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;
62 #ifdef CONFIG_GENERIC_TIME
64 * clocksource_forward_now - update clock to the current time
66 * Forward the current clock to update its state since the last call to
67 * update_wall_time(). This is useful before significant clock changes,
68 * as it avoids having to deal with this time offset explicitly.
70 static void clocksource_forward_now(void)
72 cycle_t cycle_now, cycle_delta;
75 cycle_now = clocksource_read(clock);
76 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
77 clock->cycle_last = cycle_now;
79 nsec = cyc2ns(clock, cycle_delta);
81 /* If arch requires, add in gettimeoffset() */
82 nsec += arch_gettimeoffset();
84 timespec_add_ns(&xtime, nsec);
86 nsec = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
87 clock->raw_time.tv_nsec += nsec;
91 * getnstimeofday - Returns the time of day in a timespec
92 * @ts: pointer to the timespec to be set
94 * Returns the time of day in a timespec.
96 void getnstimeofday(struct timespec *ts)
98 cycle_t cycle_now, cycle_delta;
102 WARN_ON(timekeeping_suspended);
105 seq = read_seqbegin(&xtime_lock);
109 /* read clocksource: */
110 cycle_now = clocksource_read(clock);
112 /* calculate the delta since the last update_wall_time: */
113 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
115 /* convert to nanoseconds: */
116 nsecs = cyc2ns(clock, cycle_delta);
118 /* If arch requires, add in gettimeoffset() */
119 nsecs += arch_gettimeoffset();
121 } while (read_seqretry(&xtime_lock, seq));
123 timespec_add_ns(ts, nsecs);
126 EXPORT_SYMBOL(getnstimeofday);
128 ktime_t ktime_get(void)
130 cycle_t cycle_now, cycle_delta;
134 WARN_ON(timekeeping_suspended);
137 seq = read_seqbegin(&xtime_lock);
138 secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
139 nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
141 /* read clocksource: */
142 cycle_now = clocksource_read(clock);
144 /* calculate the delta since the last update_wall_time: */
145 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
147 /* convert to nanoseconds: */
148 nsecs += cyc2ns(clock, cycle_delta);
150 } while (read_seqretry(&xtime_lock, seq));
152 * Use ktime_set/ktime_add_ns to create a proper ktime on
153 * 32-bit architectures without CONFIG_KTIME_SCALAR.
155 return ktime_add_ns(ktime_set(secs, 0), nsecs);
157 EXPORT_SYMBOL_GPL(ktime_get);
160 * ktime_get_ts - get the monotonic clock in timespec format
161 * @ts: pointer to timespec variable
163 * The function calculates the monotonic clock from the realtime
164 * clock and the wall_to_monotonic offset and stores the result
165 * in normalized timespec format in the variable pointed to by @ts.
167 void ktime_get_ts(struct timespec *ts)
169 cycle_t cycle_now, cycle_delta;
170 struct timespec tomono;
174 WARN_ON(timekeeping_suspended);
177 seq = read_seqbegin(&xtime_lock);
179 tomono = wall_to_monotonic;
181 /* read clocksource: */
182 cycle_now = clocksource_read(clock);
184 /* calculate the delta since the last update_wall_time: */
185 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
187 /* convert to nanoseconds: */
188 nsecs = cyc2ns(clock, cycle_delta);
190 } while (read_seqretry(&xtime_lock, seq));
192 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
193 ts->tv_nsec + tomono.tv_nsec + nsecs);
195 EXPORT_SYMBOL_GPL(ktime_get_ts);
198 * do_gettimeofday - Returns the time of day in a timeval
199 * @tv: pointer to the timeval to be set
201 * NOTE: Users should be converted to using getnstimeofday()
203 void do_gettimeofday(struct timeval *tv)
207 getnstimeofday(&now);
208 tv->tv_sec = now.tv_sec;
209 tv->tv_usec = now.tv_nsec/1000;
212 EXPORT_SYMBOL(do_gettimeofday);
214 * do_settimeofday - Sets the time of day
215 * @tv: pointer to the timespec variable containing the new time
217 * Sets the time of day to the new time and update NTP and notify hrtimers
219 int do_settimeofday(struct timespec *tv)
221 struct timespec ts_delta;
224 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
227 write_seqlock_irqsave(&xtime_lock, flags);
229 clocksource_forward_now();
231 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
232 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
233 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
237 update_xtime_cache(0);
242 update_vsyscall(&xtime, clock);
244 write_sequnlock_irqrestore(&xtime_lock, flags);
246 /* signal hrtimers about time change */
252 EXPORT_SYMBOL(do_settimeofday);
255 * change_clocksource - Swaps clocksources if a new one is available
257 * Accumulates current time interval and initializes new clocksource
259 static void change_clocksource(void)
261 struct clocksource *new, *old;
263 new = clocksource_get_next();
268 clocksource_forward_now();
270 if (clocksource_enable(new))
273 new->raw_time = clock->raw_time;
276 clocksource_disable(old);
278 clock->cycle_last = 0;
279 clock->cycle_last = clocksource_read(clock);
281 clock->xtime_nsec = 0;
282 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
287 * We're holding xtime lock and waking up klogd would deadlock
288 * us on enqueue. So no printing!
