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 /* The shift value of the current clocksource. */
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. */
35 /* shifted nano seconds left over when rounding cycle_interval */
37 /* Raw nano seconds accumulated per NTP interval. */
40 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
42 /* Difference between accumulated time and NTP time in ntp
43 * shifted nano seconds. */
45 /* Shift conversion between clock shifted nano seconds and
46 * ntp shifted nano seconds. */
48 /* NTP adjusted clock multiplier */
52 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
53 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
54 * at zero at system boot time, so wall_to_monotonic will be negative,
55 * however, we will ALWAYS keep the tv_nsec part positive so we can use
56 * the usual normalization.
58 * wall_to_monotonic is moved after resume from suspend for the
59 * monotonic time not to jump. We need to add total_sleep_time to
60 * wall_to_monotonic to get the real boot based time offset.
62 * - wall_to_monotonic is no longer the boot time, getboottime must be
65 struct timespec wall_to_monotonic;
66 /* time spent in suspend */
67 struct timespec total_sleep_time;
71 static struct timekeeper timekeeper;
74 * timekeeper_setup_internals - Set up internals to use clocksource clock.
76 * @clock: Pointer to clocksource.
78 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
79 * pair and interval request.
81 * Unless you're the timekeeping code, you should not be using this!
83 static void timekeeper_setup_internals(struct clocksource *clock)
88 timekeeper.clock = clock;
89 clock->cycle_last = clock->read(clock);
91 /* Do the ns -> cycle conversion first, using original mult */
92 tmp = NTP_INTERVAL_LENGTH;
96 do_div(tmp, clock->mult);
100 interval = (cycle_t) tmp;
101 timekeeper.cycle_interval = interval;
103 /* Go back from cycles -> shifted ns */
104 timekeeper.xtime_interval = (u64) interval * clock->mult;
105 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
106 timekeeper.raw_interval =
107 ((u64) interval * clock->mult) >> clock->shift;
109 timekeeper.xtime_nsec = 0;
110 timekeeper.shift = clock->shift;
112 timekeeper.ntp_error = 0;
113 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
116 * The timekeeper keeps its own mult values for the currently
117 * active clocksource. These value will be adjusted via NTP
118 * to counteract clock drifting.
120 timekeeper.mult = clock->mult;
123 /* Timekeeper helper functions. */
124 static inline s64 timekeeping_get_ns(void)
126 cycle_t cycle_now, cycle_delta;
127 struct clocksource *clock;
129 /* read clocksource: */
130 clock = timekeeper.clock;
131 cycle_now = clock->read(clock);
133 /* calculate the delta since the last update_wall_time: */
134 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
136 /* return delta convert to nanoseconds using ntp adjusted mult. */
137 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
141 static inline s64 timekeeping_get_ns_raw(void)
143 cycle_t cycle_now, cycle_delta;
144 struct clocksource *clock;
146 /* read clocksource: */
147 clock = timekeeper.clock;
148 cycle_now = clock->read(clock);
150 /* calculate the delta since the last update_wall_time: */
151 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
153 /* return delta convert to nanoseconds. */
154 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
158 * This read-write spinlock protects us from races in SMP while
159 * playing with xtime.
161 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
167 static struct timespec xtime __attribute__ ((aligned (16)));
170 * The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock.
172 static struct timespec raw_time;
174 /* flag for if timekeeping is suspended */
175 int __read_mostly timekeeping_suspended;
177 /* must hold xtime_lock */
178 void timekeeping_leap_insert(int leapsecond)
180 xtime.tv_sec += leapsecond;
181 timekeeper.wall_to_monotonic.tv_sec -= leapsecond;
182 update_vsyscall(&xtime, &timekeeper.wall_to_monotonic, timekeeper.clock,
187 * timekeeping_forward_now - update clock to the current time
189 * Forward the current clock to update its state since the last call to
190 * update_wall_time(). This is useful before significant clock changes,
191 * as it avoids having to deal with this time offset explicitly.
