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;
120 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec wtm)
125 * Verify consistency of: offset_real = -wall_to_monotonic
126 * before modifying anything
128 set_normalized_timespec(&tmp, -tk->wall_to_monotonic.tv_sec,
129 -tk->wall_to_monotonic.tv_nsec);
130 WARN_ON_ONCE(tk->offs_real.tv64 != timespec_to_ktime(tmp).tv64);
131 tk->wall_to_monotonic = wtm;
132 set_normalized_timespec(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
133 tk->offs_real = timespec_to_ktime(tmp);
136 static void tk_set_sleep_time(struct timekeeper *tk, struct timespec t)
138 /* Verify consistency before modifying */
139 WARN_ON_ONCE(tk->offs_boot.tv64 != timespec_to_ktime(tk->total_sleep_time).tv64);
141 tk->total_sleep_time = t;
142 tk->offs_boot = timespec_to_ktime(t);
146 * timekeeper_setup_internals - Set up internals to use clocksource clock.
148 * @clock: Pointer to clocksource.
150 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
151 * pair and interval request.
153 * Unless you're the timekeeping code, you should not be using this!
155 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
158 u64 tmp, ntpinterval;
159 struct clocksource *old_clock;
161 old_clock = tk->clock;
163 clock->cycle_last = clock->read(clock);
165 /* Do the ns -> cycle conversion first, using original mult */
166 tmp = NTP_INTERVAL_LENGTH;
167 tmp <<= clock->shift;
169 tmp += clock->mult/2;
170 do_div(tmp, clock->mult);
174 interval = (cycle_t) tmp;
175 tk->cycle_interval = interval;
177 /* Go back from cycles -> shifted ns */
178 tk->xtime_interval = (u64) interval * clock->mult;
179 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
181 ((u64) interval * clock->mult) >> clock->shift;
183 /* if changing clocks, convert xtime_nsec shift units */
185 int shift_change = clock->shift - old_clock->shift;
186 if (shift_change < 0)
187 tk->xtime_nsec >>= -shift_change;
189 tk->xtime_nsec <<= shift_change;
191 tk->shift = clock->shift;
194 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
197 * The timekeeper keeps its own mult values for the currently
198 * active clocksource. These value will be adjusted via NTP
199 * to counteract clock drifting.
201 tk->mult = clock->mult;
204 /* Timekeeper helper functions. */
205 static inline s64 timekeeping_get_ns(struct timekeeper *tk)
207 cycle_t cycle_now, cycle_delta;
208 struct clocksource *clock;
211 /* read clocksource: */
213 cycle_now = clock->read(clock);
215 /* calculate the delta since the last update_wall_time: */
216 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
218 nsec = cycle_delta * tk->mult + tk->xtime_nsec;
221 /* If arch requires, add in gettimeoffset() */
222 return nsec + arch_gettimeoffset();
225 static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
227 cycle_t cycle_now, cycle_delta;
228 struct clocksource *clock;
231 /* read clocksource: */
233 cycle_now = clock->read(clock);
235 /* calculate the delta since the last update_wall_time: */
236 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
238 /* convert delta to nanoseconds. */
239 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
241 /* If arch requires, add in gettimeoffset() */
242 return nsec + arch_gettimeoffset();
245 /* must hold write on timekeeper.lock */
246 static void timekeeping_update(struct timekeeper *tk, bool clearntp)
255 update_vsyscall(&xt, &tk->wall_to_monotonic, tk->clock, tk->mult);
259 * timekeeping_forward_now - update clock to the current time
261 * Forward the current clock to update its state since the last call to
262 * update_wall_time(). This is useful before significant clock changes,
263 * as it avoids having to deal with this time offset explicitly.
265 static void timekeeping_forward_now(struct timekeeper *tk)
267 cycle_t cycle_now, cycle_delta;
268 struct clocksource *clock;
272 cycle_now = clock->read(clock);
273 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
274 clock->cycle_last = cycle_now;
276 tk->xtime_nsec += cycle_delta * tk->mult;
278 /* If arch requires, add in gettimeoffset() */
279 tk->xtime_nsec += arch_gettimeoffset() << tk->shift;
281 tk_normalize_xtime(tk);
283 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
284 timespec_add_ns(&tk->raw_time, nsec);
288 * getnstimeofday - Returns the time of day in a timespec
289 * @ts: pointer to the timespec to be set
291 * Returns the time of day in a timespec.
