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);
187 void timekeeping_leap_insert(int leapsecond)
191 write_seqlock_irqsave(&timekeeper.lock, flags);
192 timekeeper.xtime.tv_sec += leapsecond;
193 timekeeper.wall_to_monotonic.tv_sec -= leapsecond;
194 timekeeping_update(false);
195 write_sequnlock_irqrestore(&timekeeper.lock, flags);
200 * timekeeping_forward_now - update clock to the current time
202 * Forward the current clock to update its state since the last call to
203 * update_wall_time(). This is useful before significant clock changes,
204 * as it avoids having to deal with this time offset explicitly.
206 static void timekeeping_forward_now(void)
208 cycle_t cycle_now, cycle_delta;
209 struct clocksource *clock;
212 clock = timekeeper.clock;
213 cycle_now = clock->read(clock);
214 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
215 clock->cycle_last = cycle_now;
217 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
220 /* If arch requires, add in gettimeoffset() */
221 nsec += arch_gettimeoffset();
223 timespec_add_ns(&timekeeper.xtime, nsec);
225 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
226 timespec_add_ns(&timekeeper.raw_time, nsec);
230 * getnstimeofday - Returns the time of day in a timespec
231 * @ts: pointer to the timespec to be set
233 * Returns the time of day in a timespec.
235 void getnstimeofday(struct timespec *ts)
240 WARN_ON(timekeeping_suspended);
243 seq = read_seqbegin(&timekeeper.lock);
245 *ts = timekeeper.xtime;
246 nsecs = timekeeping_get_ns();
248 /* If arch requires, add in gettimeoffset() */
249 nsecs += arch_gettimeoffset();
251 } while (read_seqretry(&timekeeper.lock, seq));
253 timespec_add_ns(ts, nsecs);
256 EXPORT_SYMBOL(getnstimeofday);
258 ktime_t ktime_get(void)
263 WARN_ON(timekeeping_suspended);
266 seq = read_seqbegin(&timekeeper.lock);
267 secs = timekeeper.xtime.tv_sec +
268 timekeeper.wall_to_monotonic.tv_sec;
269 nsecs = timekeeper.xtime.tv_nsec +
270 timekeeper.wall_to_monotonic.tv_nsec;
271 nsecs += timekeeping_get_ns();
272 /* If arch requires, add in gettimeoffset() */
273 nsecs += arch_gettimeoffset();
275 } while (read_seqretry(&timekeeper.lock, seq));
277 * Use ktime_set/ktime_add_ns to create a proper ktime on
278 * 32-bit architectures without CONFIG_KTIME_SCALAR.
280 return ktime_add_ns(ktime_set(secs, 0), nsecs);
282 EXPORT_SYMBOL_GPL(ktime_get);
285 * ktime_get_ts - get the monotonic clock in timespec format
286 * @ts: pointer to timespec variable
288 * The function calculates the monotonic clock from the realtime
289 * clock and the wall_to_monotonic offset and stores the result
290 * in normalized timespec format in the variable pointed to by @ts.
