2 * linux/kernel/time/timekeeping.c
4 * Kernel timekeeping code and accessor functions
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
11 #include <linux/module.h>
12 #include <linux/interrupt.h>
13 #include <linux/percpu.h>
14 #include <linux/init.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>
24 /* Structure holding internal timekeeping values. */
26 /* Current clocksource used for timekeeping. */
27 struct clocksource *clock;
28 /* NTP adjusted clock multiplier */
30 /* The shift value of the current clocksource. */
33 /* Number of clock cycles in one NTP interval. */
34 cycle_t cycle_interval;
35 /* Number of clock shifted nano seconds in one NTP interval. */
37 /* shifted nano seconds left over when rounding cycle_interval */
39 /* Raw nano seconds accumulated per NTP interval. */
42 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
44 /* Difference between accumulated time and NTP time in ntp
45 * shifted nano seconds. */
47 /* Shift conversion between clock shifted nano seconds and
48 * ntp shifted nano seconds. */
51 /* The current time */
52 struct timespec xtime;
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.
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.
64 * - wall_to_monotonic is no longer the boot time, getboottime must be
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;
73 /* Seqlock for all timekeeper values */
77 static struct timekeeper timekeeper;
80 * This read-write spinlock protects us from races in SMP while
83 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
86 /* flag for if timekeeping is suspended */
87 int __read_mostly timekeeping_suspended;
92 * timekeeper_setup_internals - Set up internals to use clocksource clock.
94 * @clock: Pointer to clocksource.
96 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
97 * pair and interval request.
99 * Unless you're the timekeeping code, you should not be using this!
101 static void timekeeper_setup_internals(struct clocksource *clock)
104 u64 tmp, ntpinterval;
106 timekeeper.clock = clock;
107 clock->cycle_last = clock->read(clock);
109 /* Do the ns -> cycle conversion first, using original mult */
110 tmp = NTP_INTERVAL_LENGTH;
111 tmp <<= clock->shift;
113 tmp += clock->mult/2;
114 do_div(tmp, clock->mult);
118 interval = (cycle_t) tmp;
119 timekeeper.cycle_interval = interval;
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;
127 timekeeper.xtime_nsec = 0;
128 timekeeper.shift = clock->shift;
130 timekeeper.ntp_error = 0;
131 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
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.
138 timekeeper.mult = clock->mult;
141 /* Timekeeper helper functions. */
142 static inline s64 timekeeping_get_ns(void)
144 cycle_t cycle_now, cycle_delta;
145 struct clocksource *clock;
147 /* read clocksource: */
148 clock = timekeeper.clock;
149 cycle_now = clock->read(clock);
151 /* calculate the delta since the last update_wall_time: */
152 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
154 /* return delta convert to nanoseconds using ntp adjusted mult. */
155 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
159 static inline s64 timekeeping_get_ns_raw(void)
161 cycle_t cycle_now, cycle_delta;
162 struct clocksource *clock;
164 /* read clocksource: */
165 clock = timekeeper.clock;
166 cycle_now = clock->read(clock);
168 /* calculate the delta since the last update_wall_time: */
169 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
171 /* return delta convert to nanoseconds. */
172 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
175 /* must hold write on timekeeper.lock */
176 static void timekeeping_update(bool clearntp)
179 timekeeper.ntp_error = 0;
182 update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
183 timekeeper.clock, timekeeper.mult);
188 * timekeeping_forward_now - update clock to the current time
190 * Forward the current clock to update its state since the last call to
191 * update_wall_time(). This is useful before significant clock changes,
192 * as it avoids having to deal with this time offset explicitly.
