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/sched.h>
  17 #include <linux/syscore_ops.h>
  18 #include <linux/clocksource.h>
  19 #include <linux/jiffies.h>
  20 #include <linux/time.h>
  21 #include <linux/tick.h>
  22 #include <linux/stop_machine.h>
  23
  24 /* Structure holding internal timekeeping values. */
  25 struct timekeeper {
  26         /* Current clocksource used for timekeeping. */
  27         struct clocksource *clock;
  28         /* The shift value of the current clocksource. */
  29         int     shift;
  30
  31         /* Number of clock cycles in one NTP interval. */
  32         cycle_t cycle_interval;
  33         /* Number of clock shifted nano seconds in one NTP interval. */
  34         u64     xtime_interval;
  35         /* shifted nano seconds left over when rounding cycle_interval */
  36         s64     xtime_remainder;
  37         /* Raw nano seconds accumulated per NTP interval. */
  38         u32     raw_interval;
  39
  40         /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
  41         u64     xtime_nsec;
  42         /* Difference between accumulated time and NTP time in ntp
  43          * shifted nano seconds. */
  44         s64     ntp_error;
  45         /* Shift conversion between clock shifted nano seconds and
  46          * ntp shifted nano seconds. */
  47         int     ntp_error_shift;
  48         /* NTP adjusted clock multiplier */
  49         u32     mult;
  50
  51         /* The current time */
  52         struct timespec xtime;
  53         /*
  54          * wall_to_monotonic is what we need to add to xtime (or xtime corrected
  55          * for sub jiffie times) to get to monotonic time.  Monotonic is pegged
  56          * at zero at system boot time, so wall_to_monotonic will be negative,
  57          * however, we will ALWAYS keep the tv_nsec part positive so we can use
  58          * the usual normalization.
  59          *
  60          * wall_to_monotonic is moved after resume from suspend for the
  61          * monotonic time not to jump. We need to add total_sleep_time to
  62          * wall_to_monotonic to get the real boot based time offset.
  63          *
  64          * - wall_to_monotonic is no longer the boot time, getboottime must be
  65          * used instead.
  66          */
  67         struct timespec wall_to_monotonic;
  68         /* time spent in suspend */
  69         struct timespec total_sleep_time;
  70         /* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. */
  71         struct timespec raw_time;
  72
  73         /* Seqlock for all timekeeper values */
  74         seqlock_t lock;
  75 };
  76
  77 static struct timekeeper timekeeper;
  78
  79 /*
  80  * This read-write spinlock protects us from races in SMP while
  81  * playing with xtime.
  82  */
  83 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
  84
  85
  86 /* flag for if timekeeping is suspended */
  87 int __read_mostly timekeeping_suspended;
  88
  89
  90
  91 /**
  92  * timekeeper_setup_internals - Set up internals to use clocksource clock.
  93  *
  94  * @clock:              Pointer to clocksource.
  95  *
  96  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
  97  * pair and interval request.
  98  *
  99  * Unless you're the timekeeping code, you should not be using this!
 100  */
 101 static void timekeeper_setup_internals(struct clocksource *clock)
 102 {
 103         cycle_t interval;
 104         u64 tmp, ntpinterval;
 105
 106         timekeeper.clock = clock;
 107         clock->cycle_last = clock->read(clock);
 108
 109         /* Do the ns -> cycle conversion first, using original mult */
 110         tmp = NTP_INTERVAL_LENGTH;
 111         tmp <<= clock->shift;
 112         ntpinterval = tmp;
 113         tmp += clock->mult/2;
 114         do_div(tmp, clock->mult);
 115         if (tmp == 0)
 116                 tmp = 1;
 117
 118         interval = (cycle_t) tmp;
 119         timekeeper.cycle_interval = interval;
 120
 121         /* Go back from cycles -> shifted ns */
 122         timekeeper.xtime_interval = (u64) interval * clock->mult;
 123         timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
 124         timekeeper.raw_interval =
 125                 ((u64) interval * clock->mult) >> clock->shift;
 126
 127         timekeeper.xtime_nsec = 0;
 128         timekeeper.shift = clock->shift;
 129
 130         timekeeper.ntp_error = 0;
 131         timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 132
 133         /*
 134          * The timekeeper keeps its own mult values for the currently
 135          * active clocksource. These value will be adjusted via NTP
 136          * to counteract clock drifting.
 137          */
 138         timekeeper.mult = clock->mult;
 139 }
 140
 141 /* Timekeeper helper functions. */
 142 static inline s64 timekeeping_get_ns(void)
 143 {
 144         cycle_t cycle_now, cycle_delta;
 145         struct clocksource *clock;
 146
 147         /* read clocksource: */
 148         clock = timekeeper.clock;
 149         cycle_now = clock->read(clock);
 150
 151         /* calculate the delta since the last update_wall_time: */
 152         cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
 153
 154         /* return delta convert to nanoseconds using ntp adjusted mult. */
 155         return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
 156                                   timekeeper.shift);
 157 }
 158
 159 static inline s64 timekeeping_get_ns_raw(void)
 160 {
 161         cycle_t cycle_now, cycle_delta;
 162         struct clocksource *clock;
 163
 164         /* read clocksource: */
 165         clock = timekeeper.clock;
 166         cycle_now = clock->read(clock);
 167
 168         /* calculate the delta since the last update_wall_time: */
 169         cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
 170
 171         /* return delta convert to nanoseconds. */
 172         return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
 173 }
 174
 175 /* must hold xtime_lock */
 176 void timekeeping_leap_insert(int leapsecond)
 177 {
 178         unsigned long flags;
 179
 180         write_seqlock_irqsave(&timekeeper.lock, flags);
 181
 182         timekeeper.xtime.tv_sec += leapsecond;
 183         timekeeper.wall_to_monotonic.tv_sec -= leapsecond;
 184         update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
 185                          timekeeper.clock, timekeeper.mult);
 186
 187         write_sequnlock_irqrestore(&timekeeper.lock, flags);
 188
 189 }
 190
 191 /**
 192  * timekeeping_forward_now - update clock to the current time
 193  *
 194  * Forward the current clock to update its state since the last call to
 195  * update_wall_time(). This is useful before significant clock changes,
 196  * as it avoids having to deal with this time offset explicitly.