293 void getnstimeofday(struct timespec *ts)
295 struct timekeeper *tk = &timekeeper;
299 WARN_ON(timekeeping_suspended);
302 seq = read_seqbegin(&tk->lock);
304 ts->tv_sec = tk->xtime_sec;
305 ts->tv_nsec = timekeeping_get_ns(tk);
307 } while (read_seqretry(&tk->lock, seq));
309 timespec_add_ns(ts, nsecs);
311 EXPORT_SYMBOL(getnstimeofday);
313 ktime_t ktime_get(void)
315 struct timekeeper *tk = &timekeeper;
319 WARN_ON(timekeeping_suspended);
322 seq = read_seqbegin(&tk->lock);
323 secs = tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
324 nsecs = timekeeping_get_ns(tk) + tk->wall_to_monotonic.tv_nsec;
326 } while (read_seqretry(&tk->lock, seq));
328 * Use ktime_set/ktime_add_ns to create a proper ktime on
329 * 32-bit architectures without CONFIG_KTIME_SCALAR.
331 return ktime_add_ns(ktime_set(secs, 0), nsecs);
333 EXPORT_SYMBOL_GPL(ktime_get);
336 * ktime_get_ts - get the monotonic clock in timespec format
337 * @ts: pointer to timespec variable
339 * The function calculates the monotonic clock from the realtime
340 * clock and the wall_to_monotonic offset and stores the result
341 * in normalized timespec format in the variable pointed to by @ts.
343 void ktime_get_ts(struct timespec *ts)
345 struct timekeeper *tk = &timekeeper;
346 struct timespec tomono;
349 WARN_ON(timekeeping_suspended);
352 seq = read_seqbegin(&tk->lock);
353 ts->tv_sec = tk->xtime_sec;
354 ts->tv_nsec = timekeeping_get_ns(tk);
355 tomono = tk->wall_to_monotonic;
357 } while (read_seqretry(&tk->lock, seq));
359 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
360 ts->tv_nsec + tomono.tv_nsec);
362 EXPORT_SYMBOL_GPL(ktime_get_ts);
364 #ifdef CONFIG_NTP_PPS
367 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
368 * @ts_raw: pointer to the timespec to be set to raw monotonic time
369 * @ts_real: pointer to the timespec to be set to the time of day
371 * This function reads both the time of day and raw monotonic time at the
372 * same time atomically and stores the resulting timestamps in timespec
375 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
377 struct timekeeper *tk = &timekeeper;
379 s64 nsecs_raw, nsecs_real;
381 WARN_ON_ONCE(timekeeping_suspended);
384 seq = read_seqbegin(&tk->lock);
386 *ts_raw = tk->raw_time;
387 ts_real->tv_sec = tk->xtime_sec;
388 ts_real->tv_nsec = 0;
390 nsecs_raw = timekeeping_get_ns_raw(tk);
391 nsecs_real = timekeeping_get_ns(tk);
393 } while (read_seqretry(&tk->lock, seq));
395 timespec_add_ns(ts_raw, nsecs_raw);
396 timespec_add_ns(ts_real, nsecs_real);
398 EXPORT_SYMBOL(getnstime_raw_and_real);
400 #endif /* CONFIG_NTP_PPS */
403 * do_gettimeofday - Returns the time of day in a timeval
404 * @tv: pointer to the timeval to be set
406 * NOTE: Users should be converted to using getnstimeofday()
408 void do_gettimeofday(struct timeval *tv)
412 getnstimeofday(&now);
413 tv->tv_sec = now.tv_sec;
414 tv->tv_usec = now.tv_nsec/1000;
416 EXPORT_SYMBOL(do_gettimeofday);
419 * do_settimeofday - Sets the time of day
420 * @tv: pointer to the timespec variable containing the new time
422 * Sets the time of day to the new time and update NTP and notify hrtimers
424 int do_settimeofday(const struct timespec *tv)
426 struct timekeeper *tk = &timekeeper;
427 struct timespec ts_delta, xt;
430 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
433 write_seqlock_irqsave(&tk->lock, flags);
435 timekeeping_forward_now(tk);
438 ts_delta.tv_sec = tv->tv_sec - xt.tv_sec;
439 ts_delta.tv_nsec = tv->tv_nsec - xt.tv_nsec;
441 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, ts_delta));
443 tk_set_xtime(tk, tv);
445 timekeeping_update(tk, true);
447 write_sequnlock_irqrestore(&tk->lock, flags);
449 /* signal hrtimers about time change */
454 EXPORT_SYMBOL(do_settimeofday);
457 * timekeeping_inject_offset - Adds or subtracts from the current time.