292 void ktime_get_ts(struct timespec *ts)
294 struct timespec tomono;
298 WARN_ON(timekeeping_suspended);
301 seq = read_seqbegin(&timekeeper.lock);
302 *ts = timekeeper.xtime;
303 tomono = timekeeper.wall_to_monotonic;
304 nsecs = timekeeping_get_ns();
305 /* If arch requires, add in gettimeoffset() */
306 nsecs += arch_gettimeoffset();
308 } while (read_seqretry(&timekeeper.lock, seq));
310 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
311 ts->tv_nsec + tomono.tv_nsec + nsecs);
313 EXPORT_SYMBOL_GPL(ktime_get_ts);
315 #ifdef CONFIG_NTP_PPS
318 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
319 * @ts_raw: pointer to the timespec to be set to raw monotonic time
320 * @ts_real: pointer to the timespec to be set to the time of day
322 * This function reads both the time of day and raw monotonic time at the
323 * same time atomically and stores the resulting timestamps in timespec
326 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
329 s64 nsecs_raw, nsecs_real;
331 WARN_ON_ONCE(timekeeping_suspended);
336 seq = read_seqbegin(&timekeeper.lock);
338 *ts_raw = timekeeper.raw_time;
339 *ts_real = timekeeper.xtime;
341 nsecs_raw = timekeeping_get_ns_raw();
342 nsecs_real = timekeeping_get_ns();
344 /* If arch requires, add in gettimeoffset() */
345 arch_offset = arch_gettimeoffset();
346 nsecs_raw += arch_offset;
347 nsecs_real += arch_offset;
349 } while (read_seqretry(&timekeeper.lock, seq));
351 timespec_add_ns(ts_raw, nsecs_raw);
352 timespec_add_ns(ts_real, nsecs_real);
354 EXPORT_SYMBOL(getnstime_raw_and_real);
356 #endif /* CONFIG_NTP_PPS */
359 * do_gettimeofday - Returns the time of day in a timeval
360 * @tv: pointer to the timeval to be set
362 * NOTE: Users should be converted to using getnstimeofday()
364 void do_gettimeofday(struct timeval *tv)
368 getnstimeofday(&now);
369 tv->tv_sec = now.tv_sec;
370 tv->tv_usec = now.tv_nsec/1000;
373 EXPORT_SYMBOL(do_gettimeofday);
375 * do_settimeofday - Sets the time of day
376 * @tv: pointer to the timespec variable containing the new time
378 * Sets the time of day to the new time and update NTP and notify hrtimers
380 int do_settimeofday(const struct timespec *tv)
382 struct timespec ts_delta;
385 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
388 write_seqlock_irqsave(&timekeeper.lock, flags);
390 timekeeping_forward_now();
392 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
393 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
394 timekeeper.wall_to_monotonic =
395 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
397 timekeeper.xtime = *tv;
398 timekeeping_update(true);
400 write_sequnlock_irqrestore(&timekeeper.lock, flags);
402 /* signal hrtimers about time change */
408 EXPORT_SYMBOL(do_settimeofday);
412 * timekeeping_inject_offset - Adds or subtracts from the current time.
413 * @tv: pointer to the timespec variable containing the offset
415 * Adds or subtracts an offset value from the current time.
417 int timekeeping_inject_offset(struct timespec *ts)
421 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
424 write_seqlock_irqsave(&timekeeper.lock, flags);
426 timekeeping_forward_now();
428 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
429 timekeeper.wall_to_monotonic =
430 timespec_sub(timekeeper.wall_to_monotonic, *ts);
432 timekeeping_update(true);
434 write_sequnlock_irqrestore(&timekeeper.lock, flags);
436 /* signal hrtimers about time change */
441 EXPORT_SYMBOL(timekeeping_inject_offset);
444 * change_clocksource - Swaps clocksources if a new one is available
446 * Accumulates current time interval and initializes new clocksource
448 static int change_clocksource(void *data)
450 struct clocksource *new, *old;
452 new = (struct clocksource *) data;
454 timekeeping_forward_now();
455 if (!new->enable || new->enable(new) == 0) {
456 old = timekeeper.clock;
457 timekeeper_setup_internals(new);
465 * timekeeping_notify - Install a new clock source
466 * @clock: pointer to the clock source
468 * This function is called from clocksource.c after a new, better clock
469 * source has been registered. The caller holds the clocksource_mutex.
