Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Kernel timekeeping code and accessor functions. Based on code from
4 * timer.c, moved in commit 8524070b7982.
5 */
6#include <linux/timekeeper_internal.h>
7#include <linux/module.h>
8#include <linux/interrupt.h>
9#include <linux/percpu.h>
10#include <linux/init.h>
11#include <linux/mm.h>
12#include <linux/nmi.h>
13#include <linux/sched.h>
14#include <linux/sched/loadavg.h>
15#include <linux/sched/clock.h>
16#include <linux/syscore_ops.h>
17#include <linux/clocksource.h>
18#include <linux/jiffies.h>
19#include <linux/time.h>
20#include <linux/tick.h>
21#include <linux/stop_machine.h>
22#include <linux/pvclock_gtod.h>
23#include <linux/compiler.h>
24#include <linux/audit.h>
25
26#include "tick-internal.h"
27#include "ntp_internal.h"
28#include "timekeeping_internal.h"
29
30#define TK_CLEAR_NTP (1 << 0)
31#define TK_MIRROR (1 << 1)
32#define TK_CLOCK_WAS_SET (1 << 2)
33
34enum timekeeping_adv_mode {
35 /* Update timekeeper when a tick has passed */
36 TK_ADV_TICK,
37
38 /* Update timekeeper on a direct frequency change */
39 TK_ADV_FREQ
40};
41
42DEFINE_RAW_SPINLOCK(timekeeper_lock);
43
44/*
45 * The most important data for readout fits into a single 64 byte
46 * cache line.
47 */
48static struct {
49 seqcount_raw_spinlock_t seq;
50 struct timekeeper timekeeper;
51} tk_core ____cacheline_aligned = {
52 .seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
53};
54
55static struct timekeeper shadow_timekeeper;
56
57/* flag for if timekeeping is suspended */
58int __read_mostly timekeeping_suspended;
59
60/**
61 * struct tk_fast - NMI safe timekeeper
62 * @seq: Sequence counter for protecting updates. The lowest bit
63 * is the index for the tk_read_base array
64 * @base: tk_read_base array. Access is indexed by the lowest bit of
65 * @seq.
66 *
67 * See @update_fast_timekeeper() below.
68 */
69struct tk_fast {
70 seqcount_latch_t seq;
71 struct tk_read_base base[2];
72};
73
74/* Suspend-time cycles value for halted fast timekeeper. */
75static u64 cycles_at_suspend;
76
77static u64 dummy_clock_read(struct clocksource *cs)
78{
79 if (timekeeping_suspended)
80 return cycles_at_suspend;
81 return local_clock();
82}
83
84static struct clocksource dummy_clock = {
85 .read = dummy_clock_read,
86};
87
88/*
89 * Boot time initialization which allows local_clock() to be utilized
90 * during early boot when clocksources are not available. local_clock()
91 * returns nanoseconds already so no conversion is required, hence mult=1
92 * and shift=0. When the first proper clocksource is installed then
93 * the fast time keepers are updated with the correct values.
94 */
95#define FAST_TK_INIT \
96 { \
97 .clock = &dummy_clock, \
98 .mask = CLOCKSOURCE_MASK(64), \
99 .mult = 1, \
100 .shift = 0, \
101 }
102
103static struct tk_fast tk_fast_mono ____cacheline_aligned = {
104 .seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
105 .base[0] = FAST_TK_INIT,
106 .base[1] = FAST_TK_INIT,
107};
108
109static struct tk_fast tk_fast_raw ____cacheline_aligned = {
110 .seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
111 .base[0] = FAST_TK_INIT,
112 .base[1] = FAST_TK_INIT,
113};
114
115static inline void tk_normalize_xtime(struct timekeeper *tk)
116{
117 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
118 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
119 tk->xtime_sec++;
120 }
121 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
122 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
123 tk->raw_sec++;
124 }
125}
126
127static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
128{
129 struct timespec64 ts;
130
131 ts.tv_sec = tk->xtime_sec;
132 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
133 return ts;
134}
135
136static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
137{
138 tk->xtime_sec = ts->tv_sec;
139 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
140}
141
142static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
143{
144 tk->xtime_sec += ts->tv_sec;
145 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
146 tk_normalize_xtime(tk);
147}
148
149static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
150{
151 struct timespec64 tmp;
152
153 /*
154 * Verify consistency of: offset_real = -wall_to_monotonic
155 * before modifying anything
156 */
157 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
158 -tk->wall_to_monotonic.tv_nsec);
159 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
160 tk->wall_to_monotonic = wtm;
161 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
162 tk->offs_real = timespec64_to_ktime(tmp);
163 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
164}
165
166static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
167{
168 tk->offs_boot = ktime_add(tk->offs_boot, delta);
169 /*
170 * Timespec representation for VDSO update to avoid 64bit division
171 * on every update.
172 */
173 tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
174}
175
176/*
177 * tk_clock_read - atomic clocksource read() helper
178 *
179 * This helper is necessary to use in the read paths because, while the
180 * seqcount ensures we don't return a bad value while structures are updated,
181 * it doesn't protect from potential crashes. There is the possibility that
182 * the tkr's clocksource may change between the read reference, and the
183 * clock reference passed to the read function. This can cause crashes if
184 * the wrong clocksource is passed to the wrong read function.
185 * This isn't necessary to use when holding the timekeeper_lock or doing
186 * a read of the fast-timekeeper tkrs (which is protected by its own locking
187 * and update logic).
188 */
189static inline u64 tk_clock_read(const struct tk_read_base *tkr)
190{
191 struct clocksource *clock = READ_ONCE(tkr->clock);
192
193 return clock->read(clock);
194}
195
196#ifdef CONFIG_DEBUG_TIMEKEEPING
197#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
198
199static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
200{
201
202 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
203 const char *name = tk->tkr_mono.clock->name;
204
205 if (offset > max_cycles) {
206 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
207 offset, name, max_cycles);
208 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
209 } else {
210 if (offset > (max_cycles >> 1)) {
211 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
212 offset, name, max_cycles >> 1);
213 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
214 }
215 }
216
217 if (tk->underflow_seen) {
218 if (jiffies - tk->last_warning > WARNING_FREQ) {
219 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
220 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
221 printk_deferred(" Your kernel is probably still fine.\n");
222 tk->last_warning = jiffies;
223 }
224 tk->underflow_seen = 0;
225 }
226
227 if (tk->overflow_seen) {
228 if (jiffies - tk->last_warning > WARNING_FREQ) {
229 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
230 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
231 printk_deferred(" Your kernel is probably still fine.\n");
232 tk->last_warning = jiffies;
233 }
234 tk->overflow_seen = 0;
235 }
236}
237
238static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
239{
240 struct timekeeper *tk = &tk_core.timekeeper;
241 u64 now, last, mask, max, delta;
242 unsigned int seq;
243
244 /*
245 * Since we're called holding a seqcount, the data may shift
246 * under us while we're doing the calculation. This can cause
247 * false positives, since we'd note a problem but throw the
248 * results away. So nest another seqcount here to atomically
249 * grab the points we are checking with.
250 */
251 do {
252 seq = read_seqcount_begin(&tk_core.seq);
253 now = tk_clock_read(tkr);
254 last = tkr->cycle_last;
255 mask = tkr->mask;
256 max = tkr->clock->max_cycles;
257 } while (read_seqcount_retry(&tk_core.seq, seq));
258
259 delta = clocksource_delta(now, last, mask);
260
261 /*
262 * Try to catch underflows by checking if we are seeing small
263 * mask-relative negative values.
264 */
265 if (unlikely((~delta & mask) < (mask >> 3))) {
266 tk->underflow_seen = 1;
267 delta = 0;
268 }
269
270 /* Cap delta value to the max_cycles values to avoid mult overflows */
271 if (unlikely(delta > max)) {
272 tk->overflow_seen = 1;
273 delta = tkr->clock->max_cycles;
274 }
275
276 return delta;
277}
278#else
279static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
280{
281}
282static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
283{
284 u64 cycle_now, delta;
285
286 /* read clocksource */
287 cycle_now = tk_clock_read(tkr);
288
289 /* calculate the delta since the last update_wall_time */
290 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
291
292 return delta;
293}
294#endif
295
296/**
297 * tk_setup_internals - Set up internals to use clocksource clock.
298 *
299 * @tk: The target timekeeper to setup.
300 * @clock: Pointer to clocksource.
301 *
302 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
303 * pair and interval request.
304 *
305 * Unless you're the timekeeping code, you should not be using this!
306 */
307static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
308{
309 u64 interval;
310 u64 tmp, ntpinterval;
311 struct clocksource *old_clock;
312
313 ++tk->cs_was_changed_seq;
314 old_clock = tk->tkr_mono.clock;
315 tk->tkr_mono.clock = clock;
316 tk->tkr_mono.mask = clock->mask;
317 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
318
319 tk->tkr_raw.clock = clock;
320 tk->tkr_raw.mask = clock->mask;
321 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
322
323 /* Do the ns -> cycle conversion first, using original mult */
324 tmp = NTP_INTERVAL_LENGTH;
325 tmp <<= clock->shift;
326 ntpinterval = tmp;
327 tmp += clock->mult/2;
328 do_div(tmp, clock->mult);
329 if (tmp == 0)
330 tmp = 1;
331
332 interval = (u64) tmp;
333 tk->cycle_interval = interval;
334
335 /* Go back from cycles -> shifted ns */
336 tk->xtime_interval = interval * clock->mult;
337 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
338 tk->raw_interval = interval * clock->mult;
339
340 /* if changing clocks, convert xtime_nsec shift units */
341 if (old_clock) {
342 int shift_change = clock->shift - old_clock->shift;
343 if (shift_change < 0) {
344 tk->tkr_mono.xtime_nsec >>= -shift_change;
345 tk->tkr_raw.xtime_nsec >>= -shift_change;
346 } else {
347 tk->tkr_mono.xtime_nsec <<= shift_change;
348 tk->tkr_raw.xtime_nsec <<= shift_change;
349 }
350 }
351
352 tk->tkr_mono.shift = clock->shift;
353 tk->tkr_raw.shift = clock->shift;
354
355 tk->ntp_error = 0;
356 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
357 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
358
359 /*
360 * The timekeeper keeps its own mult values for the currently
361 * active clocksource. These value will be adjusted via NTP
362 * to counteract clock drifting.
363 */
364 tk->tkr_mono.mult = clock->mult;
365 tk->tkr_raw.mult = clock->mult;
366 tk->ntp_err_mult = 0;
367 tk->skip_second_overflow = 0;
368}
369
370/* Timekeeper helper functions. */
371
372static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
373{
374 u64 nsec;
375
376 nsec = delta * tkr->mult + tkr->xtime_nsec;
377 nsec >>= tkr->shift;
378
379 return nsec;
380}
381
382static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
383{
384 u64 delta;
385
386 delta = timekeeping_get_delta(tkr);
387 return timekeeping_delta_to_ns(tkr, delta);
388}
389
390static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
391{
392 u64 delta;
393
394 /* calculate the delta since the last update_wall_time */
395 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
396 return timekeeping_delta_to_ns(tkr, delta);
397}
398
399/**
400 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
401 * @tkr: Timekeeping readout base from which we take the update
402 * @tkf: Pointer to NMI safe timekeeper
403 *
404 * We want to use this from any context including NMI and tracing /
405 * instrumenting the timekeeping code itself.
406 *
407 * Employ the latch technique; see @raw_write_seqcount_latch.
408 *
409 * So if a NMI hits the update of base[0] then it will use base[1]
410 * which is still consistent. In the worst case this can result is a
411 * slightly wrong timestamp (a few nanoseconds). See
412 * @ktime_get_mono_fast_ns.
413 */
414static void update_fast_timekeeper(const struct tk_read_base *tkr,
415 struct tk_fast *tkf)
416{
417 struct tk_read_base *base = tkf->base;
418
419 /* Force readers off to base[1] */
420 raw_write_seqcount_latch(&tkf->seq);
421
422 /* Update base[0] */
423 memcpy(base, tkr, sizeof(*base));
424
425 /* Force readers back to base[0] */
426 raw_write_seqcount_latch(&tkf->seq);
427
428 /* Update base[1] */
429 memcpy(base + 1, base, sizeof(*base));
430}
431
432static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
433{
434 struct tk_read_base *tkr;
435 unsigned int seq;
436 u64 now;
437
438 do {
439 seq = raw_read_seqcount_latch(&tkf->seq);
440 tkr = tkf->base + (seq & 0x01);
441 now = ktime_to_ns(tkr->base);
442
443 now += timekeeping_delta_to_ns(tkr,
444 clocksource_delta(
445 tk_clock_read(tkr),
446 tkr->cycle_last,
447 tkr->mask));
448 } while (read_seqcount_latch_retry(&tkf->seq, seq));
449
450 return now;
451}
452
453/**
454 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
455 *
456 * This timestamp is not guaranteed to be monotonic across an update.
457 * The timestamp is calculated by:
458 *
459 * now = base_mono + clock_delta * slope
460 *
461 * So if the update lowers the slope, readers who are forced to the
462 * not yet updated second array are still using the old steeper slope.