458 * @tv: pointer to the timespec variable containing the offset
460 * Adds or subtracts an offset value from the current time.
462 int timekeeping_inject_offset(struct timespec *ts)
464 struct timekeeper *tk = &timekeeper;
467 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
470 write_seqlock_irqsave(&tk->lock, flags);
472 timekeeping_forward_now(tk);
475 tk_xtime_add(tk, ts);
476 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *ts));
478 timekeeping_update(tk, true);
480 write_sequnlock_irqrestore(&tk->lock, flags);
482 /* signal hrtimers about time change */
487 EXPORT_SYMBOL(timekeeping_inject_offset);
490 * change_clocksource - Swaps clocksources if a new one is available
492 * Accumulates current time interval and initializes new clocksource
494 static int change_clocksource(void *data)
496 struct timekeeper *tk = &timekeeper;
497 struct clocksource *new, *old;
500 new = (struct clocksource *) data;
502 write_seqlock_irqsave(&tk->lock, flags);
504 timekeeping_forward_now(tk);
505 if (!new->enable || new->enable(new) == 0) {
507 tk_setup_internals(tk, new);
511 timekeeping_update(tk, true);
513 write_sequnlock_irqrestore(&tk->lock, flags);
519 * timekeeping_notify - Install a new clock source
520 * @clock: pointer to the clock source
522 * This function is called from clocksource.c after a new, better clock
523 * source has been registered. The caller holds the clocksource_mutex.
525 void timekeeping_notify(struct clocksource *clock)
527 struct timekeeper *tk = &timekeeper;
529 if (tk->clock == clock)
531 stop_machine(change_clocksource, clock, NULL);
536 * ktime_get_real - get the real (wall-) time in ktime_t format
538 * returns the time in ktime_t format
540 ktime_t ktime_get_real(void)
544 getnstimeofday(&now);
546 return timespec_to_ktime(now);
548 EXPORT_SYMBOL_GPL(ktime_get_real);
551 * getrawmonotonic - Returns the raw monotonic time in a timespec
552 * @ts: pointer to the timespec to be set
554 * Returns the raw monotonic time (completely un-modified by ntp)
556 void getrawmonotonic(struct timespec *ts)
558 struct timekeeper *tk = &timekeeper;
563 seq = read_seqbegin(&tk->lock);
564 nsecs = timekeeping_get_ns_raw(tk);
567 } while (read_seqretry(&tk->lock, seq));
569 timespec_add_ns(ts, nsecs);
571 EXPORT_SYMBOL(getrawmonotonic);
574 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
576 int timekeeping_valid_for_hres(void)
578 struct timekeeper *tk = &timekeeper;
583 seq = read_seqbegin(&tk->lock);
585 ret = tk->clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
587 } while (read_seqretry(&tk->lock, seq));
593 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
595 u64 timekeeping_max_deferment(void)
597 struct timekeeper *tk = &timekeeper;
602 seq = read_seqbegin(&tk->lock);
604 ret = tk->clock->max_idle_ns;
606 } while (read_seqretry(&tk->lock, seq));
612 * read_persistent_clock - Return time from the persistent clock.
614 * Weak dummy function for arches that do not yet support it.
615 * Reads the time from the battery backed persistent clock.
616 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
618 * XXX - Do be sure to remove it once all arches implement it.
620 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
627 * read_boot_clock - Return time of the system start.