194 static void timekeeping_forward_now(void)
196 cycle_t cycle_now, cycle_delta;
197 struct clocksource *clock;
200 clock = timekeeper.clock;
201 cycle_now = clock->read(clock);
202 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
203 clock->cycle_last = cycle_now;
205 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
208 /* If arch requires, add in gettimeoffset() */
209 nsec += arch_gettimeoffset();
211 timespec_add_ns(&timekeeper.xtime, nsec);
213 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
214 timespec_add_ns(&timekeeper.raw_time, nsec);
218 * getnstimeofday - Returns the time of day in a timespec
219 * @ts: pointer to the timespec to be set
221 * Returns the time of day in a timespec.
223 void getnstimeofday(struct timespec *ts)
228 WARN_ON(timekeeping_suspended);
231 seq = read_seqbegin(&timekeeper.lock);
233 *ts = timekeeper.xtime;
234 nsecs = timekeeping_get_ns();
236 /* If arch requires, add in gettimeoffset() */
237 nsecs += arch_gettimeoffset();
239 } while (read_seqretry(&timekeeper.lock, seq));
241 timespec_add_ns(ts, nsecs);
243 EXPORT_SYMBOL(getnstimeofday);
245 ktime_t ktime_get(void)
250 WARN_ON(timekeeping_suspended);
253 seq = read_seqbegin(&timekeeper.lock);
254 secs = timekeeper.xtime.tv_sec +
255 timekeeper.wall_to_monotonic.tv_sec;
256 nsecs = timekeeper.xtime.tv_nsec +
257 timekeeper.wall_to_monotonic.tv_nsec;
258 nsecs += timekeeping_get_ns();
259 /* If arch requires, add in gettimeoffset() */
260 nsecs += arch_gettimeoffset();
262 } while (read_seqretry(&timekeeper.lock, seq));
264 * Use ktime_set/ktime_add_ns to create a proper ktime on
265 * 32-bit architectures without CONFIG_KTIME_SCALAR.
267 return ktime_add_ns(ktime_set(secs, 0), nsecs);
269 EXPORT_SYMBOL_GPL(ktime_get);
272 * ktime_get_ts - get the monotonic clock in timespec format
273 * @ts: pointer to timespec variable
275 * The function calculates the monotonic clock from the realtime
276 * clock and the wall_to_monotonic offset and stores the result
277 * in normalized timespec format in the variable pointed to by @ts.
279 void ktime_get_ts(struct timespec *ts)
281 struct timespec tomono;
285 WARN_ON(timekeeping_suspended);
288 seq = read_seqbegin(&timekeeper.lock);
289 *ts = timekeeper.xtime;
290 tomono = timekeeper.wall_to_monotonic;
291 nsecs = timekeeping_get_ns();
292 /* If arch requires, add in gettimeoffset() */
293 nsecs += arch_gettimeoffset();
295 } while (read_seqretry(&timekeeper.lock, seq));
297 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
298 ts->tv_nsec + tomono.tv_nsec + nsecs);
300 EXPORT_SYMBOL_GPL(ktime_get_ts);
302 #ifdef CONFIG_NTP_PPS
305 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
306 * @ts_raw: pointer to the timespec to be set to raw monotonic time
307 * @ts_real: pointer to the timespec to be set to the time of day
309 * This function reads both the time of day and raw monotonic time at the
310 * same time atomically and stores the resulting timestamps in timespec
313 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
316 s64 nsecs_raw, nsecs_real;
318 WARN_ON_ONCE(timekeeping_suspended);
323 seq = read_seqbegin(&timekeeper.lock);
325 *ts_raw = timekeeper.raw_time;
326 *ts_real = timekeeper.xtime;
328 nsecs_raw = timekeeping_get_ns_raw();
329 nsecs_real = timekeeping_get_ns();
331 /* If arch requires, add in gettimeoffset() */
332 arch_offset = arch_gettimeoffset();
333 nsecs_raw += arch_offset;
334 nsecs_real += arch_offset;
336 } while (read_seqretry(&timekeeper.lock, seq));
338 timespec_add_ns(ts_raw, nsecs_raw);
339 timespec_add_ns(ts_real, nsecs_real);
341 EXPORT_SYMBOL(getnstime_raw_and_real);
343 #endif /* CONFIG_NTP_PPS */
346 * do_gettimeofday - Returns the time of day in a timeval
347 * @tv: pointer to the timeval to be set
349 * NOTE: Users should be converted to using getnstimeofday()
351 void do_gettimeofday(struct timeval *tv)
