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
25 * playing with xtime and avenrun.
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);
80 timespec_add_ns(&xtime, nsec);
82 nsec = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
83 clock->raw_time.tv_nsec += nsec;
87 * getnstimeofday - Returns the time of day in a timespec
88 * @ts: pointer to the timespec to be set
90 * Returns the time of day in a timespec.
92 void getnstimeofday(struct timespec *ts)
94 cycle_t cycle_now, cycle_delta;
98 WARN_ON(timekeeping_suspended);
101 seq = read_seqbegin(&xtime_lock);
105 /* read clocksource: */
106 cycle_now = clocksource_read(clock);
108 /* calculate the delta since the last update_wall_time: */
109 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
111 /* convert to nanoseconds: */
112 nsecs = cyc2ns(clock, cycle_delta);
114 } while (read_seqretry(&xtime_lock, seq));
116 timespec_add_ns(ts, nsecs);
119 EXPORT_SYMBOL(getnstimeofday);
122 * do_gettimeofday - Returns the time of day in a timeval
123 * @tv: pointer to the timeval to be set
125 * NOTE: Users should be converted to using getnstimeofday()
127 void do_gettimeofday(struct timeval *tv)
131 getnstimeofday(&now);
132 tv->tv_sec = now.tv_sec;
133 tv->tv_usec = now.tv_nsec/1000;
136 EXPORT_SYMBOL(do_gettimeofday);
138 * do_settimeofday - Sets the time of day
139 * @tv: pointer to the timespec variable containing the new time
141 * Sets the time of day to the new time and update NTP and notify hrtimers
143 int do_settimeofday(struct timespec *tv)
145 struct timespec ts_delta;
148 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
151 write_seqlock_irqsave(&xtime_lock, flags);
153 clocksource_forward_now();
155 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
156 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
157 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
161 update_xtime_cache(0);
166 update_vsyscall(&xtime, clock);
168 write_sequnlock_irqrestore(&xtime_lock, flags);
170 /* signal hrtimers about time change */
176 EXPORT_SYMBOL(do_settimeofday);
179 * change_clocksource - Swaps clocksources if a new one is available
181 * Accumulates current time interval and initializes new clocksource
183 static void change_clocksource(void)
185 struct clocksource *new, *old;
187 new = clocksource_get_next();
192 clocksource_forward_now();
194 if (clocksource_enable(new))
197 new->raw_time = clock->raw_time;
200 clocksource_disable(old);
202 clock->cycle_last = 0;
203 clock->cycle_last = clocksource_read(clock);
205 clock->xtime_nsec = 0;
206 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
211 * We're holding xtime lock and waking up klogd would deadlock
212 * us on enqueue. So no printing!
213 printk(KERN_INFO "Time: %s clocksource has been installed.\n",
218 static inline void clocksource_forward_now(void) { }
219 static inline void change_clocksource(void) { }
223 * getrawmonotonic - Returns the raw monotonic time in a timespec
224 * @ts: pointer to the timespec to be set
226 * Returns the raw monotonic time (completely un-modified by ntp)
228 void getrawmonotonic(struct timespec *ts)
232 cycle_t cycle_now, cycle_delta;
235 seq = read_seqbegin(&xtime_lock);
237 /* read clocksource: */
238 cycle_now = clocksource_read(clock);
240 /* calculate the delta since the last update_wall_time: */
241 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
243 /* convert to nanoseconds: */
244 nsecs = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
246 *ts = clock->raw_time;
248 } while (read_seqretry(&xtime_lock, seq));
250 timespec_add_ns(ts, nsecs);
252 EXPORT_SYMBOL(getrawmonotonic);
256 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
258 int timekeeping_valid_for_hres(void)
264 seq = read_seqbegin(&xtime_lock);
266 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
268 } while (read_seqretry(&xtime_lock, seq));
274 * read_persistent_clock - Return time in seconds from the persistent clock.
276 * Weak dummy function for arches that do not yet support it.
277 * Returns seconds from epoch using the battery backed persistent clock.
278 * Returns zero if unsupported.
280 * XXX - Do be sure to remove it once all arches implement it.
