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 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
43 /* Difference between accumulated time and NTP time in ntp
44 * shifted nano seconds. */
46 /* Shift conversion between clock shifted nano seconds and
47 * ntp shifted nano seconds. */
50 /* The current time */
51 struct timespec xtime;
53 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
54 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
55 * at zero at system boot time, so wall_to_monotonic will be negative,
56 * however, we will ALWAYS keep the tv_nsec part positive so we can use
57 * the usual normalization.
59 * wall_to_monotonic is moved after resume from suspend for the
60 * monotonic time not to jump. We need to add total_sleep_time to
61 * wall_to_monotonic to get the real boot based time offset.
63 * - wall_to_monotonic is no longer the boot time, getboottime must be
66 struct timespec wall_to_monotonic;
67 /* time spent in suspend */
68 struct timespec total_sleep_time;
69 /* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. */
70 struct timespec raw_time;
71 /* Offset clock monotonic -> clock realtime */
73 /* Offset clock monotonic -> clock boottime */
75 /* Seqlock for all timekeeper values */
79 static struct timekeeper timekeeper;
82 * This read-write spinlock protects us from races in SMP while
85 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
88 /* flag for if timekeeping is suspended */
89 int __read_mostly timekeeping_suspended;
94 * timekeeper_setup_internals - Set up internals to use clocksource clock.
96 * @clock: Pointer to clocksource.
98 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
99 * pair and interval request.
101 * Unless you're the timekeeping code, you should not be using this!
103 static void timekeeper_setup_internals(struct clocksource *clock)
106 u64 tmp, ntpinterval;
108 timekeeper.clock = clock;
109 clock->cycle_last = clock->read(clock);
111 /* Do the ns -> cycle conversion first, using original mult */
112 tmp = NTP_INTERVAL_LENGTH;
113 tmp <<= clock->shift;
115 tmp += clock->mult/2;
116 do_div(tmp, clock->mult);
120 interval = (cycle_t) tmp;
121 timekeeper.cycle_interval = interval;
123 /* Go back from cycles -> shifted ns */
124 timekeeper.xtime_interval = (u64) interval * clock->mult;
125 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
126 timekeeper.raw_interval =
127 ((u64) interval * clock->mult) >> clock->shift;
129 timekeeper.xtime_nsec = 0;
130 timekeeper.shift = clock->shift;
132 timekeeper.ntp_error = 0;
133 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
136 * The timekeeper keeps its own mult values for the currently
137 * active clocksource. These value will be adjusted via NTP
138 * to counteract clock drifting.
140 timekeeper.mult = clock->mult;
143 /* Timekeeper helper functions. */
144 static inline s64 timekeeping_get_ns(void)
146 cycle_t cycle_now, cycle_delta;
147 struct clocksource *clock;
149 /* read clocksource: */
150 clock = timekeeper.clock;
151 cycle_now = clock->read(clock);
153 /* calculate the delta since the last update_wall_time: */
154 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
156 /* return delta convert to nanoseconds using ntp adjusted mult. */
157 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
161 static inline s64 timekeeping_get_ns_raw(void)
163 cycle_t cycle_now, cycle_delta;
164 struct clocksource *clock;
166 /* read clocksource: */
167 clock = timekeeper.clock;
168 cycle_now = clock->read(clock);
170 /* calculate the delta since the last update_wall_time: */
171 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
173 /* return delta convert to nanoseconds. */
174 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
177 static void update_rt_offset(void)
179 struct timespec tmp, *wtm = &timekeeper.wall_to_monotonic;
181 set_normalized_timespec(&tmp, -wtm->tv_sec, -wtm->tv_nsec);
182 timekeeper.offs_real = timespec_to_ktime(tmp);
185 /* must hold write on timekeeper.lock */
186 static void timekeeping_update(bool clearntp)
189 timekeeper.ntp_error = 0;
193 update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
194 timekeeper.clock, timekeeper.mult);
199 * timekeeping_forward_now - update clock to the current time
201 * Forward the current clock to update its state since the last call to
202 * update_wall_time(). This is useful before significant clock changes,
203 * as it avoids having to deal with this time offset explicitly.