629 * Weak dummy function for arches that do not yet support it.
630 * Function to read the exact time the system has been started.
631 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
633 * XXX - Do be sure to remove it once all arches implement it.
635 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
642 * timekeeping_init - Initializes the clocksource and common timekeeping values
644 void __init timekeeping_init(void)
646 struct timekeeper *tk = &timekeeper;
647 struct clocksource *clock;
649 struct timespec now, boot, tmp;
651 read_persistent_clock(&now);
652 read_boot_clock(&boot);
654 seqlock_init(&tk->lock);
658 write_seqlock_irqsave(&tk->lock, flags);
659 clock = clocksource_default_clock();
661 clock->enable(clock);
662 tk_setup_internals(tk, clock);
664 tk_set_xtime(tk, &now);
665 tk->raw_time.tv_sec = 0;
666 tk->raw_time.tv_nsec = 0;
667 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
670 set_normalized_timespec(&tmp, -boot.tv_sec, -boot.tv_nsec);
671 tk_set_wall_to_mono(tk, tmp);
675 tk_set_sleep_time(tk, tmp);
677 write_sequnlock_irqrestore(&tk->lock, flags);
680 /* time in seconds when suspend began */
681 static struct timespec timekeeping_suspend_time;
684 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
685 * @delta: pointer to a timespec delta value
687 * Takes a timespec offset measuring a suspend interval and properly
688 * adds the sleep offset to the timekeeping variables.
690 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
691 struct timespec *delta)
693 if (!timespec_valid(delta)) {
694 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
695 "sleep delta value!\n");
698 tk_xtime_add(tk, delta);
699 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *delta));
700 tk_set_sleep_time(tk, timespec_add(tk->total_sleep_time, *delta));
704 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
705 * @delta: pointer to a timespec delta value
707 * This hook is for architectures that cannot support read_persistent_clock
708 * because their RTC/persistent clock is only accessible when irqs are enabled.
710 * This function should only be called by rtc_resume(), and allows
711 * a suspend offset to be injected into the timekeeping values.
713 void timekeeping_inject_sleeptime(struct timespec *delta)
715 struct timekeeper *tk = &timekeeper;
719 /* Make sure we don't set the clock twice */
720 read_persistent_clock(&ts);
721 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
724 write_seqlock_irqsave(&tk->lock, flags);
726 timekeeping_forward_now(tk);
728 __timekeeping_inject_sleeptime(tk, delta);
730 timekeeping_update(tk, true);
732 write_sequnlock_irqrestore(&tk->lock, flags);
734 /* signal hrtimers about time change */
739 * timekeeping_resume - Resumes the generic timekeeping subsystem.
741 * This is for the generic clocksource timekeeping.
742 * xtime/wall_to_monotonic/jiffies/etc are
743 * still managed by arch specific suspend/resume code.
745 static void timekeeping_resume(void)
747 struct timekeeper *tk = &timekeeper;
751 read_persistent_clock(&ts);
753 clocksource_resume();
755 write_seqlock_irqsave(&tk->lock, flags);
757 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
758 ts = timespec_sub(ts, timekeeping_suspend_time);
759 __timekeeping_inject_sleeptime(tk, &ts);
761 /* re-base the last cycle value */
762 tk->clock->cycle_last = tk->clock->read(tk->clock);
764 timekeeping_suspended = 0;
765 timekeeping_update(tk, false);
766 write_sequnlock_irqrestore(&tk->lock, flags);
768 touch_softlockup_watchdog();
770 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
772 /* Resume hrtimers */
776 static int timekeeping_suspend(void)
778 struct timekeeper *tk = &timekeeper;
780 struct timespec delta, delta_delta;
781 static struct timespec old_delta;
783 read_persistent_clock(&timekeeping_suspend_time);
785 write_seqlock_irqsave(&tk->lock, flags);
786 timekeeping_forward_now(tk);
787 timekeeping_suspended = 1;
790 * To avoid drift caused by repeated suspend/resumes,
791 * which each can add ~1 second drift error,
792 * try to compensate so the difference in system time
793 * and persistent_clock time stays close to constant.
