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)
298 WARN_ON(timekeeping_suspended);
301 seq = read_seqbegin(&timekeeper.lock);
303 ts->tv_sec = timekeeper.xtime_sec;
304 ts->tv_nsec = timekeeping_get_ns(&timekeeper);
306 } while (read_seqretry(&timekeeper.lock, seq));
308 timespec_add_ns(ts, nsecs);
310 EXPORT_SYMBOL(getnstimeofday);
312 ktime_t ktime_get(void)
317 WARN_ON(timekeeping_suspended);
320 seq = read_seqbegin(&timekeeper.lock);
321 secs = timekeeper.xtime_sec +
322 timekeeper.wall_to_monotonic.tv_sec;
323 nsecs = timekeeping_get_ns(&timekeeper) +
324 timekeeper.wall_to_monotonic.tv_nsec;
326 } while (read_seqretry(&timekeeper.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 timespec tomono;
348 WARN_ON(timekeeping_suspended);
351 seq = read_seqbegin(&timekeeper.lock);
352 ts->tv_sec = timekeeper.xtime_sec;
353 ts->tv_nsec = timekeeping_get_ns(&timekeeper);
354 tomono = timekeeper.wall_to_monotonic;
356 } while (read_seqretry(&timekeeper.lock, seq));
358 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
359 ts->tv_nsec + tomono.tv_nsec);
361 EXPORT_SYMBOL_GPL(ktime_get_ts);
363 #ifdef CONFIG_NTP_PPS
366 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
367 * @ts_raw: pointer to the timespec to be set to raw monotonic time
368 * @ts_real: pointer to the timespec to be set to the time of day
370 * This function reads both the time of day and raw monotonic time at the
371 * same time atomically and stores the resulting timestamps in timespec
374 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
377 s64 nsecs_raw, nsecs_real;
379 WARN_ON_ONCE(timekeeping_suspended);
382 seq = read_seqbegin(&timekeeper.lock);
384 *ts_raw = timekeeper.raw_time;
385 ts_real->tv_sec = timekeeper.xtime_sec;
386 ts_real->tv_nsec = 0;
388 nsecs_raw = timekeeping_get_ns_raw(&timekeeper);
389 nsecs_real = timekeeping_get_ns(&timekeeper);
391 } while (read_seqretry(&timekeeper.lock, seq));
393 timespec_add_ns(ts_raw, nsecs_raw);
394 timespec_add_ns(ts_real, nsecs_real);
396 EXPORT_SYMBOL(getnstime_raw_and_real);
398 #endif /* CONFIG_NTP_PPS */
401 * do_gettimeofday - Returns the time of day in a timeval
402 * @tv: pointer to the timeval to be set
404 * NOTE: Users should be converted to using getnstimeofday()
406 void do_gettimeofday(struct timeval *tv)
410 getnstimeofday(&now);
411 tv->tv_sec = now.tv_sec;
412 tv->tv_usec = now.tv_nsec/1000;
414 EXPORT_SYMBOL(do_gettimeofday);
417 * do_settimeofday - Sets the time of day
418 * @tv: pointer to the timespec variable containing the new time
420 * Sets the time of day to the new time and update NTP and notify hrtimers
422 int do_settimeofday(const struct timespec *tv)
424 struct timespec ts_delta, xt;
427 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
430 write_seqlock_irqsave(&timekeeper.lock, flags);
432 timekeeping_forward_now(&timekeeper);
434 xt = tk_xtime(&timekeeper);
435 ts_delta.tv_sec = tv->tv_sec - xt.tv_sec;
436 ts_delta.tv_nsec = tv->tv_nsec - xt.tv_nsec;
438 tk_set_wall_to_mono(&timekeeper,
439 timespec_sub(timekeeper.wall_to_monotonic, ts_delta));
441 tk_set_xtime(&timekeeper, tv);
443 timekeeping_update(&timekeeper, true);
445 write_sequnlock_irqrestore(&timekeeper.lock, flags);
447 /* signal hrtimers about time change */
452 EXPORT_SYMBOL(do_settimeofday);
455 * timekeeping_inject_offset - Adds or subtracts from the current time.
