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>
23 #ifdef CONFIG_XEN_PRIVILEGED_GUEST
27 /* Structure holding internal timekeeping values. */
29 /* Current clocksource used for timekeeping. */
30 struct clocksource *clock;
31 /* NTP adjusted clock multiplier */
33 /* The shift value of the current clocksource. */
36 /* Number of clock cycles in one NTP interval. */
37 cycle_t cycle_interval;
38 /* Number of clock shifted nano seconds in one NTP interval. */
40 /* shifted nano seconds left over when rounding cycle_interval */
42 /* Raw nano seconds accumulated per NTP interval. */
45 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
47 /* Difference between accumulated time and NTP time in ntp
48 * shifted nano seconds. */
50 /* Shift conversion between clock shifted nano seconds and
51 * ntp shifted nano seconds. */
54 /* The current time */
55 struct timespec xtime;
57 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
58 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
59 * at zero at system boot time, so wall_to_monotonic will be negative,
60 * however, we will ALWAYS keep the tv_nsec part positive so we can use
61 * the usual normalization.
63 * wall_to_monotonic is moved after resume from suspend for the
64 * monotonic time not to jump. We need to add total_sleep_time to
65 * wall_to_monotonic to get the real boot based time offset.
67 * - wall_to_monotonic is no longer the boot time, getboottime must be
70 struct timespec wall_to_monotonic;
71 /* time spent in suspend */
72 struct timespec total_sleep_time;
73 /* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. */
74 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);
89 /* flag for if timekeeping is suspended */
90 int __read_mostly timekeeping_suspended;
95 * timekeeper_setup_internals - Set up internals to use clocksource clock.
97 * @clock: Pointer to clocksource.
99 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
100 * pair and interval request.
102 * Unless you're the timekeeping code, you should not be using this!
104 static void timekeeper_setup_internals(struct clocksource *clock)
107 u64 tmp, ntpinterval;
109 timekeeper.clock = clock;
110 clock->cycle_last = clock->read(clock);
112 /* Do the ns -> cycle conversion first, using original mult */
113 tmp = NTP_INTERVAL_LENGTH;
114 tmp <<= clock->shift;
116 tmp += clock->mult/2;
117 do_div(tmp, clock->mult);
121 interval = (cycle_t) tmp;
122 timekeeper.cycle_interval = interval;
124 /* Go back from cycles -> shifted ns */
125 timekeeper.xtime_interval = (u64) interval * clock->mult;
126 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
127 timekeeper.raw_interval =
128 ((u64) interval * clock->mult) >> clock->shift;
130 timekeeper.xtime_nsec = 0;
131 timekeeper.shift = clock->shift;
133 timekeeper.ntp_error = 0;
134 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
137 * The timekeeper keeps its own mult values for the currently
138 * active clocksource. These value will be adjusted via NTP
139 * to counteract clock drifting.
141 timekeeper.mult = clock->mult;
144 /* Timekeeper helper functions. */
145 static inline s64 timekeeping_get_ns(void)
147 cycle_t cycle_now, cycle_delta;
148 struct clocksource *clock;
150 /* read clocksource: */
151 clock = timekeeper.clock;
152 cycle_now = clock->read(clock);
154 /* calculate the delta since the last update_wall_time: */
155 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
157 /* return delta convert to nanoseconds using ntp adjusted mult. */
158 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
162 static inline s64 timekeeping_get_ns_raw(void)
164 cycle_t cycle_now, cycle_delta;
165 struct clocksource *clock;
167 /* read clocksource: */
168 clock = timekeeper.clock;
169 cycle_now = clock->read(clock);
171 /* calculate the delta since the last update_wall_time: */
172 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
174 /* return delta convert to nanoseconds. */
175 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
178 /* must hold write on timekeeper.lock */
179 static void timekeeping_update(bool clearntp)
182 timekeeper.ntp_error = 0;
185 update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
186 timekeeper.clock, timekeeper.mult);
191 * timekeeping_forward_now - update clock to the current time
193 * Forward the current clock to update its state since the last call to
194 * update_wall_time(). This is useful before significant clock changes,
195 * as it avoids having to deal with this time offset explicitly.
