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. */
32 /* Number of clock cycles in one NTP interval. */
33 cycle_t cycle_interval;
34 /* Number of clock shifted nano seconds in one NTP interval. */
36 /* shifted nano seconds left over when rounding cycle_interval */
38 /* Raw nano seconds accumulated per NTP interval. */
41 /* Current CLOCK_REALTIME time in seconds */
43 /* Clock shifted nano seconds */
46 /* Difference between accumulated time and NTP time in ntp
47 * shifted nano seconds. */
49 /* Shift conversion between clock shifted nano seconds and
50 * ntp shifted nano seconds. */
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 /* Offset clock monotonic -> clock realtime */
70 /* time spent in suspend */
71 struct timespec total_sleep_time;
72 /* Offset clock monotonic -> clock boottime */
74 /* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. */
75 struct timespec raw_time;
76 /* Seqlock for all timekeeper values */
80 static struct timekeeper timekeeper;
83 * This read-write spinlock protects us from races in SMP while
86 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
88 /* flag for if timekeeping is suspended */
89 int __read_mostly timekeeping_suspended;
91 static inline void tk_normalize_xtime(struct timekeeper *tk)
93 while (tk->xtime_nsec >= ((u64)NSEC_PER_SEC << tk->shift)) {
94 tk->xtime_nsec -= (u64)NSEC_PER_SEC << tk->shift;
99 static struct timespec tk_xtime(struct timekeeper *tk)
103 ts.tv_sec = tk->xtime_sec;
104 ts.tv_nsec = (long)(tk->xtime_nsec >> tk->shift);
108 static void tk_set_xtime(struct timekeeper *tk, const struct timespec *ts)
110 tk->xtime_sec = ts->tv_sec;
111 tk->xtime_nsec = (u64)ts->tv_nsec << tk->shift;
114 static void tk_xtime_add(struct timekeeper *tk, const struct timespec *ts)
116 tk->xtime_sec += ts->tv_sec;
117 tk->xtime_nsec += (u64)ts->tv_nsec << tk->shift;
118 tk_normalize_xtime(tk);
121 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec wtm)
126 * Verify consistency of: offset_real = -wall_to_monotonic
127 * before modifying anything
129 set_normalized_timespec(&tmp, -tk->wall_to_monotonic.tv_sec,
130 -tk->wall_to_monotonic.tv_nsec);
131 WARN_ON_ONCE(tk->offs_real.tv64 != timespec_to_ktime(tmp).tv64);
132 tk->wall_to_monotonic = wtm;
133 set_normalized_timespec(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
134 tk->offs_real = timespec_to_ktime(tmp);
137 static void tk_set_sleep_time(struct timekeeper *tk, struct timespec t)
139 /* Verify consistency before modifying */
140 WARN_ON_ONCE(tk->offs_boot.tv64 != timespec_to_ktime(tk->total_sleep_time).tv64);
142 tk->total_sleep_time = t;
143 tk->offs_boot = timespec_to_ktime(t);
147 * timekeeper_setup_internals - Set up internals to use clocksource clock.
149 * @clock: Pointer to clocksource.
151 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
152 * pair and interval request.
154 * Unless you're the timekeeping code, you should not be using this!
156 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
159 u64 tmp, ntpinterval;
160 struct clocksource *old_clock;
162 old_clock = tk->clock;
164 clock->cycle_last = clock->read(clock);
166 /* Do the ns -> cycle conversion first, using original mult */
167 tmp = NTP_INTERVAL_LENGTH;
168 tmp <<= clock->shift;
170 tmp += clock->mult/2;
171 do_div(tmp, clock->mult);
175 interval = (cycle_t) tmp;
176 tk->cycle_interval = interval;
178 /* Go back from cycles -> shifted ns */
179 tk->xtime_interval = (u64) interval * clock->mult;
180 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
182 ((u64) interval * clock->mult) >> clock->shift;
184 /* if changing clocks, convert xtime_nsec shift units */
186 int shift_change = clock->shift - old_clock->shift;
187 if (shift_change < 0)
188 tk->xtime_nsec >>= -shift_change;
190 tk->xtime_nsec <<= shift_change;
192 tk->shift = clock->shift;
195 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
198 * The timekeeper keeps its own mult values for the currently
199 * active clocksource. These value will be adjusted via NTP
200 * to counteract clock drifting.
