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