projects/linux·blob v3.19 · refs

1884 lines · 49.6 KiB

1/*
2 * linux/kernel/time/timekeeping.c
3 *
4 * Kernel timekeeping code and accessor functions
5 *
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
8 *
9 */
10
11#include <linux/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;
62
63/* flag for if timekeeping is suspended */
64int __read_mostly timekeeping_suspended;
65
66/* Flag for if there is a persistent clock on this platform */
67bool __read_mostly persistent_clock_exist = false;
68
69static inline void tk_normalize_xtime(struct timekeeper *tk)
70{
71 while (tk->tkr.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr.shift)) {
72 tk->tkr.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr.shift;
73 tk->xtime_sec++;
74 }
75}
76
77static inline struct timespec64 tk_xtime(struct timekeeper *tk)
78{
79 struct timespec64 ts;
80
81 ts.tv_sec = tk->xtime_sec;
82 ts.tv_nsec = (long)(tk->tkr.xtime_nsec >> tk->tkr.shift);
83 return ts;
84}
85
86static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
87{
88 tk->xtime_sec = ts->tv_sec;
89 tk->tkr.xtime_nsec = (u64)ts->tv_nsec << tk->tkr.shift;
90}
91
92static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
93{
94 tk->xtime_sec += ts->tv_sec;
95 tk->tkr.xtime_nsec += (u64)ts->tv_nsec << tk->tkr.shift;
96 tk_normalize_xtime(tk);
97}
98
99static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
100{
101 struct timespec64 tmp;
102
103 /*
104 * Verify consistency of: offset_real = -wall_to_monotonic
105 * before modifying anything
106 */
107 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
108 -tk->wall_to_monotonic.tv_nsec);
109 WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
110 tk->wall_to_monotonic = wtm;
111 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
112 tk->offs_real = timespec64_to_ktime(tmp);
113 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
114}
115
116static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
117{
118 tk->offs_boot = ktime_add(tk->offs_boot, delta);
119}
120
121/**
122 * tk_setup_internals - Set up internals to use clocksource clock.
123 *
124 * @tk: The target timekeeper to setup.
125 * @clock: Pointer to clocksource.
126 *
127 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
128 * pair and interval request.
129 *
130 * Unless you're the timekeeping code, you should not be using this!
131 */
132static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
133{
134 cycle_t interval;
135 u64 tmp, ntpinterval;
136 struct clocksource *old_clock;
137
138 old_clock = tk->tkr.clock;
139 tk->tkr.clock = clock;
140 tk->tkr.read = clock->read;
141 tk->tkr.mask = clock->mask;
142 tk->tkr.cycle_last = tk->tkr.read(clock);
143
144 /* Do the ns -> cycle conversion first, using original mult */
145 tmp = NTP_INTERVAL_LENGTH;
146 tmp <<= clock->shift;
147 ntpinterval = tmp;
148 tmp += clock->mult/2;
149 do_div(tmp, clock->mult);
150 if (tmp == 0)
151 tmp = 1;
152
153 interval = (cycle_t) tmp;
154 tk->cycle_interval = interval;
155
156 /* Go back from cycles -> shifted ns */
157 tk->xtime_interval = (u64) interval * clock->mult;
158 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
159 tk->raw_interval =
160 ((u64) interval * clock->mult) >> clock->shift;
161
162 /* if changing clocks, convert xtime_nsec shift units */
163 if (old_clock) {
164 int shift_change = clock->shift - old_clock->shift;
165 if (shift_change < 0)
166 tk->tkr.xtime_nsec >>= -shift_change;
167 else
168 tk->tkr.xtime_nsec <<= shift_change;
169 }
170 tk->tkr.shift = clock->shift;
171
172 tk->ntp_error = 0;
173 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
174 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
175
176 /*
177 * The timekeeper keeps its own mult values for the currently
178 * active clocksource. These value will be adjusted via NTP
179 * to counteract clock drifting.
180 */
181 tk->tkr.mult = clock->mult;
182 tk->ntp_err_mult = 0;
183}
184
185/* Timekeeper helper functions. */
186
187#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
188static u32 default_arch_gettimeoffset(void) { return 0; }
189u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
190#else
191static inline u32 arch_gettimeoffset(void) { return 0; }
192#endif
193
194static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
195{
196 cycle_t cycle_now, delta;
197 s64 nsec;
198
199 /* read clocksource: */
200 cycle_now = tkr->read(tkr->clock);
201
202 /* calculate the delta since the last update_wall_time: */
203 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
204
205 nsec = delta * tkr->mult + tkr->xtime_nsec;
206 nsec >>= tkr->shift;
207
208 /* If arch requires, add in get_arch_timeoffset() */
209 return nsec + arch_gettimeoffset();
210}
211
212static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
213{
214 struct clocksource *clock = tk->tkr.clock;
215 cycle_t cycle_now, delta;
216 s64 nsec;
217
218 /* read clocksource: */
219 cycle_now = tk->tkr.read(clock);
220
221 /* calculate the delta since the last update_wall_time: */
222 delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
223
224 /* convert delta to nanoseconds. */
225 nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
226
227 /* If arch requires, add in get_arch_timeoffset() */
228 return nsec + arch_gettimeoffset();
229}
230
231/**
232 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
233 * @tk: The timekeeper from which we take the update
234 * @tkf: The fast timekeeper to update
235 * @tbase: The time base for the fast timekeeper (mono/raw)
236 *
237 * We want to use this from any context including NMI and tracing /
238 * instrumenting the timekeeping code itself.
239 *
240 * So we handle this differently than the other timekeeping accessor
241 * functions which retry when the sequence count has changed. The
242 * update side does:
243 *
244 * smp_wmb(); <- Ensure that the last base[1] update is visible
245 * tkf->seq++;
246 * smp_wmb(); <- Ensure that the seqcount update is visible
247 * update(tkf->base[0], tk);
248 * smp_wmb(); <- Ensure that the base[0] update is visible
249 * tkf->seq++;
250 * smp_wmb(); <- Ensure that the seqcount update is visible
251 * update(tkf->base[1], tk);
252 *
253 * The reader side does:
254 *
255 * do {
256 * seq = tkf->seq;
257 * smp_rmb();
258 * idx = seq & 0x01;
259 * now = now(tkf->base[idx]);
260 * smp_rmb();
261 * } while (seq != tkf->seq)
262 *
263 * As long as we update base[0] readers are forced off to
264 * base[1]. Once base[0] is updated readers are redirected to base[0]
265 * and the base[1] update takes place.
266 *
267 * So if a NMI hits the update of base[0] then it will use base[1]
268 * which is still consistent. In the worst case this can result is a
269 * slightly wrong timestamp (a few nanoseconds). See
270 * @ktime_get_mono_fast_ns.
271 */
272static void update_fast_timekeeper(struct timekeeper *tk)
273{
274 struct tk_read_base *base = tk_fast_mono.base;
275
276 /* Force readers off to base[1] */
277 raw_write_seqcount_latch(&tk_fast_mono.seq);
278
279 /* Update base[0] */
280 memcpy(base, &tk->tkr, sizeof(*base));
281
282 /* Force readers back to base[0] */
283 raw_write_seqcount_latch(&tk_fast_mono.seq);
284
285 /* Update base[1] */
286 memcpy(base + 1, base, sizeof(*base));
287}
288
289/**
290 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
291 *
292 * This timestamp is not guaranteed to be monotonic across an update.