282 unsigned long __attribute__((weak)) read_persistent_clock(void)
288 * timekeeping_init - Initializes the clocksource and common timekeeping values
290 void __init timekeeping_init(void)
293 unsigned long sec = read_persistent_clock();
295 write_seqlock_irqsave(&xtime_lock, flags);
299 clock = clocksource_get_next();
300 clocksource_enable(clock);
301 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
302 clock->cycle_last = clocksource_read(clock);
306 set_normalized_timespec(&wall_to_monotonic,
307 -xtime.tv_sec, -xtime.tv_nsec);
308 update_xtime_cache(0);
309 total_sleep_time = 0;
310 write_sequnlock_irqrestore(&xtime_lock, flags);
313 /* time in seconds when suspend began */
314 static unsigned long timekeeping_suspend_time;
317 * timekeeping_resume - Resumes the generic timekeeping subsystem.
320 * This is for the generic clocksource timekeeping.
321 * xtime/wall_to_monotonic/jiffies/etc are
322 * still managed by arch specific suspend/resume code.
324 static int timekeeping_resume(struct sys_device *dev)
327 unsigned long now = read_persistent_clock();
329 clocksource_resume();
331 write_seqlock_irqsave(&xtime_lock, flags);
333 if (now && (now > timekeeping_suspend_time)) {
334 unsigned long sleep_length = now - timekeeping_suspend_time;
336 xtime.tv_sec += sleep_length;
337 wall_to_monotonic.tv_sec -= sleep_length;
338 total_sleep_time += sleep_length;
340 update_xtime_cache(0);
341 /* re-base the last cycle value */
342 clock->cycle_last = 0;
343 clock->cycle_last = clocksource_read(clock);
345 timekeeping_suspended = 0;
346 write_sequnlock_irqrestore(&xtime_lock, flags);
348 touch_softlockup_watchdog();
350 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
352 /* Resume hrtimers */
353 hres_timers_resume();
358 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
362 timekeeping_suspend_time = read_persistent_clock();
364 write_seqlock_irqsave(&xtime_lock, flags);
365 clocksource_forward_now();
366 timekeeping_suspended = 1;
367 write_sequnlock_irqrestore(&xtime_lock, flags);
369 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
374 /* sysfs resume/suspend bits for timekeeping */
375 static struct sysdev_class timekeeping_sysclass = {
376 .name = "timekeeping",
377 .resume = timekeeping_resume,
378 .suspend = timekeeping_suspend,
381 static struct sys_device device_timer = {
383 .cls = &timekeeping_sysclass,
386 static int __init timekeeping_init_device(void)
388 int error = sysdev_class_register(&timekeeping_sysclass);
390 error = sysdev_register(&device_timer);
394 device_initcall(timekeeping_init_device);
397 * If the error is already larger, we look ahead even further
398 * to compensate for late or lost adjustments.
400 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
408 * Use the current error value to determine how much to look ahead.
409 * The larger the error the slower we adjust for it to avoid problems
410 * with losing too many ticks, otherwise we would overadjust and
411 * produce an even larger error. The smaller the adjustment the
412 * faster we try to adjust for it, as lost ticks can do less harm
413 * here. This is tuned so that an error of about 1 msec is adjusted
414 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
416 error2 = clock->error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
417 error2 = abs(error2);
418 for (look_ahead = 0; error2 > 0; look_ahead++)
422 * Now calculate the error in (1 << look_ahead) ticks, but first
423 * remove the single look ahead already included in the error.
425 tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift + 1);
426 tick_error -= clock->xtime_interval >> 1;
427 error = ((error - tick_error) >> look_ahead) + tick_error;
429 /* Finally calculate the adjustment shift value. */
434 *interval = -*interval;
438 for (adj = 0; error > i; adj++)
447 * Adjust the multiplier to reduce the error value,
448 * this is optimized for the most common adjustments of -1,0,1,
449 * for other values we can do a bit more work.