 197  */
 198 static void timekeeping_forward_now(void)
 199 {
 200         cycle_t cycle_now, cycle_delta;
 201         struct clocksource *clock;
 202         s64 nsec;
 203
 204         clock = timekeeper.clock;
 205         cycle_now = clock->read(clock);
 206         cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
 207         clock->cycle_last = cycle_now;
 208
 209         nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
 210                                   timekeeper.shift);
 211
 212         /* If arch requires, add in gettimeoffset() */
 213         nsec += arch_gettimeoffset();
 214
 215         timespec_add_ns(&timekeeper.xtime, nsec);
 216
 217         nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
 218         timespec_add_ns(&timekeeper.raw_time, nsec);
 219 }
 220
 221 /**
 222  * getnstimeofday - Returns the time of day in a timespec
 223  * @ts:         pointer to the timespec to be set
 224  *
 225  * Returns the time of day in a timespec.
 226  */
 227 void getnstimeofday(struct timespec *ts)
 228 {
 229         unsigned long seq;
 230         s64 nsecs;
 231
 232         WARN_ON(timekeeping_suspended);
 233
 234         do {
 235                 seq = read_seqbegin(&timekeeper.lock);
 236
 237                 *ts = timekeeper.xtime;
 238                 nsecs = timekeeping_get_ns();
 239
 240                 /* If arch requires, add in gettimeoffset() */
 241                 nsecs += arch_gettimeoffset();
 242
 243         } while (read_seqretry(&timekeeper.lock, seq));
 244
 245         timespec_add_ns(ts, nsecs);
 246 }
 247
 248 EXPORT_SYMBOL(getnstimeofday);
 249
 250 ktime_t ktime_get(void)
 251 {
 252         unsigned int seq;
 253         s64 secs, nsecs;
 254
 255         WARN_ON(timekeeping_suspended);
 256
 257         do {
 258                 seq = read_seqbegin(&timekeeper.lock);
 259                 secs = timekeeper.xtime.tv_sec +
 260                                 timekeeper.wall_to_monotonic.tv_sec;
 261                 nsecs = timekeeper.xtime.tv_nsec +
 262                                 timekeeper.wall_to_monotonic.tv_nsec;
 263                 nsecs += timekeeping_get_ns();
 264                 /* If arch requires, add in gettimeoffset() */
 265                 nsecs += arch_gettimeoffset();
 266
 267         } while (read_seqretry(&timekeeper.lock, seq));
 268         /*
 269          * Use ktime_set/ktime_add_ns to create a proper ktime on
 270          * 32-bit architectures without CONFIG_KTIME_SCALAR.
 271          */
 272         return ktime_add_ns(ktime_set(secs, 0), nsecs);
 273 }
 274 EXPORT_SYMBOL_GPL(ktime_get);
 275
 276 /**
 277  * ktime_get_ts - get the monotonic clock in timespec format
 278  * @ts:         pointer to timespec variable
 279  *
 280  * The function calculates the monotonic clock from the realtime
 281  * clock and the wall_to_monotonic offset and stores the result
 282  * in normalized timespec format in the variable pointed to by @ts.
 283  */
 284 void ktime_get_ts(struct timespec *ts)
 285 {
 286         struct timespec tomono;
 287         unsigned int seq;
 288         s64 nsecs;
 289
 290         WARN_ON(timekeeping_suspended);
 291
 292         do {
 293                 seq = read_seqbegin(&timekeeper.lock);
 294                 *ts = timekeeper.xtime;
 295                 tomono = timekeeper.wall_to_monotonic;
 296                 nsecs = timekeeping_get_ns();
 297                 /* If arch requires, add in gettimeoffset() */
 298                 nsecs += arch_gettimeoffset();
 299
 300         } while (read_seqretry(&timekeeper.lock, seq));
 301
 302         set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
 303                                 ts->tv_nsec + tomono.tv_nsec + nsecs);
 304 }
 305 EXPORT_SYMBOL_GPL(ktime_get_ts);
 306
 307 #ifdef CONFIG_NTP_PPS
 308
 309 /**
 310  * getnstime_raw_and_real - get day and raw monotonic time in timespec format
 311  * @ts_raw:     pointer to the timespec to be set to raw monotonic time
 312  * @ts_real:    pointer to the timespec to be set to the time of day
 313  *
 314  * This function reads both the time of day and raw monotonic time at the
 315  * same time atomically and stores the resulting timestamps in timespec
 316  * format.
