projects/linux·blob v3.4 · refs

1278 lines · 33.8 KiB

1/*
2 * linux/kernel/time/timekeeping.c
3 *
4 * Kernel timekeeping code and accessor functions
5 *
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
8 *
9 */
10
11#include <linux/module.h>
12#include <linux/interrupt.h>
13#include <linux/percpu.h>
14#include <linux/init.h>
15#include <linux/mm.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
24/* Structure holding internal timekeeping values. */
25struct timekeeper {
26 /* Current clocksource used for timekeeping. */
27 struct clocksource *clock;
28 /* NTP adjusted clock multiplier */
29 u32 mult;
30 /* The shift value of the current clocksource. */
31 int shift;
32
33 /* Number of clock cycles in one NTP interval. */
34 cycle_t cycle_interval;
35 /* Number of clock shifted nano seconds in one NTP interval. */
36 u64 xtime_interval;
37 /* shifted nano seconds left over when rounding cycle_interval */
38 s64 xtime_remainder;
39 /* Raw nano seconds accumulated per NTP interval. */
40 u32 raw_interval;
41
42 /* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
43 u64 xtime_nsec;
44 /* Difference between accumulated time and NTP time in ntp
45 * shifted nano seconds. */
46 s64 ntp_error;
47 /* Shift conversion between clock shifted nano seconds and
48 * ntp shifted nano seconds. */
49 int ntp_error_shift;
50
51 /* The current time */
52 struct timespec xtime;
53 /*
54 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
55 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
56 * at zero at system boot time, so wall_to_monotonic will be negative,
57 * however, we will ALWAYS keep the tv_nsec part positive so we can use
58 * the usual normalization.
59 *
60 * wall_to_monotonic is moved after resume from suspend for the
61 * monotonic time not to jump. We need to add total_sleep_time to
62 * wall_to_monotonic to get the real boot based time offset.
63 *
64 * - wall_to_monotonic is no longer the boot time, getboottime must be
65 * used instead.
66 */
67 struct timespec wall_to_monotonic;
68 /* time spent in suspend */
69 struct timespec total_sleep_time;
70 /* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock. */
71 struct timespec raw_time;
72
73 /* Seqlock for all timekeeper values */
74 seqlock_t lock;
75};
76
77static struct timekeeper timekeeper;
78
79/*
80 * This read-write spinlock protects us from races in SMP while
81 * playing with xtime.
82 */
83__cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
84
85
86/* flag for if timekeeping is suspended */
87int __read_mostly timekeeping_suspended;
88
89
90
91/**
92 * timekeeper_setup_internals - Set up internals to use clocksource clock.
93 *
94 * @clock: Pointer to clocksource.
95 *
96 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
97 * pair and interval request.
98 *
99 * Unless you're the timekeeping code, you should not be using this!
100 */
101static void timekeeper_setup_internals(struct clocksource *clock)
102{
103 cycle_t interval;
104 u64 tmp, ntpinterval;
105
106 timekeeper.clock = clock;
107 clock->cycle_last = clock->read(clock);
108
109 /* Do the ns -> cycle conversion first, using original mult */
110 tmp = NTP_INTERVAL_LENGTH;
111 tmp <<= clock->shift;
112 ntpinterval = tmp;
113 tmp += clock->mult/2;
114 do_div(tmp, clock->mult);
115 if (tmp == 0)
116 tmp = 1;
117
118 interval = (cycle_t) tmp;
119 timekeeper.cycle_interval = interval;
120
121 /* Go back from cycles -> shifted ns */
122 timekeeper.xtime_interval = (u64) interval * clock->mult;
123 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
124 timekeeper.raw_interval =
125 ((u64) interval * clock->mult) >> clock->shift;
126
127 timekeeper.xtime_nsec = 0;
128 timekeeper.shift = clock->shift;
129
130 timekeeper.ntp_error = 0;
131 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
132
133 /*
134 * The timekeeper keeps its own mult values for the currently
135 * active clocksource. These value will be adjusted via NTP
136 * to counteract clock drifting.
137 */
138 timekeeper.mult = clock->mult;
139}
140
141/* Timekeeper helper functions. */
142static inline s64 timekeeping_get_ns(void)
143{
144 cycle_t cycle_now, cycle_delta;
145 struct clocksource *clock;
146
147 /* read clocksource: */
148 clock = timekeeper.clock;
149 cycle_now = clock->read(clock);
150
151 /* calculate the delta since the last update_wall_time: */
152 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
153
154 /* return delta convert to nanoseconds using ntp adjusted mult. */
155 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
156 timekeeper.shift);
157}
158
159static inline s64 timekeeping_get_ns_raw(void)
160{
161 cycle_t cycle_now, cycle_delta;
162 struct clocksource *clock;
163
164 /* read clocksource: */
165 clock = timekeeper.clock;
166 cycle_now = clock->read(clock);
167
168 /* calculate the delta since the last update_wall_time: */
169 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
170
171 /* return delta convert to nanoseconds. */
172 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
173}
174
175/* must hold write on timekeeper.lock */
176static void timekeeping_update(bool clearntp)
177{
178 if (clearntp) {
179 timekeeper.ntp_error = 0;
180 ntp_clear();
181 }
182 update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
183 timekeeper.clock, timekeeper.mult);
184}
185
186
187/**
188 * timekeeping_forward_now - update clock to the current time
189 *
190 * Forward the current clock to update its state since the last call to
191 * update_wall_time(). This is useful before significant clock changes,
192 * as it avoids having to deal with this time offset explicitly.