197 static void timekeeping_forward_now(void)
199 cycle_t cycle_now, cycle_delta;
200 struct clocksource *clock;
203 clock = timekeeper.clock;
204 cycle_now = clock->read(clock);
205 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
206 clock->cycle_last = cycle_now;
208 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
211 /* If arch requires, add in gettimeoffset() */
212 nsec += arch_gettimeoffset();
214 timespec_add_ns(&timekeeper.xtime, nsec);
216 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
217 timespec_add_ns(&timekeeper.raw_time, nsec);
221 * getnstimeofday - Returns the time of day in a timespec
222 * @ts: pointer to the timespec to be set
224 * Returns the time of day in a timespec.
226 void getnstimeofday(struct timespec *ts)
231 WARN_ON(timekeeping_suspended);
234 seq = read_seqbegin(&timekeeper.lock);
236 *ts = timekeeper.xtime;
237 nsecs = timekeeping_get_ns();
239 /* If arch requires, add in gettimeoffset() */
240 nsecs += arch_gettimeoffset();
242 } while (read_seqretry(&timekeeper.lock, seq));
244 timespec_add_ns(ts, nsecs);
247 EXPORT_SYMBOL(getnstimeofday);
249 ktime_t ktime_get(void)
254 WARN_ON(timekeeping_suspended);
257 seq = read_seqbegin(&timekeeper.lock);
258 secs = timekeeper.xtime.tv_sec +
259 timekeeper.wall_to_monotonic.tv_sec;
260 nsecs = timekeeper.xtime.tv_nsec +
261 timekeeper.wall_to_monotonic.tv_nsec;
262 nsecs += timekeeping_get_ns();
263 /* If arch requires, add in gettimeoffset() */
264 nsecs += arch_gettimeoffset();
266 } while (read_seqretry(&timekeeper.lock, seq));
268 * Use ktime_set/ktime_add_ns to create a proper ktime on
269 * 32-bit architectures without CONFIG_KTIME_SCALAR.
271 return ktime_add_ns(ktime_set(secs, 0), nsecs);
273 EXPORT_SYMBOL_GPL(ktime_get);
276 * ktime_get_ts - get the monotonic clock in timespec format
277 * @ts: pointer to timespec variable
279 * The function calculates the monotonic clock from the realtime
280 * clock and the wall_to_monotonic offset and stores the result
281 * in normalized timespec format in the variable pointed to by @ts.
283 void ktime_get_ts(struct timespec *ts)
285 struct timespec tomono;
289 WARN_ON(timekeeping_suspended);
292 seq = read_seqbegin(&timekeeper.lock);
293 *ts = timekeeper.xtime;
294 tomono = timekeeper.wall_to_monotonic;
295 nsecs = timekeeping_get_ns();
296 /* If arch requires, add in gettimeoffset() */
297 nsecs += arch_gettimeoffset();
299 } while (read_seqretry(&timekeeper.lock, seq));
301 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
302 ts->tv_nsec + tomono.tv_nsec + nsecs);
304 EXPORT_SYMBOL_GPL(ktime_get_ts);
306 #ifdef CONFIG_NTP_PPS
309 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
310 * @ts_raw: pointer to the timespec to be set to raw monotonic time
311 * @ts_real: pointer to the timespec to be set to the time of day
313 * This function reads both the time of day and raw monotonic time at the
314 * same time atomically and stores the resulting timestamps in timespec
317 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
320 s64 nsecs_raw, nsecs_real;
322 WARN_ON_ONCE(timekeeping_suspended);
327 seq = read_seqbegin(&timekeeper.lock);
329 *ts_raw = timekeeper.raw_time;
330 *ts_real = timekeeper.xtime;
