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/timekeeper_internal.h>
12 #include <linux/module.h>
13 #include <linux/interrupt.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
17 #include <linux/sched.h>
18 #include <linux/syscore_ops.h>
19 #include <linux/clocksource.h>
20 #include <linux/jiffies.h>
21 #include <linux/time.h>
22 #include <linux/tick.h>
23 #include <linux/stop_machine.h>
24 #include <linux/pvclock_gtod.h>
27 static struct timekeeper timekeeper;
30 * This read-write spinlock protects us from races in SMP while
33 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
35 /* flag for if timekeeping is suspended */
36 int __read_mostly timekeeping_suspended;
38 static inline void tk_normalize_xtime(struct timekeeper *tk)
40 while (tk->xtime_nsec >= ((u64)NSEC_PER_SEC << tk->shift)) {
41 tk->xtime_nsec -= (u64)NSEC_PER_SEC << tk->shift;
46 static void tk_set_xtime(struct timekeeper *tk, const struct timespec *ts)
48 tk->xtime_sec = ts->tv_sec;
49 tk->xtime_nsec = (u64)ts->tv_nsec << tk->shift;
52 static void tk_xtime_add(struct timekeeper *tk, const struct timespec *ts)
54 tk->xtime_sec += ts->tv_sec;
55 tk->xtime_nsec += (u64)ts->tv_nsec << tk->shift;
56 tk_normalize_xtime(tk);
59 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec wtm)
64 * Verify consistency of: offset_real = -wall_to_monotonic
65 * before modifying anything
67 set_normalized_timespec(&tmp, -tk->wall_to_monotonic.tv_sec,
68 -tk->wall_to_monotonic.tv_nsec);
69 WARN_ON_ONCE(tk->offs_real.tv64 != timespec_to_ktime(tmp).tv64);
70 tk->wall_to_monotonic = wtm;
71 set_normalized_timespec(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
72 tk->offs_real = timespec_to_ktime(tmp);
75 static void tk_set_sleep_time(struct timekeeper *tk, struct timespec t)
77 /* Verify consistency before modifying */
78 WARN_ON_ONCE(tk->offs_boot.tv64 != timespec_to_ktime(tk->total_sleep_time).tv64);
80 tk->total_sleep_time = t;
81 tk->offs_boot = timespec_to_ktime(t);
85 * timekeeper_setup_internals - Set up internals to use clocksource clock.
87 * @clock: Pointer to clocksource.
89 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
90 * pair and interval request.
92 * Unless you're the timekeeping code, you should not be using this!
94 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
98 struct clocksource *old_clock;
100 old_clock = tk->clock;
102 clock->cycle_last = clock->read(clock);
104 /* Do the ns -> cycle conversion first, using original mult */
105 tmp = NTP_INTERVAL_LENGTH;
106 tmp <<= clock->shift;
108 tmp += clock->mult/2;
109 do_div(tmp, clock->mult);
113 interval = (cycle_t) tmp;
114 tk->cycle_interval = interval;
116 /* Go back from cycles -> shifted ns */
117 tk->xtime_interval = (u64) interval * clock->mult;
118 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
120 ((u64) interval * clock->mult) >> clock->shift;
122 /* if changing clocks, convert xtime_nsec shift units */
124 int shift_change = clock->shift - old_clock->shift;
125 if (shift_change < 0)
126 tk->xtime_nsec >>= -shift_change;
128 tk->xtime_nsec <<= shift_change;
130 tk->shift = clock->shift;
133 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
136 * The timekeeper keeps its own mult values for the currently
137 * active clocksource. These value will be adjusted via NTP
138 * to counteract clock drifting.