795 delta = timespec_sub(tk_xtime(tk), timekeeping_suspend_time);
796 delta_delta = timespec_sub(delta, old_delta);
797 if (abs(delta_delta.tv_sec) >= 2) {
799 * if delta_delta is too large, assume time correction
800 * has occured and set old_delta to the current delta.
804 /* Otherwise try to adjust old_system to compensate */
805 timekeeping_suspend_time =
806 timespec_add(timekeeping_suspend_time, delta_delta);
808 write_sequnlock_irqrestore(&tk->lock, flags);
810 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
811 clocksource_suspend();
816 /* sysfs resume/suspend bits for timekeeping */
817 static struct syscore_ops timekeeping_syscore_ops = {
818 .resume = timekeeping_resume,
819 .suspend = timekeeping_suspend,
822 static int __init timekeeping_init_ops(void)
824 register_syscore_ops(&timekeeping_syscore_ops);
828 device_initcall(timekeeping_init_ops);
831 * If the error is already larger, we look ahead even further
832 * to compensate for late or lost adjustments.
834 static __always_inline int timekeeping_bigadjust(struct timekeeper *tk,
835 s64 error, s64 *interval,
843 * Use the current error value to determine how much to look ahead.
844 * The larger the error the slower we adjust for it to avoid problems
845 * with losing too many ticks, otherwise we would overadjust and
846 * produce an even larger error. The smaller the adjustment the
847 * faster we try to adjust for it, as lost ticks can do less harm
848 * here. This is tuned so that an error of about 1 msec is adjusted
849 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
851 error2 = tk->ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
852 error2 = abs(error2);
853 for (look_ahead = 0; error2 > 0; look_ahead++)
857 * Now calculate the error in (1 << look_ahead) ticks, but first
858 * remove the single look ahead already included in the error.
860 tick_error = ntp_tick_length() >> (tk->ntp_error_shift + 1);
861 tick_error -= tk->xtime_interval >> 1;
862 error = ((error - tick_error) >> look_ahead) + tick_error;
864 /* Finally calculate the adjustment shift value. */
869 *interval = -*interval;
873 for (adj = 0; error > i; adj++)
882 * Adjust the multiplier to reduce the error value,
883 * this is optimized for the most common adjustments of -1,0,1,
884 * for other values we can do a bit more work.
886 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
888 s64 error, interval = tk->cycle_interval;
892 * The point of this is to check if the error is greater than half
895 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
897 * Note we subtract one in the shift, so that error is really error*2.
898 * This "saves" dividing(shifting) interval twice, but keeps the
899 * (error > interval) comparison as still measuring if error is
900 * larger than half an interval.
902 * Note: It does not "save" on aggravation when reading the code.
904 error = tk->ntp_error >> (tk->ntp_error_shift - 1);
905 if (error > interval) {
907 * We now divide error by 4(via shift), which checks if
908 * the error is greater than twice the interval.
909 * If it is greater, we need a bigadjust, if its smaller,
910 * we can adjust by 1.
914 * XXX - In update_wall_time, we round up to the next
915 * nanosecond, and store the amount rounded up into
916 * the error. This causes the likely below to be unlikely.
918 * The proper fix is to avoid rounding up by using
919 * the high precision tk->xtime_nsec instead of
920 * xtime.tv_nsec everywhere. Fixing this will take some
923 if (likely(error <= interval))
926 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
928 if (error < -interval) {
929 /* See comment above, this is just switched for the negative */
931 if (likely(error >= -interval)) {
933 interval = -interval;
936 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
943 if (unlikely(tk->clock->maxadj &&
944 (tk->mult + adj > tk->clock->mult + tk->clock->maxadj))) {
945 printk_once(KERN_WARNING
946 "Adjusting %s more than 11%% (%ld vs %ld)\n",
947 tk->clock->name, (long)tk->mult + adj,
948 (long)tk->clock->mult + tk->clock->maxadj);
951 * So the following can be confusing.
953 * To keep things simple, lets assume adj == 1 for now.
955 * When adj != 1, remember that the interval and offset values
956 * have been appropriately scaled so the math is the same.
