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 52 static struct timekeeper timekeeper; 53 54 /** 55 * timekeeper_setup_internals - Set up internals to use clocksource clock. 56 * 57 * @clock: Pointer to clocksource. 58 * 59 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment 60 * pair and interval request. 61 * 62 * Unless you're the timekeeping code, you should not be using this! 63 */ 64 static void timekeeper_setup_internals(struct clocksource *clock) 65 { 66 cycle_t interval; 67 u64 tmp, ntpinterval; 68 69 timekeeper.clock = clock; 70 clock->cycle_last = clock->read(clock); 71 72 /* Do the ns -> cycle conversion first, using original mult */ 73 tmp = NTP_INTERVAL_LENGTH; 74 tmp <<= clock->shift; 75 ntpinterval = tmp; 76 tmp += clock->mult/2; 77 do_div(tmp, clock->mult); 78 if (tmp == 0) 79 tmp = 1; 80 81 interval = (cycle_t) tmp; 82 timekeeper.cycle_interval = interval; 83 84 /* Go back from cycles -> shifted ns */ 85 timekeeper.xtime_interval = (u64) interval * clock->mult; 86 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval; 87 timekeeper.raw_interval = 88 ((u64) interval * clock->mult) >> clock->shift; 89 90 timekeeper.xtime_nsec = 0; 91 timekeeper.shift = clock->shift; 92 93 timekeeper.ntp_error = 0; 94 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift; 95 96 /* 97 * The timekeeper keeps its own mult values for the currently 98 * active clocksource. These value will be adjusted via NTP 99 * to counteract clock drifting. 100 */ 101 timekeeper.mult = clock->mult; 102 } 103 104 /* Timekeeper helper functions. */ 105 static inline s64 timekeeping_get_ns(void) 106 { 107 cycle_t cycle_now, cycle_delta; 108 struct clocksource *clock; 109 110 /* read clocksource: */ 111 clock = timekeeper.clock; 112 cycle_now = clock->read(clock); 113 114 /* calculate the delta since the last update_wall_time: */ 115 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; 116 117 /* return delta convert to nanoseconds using ntp adjusted mult. */ 118 return clocksource_cyc2ns(cycle_delta, timekeeper.mult, 119 timekeeper.shift); 120 } 121 122 static inline s64 timekeeping_get_ns_raw(void) 123 { 124 cycle_t cycle_now, cycle_delta; 125 struct clocksource *clock; 126 127 /* read clocksource: */ 128 clock = timekeeper.clock; 129 cycle_now = clock->read(clock); 130 131 /* calculate the delta since the last update_wall_time: */ 132 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; 133 134 /* return delta convert to nanoseconds. */ 135 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift); 136 } 137 138 /* 139 * This read-write spinlock protects us from races in SMP while 140 * playing with xtime. 141 */ 142 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock); 143 144 145 /* 146 * The current time 147 * wall_to_monotonic is what we need to add to xtime (or xtime corrected 148 * for sub jiffie times) to get to monotonic time. Monotonic is pegged 149 * at zero at system boot time, so wall_to_monotonic will be negative, 150 * however, we will ALWAYS keep the tv_nsec part positive so we can use 151 * the usual normalization. 152 * 153 * wall_to_monotonic is moved after resume from suspend for the monotonic 154 * time not to jump. We need to add total_sleep_time to wall_to_monotonic 155 * to get the real boot based time offset. 156 * 157 * - wall_to_monotonic is no longer the boot time, getboottime must be 158 * used instead. 159 */ 160 static struct timespec xtime __attribute__ ((aligned (16))); 161 static struct timespec wall_to_monotonic __attribute__ ((aligned (16))); 162 static struct timespec total_sleep_time; 163 164 /* 165 * The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. 166 */ 167 static struct timespec raw_time; 168 169 /* flag for if timekeeping is suspended */ 170 int __read_mostly timekeeping_suspended; 171 172 /* must hold xtime_lock */ 173 void timekeeping_leap_insert(int leapsecond) 174 { 175 xtime.tv_sec += leapsecond; 176 wall_to_monotonic.tv_sec -= leapsecond; 177 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock, 178 timekeeper.mult); 179 } 180 181 /** 182 * timekeeping_forward_now - update clock to the current time 183 * 184 * Forward the current clock to update its state since the last call to 185 * update_wall_time(). This is useful before significant clock changes, 186 * as it avoids having to deal with this time offset explicitly. 