140 tk->mult = clock->mult;
143 /* Timekeeper helper functions. */
144 static inline s64 timekeeping_get_ns(struct timekeeper *tk)
146 cycle_t cycle_now, cycle_delta;
147 struct clocksource *clock;
150 /* read clocksource: */
152 cycle_now = clock->read(clock);
154 /* calculate the delta since the last update_wall_time: */
155 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
157 nsec = cycle_delta * tk->mult + tk->xtime_nsec;
160 /* If arch requires, add in gettimeoffset() */
161 return nsec + arch_gettimeoffset();
164 static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
166 cycle_t cycle_now, cycle_delta;
167 struct clocksource *clock;
170 /* read clocksource: */
172 cycle_now = clock->read(clock);
174 /* calculate the delta since the last update_wall_time: */
175 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
177 /* convert delta to nanoseconds. */
178 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
180 /* If arch requires, add in gettimeoffset() */
181 return nsec + arch_gettimeoffset();
184 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
186 static void update_pvclock_gtod(struct timekeeper *tk)
188 raw_notifier_call_chain(&pvclock_gtod_chain, 0, tk);
192 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
194 * Must hold write on timekeeper.lock
196 int pvclock_gtod_register_notifier(struct notifier_block *nb)
198 struct timekeeper *tk = &timekeeper;
202 write_seqlock_irqsave(&tk->lock, flags);
203 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
204 /* update timekeeping data */
205 update_pvclock_gtod(tk);
206 write_sequnlock_irqrestore(&tk->lock, flags);
210 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
213 * pvclock_gtod_unregister_notifier - unregister a pvclock
214 * timedata update listener
216 * Must hold write on timekeeper.lock
218 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
220 struct timekeeper *tk = &timekeeper;
224 write_seqlock_irqsave(&tk->lock, flags);
225 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
226 write_sequnlock_irqrestore(&tk->lock, flags);
230 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
232 /* must hold write on timekeeper.lock */
233 static void timekeeping_update(struct timekeeper *tk, bool clearntp)
240 update_pvclock_gtod(tk);
244 * timekeeping_forward_now - update clock to the current time
246 * Forward the current clock to update its state since the last call to
247 * update_wall_time(). This is useful before significant clock changes,
248 * as it avoids having to deal with this time offset explicitly.
250 static void timekeeping_forward_now(struct timekeeper *tk)
252 cycle_t cycle_now, cycle_delta;
253 struct clocksource *clock;
257 cycle_now = clock->read(clock);
258 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
259 clock->cycle_last = cycle_now;
261 tk->xtime_nsec += cycle_delta * tk->mult;
263 /* If arch requires, add in gettimeoffset() */
264 tk->xtime_nsec += (u64)arch_gettimeoffset() << tk->shift;
266 tk_normalize_xtime(tk);
268 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
269 timespec_add_ns(&tk->raw_time, nsec);
273 * getnstimeofday - Returns the time of day in a timespec
274 * @ts: pointer to the timespec to be set
276 * Returns the time of day in a timespec.
278 void getnstimeofday(struct timespec *ts)
280 struct timekeeper *tk = &timekeeper;
284 WARN_ON(timekeeping_suspended);
287 seq = read_seqbegin(&tk->lock);
289 ts->tv_sec = tk->xtime_sec;
290 nsecs = timekeeping_get_ns(tk);
292 } while (read_seqretry(&tk->lock, seq));
295 timespec_add_ns(ts, nsecs);
297 EXPORT_SYMBOL(getnstimeofday);
299 ktime_t ktime_get(void)
301 struct timekeeper *tk = &timekeeper;
305 WARN_ON(timekeeping_suspended);
308 seq = read_seqbegin(&tk->lock);
309 secs = tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
310 nsecs = timekeeping_get_ns(tk) + tk->wall_to_monotonic.tv_nsec;
312 } while (read_seqretry(&tk->lock, seq));
314 * Use ktime_set/ktime_add_ns to create a proper ktime on
315 * 32-bit architectures without CONFIG_KTIME_SCALAR.
317 return ktime_add_ns(ktime_set(secs, 0), nsecs);
319 EXPORT_SYMBOL_GPL(ktime_get);
322 * ktime_get_ts - get the monotonic clock in timespec format
323 * @ts: pointer to timespec variable
325 * The function calculates the monotonic clock from the realtime
326 * clock and the wall_to_monotonic offset and stores the result
327 * in normalized timespec format in the variable pointed to by @ts.