 317  */
 318 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
 319 {
 320         unsigned long seq;
 321         s64 nsecs_raw, nsecs_real;
 322
 323         WARN_ON_ONCE(timekeeping_suspended);
 324
 325         do {
 326                 u32 arch_offset;
 327
 328                 seq = read_seqbegin(&timekeeper.lock);
 329
 330                 *ts_raw = timekeeper.raw_time;
 331                 *ts_real = timekeeper.xtime;
 332
 333                 nsecs_raw = timekeeping_get_ns_raw();
 334                 nsecs_real = timekeeping_get_ns();
 335
 336                 /* If arch requires, add in gettimeoffset() */
 337                 arch_offset = arch_gettimeoffset();
 338                 nsecs_raw += arch_offset;
 339                 nsecs_real += arch_offset;
 340
 341         } while (read_seqretry(&timekeeper.lock, seq));
 342
 343         timespec_add_ns(ts_raw, nsecs_raw);
 344         timespec_add_ns(ts_real, nsecs_real);
 345 }
 346 EXPORT_SYMBOL(getnstime_raw_and_real);
 347
 348 #endif /* CONFIG_NTP_PPS */
 349
 350 /**
 351  * do_gettimeofday - Returns the time of day in a timeval
 352  * @tv:         pointer to the timeval to be set
 353  *
 354  * NOTE: Users should be converted to using getnstimeofday()
 355  */
 356 void do_gettimeofday(struct timeval *tv)
 357 {
 358         struct timespec now;
 359
 360         getnstimeofday(&now);
 361         tv->tv_sec = now.tv_sec;
 362         tv->tv_usec = now.tv_nsec/1000;
 363 }
 364
 365 EXPORT_SYMBOL(do_gettimeofday);
 366 /**
 367  * do_settimeofday - Sets the time of day
 368  * @tv:         pointer to the timespec variable containing the new time
 369  *
 370  * Sets the time of day to the new time and update NTP and notify hrtimers
 371  */
 372 int do_settimeofday(const struct timespec *tv)
 373 {
 374         struct timespec ts_delta;
 375         unsigned long flags1,flags2;
 376
 377         if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
 378                 return -EINVAL;
 379
 380         write_seqlock_irqsave(&xtime_lock, flags1);
 381         write_seqlock_irqsave(&timekeeper.lock, flags2);
 382
 383         timekeeping_forward_now();
 384
 385         ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
 386         ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
 387         timekeeper.wall_to_monotonic =
 388                         timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
 389
 390         timekeeper.xtime = *tv;
 391
 392         timekeeper.ntp_error = 0;
 393         ntp_clear();
 394
 395         update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
 396                         timekeeper.clock, timekeeper.mult);
 397
 398         write_sequnlock_irqrestore(&timekeeper.lock, flags2);
 399         write_sequnlock_irqrestore(&xtime_lock, flags1);
 400
 401         /* signal hrtimers about time change */
 402         clock_was_set();
 403
 404         return 0;
 405 }
 406
 407 EXPORT_SYMBOL(do_settimeofday);
 408
 409
 410 /**
 411  * timekeeping_inject_offset - Adds or subtracts from the current time.
 412  * @tv:         pointer to the timespec variable containing the offset
 413  *
 414  * Adds or subtracts an offset value from the current time.
 415  */
 416 int timekeeping_inject_offset(struct timespec *ts)
 417 {
 418         unsigned long flags1,flags2;
 419
 420         if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
 421                 return -EINVAL;
 422
 423         write_seqlock_irqsave(&xtime_lock, flags1);
 424         write_seqlock_irqsave(&timekeeper.lock, flags2);
 425
 426         timekeeping_forward_now();
 427
 428         timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
 429         timekeeper.wall_to_monotonic =
 430                                 timespec_sub(timekeeper.wall_to_monotonic, *ts);
 431
 432         timekeeper.ntp_error = 0;
 433         ntp_clear();
 434
 435         update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
 436                         timekeeper.clock, timekeeper.mult);
 437
 438         write_sequnlock_irqrestore(&timekeeper.lock, flags2);
 439         write_sequnlock_irqrestore(&xtime_lock, flags1);
 440
 441         /* signal hrtimers about time change */
 442         clock_was_set();
 443
 444         return 0;
 445 }
 446 EXPORT_SYMBOL(timekeeping_inject_offset);
 447
 448 /**
 449  * change_clocksource - Swaps clocksources if a new one is available
 450  *
 451  * Accumulates current time interval and initializes new clocksource
 452  */
 453 static int change_clocksource(void *data)
 454 {
 455         struct clocksource *new, *old;
 456
 457         new = (struct clocksource *) data;
 458
 459         timekeeping_forward_now();
 460         if (!new->enable || new->enable(new) == 0) {
 461                 old = timekeeper.clock;
 462                 timekeeper_setup_internals(new);
 463                 if (old->disable)
 464                         old->disable(old);
 465         }
 466         return 0;
 467 }
 468
 469 /**
 470  * timekeeping_notify - Install a new clock source
 471  * @clock:              pointer to the clock source
 472  *
 473  * This function is called from clocksource.c after a new, better clock
 474  * source has been registered. The caller holds the clocksource_mutex.
