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 (!timespec_valid(tv))
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;
469 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
472 write_seqlock_irqsave(&tk->lock, flags);
474 timekeeping_forward_now(tk);
476 /* Make sure the proposed value is valid */
477 tmp = timespec_add(tk_xtime(tk), *ts);
478 if (!timespec_valid(&tmp)) {
483 tk_xtime_add(tk, ts);
484 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *ts));
486 error: /* even if we error out, we forwarded the time, so call update */
487 timekeeping_update(tk, true);
489 write_sequnlock_irqrestore(&tk->lock, flags);
491 /* signal hrtimers about time change */
496 EXPORT_SYMBOL(timekeeping_inject_offset);
499 * change_clocksource - Swaps clocksources if a new one is available
501 * Accumulates current time interval and initializes new clocksource
503 static int change_clocksource(void *data)
505 struct timekeeper *tk = &timekeeper;
506 struct clocksource *new, *old;
509 new = (struct clocksource *) data;
511 write_seqlock_irqsave(&tk->lock, flags);
513 timekeeping_forward_now(tk);
514 if (!new->enable || new->enable(new) == 0) {
516 tk_setup_internals(tk, new);
520 timekeeping_update(tk, true);
522 write_sequnlock_irqrestore(&tk->lock, flags);
528 * timekeeping_notify - Install a new clock source
529 * @clock: pointer to the clock source
531 * This function is called from clocksource.c after a new, better clock
532 * source has been registered. The caller holds the clocksource_mutex.
534 void timekeeping_notify(struct clocksource *clock)
536 struct timekeeper *tk = &timekeeper;
538 if (tk->clock == clock)
540 stop_machine(change_clocksource, clock, NULL);
545 * ktime_get_real - get the real (wall-) time in ktime_t format
547 * returns the time in ktime_t format
549 ktime_t ktime_get_real(void)
553 getnstimeofday(&now);
555 return timespec_to_ktime(now);
557 EXPORT_SYMBOL_GPL(ktime_get_real);
560 * getrawmonotonic - Returns the raw monotonic time in a timespec
561 * @ts: pointer to the timespec to be set
563 * Returns the raw monotonic time (completely un-modified by ntp)
565 void getrawmonotonic(struct timespec *ts)
567 struct timekeeper *tk = &timekeeper;
572 seq = read_seqbegin(&tk->lock);
573 nsecs = timekeeping_get_ns_raw(tk);
576 } while (read_seqretry(&tk->lock, seq));
578 timespec_add_ns(ts, nsecs);
580 EXPORT_SYMBOL(getrawmonotonic);
583 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
585 int timekeeping_valid_for_hres(void)
587 struct timekeeper *tk = &timekeeper;
592 seq = read_seqbegin(&tk->lock);
594 ret = tk->clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
596 } while (read_seqretry(&tk->lock, seq));
602 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
604 u64 timekeeping_max_deferment(void)
606 struct timekeeper *tk = &timekeeper;
611 seq = read_seqbegin(&tk->lock);
613 ret = tk->clock->max_idle_ns;
615 } while (read_seqretry(&tk->lock, seq));
621 * read_persistent_clock - Return time from the persistent clock.
623 * Weak dummy function for arches that do not yet support it.
624 * Reads the time from the battery backed persistent clock.
625 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
627 * XXX - Do be sure to remove it once all arches implement it.
629 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
636 * read_boot_clock - Return time of the system start.
638 * Weak dummy function for arches that do not yet support it.
639 * Function to read the exact time the system has been started.
640 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
642 * XXX - Do be sure to remove it once all arches implement it.