456 * @tv: pointer to the timespec variable containing the offset
458 * Adds or subtracts an offset value from the current time.
460 int timekeeping_inject_offset(struct timespec *ts)
464 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
467 write_seqlock_irqsave(&timekeeper.lock, flags);
469 timekeeping_forward_now(&timekeeper);
472 tk_xtime_add(&timekeeper, ts);
473 tk_set_wall_to_mono(&timekeeper,
474 timespec_sub(timekeeper.wall_to_monotonic, *ts));
476 timekeeping_update(&timekeeper, true);
478 write_sequnlock_irqrestore(&timekeeper.lock, flags);
480 /* signal hrtimers about time change */
485 EXPORT_SYMBOL(timekeeping_inject_offset);
488 * change_clocksource - Swaps clocksources if a new one is available
490 * Accumulates current time interval and initializes new clocksource
492 static int change_clocksource(void *data)
494 struct clocksource *new, *old;
497 new = (struct clocksource *) data;
499 write_seqlock_irqsave(&timekeeper.lock, flags);
501 timekeeping_forward_now(&timekeeper);
502 if (!new->enable || new->enable(new) == 0) {
503 old = timekeeper.clock;
504 tk_setup_internals(&timekeeper, new);
508 timekeeping_update(&timekeeper, true);
510 write_sequnlock_irqrestore(&timekeeper.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 if (timekeeper.clock == clock)
526 stop_machine(change_clocksource, clock, NULL);
531 * ktime_get_real - get the real (wall-) time in ktime_t format
533 * returns the time in ktime_t format
535 ktime_t ktime_get_real(void)
539 getnstimeofday(&now);
541 return timespec_to_ktime(now);
543 EXPORT_SYMBOL_GPL(ktime_get_real);
546 * getrawmonotonic - Returns the raw monotonic time in a timespec
547 * @ts: pointer to the timespec to be set
549 * Returns the raw monotonic time (completely un-modified by ntp)
551 void getrawmonotonic(struct timespec *ts)
557 seq = read_seqbegin(&timekeeper.lock);
558 nsecs = timekeeping_get_ns_raw(&timekeeper);
559 *ts = timekeeper.raw_time;
561 } while (read_seqretry(&timekeeper.lock, seq));
563 timespec_add_ns(ts, nsecs);
565 EXPORT_SYMBOL(getrawmonotonic);
568 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
570 int timekeeping_valid_for_hres(void)
576 seq = read_seqbegin(&timekeeper.lock);
578 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
580 } while (read_seqretry(&timekeeper.lock, seq));
586 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
588 u64 timekeeping_max_deferment(void)
594 seq = read_seqbegin(&timekeeper.lock);
596 ret = timekeeper.clock->max_idle_ns;
598 } while (read_seqretry(&timekeeper.lock, seq));
604 * read_persistent_clock - Return time from the persistent clock.
606 * Weak dummy function for arches that do not yet support it.
607 * Reads the time from the battery backed persistent clock.
608 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
610 * XXX - Do be sure to remove it once all arches implement it.
612 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
619 * read_boot_clock - Return time of the system start.
621 * Weak dummy function for arches that do not yet support it.
622 * Function to read the exact time the system has been started.
623 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
625 * XXX - Do be sure to remove it once all arches implement it.