463 *
464 * tmono
465 * ^
466 * | o n
467 * | o n
468 * | u
469 * | o
470 * |o
471 * |12345678---> reader order
472 *
473 * o = old slope
474 * u = update
475 * n = new slope
476 *
477 * So reader 6 will observe time going backwards versus reader 5.
478 *
479 * While other CPUs are likely to be able to observe that, the only way
480 * for a CPU local observation is when an NMI hits in the middle of
481 * the update. Timestamps taken from that NMI context might be ahead
482 * of the following timestamps. Callers need to be aware of that and
483 * deal with it.
484 */
485u64 ktime_get_mono_fast_ns(void)
486{
487 return __ktime_get_fast_ns(&tk_fast_mono);
488}
489EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
490
491/**
492 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
493 *
494 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
495 * conversion factor is not affected by NTP/PTP correction.
496 */
497u64 ktime_get_raw_fast_ns(void)
498{
499 return __ktime_get_fast_ns(&tk_fast_raw);
500}
501EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
502
503/**
504 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
505 *
506 * To keep it NMI safe since we're accessing from tracing, we're not using a
507 * separate timekeeper with updates to monotonic clock and boot offset
508 * protected with seqcounts. This has the following minor side effects:
509 *
510 * (1) Its possible that a timestamp be taken after the boot offset is updated
511 * but before the timekeeper is updated. If this happens, the new boot offset
512 * is added to the old timekeeping making the clock appear to update slightly
513 * earlier:
514 * CPU 0 CPU 1
515 * timekeeping_inject_sleeptime64()
516 * __timekeeping_inject_sleeptime(tk, delta);
517 * timestamp();
518 * timekeeping_update(tk, TK_CLEAR_NTP...);
519 *
520 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
521 * partially updated. Since the tk->offs_boot update is a rare event, this
522 * should be a rare occurrence which postprocessing should be able to handle.
523 *
524 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns()
525 * apply as well.
526 */
527u64 notrace ktime_get_boot_fast_ns(void)
528{
529 struct timekeeper *tk = &tk_core.timekeeper;
530
531 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
532}
533EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
534
535static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
536{
537 struct tk_read_base *tkr;
538 u64 basem, baser, delta;
539 unsigned int seq;
540
541 do {
542 seq = raw_read_seqcount_latch(&tkf->seq);
543 tkr = tkf->base + (seq & 0x01);
544 basem = ktime_to_ns(tkr->base);
545 baser = ktime_to_ns(tkr->base_real);
546
547 delta = timekeeping_delta_to_ns(tkr,
548 clocksource_delta(tk_clock_read(tkr),
549 tkr->cycle_last, tkr->mask));
550 } while (read_seqcount_latch_retry(&tkf->seq, seq));
551
552 if (mono)
553 *mono = basem + delta;
554 return baser + delta;
555}
556
557/**
558 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
559 *
560 * See ktime_get_fast_ns() for documentation of the time stamp ordering.
561 */
562u64 ktime_get_real_fast_ns(void)
563{
564 return __ktime_get_real_fast(&tk_fast_mono, NULL);
565}
566EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
567
568/**
569 * ktime_get_fast_timestamps: - NMI safe timestamps
570 * @snapshot: Pointer to timestamp storage
571 *
572 * Stores clock monotonic, boottime and realtime timestamps.
573 *
574 * Boot time is a racy access on 32bit systems if the sleep time injection
575 * happens late during resume and not in timekeeping_resume(). That could
576 * be avoided by expanding struct tk_read_base with boot offset for 32bit
577 * and adding more overhead to the update. As this is a hard to observe
578 * once per resume event which can be filtered with reasonable effort using
579 * the accurate mono/real timestamps, it's probably not worth the trouble.
580 *
581 * Aside of that it might be possible on 32 and 64 bit to observe the
582 * following when the sleep time injection happens late:
583 *
584 * CPU 0 CPU 1
585 * timekeeping_resume()
586 * ktime_get_fast_timestamps()
587 * mono, real = __ktime_get_real_fast()
588 * inject_sleep_time()
589 * update boot offset
590 * boot = mono + bootoffset;
591 *
592 * That means that boot time already has the sleep time adjustment, but
593 * real time does not. On the next readout both are in sync again.
594 *
595 * Preventing this for 64bit is not really feasible without destroying the
596 * careful cache layout of the timekeeper because the sequence count and
597 * struct tk_read_base would then need two cache lines instead of one.
598 *
599 * Access to the time keeper clock source is disabled across the innermost
600 * steps of suspend/resume. The accessors still work, but the timestamps
601 * are frozen until time keeping is resumed which happens very early.
602 *
603 * For regular suspend/resume there is no observable difference vs. sched
604 * clock, but it might affect some of the nasty low level debug printks.
605 *
606 * OTOH, access to sched clock is not guaranteed across suspend/resume on
607 * all systems either so it depends on the hardware in use.
608 *
609 * If that turns out to be a real problem then this could be mitigated by
610 * using sched clock in a similar way as during early boot. But it's not as
611 * trivial as on early boot because it needs some careful protection
612 * against the clock monotonic timestamp jumping backwards on resume.
613 */
614void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
615{
616 struct timekeeper *tk = &tk_core.timekeeper;
617
618 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
619 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
620}
621
622/**
623 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
624 * @tk: Timekeeper to snapshot.
625 *
626 * It generally is unsafe to access the clocksource after timekeeping has been
627 * suspended, so take a snapshot of the readout base of @tk and use it as the
628 * fast timekeeper's readout base while suspended. It will return the same
629 * number of cycles every time until timekeeping is resumed at which time the
630 * proper readout base for the fast timekeeper will be restored automatically.
631 */
632static void halt_fast_timekeeper(const struct timekeeper *tk)
633{
634 static struct tk_read_base tkr_dummy;
635 const struct tk_read_base *tkr = &tk->tkr_mono;
636
637 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
638 cycles_at_suspend = tk_clock_read(tkr);
639 tkr_dummy.clock = &dummy_clock;
640 tkr_dummy.base_real = tkr->base + tk->offs_real;
641 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
642
643 tkr = &tk->tkr_raw;
644 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
645 tkr_dummy.clock = &dummy_clock;
646 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
647}
648
649static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
650
651static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
652{
653 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
654}
655
656/**
657 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
658 * @nb: Pointer to the notifier block to register
659 */
660int pvclock_gtod_register_notifier(struct notifier_block *nb)
661{
662 struct timekeeper *tk = &tk_core.timekeeper;
663 unsigned long flags;
664 int ret;
665
666 raw_spin_lock_irqsave(&timekeeper_lock, flags);
667 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
668 update_pvclock_gtod(tk, true);
669 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
670
671 return ret;
672}
673EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
674
675/**
676 * pvclock_gtod_unregister_notifier - unregister a pvclock
677 * timedata update listener
678 * @nb: Pointer to the notifier block to unregister
679 */
680int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
681{
682 unsigned long flags;
683 int ret;
684
685 raw_spin_lock_irqsave(&timekeeper_lock, flags);
686 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
687 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
688
689 return ret;
690}
691EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
692
693/*
694 * tk_update_leap_state - helper to update the next_leap_ktime
695 */
696static inline void tk_update_leap_state(struct timekeeper *tk)
697{
698 tk->next_leap_ktime = ntp_get_next_leap();
699 if (tk->next_leap_ktime != KTIME_MAX)
700 /* Convert to monotonic time */
701 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
702}
703
704/*
705 * Update the ktime_t based scalar nsec members of the timekeeper
706 */
707static inline void tk_update_ktime_data(struct timekeeper *tk)
708{
709 u64 seconds;
710 u32 nsec;
711
712 /*
713 * The xtime based monotonic readout is:
714 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
715 * The ktime based monotonic readout is:
716 * nsec = base_mono + now();
717 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
718 */
719 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
720 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
721 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
722
723 /*
724 * The sum of the nanoseconds portions of xtime and
725 * wall_to_monotonic can be greater/equal one second. Take
726 * this into account before updating tk->ktime_sec.
727 */
728 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
729 if (nsec >= NSEC_PER_SEC)
730 seconds++;
731 tk->ktime_sec = seconds;
732
733 /* Update the monotonic raw base */
734 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
735}
736
737/* must hold timekeeper_lock */
738static void timekeeping_update(struct timekeeper *tk, unsigned int action)
739{
740 if (action & TK_CLEAR_NTP) {
741 tk->ntp_error = 0;
742 ntp_clear();
743 }
744
745 tk_update_leap_state(tk);
746 tk_update_ktime_data(tk);
747
748 update_vsyscall(tk);
749 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
750
751 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
752 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
753 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
754
755 if (action & TK_CLOCK_WAS_SET)
756 tk->clock_was_set_seq++;
757 /*
758 * The mirroring of the data to the shadow-timekeeper needs
759 * to happen last here to ensure we don't over-write the
760 * timekeeper structure on the next update with stale data
761 */
762 if (action & TK_MIRROR)
763 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
764 sizeof(tk_core.timekeeper));
765}
766
767/**
768 * timekeeping_forward_now - update clock to the current time
769 * @tk: Pointer to the timekeeper to update
770 *
771 * Forward the current clock to update its state since the last call to
772 * update_wall_time(). This is useful before significant clock changes,
773 * as it avoids having to deal with this time offset explicitly.
774 */
775static void timekeeping_forward_now(struct timekeeper *tk)
776{
777 u64 cycle_now, delta;
778
779 cycle_now = tk_clock_read(&tk->tkr_mono);
780 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
781 tk->tkr_mono.cycle_last = cycle_now;
782 tk->tkr_raw.cycle_last = cycle_now;
783
784 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
785 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
786
787 tk_normalize_xtime(tk);
788}
789
790/**
791 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
792 * @ts: pointer to the timespec to be set
793 *
794 * Returns the time of day in a timespec64 (WARN if suspended).
795 */
796void ktime_get_real_ts64(struct timespec64 *ts)
797{
798 struct timekeeper *tk = &tk_core.timekeeper;
799 unsigned int seq;
800 u64 nsecs;
801
802 WARN_ON(timekeeping_suspended);
803
804 do {
805 seq = read_seqcount_begin(&tk_core.seq);
806
807 ts->tv_sec = tk->xtime_sec;
808 nsecs = timekeeping_get_ns(&tk->tkr_mono);
809
810 } while (read_seqcount_retry(&tk_core.seq, seq));
811
812 ts->tv_nsec = 0;
813 timespec64_add_ns(ts, nsecs);
814}
815EXPORT_SYMBOL(ktime_get_real_ts64);
816
817ktime_t ktime_get(void)
818{
819 struct timekeeper *tk = &tk_core.timekeeper;
820 unsigned int seq;
821 ktime_t base;
822 u64 nsecs;
823
824 WARN_ON(timekeeping_suspended);
825
826 do {
827 seq = read_seqcount_begin(&tk_core.seq);
828 base = tk->tkr_mono.base;
829 nsecs = timekeeping_get_ns(&tk->tkr_mono);
830
831 } while (read_seqcount_retry(&tk_core.seq, seq));
832
833 return ktime_add_ns(base, nsecs);
834}
835EXPORT_SYMBOL_GPL(ktime_get);
836
837u32 ktime_get_resolution_ns(void)
838{
839 struct timekeeper *tk = &tk_core.timekeeper;
840 unsigned int seq;
841 u32 nsecs;
842
843 WARN_ON(timekeeping_suspended);
844
845 do {
846 seq = read_seqcount_begin(&tk_core.seq);
847 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
848 } while (read_seqcount_retry(&tk_core.seq, seq));
849
850 return nsecs;
851}
852EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
853
854static ktime_t *offsets[TK_OFFS_MAX] = {
855 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
856 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
857 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
858};
859
860ktime_t ktime_get_with_offset(enum tk_offsets offs)
861{
862 struct timekeeper *tk = &tk_core.timekeeper;
863 unsigned int seq;
864 ktime_t base, *offset = offsets[offs];
865 u64 nsecs;
866
867 WARN_ON(timekeeping_suspended);
868
869 do {
870 seq = read_seqcount_begin(&tk_core.seq);
871 base = ktime_add(tk->tkr_mono.base, *offset);
872 nsecs = timekeeping_get_ns(&tk->tkr_mono);
873
874 } while (read_seqcount_retry(&tk_core.seq, seq));
875
876 return ktime_add_ns(base, nsecs);
877
878}
879EXPORT_SYMBOL_GPL(ktime_get_with_offset);
880
881ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
882{
883 struct timekeeper *tk = &tk_core.timekeeper;
884 unsigned int seq;
885 ktime_t base, *offset = offsets[offs];
886 u64 nsecs;
887
888 WARN_ON(timekeeping_suspended);
889
890 do {
891 seq = read_seqcount_begin(&tk_core.seq);
892 base = ktime_add(tk->tkr_mono.base, *offset);
893 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
894
895 } while (read_seqcount_retry(&tk_core.seq, seq));
896
897 return ktime_add_ns(base, nsecs);
898}
899EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
900
901/**
902 * ktime_mono_to_any() - convert monotonic time to any other time
903 * @tmono: time to convert.