 475  */
 476 void timekeeping_notify(struct clocksource *clock)
 477 {
 478         if (timekeeper.clock == clock)
 479                 return;
 480         stop_machine(change_clocksource, clock, NULL);
 481         tick_clock_notify();
 482 }
 483
 484 /**
 485  * ktime_get_real - get the real (wall-) time in ktime_t format
 486  *
 487  * returns the time in ktime_t format
 488  */
 489 ktime_t ktime_get_real(void)
 490 {
 491         struct timespec now;
 492
 493         getnstimeofday(&now);
 494
 495         return timespec_to_ktime(now);
 496 }
 497 EXPORT_SYMBOL_GPL(ktime_get_real);
 498
 499 /**
 500  * getrawmonotonic - Returns the raw monotonic time in a timespec
 501  * @ts:         pointer to the timespec to be set
 502  *
 503  * Returns the raw monotonic time (completely un-modified by ntp)
 504  */
 505 void getrawmonotonic(struct timespec *ts)
 506 {
 507         unsigned long seq;
 508         s64 nsecs;
 509
 510         do {
 511                 seq = read_seqbegin(&timekeeper.lock);
 512                 nsecs = timekeeping_get_ns_raw();
 513                 *ts = timekeeper.raw_time;
 514
 515         } while (read_seqretry(&timekeeper.lock, seq));
 516
 517         timespec_add_ns(ts, nsecs);
 518 }
 519 EXPORT_SYMBOL(getrawmonotonic);
 520
 521
 522 /**
 523  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
 524  */
 525 int timekeeping_valid_for_hres(void)
 526 {
 527         unsigned long seq;
 528         int ret;
 529
 530         do {
 531                 seq = read_seqbegin(&timekeeper.lock);
 532
 533                 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
 534
 535         } while (read_seqretry(&timekeeper.lock, seq));
 536
 537         return ret;
 538 }
 539
 540 /**
 541  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
 542  */
 543 u64 timekeeping_max_deferment(void)
 544 {
 545         unsigned long seq;
 546         u64 ret;
 547         do {
 548                 seq = read_seqbegin(&timekeeper.lock);
 549
 550                 ret = timekeeper.clock->max_idle_ns;
 551
 552         } while (read_seqretry(&timekeeper.lock, seq));
 553
 554         return ret;
 555 }
 556
 557 /**
 558  * read_persistent_clock -  Return time from the persistent clock.
 559  *
 560  * Weak dummy function for arches that do not yet support it.
 561  * Reads the time from the battery backed persistent clock.
 562  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
 563  *
 564  *  XXX - Do be sure to remove it once all arches implement it.
 565  */
 566 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
 567 {
 568         ts->tv_sec = 0;
 569         ts->tv_nsec = 0;
 570 }
 571
 572 /**
 573  * read_boot_clock -  Return time of the system start.
 574  *
 575  * Weak dummy function for arches that do not yet support it.
 576  * Function to read the exact time the system has been started.
 577  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
 578  *
 579  *  XXX - Do be sure to remove it once all arches implement it.
 580  */
 581 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
 582 {
 583         ts->tv_sec = 0;
 584         ts->tv_nsec = 0;
 585 }
 586
 587 /*
 588  * timekeeping_init - Initializes the clocksource and common timekeeping values
 589  */
 590 void __init timekeeping_init(void)
 591 {
 592         struct clocksource *clock;
 593         unsigned long flags;
 594         struct timespec now, boot;
 595
 596         read_persistent_clock(&now);
 597         read_boot_clock(&boot);
 598
 599         seqlock_init(&timekeeper.lock);
 600
 601         write_seqlock_irqsave(&xtime_lock, flags);
 602         ntp_init();
 603         write_sequnlock_irqrestore(&xtime_lock, flags);
 604
 605         write_seqlock_irqsave(&timekeeper.lock, flags);
 606         clock = clocksource_default_clock();
 607         if (clock->enable)
 608                 clock->enable(clock);
 609         timekeeper_setup_internals(clock);
 610
 611         timekeeper.xtime.tv_sec = now.tv_sec;
 612         timekeeper.xtime.tv_nsec = now.tv_nsec;
 613         timekeeper.raw_time.tv_sec = 0;
 614         timekeeper.raw_time.tv_nsec = 0;
 615         if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
 616                 boot.tv_sec = timekeeper.xtime.tv_sec;
 617                 boot.tv_nsec = timekeeper.xtime.tv_nsec;
 618         }
 619         set_normalized_timespec(&timekeeper.wall_to_monotonic,
 620                                 -boot.tv_sec, -boot.tv_nsec);
 621         timekeeper.total_sleep_time.tv_sec = 0;
 622         timekeeper.total_sleep_time.tv_nsec = 0;
 623         write_sequnlock_irqrestore(&timekeeper.lock, flags);
 624 }
 625
 626 /* time in seconds when suspend began */
 627 static struct timespec timekeeping_suspend_time;
 628
 629 /**
 630  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
 631  * @delta: pointer to a timespec delta value
 632  *
 633  * Takes a timespec offset measuring a suspend interval and properly
 634  * adds the sleep offset to the timekeeping variables.
 635  */
 636 static void __timekeeping_inject_sleeptime(struct timespec *delta)
 637 {
 638         if (!timespec_valid(delta)) {
 639                 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
 640                                         "sleep delta value!\n");
 641                 return;
 642         }
 643
 644         timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
 645         timekeeper.wall_to_monotonic =
 646                         timespec_sub(timekeeper.wall_to_monotonic, *delta);
 647         timekeeper.total_sleep_time = timespec_add(
 648                                         timekeeper.total_sleep_time, *delta);
 649 }
 650
 651
 652 /**
 653  * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
 654  * @delta: pointer to a timespec delta value
 655  *
 656  * This hook is for architectures that cannot support read_persistent_clock
 657  * because their RTC/persistent clock is only accessible when irqs are enabled.
 658  *
 659  * This function should only be called by rtc_resume(), and allows
 660  * a suspend offset to be injected into the timekeeping values.