627 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
634 * timekeeping_init - Initializes the clocksource and common timekeeping values
636 void __init timekeeping_init(void)
638 struct clocksource *clock;
640 struct timespec now, boot, tmp;
642 read_persistent_clock(&now);
643 read_boot_clock(&boot);
645 seqlock_init(&timekeeper.lock);
649 write_seqlock_irqsave(&timekeeper.lock, flags);
650 clock = clocksource_default_clock();
652 clock->enable(clock);
653 tk_setup_internals(&timekeeper, clock);
655 tk_set_xtime(&timekeeper, &now);
656 timekeeper.raw_time.tv_sec = 0;
657 timekeeper.raw_time.tv_nsec = 0;
658 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
659 boot = tk_xtime(&timekeeper);
661 set_normalized_timespec(&tmp, -boot.tv_sec, -boot.tv_nsec);
662 tk_set_wall_to_mono(&timekeeper, tmp);
666 tk_set_sleep_time(&timekeeper, tmp);
668 write_sequnlock_irqrestore(&timekeeper.lock, flags);
671 /* time in seconds when suspend began */
672 static struct timespec timekeeping_suspend_time;
675 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
676 * @delta: pointer to a timespec delta value
678 * Takes a timespec offset measuring a suspend interval and properly
679 * adds the sleep offset to the timekeeping variables.
681 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
682 struct timespec *delta)
684 if (!timespec_valid(delta)) {
685 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
686 "sleep delta value!\n");
689 tk_xtime_add(tk, delta);
690 tk_set_wall_to_mono(tk, timespec_sub(tk->wall_to_monotonic, *delta));
691 tk_set_sleep_time(tk, timespec_add(tk->total_sleep_time, *delta));
695 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
696 * @delta: pointer to a timespec delta value
698 * This hook is for architectures that cannot support read_persistent_clock
699 * because their RTC/persistent clock is only accessible when irqs are enabled.
701 * This function should only be called by rtc_resume(), and allows
702 * a suspend offset to be injected into the timekeeping values.
704 void timekeeping_inject_sleeptime(struct timespec *delta)
709 /* Make sure we don't set the clock twice */
710 read_persistent_clock(&ts);
711 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
714 write_seqlock_irqsave(&timekeeper.lock, flags);
716 timekeeping_forward_now(&timekeeper);
718 __timekeeping_inject_sleeptime(&timekeeper, delta);
720 timekeeping_update(&timekeeper, true);
722 write_sequnlock_irqrestore(&timekeeper.lock, flags);
724 /* signal hrtimers about time change */
729 * timekeeping_resume - Resumes the generic timekeeping subsystem.
731 * This is for the generic clocksource timekeeping.
732 * xtime/wall_to_monotonic/jiffies/etc are
733 * still managed by arch specific suspend/resume code.
735 static void timekeeping_resume(void)
740 read_persistent_clock(&ts);
742 clocksource_resume();
744 write_seqlock_irqsave(&timekeeper.lock, flags);
746 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
747 ts = timespec_sub(ts, timekeeping_suspend_time);
748 __timekeeping_inject_sleeptime(&timekeeper, &ts);
750 /* re-base the last cycle value */
751 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
752 timekeeper.ntp_error = 0;
753 timekeeping_suspended = 0;
754 timekeeping_update(&timekeeper, false);
755 write_sequnlock_irqrestore(&timekeeper.lock, flags);
757 touch_softlockup_watchdog();
759 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
761 /* Resume hrtimers */
765 static int timekeeping_suspend(void)
768 struct timespec delta, delta_delta;
769 static struct timespec old_delta;
771 read_persistent_clock(&timekeeping_suspend_time);
773 write_seqlock_irqsave(&timekeeper.lock, flags);
774 timekeeping_forward_now(&timekeeper);
775 timekeeping_suspended = 1;
778 * To avoid drift caused by repeated suspend/resumes,
779 * which each can add ~1 second drift error,
780 * try to compensate so the difference in system time
781 * and persistent_clock time stays close to constant.
783 delta = timespec_sub(tk_xtime(&timekeeper), timekeeping_suspend_time);
784 delta_delta = timespec_sub(delta, old_delta);
785 if (abs(delta_delta.tv_sec) >= 2) {
787 * if delta_delta is too large, assume time correction
788 * has occured and set old_delta to the current delta.
