Loading...
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 /* Offset clock monotonic -> clock realtime */
74 ktime_t offs_real;
75
76 /* Offset clock monotonic -> clock boottime */
77 ktime_t offs_boot;
78
79 /* Seqlock for all timekeeper values */
80 seqlock_t lock;
81};
82
83static struct timekeeper timekeeper;
84
85/*
86 * This read-write spinlock protects us from races in SMP while
87 * playing with xtime.
88 */
89__cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
90
91
92/* flag for if timekeeping is suspended */
93int __read_mostly timekeeping_suspended;
94
95
96
97/**
98 * timekeeper_setup_internals - Set up internals to use clocksource clock.
99 *
100 * @clock: Pointer to clocksource.
101 *
102 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
103 * pair and interval request.
104 *
105 * Unless you're the timekeeping code, you should not be using this!
106 */
107static void timekeeper_setup_internals(struct clocksource *clock)
108{
109 cycle_t interval;
110 u64 tmp, ntpinterval;
111
112 timekeeper.clock = clock;
113 clock->cycle_last = clock->read(clock);
114
115 /* Do the ns -> cycle conversion first, using original mult */
116 tmp = NTP_INTERVAL_LENGTH;
117 tmp <<= clock->shift;
118 ntpinterval = tmp;
119 tmp += clock->mult/2;
120 do_div(tmp, clock->mult);
121 if (tmp == 0)
122 tmp = 1;
123
124 interval = (cycle_t) tmp;
125 timekeeper.cycle_interval = interval;
126
127 /* Go back from cycles -> shifted ns */
128 timekeeper.xtime_interval = (u64) interval * clock->mult;
129 timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
130 timekeeper.raw_interval =
131 ((u64) interval * clock->mult) >> clock->shift;
132
133 timekeeper.xtime_nsec = 0;
134 timekeeper.shift = clock->shift;
135
136 timekeeper.ntp_error = 0;
137 timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
138
139 /*
140 * The timekeeper keeps its own mult values for the currently
141 * active clocksource. These value will be adjusted via NTP
142 * to counteract clock drifting.
143 */
144 timekeeper.mult = clock->mult;
145}
146
147/* Timekeeper helper functions. */
148static inline s64 timekeeping_get_ns(void)
149{
150 cycle_t cycle_now, cycle_delta;
151 struct clocksource *clock;
152
153 /* read clocksource: */
154 clock = timekeeper.clock;
155 cycle_now = clock->read(clock);
156
157 /* calculate the delta since the last update_wall_time: */
158 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
159
160 /* return delta convert to nanoseconds using ntp adjusted mult. */
161 return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
162 timekeeper.shift);
163}
164
165static inline s64 timekeeping_get_ns_raw(void)
166{
167 cycle_t cycle_now, cycle_delta;
168 struct clocksource *clock;
169
170 /* read clocksource: */
171 clock = timekeeper.clock;
172 cycle_now = clock->read(clock);
173
174 /* calculate the delta since the last update_wall_time: */
175 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
176
177 /* return delta convert to nanoseconds. */
178 return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
179}
180
181static void update_rt_offset(void)
182{
183 struct timespec tmp, *wtm = &timekeeper.wall_to_monotonic;
184
185 set_normalized_timespec(&tmp, -wtm->tv_sec, -wtm->tv_nsec);
186 timekeeper.offs_real = timespec_to_ktime(tmp);
187}
188
189/* must hold write on timekeeper.lock */
190static void timekeeping_update(bool clearntp)
191{
192 if (clearntp) {
193 timekeeper.ntp_error = 0;
194 ntp_clear();
195 }
196 update_rt_offset();
197 update_vsyscall(&timekeeper.xtime, &timekeeper.wall_to_monotonic,
198 timekeeper.clock, timekeeper.mult);
199}
200
201
202/**
203 * timekeeping_forward_now - update clock to the current time
204 *
205 * Forward the current clock to update its state since the last call to
206 * update_wall_time(). This is useful before significant clock changes,
207 * as it avoids having to deal with this time offset explicitly.
208 */
209static void timekeeping_forward_now(void)
210{
211 cycle_t cycle_now, cycle_delta;
212 struct clocksource *clock;
213 s64 nsec;
214
215 clock = timekeeper.clock;
216 cycle_now = clock->read(clock);
217 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
218 clock->cycle_last = cycle_now;
219
220 nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
221 timekeeper.shift);
222
223 /* If arch requires, add in gettimeoffset() */
224 nsec += arch_gettimeoffset();
225
226 timespec_add_ns(&timekeeper.xtime, nsec);
227
228 nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
229 timespec_add_ns(&timekeeper.raw_time, nsec);
230}
231
232/**
233 * getnstimeofday - Returns the time of day in a timespec
234 * @ts: pointer to the timespec to be set
235 *
236 * Returns the time of day in a timespec.
237 */
238void getnstimeofday(struct timespec *ts)
239{
240 unsigned long seq;
241 s64 nsecs;
242
243 WARN_ON(timekeeping_suspended);
244
245 do {
246 seq = read_seqbegin(&timekeeper.lock);
247
248 *ts = timekeeper.xtime;
249 nsecs = timekeeping_get_ns();
250
251 /* If arch requires, add in gettimeoffset() */
252 nsecs += arch_gettimeoffset();
253
254 } while (read_seqretry(&timekeeper.lock, seq));
255
256 timespec_add_ns(ts, nsecs);
257}
258EXPORT_SYMBOL(getnstimeofday);
259
260ktime_t ktime_get(void)
261{
262 unsigned int seq;
263 s64 secs, nsecs;
264
265 WARN_ON(timekeeping_suspended);
266
267 do {
268 seq = read_seqbegin(&timekeeper.lock);
269 secs = timekeeper.xtime.tv_sec +
270 timekeeper.wall_to_monotonic.tv_sec;
271 nsecs = timekeeper.xtime.tv_nsec +
272 timekeeper.wall_to_monotonic.tv_nsec;
273 nsecs += timekeeping_get_ns();
274 /* If arch requires, add in gettimeoffset() */
275 nsecs += arch_gettimeoffset();
276
277 } while (read_seqretry(&timekeeper.lock, seq));
278 /*
279 * Use ktime_set/ktime_add_ns to create a proper ktime on
280 * 32-bit architectures without CONFIG_KTIME_SCALAR.
281 */
282 return ktime_add_ns(ktime_set(secs, 0), nsecs);
283}
284EXPORT_SYMBOL_GPL(ktime_get);
285
286/**
287 * ktime_get_ts - get the monotonic clock in timespec format
288 * @ts: pointer to timespec variable
289 *
290 * The function calculates the monotonic clock from the realtime
291 * clock and the wall_to_monotonic offset and stores the result
292 * in normalized timespec format in the variable pointed to by @ts.
293 */
294void ktime_get_ts(struct timespec *ts)
295{
296 struct timespec tomono;
297 unsigned int seq;
298 s64 nsecs;
299
300 WARN_ON(timekeeping_suspended);
301
302 do {
303 seq = read_seqbegin(&timekeeper.lock);
304 *ts = timekeeper.xtime;
305 tomono = timekeeper.wall_to_monotonic;
306 nsecs = timekeeping_get_ns();
307 /* If arch requires, add in gettimeoffset() */
308 nsecs += arch_gettimeoffset();
309
310 } while (read_seqretry(&timekeeper.lock, seq));
311
312 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
313 ts->tv_nsec + tomono.tv_nsec + nsecs);
314}
315EXPORT_SYMBOL_GPL(ktime_get_ts);
316
317#ifdef CONFIG_NTP_PPS
318
319/**
320 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
321 * @ts_raw: pointer to the timespec to be set to raw monotonic time
322 * @ts_real: pointer to the timespec to be set to the time of day
323 *
324 * This function reads both the time of day and raw monotonic time at the
325 * same time atomically and stores the resulting timestamps in timespec
326 * format.
327 */
328void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
329{
330 unsigned long seq;
331 s64 nsecs_raw, nsecs_real;
332
333 WARN_ON_ONCE(timekeeping_suspended);
334
335 do {
336 u32 arch_offset;
337
338 seq = read_seqbegin(&timekeeper.lock);
339
340 *ts_raw = timekeeper.raw_time;
341 *ts_real = timekeeper.xtime;
342
343 nsecs_raw = timekeeping_get_ns_raw();
344 nsecs_real = timekeeping_get_ns();
345
346 /* If arch requires, add in gettimeoffset() */
347 arch_offset = arch_gettimeoffset();
348 nsecs_raw += arch_offset;
349 nsecs_real += arch_offset;
350
351 } while (read_seqretry(&timekeeper.lock, seq));
352
353 timespec_add_ns(ts_raw, nsecs_raw);
354 timespec_add_ns(ts_real, nsecs_real);
355}
356EXPORT_SYMBOL(getnstime_raw_and_real);
357
358#endif /* CONFIG_NTP_PPS */
359
360/**
361 * do_gettimeofday - Returns the time of day in a timeval
362 * @tv: pointer to the timeval to be set
363 *
364 * NOTE: Users should be converted to using getnstimeofday()
365 */
366void do_gettimeofday(struct timeval *tv)
367{
368 struct timespec now;
369
370 getnstimeofday(&now);
371 tv->tv_sec = now.tv_sec;
372 tv->tv_usec = now.tv_nsec/1000;
373}
374EXPORT_SYMBOL(do_gettimeofday);
375
376/**
377 * do_settimeofday - Sets the time of day
378 * @tv: pointer to the timespec variable containing the new time
379 *
380 * Sets the time of day to the new time and update NTP and notify hrtimers
381 */
382int do_settimeofday(const struct timespec *tv)
383{
384 struct timespec ts_delta;
385 unsigned long flags;
386
387 if (!timespec_valid_strict(tv))
388 return -EINVAL;
389
390 write_seqlock_irqsave(&timekeeper.lock, flags);
391
392 timekeeping_forward_now();
393
394 ts_delta.tv_sec = tv->tv_sec - timekeeper.xtime.tv_sec;
395 ts_delta.tv_nsec = tv->tv_nsec - timekeeper.xtime.tv_nsec;
396 timekeeper.wall_to_monotonic =
397 timespec_sub(timekeeper.wall_to_monotonic, ts_delta);
398
399 timekeeper.xtime = *tv;
400 timekeeping_update(true);
401
402 write_sequnlock_irqrestore(&timekeeper.lock, flags);
403
404 /* signal hrtimers about time change */
405 clock_was_set();
406
407 return 0;
408}
409EXPORT_SYMBOL(do_settimeofday);
410
411
412/**
413 * timekeeping_inject_offset - Adds or subtracts from the current time.
414 * @tv: pointer to the timespec variable containing the offset
415 *
416 * Adds or subtracts an offset value from the current time.
417 */
418int timekeeping_inject_offset(struct timespec *ts)
419{
420 unsigned long flags;
421 struct timespec tmp;
422 int ret = 0;
423
424 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
425 return -EINVAL;
426
427 write_seqlock_irqsave(&timekeeper.lock, flags);
428
429 timekeeping_forward_now();
430
431 tmp = timespec_add(timekeeper.xtime, *ts);
432 if (!timespec_valid_strict(&tmp)) {
433 ret = -EINVAL;
434 goto error;
435 }
436
437 timekeeper.xtime = timespec_add(timekeeper.xtime, *ts);
438 timekeeper.wall_to_monotonic =
439 timespec_sub(timekeeper.wall_to_monotonic, *ts);
440
441error: /* even if we error out, we forwarded the time, so call update */
442 timekeeping_update(true);
443
444 write_sequnlock_irqrestore(&timekeeper.lock, flags);
445
446 /* signal hrtimers about time change */
447 clock_was_set();
448
449 return ret;
450}
451EXPORT_SYMBOL(timekeeping_inject_offset);
452
453/**
454 * change_clocksource - Swaps clocksources if a new one is available
455 *
456 * Accumulates current time interval and initializes new clocksource
457 */
458static int change_clocksource(void *data)
459{
460 struct clocksource *new, *old;
461 unsigned long flags;
462
463 new = (struct clocksource *) data;
464
465 write_seqlock_irqsave(&timekeeper.lock, flags);
466
467 timekeeping_forward_now();
468 if (!new->enable || new->enable(new) == 0) {
469 old = timekeeper.clock;
470 timekeeper_setup_internals(new);
471 if (old->disable)
472 old->disable(old);
473 }
474 timekeeping_update(true);
475
476 write_sequnlock_irqrestore(&timekeeper.lock, flags);
477
478 return 0;
479}
480
481/**
482 * timekeeping_notify - Install a new clock source
483 * @clock: pointer to the clock source
484 *
485 * This function is called from clocksource.c after a new, better clock
486 * source has been registered. The caller holds the clocksource_mutex.
487 */
488void timekeeping_notify(struct clocksource *clock)
489{
490 if (timekeeper.clock == clock)
491 return;
492 stop_machine(change_clocksource, clock, NULL);
493 tick_clock_notify();
494}
495
496/**
497 * ktime_get_real - get the real (wall-) time in ktime_t format
498 *
499 * returns the time in ktime_t format
500 */
501ktime_t ktime_get_real(void)
502{
503 struct timespec now;
504
505 getnstimeofday(&now);
506
507 return timespec_to_ktime(now);
508}
509EXPORT_SYMBOL_GPL(ktime_get_real);
510
511/**
512 * getrawmonotonic - Returns the raw monotonic time in a timespec
513 * @ts: pointer to the timespec to be set
514 *
515 * Returns the raw monotonic time (completely un-modified by ntp)
516 */
517void getrawmonotonic(struct timespec *ts)
518{
519 unsigned long seq;
520 s64 nsecs;
521
522 do {
523 seq = read_seqbegin(&timekeeper.lock);
524 nsecs = timekeeping_get_ns_raw();
525 *ts = timekeeper.raw_time;
526
527 } while (read_seqretry(&timekeeper.lock, seq));
528
529 timespec_add_ns(ts, nsecs);
530}
531EXPORT_SYMBOL(getrawmonotonic);
532
533
534/**
535 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
536 */
537int timekeeping_valid_for_hres(void)
538{
539 unsigned long seq;
540 int ret;
541
542 do {
543 seq = read_seqbegin(&timekeeper.lock);
544
545 ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
546
547 } while (read_seqretry(&timekeeper.lock, seq));
548
549 return ret;
550}
551
552/**
553 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
554 */
555u64 timekeeping_max_deferment(void)
556{
557 unsigned long seq;
558 u64 ret;
559 do {
560 seq = read_seqbegin(&timekeeper.lock);
561
562 ret = timekeeper.clock->max_idle_ns;
563
564 } while (read_seqretry(&timekeeper.lock, seq));
565
566 return ret;
567}
568
569/**
570 * read_persistent_clock - Return time from the persistent clock.
