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/sysdev.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 /* Raw nano seconds accumulated per NTP interval. */
38 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
40 /* Difference between accumulated time and NTP time in ntp
41 * shifted nano seconds. */
43 /* Shift conversion between clock shifted nano seconds and
44 * ntp shifted nano seconds. */
46 /* NTP adjusted clock multiplier */
50 struct timekeeper timekeeper;
53 * timekeeper_setup_internals - Set up internals to use clocksource clock.
55 * @clock: Pointer to clocksource.
57 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
58 * pair and interval request.
60 * Unless you're the timekeeping code, you should not be using this!
62 static void timekeeper_setup_internals(struct clocksource *clock)
67 timekeeper.clock = clock;
68 clock->cycle_last = clock->read(clock);
70 /* Do the ns -> cycle conversion first, using original mult */
71 tmp = NTP_INTERVAL_LENGTH;
74 do_div(tmp, clock->mult);
78 interval = (cycle_t) tmp;
79 timekeeper.cycle_interval = interval;
81 /* Go back from cycles -> shifted ns */
82 timekeeper.xtime_interval = (u64) interval * clock->mult;
83 timekeeper.raw_interval =
84 ((u64) interval * clock->mult) >> clock->shift;
86 timekeeper.xtime_nsec = 0;
87 timekeeper.shift = clock->shift;
89 timekeeper.ntp_error = 0;
90 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
93 * The timekeeper keeps its own mult values for the currently
94 * active clocksource. These value will be adjusted via NTP
95 * to counteract clock drifting.
97 timekeeper.mult = clock->mult;
100 /* Timekeeper helper functions. */
101 static inline s64 timekeeping_get_ns(void)
103 cycle_t cycle_now, cycle_delta;
104 struct clocksource *clock;
106 /* read clocksource: */
107 clock = timekeeper.clock;
108 cycle_now = clock->read(clock);
110 /* calculate the delta since the last update_wall_time: */
111 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
113 /* return delta convert to nanoseconds using ntp adjusted mult. */
114 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
118 static inline s64 timekeeping_get_ns_raw(void)
120 cycle_t cycle_now, cycle_delta;
121 struct clocksource *clock;
123 /* read clocksource: */
124 clock = timekeeper.clock;
125 cycle_now = clock->read(clock);
127 /* calculate the delta since the last update_wall_time: */
128 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
130 /* return delta convert to nanoseconds using ntp adjusted mult. */
131 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
135 * This read-write spinlock protects us from races in SMP while
136 * playing with xtime.
138 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
143 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
144 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
145 * at zero at system boot time, so wall_to_monotonic will be negative,
146 * however, we will ALWAYS keep the tv_nsec part positive so we can use
147 * the usual normalization.
149 * wall_to_monotonic is moved after resume from suspend for the monotonic
150 * time not to jump. We need to add total_sleep_time to wall_to_monotonic
151 * to get the real boot based time offset.
153 * - wall_to_monotonic is no longer the boot time, getboottime must be
156 static struct timespec xtime __attribute__ ((aligned (16)));
157 static struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
158 static struct timespec total_sleep_time;
161 * The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock.
163 struct timespec raw_time;
165 /* flag for if timekeeping is suspended */
166 int __read_mostly timekeeping_suspended;
168 /* must hold xtime_lock */
169 void timekeeping_leap_insert(int leapsecond)
171 xtime.tv_sec += leapsecond;
172 wall_to_monotonic.tv_sec -= leapsecond;
173 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
178 * timekeeping_forward_now - update clock to the current time
180 * Forward the current clock to update its state since the last call to
181 * update_wall_time(). This is useful before significant clock changes,
182 * as it avoids having to deal with this time offset explicitly.
184 static void timekeeping_forward_now(void)
186 cycle_t cycle_now, cycle_delta;
187 struct clocksource *clock;
190 clock = timekeeper.clock;
191 cycle_now = clock->read(clock);
192 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
193 clock->cycle_last = cycle_now;
195 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
198 /* If arch requires, add in gettimeoffset() */
199 nsec += arch_gettimeoffset();
201 timespec_add_ns(&xtime, nsec);
203 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
204 timespec_add_ns(&raw_time, nsec);
208 * getnstimeofday - Returns the time of day in a timespec
209 * @ts: pointer to the timespec to be set
211 * Returns the time of day in a timespec.