471 void timekeeping_notify(struct clocksource *clock)
473 if (timekeeper.clock == clock)
475 stop_machine(change_clocksource, clock, NULL);
480 * ktime_get_real - get the real (wall-) time in ktime_t format
482 * returns the time in ktime_t format
484 ktime_t ktime_get_real(void)
488 getnstimeofday(&now);
490 return timespec_to_ktime(now);
492 EXPORT_SYMBOL_GPL(ktime_get_real);
495 * getrawmonotonic - Returns the raw monotonic time in a timespec
496 * @ts: pointer to the timespec to be set
498 * Returns the raw monotonic time (completely un-modified by ntp)
500 void getrawmonotonic(struct timespec *ts)
506 seq = read_seqbegin(&timekeeper.lock);
507 nsecs = timekeeping_get_ns_raw();
508 *ts = timekeeper.raw_time;
510 } while (read_seqretry(&timekeeper.lock, seq));
512 timespec_add_ns(ts, nsecs);
514 EXPORT_SYMBOL(getrawmonotonic);
518 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
520 int timekeeping_valid_for_hres(void)
526 seq = read_seqbegin(&timekeeper.lock);
528 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
530 } while (read_seqretry(&timekeeper.lock, seq));
536 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
538 u64 timekeeping_max_deferment(void)
543 seq = read_seqbegin(&timekeeper.lock);
545 ret = timekeeper.clock->max_idle_ns;
547 } while (read_seqretry(&timekeeper.lock, seq));
553 * read_persistent_clock - Return time from the persistent clock.
555 * Weak dummy function for arches that do not yet support it.
556 * Reads the time from the battery backed persistent clock.
557 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
559 * XXX - Do be sure to remove it once all arches implement it.
561 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
568 * read_boot_clock - Return time of the system start.
570 * Weak dummy function for arches that do not yet support it.
571 * Function to read the exact time the system has been started.
572 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
574 * XXX - Do be sure to remove it once all arches implement it.
576 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
583 * timekeeping_init - Initializes the clocksource and common timekeeping values
585 void __init timekeeping_init(void)
587 struct clocksource *clock;
589 struct timespec now, boot;
591 read_persistent_clock(&now);
592 read_boot_clock(&boot);
594 seqlock_init(&timekeeper.lock);
598 write_seqlock_irqsave(&timekeeper.lock, flags);
599 clock = clocksource_default_clock();
601 clock->enable(clock);
602 timekeeper_setup_internals(clock);
604 timekeeper.xtime.tv_sec = now.tv_sec;
605 timekeeper.xtime.tv_nsec = now.tv_nsec;
606 timekeeper.raw_time.tv_sec = 0;
607 timekeeper.raw_time.tv_nsec = 0;
608 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
609 boot.tv_sec = timekeeper.xtime.tv_sec;
610 boot.tv_nsec = timekeeper.xtime.tv_nsec;
612 set_normalized_timespec(&timekeeper.wall_to_monotonic,
613 -boot.tv_sec, -boot.tv_nsec);
614 timekeeper.total_sleep_time.tv_sec = 0;
615 timekeeper.total_sleep_time.tv_nsec = 0;
616 write_sequnlock_irqrestore(&timekeeper.lock, flags);
619 /* time in seconds when suspend began */
620 static struct timespec timekeeping_suspend_time;
623 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
624 * @delta: pointer to a timespec delta value
626 * Takes a timespec offset measuring a suspend interval and properly
627 * adds the sleep offset to the timekeeping variables.
629 static void __timekeeping_inject_sleeptime(struct timespec *delta)
631 if (!timespec_valid(delta)) {
632 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
633 "sleep delta value!\n");
637 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
638 timekeeper.wall_to_monotonic =
639 timespec_sub(timekeeper.wall_to_monotonic, *delta);
640 timekeeper.total_sleep_time = timespec_add(
641 timekeeper.total_sleep_time, *delta);
646 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
647 * @delta: pointer to a timespec delta value
649 * This hook is for architectures that cannot support read_persistent_clock
650 * because their RTC/persistent clock is only accessible when irqs are enabled.
652 * This function should only be called by rtc_resume(), and allows
653 * a suspend offset to be injected into the timekeeping values.