202 tk->mult = clock->mult;
205 /* Timekeeper helper functions. */
206 static inline s64 timekeeping_get_ns(struct timekeeper *tk)
208 cycle_t cycle_now, cycle_delta;
209 struct clocksource *clock;
212 /* read clocksource: */
214 cycle_now = clock->read(clock);
216 /* calculate the delta since the last update_wall_time: */
217 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
219 nsec = cycle_delta * tk->mult + tk->xtime_nsec;
222 /* If arch requires, add in gettimeoffset() */
223 return nsec + arch_gettimeoffset();
226 static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
228 cycle_t cycle_now, cycle_delta;
229 struct clocksource *clock;
232 /* read clocksource: */
234 cycle_now = clock->read(clock);
236 /* calculate the delta since the last update_wall_time: */
237 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
239 /* convert delta to nanoseconds. */
240 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
242 /* If arch requires, add in gettimeoffset() */
243 return nsec + arch_gettimeoffset();
246 /* must hold write on timekeeper.lock */
247 static void timekeeping_update(struct timekeeper *tk, bool clearntp)
256 update_vsyscall(&xt, &tk->wall_to_monotonic, tk->clock, tk->mult);
260 * timekeeping_forward_now - update clock to the current time
262 * Forward the current clock to update its state since the last call to
263 * update_wall_time(). This is useful before significant clock changes,
264 * as it avoids having to deal with this time offset explicitly.
266 static void timekeeping_forward_now(struct timekeeper *tk)
268 cycle_t cycle_now, cycle_delta;
269 struct clocksource *clock;
273 cycle_now = clock->read(clock);
274 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
275 clock->cycle_last = cycle_now;
277 tk->xtime_nsec += cycle_delta * tk->mult;
279 /* If arch requires, add in gettimeoffset() */
280 tk->xtime_nsec += (u64)arch_gettimeoffset() << tk->shift;
282 tk_normalize_xtime(tk);
284 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
285 timespec_add_ns(&tk->raw_time, nsec);
289 * getnstimeofday - Returns the time of day in a timespec
290 * @ts: pointer to the timespec to be set
292 * Returns the time of day in a timespec.
294 void getnstimeofday(struct timespec *ts)
296 struct timekeeper *tk = &timekeeper;
300 WARN_ON(timekeeping_suspended);
303 seq = read_seqbegin(&tk->lock);
305 ts->tv_sec = tk->xtime_sec;
306 nsecs = timekeeping_get_ns(tk);
308 } while (read_seqretry(&tk->lock, seq));
311 timespec_add_ns(ts, nsecs);
313 EXPORT_SYMBOL(getnstimeofday);
315 ktime_t ktime_get(void)
317 struct timekeeper *tk = &timekeeper;
321 WARN_ON(timekeeping_suspended);
324 seq = read_seqbegin(&tk->lock);
325 secs = tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
326 nsecs = timekeeping_get_ns(tk) + tk->wall_to_monotonic.tv_nsec;
328 } while (read_seqretry(&tk->lock, seq));
330 * Use ktime_set/ktime_add_ns to create a proper ktime on
331 * 32-bit architectures without CONFIG_KTIME_SCALAR.
333 return ktime_add_ns(ktime_set(secs, 0), nsecs);
335 EXPORT_SYMBOL_GPL(ktime_get);
338 * ktime_get_ts - get the monotonic clock in timespec format
339 * @ts: pointer to timespec variable
341 * The function calculates the monotonic clock from the realtime
342 * clock and the wall_to_monotonic offset and stores the result
343 * in normalized timespec format in the variable pointed to by @ts.
