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/sysdev.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/tick.h>
22 /* Structure holding internal timekeeping values. */
24 /* Current clocksource used for timekeeping. */
25 struct clocksource *clock;
27 /* Number of clock cycles in one NTP interval. */
28 cycle_t cycle_interval;
29 /* Number of clock shifted nano seconds in one NTP interval. */
31 /* Raw nano seconds accumulated per NTP interval. */
34 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
36 /* Difference between accumulated time and NTP time in ntp
37 * shifted nano seconds. */
41 struct timekeeper timekeeper;
44 * timekeeper_setup_internals - Set up internals to use clocksource clock.
46 * @clock: Pointer to clocksource.
48 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
49 * pair and interval request.
51 * Unless you're the timekeeping code, you should not be using this!
53 static void timekeeper_setup_internals(struct clocksource *clock)
58 timekeeper.clock = clock;
59 clock->cycle_last = clock->read(clock);
61 /* Do the ns -> cycle conversion first, using original mult */
62 tmp = NTP_INTERVAL_LENGTH;
64 tmp += clock->mult_orig/2;
65 do_div(tmp, clock->mult_orig);
69 interval = (cycle_t) tmp;
70 timekeeper.cycle_interval = interval;
72 /* Go back from cycles -> shifted ns */
73 timekeeper.xtime_interval = (u64) interval * clock->mult;
74 timekeeper.raw_interval =
75 ((u64) interval * clock->mult_orig) >> clock->shift;
77 timekeeper.xtime_nsec = 0;
79 timekeeper.ntp_error = 0;
83 * This read-write spinlock protects us from races in SMP while
86 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
91 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
92 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
93 * at zero at system boot time, so wall_to_monotonic will be negative,
94 * however, we will ALWAYS keep the tv_nsec part positive so we can use
95 * the usual normalization.
97 * wall_to_monotonic is moved after resume from suspend for the monotonic
98 * time not to jump. We need to add total_sleep_time to wall_to_monotonic
99 * to get the real boot based time offset.
101 * - wall_to_monotonic is no longer the boot time, getboottime must be
104 struct timespec xtime __attribute__ ((aligned (16)));
105 struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
106 static unsigned long total_sleep_time; /* seconds */
109 * The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock.
111 struct timespec raw_time;
113 /* flag for if timekeeping is suspended */
114 int __read_mostly timekeeping_suspended;
116 static struct timespec xtime_cache __attribute__ ((aligned (16)));
117 void update_xtime_cache(u64 nsec)
120 timespec_add_ns(&xtime_cache, nsec);
123 /* must hold xtime_lock */
124 void timekeeping_leap_insert(int leapsecond)
126 xtime.tv_sec += leapsecond;
127 wall_to_monotonic.tv_sec -= leapsecond;
128 update_vsyscall(&xtime, timekeeper.clock);
131 #ifdef CONFIG_GENERIC_TIME
133 * timekeeping_forward_now - update clock to the current time
135 * Forward the current clock to update its state since the last call to
136 * update_wall_time(). This is useful before significant clock changes,
137 * as it avoids having to deal with this time offset explicitly.
139 static void timekeeping_forward_now(void)
141 cycle_t cycle_now, cycle_delta;
142 struct clocksource *clock;
145 clock = timekeeper.clock;
146 cycle_now = clock->read(clock);
147 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
148 clock->cycle_last = cycle_now;
150 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
152 /* If arch requires, add in gettimeoffset() */
153 nsec += arch_gettimeoffset();
155 timespec_add_ns(&xtime, nsec);
157 nsec = clocksource_cyc2ns(cycle_delta, clock->mult_orig, clock->shift);
158 timespec_add_ns(&raw_time, nsec);
162 * getnstimeofday - Returns the time of day in a timespec
163 * @ts: pointer to the timespec to be set
165 * Returns the time of day in a timespec.
