 310
 311         clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
 312
 313         /* Resume hrtimers */
 314         hres_timers_resume();
 315
 316         return 0;
 317 }
 318
 319 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
 320 {
 321         unsigned long flags;
 322
 323         timekeeping_suspend_time = read_persistent_clock();
 324
 325         write_seqlock_irqsave(&xtime_lock, flags);
 326         /* Get the current xtime offset */
 327         timekeeping_suspend_nsecs = __get_nsec_offset();
 328         timekeeping_suspended = 1;
 329         write_sequnlock_irqrestore(&xtime_lock, flags);
 330
 331         clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
 332
 333         return 0;
 334 }
 335
 336 /* sysfs resume/suspend bits for timekeeping */
 337 static struct sysdev_class timekeeping_sysclass = {
 338         .name           = "timekeeping",
 339         .resume         = timekeeping_resume,
 340         .suspend        = timekeeping_suspend,
 341 };
 342
 343 static struct sys_device device_timer = {
 344         .id             = 0,
 345         .cls            = &timekeeping_sysclass,
 346 };
 347
 348 static int __init timekeeping_init_device(void)
 349 {
 350         int error = sysdev_class_register(&timekeeping_sysclass);
 351         if (!error)
 352                 error = sysdev_register(&device_timer);
 353         return error;
 354 }
 355
 356 device_initcall(timekeeping_init_device);
 357
 358 /*
 359  * If the error is already larger, we look ahead even further
 360  * to compensate for late or lost adjustments.
 361  */
 362 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
 363                                                  s64 *offset)
 364 {
 365         s64 tick_error, i;
 366         u32 look_ahead, adj;
 367         s32 error2, mult;
 368
 369         /*
 370          * Use the current error value to determine how much to look ahead.
 371          * The larger the error the slower we adjust for it to avoid problems
 372          * with losing too many ticks, otherwise we would overadjust and
 373          * produce an even larger error.  The smaller the adjustment the
 374          * faster we try to adjust for it, as lost ticks can do less harm
 375          * here.  This is tuned so that an error of about 1 msec is adusted
 376          * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
 377          */
 378         error2 = clock->error >> (TICK_LENGTH_SHIFT + 22 - 2 * SHIFT_HZ);
 379         error2 = abs(error2);
 380         for (look_ahead = 0; error2 > 0; look_ahead++)
 381                 error2 >>= 2;
 382
 383         /*
 384          * Now calculate the error in (1 << look_ahead) ticks, but first
 385          * remove the single look ahead already included in the error.
 386          */
 387         tick_error = current_tick_length() >>
 388                 (TICK_LENGTH_SHIFT - clock->shift + 1);
 389         tick_error -= clock->xtime_interval >> 1;
 390         error = ((error - tick_error) >> look_ahead) + tick_error;
 391
 392         /* Finally calculate the adjustment shift value.  */
 393         i = *interval;
 394         mult = 1;
 395         if (error < 0) {
 396                 error = -error;
 397                 *interval = -*interval;
 398                 *offset = -*offset;
 399                 mult = -1;
 400         }
 401         for (adj = 0; error > i; adj++)
 402                 error >>= 1;
 403
 404         *interval <<= adj;
 405         *offset <<= adj;
 406         return mult << adj;
 407 }
 408
 409 /*
 410  * Adjust the multiplier to reduce the error value,
 411  * this is optimized for the most common adjustments of -1,0,1,
 412  * for other values we can do a bit more work.