655 void timekeeping_inject_sleeptime(struct timespec *delta)
660 /* Make sure we don't set the clock twice */
661 read_persistent_clock(&ts);
662 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
665 write_seqlock_irqsave(&timekeeper.lock, flags);
667 timekeeping_forward_now();
669 __timekeeping_inject_sleeptime(delta);
671 timekeeping_update(true);
673 write_sequnlock_irqrestore(&timekeeper.lock, flags);
675 /* signal hrtimers about time change */
681 * timekeeping_resume - Resumes the generic timekeeping subsystem.
683 * This is for the generic clocksource timekeeping.
684 * xtime/wall_to_monotonic/jiffies/etc are
685 * still managed by arch specific suspend/resume code.
687 static void timekeeping_resume(void)
692 read_persistent_clock(&ts);
694 clocksource_resume();
696 write_seqlock_irqsave(&timekeeper.lock, flags);
698 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
699 ts = timespec_sub(ts, timekeeping_suspend_time);
700 __timekeeping_inject_sleeptime(&ts);
702 /* re-base the last cycle value */
703 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
704 timekeeper.ntp_error = 0;
705 timekeeping_suspended = 0;
706 write_sequnlock_irqrestore(&timekeeper.lock, flags);
708 touch_softlockup_watchdog();
710 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
712 /* Resume hrtimers */
716 static int timekeeping_suspend(void)
719 struct timespec delta, delta_delta;
720 static struct timespec old_delta;
722 read_persistent_clock(&timekeeping_suspend_time);
724 write_seqlock_irqsave(&timekeeper.lock, flags);
725 timekeeping_forward_now();
726 timekeeping_suspended = 1;
729 * To avoid drift caused by repeated suspend/resumes,
730 * which each can add ~1 second drift error,
731 * try to compensate so the difference in system time
732 * and persistent_clock time stays close to constant.
734 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
735 delta_delta = timespec_sub(delta, old_delta);
736 if (abs(delta_delta.tv_sec) >= 2) {
738 * if delta_delta is too large, assume time correction
739 * has occured and set old_delta to the current delta.
743 /* Otherwise try to adjust old_system to compensate */
744 timekeeping_suspend_time =
745 timespec_add(timekeeping_suspend_time, delta_delta);
747 write_sequnlock_irqrestore(&timekeeper.lock, flags);
749 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
750 clocksource_suspend();
755 /* sysfs resume/suspend bits for timekeeping */
756 static struct syscore_ops timekeeping_syscore_ops = {
757 .resume = timekeeping_resume,
758 .suspend = timekeeping_suspend,
761 static int __init timekeeping_init_ops(void)
763 register_syscore_ops(&timekeeping_syscore_ops);
767 device_initcall(timekeeping_init_ops);
770 * If the error is already larger, we look ahead even further
771 * to compensate for late or lost adjustments.
773 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
781 * Use the current error value to determine how much to look ahead.
782 * The larger the error the slower we adjust for it to avoid problems
783 * with losing too many ticks, otherwise we would overadjust and
784 * produce an even larger error. The smaller the adjustment the
785 * faster we try to adjust for it, as lost ticks can do less harm
786 * here. This is tuned so that an error of about 1 msec is adjusted
787 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
789 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
790 error2 = abs(error2);
791 for (look_ahead = 0; error2 > 0; look_ahead++)
795 * Now calculate the error in (1 << look_ahead) ticks, but first
796 * remove the single look ahead already included in the error.
798 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
799 tick_error -= timekeeper.xtime_interval >> 1;
800 error = ((error - tick_error) >> look_ahead) + tick_error;
802 /* Finally calculate the adjustment shift value. */
807 *interval = -*interval;
811 for (adj = 0; error > i; adj++)
820 * Adjust the multiplier to reduce the error value,
821 * this is optimized for the most common adjustments of -1,0,1,
822 * for other values we can do a bit more work.
824 static void timekeeping_adjust(s64 offset)
826 s64 error, interval = timekeeper.cycle_interval;
830 * The point of this is to check if the error is greater then half
833 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
835 * Note we subtract one in the shift, so that error is really error*2.