167 void getnstimeofday(struct timespec *ts)
169 cycle_t cycle_now, cycle_delta;
170 struct clocksource *clock;
174 WARN_ON(timekeeping_suspended);
177 seq = read_seqbegin(&xtime_lock);
181 /* read clocksource: */
182 clock = timekeeper.clock;
183 cycle_now = clock->read(clock);
185 /* calculate the delta since the last update_wall_time: */
186 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
188 /* convert to nanoseconds: */
189 nsecs = clocksource_cyc2ns(cycle_delta, clock->mult,
192 /* If arch requires, add in gettimeoffset() */
193 nsecs += arch_gettimeoffset();
195 } while (read_seqretry(&xtime_lock, seq));
197 timespec_add_ns(ts, nsecs);
200 EXPORT_SYMBOL(getnstimeofday);
202 ktime_t ktime_get(void)
204 cycle_t cycle_now, cycle_delta;
205 struct clocksource *clock;
209 WARN_ON(timekeeping_suspended);
212 seq = read_seqbegin(&xtime_lock);
213 secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
214 nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
216 /* read clocksource: */
217 clock = timekeeper.clock;
218 cycle_now = clock->read(clock);
220 /* calculate the delta since the last update_wall_time: */
221 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
223 /* convert to nanoseconds: */
224 nsecs += clocksource_cyc2ns(cycle_delta, clock->mult,
227 } while (read_seqretry(&xtime_lock, seq));
229 * Use ktime_set/ktime_add_ns to create a proper ktime on
230 * 32-bit architectures without CONFIG_KTIME_SCALAR.
232 return ktime_add_ns(ktime_set(secs, 0), nsecs);
234 EXPORT_SYMBOL_GPL(ktime_get);
237 * ktime_get_ts - get the monotonic clock in timespec format
238 * @ts: pointer to timespec variable
240 * The function calculates the monotonic clock from the realtime
241 * clock and the wall_to_monotonic offset and stores the result
242 * in normalized timespec format in the variable pointed to by @ts.
244 void ktime_get_ts(struct timespec *ts)
246 cycle_t cycle_now, cycle_delta;
247 struct clocksource *clock;
248 struct timespec tomono;
252 WARN_ON(timekeeping_suspended);
255 seq = read_seqbegin(&xtime_lock);
257 tomono = wall_to_monotonic;
259 /* read clocksource: */
260 clock = timekeeper.clock;
261 cycle_now = clock->read(clock);
263 /* calculate the delta since the last update_wall_time: */
264 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
266 /* convert to nanoseconds: */
267 nsecs = clocksource_cyc2ns(cycle_delta, clock->mult,
270 } while (read_seqretry(&xtime_lock, seq));
272 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
273 ts->tv_nsec + tomono.tv_nsec + nsecs);
275 EXPORT_SYMBOL_GPL(ktime_get_ts);
278 * do_gettimeofday - Returns the time of day in a timeval
279 * @tv: pointer to the timeval to be set
281 * NOTE: Users should be converted to using getnstimeofday()
283 void do_gettimeofday(struct timeval *tv)
287 getnstimeofday(&now);
288 tv->tv_sec = now.tv_sec;
289 tv->tv_usec = now.tv_nsec/1000;
292 EXPORT_SYMBOL(do_gettimeofday);
294 * do_settimeofday - Sets the time of day
295 * @tv: pointer to the timespec variable containing the new time
297 * Sets the time of day to the new time and update NTP and notify hrtimers
299 int do_settimeofday(struct timespec *tv)
301 struct timespec ts_delta;
304 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
307 write_seqlock_irqsave(&xtime_lock, flags);
309 timekeeping_forward_now();
311 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
312 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
313 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
317 update_xtime_cache(0);
319 timekeeper.ntp_error = 0;
322 update_vsyscall(&xtime, timekeeper.clock);
324 write_sequnlock_irqrestore(&xtime_lock, flags);
326 /* signal hrtimers about time change */
332 EXPORT_SYMBOL(do_settimeofday);
335 * change_clocksource - Swaps clocksources if a new one is available
337 * Accumulates current time interval and initializes new clocksource
339 static void change_clocksource(void)
341 struct clocksource *new, *old;
343 new = clocksource_get_next();
345 if (!new || timekeeper.clock == new)
348 timekeeping_forward_now();
350 if (new->enable && !new->enable(new))
353 * The frequency may have changed while the clocksource
354 * was disabled. If so the code in ->enable() must update
355 * the mult value to reflect the new frequency. Make sure
356 * mult_orig follows this change.