 661  */
 662 void timekeeping_inject_sleeptime(struct timespec *delta)
 663 {
 664         unsigned long flags1,flags2;
 665         struct timespec ts;
 666
 667         /* Make sure we don't set the clock twice */
 668         read_persistent_clock(&ts);
 669         if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
 670                 return;
 671
 672         write_seqlock_irqsave(&xtime_lock, flags1);
 673         write_seqlock_irqsave(&timekeeper.lock, flags2);
 674
 675         timekeeping_forward_now();
 676
 677         __timekeeping_inject_sleeptime(delta);
 678
 679         timekeeper.ntp_error = 0;
 680         ntp_clear();
 681         update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
 682                         timekeeper.clock, timekeeper.mult);
 683
 684         write_sequnlock_irqrestore(&timekeeper.lock, flags2);
 685         write_sequnlock_irqrestore(&xtime_lock, flags1);
 686
 687         /* signal hrtimers about time change */
 688         clock_was_set();
 689 }
 690
 691
 692 /**
 693  * timekeeping_resume - Resumes the generic timekeeping subsystem.
 694  *
 695  * This is for the generic clocksource timekeeping.
 696  * xtime/wall_to_monotonic/jiffies/etc are
 697  * still managed by arch specific suspend/resume code.
 698  */
 699 static void timekeeping_resume(void)
 700 {
 701         unsigned long flags1,flags2;
 702         struct timespec ts;
 703
 704         read_persistent_clock(&ts);
 705
 706         clocksource_resume();
 707
 708         write_seqlock_irqsave(&xtime_lock, flags1);
 709         write_seqlock_irqsave(&timekeeper.lock, flags2);
 710
 711         if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
 712                 ts = timespec_sub(ts, timekeeping_suspend_time);
 713                 __timekeeping_inject_sleeptime(&ts);
 714         }
 715         /* re-base the last cycle value */
 716         timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
 717         timekeeper.ntp_error = 0;
 718         timekeeping_suspended = 0;
 719         write_sequnlock_irqrestore(&timekeeper.lock, flags2);
 720         write_sequnlock_irqrestore(&xtime_lock, flags1);
 721
 722         touch_softlockup_watchdog();
 723
 724         clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
 725
 726         /* Resume hrtimers */
 727         hrtimers_resume();
 728 }
 729
 730 static int timekeeping_suspend(void)
 731 {
 732         unsigned long flags1,flags2;
 733         struct timespec         delta, delta_delta;
 734         static struct timespec  old_delta;
 735
 736         read_persistent_clock(&timekeeping_suspend_time);
 737
 738         write_seqlock_irqsave(&xtime_lock, flags1);
 739         write_seqlock_irqsave(&timekeeper.lock, flags2);
 740         timekeeping_forward_now();
 741         timekeeping_suspended = 1;
 742
 743         /*
 744          * To avoid drift caused by repeated suspend/resumes,
 745          * which each can add ~1 second drift error,
 746          * try to compensate so the difference in system time
 747          * and persistent_clock time stays close to constant.
 748          */
 749         delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
 750         delta_delta = timespec_sub(delta, old_delta);
 751         if (abs(delta_delta.tv_sec)  >= 2) {
 752                 /*
 753                  * if delta_delta is too large, assume time correction
 754                  * has occured and set old_delta to the current delta.
 755                  */
 756                 old_delta = delta;
 757         } else {
 758                 /* Otherwise try to adjust old_system to compensate */
 759                 timekeeping_suspend_time =
 760                         timespec_add(timekeeping_suspend_time, delta_delta);
 761         }
 762         write_sequnlock_irqrestore(&timekeeper.lock, flags2);
 763         write_sequnlock_irqrestore(&xtime_lock, flags1);
 764
 765         clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
 766         clocksource_suspend();
 767
 768         return 0;
 769 }
 770
 771 /* sysfs resume/suspend bits for timekeeping */
 772 static struct syscore_ops timekeeping_syscore_ops = {
 773         .resume         = timekeeping_resume,
 774         .suspend        = timekeeping_suspend,
 775 };
 776
 777 static int __init timekeeping_init_ops(void)
 778 {
 779         register_syscore_ops(&timekeeping_syscore_ops);
 780         return 0;
 781 }
 782
 783 device_initcall(timekeeping_init_ops);
 784
 785 /*
 786  * If the error is already larger, we look ahead even further
 787  * to compensate for late or lost adjustments.
 788  */
 789 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
 790                                                  s64 *offset)
 791 {
 792         s64 tick_error, i;
 793         u32 look_ahead, adj;
 794         s32 error2, mult;
 795
 796         /*
 797          * Use the current error value to determine how much to look ahead.
 798          * The larger the error the slower we adjust for it to avoid problems
 799          * with losing too many ticks, otherwise we would overadjust and
 800          * produce an even larger error.  The smaller the adjustment the
 801          * faster we try to adjust for it, as lost ticks can do less harm
 802          * here.  This is tuned so that an error of about 1 msec is adjusted
 803          * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
 804          */
 805         error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
 806         error2 = abs(error2);
 807         for (look_ahead = 0; error2 > 0; look_ahead++)
 808                 error2 >>= 2;
 809
 810         /*
 811          * Now calculate the error in (1 << look_ahead) ticks, but first
 812          * remove the single look ahead already included in the error.