213 void getnstimeofday(struct timespec *ts)
218 WARN_ON(timekeeping_suspended);
221 seq = read_seqbegin(&xtime_lock);
224 nsecs = timekeeping_get_ns();
226 /* If arch requires, add in gettimeoffset() */
227 nsecs += arch_gettimeoffset();
229 } while (read_seqretry(&xtime_lock, seq));
231 timespec_add_ns(ts, nsecs);
234 EXPORT_SYMBOL(getnstimeofday);
236 ktime_t ktime_get(void)
241 WARN_ON(timekeeping_suspended);
244 seq = read_seqbegin(&xtime_lock);
245 secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
246 nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
247 nsecs += timekeeping_get_ns();
249 } while (read_seqretry(&xtime_lock, seq));
251 * Use ktime_set/ktime_add_ns to create a proper ktime on
252 * 32-bit architectures without CONFIG_KTIME_SCALAR.
254 return ktime_add_ns(ktime_set(secs, 0), nsecs);
256 EXPORT_SYMBOL_GPL(ktime_get);
259 * ktime_get_ts - get the monotonic clock in timespec format
260 * @ts: pointer to timespec variable
262 * The function calculates the monotonic clock from the realtime
263 * clock and the wall_to_monotonic offset and stores the result
264 * in normalized timespec format in the variable pointed to by @ts.
266 void ktime_get_ts(struct timespec *ts)
268 struct timespec tomono;
272 WARN_ON(timekeeping_suspended);
275 seq = read_seqbegin(&xtime_lock);
277 tomono = wall_to_monotonic;
278 nsecs = timekeeping_get_ns();
280 } while (read_seqretry(&xtime_lock, seq));
282 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
283 ts->tv_nsec + tomono.tv_nsec + nsecs);
285 EXPORT_SYMBOL_GPL(ktime_get_ts);
288 * do_gettimeofday - Returns the time of day in a timeval
289 * @tv: pointer to the timeval to be set
291 * NOTE: Users should be converted to using getnstimeofday()
293 void do_gettimeofday(struct timeval *tv)
297 getnstimeofday(&now);
298 tv->tv_sec = now.tv_sec;
299 tv->tv_usec = now.tv_nsec/1000;
302 EXPORT_SYMBOL(do_gettimeofday);
304 * do_settimeofday - Sets the time of day
305 * @tv: pointer to the timespec variable containing the new time
307 * Sets the time of day to the new time and update NTP and notify hrtimers
309 int do_settimeofday(struct timespec *tv)
311 struct timespec ts_delta;
314 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
317 write_seqlock_irqsave(&xtime_lock, flags);
319 timekeeping_forward_now();
321 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
322 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
323 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
327 timekeeper.ntp_error = 0;
330 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
333 write_sequnlock_irqrestore(&xtime_lock, flags);
335 /* signal hrtimers about time change */
341 EXPORT_SYMBOL(do_settimeofday);
344 * change_clocksource - Swaps clocksources if a new one is available
346 * Accumulates current time interval and initializes new clocksource
348 static int change_clocksource(void *data)
350 struct clocksource *new, *old;
352 new = (struct clocksource *) data;
354 timekeeping_forward_now();
355 if (!new->enable || new->enable(new) == 0) {
356 old = timekeeper.clock;
357 timekeeper_setup_internals(new);
365 * timekeeping_notify - Install a new clock source
366 * @clock: pointer to the clock source
368 * This function is called from clocksource.c after a new, better clock
369 * source has been registered. The caller holds the clocksource_mutex.