836 * This "saves" dividing(shifting) interval twice, but keeps the
837 * (error > interval) comparison as still measuring if error is
838 * larger then half an interval.
840 * Note: It does not "save" on aggravation when reading the code.
842 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
843 if (error > interval) {
845 * We now divide error by 4(via shift), which checks if
846 * the error is greater then twice the interval.
847 * If it is greater, we need a bigadjust, if its smaller,
848 * we can adjust by 1.
852 * XXX - In update_wall_time, we round up to the next
853 * nanosecond, and store the amount rounded up into
854 * the error. This causes the likely below to be unlikely.
856 * The proper fix is to avoid rounding up by using
857 * the high precision timekeeper.xtime_nsec instead of
858 * xtime.tv_nsec everywhere. Fixing this will take some
861 if (likely(error <= interval))
864 adj = timekeeping_bigadjust(error, &interval, &offset);
865 } else if (error < -interval) {
866 /* See comment above, this is just switched for the negative */
868 if (likely(error >= -interval)) {
870 interval = -interval;
873 adj = timekeeping_bigadjust(error, &interval, &offset);
874 } else /* No adjustment needed */
877 WARN_ONCE(timekeeper.clock->maxadj &&
878 (timekeeper.mult + adj > timekeeper.clock->mult +
879 timekeeper.clock->maxadj),
880 "Adjusting %s more then 11%% (%ld vs %ld)\n",
881 timekeeper.clock->name, (long)timekeeper.mult + adj,
882 (long)timekeeper.clock->mult +
883 timekeeper.clock->maxadj);
885 * So the following can be confusing.
887 * To keep things simple, lets assume adj == 1 for now.
889 * When adj != 1, remember that the interval and offset values
890 * have been appropriately scaled so the math is the same.
892 * The basic idea here is that we're increasing the multiplier
893 * by one, this causes the xtime_interval to be incremented by
894 * one cycle_interval. This is because:
895 * xtime_interval = cycle_interval * mult
896 * So if mult is being incremented by one:
897 * xtime_interval = cycle_interval * (mult + 1)
899 * xtime_interval = (cycle_interval * mult) + cycle_interval
900 * Which can be shortened to:
901 * xtime_interval += cycle_interval
903 * So offset stores the non-accumulated cycles. Thus the current
904 * time (in shifted nanoseconds) is:
905 * now = (offset * adj) + xtime_nsec
906 * Now, even though we're adjusting the clock frequency, we have
907 * to keep time consistent. In other words, we can't jump back
908 * in time, and we also want to avoid jumping forward in time.
910 * So given the same offset value, we need the time to be the same
911 * both before and after the freq adjustment.
912 * now = (offset * adj_1) + xtime_nsec_1
913 * now = (offset * adj_2) + xtime_nsec_2
915 * (offset * adj_1) + xtime_nsec_1 =
916 * (offset * adj_2) + xtime_nsec_2
920 * (offset * adj_1) + xtime_nsec_1 =
921 * (offset * (adj_1+1)) + xtime_nsec_2
922 * (offset * adj_1) + xtime_nsec_1 =
923 * (offset * adj_1) + offset + xtime_nsec_2
924 * Canceling the sides:
925 * xtime_nsec_1 = offset + xtime_nsec_2
927 * xtime_nsec_2 = xtime_nsec_1 - offset
928 * Which simplfies to:
929 * xtime_nsec -= offset
931 * XXX - TODO: Doc ntp_error calculation.
933 timekeeper.mult += adj;
934 timekeeper.xtime_interval += interval;
935 timekeeper.xtime_nsec -= offset;
936 timekeeper.ntp_error -= (interval - offset) <<
937 timekeeper.ntp_error_shift;
942 * logarithmic_accumulation - shifted accumulation of cycles
944 * This functions accumulates a shifted interval of cycles into
945 * into a shifted interval nanoseconds. Allows for O(log) accumulation
948 * Returns the unconsumed cycles.