205 static void timekeeping_forward_now(void)
207 cycle_t cycle_now, cycle_delta;
208 struct clocksource *clock;
211 clock = timekeeper.clock;
212 cycle_now = clock->read(clock);
213 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
214 clock->cycle_last = cycle_now;
216 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
219 /* If arch requires, add in gettimeoffset() */
220 nsec += arch_gettimeoffset();
222 timespec_add_ns(&timekeeper.xtime, nsec);
224 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
225 timespec_add_ns(&timekeeper.raw_time, nsec);
229 * getnstimeofday - Returns the time of day in a timespec
230 * @ts: pointer to the timespec to be set
232 * Returns the time of day in a timespec.
234 void getnstimeofday(struct timespec *ts)
239 WARN_ON(timekeeping_suspended);
242 seq = read_seqbegin(&timekeeper.lock);
244 *ts = timekeeper.xtime;
245 nsecs = timekeeping_get_ns();
247 /* If arch requires, add in gettimeoffset() */
248 nsecs += arch_gettimeoffset();
250 } while (read_seqretry(&timekeeper.lock, seq));
252 timespec_add_ns(ts, nsecs);
254 EXPORT_SYMBOL(getnstimeofday);
256 ktime_t ktime_get(void)
261 WARN_ON(timekeeping_suspended);
264 seq = read_seqbegin(&timekeeper.lock);
265 secs = timekeeper.xtime.tv_sec +
266 timekeeper.wall_to_monotonic.tv_sec;
267 nsecs = timekeeper.xtime.tv_nsec +
268 timekeeper.wall_to_monotonic.tv_nsec;
269 nsecs += timekeeping_get_ns();
270 /* If arch requires, add in gettimeoffset() */
271 nsecs += arch_gettimeoffset();
273 } while (read_seqretry(&timekeeper.lock, seq));
275 * Use ktime_set/ktime_add_ns to create a proper ktime on
276 * 32-bit architectures without CONFIG_KTIME_SCALAR.
278 return ktime_add_ns(ktime_set(secs, 0), nsecs);
280 EXPORT_SYMBOL_GPL(ktime_get);
283 * ktime_get_ts - get the monotonic clock in timespec format
284 * @ts: pointer to timespec variable
286 * The function calculates the monotonic clock from the realtime
287 * clock and the wall_to_monotonic offset and stores the result
288 * in normalized timespec format in the variable pointed to by @ts.
290 void ktime_get_ts(struct timespec *ts)
292 struct timespec tomono;
296 WARN_ON(timekeeping_suspended);
299 seq = read_seqbegin(&timekeeper.lock);
300 *ts = timekeeper.xtime;
301 tomono = timekeeper.wall_to_monotonic;
302 nsecs = timekeeping_get_ns();
303 /* If arch requires, add in gettimeoffset() */
304 nsecs += arch_gettimeoffset();
306 } while (read_seqretry(&timekeeper.lock, seq));
308 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
309 ts->tv_nsec + tomono.tv_nsec + nsecs);
311 EXPORT_SYMBOL_GPL(ktime_get_ts);
313 #ifdef CONFIG_NTP_PPS
316 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
317 * @ts_raw: pointer to the timespec to be set to raw monotonic time
318 * @ts_real: pointer to the timespec to be set to the time of day
320 * This function reads both the time of day and raw monotonic time at the
321 * same time atomically and stores the resulting timestamps in timespec
324 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
327 s64 nsecs_raw, nsecs_real;
329 WARN_ON_ONCE(timekeeping_suspended);
334 seq = read_seqbegin(&timekeeper.lock);
336 *ts_raw = timekeeper.raw_time;
337 *ts_real = timekeeper.xtime;
339 nsecs_raw = timekeeping_get_ns_raw();
340 nsecs_real = timekeeping_get_ns();
342 /* If arch requires, add in gettimeoffset() */
343 arch_offset = arch_gettimeoffset();
344 nsecs_raw += arch_offset;
345 nsecs_real += arch_offset;
347 } while (read_seqretry(&timekeeper.lock, seq));
349 timespec_add_ns(ts_raw, nsecs_raw);
350 timespec_add_ns(ts_real, nsecs_real);
352 EXPORT_SYMBOL(getnstime_raw_and_real);
354 #endif /* CONFIG_NTP_PPS */
357 * do_gettimeofday - Returns the time of day in a timeval
358 * @tv: pointer to the timeval to be set
360 * NOTE: Users should be converted to using getnstimeofday()
362 void do_gettimeofday(struct timeval *tv)