904 * @offs: which offset to use
905 */
906ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
907{
908 ktime_t *offset = offsets[offs];
909 unsigned int seq;
910 ktime_t tconv;
911
912 do {
913 seq = read_seqcount_begin(&tk_core.seq);
914 tconv = ktime_add(tmono, *offset);
915 } while (read_seqcount_retry(&tk_core.seq, seq));
916
917 return tconv;
918}
919EXPORT_SYMBOL_GPL(ktime_mono_to_any);
920
921/**
922 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
923 */
924ktime_t ktime_get_raw(void)
925{
926 struct timekeeper *tk = &tk_core.timekeeper;
927 unsigned int seq;
928 ktime_t base;
929 u64 nsecs;
930
931 do {
932 seq = read_seqcount_begin(&tk_core.seq);
933 base = tk->tkr_raw.base;
934 nsecs = timekeeping_get_ns(&tk->tkr_raw);
935
936 } while (read_seqcount_retry(&tk_core.seq, seq));
937
938 return ktime_add_ns(base, nsecs);
939}
940EXPORT_SYMBOL_GPL(ktime_get_raw);
941
942/**
943 * ktime_get_ts64 - get the monotonic clock in timespec64 format
944 * @ts: pointer to timespec variable
945 *
946 * The function calculates the monotonic clock from the realtime
947 * clock and the wall_to_monotonic offset and stores the result
948 * in normalized timespec64 format in the variable pointed to by @ts.
949 */
950void ktime_get_ts64(struct timespec64 *ts)
951{
952 struct timekeeper *tk = &tk_core.timekeeper;
953 struct timespec64 tomono;
954 unsigned int seq;
955 u64 nsec;
956
957 WARN_ON(timekeeping_suspended);
958
959 do {
960 seq = read_seqcount_begin(&tk_core.seq);
961 ts->tv_sec = tk->xtime_sec;
962 nsec = timekeeping_get_ns(&tk->tkr_mono);
963 tomono = tk->wall_to_monotonic;
964
965 } while (read_seqcount_retry(&tk_core.seq, seq));
966
967 ts->tv_sec += tomono.tv_sec;
968 ts->tv_nsec = 0;
969 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
970}
971EXPORT_SYMBOL_GPL(ktime_get_ts64);
972
973/**
974 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
975 *
976 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
977 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
978 * works on both 32 and 64 bit systems. On 32 bit systems the readout
979 * covers ~136 years of uptime which should be enough to prevent
980 * premature wrap arounds.
981 */
982time64_t ktime_get_seconds(void)
983{
984 struct timekeeper *tk = &tk_core.timekeeper;
985
986 WARN_ON(timekeeping_suspended);
987 return tk->ktime_sec;
988}
989EXPORT_SYMBOL_GPL(ktime_get_seconds);
990
991/**
992 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
993 *
994 * Returns the wall clock seconds since 1970.
995 *
996 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
997 * 32bit systems the access must be protected with the sequence
998 * counter to provide "atomic" access to the 64bit tk->xtime_sec
999 * value.
1000 */
1001time64_t ktime_get_real_seconds(void)
1002{
1003 struct timekeeper *tk = &tk_core.timekeeper;
1004 time64_t seconds;
1005 unsigned int seq;
1006
1007 if (IS_ENABLED(CONFIG_64BIT))
1008 return tk->xtime_sec;
1009
1010 do {
1011 seq = read_seqcount_begin(&tk_core.seq);
1012 seconds = tk->xtime_sec;
1013
1014 } while (read_seqcount_retry(&tk_core.seq, seq));
1015
1016 return seconds;
1017}
1018EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1019
1020/**
1021 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1022 * but without the sequence counter protect. This internal function
1023 * is called just when timekeeping lock is already held.
1024 */
1025noinstr time64_t __ktime_get_real_seconds(void)
1026{
1027 struct timekeeper *tk = &tk_core.timekeeper;
1028
1029 return tk->xtime_sec;
1030}
1031
1032/**
1033 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1034 * @systime_snapshot: pointer to struct receiving the system time snapshot
1035 */
1036void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1037{
1038 struct timekeeper *tk = &tk_core.timekeeper;
1039 unsigned int seq;
1040 ktime_t base_raw;
1041 ktime_t base_real;
1042 u64 nsec_raw;
1043 u64 nsec_real;
1044 u64 now;
1045
1046 WARN_ON_ONCE(timekeeping_suspended);
1047
1048 do {
1049 seq = read_seqcount_begin(&tk_core.seq);
1050 now = tk_clock_read(&tk->tkr_mono);
1051 systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1052 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1053 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1054 base_real = ktime_add(tk->tkr_mono.base,
1055 tk_core.timekeeper.offs_real);
1056 base_raw = tk->tkr_raw.base;
1057 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1058 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1059 } while (read_seqcount_retry(&tk_core.seq, seq));
1060
1061 systime_snapshot->cycles = now;
1062 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1063 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1064}
1065EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1066
1067/* Scale base by mult/div checking for overflow */
1068static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1069{
1070 u64 tmp, rem;
1071
1072 tmp = div64_u64_rem(*base, div, &rem);
1073
1074 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1075 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1076 return -EOVERFLOW;
1077 tmp *= mult;
1078
1079 rem = div64_u64(rem * mult, div);
1080 *base = tmp + rem;
1081 return 0;
1082}
1083
1084/**
1085 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1086 * @history: Snapshot representing start of history
1087 * @partial_history_cycles: Cycle offset into history (fractional part)
1088 * @total_history_cycles: Total history length in cycles
1089 * @discontinuity: True indicates clock was set on history period
1090 * @ts: Cross timestamp that should be adjusted using
1091 * partial/total ratio
1092 *
1093 * Helper function used by get_device_system_crosststamp() to correct the
1094 * crosstimestamp corresponding to the start of the current interval to the
1095 * system counter value (timestamp point) provided by the driver. The
1096 * total_history_* quantities are the total history starting at the provided
1097 * reference point and ending at the start of the current interval. The cycle
1098 * count between the driver timestamp point and the start of the current
1099 * interval is partial_history_cycles.
1100 */
1101static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1102 u64 partial_history_cycles,
1103 u64 total_history_cycles,
1104 bool discontinuity,
1105 struct system_device_crosststamp *ts)
1106{
1107 struct timekeeper *tk = &tk_core.timekeeper;
1108 u64 corr_raw, corr_real;
1109 bool interp_forward;
1110 int ret;
1111
1112 if (total_history_cycles == 0 || partial_history_cycles == 0)
1113 return 0;
1114
1115 /* Interpolate shortest distance from beginning or end of history */
1116 interp_forward = partial_history_cycles > total_history_cycles / 2;
1117 partial_history_cycles = interp_forward ?
1118 total_history_cycles - partial_history_cycles :
1119 partial_history_cycles;
1120
1121 /*
1122 * Scale the monotonic raw time delta by:
1123 * partial_history_cycles / total_history_cycles
1124 */
1125 corr_raw = (u64)ktime_to_ns(
1126 ktime_sub(ts->sys_monoraw, history->raw));
1127 ret = scale64_check_overflow(partial_history_cycles,
1128 total_history_cycles, &corr_raw);
1129 if (ret)
1130 return ret;
1131
1132 /*
1133 * If there is a discontinuity in the history, scale monotonic raw
1134 * correction by:
1135 * mult(real)/mult(raw) yielding the realtime correction
1136 * Otherwise, calculate the realtime correction similar to monotonic
1137 * raw calculation
1138 */
1139 if (discontinuity) {
1140 corr_real = mul_u64_u32_div
1141 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1142 } else {
1143 corr_real = (u64)ktime_to_ns(
1144 ktime_sub(ts->sys_realtime, history->real));
1145 ret = scale64_check_overflow(partial_history_cycles,
1146 total_history_cycles, &corr_real);
1147 if (ret)
1148 return ret;
1149 }
1150
1151 /* Fixup monotonic raw and real time time values */
1152 if (interp_forward) {
1153 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1154 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1155 } else {
1156 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1157 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1158 }
1159
1160 return 0;
1161}
1162
1163/*
1164 * cycle_between - true if test occurs chronologically between before and after
1165 */
1166static bool cycle_between(u64 before, u64 test, u64 after)
1167{
1168 if (test > before && test < after)
1169 return true;
1170 if (test < before && before > after)
1171 return true;
1172 return false;
1173}
1174
1175/**
1176 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1177 * @get_time_fn: Callback to get simultaneous device time and
1178 * system counter from the device driver
1179 * @ctx: Context passed to get_time_fn()
1180 * @history_begin: Historical reference point used to interpolate system
1181 * time when counter provided by the driver is before the current interval
1182 * @xtstamp: Receives simultaneously captured system and device time
1183 *
1184 * Reads a timestamp from a device and correlates it to system time
1185 */
1186int get_device_system_crosststamp(int (*get_time_fn)
1187 (ktime_t *device_time,
1188 struct system_counterval_t *sys_counterval,
1189 void *ctx),
1190 void *ctx,
1191 struct system_time_snapshot *history_begin,
1192 struct system_device_crosststamp *xtstamp)
1193{
1194 struct system_counterval_t system_counterval;
1195 struct timekeeper *tk = &tk_core.timekeeper;
1196 u64 cycles, now, interval_start;
1197 unsigned int clock_was_set_seq = 0;
1198 ktime_t base_real, base_raw;
1199 u64 nsec_real, nsec_raw;
1200 u8 cs_was_changed_seq;
1201 unsigned int seq;
1202 bool do_interp;
1203 int ret;
1204
1205 do {
1206 seq = read_seqcount_begin(&tk_core.seq);
1207 /*
1208 * Try to synchronously capture device time and a system
1209 * counter value calling back into the device driver
1210 */
1211 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1212 if (ret)
1213 return ret;
1214
1215 /*
1216 * Verify that the clocksource associated with the captured
1217 * system counter value is the same as the currently installed
1218 * timekeeper clocksource
1219 */
1220 if (tk->tkr_mono.clock != system_counterval.cs)
1221 return -ENODEV;
1222 cycles = system_counterval.cycles;
1223
1224 /*
1225 * Check whether the system counter value provided by the
1226 * device driver is on the current timekeeping interval.
1227 */
1228 now = tk_clock_read(&tk->tkr_mono);
1229 interval_start = tk->tkr_mono.cycle_last;
1230 if (!cycle_between(interval_start, cycles, now)) {
1231 clock_was_set_seq = tk->clock_was_set_seq;
1232 cs_was_changed_seq = tk->cs_was_changed_seq;
1233 cycles = interval_start;
1234 do_interp = true;
1235 } else {
1236 do_interp = false;
1237 }
1238
1239 base_real = ktime_add(tk->tkr_mono.base,
1240 tk_core.timekeeper.offs_real);
1241 base_raw = tk->tkr_raw.base;
1242
1243 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1244 system_counterval.cycles);
1245 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1246 system_counterval.cycles);
1247 } while (read_seqcount_retry(&tk_core.seq, seq));
1248
1249 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1250 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1251
1252 /*
1253 * Interpolate if necessary, adjusting back from the start of the
1254 * current interval
1255 */
1256 if (do_interp) {
1257 u64 partial_history_cycles, total_history_cycles;
1258 bool discontinuity;
1259
1260 /*
1261 * Check that the counter value occurs after the provided
1262 * history reference and that the history doesn't cross a
1263 * clocksource change
1264 */
1265 if (!history_begin ||
1266 !cycle_between(history_begin->cycles,
1267 system_counterval.cycles, cycles) ||
1268 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1269 return -EINVAL;
1270 partial_history_cycles = cycles - system_counterval.cycles;
1271 total_history_cycles = cycles - history_begin->cycles;
1272 discontinuity =
1273 history_begin->clock_was_set_seq != clock_was_set_seq;
1274
1275 ret = adjust_historical_crosststamp(history_begin,
1276 partial_history_cycles,
1277 total_history_cycles,
1278 discontinuity, xtstamp);
1279 if (ret)
1280 return ret;
1281 }
1282
1283 return 0;
1284}
1285EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1286
1287/**
1288 * do_settimeofday64 - Sets the time of day.
1289 * @ts: pointer to the timespec64 variable containing the new time
1290 *
1291 * Sets the time of day to the new time and update NTP and notify hrtimers
1292 */
1293int do_settimeofday64(const struct timespec64 *ts)
1294{
1295 struct timekeeper *tk = &tk_core.timekeeper;
1296 struct timespec64 ts_delta, xt;
1297 unsigned long flags;
1298 int ret = 0;
1299
1300 if (!timespec64_valid_settod(ts))
1301 return -EINVAL;
1302
1303 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1304 write_seqcount_begin(&tk_core.seq);
1305
1306 timekeeping_forward_now(tk);
1307
1308 xt = tk_xtime(tk);
1309 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1310 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1311
1312 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1313 ret = -EINVAL;
1314 goto out;
1315 }
1316
1317 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1318
1319 tk_set_xtime(tk, ts);
1320out:
1321 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1322
1323 write_seqcount_end(&tk_core.seq);
1324 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1325
1326 /* signal hrtimers about time change */
1327 clock_was_set();
1328
1329 if (!ret)
1330 audit_tk_injoffset(ts_delta);
1331
1332 return ret;
1333}
1334EXPORT_SYMBOL(do_settimeofday64);
1335
1336/**
1337 * timekeeping_inject_offset - Adds or subtracts from the current time.