371 void timekeeping_notify(struct clocksource *clock)
373 if (timekeeper.clock == clock)
375 stop_machine(change_clocksource, clock, NULL);
380 * ktime_get_real - get the real (wall-) time in ktime_t format
382 * returns the time in ktime_t format
384 ktime_t ktime_get_real(void)
388 getnstimeofday(&now);
390 return timespec_to_ktime(now);
392 EXPORT_SYMBOL_GPL(ktime_get_real);
395 * getrawmonotonic - Returns the raw monotonic time in a timespec
396 * @ts: pointer to the timespec to be set
398 * Returns the raw monotonic time (completely un-modified by ntp)
400 void getrawmonotonic(struct timespec *ts)
406 seq = read_seqbegin(&xtime_lock);
407 nsecs = timekeeping_get_ns_raw();
410 } while (read_seqretry(&xtime_lock, seq));
412 timespec_add_ns(ts, nsecs);
414 EXPORT_SYMBOL(getrawmonotonic);
418 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
420 int timekeeping_valid_for_hres(void)
426 seq = read_seqbegin(&xtime_lock);
428 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
430 } while (read_seqretry(&xtime_lock, seq));
436 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
438 * Caller must observe xtime_lock via read_seqbegin/read_seqretry to
439 * ensure that the clocksource does not change!
441 u64 timekeeping_max_deferment(void)
443 return timekeeper.clock->max_idle_ns;
447 * read_persistent_clock - Return time from the persistent clock.
449 * Weak dummy function for arches that do not yet support it.
450 * Reads the time from the battery backed persistent clock.
451 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
453 * XXX - Do be sure to remove it once all arches implement it.
455 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
462 * read_boot_clock - Return time of the system start.
464 * Weak dummy function for arches that do not yet support it.
465 * Function to read the exact time the system has been started.
466 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
468 * XXX - Do be sure to remove it once all arches implement it.
470 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
477 * timekeeping_init - Initializes the clocksource and common timekeeping values
479 void __init timekeeping_init(void)
481 struct clocksource *clock;
483 struct timespec now, boot;
485 read_persistent_clock(&now);
486 read_boot_clock(&boot);
488 write_seqlock_irqsave(&xtime_lock, flags);
492 clock = clocksource_default_clock();
494 clock->enable(clock);
495 timekeeper_setup_internals(clock);
497 xtime.tv_sec = now.tv_sec;
498 xtime.tv_nsec = now.tv_nsec;
500 raw_time.tv_nsec = 0;
501 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
502 boot.tv_sec = xtime.tv_sec;
503 boot.tv_nsec = xtime.tv_nsec;
505 set_normalized_timespec(&wall_to_monotonic,
506 -boot.tv_sec, -boot.tv_nsec);
507 total_sleep_time.tv_sec = 0;
508 total_sleep_time.tv_nsec = 0;
509 write_sequnlock_irqrestore(&xtime_lock, flags);
512 /* time in seconds when suspend began */
513 static struct timespec timekeeping_suspend_time;
516 * timekeeping_resume - Resumes the generic timekeeping subsystem.
519 * This is for the generic clocksource timekeeping.
520 * xtime/wall_to_monotonic/jiffies/etc are
521 * still managed by arch specific suspend/resume code.
523 static int timekeeping_resume(struct sys_device *dev)
528 read_persistent_clock(&ts);
530 clocksource_resume();
532 write_seqlock_irqsave(&xtime_lock, flags);
534 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
535 ts = timespec_sub(ts, timekeeping_suspend_time);
536 xtime = timespec_add(xtime, ts);
537 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts);
538 total_sleep_time = timespec_add(total_sleep_time, ts);
540 /* re-base the last cycle value */
541 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
542 timekeeper.ntp_error = 0;
543 timekeeping_suspended = 0;
544 write_sequnlock_irqrestore(&xtime_lock, flags);
546 touch_softlockup_watchdog();
548 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
550 /* Resume hrtimers */
551 hres_timers_resume();
556 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
560 read_persistent_clock(&timekeeping_suspend_time);
562 write_seqlock_irqsave(&xtime_lock, flags);
563 timekeeping_forward_now();
564 timekeeping_suspended = 1;
565 write_sequnlock_irqrestore(&xtime_lock, flags);
567 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
568 clocksource_suspend();
573 /* sysfs resume/suspend bits for timekeeping */
574 static struct sysdev_class timekeeping_sysclass = {
575 .name = "timekeeping",
576 .resume = timekeeping_resume,
577 .suspend = timekeeping_suspend,
580 static struct sys_device device_timer = {
582 .cls = &timekeeping_sysclass,
585 static int __init timekeeping_init_device(void)
587 int error = sysdev_class_register(&timekeeping_sysclass);
589 error = sysdev_register(&device_timer);
593 device_initcall(timekeeping_init_device);
596 * If the error is already larger, we look ahead even further
597 * to compensate for late or lost adjustments.