329 void ktime_get_ts(struct timespec *ts)
331 struct timekeeper *tk = &timekeeper;
332 struct timespec tomono;
336 WARN_ON(timekeeping_suspended);
339 seq = read_seqbegin(&tk->lock);
340 ts->tv_sec = tk->xtime_sec;
341 nsec = timekeeping_get_ns(tk);
342 tomono = tk->wall_to_monotonic;
344 } while (read_seqretry(&tk->lock, seq));
346 ts->tv_sec += tomono.tv_sec;
348 timespec_add_ns(ts, nsec + tomono.tv_nsec);
350 EXPORT_SYMBOL_GPL(ktime_get_ts);
352 #ifdef CONFIG_NTP_PPS
355 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
356 * @ts_raw: pointer to the timespec to be set to raw monotonic time
357 * @ts_real: pointer to the timespec to be set to the time of day
359 * This function reads both the time of day and raw monotonic time at the
360 * same time atomically and stores the resulting timestamps in timespec
363 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
365 struct timekeeper *tk = &timekeeper;
367 s64 nsecs_raw, nsecs_real;
369 WARN_ON_ONCE(timekeeping_suspended);
372 seq = read_seqbegin(&tk->lock);
374 *ts_raw = tk->raw_time;
375 ts_real->tv_sec = tk->xtime_sec;
376 ts_real->tv_nsec = 0;
378 nsecs_raw = timekeeping_get_ns_raw(tk);
379 nsecs_real = timekeeping_get_ns(tk);
381 } while (read_seqretry(&tk->lock, seq));
383 timespec_add_ns(ts_raw, nsecs_raw);
384 timespec_add_ns(ts_real, nsecs_real);
386 EXPORT_SYMBOL(getnstime_raw_and_real);
388 #endif /* CONFIG_NTP_PPS */
391 * do_gettimeofday - Returns the time of day in a timeval
392 * @tv: pointer to the timeval to be set
394 * NOTE: Users should be converted to using getnstimeofday()
396 void do_gettimeofday(struct timeval *tv)
400 getnstimeofday(&now);
401 tv->tv_sec = now.tv_sec;
402 tv->tv_usec = now.tv_nsec/1000;
404 EXPORT_SYMBOL(do_gettimeofday);
407 * do_settimeofday - Sets the time of day
408 * @tv: pointer to the timespec variable containing the new time
410 * Sets the time of day to the new time and update NTP and notify hrtimers
412 int do_settimeofday(const struct timespec *tv)
414 struct timekeeper *tk = &timekeeper;
415 struct timespec ts_delta, xt;
418 if (!timespec_valid_strict(tv))
421 write_seqlock_irqsave(&tk->lock, flags);
423 timekeeping_forward_now(tk);
426 ts_delta.tv_sec = tv->tv_sec - xt.tv_sec;
427 ts_delta.tv_nsec = tv->tv_nsec - xt.tv_nsec;
429 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, ts_delta));
431 tk_set_xtime(tk, tv);
433 timekeeping_update(tk, true);
435 write_sequnlock_irqrestore(&tk->lock, flags);
437 /* signal hrtimers about time change */
442 EXPORT_SYMBOL(do_settimeofday);
445 * timekeeping_inject_offset - Adds or subtracts from the current time.
446 * @tv: pointer to the timespec variable containing the offset
448 * Adds or subtracts an offset value from the current time.
450 int timekeeping_inject_offset(struct timespec *ts)
452 struct timekeeper *tk = &timekeeper;
457 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
460 write_seqlock_irqsave(&tk->lock, flags);
462 timekeeping_forward_now(tk);
464 /* Make sure the proposed value is valid */
465 tmp = timespec_add(tk_xtime(tk), *ts);
466 if (!timespec_valid_strict(&tmp)) {
471 tk_xtime_add(tk, ts);
472 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *ts));
474 error: /* even if we error out, we forwarded the time, so call update */
475 timekeeping_update(tk, true);
477 write_sequnlock_irqrestore(&tk->lock, flags);
479 /* signal hrtimers about time change */
484 EXPORT_SYMBOL(timekeeping_inject_offset);
487 * change_clocksource - Swaps clocksources if a new one is available
489 * Accumulates current time interval and initializes new clocksource
491 static int change_clocksource(void *data)
493 struct timekeeper *tk = &timekeeper;
494 struct clocksource *new, *old;
497 new = (struct clocksource *) data;
499 write_seqlock_irqsave(&tk->lock, flags);
501 timekeeping_forward_now(tk);
502 if (!new->enable || new->enable(new) == 0) {
504 tk_setup_internals(tk, new);
508 timekeeping_update(tk, true);
510 write_sequnlock_irqrestore(&tk->lock, flags);
516 * timekeeping_notify - Install a new clock source
517 * @clock: pointer to the clock source
519 * This function is called from clocksource.c after a new, better clock
520 * source has been registered. The caller holds the clocksource_mutex.