571 *
572 * Weak dummy function for arches that do not yet support it.
573 * Reads the time from the battery backed persistent clock.
574 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
575 *
576 * XXX - Do be sure to remove it once all arches implement it.
577 */
578void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
579{
580 ts->tv_sec = 0;
581 ts->tv_nsec = 0;
582}
583
584/**
585 * read_boot_clock - Return time of the system start.
586 *
587 * Weak dummy function for arches that do not yet support it.
588 * Function to read the exact time the system has been started.
589 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
590 *
591 * XXX - Do be sure to remove it once all arches implement it.
592 */
593void __attribute__((weak)) read_boot_clock(struct timespec *ts)
594{
595 ts->tv_sec = 0;
596 ts->tv_nsec = 0;
597}
598
599/*
600 * timekeeping_init - Initializes the clocksource and common timekeeping values
601 */
602void __init timekeeping_init(void)
603{
604 struct clocksource *clock;
605 unsigned long flags;
606 struct timespec now, boot;
607
608 read_persistent_clock(&now);
609 if (!timespec_valid_strict(&now)) {
610 pr_warn("WARNING: Persistent clock returned invalid value!\n"
611 " Check your CMOS/BIOS settings.\n");
612 now.tv_sec = 0;
613 now.tv_nsec = 0;
614 }
615
616 read_boot_clock(&boot);
617 if (!timespec_valid_strict(&boot)) {
618 pr_warn("WARNING: Boot clock returned invalid value!\n"
619 " Check your CMOS/BIOS settings.\n");
620 boot.tv_sec = 0;
621 boot.tv_nsec = 0;
622 }
623
624 seqlock_init(&timekeeper.lock);
625
626 ntp_init();
627
628 write_seqlock_irqsave(&timekeeper.lock, flags);
629 clock = clocksource_default_clock();
630 if (clock->enable)
631 clock->enable(clock);
632 timekeeper_setup_internals(clock);
633
634 timekeeper.xtime.tv_sec = now.tv_sec;
635 timekeeper.xtime.tv_nsec = now.tv_nsec;
636 timekeeper.raw_time.tv_sec = 0;
637 timekeeper.raw_time.tv_nsec = 0;
638 if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
639 boot.tv_sec = timekeeper.xtime.tv_sec;
640 boot.tv_nsec = timekeeper.xtime.tv_nsec;
641 }
642 set_normalized_timespec(&timekeeper.wall_to_monotonic,
643 -boot.tv_sec, -boot.tv_nsec);
644 update_rt_offset();
645 timekeeper.total_sleep_time.tv_sec = 0;
646 timekeeper.total_sleep_time.tv_nsec = 0;
647 write_sequnlock_irqrestore(&timekeeper.lock, flags);
648}
649
650/* time in seconds when suspend began */
651static struct timespec timekeeping_suspend_time;
652
653static void update_sleep_time(struct timespec t)
654{
655 timekeeper.total_sleep_time = t;
656 timekeeper.offs_boot = timespec_to_ktime(t);
657}
658
659/**
660 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
661 * @delta: pointer to a timespec delta value
662 *
663 * Takes a timespec offset measuring a suspend interval and properly
664 * adds the sleep offset to the timekeeping variables.
665 */
666static void __timekeeping_inject_sleeptime(struct timespec *delta)
667{
668 if (!timespec_valid_strict(delta)) {
669 printk(KERN_WARNING "__timekeeping_inject_sleeptime: Invalid "
670 "sleep delta value!\n");
671 return;
672 }
673
674 timekeeper.xtime = timespec_add(timekeeper.xtime, *delta);
675 timekeeper.wall_to_monotonic =
676 timespec_sub(timekeeper.wall_to_monotonic, *delta);
677 update_sleep_time(timespec_add(timekeeper.total_sleep_time, *delta));
678}
679
680
681/**
682 * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
683 * @delta: pointer to a timespec delta value
684 *
685 * This hook is for architectures that cannot support read_persistent_clock
686 * because their RTC/persistent clock is only accessible when irqs are enabled.
687 *
688 * This function should only be called by rtc_resume(), and allows
689 * a suspend offset to be injected into the timekeeping values.
690 */
691void timekeeping_inject_sleeptime(struct timespec *delta)
692{
693 unsigned long flags;
694 struct timespec ts;
695
696 /* Make sure we don't set the clock twice */
697 read_persistent_clock(&ts);
698 if (!(ts.tv_sec == 0 && ts.tv_nsec == 0))
699 return;
700
701 write_seqlock_irqsave(&timekeeper.lock, flags);
702
703 timekeeping_forward_now();
704
705 __timekeeping_inject_sleeptime(delta);
706
707 timekeeping_update(true);
708
709 write_sequnlock_irqrestore(&timekeeper.lock, flags);
710
711 /* signal hrtimers about time change */
712 clock_was_set();
713}
714
715
716/**
717 * timekeeping_resume - Resumes the generic timekeeping subsystem.
718 *
719 * This is for the generic clocksource timekeeping.
720 * xtime/wall_to_monotonic/jiffies/etc are
721 * still managed by arch specific suspend/resume code.
722 */
723static void timekeeping_resume(void)
724{
725 unsigned long flags;
726 struct timespec ts;
727
728 read_persistent_clock(&ts);
729
730 clocksource_resume();
731
732 write_seqlock_irqsave(&timekeeper.lock, flags);
733
734 if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
735 ts = timespec_sub(ts, timekeeping_suspend_time);
736 __timekeeping_inject_sleeptime(&ts);
737 }
738 /* re-base the last cycle value */
739 timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
740 timekeeper.ntp_error = 0;
741 timekeeping_suspended = 0;
742 timekeeping_update(false);
743 write_sequnlock_irqrestore(&timekeeper.lock, flags);
744
745 touch_softlockup_watchdog();
746
747 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
748
749 /* Resume hrtimers */
750 hrtimers_resume();
751}
752
753static int timekeeping_suspend(void)
754{
755 unsigned long flags;
756 struct timespec delta, delta_delta;
757 static struct timespec old_delta;
758
759 read_persistent_clock(&timekeeping_suspend_time);
760
761 write_seqlock_irqsave(&timekeeper.lock, flags);
762 timekeeping_forward_now();
763 timekeeping_suspended = 1;
764
765 /*
766 * To avoid drift caused by repeated suspend/resumes,
767 * which each can add ~1 second drift error,
768 * try to compensate so the difference in system time
769 * and persistent_clock time stays close to constant.
770 */
771 delta = timespec_sub(timekeeper.xtime, timekeeping_suspend_time);
772 delta_delta = timespec_sub(delta, old_delta);
773 if (abs(delta_delta.tv_sec) >= 2) {
774 /*
775 * if delta_delta is too large, assume time correction
776 * has occured and set old_delta to the current delta.
777 */
778 old_delta = delta;
779 } else {
780 /* Otherwise try to adjust old_system to compensate */
781 timekeeping_suspend_time =
782 timespec_add(timekeeping_suspend_time, delta_delta);
783 }
784 write_sequnlock_irqrestore(&timekeeper.lock, flags);
785
786 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
787 clocksource_suspend();
788
789 return 0;
790}
791
792/* sysfs resume/suspend bits for timekeeping */
793static struct syscore_ops timekeeping_syscore_ops = {
794 .resume = timekeeping_resume,
795 .suspend = timekeeping_suspend,
796};
797
798static int __init timekeeping_init_ops(void)
799{
800 register_syscore_ops(&timekeeping_syscore_ops);
801 return 0;
802}
803
804device_initcall(timekeeping_init_ops);
805
806/*
807 * If the error is already larger, we look ahead even further
808 * to compensate for late or lost adjustments.
809 */
810static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
811 s64 *offset)
812{
813 s64 tick_error, i;
814 u32 look_ahead, adj;
815 s32 error2, mult;
816
817 /*
818 * Use the current error value to determine how much to look ahead.
819 * The larger the error the slower we adjust for it to avoid problems
820 * with losing too many ticks, otherwise we would overadjust and
821 * produce an even larger error. The smaller the adjustment the
822 * faster we try to adjust for it, as lost ticks can do less harm
823 * here. This is tuned so that an error of about 1 msec is adjusted
824 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
825 */
826 error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
827 error2 = abs(error2);
828 for (look_ahead = 0; error2 > 0; look_ahead++)
829 error2 >>= 2;
830
831 /*
832 * Now calculate the error in (1 << look_ahead) ticks, but first
833 * remove the single look ahead already included in the error.
834 */
835 tick_error = ntp_tick_length() >> (timekeeper.ntp_error_shift + 1);
836 tick_error -= timekeeper.xtime_interval >> 1;
837 error = ((error - tick_error) >> look_ahead) + tick_error;
838
839 /* Finally calculate the adjustment shift value. */
840 i = *interval;
841 mult = 1;
842 if (error < 0) {
843 error = -error;
844 *interval = -*interval;
845 *offset = -*offset;
846 mult = -1;
847 }
848 for (adj = 0; error > i; adj++)
849 error >>= 1;
850
851 *interval <<= adj;
852 *offset <<= adj;
853 return mult << adj;
854}
855
856/*
857 * Adjust the multiplier to reduce the error value,
858 * this is optimized for the most common adjustments of -1,0,1,
859 * for other values we can do a bit more work.
860 */
861static void timekeeping_adjust(s64 offset)
862{
863 s64 error, interval = timekeeper.cycle_interval;
864 int adj;
865
866 /*
867 * The point of this is to check if the error is greater than half
868 * an interval.
869 *
870 * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs.
871 *
872 * Note we subtract one in the shift, so that error is really error*2.
873 * This "saves" dividing(shifting) interval twice, but keeps the
874 * (error > interval) comparison as still measuring if error is
875 * larger than half an interval.
876 *
877 * Note: It does not "save" on aggravation when reading the code.
878 */
879 error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
880 if (error > interval) {
881 /*
882 * We now divide error by 4(via shift), which checks if
883 * the error is greater than twice the interval.
884 * If it is greater, we need a bigadjust, if its smaller,
885 * we can adjust by 1.
886 */
887 error >>= 2;
888 /*
889 * XXX - In update_wall_time, we round up to the next
890 * nanosecond, and store the amount rounded up into
891 * the error. This causes the likely below to be unlikely.
892 *
893 * The proper fix is to avoid rounding up by using
894 * the high precision timekeeper.xtime_nsec instead of
895 * xtime.tv_nsec everywhere. Fixing this will take some
896 * time.
897 */
898 if (likely(error <= interval))
899 adj = 1;
900 else
901 adj = timekeeping_bigadjust(error, &interval, &offset);
902 } else if (error < -interval) {
903 /* See comment above, this is just switched for the negative */
904 error >>= 2;
905 if (likely(error >= -interval)) {
906 adj = -1;
907 interval = -interval;
908 offset = -offset;
909 } else
910 adj = timekeeping_bigadjust(error, &interval, &offset);
911 } else /* No adjustment needed */
912 return;
913
914 if (unlikely(timekeeper.clock->maxadj &&
915 (timekeeper.mult + adj >
916 timekeeper.clock->mult + timekeeper.clock->maxadj))) {
917 printk_once(KERN_WARNING
918 "Adjusting %s more than 11%% (%ld vs %ld)\n",
919 timekeeper.clock->name, (long)timekeeper.mult + adj,
920 (long)timekeeper.clock->mult +
921 timekeeper.clock->maxadj);
922 }
923 /*
924 * So the following can be confusing.
925 *
926 * To keep things simple, lets assume adj == 1 for now.
927 *
928 * When adj != 1, remember that the interval and offset values
929 * have been appropriately scaled so the math is the same.
930 *
931 * The basic idea here is that we're increasing the multiplier
932 * by one, this causes the xtime_interval to be incremented by
933 * one cycle_interval. This is because:
934 * xtime_interval = cycle_interval * mult
935 * So if mult is being incremented by one:
936 * xtime_interval = cycle_interval * (mult + 1)
937 * Its the same as:
938 * xtime_interval = (cycle_interval * mult) + cycle_interval
939 * Which can be shortened to:
940 * xtime_interval += cycle_interval
941 *
942 * So offset stores the non-accumulated cycles. Thus the current
943 * time (in shifted nanoseconds) is:
944 * now = (offset * adj) + xtime_nsec
945 * Now, even though we're adjusting the clock frequency, we have
946 * to keep time consistent. In other words, we can't jump back
947 * in time, and we also want to avoid jumping forward in time.
948 *
949 * So given the same offset value, we need the time to be the same
950 * both before and after the freq adjustment.
951 * now = (offset * adj_1) + xtime_nsec_1
952 * now = (offset * adj_2) + xtime_nsec_2
953 * So:
954 * (offset * adj_1) + xtime_nsec_1 =
955 * (offset * adj_2) + xtime_nsec_2
956 * And we know:
957 * adj_2 = adj_1 + 1
958 * So:
959 * (offset * adj_1) + xtime_nsec_1 =
960 * (offset * (adj_1+1)) + xtime_nsec_2
961 * (offset * adj_1) + xtime_nsec_1 =
962 * (offset * adj_1) + offset + xtime_nsec_2
963 * Canceling the sides:
964 * xtime_nsec_1 = offset + xtime_nsec_2
965 * Which gives us:
966 * xtime_nsec_2 = xtime_nsec_1 - offset
967 * Which simplfies to:
968 * xtime_nsec -= offset
969 *
970 * XXX - TODO: Doc ntp_error calculation.