644 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
651 * timekeeping_init - Initializes the clocksource and common timekeeping values
653 void __init timekeeping_init(void)
655 struct timekeeper *tk = &timekeeper;
656 struct clocksource *clock;
658 struct timespec now, boot, tmp;
660 read_persistent_clock(&now);
661 if (!timespec_valid(&now)) {
662 pr_warn("WARNING: Persistent clock returned invalid value!\n"
663 " Check your CMOS/BIOS settings.\n");
668 read_boot_clock(&boot);
669 if (!timespec_valid(&boot)) {
670 pr_warn("WARNING: Boot clock returned invalid value!\n"
671 " Check your CMOS/BIOS settings.\n");
676 seqlock_init(&tk->lock);
680 write_seqlock_irqsave(&tk->lock, flags);
681 clock = clocksource_default_clock();
683 clock->enable(clock);
684 tk_setup_internals(tk, clock);
686 tk_set_xtime(tk, &now);
687 tk->raw_time.tv_sec = 0;
688 tk->raw_time.tv_nsec = 0;
689 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
692 set_normalized_timespec(&tmp, -boot.tv_sec, -boot.tv_nsec);
693 tk_set_wall_to_mono(tk, tmp);
697 tk_set_sleep_time(tk, tmp);
699 write_sequnlock_irqrestore(&tk->lock, flags);
702 /* time in seconds when suspend began */
703 static struct timespec timekeeping_suspend_time;
706 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
707 * @delta: pointer to a timespec delta value
709 * Takes a timespec offset measuring a suspend interval and properly
710 * adds the sleep offset to the timekeeping variables.
712 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
713 struct timespec *delta)
715 if (!timespec_valid(delta)) {
716 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
717 "sleep delta value!\n");
720 tk_xtime_add(tk, delta);
721 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *delta));
722 tk_set_sleep_time(tk, timespec_add(tk->total_sleep_time, *delta));
726 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
727 * @delta: pointer to a timespec delta value
729 * This hook is for architectures that cannot support read_persistent_clock
730 * because their RTC/persistent clock is only accessible when irqs are enabled.
732 * This function should only be called by rtc_resume(), and allows
733 * a suspend offset to be injected into the timekeeping values.
735 void timekeeping_inject_sleeptime(struct timespec *delta)
737 struct timekeeper *tk = &timekeeper;
741 /* Make sure we don't set the clock twice */
742 read_persistent_clock(&ts);
743 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
746 write_seqlock_irqsave(&tk->lock, flags);
748 timekeeping_forward_now(tk);
750 __timekeeping_inject_sleeptime(tk, delta);
752 timekeeping_update(tk, true);
754 write_sequnlock_irqrestore(&tk->lock, flags);
756 /* signal hrtimers about time change */
761 * timekeeping_resume - Resumes the generic timekeeping subsystem.
763 * This is for the generic clocksource timekeeping.
764 * xtime/wall_to_monotonic/jiffies/etc are
765 * still managed by arch specific suspend/resume code.
767 static void timekeeping_resume(void)
769 struct timekeeper *tk = &timekeeper;
773 read_persistent_clock(&ts);
775 clocksource_resume();
777 write_seqlock_irqsave(&tk->lock, flags);
779 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
780 ts = timespec_sub(ts, timekeeping_suspend_time);
781 __timekeeping_inject_sleeptime(tk, &ts);
783 /* re-base the last cycle value */
784 tk->clock->cycle_last = tk->clock->read(tk->clock);
786 timekeeping_suspended = 0;
787 timekeeping_update(tk, false);
788 write_sequnlock_irqrestore(&tk->lock, flags);
790 touch_softlockup_watchdog();
792 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
794 /* Resume hrtimers */
798 static int timekeeping_suspend(void)
800 struct timekeeper *tk = &timekeeper;
802 struct timespec delta, delta_delta;
803 static struct timespec old_delta;
805 read_persistent_clock(&timekeeping_suspend_time);
807 write_seqlock_irqsave(&tk->lock, flags);
808 timekeeping_forward_now(tk);
809 timekeeping_suspended = 1;
812 * To avoid drift caused by repeated suspend/resumes,
813 * which each can add ~1 second drift error,
814 * try to compensate so the difference in system time
815 * and persistent_clock time stays close to constant.
817 delta = timespec_sub(tk_xtime(tk), timekeeping_suspend_time);
818 delta_delta = timespec_sub(delta, old_delta);
819 if (abs(delta_delta.tv_sec) >= 2) {
821 * if delta_delta is too large, assume time correction
822 * has occured and set old_delta to the current delta.