522 void timekeeping_notify(struct clocksource *clock)
524 struct timekeeper *tk = &timekeeper;
526 if (tk->clock == clock)
528 stop_machine(change_clocksource, clock, NULL);
533 * ktime_get_real - get the real (wall-) time in ktime_t format
535 * returns the time in ktime_t format
537 ktime_t ktime_get_real(void)
541 getnstimeofday(&now);
543 return timespec_to_ktime(now);
545 EXPORT_SYMBOL_GPL(ktime_get_real);
548 * getrawmonotonic - Returns the raw monotonic time in a timespec
549 * @ts: pointer to the timespec to be set
551 * Returns the raw monotonic time (completely un-modified by ntp)
553 void getrawmonotonic(struct timespec *ts)
555 struct timekeeper *tk = &timekeeper;
560 seq = read_seqbegin(&tk->lock);
561 nsecs = timekeeping_get_ns_raw(tk);
564 } while (read_seqretry(&tk->lock, seq));
566 timespec_add_ns(ts, nsecs);
568 EXPORT_SYMBOL(getrawmonotonic);
571 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
573 int timekeeping_valid_for_hres(void)
575 struct timekeeper *tk = &timekeeper;
580 seq = read_seqbegin(&tk->lock);
582 ret = tk->clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
584 } while (read_seqretry(&tk->lock, seq));
590 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
592 u64 timekeeping_max_deferment(void)
594 struct timekeeper *tk = &timekeeper;
599 seq = read_seqbegin(&tk->lock);
601 ret = tk->clock->max_idle_ns;
603 } while (read_seqretry(&tk->lock, seq));
609 * read_persistent_clock - Return time from the persistent clock.
611 * Weak dummy function for arches that do not yet support it.
612 * Reads the time from the battery backed persistent clock.
613 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
615 * XXX - Do be sure to remove it once all arches implement it.
617 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
624 * read_boot_clock - Return time of the system start.
626 * Weak dummy function for arches that do not yet support it.
627 * Function to read the exact time the system has been started.
628 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
630 * XXX - Do be sure to remove it once all arches implement it.
632 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
639 * timekeeping_init - Initializes the clocksource and common timekeeping values
641 void __init timekeeping_init(void)
643 struct timekeeper *tk = &timekeeper;
644 struct clocksource *clock;
646 struct timespec now, boot, tmp;
648 read_persistent_clock(&now);
649 if (!timespec_valid_strict(&now)) {
650 pr_warn("WARNING: Persistent clock returned invalid value!\n"
651 " Check your CMOS/BIOS settings.\n");
656 read_boot_clock(&boot);
657 if (!timespec_valid_strict(&boot)) {
658 pr_warn("WARNING: Boot clock returned invalid value!\n"
659 " Check your CMOS/BIOS settings.\n");
664 seqlock_init(&tk->lock);
668 write_seqlock_irqsave(&tk->lock, flags);
669 clock = clocksource_default_clock();
671 clock->enable(clock);
672 tk_setup_internals(tk, clock);
674 tk_set_xtime(tk, &now);
675 tk->raw_time.tv_sec = 0;
676 tk->raw_time.tv_nsec = 0;
677 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
680 set_normalized_timespec(&tmp, -boot.tv_sec, -boot.tv_nsec);
681 tk_set_wall_to_mono(tk, tmp);
685 tk_set_sleep_time(tk, tmp);
687 write_sequnlock_irqrestore(&tk->lock, flags);
690 /* time in seconds when suspend began */
691 static struct timespec timekeeping_suspend_time;
694 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
695 * @delta: pointer to a timespec delta value
697 * Takes a timespec offset measuring a suspend interval and properly
698 * adds the sleep offset to the timekeeping variables.
700 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
701 struct timespec *delta)
703 if (!timespec_valid_strict(delta)) {
704 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
705 "sleep delta value!\n");
708 tk_xtime_add(tk, delta);
709 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *delta));
710 tk_set_sleep_time(tk, timespec_add(tk->total_sleep_time, *delta));
714 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
715 * @delta: pointer to a timespec delta value
717 * This hook is for architectures that cannot support read_persistent_clock
718 * because their RTC/persistent clock is only accessible when irqs are enabled.
720 * This function should only be called by rtc_resume(), and allows
721 * a suspend offset to be injected into the timekeeping values.
723 void timekeeping_inject_sleeptime(struct timespec *delta)
725 struct timekeeper *tk = &timekeeper;
729 /* Make sure we don't set the clock twice */
730 read_persistent_clock(&ts);
731 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
734 write_seqlock_irqsave(&tk->lock, flags);
736 timekeeping_forward_now(tk);
738 __timekeeping_inject_sleeptime(tk, delta);
740 timekeeping_update(tk, true);
742 write_sequnlock_irqrestore(&tk->lock, flags);
744 /* signal hrtimers about time change */
749 * timekeeping_resume - Resumes the generic timekeeping subsystem.
751 * This is for the generic clocksource timekeeping.
752 * xtime/wall_to_monotonic/jiffies/etc are
753 * still managed by arch specific suspend/resume code.