366 getnstimeofday(&now);
367 tv->tv_sec = now.tv_sec;
368 tv->tv_usec = now.tv_nsec/1000;
370 EXPORT_SYMBOL(do_gettimeofday);
373 * do_settimeofday - Sets the time of day
374 * @tv: pointer to the timespec variable containing the new time
376 * Sets the time of day to the new time and update NTP and notify hrtimers
378 int do_settimeofday(const struct timespec *tv)
380 struct timespec ts_delta;
383 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
386 write_seqlock_irqsave(&timekeeper.lock, flags);
388 timekeeping_forward_now();
390 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
391 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
392 timekeeper.wall_to_monotonic =
393 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
395 timekeeper.xtime = *tv;
396 timekeeping_update(true);
398 write_sequnlock_irqrestore(&timekeeper.lock, flags);
400 /* signal hrtimers about time change */
405 EXPORT_SYMBOL(do_settimeofday);
409 * timekeeping_inject_offset - Adds or subtracts from the current time.
410 * @tv: pointer to the timespec variable containing the offset
412 * Adds or subtracts an offset value from the current time.
414 int timekeeping_inject_offset(struct timespec *ts)
418 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
421 write_seqlock_irqsave(&timekeeper.lock, flags);
423 timekeeping_forward_now();
425 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
426 timekeeper.wall_to_monotonic =
427 timespec_sub(timekeeper.wall_to_monotonic, *ts);
429 timekeeping_update(true);
431 write_sequnlock_irqrestore(&timekeeper.lock, flags);
433 /* signal hrtimers about time change */
438 EXPORT_SYMBOL(timekeeping_inject_offset);
441 * change_clocksource - Swaps clocksources if a new one is available
443 * Accumulates current time interval and initializes new clocksource
445 static int change_clocksource(void *data)
447 struct clocksource *new, *old;
450 new = (struct clocksource *) data;
452 write_seqlock_irqsave(&timekeeper.lock, flags);
454 timekeeping_forward_now();
455 if (!new->enable || new->enable(new) == 0) {
456 old = timekeeper.clock;
457 timekeeper_setup_internals(new);
461 timekeeping_update(true);
463 write_sequnlock_irqrestore(&timekeeper.lock, flags);
469 * timekeeping_notify - Install a new clock source
470 * @clock: pointer to the clock source
472 * This function is called from clocksource.c after a new, better clock
473 * source has been registered. The caller holds the clocksource_mutex.
475 void timekeeping_notify(struct clocksource *clock)
477 if (timekeeper.clock == clock)
479 stop_machine(change_clocksource, clock, NULL);
484 * ktime_get_real - get the real (wall-) time in ktime_t format
486 * returns the time in ktime_t format
488 ktime_t ktime_get_real(void)
492 getnstimeofday(&now);
494 return timespec_to_ktime(now);
496 EXPORT_SYMBOL_GPL(ktime_get_real);
499 * getrawmonotonic - Returns the raw monotonic time in a timespec
500 * @ts: pointer to the timespec to be set
502 * Returns the raw monotonic time (completely un-modified by ntp)
504 void getrawmonotonic(struct timespec *ts)
510 seq = read_seqbegin(&timekeeper.lock);
511 nsecs = timekeeping_get_ns_raw();
512 *ts = timekeeper.raw_time;
514 } while (read_seqretry(&timekeeper.lock, seq));
516 timespec_add_ns(ts, nsecs);
518 EXPORT_SYMBOL(getrawmonotonic);
522 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
524 int timekeeping_valid_for_hres(void)
530 seq = read_seqbegin(&timekeeper.lock);
532 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
534 } while (read_seqretry(&timekeeper.lock, seq));
540 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
542 u64 timekeeping_max_deferment(void)
548 seq = read_seqbegin(&timekeeper.lock);
550 ret = timekeeper.clock->max_idle_ns;
552 } while (read_seqretry(&timekeeper.lock, seq));
558 * read_persistent_clock - Return time from the persistent clock.