293 * The timestamp is calculated by:
294 *
295 * now = base_mono + clock_delta * slope
296 *
297 * So if the update lowers the slope, readers who are forced to the
298 * not yet updated second array are still using the old steeper slope.
299 *
300 * tmono
301 * ^
302 * | o n
303 * | o n
304 * | u
305 * | o
306 * |o
307 * |12345678---> reader order
308 *
309 * o = old slope
310 * u = update
311 * n = new slope
312 *
313 * So reader 6 will observe time going backwards versus reader 5.
314 *
315 * While other CPUs are likely to be able observe that, the only way
316 * for a CPU local observation is when an NMI hits in the middle of
317 * the update. Timestamps taken from that NMI context might be ahead
318 * of the following timestamps. Callers need to be aware of that and
319 * deal with it.
320 */
321u64 notrace ktime_get_mono_fast_ns(void)
322{
323 struct tk_read_base *tkr;
324 unsigned int seq;
325 u64 now;
326
327 do {
328 seq = raw_read_seqcount(&tk_fast_mono.seq);
329 tkr = tk_fast_mono.base + (seq & 0x01);
330 now = ktime_to_ns(tkr->base_mono) + timekeeping_get_ns(tkr);
331
332 } while (read_seqcount_retry(&tk_fast_mono.seq, seq));
333 return now;
334}
335EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
336
337#ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
338
339static inline void update_vsyscall(struct timekeeper *tk)
340{
341 struct timespec xt, wm;
342
343 xt = timespec64_to_timespec(tk_xtime(tk));
344 wm = timespec64_to_timespec(tk->wall_to_monotonic);
345 update_vsyscall_old(&xt, &wm, tk->tkr.clock, tk->tkr.mult,
346 tk->tkr.cycle_last);
347}
348
349static inline void old_vsyscall_fixup(struct timekeeper *tk)
350{
351 s64 remainder;
352
353 /*
354 * Store only full nanoseconds into xtime_nsec after rounding
355 * it up and add the remainder to the error difference.
356 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
357 * by truncating the remainder in vsyscalls. However, it causes
358 * additional work to be done in timekeeping_adjust(). Once
359 * the vsyscall implementations are converted to use xtime_nsec
360 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
361 * users are removed, this can be killed.
362 */
363 remainder = tk->tkr.xtime_nsec & ((1ULL << tk->tkr.shift) - 1);
364 tk->tkr.xtime_nsec -= remainder;
365 tk->tkr.xtime_nsec += 1ULL << tk->tkr.shift;
366 tk->ntp_error += remainder << tk->ntp_error_shift;
367 tk->ntp_error -= (1ULL << tk->tkr.shift) << tk->ntp_error_shift;
368}
369#else
370#define old_vsyscall_fixup(tk)
371#endif
372
373static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
374
375static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
376{
377 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
378}
379
380/**
381 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
382 */
383int pvclock_gtod_register_notifier(struct notifier_block *nb)
384{
385 struct timekeeper *tk = &tk_core.timekeeper;
386 unsigned long flags;
387 int ret;
388
389 raw_spin_lock_irqsave(&timekeeper_lock, flags);
390 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
391 update_pvclock_gtod(tk, true);
392 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
393
394 return ret;
395}
396EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
397
398/**
399 * pvclock_gtod_unregister_notifier - unregister a pvclock
400 * timedata update listener
401 */
402int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
403{
404 unsigned long flags;
405 int ret;
406
407 raw_spin_lock_irqsave(&timekeeper_lock, flags);
408 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
409 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
410
411 return ret;
412}
413EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
414
415/*
416 * Update the ktime_t based scalar nsec members of the timekeeper
417 */
418static inline void tk_update_ktime_data(struct timekeeper *tk)
419{
420 u64 seconds;
421 u32 nsec;
422
423 /*
424 * The xtime based monotonic readout is:
425 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
426 * The ktime based monotonic readout is:
427 * nsec = base_mono + now();
428 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
429 */
430 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
431 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
432 tk->tkr.base_mono = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
433
434 /* Update the monotonic raw base */
435 tk->base_raw = timespec64_to_ktime(tk->raw_time);
436
437 /*
438 * The sum of the nanoseconds portions of xtime and
439 * wall_to_monotonic can be greater/equal one second. Take
440 * this into account before updating tk->ktime_sec.
441 */
442 nsec += (u32)(tk->tkr.xtime_nsec >> tk->tkr.shift);
443 if (nsec >= NSEC_PER_SEC)
444 seconds++;
445 tk->ktime_sec = seconds;
446}
447
448/* must hold timekeeper_lock */
449static void timekeeping_update(struct timekeeper *tk, unsigned int action)
450{
451 if (action & TK_CLEAR_NTP) {
452 tk->ntp_error = 0;
453 ntp_clear();
454 }
455
456 tk_update_ktime_data(tk);
457
458 update_vsyscall(tk);
459 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
460
461 if (action & TK_MIRROR)
462 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
463 sizeof(tk_core.timekeeper));
464
465 update_fast_timekeeper(tk);
466}
467
468/**
469 * timekeeping_forward_now - update clock to the current time
470 *
471 * Forward the current clock to update its state since the last call to
472 * update_wall_time(). This is useful before significant clock changes,
473 * as it avoids having to deal with this time offset explicitly.
474 */
475static void timekeeping_forward_now(struct timekeeper *tk)
476{
477 struct clocksource *clock = tk->tkr.clock;
478 cycle_t cycle_now, delta;
479 s64 nsec;
480
481 cycle_now = tk->tkr.read(clock);
482 delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
483 tk->tkr.cycle_last = cycle_now;
484
485 tk->tkr.xtime_nsec += delta * tk->tkr.mult;
486
487 /* If arch requires, add in get_arch_timeoffset() */
488 tk->tkr.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr.shift;
489
490 tk_normalize_xtime(tk);
491
492 nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
493 timespec64_add_ns(&tk->raw_time, nsec);
494}
495
496/**
497 * __getnstimeofday64 - Returns the time of day in a timespec64.
498 * @ts: pointer to the timespec to be set
499 *
500 * Updates the time of day in the timespec.
501 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
502 */
503int __getnstimeofday64(struct timespec64 *ts)
504{
505 struct timekeeper *tk = &tk_core.timekeeper;
506 unsigned long seq;
507 s64 nsecs = 0;
508
509 do {
510 seq = read_seqcount_begin(&tk_core.seq);
511
512 ts->tv_sec = tk->xtime_sec;
513 nsecs = timekeeping_get_ns(&tk->tkr);
514
515 } while (read_seqcount_retry(&tk_core.seq, seq));
516
517 ts->tv_nsec = 0;
518 timespec64_add_ns(ts, nsecs);
519
520 /*
521 * Do not bail out early, in case there were callers still using
522 * the value, even in the face of the WARN_ON.
523 */
524 if (unlikely(timekeeping_suspended))
525 return -EAGAIN;
526 return 0;
527}
528EXPORT_SYMBOL(__getnstimeofday64);
529
530/**
531 * getnstimeofday64 - Returns the time of day in a timespec64.
532 * @ts: pointer to the timespec64 to be set
533 *
534 * Returns the time of day in a timespec64 (WARN if suspended).