332 nsecs_raw = timekeeping_get_ns_raw();
333 nsecs_real = timekeeping_get_ns();
335 /* If arch requires, add in gettimeoffset() */
336 arch_offset = arch_gettimeoffset();
337 nsecs_raw += arch_offset;
338 nsecs_real += arch_offset;
340 } while (read_seqretry(&timekeeper.lock, seq));
342 timespec_add_ns(ts_raw, nsecs_raw);
343 timespec_add_ns(ts_real, nsecs_real);
345 EXPORT_SYMBOL(getnstime_raw_and_real);
347 #endif /* CONFIG_NTP_PPS */
350 * do_gettimeofday - Returns the time of day in a timeval
351 * @tv: pointer to the timeval to be set
353 * NOTE: Users should be converted to using getnstimeofday()
355 void do_gettimeofday(struct timeval *tv)
359 getnstimeofday(&now);
360 tv->tv_sec = now.tv_sec;
361 tv->tv_usec = now.tv_nsec/1000;
364 EXPORT_SYMBOL(do_gettimeofday);
366 * do_settimeofday - Sets the time of day
367 * @tv: pointer to the timespec variable containing the new time
369 * Sets the time of day to the new time and update NTP and notify hrtimers
371 int do_settimeofday(const struct timespec *tv)
373 struct timespec ts_delta;
376 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
379 write_seqlock_irqsave(&timekeeper.lock, flags);
381 timekeeping_forward_now();
383 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
384 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
385 timekeeper.wall_to_monotonic =
386 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
388 timekeeper.xtime = *tv;
389 timekeeping_update(true);
390 #ifdef CONFIG_XEN_PRIVILEGED_GUEST
391 xen_update_wallclock(tv);
394 write_sequnlock_irqrestore(&timekeeper.lock, flags);
396 /* signal hrtimers about time change */
402 EXPORT_SYMBOL(do_settimeofday);
406 * timekeeping_inject_offset - Adds or subtracts from the current time.
407 * @tv: pointer to the timespec variable containing the offset
409 * Adds or subtracts an offset value from the current time.
411 int timekeeping_inject_offset(struct timespec *ts)
415 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
418 write_seqlock_irqsave(&timekeeper.lock, flags);
420 timekeeping_forward_now();
422 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
423 timekeeper.wall_to_monotonic =
424 timespec_sub(timekeeper.wall_to_monotonic, *ts);
426 timekeeping_update(true);
428 write_sequnlock_irqrestore(&timekeeper.lock, flags);
430 /* signal hrtimers about time change */
435 EXPORT_SYMBOL(timekeeping_inject_offset);
438 * change_clocksource - Swaps clocksources if a new one is available
440 * Accumulates current time interval and initializes new clocksource
442 static int change_clocksource(void *data)
444 struct clocksource *new, *old;
447 new = (struct clocksource *) data;
449 write_seqlock_irqsave(&timekeeper.lock, flags);
451 timekeeping_forward_now();
452 if (!new->enable || new->enable(new) == 0) {
453 old = timekeeper.clock;
454 timekeeper_setup_internals(new);
458 timekeeping_update(true);
460 write_sequnlock_irqrestore(&timekeeper.lock, flags);
466 * timekeeping_notify - Install a new clock source
467 * @clock: pointer to the clock source
469 * This function is called from clocksource.c after a new, better clock
470 * source has been registered. The caller holds the clocksource_mutex.
