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/sysdev.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/tick.h>
24 * This read-write spinlock protects us from races in SMP while
27 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
32 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
33 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
34 * at zero at system boot time, so wall_to_monotonic will be negative,
35 * however, we will ALWAYS keep the tv_nsec part positive so we can use
36 * the usual normalization.
38 * wall_to_monotonic is moved after resume from suspend for the monotonic
39 * time not to jump. We need to add total_sleep_time to wall_to_monotonic
40 * to get the real boot based time offset.
42 * - wall_to_monotonic is no longer the boot time, getboottime must be
45 struct timespec xtime __attribute__ ((aligned (16)));
46 struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
47 static unsigned long total_sleep_time; /* seconds */
49 /* flag for if timekeeping is suspended */
50 int __read_mostly timekeeping_suspended;
52 static struct timespec xtime_cache __attribute__ ((aligned (16)));
53 void update_xtime_cache(u64 nsec)
56 timespec_add_ns(&xtime_cache, nsec);
59 struct clocksource *clock;
61 /* must hold xtime_lock */
62 void timekeeping_leap_insert(int leapsecond)
64 xtime.tv_sec += leapsecond;
65 wall_to_monotonic.tv_sec -= leapsecond;
66 update_vsyscall(&xtime, clock);
69 #ifdef CONFIG_GENERIC_TIME
71 * clocksource_forward_now - update clock to the current time
73 * Forward the current clock to update its state since the last call to
74 * update_wall_time(). This is useful before significant clock changes,
75 * as it avoids having to deal with this time offset explicitly.
77 static void clocksource_forward_now(void)
79 cycle_t cycle_now, cycle_delta;
82 cycle_now = clock->read(clock);
83 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
84 clock->cycle_last = cycle_now;
86 nsec = cyc2ns(clock, cycle_delta);
88 /* If arch requires, add in gettimeoffset() */
89 nsec += arch_gettimeoffset();
91 timespec_add_ns(&xtime, nsec);
93 nsec = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
94 clock->raw_time.tv_nsec += nsec;
98 * getnstimeofday - Returns the time of day in a timespec
99 * @ts: pointer to the timespec to be set
101 * Returns the time of day in a timespec.
103 void getnstimeofday(struct timespec *ts)
105 cycle_t cycle_now, cycle_delta;
109 WARN_ON(timekeeping_suspended);
112 seq = read_seqbegin(&xtime_lock);
116 /* read clocksource: */
117 cycle_now = clock->read(clock);
119 /* calculate the delta since the last update_wall_time: */
120 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
122 /* convert to nanoseconds: */
123 nsecs = cyc2ns(clock, cycle_delta);
125 /* If arch requires, add in gettimeoffset() */
126 nsecs += arch_gettimeoffset();
128 } while (read_seqretry(&xtime_lock, seq));
130 timespec_add_ns(ts, nsecs);
133 EXPORT_SYMBOL(getnstimeofday);
135 ktime_t ktime_get(void)
137 cycle_t cycle_now, cycle_delta;
141 WARN_ON(timekeeping_suspended);
144 seq = read_seqbegin(&xtime_lock);
145 secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
146 nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
148 /* read clocksource: */
149 cycle_now = clock->read(clock);
151 /* calculate the delta since the last update_wall_time: */
152 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
154 /* convert to nanoseconds: */
155 nsecs += cyc2ns(clock, cycle_delta);
157 } while (read_seqretry(&xtime_lock, seq));
159 * Use ktime_set/ktime_add_ns to create a proper ktime on
160 * 32-bit architectures without CONFIG_KTIME_SCALAR.
162 return ktime_add_ns(ktime_set(secs, 0), nsecs);
164 EXPORT_SYMBOL_GPL(ktime_get);
167 * ktime_get_ts - get the monotonic clock in timespec format
168 * @ts: pointer to timespec variable
170 * The function calculates the monotonic clock from the realtime
171 * clock and the wall_to_monotonic offset and stores the result
172 * in normalized timespec format in the variable pointed to by @ts.