755 static void timekeeping_resume(void)
757 struct timekeeper *tk = &timekeeper;
761 read_persistent_clock(&ts);
763 clockevents_resume();
764 clocksource_resume();
766 write_seqlock_irqsave(&tk->lock, flags);
768 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
769 ts = timespec_sub(ts, timekeeping_suspend_time);
770 __timekeeping_inject_sleeptime(tk, &ts);
772 /* re-base the last cycle value */
773 tk->clock->cycle_last = tk->clock->read(tk->clock);
775 timekeeping_suspended = 0;
776 timekeeping_update(tk, false);
777 write_sequnlock_irqrestore(&tk->lock, flags);
779 touch_softlockup_watchdog();
781 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
783 /* Resume hrtimers */
787 static int timekeeping_suspend(void)
789 struct timekeeper *tk = &timekeeper;
791 struct timespec delta, delta_delta;
792 static struct timespec old_delta;
794 read_persistent_clock(&timekeeping_suspend_time);
796 write_seqlock_irqsave(&tk->lock, flags);
797 timekeeping_forward_now(tk);
798 timekeeping_suspended = 1;
801 * To avoid drift caused by repeated suspend/resumes,
802 * which each can add ~1 second drift error,
803 * try to compensate so the difference in system time
804 * and persistent_clock time stays close to constant.
806 delta = timespec_sub(tk_xtime(tk), timekeeping_suspend_time);
807 delta_delta = timespec_sub(delta, old_delta);
808 if (abs(delta_delta.tv_sec) >= 2) {
810 * if delta_delta is too large, assume time correction
811 * has occured and set old_delta to the current delta.
815 /* Otherwise try to adjust old_system to compensate */
816 timekeeping_suspend_time =
817 timespec_add(timekeeping_suspend_time, delta_delta);
819 write_sequnlock_irqrestore(&tk->lock, flags);
821 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
822 clocksource_suspend();
823 clockevents_suspend();
828 /* sysfs resume/suspend bits for timekeeping */
829 static struct syscore_ops timekeeping_syscore_ops = {
830 .resume = timekeeping_resume,
831 .suspend = timekeeping_suspend,
834 static int __init timekeeping_init_ops(void)
836 register_syscore_ops(&timekeeping_syscore_ops);
840 device_initcall(timekeeping_init_ops);
843 * If the error is already larger, we look ahead even further
844 * to compensate for late or lost adjustments.
846 static __always_inline int timekeeping_bigadjust(struct timekeeper *tk,
847 s64 error, s64 *interval,
855 * Use the current error value to determine how much to look ahead.
856 * The larger the error the slower we adjust for it to avoid problems
857 * with losing too many ticks, otherwise we would overadjust and
858 * produce an even larger error. The smaller the adjustment the
859 * faster we try to adjust for it, as lost ticks can do less harm
860 * here. This is tuned so that an error of about 1 msec is adjusted
861 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
863 error2 = tk->ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
864 error2 = abs(error2);
865 for (look_ahead = 0; error2 > 0; look_ahead++)
869 * Now calculate the error in (1 << look_ahead) ticks, but first
870 * remove the single look ahead already included in the error.
872 tick_error = ntp_tick_length() >> (tk->ntp_error_shift + 1);
873 tick_error -= tk->xtime_interval >> 1;
874 error = ((error - tick_error) >> look_ahead) + tick_error;
876 /* Finally calculate the adjustment shift value. */
881 *interval = -*interval;
885 for (adj = 0; error > i; adj++)
894 * Adjust the multiplier to reduce the error value,
895 * this is optimized for the most common adjustments of -1,0,1,
896 * for other values we can do a bit more work.
898 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
900 s64 error, interval = tk->cycle_interval;
904 * The point of this is to check if the error is greater than half
907 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
909 * Note we subtract one in the shift, so that error is really error*2.
910 * This "saves" dividing(shifting) interval twice, but keeps the
911 * (error > interval) comparison as still measuring if error is
912 * larger than half an interval.
914 * Note: It does not "save" on aggravation when reading the code.
916 error = tk->ntp_error >> (tk->ntp_error_shift - 1);
917 if (error > interval) {
919 * We now divide error by 4(via shift), which checks if
920 * the error is greater than twice the interval.
921 * If it is greater, we need a bigadjust, if its smaller,
922 * we can adjust by 1.
926 * XXX - In update_wall_time, we round up to the next
927 * nanosecond, and store the amount rounded up into
928 * the error. This causes the likely below to be unlikely.