355 getnstimeofday(&now);
356 tv->tv_sec = now.tv_sec;
357 tv->tv_usec = now.tv_nsec/1000;
359 EXPORT_SYMBOL(do_gettimeofday);
362 * do_settimeofday - Sets the time of day
363 * @tv: pointer to the timespec variable containing the new time
365 * Sets the time of day to the new time and update NTP and notify hrtimers
367 int do_settimeofday(const struct timespec *tv)
369 struct timespec ts_delta;
372 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
375 write_seqlock_irqsave(&timekeeper.lock, flags);
377 timekeeping_forward_now();
379 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
380 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
381 timekeeper.wall_to_monotonic =
382 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
384 timekeeper.xtime = *tv;
385 timekeeping_update(true);
387 write_sequnlock_irqrestore(&timekeeper.lock, flags);
389 /* signal hrtimers about time change */
394 EXPORT_SYMBOL(do_settimeofday);
398 * timekeeping_inject_offset - Adds or subtracts from the current time.
399 * @tv: pointer to the timespec variable containing the offset
401 * Adds or subtracts an offset value from the current time.
403 int timekeeping_inject_offset(struct timespec *ts)
407 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
410 write_seqlock_irqsave(&timekeeper.lock, flags);
412 timekeeping_forward_now();
414 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
415 timekeeper.wall_to_monotonic =
416 timespec_sub(timekeeper.wall_to_monotonic, *ts);
418 timekeeping_update(true);
420 write_sequnlock_irqrestore(&timekeeper.lock, flags);
422 /* signal hrtimers about time change */
427 EXPORT_SYMBOL(timekeeping_inject_offset);
430 * change_clocksource - Swaps clocksources if a new one is available
432 * Accumulates current time interval and initializes new clocksource
434 static int change_clocksource(void *data)
436 struct clocksource *new, *old;
439 new = (struct clocksource *) data;
441 write_seqlock_irqsave(&timekeeper.lock, flags);
443 timekeeping_forward_now();
444 if (!new->enable || new->enable(new) == 0) {
445 old = timekeeper.clock;
446 timekeeper_setup_internals(new);
450 timekeeping_update(true);
452 write_sequnlock_irqrestore(&timekeeper.lock, flags);
458 * timekeeping_notify - Install a new clock source
459 * @clock: pointer to the clock source
461 * This function is called from clocksource.c after a new, better clock
462 * source has been registered. The caller holds the clocksource_mutex.
464 void timekeeping_notify(struct clocksource *clock)
466 if (timekeeper.clock == clock)
468 stop_machine(change_clocksource, clock, NULL);
473 * ktime_get_real - get the real (wall-) time in ktime_t format
475 * returns the time in ktime_t format
477 ktime_t ktime_get_real(void)
481 getnstimeofday(&now);
483 return timespec_to_ktime(now);
485 EXPORT_SYMBOL_GPL(ktime_get_real);
488 * getrawmonotonic - Returns the raw monotonic time in a timespec
489 * @ts: pointer to the timespec to be set
491 * Returns the raw monotonic time (completely un-modified by ntp)
493 void getrawmonotonic(struct timespec *ts)
499 seq = read_seqbegin(&timekeeper.lock);
500 nsecs = timekeeping_get_ns_raw();
501 *ts = timekeeper.raw_time;
503 } while (read_seqretry(&timekeeper.lock, seq));
505 timespec_add_ns(ts, nsecs);
507 EXPORT_SYMBOL(getrawmonotonic);
511 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
513 int timekeeping_valid_for_hres(void)
519 seq = read_seqbegin(&timekeeper.lock);
521 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
523 } while (read_seqretry(&timekeeper.lock, seq));
529 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
531 u64 timekeeping_max_deferment(void)
536 seq = read_seqbegin(&timekeeper.lock);
538 ret = timekeeper.clock->max_idle_ns;
540 } while (read_seqretry(&timekeeper.lock, seq));
546 * read_persistent_clock - Return time from the persistent clock.