535 */
536void getnstimeofday64(struct timespec64 *ts)
537{
538 WARN_ON(__getnstimeofday64(ts));
539}
540EXPORT_SYMBOL(getnstimeofday64);
541
542ktime_t ktime_get(void)
543{
544 struct timekeeper *tk = &tk_core.timekeeper;
545 unsigned int seq;
546 ktime_t base;
547 s64 nsecs;
548
549 WARN_ON(timekeeping_suspended);
550
551 do {
552 seq = read_seqcount_begin(&tk_core.seq);
553 base = tk->tkr.base_mono;
554 nsecs = timekeeping_get_ns(&tk->tkr);
555
556 } while (read_seqcount_retry(&tk_core.seq, seq));
557
558 return ktime_add_ns(base, nsecs);
559}
560EXPORT_SYMBOL_GPL(ktime_get);
561
562static ktime_t *offsets[TK_OFFS_MAX] = {
563 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
564 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
565 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
566};
567
568ktime_t ktime_get_with_offset(enum tk_offsets offs)
569{
570 struct timekeeper *tk = &tk_core.timekeeper;
571 unsigned int seq;
572 ktime_t base, *offset = offsets[offs];
573 s64 nsecs;
574
575 WARN_ON(timekeeping_suspended);
576
577 do {
578 seq = read_seqcount_begin(&tk_core.seq);
579 base = ktime_add(tk->tkr.base_mono, *offset);
580 nsecs = timekeeping_get_ns(&tk->tkr);
581
582 } while (read_seqcount_retry(&tk_core.seq, seq));
583
584 return ktime_add_ns(base, nsecs);
585
586}
587EXPORT_SYMBOL_GPL(ktime_get_with_offset);
588
589/**
590 * ktime_mono_to_any() - convert mononotic time to any other time
591 * @tmono: time to convert.
592 * @offs: which offset to use
593 */
594ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
595{
596 ktime_t *offset = offsets[offs];
597 unsigned long seq;
598 ktime_t tconv;
599
600 do {
601 seq = read_seqcount_begin(&tk_core.seq);
602 tconv = ktime_add(tmono, *offset);
603 } while (read_seqcount_retry(&tk_core.seq, seq));
604
605 return tconv;
606}
607EXPORT_SYMBOL_GPL(ktime_mono_to_any);
608
609/**
610 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
611 */
612ktime_t ktime_get_raw(void)
613{
614 struct timekeeper *tk = &tk_core.timekeeper;
615 unsigned int seq;
616 ktime_t base;
617 s64 nsecs;
618
619 do {
620 seq = read_seqcount_begin(&tk_core.seq);
621 base = tk->base_raw;
622 nsecs = timekeeping_get_ns_raw(tk);
623
624 } while (read_seqcount_retry(&tk_core.seq, seq));
625
626 return ktime_add_ns(base, nsecs);
627}
628EXPORT_SYMBOL_GPL(ktime_get_raw);
629
630/**
631 * ktime_get_ts64 - get the monotonic clock in timespec64 format
632 * @ts: pointer to timespec variable
633 *
634 * The function calculates the monotonic clock from the realtime
635 * clock and the wall_to_monotonic offset and stores the result
636 * in normalized timespec64 format in the variable pointed to by @ts.
637 */
638void ktime_get_ts64(struct timespec64 *ts)
639{
640 struct timekeeper *tk = &tk_core.timekeeper;
641 struct timespec64 tomono;
642 s64 nsec;
643 unsigned int seq;
644
645 WARN_ON(timekeeping_suspended);
646
647 do {
648 seq = read_seqcount_begin(&tk_core.seq);
649 ts->tv_sec = tk->xtime_sec;
650 nsec = timekeeping_get_ns(&tk->tkr);
651 tomono = tk->wall_to_monotonic;
652
653 } while (read_seqcount_retry(&tk_core.seq, seq));
654
655 ts->tv_sec += tomono.tv_sec;
656 ts->tv_nsec = 0;
657 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
658}
659EXPORT_SYMBOL_GPL(ktime_get_ts64);
660
661/**
662 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
663 *
664 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
665 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
666 * works on both 32 and 64 bit systems. On 32 bit systems the readout
667 * covers ~136 years of uptime which should be enough to prevent
668 * premature wrap arounds.
669 */
670time64_t ktime_get_seconds(void)
671{
672 struct timekeeper *tk = &tk_core.timekeeper;
673
674 WARN_ON(timekeeping_suspended);
675 return tk->ktime_sec;
676}
677EXPORT_SYMBOL_GPL(ktime_get_seconds);
678
679/**
680 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
681 *
682 * Returns the wall clock seconds since 1970. This replaces the
683 * get_seconds() interface which is not y2038 safe on 32bit systems.
684 *
685 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
686 * 32bit systems the access must be protected with the sequence
687 * counter to provide "atomic" access to the 64bit tk->xtime_sec
688 * value.
689 */
690time64_t ktime_get_real_seconds(void)
691{
692 struct timekeeper *tk = &tk_core.timekeeper;
693 time64_t seconds;
694 unsigned int seq;
695
696 if (IS_ENABLED(CONFIG_64BIT))
697 return tk->xtime_sec;
698
699 do {
700 seq = read_seqcount_begin(&tk_core.seq);
701 seconds = tk->xtime_sec;
702
703 } while (read_seqcount_retry(&tk_core.seq, seq));
704
705 return seconds;
706}
707EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
708
709#ifdef CONFIG_NTP_PPS
710
711/**
712 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
713 * @ts_raw: pointer to the timespec to be set to raw monotonic time
714 * @ts_real: pointer to the timespec to be set to the time of day
715 *
716 * This function reads both the time of day and raw monotonic time at the
717 * same time atomically and stores the resulting timestamps in timespec
718 * format.
719 */
720void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
721{
722 struct timekeeper *tk = &tk_core.timekeeper;
723 unsigned long seq;
724 s64 nsecs_raw, nsecs_real;
725
726 WARN_ON_ONCE(timekeeping_suspended);
727
728 do {
729 seq = read_seqcount_begin(&tk_core.seq);
730
731 *ts_raw = timespec64_to_timespec(tk->raw_time);
732 ts_real->tv_sec = tk->xtime_sec;
733 ts_real->tv_nsec = 0;
734
735 nsecs_raw = timekeeping_get_ns_raw(tk);
736 nsecs_real = timekeeping_get_ns(&tk->tkr);
737
738 } while (read_seqcount_retry(&tk_core.seq, seq));
739
740 timespec_add_ns(ts_raw, nsecs_raw);
741 timespec_add_ns(ts_real, nsecs_real);
742}
743EXPORT_SYMBOL(getnstime_raw_and_real);
744
745#endif /* CONFIG_NTP_PPS */
746
747/**
748 * do_gettimeofday - Returns the time of day in a timeval
749 * @tv: pointer to the timeval to be set
750 *
751 * NOTE: Users should be converted to using getnstimeofday()
752 */
753void do_gettimeofday(struct timeval *tv)
754{
755 struct timespec64 now;
756
757 getnstimeofday64(&now);
758 tv->tv_sec = now.tv_sec;
759 tv->tv_usec = now.tv_nsec/1000;
760}
761EXPORT_SYMBOL(do_gettimeofday);
762
763/**
764 * do_settimeofday64 - Sets the time of day.