289 printk(KERN_INFO "Time: %s clocksource has been installed.\n",
293 #else /* GENERIC_TIME */
294 static inline void clocksource_forward_now(void) { }
295 static inline void change_clocksource(void) { }
298 * ktime_get - get the monotonic time in ktime_t format
300 * returns the time in ktime_t format
302 ktime_t ktime_get(void)
308 return timespec_to_ktime(now);
310 EXPORT_SYMBOL_GPL(ktime_get);
313 * ktime_get_ts - get the monotonic clock in timespec format
314 * @ts: pointer to timespec variable
316 * The function calculates the monotonic clock from the realtime
317 * clock and the wall_to_monotonic offset and stores the result
318 * in normalized timespec format in the variable pointed to by @ts.
320 void ktime_get_ts(struct timespec *ts)
322 struct timespec tomono;
326 seq = read_seqbegin(&xtime_lock);
328 tomono = wall_to_monotonic;
330 } while (read_seqretry(&xtime_lock, seq));
332 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
333 ts->tv_nsec + tomono.tv_nsec);
335 EXPORT_SYMBOL_GPL(ktime_get_ts);
336 #endif /* !GENERIC_TIME */
339 * ktime_get_real - get the real (wall-) time in ktime_t format
341 * returns the time in ktime_t format
343 ktime_t ktime_get_real(void)
347 getnstimeofday(&now);
349 return timespec_to_ktime(now);
351 EXPORT_SYMBOL_GPL(ktime_get_real);
354 * getrawmonotonic - Returns the raw monotonic time in a timespec
355 * @ts: pointer to the timespec to be set
357 * Returns the raw monotonic time (completely un-modified by ntp)
359 void getrawmonotonic(struct timespec *ts)
363 cycle_t cycle_now, cycle_delta;
366 seq = read_seqbegin(&xtime_lock);
368 /* read clocksource: */
369 cycle_now = clocksource_read(clock);
371 /* calculate the delta since the last update_wall_time: */
372 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
374 /* convert to nanoseconds: */
375 nsecs = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
377 *ts = clock->raw_time;
379 } while (read_seqretry(&xtime_lock, seq));
381 timespec_add_ns(ts, nsecs);
383 EXPORT_SYMBOL(getrawmonotonic);
387 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
389 int timekeeping_valid_for_hres(void)
395 seq = read_seqbegin(&xtime_lock);
397 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
399 } while (read_seqretry(&xtime_lock, seq));
405 * read_persistent_clock - Return time in seconds from the persistent clock.
407 * Weak dummy function for arches that do not yet support it.
408 * Returns seconds from epoch using the battery backed persistent clock.
409 * Returns zero if unsupported.
411 * XXX - Do be sure to remove it once all arches implement it.