193 */
194static void timekeeping_forward_now(void)
195{
196 cycle_t cycle_now, cycle_delta;
197 struct clocksource *clock;
198 s64 nsec;
199
200 clock = timekeeper.clock;
201 cycle_now = clock->read(clock);
202 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
203 clock->cycle_last = cycle_now;
204
205 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
206 timekeeper.shift);
207
208 /* If arch requires, add in gettimeoffset() */
209 nsec += arch_gettimeoffset();
210
211 timespec_add_ns(&timekeeper.xtime, nsec);
212
213 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
214 timespec_add_ns(&timekeeper.raw_time, nsec);
215}
216
217/**
218 * getnstimeofday - Returns the time of day in a timespec
219 * @ts: pointer to the timespec to be set
220 *
221 * Returns the time of day in a timespec.
222 */
223void getnstimeofday(struct timespec *ts)
224{
225 unsigned long seq;
226 s64 nsecs;
227
228 WARN_ON(timekeeping_suspended);
229
230 do {
231 seq = read_seqbegin(&timekeeper.lock);
232
233 *ts = timekeeper.xtime;
234 nsecs = timekeeping_get_ns();
235
236 /* If arch requires, add in gettimeoffset() */
237 nsecs += arch_gettimeoffset();
238
239 } while (read_seqretry(&timekeeper.lock, seq));
240
241 timespec_add_ns(ts, nsecs);
242}
243
244EXPORT_SYMBOL(getnstimeofday);
245
246ktime_t ktime_get(void)
247{
248 unsigned int seq;
249 s64 secs, nsecs;
250
251 WARN_ON(timekeeping_suspended);
252
253 do {
254 seq = read_seqbegin(&timekeeper.lock);
255 secs = timekeeper.xtime.tv_sec +
256 timekeeper.wall_to_monotonic.tv_sec;
257 nsecs = timekeeper.xtime.tv_nsec +
258 timekeeper.wall_to_monotonic.tv_nsec;
259 nsecs += timekeeping_get_ns();
260 /* If arch requires, add in gettimeoffset() */
261 nsecs += arch_gettimeoffset();
262
263 } while (read_seqretry(&timekeeper.lock, seq));
264 /*
265 * Use ktime_set/ktime_add_ns to create a proper ktime on
266 * 32-bit architectures without CONFIG_KTIME_SCALAR.
267 */
268 return ktime_add_ns(ktime_set(secs, 0), nsecs);
269}
270EXPORT_SYMBOL_GPL(ktime_get);
271
272/**
273 * ktime_get_ts - get the monotonic clock in timespec format
274 * @ts: pointer to timespec variable
275 *
276 * The function calculates the monotonic clock from the realtime
277 * clock and the wall_to_monotonic offset and stores the result
278 * in normalized timespec format in the variable pointed to by @ts.
279 */
280void ktime_get_ts(struct timespec *ts)
281{
282 struct timespec tomono;
283 unsigned int seq;
284 s64 nsecs;
285
286 WARN_ON(timekeeping_suspended);
287
288 do {
289 seq = read_seqbegin(&timekeeper.lock);
290 *ts = timekeeper.xtime;
291 tomono = timekeeper.wall_to_monotonic;
292 nsecs = timekeeping_get_ns();
293 /* If arch requires, add in gettimeoffset() */
294 nsecs += arch_gettimeoffset();
295
296 } while (read_seqretry(&timekeeper.lock, seq));
297
298 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
299 ts->tv_nsec + tomono.tv_nsec + nsecs);
300}
301EXPORT_SYMBOL_GPL(ktime_get_ts);
302
303#ifdef CONFIG_NTP_PPS
304
305/**
306 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
307 * @ts_raw: pointer to the timespec to be set to raw monotonic time
308 * @ts_real: pointer to the timespec to be set to the time of day
309 *
310 * This function reads both the time of day and raw monotonic time at the
311 * same time atomically and stores the resulting timestamps in timespec
312 * format.