560 * Weak dummy function for arches that do not yet support it.
561 * Reads the time from the battery backed persistent clock.
562 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
564 * XXX - Do be sure to remove it once all arches implement it.
566 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
573 * read_boot_clock - Return time of the system start.
575 * Weak dummy function for arches that do not yet support it.
576 * Function to read the exact time the system has been started.
577 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
579 * XXX - Do be sure to remove it once all arches implement it.
581 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
588 * timekeeping_init - Initializes the clocksource and common timekeeping values
590 void __init timekeeping_init(void)
592 struct clocksource *clock;
594 struct timespec now, boot;
596 read_persistent_clock(&now);
597 read_boot_clock(&boot);
599 seqlock_init(&timekeeper.lock);
603 write_seqlock_irqsave(&timekeeper.lock, flags);
604 clock = clocksource_default_clock();
606 clock->enable(clock);
607 timekeeper_setup_internals(clock);
609 timekeeper.xtime.tv_sec = now.tv_sec;
610 timekeeper.xtime.tv_nsec = now.tv_nsec;
611 timekeeper.raw_time.tv_sec = 0;
612 timekeeper.raw_time.tv_nsec = 0;
613 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
614 boot.tv_sec = timekeeper.xtime.tv_sec;
615 boot.tv_nsec = timekeeper.xtime.tv_nsec;
617 set_normalized_timespec(&timekeeper.wall_to_monotonic,
618 -boot.tv_sec, -boot.tv_nsec);
620 timekeeper.total_sleep_time.tv_sec = 0;
621 timekeeper.total_sleep_time.tv_nsec = 0;
622 write_sequnlock_irqrestore(&timekeeper.lock, flags);
625 /* time in seconds when suspend began */
626 static struct timespec timekeeping_suspend_time;
628 static void update_sleep_time(struct timespec t)
630 timekeeper.total_sleep_time = t;
631 timekeeper.offs_boot = timespec_to_ktime(t);
635 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
636 * @delta: pointer to a timespec delta value
638 * Takes a timespec offset measuring a suspend interval and properly
639 * adds the sleep offset to the timekeeping variables.
641 static void __timekeeping_inject_sleeptime(struct timespec *delta)
643 if (!timespec_valid(delta)) {
644 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
645 "sleep delta value!\n");
649 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
650 timekeeper.wall_to_monotonic =
651 timespec_sub(timekeeper.wall_to_monotonic, *delta);
652 update_sleep_time(timespec_add(timekeeper.total_sleep_time, *delta));
657 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
658 * @delta: pointer to a timespec delta value
660 * This hook is for architectures that cannot support read_persistent_clock
661 * because their RTC/persistent clock is only accessible when irqs are enabled.
663 * This function should only be called by rtc_resume(), and allows
664 * a suspend offset to be injected into the timekeeping values.
666 void timekeeping_inject_sleeptime(struct timespec *delta)
671 /* Make sure we don't set the clock twice */
672 read_persistent_clock(&ts);
673 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
676 write_seqlock_irqsave(&timekeeper.lock, flags);
678 timekeeping_forward_now();
680 __timekeeping_inject_sleeptime(delta);
682 timekeeping_update(true);
684 write_sequnlock_irqrestore(&timekeeper.lock, flags);
686 /* signal hrtimers about time change */
692 * timekeeping_resume - Resumes the generic timekeeping subsystem.
694 * This is for the generic clocksource timekeeping.
695 * xtime/wall_to_monotonic/jiffies/etc are
696 * still managed by arch specific suspend/resume code.