413 unsigned long __attribute__((weak)) read_persistent_clock(void)
419 * timekeeping_init - Initializes the clocksource and common timekeeping values
421 void __init timekeeping_init(void)
424 unsigned long sec = read_persistent_clock();
426 write_seqlock_irqsave(&xtime_lock, flags);
430 clock = clocksource_get_next();
431 clocksource_enable(clock);
432 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
433 clock->cycle_last = clocksource_read(clock);
437 set_normalized_timespec(&wall_to_monotonic,
438 -xtime.tv_sec, -xtime.tv_nsec);
439 update_xtime_cache(0);
440 total_sleep_time = 0;
441 write_sequnlock_irqrestore(&xtime_lock, flags);
444 /* time in seconds when suspend began */
445 static unsigned long timekeeping_suspend_time;
448 * timekeeping_resume - Resumes the generic timekeeping subsystem.
451 * This is for the generic clocksource timekeeping.
452 * xtime/wall_to_monotonic/jiffies/etc are
453 * still managed by arch specific suspend/resume code.
455 static int timekeeping_resume(struct sys_device *dev)
458 unsigned long now = read_persistent_clock();
460 clocksource_resume();
462 write_seqlock_irqsave(&xtime_lock, flags);
464 if (now && (now > timekeeping_suspend_time)) {
465 unsigned long sleep_length = now - timekeeping_suspend_time;
467 xtime.tv_sec += sleep_length;
468 wall_to_monotonic.tv_sec -= sleep_length;
469 total_sleep_time += sleep_length;
471 update_xtime_cache(0);
472 /* re-base the last cycle value */
473 clock->cycle_last = 0;
474 clock->cycle_last = clocksource_read(clock);
476 timekeeping_suspended = 0;
477 write_sequnlock_irqrestore(&xtime_lock, flags);
479 touch_softlockup_watchdog();
481 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
483 /* Resume hrtimers */
484 hres_timers_resume();
489 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
493 timekeeping_suspend_time = read_persistent_clock();
495 write_seqlock_irqsave(&xtime_lock, flags);
496 clocksource_forward_now();
497 timekeeping_suspended = 1;
498 write_sequnlock_irqrestore(&xtime_lock, flags);
500 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
505 /* sysfs resume/suspend bits for timekeeping */
506 static struct sysdev_class timekeeping_sysclass = {
507 .name = "timekeeping",
508 .resume = timekeeping_resume,
509 .suspend = timekeeping_suspend,
512 static struct sys_device device_timer = {
514 .cls = &timekeeping_sysclass,
517 static int __init timekeeping_init_device(void)
519 int error = sysdev_class_register(&timekeeping_sysclass);
521 error = sysdev_register(&device_timer);
525 device_initcall(timekeeping_init_device);
528 * If the error is already larger, we look ahead even further
529 * to compensate for late or lost adjustments.
531 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
539 * Use the current error value to determine how much to look ahead.
540 * The larger the error the slower we adjust for it to avoid problems
541 * with losing too many ticks, otherwise we would overadjust and
542 * produce an even larger error. The smaller the adjustment the
543 * faster we try to adjust for it, as lost ticks can do less harm
544 * here. This is tuned so that an error of about 1 msec is adjusted
545 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
547 error2 = clock->error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
548 error2 = abs(error2);
549 for (look_ahead = 0; error2 > 0; look_ahead++)
553 * Now calculate the error in (1 << look_ahead) ticks, but first
554 * remove the single look ahead already included in the error.
556 tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift + 1);
557 tick_error -= clock->xtime_interval >> 1;
558 error = ((error - tick_error) >> look_ahead) + tick_error;
560 /* Finally calculate the adjustment shift value. */
565 *interval = -*interval;
569 for (adj = 0; error > i; adj++)
578 * Adjust the multiplier to reduce the error value,
579 * this is optimized for the most common adjustments of -1,0,1,
580 * for other values we can do a bit more work.
582 static void clocksource_adjust(s64 offset)
584 s64 error, interval = clock->cycle_interval;
587 error = clock->error >> (NTP_SCALE_SHIFT - clock->shift - 1);
588 if (error > interval) {
590 if (likely(error <= interval))
593 adj = clocksource_bigadjust(error, &interval, &offset);
594 } else if (error < -interval) {
596 if (likely(error >= -interval)) {
598 interval = -interval;
601 adj = clocksource_bigadjust(error, &interval, &offset);
606 clock->xtime_interval += interval;
607 clock->xtime_nsec -= offset;
608 clock->error -= (interval - offset) <<
609 (NTP_SCALE_SHIFT - clock->shift);
613 * update_wall_time - Uses the current clocksource to increment the wall time
615 * Called from the timer interrupt, must hold a write on xtime_lock.