 813          */
 814         tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
 815         tick_error -= timekeeper.xtime_interval >> 1;
 816         error = ((error - tick_error) >> look_ahead) + tick_error;
 817
 818         /* Finally calculate the adjustment shift value.  */
 819         i = *interval;
 820         mult = 1;
 821         if (error < 0) {
 822                 error = -error;
 823                 *interval = -*interval;
 824                 *offset = -*offset;
 825                 mult = -1;
 826         }
 827         for (adj = 0; error > i; adj++)
 828                 error >>= 1;
 829
 830         *interval <<= adj;
 831         *offset <<= adj;
 832         return mult << adj;
 833 }
 834
 835 /*
 836  * Adjust the multiplier to reduce the error value,
 837  * this is optimized for the most common adjustments of -1,0,1,
 838  * for other values we can do a bit more work.
 839  */
 840 static void timekeeping_adjust(s64 offset)
 841 {
 842         s64 error, interval = timekeeper.cycle_interval;
 843         int adj;
 844
 845         /*
 846          * The point of this is to check if the error is greater then half
 847          * an interval.
 848          *
 849          * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
 850          *
 851          * Note we subtract one in the shift, so that error is really error*2.
 852          * This "saves" dividing(shifting) interval twice, but keeps the
 853          * (error > interval) comparison as still measuring if error is
 854          * larger then half an interval.
 855          *
 856          * Note: It does not "save" on aggravation when reading the code.
 857          */
 858         error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
 859         if (error > interval) {
 860                 /*
 861                  * We now divide error by 4(via shift), which checks if
 862                  * the error is greater then twice the interval.
 863                  * If it is greater, we need a bigadjust, if its smaller,
 864                  * we can adjust by 1.
 865                  */
 866                 error >>= 2;
 867                 /*
 868                  * XXX - In update_wall_time, we round up to the next
 869                  * nanosecond, and store the amount rounded up into
 870                  * the error. This causes the likely below to be unlikely.
 871                  *
 872                  * The proper fix is to avoid rounding up by using
 873                  * the high precision timekeeper.xtime_nsec instead of
 874                  * xtime.tv_nsec everywhere. Fixing this will take some
 875                  * time.
 876                  */
 877                 if (likely(error <= interval))
 878                         adj = 1;
 879                 else
 880                         adj = timekeeping_bigadjust(error, &interval, &offset);
 881         } else if (error < -interval) {
 882                 /* See comment above, this is just switched for the negative */
 883                 error >>= 2;
 884                 if (likely(error >= -interval)) {
 885                         adj = -1;
 886                         interval = -interval;
 887                         offset = -offset;
 888                 } else
 889                         adj = timekeeping_bigadjust(error, &interval, &offset);
 890         } else /* No adjustment needed */
 891                 return;
 892
 893         WARN_ONCE(timekeeper.clock->maxadj &&
 894                         (timekeeper.mult + adj > timekeeper.clock->mult +
 895                                                 timekeeper.clock->maxadj),
 896                         "Adjusting %s more then 11%% (%ld vs %ld)\n",
 897                         timekeeper.clock->name, (long)timekeeper.mult + adj,
 898                         (long)timekeeper.clock->mult +
 899                                 timekeeper.clock->maxadj);
 900         /*
 901          * So the following can be confusing.
 902          *
 903          * To keep things simple, lets assume adj == 1 for now.
 904          *
 905          * When adj != 1, remember that the interval and offset values
 906          * have been appropriately scaled so the math is the same.
 907          *
 908          * The basic idea here is that we're increasing the multiplier
 909          * by one, this causes the xtime_interval to be incremented by
 910          * one cycle_interval. This is because:
 911          *      xtime_interval = cycle_interval * mult
 912          * So if mult is being incremented by one:
 913          *      xtime_interval = cycle_interval * (mult + 1)
 914          * Its the same as:
 915          *      xtime_interval = (cycle_interval * mult) + cycle_interval
 916          * Which can be shortened to:
 917          *      xtime_interval += cycle_interval
 918          *
 919          * So offset stores the non-accumulated cycles. Thus the current
 920          * time (in shifted nanoseconds) is:
 921          *      now = (offset * adj) + xtime_nsec
 922          * Now, even though we're adjusting the clock frequency, we have
 923          * to keep time consistent. In other words, we can't jump back
 924          * in time, and we also want to avoid jumping forward in time.
 925          *
 926          * So given the same offset value, we need the time to be the same
 927          * both before and after the freq adjustment.
 928          *      now = (offset * adj_1) + xtime_nsec_1
 929          *      now = (offset * adj_2) + xtime_nsec_2
 930          * So:
 931          *      (offset * adj_1) + xtime_nsec_1 =
 932          *              (offset * adj_2) + xtime_nsec_2
 933          * And we know:
 934          *      adj_2 = adj_1 + 1
 935          * So:
 936          *      (offset * adj_1) + xtime_nsec_1 =
 937          *              (offset * (adj_1+1)) + xtime_nsec_2
 938          *      (offset * adj_1) + xtime_nsec_1 =
 939          *              (offset * adj_1) + offset + xtime_nsec_2
 940          * Canceling the sides:
 941          *      xtime_nsec_1 = offset + xtime_nsec_2
 942          * Which gives us:
 943          *      xtime_nsec_2 = xtime_nsec_1 - offset
 944          * Which simplfies to:
 945          *      xtime_nsec -= offset
 946          *
 947          * XXX - TODO: Doc ntp_error calculation.
 948          */
 949         timekeeper.mult += adj;
 950         timekeeper.xtime_interval += interval;
 951         timekeeper.xtime_nsec -= offset;
 952         timekeeper.ntp_error -= (interval - offset) <<
 953                                 timekeeper.ntp_error_shift;
 954 }
 955
 956
 957 /**
 958  * logarithmic_accumulation - shifted accumulation of cycles
 959  *
 960  * This functions accumulates a shifted interval of cycles into
 961  * into a shifted interval nanoseconds. Allows for O(log) accumulation
 962  * loop.