472 void timekeeping_notify(struct clocksource *clock)
474 if (timekeeper.clock == clock)
476 stop_machine(change_clocksource, clock, NULL);
481 * ktime_get_real - get the real (wall-) time in ktime_t format
483 * returns the time in ktime_t format
485 ktime_t ktime_get_real(void)
489 getnstimeofday(&now);
491 return timespec_to_ktime(now);
493 EXPORT_SYMBOL_GPL(ktime_get_real);
496 * getrawmonotonic - Returns the raw monotonic time in a timespec
497 * @ts: pointer to the timespec to be set
499 * Returns the raw monotonic time (completely un-modified by ntp)
501 void getrawmonotonic(struct timespec *ts)
507 seq = read_seqbegin(&timekeeper.lock);
508 nsecs = timekeeping_get_ns_raw();
509 *ts = timekeeper.raw_time;
511 } while (read_seqretry(&timekeeper.lock, seq));
513 timespec_add_ns(ts, nsecs);
515 EXPORT_SYMBOL(getrawmonotonic);
519 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
521 int timekeeping_valid_for_hres(void)
527 seq = read_seqbegin(&timekeeper.lock);
529 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
531 } while (read_seqretry(&timekeeper.lock, seq));
537 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
539 u64 timekeeping_max_deferment(void)
544 seq = read_seqbegin(&timekeeper.lock);
546 ret = timekeeper.clock->max_idle_ns;
548 } while (read_seqretry(&timekeeper.lock, seq));
554 * read_persistent_clock - Return time from the persistent clock.
556 * Weak dummy function for arches that do not yet support it.
557 * Reads the time from the battery backed persistent clock.
558 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
560 * XXX - Do be sure to remove it once all arches implement it.
562 void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
569 * read_boot_clock - Return time of the system start.
571 * Weak dummy function for arches that do not yet support it.
572 * Function to read the exact time the system has been started.
573 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
575 * XXX - Do be sure to remove it once all arches implement it.
577 void __attribute__((weak)) read_boot_clock(struct timespec *ts)
584 * timekeeping_init - Initializes the clocksource and common timekeeping values
586 void __init timekeeping_init(void)
588 struct clocksource *clock;
590 struct timespec now, boot;
592 read_persistent_clock(&now);
593 read_boot_clock(&boot);
595 seqlock_init(&timekeeper.lock);
599 write_seqlock_irqsave(&timekeeper.lock, flags);
600 clock = clocksource_default_clock();
602 clock->enable(clock);
603 timekeeper_setup_internals(clock);
605 timekeeper.xtime.tv_sec = now.tv_sec;
606 timekeeper.xtime.tv_nsec = now.tv_nsec;
607 timekeeper.raw_time.tv_sec = 0;
608 timekeeper.raw_time.tv_nsec = 0;
609 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
610 boot.tv_sec = timekeeper.xtime.tv_sec;
611 boot.tv_nsec = timekeeper.xtime.tv_nsec;
613 set_normalized_timespec(&timekeeper.wall_to_monotonic,
614 -boot.tv_sec, -boot.tv_nsec);
615 timekeeper.total_sleep_time.tv_sec = 0;
616 timekeeper.total_sleep_time.tv_nsec = 0;
617 write_sequnlock_irqrestore(&timekeeper.lock, flags);
620 /* time in seconds when suspend began */
621 static struct timespec timekeeping_suspend_time;
624 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
625 * @delta: pointer to a timespec delta value
627 * Takes a timespec offset measuring a suspend interval and properly
628 * adds the sleep offset to the timekeeping variables.
630 static void __timekeeping_inject_sleeptime(struct timespec *delta)
632 if (!timespec_valid(delta)) {
633 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
634 "sleep delta value!\n");
638 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
639 timekeeper.wall_to_monotonic =
640 timespec_sub(timekeeper.wall_to_monotonic, *delta);
641 timekeeper.total_sleep_time = timespec_add(
642 timekeeper.total_sleep_time, *delta);
647 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
648 * @delta: pointer to a timespec delta value
650 * This hook is for architectures that cannot support read_persistent_clock
651 * because their RTC/persistent clock is only accessible when irqs are enabled.
653 * This function should only be called by rtc_resume(), and allows
654 * a suspend offset to be injected into the timekeeping values.
656 void timekeeping_inject_sleeptime(struct timespec *delta)
661 /* Make sure we don't set the clock twice */
662 read_persistent_clock(&ts);
663 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
666 write_seqlock_irqsave(&timekeeper.lock, flags);
668 timekeeping_forward_now();
670 __timekeeping_inject_sleeptime(delta);
672 timekeeping_update(true);
674 write_sequnlock_irqrestore(&timekeeper.lock, flags);
676 /* signal hrtimers about time change */
682 * timekeeping_resume - Resumes the generic timekeeping subsystem.