548 * Weak dummy function for arches that do not yet support it.
549 * Reads the time from the battery backed persistent clock.
550 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
552 * XXX - Do be sure to remove it once all arches implement it.
554 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
561 * read_boot_clock - Return time of the system start.
563 * Weak dummy function for arches that do not yet support it.
564 * Function to read the exact time the system has been started.
565 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
567 * XXX - Do be sure to remove it once all arches implement it.
569 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
576 * timekeeping_init - Initializes the clocksource and common timekeeping values
578 void __init timekeeping_init(void)
580 struct clocksource *clock;
582 struct timespec now, boot;
584 read_persistent_clock(&now);
585 read_boot_clock(&boot);
587 seqlock_init(&timekeeper.lock);
591 write_seqlock_irqsave(&timekeeper.lock, flags);
592 clock = clocksource_default_clock();
594 clock->enable(clock);
595 timekeeper_setup_internals(clock);
597 timekeeper.xtime.tv_sec = now.tv_sec;
598 timekeeper.xtime.tv_nsec = now.tv_nsec;
599 timekeeper.raw_time.tv_sec = 0;
600 timekeeper.raw_time.tv_nsec = 0;
601 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
602 boot.tv_sec = timekeeper.xtime.tv_sec;
603 boot.tv_nsec = timekeeper.xtime.tv_nsec;
605 set_normalized_timespec(&timekeeper.wall_to_monotonic,
606 -boot.tv_sec, -boot.tv_nsec);
607 timekeeper.total_sleep_time.tv_sec = 0;
608 timekeeper.total_sleep_time.tv_nsec = 0;
609 write_sequnlock_irqrestore(&timekeeper.lock, flags);
612 /* time in seconds when suspend began */
613 static struct timespec timekeeping_suspend_time;
616 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
617 * @delta: pointer to a timespec delta value
619 * Takes a timespec offset measuring a suspend interval and properly
620 * adds the sleep offset to the timekeeping variables.
622 static void __timekeeping_inject_sleeptime(struct timespec *delta)
624 if (!timespec_valid(delta)) {
625 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
626 "sleep delta value!\n");
630 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
631 timekeeper.wall_to_monotonic =
632 timespec_sub(timekeeper.wall_to_monotonic, *delta);
633 timekeeper.total_sleep_time = timespec_add(
634 timekeeper.total_sleep_time, *delta);
639 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
640 * @delta: pointer to a timespec delta value
642 * This hook is for architectures that cannot support read_persistent_clock
643 * because their RTC/persistent clock is only accessible when irqs are enabled.
645 * This function should only be called by rtc_resume(), and allows
646 * a suspend offset to be injected into the timekeeping values.
648 void timekeeping_inject_sleeptime(struct timespec *delta)
653 /* Make sure we don't set the clock twice */
654 read_persistent_clock(&ts);
655 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
658 write_seqlock_irqsave(&timekeeper.lock, flags);
660 timekeeping_forward_now();
662 __timekeeping_inject_sleeptime(delta);
664 timekeeping_update(true);
666 write_sequnlock_irqrestore(&timekeeper.lock, flags);
668 /* signal hrtimers about time change */
674 * timekeeping_resume - Resumes the generic timekeeping subsystem.
676 * This is for the generic clocksource timekeeping.
677 * xtime/wall_to_monotonic/jiffies/etc are
678 * still managed by arch specific suspend/resume code.