930 * The proper fix is to avoid rounding up by using
931 * the high precision tk->xtime_nsec instead of
932 * xtime.tv_nsec everywhere. Fixing this will take some
935 if (likely(error <= interval))
938 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
940 if (error < -interval) {
941 /* See comment above, this is just switched for the negative */
943 if (likely(error >= -interval)) {
945 interval = -interval;
948 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
955 if (unlikely(tk->clock->maxadj &&
956 (tk->mult + adj > tk->clock->mult + tk->clock->maxadj))) {
957 printk_once(KERN_WARNING
958 "Adjusting %s more than 11%% (%ld vs %ld)\n",
959 tk->clock->name, (long)tk->mult + adj,
960 (long)tk->clock->mult + tk->clock->maxadj);
963 * So the following can be confusing.
965 * To keep things simple, lets assume adj == 1 for now.
967 * When adj != 1, remember that the interval and offset values
968 * have been appropriately scaled so the math is the same.
970 * The basic idea here is that we're increasing the multiplier
971 * by one, this causes the xtime_interval to be incremented by
972 * one cycle_interval. This is because:
973 * xtime_interval = cycle_interval * mult
974 * So if mult is being incremented by one:
975 * xtime_interval = cycle_interval * (mult + 1)
977 * xtime_interval = (cycle_interval * mult) + cycle_interval
978 * Which can be shortened to:
979 * xtime_interval += cycle_interval
981 * So offset stores the non-accumulated cycles. Thus the current
982 * time (in shifted nanoseconds) is:
983 * now = (offset * adj) + xtime_nsec
984 * Now, even though we're adjusting the clock frequency, we have
985 * to keep time consistent. In other words, we can't jump back
986 * in time, and we also want to avoid jumping forward in time.
988 * So given the same offset value, we need the time to be the same
989 * both before and after the freq adjustment.
990 * now = (offset * adj_1) + xtime_nsec_1
991 * now = (offset * adj_2) + xtime_nsec_2
993 * (offset * adj_1) + xtime_nsec_1 =
994 * (offset * adj_2) + xtime_nsec_2
998 * (offset * adj_1) + xtime_nsec_1 =
999 * (offset * (adj_1+1)) + xtime_nsec_2
1000 * (offset * adj_1) + xtime_nsec_1 =
1001 * (offset * adj_1) + offset + xtime_nsec_2
1002 * Canceling the sides:
1003 * xtime_nsec_1 = offset + xtime_nsec_2
1005 * xtime_nsec_2 = xtime_nsec_1 - offset
1006 * Which simplfies to:
1007 * xtime_nsec -= offset
1009 * XXX - TODO: Doc ntp_error calculation.
1012 tk->xtime_interval += interval;
1013 tk->xtime_nsec -= offset;
1014 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1018 * It may be possible that when we entered this function, xtime_nsec
1019 * was very small. Further, if we're slightly speeding the clocksource
1020 * in the code above, its possible the required corrective factor to
1021 * xtime_nsec could cause it to underflow.
1023 * Now, since we already accumulated the second, cannot simply roll
1024 * the accumulated second back, since the NTP subsystem has been
1025 * notified via second_overflow. So instead we push xtime_nsec forward
1026 * by the amount we underflowed, and add that amount into the error.
1028 * We'll correct this error next time through this function, when
1029 * xtime_nsec is not as small.
1031 if (unlikely((s64)tk->xtime_nsec < 0)) {
1032 s64 neg = -(s64)tk->xtime_nsec;
1034 tk->ntp_error += neg << tk->ntp_error_shift;
1040 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1042 * Helper function that accumulates a the nsecs greater then a second
1043 * from the xtime_nsec field to the xtime_secs field.
1044 * It also calls into the NTP code to handle leapsecond processing.
1047 static inline void accumulate_nsecs_to_secs(struct timekeeper *tk)
1049 u64 nsecps = (u64)NSEC_PER_SEC << tk->shift;
1051 while (tk->xtime_nsec >= nsecps) {
1054 tk->xtime_nsec -= nsecps;
1057 /* Figure out if its a leap sec and apply if needed */
1058 leap = second_overflow(tk->xtime_sec);
1059 if (unlikely(leap)) {
1062 tk->xtime_sec += leap;
1066 tk_set_wall_to_mono(tk,
1067 timespec_sub(tk->wall_to_monotonic, ts));
1069 clock_was_set_delayed();
1075 * logarithmic_accumulation - shifted accumulation of cycles
1077 * This functions accumulates a shifted interval of cycles into
1078 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1081 * Returns the unconsumed cycles.