958 * The basic idea here is that we're increasing the multiplier
959 * by one, this causes the xtime_interval to be incremented by
960 * one cycle_interval. This is because:
961 * xtime_interval = cycle_interval * mult
962 * So if mult is being incremented by one:
963 * xtime_interval = cycle_interval * (mult + 1)
965 * xtime_interval = (cycle_interval * mult) + cycle_interval
966 * Which can be shortened to:
967 * xtime_interval += cycle_interval
969 * So offset stores the non-accumulated cycles. Thus the current
970 * time (in shifted nanoseconds) is:
971 * now = (offset * adj) + xtime_nsec
972 * Now, even though we're adjusting the clock frequency, we have
973 * to keep time consistent. In other words, we can't jump back
974 * in time, and we also want to avoid jumping forward in time.
976 * So given the same offset value, we need the time to be the same
977 * both before and after the freq adjustment.
978 * now = (offset * adj_1) + xtime_nsec_1
979 * now = (offset * adj_2) + xtime_nsec_2
981 * (offset * adj_1) + xtime_nsec_1 =
982 * (offset * adj_2) + xtime_nsec_2
986 * (offset * adj_1) + xtime_nsec_1 =
987 * (offset * (adj_1+1)) + xtime_nsec_2
988 * (offset * adj_1) + xtime_nsec_1 =
989 * (offset * adj_1) + offset + xtime_nsec_2
990 * Canceling the sides:
991 * xtime_nsec_1 = offset + xtime_nsec_2
993 * xtime_nsec_2 = xtime_nsec_1 - offset
994 * Which simplfies to:
995 * xtime_nsec -= offset
997 * XXX - TODO: Doc ntp_error calculation.
1000 tk->xtime_interval += interval;
1001 tk->xtime_nsec -= offset;
1002 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1006 * It may be possible that when we entered this function, xtime_nsec
1007 * was very small. Further, if we're slightly speeding the clocksource
1008 * in the code above, its possible the required corrective factor to
1009 * xtime_nsec could cause it to underflow.
1011 * Now, since we already accumulated the second, cannot simply roll
1012 * the accumulated second back, since the NTP subsystem has been
1013 * notified via second_overflow. So instead we push xtime_nsec forward
1014 * by the amount we underflowed, and add that amount into the error.
1016 * We'll correct this error next time through this function, when
1017 * xtime_nsec is not as small.
1019 if (unlikely((s64)tk->xtime_nsec < 0)) {
1020 s64 neg = -(s64)tk->xtime_nsec;
1022 tk->ntp_error += neg << tk->ntp_error_shift;
1028 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1030 * Helper function that accumulates a the nsecs greater then a second
1031 * from the xtime_nsec field to the xtime_secs field.
1032 * It also calls into the NTP code to handle leapsecond processing.
1035 static inline void accumulate_nsecs_to_secs(struct timekeeper *tk)
1037 u64 nsecps = (u64)NSEC_PER_SEC << tk->shift;
1039 while (tk->xtime_nsec >= nsecps) {
1042 tk->xtime_nsec -= nsecps;
1045 /* Figure out if its a leap sec and apply if needed */
1046 leap = second_overflow(tk->xtime_sec);
1047 if (unlikely(leap)) {
1050 tk->xtime_sec += leap;
1054 tk_set_wall_to_mono(tk,
1055 timespec_sub(tk->wall_to_monotonic, ts));
1057 clock_was_set_delayed();
1063 * logarithmic_accumulation - shifted accumulation of cycles
1065 * This functions accumulates a shifted interval of cycles into
1066 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1069 * Returns the unconsumed cycles.