358 new->mult_orig = new->mult;
360 old = timekeeper.clock;
361 timekeeper_setup_internals(new);
364 * Save mult_orig in mult so that the value can be restored
365 * regardless if ->enable() updates the value of mult or not.
367 old->mult = old->mult_orig;
373 #else /* GENERIC_TIME */
374 static inline void timekeeping_forward_now(void) { }
375 static inline void change_clocksource(void) { }
378 * ktime_get - get the monotonic time in ktime_t format
380 * returns the time in ktime_t format
382 ktime_t ktime_get(void)
388 return timespec_to_ktime(now);
390 EXPORT_SYMBOL_GPL(ktime_get);
393 * ktime_get_ts - get the monotonic clock in timespec format
394 * @ts: pointer to timespec variable
396 * The function calculates the monotonic clock from the realtime
397 * clock and the wall_to_monotonic offset and stores the result
398 * in normalized timespec format in the variable pointed to by @ts.
400 void ktime_get_ts(struct timespec *ts)
402 struct timespec tomono;
406 seq = read_seqbegin(&xtime_lock);
408 tomono = wall_to_monotonic;
410 } while (read_seqretry(&xtime_lock, seq));
412 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
413 ts->tv_nsec + tomono.tv_nsec);
415 EXPORT_SYMBOL_GPL(ktime_get_ts);
416 #endif /* !GENERIC_TIME */
419 * ktime_get_real - get the real (wall-) time in ktime_t format
421 * returns the time in ktime_t format
423 ktime_t ktime_get_real(void)
427 getnstimeofday(&now);
429 return timespec_to_ktime(now);
431 EXPORT_SYMBOL_GPL(ktime_get_real);
434 * getrawmonotonic - Returns the raw monotonic time in a timespec
435 * @ts: pointer to the timespec to be set
437 * Returns the raw monotonic time (completely un-modified by ntp)
439 void getrawmonotonic(struct timespec *ts)
443 cycle_t cycle_now, cycle_delta;
444 struct clocksource *clock;
447 seq = read_seqbegin(&xtime_lock);
449 /* read clocksource: */
450 clock = timekeeper.clock;
451 cycle_now = clock->read(clock);
453 /* calculate the delta since the last update_wall_time: */
454 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
456 /* convert to nanoseconds: */
457 nsecs = clocksource_cyc2ns(cycle_delta, clock->mult_orig,
462 } while (read_seqretry(&xtime_lock, seq));
464 timespec_add_ns(ts, nsecs);
466 EXPORT_SYMBOL(getrawmonotonic);
470 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
472 int timekeeping_valid_for_hres(void)
478 seq = read_seqbegin(&xtime_lock);
480 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
482 } while (read_seqretry(&xtime_lock, seq));
488 * read_persistent_clock - Return time in seconds from the persistent clock.
490 * Weak dummy function for arches that do not yet support it.
491 * Returns seconds from epoch using the battery backed persistent clock.
492 * Returns zero if unsupported.
494 * XXX - Do be sure to remove it once all arches implement it.
496 unsigned long __attribute__((weak)) read_persistent_clock(void)
502 * timekeeping_init - Initializes the clocksource and common timekeeping values
504 void __init timekeeping_init(void)
506 struct clocksource *clock;
508 unsigned long sec = read_persistent_clock();
510 write_seqlock_irqsave(&xtime_lock, flags);
514 clock = clocksource_default_clock();
516 clock->enable(clock);
517 /* set mult_orig on enable */
518 clock->mult_orig = clock->mult;
520 timekeeper_setup_internals(clock);
525 raw_time.tv_nsec = 0;
526 set_normalized_timespec(&wall_to_monotonic,
527 -xtime.tv_sec, -xtime.tv_nsec);
528 update_xtime_cache(0);
529 total_sleep_time = 0;
530 write_sequnlock_irqrestore(&xtime_lock, flags);
533 /* time in seconds when suspend began */
534 static unsigned long timekeeping_suspend_time;
537 * timekeeping_resume - Resumes the generic timekeeping subsystem.