599 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
607 * Use the current error value to determine how much to look ahead.
608 * The larger the error the slower we adjust for it to avoid problems
609 * with losing too many ticks, otherwise we would overadjust and
610 * produce an even larger error. The smaller the adjustment the
611 * faster we try to adjust for it, as lost ticks can do less harm
612 * here. This is tuned so that an error of about 1 msec is adjusted
613 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
615 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
616 error2 = abs(error2);
617 for (look_ahead = 0; error2 > 0; look_ahead++)
621 * Now calculate the error in (1 << look_ahead) ticks, but first
622 * remove the single look ahead already included in the error.
624 tick_error = tick_length >> (timekeeper.ntp_error_shift + 1);
625 tick_error -= timekeeper.xtime_interval >> 1;
626 error = ((error - tick_error) >> look_ahead) + tick_error;
628 /* Finally calculate the adjustment shift value. */
633 *interval = -*interval;
637 for (adj = 0; error > i; adj++)
646 * Adjust the multiplier to reduce the error value,
647 * this is optimized for the most common adjustments of -1,0,1,
648 * for other values we can do a bit more work.
650 static void timekeeping_adjust(s64 offset)
652 s64 error, interval = timekeeper.cycle_interval;
655 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
656 if (error > interval) {
658 if (likely(error <= interval))
661 adj = timekeeping_bigadjust(error, &interval, &offset);
662 } else if (error < -interval) {
664 if (likely(error >= -interval)) {
666 interval = -interval;
669 adj = timekeeping_bigadjust(error, &interval, &offset);
673 timekeeper.mult += adj;
674 timekeeper.xtime_interval += interval;
675 timekeeper.xtime_nsec -= offset;
676 timekeeper.ntp_error -= (interval - offset) <<
677 timekeeper.ntp_error_shift;
682 * logarithmic_accumulation - shifted accumulation of cycles
684 * This functions accumulates a shifted interval of cycles into
685 * into a shifted interval nanoseconds. Allows for O(log) accumulation
688 * Returns the unconsumed cycles.
690 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
692 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
694 /* If the offset is smaller then a shifted interval, do nothing */
695 if (offset < timekeeper.cycle_interval<<shift)
698 /* Accumulate one shifted interval */
699 offset -= timekeeper.cycle_interval << shift;
700 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
702 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
703 while (timekeeper.xtime_nsec >= nsecps) {
704 timekeeper.xtime_nsec -= nsecps;
709 /* Accumulate into raw time */
710 raw_time.tv_nsec += timekeeper.raw_interval << shift;;
711 while (raw_time.tv_nsec >= NSEC_PER_SEC) {
712 raw_time.tv_nsec -= NSEC_PER_SEC;
716 /* Accumulate error between NTP and clock interval */
717 timekeeper.ntp_error += tick_length << shift;
718 timekeeper.ntp_error -= timekeeper.xtime_interval <<
719 (timekeeper.ntp_error_shift + shift);
726 * update_wall_time - Uses the current clocksource to increment the wall time
728 * Called from the timer interrupt, must hold a write on xtime_lock.