680 static void timekeeping_resume(void)
685 read_persistent_clock(&ts);
687 clocksource_resume();
689 write_seqlock_irqsave(&timekeeper.lock, flags);
691 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
692 ts = timespec_sub(ts, timekeeping_suspend_time);
693 __timekeeping_inject_sleeptime(&ts);
695 /* re-base the last cycle value */
696 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
697 timekeeper.ntp_error = 0;
698 timekeeping_suspended = 0;
699 write_sequnlock_irqrestore(&timekeeper.lock, flags);
701 touch_softlockup_watchdog();
703 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
705 /* Resume hrtimers */
709 static int timekeeping_suspend(void)
712 struct timespec delta, delta_delta;
713 static struct timespec old_delta;
715 read_persistent_clock(&timekeeping_suspend_time);
717 write_seqlock_irqsave(&timekeeper.lock, flags);
718 timekeeping_forward_now();
719 timekeeping_suspended = 1;
722 * To avoid drift caused by repeated suspend/resumes,
723 * which each can add ~1 second drift error,
724 * try to compensate so the difference in system time
725 * and persistent_clock time stays close to constant.
727 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
728 delta_delta = timespec_sub(delta, old_delta);
729 if (abs(delta_delta.tv_sec) >= 2) {
731 * if delta_delta is too large, assume time correction
732 * has occured and set old_delta to the current delta.
736 /* Otherwise try to adjust old_system to compensate */
737 timekeeping_suspend_time =
738 timespec_add(timekeeping_suspend_time, delta_delta);
740 write_sequnlock_irqrestore(&timekeeper.lock, flags);
742 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
743 clocksource_suspend();
748 /* sysfs resume/suspend bits for timekeeping */
749 static struct syscore_ops timekeeping_syscore_ops = {
750 .resume = timekeeping_resume,
751 .suspend = timekeeping_suspend,
754 static int __init timekeeping_init_ops(void)
756 register_syscore_ops(&timekeeping_syscore_ops);
760 device_initcall(timekeeping_init_ops);
763 * If the error is already larger, we look ahead even further
764 * to compensate for late or lost adjustments.
766 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
774 * Use the current error value to determine how much to look ahead.
775 * The larger the error the slower we adjust for it to avoid problems
776 * with losing too many ticks, otherwise we would overadjust and
777 * produce an even larger error. The smaller the adjustment the
778 * faster we try to adjust for it, as lost ticks can do less harm
779 * here. This is tuned so that an error of about 1 msec is adjusted
780 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
782 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
783 error2 = abs(error2);
784 for (look_ahead = 0; error2 > 0; look_ahead++)
788 * Now calculate the error in (1 << look_ahead) ticks, but first
789 * remove the single look ahead already included in the error.
791 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
792 tick_error -= timekeeper.xtime_interval >> 1;
793 error = ((error - tick_error) >> look_ahead) + tick_error;
795 /* Finally calculate the adjustment shift value. */
800 *interval = -*interval;
804 for (adj = 0; error > i; adj++)
813 * Adjust the multiplier to reduce the error value,
814 * this is optimized for the most common adjustments of -1,0,1,
815 * for other values we can do a bit more work.
817 static void timekeeping_adjust(s64 offset)
819 s64 error, interval = timekeeper.cycle_interval;
823 * The point of this is to check if the error is greater than half
826 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
828 * Note we subtract one in the shift, so that error is really error*2.
829 * This "saves" dividing(shifting) interval twice, but keeps the
830 * (error > interval) comparison as still measuring if error is
831 * larger than half an interval.
833 * Note: It does not "save" on aggravation when reading the code.
835 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
836 if (error > interval) {
838 * We now divide error by 4(via shift), which checks if
839 * the error is greater than twice the interval.
840 * If it is greater, we need a bigadjust, if its smaller,
841 * we can adjust by 1.