313 */
314void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
315{
316 unsigned long seq;
317 s64 nsecs_raw, nsecs_real;
318
319 WARN_ON_ONCE(timekeeping_suspended);
320
321 do {
322 u32 arch_offset;
323
324 seq = read_seqbegin(&timekeeper.lock);
325
326 *ts_raw = timekeeper.raw_time;
327 *ts_real = timekeeper.xtime;
328
329 nsecs_raw = timekeeping_get_ns_raw();
330 nsecs_real = timekeeping_get_ns();
331
332 /* If arch requires, add in gettimeoffset() */
333 arch_offset = arch_gettimeoffset();
334 nsecs_raw += arch_offset;
335 nsecs_real += arch_offset;
336
337 } while (read_seqretry(&timekeeper.lock, seq));
338
339 timespec_add_ns(ts_raw, nsecs_raw);
340 timespec_add_ns(ts_real, nsecs_real);
341}
342EXPORT_SYMBOL(getnstime_raw_and_real);
343
344#endif /* CONFIG_NTP_PPS */
345
346/**
347 * do_gettimeofday - Returns the time of day in a timeval
348 * @tv: pointer to the timeval to be set
349 *
350 * NOTE: Users should be converted to using getnstimeofday()
351 */
352void do_gettimeofday(struct timeval *tv)
353{
354 struct timespec now;
355
356 getnstimeofday(&now);
357 tv->tv_sec = now.tv_sec;
358 tv->tv_usec = now.tv_nsec/1000;
359}
360
361EXPORT_SYMBOL(do_gettimeofday);
362/**
363 * do_settimeofday - Sets the time of day
364 * @tv: pointer to the timespec variable containing the new time
365 *
366 * Sets the time of day to the new time and update NTP and notify hrtimers
367 */
368int do_settimeofday(const struct timespec *tv)
369{
370 struct timespec ts_delta;
371 unsigned long flags;
372
373 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
374 return -EINVAL;
375
376 write_seqlock_irqsave(&timekeeper.lock, flags);
377
378 timekeeping_forward_now();
379
380 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
381 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
382 timekeeper.wall_to_monotonic =
383 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
384
385 timekeeper.xtime = *tv;
386 timekeeping_update(true);
387
388 write_sequnlock_irqrestore(&timekeeper.lock, flags);
389
390 /* signal hrtimers about time change */
391 clock_was_set();
392
393 return 0;
394}
395
396EXPORT_SYMBOL(do_settimeofday);
397
398
399/**
400 * timekeeping_inject_offset - Adds or subtracts from the current time.
401 * @tv: pointer to the timespec variable containing the offset
402 *
403 * Adds or subtracts an offset value from the current time.
404 */
405int timekeeping_inject_offset(struct timespec *ts)
406{
407 unsigned long flags;
408
409 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
410 return -EINVAL;
411
412 write_seqlock_irqsave(&timekeeper.lock, flags);
413
414 timekeeping_forward_now();
415
416 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
417 timekeeper.wall_to_monotonic =
418 timespec_sub(timekeeper.wall_to_monotonic, *ts);
419
420 timekeeping_update(true);
421
422 write_sequnlock_irqrestore(&timekeeper.lock, flags);
423
424 /* signal hrtimers about time change */
425 clock_was_set();
426
427 return 0;
428}
429EXPORT_SYMBOL(timekeeping_inject_offset);
430
431/**
432 * change_clocksource - Swaps clocksources if a new one is available
433 *
434 * Accumulates current time interval and initializes new clocksource
435 */
436static int change_clocksource(void *data)
437{
438 struct clocksource *new, *old;
439 unsigned long flags;
440
441 new = (struct clocksource *) data;
442
443 write_seqlock_irqsave(&timekeeper.lock, flags);
444
445 timekeeping_forward_now();
446 if (!new->enable || new->enable(new) == 0) {
447 old = timekeeper.clock;
448 timekeeper_setup_internals(new);
449 if (old->disable)
450 old->disable(old);
451 }
452 timekeeping_update(true);
453
454 write_sequnlock_irqrestore(&timekeeper.lock, flags);
455
456 return 0;
457}
458
459/**
460 * timekeeping_notify - Install a new clock source
461 * @clock: pointer to the clock source
462 *
463 * This function is called from clocksource.c after a new, better clock
464 * source has been registered. The caller holds the clocksource_mutex.