971 */
972 timekeeper.mult += adj;
973 timekeeper.xtime_interval += interval;
974 timekeeper.xtime_nsec -= offset;
975 timekeeper.ntp_error -= (interval - offset) <<
976 timekeeper.ntp_error_shift;
977}
978
979
980/**
981 * logarithmic_accumulation - shifted accumulation of cycles
982 *
983 * This functions accumulates a shifted interval of cycles into
984 * into a shifted interval nanoseconds. Allows for O(log) accumulation
985 * loop.
986 *
987 * Returns the unconsumed cycles.
988 */
989static cycle_t logarithmic_accumulation(cycle_t offset, int shift)
990{
991 u64 nsecps = (u64)NSEC_PER_SEC << timekeeper.shift;
992 u64 raw_nsecs;
993
994 /* If the offset is smaller than a shifted interval, do nothing */
995 if (offset < timekeeper.cycle_interval<<shift)
996 return offset;
997
998 /* Accumulate one shifted interval */
999 offset -= timekeeper.cycle_interval << shift;
1000 timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift;
1001
1002 timekeeper.xtime_nsec += timekeeper.xtime_interval << shift;
1003 while (timekeeper.xtime_nsec >= nsecps) {
1004 int leap;
1005 timekeeper.xtime_nsec -= nsecps;
1006 timekeeper.xtime.tv_sec++;
1007 leap = second_overflow(timekeeper.xtime.tv_sec);
1008 timekeeper.xtime.tv_sec += leap;
1009 timekeeper.wall_to_monotonic.tv_sec -= leap;
1010 if (leap)
1011 clock_was_set_delayed();
1012 }
1013
1014 /* Accumulate raw time */
1015 raw_nsecs = timekeeper.raw_interval << shift;
1016 raw_nsecs += timekeeper.raw_time.tv_nsec;
1017 if (raw_nsecs >= NSEC_PER_SEC) {
1018 u64 raw_secs = raw_nsecs;
1019 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1020 timekeeper.raw_time.tv_sec += raw_secs;
1021 }
1022 timekeeper.raw_time.tv_nsec = raw_nsecs;
1023
1024 /* Accumulate error between NTP and clock interval */
1025 timekeeper.ntp_error += ntp_tick_length() << shift;
1026 timekeeper.ntp_error -=
1027 (timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
1028 (timekeeper.ntp_error_shift + shift);
1029
1030 return offset;
1031}
1032
1033
1034/**
1035 * update_wall_time - Uses the current clocksource to increment the wall time
1036 *
1037 */
1038static void update_wall_time(void)
1039{
1040 struct clocksource *clock;
1041 cycle_t offset;
1042 int shift = 0, maxshift;
1043 unsigned long flags;
1044
1045 write_seqlock_irqsave(&timekeeper.lock, flags);
1046
1047 /* Make sure we're fully resumed: */
1048 if (unlikely(timekeeping_suspended))
1049 goto out;
1050
1051 clock = timekeeper.clock;
1052
1053#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1054 offset = timekeeper.cycle_interval;
1055#else
1056 offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
1057#endif
1058 /* Check if there's really nothing to do */
1059 if (offset < timekeeper.cycle_interval)
1060 goto out;
1061
1062 timekeeper.xtime_nsec = (s64)timekeeper.xtime.tv_nsec <<
1063 timekeeper.shift;
1064
1065 /*
1066 * With NO_HZ we may have to accumulate many cycle_intervals
1067 * (think "ticks") worth of time at once. To do this efficiently,
1068 * we calculate the largest doubling multiple of cycle_intervals
1069 * that is smaller than the offset. We then accumulate that
1070 * chunk in one go, and then try to consume the next smaller
1071 * doubled multiple.
1072 */
1073 shift = ilog2(offset) - ilog2(timekeeper.cycle_interval);
1074 shift = max(0, shift);
1075 /* Bound shift to one less than what overflows tick_length */
1076 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1077 shift = min(shift, maxshift);
1078 while (offset >= timekeeper.cycle_interval) {
1079 offset = logarithmic_accumulation(offset, shift);
1080 if(offset < timekeeper.cycle_interval<<shift)
1081 shift--;
1082 }
1083
1084 /* correct the clock when NTP error is too big */
1085 timekeeping_adjust(offset);
1086
1087 /*
1088 * Since in the loop above, we accumulate any amount of time
1089 * in xtime_nsec over a second into xtime.tv_sec, its possible for
1090 * xtime_nsec to be fairly small after the loop. Further, if we're
1091 * slightly speeding the clocksource up in timekeeping_adjust(),
1092 * its possible the required corrective factor to xtime_nsec could
1093 * cause it to underflow.
1094 *
1095 * Now, we cannot simply roll the accumulated second back, since
1096 * the NTP subsystem has been notified via second_overflow. So
1097 * instead we push xtime_nsec forward by the amount we underflowed,
1098 * and add that amount into the error.
1099 *
1100 * We'll correct this error next time through this function, when
1101 * xtime_nsec is not as small.
1102 */
1103 if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
1104 s64 neg = -(s64)timekeeper.xtime_nsec;
1105 timekeeper.xtime_nsec = 0;
1106 timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
1107 }
1108
1109
1110 /*
1111 * Store full nanoseconds into xtime after rounding it up and
1112 * add the remainder to the error difference.
1113 */
1114 timekeeper.xtime.tv_nsec = ((s64)timekeeper.xtime_nsec >>
1115 timekeeper.shift) + 1;
1116 timekeeper.xtime_nsec -= (s64)timekeeper.xtime.tv_nsec <<
1117 timekeeper.shift;
1118 timekeeper.ntp_error += timekeeper.xtime_nsec <<
1119 timekeeper.ntp_error_shift;
1120
1121 /*
1122 * Finally, make sure that after the rounding
1123 * xtime.tv_nsec isn't larger than NSEC_PER_SEC
1124 */
1125 if (unlikely(timekeeper.xtime.tv_nsec >= NSEC_PER_SEC)) {
1126 int leap;
1127 timekeeper.xtime.tv_nsec -= NSEC_PER_SEC;
1128 timekeeper.xtime.tv_sec++;
1129 leap = second_overflow(timekeeper.xtime.tv_sec);
1130 timekeeper.xtime.tv_sec += leap;
1131 timekeeper.wall_to_monotonic.tv_sec -= leap;
1132 if (leap)
1133 clock_was_set_delayed();
1134 }
1135
1136 timekeeping_update(false);
1137
1138out:
1139 write_sequnlock_irqrestore(&timekeeper.lock, flags);
1140
1141}
1142
1143/**
1144 * getboottime - Return the real time of system boot.
1145 * @ts: pointer to the timespec to be set
1146 *
1147 * Returns the wall-time of boot in a timespec.
1148 *
1149 * This is based on the wall_to_monotonic offset and the total suspend
1150 * time. Calls to settimeofday will affect the value returned (which
1151 * basically means that however wrong your real time clock is at boot time,
1152 * you get the right time here).
1153 */
1154void getboottime(struct timespec *ts)
1155{
1156 struct timespec boottime = {
1157 .tv_sec = timekeeper.wall_to_monotonic.tv_sec +
1158 timekeeper.total_sleep_time.tv_sec,
1159 .tv_nsec = timekeeper.wall_to_monotonic.tv_nsec +
1160 timekeeper.total_sleep_time.tv_nsec
1161 };
1162
1163 set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
1164}
1165EXPORT_SYMBOL_GPL(getboottime);
1166
1167
1168/**
1169 * get_monotonic_boottime - Returns monotonic time since boot
1170 * @ts: pointer to the timespec to be set
1171 *
1172 * Returns the monotonic time since boot in a timespec.
1173 *
1174 * This is similar to CLOCK_MONTONIC/ktime_get_ts, but also
1175 * includes the time spent in suspend.
1176 */
1177void get_monotonic_boottime(struct timespec *ts)
1178{
1179 struct timespec tomono, sleep;
1180 unsigned int seq;
1181 s64 nsecs;
1182
1183 WARN_ON(timekeeping_suspended);
1184
1185 do {
1186 seq = read_seqbegin(&timekeeper.lock);
1187 *ts = timekeeper.xtime;
1188 tomono = timekeeper.wall_to_monotonic;
1189 sleep = timekeeper.total_sleep_time;
1190 nsecs = timekeeping_get_ns();
1191
1192 } while (read_seqretry(&timekeeper.lock, seq));
1193
1194 set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec + sleep.tv_sec,
1195 ts->tv_nsec + tomono.tv_nsec + sleep.tv_nsec + nsecs);
1196}
1197EXPORT_SYMBOL_GPL(get_monotonic_boottime);
1198
1199/**
1200 * ktime_get_boottime - Returns monotonic time since boot in a ktime
1201 *
1202 * Returns the monotonic time since boot in a ktime
1203 *
1204 * This is similar to CLOCK_MONTONIC/ktime_get, but also
1205 * includes the time spent in suspend.
1206 */
1207ktime_t ktime_get_boottime(void)
1208{
1209 struct timespec ts;
1210
1211 get_monotonic_boottime(&ts);
1212 return timespec_to_ktime(ts);
1213}
1214EXPORT_SYMBOL_GPL(ktime_get_boottime);
1215
1216/**
1217 * monotonic_to_bootbased - Convert the monotonic time to boot based.
1218 * @ts: pointer to the timespec to be converted
1219 */
1220void monotonic_to_bootbased(struct timespec *ts)
1221{
1222 *ts = timespec_add(*ts, timekeeper.total_sleep_time);
1223}
1224EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
1225
1226unsigned long get_seconds(void)
1227{
1228 return timekeeper.xtime.tv_sec;
1229}
1230EXPORT_SYMBOL(get_seconds);
1231
1232struct timespec __current_kernel_time(void)
1233{
1234 return timekeeper.xtime;
1235}
1236
1237struct timespec current_kernel_time(void)
1238{
1239 struct timespec now;
1240 unsigned long seq;
1241
1242 do {
1243 seq = read_seqbegin(&timekeeper.lock);
1244
1245 now = timekeeper.xtime;
1246 } while (read_seqretry(&timekeeper.lock, seq));
1247
1248 return now;
1249}
1250EXPORT_SYMBOL(current_kernel_time);
1251
1252struct timespec get_monotonic_coarse(void)
1253{
1254 struct timespec now, mono;
1255 unsigned long seq;
1256
1257 do {
1258 seq = read_seqbegin(&timekeeper.lock);
1259
1260 now = timekeeper.xtime;
1261 mono = timekeeper.wall_to_monotonic;
1262 } while (read_seqretry(&timekeeper.lock, seq));
1263
1264 set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
1265 now.tv_nsec + mono.tv_nsec);
1266 return now;
1267}
1268
1269/*
1270 * The 64-bit jiffies value is not atomic - you MUST NOT read it
1271 * without sampling the sequence number in xtime_lock.
1272 * jiffies is defined in the linker script...
1273 */
1274void do_timer(unsigned long ticks)
1275{
1276 jiffies_64 += ticks;
1277 update_wall_time();
1278 calc_global_load(ticks);
1279}
1280
1281/**
1282 * get_xtime_and_monotonic_and_sleep_offset() - get xtime, wall_to_monotonic,
1283 * and sleep offsets.
1284 * @xtim: pointer to timespec to be set with xtime
1285 * @wtom: pointer to timespec to be set with wall_to_monotonic
1286 * @sleep: pointer to timespec to be set with time in suspend
1287 */
1288void get_xtime_and_monotonic_and_sleep_offset(struct timespec *xtim,
1289 struct timespec *wtom, struct timespec *sleep)
1290{
1291 unsigned long seq;
1292
1293 do {
1294 seq = read_seqbegin(&timekeeper.lock);
1295 *xtim = timekeeper.xtime;
1296 *wtom = timekeeper.wall_to_monotonic;
1297 *sleep = timekeeper.total_sleep_time;
1298 } while (read_seqretry(&timekeeper.lock, seq));
1299}
1300
1301#ifdef CONFIG_HIGH_RES_TIMERS
1302/**
1303 * ktime_get_update_offsets - hrtimer helper
1304 * @offs_real: pointer to storage for monotonic -> realtime offset
1305 * @offs_boot: pointer to storage for monotonic -> boottime offset
1306 *
1307 * Returns current monotonic time and updates the offsets
1308 * Called from hrtimer_interupt() or retrigger_next_event()
1309 */
1310ktime_t ktime_get_update_offsets(ktime_t *offs_real, ktime_t *offs_boot)
1311{
1312 ktime_t now;
1313 unsigned int seq;
1314 u64 secs, nsecs;
1315
1316 do {
1317 seq = read_seqbegin(&timekeeper.lock);
1318
1319 secs = timekeeper.xtime.tv_sec;
1320 nsecs = timekeeper.xtime.tv_nsec;
1321 nsecs += timekeeping_get_ns();
1322 /* If arch requires, add in gettimeoffset() */
1323 nsecs += arch_gettimeoffset();
1324
1325 *offs_real = timekeeper.offs_real;
1326 *offs_boot = timekeeper.offs_boot;
1327 } while (read_seqretry(&timekeeper.lock, seq));
1328
1329 now = ktime_add_ns(ktime_set(secs, 0), nsecs);
1330 now = ktime_sub(now, *offs_real);
1331 return now;
1332}
1333#endif
1334
1335/**
1336 * ktime_get_monotonic_offset() - get wall_to_monotonic in ktime_t format
1337 */
1338ktime_t ktime_get_monotonic_offset(void)
1339{
1340 unsigned long seq;
1341 struct timespec wtom;
1342
1343 do {
1344 seq = read_seqbegin(&timekeeper.lock);
1345 wtom = timekeeper.wall_to_monotonic;
1346 } while (read_seqretry(&timekeeper.lock, seq));
1347
1348 return timespec_to_ktime(wtom);
1349}
1350EXPORT_SYMBOL_GPL(ktime_get_monotonic_offset);
1351
1352
1353/**
1354 * xtime_update() - advances the timekeeping infrastructure
1355 * @ticks: number of ticks, that have elapsed since the last call.
1356 *
1357 * Must be called with interrupts disabled.