193 static void timekeeping_forward_now(void)
195 cycle_t cycle_now, cycle_delta;
196 struct clocksource *clock;
199 clock = timekeeper.clock;
200 cycle_now = clock->read(clock);
201 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
202 clock->cycle_last = cycle_now;
204 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
207 /* If arch requires, add in gettimeoffset() */
208 nsec += arch_gettimeoffset();
210 timespec_add_ns(&xtime, nsec);
212 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
213 timespec_add_ns(&raw_time, nsec);
217 * getnstimeofday - Returns the time of day in a timespec
218 * @ts: pointer to the timespec to be set
220 * Returns the time of day in a timespec.
222 void getnstimeofday(struct timespec *ts)
227 WARN_ON(timekeeping_suspended);
230 seq = read_seqbegin(&xtime_lock);
233 nsecs = timekeeping_get_ns();
235 /* If arch requires, add in gettimeoffset() */
236 nsecs += arch_gettimeoffset();
238 } while (read_seqretry(&xtime_lock, seq));
240 timespec_add_ns(ts, nsecs);
243 EXPORT_SYMBOL(getnstimeofday);
245 ktime_t ktime_get(void)
250 WARN_ON(timekeeping_suspended);
253 seq = read_seqbegin(&xtime_lock);
254 secs = xtime.tv_sec + timekeeper.wall_to_monotonic.tv_sec;
255 nsecs = xtime.tv_nsec + timekeeper.wall_to_monotonic.tv_nsec;
256 nsecs += timekeeping_get_ns();
257 /* If arch requires, add in gettimeoffset() */
258 nsecs += arch_gettimeoffset();
260 } while (read_seqretry(&xtime_lock, seq));
262 * Use ktime_set/ktime_add_ns to create a proper ktime on
263 * 32-bit architectures without CONFIG_KTIME_SCALAR.
265 return ktime_add_ns(ktime_set(secs, 0), nsecs);
267 EXPORT_SYMBOL_GPL(ktime_get);
270 * ktime_get_ts - get the monotonic clock in timespec format
271 * @ts: pointer to timespec variable
273 * The function calculates the monotonic clock from the realtime
274 * clock and the wall_to_monotonic offset and stores the result
275 * in normalized timespec format in the variable pointed to by @ts.
277 void ktime_get_ts(struct timespec *ts)
279 struct timespec tomono;
283 WARN_ON(timekeeping_suspended);
286 seq = read_seqbegin(&xtime_lock);
288 tomono = timekeeper.wall_to_monotonic;
289 nsecs = timekeeping_get_ns();
290 /* If arch requires, add in gettimeoffset() */
291 nsecs += arch_gettimeoffset();
293 } while (read_seqretry(&xtime_lock, seq));
295 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
296 ts->tv_nsec + tomono.tv_nsec + nsecs);
298 EXPORT_SYMBOL_GPL(ktime_get_ts);
300 #ifdef CONFIG_NTP_PPS
303 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
304 * @ts_raw: pointer to the timespec to be set to raw monotonic time
305 * @ts_real: pointer to the timespec to be set to the time of day
307 * This function reads both the time of day and raw monotonic time at the
308 * same time atomically and stores the resulting timestamps in timespec
311 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
314 s64 nsecs_raw, nsecs_real;
316 WARN_ON_ONCE(timekeeping_suspended);
321 seq = read_seqbegin(&xtime_lock);
326 nsecs_raw = timekeeping_get_ns_raw();
327 nsecs_real = timekeeping_get_ns();
329 /* If arch requires, add in gettimeoffset() */
330 arch_offset = arch_gettimeoffset();
331 nsecs_raw += arch_offset;
332 nsecs_real += arch_offset;
334 } while (read_seqretry(&xtime_lock, seq));
336 timespec_add_ns(ts_raw, nsecs_raw);
337 timespec_add_ns(ts_real, nsecs_real);
339 EXPORT_SYMBOL(getnstime_raw_and_real);
341 #endif /* CONFIG_NTP_PPS */
344 * do_gettimeofday - Returns the time of day in a timeval
345 * @tv: pointer to the timeval to be set
347 * NOTE: Users should be converted to using getnstimeofday()
349 void do_gettimeofday(struct timeval *tv)