1338 * @ts: Pointer to the timespec variable containing the offset
1339 *
1340 * Adds or subtracts an offset value from the current time.
1341 */
1342static int timekeeping_inject_offset(const struct timespec64 *ts)
1343{
1344 struct timekeeper *tk = &tk_core.timekeeper;
1345 unsigned long flags;
1346 struct timespec64 tmp;
1347 int ret = 0;
1348
1349 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1350 return -EINVAL;
1351
1352 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1353 write_seqcount_begin(&tk_core.seq);
1354
1355 timekeeping_forward_now(tk);
1356
1357 /* Make sure the proposed value is valid */
1358 tmp = timespec64_add(tk_xtime(tk), *ts);
1359 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1360 !timespec64_valid_settod(&tmp)) {
1361 ret = -EINVAL;
1362 goto error;
1363 }
1364
1365 tk_xtime_add(tk, ts);
1366 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1367
1368error: /* even if we error out, we forwarded the time, so call update */
1369 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1370
1371 write_seqcount_end(&tk_core.seq);
1372 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1373
1374 /* signal hrtimers about time change */
1375 clock_was_set();
1376
1377 return ret;
1378}
1379
1380/*
1381 * Indicates if there is an offset between the system clock and the hardware
1382 * clock/persistent clock/rtc.
1383 */
1384int persistent_clock_is_local;
1385
1386/*
1387 * Adjust the time obtained from the CMOS to be UTC time instead of
1388 * local time.
1389 *
1390 * This is ugly, but preferable to the alternatives. Otherwise we
1391 * would either need to write a program to do it in /etc/rc (and risk
1392 * confusion if the program gets run more than once; it would also be
1393 * hard to make the program warp the clock precisely n hours) or
1394 * compile in the timezone information into the kernel. Bad, bad....
1395 *
1396 * - TYT, 1992-01-01
1397 *
1398 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1399 * as real UNIX machines always do it. This avoids all headaches about
1400 * daylight saving times and warping kernel clocks.
1401 */
1402void timekeeping_warp_clock(void)
1403{
1404 if (sys_tz.tz_minuteswest != 0) {
1405 struct timespec64 adjust;
1406
1407 persistent_clock_is_local = 1;
1408 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1409 adjust.tv_nsec = 0;
1410 timekeeping_inject_offset(&adjust);
1411 }
1412}
1413
1414/*
1415 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1416 */
1417static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1418{
1419 tk->tai_offset = tai_offset;
1420 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1421}
1422
1423/*
1424 * change_clocksource - Swaps clocksources if a new one is available
1425 *
1426 * Accumulates current time interval and initializes new clocksource
1427 */
1428static int change_clocksource(void *data)
1429{
1430 struct timekeeper *tk = &tk_core.timekeeper;
1431 struct clocksource *new, *old = NULL;
1432 unsigned long flags;
1433 bool change = false;
1434
1435 new = (struct clocksource *) data;
1436
1437 /*
1438 * If the cs is in module, get a module reference. Succeeds
1439 * for built-in code (owner == NULL) as well.
1440 */
1441 if (try_module_get(new->owner)) {
1442 if (!new->enable || new->enable(new) == 0)
1443 change = true;
1444 else
1445 module_put(new->owner);
1446 }
1447
1448 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1449 write_seqcount_begin(&tk_core.seq);
1450
1451 timekeeping_forward_now(tk);
1452
1453 if (change) {
1454 old = tk->tkr_mono.clock;
1455 tk_setup_internals(tk, new);
1456 }
1457
1458 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1459
1460 write_seqcount_end(&tk_core.seq);
1461 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1462
1463 if (old) {
1464 if (old->disable)
1465 old->disable(old);
1466
1467 module_put(old->owner);
1468 }
1469
1470 return 0;
1471}
1472
1473/**
1474 * timekeeping_notify - Install a new clock source
1475 * @clock: pointer to the clock source
1476 *
1477 * This function is called from clocksource.c after a new, better clock
1478 * source has been registered. The caller holds the clocksource_mutex.
1479 */
1480int timekeeping_notify(struct clocksource *clock)
1481{
1482 struct timekeeper *tk = &tk_core.timekeeper;
1483
1484 if (tk->tkr_mono.clock == clock)
1485 return 0;
1486 stop_machine(change_clocksource, clock, NULL);
1487 tick_clock_notify();
1488 return tk->tkr_mono.clock == clock ? 0 : -1;
1489}
1490
1491/**
1492 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1493 * @ts: pointer to the timespec64 to be set
1494 *
1495 * Returns the raw monotonic time (completely un-modified by ntp)
1496 */
1497void ktime_get_raw_ts64(struct timespec64 *ts)
1498{
1499 struct timekeeper *tk = &tk_core.timekeeper;
1500 unsigned int seq;
1501 u64 nsecs;
1502
1503 do {
1504 seq = read_seqcount_begin(&tk_core.seq);
1505 ts->tv_sec = tk->raw_sec;
1506 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1507
1508 } while (read_seqcount_retry(&tk_core.seq, seq));
1509
1510 ts->tv_nsec = 0;
1511 timespec64_add_ns(ts, nsecs);
1512}
1513EXPORT_SYMBOL(ktime_get_raw_ts64);
1514
1515
1516/**
1517 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1518 */
1519int timekeeping_valid_for_hres(void)
1520{
1521 struct timekeeper *tk = &tk_core.timekeeper;
1522 unsigned int seq;
1523 int ret;
1524
1525 do {
1526 seq = read_seqcount_begin(&tk_core.seq);
1527
1528 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1529
1530 } while (read_seqcount_retry(&tk_core.seq, seq));
1531
1532 return ret;
1533}
1534
1535/**
1536 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1537 */
1538u64 timekeeping_max_deferment(void)
1539{
1540 struct timekeeper *tk = &tk_core.timekeeper;
1541 unsigned int seq;
1542 u64 ret;
1543
1544 do {
1545 seq = read_seqcount_begin(&tk_core.seq);
1546
1547 ret = tk->tkr_mono.clock->max_idle_ns;
1548
1549 } while (read_seqcount_retry(&tk_core.seq, seq));
1550
1551 return ret;
1552}
1553
1554/**
1555 * read_persistent_clock64 - Return time from the persistent clock.
1556 * @ts: Pointer to the storage for the readout value
1557 *
1558 * Weak dummy function for arches that do not yet support it.
1559 * Reads the time from the battery backed persistent clock.
1560 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1561 *
1562 * XXX - Do be sure to remove it once all arches implement it.
1563 */
1564void __weak read_persistent_clock64(struct timespec64 *ts)
1565{
1566 ts->tv_sec = 0;
1567 ts->tv_nsec = 0;
1568}
1569
1570/**
1571 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1572 * from the boot.
1573 *
1574 * Weak dummy function for arches that do not yet support it.
1575 * @wall_time: - current time as returned by persistent clock
1576 * @boot_offset: - offset that is defined as wall_time - boot_time
1577 *
1578 * The default function calculates offset based on the current value of
1579 * local_clock(). This way architectures that support sched_clock() but don't
1580 * support dedicated boot time clock will provide the best estimate of the
1581 * boot time.
1582 */
1583void __weak __init
1584read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1585 struct timespec64 *boot_offset)
1586{
1587 read_persistent_clock64(wall_time);
1588 *boot_offset = ns_to_timespec64(local_clock());
1589}
1590
1591/*
1592 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1593 *
1594 * The flag starts of false and is only set when a suspend reaches
1595 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1596 * timekeeper clocksource is not stopping across suspend and has been
1597 * used to update sleep time. If the timekeeper clocksource has stopped
1598 * then the flag stays true and is used by the RTC resume code to decide
1599 * whether sleeptime must be injected and if so the flag gets false then.
1600 *
1601 * If a suspend fails before reaching timekeeping_resume() then the flag
1602 * stays false and prevents erroneous sleeptime injection.
1603 */
1604static bool suspend_timing_needed;
1605
1606/* Flag for if there is a persistent clock on this platform */
1607static bool persistent_clock_exists;
1608
1609/*
1610 * timekeeping_init - Initializes the clocksource and common timekeeping values
1611 */
1612void __init timekeeping_init(void)
1613{
1614 struct timespec64 wall_time, boot_offset, wall_to_mono;
1615 struct timekeeper *tk = &tk_core.timekeeper;
1616 struct clocksource *clock;
1617 unsigned long flags;
1618
1619 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1620 if (timespec64_valid_settod(&wall_time) &&
1621 timespec64_to_ns(&wall_time) > 0) {
1622 persistent_clock_exists = true;
1623 } else if (timespec64_to_ns(&wall_time) != 0) {
1624 pr_warn("Persistent clock returned invalid value");
1625 wall_time = (struct timespec64){0};
1626 }
1627
1628 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1629 boot_offset = (struct timespec64){0};
1630
1631 /*
1632 * We want set wall_to_mono, so the following is true:
1633 * wall time + wall_to_mono = boot time
1634 */
1635 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1636
1637 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1638 write_seqcount_begin(&tk_core.seq);
1639 ntp_init();
1640
1641 clock = clocksource_default_clock();
1642 if (clock->enable)
1643 clock->enable(clock);
1644 tk_setup_internals(tk, clock);
1645
1646 tk_set_xtime(tk, &wall_time);
1647 tk->raw_sec = 0;
1648
1649 tk_set_wall_to_mono(tk, wall_to_mono);
1650
1651 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1652
1653 write_seqcount_end(&tk_core.seq);
1654 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1655}
1656
1657/* time in seconds when suspend began for persistent clock */
1658static struct timespec64 timekeeping_suspend_time;
1659
1660/**
1661 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1662 * @tk: Pointer to the timekeeper to be updated
1663 * @delta: Pointer to the delta value in timespec64 format
1664 *
1665 * Takes a timespec offset measuring a suspend interval and properly
1666 * adds the sleep offset to the timekeeping variables.
1667 */
1668static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1669 const struct timespec64 *delta)
1670{
1671 if (!timespec64_valid_strict(delta)) {
1672 printk_deferred(KERN_WARNING
1673 "__timekeeping_inject_sleeptime: Invalid "
1674 "sleep delta value!\n");
1675 return;
1676 }
1677 tk_xtime_add(tk, delta);
1678 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1679 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1680 tk_debug_account_sleep_time(delta);
1681}
1682
1683#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1684/**
1685 * We have three kinds of time sources to use for sleep time
1686 * injection, the preference order is:
1687 * 1) non-stop clocksource
1688 * 2) persistent clock (ie: RTC accessible when irqs are off)
1689 * 3) RTC
1690 *
1691 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1692 * If system has neither 1) nor 2), 3) will be used finally.
1693 *
1694 *
1695 * If timekeeping has injected sleeptime via either 1) or 2),
1696 * 3) becomes needless, so in this case we don't need to call
1697 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1698 * means.
1699 */
1700bool timekeeping_rtc_skipresume(void)
1701{
1702 return !suspend_timing_needed;
1703}
1704
1705/**
1706 * 1) can be determined whether to use or not only when doing
1707 * timekeeping_resume() which is invoked after rtc_suspend(),
1708 * so we can't skip rtc_suspend() surely if system has 1).
1709 *
1710 * But if system has 2), 2) will definitely be used, so in this
1711 * case we don't need to call rtc_suspend(), and this is what
1712 * timekeeping_rtc_skipsuspend() means.
1713 */
1714bool timekeeping_rtc_skipsuspend(void)
1715{
1716 return persistent_clock_exists;
1717}
1718
1719/**
1720 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1721 * @delta: pointer to a timespec64 delta value
1722 *
1723 * This hook is for architectures that cannot support read_persistent_clock64
1724 * because their RTC/persistent clock is only accessible when irqs are enabled.
1725 * and also don't have an effective nonstop clocksource.
1726 *
1727 * This function should only be called by rtc_resume(), and allows
1728 * a suspend offset to be injected into the timekeeping values.
1729 */
1730void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1731{
1732 struct timekeeper *tk = &tk_core.timekeeper;
1733 unsigned long flags;
1734
1735 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1736 write_seqcount_begin(&tk_core.seq);
1737
1738 suspend_timing_needed = false;
1739
1740 timekeeping_forward_now(tk);
1741
1742 __timekeeping_inject_sleeptime(tk, delta);
1743
1744 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1745
1746 write_seqcount_end(&tk_core.seq);
1747 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1748
1749 /* signal hrtimers about time change */
1750 clock_was_set();
1751}
1752#endif
1753
1754/**
1755 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1756 */
1757void timekeeping_resume(void)
1758{
1759 struct timekeeper *tk = &tk_core.timekeeper;
1760 struct clocksource *clock = tk->tkr_mono.clock;
1761 unsigned long flags;
1762 struct timespec64 ts_new, ts_delta;
1763 u64 cycle_now, nsec;
1764 bool inject_sleeptime = false;
1765
1766 read_persistent_clock64(&ts_new);
1767
1768 clockevents_resume();
1769 clocksource_resume();
1770
1771 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1772 write_seqcount_begin(&tk_core.seq);
1773
1774 /*
1775 * After system resumes, we need to calculate the suspended time and
1776 * compensate it for the OS time. There are 3 sources that could be
1777 * used: Nonstop clocksource during suspend, persistent clock and rtc
1778 * device.