765 * @ts: pointer to the timespec64 variable containing the new time
766 *
767 * Sets the time of day to the new time and update NTP and notify hrtimers
768 */
769int do_settimeofday64(const struct timespec64 *ts)
770{
771 struct timekeeper *tk = &tk_core.timekeeper;
772 struct timespec64 ts_delta, xt;
773 unsigned long flags;
774
775 if (!timespec64_valid_strict(ts))
776 return -EINVAL;
777
778 raw_spin_lock_irqsave(&timekeeper_lock, flags);
779 write_seqcount_begin(&tk_core.seq);
780
781 timekeeping_forward_now(tk);
782
783 xt = tk_xtime(tk);
784 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
785 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
786
787 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
788
789 tk_set_xtime(tk, ts);
790
791 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
792
793 write_seqcount_end(&tk_core.seq);
794 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
795
796 /* signal hrtimers about time change */
797 clock_was_set();
798
799 return 0;
800}
801EXPORT_SYMBOL(do_settimeofday64);
802
803/**
804 * timekeeping_inject_offset - Adds or subtracts from the current time.
805 * @tv: pointer to the timespec variable containing the offset
806 *
807 * Adds or subtracts an offset value from the current time.
808 */
809int timekeeping_inject_offset(struct timespec *ts)
810{
811 struct timekeeper *tk = &tk_core.timekeeper;
812 unsigned long flags;
813 struct timespec64 ts64, tmp;
814 int ret = 0;
815
816 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
817 return -EINVAL;
818
819 ts64 = timespec_to_timespec64(*ts);
820
821 raw_spin_lock_irqsave(&timekeeper_lock, flags);
822 write_seqcount_begin(&tk_core.seq);
823
824 timekeeping_forward_now(tk);
825
826 /* Make sure the proposed value is valid */
827 tmp = timespec64_add(tk_xtime(tk), ts64);
828 if (!timespec64_valid_strict(&tmp)) {
829 ret = -EINVAL;
830 goto error;
831 }
832
833 tk_xtime_add(tk, &ts64);
834 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
835
836error: /* even if we error out, we forwarded the time, so call update */
837 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
838
839 write_seqcount_end(&tk_core.seq);
840 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
841
842 /* signal hrtimers about time change */
843 clock_was_set();
844
845 return ret;
846}
847EXPORT_SYMBOL(timekeeping_inject_offset);
848
849
850/**
851 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
852 *
853 */
854s32 timekeeping_get_tai_offset(void)
855{
856 struct timekeeper *tk = &tk_core.timekeeper;
857 unsigned int seq;
858 s32 ret;
859
860 do {
861 seq = read_seqcount_begin(&tk_core.seq);
862 ret = tk->tai_offset;
863 } while (read_seqcount_retry(&tk_core.seq, seq));
864
865 return ret;
866}
867
868/**
869 * __timekeeping_set_tai_offset - Lock free worker function
870 *
871 */
872static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
873{
874 tk->tai_offset = tai_offset;
875 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
876}
877
878/**
879 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
880 *
881 */
882void timekeeping_set_tai_offset(s32 tai_offset)
883{
884 struct timekeeper *tk = &tk_core.timekeeper;
885 unsigned long flags;
886
887 raw_spin_lock_irqsave(&timekeeper_lock, flags);
888 write_seqcount_begin(&tk_core.seq);
889 __timekeeping_set_tai_offset(tk, tai_offset);
890 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
891 write_seqcount_end(&tk_core.seq);
892 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
893 clock_was_set();
894}
895
896/**
897 * change_clocksource - Swaps clocksources if a new one is available
898 *
899 * Accumulates current time interval and initializes new clocksource
900 */
901static int change_clocksource(void *data)
902{
903 struct timekeeper *tk = &tk_core.timekeeper;
904 struct clocksource *new, *old;
905 unsigned long flags;
906
907 new = (struct clocksource *) data;
908
909 raw_spin_lock_irqsave(&timekeeper_lock, flags);
910 write_seqcount_begin(&tk_core.seq);
911
912 timekeeping_forward_now(tk);
913 /*
914 * If the cs is in module, get a module reference. Succeeds
915 * for built-in code (owner == NULL) as well.
916 */
917 if (try_module_get(new->owner)) {
918 if (!new->enable || new->enable(new) == 0) {
919 old = tk->tkr.clock;
920 tk_setup_internals(tk, new);
921 if (old->disable)
922 old->disable(old);
923 module_put(old->owner);
924 } else {
925 module_put(new->owner);
926 }
927 }
928 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
929
930 write_seqcount_end(&tk_core.seq);
931 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
932
933 return 0;
934}
935
936/**
937 * timekeeping_notify - Install a new clock source
938 * @clock: pointer to the clock source
939 *
940 * This function is called from clocksource.c after a new, better clock
941 * source has been registered. The caller holds the clocksource_mutex.
942 */
943int timekeeping_notify(struct clocksource *clock)
944{
945 struct timekeeper *tk = &tk_core.timekeeper;
946
947 if (tk->tkr.clock == clock)
948 return 0;
949 stop_machine(change_clocksource, clock, NULL);
950 tick_clock_notify();
951 return tk->tkr.clock == clock ? 0 : -1;
952}
953
954/**
955 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
956 * @ts: pointer to the timespec64 to be set
957 *
958 * Returns the raw monotonic time (completely un-modified by ntp)
959 */
960void getrawmonotonic64(struct timespec64 *ts)
961{
962 struct timekeeper *tk = &tk_core.timekeeper;
963 struct timespec64 ts64;
964 unsigned long seq;
965 s64 nsecs;
966
967 do {
968 seq = read_seqcount_begin(&tk_core.seq);
969 nsecs = timekeeping_get_ns_raw(tk);
970 ts64 = tk->raw_time;
971
972 } while (read_seqcount_retry(&tk_core.seq, seq));
973
974 timespec64_add_ns(&ts64, nsecs);
975 *ts = ts64;
976}
977EXPORT_SYMBOL(getrawmonotonic64);
978
979
980/**
981 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
982 */
983int timekeeping_valid_for_hres(void)
984{
985 struct timekeeper *tk = &tk_core.timekeeper;
986 unsigned long seq;
987 int ret;
988
989 do {
990 seq = read_seqcount_begin(&tk_core.seq);
991
992 ret = tk->tkr.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
993
994 } while (read_seqcount_retry(&tk_core.seq, seq));
995
996 return ret;
997}
998
999/**
1000 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1001 */
1002u64 timekeeping_max_deferment(void)
1003{
1004 struct timekeeper *tk = &tk_core.timekeeper;
1005 unsigned long seq;
1006 u64 ret;
1007
1008 do {
1009 seq = read_seqcount_begin(&tk_core.seq);
1010
1011 ret = tk->tkr.clock->max_idle_ns;
1012
1013 } while (read_seqcount_retry(&tk_core.seq, seq));
1014
1015 return ret;
1016}
1017
1018/**
1019 * read_persistent_clock - Return time from the persistent clock.
1020 *
1021 * Weak dummy function for arches that do not yet support it.
1022 * Reads the time from the battery backed persistent clock.
1023 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1024 *
1025 * XXX - Do be sure to remove it once all arches implement it.