698 static void timekeeping_resume(void)
703 read_persistent_clock(&ts);
705 clocksource_resume();
707 write_seqlock_irqsave(&timekeeper.lock, flags);
709 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
710 ts = timespec_sub(ts, timekeeping_suspend_time);
711 __timekeeping_inject_sleeptime(&ts);
713 /* re-base the last cycle value */
714 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
715 timekeeper.ntp_error = 0;
716 timekeeping_suspended = 0;
717 write_sequnlock_irqrestore(&timekeeper.lock, flags);
719 touch_softlockup_watchdog();
721 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
723 /* Resume hrtimers */
727 static int timekeeping_suspend(void)
730 struct timespec delta, delta_delta;
731 static struct timespec old_delta;
733 read_persistent_clock(&timekeeping_suspend_time);
735 write_seqlock_irqsave(&timekeeper.lock, flags);
736 timekeeping_forward_now();
737 timekeeping_suspended = 1;
740 * To avoid drift caused by repeated suspend/resumes,
741 * which each can add ~1 second drift error,
742 * try to compensate so the difference in system time
743 * and persistent_clock time stays close to constant.
745 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
746 delta_delta = timespec_sub(delta, old_delta);
747 if (abs(delta_delta.tv_sec) >= 2) {
749 * if delta_delta is too large, assume time correction
750 * has occured and set old_delta to the current delta.
754 /* Otherwise try to adjust old_system to compensate */
755 timekeeping_suspend_time =
756 timespec_add(timekeeping_suspend_time, delta_delta);
758 write_sequnlock_irqrestore(&timekeeper.lock, flags);
760 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
761 clocksource_suspend();
766 /* sysfs resume/suspend bits for timekeeping */
767 static struct syscore_ops timekeeping_syscore_ops = {
768 .resume = timekeeping_resume,
769 .suspend = timekeeping_suspend,
772 static int __init timekeeping_init_ops(void)
774 register_syscore_ops(&timekeeping_syscore_ops);
778 device_initcall(timekeeping_init_ops);
781 * If the error is already larger, we look ahead even further
782 * to compensate for late or lost adjustments.
784 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
792 * Use the current error value to determine how much to look ahead.
793 * The larger the error the slower we adjust for it to avoid problems
794 * with losing too many ticks, otherwise we would overadjust and
795 * produce an even larger error. The smaller the adjustment the
796 * faster we try to adjust for it, as lost ticks can do less harm
797 * here. This is tuned so that an error of about 1 msec is adjusted
798 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
800 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
801 error2 = abs(error2);
802 for (look_ahead = 0; error2 > 0; look_ahead++)
806 * Now calculate the error in (1 << look_ahead) ticks, but first
807 * remove the single look ahead already included in the error.
809 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
810 tick_error -= timekeeper.xtime_interval >> 1;
811 error = ((error - tick_error) >> look_ahead) + tick_error;
813 /* Finally calculate the adjustment shift value. */
818 *interval = -*interval;
822 for (adj = 0; error > i; adj++)
831 * Adjust the multiplier to reduce the error value,
832 * this is optimized for the most common adjustments of -1,0,1,
833 * for other values we can do a bit more work.
835 static void timekeeping_adjust(s64 offset)
837 s64 error, interval = timekeeper.cycle_interval;
841 * The point of this is to check if the error is greater than half
844 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
846 * Note we subtract one in the shift, so that error is really error*2.
847 * This "saves" dividing(shifting) interval twice, but keeps the
848 * (error > interval) comparison as still measuring if error is
849 * larger than half an interval.
851 * Note: It does not "save" on aggravation when reading the code.
853 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
854 if (error > interval) {
856 * We now divide error by 4(via shift), which checks if
857 * the error is greater than twice the interval.
858 * If it is greater, we need a bigadjust, if its smaller,
859 * we can adjust by 1.
863 * XXX - In update_wall_time, we round up to the next
864 * nanosecond, and store the amount rounded up into
865 * the error. This causes the likely below to be unlikely.