345 void ktime_get_ts(struct timespec *ts)
347 struct timekeeper *tk = &timekeeper;
348 struct timespec tomono;
352 WARN_ON(timekeeping_suspended);
355 seq = read_seqbegin(&tk->lock);
356 ts->tv_sec = tk->xtime_sec;
357 nsec = timekeeping_get_ns(tk);
358 tomono = tk->wall_to_monotonic;
360 } while (read_seqretry(&tk->lock, seq));
362 ts->tv_sec += tomono.tv_sec;
364 timespec_add_ns(ts, nsec + tomono.tv_nsec);
366 EXPORT_SYMBOL_GPL(ktime_get_ts);
368 #ifdef CONFIG_NTP_PPS
371 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
372 * @ts_raw: pointer to the timespec to be set to raw monotonic time
373 * @ts_real: pointer to the timespec to be set to the time of day
375 * This function reads both the time of day and raw monotonic time at the
376 * same time atomically and stores the resulting timestamps in timespec
379 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
381 struct timekeeper *tk = &timekeeper;
383 s64 nsecs_raw, nsecs_real;
385 WARN_ON_ONCE(timekeeping_suspended);
388 seq = read_seqbegin(&tk->lock);
390 *ts_raw = tk->raw_time;
391 ts_real->tv_sec = tk->xtime_sec;
392 ts_real->tv_nsec = 0;
394 nsecs_raw = timekeeping_get_ns_raw(tk);
395 nsecs_real = timekeeping_get_ns(tk);
397 } while (read_seqretry(&tk->lock, seq));
399 timespec_add_ns(ts_raw, nsecs_raw);
400 timespec_add_ns(ts_real, nsecs_real);
402 EXPORT_SYMBOL(getnstime_raw_and_real);
404 #endif /* CONFIG_NTP_PPS */
407 * do_gettimeofday - Returns the time of day in a timeval
408 * @tv: pointer to the timeval to be set
410 * NOTE: Users should be converted to using getnstimeofday()
412 void do_gettimeofday(struct timeval *tv)
416 getnstimeofday(&now);
417 tv->tv_sec = now.tv_sec;
418 tv->tv_usec = now.tv_nsec/1000;
420 EXPORT_SYMBOL(do_gettimeofday);
423 * do_settimeofday - Sets the time of day
424 * @tv: pointer to the timespec variable containing the new time
426 * Sets the time of day to the new time and update NTP and notify hrtimers
428 int do_settimeofday(const struct timespec *tv)
430 struct timekeeper *tk = &timekeeper;
431 struct timespec ts_delta, xt;
434 if (!timespec_valid_strict(tv))
437 write_seqlock_irqsave(&tk->lock, flags);
439 timekeeping_forward_now(tk);
442 ts_delta.tv_sec = tv->tv_sec - xt.tv_sec;
443 ts_delta.tv_nsec = tv->tv_nsec - xt.tv_nsec;
445 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, ts_delta));
447 tk_set_xtime(tk, tv);
449 timekeeping_update(tk, true);
451 write_sequnlock_irqrestore(&tk->lock, flags);
453 /* signal hrtimers about time change */
458 EXPORT_SYMBOL(do_settimeofday);
461 * timekeeping_inject_offset - Adds or subtracts from the current time.
462 * @tv: pointer to the timespec variable containing the offset
464 * Adds or subtracts an offset value from the current time.
466 int timekeeping_inject_offset(struct timespec *ts)
468 struct timekeeper *tk = &timekeeper;
473 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
476 write_seqlock_irqsave(&tk->lock, flags);
478 timekeeping_forward_now(tk);
480 /* Make sure the proposed value is valid */
481 tmp = timespec_add(tk_xtime(tk), *ts);
482 if (!timespec_valid_strict(&tmp)) {
487 tk_xtime_add(tk, ts);
488 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *ts));
490 error: /* even if we error out, we forwarded the time, so call update */
491 timekeeping_update(tk, true);
493 write_sequnlock_irqrestore(&tk->lock, flags);
495 /* signal hrtimers about time change */
500 EXPORT_SYMBOL(timekeeping_inject_offset);
503 * change_clocksource - Swaps clocksources if a new one is available
505 * Accumulates current time interval and initializes new clocksource
507 static int change_clocksource(void *data)
509 struct timekeeper *tk = &timekeeper;
510 struct clocksource *new, *old;
513 new = (struct clocksource *) data;
515 write_seqlock_irqsave(&tk->lock, flags);
517 timekeeping_forward_now(tk);
518 if (!new->enable || new->enable(new) == 0) {
520 tk_setup_internals(tk, new);
524 timekeeping_update(tk, true);
526 write_sequnlock_irqrestore(&tk->lock, flags);
532 * timekeeping_notify - Install a new clock source
533 * @clock: pointer to the clock source
535 * This function is called from clocksource.c after a new, better clock
536 * source has been registered. The caller holds the clocksource_mutex.