 413  */
 414 static void clocksource_adjust(s64 offset)
 415 {
 416         s64 error, interval = clock->cycle_interval;
 417         int adj;
 418
 419         error = clock->error >> (TICK_LENGTH_SHIFT - clock->shift - 1);
 420         if (error > interval) {
 421                 error >>= 2;
 422                 if (likely(error <= interval))
 423                         adj = 1;
 424                 else
 425                         adj = clocksource_bigadjust(error, &interval, &offset);
 426         } else if (error < -interval) {
 427                 error >>= 2;
 428                 if (likely(error >= -interval)) {
 429                         adj = -1;
 430                         interval = -interval;
 431                         offset = -offset;
 432                 } else
 433                         adj = clocksource_bigadjust(error, &interval, &offset);
 434         } else
 435                 return;
 436
 437         clock->mult += adj;
 438         clock->xtime_interval += interval;
 439         clock->xtime_nsec -= offset;
 440         clock->error -= (interval - offset) <<
 441                         (TICK_LENGTH_SHIFT - clock->shift);
 442 }
 443
 444 /**
 445  * update_wall_time - Uses the current clocksource to increment the wall time
 446  *
 447  * Called from the timer interrupt, must hold a write on xtime_lock.
 448  */
 449 void update_wall_time(void)
 450 {
 451         cycle_t offset;
 452
 453         /* Make sure we're fully resumed: */
 454         if (unlikely(timekeeping_suspended))
 455                 return;
 456
 457 #ifdef CONFIG_GENERIC_TIME
 458         offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask;
 459 #else
 460         offset = clock->cycle_interval;
 461 #endif
 462         clock->xtime_nsec += (s64)xtime.tv_nsec << clock->shift;
 463
 464         /* normally this loop will run just once, however in the
 465          * case of lost or late ticks, it will accumulate correctly.
 466          */
 467         while (offset >= clock->cycle_interval) {
 468                 /* accumulate one interval */
 469                 clock->xtime_nsec += clock->xtime_interval;
 470                 clock->cycle_last += clock->cycle_interval;
 471                 offset -= clock->cycle_interval;
 472
 473                 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
 474                         clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
 475                         xtime.tv_sec++;
 476                         second_overflow();
 477                 }
 478
 479                 /* accumulate error between NTP and clock interval */
 480                 clock->error += current_tick_length();
 481                 clock->error -= clock->xtime_interval << (TICK_LENGTH_SHIFT - clock->shift);
 482         }
 483
 484         /* correct the clock when NTP error is too big */
 485         clocksource_adjust(offset);
 486
 487         /* store full nanoseconds into xtime */
 488         xtime.tv_nsec = (s64)clock->xtime_nsec >> clock->shift;
 489         clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
 490
 491         update_xtime_cache(cyc2ns(clock, offset));
 492
 493         /* check to see if there is a new clocksource to use */
 494         change_clocksource();
 495         update_vsyscall(&xtime, clock);
 496 }
 497
 498 /**
 499  * getboottime - Return the real time of system boot.
 500  * @ts:         pointer to the timespec to be set
 501  *
 502  * Returns the time of day in a timespec.
 503  *
 504  * This is based on the wall_to_monotonic offset and the total suspend
 505  * time. Calls to settimeofday will affect the value returned (which
 506  * basically means that however wrong your real time clock is at boot time,
 507  * you get the right time here).
 508  */
 509 void getboottime(struct timespec *ts)
 510 {
 511         set_normalized_timespec(ts,
 512                 - (wall_to_monotonic.tv_sec + total_sleep_time),
 513                 - wall_to_monotonic.tv_nsec);
 514 }
 515
 516 /**
 517  * monotonic_to_bootbased - Convert the monotonic time to boot based.
 518  * @ts:         pointer to the timespec to be converted
 519  */
 520 void monotonic_to_bootbased(struct timespec *ts)
 521 {
 522         ts->tv_sec += total_sleep_time;
 523 }
 524
 525 unsigned long get_seconds(void)
 526 {
 527         return xtime_cache.tv_sec;
 528 }
 529 EXPORT_SYMBOL(get_seconds);
 530
 531
 532 struct timespec current_kernel_time(void)
 533 {
 534         struct timespec now;
 535         unsigned long seq;
 536
 537         do {
 538                 seq = read_seqbegin(&xtime_lock);
 539
 540                 now = xtime_cache;
 541         } while (read_seqretry(&xtime_lock, seq));
 542
 543         return now;
 544 }
 545 EXPORT_SYMBOL(current_kernel_time);