950 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
952 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
955 /* If the offset is smaller then a shifted interval, do nothing */
956 if (offset < timekeeper.cycle_interval<<shift)
959 /* Accumulate one shifted interval */
960 offset -= timekeeper.cycle_interval << shift;
961 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
963 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
964 while (timekeeper.xtime_nsec >= nsecps) {
965 timekeeper.xtime_nsec -= nsecps;
966 timekeeper.xtime.tv_sec++;
970 /* Accumulate raw time */
971 raw_nsecs = timekeeper.raw_interval << shift;
972 raw_nsecs += timekeeper.raw_time.tv_nsec;
973 if (raw_nsecs >= NSEC_PER_SEC) {
974 u64 raw_secs = raw_nsecs;
975 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
976 timekeeper.raw_time.tv_sec += raw_secs;
978 timekeeper.raw_time.tv_nsec = raw_nsecs;
980 /* Accumulate error between NTP and clock interval */
981 timekeeper.ntp_error += ntp_tick_length() << shift;
982 timekeeper.ntp_error -=
983 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
984 (timekeeper.ntp_error_shift + shift);
991 * update_wall_time - Uses the current clocksource to increment the wall time
994 static void update_wall_time(void)
996 struct clocksource *clock;
998 int shift = 0, maxshift;
1001 write_seqlock_irqsave(&timekeeper.lock, flags);
1003 /* Make sure we're fully resumed: */
1004 if (unlikely(timekeeping_suspended))
1007 clock = timekeeper.clock;
1009 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1010 offset = timekeeper.cycle_interval;
1012 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1014 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1018 * With NO_HZ we may have to accumulate many cycle_intervals
1019 * (think "ticks") worth of time at once. To do this efficiently,
1020 * we calculate the largest doubling multiple of cycle_intervals
1021 * that is smaller then the offset. We then accumulate that
1022 * chunk in one go, and then try to consume the next smaller
1025 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1026 shift = max(0, shift);
1027 /* Bound shift to one less then what overflows tick_length */
1028 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1029 shift = min(shift, maxshift);
1030 while (offset >= timekeeper.cycle_interval) {
1031 offset = logarithmic_accumulation(offset, shift);
1032 if(offset < timekeeper.cycle_interval<<shift)
1036 /* correct the clock when NTP error is too big */
1037 timekeeping_adjust(offset);
1040 * Since in the loop above, we accumulate any amount of time
1041 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1042 * xtime_nsec to be fairly small after the loop. Further, if we're
1043 * slightly speeding the clocksource up in timekeeping_adjust(),
1044 * its possible the required corrective factor to xtime_nsec could
1045 * cause it to underflow.
1047 * Now, we cannot simply roll the accumulated second back, since
1048 * the NTP subsystem has been notified via second_overflow. So
1049 * instead we push xtime_nsec forward by the amount we underflowed,
1050 * and add that amount into the error.
1052 * We'll correct this error next time through this function, when
1053 * xtime_nsec is not as small.
1055 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1056 s64 neg = -(s64)timekeeper.xtime_nsec;
1057 timekeeper.xtime_nsec = 0;
1058 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1063 * Store full nanoseconds into xtime after rounding it up and
1064 * add the remainder to the error difference.
1066 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1067 timekeeper.shift) + 1;
1068 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1070 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1071 timekeeper.ntp_error_shift;
1074 * Finally, make sure that after the rounding
1075 * xtime.tv_nsec isn't larger then NSEC_PER_SEC
1077 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1078 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1079 timekeeper.xtime.tv_sec++;
1083 timekeeping_update(false);
1086 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1091 * getboottime - Return the real time of system boot.
1092 * @ts: pointer to the timespec to be set
1094 * Returns the wall-time of boot in a timespec.
1096 * This is based on the wall_to_monotonic offset and the total suspend
1097 * time. Calls to settimeofday will affect the value returned (which
1098 * basically means that however wrong your real time clock is at boot time,
1099 * you get the right time here).