 963  *
 964  * Returns the unconsumed cycles.
 965  */
 966 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
 967 {
 968         u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
 969         u64 raw_nsecs;
 970
 971         /* If the offset is smaller then a shifted interval, do nothing */
 972         if (offset < timekeeper.cycle_interval<<shift)
 973                 return offset;
 974
 975         /* Accumulate one shifted interval */
 976         offset -= timekeeper.cycle_interval << shift;
 977         timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
 978
 979         timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
 980         while (timekeeper.xtime_nsec >= nsecps) {
 981                 timekeeper.xtime_nsec -= nsecps;
 982                 timekeeper.xtime.tv_sec++;
 983                 second_overflow();
 984         }
 985
 986         /* Accumulate raw time */
 987         raw_nsecs = timekeeper.raw_interval << shift;
 988         raw_nsecs += timekeeper.raw_time.tv_nsec;
 989         if (raw_nsecs >= NSEC_PER_SEC) {
 990                 u64 raw_secs = raw_nsecs;
 991                 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
 992                 timekeeper.raw_time.tv_sec += raw_secs;
 993         }
 994         timekeeper.raw_time.tv_nsec = raw_nsecs;
 995
 996         /* Accumulate error between NTP and clock interval */
 997         timekeeper.ntp_error += ntp_tick_length() << shift;
 998         timekeeper.ntp_error -=
 999             (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
1000                                 (timekeeper.ntp_error_shift + shift);
1001
1002         return offset;
1003 }
1004
1005
1006 /**
1007  * update_wall_time - Uses the current clocksource to increment the wall time
1008  *
1009  * Called from the timer interrupt, must hold a write on xtime_lock.
1010  */
1011 static void update_wall_time(void)
1012 {
1013         struct clocksource *clock;
1014         cycle_t offset;
1015         int shift = 0, maxshift;
1016         unsigned long flags;
1017
1018         write_seqlock_irqsave(&timekeeper.lock, flags);
1019
1020         /* Make sure we're fully resumed: */
1021         if (unlikely(timekeeping_suspended))
1022                 goto out;
1023
1024         clock = timekeeper.clock;
1025
1026 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1027         offset = timekeeper.cycle_interval;
1028 #else
1029         offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1030 #endif
1031         timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1032                                                 timekeeper.shift;
1033
1034         /*
1035          * With NO_HZ we may have to accumulate many cycle_intervals
1036          * (think "ticks") worth of time at once. To do this efficiently,
1037          * we calculate the largest doubling multiple of cycle_intervals
1038          * that is smaller then the offset. We then accumulate that
1039          * chunk in one go, and then try to consume the next smaller
1040          * doubled multiple.
1041          */
1042         shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1043         shift = max(0, shift);
1044         /* Bound shift to one less then what overflows tick_length */
1045         maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1046         shift = min(shift, maxshift);
1047         while (offset >= timekeeper.cycle_interval) {
1048                 offset = logarithmic_accumulation(offset, shift);
1049                 if(offset < timekeeper.cycle_interval<<shift)
1050                         shift--;
1051         }
1052
1053         /* correct the clock when NTP error is too big */
1054         timekeeping_adjust(offset);
1055
1056         /*
1057          * Since in the loop above, we accumulate any amount of time
1058          * in xtime_nsec over a second into xtime.tv_sec, its possible for
1059          * xtime_nsec to be fairly small after the loop. Further, if we're
1060          * slightly speeding the clocksource up in timekeeping_adjust(),
1061          * its possible the required corrective factor to xtime_nsec could
1062          * cause it to underflow.
1063          *
1064          * Now, we cannot simply roll the accumulated second back, since
1065          * the NTP subsystem has been notified via second_overflow. So
1066          * instead we push xtime_nsec forward by the amount we underflowed,
1067          * and add that amount into the error.
1068          *
1069          * We'll correct this error next time through this function, when
1070          * xtime_nsec is not as small.
1071          */
1072         if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1073                 s64 neg = -(s64)timekeeper.xtime_nsec;
1074                 timekeeper.xtime_nsec = 0;
1075                 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1076         }
1077
1078
1079         /*
1080          * Store full nanoseconds into xtime after rounding it up and
1081          * add the remainder to the error difference.
1082          */
1083         timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1084                                                 timekeeper.shift) + 1;
1085         timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1086                                                 timekeeper.shift;
1087         timekeeper.ntp_error += timekeeper.xtime_nsec <<
1088                                 timekeeper.ntp_error_shift;
1089
1090         /*
1091          * Finally, make sure that after the rounding
1092          * xtime.tv_nsec isn't larger then NSEC_PER_SEC
1093          */
1094         if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1095                 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1096                 timekeeper.xtime.tv_sec++;
1097                 second_overflow();
1098         }
1099
1100         /* check to see if there is a new clocksource to use */
1101         update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
1102                         timekeeper.clock, timekeeper.mult);
1103
1104 out:
1105         write_sequnlock_irqrestore(&timekeeper.lock, flags);
1106
1107 }
1108
1109 /**
1110  * getboottime - Return the real time of system boot.
1111  * @ts:         pointer to the timespec to be set
1112  *
1113  * Returns the wall-time of boot in a timespec.
1114  *
1115  * This is based on the wall_to_monotonic offset and the total suspend
1116  * time. Calls to settimeofday will affect the value returned (which
1117  * basically means that however wrong your real time clock is at boot time,
1118  * you get the right time here).