465 */
466void timekeeping_notify(struct clocksource *clock)
467{
468 if (timekeeper.clock == clock)
469 return;
470 stop_machine(change_clocksource, clock, NULL);
471 tick_clock_notify();
472}
473
474/**
475 * ktime_get_real - get the real (wall-) time in ktime_t format
476 *
477 * returns the time in ktime_t format
478 */
479ktime_t ktime_get_real(void)
480{
481 struct timespec now;
482
483 getnstimeofday(&now);
484
485 return timespec_to_ktime(now);
486}
487EXPORT_SYMBOL_GPL(ktime_get_real);
488
489/**
490 * getrawmonotonic - Returns the raw monotonic time in a timespec
491 * @ts: pointer to the timespec to be set
492 *
493 * Returns the raw monotonic time (completely un-modified by ntp)
494 */
495void getrawmonotonic(struct timespec *ts)
496{
497 unsigned long seq;
498 s64 nsecs;
499
500 do {
501 seq = read_seqbegin(&timekeeper.lock);
502 nsecs = timekeeping_get_ns_raw();
503 *ts = timekeeper.raw_time;
504
505 } while (read_seqretry(&timekeeper.lock, seq));
506
507 timespec_add_ns(ts, nsecs);
508}
509EXPORT_SYMBOL(getrawmonotonic);
510
511
512/**
513 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
514 */
515int timekeeping_valid_for_hres(void)
516{
517 unsigned long seq;
518 int ret;
519
520 do {
521 seq = read_seqbegin(&timekeeper.lock);
522
523 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
524
525 } while (read_seqretry(&timekeeper.lock, seq));
526
527 return ret;
528}
529
530/**
531 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
532 */
533u64 timekeeping_max_deferment(void)
534{
535 unsigned long seq;
536 u64 ret;
537 do {
538 seq = read_seqbegin(&timekeeper.lock);
539
540 ret = timekeeper.clock->max_idle_ns;
541
542 } while (read_seqretry(&timekeeper.lock, seq));
543
544 return ret;
545}
546
547/**
548 * read_persistent_clock - Return time from the persistent clock.
549 *
550 * Weak dummy function for arches that do not yet support it.
551 * Reads the time from the battery backed persistent clock.
552 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
553 *
554 * XXX - Do be sure to remove it once all arches implement it.
555 */
556void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
557{
558 ts->tv_sec = 0;
559 ts->tv_nsec = 0;
560}
561
562/**
563 * read_boot_clock - Return time of the system start.
564 *
565 * Weak dummy function for arches that do not yet support it.
566 * Function to read the exact time the system has been started.
567 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
568 *
569 * XXX - Do be sure to remove it once all arches implement it.
570 */
571void __attribute__((weak)) read_boot_clock(struct timespec *ts)
572{
573 ts->tv_sec = 0;
574 ts->tv_nsec = 0;
575}
576
577/*
578 * timekeeping_init - Initializes the clocksource and common timekeeping values
579 */
580void __init timekeeping_init(void)
581{
582 struct clocksource *clock;
583 unsigned long flags;
584 struct timespec now, boot;
585
586 read_persistent_clock(&now);
587 read_boot_clock(&boot);
588
589 seqlock_init(&timekeeper.lock);
590
591 ntp_init();
592
593 write_seqlock_irqsave(&timekeeper.lock, flags);
594 clock = clocksource_default_clock();
595 if (clock->enable)
596 clock->enable(clock);
597 timekeeper_setup_internals(clock);
598
599 timekeeper.xtime.tv_sec = now.tv_sec;
600 timekeeper.xtime.tv_nsec = now.tv_nsec;
601 timekeeper.raw_time.tv_sec = 0;
602 timekeeper.raw_time.tv_nsec = 0;
603 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
604 boot.tv_sec = timekeeper.xtime.tv_sec;
605 boot.tv_nsec = timekeeper.xtime.tv_nsec;
606 }
607 set_normalized_timespec(&timekeeper.wall_to_monotonic,
608 -boot.tv_sec, -boot.tv_nsec);
609 timekeeper.total_sleep_time.tv_sec = 0;
610 timekeeper.total_sleep_time.tv_nsec = 0;
611 write_sequnlock_irqrestore(&timekeeper.lock, flags);
612}
613
614/* time in seconds when suspend began */
615static struct timespec timekeeping_suspend_time;
616
617/**
618 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
619 * @delta: pointer to a timespec delta value
620 *
621 * Takes a timespec offset measuring a suspend interval and properly
622 * adds the sleep offset to the timekeeping variables.
623 */
624static void __timekeeping_inject_sleeptime(struct timespec *delta)
625{
626 if (!timespec_valid(delta)) {
627 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
628 "sleep delta value!\n");
629 return;
630 }
631
632 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
633 timekeeper.wall_to_monotonic =
634 timespec_sub(timekeeper.wall_to_monotonic, *delta);
635 timekeeper.total_sleep_time = timespec_add(
636 timekeeper.total_sleep_time, *delta);
637}
638
639
640/**
641 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
642 * @delta: pointer to a timespec delta value
643 *
644 * This hook is for architectures that cannot support read_persistent_clock
645 * because their RTC/persistent clock is only accessible when irqs are enabled.
646 *
647 * This function should only be called by rtc_resume(), and allows
648 * a suspend offset to be injected into the timekeeping values.