684 * This is for the generic clocksource timekeeping.
685 * xtime/wall_to_monotonic/jiffies/etc are
686 * still managed by arch specific suspend/resume code.
688 static void timekeeping_resume(void)
693 read_persistent_clock(&ts);
695 clocksource_resume();
697 write_seqlock_irqsave(&timekeeper.lock, flags);
699 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
700 ts = timespec_sub(ts, timekeeping_suspend_time);
701 __timekeeping_inject_sleeptime(&ts);
703 /* re-base the last cycle value */
704 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
705 timekeeper.ntp_error = 0;
706 timekeeping_suspended = 0;
707 write_sequnlock_irqrestore(&timekeeper.lock, flags);
709 touch_softlockup_watchdog();
711 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
713 /* Resume hrtimers */
717 static int timekeeping_suspend(void)
720 struct timespec delta, delta_delta;
721 static struct timespec old_delta;
723 read_persistent_clock(&timekeeping_suspend_time);
725 write_seqlock_irqsave(&timekeeper.lock, flags);
726 timekeeping_forward_now();
727 timekeeping_suspended = 1;
730 * To avoid drift caused by repeated suspend/resumes,
731 * which each can add ~1 second drift error,
732 * try to compensate so the difference in system time
733 * and persistent_clock time stays close to constant.
735 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
736 delta_delta = timespec_sub(delta, old_delta);
737 if (abs(delta_delta.tv_sec) >= 2) {
739 * if delta_delta is too large, assume time correction
740 * has occured and set old_delta to the current delta.
744 /* Otherwise try to adjust old_system to compensate */
745 timekeeping_suspend_time =
746 timespec_add(timekeeping_suspend_time, delta_delta);
748 write_sequnlock_irqrestore(&timekeeper.lock, flags);
750 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
751 clocksource_suspend();
756 /* sysfs resume/suspend bits for timekeeping */
757 static struct syscore_ops timekeeping_syscore_ops = {
758 .resume = timekeeping_resume,
759 .suspend = timekeeping_suspend,
762 static int __init timekeeping_init_ops(void)
764 register_syscore_ops(&timekeeping_syscore_ops);
768 device_initcall(timekeeping_init_ops);
771 * If the error is already larger, we look ahead even further
772 * to compensate for late or lost adjustments.
774 static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
782 * Use the current error value to determine how much to look ahead.
783 * The larger the error the slower we adjust for it to avoid problems
784 * with losing too many ticks, otherwise we would overadjust and
785 * produce an even larger error. The smaller the adjustment the
786 * faster we try to adjust for it, as lost ticks can do less harm
787 * here. This is tuned so that an error of about 1 msec is adjusted
788 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
790 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
791 error2 = abs(error2);
792 for (look_ahead = 0; error2 > 0; look_ahead++)
796 * Now calculate the error in (1 << look_ahead) ticks, but first
797 * remove the single look ahead already included in the error.
799 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
800 tick_error -= timekeeper.xtime_interval >> 1;
801 error = ((error - tick_error) >> look_ahead) + tick_error;
803 /* Finally calculate the adjustment shift value. */
808 *interval = -*interval;
812 for (adj = 0; error > i; adj++)
821 * Adjust the multiplier to reduce the error value,
822 * this is optimized for the most common adjustments of -1,0,1,
823 * for other values we can do a bit more work.
825 static void timekeeping_adjust(s64 offset)
827 s64 error, interval = timekeeper.cycle_interval;
831 * The point of this is to check if the error is greater than half
834 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
836 * Note we subtract one in the shift, so that error is really error*2.
837 * This "saves" dividing(shifting) interval twice, but keeps the
838 * (error > interval) comparison as still measuring if error is
839 * larger than half an interval.
841 * Note: It does not "save" on aggravation when reading the code.
843 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
844 if (error > interval) {
846 * We now divide error by 4(via shift), which checks if
847 * the error is greater than twice the interval.