792 /* Otherwise try to adjust old_system to compensate */
793 timekeeping_suspend_time =
794 timespec_add(timekeeping_suspend_time, delta_delta);
796 write_sequnlock_irqrestore(&timekeeper.lock, flags);
798 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
799 clocksource_suspend();
804 /* sysfs resume/suspend bits for timekeeping */
805 static struct syscore_ops timekeeping_syscore_ops = {
806 .resume = timekeeping_resume,
807 .suspend = timekeeping_suspend,
810 static int __init timekeeping_init_ops(void)
812 register_syscore_ops(&timekeeping_syscore_ops);
816 device_initcall(timekeeping_init_ops);
819 * If the error is already larger, we look ahead even further
820 * to compensate for late or lost adjustments.
822 static __always_inline int timekeeping_bigadjust(struct timekeeper *tk,
823 s64 error, s64 *interval,
831 * Use the current error value to determine how much to look ahead.
832 * The larger the error the slower we adjust for it to avoid problems
833 * with losing too many ticks, otherwise we would overadjust and
834 * produce an even larger error. The smaller the adjustment the
835 * faster we try to adjust for it, as lost ticks can do less harm
836 * here. This is tuned so that an error of about 1 msec is adjusted
837 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
839 error2 = tk->ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
840 error2 = abs(error2);
841 for (look_ahead = 0; error2 > 0; look_ahead++)
845 * Now calculate the error in (1 << look_ahead) ticks, but first
846 * remove the single look ahead already included in the error.
848 tick_error = ntp_tick_length() >> (tk->ntp_error_shift + 1);
849 tick_error -= tk->xtime_interval >> 1;
850 error = ((error - tick_error) >> look_ahead) + tick_error;
852 /* Finally calculate the adjustment shift value. */
857 *interval = -*interval;
861 for (adj = 0; error > i; adj++)
870 * Adjust the multiplier to reduce the error value,
871 * this is optimized for the most common adjustments of -1,0,1,
872 * for other values we can do a bit more work.
874 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
876 s64 error, interval = tk->cycle_interval;
880 * The point of this is to check if the error is greater than half
883 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
885 * Note we subtract one in the shift, so that error is really error*2.
886 * This "saves" dividing(shifting) interval twice, but keeps the
887 * (error > interval) comparison as still measuring if error is
888 * larger than half an interval.
890 * Note: It does not "save" on aggravation when reading the code.
892 error = tk->ntp_error >> (tk->ntp_error_shift - 1);
893 if (error > interval) {
895 * We now divide error by 4(via shift), which checks if
896 * the error is greater than twice the interval.
897 * If it is greater, we need a bigadjust, if its smaller,
898 * we can adjust by 1.
902 * XXX - In update_wall_time, we round up to the next
903 * nanosecond, and store the amount rounded up into
904 * the error. This causes the likely below to be unlikely.
906 * The proper fix is to avoid rounding up by using
907 * the high precision timekeeper.xtime_nsec instead of
908 * xtime.tv_nsec everywhere. Fixing this will take some
911 if (likely(error <= interval))
914 adj = timekeeping_bigadjust(tk, error, &interval,
916 } else if (error < -interval) {
917 /* See comment above, this is just switched for the negative */
919 if (likely(error >= -interval)) {
921 interval = -interval;
924 adj = timekeeping_bigadjust(tk, error, &interval,
929 if (unlikely(tk->clock->maxadj &&
930 (tk->mult + adj > tk->clock->mult + tk->clock->maxadj))) {
931 printk_once(KERN_WARNING
932 "Adjusting %s more than 11%% (%ld vs %ld)\n",
933 tk->clock->name, (long)tk->mult + adj,
934 (long)tk->clock->mult + tk->clock->maxadj);
937 * So the following can be confusing.
939 * To keep things simple, lets assume adj == 1 for now.
941 * When adj != 1, remember that the interval and offset values
942 * have been appropriately scaled so the math is the same.