174 void ktime_get_ts(struct timespec *ts)
176 cycle_t cycle_now, cycle_delta;
177 struct timespec tomono;
181 WARN_ON(timekeeping_suspended);
184 seq = read_seqbegin(&xtime_lock);
186 tomono = wall_to_monotonic;
188 /* read clocksource: */
189 cycle_now = clock->read(clock);
191 /* calculate the delta since the last update_wall_time: */
192 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
194 /* convert to nanoseconds: */
195 nsecs = cyc2ns(clock, cycle_delta);
197 } while (read_seqretry(&xtime_lock, seq));
199 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
200 ts->tv_nsec + tomono.tv_nsec + nsecs);
202 EXPORT_SYMBOL_GPL(ktime_get_ts);
205 * do_gettimeofday - Returns the time of day in a timeval
206 * @tv: pointer to the timeval to be set
208 * NOTE: Users should be converted to using getnstimeofday()
210 void do_gettimeofday(struct timeval *tv)
214 getnstimeofday(&now);
215 tv->tv_sec = now.tv_sec;
216 tv->tv_usec = now.tv_nsec/1000;
219 EXPORT_SYMBOL(do_gettimeofday);
221 * do_settimeofday - Sets the time of day
222 * @tv: pointer to the timespec variable containing the new time
224 * Sets the time of day to the new time and update NTP and notify hrtimers
226 int do_settimeofday(struct timespec *tv)
228 struct timespec ts_delta;
231 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
234 write_seqlock_irqsave(&xtime_lock, flags);
236 clocksource_forward_now();
238 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
239 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
240 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
244 update_xtime_cache(0);
249 update_vsyscall(&xtime, clock);
251 write_sequnlock_irqrestore(&xtime_lock, flags);
253 /* signal hrtimers about time change */
259 EXPORT_SYMBOL(do_settimeofday);
262 * change_clocksource - Swaps clocksources if a new one is available
264 * Accumulates current time interval and initializes new clocksource
266 static void change_clocksource(void)
268 struct clocksource *new, *old;
270 new = clocksource_get_next();
275 clocksource_forward_now();
277 if (new->enable && !new->enable(new))
280 * The frequency may have changed while the clocksource
281 * was disabled. If so the code in ->enable() must update
282 * the mult value to reflect the new frequency. Make sure
283 * mult_orig follows this change.
285 new->mult_orig = new->mult;
287 new->raw_time = clock->raw_time;
291 * Save mult_orig in mult so that the value can be restored
292 * regardless if ->enable() updates the value of mult or not.
294 old->mult = old->mult_orig;
298 clock->cycle_last = clock->read(clock);
300 clock->xtime_nsec = 0;
301 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
306 * We're holding xtime lock and waking up klogd would deadlock
307 * us on enqueue. So no printing!
308 printk(KERN_INFO "Time: %s clocksource has been installed.\n",
312 #else /* GENERIC_TIME */
313 static inline void clocksource_forward_now(void) { }
314 static inline void change_clocksource(void) { }
317 * ktime_get - get the monotonic time in ktime_t format
319 * returns the time in ktime_t format
321 ktime_t ktime_get(void)
327 return timespec_to_ktime(now);
329 EXPORT_SYMBOL_GPL(ktime_get);
332 * ktime_get_ts - get the monotonic clock in timespec format
333 * @ts: pointer to timespec variable
335 * The function calculates the monotonic clock from the realtime
336 * clock and the wall_to_monotonic offset and stores the result
337 * in normalized timespec format in the variable pointed to by @ts.
339 void ktime_get_ts(struct timespec *ts)
341 struct timespec tomono;
345 seq = read_seqbegin(&xtime_lock);
347 tomono = wall_to_monotonic;
349 } while (read_seqretry(&xtime_lock, seq));
351 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
352 ts->tv_nsec + tomono.tv_nsec);
354 EXPORT_SYMBOL_GPL(ktime_get_ts);
355 #endif /* !GENERIC_TIME */
358 * ktime_get_real - get the real (wall-) time in ktime_t format
360 * returns the time in ktime_t format
362 ktime_t ktime_get_real(void)
366 getnstimeofday(&now);
368 return timespec_to_ktime(now);
370 EXPORT_SYMBOL_GPL(ktime_get_real);
373 * getrawmonotonic - Returns the raw monotonic time in a timespec
374 * @ts: pointer to the timespec to be set
376 * Returns the raw monotonic time (completely un-modified by ntp)
378 void getrawmonotonic(struct timespec *ts)
382 cycle_t cycle_now, cycle_delta;
385 seq = read_seqbegin(&xtime_lock);
387 /* read clocksource: */
388 cycle_now = clock->read(clock);
390 /* calculate the delta since the last update_wall_time: */
391 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
393 /* convert to nanoseconds: */
394 nsecs = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
396 *ts = clock->raw_time;
398 } while (read_seqretry(&xtime_lock, seq));
400 timespec_add_ns(ts, nsecs);
402 EXPORT_SYMBOL(getrawmonotonic);
406 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
408 int timekeeping_valid_for_hres(void)
414 seq = read_seqbegin(&xtime_lock);
416 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
418 } while (read_seqretry(&xtime_lock, seq));
424 * read_persistent_clock - Return time in seconds from the persistent clock.