848 * If it is greater, we need a bigadjust, if its smaller,
849 * we can adjust by 1.
853 * XXX - In update_wall_time, we round up to the next
854 * nanosecond, and store the amount rounded up into
855 * the error. This causes the likely below to be unlikely.
857 * The proper fix is to avoid rounding up by using
858 * the high precision timekeeper.xtime_nsec instead of
859 * xtime.tv_nsec everywhere. Fixing this will take some
862 if (likely(error <= interval))
865 adj = timekeeping_bigadjust(error, &interval, &offset);
866 } else if (error < -interval) {
867 /* See comment above, this is just switched for the negative */
869 if (likely(error >= -interval)) {
871 interval = -interval;
874 adj = timekeeping_bigadjust(error, &interval, &offset);
875 } else /* No adjustment needed */
878 if (unlikely(timekeeper.clock->maxadj &&
879 (timekeeper.mult + adj >
880 timekeeper.clock->mult + timekeeper.clock->maxadj))) {
881 printk_once(KERN_WARNING
882 "Adjusting %s more than 11%% (%ld vs %ld)\n",
883 timekeeper.clock->name, (long)timekeeper.mult + adj,
884 (long)timekeeper.clock->mult +
885 timekeeper.clock->maxadj);
888 * So the following can be confusing.
890 * To keep things simple, lets assume adj == 1 for now.
892 * When adj != 1, remember that the interval and offset values
893 * have been appropriately scaled so the math is the same.
895 * The basic idea here is that we're increasing the multiplier
896 * by one, this causes the xtime_interval to be incremented by
897 * one cycle_interval. This is because:
898 * xtime_interval = cycle_interval * mult
899 * So if mult is being incremented by one:
900 * xtime_interval = cycle_interval * (mult + 1)
902 * xtime_interval = (cycle_interval * mult) + cycle_interval
903 * Which can be shortened to:
904 * xtime_interval += cycle_interval
906 * So offset stores the non-accumulated cycles. Thus the current
907 * time (in shifted nanoseconds) is:
908 * now = (offset * adj) + xtime_nsec
909 * Now, even though we're adjusting the clock frequency, we have
910 * to keep time consistent. In other words, we can't jump back
911 * in time, and we also want to avoid jumping forward in time.
913 * So given the same offset value, we need the time to be the same
914 * both before and after the freq adjustment.
915 * now = (offset * adj_1) + xtime_nsec_1
916 * now = (offset * adj_2) + xtime_nsec_2
918 * (offset * adj_1) + xtime_nsec_1 =
919 * (offset * adj_2) + xtime_nsec_2
923 * (offset * adj_1) + xtime_nsec_1 =
924 * (offset * (adj_1+1)) + xtime_nsec_2
925 * (offset * adj_1) + xtime_nsec_1 =
926 * (offset * adj_1) + offset + xtime_nsec_2
927 * Canceling the sides:
928 * xtime_nsec_1 = offset + xtime_nsec_2
930 * xtime_nsec_2 = xtime_nsec_1 - offset
931 * Which simplfies to:
932 * xtime_nsec -= offset
934 * XXX - TODO: Doc ntp_error calculation.
936 timekeeper.mult += adj;
937 timekeeper.xtime_interval += interval;
938 timekeeper.xtime_nsec -= offset;
939 timekeeper.ntp_error -= (interval - offset) <<
940 timekeeper.ntp_error_shift;
945 * logarithmic_accumulation - shifted accumulation of cycles
947 * This functions accumulates a shifted interval of cycles into
948 * into a shifted interval nanoseconds. Allows for O(log) accumulation
951 * Returns the unconsumed cycles.