649 */
650void timekeeping_inject_sleeptime(struct timespec *delta)
651{
652 unsigned long flags;
653 struct timespec ts;
654
655 /* Make sure we don't set the clock twice */
656 read_persistent_clock(&ts);
657 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
658 return;
659
660 write_seqlock_irqsave(&timekeeper.lock, flags);
661
662 timekeeping_forward_now();
663
664 __timekeeping_inject_sleeptime(delta);
665
666 timekeeping_update(true);
667
668 write_sequnlock_irqrestore(&timekeeper.lock, flags);
669
670 /* signal hrtimers about time change */
671 clock_was_set();
672}
673
674
675/**
676 * timekeeping_resume - Resumes the generic timekeeping subsystem.
677 *
678 * This is for the generic clocksource timekeeping.
679 * xtime/wall_to_monotonic/jiffies/etc are
680 * still managed by arch specific suspend/resume code.
681 */
682static void timekeeping_resume(void)
683{
684 unsigned long flags;
685 struct timespec ts;
686
687 read_persistent_clock(&ts);
688
689 clocksource_resume();
690
691 write_seqlock_irqsave(&timekeeper.lock, flags);
692
693 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
694 ts = timespec_sub(ts, timekeeping_suspend_time);
695 __timekeeping_inject_sleeptime(&ts);
696 }
697 /* re-base the last cycle value */
698 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
699 timekeeper.ntp_error = 0;
700 timekeeping_suspended = 0;
701 write_sequnlock_irqrestore(&timekeeper.lock, flags);
702
703 touch_softlockup_watchdog();
704
705 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
706
707 /* Resume hrtimers */
708 hrtimers_resume();
709}
710
711static int timekeeping_suspend(void)
712{
713 unsigned long flags;
714 struct timespec delta, delta_delta;
715 static struct timespec old_delta;
716
717 read_persistent_clock(&timekeeping_suspend_time);
718
719 write_seqlock_irqsave(&timekeeper.lock, flags);
720 timekeeping_forward_now();
721 timekeeping_suspended = 1;
722
723 /*
724 * To avoid drift caused by repeated suspend/resumes,
725 * which each can add ~1 second drift error,
726 * try to compensate so the difference in system time
727 * and persistent_clock time stays close to constant.
728 */
729 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
730 delta_delta = timespec_sub(delta, old_delta);
731 if (abs(delta_delta.tv_sec) >= 2) {
732 /*
733 * if delta_delta is too large, assume time correction
734 * has occured and set old_delta to the current delta.
735 */
736 old_delta = delta;
737 } else {
738 /* Otherwise try to adjust old_system to compensate */
739 timekeeping_suspend_time =
740 timespec_add(timekeeping_suspend_time, delta_delta);
741 }
742 write_sequnlock_irqrestore(&timekeeper.lock, flags);
743
744 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
745 clocksource_suspend();
746
747 return 0;
748}
749
750/* sysfs resume/suspend bits for timekeeping */
751static struct syscore_ops timekeeping_syscore_ops = {
752 .resume = timekeeping_resume,
753 .suspend = timekeeping_suspend,
754};
755
756static int __init timekeeping_init_ops(void)
757{
758 register_syscore_ops(&timekeeping_syscore_ops);
759 return 0;
760}
761
762device_initcall(timekeeping_init_ops);
763
764/*
765 * If the error is already larger, we look ahead even further
766 * to compensate for late or lost adjustments.
767 */
768static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
769 s64 *offset)
770{
771 s64 tick_error, i;
772 u32 look_ahead, adj;
773 s32 error2, mult;
774
775 /*
776 * Use the current error value to determine how much to look ahead.
777 * The larger the error the slower we adjust for it to avoid problems
778 * with losing too many ticks, otherwise we would overadjust and
779 * produce an even larger error. The smaller the adjustment the
780 * faster we try to adjust for it, as lost ticks can do less harm
781 * here. This is tuned so that an error of about 1 msec is adjusted
782 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
783 */
784 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
785 error2 = abs(error2);
786 for (look_ahead = 0; error2 > 0; look_ahead++)
787 error2 >>= 2;
788
789 /*
790 * Now calculate the error in (1 << look_ahead) ticks, but first
791 * remove the single look ahead already included in the error.
792 */
793 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
794 tick_error -= timekeeper.xtime_interval >> 1;
795 error = ((error - tick_error) >> look_ahead) + tick_error;
796
797 /* Finally calculate the adjustment shift value. */
798 i = *interval;
799 mult = 1;
800 if (error < 0) {
801 error = -error;
802 *interval = -*interval;
803 *offset = -*offset;
804 mult = -1;
805 }
806 for (adj = 0; error > i; adj++)
807 error >>= 1;
808
809 *interval <<= adj;
810 *offset <<= adj;
811 return mult << adj;
812}
813
814/*
815 * Adjust the multiplier to reduce the error value,
816 * this is optimized for the most common adjustments of -1,0,1,
817 * for other values we can do a bit more work.
818 */
819static void timekeeping_adjust(s64 offset)
820{
821 s64 error, interval = timekeeper.cycle_interval;
822 int adj;
823
824 /*
825 * The point of this is to check if the error is greater than half
826 * an interval.