1358 */
1359void xtime_update(unsigned long ticks)
1360{
1361 write_seqlock(&xtime_lock);
1362 do_timer(ticks);
1363 write_sequnlock(&xtime_lock);
1364}
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/timekeeper_internal.h>
12#include <linux/module.h>
13#include <linux/interrupt.h>
14#include <linux/percpu.h>
15#include <linux/init.h>
16#include <linux/mm.h>
17#include <linux/sched.h>
18#include <linux/syscore_ops.h>
19#include <linux/clocksource.h>
20#include <linux/jiffies.h>
21#include <linux/time.h>
22#include <linux/tick.h>
23#include <linux/stop_machine.h>
24#include <linux/pvclock_gtod.h>
25#include <linux/compiler.h>
26
27#include "tick-internal.h"
28#include "ntp_internal.h"
29#include "timekeeping_internal.h"
30
31#define TK_CLEAR_NTP (1 << 0)
32#define TK_MIRROR (1 << 1)
33#define TK_CLOCK_WAS_SET (1 << 2)
34
35/*
36 * The most important data for readout fits into a single 64 byte
37 * cache line.
38 */
39static struct {
40 seqcount_t seq;
41 struct timekeeper timekeeper;
42} tk_core ____cacheline_aligned;
43
44static DEFINE_RAW_SPINLOCK(timekeeper_lock);
45static struct timekeeper shadow_timekeeper;
46
47/**
48 * struct tk_fast - NMI safe timekeeper
49 * @seq: Sequence counter for protecting updates. The lowest bit
50 * is the index for the tk_read_base array
51 * @base: tk_read_base array. Access is indexed by the lowest bit of
52 * @seq.
53 *
54 * See @update_fast_timekeeper() below.
55 */
56struct tk_fast {
57 seqcount_t seq;
58 struct tk_read_base base[2];
59};
60
61static struct tk_fast tk_fast_mono ____cacheline_aligned;
62static struct tk_fast tk_fast_raw ____cacheline_aligned;
63
64/* flag for if timekeeping is suspended */
65int __read_mostly timekeeping_suspended;
66
67static inline void tk_normalize_xtime(struct timekeeper *tk)
68{
69 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
70 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
71 tk->xtime_sec++;
72 }
73}
74
75static inline struct timespec64 tk_xtime(struct timekeeper *tk)
76{
77 struct timespec64 ts;
78
79 ts.tv_sec = tk->xtime_sec;
80 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
81 return ts;
82}
83
84static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
85{
86 tk->xtime_sec = ts->tv_sec;
87 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
88}
89
90static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
91{
92 tk->xtime_sec += ts->tv_sec;
93 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
94 tk_normalize_xtime(tk);
95}
96
97static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
98{
99 struct timespec64 tmp;
100
101 /*
102 * Verify consistency of: offset_real = -wall_to_monotonic
103 * before modifying anything
104 */
105 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
106 -tk->wall_to_monotonic.tv_nsec);
107 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
108 tk->wall_to_monotonic = wtm;
109 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
110 tk->offs_real = timespec64_to_ktime(tmp);
111 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
112}
113
114static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
115{
116 tk->offs_boot = ktime_add(tk->offs_boot, delta);
117}
118
119#ifdef CONFIG_DEBUG_TIMEKEEPING
120#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
121
122static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
123{
124
125 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
126 const char *name = tk->tkr_mono.clock->name;
127
128 if (offset > max_cycles) {
129 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
130 offset, name, max_cycles);
131 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
132 } else {
133 if (offset > (max_cycles >> 1)) {
134 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
135 offset, name, max_cycles >> 1);
136 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
137 }
138 }
139
140 if (tk->underflow_seen) {
141 if (jiffies - tk->last_warning > WARNING_FREQ) {
142 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
143 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
144 printk_deferred(" Your kernel is probably still fine.\n");
145 tk->last_warning = jiffies;
146 }
147 tk->underflow_seen = 0;
148 }
149
150 if (tk->overflow_seen) {
151 if (jiffies - tk->last_warning > WARNING_FREQ) {
152 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
153 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
154 printk_deferred(" Your kernel is probably still fine.\n");
155 tk->last_warning = jiffies;
156 }
157 tk->overflow_seen = 0;
158 }
159}
160
161static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
162{
163 struct timekeeper *tk = &tk_core.timekeeper;
164 u64 now, last, mask, max, delta;
165 unsigned int seq;
166
167 /*
168 * Since we're called holding a seqlock, the data may shift
169 * under us while we're doing the calculation. This can cause
170 * false positives, since we'd note a problem but throw the
171 * results away. So nest another seqlock here to atomically
172 * grab the points we are checking with.
173 */
174 do {
175 seq = read_seqcount_begin(&tk_core.seq);
176 now = tkr->read(tkr->clock);
177 last = tkr->cycle_last;
178 mask = tkr->mask;
179 max = tkr->clock->max_cycles;
180 } while (read_seqcount_retry(&tk_core.seq, seq));
181
182 delta = clocksource_delta(now, last, mask);
183
184 /*
185 * Try to catch underflows by checking if we are seeing small
186 * mask-relative negative values.
187 */
188 if (unlikely((~delta & mask) < (mask >> 3))) {
189 tk->underflow_seen = 1;
190 delta = 0;
191 }
192
193 /* Cap delta value to the max_cycles values to avoid mult overflows */
194 if (unlikely(delta > max)) {
195 tk->overflow_seen = 1;
196 delta = tkr->clock->max_cycles;
197 }
198
199 return delta;
200}
201#else
202static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
203{
204}
205static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
206{
207 u64 cycle_now, delta;
208
209 /* read clocksource */
210 cycle_now = tkr->read(tkr->clock);
211
212 /* calculate the delta since the last update_wall_time */
213 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
214
215 return delta;
216}
217#endif
218
219/**
220 * tk_setup_internals - Set up internals to use clocksource clock.
221 *
222 * @tk: The target timekeeper to setup.
223 * @clock: Pointer to clocksource.
224 *
225 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
226 * pair and interval request.
227 *
228 * Unless you're the timekeeping code, you should not be using this!
229 */
230static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
231{
232 u64 interval;
233 u64 tmp, ntpinterval;
234 struct clocksource *old_clock;
235
236 ++tk->cs_was_changed_seq;
237 old_clock = tk->tkr_mono.clock;
238 tk->tkr_mono.clock = clock;
239 tk->tkr_mono.read = clock->read;
240 tk->tkr_mono.mask = clock->mask;
241 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
242
243 tk->tkr_raw.clock = clock;
244 tk->tkr_raw.read = clock->read;
245 tk->tkr_raw.mask = clock->mask;
246 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
247
248 /* Do the ns -> cycle conversion first, using original mult */
249 tmp = NTP_INTERVAL_LENGTH;
250 tmp <<= clock->shift;
251 ntpinterval = tmp;
252 tmp += clock->mult/2;
253 do_div(tmp, clock->mult);
254 if (tmp == 0)
255 tmp = 1;
256
257 interval = (u64) tmp;
258 tk->cycle_interval = interval;
259
260 /* Go back from cycles -> shifted ns */
261 tk->xtime_interval = interval * clock->mult;
262 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
263 tk->raw_interval = (interval * clock->mult) >> clock->shift;
264
265 /* if changing clocks, convert xtime_nsec shift units */
266 if (old_clock) {
267 int shift_change = clock->shift - old_clock->shift;
268 if (shift_change < 0)
269 tk->tkr_mono.xtime_nsec >>= -shift_change;
270 else
271 tk->tkr_mono.xtime_nsec <<= shift_change;
272 }
273 tk->tkr_raw.xtime_nsec = 0;
274
275 tk->tkr_mono.shift = clock->shift;
276 tk->tkr_raw.shift = clock->shift;
277
278 tk->ntp_error = 0;
279 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
280 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
281
282 /*
283 * The timekeeper keeps its own mult values for the currently
284 * active clocksource. These value will be adjusted via NTP
285 * to counteract clock drifting.
286 */
287 tk->tkr_mono.mult = clock->mult;
288 tk->tkr_raw.mult = clock->mult;
289 tk->ntp_err_mult = 0;
290}
291
292/* Timekeeper helper functions. */
293
294#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
295static u32 default_arch_gettimeoffset(void) { return 0; }
296u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
297#else
298static inline u32 arch_gettimeoffset(void) { return 0; }
299#endif
300
301static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta)
302{
303 u64 nsec;
304
305 nsec = delta * tkr->mult + tkr->xtime_nsec;
306 nsec >>= tkr->shift;
307
308 /* If arch requires, add in get_arch_timeoffset() */
309 return nsec + arch_gettimeoffset();
310}
311
312static inline u64 timekeeping_get_ns(struct tk_read_base *tkr)
313{
314 u64 delta;
315
316 delta = timekeeping_get_delta(tkr);
317 return timekeeping_delta_to_ns(tkr, delta);
318}
319
320static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles)
321{
322 u64 delta;
323
324 /* calculate the delta since the last update_wall_time */
325 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
326 return timekeeping_delta_to_ns(tkr, delta);
327}
328
329/**
330 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
331 * @tkr: Timekeeping readout base from which we take the update
332 *
333 * We want to use this from any context including NMI and tracing /
334 * instrumenting the timekeeping code itself.
335 *
336 * Employ the latch technique; see @raw_write_seqcount_latch.
337 *
338 * So if a NMI hits the update of base[0] then it will use base[1]
339 * which is still consistent. In the worst case this can result is a
340 * slightly wrong timestamp (a few nanoseconds). See
341 * @ktime_get_mono_fast_ns.
342 */
343static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
344{
345 struct tk_read_base *base = tkf->base;
346
347 /* Force readers off to base[1] */
348 raw_write_seqcount_latch(&tkf->seq);
349
350 /* Update base[0] */
351 memcpy(base, tkr, sizeof(*base));
352
353 /* Force readers back to base[0] */
354 raw_write_seqcount_latch(&tkf->seq);
355
356 /* Update base[1] */
357 memcpy(base + 1, base, sizeof(*base));
358}
359
360/**
361 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
362 *
363 * This timestamp is not guaranteed to be monotonic across an update.
364 * The timestamp is calculated by:
365 *
366 * now = base_mono + clock_delta * slope
367 *
368 * So if the update lowers the slope, readers who are forced to the
369 * not yet updated second array are still using the old steeper slope.
370 *
371 * tmono
372 * ^
373 * | o n
374 * | o n
375 * | u
376 * | o
377 * |o
378 * |12345678---> reader order
379 *
380 * o = old slope
381 * u = update
382 * n = new slope
383 *
384 * So reader 6 will observe time going backwards versus reader 5.
385 *
386 * While other CPUs are likely to be able observe that, the only way
387 * for a CPU local observation is when an NMI hits in the middle of
388 * the update. Timestamps taken from that NMI context might be ahead
389 * of the following timestamps. Callers need to be aware of that and
390 * deal with it.
391 */
392static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
393{
394 struct tk_read_base *tkr;
395 unsigned int seq;
396 u64 now;
397
398 do {
399 seq = raw_read_seqcount_latch(&tkf->seq);
400 tkr = tkf->base + (seq & 0x01);
401 now = ktime_to_ns(tkr->base);
402
403 now += timekeeping_delta_to_ns(tkr,
404 clocksource_delta(
405 tkr->read(tkr->clock),
406 tkr->cycle_last,
407 tkr->mask));
408 } while (read_seqcount_retry(&tkf->seq, seq));
409
410 return now;
411}
412
413u64 ktime_get_mono_fast_ns(void)
414{
415 return __ktime_get_fast_ns(&tk_fast_mono);
416}
417EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
418
419u64 ktime_get_raw_fast_ns(void)
420{
421 return __ktime_get_fast_ns(&tk_fast_raw);
422}
423EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
424
425/**
426 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
427 *
428 * To keep it NMI safe since we're accessing from tracing, we're not using a
429 * separate timekeeper with updates to monotonic clock and boot offset
430 * protected with seqlocks. This has the following minor side effects:
431 *
432 * (1) Its possible that a timestamp be taken after the boot offset is updated
433 * but before the timekeeper is updated. If this happens, the new boot offset
434 * is added to the old timekeeping making the clock appear to update slightly
435 * earlier:
436 * CPU 0 CPU 1
437 * timekeeping_inject_sleeptime64()
438 * __timekeeping_inject_sleeptime(tk, delta);
439 * timestamp();
440 * timekeeping_update(tk, TK_CLEAR_NTP...);
441 *
442 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
443 * partially updated. Since the tk->offs_boot update is a rare event, this
444 * should be a rare occurrence which postprocessing should be able to handle.
445 */
446u64 notrace ktime_get_boot_fast_ns(void)
447{
448 struct timekeeper *tk = &tk_core.timekeeper;
449
450 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
451}
452EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
453
454/* Suspend-time cycles value for halted fast timekeeper. */
455static u64 cycles_at_suspend;
456
457static u64 dummy_clock_read(struct clocksource *cs)
458{
459 return cycles_at_suspend;
460}
461
462/**
463 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
464 * @tk: Timekeeper to snapshot.
465 *
466 * It generally is unsafe to access the clocksource after timekeeping has been
467 * suspended, so take a snapshot of the readout base of @tk and use it as the
468 * fast timekeeper's readout base while suspended. It will return the same
469 * number of cycles every time until timekeeping is resumed at which time the
470 * proper readout base for the fast timekeeper will be restored automatically.
471 */
472static void halt_fast_timekeeper(struct timekeeper *tk)
473{
474 static struct tk_read_base tkr_dummy;
475 struct tk_read_base *tkr = &tk->tkr_mono;
476
477 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
478 cycles_at_suspend = tkr->read(tkr->clock);
479 tkr_dummy.read = dummy_clock_read;
480 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
481
482 tkr = &tk->tkr_raw;
483 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
484 tkr_dummy.read = dummy_clock_read;
485 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
486}
487
488#ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
489
490static inline void update_vsyscall(struct timekeeper *tk)
491{
492 struct timespec xt, wm;
493
494 xt = timespec64_to_timespec(tk_xtime(tk));
495 wm = timespec64_to_timespec(tk->wall_to_monotonic);
496 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
497 tk->tkr_mono.cycle_last);
498}
499
500static inline void old_vsyscall_fixup(struct timekeeper *tk)
501{
502 s64 remainder;
503
504 /*
505 * Store only full nanoseconds into xtime_nsec after rounding
506 * it up and add the remainder to the error difference.