953 static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
955 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
958 /* If the offset is smaller than a shifted interval, do nothing */
959 if (offset < timekeeper.cycle_interval<<shift)
962 /* Accumulate one shifted interval */
963 offset -= timekeeper.cycle_interval << shift;
964 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
966 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
967 while (timekeeper.xtime_nsec >= nsecps) {
969 timekeeper.xtime_nsec -= nsecps;
970 timekeeper.xtime.tv_sec++;
971 leap = second_overflow(timekeeper.xtime.tv_sec);
972 timekeeper.xtime.tv_sec += leap;
975 /* Accumulate raw time */
976 raw_nsecs = timekeeper.raw_interval << shift;
977 raw_nsecs += timekeeper.raw_time.tv_nsec;
978 if (raw_nsecs >= NSEC_PER_SEC) {
979 u64 raw_secs = raw_nsecs;
980 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
981 timekeeper.raw_time.tv_sec += raw_secs;
983 timekeeper.raw_time.tv_nsec = raw_nsecs;
985 /* Accumulate error between NTP and clock interval */
986 timekeeper.ntp_error += ntp_tick_length() << shift;
987 timekeeper.ntp_error -=
988 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
989 (timekeeper.ntp_error_shift + shift);
996 * update_wall_time - Uses the current clocksource to increment the wall time
999 static void update_wall_time(void)
1001 struct clocksource *clock;
1003 int shift = 0, maxshift;
1004 unsigned long flags;
1006 write_seqlock_irqsave(&timekeeper.lock, flags);
1008 /* Make sure we're fully resumed: */
1009 if (unlikely(timekeeping_suspended))
1012 clock = timekeeper.clock;
1014 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1015 offset = timekeeper.cycle_interval;
1017 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1019 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1023 * With NO_HZ we may have to accumulate many cycle_intervals
1024 * (think "ticks") worth of time at once. To do this efficiently,
1025 * we calculate the largest doubling multiple of cycle_intervals
1026 * that is smaller than the offset. We then accumulate that
1027 * chunk in one go, and then try to consume the next smaller
1030 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1031 shift = max(0, shift);
1032 /* Bound shift to one less than what overflows tick_length */
1033 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1034 shift = min(shift, maxshift);
1035 while (offset >= timekeeper.cycle_interval) {
1036 offset = logarithmic_accumulation(offset, shift);
1037 if(offset < timekeeper.cycle_interval<<shift)
1041 /* correct the clock when NTP error is too big */
1042 timekeeping_adjust(offset);
1045 * Since in the loop above, we accumulate any amount of time
1046 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1047 * xtime_nsec to be fairly small after the loop. Further, if we're
1048 * slightly speeding the clocksource up in timekeeping_adjust(),
1049 * its possible the required corrective factor to xtime_nsec could
1050 * cause it to underflow.
1052 * Now, we cannot simply roll the accumulated second back, since
1053 * the NTP subsystem has been notified via second_overflow. So
1054 * instead we push xtime_nsec forward by the amount we underflowed,
1055 * and add that amount into the error.
1057 * We'll correct this error next time through this function, when
1058 * xtime_nsec is not as small.
1060 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1061 s64 neg = -(s64)timekeeper.xtime_nsec;
1062 timekeeper.xtime_nsec = 0;
1063 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1068 * Store full nanoseconds into xtime after rounding it up and
1069 * add the remainder to the error difference.
1071 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1072 timekeeper.shift) + 1;
1073 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1075 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1076 timekeeper.ntp_error_shift;
1079 * Finally, make sure that after the rounding
1080 * xtime.tv_nsec isn't larger than NSEC_PER_SEC
1082 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1084 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1085 timekeeper.xtime.tv_sec++;
1086 leap = second_overflow(timekeeper.xtime.tv_sec);
1087 timekeeper.xtime.tv_sec += leap;
1090 timekeeping_update(false);
1093 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1098 * getboottime - Return the real time of system boot.
1099 * @ts: pointer to the timespec to be set
1101 * Returns the wall-time of boot in a timespec.
1103 * This is based on the wall_to_monotonic offset and the total suspend
1104 * time. Calls to settimeofday will affect the value returned (which
1105 * basically means that however wrong your real time clock is at boot time,
1106 * you get the right time here).