845 * XXX - In update_wall_time, we round up to the next
846 * nanosecond, and store the amount rounded up into
847 * the error. This causes the likely below to be unlikely.
849 * The proper fix is to avoid rounding up by using
850 * the high precision timekeeper.xtime_nsec instead of
851 * xtime.tv_nsec everywhere. Fixing this will take some
854 if (likely(error <= interval))
857 adj = timekeeping_bigadjust(error, &interval, &offset);
858 } else if (error < -interval) {
859 /* See comment above, this is just switched for the negative */
861 if (likely(error >= -interval)) {
863 interval = -interval;
866 adj = timekeeping_bigadjust(error, &interval, &offset);
867 } else /* No adjustment needed */
870 if (unlikely(timekeeper.clock->maxadj &&
871 (timekeeper.mult + adj >
872 timekeeper.clock->mult + timekeeper.clock->maxadj))) {
873 printk_once(KERN_WARNING
874 "Adjusting %s more than 11%% (%ld vs %ld)\n",
875 timekeeper.clock->name, (long)timekeeper.mult + adj,
876 (long)timekeeper.clock->mult +
877 timekeeper.clock->maxadj);
880 * So the following can be confusing.
882 * To keep things simple, lets assume adj == 1 for now.
884 * When adj != 1, remember that the interval and offset values
885 * have been appropriately scaled so the math is the same.
887 * The basic idea here is that we're increasing the multiplier
888 * by one, this causes the xtime_interval to be incremented by
889 * one cycle_interval. This is because:
890 * xtime_interval = cycle_interval * mult
891 * So if mult is being incremented by one:
892 * xtime_interval = cycle_interval * (mult + 1)
894 * xtime_interval = (cycle_interval * mult) + cycle_interval
895 * Which can be shortened to:
896 * xtime_interval += cycle_interval
898 * So offset stores the non-accumulated cycles. Thus the current
899 * time (in shifted nanoseconds) is:
900 * now = (offset * adj) + xtime_nsec
901 * Now, even though we're adjusting the clock frequency, we have
902 * to keep time consistent. In other words, we can't jump back
903 * in time, and we also want to avoid jumping forward in time.
905 * So given the same offset value, we need the time to be the same
906 * both before and after the freq adjustment.
907 * now = (offset * adj_1) + xtime_nsec_1
908 * now = (offset * adj_2) + xtime_nsec_2
910 * (offset * adj_1) + xtime_nsec_1 =
911 * (offset * adj_2) + xtime_nsec_2
915 * (offset * adj_1) + xtime_nsec_1 =
916 * (offset * (adj_1+1)) + xtime_nsec_2
917 * (offset * adj_1) + xtime_nsec_1 =
918 * (offset * adj_1) + offset + xtime_nsec_2
919 * Canceling the sides:
920 * xtime_nsec_1 = offset + xtime_nsec_2
922 * xtime_nsec_2 = xtime_nsec_1 - offset
923 * Which simplfies to:
924 * xtime_nsec -= offset
926 * XXX - TODO: Doc ntp_error calculation.
928 timekeeper.mult += adj;
929 timekeeper.xtime_interval += interval;
930 timekeeper.xtime_nsec -= offset;
931 timekeeper.ntp_error -= (interval - offset) <<
932 timekeeper.ntp_error_shift;
937 * logarithmic_accumulation - shifted accumulation of cycles
939 * This functions accumulates a shifted interval of cycles into
940 * into a shifted interval nanoseconds. Allows for O(log) accumulation
943 * Returns the unconsumed cycles.