826 /* Otherwise try to adjust old_system to compensate */
827 timekeeping_suspend_time =
828 timespec_add(timekeeping_suspend_time, delta_delta);
830 write_sequnlock_irqrestore(&tk->lock, flags);
832 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
833 clocksource_suspend();
838 /* sysfs resume/suspend bits for timekeeping */
839 static struct syscore_ops timekeeping_syscore_ops = {
840 .resume = timekeeping_resume,
841 .suspend = timekeeping_suspend,
844 static int __init timekeeping_init_ops(void)
846 register_syscore_ops(&timekeeping_syscore_ops);
850 device_initcall(timekeeping_init_ops);
853 * If the error is already larger, we look ahead even further
854 * to compensate for late or lost adjustments.
856 static __always_inline int timekeeping_bigadjust(struct timekeeper *tk,
857 s64 error, s64 *interval,
865 * Use the current error value to determine how much to look ahead.
866 * The larger the error the slower we adjust for it to avoid problems
867 * with losing too many ticks, otherwise we would overadjust and
868 * produce an even larger error. The smaller the adjustment the
869 * faster we try to adjust for it, as lost ticks can do less harm
870 * here. This is tuned so that an error of about 1 msec is adjusted
871 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
873 error2 = tk->ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
874 error2 = abs(error2);
875 for (look_ahead = 0; error2 > 0; look_ahead++)
879 * Now calculate the error in (1 << look_ahead) ticks, but first
880 * remove the single look ahead already included in the error.
882 tick_error = ntp_tick_length() >> (tk->ntp_error_shift + 1);
883 tick_error -= tk->xtime_interval >> 1;
884 error = ((error - tick_error) >> look_ahead) + tick_error;
886 /* Finally calculate the adjustment shift value. */
891 *interval = -*interval;
895 for (adj = 0; error > i; adj++)
904 * Adjust the multiplier to reduce the error value,
905 * this is optimized for the most common adjustments of -1,0,1,
906 * for other values we can do a bit more work.
908 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
910 s64 error, interval = tk->cycle_interval;
914 * The point of this is to check if the error is greater than half
917 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
919 * Note we subtract one in the shift, so that error is really error*2.
920 * This "saves" dividing(shifting) interval twice, but keeps the
921 * (error > interval) comparison as still measuring if error is
922 * larger than half an interval.
924 * Note: It does not "save" on aggravation when reading the code.
926 error = tk->ntp_error >> (tk->ntp_error_shift - 1);
927 if (error > interval) {
929 * We now divide error by 4(via shift), which checks if
930 * the error is greater than twice the interval.
931 * If it is greater, we need a bigadjust, if its smaller,
932 * we can adjust by 1.
936 * XXX - In update_wall_time, we round up to the next
937 * nanosecond, and store the amount rounded up into
938 * the error. This causes the likely below to be unlikely.
940 * The proper fix is to avoid rounding up by using
941 * the high precision tk->xtime_nsec instead of
942 * xtime.tv_nsec everywhere. Fixing this will take some
945 if (likely(error <= interval))
948 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
950 if (error < -interval) {
951 /* See comment above, this is just switched for the negative */
953 if (likely(error >= -interval)) {
955 interval = -interval;
958 adj = timekeeping_bigadjust(tk, error, &interval, &offset);
965 if (unlikely(tk->clock->maxadj &&
966 (tk->mult + adj > tk->clock->mult + tk->clock->maxadj))) {
967 printk_once(KERN_WARNING
968 "Adjusting %s more than 11%% (%ld vs %ld)\n",
969 tk->clock->name, (long)tk->mult + adj,
970 (long)tk->clock->mult + tk->clock->maxadj);
973 * So the following can be confusing.
975 * To keep things simple, lets assume adj == 1 for now.
977 * When adj != 1, remember that the interval and offset values
978 * have been appropriately scaled so the math is the same.