867 * The proper fix is to avoid rounding up by using
868 * the high precision timekeeper.xtime_nsec instead of
869 * xtime.tv_nsec everywhere. Fixing this will take some
872 if (likely(error <= interval))
875 adj = timekeeping_bigadjust(error, &interval, &offset);
876 } else if (error < -interval) {
877 /* See comment above, this is just switched for the negative */
879 if (likely(error >= -interval)) {
881 interval = -interval;
884 adj = timekeeping_bigadjust(error, &interval, &offset);
885 } else /* No adjustment needed */
888 if (unlikely(timekeeper.clock->maxadj &&
889 (timekeeper.mult + adj >
890 timekeeper.clock->mult + timekeeper.clock->maxadj))) {
891 printk_once(KERN_WARNING
892 "Adjusting %s more than 11%% (%ld vs %ld)\n",
893 timekeeper.clock->name, (long)timekeeper.mult + adj,
894 (long)timekeeper.clock->mult +
895 timekeeper.clock->maxadj);
898 * So the following can be confusing.
900 * To keep things simple, lets assume adj == 1 for now.
902 * When adj != 1, remember that the interval and offset values
903 * have been appropriately scaled so the math is the same.
905 * The basic idea here is that we're increasing the multiplier
906 * by one, this causes the xtime_interval to be incremented by
907 * one cycle_interval. This is because:
908 * xtime_interval = cycle_interval * mult
909 * So if mult is being incremented by one:
910 * xtime_interval = cycle_interval * (mult + 1)
912 * xtime_interval = (cycle_interval * mult) + cycle_interval
913 * Which can be shortened to:
914 * xtime_interval += cycle_interval
916 * So offset stores the non-accumulated cycles. Thus the current
917 * time (in shifted nanoseconds) is:
918 * now = (offset * adj) + xtime_nsec
919 * Now, even though we're adjusting the clock frequency, we have
920 * to keep time consistent. In other words, we can't jump back
921 * in time, and we also want to avoid jumping forward in time.
923 * So given the same offset value, we need the time to be the same
924 * both before and after the freq adjustment.
925 * now = (offset * adj_1) + xtime_nsec_1
926 * now = (offset * adj_2) + xtime_nsec_2
928 * (offset * adj_1) + xtime_nsec_1 =
929 * (offset * adj_2) + xtime_nsec_2
933 * (offset * adj_1) + xtime_nsec_1 =
934 * (offset * (adj_1+1)) + xtime_nsec_2
935 * (offset * adj_1) + xtime_nsec_1 =
936 * (offset * adj_1) + offset + xtime_nsec_2
937 * Canceling the sides:
938 * xtime_nsec_1 = offset + xtime_nsec_2
940 * xtime_nsec_2 = xtime_nsec_1 - offset
941 * Which simplfies to:
942 * xtime_nsec -= offset
944 * XXX - TODO: Doc ntp_error calculation.
946 timekeeper.mult += adj;
947 timekeeper.xtime_interval += interval;
948 timekeeper.xtime_nsec -= offset;
949 timekeeper.ntp_error -= (interval - offset) <<
950 timekeeper.ntp_error_shift;
955 * logarithmic_accumulation - shifted accumulation of cycles
957 * This functions accumulates a shifted interval of cycles into
958 * into a shifted interval nanoseconds. Allows for O(log) accumulation
961 * Returns the unconsumed cycles.
963 static cycle_t logarithmic_accumulation(cycle_t offset, u32 shift)
965 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
968 /* If the offset is smaller than a shifted interval, do nothing */
969 if (offset < timekeeper.cycle_interval<<shift)
972 /* Accumulate one shifted interval */
973 offset -= timekeeper.cycle_interval << shift;
974 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
976 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
977 while (timekeeper.xtime_nsec >= nsecps) {
979 timekeeper.xtime_nsec -= nsecps;
980 timekeeper.xtime.tv_sec++;
981 leap = second_overflow(timekeeper.xtime.tv_sec);
982 timekeeper.xtime.tv_sec += leap;
983 timekeeper.wall_to_monotonic.tv_sec -= leap;
985 clock_was_set_delayed();
988 /* Accumulate raw time */
989 raw_nsecs = timekeeper.raw_interval << shift;
990 raw_nsecs += timekeeper.raw_time.tv_nsec;
991 if (raw_nsecs >= NSEC_PER_SEC) {
992 u64 raw_secs = raw_nsecs;
993 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
994 timekeeper.raw_time.tv_sec += raw_secs;
996 timekeeper.raw_time.tv_nsec = raw_nsecs;
998 /* Accumulate error between NTP and clock interval */
999 timekeeper.ntp_error += ntp_tick_length() << shift;
1000 timekeeper.ntp_error -=
1001 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