426 * Weak dummy function for arches that do not yet support it.
427 * Returns seconds from epoch using the battery backed persistent clock.
428 * Returns zero if unsupported.
430 * XXX - Do be sure to remove it once all arches implement it.
432 unsigned long __attribute__((weak)) read_persistent_clock(void)
438 * timekeeping_init - Initializes the clocksource and common timekeeping values
440 void __init timekeeping_init(void)
443 unsigned long sec = read_persistent_clock();
445 write_seqlock_irqsave(&xtime_lock, flags);
449 clock = clocksource_get_next();
451 clock->enable(clock);
452 /* set mult_orig on enable */
453 clock->mult_orig = clock->mult;
454 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
455 clock->cycle_last = clock->read(clock);
459 set_normalized_timespec(&wall_to_monotonic,
460 -xtime.tv_sec, -xtime.tv_nsec);
461 update_xtime_cache(0);
462 total_sleep_time = 0;
463 write_sequnlock_irqrestore(&xtime_lock, flags);
466 /* time in seconds when suspend began */
467 static unsigned long timekeeping_suspend_time;
470 * timekeeping_resume - Resumes the generic timekeeping subsystem.
473 * This is for the generic clocksource timekeeping.
474 * xtime/wall_to_monotonic/jiffies/etc are
475 * still managed by arch specific suspend/resume code.
477 static int timekeeping_resume(struct sys_device *dev)
480 unsigned long now = read_persistent_clock();
482 clocksource_resume();
484 write_seqlock_irqsave(&xtime_lock, flags);
486 if (now && (now > timekeeping_suspend_time)) {
487 unsigned long sleep_length = now - timekeeping_suspend_time;
489 xtime.tv_sec += sleep_length;
490 wall_to_monotonic.tv_sec -= sleep_length;
491 total_sleep_time += sleep_length;
493 update_xtime_cache(0);
494 /* re-base the last cycle value */
495 clock->cycle_last = clock->read(clock);
497 timekeeping_suspended = 0;
498 write_sequnlock_irqrestore(&xtime_lock, flags);
500 touch_softlockup_watchdog();
502 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
504 /* Resume hrtimers */
505 hres_timers_resume();
510 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
514 timekeeping_suspend_time = read_persistent_clock();
516 write_seqlock_irqsave(&xtime_lock, flags);
517 clocksource_forward_now();
518 timekeeping_suspended = 1;
519 write_sequnlock_irqrestore(&xtime_lock, flags);
521 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
526 /* sysfs resume/suspend bits for timekeeping */
527 static struct sysdev_class timekeeping_sysclass = {
528 .name = "timekeeping",
529 .resume = timekeeping_resume,
530 .suspend = timekeeping_suspend,
533 static struct sys_device device_timer = {
535 .cls = &timekeeping_sysclass,
538 static int __init timekeeping_init_device(void)
540 int error = sysdev_class_register(&timekeeping_sysclass);
542 error = sysdev_register(&device_timer);
546 device_initcall(timekeeping_init_device);
549 * If the error is already larger, we look ahead even further
550 * to compensate for late or lost adjustments.
552 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
560 * Use the current error value to determine how much to look ahead.
561 * The larger the error the slower we adjust for it to avoid problems
562 * with losing too many ticks, otherwise we would overadjust and
563 * produce an even larger error. The smaller the adjustment the
564 * faster we try to adjust for it, as lost ticks can do less harm
565 * here. This is tuned so that an error of about 1 msec is adjusted
566 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
568 error2 = clock->error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
569 error2 = abs(error2);
570 for (look_ahead = 0; error2 > 0; look_ahead++)
574 * Now calculate the error in (1 << look_ahead) ticks, but first
575 * remove the single look ahead already included in the error.
577 tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift + 1);
578 tick_error -= clock->xtime_interval >> 1;
579 error = ((error - tick_error) >> look_ahead) + tick_error;
581 /* Finally calculate the adjustment shift value. */
586 *interval = -*interval;
590 for (adj = 0; error > i; adj++)
599 * Adjust the multiplier to reduce the error value,
600 * this is optimized for the most common adjustments of -1,0,1,
601 * for other values we can do a bit more work.