980 * The basic idea here is that we're increasing the multiplier
981 * by one, this causes the xtime_interval to be incremented by
982 * one cycle_interval. This is because:
983 * xtime_interval = cycle_interval * mult
984 * So if mult is being incremented by one:
985 * xtime_interval = cycle_interval * (mult + 1)
987 * xtime_interval = (cycle_interval * mult) + cycle_interval
988 * Which can be shortened to:
989 * xtime_interval += cycle_interval
991 * So offset stores the non-accumulated cycles. Thus the current
992 * time (in shifted nanoseconds) is:
993 * now = (offset * adj) + xtime_nsec
994 * Now, even though we're adjusting the clock frequency, we have
995 * to keep time consistent. In other words, we can't jump back
996 * in time, and we also want to avoid jumping forward in time.
998 * So given the same offset value, we need the time to be the same
999 * both before and after the freq adjustment.
1000 * now = (offset * adj_1) + xtime_nsec_1
1001 * now = (offset * adj_2) + xtime_nsec_2
1003 * (offset * adj_1) + xtime_nsec_1 =
1004 * (offset * adj_2) + xtime_nsec_2
1008 * (offset * adj_1) + xtime_nsec_1 =
1009 * (offset * (adj_1+1)) + xtime_nsec_2
1010 * (offset * adj_1) + xtime_nsec_1 =
1011 * (offset * adj_1) + offset + xtime_nsec_2
1012 * Canceling the sides:
1013 * xtime_nsec_1 = offset + xtime_nsec_2
1015 * xtime_nsec_2 = xtime_nsec_1 - offset
1016 * Which simplfies to:
1017 * xtime_nsec -= offset
1019 * XXX - TODO: Doc ntp_error calculation.
1022 tk->xtime_interval += interval;
1023 tk->xtime_nsec -= offset;
1024 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1028 * It may be possible that when we entered this function, xtime_nsec
1029 * was very small. Further, if we're slightly speeding the clocksource
1030 * in the code above, its possible the required corrective factor to
1031 * xtime_nsec could cause it to underflow.
1033 * Now, since we already accumulated the second, cannot simply roll
1034 * the accumulated second back, since the NTP subsystem has been
1035 * notified via second_overflow. So instead we push xtime_nsec forward
1036 * by the amount we underflowed, and add that amount into the error.
1038 * We'll correct this error next time through this function, when
1039 * xtime_nsec is not as small.
1041 if (unlikely((s64)tk->xtime_nsec < 0)) {
1042 s64 neg = -(s64)tk->xtime_nsec;
1044 tk->ntp_error += neg << tk->ntp_error_shift;
1050 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1052 * Helper function that accumulates a the nsecs greater then a second
1053 * from the xtime_nsec field to the xtime_secs field.
1054 * It also calls into the NTP code to handle leapsecond processing.
1057 static inline void accumulate_nsecs_to_secs(struct timekeeper *tk)
1059 u64 nsecps = (u64)NSEC_PER_SEC << tk->shift;
1061 while (tk->xtime_nsec >= nsecps) {
1064 tk->xtime_nsec -= nsecps;
1067 /* Figure out if its a leap sec and apply if needed */
1068 leap = second_overflow(tk->xtime_sec);
1069 if (unlikely(leap)) {
1072 tk->xtime_sec += leap;
1076 tk_set_wall_to_mono(tk,
1077 timespec_sub(tk->wall_to_monotonic, ts));
1079 clock_was_set_delayed();
1085 * logarithmic_accumulation - shifted accumulation of cycles
1087 * This functions accumulates a shifted interval of cycles into
1088 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1091 * Returns the unconsumed cycles.