827 *
828 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
829 *
830 * Note we subtract one in the shift, so that error is really error*2.
831 * This "saves" dividing(shifting) interval twice, but keeps the
832 * (error > interval) comparison as still measuring if error is
833 * larger than half an interval.
834 *
835 * Note: It does not "save" on aggravation when reading the code.
836 */
837 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
838 if (error > interval) {
839 /*
840 * We now divide error by 4(via shift), which checks if
841 * the error is greater than twice the interval.
842 * If it is greater, we need a bigadjust, if its smaller,
843 * we can adjust by 1.
844 */
845 error >>= 2;
846 /*
847 * XXX - In update_wall_time, we round up to the next
848 * nanosecond, and store the amount rounded up into
849 * the error. This causes the likely below to be unlikely.
850 *
851 * The proper fix is to avoid rounding up by using
852 * the high precision timekeeper.xtime_nsec instead of
853 * xtime.tv_nsec everywhere. Fixing this will take some
854 * time.
855 */
856 if (likely(error <= interval))
857 adj = 1;
858 else
859 adj = timekeeping_bigadjust(error, &interval, &offset);
860 } else if (error < -interval) {
861 /* See comment above, this is just switched for the negative */
862 error >>= 2;
863 if (likely(error >= -interval)) {
864 adj = -1;
865 interval = -interval;
866 offset = -offset;
867 } else
868 adj = timekeeping_bigadjust(error, &interval, &offset);
869 } else /* No adjustment needed */
870 return;
871
872 if (unlikely(timekeeper.clock->maxadj &&
873 (timekeeper.mult + adj >
874 timekeeper.clock->mult + timekeeper.clock->maxadj))) {
875 printk_once(KERN_WARNING
876 "Adjusting %s more than 11%% (%ld vs %ld)\n",
877 timekeeper.clock->name, (long)timekeeper.mult + adj,
878 (long)timekeeper.clock->mult +
879 timekeeper.clock->maxadj);
880 }
881 /*
882 * So the following can be confusing.
883 *
884 * To keep things simple, lets assume adj == 1 for now.
885 *
886 * When adj != 1, remember that the interval and offset values
887 * have been appropriately scaled so the math is the same.
888 *
889 * The basic idea here is that we're increasing the multiplier
890 * by one, this causes the xtime_interval to be incremented by
891 * one cycle_interval. This is because:
892 * xtime_interval = cycle_interval * mult
893 * So if mult is being incremented by one:
894 * xtime_interval = cycle_interval * (mult + 1)
895 * Its the same as:
896 * xtime_interval = (cycle_interval * mult) + cycle_interval
897 * Which can be shortened to:
898 * xtime_interval += cycle_interval
899 *
900 * So offset stores the non-accumulated cycles. Thus the current
901 * time (in shifted nanoseconds) is:
902 * now = (offset * adj) + xtime_nsec
903 * Now, even though we're adjusting the clock frequency, we have
904 * to keep time consistent. In other words, we can't jump back
905 * in time, and we also want to avoid jumping forward in time.
906 *
907 * So given the same offset value, we need the time to be the same
908 * both before and after the freq adjustment.
909 * now = (offset * adj_1) + xtime_nsec_1
910 * now = (offset * adj_2) + xtime_nsec_2
911 * So:
912 * (offset * adj_1) + xtime_nsec_1 =
913 * (offset * adj_2) + xtime_nsec_2
914 * And we know:
915 * adj_2 = adj_1 + 1
916 * So:
917 * (offset * adj_1) + xtime_nsec_1 =
918 * (offset * (adj_1+1)) + xtime_nsec_2
919 * (offset * adj_1) + xtime_nsec_1 =
920 * (offset * adj_1) + offset + xtime_nsec_2
921 * Canceling the sides:
922 * xtime_nsec_1 = offset + xtime_nsec_2
923 * Which gives us:
924 * xtime_nsec_2 = xtime_nsec_1 - offset
925 * Which simplfies to:
926 * xtime_nsec -= offset
927 *
928 * XXX - TODO: Doc ntp_error calculation.
929 */
930 timekeeper.mult += adj;
931 timekeeper.xtime_interval += interval;
932 timekeeper.xtime_nsec -= offset;
933 timekeeper.ntp_error -= (interval - offset) <<
934 timekeeper.ntp_error_shift;
935}
936
937
938/**
939 * logarithmic_accumulation - shifted accumulation of cycles
940 *
941 * This functions accumulates a shifted interval of cycles into
942 * into a shifted interval nanoseconds. Allows for O(log) accumulation
943 * loop.
944 *
945 * Returns the unconsumed cycles.