1083 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1088 /* If the offset is smaller then a shifted interval, do nothing */
1089 if (offset < tk->cycle_interval<<shift)
1092 /* Accumulate one shifted interval */
1093 offset -= tk->cycle_interval << shift;
1094 tk->clock->cycle_last += tk->cycle_interval << shift;
1096 tk->xtime_nsec += tk->xtime_interval << shift;
1097 accumulate_nsecs_to_secs(tk);
1099 /* Accumulate raw time */
1100 raw_nsecs = (u64)tk->raw_interval << shift;
1101 raw_nsecs += tk->raw_time.tv_nsec;
1102 if (raw_nsecs >= NSEC_PER_SEC) {
1103 u64 raw_secs = raw_nsecs;
1104 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1105 tk->raw_time.tv_sec += raw_secs;
1107 tk->raw_time.tv_nsec = raw_nsecs;
1109 /* Accumulate error between NTP and clock interval */
1110 tk->ntp_error += ntp_tick_length() << shift;
1111 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1112 (tk->ntp_error_shift + shift);
1117 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
1118 static inline void old_vsyscall_fixup(struct timekeeper *tk)
1123 * Store only full nanoseconds into xtime_nsec after rounding
1124 * it up and add the remainder to the error difference.
1125 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
1126 * by truncating the remainder in vsyscalls. However, it causes
1127 * additional work to be done in timekeeping_adjust(). Once
1128 * the vsyscall implementations are converted to use xtime_nsec
1129 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
1130 * users are removed, this can be killed.
1132 remainder = tk->xtime_nsec & ((1ULL << tk->shift) - 1);
1133 tk->xtime_nsec -= remainder;
1134 tk->xtime_nsec += 1ULL << tk->shift;
1135 tk->ntp_error += remainder << tk->ntp_error_shift;
1139 #define old_vsyscall_fixup(tk)
1145 * update_wall_time - Uses the current clocksource to increment the wall time
1148 static void update_wall_time(void)
1150 struct clocksource *clock;
1151 struct timekeeper *tk = &timekeeper;
1153 int shift = 0, maxshift;
1154 unsigned long flags;
1156 write_seqlock_irqsave(&tk->lock, flags);
1158 /* Make sure we're fully resumed: */
1159 if (unlikely(timekeeping_suspended))
1164 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1165 offset = tk->cycle_interval;
1167 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1170 /* Check if there's really nothing to do */
1171 if (offset < tk->cycle_interval)
1175 * With NO_HZ we may have to accumulate many cycle_intervals
1176 * (think "ticks") worth of time at once. To do this efficiently,
1177 * we calculate the largest doubling multiple of cycle_intervals
1178 * that is smaller than the offset. We then accumulate that
1179 * chunk in one go, and then try to consume the next smaller
1182 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1183 shift = max(0, shift);
1184 /* Bound shift to one less than what overflows tick_length */
1185 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1186 shift = min(shift, maxshift);
1187 while (offset >= tk->cycle_interval) {
1188 offset = logarithmic_accumulation(tk, offset, shift);
1189 if (offset < tk->cycle_interval<<shift)
1193 /* correct the clock when NTP error is too big */
1194 timekeeping_adjust(tk, offset);
1197 * XXX This can be killed once everyone converts
1198 * to the new update_vsyscall.
1200 old_vsyscall_fixup(tk);
1203 * Finally, make sure that after the rounding
1204 * xtime_nsec isn't larger than NSEC_PER_SEC
1206 accumulate_nsecs_to_secs(tk);
1208 timekeeping_update(tk, false);
1211 write_sequnlock_irqrestore(&tk->lock, flags);
1216 * getboottime - Return the real time of system boot.
1217 * @ts: pointer to the timespec to be set
1219 * Returns the wall-time of boot in a timespec.
1221 * This is based on the wall_to_monotonic offset and the total suspend
1222 * time. Calls to settimeofday will affect the value returned (which
1223 * basically means that however wrong your real time clock is at boot time,
1224 * you get the right time here).
1226 void getboottime(struct timespec *ts)
1228 struct timekeeper *tk = &timekeeper;
1229 struct timespec boottime = {
1230 .tv_sec = tk->wall_to_monotonic.tv_sec +
1231 tk->total_sleep_time.tv_sec,
1232 .tv_nsec = tk->wall_to_monotonic.tv_nsec +
1233 tk->total_sleep_time.tv_nsec
1236 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1238 EXPORT_SYMBOL_GPL(getboottime);
1241 * get_monotonic_boottime - Returns monotonic time since boot
1242 * @ts: pointer to the timespec to be set
1244 * Returns the monotonic time since boot in a timespec.
1246 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1247 * includes the time spent in suspend.