538 void timekeeping_notify(struct clocksource *clock)
540 struct timekeeper *tk = &timekeeper;
542 if (tk->clock == clock)
544 stop_machine(change_clocksource, clock, NULL);
549 * ktime_get_real - get the real (wall-) time in ktime_t format
551 * returns the time in ktime_t format
553 ktime_t ktime_get_real(void)
557 getnstimeofday(&now);
559 return timespec_to_ktime(now);
561 EXPORT_SYMBOL_GPL(ktime_get_real);
564 * getrawmonotonic - Returns the raw monotonic time in a timespec
565 * @ts: pointer to the timespec to be set
567 * Returns the raw monotonic time (completely un-modified by ntp)
569 void getrawmonotonic(struct timespec *ts)
571 struct timekeeper *tk = &timekeeper;
576 seq = read_seqbegin(&tk->lock);
577 nsecs = timekeeping_get_ns_raw(tk);
580 } while (read_seqretry(&tk->lock, seq));
582 timespec_add_ns(ts, nsecs);
584 EXPORT_SYMBOL(getrawmonotonic);
587 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
589 int timekeeping_valid_for_hres(void)
591 struct timekeeper *tk = &timekeeper;
596 seq = read_seqbegin(&tk->lock);
598 ret = tk->clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
600 } while (read_seqretry(&tk->lock, seq));
606 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
608 u64 timekeeping_max_deferment(void)
610 struct timekeeper *tk = &timekeeper;
615 seq = read_seqbegin(&tk->lock);
617 ret = tk->clock->max_idle_ns;
619 } while (read_seqretry(&tk->lock, seq));
625 * read_persistent_clock - Return time from the persistent clock.
627 * Weak dummy function for arches that do not yet support it.
628 * Reads the time from the battery backed persistent clock.
629 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
631 * XXX - Do be sure to remove it once all arches implement it.
633 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
640 * read_boot_clock - Return time of the system start.
642 * Weak dummy function for arches that do not yet support it.
643 * Function to read the exact time the system has been started.
644 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
646 * XXX - Do be sure to remove it once all arches implement it.
648 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
655 * timekeeping_init - Initializes the clocksource and common timekeeping values
657 void __init timekeeping_init(void)
659 struct timekeeper *tk = &timekeeper;
660 struct clocksource *clock;
662 struct timespec now, boot, tmp;
664 read_persistent_clock(&now);
665 if (!timespec_valid_strict(&now)) {
666 pr_warn("WARNING: Persistent clock returned invalid value!\n"
667 " Check your CMOS/BIOS settings.\n");
672 read_boot_clock(&boot);
673 if (!timespec_valid_strict(&boot)) {
674 pr_warn("WARNING: Boot clock returned invalid value!\n"
675 " Check your CMOS/BIOS settings.\n");
680 seqlock_init(&tk->lock);
684 write_seqlock_irqsave(&tk->lock, flags);
685 clock = clocksource_default_clock();
687 clock->enable(clock);
688 tk_setup_internals(tk, clock);
690 tk_set_xtime(tk, &now);
691 tk->raw_time.tv_sec = 0;
692 tk->raw_time.tv_nsec = 0;
693 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
696 set_normalized_timespec(&tmp, -boot.tv_sec, -boot.tv_nsec);
697 tk_set_wall_to_mono(tk, tmp);
701 tk_set_sleep_time(tk, tmp);
703 write_sequnlock_irqrestore(&tk->lock, flags);
706 /* time in seconds when suspend began */
707 static struct timespec timekeeping_suspend_time;
710 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
711 * @delta: pointer to a timespec delta value
713 * Takes a timespec offset measuring a suspend interval and properly
714 * adds the sleep offset to the timekeeping variables.
716 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
717 struct timespec *delta)
719 if (!timespec_valid_strict(delta)) {
720 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
721 "sleep delta value!\n");
724 tk_xtime_add(tk, delta);
725 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *delta));
726 tk_set_sleep_time(tk, timespec_add(tk->total_sleep_time, *delta));
730 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
731 * @delta: pointer to a timespec delta value
733 * This hook is for architectures that cannot support read_persistent_clock
734 * because their RTC/persistent clock is only accessible when irqs are enabled.
736 * This function should only be called by rtc_resume(), and allows
737 * a suspend offset to be injected into the timekeeping values.
739 void timekeeping_inject_sleeptime(struct timespec *delta)
741 struct timekeeper *tk = &timekeeper;
745 /* Make sure we don't set the clock twice */
746 read_persistent_clock(&ts);
747 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
750 write_seqlock_irqsave(&tk->lock, flags);
752 timekeeping_forward_now(tk);
754 __timekeeping_inject_sleeptime(tk, delta);
756 timekeeping_update(tk, true);
758 write_sequnlock_irqrestore(&tk->lock, flags);
760 /* signal hrtimers about time change */
765 * timekeeping_resume - Resumes the generic timekeeping subsystem.
767 * This is for the generic clocksource timekeeping.
768 * xtime/wall_to_monotonic/jiffies/etc are
769 * still managed by arch specific suspend/resume code.