946 */
947static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
948{
949 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
950 u64 raw_nsecs;
951
952 /* If the offset is smaller than a shifted interval, do nothing */
953 if (offset < timekeeper.cycle_interval<<shift)
954 return offset;
955
956 /* Accumulate one shifted interval */
957 offset -= timekeeper.cycle_interval << shift;
958 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
959
960 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
961 while (timekeeper.xtime_nsec >= nsecps) {
962 int leap;
963 timekeeper.xtime_nsec -= nsecps;
964 timekeeper.xtime.tv_sec++;
965 leap = second_overflow(timekeeper.xtime.tv_sec);
966 timekeeper.xtime.tv_sec += leap;
967 }
968
969 /* Accumulate raw time */
970 raw_nsecs = timekeeper.raw_interval << shift;
971 raw_nsecs += timekeeper.raw_time.tv_nsec;
972 if (raw_nsecs >= NSEC_PER_SEC) {
973 u64 raw_secs = raw_nsecs;
974 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
975 timekeeper.raw_time.tv_sec += raw_secs;
976 }
977 timekeeper.raw_time.tv_nsec = raw_nsecs;
978
979 /* Accumulate error between NTP and clock interval */
980 timekeeper.ntp_error += ntp_tick_length() << shift;
981 timekeeper.ntp_error -=
982 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
983 (timekeeper.ntp_error_shift + shift);
984
985 return offset;
986}
987
988
989/**
990 * update_wall_time - Uses the current clocksource to increment the wall time
991 *
992 */
993static void update_wall_time(void)
994{
995 struct clocksource *clock;
996 cycle_t offset;
997 int shift = 0, maxshift;
998 unsigned long flags;
999
1000 write_seqlock_irqsave(&timekeeper.lock, flags);
1001
1002 /* Make sure we're fully resumed: */
1003 if (unlikely(timekeeping_suspended))
1004 goto out;
1005
1006 clock = timekeeper.clock;
1007
1008#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1009 offset = timekeeper.cycle_interval;
1010#else
1011 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1012#endif
1013 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1014 timekeeper.shift;
1015
1016 /*
1017 * With NO_HZ we may have to accumulate many cycle_intervals
1018 * (think "ticks") worth of time at once. To do this efficiently,
1019 * we calculate the largest doubling multiple of cycle_intervals
1020 * that is smaller than the offset. We then accumulate that
1021 * chunk in one go, and then try to consume the next smaller
1022 * doubled multiple.
1023 */
1024 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1025 shift = max(0, shift);
1026 /* Bound shift to one less than what overflows tick_length */
1027 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1028 shift = min(shift, maxshift);
1029 while (offset >= timekeeper.cycle_interval) {
1030 offset = logarithmic_accumulation(offset, shift);
1031 if(offset < timekeeper.cycle_interval<<shift)
1032 shift--;
1033 }
1034
1035 /* correct the clock when NTP error is too big */
1036 timekeeping_adjust(offset);
1037
1038 /*
1039 * Since in the loop above, we accumulate any amount of time
1040 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1041 * xtime_nsec to be fairly small after the loop. Further, if we're
1042 * slightly speeding the clocksource up in timekeeping_adjust(),
1043 * its possible the required corrective factor to xtime_nsec could
1044 * cause it to underflow.
1045 *
1046 * Now, we cannot simply roll the accumulated second back, since
1047 * the NTP subsystem has been notified via second_overflow. So
1048 * instead we push xtime_nsec forward by the amount we underflowed,
1049 * and add that amount into the error.
1050 *
1051 * We'll correct this error next time through this function, when
1052 * xtime_nsec is not as small.
1053 */
1054 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1055 s64 neg = -(s64)timekeeper.xtime_nsec;
1056 timekeeper.xtime_nsec = 0;
1057 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1058 }
1059
1060
1061 /*
1062 * Store full nanoseconds into xtime after rounding it up and
1063 * add the remainder to the error difference.
1064 */
1065 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1066 timekeeper.shift) + 1;
1067 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1068 timekeeper.shift;
1069 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1070 timekeeper.ntp_error_shift;
1071
1072 /*
1073 * Finally, make sure that after the rounding
1074 * xtime.tv_nsec isn't larger than NSEC_PER_SEC
1075 */
1076 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1077 int leap;
1078 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1079 timekeeper.xtime.tv_sec++;
1080 leap = second_overflow(timekeeper.xtime.tv_sec);
1081 timekeeper.xtime.tv_sec += leap;
1082 }
1083
1084 timekeeping_update(false);
1085
1086out:
1087 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1088
1089}
1090
1091/**
1092 * getboottime - Return the real time of system boot.
1093 * @ts: pointer to the timespec to be set
1094 *
1095 * Returns the wall-time of boot in a timespec.
1096 *
1097 * This is based on the wall_to_monotonic offset and the total suspend
1098 * time. Calls to settimeofday will affect the value returned (which
1099 * basically means that however wrong your real time clock is at boot time,
1100 * you get the right time here).