1249 void get_monotonic_boottime(struct timespec *ts)
1251 struct timekeeper *tk = &timekeeper;
1252 struct timespec tomono, sleep;
1256 WARN_ON(timekeeping_suspended);
1259 seq = read_seqbegin(&tk->lock);
1260 ts->tv_sec = tk->xtime_sec;
1261 nsec = timekeeping_get_ns(tk);
1262 tomono = tk->wall_to_monotonic;
1263 sleep = tk->total_sleep_time;
1265 } while (read_seqretry(&tk->lock, seq));
1267 ts->tv_sec += tomono.tv_sec + sleep.tv_sec;
1269 timespec_add_ns(ts, nsec + tomono.tv_nsec + sleep.tv_nsec);
1271 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1274 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1276 * Returns the monotonic time since boot in a ktime
1278 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1279 * includes the time spent in suspend.
1281 ktime_t ktime_get_boottime(void)
1285 get_monotonic_boottime(&ts);
1286 return timespec_to_ktime(ts);
1288 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1291 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1292 * @ts: pointer to the timespec to be converted
1294 void monotonic_to_bootbased(struct timespec *ts)
1296 struct timekeeper *tk = &timekeeper;
1298 *ts = timespec_add(*ts, tk->total_sleep_time);
1300 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1302 unsigned long get_seconds(void)
1304 struct timekeeper *tk = &timekeeper;
1306 return tk->xtime_sec;
1308 EXPORT_SYMBOL(get_seconds);
1310 struct timespec __current_kernel_time(void)
1312 struct timekeeper *tk = &timekeeper;
1314 return tk_xtime(tk);
1317 struct timespec current_kernel_time(void)
1319 struct timekeeper *tk = &timekeeper;
1320 struct timespec now;
1324 seq = read_seqbegin(&tk->lock);
1327 } while (read_seqretry(&tk->lock, seq));
1331 EXPORT_SYMBOL(current_kernel_time);
1333 struct timespec get_monotonic_coarse(void)
1335 struct timekeeper *tk = &timekeeper;
1336 struct timespec now, mono;
1340 seq = read_seqbegin(&tk->lock);
1343 mono = tk->wall_to_monotonic;
1344 } while (read_seqretry(&tk->lock, seq));
1346 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1347 now.tv_nsec + mono.tv_nsec);
1352 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1353 * without sampling the sequence number in xtime_lock.
1354 * jiffies is defined in the linker script...
1356 void do_timer(unsigned long ticks)
1358 jiffies_64 += ticks;
1360 calc_global_load(ticks);
1364 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1365 * and sleep offsets.
1366 * @xtim: pointer to timespec to be set with xtime
1367 * @wtom: pointer to timespec to be set with wall_to_monotonic
1368 * @sleep: pointer to timespec to be set with time in suspend
1370 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1371 struct timespec *wtom, struct timespec *sleep)
1373 struct timekeeper *tk = &timekeeper;
1377 seq = read_seqbegin(&tk->lock);
1378 *xtim = tk_xtime(tk);
1379 *wtom = tk->wall_to_monotonic;
1380 *sleep = tk->total_sleep_time;
1381 } while (read_seqretry(&tk->lock, seq));
1384 #ifdef CONFIG_HIGH_RES_TIMERS
1386 * ktime_get_update_offsets - hrtimer helper
1387 * @offs_real: pointer to storage for monotonic -> realtime offset
1388 * @offs_boot: pointer to storage for monotonic -> boottime offset
1390 * Returns current monotonic time and updates the offsets
1391 * Called from hrtimer_interupt() or retrigger_next_event()
1393 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1395 struct timekeeper *tk = &timekeeper;
1401 seq = read_seqbegin(&tk->lock);
1403 secs = tk->xtime_sec;
1404 nsecs = timekeeping_get_ns(tk);
1406 *offs_real = tk->offs_real;
1407 *offs_boot = tk->offs_boot;
1408 } while (read_seqretry(&tk->lock, seq));
1410 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1411 now = ktime_sub(now, *offs_real);
1417 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1419 ktime_t ktime_get_monotonic_offset(void)
1421 struct timekeeper *tk = &timekeeper;
1423 struct timespec wtom;
1426 seq = read_seqbegin(&tk->lock);
1427 wtom = tk->wall_to_monotonic;
1428 } while (read_seqretry(&tk->lock, seq));
1430 return timespec_to_ktime(wtom);
1432 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1435 * xtime_update() - advances the timekeeping infrastructure
1436 * @ticks: number of ticks, that have elapsed since the last call.
1438 * Must be called with interrupts disabled.
1440 void xtime_update(unsigned long ticks)
1442 write_seqlock(&xtime_lock);
1444 write_sequnlock(&xtime_lock);