187 */ 188 static void timekeeping_forward_now(void) 189 { 190 cycle_t cycle_now, cycle_delta; 191 struct clocksource *clock; 192 s64 nsec; 193 194 clock = timekeeper.clock; 195 cycle_now = clock->read(clock); 196 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; 197 clock->cycle_last = cycle_now; 198 199 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult, 200 timekeeper.shift); 201 202 /* If arch requires, add in gettimeoffset() */ 203 nsec += arch_gettimeoffset(); 204 205 timespec_add_ns(&xtime, nsec); 206 207 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift); 208 timespec_add_ns(&raw_time, nsec); 209 } 210 211 /** 212 * getnstimeofday - Returns the time of day in a timespec 213 * @ts: pointer to the timespec to be set 214 * 215 * Returns the time of day in a timespec. 216 */ 217 void getnstimeofday(struct timespec *ts) 218 { 219 unsigned long seq; 220 s64 nsecs; 221 222 WARN_ON(timekeeping_suspended); 223 224 do { 225 seq = read_seqbegin(&xtime_lock); 226 227 *ts = xtime; 228 nsecs = timekeeping_get_ns(); 229 230 /* If arch requires, add in gettimeoffset() */ 231 nsecs += arch_gettimeoffset(); 232 233 } while (read_seqretry(&xtime_lock, seq)); 234 235 timespec_add_ns(ts, nsecs); 236 } 237 238 EXPORT_SYMBOL(getnstimeofday); 239 240 ktime_t ktime_get(void) 241 { 242 unsigned int seq; 243 s64 secs, nsecs; 244 245 WARN_ON(timekeeping_suspended); 246 247 do { 248 seq = read_seqbegin(&xtime_lock); 249 secs = xtime.tv_sec + wall_to_monotonic.tv_sec; 250 nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec; 251 nsecs += timekeeping_get_ns(); 252 /* If arch requires, add in gettimeoffset() */ 253 nsecs += arch_gettimeoffset(); 254 255 } while (read_seqretry(&xtime_lock, seq)); 256 /* 257 * Use ktime_set/ktime_add_ns to create a proper ktime on 258 * 32-bit architectures without CONFIG_KTIME_SCALAR. 259 */ 260 return ktime_add_ns(ktime_set(secs, 0), nsecs); 261 } 262 EXPORT_SYMBOL_GPL(ktime_get); 263 264 /** 265 * ktime_get_ts - get the monotonic clock in timespec format 266 * @ts: pointer to timespec variable 267 * 268 * The function calculates the monotonic clock from the realtime 269 * clock and the wall_to_monotonic offset and stores the result 270 * in normalized timespec format in the variable pointed to by @ts. 271 */ 272 void ktime_get_ts(struct timespec *ts) 273 { 274 struct timespec tomono; 275 unsigned int seq; 276 s64 nsecs; 277 278 WARN_ON(timekeeping_suspended); 279 280 do { 281 seq = read_seqbegin(&xtime_lock); 282 *ts = xtime; 283 tomono = wall_to_monotonic; 284 nsecs = timekeeping_get_ns(); 285 /* If arch requires, add in gettimeoffset() */ 286 nsecs += arch_gettimeoffset(); 287 288 } while (read_seqretry(&xtime_lock, seq)); 289 290 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec, 291 ts->tv_nsec + tomono.tv_nsec + nsecs); 292 } 293 EXPORT_SYMBOL_GPL(ktime_get_ts); 294 295 #ifdef CONFIG_NTP_PPS 296 297 /** 298 * getnstime_raw_and_real - get day and raw monotonic time in timespec format 299 * @ts_raw: pointer to the timespec to be set to raw monotonic time 300 * @ts_real: pointer to the timespec to be set to the time of day 301 * 302 * This function reads both the time of day and raw monotonic time at the 303 * same time atomically and stores the resulting timestamps in timespec 304 * format. 305 */ 306 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real) 307 { 308 unsigned long seq; 309 s64 nsecs_raw, nsecs_real; 310 311 WARN_ON_ONCE(timekeeping_suspended); 312 313 do { 314 u32 arch_offset; 315 316 seq = read_seqbegin(&xtime_lock); 317 318 *ts_raw = raw_time; 319 *ts_real = xtime; 320 321 nsecs_raw = timekeeping_get_ns_raw(); 322 nsecs_real = timekeeping_get_ns(); 323 324 /* If arch requires, add in gettimeoffset() */ 325 arch_offset = arch_gettimeoffset(); 326 nsecs_raw += arch_offset; 327 nsecs_real += arch_offset; 328 329 } while (read_seqretry(&xtime_lock, seq)); 330 331 timespec_add_ns(ts_raw, nsecs_raw); 332 timespec_add_ns(ts_real, nsecs_real); 333 } 334 EXPORT_SYMBOL(getnstime_raw_and_real); 335 336 #endif /* CONFIG_NTP_PPS */ 337 338 /** 339 * do_gettimeofday - Returns the time of day in a timeval 340 * @tv: pointer to the timeval to be set 341 * 342 * NOTE: Users should be converted to using getnstimeofday() 343 */ 344 void do_gettimeofday(struct timeval *tv) 345 { 346 struct timespec now; 347 348 getnstimeofday(&now); 349 tv->tv_sec = now.tv_sec; 350 tv->tv_usec = now.tv_nsec/1000; 351 } 352 353 EXPORT_SYMBOL(do_gettimeofday); 354 /** 355 * do_settimeofday - Sets the time of day 356 * @tv: pointer