1093 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1098 /* If the offset is smaller then a shifted interval, do nothing */
1099 if (offset < tk->cycle_interval<<shift)
1102 /* Accumulate one shifted interval */
1103 offset -= tk->cycle_interval << shift;
1104 tk->clock->cycle_last += tk->cycle_interval << shift;
1106 tk->xtime_nsec += tk->xtime_interval << shift;
1107 accumulate_nsecs_to_secs(tk);
1109 /* Accumulate raw time */
1110 raw_nsecs = tk->raw_interval << shift;
1111 raw_nsecs += tk->raw_time.tv_nsec;
1112 if (raw_nsecs >= NSEC_PER_SEC) {
1113 u64 raw_secs = raw_nsecs;
1114 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1115 tk->raw_time.tv_sec += raw_secs;
1117 tk->raw_time.tv_nsec = raw_nsecs;
1119 /* Accumulate error between NTP and clock interval */
1120 tk->ntp_error += ntp_tick_length() << shift;
1121 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1122 (tk->ntp_error_shift + shift);
1128 * update_wall_time - Uses the current clocksource to increment the wall time
1131 static void update_wall_time(void)
1133 struct clocksource *clock;
1134 struct timekeeper *tk = &timekeeper;
1136 int shift = 0, maxshift;
1137 unsigned long flags;
1140 write_seqlock_irqsave(&tk->lock, flags);
1142 /* Make sure we're fully resumed: */
1143 if (unlikely(timekeeping_suspended))
1148 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1149 offset = tk->cycle_interval;
1151 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1155 * With NO_HZ we may have to accumulate many cycle_intervals
1156 * (think "ticks") worth of time at once. To do this efficiently,
1157 * we calculate the largest doubling multiple of cycle_intervals
1158 * that is smaller than the offset. We then accumulate that
1159 * chunk in one go, and then try to consume the next smaller
1162 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1163 shift = max(0, shift);
1164 /* Bound shift to one less than what overflows tick_length */
1165 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1166 shift = min(shift, maxshift);
1167 while (offset >= tk->cycle_interval) {
1168 offset = logarithmic_accumulation(tk, offset, shift);
1169 if (offset < tk->cycle_interval<<shift)
1173 /* correct the clock when NTP error is too big */
1174 timekeeping_adjust(tk, offset);
1178 * Store only full nanoseconds into xtime_nsec after rounding
1179 * it up and add the remainder to the error difference.
1180 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
1181 * by truncating the remainder in vsyscalls. However, it causes
1182 * additional work to be done in timekeeping_adjust(). Once
1183 * the vsyscall implementations are converted to use xtime_nsec
1184 * (shifted nanoseconds), this can be killed.
1186 remainder = tk->xtime_nsec & ((1 << tk->shift) - 1);
1187 tk->xtime_nsec -= remainder;
1188 tk->xtime_nsec += 1 << tk->shift;
1189 tk->ntp_error += remainder << tk->ntp_error_shift;
1192 * Finally, make sure that after the rounding
1193 * xtime_nsec isn't larger than NSEC_PER_SEC
1195 accumulate_nsecs_to_secs(tk);
1197 timekeeping_update(tk, false);
1200 write_sequnlock_irqrestore(&tk->lock, flags);
1205 * getboottime - Return the real time of system boot.
1206 * @ts: pointer to the timespec to be set
1208 * Returns the wall-time of boot in a timespec.
1210 * This is based on the wall_to_monotonic offset and the total suspend
1211 * time. Calls to settimeofday will affect the value returned (which
1212 * basically means that however wrong your real time clock is at boot time,
1213 * you get the right time here).
1215 void getboottime(struct timespec *ts)
1217 struct timekeeper *tk = &timekeeper;
1218 struct timespec boottime = {
1219 .tv_sec = tk->wall_to_monotonic.tv_sec +
1220 tk->total_sleep_time.tv_sec,
1221 .tv_nsec = tk->wall_to_monotonic.tv_nsec +
1222 tk->total_sleep_time.tv_nsec
1225 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1227 EXPORT_SYMBOL_GPL(getboottime);
1230 * get_monotonic_boottime - Returns monotonic time since boot
1231 * @ts: pointer to the timespec to be set
1233 * Returns the monotonic time since boot in a timespec.
1235 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1236 * includes the time spent in suspend.
1238 void get_monotonic_boottime(struct timespec *ts)
1240 struct timekeeper *tk = &timekeeper;
1241 struct timespec tomono, sleep;
1244 WARN_ON(timekeeping_suspended);
1247 seq = read_seqbegin(&tk->lock);
1248 ts->tv_sec = tk->xtime_sec;
1249 ts->tv_nsec = timekeeping_get_ns(tk);
1250 tomono = tk->wall_to_monotonic;
1251 sleep = tk->total_sleep_time;
1253 } while (read_seqretry(&tk->lock, seq));
1255 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1256 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec);
1258 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1261 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1263 * Returns the monotonic time since boot in a ktime
1265 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1266 * includes the time spent in suspend.