1026 */
1027void __weak read_persistent_clock(struct timespec *ts)
1028{
1029 ts->tv_sec = 0;
1030 ts->tv_nsec = 0;
1031}
1032
1033/**
1034 * read_boot_clock - Return time of the system start.
1035 *
1036 * Weak dummy function for arches that do not yet support it.
1037 * Function to read the exact time the system has been started.
1038 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1039 *
1040 * XXX - Do be sure to remove it once all arches implement it.
1041 */
1042void __weak read_boot_clock(struct timespec *ts)
1043{
1044 ts->tv_sec = 0;
1045 ts->tv_nsec = 0;
1046}
1047
1048/*
1049 * timekeeping_init - Initializes the clocksource and common timekeeping values
1050 */
1051void __init timekeeping_init(void)
1052{
1053 struct timekeeper *tk = &tk_core.timekeeper;
1054 struct clocksource *clock;
1055 unsigned long flags;
1056 struct timespec64 now, boot, tmp;
1057 struct timespec ts;
1058
1059 read_persistent_clock(&ts);
1060 now = timespec_to_timespec64(ts);
1061 if (!timespec64_valid_strict(&now)) {
1062 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1063 " Check your CMOS/BIOS settings.\n");
1064 now.tv_sec = 0;
1065 now.tv_nsec = 0;
1066 } else if (now.tv_sec || now.tv_nsec)
1067 persistent_clock_exist = true;
1068
1069 read_boot_clock(&ts);
1070 boot = timespec_to_timespec64(ts);
1071 if (!timespec64_valid_strict(&boot)) {
1072 pr_warn("WARNING: Boot clock returned invalid value!\n"
1073 " Check your CMOS/BIOS settings.\n");
1074 boot.tv_sec = 0;
1075 boot.tv_nsec = 0;
1076 }
1077
1078 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1079 write_seqcount_begin(&tk_core.seq);
1080 ntp_init();
1081
1082 clock = clocksource_default_clock();
1083 if (clock->enable)
1084 clock->enable(clock);
1085 tk_setup_internals(tk, clock);
1086
1087 tk_set_xtime(tk, &now);
1088 tk->raw_time.tv_sec = 0;
1089 tk->raw_time.tv_nsec = 0;
1090 tk->base_raw.tv64 = 0;
1091 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1092 boot = tk_xtime(tk);
1093
1094 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1095 tk_set_wall_to_mono(tk, tmp);
1096
1097 timekeeping_update(tk, TK_MIRROR);
1098
1099 write_seqcount_end(&tk_core.seq);
1100 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1101}
1102
1103/* time in seconds when suspend began */
1104static struct timespec64 timekeeping_suspend_time;
1105
1106/**
1107 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1108 * @delta: pointer to a timespec delta value
1109 *
1110 * Takes a timespec offset measuring a suspend interval and properly
1111 * adds the sleep offset to the timekeeping variables.
1112 */
1113static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1114 struct timespec64 *delta)
1115{
1116 if (!timespec64_valid_strict(delta)) {
1117 printk_deferred(KERN_WARNING
1118 "__timekeeping_inject_sleeptime: Invalid "
1119 "sleep delta value!\n");
1120 return;
1121 }
1122 tk_xtime_add(tk, delta);
1123 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1124 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1125 tk_debug_account_sleep_time(delta);
1126}
1127
1128/**
1129 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1130 * @delta: pointer to a timespec64 delta value
1131 *
1132 * This hook is for architectures that cannot support read_persistent_clock
1133 * because their RTC/persistent clock is only accessible when irqs are enabled.
1134 *
1135 * This function should only be called by rtc_resume(), and allows
1136 * a suspend offset to be injected into the timekeeping values.
1137 */
1138void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1139{
1140 struct timekeeper *tk = &tk_core.timekeeper;
1141 unsigned long flags;
1142
1143 /*
1144 * Make sure we don't set the clock twice, as timekeeping_resume()
1145 * already did it
1146 */
1147 if (has_persistent_clock())
1148 return;
1149
1150 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1151 write_seqcount_begin(&tk_core.seq);
1152
1153 timekeeping_forward_now(tk);
1154
1155 __timekeeping_inject_sleeptime(tk, delta);
1156
1157 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1158
1159 write_seqcount_end(&tk_core.seq);
1160 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1161
1162 /* signal hrtimers about time change */
1163 clock_was_set();
1164}
1165
1166/**
1167 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1168 *
1169 * This is for the generic clocksource timekeeping.
1170 * xtime/wall_to_monotonic/jiffies/etc are
1171 * still managed by arch specific suspend/resume code.
1172 */
1173static void timekeeping_resume(void)
1174{
1175 struct timekeeper *tk = &tk_core.timekeeper;
1176 struct clocksource *clock = tk->tkr.clock;
1177 unsigned long flags;
1178 struct timespec64 ts_new, ts_delta;
1179 struct timespec tmp;
1180 cycle_t cycle_now, cycle_delta;
1181 bool suspendtime_found = false;
1182
1183 read_persistent_clock(&tmp);
1184 ts_new = timespec_to_timespec64(tmp);
1185
1186 clockevents_resume();
1187 clocksource_resume();
1188
1189 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1190 write_seqcount_begin(&tk_core.seq);
1191
1192 /*
1193 * After system resumes, we need to calculate the suspended time and
1194 * compensate it for the OS time. There are 3 sources that could be
1195 * used: Nonstop clocksource during suspend, persistent clock and rtc
1196 * device.
1197 *
1198 * One specific platform may have 1 or 2 or all of them, and the
1199 * preference will be:
1200 * suspend-nonstop clocksource -> persistent clock -> rtc
1201 * The less preferred source will only be tried if there is no better
1202 * usable source. The rtc part is handled separately in rtc core code.
1203 */
1204 cycle_now = tk->tkr.read(clock);
1205 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1206 cycle_now > tk->tkr.cycle_last) {
1207 u64 num, max = ULLONG_MAX;
1208 u32 mult = clock->mult;
1209 u32 shift = clock->shift;
1210 s64 nsec = 0;
1211
1212 cycle_delta = clocksource_delta(cycle_now, tk->tkr.cycle_last,
1213 tk->tkr.mask);
1214
1215 /*
1216 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1217 * suspended time is too long. In that case we need do the
1218 * 64 bits math carefully
1219 */
1220 do_div(max, mult);
1221 if (cycle_delta > max) {
1222 num = div64_u64(cycle_delta, max);
1223 nsec = (((u64) max * mult) >> shift) * num;
1224 cycle_delta -= num * max;
1225 }
1226 nsec += ((u64) cycle_delta * mult) >> shift;
1227
1228 ts_delta = ns_to_timespec64(nsec);
1229 suspendtime_found = true;
1230 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1231 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1232 suspendtime_found = true;
1233 }
1234
1235 if (suspendtime_found)
1236 __timekeeping_inject_sleeptime(tk, &ts_delta);
1237
1238 /* Re-base the last cycle value */
1239 tk->tkr.cycle_last = cycle_now;
1240 tk->ntp_error = 0;
1241 timekeeping_suspended = 0;
1242 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1243 write_seqcount_end(&tk_core.seq);
1244 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1245
1246 touch_softlockup_watchdog();
1247
1248 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
1249
1250 /* Resume hrtimers */
1251 hrtimers_resume();
1252}
1253
1254static int timekeeping_suspend(void)
1255{
1256 struct timekeeper *tk = &tk_core.timekeeper;
1257 unsigned long flags;
1258 struct timespec64 delta, delta_delta;
1259 static struct timespec64 old_delta;
1260 struct timespec tmp;
1261
1262 read_persistent_clock(&tmp);
1263 timekeeping_suspend_time = timespec_to_timespec64(tmp);
1264
1265 /*
1266 * On some systems the persistent_clock can not be detected at
1267 * timekeeping_init by its return value, so if we see a valid
1268 * value returned, update the persistent_clock_exists flag.