to the timespec variable containing the new time 357 * 358 * Sets the time of day to the new time and update NTP and notify hrtimers 359 */ 360 int do_settimeofday(const struct timespec *tv) 361 { 362 struct timespec ts_delta; 363 unsigned long flags; 364 365 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC) 366 return -EINVAL; 367 368 write_seqlock_irqsave(&xtime_lock, flags); 369 370 timekeeping_forward_now(); 371 372 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec; 373 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec; 374 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta); 375 376 xtime = *tv; 377 378 timekeeper.ntp_error = 0; 379 ntp_clear(); 380 381 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock, 382 timekeeper.mult); 383 384 write_sequnlock_irqrestore(&xtime_lock, flags); 385 386 /* signal hrtimers about time change */ 387 clock_was_set(); 388 389 return 0; 390 } 391 392 EXPORT_SYMBOL(do_settimeofday); 393 394 395 /** 396 * timekeeping_inject_offset - Adds or subtracts from the current time. 397 * @tv: pointer to the timespec variable containing the offset 398 * 399 * Adds or subtracts an offset value from the current time. 400 */ 401 int timekeeping_inject_offset(struct timespec *ts) 402 { 403 unsigned long flags; 404 405 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC) 406 return -EINVAL; 407 408 write_seqlock_irqsave(&xtime_lock, flags); 409 410 timekeeping_forward_now(); 411 412 xtime = timespec_add(xtime, *ts); 413 wall_to_monotonic = timespec_sub(wall_to_monotonic, *ts); 414 415 timekeeper.ntp_error = 0; 416 ntp_clear(); 417 418 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock, 419 timekeeper.mult); 420 421 write_sequnlock_irqrestore(&xtime_lock, flags); 422 423 /* signal hrtimers about time change */ 424 clock_was_set(); 425 426 return 0; 427 } 428 EXPORT_SYMBOL(timekeeping_inject_offset); 429 430 /** 431 * change_clocksource - Swaps clocksources if a new one is available 432 * 433 * Accumulates current time interval and initializes new clocksource 434 */ 435 static int change_clocksource(void *data) 436 { 437 struct clocksource *new, *old; 438 439 new = (struct clocksource *) data; 440 441 timekeeping_forward_now(); 442 if (!new->enable || new->enable(new) == 0) { 443 old = timekeeper.clock; 444 timekeeper_setup_internals(new); 445 if (old->disable) 446 old->disable(old); 447 } 448 return 0; 449 } 450 451 /** 452 * timekeeping_notify - Install a new clock source 453 * @clock: pointer to the clock source 454 * 455 * This function is called from clocksource.c after a new, better clock 456 * source has been registered. The caller holds the clocksource_mutex. 457 */ 458 void timekeeping_notify(struct clocksource *clock) 459 { 460 if (timekeeper.clock == clock) 461 return; 462 stop_machine(change_clocksource, clock, NULL); 463 tick_clock_notify(); 464 } 465 466 /** 467 * ktime_get_real - get the real (wall-) time in ktime_t format 468 * 469 * returns the time in ktime_t format 470 */ 471 ktime_t ktime_get_real(void) 472 { 473 struct timespec now; 474 475 getnstimeofday(&now); 476 477 return timespec_to_ktime(now); 478 } 479 EXPORT_SYMBOL_GPL(ktime_get_real); 480 481 /** 482 * getrawmonotonic - Returns the raw monotonic time in a timespec 483 * @ts: pointer to the timespec to be set 484 * 485 * Returns the raw monotonic time (completely un-modified by ntp) 486 */ 487 void getrawmonotonic(struct timespec *ts) 488 { 489 unsigned long seq; 490 s64 nsecs; 491 492 do { 493 seq = read_seqbegin(&xtime_lock); 494 nsecs = timekeeping_get_ns_raw(); 495 *ts = raw_time; 496 497 } while (read_seqretry(&xtime_lock, seq)); 498 499 timespec_add_ns(ts, nsecs); 500 } 501 EXPORT_SYMBOL(getrawmonotonic); 502 503 504 /** 505 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres 506 */ 507 int timekeeping_valid_for_hres(void) 508 { 509 unsigned long seq; 510 int ret; 511 512 do { 513 seq = read_seqbegin(&xtime_lock); 514 515 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES; 516 517 } while (read_seqretry(&xtime_lock, seq)); 518 519 return ret; 520 } 521 522 /** 523 * timekeeping_max_deferment - Returns max time the clocksource can be deferred 524 * 525 * Caller must observe xtime_lock via read_seqbegin/read_seqretry to 526 * ensure that the clocksource does not change! 527 */ 528 u64 timekeeping_max_deferment(void) 529 { 530 return timekeeper.clock->max_idle_ns; 531 } 532 533 /** 534 * read_persistent_clock - Return time from the persistent clock. 