1002 (timekeeper.ntp_error_shift + shift);
1009 * update_wall_time - Uses the current clocksource to increment the wall time
1012 static void update_wall_time(void)
1014 struct clocksource *clock;
1016 int shift = 0, maxshift;
1017 unsigned long flags;
1019 write_seqlock_irqsave(&timekeeper.lock, flags);
1021 /* Make sure we're fully resumed: */
1022 if (unlikely(timekeeping_suspended))
1025 clock = timekeeper.clock;
1027 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1028 offset = timekeeper.cycle_interval;
1030 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1032 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1036 * With NO_HZ we may have to accumulate many cycle_intervals
1037 * (think "ticks") worth of time at once. To do this efficiently,
1038 * we calculate the largest doubling multiple of cycle_intervals
1039 * that is smaller than the offset. We then accumulate that
1040 * chunk in one go, and then try to consume the next smaller
1043 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1044 shift = max(0, shift);
1045 /* Bound shift to one less than what overflows tick_length */
1046 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1047 shift = min(shift, maxshift);
1048 while (offset >= timekeeper.cycle_interval) {
1049 offset = logarithmic_accumulation(offset, shift);
1050 if(offset < timekeeper.cycle_interval<<shift)
1054 /* correct the clock when NTP error is too big */
1055 timekeeping_adjust(offset);
1058 * Since in the loop above, we accumulate any amount of time
1059 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1060 * xtime_nsec to be fairly small after the loop. Further, if we're
1061 * slightly speeding the clocksource up in timekeeping_adjust(),
1062 * its possible the required corrective factor to xtime_nsec could
1063 * cause it to underflow.
1065 * Now, we cannot simply roll the accumulated second back, since
1066 * the NTP subsystem has been notified via second_overflow. So
1067 * instead we push xtime_nsec forward by the amount we underflowed,
1068 * and add that amount into the error.
1070 * We'll correct this error next time through this function, when
1071 * xtime_nsec is not as small.
1073 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1074 s64 neg = -(s64)timekeeper.xtime_nsec;
1075 timekeeper.xtime_nsec = 0;
1076 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1081 * Store full nanoseconds into xtime after rounding it up and
1082 * add the remainder to the error difference.
1084 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1085 timekeeper.shift) + 1;
1086 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1088 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1089 timekeeper.ntp_error_shift;
1092 * Finally, make sure that after the rounding
1093 * xtime.tv_nsec isn't larger than NSEC_PER_SEC
1095 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1097 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1098 timekeeper.xtime.tv_sec++;
1099 leap = second_overflow(timekeeper.xtime.tv_sec);
1100 timekeeper.xtime.tv_sec += leap;
1101 timekeeper.wall_to_monotonic.tv_sec -= leap;
1103 clock_was_set_delayed();
1106 timekeeping_update(false);
1109 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1114 * getboottime - Return the real time of system boot.
1115 * @ts: pointer to the timespec to be set
1117 * Returns the wall-time of boot in a timespec.
1119 * This is based on the wall_to_monotonic offset and the total suspend
1120 * time. Calls to settimeofday will affect the value returned (which
1121 * basically means that however wrong your real time clock is at boot time,
1122 * you get the right time here).
1124 void getboottime(struct timespec *ts)
1126 struct timespec boottime = {
1127 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1128 timekeeper.total_sleep_time.tv_sec,
1129 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1130 timekeeper.total_sleep_time.tv_nsec
1133 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1135 EXPORT_SYMBOL_GPL(getboottime);
1139 * get_monotonic_boottime - Returns monotonic time since boot
1140 * @ts: pointer to the timespec to be set
1142 * Returns the monotonic time since boot in a timespec.
1144 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1145 * includes the time spent in suspend.