540 * This is for the generic clocksource timekeeping.
541 * xtime/wall_to_monotonic/jiffies/etc are
542 * still managed by arch specific suspend/resume code.
544 static int timekeeping_resume(struct sys_device *dev)
547 unsigned long now = read_persistent_clock();
549 clocksource_resume();
551 write_seqlock_irqsave(&xtime_lock, flags);
553 if (now && (now > timekeeping_suspend_time)) {
554 unsigned long sleep_length = now - timekeeping_suspend_time;
556 xtime.tv_sec += sleep_length;
557 wall_to_monotonic.tv_sec -= sleep_length;
558 total_sleep_time += sleep_length;
560 update_xtime_cache(0);
561 /* re-base the last cycle value */
562 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
563 timekeeper.ntp_error = 0;
564 timekeeping_suspended = 0;
565 write_sequnlock_irqrestore(&xtime_lock, flags);
567 touch_softlockup_watchdog();
569 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
571 /* Resume hrtimers */
572 hres_timers_resume();
577 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
581 timekeeping_suspend_time = read_persistent_clock();
583 write_seqlock_irqsave(&xtime_lock, flags);
584 timekeeping_forward_now();
585 timekeeping_suspended = 1;
586 write_sequnlock_irqrestore(&xtime_lock, flags);
588 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
593 /* sysfs resume/suspend bits for timekeeping */
594 static struct sysdev_class timekeeping_sysclass = {
595 .name = "timekeeping",
596 .resume = timekeeping_resume,
597 .suspend = timekeeping_suspend,
600 static struct sys_device device_timer = {
602 .cls = &timekeeping_sysclass,
605 static int __init timekeeping_init_device(void)
607 int error = sysdev_class_register(&timekeeping_sysclass);
609 error = sysdev_register(&device_timer);
613 device_initcall(timekeeping_init_device);
616 * If the error is already larger, we look ahead even further
617 * to compensate for late or lost adjustments.
619 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
627 * Use the current error value to determine how much to look ahead.
628 * The larger the error the slower we adjust for it to avoid problems
629 * with losing too many ticks, otherwise we would overadjust and
630 * produce an even larger error. The smaller the adjustment the
631 * faster we try to adjust for it, as lost ticks can do less harm
632 * here. This is tuned so that an error of about 1 msec is adjusted
633 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
635 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
636 error2 = abs(error2);
637 for (look_ahead = 0; error2 > 0; look_ahead++)
641 * Now calculate the error in (1 << look_ahead) ticks, but first
642 * remove the single look ahead already included in the error.
644 tick_error = tick_length >>
645 (NTP_SCALE_SHIFT - timekeeper.clock->shift + 1);
646 tick_error -= timekeeper.xtime_interval >> 1;
647 error = ((error - tick_error) >> look_ahead) + tick_error;
649 /* Finally calculate the adjustment shift value. */
654 *interval = -*interval;
658 for (adj = 0; error > i; adj++)
667 * Adjust the multiplier to reduce the error value,
668 * this is optimized for the most common adjustments of -1,0,1,
669 * for other values we can do a bit more work.
671 static void timekeeping_adjust(s64 offset)
673 s64 error, interval = timekeeper.cycle_interval;
676 error = timekeeper.ntp_error >>
677 (NTP_SCALE_SHIFT - timekeeper.clock->shift - 1);
678 if (error > interval) {
680 if (likely(error <= interval))
683 adj = timekeeping_bigadjust(error, &interval, &offset);
684 } else if (error < -interval) {
686 if (likely(error >= -interval)) {
688 interval = -interval;
691 adj = timekeeping_bigadjust(error, &interval, &offset);
695 timekeeper.clock->mult += adj;
696 timekeeper.xtime_interval += interval;
697 timekeeper.xtime_nsec -= offset;
698 timekeeper.ntp_error -= (interval - offset) <<
699 (NTP_SCALE_SHIFT - timekeeper.clock->shift);
703 * update_wall_time - Uses the current clocksource to increment the wall time
705 * Called from the timer interrupt, must hold a write on xtime_lock.