535 * 536 * Weak dummy function for arches that do not yet support it. 537 * Reads the time from the battery backed persistent clock. 538 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported. 539 * 540 * XXX - Do be sure to remove it once all arches implement it. 541 */ 542 void __attribute__((weak)) read_persistent_clock(struct timespec *ts) 543 { 544 ts->tv_sec = 0; 545 ts->tv_nsec = 0; 546 } 547 548 /** 549 * read_boot_clock - Return time of the system start. 550 * 551 * Weak dummy function for arches that do not yet support it. 552 * Function to read the exact time the system has been started. 553 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported. 554 * 555 * XXX - Do be sure to remove it once all arches implement it. 556 */ 557 void __attribute__((weak)) read_boot_clock(struct timespec *ts) 558 { 559 ts->tv_sec = 0; 560 ts->tv_nsec = 0; 561 } 562 563 /* 564 * timekeeping_init - Initializes the clocksource and common timekeeping values 565 */ 566 void __init timekeeping_init(void) 567 { 568 struct clocksource *clock; 569 unsigned long flags; 570 struct timespec now, boot; 571 572 read_persistent_clock(&now); 573 read_boot_clock(&boot); 574 575 write_seqlock_irqsave(&xtime_lock, flags); 576 577 ntp_init(); 578 579 clock = clocksource_default_clock(); 580 if (clock->enable) 581 clock->enable(clock); 582 timekeeper_setup_internals(clock); 583 584 xtime.tv_sec = now.tv_sec; 585 xtime.tv_nsec = now.tv_nsec; 586 raw_time.tv_sec = 0; 587 raw_time.tv_nsec = 0; 588 if (boot.tv_sec == 0 && boot.tv_nsec == 0) { 589 boot.tv_sec = xtime.tv_sec; 590 boot.tv_nsec = xtime.tv_nsec; 591 } 592 set_normalized_timespec(&wall_to_monotonic, 593 -boot.tv_sec, -boot.tv_nsec); 594 total_sleep_time.tv_sec = 0; 595 total_sleep_time.tv_nsec = 0; 596 write_sequnlock_irqrestore(&xtime_lock, flags); 597 } 598 599 /* time in seconds when suspend began */ 600 static struct timespec timekeeping_suspend_time; 601 602 /** 603 * __timekeeping_inject_sleeptime - Internal function to add sleep interval 604 * @delta: pointer to a timespec delta value 605 * 606 * Takes a timespec offset measuring a suspend interval and properly 607 * adds the sleep offset to the timekeeping variables. 608 */ 609 static void __timekeeping_inject_sleeptime(struct timespec *delta) 610 { 611 if (!timespec_valid(delta)) { 612 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid " 613 "sleep delta value!\n"); 614 return; 615 } 616 617 xtime = timespec_add(xtime, *delta); 618 wall_to_monotonic = timespec_sub(wall_to_monotonic, *delta); 619 total_sleep_time = timespec_add(total_sleep_time, *delta); 620 } 621 622 623 /** 624 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values 625 * @delta: pointer to a timespec delta value 626 * 627 * This hook is for architectures that cannot support read_persistent_clock 628 * because their RTC/persistent clock is only accessible when irqs are enabled. 629 * 630 * This function should only be called by rtc_resume(), and allows 631 * a suspend offset to be injected into the timekeeping values. 632 */ 633 void timekeeping_inject_sleeptime(struct timespec *delta) 634 { 635 unsigned long flags; 636 struct timespec ts; 637 638 /* Make sure we don't set the clock twice */ 639 read_persistent_clock(&ts); 640 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0)) 641 return; 642 643 write_seqlock_irqsave(&xtime_lock, flags); 644 timekeeping_forward_now(); 645 646 __timekeeping_inject_sleeptime(delta); 647 648 timekeeper.ntp_error = 0; 649 ntp_clear(); 650 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock, 651 timekeeper.mult); 652 653 write_sequnlock_irqrestore(&xtime_lock, flags); 654 655 /* signal hrtimers about time change */ 656 clock_was_set(); 657 } 658 659 660 /** 661 * timekeeping_resume - Resumes the generic timekeeping subsystem. 662 * 663 * This is for the generic clocksource timekeeping. 