507 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
508 * by truncating the remainder in vsyscalls. However, it causes
509 * additional work to be done in timekeeping_adjust(). Once
510 * the vsyscall implementations are converted to use xtime_nsec
511 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
512 * users are removed, this can be killed.
513 */
514 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
515 if (remainder != 0) {
516 tk->tkr_mono.xtime_nsec -= remainder;
517 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
518 tk->ntp_error += remainder << tk->ntp_error_shift;
519 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
520 }
521}
522#else
523#define old_vsyscall_fixup(tk)
524#endif
525
526static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
527
528static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
529{
530 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
531}
532
533/**
534 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
535 */
536int pvclock_gtod_register_notifier(struct notifier_block *nb)
537{
538 struct timekeeper *tk = &tk_core.timekeeper;
539 unsigned long flags;
540 int ret;
541
542 raw_spin_lock_irqsave(&timekeeper_lock, flags);
543 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
544 update_pvclock_gtod(tk, true);
545 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
546
547 return ret;
548}
549EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
550
551/**
552 * pvclock_gtod_unregister_notifier - unregister a pvclock
553 * timedata update listener
554 */
555int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
556{
557 unsigned long flags;
558 int ret;
559
560 raw_spin_lock_irqsave(&timekeeper_lock, flags);
561 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
562 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
563
564 return ret;
565}
566EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
567
568/*
569 * tk_update_leap_state - helper to update the next_leap_ktime
570 */
571static inline void tk_update_leap_state(struct timekeeper *tk)
572{
573 tk->next_leap_ktime = ntp_get_next_leap();
574 if (tk->next_leap_ktime != KTIME_MAX)
575 /* Convert to monotonic time */
576 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
577}
578
579/*
580 * Update the ktime_t based scalar nsec members of the timekeeper
581 */
582static inline void tk_update_ktime_data(struct timekeeper *tk)
583{
584 u64 seconds;
585 u32 nsec;
586
587 /*
588 * The xtime based monotonic readout is:
589 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
590 * The ktime based monotonic readout is:
591 * nsec = base_mono + now();
592 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
593 */
594 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
595 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
596 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
597
598 /* Update the monotonic raw base */
599 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
600
601 /*
602 * The sum of the nanoseconds portions of xtime and
603 * wall_to_monotonic can be greater/equal one second. Take
604 * this into account before updating tk->ktime_sec.
605 */
606 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
607 if (nsec >= NSEC_PER_SEC)
608 seconds++;
609 tk->ktime_sec = seconds;
610}
611
612/* must hold timekeeper_lock */
613static void timekeeping_update(struct timekeeper *tk, unsigned int action)
614{
615 if (action & TK_CLEAR_NTP) {
616 tk->ntp_error = 0;
617 ntp_clear();
618 }
619
620 tk_update_leap_state(tk);
621 tk_update_ktime_data(tk);
622
623 update_vsyscall(tk);
624 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
625
626 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
627 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
628
629 if (action & TK_CLOCK_WAS_SET)
630 tk->clock_was_set_seq++;
631 /*
632 * The mirroring of the data to the shadow-timekeeper needs
633 * to happen last here to ensure we don't over-write the
634 * timekeeper structure on the next update with stale data
635 */
636 if (action & TK_MIRROR)
637 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
638 sizeof(tk_core.timekeeper));
639}
640
641/**
642 * timekeeping_forward_now - update clock to the current time
643 *
644 * Forward the current clock to update its state since the last call to
645 * update_wall_time(). This is useful before significant clock changes,
646 * as it avoids having to deal with this time offset explicitly.
647 */
648static void timekeeping_forward_now(struct timekeeper *tk)
649{
650 struct clocksource *clock = tk->tkr_mono.clock;
651 u64 cycle_now, delta;
652 u64 nsec;
653
654 cycle_now = tk->tkr_mono.read(clock);
655 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
656 tk->tkr_mono.cycle_last = cycle_now;
657 tk->tkr_raw.cycle_last = cycle_now;
658
659 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
660
661 /* If arch requires, add in get_arch_timeoffset() */
662 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
663
664 tk_normalize_xtime(tk);
665
666 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
667 timespec64_add_ns(&tk->raw_time, nsec);
668}
669
670/**
671 * __getnstimeofday64 - Returns the time of day in a timespec64.
672 * @ts: pointer to the timespec to be set
673 *
674 * Updates the time of day in the timespec.
675 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
676 */
677int __getnstimeofday64(struct timespec64 *ts)
678{
679 struct timekeeper *tk = &tk_core.timekeeper;
680 unsigned long seq;
681 u64 nsecs;
682
683 do {
684 seq = read_seqcount_begin(&tk_core.seq);
685
686 ts->tv_sec = tk->xtime_sec;
687 nsecs = timekeeping_get_ns(&tk->tkr_mono);
688
689 } while (read_seqcount_retry(&tk_core.seq, seq));
690
691 ts->tv_nsec = 0;
692 timespec64_add_ns(ts, nsecs);
693
694 /*
695 * Do not bail out early, in case there were callers still using
696 * the value, even in the face of the WARN_ON.
697 */
698 if (unlikely(timekeeping_suspended))
699 return -EAGAIN;
700 return 0;
701}
702EXPORT_SYMBOL(__getnstimeofday64);
703
704/**
705 * getnstimeofday64 - Returns the time of day in a timespec64.
706 * @ts: pointer to the timespec64 to be set
707 *
708 * Returns the time of day in a timespec64 (WARN if suspended).
709 */
710void getnstimeofday64(struct timespec64 *ts)
711{
712 WARN_ON(__getnstimeofday64(ts));
713}
714EXPORT_SYMBOL(getnstimeofday64);
715
716ktime_t ktime_get(void)
717{
718 struct timekeeper *tk = &tk_core.timekeeper;
719 unsigned int seq;
720 ktime_t base;
721 u64 nsecs;
722
723 WARN_ON(timekeeping_suspended);
724
725 do {
726 seq = read_seqcount_begin(&tk_core.seq);
727 base = tk->tkr_mono.base;
728 nsecs = timekeeping_get_ns(&tk->tkr_mono);
729
730 } while (read_seqcount_retry(&tk_core.seq, seq));
731
732 return ktime_add_ns(base, nsecs);
733}
734EXPORT_SYMBOL_GPL(ktime_get);
735
736u32 ktime_get_resolution_ns(void)
737{
738 struct timekeeper *tk = &tk_core.timekeeper;
739 unsigned int seq;
740 u32 nsecs;
741
742 WARN_ON(timekeeping_suspended);
743
744 do {
745 seq = read_seqcount_begin(&tk_core.seq);
746 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
747 } while (read_seqcount_retry(&tk_core.seq, seq));
748
749 return nsecs;
750}
751EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
752
753static ktime_t *offsets[TK_OFFS_MAX] = {
754 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
755 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
756 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
757};
758
759ktime_t ktime_get_with_offset(enum tk_offsets offs)
760{
761 struct timekeeper *tk = &tk_core.timekeeper;
762 unsigned int seq;
763 ktime_t base, *offset = offsets[offs];
764 u64 nsecs;
765
766 WARN_ON(timekeeping_suspended);
767
768 do {
769 seq = read_seqcount_begin(&tk_core.seq);
770 base = ktime_add(tk->tkr_mono.base, *offset);
771 nsecs = timekeeping_get_ns(&tk->tkr_mono);
772
773 } while (read_seqcount_retry(&tk_core.seq, seq));
774
775 return ktime_add_ns(base, nsecs);
776
777}
778EXPORT_SYMBOL_GPL(ktime_get_with_offset);
779
780/**
781 * ktime_mono_to_any() - convert mononotic time to any other time
782 * @tmono: time to convert.
783 * @offs: which offset to use
784 */
785ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
786{
787 ktime_t *offset = offsets[offs];
788 unsigned long seq;
789 ktime_t tconv;
790
791 do {
792 seq = read_seqcount_begin(&tk_core.seq);
793 tconv = ktime_add(tmono, *offset);
794 } while (read_seqcount_retry(&tk_core.seq, seq));
795
796 return tconv;
797}
798EXPORT_SYMBOL_GPL(ktime_mono_to_any);
799
800/**
801 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
802 */
803ktime_t ktime_get_raw(void)
804{
805 struct timekeeper *tk = &tk_core.timekeeper;
806 unsigned int seq;
807 ktime_t base;
808 u64 nsecs;
809
810 do {
811 seq = read_seqcount_begin(&tk_core.seq);
812 base = tk->tkr_raw.base;
813 nsecs = timekeeping_get_ns(&tk->tkr_raw);
814
815 } while (read_seqcount_retry(&tk_core.seq, seq));
816
817 return ktime_add_ns(base, nsecs);
818}
819EXPORT_SYMBOL_GPL(ktime_get_raw);
820
821/**
822 * ktime_get_ts64 - get the monotonic clock in timespec64 format
823 * @ts: pointer to timespec variable
824 *
825 * The function calculates the monotonic clock from the realtime
826 * clock and the wall_to_monotonic offset and stores the result
827 * in normalized timespec64 format in the variable pointed to by @ts.
828 */
829void ktime_get_ts64(struct timespec64 *ts)
830{
831 struct timekeeper *tk = &tk_core.timekeeper;
832 struct timespec64 tomono;
833 unsigned int seq;
834 u64 nsec;
835
836 WARN_ON(timekeeping_suspended);
837
838 do {
839 seq = read_seqcount_begin(&tk_core.seq);
840 ts->tv_sec = tk->xtime_sec;
841 nsec = timekeeping_get_ns(&tk->tkr_mono);
842 tomono = tk->wall_to_monotonic;
843
844 } while (read_seqcount_retry(&tk_core.seq, seq));
845
846 ts->tv_sec += tomono.tv_sec;
847 ts->tv_nsec = 0;
848 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
849}
850EXPORT_SYMBOL_GPL(ktime_get_ts64);
851
852/**
853 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
854 *
855 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
856 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
857 * works on both 32 and 64 bit systems. On 32 bit systems the readout
858 * covers ~136 years of uptime which should be enough to prevent
859 * premature wrap arounds.
860 */
861time64_t ktime_get_seconds(void)
862{
863 struct timekeeper *tk = &tk_core.timekeeper;
864
865 WARN_ON(timekeeping_suspended);
866 return tk->ktime_sec;
867}
868EXPORT_SYMBOL_GPL(ktime_get_seconds);
869
870/**
871 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
872 *
873 * Returns the wall clock seconds since 1970. This replaces the
874 * get_seconds() interface which is not y2038 safe on 32bit systems.
875 *
876 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
877 * 32bit systems the access must be protected with the sequence
878 * counter to provide "atomic" access to the 64bit tk->xtime_sec
879 * value.
880 */
881time64_t ktime_get_real_seconds(void)
882{
883 struct timekeeper *tk = &tk_core.timekeeper;
884 time64_t seconds;
885 unsigned int seq;
886
887 if (IS_ENABLED(CONFIG_64BIT))
888 return tk->xtime_sec;
889
890 do {
891 seq = read_seqcount_begin(&tk_core.seq);
892 seconds = tk->xtime_sec;
893
894 } while (read_seqcount_retry(&tk_core.seq, seq));
895
896 return seconds;
897}
898EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
899
900/**
901 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
902 * but without the sequence counter protect. This internal function
903 * is called just when timekeeping lock is already held.
904 */
905time64_t __ktime_get_real_seconds(void)
906{
907 struct timekeeper *tk = &tk_core.timekeeper;
908
909 return tk->xtime_sec;
910}
911
912/**
913 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
914 * @systime_snapshot: pointer to struct receiving the system time snapshot
915 */
916void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
917{
918 struct timekeeper *tk = &tk_core.timekeeper;
919 unsigned long seq;
920 ktime_t base_raw;
921 ktime_t base_real;
922 u64 nsec_raw;
923 u64 nsec_real;
924 u64 now;
925
926 WARN_ON_ONCE(timekeeping_suspended);
927
928 do {
929 seq = read_seqcount_begin(&tk_core.seq);
930
931 now = tk->tkr_mono.read(tk->tkr_mono.clock);
932 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
933 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
934 base_real = ktime_add(tk->tkr_mono.base,
935 tk_core.timekeeper.offs_real);
936 base_raw = tk->tkr_raw.base;
937 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
938 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
939 } while (read_seqcount_retry(&tk_core.seq, seq));
940
941 systime_snapshot->cycles = now;
942 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
943 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
944}
945EXPORT_SYMBOL_GPL(ktime_get_snapshot);
946
947/* Scale base by mult/div checking for overflow */
948static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
949{
950 u64 tmp, rem;
951
952 tmp = div64_u64_rem(*base, div, &rem);
953
954 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
955 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
956 return -EOVERFLOW;
957 tmp *= mult;
958 rem *= mult;
959
960 do_div(rem, div);
961 *base = tmp + rem;
962 return 0;
963}
964
965/**
966 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
967 * @history: Snapshot representing start of history
968 * @partial_history_cycles: Cycle offset into history (fractional part)
969 * @total_history_cycles: Total history length in cycles
970 * @discontinuity: True indicates clock was set on history period
971 * @ts: Cross timestamp that should be adjusted using
972 * partial/total ratio
973 *
974 * Helper function used by get_device_system_crosststamp() to correct the
975 * crosstimestamp corresponding to the start of the current interval to the
976 * system counter value (timestamp point) provided by the driver. The
977 * total_history_* quantities are the total history starting at the provided
978 * reference point and ending at the start of the current interval. The cycle
979 * count between the driver timestamp point and the start of the current
980 * interval is partial_history_cycles.
981 */
982static int adjust_historical_crosststamp(struct system_time_snapshot *history,
983 u64 partial_history_cycles,
984 u64 total_history_cycles,
985 bool discontinuity,
986 struct system_device_crosststamp *ts)
987{
988 struct timekeeper *tk = &tk_core.timekeeper;
989 u64 corr_raw, corr_real;
990 bool interp_forward;
991 int ret;
992
993 if (total_history_cycles == 0 || partial_history_cycles == 0)
994 return 0;
995
996 /* Interpolate shortest distance from beginning or end of history */
997 interp_forward = partial_history_cycles > total_history_cycles/2 ?