1147 void get_monotonic_boottime(struct timespec *ts)
1149 struct timespec tomono, sleep;
1153 WARN_ON(timekeeping_suspended);
1156 seq = read_seqbegin(&timekeeper.lock);
1157 *ts = timekeeper.xtime;
1158 tomono = timekeeper.wall_to_monotonic;
1159 sleep = timekeeper.total_sleep_time;
1160 nsecs = timekeeping_get_ns();
1162 } while (read_seqretry(&timekeeper.lock, seq));
1164 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1165 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1167 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1170 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1172 * Returns the monotonic time since boot in a ktime
1174 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1175 * includes the time spent in suspend.
1177 ktime_t ktime_get_boottime(void)
1181 get_monotonic_boottime(&ts);
1182 return timespec_to_ktime(ts);
1184 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1187 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1188 * @ts: pointer to the timespec to be converted
1190 void monotonic_to_bootbased(struct timespec *ts)
1192 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1194 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1196 unsigned long get_seconds(void)
1198 return timekeeper.xtime.tv_sec;
1200 EXPORT_SYMBOL(get_seconds);
1202 struct timespec __current_kernel_time(void)
1204 return timekeeper.xtime;
1207 struct timespec current_kernel_time(void)
1209 struct timespec now;
1213 seq = read_seqbegin(&timekeeper.lock);
1215 now = timekeeper.xtime;
1216 } while (read_seqretry(&timekeeper.lock, seq));
1220 EXPORT_SYMBOL(current_kernel_time);
1222 struct timespec get_monotonic_coarse(void)
1224 struct timespec now, mono;
1228 seq = read_seqbegin(&timekeeper.lock);
1230 now = timekeeper.xtime;
1231 mono = timekeeper.wall_to_monotonic;
1232 } while (read_seqretry(&timekeeper.lock, seq));
1234 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1235 now.tv_nsec + mono.tv_nsec);
1240 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1241 * without sampling the sequence number in xtime_lock.
1242 * jiffies is defined in the linker script...
1244 void do_timer(unsigned long ticks)
1246 jiffies_64 += ticks;
1248 calc_global_load(ticks);
1252 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1253 * and sleep offsets.
1254 * @xtim: pointer to timespec to be set with xtime
1255 * @wtom: pointer to timespec to be set with wall_to_monotonic
1256 * @sleep: pointer to timespec to be set with time in suspend
1258 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1259 struct timespec *wtom, struct timespec *sleep)
1264 seq = read_seqbegin(&timekeeper.lock);
1265 *xtim = timekeeper.xtime;
1266 *wtom = timekeeper.wall_to_monotonic;
1267 *sleep = timekeeper.total_sleep_time;
1268 } while (read_seqretry(&timekeeper.lock, seq));
1271 #ifdef CONFIG_HIGH_RES_TIMERS
1273 * ktime_get_update_offsets - hrtimer helper
1274 * @offs_real: pointer to storage for monotonic -> realtime offset
1275 * @offs_boot: pointer to storage for monotonic -> boottime offset
1277 * Returns current monotonic time and updates the offsets
1278 * Called from hrtimer_interupt() or retrigger_next_event()
1280 ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1287 seq = read_seqbegin(&timekeeper.lock);
1289 secs = timekeeper.xtime.tv_sec;
1290 nsecs = timekeeper.xtime.tv_nsec;
1291 nsecs += timekeeping_get_ns();
1292 /* If arch requires, add in gettimeoffset() */
1293 nsecs += arch_gettimeoffset();
1295 *offs_real = timekeeper.offs_real;
1296 *offs_boot = timekeeper.offs_boot;
1297 } while (read_seqretry(&timekeeper.lock, seq));
1299 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1300 now = ktime_sub(now, *offs_real);
1306 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1308 ktime_t ktime_get_monotonic_offset(void)
1311 struct timespec wtom;
1314 seq = read_seqbegin(&timekeeper.lock);
1315 wtom = timekeeper.wall_to_monotonic;
1316 } while (read_seqretry(&timekeeper.lock, seq));
1318 return timespec_to_ktime(wtom);
1320 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1324 * xtime_update() - advances the timekeeping infrastructure
1325 * @ticks: number of ticks, that have elapsed since the last call.
1327 * Must be called with interrupts disabled.
1329 void xtime_update(unsigned long ticks)
1331 write_seqlock(&xtime_lock);
1333 write_sequnlock(&xtime_lock);