603 static void clocksource_adjust(s64 offset)
605 s64 error, interval = clock->cycle_interval;
608 error = clock->error >> (NTP_SCALE_SHIFT - clock->shift - 1);
609 if (error > interval) {
611 if (likely(error <= interval))
614 adj = clocksource_bigadjust(error, &interval, &offset);
615 } else if (error < -interval) {
617 if (likely(error >= -interval)) {
619 interval = -interval;
622 adj = clocksource_bigadjust(error, &interval, &offset);
627 clock->xtime_interval += interval;
628 clock->xtime_nsec -= offset;
629 clock->error -= (interval - offset) <<
630 (NTP_SCALE_SHIFT - clock->shift);
634 * update_wall_time - Uses the current clocksource to increment the wall time
636 * Called from the timer interrupt, must hold a write on xtime_lock.
638 void update_wall_time(void)
642 /* Make sure we're fully resumed: */
643 if (unlikely(timekeeping_suspended))
646 #ifdef CONFIG_GENERIC_TIME
647 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
649 offset = clock->cycle_interval;
651 clock->xtime_nsec = (s64)xtime.tv_nsec << clock->shift;
653 /* normally this loop will run just once, however in the
654 * case of lost or late ticks, it will accumulate correctly.
656 while (offset >= clock->cycle_interval) {
657 /* accumulate one interval */
658 offset -= clock->cycle_interval;
659 clock->cycle_last += clock->cycle_interval;
661 clock->xtime_nsec += clock->xtime_interval;
662 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
663 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
668 clock->raw_time.tv_nsec += clock->raw_interval;
669 if (clock->raw_time.tv_nsec >= NSEC_PER_SEC) {
670 clock->raw_time.tv_nsec -= NSEC_PER_SEC;
671 clock->raw_time.tv_sec++;
674 /* accumulate error between NTP and clock interval */
675 clock->error += tick_length;
676 clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift);
679 /* correct the clock when NTP error is too big */
680 clocksource_adjust(offset);
683 * Since in the loop above, we accumulate any amount of time
684 * in xtime_nsec over a second into xtime.tv_sec, its possible for
685 * xtime_nsec to be fairly small after the loop. Further, if we're
686 * slightly speeding the clocksource up in clocksource_adjust(),
687 * its possible the required corrective factor to xtime_nsec could
688 * cause it to underflow.
690 * Now, we cannot simply roll the accumulated second back, since
691 * the NTP subsystem has been notified via second_overflow. So
692 * instead we push xtime_nsec forward by the amount we underflowed,
693 * and add that amount into the error.
695 * We'll correct this error next time through this function, when
696 * xtime_nsec is not as small.
698 if (unlikely((s64)clock->xtime_nsec < 0)) {
699 s64 neg = -(s64)clock->xtime_nsec;
700 clock->xtime_nsec = 0;
701 clock->error += neg << (NTP_SCALE_SHIFT - clock->shift);
704 /* store full nanoseconds into xtime after rounding it up and
705 * add the remainder to the error difference.
707 xtime.tv_nsec = ((s64)clock->xtime_nsec >> clock->shift) + 1;
708 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
709 clock->error += clock->xtime_nsec << (NTP_SCALE_SHIFT - clock->shift);
711 update_xtime_cache(cyc2ns(clock, offset));
713 /* check to see if there is a new clocksource to use */
714 change_clocksource();
715 update_vsyscall(&xtime, clock);
719 * getboottime - Return the real time of system boot.
720 * @ts: pointer to the timespec to be set
722 * Returns the time of day in a timespec.
724 * This is based on the wall_to_monotonic offset and the total suspend
725 * time. Calls to settimeofday will affect the value returned (which
726 * basically means that however wrong your real time clock is at boot time,
727 * you get the right time here).
729 void getboottime(struct timespec *ts)
731 set_normalized_timespec(ts,
732 - (wall_to_monotonic.tv_sec + total_sleep_time),
733 - wall_to_monotonic.tv_nsec);
737 * monotonic_to_bootbased - Convert the monotonic time to boot based.
738 * @ts: pointer to the timespec to be converted
740 void monotonic_to_bootbased(struct timespec *ts)
742 ts->tv_sec += total_sleep_time;
745 unsigned long get_seconds(void)
747 return xtime_cache.tv_sec;
749 EXPORT_SYMBOL(get_seconds);
752 struct timespec current_kernel_time(void)
758 seq = read_seqbegin(&xtime_lock);
761 } while (read_seqretry(&xtime_lock, seq));
765 EXPORT_SYMBOL(current_kernel_time);