1269 */
1270 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1271 persistent_clock_exist = true;
1272
1273 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1274 write_seqcount_begin(&tk_core.seq);
1275 timekeeping_forward_now(tk);
1276 timekeeping_suspended = 1;
1277
1278 /*
1279 * To avoid drift caused by repeated suspend/resumes,
1280 * which each can add ~1 second drift error,
1281 * try to compensate so the difference in system time
1282 * and persistent_clock time stays close to constant.
1283 */
1284 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1285 delta_delta = timespec64_sub(delta, old_delta);
1286 if (abs(delta_delta.tv_sec) >= 2) {
1287 /*
1288 * if delta_delta is too large, assume time correction
1289 * has occured and set old_delta to the current delta.
1290 */
1291 old_delta = delta;
1292 } else {
1293 /* Otherwise try to adjust old_system to compensate */
1294 timekeeping_suspend_time =
1295 timespec64_add(timekeeping_suspend_time, delta_delta);
1296 }
1297
1298 timekeeping_update(tk, TK_MIRROR);
1299 write_seqcount_end(&tk_core.seq);
1300 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1301
1302 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
1303 clocksource_suspend();
1304 clockevents_suspend();
1305
1306 return 0;
1307}
1308
1309/* sysfs resume/suspend bits for timekeeping */
1310static struct syscore_ops timekeeping_syscore_ops = {
1311 .resume = timekeeping_resume,
1312 .suspend = timekeeping_suspend,
1313};
1314
1315static int __init timekeeping_init_ops(void)
1316{
1317 register_syscore_ops(&timekeeping_syscore_ops);
1318 return 0;
1319}
1320device_initcall(timekeeping_init_ops);
1321
1322/*
1323 * Apply a multiplier adjustment to the timekeeper
1324 */
1325static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1326 s64 offset,
1327 bool negative,
1328 int adj_scale)
1329{
1330 s64 interval = tk->cycle_interval;
1331 s32 mult_adj = 1;
1332
1333 if (negative) {
1334 mult_adj = -mult_adj;
1335 interval = -interval;
1336 offset = -offset;
1337 }
1338 mult_adj <<= adj_scale;
1339 interval <<= adj_scale;
1340 offset <<= adj_scale;
1341
1342 /*
1343 * So the following can be confusing.
1344 *
1345 * To keep things simple, lets assume mult_adj == 1 for now.
1346 *
1347 * When mult_adj != 1, remember that the interval and offset values
1348 * have been appropriately scaled so the math is the same.
1349 *
1350 * The basic idea here is that we're increasing the multiplier
1351 * by one, this causes the xtime_interval to be incremented by
1352 * one cycle_interval. This is because:
1353 * xtime_interval = cycle_interval * mult
1354 * So if mult is being incremented by one:
1355 * xtime_interval = cycle_interval * (mult + 1)
1356 * Its the same as:
1357 * xtime_interval = (cycle_interval * mult) + cycle_interval
1358 * Which can be shortened to:
1359 * xtime_interval += cycle_interval
1360 *
1361 * So offset stores the non-accumulated cycles. Thus the current
1362 * time (in shifted nanoseconds) is:
1363 * now = (offset * adj) + xtime_nsec
1364 * Now, even though we're adjusting the clock frequency, we have
1365 * to keep time consistent. In other words, we can't jump back
1366 * in time, and we also want to avoid jumping forward in time.
1367 *
1368 * So given the same offset value, we need the time to be the same
1369 * both before and after the freq adjustment.
1370 * now = (offset * adj_1) + xtime_nsec_1
1371 * now = (offset * adj_2) + xtime_nsec_2
1372 * So:
1373 * (offset * adj_1) + xtime_nsec_1 =
1374 * (offset * adj_2) + xtime_nsec_2
1375 * And we know:
1376 * adj_2 = adj_1 + 1
1377 * So:
1378 * (offset * adj_1) + xtime_nsec_1 =
1379 * (offset * (adj_1+1)) + xtime_nsec_2
1380 * (offset * adj_1) + xtime_nsec_1 =
1381 * (offset * adj_1) + offset + xtime_nsec_2
1382 * Canceling the sides:
1383 * xtime_nsec_1 = offset + xtime_nsec_2
1384 * Which gives us:
1385 * xtime_nsec_2 = xtime_nsec_1 - offset
1386 * Which simplfies to:
1387 * xtime_nsec -= offset
1388 *
1389 * XXX - TODO: Doc ntp_error calculation.
1390 */
1391 if ((mult_adj > 0) && (tk->tkr.mult + mult_adj < mult_adj)) {
1392 /* NTP adjustment caused clocksource mult overflow */
1393 WARN_ON_ONCE(1);
1394 return;
1395 }
1396
1397 tk->tkr.mult += mult_adj;
1398 tk->xtime_interval += interval;
1399 tk->tkr.xtime_nsec -= offset;
1400 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1401}
1402
1403/*
1404 * Calculate the multiplier adjustment needed to match the frequency
1405 * specified by NTP
1406 */
1407static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1408 s64 offset)
1409{
1410 s64 interval = tk->cycle_interval;
1411 s64 xinterval = tk->xtime_interval;
1412 s64 tick_error;
1413 bool negative;
1414 u32 adj;
1415
1416 /* Remove any current error adj from freq calculation */
1417 if (tk->ntp_err_mult)
1418 xinterval -= tk->cycle_interval;
1419
1420 tk->ntp_tick = ntp_tick_length();
1421
1422 /* Calculate current error per tick */
1423 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1424 tick_error -= (xinterval + tk->xtime_remainder);
1425
1426 /* Don't worry about correcting it if its small */
1427 if (likely((tick_error >= 0) && (tick_error <= interval)))
1428 return;
1429
1430 /* preserve the direction of correction */
1431 negative = (tick_error < 0);
1432
1433 /* Sort out the magnitude of the correction */
1434 tick_error = abs(tick_error);
1435 for (adj = 0; tick_error > interval; adj++)
1436 tick_error >>= 1;
1437
1438 /* scale the corrections */
1439 timekeeping_apply_adjustment(tk, offset, negative, adj);
1440}
1441
1442/*
1443 * Adjust the timekeeper's multiplier to the correct frequency
1444 * and also to reduce the accumulated error value.