1779 *
1780 * One specific platform may have 1 or 2 or all of them, and the
1781 * preference will be:
1782 * suspend-nonstop clocksource -> persistent clock -> rtc
1783 * The less preferred source will only be tried if there is no better
1784 * usable source. The rtc part is handled separately in rtc core code.
1785 */
1786 cycle_now = tk_clock_read(&tk->tkr_mono);
1787 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1788 if (nsec > 0) {
1789 ts_delta = ns_to_timespec64(nsec);
1790 inject_sleeptime = true;
1791 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1792 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1793 inject_sleeptime = true;
1794 }
1795
1796 if (inject_sleeptime) {
1797 suspend_timing_needed = false;
1798 __timekeeping_inject_sleeptime(tk, &ts_delta);
1799 }
1800
1801 /* Re-base the last cycle value */
1802 tk->tkr_mono.cycle_last = cycle_now;
1803 tk->tkr_raw.cycle_last = cycle_now;
1804
1805 tk->ntp_error = 0;
1806 timekeeping_suspended = 0;
1807 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1808 write_seqcount_end(&tk_core.seq);
1809 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1810
1811 touch_softlockup_watchdog();
1812
1813 tick_resume();
1814 hrtimers_resume();
1815}
1816
1817int timekeeping_suspend(void)
1818{
1819 struct timekeeper *tk = &tk_core.timekeeper;
1820 unsigned long flags;
1821 struct timespec64 delta, delta_delta;
1822 static struct timespec64 old_delta;
1823 struct clocksource *curr_clock;
1824 u64 cycle_now;
1825
1826 read_persistent_clock64(&timekeeping_suspend_time);
1827
1828 /*
1829 * On some systems the persistent_clock can not be detected at
1830 * timekeeping_init by its return value, so if we see a valid
1831 * value returned, update the persistent_clock_exists flag.
1832 */
1833 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1834 persistent_clock_exists = true;
1835
1836 suspend_timing_needed = true;
1837
1838 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1839 write_seqcount_begin(&tk_core.seq);
1840 timekeeping_forward_now(tk);
1841 timekeeping_suspended = 1;
1842
1843 /*
1844 * Since we've called forward_now, cycle_last stores the value
1845 * just read from the current clocksource. Save this to potentially
1846 * use in suspend timing.
1847 */
1848 curr_clock = tk->tkr_mono.clock;
1849 cycle_now = tk->tkr_mono.cycle_last;
1850 clocksource_start_suspend_timing(curr_clock, cycle_now);
1851
1852 if (persistent_clock_exists) {
1853 /*
1854 * To avoid drift caused by repeated suspend/resumes,
1855 * which each can add ~1 second drift error,
1856 * try to compensate so the difference in system time
1857 * and persistent_clock time stays close to constant.
1858 */
1859 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1860 delta_delta = timespec64_sub(delta, old_delta);
1861 if (abs(delta_delta.tv_sec) >= 2) {
1862 /*
1863 * if delta_delta is too large, assume time correction
1864 * has occurred and set old_delta to the current delta.
1865 */
1866 old_delta = delta;
1867 } else {
1868 /* Otherwise try to adjust old_system to compensate */
1869 timekeeping_suspend_time =
1870 timespec64_add(timekeeping_suspend_time, delta_delta);
1871 }
1872 }
1873
1874 timekeeping_update(tk, TK_MIRROR);
1875 halt_fast_timekeeper(tk);
1876 write_seqcount_end(&tk_core.seq);
1877 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1878
1879 tick_suspend();
1880 clocksource_suspend();
1881 clockevents_suspend();
1882
1883 return 0;
1884}
1885
1886/* sysfs resume/suspend bits for timekeeping */
1887static struct syscore_ops timekeeping_syscore_ops = {
1888 .resume = timekeeping_resume,
1889 .suspend = timekeeping_suspend,
1890};
1891
1892static int __init timekeeping_init_ops(void)
1893{
1894 register_syscore_ops(&timekeeping_syscore_ops);
1895 return 0;
1896}
1897device_initcall(timekeeping_init_ops);
1898
1899/*
1900 * Apply a multiplier adjustment to the timekeeper
1901 */
1902static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1903 s64 offset,
1904 s32 mult_adj)
1905{
1906 s64 interval = tk->cycle_interval;
1907
1908 if (mult_adj == 0) {
1909 return;
1910 } else if (mult_adj == -1) {
1911 interval = -interval;
1912 offset = -offset;
1913 } else if (mult_adj != 1) {
1914 interval *= mult_adj;
1915 offset *= mult_adj;
1916 }
1917
1918 /*
1919 * So the following can be confusing.
1920 *
1921 * To keep things simple, lets assume mult_adj == 1 for now.
1922 *
1923 * When mult_adj != 1, remember that the interval and offset values
1924 * have been appropriately scaled so the math is the same.
1925 *
1926 * The basic idea here is that we're increasing the multiplier
1927 * by one, this causes the xtime_interval to be incremented by
1928 * one cycle_interval. This is because:
1929 * xtime_interval = cycle_interval * mult
1930 * So if mult is being incremented by one:
1931 * xtime_interval = cycle_interval * (mult + 1)
1932 * Its the same as:
1933 * xtime_interval = (cycle_interval * mult) + cycle_interval
1934 * Which can be shortened to:
1935 * xtime_interval += cycle_interval
1936 *
1937 * So offset stores the non-accumulated cycles. Thus the current
1938 * time (in shifted nanoseconds) is:
1939 * now = (offset * adj) + xtime_nsec
1940 * Now, even though we're adjusting the clock frequency, we have
1941 * to keep time consistent. In other words, we can't jump back
1942 * in time, and we also want to avoid jumping forward in time.
1943 *
1944 * So given the same offset value, we need the time to be the same
1945 * both before and after the freq adjustment.
1946 * now = (offset * adj_1) + xtime_nsec_1
1947 * now = (offset * adj_2) + xtime_nsec_2
1948 * So:
1949 * (offset * adj_1) + xtime_nsec_1 =
1950 * (offset * adj_2) + xtime_nsec_2
1951 * And we know:
1952 * adj_2 = adj_1 + 1
1953 * So:
1954 * (offset * adj_1) + xtime_nsec_1 =
1955 * (offset * (adj_1+1)) + xtime_nsec_2
1956 * (offset * adj_1) + xtime_nsec_1 =
1957 * (offset * adj_1) + offset + xtime_nsec_2
1958 * Canceling the sides:
1959 * xtime_nsec_1 = offset + xtime_nsec_2
1960 * Which gives us:
1961 * xtime_nsec_2 = xtime_nsec_1 - offset
1962 * Which simplifies to:
1963 * xtime_nsec -= offset
1964 */
1965 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1966 /* NTP adjustment caused clocksource mult overflow */
1967 WARN_ON_ONCE(1);
1968 return;
1969 }
1970
1971 tk->tkr_mono.mult += mult_adj;
1972 tk->xtime_interval += interval;
1973 tk->tkr_mono.xtime_nsec -= offset;
1974}
1975
1976/*
1977 * Adjust the timekeeper's multiplier to the correct frequency
1978 * and also to reduce the accumulated error value.
1979 */
1980static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1981{
1982 u32 mult;
1983
1984 /*
1985 * Determine the multiplier from the current NTP tick length.
1986 * Avoid expensive division when the tick length doesn't change.
1987 */
1988 if (likely(tk->ntp_tick == ntp_tick_length())) {
1989 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1990 } else {
1991 tk->ntp_tick = ntp_tick_length();
1992 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1993 tk->xtime_remainder, tk->cycle_interval);
1994 }
1995
1996 /*
1997 * If the clock is behind the NTP time, increase the multiplier by 1
1998 * to catch up with it. If it's ahead and there was a remainder in the
1999 * tick division, the clock will slow down. Otherwise it will stay
2000 * ahead until the tick length changes to a non-divisible value.
2001 */
2002 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2003 mult += tk->ntp_err_mult;
2004
2005 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2006
2007 if (unlikely(tk->tkr_mono.clock->maxadj &&
2008 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2009 > tk->tkr_mono.clock->maxadj))) {
2010 printk_once(KERN_WARNING
2011 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2012 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2013 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2014 }
2015
2016 /*
2017 * It may be possible that when we entered this function, xtime_nsec
2018 * was very small. Further, if we're slightly speeding the clocksource
2019 * in the code above, its possible the required corrective factor to
2020 * xtime_nsec could cause it to underflow.
2021 *
2022 * Now, since we have already accumulated the second and the NTP
2023 * subsystem has been notified via second_overflow(), we need to skip
2024 * the next update.
2025 */
2026 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2027 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2028 tk->tkr_mono.shift;
2029 tk->xtime_sec--;
2030 tk->skip_second_overflow = 1;
2031 }
2032}
2033
2034/*
2035 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2036 *
2037 * Helper function that accumulates the nsecs greater than a second
2038 * from the xtime_nsec field to the xtime_secs field.
2039 * It also calls into the NTP code to handle leapsecond processing.
2040 */
2041static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2042{
2043 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2044 unsigned int clock_set = 0;
2045
2046 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2047 int leap;
2048
2049 tk->tkr_mono.xtime_nsec -= nsecps;
2050 tk->xtime_sec++;
2051
2052 /*
2053 * Skip NTP update if this second was accumulated before,
2054 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2055 */
2056 if (unlikely(tk->skip_second_overflow)) {
2057 tk->skip_second_overflow = 0;
2058 continue;
2059 }
2060
2061 /* Figure out if its a leap sec and apply if needed */
2062 leap = second_overflow(tk->xtime_sec);
2063 if (unlikely(leap)) {
2064 struct timespec64 ts;
2065
2066 tk->xtime_sec += leap;
2067
2068 ts.tv_sec = leap;
2069 ts.tv_nsec = 0;
2070 tk_set_wall_to_mono(tk,
2071 timespec64_sub(tk->wall_to_monotonic, ts));
2072
2073 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2074
2075 clock_set = TK_CLOCK_WAS_SET;
2076 }
2077 }
2078 return clock_set;
2079}
2080
2081/*
2082 * logarithmic_accumulation - shifted accumulation of cycles
2083 *
2084 * This functions accumulates a shifted interval of cycles into
2085 * a shifted interval nanoseconds. Allows for O(log) accumulation
2086 * loop.
2087 *
2088 * Returns the unconsumed cycles.
2089 */
2090static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2091 u32 shift, unsigned int *clock_set)
2092{
2093 u64 interval = tk->cycle_interval << shift;
2094 u64 snsec_per_sec;
2095
2096 /* If the offset is smaller than a shifted interval, do nothing */
2097 if (offset < interval)
2098 return offset;
2099
2100 /* Accumulate one shifted interval */
2101 offset -= interval;
2102 tk->tkr_mono.cycle_last += interval;
2103 tk->tkr_raw.cycle_last += interval;
2104
2105 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2106 *clock_set |= accumulate_nsecs_to_secs(tk);
2107
2108 /* Accumulate raw time */
2109 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2110 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2111 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2112 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2113 tk->raw_sec++;
2114 }
2115
2116 /* Accumulate error between NTP and clock interval */
2117 tk->ntp_error += tk->ntp_tick << shift;
2118 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2119 (tk->ntp_error_shift + shift);
2120
2121 return offset;
2122}
2123
2124/*
2125 * timekeeping_advance - Updates the timekeeper to the current time and
2126 * current NTP tick length
2127 */
2128static void timekeeping_advance(enum timekeeping_adv_mode mode)
2129{
2130 struct timekeeper *real_tk = &tk_core.timekeeper;
2131 struct timekeeper *tk = &shadow_timekeeper;
2132 u64 offset;
2133 int shift = 0, maxshift;
2134 unsigned int clock_set = 0;
2135 unsigned long flags;
2136
2137 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2138
2139 /* Make sure we're fully resumed: */
2140 if (unlikely(timekeeping_suspended))
2141 goto out;
2142
2143 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2144 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2145
2146 /* Check if there's really nothing to do */
2147 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2148 goto out;
2149
2150 /* Do some additional sanity checking */
2151 timekeeping_check_update(tk, offset);
2152
2153 /*
2154 * With NO_HZ we may have to accumulate many cycle_intervals
2155 * (think "ticks") worth of time at once. To do this efficiently,
2156 * we calculate the largest doubling multiple of cycle_intervals
2157 * that is smaller than the offset. We then accumulate that
2158 * chunk in one go, and then try to consume the next smaller
2159 * doubled multiple.