730 void update_wall_time(void)
732 struct clocksource *clock;
734 int shift = 0, maxshift;
736 /* Make sure we're fully resumed: */
737 if (unlikely(timekeeping_suspended))
740 clock = timekeeper.clock;
742 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
743 offset = timekeeper.cycle_interval;
745 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
747 timekeeper.xtime_nsec = (s64)xtime.tv_nsec << timekeeper.shift;
750 * With NO_HZ we may have to accumulate many cycle_intervals
751 * (think "ticks") worth of time at once. To do this efficiently,
752 * we calculate the largest doubling multiple of cycle_intervals
753 * that is smaller then the offset. We then accumulate that
754 * chunk in one go, and then try to consume the next smaller
757 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
758 shift = max(0, shift);
759 /* Bound shift to one less then what overflows tick_length */
760 maxshift = (8*sizeof(tick_length) - (ilog2(tick_length)+1)) - 1;
761 shift = min(shift, maxshift);
762 while (offset >= timekeeper.cycle_interval) {
763 offset = logarithmic_accumulation(offset, shift);
764 if(offset < timekeeper.cycle_interval<<shift)
768 /* correct the clock when NTP error is too big */
769 timekeeping_adjust(offset);
772 * Since in the loop above, we accumulate any amount of time
773 * in xtime_nsec over a second into xtime.tv_sec, its possible for
774 * xtime_nsec to be fairly small after the loop. Further, if we're
775 * slightly speeding the clocksource up in timekeeping_adjust(),
776 * its possible the required corrective factor to xtime_nsec could
777 * cause it to underflow.
779 * Now, we cannot simply roll the accumulated second back, since
780 * the NTP subsystem has been notified via second_overflow. So
781 * instead we push xtime_nsec forward by the amount we underflowed,
782 * and add that amount into the error.
784 * We'll correct this error next time through this function, when
785 * xtime_nsec is not as small.
787 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
788 s64 neg = -(s64)timekeeper.xtime_nsec;
789 timekeeper.xtime_nsec = 0;
790 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
795 * Store full nanoseconds into xtime after rounding it up and
796 * add the remainder to the error difference.
798 xtime.tv_nsec = ((s64) timekeeper.xtime_nsec >> timekeeper.shift) + 1;
799 timekeeper.xtime_nsec -= (s64) xtime.tv_nsec << timekeeper.shift;
800 timekeeper.ntp_error += timekeeper.xtime_nsec <<
801 timekeeper.ntp_error_shift;
804 * Finally, make sure that after the rounding
805 * xtime.tv_nsec isn't larger then NSEC_PER_SEC
807 if (unlikely(xtime.tv_nsec >= NSEC_PER_SEC)) {
808 xtime.tv_nsec -= NSEC_PER_SEC;
813 /* check to see if there is a new clocksource to use */
814 update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
819 * getboottime - Return the real time of system boot.
820 * @ts: pointer to the timespec to be set
822 * Returns the time of day in a timespec.
824 * This is based on the wall_to_monotonic offset and the total suspend
825 * time. Calls to settimeofday will affect the value returned (which
826 * basically means that however wrong your real time clock is at boot time,
827 * you get the right time here).
829 void getboottime(struct timespec *ts)
831 struct timespec boottime = {
832 .tv_sec = wall_to_monotonic.tv_sec + total_sleep_time.tv_sec,
833 .tv_nsec = wall_to_monotonic.tv_nsec + total_sleep_time.tv_nsec
836 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
838 EXPORT_SYMBOL_GPL(getboottime);
841 * monotonic_to_bootbased - Convert the monotonic time to boot based.
842 * @ts: pointer to the timespec to be converted
844 void monotonic_to_bootbased(struct timespec *ts)
846 *ts = timespec_add(*ts, total_sleep_time);
848 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
850 unsigned long get_seconds(void)
854 EXPORT_SYMBOL(get_seconds);
856 struct timespec __current_kernel_time(void)
861 struct timespec __get_wall_to_monotonic(void)
863 return wall_to_monotonic;
866 struct timespec current_kernel_time(void)
872 seq = read_seqbegin(&xtime_lock);
875 } while (read_seqretry(&xtime_lock, seq));
879 EXPORT_SYMBOL(current_kernel_time);
881 struct timespec get_monotonic_coarse(void)
883 struct timespec now, mono;
887 seq = read_seqbegin(&xtime_lock);
890 mono = wall_to_monotonic;
891 } while (read_seqretry(&xtime_lock, seq));
893 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
894 now.tv_nsec + mono.tv_nsec);