944 * The basic idea here is that we're increasing the multiplier
945 * by one, this causes the xtime_interval to be incremented by
946 * one cycle_interval. This is because:
947 * xtime_interval = cycle_interval * mult
948 * So if mult is being incremented by one:
949 * xtime_interval = cycle_interval * (mult + 1)
951 * xtime_interval = (cycle_interval * mult) + cycle_interval
952 * Which can be shortened to:
953 * xtime_interval += cycle_interval
955 * So offset stores the non-accumulated cycles. Thus the current
956 * time (in shifted nanoseconds) is:
957 * now = (offset * adj) + xtime_nsec
958 * Now, even though we're adjusting the clock frequency, we have
959 * to keep time consistent. In other words, we can't jump back
960 * in time, and we also want to avoid jumping forward in time.
962 * So given the same offset value, we need the time to be the same
963 * both before and after the freq adjustment.
964 * now = (offset * adj_1) + xtime_nsec_1
965 * now = (offset * adj_2) + xtime_nsec_2
967 * (offset * adj_1) + xtime_nsec_1 =
968 * (offset * adj_2) + xtime_nsec_2
972 * (offset * adj_1) + xtime_nsec_1 =
973 * (offset * (adj_1+1)) + xtime_nsec_2
974 * (offset * adj_1) + xtime_nsec_1 =
975 * (offset * adj_1) + offset + xtime_nsec_2
976 * Canceling the sides:
977 * xtime_nsec_1 = offset + xtime_nsec_2
979 * xtime_nsec_2 = xtime_nsec_1 - offset
980 * Which simplfies to:
981 * xtime_nsec -= offset
983 * XXX - TODO: Doc ntp_error calculation.
986 tk->xtime_interval += interval;
987 tk->xtime_nsec -= offset;
988 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
991 * It may be possible that when we entered this function, xtime_nsec
992 * was very small. Further, if we're slightly speeding the clocksource
993 * in the code above, its possible the required corrective factor to
994 * xtime_nsec could cause it to underflow.
996 * Now, since we already accumulated the second, cannot simply roll
997 * the accumulated second back, since the NTP subsystem has been
998 * notified via second_overflow. So instead we push xtime_nsec forward
999 * by the amount we underflowed, and add that amount into the error.
1001 * We'll correct this error next time through this function, when
1002 * xtime_nsec is not as small.
1004 if (unlikely((s64)tk->xtime_nsec < 0)) {
1005 s64 neg = -(s64)tk->xtime_nsec;
1007 tk->ntp_error += neg << tk->ntp_error_shift;
1013 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1015 * Helper function that accumulates a the nsecs greater then a second
1016 * from the xtime_nsec field to the xtime_secs field.
1017 * It also calls into the NTP code to handle leapsecond processing.
1020 static inline void accumulate_nsecs_to_secs(struct timekeeper *tk)
1022 u64 nsecps = (u64)NSEC_PER_SEC << tk->shift;
1024 while (tk->xtime_nsec >= nsecps) {
1027 tk->xtime_nsec -= nsecps;
1030 /* Figure out if its a leap sec and apply if needed */
1031 leap = second_overflow(tk->xtime_sec);
1032 if (unlikely(leap)) {
1035 tk->xtime_sec += leap;
1039 tk_set_wall_to_mono(tk,
1040 timespec_sub(tk->wall_to_monotonic, ts));
1042 clock_was_set_delayed();
1048 * logarithmic_accumulation - shifted accumulation of cycles
1050 * This functions accumulates a shifted interval of cycles into
1051 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1054 * Returns the unconsumed cycles.