2160 */
2161 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2162 shift = max(0, shift);
2163 /* Bound shift to one less than what overflows tick_length */
2164 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2165 shift = min(shift, maxshift);
2166 while (offset >= tk->cycle_interval) {
2167 offset = logarithmic_accumulation(tk, offset, shift,
2168 &clock_set);
2169 if (offset < tk->cycle_interval<<shift)
2170 shift--;
2171 }
2172
2173 /* Adjust the multiplier to correct NTP error */
2174 timekeeping_adjust(tk, offset);
2175
2176 /*
2177 * Finally, make sure that after the rounding
2178 * xtime_nsec isn't larger than NSEC_PER_SEC
2179 */
2180 clock_set |= accumulate_nsecs_to_secs(tk);
2181
2182 write_seqcount_begin(&tk_core.seq);
2183 /*
2184 * Update the real timekeeper.
2185 *
2186 * We could avoid this memcpy by switching pointers, but that
2187 * requires changes to all other timekeeper usage sites as
2188 * well, i.e. move the timekeeper pointer getter into the
2189 * spinlocked/seqcount protected sections. And we trade this
2190 * memcpy under the tk_core.seq against one before we start
2191 * updating.
2192 */
2193 timekeeping_update(tk, clock_set);
2194 memcpy(real_tk, tk, sizeof(*tk));
2195 /* The memcpy must come last. Do not put anything here! */
2196 write_seqcount_end(&tk_core.seq);
2197out:
2198 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2199 if (clock_set)
2200 /* Have to call _delayed version, since in irq context*/
2201 clock_was_set_delayed();
2202}
2203
2204/**
2205 * update_wall_time - Uses the current clocksource to increment the wall time
2206 *
2207 */
2208void update_wall_time(void)
2209{
2210 timekeeping_advance(TK_ADV_TICK);
2211}
2212
2213/**
2214 * getboottime64 - Return the real time of system boot.
2215 * @ts: pointer to the timespec64 to be set
2216 *
2217 * Returns the wall-time of boot in a timespec64.
2218 *
2219 * This is based on the wall_to_monotonic offset and the total suspend
2220 * time. Calls to settimeofday will affect the value returned (which
2221 * basically means that however wrong your real time clock is at boot time,
2222 * you get the right time here).
2223 */
2224void getboottime64(struct timespec64 *ts)
2225{
2226 struct timekeeper *tk = &tk_core.timekeeper;
2227 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2228
2229 *ts = ktime_to_timespec64(t);
2230}
2231EXPORT_SYMBOL_GPL(getboottime64);
2232
2233void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2234{
2235 struct timekeeper *tk = &tk_core.timekeeper;
2236 unsigned int seq;
2237
2238 do {
2239 seq = read_seqcount_begin(&tk_core.seq);
2240
2241 *ts = tk_xtime(tk);
2242 } while (read_seqcount_retry(&tk_core.seq, seq));
2243}
2244EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2245
2246void ktime_get_coarse_ts64(struct timespec64 *ts)
2247{
2248 struct timekeeper *tk = &tk_core.timekeeper;
2249 struct timespec64 now, mono;
2250 unsigned int seq;
2251
2252 do {
2253 seq = read_seqcount_begin(&tk_core.seq);
2254
2255 now = tk_xtime(tk);
2256 mono = tk->wall_to_monotonic;
2257 } while (read_seqcount_retry(&tk_core.seq, seq));
2258
2259 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2260 now.tv_nsec + mono.tv_nsec);
2261}
2262EXPORT_SYMBOL(ktime_get_coarse_ts64);
2263
2264/*
2265 * Must hold jiffies_lock
2266 */
2267void do_timer(unsigned long ticks)
2268{
2269 jiffies_64 += ticks;
2270 calc_global_load();
2271}
2272
2273/**
2274 * ktime_get_update_offsets_now - hrtimer helper
2275 * @cwsseq: pointer to check and store the clock was set sequence number
2276 * @offs_real: pointer to storage for monotonic -> realtime offset
2277 * @offs_boot: pointer to storage for monotonic -> boottime offset
2278 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2279 *
2280 * Returns current monotonic time and updates the offsets if the
2281 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2282 * different.
2283 *
2284 * Called from hrtimer_interrupt() or retrigger_next_event()
2285 */
2286ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2287 ktime_t *offs_boot, ktime_t *offs_tai)
2288{
2289 struct timekeeper *tk = &tk_core.timekeeper;
2290 unsigned int seq;
2291 ktime_t base;
2292 u64 nsecs;
2293
2294 do {
2295 seq = read_seqcount_begin(&tk_core.seq);
2296
2297 base = tk->tkr_mono.base;
2298 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2299 base = ktime_add_ns(base, nsecs);
2300
2301 if (*cwsseq != tk->clock_was_set_seq) {
2302 *cwsseq = tk->clock_was_set_seq;
2303 *offs_real = tk->offs_real;
2304 *offs_boot = tk->offs_boot;
2305 *offs_tai = tk->offs_tai;
2306 }
2307
2308 /* Handle leapsecond insertion adjustments */
2309 if (unlikely(base >= tk->next_leap_ktime))
2310 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2311
2312 } while (read_seqcount_retry(&tk_core.seq, seq));
2313
2314 return base;
2315}
2316
2317/*
2318 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2319 */
2320static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2321{
2322 if (txc->modes & ADJ_ADJTIME) {
2323 /* singleshot must not be used with any other mode bits */
2324 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2325 return -EINVAL;
2326 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2327 !capable(CAP_SYS_TIME))
2328 return -EPERM;
2329 } else {
2330 /* In order to modify anything, you gotta be super-user! */
2331 if (txc->modes && !capable(CAP_SYS_TIME))
2332 return -EPERM;
2333 /*
2334 * if the quartz is off by more than 10% then
2335 * something is VERY wrong!
2336 */
2337 if (txc->modes & ADJ_TICK &&
2338 (txc->tick < 900000/USER_HZ ||
2339 txc->tick > 1100000/USER_HZ))
2340 return -EINVAL;
2341 }
2342
2343 if (txc->modes & ADJ_SETOFFSET) {
2344 /* In order to inject time, you gotta be super-user! */
2345 if (!capable(CAP_SYS_TIME))
2346 return -EPERM;
2347
2348 /*
2349 * Validate if a timespec/timeval used to inject a time
2350 * offset is valid. Offsets can be positive or negative, so
2351 * we don't check tv_sec. The value of the timeval/timespec
2352 * is the sum of its fields,but *NOTE*:
2353 * The field tv_usec/tv_nsec must always be non-negative and
2354 * we can't have more nanoseconds/microseconds than a second.
2355 */
2356 if (txc->time.tv_usec < 0)
2357 return -EINVAL;
2358
2359 if (txc->modes & ADJ_NANO) {
2360 if (txc->time.tv_usec >= NSEC_PER_SEC)
2361 return -EINVAL;
2362 } else {
2363 if (txc->time.tv_usec >= USEC_PER_SEC)
2364 return -EINVAL;
2365 }
2366 }
2367
2368 /*
2369 * Check for potential multiplication overflows that can
2370 * only happen on 64-bit systems:
2371 */
2372 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2373 if (LLONG_MIN / PPM_SCALE > txc->freq)
2374 return -EINVAL;
2375 if (LLONG_MAX / PPM_SCALE < txc->freq)
2376 return -EINVAL;
2377 }
2378
2379 return 0;
2380}
2381
2382
2383/**
2384 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2385 */
2386int do_adjtimex(struct __kernel_timex *txc)
2387{
2388 struct timekeeper *tk = &tk_core.timekeeper;
2389 struct audit_ntp_data ad;
2390 unsigned long flags;
2391 struct timespec64 ts;
2392 s32 orig_tai, tai;
2393 int ret;
2394
2395 /* Validate the data before disabling interrupts */
2396 ret = timekeeping_validate_timex(txc);
2397 if (ret)
2398 return ret;
2399
2400 if (txc->modes & ADJ_SETOFFSET) {
2401 struct timespec64 delta;
2402 delta.tv_sec = txc->time.tv_sec;
2403 delta.tv_nsec = txc->time.tv_usec;
2404 if (!(txc->modes & ADJ_NANO))
2405 delta.tv_nsec *= 1000;
2406 ret = timekeeping_inject_offset(&delta);
2407 if (ret)
2408 return ret;
2409
2410 audit_tk_injoffset(delta);
2411 }
2412
2413 audit_ntp_init(&ad);
2414
2415 ktime_get_real_ts64(&ts);
2416
2417 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2418 write_seqcount_begin(&tk_core.seq);
2419
2420 orig_tai = tai = tk->tai_offset;
2421 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2422
2423 if (tai != orig_tai) {
2424 __timekeeping_set_tai_offset(tk, tai);
2425 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2426 }
2427 tk_update_leap_state(tk);
2428
2429 write_seqcount_end(&tk_core.seq);
2430 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2431
2432 audit_ntp_log(&ad);
2433
2434 /* Update the multiplier immediately if frequency was set directly */
2435 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2436 timekeeping_advance(TK_ADV_FREQ);
2437
2438 if (tai != orig_tai)
2439 clock_was_set();
2440
2441 ntp_notify_cmos_timer();
2442
2443 return ret;
2444}
2445
2446#ifdef CONFIG_NTP_PPS
2447/**
2448 * hardpps() - Accessor function to NTP __hardpps function
2449 */
2450void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2451{
2452 unsigned long flags;
2453
2454 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2455 write_seqcount_begin(&tk_core.seq);
2456
2457 __hardpps(phase_ts, raw_ts);
2458
2459 write_seqcount_end(&tk_core.seq);
2460 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2461}
2462EXPORT_SYMBOL(hardpps);
2463#endif /* CONFIG_NTP_PPS */
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/nmi.h>
18#include <linux/sched.h>
19#include <linux/sched/loadavg.h>
20#include <linux/syscore_ops.h>
21#include <linux/clocksource.h>
22#include <linux/jiffies.h>
23#include <linux/time.h>
24#include <linux/tick.h>
25#include <linux/stop_machine.h>
26#include <linux/pvclock_gtod.h>
27#include <linux/compiler.h>
28
29#include "tick-internal.h"
30#include "ntp_internal.h"
31#include "timekeeping_internal.h"
32
33#define TK_CLEAR_NTP (1 << 0)
34#define TK_MIRROR (1 << 1)
35#define TK_CLOCK_WAS_SET (1 << 2)
36
37/*
38 * The most important data for readout fits into a single 64 byte
39 * cache line.
40 */
41static struct {
42 seqcount_t seq;
43 struct timekeeper timekeeper;
44} tk_core ____cacheline_aligned;
45
46static DEFINE_RAW_SPINLOCK(timekeeper_lock);
47static struct timekeeper shadow_timekeeper;
48
49/**
50 * struct tk_fast - NMI safe timekeeper
51 * @seq: Sequence counter for protecting updates. The lowest bit
52 * is the index for the tk_read_base array
53 * @base: tk_read_base array. Access is indexed by the lowest bit of
54 * @seq.
55 *
56 * See @update_fast_timekeeper() below.
57 */
58struct tk_fast {
59 seqcount_t seq;
60 struct tk_read_base base[2];
61};
62
63/* Suspend-time cycles value for halted fast timekeeper. */
64static u64 cycles_at_suspend;
65
66static u64 dummy_clock_read(struct clocksource *cs)
67{
68 return cycles_at_suspend;
69}
70
71static struct clocksource dummy_clock = {
72 .read = dummy_clock_read,
73};
74
75static struct tk_fast tk_fast_mono ____cacheline_aligned = {
76 .base[0] = { .clock = &dummy_clock, },
77 .base[1] = { .clock = &dummy_clock, },
78};
79
80static struct tk_fast tk_fast_raw ____cacheline_aligned = {
81 .base[0] = { .clock = &dummy_clock, },
82 .base[1] = { .clock = &dummy_clock, },
83};
84
85/* flag for if timekeeping is suspended */
86int __read_mostly timekeeping_suspended;
87
88static inline void tk_normalize_xtime(struct timekeeper *tk)
89{
90 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
91 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
92 tk->xtime_sec++;
93 }
94 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
95 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
96 tk->raw_sec++;
97 }
98}
99
100static inline struct timespec64 tk_xtime(struct timekeeper *tk)
101{
102 struct timespec64 ts;
103
104 ts.tv_sec = tk->xtime_sec;
105 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
106 return ts;
107}
108
109static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
110{
111 tk->xtime_sec = ts->tv_sec;
112 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
113}
114
115static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
116{
117 tk->xtime_sec += ts->tv_sec;
118 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
119 tk_normalize_xtime(tk);
120}
121
122static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
123{
124 struct timespec64 tmp;
125
126 /*
127 * Verify consistency of: offset_real = -wall_to_monotonic
128 * before modifying anything
129 */
130 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
131 -tk->wall_to_monotonic.tv_nsec);
132 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
133 tk->wall_to_monotonic = wtm;
134 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
135 tk->offs_real = timespec64_to_ktime(tmp);
136 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
137}
138
139static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
140{
141 tk->offs_boot = ktime_add(tk->offs_boot, delta);
142}
143
144/*
145 * tk_clock_read - atomic clocksource read() helper
146 *
147 * This helper is necessary to use in the read paths because, while the
148 * seqlock ensures we don't return a bad value while structures are updated,
149 * it doesn't protect from potential crashes. There is the possibility that
150 * the tkr's clocksource may change between the read reference, and the
151 * clock reference passed to the read function. This can cause crashes if
152 * the wrong clocksource is passed to the wrong read function.