451 static void clocksource_adjust(s64 offset)
453 s64 error, interval = clock->cycle_interval;
456 error = clock->error >> (NTP_SCALE_SHIFT - clock->shift - 1);
457 if (error > interval) {
459 if (likely(error <= interval))
462 adj = clocksource_bigadjust(error, &interval, &offset);
463 } else if (error < -interval) {
465 if (likely(error >= -interval)) {
467 interval = -interval;
470 adj = clocksource_bigadjust(error, &interval, &offset);
475 clock->xtime_interval += interval;
476 clock->xtime_nsec -= offset;
477 clock->error -= (interval - offset) <<
478 (NTP_SCALE_SHIFT - clock->shift);
482 * update_wall_time - Uses the current clocksource to increment the wall time
484 * Called from the timer interrupt, must hold a write on xtime_lock.
486 void update_wall_time(void)
490 /* Make sure we're fully resumed: */
491 if (unlikely(timekeeping_suspended))
494 #ifdef CONFIG_GENERIC_TIME
495 offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask;
497 offset = clock->cycle_interval;
499 clock->xtime_nsec = (s64)xtime.tv_nsec << clock->shift;
501 /* normally this loop will run just once, however in the
502 * case of lost or late ticks, it will accumulate correctly.
504 while (offset >= clock->cycle_interval) {
505 /* accumulate one interval */
506 offset -= clock->cycle_interval;
507 clock->cycle_last += clock->cycle_interval;
509 clock->xtime_nsec += clock->xtime_interval;
510 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
511 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
516 clock->raw_time.tv_nsec += clock->raw_interval;
517 if (clock->raw_time.tv_nsec >= NSEC_PER_SEC) {
518 clock->raw_time.tv_nsec -= NSEC_PER_SEC;
519 clock->raw_time.tv_sec++;
522 /* accumulate error between NTP and clock interval */
523 clock->error += tick_length;
524 clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift);
527 /* correct the clock when NTP error is too big */
528 clocksource_adjust(offset);
531 * Since in the loop above, we accumulate any amount of time
532 * in xtime_nsec over a second into xtime.tv_sec, its possible for
533 * xtime_nsec to be fairly small after the loop. Further, if we're
534 * slightly speeding the clocksource up in clocksource_adjust(),
535 * its possible the required corrective factor to xtime_nsec could
536 * cause it to underflow.
538 * Now, we cannot simply roll the accumulated second back, since
539 * the NTP subsystem has been notified via second_overflow. So
540 * instead we push xtime_nsec forward by the amount we underflowed,
541 * and add that amount into the error.
543 * We'll correct this error next time through this function, when
544 * xtime_nsec is not as small.
546 if (unlikely((s64)clock->xtime_nsec < 0)) {
547 s64 neg = -(s64)clock->xtime_nsec;
548 clock->xtime_nsec = 0;
549 clock->error += neg << (NTP_SCALE_SHIFT - clock->shift);
552 /* store full nanoseconds into xtime after rounding it up and
553 * add the remainder to the error difference.
555 xtime.tv_nsec = ((s64)clock->xtime_nsec >> clock->shift) + 1;
556 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
557 clock->error += clock->xtime_nsec << (NTP_SCALE_SHIFT - clock->shift);
559 update_xtime_cache(cyc2ns(clock, offset));
561 /* check to see if there is a new clocksource to use */
562 change_clocksource();
563 update_vsyscall(&xtime, clock);
567 * getboottime - Return the real time of system boot.
568 * @ts: pointer to the timespec to be set
570 * Returns the time of day in a timespec.
572 * This is based on the wall_to_monotonic offset and the total suspend
573 * time. Calls to settimeofday will affect the value returned (which
574 * basically means that however wrong your real time clock is at boot time,
575 * you get the right time here).
577 void getboottime(struct timespec *ts)
579 set_normalized_timespec(ts,
580 - (wall_to_monotonic.tv_sec + total_sleep_time),
581 - wall_to_monotonic.tv_nsec);
585 * monotonic_to_bootbased - Convert the monotonic time to boot based.
586 * @ts: pointer to the timespec to be converted
588 void monotonic_to_bootbased(struct timespec *ts)
590 ts->tv_sec += total_sleep_time;
593 unsigned long get_seconds(void)
595 return xtime_cache.tv_sec;
597 EXPORT_SYMBOL(get_seconds);
600 struct timespec current_kernel_time(void)
606 seq = read_seqbegin(&xtime_lock);
609 } while (read_seqretry(&xtime_lock, seq));
613 EXPORT_SYMBOL(current_kernel_time);