353 getnstimeofday(&now);
354 tv->tv_sec = now.tv_sec;
355 tv->tv_usec = now.tv_nsec/1000;
358 EXPORT_SYMBOL(do_gettimeofday);
360 * do_settimeofday - Sets the time of day
361 * @tv: pointer to the timespec variable containing the new time
363 * Sets the time of day to the new time and update NTP and notify hrtimers
365 int do_settimeofday(const struct timespec *tv)
367 struct timespec ts_delta;
370 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
373 write_seqlock_irqsave(&xtime_lock, flags);
375 timekeeping_forward_now();
377 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
378 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
379 timekeeper.wall_to_monotonic =
380 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
384 timekeeper.ntp_error = 0;
387 update_vsyscall(&xtime, &timekeeper.wall_to_monotonic, timekeeper.clock,
390 write_sequnlock_irqrestore(&xtime_lock, flags);
392 /* signal hrtimers about time change */
398 EXPORT_SYMBOL(do_settimeofday);
402 * timekeeping_inject_offset - Adds or subtracts from the current time.
403 * @tv: pointer to the timespec variable containing the offset
405 * Adds or subtracts an offset value from the current time.
407 int timekeeping_inject_offset(struct timespec *ts)
411 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
414 write_seqlock_irqsave(&xtime_lock, flags);
416 timekeeping_forward_now();
418 xtime = timespec_add(xtime, *ts);
419 timekeeper.wall_to_monotonic =
420 timespec_sub(timekeeper.wall_to_monotonic, *ts);
422 timekeeper.ntp_error = 0;
425 update_vsyscall(&xtime, &timekeeper.wall_to_monotonic, timekeeper.clock,
428 write_sequnlock_irqrestore(&xtime_lock, flags);
430 /* signal hrtimers about time change */
435 EXPORT_SYMBOL(timekeeping_inject_offset);
438 * change_clocksource - Swaps clocksources if a new one is available
440 * Accumulates current time interval and initializes new clocksource
442 static int change_clocksource(void *data)
444 struct clocksource *new, *old;
446 new = (struct clocksource *) data;
448 timekeeping_forward_now();
449 if (!new->enable || new->enable(new) == 0) {
450 old = timekeeper.clock;
451 timekeeper_setup_internals(new);
459 * timekeeping_notify - Install a new clock source
460 * @clock: pointer to the clock source
462 * This function is called from clocksource.c after a new, better clock
463 * source has been registered. The caller holds the clocksource_mutex.
465 void timekeeping_notify(struct clocksource *clock)
467 if (timekeeper.clock == clock)
469 stop_machine(change_clocksource, clock, NULL);
474 * ktime_get_real - get the real (wall-) time in ktime_t format
476 * returns the time in ktime_t format
478 ktime_t ktime_get_real(void)
482 getnstimeofday(&now);
484 return timespec_to_ktime(now);
486 EXPORT_SYMBOL_GPL(ktime_get_real);
489 * getrawmonotonic - Returns the raw monotonic time in a timespec
490 * @ts: pointer to the timespec to be set
492 * Returns the raw monotonic time (completely un-modified by ntp)
494 void getrawmonotonic(struct timespec *ts)
500 seq = read_seqbegin(&xtime_lock);
501 nsecs = timekeeping_get_ns_raw();
504 } while (read_seqretry(&xtime_lock, seq));
506 timespec_add_ns(ts, nsecs);
508 EXPORT_SYMBOL(getrawmonotonic);
512 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
514 int timekeeping_valid_for_hres(void)
520 seq = read_seqbegin(&xtime_lock);
522 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
524 } while (read_seqretry(&xtime_lock, seq));
530 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
532 * Caller must observe xtime_lock via read_seqbegin/read_seqretry to
533 * ensure that the clocksource does not change!