1108 void getboottime(struct timespec *ts)
1110 struct timespec boottime = {
1111 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1112 timekeeper.total_sleep_time.tv_sec,
1113 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1114 timekeeper.total_sleep_time.tv_nsec
1117 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1119 EXPORT_SYMBOL_GPL(getboottime);
1123 * get_monotonic_boottime - Returns monotonic time since boot
1124 * @ts: pointer to the timespec to be set
1126 * Returns the monotonic time since boot in a timespec.
1128 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1129 * includes the time spent in suspend.
1131 void get_monotonic_boottime(struct timespec *ts)
1133 struct timespec tomono, sleep;
1137 WARN_ON(timekeeping_suspended);
1140 seq = read_seqbegin(&timekeeper.lock);
1141 *ts = timekeeper.xtime;
1142 tomono = timekeeper.wall_to_monotonic;
1143 sleep = timekeeper.total_sleep_time;
1144 nsecs = timekeeping_get_ns();
1146 } while (read_seqretry(&timekeeper.lock, seq));
1148 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1149 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1151 EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1154 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1156 * Returns the monotonic time since boot in a ktime
1158 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1159 * includes the time spent in suspend.
1161 ktime_t ktime_get_boottime(void)
1165 get_monotonic_boottime(&ts);
1166 return timespec_to_ktime(ts);
1168 EXPORT_SYMBOL_GPL(ktime_get_boottime);
1171 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1172 * @ts: pointer to the timespec to be converted
1174 void monotonic_to_bootbased(struct timespec *ts)
1176 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1178 EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1180 unsigned long get_seconds(void)
1182 return timekeeper.xtime.tv_sec;
1184 EXPORT_SYMBOL(get_seconds);
1186 struct timespec __current_kernel_time(void)
1188 return timekeeper.xtime;
1191 struct timespec current_kernel_time(void)
1193 struct timespec now;
1197 seq = read_seqbegin(&timekeeper.lock);
1199 now = timekeeper.xtime;
1200 } while (read_seqretry(&timekeeper.lock, seq));
1204 EXPORT_SYMBOL(current_kernel_time);
1206 struct timespec get_monotonic_coarse(void)
1208 struct timespec now, mono;
1212 seq = read_seqbegin(&timekeeper.lock);
1214 now = timekeeper.xtime;
1215 mono = timekeeper.wall_to_monotonic;
1216 } while (read_seqretry(&timekeeper.lock, seq));
1218 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1219 now.tv_nsec + mono.tv_nsec);
1224 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1225 * without sampling the sequence number in xtime_lock.
1226 * jiffies is defined in the linker script...
1228 void do_timer(unsigned long ticks)
1230 jiffies_64 += ticks;
1232 calc_global_load(ticks);
1236 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1237 * and sleep offsets.
1238 * @xtim: pointer to timespec to be set with xtime
1239 * @wtom: pointer to timespec to be set with wall_to_monotonic
1240 * @sleep: pointer to timespec to be set with time in suspend
1242 void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1243 struct timespec *wtom, struct timespec *sleep)
1248 seq = read_seqbegin(&timekeeper.lock);
1249 *xtim = timekeeper.xtime;
1250 *wtom = timekeeper.wall_to_monotonic;
1251 *sleep = timekeeper.total_sleep_time;
1252 } while (read_seqretry(&timekeeper.lock, seq));
1256 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1258 ktime_t ktime_get_monotonic_offset(void)
1261 struct timespec wtom;
1264 seq = read_seqbegin(&timekeeper.lock);
1265 wtom = timekeeper.wall_to_monotonic;
1266 } while (read_seqretry(&timekeeper.lock, seq));
1268 return timespec_to_ktime(wtom);
1270 EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1274 * xtime_update() - advances the timekeeping infrastructure
1275 * @ticks: number of ticks, that have elapsed since the last call.
1277 * Must be called with interrupts disabled.
1279 void xtime_update(unsigned long ticks)
1281 write_seqlock(&xtime_lock);
1283 write_sequnlock(&xtime_lock);