1101 void getboottime(struct timespec *ts)
1103 struct timespec boottime = {
1104 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1105 timekeeper.total_sleep_time.tv_sec,
1106 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1107 timekeeper.total_sleep_time.tv_nsec
1110 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1112 EXPORT_SYMBOL_GPL(getboottime);
1116 * get_monotonic_boottime - Returns monotonic time since boot
1117 * @ts: pointer to the timespec to be set
1119 * Returns the monotonic time since boot in a timespec.
1121 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1122 * includes the time spent in suspend.
1124 void get_monotonic_boottime(struct timespec *ts)
1126 struct timespec tomono, sleep;
1130 WARN_ON(timekeeping_suspended);
1133 seq = read_seqbegin(&timekeeper.lock);
1134 *ts = timekeeper.xtime;
1135 tomono = timekeeper.wall_to_monotonic;
1136 sleep = timekeeper.total_sleep_time;
1137 nsecs = timekeeping_get_ns();
1139 } while (read_seqretry(&timekeeper.lock, seq));
1141 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1142 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1144 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1147 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1149 * Returns the monotonic time since boot in a ktime
1151 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1152 * includes the time spent in suspend.
1154 ktime_t ktime_get_boottime(void)
1158 get_monotonic_boottime(&ts);
1159 return timespec_to_ktime(ts);
1161 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1164 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1165 * @ts: pointer to the timespec to be converted
1167 void monotonic_to_bootbased(struct timespec *ts)
1169 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1171 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1173 unsigned long get_seconds(void)
1175 return timekeeper.xtime.tv_sec;
1177 EXPORT_SYMBOL(get_seconds);
1179 struct timespec __current_kernel_time(void)
1181 return timekeeper.xtime;
1184 struct timespec current_kernel_time(void)
1186 struct timespec now;
1190 seq = read_seqbegin(&timekeeper.lock);
1192 now = timekeeper.xtime;
1193 } while (read_seqretry(&timekeeper.lock, seq));
1197 EXPORT_SYMBOL(current_kernel_time);
1199 struct timespec get_monotonic_coarse(void)
1201 struct timespec now, mono;
1205 seq = read_seqbegin(&timekeeper.lock);
1207 now = timekeeper.xtime;
1208 mono = timekeeper.wall_to_monotonic;
1209 } while (read_seqretry(&timekeeper.lock, seq));
1211 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1212 now.tv_nsec + mono.tv_nsec);
1217 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1218 * without sampling the sequence number in xtime_lock.
1219 * jiffies is defined in the linker script...
1221 void do_timer(unsigned long ticks)
1223 jiffies_64 += ticks;
1225 calc_global_load(ticks);
1229 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1230 * and sleep offsets.
1231 * @xtim: pointer to timespec to be set with xtime
1232 * @wtom: pointer to timespec to be set with wall_to_monotonic
1233 * @sleep: pointer to timespec to be set with time in suspend
1235 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1236 struct timespec *wtom, struct timespec *sleep)
1241 seq = read_seqbegin(&timekeeper.lock);
1242 *xtim = timekeeper.xtime;
1243 *wtom = timekeeper.wall_to_monotonic;
1244 *sleep = timekeeper.total_sleep_time;
1245 } while (read_seqretry(&timekeeper.lock, seq));
1249 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1251 ktime_t ktime_get_monotonic_offset(void)
1254 struct timespec wtom;
1257 seq = read_seqbegin(&timekeeper.lock);
1258 wtom = timekeeper.wall_to_monotonic;
1259 } while (read_seqretry(&timekeeper.lock, seq));
1261 return timespec_to_ktime(wtom);
1263 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1267 * xtime_update() - advances the timekeeping infrastructure
1268 * @ticks: number of ticks, that have elapsed since the last call.
1270 * Must be called with interrupts disabled.
1272 void xtime_update(unsigned long ticks)
1274 write_seqlock(&xtime_lock);
1276 write_sequnlock(&xtime_lock);