535 u64 timekeeping_max_deferment(void)
537 return timekeeper.clock->max_idle_ns;
541 * read_persistent_clock - Return time from the persistent clock.
543 * Weak dummy function for arches that do not yet support it.
544 * Reads the time from the battery backed persistent clock.
545 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
547 * XXX - Do be sure to remove it once all arches implement it.
549 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
556 * read_boot_clock - Return time of the system start.
558 * Weak dummy function for arches that do not yet support it.
559 * Function to read the exact time the system has been started.
560 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
562 * XXX - Do be sure to remove it once all arches implement it.
564 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
571 * timekeeping_init - Initializes the clocksource and common timekeeping values
573 void __init timekeeping_init(void)
575 struct clocksource *clock;
577 struct timespec now, boot;
579 read_persistent_clock(&now);
580 read_boot_clock(&boot);
582 write_seqlock_irqsave(&xtime_lock, flags);
586 clock = clocksource_default_clock();
588 clock->enable(clock);
589 timekeeper_setup_internals(clock);
591 xtime.tv_sec = now.tv_sec;
592 xtime.tv_nsec = now.tv_nsec;
594 raw_time.tv_nsec = 0;
595 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
596 boot.tv_sec = xtime.tv_sec;
597 boot.tv_nsec = xtime.tv_nsec;
599 set_normalized_timespec(&timekeeper.wall_to_monotonic,
600 -boot.tv_sec, -boot.tv_nsec);
601 timekeeper.total_sleep_time.tv_sec = 0;
602 timekeeper.total_sleep_time.tv_nsec = 0;
603 write_sequnlock_irqrestore(&xtime_lock, flags);
606 /* time in seconds when suspend began */
607 static struct timespec timekeeping_suspend_time;
610 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
611 * @delta: pointer to a timespec delta value
613 * Takes a timespec offset measuring a suspend interval and properly
614 * adds the sleep offset to the timekeeping variables.
616 static void __timekeeping_inject_sleeptime(struct timespec *delta)
618 if (!timespec_valid(delta)) {
619 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
620 "sleep delta value!\n");
624 xtime = timespec_add(xtime, *delta);
625 timekeeper.wall_to_monotonic =
626 timespec_sub(timekeeper.wall_to_monotonic, *delta);
627 timekeeper.total_sleep_time = timespec_add(
628 timekeeper.total_sleep_time, *delta);
633 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
634 * @delta: pointer to a timespec delta value
636 * This hook is for architectures that cannot support read_persistent_clock
637 * because their RTC/persistent clock is only accessible when irqs are enabled.
639 * This function should only be called by rtc_resume(), and allows
640 * a suspend offset to be injected into the timekeeping values.
642 void timekeeping_inject_sleeptime(struct timespec *delta)
647 /* Make sure we don't set the clock twice */
648 read_persistent_clock(&ts);
649 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
652 write_seqlock_irqsave(&xtime_lock, flags);
653 timekeeping_forward_now();
655 __timekeeping_inject_sleeptime(delta);
657 timekeeper.ntp_error = 0;
659 update_vsyscall(&xtime, &timekeeper.wall_to_monotonic, timekeeper.clock,
662 write_sequnlock_irqrestore(&xtime_lock, flags);
664 /* signal hrtimers about time change */
670 * timekeeping_resume - Resumes the generic timekeeping subsystem.
672 * This is for the generic clocksource timekeeping.
673 * xtime/wall_to_monotonic/jiffies/etc are
674 * still managed by arch specific suspend/resume code.