1119  */
1120 void getboottime(struct timespec *ts)
1121 {
1122         struct timespec boottime = {
1123                 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1124                                 timekeeper.total_sleep_time.tv_sec,
1125                 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1126                                 timekeeper.total_sleep_time.tv_nsec
1127         };
1128
1129         set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1130 }
1131 EXPORT_SYMBOL_GPL(getboottime);
1132
1133
1134 /**
1135  * get_monotonic_boottime - Returns monotonic time since boot
1136  * @ts:         pointer to the timespec to be set
1137  *
1138  * Returns the monotonic time since boot in a timespec.
1139  *
1140  * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1141  * includes the time spent in suspend.
1142  */
1143 void get_monotonic_boottime(struct timespec *ts)
1144 {
1145         struct timespec tomono, sleep;
1146         unsigned int seq;
1147         s64 nsecs;
1148
1149         WARN_ON(timekeeping_suspended);
1150
1151         do {
1152                 seq = read_seqbegin(&timekeeper.lock);
1153                 *ts = timekeeper.xtime;
1154                 tomono = timekeeper.wall_to_monotonic;
1155                 sleep = timekeeper.total_sleep_time;
1156                 nsecs = timekeeping_get_ns();
1157
1158         } while (read_seqretry(&timekeeper.lock, seq));
1159
1160         set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1161                         ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1162 }
1163 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1164
1165 /**
1166  * ktime_get_boottime - Returns monotonic time since boot in a ktime
1167  *
1168  * Returns the monotonic time since boot in a ktime
1169  *
1170  * This is similar to CLOCK_MONTONIC/ktime_get, but also
1171  * includes the time spent in suspend.
1172  */
1173 ktime_t ktime_get_boottime(void)
1174 {
1175         struct timespec ts;
1176
1177         get_monotonic_boottime(&ts);
1178         return timespec_to_ktime(ts);
1179 }
1180 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1181
1182 /**
1183  * monotonic_to_bootbased - Convert the monotonic time to boot based.
1184  * @ts:         pointer to the timespec to be converted
1185  */
1186 void monotonic_to_bootbased(struct timespec *ts)
1187 {
1188         *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1189 }
1190 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1191
1192 unsigned long get_seconds(void)
1193 {
1194         return timekeeper.xtime.tv_sec;
1195 }
1196 EXPORT_SYMBOL(get_seconds);
1197
1198 struct timespec __current_kernel_time(void)
1199 {
1200         return timekeeper.xtime;
1201 }
1202
1203 struct timespec current_kernel_time(void)
1204 {
1205         struct timespec now;
1206         unsigned long seq;
1207
1208         do {
1209                 seq = read_seqbegin(&timekeeper.lock);
1210
1211                 now = timekeeper.xtime;
1212         } while (read_seqretry(&timekeeper.lock, seq));
1213
1214         return now;
1215 }
1216 EXPORT_SYMBOL(current_kernel_time);
1217
1218 struct timespec get_monotonic_coarse(void)
1219 {
1220         struct timespec now, mono;
1221         unsigned long seq;
1222
1223         do {
1224                 seq = read_seqbegin(&timekeeper.lock);
1225
1226                 now = timekeeper.xtime;
1227                 mono = timekeeper.wall_to_monotonic;
1228         } while (read_seqretry(&timekeeper.lock, seq));
1229
1230         set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1231                                 now.tv_nsec + mono.tv_nsec);
1232         return now;
1233 }
1234
1235 /*
1236  * The 64-bit jiffies value is not atomic - you MUST NOT read it
1237  * without sampling the sequence number in xtime_lock.
1238  * jiffies is defined in the linker script...
1239  */
1240 void do_timer(unsigned long ticks)
1241 {
1242         jiffies_64 += ticks;
1243         update_wall_time();
1244         calc_global_load(ticks);
1245 }
1246
1247 /**
1248  * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1249  *    and sleep offsets.
1250  * @xtim:       pointer to timespec to be set with xtime
1251  * @wtom:       pointer to timespec to be set with wall_to_monotonic
1252  * @sleep:      pointer to timespec to be set with time in suspend
1253  */
1254 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1255                                 struct timespec *wtom, struct timespec *sleep)
1256 {
1257         unsigned long seq;
1258
1259         do {
1260                 seq = read_seqbegin(&timekeeper.lock);
1261                 *xtim = timekeeper.xtime;
1262                 *wtom = timekeeper.wall_to_monotonic;
1263                 *sleep = timekeeper.total_sleep_time;
1264         } while (read_seqretry(&timekeeper.lock, seq));
1265 }
1266
1267 /**
1268  * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1269  */
1270 ktime_t ktime_get_monotonic_offset(void)
1271 {
1272         unsigned long seq;
1273         struct timespec wtom;
1274
1275         do {
1276                 seq = read_seqbegin(&timekeeper.lock);
1277                 wtom = timekeeper.wall_to_monotonic;
1278         } while (read_seqretry(&timekeeper.lock, seq));
1279
1280         return timespec_to_ktime(wtom);
1281 }
1282
1283 /**
1284  * xtime_update() - advances the timekeeping infrastructure
1285  * @ticks:      number of ticks, that have elapsed since the last call.
1286  *
1287  * Must be called with interrupts disabled.
1288  */
1289 void xtime_update(unsigned long ticks)
1290 {
1291         write_seqlock(&xtime_lock);
1292         do_timer(ticks);
1293         write_sequnlock(&xtime_lock);
1294 }