664 * xtime/wall_to_monotonic/jiffies/etc are 665 * still managed by arch specific suspend/resume code. 666 */ 667 static void timekeeping_resume(void) 668 { 669 unsigned long flags; 670 struct timespec ts; 671 672 read_persistent_clock(&ts); 673 674 clocksource_resume(); 675 676 write_seqlock_irqsave(&xtime_lock, flags); 677 678 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) { 679 ts = timespec_sub(ts, timekeeping_suspend_time); 680 __timekeeping_inject_sleeptime(&ts); 681 } 682 /* re-base the last cycle value */ 683 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock); 684 timekeeper.ntp_error = 0; 685 timekeeping_suspended = 0; 686 write_sequnlock_irqrestore(&xtime_lock, flags); 687 688 touch_softlockup_watchdog(); 689 690 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL); 691 692 /* Resume hrtimers */ 693 hrtimers_resume(); 694 } 695 696 static int timekeeping_suspend(void) 697 { 698 unsigned long flags; 699 struct timespec delta, delta_delta; 700 static struct timespec old_delta; 701 702 read_persistent_clock(&timekeeping_suspend_time); 703 704 write_seqlock_irqsave(&xtime_lock, flags); 705 timekeeping_forward_now(); 706 timekeeping_suspended = 1; 707 708 /* 709 * To avoid drift caused by repeated suspend/resumes, 710 * which each can add ~1 second drift error, 711 * try to compensate so the difference in system time 712 * and persistent_clock time stays close to constant. 713 */ 714 delta = timespec_sub(xtime, timekeeping_suspend_time); 715 delta_delta = timespec_sub(delta, old_delta); 716 if (abs(delta_delta.tv_sec) >= 2) { 717 /* 718 * if delta_delta is too large, assume time correction 719 * has occured and set old_delta to the current delta. 720 */ 721 old_delta = delta; 722 } else { 723 /* Otherwise try to adjust old_system to compensate */ 724 timekeeping_suspend_time = 725 timespec_add(timekeeping_suspend_time, delta_delta); 726 } 727 write_sequnlock_irqrestore(&xtime_lock, flags); 728 729 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL); 730 clocksource_suspend(); 731 732 return 0; 733 } 734 735 /* sysfs resume/suspend bits for timekeeping */ 736 static struct syscore_ops timekeeping_syscore_ops = { 737 .resume = timekeeping_resume, 738 .suspend = timekeeping_suspend, 739 }; 740 741 static int __init timekeeping_init_ops(void) 742 { 743 register_syscore_ops(&timekeeping_syscore_ops); 744 return 0; 745 } 746 747 device_initcall(timekeeping_init_ops); 748 749 /* 750 * If the error is already larger, we look ahead even further 751 * to compensate for late or lost adjustments. 752 */ 753 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval, 754 s64 *offset) 755 { 756 s64 tick_error, i; 757 u32 look_ahead, adj; 758 s32 error2, mult; 759 760 /* 761 * Use the current error value to determine how much to look ahead. 762 * The larger the error the slower we adjust for it to avoid problems 763 * with losing too many ticks, otherwise we would overadjust and 764 * produce an even larger error. The smaller the adjustment the 765 * faster we try to adjust for it, as lost ticks can do less harm 766 * here. This is tuned so that an error of about 1 msec is adjusted 767 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks). 768 */ 769 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ); 770 error2 = abs(error2); 771 for (look_ahead = 0; error2 > 0; look_ahead++) 772 error2 >>= 2; 773 774 /* 775 * Now calculate the error in (1 << look_ahead) ticks, but first 776 * remove the single look ahead already included in the error. 777 */ 778 tick_error = tick_length >> (timekeeper.ntp_error_shift + 1); 779 tick_error -= timekeeper.xtime_interval >> 1; 780 error = ((error - tick_error) >> look_ahead) + tick_error; 781 782 /* Finally calculate the adjustment shift value. */ 783 i = *interval; 784 mult = 1; 785 if (error < 0) { 786 error = -error; 787 *interval = -*interval; 788 *offset = -*offset; 789 mult = -1; 790 } 791 for (adj = 0; error > i; adj++) 792 error >>= 1; 793 794 *interval <<= adj; 795 *offset <<= adj; 796 return mult << adj; 797 } 798 799 /* 800 * Adjust the multiplier to reduce the error value, 801 * this is optimized for the most common adjustments of -1,0,1, 802 * for other values we can do a bit more work. 803 */ 804 static void timekeeping_adjust(s64 offset) 805 { 806 s64 error, interval = timekeeper.cycle_interval; 807 int adj; 808 809 /* 810 * The point of this is to check if the error is greater then half 811 * an interval. 812 * 813 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs. 814 * 815 * Note we subtract one in the shift, so that error is really error*2. 816 * This "saves" dividing(shifting) interval twice, but keeps the 817 * (error > interval) comparison as still measuring if error is 818 * larger then half an interval. 819 * 820 * Note: It does not "save" on aggravation when reading the code. 821 */ 822 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1); 823 if (error > interval) { 824 /* 825 * We now divide error by 4(via shift), which checks if 826 * the error is greater then twice the interval. 