676 static void timekeeping_resume(void)
681 read_persistent_clock(&ts);
683 clocksource_resume();
685 write_seqlock_irqsave(&xtime_lock, flags);
687 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
688 ts = timespec_sub(ts, timekeeping_suspend_time);
689 __timekeeping_inject_sleeptime(&ts);
691 /* re-base the last cycle value */
692 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
693 timekeeper.ntp_error = 0;
694 timekeeping_suspended = 0;
695 write_sequnlock_irqrestore(&xtime_lock, flags);
697 touch_softlockup_watchdog();
699 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
701 /* Resume hrtimers */
705 static int timekeeping_suspend(void)
708 struct timespec delta, delta_delta;
709 static struct timespec old_delta;
711 read_persistent_clock(&timekeeping_suspend_time);
713 write_seqlock_irqsave(&xtime_lock, flags);
714 timekeeping_forward_now();
715 timekeeping_suspended = 1;
718 * To avoid drift caused by repeated suspend/resumes,
719 * which each can add ~1 second drift error,
720 * try to compensate so the difference in system time
721 * and persistent_clock time stays close to constant.
723 delta = timespec_sub(xtime, timekeeping_suspend_time);
724 delta_delta = timespec_sub(delta, old_delta);
725 if (abs(delta_delta.tv_sec) >= 2) {
727 * if delta_delta is too large, assume time correction
728 * has occured and set old_delta to the current delta.
732 /* Otherwise try to adjust old_system to compensate */
733 timekeeping_suspend_time =
734 timespec_add(timekeeping_suspend_time, delta_delta);
736 write_sequnlock_irqrestore(&xtime_lock, flags);
738 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
739 clocksource_suspend();
744 /* sysfs resume/suspend bits for timekeeping */
745 static struct syscore_ops timekeeping_syscore_ops = {
746 .resume = timekeeping_resume,
747 .suspend = timekeeping_suspend,
750 static int __init timekeeping_init_ops(void)
752 register_syscore_ops(&timekeeping_syscore_ops);
756 device_initcall(timekeeping_init_ops);
759 * If the error is already larger, we look ahead even further
760 * to compensate for late or lost adjustments.
762 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
770 * Use the current error value to determine how much to look ahead.
771 * The larger the error the slower we adjust for it to avoid problems
772 * with losing too many ticks, otherwise we would overadjust and
773 * produce an even larger error. The smaller the adjustment the
774 * faster we try to adjust for it, as lost ticks can do less harm
775 * here. This is tuned so that an error of about 1 msec is adjusted
776 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
778 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
779 error2 = abs(error2);
780 for (look_ahead = 0; error2 > 0; look_ahead++)
784 * Now calculate the error in (1 << look_ahead) ticks, but first
785 * remove the single look ahead already included in the error.
787 tick_error = tick_length >> (timekeeper.ntp_error_shift + 1);
788 tick_error -= timekeeper.xtime_interval >> 1;
789 error = ((error - tick_error) >> look_ahead) + tick_error;
791 /* Finally calculate the adjustment shift value. */
796 *interval = -*interval;
800 for (adj = 0; error > i; adj++)
809 * Adjust the multiplier to reduce the error value,
810 * this is optimized for the most common adjustments of -1,0,1,
811 * for other values we can do a bit more work.
813 static void timekeeping_adjust(s64 offset)
815 s64 error, interval = timekeeper.cycle_interval;
819 * The point of this is to check if the error is greater then half
822 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
824 * Note we subtract one in the shift, so that error is really error*2.
825 * This "saves" dividing(shifting) interval twice, but keeps the
826 * (error > interval) comparison as still measuring if error is
827 * larger then half an interval.
829 * Note: It does not "save" on aggravation when reading the code.
831 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
832 if (error > interval) {
834 * We now divide error by 4(via shift), which checks if
835 * the error is greater then twice the interval.