945 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
947 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
950 /* If the offset is smaller than a shifted interval, do nothing */
951 if (offset < timekeeper.cycle_interval<<shift)
954 /* Accumulate one shifted interval */
955 offset -= timekeeper.cycle_interval << shift;
956 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
958 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
959 while (timekeeper.xtime_nsec >= nsecps) {
961 timekeeper.xtime_nsec -= nsecps;
962 timekeeper.xtime.tv_sec++;
963 leap = second_overflow(timekeeper.xtime.tv_sec);
964 timekeeper.xtime.tv_sec += leap;
967 /* Accumulate raw time */
968 raw_nsecs = timekeeper.raw_interval << shift;
969 raw_nsecs += timekeeper.raw_time.tv_nsec;
970 if (raw_nsecs >= NSEC_PER_SEC) {
971 u64 raw_secs = raw_nsecs;
972 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
973 timekeeper.raw_time.tv_sec += raw_secs;
975 timekeeper.raw_time.tv_nsec = raw_nsecs;
977 /* Accumulate error between NTP and clock interval */
978 timekeeper.ntp_error += ntp_tick_length() << shift;
979 timekeeper.ntp_error -=
980 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
981 (timekeeper.ntp_error_shift + shift);
988 * update_wall_time - Uses the current clocksource to increment the wall time
991 static void update_wall_time(void)
993 struct clocksource *clock;
995 int shift = 0, maxshift;
998 write_seqlock_irqsave(&timekeeper.lock, flags);
1000 /* Make sure we're fully resumed: */
1001 if (unlikely(timekeeping_suspended))
1004 clock = timekeeper.clock;
1006 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1007 offset = timekeeper.cycle_interval;
1009 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1011 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1015 * With NO_HZ we may have to accumulate many cycle_intervals
1016 * (think "ticks") worth of time at once. To do this efficiently,
1017 * we calculate the largest doubling multiple of cycle_intervals
1018 * that is smaller than the offset. We then accumulate that
1019 * chunk in one go, and then try to consume the next smaller
1022 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1023 shift = max(0, shift);
1024 /* Bound shift to one less than what overflows tick_length */
1025 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1026 shift = min(shift, maxshift);
1027 while (offset >= timekeeper.cycle_interval) {
1028 offset = logarithmic_accumulation(offset, shift);
1029 if(offset < timekeeper.cycle_interval<<shift)
1033 /* correct the clock when NTP error is too big */
1034 timekeeping_adjust(offset);
1037 * Since in the loop above, we accumulate any amount of time
1038 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1039 * xtime_nsec to be fairly small after the loop. Further, if we're
1040 * slightly speeding the clocksource up in timekeeping_adjust(),
1041 * its possible the required corrective factor to xtime_nsec could
1042 * cause it to underflow.
1044 * Now, we cannot simply roll the accumulated second back, since
1045 * the NTP subsystem has been notified via second_overflow. So
1046 * instead we push xtime_nsec forward by the amount we underflowed,
1047 * and add that amount into the error.
1049 * We'll correct this error next time through this function, when
1050 * xtime_nsec is not as small.
1052 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1053 s64 neg = -(s64)timekeeper.xtime_nsec;
1054 timekeeper.xtime_nsec = 0;
1055 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1060 * Store full nanoseconds into xtime after rounding it up and
1061 * add the remainder to the error difference.
1063 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1064 timekeeper.shift) + 1;
1065 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1067 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1068 timekeeper.ntp_error_shift;
1071 * Finally, make sure that after the rounding
1072 * xtime.tv_nsec isn't larger than NSEC_PER_SEC
1074 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1076 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1077 timekeeper.xtime.tv_sec++;
1078 leap = second_overflow(timekeeper.xtime.tv_sec);
1079 timekeeper.xtime.tv_sec += leap;
1082 timekeeping_update(false);
1085 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1090 * getboottime - Return the real time of system boot.
1091 * @ts: pointer to the timespec to be set
1093 * Returns the wall-time of boot in a timespec.
1095 * This is based on the wall_to_monotonic offset and the total suspend
1096 * time. Calls to settimeofday will affect the value returned (which
1097 * basically means that however wrong your real time clock is at boot time,
1098 * you get the right time here).