153 * This isn't necessary to use when holding the timekeeper_lock or doing
154 * a read of the fast-timekeeper tkrs (which is protected by its own locking
155 * and update logic).
156 */
157static inline u64 tk_clock_read(struct tk_read_base *tkr)
158{
159 struct clocksource *clock = READ_ONCE(tkr->clock);
160
161 return clock->read(clock);
162}
163
164#ifdef CONFIG_DEBUG_TIMEKEEPING
165#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
166
167static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
168{
169
170 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
171 const char *name = tk->tkr_mono.clock->name;
172
173 if (offset > max_cycles) {
174 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
175 offset, name, max_cycles);
176 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
177 } else {
178 if (offset > (max_cycles >> 1)) {
179 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
180 offset, name, max_cycles >> 1);
181 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
182 }
183 }
184
185 if (tk->underflow_seen) {
186 if (jiffies - tk->last_warning > WARNING_FREQ) {
187 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
188 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
189 printk_deferred(" Your kernel is probably still fine.\n");
190 tk->last_warning = jiffies;
191 }
192 tk->underflow_seen = 0;
193 }
194
195 if (tk->overflow_seen) {
196 if (jiffies - tk->last_warning > WARNING_FREQ) {
197 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
198 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
199 printk_deferred(" Your kernel is probably still fine.\n");
200 tk->last_warning = jiffies;
201 }
202 tk->overflow_seen = 0;
203 }
204}
205
206static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
207{
208 struct timekeeper *tk = &tk_core.timekeeper;
209 u64 now, last, mask, max, delta;
210 unsigned int seq;
211
212 /*
213 * Since we're called holding a seqlock, the data may shift
214 * under us while we're doing the calculation. This can cause
215 * false positives, since we'd note a problem but throw the
216 * results away. So nest another seqlock here to atomically
217 * grab the points we are checking with.
218 */
219 do {
220 seq = read_seqcount_begin(&tk_core.seq);
221 now = tk_clock_read(tkr);
222 last = tkr->cycle_last;
223 mask = tkr->mask;
224 max = tkr->clock->max_cycles;
225 } while (read_seqcount_retry(&tk_core.seq, seq));
226
227 delta = clocksource_delta(now, last, mask);
228
229 /*
230 * Try to catch underflows by checking if we are seeing small
231 * mask-relative negative values.
232 */
233 if (unlikely((~delta & mask) < (mask >> 3))) {
234 tk->underflow_seen = 1;
235 delta = 0;
236 }
237
238 /* Cap delta value to the max_cycles values to avoid mult overflows */
239 if (unlikely(delta > max)) {
240 tk->overflow_seen = 1;
241 delta = tkr->clock->max_cycles;
242 }
243
244 return delta;
245}
246#else
247static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
248{
249}
250static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
251{
252 u64 cycle_now, delta;
253
254 /* read clocksource */
255 cycle_now = tk_clock_read(tkr);
256
257 /* calculate the delta since the last update_wall_time */
258 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
259
260 return delta;
261}
262#endif
263
264/**
265 * tk_setup_internals - Set up internals to use clocksource clock.
266 *
267 * @tk: The target timekeeper to setup.
268 * @clock: Pointer to clocksource.
269 *
270 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
271 * pair and interval request.
272 *
273 * Unless you're the timekeeping code, you should not be using this!
274 */
275static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
276{
277 u64 interval;
278 u64 tmp, ntpinterval;
279 struct clocksource *old_clock;
280
281 ++tk->cs_was_changed_seq;
282 old_clock = tk->tkr_mono.clock;
283 tk->tkr_mono.clock = clock;
284 tk->tkr_mono.mask = clock->mask;
285 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
286
287 tk->tkr_raw.clock = clock;
288 tk->tkr_raw.mask = clock->mask;
289 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
290
291 /* Do the ns -> cycle conversion first, using original mult */
292 tmp = NTP_INTERVAL_LENGTH;
293 tmp <<= clock->shift;
294 ntpinterval = tmp;
295 tmp += clock->mult/2;
296 do_div(tmp, clock->mult);
297 if (tmp == 0)
298 tmp = 1;
299
300 interval = (u64) tmp;
301 tk->cycle_interval = interval;
302
303 /* Go back from cycles -> shifted ns */
304 tk->xtime_interval = interval * clock->mult;
305 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
306 tk->raw_interval = interval * clock->mult;
307
308 /* if changing clocks, convert xtime_nsec shift units */
309 if (old_clock) {
310 int shift_change = clock->shift - old_clock->shift;
311 if (shift_change < 0) {
312 tk->tkr_mono.xtime_nsec >>= -shift_change;
313 tk->tkr_raw.xtime_nsec >>= -shift_change;
314 } else {
315 tk->tkr_mono.xtime_nsec <<= shift_change;
316 tk->tkr_raw.xtime_nsec <<= shift_change;
317 }
318 }
319
320 tk->tkr_mono.shift = clock->shift;
321 tk->tkr_raw.shift = clock->shift;
322
323 tk->ntp_error = 0;
324 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
325 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
326
327 /*
328 * The timekeeper keeps its own mult values for the currently
329 * active clocksource. These value will be adjusted via NTP
330 * to counteract clock drifting.
331 */
332 tk->tkr_mono.mult = clock->mult;
333 tk->tkr_raw.mult = clock->mult;
334 tk->ntp_err_mult = 0;
335 tk->skip_second_overflow = 0;
336}
337
338/* Timekeeper helper functions. */
339
340#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
341static u32 default_arch_gettimeoffset(void) { return 0; }
342u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
343#else
344static inline u32 arch_gettimeoffset(void) { return 0; }
345#endif
346
347static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta)
348{
349 u64 nsec;
350
351 nsec = delta * tkr->mult + tkr->xtime_nsec;
352 nsec >>= tkr->shift;
353
354 /* If arch requires, add in get_arch_timeoffset() */
355 return nsec + arch_gettimeoffset();
356}
357
358static inline u64 timekeeping_get_ns(struct tk_read_base *tkr)
359{
360 u64 delta;
361
362 delta = timekeeping_get_delta(tkr);
363 return timekeeping_delta_to_ns(tkr, delta);
364}
365
366static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles)
367{
368 u64 delta;
369
370 /* calculate the delta since the last update_wall_time */
371 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
372 return timekeeping_delta_to_ns(tkr, delta);
373}
374
375/**
376 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
377 * @tkr: Timekeeping readout base from which we take the update
378 *
379 * We want to use this from any context including NMI and tracing /
380 * instrumenting the timekeeping code itself.
381 *
382 * Employ the latch technique; see @raw_write_seqcount_latch.
383 *
384 * So if a NMI hits the update of base[0] then it will use base[1]
385 * which is still consistent. In the worst case this can result is a
386 * slightly wrong timestamp (a few nanoseconds). See
387 * @ktime_get_mono_fast_ns.
388 */
389static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
390{
391 struct tk_read_base *base = tkf->base;
392
393 /* Force readers off to base[1] */
394 raw_write_seqcount_latch(&tkf->seq);
395
396 /* Update base[0] */
397 memcpy(base, tkr, sizeof(*base));
398
399 /* Force readers back to base[0] */
400 raw_write_seqcount_latch(&tkf->seq);
401
402 /* Update base[1] */
403 memcpy(base + 1, base, sizeof(*base));
404}
405
406/**
407 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
408 *
409 * This timestamp is not guaranteed to be monotonic across an update.
410 * The timestamp is calculated by:
411 *
412 * now = base_mono + clock_delta * slope
413 *
414 * So if the update lowers the slope, readers who are forced to the
415 * not yet updated second array are still using the old steeper slope.
416 *
417 * tmono
418 * ^
419 * | o n
420 * | o n
421 * | u
422 * | o
423 * |o
424 * |12345678---> reader order
425 *
426 * o = old slope
427 * u = update
428 * n = new slope
429 *
430 * So reader 6 will observe time going backwards versus reader 5.
431 *
432 * While other CPUs are likely to be able observe that, the only way
433 * for a CPU local observation is when an NMI hits in the middle of
434 * the update. Timestamps taken from that NMI context might be ahead
435 * of the following timestamps. Callers need to be aware of that and
436 * deal with it.
437 */
438static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
439{
440 struct tk_read_base *tkr;
441 unsigned int seq;
442 u64 now;
443
444 do {
445 seq = raw_read_seqcount_latch(&tkf->seq);
446 tkr = tkf->base + (seq & 0x01);
447 now = ktime_to_ns(tkr->base);
448
449 now += timekeeping_delta_to_ns(tkr,
450 clocksource_delta(
451 tk_clock_read(tkr),
452 tkr->cycle_last,
453 tkr->mask));
454 } while (read_seqcount_retry(&tkf->seq, seq));
455
456 return now;
457}
458
459u64 ktime_get_mono_fast_ns(void)
460{
461 return __ktime_get_fast_ns(&tk_fast_mono);
462}
463EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
464
465u64 ktime_get_raw_fast_ns(void)
466{
467 return __ktime_get_fast_ns(&tk_fast_raw);
468}
469EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
470
471/**
472 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
473 *
474 * To keep it NMI safe since we're accessing from tracing, we're not using a
475 * separate timekeeper with updates to monotonic clock and boot offset
476 * protected with seqlocks. This has the following minor side effects:
477 *
478 * (1) Its possible that a timestamp be taken after the boot offset is updated
479 * but before the timekeeper is updated. If this happens, the new boot offset
480 * is added to the old timekeeping making the clock appear to update slightly
481 * earlier:
482 * CPU 0 CPU 1
483 * timekeeping_inject_sleeptime64()
484 * __timekeeping_inject_sleeptime(tk, delta);
485 * timestamp();
486 * timekeeping_update(tk, TK_CLEAR_NTP...);
487 *
488 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
489 * partially updated. Since the tk->offs_boot update is a rare event, this
490 * should be a rare occurrence which postprocessing should be able to handle.
491 */
492u64 notrace ktime_get_boot_fast_ns(void)
493{
494 struct timekeeper *tk = &tk_core.timekeeper;
495
496 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
497}
498EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
499
500
501/*
502 * See comment for __ktime_get_fast_ns() vs. timestamp ordering
503 */
504static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf)
505{
506 struct tk_read_base *tkr;
507 unsigned int seq;
508 u64 now;
509
510 do {
511 seq = raw_read_seqcount_latch(&tkf->seq);
512 tkr = tkf->base + (seq & 0x01);
513 now = ktime_to_ns(tkr->base_real);
514
515 now += timekeeping_delta_to_ns(tkr,
516 clocksource_delta(
517 tk_clock_read(tkr),
518 tkr->cycle_last,
519 tkr->mask));
520 } while (read_seqcount_retry(&tkf->seq, seq));
521
522 return now;
523}
524
525/**
526 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
527 */
528u64 ktime_get_real_fast_ns(void)
529{
530 return __ktime_get_real_fast_ns(&tk_fast_mono);
531}
532EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
533
534/**
535 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
536 * @tk: Timekeeper to snapshot.
537 *
538 * It generally is unsafe to access the clocksource after timekeeping has been
539 * suspended, so take a snapshot of the readout base of @tk and use it as the
540 * fast timekeeper's readout base while suspended. It will return the same
541 * number of cycles every time until timekeeping is resumed at which time the
542 * proper readout base for the fast timekeeper will be restored automatically.
543 */
544static void halt_fast_timekeeper(struct timekeeper *tk)
545{
546 static struct tk_read_base tkr_dummy;
547 struct tk_read_base *tkr = &tk->tkr_mono;
548
549 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
550 cycles_at_suspend = tk_clock_read(tkr);
551 tkr_dummy.clock = &dummy_clock;
552 tkr_dummy.base_real = tkr->base + tk->offs_real;
553 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
554
555 tkr = &tk->tkr_raw;
556 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
557 tkr_dummy.clock = &dummy_clock;
558 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
559}
560
561static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
562
563static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
564{
565 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
566}
567
568/**
569 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
570 */
571int pvclock_gtod_register_notifier(struct notifier_block *nb)
572{
573 struct timekeeper *tk = &tk_core.timekeeper;
574 unsigned long flags;
575 int ret;
576
577 raw_spin_lock_irqsave(&timekeeper_lock, flags);
578 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
579 update_pvclock_gtod(tk, true);
580 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
581
582 return ret;
583}
584EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
585
586/**
587 * pvclock_gtod_unregister_notifier - unregister a pvclock
588 * timedata update listener
589 */
590int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
591{
592 unsigned long flags;
593 int ret;
594
595 raw_spin_lock_irqsave(&timekeeper_lock, flags);
596 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
597 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
598
599 return ret;
600}
601EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
602
603/*
604 * tk_update_leap_state - helper to update the next_leap_ktime
605 */
606static inline void tk_update_leap_state(struct timekeeper *tk)
607{
608 tk->next_leap_ktime = ntp_get_next_leap();
609 if (tk->next_leap_ktime != KTIME_MAX)
610 /* Convert to monotonic time */
611 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
612}
613
614/*
615 * Update the ktime_t based scalar nsec members of the timekeeper
616 */
617static inline void tk_update_ktime_data(struct timekeeper *tk)
618{
619 u64 seconds;
620 u32 nsec;
621
622 /*
623 * The xtime based monotonic readout is:
624 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
625 * The ktime based monotonic readout is:
626 * nsec = base_mono + now();
627 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
628 */
629 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
630 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
631 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
632
633 /*
634 * The sum of the nanoseconds portions of xtime and
635 * wall_to_monotonic can be greater/equal one second. Take
636 * this into account before updating tk->ktime_sec.