1445 */
1446static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1447{
1448 /* Correct for the current frequency error */
1449 timekeeping_freqadjust(tk, offset);
1450
1451 /* Next make a small adjustment to fix any cumulative error */
1452 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1453 tk->ntp_err_mult = 1;
1454 timekeeping_apply_adjustment(tk, offset, 0, 0);
1455 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1456 /* Undo any existing error adjustment */
1457 timekeeping_apply_adjustment(tk, offset, 1, 0);
1458 tk->ntp_err_mult = 0;
1459 }
1460
1461 if (unlikely(tk->tkr.clock->maxadj &&
1462 (abs(tk->tkr.mult - tk->tkr.clock->mult)
1463 > tk->tkr.clock->maxadj))) {
1464 printk_once(KERN_WARNING
1465 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1466 tk->tkr.clock->name, (long)tk->tkr.mult,
1467 (long)tk->tkr.clock->mult + tk->tkr.clock->maxadj);
1468 }
1469
1470 /*
1471 * It may be possible that when we entered this function, xtime_nsec
1472 * was very small. Further, if we're slightly speeding the clocksource
1473 * in the code above, its possible the required corrective factor to
1474 * xtime_nsec could cause it to underflow.
1475 *
1476 * Now, since we already accumulated the second, cannot simply roll
1477 * the accumulated second back, since the NTP subsystem has been
1478 * notified via second_overflow. So instead we push xtime_nsec forward
1479 * by the amount we underflowed, and add that amount into the error.
1480 *
1481 * We'll correct this error next time through this function, when
1482 * xtime_nsec is not as small.
1483 */
1484 if (unlikely((s64)tk->tkr.xtime_nsec < 0)) {
1485 s64 neg = -(s64)tk->tkr.xtime_nsec;
1486 tk->tkr.xtime_nsec = 0;
1487 tk->ntp_error += neg << tk->ntp_error_shift;
1488 }
1489}
1490
1491/**
1492 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1493 *
1494 * Helper function that accumulates a the nsecs greater then a second
1495 * from the xtime_nsec field to the xtime_secs field.
1496 * It also calls into the NTP code to handle leapsecond processing.
1497 *
1498 */
1499static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1500{
1501 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr.shift;
1502 unsigned int clock_set = 0;
1503
1504 while (tk->tkr.xtime_nsec >= nsecps) {
1505 int leap;
1506
1507 tk->tkr.xtime_nsec -= nsecps;
1508 tk->xtime_sec++;
1509
1510 /* Figure out if its a leap sec and apply if needed */
1511 leap = second_overflow(tk->xtime_sec);
1512 if (unlikely(leap)) {
1513 struct timespec64 ts;
1514
1515 tk->xtime_sec += leap;
1516
1517 ts.tv_sec = leap;
1518 ts.tv_nsec = 0;
1519 tk_set_wall_to_mono(tk,
1520 timespec64_sub(tk->wall_to_monotonic, ts));
1521
1522 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1523
1524 clock_set = TK_CLOCK_WAS_SET;
1525 }
1526 }
1527 return clock_set;
1528}
1529
1530/**
1531 * logarithmic_accumulation - shifted accumulation of cycles
1532 *
1533 * This functions accumulates a shifted interval of cycles into
1534 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1535 * loop.
1536 *
1537 * Returns the unconsumed cycles.
1538 */
1539static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1540 u32 shift,
1541 unsigned int *clock_set)
1542{
1543 cycle_t interval = tk->cycle_interval << shift;
1544 u64 raw_nsecs;
1545
1546 /* If the offset is smaller then a shifted interval, do nothing */
1547 if (offset < interval)
1548 return offset;
1549
1550 /* Accumulate one shifted interval */
1551 offset -= interval;
1552 tk->tkr.cycle_last += interval;
1553
1554 tk->tkr.xtime_nsec += tk->xtime_interval << shift;
1555 *clock_set |= accumulate_nsecs_to_secs(tk);
1556
1557 /* Accumulate raw time */
1558 raw_nsecs = (u64)tk->raw_interval << shift;
1559 raw_nsecs += tk->raw_time.tv_nsec;
1560 if (raw_nsecs >= NSEC_PER_SEC) {
1561 u64 raw_secs = raw_nsecs;
1562 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1563 tk->raw_time.tv_sec += raw_secs;
1564 }
1565 tk->raw_time.tv_nsec = raw_nsecs;
1566
1567 /* Accumulate error between NTP and clock interval */
1568 tk->ntp_error += tk->ntp_tick << shift;
1569 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1570 (tk->ntp_error_shift + shift);
1571
1572 return offset;
1573}
1574
1575/**
1576 * update_wall_time - Uses the current clocksource to increment the wall time
1577 *
1578 */
1579void update_wall_time(void)
1580{
1581 struct timekeeper *real_tk = &tk_core.timekeeper;
1582 struct timekeeper *tk = &shadow_timekeeper;
1583 cycle_t offset;
1584 int shift = 0, maxshift;
1585 unsigned int clock_set = 0;
1586 unsigned long flags;
1587
1588 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1589
1590 /* Make sure we're fully resumed: */
1591 if (unlikely(timekeeping_suspended))
1592 goto out;
1593
1594#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1595 offset = real_tk->cycle_interval;
1596#else
1597 offset = clocksource_delta(tk->tkr.read(tk->tkr.clock),
1598 tk->tkr.cycle_last, tk->tkr.mask);
1599#endif
1600
1601 /* Check if there's really nothing to do */
1602 if (offset < real_tk->cycle_interval)
1603 goto out;
1604
1605 /*
1606 * With NO_HZ we may have to accumulate many cycle_intervals
1607 * (think "ticks") worth of time at once. To do this efficiently,
1608 * we calculate the largest doubling multiple of cycle_intervals
1609 * that is smaller than the offset. We then accumulate that
1610 * chunk in one go, and then try to consume the next smaller
1611 * doubled multiple.
1612 */
1613 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1614 shift = max(0, shift);
1615 /* Bound shift to one less than what overflows tick_length */
1616 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1617 shift = min(shift, maxshift);
1618 while (offset >= tk->cycle_interval) {
1619 offset = logarithmic_accumulation(tk, offset, shift,
1620 &clock_set);
1621 if (offset < tk->cycle_interval<<shift)
1622 shift--;
1623 }
1624
1625 /* correct the clock when NTP error is too big */
1626 timekeeping_adjust(tk, offset);
1627
1628 /*
1629 * XXX This can be killed once everyone converts
1630 * to the new update_vsyscall.
1631 */
1632 old_vsyscall_fixup(tk);
1633
1634 /*
1635 * Finally, make sure that after the rounding
1636 * xtime_nsec isn't larger than NSEC_PER_SEC
1637 */
1638 clock_set |= accumulate_nsecs_to_secs(tk);
1639
1640 write_seqcount_begin(&tk_core.seq);
1641 /*
1642 * Update the real timekeeper.
1643 *
1644 * We could avoid this memcpy by switching pointers, but that
1645 * requires changes to all other timekeeper usage sites as
1646 * well, i.e. move the timekeeper pointer getter into the
1647 * spinlocked/seqcount protected sections. And we trade this
1648 * memcpy under the tk_core.seq against one before we start
1649 * updating.
1650 */
1651 memcpy(real_tk, tk, sizeof(*tk));
1652 timekeeping_update(real_tk, clock_set);
1653 write_seqcount_end(&tk_core.seq);
1654out:
1655 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1656 if (clock_set)
1657 /* Have to call _delayed version, since in irq context*/
1658 clock_was_set_delayed();
1659}
1660
1661/**
1662 * getboottime - Return the real time of system boot.
1663 * @ts: pointer to the timespec to be set
1664 *
1665 * Returns the wall-time of boot in a timespec.