836 * If it is greater, we need a bigadjust, if its smaller,
837 * we can adjust by 1.
841 * XXX - In update_wall_time, we round up to the next
842 * nanosecond, and store the amount rounded up into
843 * the error. This causes the likely below to be unlikely.
845 * The proper fix is to avoid rounding up by using
846 * the high precision timekeeper.xtime_nsec instead of
847 * xtime.tv_nsec everywhere. Fixing this will take some
850 if (likely(error <= interval))
853 adj = timekeeping_bigadjust(error, &interval, &offset);
854 } else if (error < -interval) {
855 /* See comment above, this is just switched for the negative */
857 if (likely(error >= -interval)) {
859 interval = -interval;
862 adj = timekeeping_bigadjust(error, &interval, &offset);
863 } else /* No adjustment needed */
866 WARN_ONCE(timekeeper.clock->maxadj &&
867 (timekeeper.mult + adj > timekeeper.clock->mult +
868 timekeeper.clock->maxadj),
869 "Adjusting %s more then 11%% (%ld vs %ld)\n",
870 timekeeper.clock->name, (long)timekeeper.mult + adj,
871 (long)timekeeper.clock->mult +
872 timekeeper.clock->maxadj);
874 * So the following can be confusing.
876 * To keep things simple, lets assume adj == 1 for now.
878 * When adj != 1, remember that the interval and offset values
879 * have been appropriately scaled so the math is the same.
881 * The basic idea here is that we're increasing the multiplier
882 * by one, this causes the xtime_interval to be incremented by
883 * one cycle_interval. This is because:
884 * xtime_interval = cycle_interval * mult
885 * So if mult is being incremented by one:
886 * xtime_interval = cycle_interval * (mult + 1)
888 * xtime_interval = (cycle_interval * mult) + cycle_interval
889 * Which can be shortened to:
890 * xtime_interval += cycle_interval
892 * So offset stores the non-accumulated cycles. Thus the current
893 * time (in shifted nanoseconds) is:
894 * now = (offset * adj) + xtime_nsec
895 * Now, even though we're adjusting the clock frequency, we have
896 * to keep time consistent. In other words, we can't jump back
897 * in time, and we also want to avoid jumping forward in time.
899 * So given the same offset value, we need the time to be the same
900 * both before and after the freq adjustment.
901 * now = (offset * adj_1) + xtime_nsec_1
902 * now = (offset * adj_2) + xtime_nsec_2
904 * (offset * adj_1) + xtime_nsec_1 =
905 * (offset * adj_2) + xtime_nsec_2
909 * (offset * adj_1) + xtime_nsec_1 =
910 * (offset * (adj_1+1)) + xtime_nsec_2
911 * (offset * adj_1) + xtime_nsec_1 =
912 * (offset * adj_1) + offset + xtime_nsec_2
913 * Canceling the sides:
914 * xtime_nsec_1 = offset + xtime_nsec_2
916 * xtime_nsec_2 = xtime_nsec_1 - offset
917 * Which simplfies to:
918 * xtime_nsec -= offset
920 * XXX - TODO: Doc ntp_error calculation.
922 timekeeper.mult += adj;
923 timekeeper.xtime_interval += interval;
924 timekeeper.xtime_nsec -= offset;
925 timekeeper.ntp_error -= (interval - offset) <<
926 timekeeper.ntp_error_shift;
931 * logarithmic_accumulation - shifted accumulation of cycles
933 * This functions accumulates a shifted interval of cycles into
934 * into a shifted interval nanoseconds. Allows for O(log) accumulation
937 * Returns the unconsumed cycles.