771 static void timekeeping_resume(void)
773 struct timekeeper *tk = &timekeeper;
777 read_persistent_clock(&ts);
779 clocksource_resume();
781 write_seqlock_irqsave(&tk->lock, flags);
783 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
784 ts = timespec_sub(ts, timekeeping_suspend_time);
785 __timekeeping_inject_sleeptime(tk, &ts);
787 /* re-base the last cycle value */
788 tk->clock->cycle_last = tk->clock->read(tk->clock);
790 timekeeping_suspended = 0;
791 timekeeping_update(tk, false);
792 write_sequnlock_irqrestore(&tk->lock, flags);
794 touch_softlockup_watchdog();
796 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
798 /* Resume hrtimers */
802 static int timekeeping_suspend(void)
804 struct timekeeper *tk = &timekeeper;
806 struct timespec delta, delta_delta;
807 static struct timespec old_delta;
809 read_persistent_clock(&timekeeping_suspend_time);
811 write_seqlock_irqsave(&tk->lock, flags);
812 timekeeping_forward_now(tk);
813 timekeeping_suspended = 1;
816 * To avoid drift caused by repeated suspend/resumes,
817 * which each can add ~1 second drift error,
818 * try to compensate so the difference in system time
819 * and persistent_clock time stays close to constant.
821 delta = timespec_sub(tk_xtime(tk), timekeeping_suspend_time);
822 delta_delta = timespec_sub(delta, old_delta);
823 if (abs(delta_delta.tv_sec) >= 2) {
825 * if delta_delta is too large, assume time correction
826 * has occured and set old_delta to the current delta.
830 /* Otherwise try to adjust old_system to compensate */
831 timekeeping_suspend_time =
832 timespec_add(timekeeping_suspend_time, delta_delta);
834 write_sequnlock_irqrestore(&tk->lock, flags);
836 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
837 clocksource_suspend();
842 /* sysfs resume/suspend bits for timekeeping */
843 static struct syscore_ops timekeeping_syscore_ops = {
844 .resume = timekeeping_resume,
845 .suspend = timekeeping_suspend,
848 static int __init timekeeping_init_ops(void)
850 register_syscore_ops(&timekeeping_syscore_ops);
854 device_initcall(timekeeping_init_ops);
857 * If the error is already larger, we look ahead even further
858 * to compensate for late or lost adjustments.
860 static __always_inline int timekeeping_bigadjust(struct timekeeper *tk,
861 s64 error, s64 *interval,
869 * Use the current error value to determine how much to look ahead.
870 * The larger the error the slower we adjust for it to avoid problems
871 * with losing too many ticks, otherwise we would overadjust and
872 * produce an even larger error. The smaller the adjustment the
873 * faster we try to adjust for it, as lost ticks can do less harm
874 * here. This is tuned so that an error of about 1 msec is adjusted
875 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
877 error2 = tk->ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
878 error2 = abs(error2);
879 for (look_ahead = 0; error2 > 0; look_ahead++)
883 * Now calculate the error in (1 << look_ahead) ticks, but first
884 * remove the single look ahead already included in the error.
886 tick_error = ntp_tick_length() >> (tk->ntp_error_shift + 1);
887 tick_error -= tk->xtime_interval >> 1;
888 error = ((error - tick_error) >> look_ahead) + tick_error;
890 /* Finally calculate the adjustment shift value. */
895 *interval = -*interval;
899 for (adj = 0; error > i; adj++)
908 * Adjust the multiplier to reduce the error value,
909 * this is optimized for the most common adjustments of -1,0,1,
910 * for other values we can do a bit more work.
912 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
914 s64 error, interval = tk->cycle_interval;
918 * The point of this is to check if the error is greater than half
921 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
923 * Note we subtract one in the shift, so that error is really error*2.
924 * This "saves" dividing(shifting) interval twice, but keeps the
925 * (error > interval) comparison as still measuring if error is
926 * larger than half an interval.
928 * Note: It does not "save" on aggravation when reading the code.
930 error = tk->ntp_error >> (tk->ntp_error_shift - 1);
931 if (error > interval) {
933 * We now divide error by 4(via shift), which checks if
934 * the error is greater than twice the interval.
935 * If it is greater, we need a bigadjust, if its smaller,
936 * we can adjust by 1.
940 * XXX - In update_wall_time, we round up to the next
941 * nanosecond, and store the amount rounded up into
942 * the error. This causes the likely below to be unlikely.