1666 *
1667 * This is based on the wall_to_monotonic offset and the total suspend
1668 * time. Calls to settimeofday will affect the value returned (which
1669 * basically means that however wrong your real time clock is at boot time,
1670 * you get the right time here).
1671 */
1672void getboottime(struct timespec *ts)
1673{
1674 struct timekeeper *tk = &tk_core.timekeeper;
1675 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1676
1677 *ts = ktime_to_timespec(t);
1678}
1679EXPORT_SYMBOL_GPL(getboottime);
1680
1681unsigned long get_seconds(void)
1682{
1683 struct timekeeper *tk = &tk_core.timekeeper;
1684
1685 return tk->xtime_sec;
1686}
1687EXPORT_SYMBOL(get_seconds);
1688
1689struct timespec __current_kernel_time(void)
1690{
1691 struct timekeeper *tk = &tk_core.timekeeper;
1692
1693 return timespec64_to_timespec(tk_xtime(tk));
1694}
1695
1696struct timespec current_kernel_time(void)
1697{
1698 struct timekeeper *tk = &tk_core.timekeeper;
1699 struct timespec64 now;
1700 unsigned long seq;
1701
1702 do {
1703 seq = read_seqcount_begin(&tk_core.seq);
1704
1705 now = tk_xtime(tk);
1706 } while (read_seqcount_retry(&tk_core.seq, seq));
1707
1708 return timespec64_to_timespec(now);
1709}
1710EXPORT_SYMBOL(current_kernel_time);
1711
1712struct timespec64 get_monotonic_coarse64(void)
1713{
1714 struct timekeeper *tk = &tk_core.timekeeper;
1715 struct timespec64 now, mono;
1716 unsigned long seq;
1717
1718 do {
1719 seq = read_seqcount_begin(&tk_core.seq);
1720
1721 now = tk_xtime(tk);
1722 mono = tk->wall_to_monotonic;
1723 } while (read_seqcount_retry(&tk_core.seq, seq));
1724
1725 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1726 now.tv_nsec + mono.tv_nsec);
1727
1728 return now;
1729}
1730
1731/*
1732 * Must hold jiffies_lock
1733 */
1734void do_timer(unsigned long ticks)
1735{
1736 jiffies_64 += ticks;
1737 calc_global_load(ticks);
1738}
1739
1740/**
1741 * ktime_get_update_offsets_tick - hrtimer helper
1742 * @offs_real: pointer to storage for monotonic -> realtime offset
1743 * @offs_boot: pointer to storage for monotonic -> boottime offset
1744 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1745 *
1746 * Returns monotonic time at last tick and various offsets
1747 */
1748ktime_t ktime_get_update_offsets_tick(ktime_t *offs_real, ktime_t *offs_boot,
1749 ktime_t *offs_tai)
1750{
1751 struct timekeeper *tk = &tk_core.timekeeper;
1752 unsigned int seq;
1753 ktime_t base;
1754 u64 nsecs;
1755
1756 do {
1757 seq = read_seqcount_begin(&tk_core.seq);
1758
1759 base = tk->tkr.base_mono;
1760 nsecs = tk->tkr.xtime_nsec >> tk->tkr.shift;
1761
1762 *offs_real = tk->offs_real;
1763 *offs_boot = tk->offs_boot;
1764 *offs_tai = tk->offs_tai;
1765 } while (read_seqcount_retry(&tk_core.seq, seq));
1766
1767 return ktime_add_ns(base, nsecs);
1768}
1769
1770#ifdef CONFIG_HIGH_RES_TIMERS
1771/**
1772 * ktime_get_update_offsets_now - hrtimer helper
1773 * @offs_real: pointer to storage for monotonic -> realtime offset
1774 * @offs_boot: pointer to storage for monotonic -> boottime offset
1775 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1776 *
1777 * Returns current monotonic time and updates the offsets
1778 * Called from hrtimer_interrupt() or retrigger_next_event()
1779 */
1780ktime_t ktime_get_update_offsets_now(ktime_t *offs_real, ktime_t *offs_boot,
1781 ktime_t *offs_tai)
1782{
1783 struct timekeeper *tk = &tk_core.timekeeper;
1784 unsigned int seq;
1785 ktime_t base;
1786 u64 nsecs;
1787
1788 do {
1789 seq = read_seqcount_begin(&tk_core.seq);
1790
1791 base = tk->tkr.base_mono;
1792 nsecs = timekeeping_get_ns(&tk->tkr);
1793
1794 *offs_real = tk->offs_real;
1795 *offs_boot = tk->offs_boot;
1796 *offs_tai = tk->offs_tai;
1797 } while (read_seqcount_retry(&tk_core.seq, seq));
1798
1799 return ktime_add_ns(base, nsecs);
1800}
1801#endif
1802
1803/**
1804 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1805 */
1806int do_adjtimex(struct timex *txc)
1807{
1808 struct timekeeper *tk = &tk_core.timekeeper;
1809 unsigned long flags;
1810 struct timespec64 ts;
1811 s32 orig_tai, tai;
1812 int ret;
1813
1814 /* Validate the data before disabling interrupts */
1815 ret = ntp_validate_timex(txc);
1816 if (ret)
1817 return ret;
1818
1819 if (txc->modes & ADJ_SETOFFSET) {
1820 struct timespec delta;
1821 delta.tv_sec = txc->time.tv_sec;
1822 delta.tv_nsec = txc->time.tv_usec;
1823 if (!(txc->modes & ADJ_NANO))
1824 delta.tv_nsec *= 1000;
1825 ret = timekeeping_inject_offset(&delta);
1826 if (ret)
1827 return ret;
1828 }
1829
1830 getnstimeofday64(&ts);
1831
1832 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1833 write_seqcount_begin(&tk_core.seq);
1834
1835 orig_tai = tai = tk->tai_offset;
1836 ret = __do_adjtimex(txc, &ts, &tai);
1837
1838 if (tai != orig_tai) {
1839 __timekeeping_set_tai_offset(tk, tai);
1840 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1841 }
1842 write_seqcount_end(&tk_core.seq);
1843 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1844
1845 if (tai != orig_tai)
1846 clock_was_set();
1847
1848 ntp_notify_cmos_timer();
1849
1850 return ret;
1851}
1852
1853#ifdef CONFIG_NTP_PPS
1854/**
1855 * hardpps() - Accessor function to NTP __hardpps function
1856 */
1857void hardpps(const struct timespec *phase_ts, const struct timespec *raw_ts)
1858{
1859 unsigned long flags;
1860
1861 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1862 write_seqcount_begin(&tk_core.seq);
1863
1864 __hardpps(phase_ts, raw_ts);
1865
1866 write_seqcount_end(&tk_core.seq);
1867 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1868}
1869EXPORT_SYMBOL(hardpps);
1870#endif
1871
1872/**
1873 * xtime_update() - advances the timekeeping infrastructure
1874 * @ticks: number of ticks, that have elapsed since the last call.
1875 *
1876 * Must be called with interrupts disabled.
1877 */
1878void xtime_update(unsigned long ticks)
1879{
1880 write_seqlock(&jiffies_lock);
1881 do_timer(ticks);
1882 write_sequnlock(&jiffies_lock);
1883 update_wall_time();
1884}