1101 */
1102void getboottime(struct timespec *ts)
1103{
1104 struct timespec boottime = {
1105 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1106 timekeeper.total_sleep_time.tv_sec,
1107 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1108 timekeeper.total_sleep_time.tv_nsec
1109 };
1110
1111 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1112}
1113EXPORT_SYMBOL_GPL(getboottime);
1114
1115
1116/**
1117 * get_monotonic_boottime - Returns monotonic time since boot
1118 * @ts: pointer to the timespec to be set
1119 *
1120 * Returns the monotonic time since boot in a timespec.
1121 *
1122 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1123 * includes the time spent in suspend.
1124 */
1125void get_monotonic_boottime(struct timespec *ts)
1126{
1127 struct timespec tomono, sleep;
1128 unsigned int seq;
1129 s64 nsecs;
1130
1131 WARN_ON(timekeeping_suspended);
1132
1133 do {
1134 seq = read_seqbegin(&timekeeper.lock);
1135 *ts = timekeeper.xtime;
1136 tomono = timekeeper.wall_to_monotonic;
1137 sleep = timekeeper.total_sleep_time;
1138 nsecs = timekeeping_get_ns();
1139
1140 } while (read_seqretry(&timekeeper.lock, seq));
1141
1142 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1143 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1144}
1145EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1146
1147/**
1148 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1149 *
1150 * Returns the monotonic time since boot in a ktime
1151 *
1152 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1153 * includes the time spent in suspend.
1154 */
1155ktime_t ktime_get_boottime(void)
1156{
1157 struct timespec ts;
1158
1159 get_monotonic_boottime(&ts);
1160 return timespec_to_ktime(ts);
1161}
1162EXPORT_SYMBOL_GPL(ktime_get_boottime);
1163
1164/**
1165 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1166 * @ts: pointer to the timespec to be converted
1167 */
1168void monotonic_to_bootbased(struct timespec *ts)
1169{
1170 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1171}
1172EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1173
1174unsigned long get_seconds(void)
1175{
1176 return timekeeper.xtime.tv_sec;
1177}
1178EXPORT_SYMBOL(get_seconds);
1179
1180struct timespec __current_kernel_time(void)
1181{
1182 return timekeeper.xtime;
1183}
1184
1185struct timespec current_kernel_time(void)
1186{
1187 struct timespec now;
1188 unsigned long seq;
1189
1190 do {
1191 seq = read_seqbegin(&timekeeper.lock);
1192
1193 now = timekeeper.xtime;
1194 } while (read_seqretry(&timekeeper.lock, seq));
1195
1196 return now;
1197}
1198EXPORT_SYMBOL(current_kernel_time);
1199
1200struct timespec get_monotonic_coarse(void)
1201{
1202 struct timespec now, mono;
1203 unsigned long seq;
1204
1205 do {
1206 seq = read_seqbegin(&timekeeper.lock);
1207
1208 now = timekeeper.xtime;
1209 mono = timekeeper.wall_to_monotonic;
1210 } while (read_seqretry(&timekeeper.lock, seq));
1211
1212 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1213 now.tv_nsec + mono.tv_nsec);
1214 return now;
1215}
1216
1217/*
1218 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1219 * without sampling the sequence number in xtime_lock.
1220 * jiffies is defined in the linker script...
1221 */
1222void do_timer(unsigned long ticks)
1223{
1224 jiffies_64 += ticks;
1225 update_wall_time();
1226 calc_global_load(ticks);
1227}
1228
1229/**
1230 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1231 * and sleep offsets.
1232 * @xtim: pointer to timespec to be set with xtime
1233 * @wtom: pointer to timespec to be set with wall_to_monotonic
1234 * @sleep: pointer to timespec to be set with time in suspend
1235 */
1236void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1237 struct timespec *wtom, struct timespec *sleep)
1238{
1239 unsigned long seq;
1240
1241 do {
1242 seq = read_seqbegin(&timekeeper.lock);
1243 *xtim = timekeeper.xtime;
1244 *wtom = timekeeper.wall_to_monotonic;
1245 *sleep = timekeeper.total_sleep_time;
1246 } while (read_seqretry(&timekeeper.lock, seq));
1247}
1248
1249/**
1250 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1251 */
1252ktime_t ktime_get_monotonic_offset(void)
1253{
1254 unsigned long seq;
1255 struct timespec wtom;
1256
1257 do {
1258 seq = read_seqbegin(&timekeeper.lock);
1259 wtom = timekeeper.wall_to_monotonic;
1260 } while (read_seqretry(&timekeeper.lock, seq));
1261
1262 return timespec_to_ktime(wtom);
1263}
1264EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1265
1266
1267/**
1268 * xtime_update() - advances the timekeeping infrastructure
1269 * @ticks: number of ticks, that have elapsed since the last call.
1270 *
1271 * Must be called with interrupts disabled.
1272 */
1273void xtime_update(unsigned long ticks)
1274{
1275 write_seqlock(&xtime_lock);
1276 do_timer(ticks);
1277 write_sequnlock(&xtime_lock);
1278}