1056 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1061 /* If the offset is smaller then a shifted interval, do nothing */
1062 if (offset < tk->cycle_interval<<shift)
1065 /* Accumulate one shifted interval */
1066 offset -= tk->cycle_interval << shift;
1067 tk->clock->cycle_last += tk->cycle_interval << shift;
1069 tk->xtime_nsec += tk->xtime_interval << shift;
1070 accumulate_nsecs_to_secs(tk);
1072 /* Accumulate raw time */
1073 raw_nsecs = tk->raw_interval << shift;
1074 raw_nsecs += tk->raw_time.tv_nsec;
1075 if (raw_nsecs >= NSEC_PER_SEC) {
1076 u64 raw_secs = raw_nsecs;
1077 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1078 tk->raw_time.tv_sec += raw_secs;
1080 tk->raw_time.tv_nsec = raw_nsecs;
1082 /* Accumulate error between NTP and clock interval */
1083 tk->ntp_error += ntp_tick_length() << shift;
1084 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1085 (tk->ntp_error_shift + shift);
1091 * update_wall_time - Uses the current clocksource to increment the wall time
1094 static void update_wall_time(void)
1096 struct clocksource *clock;
1098 int shift = 0, maxshift;
1099 unsigned long flags;
1102 write_seqlock_irqsave(&timekeeper.lock, flags);
1104 /* Make sure we're fully resumed: */
1105 if (unlikely(timekeeping_suspended))
1108 clock = timekeeper.clock;
1110 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1111 offset = timekeeper.cycle_interval;
1113 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1117 * With NO_HZ we may have to accumulate many cycle_intervals
1118 * (think "ticks") worth of time at once. To do this efficiently,
1119 * we calculate the largest doubling multiple of cycle_intervals
1120 * that is smaller than the offset. We then accumulate that
1121 * chunk in one go, and then try to consume the next smaller
1124 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1125 shift = max(0, shift);
1126 /* Bound shift to one less than what overflows tick_length */
1127 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1128 shift = min(shift, maxshift);
1129 while (offset >= timekeeper.cycle_interval) {
1130 offset = logarithmic_accumulation(&timekeeper, offset, shift);
1131 if(offset < timekeeper.cycle_interval<<shift)
1135 /* correct the clock when NTP error is too big */
1136 timekeeping_adjust(&timekeeper, offset);
1140 * Store only full nanoseconds into xtime_nsec after rounding
1141 * it up and add the remainder to the error difference.
1142 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
1143 * by truncating the remainder in vsyscalls. However, it causes
1144 * additional work to be done in timekeeping_adjust(). Once
1145 * the vsyscall implementations are converted to use xtime_nsec
1146 * (shifted nanoseconds), this can be killed.
1148 remainder = timekeeper.xtime_nsec & ((1 << timekeeper.shift) - 1);
1149 timekeeper.xtime_nsec -= remainder;
1150 timekeeper.xtime_nsec += 1 << timekeeper.shift;
1151 timekeeper.ntp_error += remainder << timekeeper.ntp_error_shift;
1154 * Finally, make sure that after the rounding
1155 * xtime_nsec isn't larger than NSEC_PER_SEC
1157 accumulate_nsecs_to_secs(&timekeeper);
1159 timekeeping_update(&timekeeper, false);
1162 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1167 * getboottime - Return the real time of system boot.
1168 * @ts: pointer to the timespec to be set
1170 * Returns the wall-time of boot in a timespec.
1172 * This is based on the wall_to_monotonic offset and the total suspend
1173 * time. Calls to settimeofday will affect the value returned (which
1174 * basically means that however wrong your real time clock is at boot time,
1175 * you get the right time here).
1177 void getboottime(struct timespec *ts)
1179 struct timespec boottime = {
1180 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1181 timekeeper.total_sleep_time.tv_sec,
1182 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1183 timekeeper.total_sleep_time.tv_nsec
1186 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1188 EXPORT_SYMBOL_GPL(getboottime);
1191 * get_monotonic_boottime - Returns monotonic time since boot
1192 * @ts: pointer to the timespec to be set
1194 * Returns the monotonic time since boot in a timespec.
1196 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1197 * includes the time spent in suspend.
1199 void get_monotonic_boottime(struct timespec *ts)
1201 struct timespec tomono, sleep;
1204 WARN_ON(timekeeping_suspended);
1207 seq = read_seqbegin(&timekeeper.lock);
1208 ts->tv_sec = timekeeper.xtime_sec;
1209 ts->tv_nsec = timekeeping_get_ns(&timekeeper);
1210 tomono = timekeeper.wall_to_monotonic;
1211 sleep = timekeeper.total_sleep_time;
1213 } while (read_seqretry(&timekeeper.lock, seq));
1215 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1216 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec);
1218 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1221 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1223 * Returns the monotonic time since boot in a ktime
1225 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1226 * includes the time spent in suspend.