827 * If it is greater, we need a bigadjust, if its smaller, 828 * we can adjust by 1. 829 */ 830 error >>= 2; 831 /* 832 * XXX - In update_wall_time, we round up to the next 833 * nanosecond, and store the amount rounded up into 834 * the error. This causes the likely below to be unlikely. 835 * 836 * The proper fix is to avoid rounding up by using 837 * the high precision timekeeper.xtime_nsec instead of 838 * xtime.tv_nsec everywhere. Fixing this will take some 839 * time. 840 */ 841 if (likely(error <= interval)) 842 adj = 1; 843 else 844 adj = timekeeping_bigadjust(error, &interval, &offset); 845 } else if (error < -interval) { 846 /* See comment above, this is just switched for the negative */ 847 error >>= 2; 848 if (likely(error >= -interval)) { 849 adj = -1; 850 interval = -interval; 851 offset = -offset; 852 } else 853 adj = timekeeping_bigadjust(error, &interval, &offset); 854 } else /* No adjustment needed */ 855 return; 856 857 WARN_ONCE(timekeeper.clock->maxadj && 858 (timekeeper.mult + adj > timekeeper.clock->mult + 859 timekeeper.clock->maxadj), 860 "Adjusting %s more then 11%% (%ld vs %ld)\n", 861 timekeeper.clock->name, (long)timekeeper.mult + adj, 862 (long)timekeeper.clock->mult + 863 timekeeper.clock->maxadj); 864 /* 865 * So the following can be confusing. 866 * 867 * To keep things simple, lets assume adj == 1 for now. 868 * 869 * When adj != 1, remember that the interval and offset values 870 * have been appropriately scaled so the math is the same. 871 * 872 * The basic idea here is that we're increasing the multiplier 873 * by one, this causes the xtime_interval to be incremented by 874 * one cycle_interval. This is because: 875 * xtime_interval = cycle_interval * mult 876 * So if mult is being incremented by one: 877 * xtime_interval = cycle_interval * (mult + 1) 878 * Its the same as: 879 * xtime_interval = (cycle_interval * mult) + cycle_interval 880 * Which can be shortened to: 881 * xtime_interval += cycle_interval 882 * 883 * So offset stores the non-accumulated cycles. Thus the current 884 * time (in shifted nanoseconds) is: 885 * now = (offset * adj) + xtime_nsec 886 * Now, even though we're adjusting the clock frequency, we have 887 * to keep time consistent. In other words, we can't jump back 888 * in time, and we also want to avoid jumping forward in time. 889 * 890 * So given the same offset value, we need the time to be the same 891 * both before and after the freq adjustment. 892 * now = (offset * adj_1) + xtime_nsec_1 893 * now = (offset * adj_2) + xtime_nsec_2 894 * So: 895 * (offset * adj_1) + xtime_nsec_1 = 896 * (offset * adj_2) + xtime_nsec_2 897 * And we know: 898 * adj_2 = adj_1 + 1 899 * So: 900 * (offset * adj_1) + xtime_nsec_1 = 901 * (offset * (adj_1+1)) + xtime_nsec_2 902 * (offset * adj_1) + xtime_nsec_1 = 903 * (offset * adj_1) + offset + xtime_nsec_2 904 * Canceling the sides: 905 * xtime_nsec_1 = offset + xtime_nsec_2 906 * Which gives us: 907 * xtime_nsec_2 = xtime_nsec_1 - offset 908 * Which simplfies to: 909 * xtime_nsec -= offset 910 * 911 * XXX - TODO: Doc ntp_error calculation. 912 */ 913 timekeeper.mult += adj; 914 timekeeper.xtime_interval += interval; 915 timekeeper.xtime_nsec -= offset; 916 timekeeper.ntp_error -= (interval - offset) << 917 timekeeper.ntp_error_shift; 918 } 919 920 921 /** 922 * logarithmic_accumulation - shifted accumulation of cycles 923 * 924 * This functions accumulates a shifted interval of cycles into 925 * into a shifted interval nanoseconds. Allows for O(log) accumulation 926 * loop. 927 * 928 * Returns the unconsumed cycles. 929 */ 930 static cycle_t logarithmic_accumulation(cycle_t offset, int shift) 931 { 932 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift; 933 u64 raw_nsecs; 934 935 /* If the offset is smaller then a shifted interval, do nothing */ 936 if (offset < timekeeper.cycle_interval<<shift) 937 return offset; 938 939 /* Accumulate one shifted interval */ 940 offset -= timekeeper.cycle_interval << shift; 941 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift; 942 943 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift; 944 while (timekeeper.xtime_nsec >= nsecps) { 945 timekeeper.xtime_nsec -= nsecps; 946 xtime.tv_sec++; 947 second_overflow(); 948 } 949 950 /* Accumulate raw time */ 951 raw_nsecs = timekeeper.raw_interval << shift; 952 raw_nsecs += raw_time.tv_nsec; 953 if (raw_nsecs >= NSEC_PER_SEC) { 