998 true : false;
999 partial_history_cycles = interp_forward ?
1000 total_history_cycles - partial_history_cycles :
1001 partial_history_cycles;
1002
1003 /*
1004 * Scale the monotonic raw time delta by:
1005 * partial_history_cycles / total_history_cycles
1006 */
1007 corr_raw = (u64)ktime_to_ns(
1008 ktime_sub(ts->sys_monoraw, history->raw));
1009 ret = scale64_check_overflow(partial_history_cycles,
1010 total_history_cycles, &corr_raw);
1011 if (ret)
1012 return ret;
1013
1014 /*
1015 * If there is a discontinuity in the history, scale monotonic raw
1016 * correction by:
1017 * mult(real)/mult(raw) yielding the realtime correction
1018 * Otherwise, calculate the realtime correction similar to monotonic
1019 * raw calculation
1020 */
1021 if (discontinuity) {
1022 corr_real = mul_u64_u32_div
1023 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1024 } else {
1025 corr_real = (u64)ktime_to_ns(
1026 ktime_sub(ts->sys_realtime, history->real));
1027 ret = scale64_check_overflow(partial_history_cycles,
1028 total_history_cycles, &corr_real);
1029 if (ret)
1030 return ret;
1031 }
1032
1033 /* Fixup monotonic raw and real time time values */
1034 if (interp_forward) {
1035 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1036 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1037 } else {
1038 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1039 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1040 }
1041
1042 return 0;
1043}
1044
1045/*
1046 * cycle_between - true if test occurs chronologically between before and after
1047 */
1048static bool cycle_between(u64 before, u64 test, u64 after)
1049{
1050 if (test > before && test < after)
1051 return true;
1052 if (test < before && before > after)
1053 return true;
1054 return false;
1055}
1056
1057/**
1058 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1059 * @get_time_fn: Callback to get simultaneous device time and
1060 * system counter from the device driver
1061 * @ctx: Context passed to get_time_fn()
1062 * @history_begin: Historical reference point used to interpolate system
1063 * time when counter provided by the driver is before the current interval
1064 * @xtstamp: Receives simultaneously captured system and device time
1065 *
1066 * Reads a timestamp from a device and correlates it to system time
1067 */
1068int get_device_system_crosststamp(int (*get_time_fn)
1069 (ktime_t *device_time,
1070 struct system_counterval_t *sys_counterval,
1071 void *ctx),
1072 void *ctx,
1073 struct system_time_snapshot *history_begin,
1074 struct system_device_crosststamp *xtstamp)
1075{
1076 struct system_counterval_t system_counterval;
1077 struct timekeeper *tk = &tk_core.timekeeper;
1078 u64 cycles, now, interval_start;
1079 unsigned int clock_was_set_seq = 0;
1080 ktime_t base_real, base_raw;
1081 u64 nsec_real, nsec_raw;
1082 u8 cs_was_changed_seq;
1083 unsigned long seq;
1084 bool do_interp;
1085 int ret;
1086
1087 do {
1088 seq = read_seqcount_begin(&tk_core.seq);
1089 /*
1090 * Try to synchronously capture device time and a system
1091 * counter value calling back into the device driver
1092 */
1093 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1094 if (ret)
1095 return ret;
1096
1097 /*
1098 * Verify that the clocksource associated with the captured
1099 * system counter value is the same as the currently installed
1100 * timekeeper clocksource
1101 */
1102 if (tk->tkr_mono.clock != system_counterval.cs)
1103 return -ENODEV;
1104 cycles = system_counterval.cycles;
1105
1106 /*
1107 * Check whether the system counter value provided by the
1108 * device driver is on the current timekeeping interval.
1109 */
1110 now = tk->tkr_mono.read(tk->tkr_mono.clock);
1111 interval_start = tk->tkr_mono.cycle_last;
1112 if (!cycle_between(interval_start, cycles, now)) {
1113 clock_was_set_seq = tk->clock_was_set_seq;
1114 cs_was_changed_seq = tk->cs_was_changed_seq;
1115 cycles = interval_start;
1116 do_interp = true;
1117 } else {
1118 do_interp = false;
1119 }
1120
1121 base_real = ktime_add(tk->tkr_mono.base,
1122 tk_core.timekeeper.offs_real);
1123 base_raw = tk->tkr_raw.base;
1124
1125 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1126 system_counterval.cycles);
1127 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1128 system_counterval.cycles);
1129 } while (read_seqcount_retry(&tk_core.seq, seq));
1130
1131 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1132 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1133
1134 /*
1135 * Interpolate if necessary, adjusting back from the start of the
1136 * current interval
1137 */
1138 if (do_interp) {
1139 u64 partial_history_cycles, total_history_cycles;
1140 bool discontinuity;
1141
1142 /*
1143 * Check that the counter value occurs after the provided
1144 * history reference and that the history doesn't cross a
1145 * clocksource change
1146 */
1147 if (!history_begin ||
1148 !cycle_between(history_begin->cycles,
1149 system_counterval.cycles, cycles) ||
1150 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1151 return -EINVAL;
1152 partial_history_cycles = cycles - system_counterval.cycles;
1153 total_history_cycles = cycles - history_begin->cycles;
1154 discontinuity =
1155 history_begin->clock_was_set_seq != clock_was_set_seq;
1156
1157 ret = adjust_historical_crosststamp(history_begin,
1158 partial_history_cycles,
1159 total_history_cycles,
1160 discontinuity, xtstamp);
1161 if (ret)
1162 return ret;
1163 }
1164
1165 return 0;
1166}
1167EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1168
1169/**
1170 * do_gettimeofday - Returns the time of day in a timeval
1171 * @tv: pointer to the timeval to be set
1172 *
1173 * NOTE: Users should be converted to using getnstimeofday()
1174 */
1175void do_gettimeofday(struct timeval *tv)
1176{
1177 struct timespec64 now;
1178
1179 getnstimeofday64(&now);
1180 tv->tv_sec = now.tv_sec;
1181 tv->tv_usec = now.tv_nsec/1000;
1182}
1183EXPORT_SYMBOL(do_gettimeofday);
1184
1185/**
1186 * do_settimeofday64 - Sets the time of day.
1187 * @ts: pointer to the timespec64 variable containing the new time
1188 *
1189 * Sets the time of day to the new time and update NTP and notify hrtimers
1190 */
1191int do_settimeofday64(const struct timespec64 *ts)
1192{
1193 struct timekeeper *tk = &tk_core.timekeeper;
1194 struct timespec64 ts_delta, xt;
1195 unsigned long flags;
1196 int ret = 0;
1197
1198 if (!timespec64_valid_strict(ts))
1199 return -EINVAL;
1200
1201 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1202 write_seqcount_begin(&tk_core.seq);
1203
1204 timekeeping_forward_now(tk);
1205
1206 xt = tk_xtime(tk);
1207 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1208 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1209
1210 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1211 ret = -EINVAL;
1212 goto out;
1213 }
1214
1215 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1216
1217 tk_set_xtime(tk, ts);
1218out:
1219 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1220
1221 write_seqcount_end(&tk_core.seq);
1222 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1223
1224 /* signal hrtimers about time change */
1225 clock_was_set();
1226
1227 return ret;
1228}
1229EXPORT_SYMBOL(do_settimeofday64);
1230
1231/**
1232 * timekeeping_inject_offset - Adds or subtracts from the current time.
1233 * @tv: pointer to the timespec variable containing the offset
1234 *
1235 * Adds or subtracts an offset value from the current time.
1236 */
1237int timekeeping_inject_offset(struct timespec *ts)
1238{
1239 struct timekeeper *tk = &tk_core.timekeeper;
1240 unsigned long flags;
1241 struct timespec64 ts64, tmp;
1242 int ret = 0;
1243
1244 if (!timespec_inject_offset_valid(ts))
1245 return -EINVAL;
1246
1247 ts64 = timespec_to_timespec64(*ts);
1248
1249 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1250 write_seqcount_begin(&tk_core.seq);
1251
1252 timekeeping_forward_now(tk);
1253
1254 /* Make sure the proposed value is valid */
1255 tmp = timespec64_add(tk_xtime(tk), ts64);
1256 if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1257 !timespec64_valid_strict(&tmp)) {
1258 ret = -EINVAL;
1259 goto error;
1260 }
1261
1262 tk_xtime_add(tk, &ts64);
1263 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1264
1265error: /* even if we error out, we forwarded the time, so call update */
1266 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1267
1268 write_seqcount_end(&tk_core.seq);
1269 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1270
1271 /* signal hrtimers about time change */
1272 clock_was_set();
1273
1274 return ret;
1275}
1276EXPORT_SYMBOL(timekeeping_inject_offset);
1277
1278
1279/**
1280 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1281 *
1282 */
1283s32 timekeeping_get_tai_offset(void)
1284{
1285 struct timekeeper *tk = &tk_core.timekeeper;
1286 unsigned int seq;
1287 s32 ret;
1288
1289 do {
1290 seq = read_seqcount_begin(&tk_core.seq);
1291 ret = tk->tai_offset;
1292 } while (read_seqcount_retry(&tk_core.seq, seq));
1293
1294 return ret;
1295}
1296
1297/**
1298 * __timekeeping_set_tai_offset - Lock free worker function
1299 *
1300 */
1301static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1302{
1303 tk->tai_offset = tai_offset;
1304 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1305}
1306
1307/**
1308 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1309 *
1310 */
1311void timekeeping_set_tai_offset(s32 tai_offset)
1312{
1313 struct timekeeper *tk = &tk_core.timekeeper;
1314 unsigned long flags;
1315
1316 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1317 write_seqcount_begin(&tk_core.seq);
1318 __timekeeping_set_tai_offset(tk, tai_offset);
1319 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1320 write_seqcount_end(&tk_core.seq);
1321 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1322 clock_was_set();
1323}
1324
1325/**
1326 * change_clocksource - Swaps clocksources if a new one is available
1327 *
1328 * Accumulates current time interval and initializes new clocksource
1329 */
1330static int change_clocksource(void *data)
1331{
1332 struct timekeeper *tk = &tk_core.timekeeper;
1333 struct clocksource *new, *old;
1334 unsigned long flags;
1335
1336 new = (struct clocksource *) data;
1337
1338 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1339 write_seqcount_begin(&tk_core.seq);
1340
1341 timekeeping_forward_now(tk);
1342 /*
1343 * If the cs is in module, get a module reference. Succeeds
1344 * for built-in code (owner == NULL) as well.
1345 */
1346 if (try_module_get(new->owner)) {
1347 if (!new->enable || new->enable(new) == 0) {
1348 old = tk->tkr_mono.clock;
1349 tk_setup_internals(tk, new);
1350 if (old->disable)
1351 old->disable(old);
1352 module_put(old->owner);
1353 } else {
1354 module_put(new->owner);
1355 }
1356 }
1357 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1358
1359 write_seqcount_end(&tk_core.seq);
1360 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1361
1362 return 0;
1363}
1364
1365/**
1366 * timekeeping_notify - Install a new clock source
1367 * @clock: pointer to the clock source
1368 *
1369 * This function is called from clocksource.c after a new, better clock
1370 * source has been registered. The caller holds the clocksource_mutex.
1371 */
1372int timekeeping_notify(struct clocksource *clock)
1373{
1374 struct timekeeper *tk = &tk_core.timekeeper;
1375
1376 if (tk->tkr_mono.clock == clock)
1377 return 0;
1378 stop_machine(change_clocksource, clock, NULL);
1379 tick_clock_notify();
1380 return tk->tkr_mono.clock == clock ? 0 : -1;
1381}
1382
1383/**
1384 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1385 * @ts: pointer to the timespec64 to be set
1386 *
1387 * Returns the raw monotonic time (completely un-modified by ntp)
1388 */
1389void getrawmonotonic64(struct timespec64 *ts)
1390{
1391 struct timekeeper *tk = &tk_core.timekeeper;
1392 struct timespec64 ts64;
1393 unsigned long seq;
1394 u64 nsecs;
1395
1396 do {
1397 seq = read_seqcount_begin(&tk_core.seq);
1398 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1399 ts64 = tk->raw_time;
1400
1401 } while (read_seqcount_retry(&tk_core.seq, seq));
1402
1403 timespec64_add_ns(&ts64, nsecs);
1404 *ts = ts64;
1405}
1406EXPORT_SYMBOL(getrawmonotonic64);
1407
1408
1409/**
1410 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1411 */
1412int timekeeping_valid_for_hres(void)
1413{
1414 struct timekeeper *tk = &tk_core.timekeeper;
1415 unsigned long seq;
1416 int ret;
1417
1418 do {
1419 seq = read_seqcount_begin(&tk_core.seq);
1420
1421 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1422
1423 } while (read_seqcount_retry(&tk_core.seq, seq));
1424
1425 return ret;
1426}
1427
1428/**
1429 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1430 */
1431u64 timekeeping_max_deferment(void)
1432{
1433 struct timekeeper *tk = &tk_core.timekeeper;
1434 unsigned long seq;
1435 u64 ret;
1436
1437 do {
1438 seq = read_seqcount_begin(&tk_core.seq);
1439
1440 ret = tk->tkr_mono.clock->max_idle_ns;
1441
1442 } while (read_seqcount_retry(&tk_core.seq, seq));
1443
1444 return ret;
1445}
1446
1447/**
1448 * read_persistent_clock - Return time from the persistent clock.
1449 *
1450 * Weak dummy function for arches that do not yet support it.
1451 * Reads the time from the battery backed persistent clock.
1452 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1453 *
1454 * XXX - Do be sure to remove it once all arches implement it.
1455 */
1456void __weak read_persistent_clock(struct timespec *ts)
1457{
1458 ts->tv_sec = 0;
1459 ts->tv_nsec = 0;
1460}
1461
1462void __weak read_persistent_clock64(struct timespec64 *ts64)
1463{
1464 struct timespec ts;
1465
1466 read_persistent_clock(&ts);
1467 *ts64 = timespec_to_timespec64(ts);
1468}
1469
1470/**
1471 * read_boot_clock64 - Return time of the system start.
1472 *
1473 * Weak dummy function for arches that do not yet support it.