1228 ktime_t ktime_get_boottime(void)
1232 get_monotonic_boottime(&ts);
1233 return timespec_to_ktime(ts);
1235 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1238 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1239 * @ts: pointer to the timespec to be converted
1241 void monotonic_to_bootbased(struct timespec *ts)
1243 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1245 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1247 unsigned long get_seconds(void)
1249 return timekeeper.xtime_sec;
1251 EXPORT_SYMBOL(get_seconds);
1253 struct timespec __current_kernel_time(void)
1255 return tk_xtime(&timekeeper);
1258 struct timespec current_kernel_time(void)
1260 struct timespec now;
1264 seq = read_seqbegin(&timekeeper.lock);
1266 now = tk_xtime(&timekeeper);
1267 } while (read_seqretry(&timekeeper.lock, seq));
1271 EXPORT_SYMBOL(current_kernel_time);
1273 struct timespec get_monotonic_coarse(void)
1275 struct timespec now, mono;
1279 seq = read_seqbegin(&timekeeper.lock);
1281 now = tk_xtime(&timekeeper);
1282 mono = timekeeper.wall_to_monotonic;
1283 } while (read_seqretry(&timekeeper.lock, seq));
1285 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1286 now.tv_nsec + mono.tv_nsec);
1291 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1292 * without sampling the sequence number in xtime_lock.
1293 * jiffies is defined in the linker script...
1295 void do_timer(unsigned long ticks)
1297 jiffies_64 += ticks;
1299 calc_global_load(ticks);
1303 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1304 * and sleep offsets.
1305 * @xtim: pointer to timespec to be set with xtime
1306 * @wtom: pointer to timespec to be set with wall_to_monotonic
1307 * @sleep: pointer to timespec to be set with time in suspend
1309 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1310 struct timespec *wtom, struct timespec *sleep)
1315 seq = read_seqbegin(&timekeeper.lock);
1316 *xtim = tk_xtime(&timekeeper);
1317 *wtom = timekeeper.wall_to_monotonic;
1318 *sleep = timekeeper.total_sleep_time;
1319 } while (read_seqretry(&timekeeper.lock, seq));
1322 #ifdef CONFIG_HIGH_RES_TIMERS
1324 * ktime_get_update_offsets - hrtimer helper
1325 * @offs_real: pointer to storage for monotonic -> realtime offset
1326 * @offs_boot: pointer to storage for monotonic -> boottime offset
1328 * Returns current monotonic time and updates the offsets
1329 * Called from hrtimer_interupt() or retrigger_next_event()
1331 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1338 seq = read_seqbegin(&timekeeper.lock);
1340 secs = timekeeper.xtime_sec;
1341 nsecs = timekeeping_get_ns(&timekeeper);
1343 *offs_real = timekeeper.offs_real;
1344 *offs_boot = timekeeper.offs_boot;
1345 } while (read_seqretry(&timekeeper.lock, seq));
1347 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1348 now = ktime_sub(now, *offs_real);
1354 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1356 ktime_t ktime_get_monotonic_offset(void)
1359 struct timespec wtom;
1362 seq = read_seqbegin(&timekeeper.lock);
1363 wtom = timekeeper.wall_to_monotonic;
1364 } while (read_seqretry(&timekeeper.lock, seq));
1366 return timespec_to_ktime(wtom);
1368 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1371 * xtime_update() - advances the timekeeping infrastructure
1372 * @ticks: number of ticks, that have elapsed since the last call.
1374 * Must be called with interrupts disabled.
1376 void xtime_update(unsigned long ticks)
1378 write_seqlock(&xtime_lock);
1380 write_sequnlock(&xtime_lock);