954 u64 raw_secs = raw_nsecs; 955 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC); 956 raw_time.tv_sec += raw_secs; 957 } 958 raw_time.tv_nsec = raw_nsecs; 959 960 /* Accumulate error between NTP and clock interval */ 961 timekeeper.ntp_error += tick_length << shift; 962 timekeeper.ntp_error -= 963 (timekeeper.xtime_interval + timekeeper.xtime_remainder) << 964 (timekeeper.ntp_error_shift + shift); 965 966 return offset; 967 } 968 969 970 /** 971 * update_wall_time - Uses the current clocksource to increment the wall time 972 * 973 * Called from the timer interrupt, must hold a write on xtime_lock. 974 */ 975 static void update_wall_time(void) 976 { 977 struct clocksource *clock; 978 cycle_t offset; 979 int shift = 0, maxshift; 980 981 /* Make sure we're fully resumed: */ 982 if (unlikely(timekeeping_suspended)) 983 return; 984 985 clock = timekeeper.clock; 986 987 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET 988 offset = timekeeper.cycle_interval; 989 #else 990 offset = (clock->read(clock) - clock->cycle_last) & clock->mask; 991 #endif 992 timekeeper.xtime_nsec = (s64)xtime.tv_nsec << timekeeper.shift; 993 994 /* 995 * With NO_HZ we may have to accumulate many cycle_intervals 996 * (think "ticks") worth of time at once. To do this efficiently, 997 * we calculate the largest doubling multiple of cycle_intervals 998 * that is smaller then the offset. We then accumulate that 999 * chunk in one go, and then try to consume the next smaller 1000 * doubled multiple. 1001 */ 1002 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval); 1003 shift = max(0, shift); 1004 /* Bound shift to one less then what overflows tick_length */ 1005 maxshift = (8*sizeof(tick_length) - (ilog2(tick_length)+1)) - 1; 1006 shift = min(shift, maxshift); 1007 while (offset >= timekeeper.cycle_interval) { 1008 offset = logarithmic_accumulation(offset, shift); 1009 if(offset < timekeeper.cycle_interval<<shift) 1010 shift--; 1011 } 1012 1013 /* correct the clock when NTP error is too big */ 1014 timekeeping_adjust(offset); 1015 1016 /* 1017 * Since in the loop above, we accumulate any amount of time 1018 * in xtime_nsec over a second into xtime.tv_sec, its possible for 1019 * xtime_nsec to be fairly small after the loop. Further, if we're 1020 * slightly speeding the clocksource up in timekeeping_adjust(), 1021 * its possible the required corrective factor to xtime_nsec could 1022 * cause it to underflow. 1023 * 1024 * Now, we cannot simply roll the accumulated second back, since 1025 * the NTP subsystem has been notified via second_overflow. So 1026 * instead we push xtime_nsec forward by the amount we underflowed, 1027 * and add that amount into the error. 1028 * 1029 * We'll correct this error next time through this function, when 1030 * xtime_nsec is not as small. 1031 */ 1032 if (unlikely((s64)timekeeper.xtime_nsec < 0)) { 1033 s64 neg = -(s64)timekeeper.xtime_nsec; 1034 timekeeper.xtime_nsec = 0; 1035 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift; 1036 } 1037 1038 1039 /* 1040 * Store full nanoseconds into xtime after rounding it up and 1041 * add the remainder to the error difference. 1042 */ 1043 xtime.tv_nsec = ((s64) timekeeper.xtime_nsec >> timekeeper.shift) + 1; 1044 timekeeper.xtime_nsec -= (s64) xtime.tv_nsec << timekeeper.shift; 1045 timekeeper.ntp_error += timekeeper.xtime_nsec << 1046 timekeeper.ntp_error_shift; 1047 1048 /* 1049 * Finally, make sure that after the rounding 1050 * xtime.tv_nsec isn't larger then NSEC_PER_SEC 1051 */ 1052 if (unlikely(xtime.tv_nsec >= NSEC_PER_SEC)) { 1053 xtime.tv_nsec -= NSEC_PER_SEC; 1054 xtime.tv_sec++; 1055 second_overflow(); 1056 } 1057 1058 /* check to see if there is a new clocksource to use */ 1059 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock, 1060 timekeeper.mult); 1061 } 1062 1063 /** 1064 * getboottime - Return the real time of system boot. 1065 * @ts: pointer to the timespec to be set 1066 * 1067 * Returns the wall-time of boot in a timespec. 1068 * 1069 * This is based on the wall_to_monotonic offset and the total suspend 1070 * time. Calls to settimeofday will affect the value returned (which 1071 * basically means that however wrong your real time clock is at boot time, 1072 * you get the right time here). 