939 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
941 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
944 /* If the offset is smaller then a shifted interval, do nothing */
945 if (offset < timekeeper.cycle_interval<<shift)
948 /* Accumulate one shifted interval */
949 offset -= timekeeper.cycle_interval << shift;
950 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
952 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
953 while (timekeeper.xtime_nsec >= nsecps) {
954 timekeeper.xtime_nsec -= nsecps;
959 /* Accumulate raw time */
960 raw_nsecs = timekeeper.raw_interval << shift;
961 raw_nsecs += raw_time.tv_nsec;
962 if (raw_nsecs >= NSEC_PER_SEC) {
963 u64 raw_secs = raw_nsecs;
964 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
965 raw_time.tv_sec += raw_secs;
967 raw_time.tv_nsec = raw_nsecs;
969 /* Accumulate error between NTP and clock interval */
970 timekeeper.ntp_error += tick_length << shift;
971 timekeeper.ntp_error -=
972 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
973 (timekeeper.ntp_error_shift + shift);
980 * update_wall_time - Uses the current clocksource to increment the wall time
982 * Called from the timer interrupt, must hold a write on xtime_lock.
984 static void update_wall_time(void)
986 struct clocksource *clock;
988 int shift = 0, maxshift;
990 /* Make sure we're fully resumed: */
991 if (unlikely(timekeeping_suspended))
994 clock = timekeeper.clock;
996 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
997 offset = timekeeper.cycle_interval;
999 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1001 timekeeper.xtime_nsec = (s64)xtime.tv_nsec << timekeeper.shift;
1004 * With NO_HZ we may have to accumulate many cycle_intervals
1005 * (think "ticks") worth of time at once. To do this efficiently,
1006 * we calculate the largest doubling multiple of cycle_intervals
1007 * that is smaller then the offset. We then accumulate that
1008 * chunk in one go, and then try to consume the next smaller
1011 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1012 shift = max(0, shift);
1013 /* Bound shift to one less then what overflows tick_length */
1014 maxshift = (8*sizeof(tick_length) - (ilog2(tick_length)+1)) - 1;
1015 shift = min(shift, maxshift);
1016 while (offset >= timekeeper.cycle_interval) {
1017 offset = logarithmic_accumulation(offset, shift);
1018 if(offset < timekeeper.cycle_interval<<shift)
1022 /* correct the clock when NTP error is too big */
1023 timekeeping_adjust(offset);
1026 * Since in the loop above, we accumulate any amount of time
1027 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1028 * xtime_nsec to be fairly small after the loop. Further, if we're
1029 * slightly speeding the clocksource up in timekeeping_adjust(),
1030 * its possible the required corrective factor to xtime_nsec could
1031 * cause it to underflow.
1033 * Now, we cannot simply roll the accumulated second back, since
1034 * the NTP subsystem has been notified via second_overflow. So
1035 * instead we push xtime_nsec forward by the amount we underflowed,
1036 * and add that amount into the error.
1038 * We'll correct this error next time through this function, when
1039 * xtime_nsec is not as small.
1041 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1042 s64 neg = -(s64)timekeeper.xtime_nsec;
1043 timekeeper.xtime_nsec = 0;
1044 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1049 * Store full nanoseconds into xtime after rounding it up and
1050 * add the remainder to the error difference.
1052 xtime.tv_nsec = ((s64) timekeeper.xtime_nsec >> timekeeper.shift) + 1;
1053 timekeeper.xtime_nsec -= (s64) xtime.tv_nsec << timekeeper.shift;
1054 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1055 timekeeper.ntp_error_shift;
1058 * Finally, make sure that after the rounding
1059 * xtime.tv_nsec isn't larger then NSEC_PER_SEC
1061 if (unlikely(xtime.tv_nsec >= NSEC_PER_SEC)) {
1062 xtime.tv_nsec -= NSEC_PER_SEC;
1067 /* check to see if there is a new clocksource to use */
1068 update_vsyscall(&xtime, &timekeeper.wall_to_monotonic, timekeeper.clock,
1073 * getboottime - Return the real time of system boot.
1074 * @ts: pointer to the timespec to be set
1076 * Returns the wall-time of boot in a timespec.