617 void update_wall_time(void)
621 /* Make sure we're fully resumed: */
622 if (unlikely(timekeeping_suspended))
625 #ifdef CONFIG_GENERIC_TIME
626 offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask;
628 offset = clock->cycle_interval;
630 clock->xtime_nsec = (s64)xtime.tv_nsec << clock->shift;
632 /* normally this loop will run just once, however in the
633 * case of lost or late ticks, it will accumulate correctly.
635 while (offset >= clock->cycle_interval) {
636 /* accumulate one interval */
637 offset -= clock->cycle_interval;
638 clock->cycle_last += clock->cycle_interval;
640 clock->xtime_nsec += clock->xtime_interval;
641 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
642 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
647 clock->raw_time.tv_nsec += clock->raw_interval;
648 if (clock->raw_time.tv_nsec >= NSEC_PER_SEC) {
649 clock->raw_time.tv_nsec -= NSEC_PER_SEC;
650 clock->raw_time.tv_sec++;
653 /* accumulate error between NTP and clock interval */
654 clock->error += tick_length;
655 clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift);
658 /* correct the clock when NTP error is too big */
659 clocksource_adjust(offset);
662 * Since in the loop above, we accumulate any amount of time
663 * in xtime_nsec over a second into xtime.tv_sec, its possible for
664 * xtime_nsec to be fairly small after the loop. Further, if we're
665 * slightly speeding the clocksource up in clocksource_adjust(),
666 * its possible the required corrective factor to xtime_nsec could
667 * cause it to underflow.
669 * Now, we cannot simply roll the accumulated second back, since
670 * the NTP subsystem has been notified via second_overflow. So
671 * instead we push xtime_nsec forward by the amount we underflowed,
672 * and add that amount into the error.
674 * We'll correct this error next time through this function, when
675 * xtime_nsec is not as small.
677 if (unlikely((s64)clock->xtime_nsec < 0)) {
678 s64 neg = -(s64)clock->xtime_nsec;
679 clock->xtime_nsec = 0;
680 clock->error += neg << (NTP_SCALE_SHIFT - clock->shift);
683 /* store full nanoseconds into xtime after rounding it up and
684 * add the remainder to the error difference.
686 xtime.tv_nsec = ((s64)clock->xtime_nsec >> clock->shift) + 1;
687 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
688 clock->error += clock->xtime_nsec << (NTP_SCALE_SHIFT - clock->shift);
690 update_xtime_cache(cyc2ns(clock, offset));
692 /* check to see if there is a new clocksource to use */
693 change_clocksource();
694 update_vsyscall(&xtime, clock);
698 * getboottime - Return the real time of system boot.
699 * @ts: pointer to the timespec to be set
701 * Returns the time of day in a timespec.
703 * This is based on the wall_to_monotonic offset and the total suspend
704 * time. Calls to settimeofday will affect the value returned (which
705 * basically means that however wrong your real time clock is at boot time,
706 * you get the right time here).
708 void getboottime(struct timespec *ts)
710 set_normalized_timespec(ts,
711 - (wall_to_monotonic.tv_sec + total_sleep_time),
712 - wall_to_monotonic.tv_nsec);
716 * monotonic_to_bootbased - Convert the monotonic time to boot based.
717 * @ts: pointer to the timespec to be converted
719 void monotonic_to_bootbased(struct timespec *ts)
721 ts->tv_sec += total_sleep_time;
724 unsigned long get_seconds(void)
726 return xtime_cache.tv_sec;
728 EXPORT_SYMBOL(get_seconds);
731 struct timespec current_kernel_time(void)
737 seq = read_seqbegin(&xtime_lock);
740 } while (read_seqretry(&xtime_lock, seq));
744 EXPORT_SYMBOL(current_kernel_time);