1073 */ 1074 void getboottime(struct timespec *ts) 1075 { 1076 struct timespec boottime = { 1077 .tv_sec = wall_to_monotonic.tv_sec + total_sleep_time.tv_sec, 1078 .tv_nsec = wall_to_monotonic.tv_nsec + total_sleep_time.tv_nsec 1079 }; 1080 1081 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec); 1082 } 1083 EXPORT_SYMBOL_GPL(getboottime); 1084 1085 1086 /** 1087 * get_monotonic_boottime - Returns monotonic time since boot 1088 * @ts: pointer to the timespec to be set 1089 * 1090 * Returns the monotonic time since boot in a timespec. 1091 * 1092 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also 1093 * includes the time spent in suspend. 1094 */ 1095 void get_monotonic_boottime(struct timespec *ts) 1096 { 1097 struct timespec tomono, sleep; 1098 unsigned int seq; 1099 s64 nsecs; 1100 1101 WARN_ON(timekeeping_suspended); 1102 1103 do { 1104 seq = read_seqbegin(&xtime_lock); 1105 *ts = xtime; 1106 tomono = wall_to_monotonic; 1107 sleep = total_sleep_time; 1108 nsecs = timekeeping_get_ns(); 1109 1110 } while (read_seqretry(&xtime_lock, seq)); 1111 1112 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec, 1113 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs); 1114 } 1115 EXPORT_SYMBOL_GPL(get_monotonic_boottime); 1116 1117 /** 1118 * ktime_get_boottime - Returns monotonic time since boot in a ktime 1119 * 1120 * Returns the monotonic time since boot in a ktime 1121 * 1122 * This is similar to CLOCK_MONTONIC/ktime_get, but also 1123 * includes the time spent in suspend. 1124 */ 1125 ktime_t ktime_get_boottime(void) 1126 { 1127 struct timespec ts; 1128 1129 get_monotonic_boottime(&ts); 1130 return timespec_to_ktime(ts); 1131 } 1132 EXPORT_SYMBOL_GPL(ktime_get_boottime); 1133 1134 /** 1135 * monotonic_to_bootbased - Convert the monotonic time to boot based. 1136 * @ts: pointer to the timespec to be converted 1137 */ 1138 void monotonic_to_bootbased(struct timespec *ts) 1139 { 1140 *ts = timespec_add(*ts, total_sleep_time); 1141 } 1142 EXPORT_SYMBOL_GPL(monotonic_to_bootbased); 1143 1144 unsigned long get_seconds(void) 1145 { 1146 return xtime.tv_sec; 1147 } 1148 EXPORT_SYMBOL(get_seconds); 1149 1150 struct timespec __current_kernel_time(void) 1151 { 1152 return xtime; 1153 } 1154 1155 struct timespec current_kernel_time(void) 1156 { 1157 struct timespec now; 1158 unsigned long seq; 1159 1160 do { 1161 seq = read_seqbegin(&xtime_lock); 1162 1163 now = xtime; 1164 } while (read_seqretry(&xtime_lock, seq)); 1165 1166 return now; 1167 } 1168 EXPORT_SYMBOL(current_kernel_time); 1169 1170 struct timespec get_monotonic_coarse(void) 1171 { 1172 struct timespec now, mono; 1173 unsigned long seq; 1174 1175 do { 1176 seq = read_seqbegin(&xtime_lock); 1177 1178 now = xtime; 1179 mono = wall_to_monotonic; 1180 } while (read_seqretry(&xtime_lock, seq)); 1181 1182 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec, 1183 now.tv_nsec + mono.tv_nsec); 1184 return now; 1185 } 1186 1187 /* 1188 * The 64-bit jiffies value is not atomic - you MUST NOT read it 1189 * without sampling the sequence number in xtime_lock. 1190 * jiffies is defined in the linker script... 1191 */ 1192 void do_timer(unsigned long ticks) 1193 { 1194 jiffies_64 += ticks; 1195 update_wall_time(); 1196 calc_global_load(ticks); 1197 } 1198 1199 /** 1200 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic, 1201 * and sleep offsets. 1202 * @xtim: pointer to timespec to be set with xtime 1203 * @wtom: pointer to timespec to be set with wall_to_monotonic 1204 * @sleep: pointer to timespec to be set with time in suspend 1205 */ 1206 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim, 1207 struct timespec *wtom, struct timespec *sleep) 1208 { 1209 unsigned long seq; 1210 1211 do { 1212 seq = read_seqbegin(&xtime_lock); 1213 *xtim = xtime; 1214 *wtom = wall_to_monotonic; 1215 *sleep = total_sleep_time; 1216 } while (read_seqretry(&xtime_lock, seq)); 1217 } 1218 1219 /** 1220 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format 1221 */ 1222 ktime_t ktime_get_monotonic_offset(void) 1223 { 1224 unsigned long seq; 1225 struct timespec wtom; 1226 1227 do { 1228 seq = read_seqbegin(&xtime_lock); 1229 wtom = wall_to_monotonic; 1230 } while (read_seqretry(&xtime_lock, seq)); 1231 return timespec_to_ktime(wtom); 1232 } 1233 1234 /** 1235 * xtime_update() - advances the timekeeping infrastructure 1236 * @ticks: number of ticks, that have elapsed since the last call. 1237 * 1238 * Must be called with interrupts disabled. 1239 */ 1240 void xtime_update(unsigned long ticks) 1241 { 1242 write_seqlock(&xtime_lock); 1243 do_timer(ticks); 1244 write_sequnlock(&xtime_lock); 1245 } 1246