1474 * Function to read the exact time the system has been started.
1475 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1476 *
1477 * XXX - Do be sure to remove it once all arches implement it.
1478 */
1479void __weak read_boot_clock64(struct timespec64 *ts)
1480{
1481 ts->tv_sec = 0;
1482 ts->tv_nsec = 0;
1483}
1484
1485/* Flag for if timekeeping_resume() has injected sleeptime */
1486static bool sleeptime_injected;
1487
1488/* Flag for if there is a persistent clock on this platform */
1489static bool persistent_clock_exists;
1490
1491/*
1492 * timekeeping_init - Initializes the clocksource and common timekeeping values
1493 */
1494void __init timekeeping_init(void)
1495{
1496 struct timekeeper *tk = &tk_core.timekeeper;
1497 struct clocksource *clock;
1498 unsigned long flags;
1499 struct timespec64 now, boot, tmp;
1500
1501 read_persistent_clock64(&now);
1502 if (!timespec64_valid_strict(&now)) {
1503 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1504 " Check your CMOS/BIOS settings.\n");
1505 now.tv_sec = 0;
1506 now.tv_nsec = 0;
1507 } else if (now.tv_sec || now.tv_nsec)
1508 persistent_clock_exists = true;
1509
1510 read_boot_clock64(&boot);
1511 if (!timespec64_valid_strict(&boot)) {
1512 pr_warn("WARNING: Boot clock returned invalid value!\n"
1513 " Check your CMOS/BIOS settings.\n");
1514 boot.tv_sec = 0;
1515 boot.tv_nsec = 0;
1516 }
1517
1518 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1519 write_seqcount_begin(&tk_core.seq);
1520 ntp_init();
1521
1522 clock = clocksource_default_clock();
1523 if (clock->enable)
1524 clock->enable(clock);
1525 tk_setup_internals(tk, clock);
1526
1527 tk_set_xtime(tk, &now);
1528 tk->raw_time.tv_sec = 0;
1529 tk->raw_time.tv_nsec = 0;
1530 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1531 boot = tk_xtime(tk);
1532
1533 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1534 tk_set_wall_to_mono(tk, tmp);
1535
1536 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1537
1538 write_seqcount_end(&tk_core.seq);
1539 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1540}
1541
1542/* time in seconds when suspend began for persistent clock */
1543static struct timespec64 timekeeping_suspend_time;
1544
1545/**
1546 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1547 * @delta: pointer to a timespec delta value
1548 *
1549 * Takes a timespec offset measuring a suspend interval and properly
1550 * adds the sleep offset to the timekeeping variables.
1551 */
1552static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1553 struct timespec64 *delta)
1554{
1555 if (!timespec64_valid_strict(delta)) {
1556 printk_deferred(KERN_WARNING
1557 "__timekeeping_inject_sleeptime: Invalid "
1558 "sleep delta value!\n");
1559 return;
1560 }
1561 tk_xtime_add(tk, delta);
1562 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1563 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1564 tk_debug_account_sleep_time(delta);
1565}
1566
1567#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1568/**
1569 * We have three kinds of time sources to use for sleep time
1570 * injection, the preference order is:
1571 * 1) non-stop clocksource
1572 * 2) persistent clock (ie: RTC accessible when irqs are off)
1573 * 3) RTC
1574 *
1575 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1576 * If system has neither 1) nor 2), 3) will be used finally.
1577 *
1578 *
1579 * If timekeeping has injected sleeptime via either 1) or 2),
1580 * 3) becomes needless, so in this case we don't need to call
1581 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1582 * means.
1583 */
1584bool timekeeping_rtc_skipresume(void)
1585{
1586 return sleeptime_injected;
1587}
1588
1589/**
1590 * 1) can be determined whether to use or not only when doing
1591 * timekeeping_resume() which is invoked after rtc_suspend(),
1592 * so we can't skip rtc_suspend() surely if system has 1).
1593 *
1594 * But if system has 2), 2) will definitely be used, so in this
1595 * case we don't need to call rtc_suspend(), and this is what
1596 * timekeeping_rtc_skipsuspend() means.
1597 */
1598bool timekeeping_rtc_skipsuspend(void)
1599{
1600 return persistent_clock_exists;
1601}
1602
1603/**
1604 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1605 * @delta: pointer to a timespec64 delta value
1606 *
1607 * This hook is for architectures that cannot support read_persistent_clock64
1608 * because their RTC/persistent clock is only accessible when irqs are enabled.
1609 * and also don't have an effective nonstop clocksource.
1610 *
1611 * This function should only be called by rtc_resume(), and allows
1612 * a suspend offset to be injected into the timekeeping values.
1613 */
1614void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1615{
1616 struct timekeeper *tk = &tk_core.timekeeper;
1617 unsigned long flags;
1618
1619 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1620 write_seqcount_begin(&tk_core.seq);
1621
1622 timekeeping_forward_now(tk);
1623
1624 __timekeeping_inject_sleeptime(tk, delta);
1625
1626 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1627
1628 write_seqcount_end(&tk_core.seq);
1629 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1630
1631 /* signal hrtimers about time change */
1632 clock_was_set();
1633}
1634#endif
1635
1636/**
1637 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1638 */
1639void timekeeping_resume(void)
1640{
1641 struct timekeeper *tk = &tk_core.timekeeper;
1642 struct clocksource *clock = tk->tkr_mono.clock;
1643 unsigned long flags;
1644 struct timespec64 ts_new, ts_delta;
1645 u64 cycle_now;
1646
1647 sleeptime_injected = false;
1648 read_persistent_clock64(&ts_new);
1649
1650 clockevents_resume();
1651 clocksource_resume();
1652
1653 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1654 write_seqcount_begin(&tk_core.seq);
1655
1656 /*
1657 * After system resumes, we need to calculate the suspended time and
1658 * compensate it for the OS time. There are 3 sources that could be
1659 * used: Nonstop clocksource during suspend, persistent clock and rtc
1660 * device.
1661 *
1662 * One specific platform may have 1 or 2 or all of them, and the
1663 * preference will be:
1664 * suspend-nonstop clocksource -> persistent clock -> rtc
1665 * The less preferred source will only be tried if there is no better
1666 * usable source. The rtc part is handled separately in rtc core code.
1667 */
1668 cycle_now = tk->tkr_mono.read(clock);
1669 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1670 cycle_now > tk->tkr_mono.cycle_last) {
1671 u64 nsec, cyc_delta;
1672
1673 cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1674 tk->tkr_mono.mask);
1675 nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift);
1676 ts_delta = ns_to_timespec64(nsec);
1677 sleeptime_injected = true;
1678 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1679 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1680 sleeptime_injected = true;
1681 }
1682
1683 if (sleeptime_injected)
1684 __timekeeping_inject_sleeptime(tk, &ts_delta);
1685
1686 /* Re-base the last cycle value */
1687 tk->tkr_mono.cycle_last = cycle_now;
1688 tk->tkr_raw.cycle_last = cycle_now;
1689
1690 tk->ntp_error = 0;
1691 timekeeping_suspended = 0;
1692 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1693 write_seqcount_end(&tk_core.seq);
1694 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1695
1696 touch_softlockup_watchdog();
1697
1698 tick_resume();
1699 hrtimers_resume();
1700}
1701
1702int timekeeping_suspend(void)
1703{
1704 struct timekeeper *tk = &tk_core.timekeeper;
1705 unsigned long flags;
1706 struct timespec64 delta, delta_delta;
1707 static struct timespec64 old_delta;
1708
1709 read_persistent_clock64(&timekeeping_suspend_time);
1710
1711 /*
1712 * On some systems the persistent_clock can not be detected at
1713 * timekeeping_init by its return value, so if we see a valid
1714 * value returned, update the persistent_clock_exists flag.
1715 */
1716 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1717 persistent_clock_exists = true;
1718
1719 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1720 write_seqcount_begin(&tk_core.seq);
1721 timekeeping_forward_now(tk);
1722 timekeeping_suspended = 1;
1723
1724 if (persistent_clock_exists) {
1725 /*
1726 * To avoid drift caused by repeated suspend/resumes,
1727 * which each can add ~1 second drift error,
1728 * try to compensate so the difference in system time
1729 * and persistent_clock time stays close to constant.
1730 */
1731 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1732 delta_delta = timespec64_sub(delta, old_delta);
1733 if (abs(delta_delta.tv_sec) >= 2) {
1734 /*
1735 * if delta_delta is too large, assume time correction
1736 * has occurred and set old_delta to the current delta.
1737 */
1738 old_delta = delta;
1739 } else {
1740 /* Otherwise try to adjust old_system to compensate */
1741 timekeeping_suspend_time =
1742 timespec64_add(timekeeping_suspend_time, delta_delta);
1743 }
1744 }
1745
1746 timekeeping_update(tk, TK_MIRROR);
1747 halt_fast_timekeeper(tk);
1748 write_seqcount_end(&tk_core.seq);
1749 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1750
1751 tick_suspend();
1752 clocksource_suspend();
1753 clockevents_suspend();
1754
1755 return 0;
1756}
1757
1758/* sysfs resume/suspend bits for timekeeping */
1759static struct syscore_ops timekeeping_syscore_ops = {
1760 .resume = timekeeping_resume,
1761 .suspend = timekeeping_suspend,
1762};
1763
1764static int __init timekeeping_init_ops(void)
1765{
1766 register_syscore_ops(&timekeeping_syscore_ops);
1767 return 0;
1768}
1769device_initcall(timekeeping_init_ops);
1770
1771/*
1772 * Apply a multiplier adjustment to the timekeeper
1773 */
1774static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1775 s64 offset,
1776 bool negative,
1777 int adj_scale)
1778{
1779 s64 interval = tk->cycle_interval;
1780 s32 mult_adj = 1;
1781
1782 if (negative) {
1783 mult_adj = -mult_adj;
1784 interval = -interval;
1785 offset = -offset;
1786 }
1787 mult_adj <<= adj_scale;
1788 interval <<= adj_scale;
1789 offset <<= adj_scale;
1790
1791 /*
1792 * So the following can be confusing.
1793 *
1794 * To keep things simple, lets assume mult_adj == 1 for now.
1795 *
1796 * When mult_adj != 1, remember that the interval and offset values
1797 * have been appropriately scaled so the math is the same.
1798 *
1799 * The basic idea here is that we're increasing the multiplier
1800 * by one, this causes the xtime_interval to be incremented by
1801 * one cycle_interval. This is because:
1802 * xtime_interval = cycle_interval * mult
1803 * So if mult is being incremented by one:
1804 * xtime_interval = cycle_interval * (mult + 1)
1805 * Its the same as:
1806 * xtime_interval = (cycle_interval * mult) + cycle_interval
1807 * Which can be shortened to:
1808 * xtime_interval += cycle_interval
1809 *
1810 * So offset stores the non-accumulated cycles. Thus the current
1811 * time (in shifted nanoseconds) is:
1812 * now = (offset * adj) + xtime_nsec
1813 * Now, even though we're adjusting the clock frequency, we have
1814 * to keep time consistent. In other words, we can't jump back
1815 * in time, and we also want to avoid jumping forward in time.
1816 *
1817 * So given the same offset value, we need the time to be the same
1818 * both before and after the freq adjustment.
1819 * now = (offset * adj_1) + xtime_nsec_1
1820 * now = (offset * adj_2) + xtime_nsec_2
1821 * So:
1822 * (offset * adj_1) + xtime_nsec_1 =
1823 * (offset * adj_2) + xtime_nsec_2
1824 * And we know:
1825 * adj_2 = adj_1 + 1
1826 * So:
1827 * (offset * adj_1) + xtime_nsec_1 =
1828 * (offset * (adj_1+1)) + xtime_nsec_2
1829 * (offset * adj_1) + xtime_nsec_1 =
1830 * (offset * adj_1) + offset + xtime_nsec_2
1831 * Canceling the sides:
1832 * xtime_nsec_1 = offset + xtime_nsec_2
1833 * Which gives us:
1834 * xtime_nsec_2 = xtime_nsec_1 - offset
1835 * Which simplfies to:
1836 * xtime_nsec -= offset
1837 *
1838 * XXX - TODO: Doc ntp_error calculation.
1839 */
1840 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1841 /* NTP adjustment caused clocksource mult overflow */
1842 WARN_ON_ONCE(1);
1843 return;
1844 }
1845
1846 tk->tkr_mono.mult += mult_adj;
1847 tk->xtime_interval += interval;
1848 tk->tkr_mono.xtime_nsec -= offset;
1849 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1850}
1851
1852/*
1853 * Calculate the multiplier adjustment needed to match the frequency
1854 * specified by NTP
1855 */
1856static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1857 s64 offset)
1858{
1859 s64 interval = tk->cycle_interval;
1860 s64 xinterval = tk->xtime_interval;
1861 u32 base = tk->tkr_mono.clock->mult;
1862 u32 max = tk->tkr_mono.clock->maxadj;
1863 u32 cur_adj = tk->tkr_mono.mult;
1864 s64 tick_error;
1865 bool negative;
1866 u32 adj_scale;
1867
1868 /* Remove any current error adj from freq calculation */
1869 if (tk->ntp_err_mult)
1870 xinterval -= tk->cycle_interval;
1871
1872 tk->ntp_tick = ntp_tick_length();
1873
1874 /* Calculate current error per tick */
1875 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1876 tick_error -= (xinterval + tk->xtime_remainder);
1877
1878 /* Don't worry about correcting it if its small */
1879 if (likely((tick_error >= 0) && (tick_error <= interval)))
1880 return;
1881
1882 /* preserve the direction of correction */
1883 negative = (tick_error < 0);
1884
1885 /* If any adjustment would pass the max, just return */
1886 if (negative && (cur_adj - 1) <= (base - max))
1887 return;
1888 if (!negative && (cur_adj + 1) >= (base + max))
1889 return;
1890 /*
1891 * Sort out the magnitude of the correction, but
1892 * avoid making so large a correction that we go
1893 * over the max adjustment.