1100 void getboottime(struct timespec *ts)
1102 struct timespec boottime = {
1103 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1104 timekeeper.total_sleep_time.tv_sec,
1105 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1106 timekeeper.total_sleep_time.tv_nsec
1109 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1111 EXPORT_SYMBOL_GPL(getboottime);
1115 * get_monotonic_boottime - Returns monotonic time since boot
1116 * @ts: pointer to the timespec to be set
1118 * Returns the monotonic time since boot in a timespec.
1120 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1121 * includes the time spent in suspend.
1123 void get_monotonic_boottime(struct timespec *ts)
1125 struct timespec tomono, sleep;
1129 WARN_ON(timekeeping_suspended);
1132 seq = read_seqbegin(&timekeeper.lock);
1133 *ts = timekeeper.xtime;
1134 tomono = timekeeper.wall_to_monotonic;
1135 sleep = timekeeper.total_sleep_time;
1136 nsecs = timekeeping_get_ns();
1138 } while (read_seqretry(&timekeeper.lock, seq));
1140 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1141 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1143 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1146 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1148 * Returns the monotonic time since boot in a ktime
1150 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1151 * includes the time spent in suspend.
1153 ktime_t ktime_get_boottime(void)
1157 get_monotonic_boottime(&ts);
1158 return timespec_to_ktime(ts);
1160 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1163 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1164 * @ts: pointer to the timespec to be converted
1166 void monotonic_to_bootbased(struct timespec *ts)
1168 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1170 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1172 unsigned long get_seconds(void)
1174 return timekeeper.xtime.tv_sec;
1176 EXPORT_SYMBOL(get_seconds);
1178 struct timespec __current_kernel_time(void)
1180 return timekeeper.xtime;
1183 struct timespec current_kernel_time(void)
1185 struct timespec now;
1189 seq = read_seqbegin(&timekeeper.lock);
1191 now = timekeeper.xtime;
1192 } while (read_seqretry(&timekeeper.lock, seq));
1196 EXPORT_SYMBOL(current_kernel_time);
1198 struct timespec get_monotonic_coarse(void)
1200 struct timespec now, mono;
1204 seq = read_seqbegin(&timekeeper.lock);
1206 now = timekeeper.xtime;
1207 mono = timekeeper.wall_to_monotonic;
1208 } while (read_seqretry(&timekeeper.lock, seq));
1210 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1211 now.tv_nsec + mono.tv_nsec);
1216 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1217 * without sampling the sequence number in xtime_lock.
1218 * jiffies is defined in the linker script...
1220 void do_timer(unsigned long ticks)
1222 jiffies_64 += ticks;
1224 calc_global_load(ticks);
1228 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1229 * and sleep offsets.
1230 * @xtim: pointer to timespec to be set with xtime
1231 * @wtom: pointer to timespec to be set with wall_to_monotonic
1232 * @sleep: pointer to timespec to be set with time in suspend
1234 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1235 struct timespec *wtom, struct timespec *sleep)
1240 seq = read_seqbegin(&timekeeper.lock);
1241 *xtim = timekeeper.xtime;
1242 *wtom = timekeeper.wall_to_monotonic;
1243 *sleep = timekeeper.total_sleep_time;
1244 } while (read_seqretry(&timekeeper.lock, seq));
1248 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1250 ktime_t ktime_get_monotonic_offset(void)
1253 struct timespec wtom;
1256 seq = read_seqbegin(&timekeeper.lock);
1257 wtom = timekeeper.wall_to_monotonic;
1258 } while (read_seqretry(&timekeeper.lock, seq));
1260 return timespec_to_ktime(wtom);
1262 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1266 * xtime_update() - advances the timekeeping infrastructure
1267 * @ticks: number of ticks, that have elapsed since the last call.
1269 * Must be called with interrupts disabled.
1271 void xtime_update(unsigned long ticks)
1273 write_seqlock(&xtime_lock);
1275 write_sequnlock(&xtime_lock);