1071 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1076 /* If the offset is smaller then a shifted interval, do nothing */
1077 if (offset < tk->cycle_interval<<shift)
1080 /* Accumulate one shifted interval */
1081 offset -= tk->cycle_interval << shift;
1082 tk->clock->cycle_last += tk->cycle_interval << shift;
1084 tk->xtime_nsec += tk->xtime_interval << shift;
1085 accumulate_nsecs_to_secs(tk);
1087 /* Accumulate raw time */
1088 raw_nsecs = tk->raw_interval << shift;
1089 raw_nsecs += tk->raw_time.tv_nsec;
1090 if (raw_nsecs >= NSEC_PER_SEC) {
1091 u64 raw_secs = raw_nsecs;
1092 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1093 tk->raw_time.tv_sec += raw_secs;
1095 tk->raw_time.tv_nsec = raw_nsecs;
1097 /* Accumulate error between NTP and clock interval */
1098 tk->ntp_error += ntp_tick_length() << shift;
1099 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1100 (tk->ntp_error_shift + shift);
1106 * update_wall_time - Uses the current clocksource to increment the wall time
1109 static void update_wall_time(void)
1111 struct clocksource *clock;
1112 struct timekeeper *tk = &timekeeper;
1114 int shift = 0, maxshift;
1115 unsigned long flags;
1118 write_seqlock_irqsave(&tk->lock, flags);
1120 /* Make sure we're fully resumed: */
1121 if (unlikely(timekeeping_suspended))
1126 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1127 offset = tk->cycle_interval;
1129 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1133 * With NO_HZ we may have to accumulate many cycle_intervals
1134 * (think "ticks") worth of time at once. To do this efficiently,
1135 * we calculate the largest doubling multiple of cycle_intervals
1136 * that is smaller than the offset. We then accumulate that
1137 * chunk in one go, and then try to consume the next smaller
1140 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1141 shift = max(0, shift);
1142 /* Bound shift to one less than what overflows tick_length */
1143 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1144 shift = min(shift, maxshift);
1145 while (offset >= tk->cycle_interval) {
1146 offset = logarithmic_accumulation(tk, offset, shift);
1147 if (offset < tk->cycle_interval<<shift)
1151 /* correct the clock when NTP error is too big */
1152 timekeeping_adjust(tk, offset);
1156 * Store only full nanoseconds into xtime_nsec after rounding
1157 * it up and add the remainder to the error difference.
1158 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
1159 * by truncating the remainder in vsyscalls. However, it causes
1160 * additional work to be done in timekeeping_adjust(). Once
1161 * the vsyscall implementations are converted to use xtime_nsec
1162 * (shifted nanoseconds), this can be killed.
1164 remainder = tk->xtime_nsec & ((1 << tk->shift) - 1);
1165 tk->xtime_nsec -= remainder;
1166 tk->xtime_nsec += 1 << tk->shift;
1167 tk->ntp_error += remainder << tk->ntp_error_shift;
1170 * Finally, make sure that after the rounding
1171 * xtime_nsec isn't larger than NSEC_PER_SEC
1173 accumulate_nsecs_to_secs(tk);
1175 timekeeping_update(tk, false);
1178 write_sequnlock_irqrestore(&tk->lock, flags);
1183 * getboottime - Return the real time of system boot.
1184 * @ts: pointer to the timespec to be set
1186 * Returns the wall-time of boot in a timespec.
1188 * This is based on the wall_to_monotonic offset and the total suspend
1189 * time. Calls to settimeofday will affect the value returned (which
1190 * basically means that however wrong your real time clock is at boot time,
1191 * you get the right time here).
1193 void getboottime(struct timespec *ts)
1195 struct timekeeper *tk = &timekeeper;
1196 struct timespec boottime = {
1197 .tv_sec = tk->wall_to_monotonic.tv_sec +
1198 tk->total_sleep_time.tv_sec,
1199 .tv_nsec = tk->wall_to_monotonic.tv_nsec +
1200 tk->total_sleep_time.tv_nsec
1203 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1205 EXPORT_SYMBOL_GPL(getboottime);
1208 * get_monotonic_boottime - Returns monotonic time since boot
1209 * @ts: pointer to the timespec to be set
1211 * Returns the monotonic time since boot in a timespec.
1213 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1214 * includes the time spent in suspend.
1216 void get_monotonic_boottime(struct timespec *ts)
1218 struct timekeeper *tk = &timekeeper;
1219 struct timespec tomono, sleep;
1222 WARN_ON(timekeeping_suspended);
1225 seq = read_seqbegin(&tk->lock);
1226 ts->tv_sec = tk->xtime_sec;
1227 ts->tv_nsec = timekeeping_get_ns(tk);
1228 tomono = tk->wall_to_monotonic;
1229 sleep = tk->total_sleep_time;
1231 } while (read_seqretry(&tk->lock, seq));
1233 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1234 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec);
1236 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1239 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1241 * Returns the monotonic time since boot in a ktime
1243 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1244 * includes the time spent in suspend.