1268 ktime_t ktime_get_boottime(void)
1272 get_monotonic_boottime(&ts);
1273 return timespec_to_ktime(ts);
1275 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1278 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1279 * @ts: pointer to the timespec to be converted
1281 void monotonic_to_bootbased(struct timespec *ts)
1283 struct timekeeper *tk = &timekeeper;
1285 *ts = timespec_add(*ts, tk->total_sleep_time);
1287 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1289 unsigned long get_seconds(void)
1291 struct timekeeper *tk = &timekeeper;
1293 return tk->xtime_sec;
1295 EXPORT_SYMBOL(get_seconds);
1297 struct timespec __current_kernel_time(void)
1299 struct timekeeper *tk = &timekeeper;
1301 return tk_xtime(tk);
1304 struct timespec current_kernel_time(void)
1306 struct timekeeper *tk = &timekeeper;
1307 struct timespec now;
1311 seq = read_seqbegin(&tk->lock);
1314 } while (read_seqretry(&tk->lock, seq));
1318 EXPORT_SYMBOL(current_kernel_time);
1320 struct timespec get_monotonic_coarse(void)
1322 struct timekeeper *tk = &timekeeper;
1323 struct timespec now, mono;
1327 seq = read_seqbegin(&tk->lock);
1330 mono = tk->wall_to_monotonic;
1331 } while (read_seqretry(&tk->lock, seq));
1333 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1334 now.tv_nsec + mono.tv_nsec);
1339 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1340 * without sampling the sequence number in xtime_lock.
1341 * jiffies is defined in the linker script...
1343 void do_timer(unsigned long ticks)
1345 jiffies_64 += ticks;
1347 calc_global_load(ticks);
1351 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1352 * and sleep offsets.
1353 * @xtim: pointer to timespec to be set with xtime
1354 * @wtom: pointer to timespec to be set with wall_to_monotonic
1355 * @sleep: pointer to timespec to be set with time in suspend
1357 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1358 struct timespec *wtom, struct timespec *sleep)
1360 struct timekeeper *tk = &timekeeper;
1364 seq = read_seqbegin(&tk->lock);
1365 *xtim = tk_xtime(tk);
1366 *wtom = tk->wall_to_monotonic;
1367 *sleep = tk->total_sleep_time;
1368 } while (read_seqretry(&tk->lock, seq));
1371 #ifdef CONFIG_HIGH_RES_TIMERS
1373 * ktime_get_update_offsets - hrtimer helper
1374 * @offs_real: pointer to storage for monotonic -> realtime offset
1375 * @offs_boot: pointer to storage for monotonic -> boottime offset
1377 * Returns current monotonic time and updates the offsets
1378 * Called from hrtimer_interupt() or retrigger_next_event()
1380 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1382 struct timekeeper *tk = &timekeeper;
1388 seq = read_seqbegin(&tk->lock);
1390 secs = tk->xtime_sec;
1391 nsecs = timekeeping_get_ns(tk);
1393 *offs_real = tk->offs_real;
1394 *offs_boot = tk->offs_boot;
1395 } while (read_seqretry(&tk->lock, seq));
1397 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1398 now = ktime_sub(now, *offs_real);
1404 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1406 ktime_t ktime_get_monotonic_offset(void)
1408 struct timekeeper *tk = &timekeeper;
1410 struct timespec wtom;
1413 seq = read_seqbegin(&tk->lock);
1414 wtom = tk->wall_to_monotonic;
1415 } while (read_seqretry(&tk->lock, seq));
1417 return timespec_to_ktime(wtom);
1419 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1422 * xtime_update() - advances the timekeeping infrastructure
1423 * @ticks: number of ticks, that have elapsed since the last call.
1425 * Must be called with interrupts disabled.
1427 void xtime_update(unsigned long ticks)
1429 write_seqlock(&xtime_lock);
1431 write_sequnlock(&xtime_lock);