944 * The proper fix is to avoid rounding up by using
945 * the high precision tk->xtime_nsec instead of
946 * xtime.tv_nsec everywhere. Fixing this will take some
949 if (likely(error <= interval))
952 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
954 if (error < -interval) {
955 /* See comment above, this is just switched for the negative */
957 if (likely(error >= -interval)) {
959 interval = -interval;
962 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
969 if (unlikely(tk->clock->maxadj &&
970 (tk->mult + adj > tk->clock->mult + tk->clock->maxadj))) {
971 printk_once(KERN_WARNING
972 "Adjusting %s more than 11%% (%ld vs %ld)\n",
973 tk->clock->name, (long)tk->mult + adj,
974 (long)tk->clock->mult + tk->clock->maxadj);
977 * So the following can be confusing.
979 * To keep things simple, lets assume adj == 1 for now.
981 * When adj != 1, remember that the interval and offset values
982 * have been appropriately scaled so the math is the same.
984 * The basic idea here is that we're increasing the multiplier
985 * by one, this causes the xtime_interval to be incremented by
986 * one cycle_interval. This is because:
987 * xtime_interval = cycle_interval * mult
988 * So if mult is being incremented by one:
989 * xtime_interval = cycle_interval * (mult + 1)
991 * xtime_interval = (cycle_interval * mult) + cycle_interval
992 * Which can be shortened to:
993 * xtime_interval += cycle_interval
995 * So offset stores the non-accumulated cycles. Thus the current
996 * time (in shifted nanoseconds) is:
997 * now = (offset * adj) + xtime_nsec
998 * Now, even though we're adjusting the clock frequency, we have
999 * to keep time consistent. In other words, we can't jump back
1000 * in time, and we also want to avoid jumping forward in time.
1002 * So given the same offset value, we need the time to be the same
1003 * both before and after the freq adjustment.
1004 * now = (offset * adj_1) + xtime_nsec_1
1005 * now = (offset * adj_2) + xtime_nsec_2
1007 * (offset * adj_1) + xtime_nsec_1 =
1008 * (offset * adj_2) + xtime_nsec_2
1012 * (offset * adj_1) + xtime_nsec_1 =
1013 * (offset * (adj_1+1)) + xtime_nsec_2
1014 * (offset * adj_1) + xtime_nsec_1 =
1015 * (offset * adj_1) + offset + xtime_nsec_2
1016 * Canceling the sides:
1017 * xtime_nsec_1 = offset + xtime_nsec_2
1019 * xtime_nsec_2 = xtime_nsec_1 - offset
1020 * Which simplfies to:
1021 * xtime_nsec -= offset
1023 * XXX - TODO: Doc ntp_error calculation.
1026 tk->xtime_interval += interval;
1027 tk->xtime_nsec -= offset;
1028 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1032 * It may be possible that when we entered this function, xtime_nsec
1033 * was very small. Further, if we're slightly speeding the clocksource
1034 * in the code above, its possible the required corrective factor to
1035 * xtime_nsec could cause it to underflow.
1037 * Now, since we already accumulated the second, cannot simply roll
1038 * the accumulated second back, since the NTP subsystem has been
1039 * notified via second_overflow. So instead we push xtime_nsec forward
1040 * by the amount we underflowed, and add that amount into the error.
1042 * We'll correct this error next time through this function, when
1043 * xtime_nsec is not as small.
1045 if (unlikely((s64)tk->xtime_nsec < 0)) {
1046 s64 neg = -(s64)tk->xtime_nsec;
1048 tk->ntp_error += neg << tk->ntp_error_shift;
1054 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1056 * Helper function that accumulates a the nsecs greater then a second
1057 * from the xtime_nsec field to the xtime_secs field.
1058 * It also calls into the NTP code to handle leapsecond processing.
1061 static inline void accumulate_nsecs_to_secs(struct timekeeper *tk)
1063 u64 nsecps = (u64)NSEC_PER_SEC << tk->shift;
1065 while (tk->xtime_nsec >= nsecps) {
1068 tk->xtime_nsec -= nsecps;
1071 /* Figure out if its a leap sec and apply if needed */
1072 leap = second_overflow(tk->xtime_sec);
1073 if (unlikely(leap)) {
1076 tk->xtime_sec += leap;
1080 tk_set_wall_to_mono(tk,
1081 timespec_sub(tk->wall_to_monotonic, ts));
1083 clock_was_set_delayed();
1089 * logarithmic_accumulation - shifted accumulation of cycles
1091 * This functions accumulates a shifted interval of cycles into
1092 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1095 * Returns the unconsumed cycles.