1894 */
1895 adj_scale = 0;
1896 tick_error = abs(tick_error);
1897 while (tick_error > interval) {
1898 u32 adj = 1 << (adj_scale + 1);
1899
1900 /* Check if adjustment gets us within 1 unit from the max */
1901 if (negative && (cur_adj - adj) <= (base - max))
1902 break;
1903 if (!negative && (cur_adj + adj) >= (base + max))
1904 break;
1905
1906 adj_scale++;
1907 tick_error >>= 1;
1908 }
1909
1910 /* scale the corrections */
1911 timekeeping_apply_adjustment(tk, offset, negative, adj_scale);
1912}
1913
1914/*
1915 * Adjust the timekeeper's multiplier to the correct frequency
1916 * and also to reduce the accumulated error value.
1917 */
1918static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1919{
1920 /* Correct for the current frequency error */
1921 timekeeping_freqadjust(tk, offset);
1922
1923 /* Next make a small adjustment to fix any cumulative error */
1924 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1925 tk->ntp_err_mult = 1;
1926 timekeeping_apply_adjustment(tk, offset, 0, 0);
1927 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1928 /* Undo any existing error adjustment */
1929 timekeeping_apply_adjustment(tk, offset, 1, 0);
1930 tk->ntp_err_mult = 0;
1931 }
1932
1933 if (unlikely(tk->tkr_mono.clock->maxadj &&
1934 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1935 > tk->tkr_mono.clock->maxadj))) {
1936 printk_once(KERN_WARNING
1937 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1938 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1939 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1940 }
1941
1942 /*
1943 * It may be possible that when we entered this function, xtime_nsec
1944 * was very small. Further, if we're slightly speeding the clocksource
1945 * in the code above, its possible the required corrective factor to
1946 * xtime_nsec could cause it to underflow.
1947 *
1948 * Now, since we already accumulated the second, cannot simply roll
1949 * the accumulated second back, since the NTP subsystem has been
1950 * notified via second_overflow. So instead we push xtime_nsec forward
1951 * by the amount we underflowed, and add that amount into the error.
1952 *
1953 * We'll correct this error next time through this function, when
1954 * xtime_nsec is not as small.
1955 */
1956 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1957 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1958 tk->tkr_mono.xtime_nsec = 0;
1959 tk->ntp_error += neg << tk->ntp_error_shift;
1960 }
1961}
1962
1963/**
1964 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1965 *
1966 * Helper function that accumulates the nsecs greater than a second
1967 * from the xtime_nsec field to the xtime_secs field.
1968 * It also calls into the NTP code to handle leapsecond processing.
1969 *
1970 */
1971static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1972{
1973 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1974 unsigned int clock_set = 0;
1975
1976 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1977 int leap;
1978
1979 tk->tkr_mono.xtime_nsec -= nsecps;
1980 tk->xtime_sec++;
1981
1982 /* Figure out if its a leap sec and apply if needed */
1983 leap = second_overflow(tk->xtime_sec);
1984 if (unlikely(leap)) {
1985 struct timespec64 ts;
1986
1987 tk->xtime_sec += leap;
1988
1989 ts.tv_sec = leap;
1990 ts.tv_nsec = 0;
1991 tk_set_wall_to_mono(tk,
1992 timespec64_sub(tk->wall_to_monotonic, ts));
1993
1994 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1995
1996 clock_set = TK_CLOCK_WAS_SET;
1997 }
1998 }
1999 return clock_set;
2000}
2001
2002/**
2003 * logarithmic_accumulation - shifted accumulation of cycles
2004 *
2005 * This functions accumulates a shifted interval of cycles into
2006 * into a shifted interval nanoseconds. Allows for O(log) accumulation
2007 * loop.
2008 *
2009 * Returns the unconsumed cycles.
2010 */
2011static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2012 u32 shift, unsigned int *clock_set)
2013{
2014 u64 interval = tk->cycle_interval << shift;
2015 u64 raw_nsecs;
2016
2017 /* If the offset is smaller than a shifted interval, do nothing */
2018 if (offset < interval)
2019 return offset;
2020
2021 /* Accumulate one shifted interval */
2022 offset -= interval;
2023 tk->tkr_mono.cycle_last += interval;
2024 tk->tkr_raw.cycle_last += interval;
2025
2026 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2027 *clock_set |= accumulate_nsecs_to_secs(tk);
2028
2029 /* Accumulate raw time */
2030 raw_nsecs = (u64)tk->raw_interval << shift;
2031 raw_nsecs += tk->raw_time.tv_nsec;
2032 if (raw_nsecs >= NSEC_PER_SEC) {
2033 u64 raw_secs = raw_nsecs;
2034 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
2035 tk->raw_time.tv_sec += raw_secs;
2036 }
2037 tk->raw_time.tv_nsec = raw_nsecs;
2038
2039 /* Accumulate error between NTP and clock interval */
2040 tk->ntp_error += tk->ntp_tick << shift;
2041 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2042 (tk->ntp_error_shift + shift);
2043
2044 return offset;
2045}
2046
2047/**
2048 * update_wall_time - Uses the current clocksource to increment the wall time
2049 *
2050 */
2051void update_wall_time(void)
2052{
2053 struct timekeeper *real_tk = &tk_core.timekeeper;
2054 struct timekeeper *tk = &shadow_timekeeper;
2055 u64 offset;
2056 int shift = 0, maxshift;
2057 unsigned int clock_set = 0;
2058 unsigned long flags;
2059
2060 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2061
2062 /* Make sure we're fully resumed: */
2063 if (unlikely(timekeeping_suspended))
2064 goto out;
2065
2066#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2067 offset = real_tk->cycle_interval;
2068#else
2069 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
2070 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2071#endif
2072
2073 /* Check if there's really nothing to do */
2074 if (offset < real_tk->cycle_interval)
2075 goto out;
2076
2077 /* Do some additional sanity checking */
2078 timekeeping_check_update(real_tk, offset);
2079
2080 /*
2081 * With NO_HZ we may have to accumulate many cycle_intervals
2082 * (think "ticks") worth of time at once. To do this efficiently,
2083 * we calculate the largest doubling multiple of cycle_intervals
2084 * that is smaller than the offset. We then accumulate that
2085 * chunk in one go, and then try to consume the next smaller
2086 * doubled multiple.
2087 */
2088 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2089 shift = max(0, shift);
2090 /* Bound shift to one less than what overflows tick_length */
2091 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2092 shift = min(shift, maxshift);
2093 while (offset >= tk->cycle_interval) {
2094 offset = logarithmic_accumulation(tk, offset, shift,
2095 &clock_set);
2096 if (offset < tk->cycle_interval<<shift)
2097 shift--;
2098 }
2099
2100 /* correct the clock when NTP error is too big */
2101 timekeeping_adjust(tk, offset);
2102
2103 /*
2104 * XXX This can be killed once everyone converts
2105 * to the new update_vsyscall.
2106 */
2107 old_vsyscall_fixup(tk);
2108
2109 /*
2110 * Finally, make sure that after the rounding
2111 * xtime_nsec isn't larger than NSEC_PER_SEC
2112 */
2113 clock_set |= accumulate_nsecs_to_secs(tk);
2114
2115 write_seqcount_begin(&tk_core.seq);
2116 /*
2117 * Update the real timekeeper.
2118 *
2119 * We could avoid this memcpy by switching pointers, but that
2120 * requires changes to all other timekeeper usage sites as
2121 * well, i.e. move the timekeeper pointer getter into the
2122 * spinlocked/seqcount protected sections. And we trade this
2123 * memcpy under the tk_core.seq against one before we start
2124 * updating.
2125 */
2126 timekeeping_update(tk, clock_set);
2127 memcpy(real_tk, tk, sizeof(*tk));
2128 /* The memcpy must come last. Do not put anything here! */
2129 write_seqcount_end(&tk_core.seq);
2130out:
2131 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2132 if (clock_set)
2133 /* Have to call _delayed version, since in irq context*/
2134 clock_was_set_delayed();
2135}
2136
2137/**
2138 * getboottime64 - Return the real time of system boot.
2139 * @ts: pointer to the timespec64 to be set
2140 *
2141 * Returns the wall-time of boot in a timespec64.
2142 *
2143 * This is based on the wall_to_monotonic offset and the total suspend
2144 * time. Calls to settimeofday will affect the value returned (which
2145 * basically means that however wrong your real time clock is at boot time,
2146 * you get the right time here).
2147 */
2148void getboottime64(struct timespec64 *ts)
2149{
2150 struct timekeeper *tk = &tk_core.timekeeper;
2151 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2152
2153 *ts = ktime_to_timespec64(t);
2154}
2155EXPORT_SYMBOL_GPL(getboottime64);
2156
2157unsigned long get_seconds(void)
2158{
2159 struct timekeeper *tk = &tk_core.timekeeper;
2160
2161 return tk->xtime_sec;
2162}
2163EXPORT_SYMBOL(get_seconds);
2164
2165struct timespec __current_kernel_time(void)
2166{
2167 struct timekeeper *tk = &tk_core.timekeeper;
2168
2169 return timespec64_to_timespec(tk_xtime(tk));
2170}
2171
2172struct timespec64 current_kernel_time64(void)
2173{
2174 struct timekeeper *tk = &tk_core.timekeeper;
2175 struct timespec64 now;
2176 unsigned long seq;
2177
2178 do {
2179 seq = read_seqcount_begin(&tk_core.seq);
2180
2181 now = tk_xtime(tk);
2182 } while (read_seqcount_retry(&tk_core.seq, seq));
2183
2184 return now;
2185}
2186EXPORT_SYMBOL(current_kernel_time64);
2187
2188struct timespec64 get_monotonic_coarse64(void)
2189{
2190 struct timekeeper *tk = &tk_core.timekeeper;
2191 struct timespec64 now, mono;
2192 unsigned long seq;
2193
2194 do {
2195 seq = read_seqcount_begin(&tk_core.seq);
2196
2197 now = tk_xtime(tk);
2198 mono = tk->wall_to_monotonic;
2199 } while (read_seqcount_retry(&tk_core.seq, seq));
2200
2201 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2202 now.tv_nsec + mono.tv_nsec);
2203
2204 return now;
2205}
2206EXPORT_SYMBOL(get_monotonic_coarse64);
2207
2208/*
2209 * Must hold jiffies_lock
2210 */
2211void do_timer(unsigned long ticks)
2212{
2213 jiffies_64 += ticks;
2214 calc_global_load(ticks);
2215}
2216
2217/**
2218 * ktime_get_update_offsets_now - hrtimer helper
2219 * @cwsseq: pointer to check and store the clock was set sequence number
2220 * @offs_real: pointer to storage for monotonic -> realtime offset
2221 * @offs_boot: pointer to storage for monotonic -> boottime offset
2222 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2223 *
2224 * Returns current monotonic time and updates the offsets if the
2225 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2226 * different.
2227 *
2228 * Called from hrtimer_interrupt() or retrigger_next_event()
2229 */
2230ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2231 ktime_t *offs_boot, ktime_t *offs_tai)
2232{
2233 struct timekeeper *tk = &tk_core.timekeeper;
2234 unsigned int seq;
2235 ktime_t base;
2236 u64 nsecs;
2237
2238 do {
2239 seq = read_seqcount_begin(&tk_core.seq);
2240
2241 base = tk->tkr_mono.base;
2242 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2243 base = ktime_add_ns(base, nsecs);
2244
2245 if (*cwsseq != tk->clock_was_set_seq) {
2246 *cwsseq = tk->clock_was_set_seq;
2247 *offs_real = tk->offs_real;
2248 *offs_boot = tk->offs_boot;
2249 *offs_tai = tk->offs_tai;
2250 }
2251
2252 /* Handle leapsecond insertion adjustments */
2253 if (unlikely(base >= tk->next_leap_ktime))
2254 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2255
2256 } while (read_seqcount_retry(&tk_core.seq, seq));
2257
2258 return base;
2259}
2260
2261/**
2262 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2263 */
2264int do_adjtimex(struct timex *txc)
2265{
2266 struct timekeeper *tk = &tk_core.timekeeper;
2267 unsigned long flags;
2268 struct timespec64 ts;
2269 s32 orig_tai, tai;
2270 int ret;
2271
2272 /* Validate the data before disabling interrupts */
2273 ret = ntp_validate_timex(txc);
2274 if (ret)
2275 return ret;
2276
2277 if (txc->modes & ADJ_SETOFFSET) {
2278 struct timespec delta;
2279 delta.tv_sec = txc->time.tv_sec;
2280 delta.tv_nsec = txc->time.tv_usec;
2281 if (!(txc->modes & ADJ_NANO))
2282 delta.tv_nsec *= 1000;
2283 ret = timekeeping_inject_offset(&delta);
2284 if (ret)
2285 return ret;
2286 }
2287
2288 getnstimeofday64(&ts);
2289
2290 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2291 write_seqcount_begin(&tk_core.seq);
2292
2293 orig_tai = tai = tk->tai_offset;
2294 ret = __do_adjtimex(txc, &ts, &tai);
2295
2296 if (tai != orig_tai) {
2297 __timekeeping_set_tai_offset(tk, tai);
2298 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2299 }
2300 tk_update_leap_state(tk);
2301
2302 write_seqcount_end(&tk_core.seq);
2303 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2304
2305 if (tai != orig_tai)
2306 clock_was_set();
2307
2308 ntp_notify_cmos_timer();
2309
2310 return ret;
2311}
2312
2313#ifdef CONFIG_NTP_PPS
2314/**
2315 * hardpps() - Accessor function to NTP __hardpps function
2316 */
2317void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2318{
2319 unsigned long flags;
2320
2321 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2322 write_seqcount_begin(&tk_core.seq);
2323
2324 __hardpps(phase_ts, raw_ts);
2325
2326 write_seqcount_end(&tk_core.seq);
2327 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2328}
2329EXPORT_SYMBOL(hardpps);
2330#endif
2331
2332/**
2333 * xtime_update() - advances the timekeeping infrastructure
2334 * @ticks: number of ticks, that have elapsed since the last call.
2335 *
2336 * Must be called with interrupts disabled.
2337 */
2338void xtime_update(unsigned long ticks)
2339{
2340 write_seqlock(&jiffies_lock);
2341 do_timer(ticks);
2342 write_sequnlock(&jiffies_lock);
2343 update_wall_time();
2344}