1078 * This is based on the wall_to_monotonic offset and the total suspend
1079 * time. Calls to settimeofday will affect the value returned (which
1080 * basically means that however wrong your real time clock is at boot time,
1081 * you get the right time here).
1083 void getboottime(struct timespec *ts)
1085 struct timespec boottime = {
1086 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1087 timekeeper.total_sleep_time.tv_sec,
1088 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1089 timekeeper.total_sleep_time.tv_nsec
1092 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1094 EXPORT_SYMBOL_GPL(getboottime);
1098 * get_monotonic_boottime - Returns monotonic time since boot
1099 * @ts: pointer to the timespec to be set
1101 * Returns the monotonic time since boot in a timespec.
1103 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1104 * includes the time spent in suspend.
1106 void get_monotonic_boottime(struct timespec *ts)
1108 struct timespec tomono, sleep;
1112 WARN_ON(timekeeping_suspended);
1115 seq = read_seqbegin(&xtime_lock);
1117 tomono = timekeeper.wall_to_monotonic;
1118 sleep = timekeeper.total_sleep_time;
1119 nsecs = timekeeping_get_ns();
1121 } while (read_seqretry(&xtime_lock, seq));
1123 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1124 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1126 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1129 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1131 * Returns the monotonic time since boot in a ktime
1133 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1134 * includes the time spent in suspend.
1136 ktime_t ktime_get_boottime(void)
1140 get_monotonic_boottime(&ts);
1141 return timespec_to_ktime(ts);
1143 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1146 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1147 * @ts: pointer to the timespec to be converted
1149 void monotonic_to_bootbased(struct timespec *ts)
1151 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1153 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1155 unsigned long get_seconds(void)
1157 return xtime.tv_sec;
1159 EXPORT_SYMBOL(get_seconds);
1161 struct timespec __current_kernel_time(void)
1166 struct timespec current_kernel_time(void)
1168 struct timespec now;
1172 seq = read_seqbegin(&xtime_lock);
1175 } while (read_seqretry(&xtime_lock, seq));
1179 EXPORT_SYMBOL(current_kernel_time);
1181 struct timespec get_monotonic_coarse(void)
1183 struct timespec now, mono;
1187 seq = read_seqbegin(&xtime_lock);
1190 mono = timekeeper.wall_to_monotonic;
1191 } while (read_seqretry(&xtime_lock, seq));
1193 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1194 now.tv_nsec + mono.tv_nsec);
1199 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1200 * without sampling the sequence number in xtime_lock.
1201 * jiffies is defined in the linker script...
1203 void do_timer(unsigned long ticks)
1205 jiffies_64 += ticks;
1207 calc_global_load(ticks);
1211 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1212 * and sleep offsets.
1213 * @xtim: pointer to timespec to be set with xtime
1214 * @wtom: pointer to timespec to be set with wall_to_monotonic
1215 * @sleep: pointer to timespec to be set with time in suspend
1217 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1218 struct timespec *wtom, struct timespec *sleep)
1223 seq = read_seqbegin(&xtime_lock);
1225 *wtom = timekeeper.wall_to_monotonic;
1226 *sleep = timekeeper.total_sleep_time;
1227 } while (read_seqretry(&xtime_lock, seq));
1231 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1233 ktime_t ktime_get_monotonic_offset(void)
1236 struct timespec wtom;
1239 seq = read_seqbegin(&xtime_lock);
1240 wtom = timekeeper.wall_to_monotonic;
1241 } while (read_seqretry(&xtime_lock, seq));
1242 return timespec_to_ktime(wtom);
1246 * xtime_update() - advances the timekeeping infrastructure
1247 * @ticks: number of ticks, that have elapsed since the last call.
1249 * Must be called with interrupts disabled.
1251 void xtime_update(unsigned long ticks)
1253 write_seqlock(&xtime_lock);
1255 write_sequnlock(&xtime_lock);