1097 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1102 /* If the offset is smaller then a shifted interval, do nothing */
1103 if (offset < tk->cycle_interval<<shift)
1106 /* Accumulate one shifted interval */
1107 offset -= tk->cycle_interval << shift;
1108 tk->clock->cycle_last += tk->cycle_interval << shift;
1110 tk->xtime_nsec += tk->xtime_interval << shift;
1111 accumulate_nsecs_to_secs(tk);
1113 /* Accumulate raw time */
1114 raw_nsecs = tk->raw_interval << shift;
1115 raw_nsecs += tk->raw_time.tv_nsec;
1116 if (raw_nsecs >= NSEC_PER_SEC) {
1117 u64 raw_secs = raw_nsecs;
1118 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1119 tk->raw_time.tv_sec += raw_secs;
1121 tk->raw_time.tv_nsec = raw_nsecs;
1123 /* Accumulate error between NTP and clock interval */
1124 tk->ntp_error += ntp_tick_length() << shift;
1125 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1126 (tk->ntp_error_shift + shift);
1132 * update_wall_time - Uses the current clocksource to increment the wall time
1135 static void update_wall_time(void)
1137 struct clocksource *clock;
1138 struct timekeeper *tk = &timekeeper;
1140 int shift = 0, maxshift;
1141 unsigned long flags;
1144 write_seqlock_irqsave(&tk->lock, flags);
1146 /* Make sure we're fully resumed: */
1147 if (unlikely(timekeeping_suspended))
1152 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1153 offset = tk->cycle_interval;
1155 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1158 /* Check if there's really nothing to do */
1159 if (offset < tk->cycle_interval)
1163 * With NO_HZ we may have to accumulate many cycle_intervals
1164 * (think "ticks") worth of time at once. To do this efficiently,
1165 * we calculate the largest doubling multiple of cycle_intervals
1166 * that is smaller than the offset. We then accumulate that
1167 * chunk in one go, and then try to consume the next smaller
1170 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1171 shift = max(0, shift);
1172 /* Bound shift to one less than what overflows tick_length */
1173 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1174 shift = min(shift, maxshift);
1175 while (offset >= tk->cycle_interval) {
1176 offset = logarithmic_accumulation(tk, offset, shift);
1177 if (offset < tk->cycle_interval<<shift)
1181 /* correct the clock when NTP error is too big */
1182 timekeeping_adjust(tk, offset);
1186 * Store only full nanoseconds into xtime_nsec after rounding
1187 * it up and add the remainder to the error difference.
1188 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
1189 * by truncating the remainder in vsyscalls. However, it causes
1190 * additional work to be done in timekeeping_adjust(). Once
1191 * the vsyscall implementations are converted to use xtime_nsec
1192 * (shifted nanoseconds), this can be killed.
1194 remainder = tk->xtime_nsec & ((1ULL << tk->shift) - 1);
1195 tk->xtime_nsec -= remainder;
1196 tk->xtime_nsec += 1ULL << tk->shift;
1197 tk->ntp_error += remainder << tk->ntp_error_shift;
1200 * Finally, make sure that after the rounding
1201 * xtime_nsec isn't larger than NSEC_PER_SEC
1203 accumulate_nsecs_to_secs(tk);
1205 timekeeping_update(tk, false);
1208 write_sequnlock_irqrestore(&tk->lock, flags);
1213 * getboottime - Return the real time of system boot.
1214 * @ts: pointer to the timespec to be set
1216 * Returns the wall-time of boot in a timespec.
1218 * This is based on the wall_to_monotonic offset and the total suspend
1219 * time. Calls to settimeofday will affect the value returned (which
1220 * basically means that however wrong your real time clock is at boot time,
1221 * you get the right time here).
1223 void getboottime(struct timespec *ts)
1225 struct timekeeper *tk = &timekeeper;
1226 struct timespec boottime = {
1227 .tv_sec = tk->wall_to_monotonic.tv_sec +
1228 tk->total_sleep_time.tv_sec,
1229 .tv_nsec = tk->wall_to_monotonic.tv_nsec +
1230 tk->total_sleep_time.tv_nsec
1233 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1235 EXPORT_SYMBOL_GPL(getboottime);
1238 * get_monotonic_boottime - Returns monotonic time since boot
1239 * @ts: pointer to the timespec to be set
1241 * Returns the monotonic time since boot in a timespec.
1243 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1244 * includes the time spent in suspend.