1246 ktime_t ktime_get_boottime(void)
1250 get_monotonic_boottime(&ts);
1251 return timespec_to_ktime(ts);
1253 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1256 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1257 * @ts: pointer to the timespec to be converted
1259 void monotonic_to_bootbased(struct timespec *ts)
1261 struct timekeeper *tk = &timekeeper;
1263 *ts = timespec_add(*ts, tk->total_sleep_time);
1265 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1267 unsigned long get_seconds(void)
1269 struct timekeeper *tk = &timekeeper;
1271 return tk->xtime_sec;
1273 EXPORT_SYMBOL(get_seconds);
1275 struct timespec __current_kernel_time(void)
1277 struct timekeeper *tk = &timekeeper;
1279 return tk_xtime(tk);
1282 struct timespec current_kernel_time(void)
1284 struct timekeeper *tk = &timekeeper;
1285 struct timespec now;
1289 seq = read_seqbegin(&tk->lock);
1292 } while (read_seqretry(&tk->lock, seq));
1296 EXPORT_SYMBOL(current_kernel_time);
1298 struct timespec get_monotonic_coarse(void)
1300 struct timekeeper *tk = &timekeeper;
1301 struct timespec now, mono;
1305 seq = read_seqbegin(&tk->lock);
1308 mono = tk->wall_to_monotonic;
1309 } while (read_seqretry(&tk->lock, seq));
1311 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1312 now.tv_nsec + mono.tv_nsec);
1317 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1318 * without sampling the sequence number in xtime_lock.
1319 * jiffies is defined in the linker script...
1321 void do_timer(unsigned long ticks)
1323 jiffies_64 += ticks;
1325 calc_global_load(ticks);
1329 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1330 * and sleep offsets.
1331 * @xtim: pointer to timespec to be set with xtime
1332 * @wtom: pointer to timespec to be set with wall_to_monotonic
1333 * @sleep: pointer to timespec to be set with time in suspend
1335 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1336 struct timespec *wtom, struct timespec *sleep)
1338 struct timekeeper *tk = &timekeeper;
1342 seq = read_seqbegin(&tk->lock);
1343 *xtim = tk_xtime(tk);
1344 *wtom = tk->wall_to_monotonic;
1345 *sleep = tk->total_sleep_time;
1346 } while (read_seqretry(&tk->lock, seq));
1349 #ifdef CONFIG_HIGH_RES_TIMERS
1351 * ktime_get_update_offsets - hrtimer helper
1352 * @offs_real: pointer to storage for monotonic -> realtime offset
1353 * @offs_boot: pointer to storage for monotonic -> boottime offset
1355 * Returns current monotonic time and updates the offsets
1356 * Called from hrtimer_interupt() or retrigger_next_event()
1358 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1360 struct timekeeper *tk = &timekeeper;
1366 seq = read_seqbegin(&tk->lock);
1368 secs = tk->xtime_sec;
1369 nsecs = timekeeping_get_ns(tk);
1371 *offs_real = tk->offs_real;
1372 *offs_boot = tk->offs_boot;
1373 } while (read_seqretry(&tk->lock, seq));
1375 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1376 now = ktime_sub(now, *offs_real);
1382 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1384 ktime_t ktime_get_monotonic_offset(void)
1386 struct timekeeper *tk = &timekeeper;
1388 struct timespec wtom;
1391 seq = read_seqbegin(&tk->lock);
1392 wtom = tk->wall_to_monotonic;
1393 } while (read_seqretry(&tk->lock, seq));
1395 return timespec_to_ktime(wtom);
1397 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1400 * xtime_update() - advances the timekeeping infrastructure
1401 * @ticks: number of ticks, that have elapsed since the last call.
1403 * Must be called with interrupts disabled.
1405 void xtime_update(unsigned long ticks)
1407 write_seqlock(&xtime_lock);
1409 write_sequnlock(&xtime_lock);