707 void update_wall_time(void)
709 struct clocksource *clock;
713 /* Make sure we're fully resumed: */
714 if (unlikely(timekeeping_suspended))
717 clock = timekeeper.clock;
718 #ifdef CONFIG_GENERIC_TIME
719 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
721 offset = timekeeper.cycle_interval;
723 timekeeper.xtime_nsec = (s64)xtime.tv_nsec << clock->shift;
725 /* normally this loop will run just once, however in the
726 * case of lost or late ticks, it will accumulate correctly.
728 while (offset >= timekeeper.cycle_interval) {
729 u64 nsecps = (u64)NSEC_PER_SEC << clock->shift;
731 /* accumulate one interval */
732 offset -= timekeeper.cycle_interval;
733 clock->cycle_last += timekeeper.cycle_interval;
735 timekeeper.xtime_nsec += timekeeper.xtime_interval;
736 if (timekeeper.xtime_nsec >= nsecps) {
737 timekeeper.xtime_nsec -= nsecps;
742 raw_time.tv_nsec += timekeeper.raw_interval;
743 if (raw_time.tv_nsec >= NSEC_PER_SEC) {
744 raw_time.tv_nsec -= NSEC_PER_SEC;
748 /* accumulate error between NTP and clock interval */
749 timekeeper.ntp_error += tick_length;
750 timekeeper.ntp_error -= timekeeper.xtime_interval <<
751 (NTP_SCALE_SHIFT - clock->shift);
754 /* correct the clock when NTP error is too big */
755 timekeeping_adjust(offset);
758 * Since in the loop above, we accumulate any amount of time
759 * in xtime_nsec over a second into xtime.tv_sec, its possible for
760 * xtime_nsec to be fairly small after the loop. Further, if we're
761 * slightly speeding the clocksource up in timekeeping_adjust(),
762 * its possible the required corrective factor to xtime_nsec could
763 * cause it to underflow.
765 * Now, we cannot simply roll the accumulated second back, since
766 * the NTP subsystem has been notified via second_overflow. So
767 * instead we push xtime_nsec forward by the amount we underflowed,
768 * and add that amount into the error.
770 * We'll correct this error next time through this function, when
771 * xtime_nsec is not as small.
773 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
774 s64 neg = -(s64)timekeeper.xtime_nsec;
775 timekeeper.xtime_nsec = 0;
776 timekeeper.ntp_error += neg << (NTP_SCALE_SHIFT - clock->shift);
779 /* store full nanoseconds into xtime after rounding it up and
780 * add the remainder to the error difference.
782 xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >> clock->shift) + 1;
783 timekeeper.xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
784 timekeeper.ntp_error += timekeeper.xtime_nsec <<
785 (NTP_SCALE_SHIFT - clock->shift);
787 nsecs = clocksource_cyc2ns(offset, clock->mult, clock->shift);
788 update_xtime_cache(nsecs);
790 /* check to see if there is a new clocksource to use */
791 change_clocksource();
792 update_vsyscall(&xtime, timekeeper.clock);
796 * getboottime - Return the real time of system boot.
797 * @ts: pointer to the timespec to be set
799 * Returns the time of day in a timespec.
801 * This is based on the wall_to_monotonic offset and the total suspend
802 * time. Calls to settimeofday will affect the value returned (which
803 * basically means that however wrong your real time clock is at boot time,
804 * you get the right time here).
806 void getboottime(struct timespec *ts)
808 set_normalized_timespec(ts,
809 - (wall_to_monotonic.tv_sec + total_sleep_time),
810 - wall_to_monotonic.tv_nsec);
814 * monotonic_to_bootbased - Convert the monotonic time to boot based.
815 * @ts: pointer to the timespec to be converted
817 void monotonic_to_bootbased(struct timespec *ts)
819 ts->tv_sec += total_sleep_time;
822 unsigned long get_seconds(void)
824 return xtime_cache.tv_sec;
826 EXPORT_SYMBOL(get_seconds);
829 struct timespec current_kernel_time(void)
835 seq = read_seqbegin(&xtime_lock);
838 } while (read_seqretry(&xtime_lock, seq));
842 EXPORT_SYMBOL(current_kernel_time);