1246 void get_monotonic_boottime(struct timespec *ts)
1248 struct timekeeper *tk = &timekeeper;
1249 struct timespec tomono, sleep;
1253 WARN_ON(timekeeping_suspended);
1256 seq = read_seqbegin(&tk->lock);
1257 ts->tv_sec = tk->xtime_sec;
1258 nsec = timekeeping_get_ns(tk);
1259 tomono = tk->wall_to_monotonic;
1260 sleep = tk->total_sleep_time;
1262 } while (read_seqretry(&tk->lock, seq));
1264 ts->tv_sec += tomono.tv_sec + sleep.tv_sec;
1266 timespec_add_ns(ts, nsec + tomono.tv_nsec + sleep.tv_nsec);
1268 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1271 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1273 * Returns the monotonic time since boot in a ktime
1275 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1276 * includes the time spent in suspend.
1278 ktime_t ktime_get_boottime(void)
1282 get_monotonic_boottime(&ts);
1283 return timespec_to_ktime(ts);
1285 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1288 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1289 * @ts: pointer to the timespec to be converted
1291 void monotonic_to_bootbased(struct timespec *ts)
1293 struct timekeeper *tk = &timekeeper;
1295 *ts = timespec_add(*ts, tk->total_sleep_time);
1297 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1299 unsigned long get_seconds(void)
1301 struct timekeeper *tk = &timekeeper;
1303 return tk->xtime_sec;
1305 EXPORT_SYMBOL(get_seconds);
1307 struct timespec __current_kernel_time(void)
1309 struct timekeeper *tk = &timekeeper;
1311 return tk_xtime(tk);
1314 struct timespec current_kernel_time(void)
1316 struct timekeeper *tk = &timekeeper;
1317 struct timespec now;
1321 seq = read_seqbegin(&tk->lock);
1324 } while (read_seqretry(&tk->lock, seq));
1328 EXPORT_SYMBOL(current_kernel_time);
1330 struct timespec get_monotonic_coarse(void)
1332 struct timekeeper *tk = &timekeeper;
1333 struct timespec now, mono;
1337 seq = read_seqbegin(&tk->lock);
1340 mono = tk->wall_to_monotonic;
1341 } while (read_seqretry(&tk->lock, seq));
1343 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1344 now.tv_nsec + mono.tv_nsec);
1349 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1350 * without sampling the sequence number in xtime_lock.
1351 * jiffies is defined in the linker script...
1353 void do_timer(unsigned long ticks)
1355 jiffies_64 += ticks;
1357 calc_global_load(ticks);
1361 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1362 * and sleep offsets.
1363 * @xtim: pointer to timespec to be set with xtime
1364 * @wtom: pointer to timespec to be set with wall_to_monotonic
1365 * @sleep: pointer to timespec to be set with time in suspend
1367 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1368 struct timespec *wtom, struct timespec *sleep)
1370 struct timekeeper *tk = &timekeeper;
1374 seq = read_seqbegin(&tk->lock);
1375 *xtim = tk_xtime(tk);
1376 *wtom = tk->wall_to_monotonic;
1377 *sleep = tk->total_sleep_time;
1378 } while (read_seqretry(&tk->lock, seq));
1381 #ifdef CONFIG_HIGH_RES_TIMERS
1383 * ktime_get_update_offsets - hrtimer helper
1384 * @offs_real: pointer to storage for monotonic -> realtime offset
1385 * @offs_boot: pointer to storage for monotonic -> boottime offset
1387 * Returns current monotonic time and updates the offsets
1388 * Called from hrtimer_interupt() or retrigger_next_event()
1390 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1392 struct timekeeper *tk = &timekeeper;
1398 seq = read_seqbegin(&tk->lock);
1400 secs = tk->xtime_sec;
1401 nsecs = timekeeping_get_ns(tk);
1403 *offs_real = tk->offs_real;
1404 *offs_boot = tk->offs_boot;
1405 } while (read_seqretry(&tk->lock, seq));
1407 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1408 now = ktime_sub(now, *offs_real);
1414 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1416 ktime_t ktime_get_monotonic_offset(void)
1418 struct timekeeper *tk = &timekeeper;
1420 struct timespec wtom;
1423 seq = read_seqbegin(&tk->lock);
1424 wtom = tk->wall_to_monotonic;
1425 } while (read_seqretry(&tk->lock, seq));
1427 return timespec_to_ktime(wtom);
1429 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1432 * xtime_update() - advances the timekeeping infrastructure
1433 * @ticks: number of ticks, that have elapsed since the last call.
1435 * Must be called with interrupts disabled.
1437 void xtime_update(unsigned long ticks)
1439 write_seqlock(&xtime_lock);
1441 write_sequnlock(&xtime_lock);