637 */
638 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
639 if (nsec >= NSEC_PER_SEC)
640 seconds++;
641 tk->ktime_sec = seconds;
642
643 /* Update the monotonic raw base */
644 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
645}
646
647/* must hold timekeeper_lock */
648static void timekeeping_update(struct timekeeper *tk, unsigned int action)
649{
650 if (action & TK_CLEAR_NTP) {
651 tk->ntp_error = 0;
652 ntp_clear();
653 }
654
655 tk_update_leap_state(tk);
656 tk_update_ktime_data(tk);
657
658 update_vsyscall(tk);
659 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
660
661 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
662 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
663 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
664
665 if (action & TK_CLOCK_WAS_SET)
666 tk->clock_was_set_seq++;
667 /*
668 * The mirroring of the data to the shadow-timekeeper needs
669 * to happen last here to ensure we don't over-write the
670 * timekeeper structure on the next update with stale data
671 */
672 if (action & TK_MIRROR)
673 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
674 sizeof(tk_core.timekeeper));
675}
676
677/**
678 * timekeeping_forward_now - update clock to the current time
679 *
680 * Forward the current clock to update its state since the last call to
681 * update_wall_time(). This is useful before significant clock changes,
682 * as it avoids having to deal with this time offset explicitly.
683 */
684static void timekeeping_forward_now(struct timekeeper *tk)
685{
686 u64 cycle_now, delta;
687
688 cycle_now = tk_clock_read(&tk->tkr_mono);
689 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
690 tk->tkr_mono.cycle_last = cycle_now;
691 tk->tkr_raw.cycle_last = cycle_now;
692
693 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
694
695 /* If arch requires, add in get_arch_timeoffset() */
696 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
697
698
699 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
700
701 /* If arch requires, add in get_arch_timeoffset() */
702 tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
703
704 tk_normalize_xtime(tk);
705}
706
707/**
708 * __getnstimeofday64 - Returns the time of day in a timespec64.
709 * @ts: pointer to the timespec to be set
710 *
711 * Updates the time of day in the timespec.
712 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
713 */
714int __getnstimeofday64(struct timespec64 *ts)
715{
716 struct timekeeper *tk = &tk_core.timekeeper;
717 unsigned long seq;
718 u64 nsecs;
719
720 do {
721 seq = read_seqcount_begin(&tk_core.seq);
722
723 ts->tv_sec = tk->xtime_sec;
724 nsecs = timekeeping_get_ns(&tk->tkr_mono);
725
726 } while (read_seqcount_retry(&tk_core.seq, seq));
727
728 ts->tv_nsec = 0;
729 timespec64_add_ns(ts, nsecs);
730
731 /*
732 * Do not bail out early, in case there were callers still using
733 * the value, even in the face of the WARN_ON.
734 */
735 if (unlikely(timekeeping_suspended))
736 return -EAGAIN;
737 return 0;
738}
739EXPORT_SYMBOL(__getnstimeofday64);
740
741/**
742 * getnstimeofday64 - Returns the time of day in a timespec64.
743 * @ts: pointer to the timespec64 to be set
744 *
745 * Returns the time of day in a timespec64 (WARN if suspended).
746 */
747void getnstimeofday64(struct timespec64 *ts)
748{
749 WARN_ON(__getnstimeofday64(ts));
750}
751EXPORT_SYMBOL(getnstimeofday64);
752
753ktime_t ktime_get(void)
754{
755 struct timekeeper *tk = &tk_core.timekeeper;
756 unsigned int seq;
757 ktime_t base;
758 u64 nsecs;
759
760 WARN_ON(timekeeping_suspended);
761
762 do {
763 seq = read_seqcount_begin(&tk_core.seq);
764 base = tk->tkr_mono.base;
765 nsecs = timekeeping_get_ns(&tk->tkr_mono);
766
767 } while (read_seqcount_retry(&tk_core.seq, seq));
768
769 return ktime_add_ns(base, nsecs);
770}
771EXPORT_SYMBOL_GPL(ktime_get);
772
773u32 ktime_get_resolution_ns(void)
774{
775 struct timekeeper *tk = &tk_core.timekeeper;
776 unsigned int seq;
777 u32 nsecs;
778
779 WARN_ON(timekeeping_suspended);
780
781 do {
782 seq = read_seqcount_begin(&tk_core.seq);
783 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
784 } while (read_seqcount_retry(&tk_core.seq, seq));
785
786 return nsecs;
787}
788EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
789
790static ktime_t *offsets[TK_OFFS_MAX] = {
791 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
792 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
793 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
794};
795
796ktime_t ktime_get_with_offset(enum tk_offsets offs)
797{
798 struct timekeeper *tk = &tk_core.timekeeper;
799 unsigned int seq;
800 ktime_t base, *offset = offsets[offs];
801 u64 nsecs;
802
803 WARN_ON(timekeeping_suspended);
804
805 do {
806 seq = read_seqcount_begin(&tk_core.seq);
807 base = ktime_add(tk->tkr_mono.base, *offset);
808 nsecs = timekeeping_get_ns(&tk->tkr_mono);
809
810 } while (read_seqcount_retry(&tk_core.seq, seq));
811
812 return ktime_add_ns(base, nsecs);
813
814}
815EXPORT_SYMBOL_GPL(ktime_get_with_offset);
816
817/**
818 * ktime_mono_to_any() - convert mononotic time to any other time
819 * @tmono: time to convert.
820 * @offs: which offset to use
821 */
822ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
823{
824 ktime_t *offset = offsets[offs];
825 unsigned long seq;
826 ktime_t tconv;
827
828 do {
829 seq = read_seqcount_begin(&tk_core.seq);
830 tconv = ktime_add(tmono, *offset);
831 } while (read_seqcount_retry(&tk_core.seq, seq));
832
833 return tconv;
834}
835EXPORT_SYMBOL_GPL(ktime_mono_to_any);
836
837/**
838 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
839 */
840ktime_t ktime_get_raw(void)
841{
842 struct timekeeper *tk = &tk_core.timekeeper;
843 unsigned int seq;
844 ktime_t base;
845 u64 nsecs;
846
847 do {
848 seq = read_seqcount_begin(&tk_core.seq);
849 base = tk->tkr_raw.base;
850 nsecs = timekeeping_get_ns(&tk->tkr_raw);
851
852 } while (read_seqcount_retry(&tk_core.seq, seq));
853
854 return ktime_add_ns(base, nsecs);
855}
856EXPORT_SYMBOL_GPL(ktime_get_raw);
857
858/**
859 * ktime_get_ts64 - get the monotonic clock in timespec64 format
860 * @ts: pointer to timespec variable
861 *
862 * The function calculates the monotonic clock from the realtime
863 * clock and the wall_to_monotonic offset and stores the result
864 * in normalized timespec64 format in the variable pointed to by @ts.
865 */
866void ktime_get_ts64(struct timespec64 *ts)
867{
868 struct timekeeper *tk = &tk_core.timekeeper;
869 struct timespec64 tomono;
870 unsigned int seq;
871 u64 nsec;
872
873 WARN_ON(timekeeping_suspended);
874
875 do {
876 seq = read_seqcount_begin(&tk_core.seq);
877 ts->tv_sec = tk->xtime_sec;
878 nsec = timekeeping_get_ns(&tk->tkr_mono);
879 tomono = tk->wall_to_monotonic;
880
881 } while (read_seqcount_retry(&tk_core.seq, seq));
882
883 ts->tv_sec += tomono.tv_sec;
884 ts->tv_nsec = 0;
885 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
886}
887EXPORT_SYMBOL_GPL(ktime_get_ts64);
888
889/**
890 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
891 *
892 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
893 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
894 * works on both 32 and 64 bit systems. On 32 bit systems the readout
895 * covers ~136 years of uptime which should be enough to prevent
896 * premature wrap arounds.
897 */
898time64_t ktime_get_seconds(void)
899{
900 struct timekeeper *tk = &tk_core.timekeeper;
901
902 WARN_ON(timekeeping_suspended);
903 return tk->ktime_sec;
904}
905EXPORT_SYMBOL_GPL(ktime_get_seconds);
906
907/**
908 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
909 *
910 * Returns the wall clock seconds since 1970. This replaces the
911 * get_seconds() interface which is not y2038 safe on 32bit systems.
912 *
913 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
914 * 32bit systems the access must be protected with the sequence
915 * counter to provide "atomic" access to the 64bit tk->xtime_sec
916 * value.
917 */
918time64_t ktime_get_real_seconds(void)
919{
920 struct timekeeper *tk = &tk_core.timekeeper;
921 time64_t seconds;
922 unsigned int seq;
923
924 if (IS_ENABLED(CONFIG_64BIT))
925 return tk->xtime_sec;
926
927 do {
928 seq = read_seqcount_begin(&tk_core.seq);
929 seconds = tk->xtime_sec;
930
931 } while (read_seqcount_retry(&tk_core.seq, seq));
932
933 return seconds;
934}
935EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
936
937/**
938 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
939 * but without the sequence counter protect. This internal function
940 * is called just when timekeeping lock is already held.
941 */
942time64_t __ktime_get_real_seconds(void)
943{
944 struct timekeeper *tk = &tk_core.timekeeper;
945
946 return tk->xtime_sec;
947}
948
949/**
950 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
951 * @systime_snapshot: pointer to struct receiving the system time snapshot
952 */
953void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
954{
955 struct timekeeper *tk = &tk_core.timekeeper;
956 unsigned long seq;
957 ktime_t base_raw;
958 ktime_t base_real;
959 u64 nsec_raw;
960 u64 nsec_real;
961 u64 now;
962
963 WARN_ON_ONCE(timekeeping_suspended);
964
965 do {
966 seq = read_seqcount_begin(&tk_core.seq);
967 now = tk_clock_read(&tk->tkr_mono);
968 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
969 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
970 base_real = ktime_add(tk->tkr_mono.base,
971 tk_core.timekeeper.offs_real);
972 base_raw = tk->tkr_raw.base;
973 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
974 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
975 } while (read_seqcount_retry(&tk_core.seq, seq));
976
977 systime_snapshot->cycles = now;
978 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
979 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
980}
981EXPORT_SYMBOL_GPL(ktime_get_snapshot);
982
983/* Scale base by mult/div checking for overflow */
984static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
985{
986 u64 tmp, rem;
987
988 tmp = div64_u64_rem(*base, div, &rem);
989
990 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
991 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
992 return -EOVERFLOW;
993 tmp *= mult;
994 rem *= mult;
995
996 do_div(rem, div);
997 *base = tmp + rem;
998 return 0;
999}
1000
1001/**
1002 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1003 * @history: Snapshot representing start of history
1004 * @partial_history_cycles: Cycle offset into history (fractional part)
1005 * @total_history_cycles: Total history length in cycles
1006 * @discontinuity: True indicates clock was set on history period
1007 * @ts: Cross timestamp that should be adjusted using
1008 * partial/total ratio
1009 *
1010 * Helper function used by get_device_system_crosststamp() to correct the
1011 * crosstimestamp corresponding to the start of the current interval to the
1012 * system counter value (timestamp point) provided by the driver. The
1013 * total_history_* quantities are the total history starting at the provided
1014 * reference point and ending at the start of the current interval. The cycle
1015 * count between the driver timestamp point and the start of the current
1016 * interval is partial_history_cycles.
1017 */
1018static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1019 u64 partial_history_cycles,
1020 u64 total_history_cycles,
1021 bool discontinuity,
1022 struct system_device_crosststamp *ts)
1023{
1024 struct timekeeper *tk = &tk_core.timekeeper;
1025 u64 corr_raw, corr_real;
1026 bool interp_forward;
1027 int ret;
1028
1029 if (total_history_cycles == 0 || partial_history_cycles == 0)
1030 return 0;
1031
1032 /* Interpolate shortest distance from beginning or end of history */
1033 interp_forward = partial_history_cycles > total_history_cycles / 2;
1034 partial_history_cycles = interp_forward ?
1035 total_history_cycles - partial_history_cycles :
1036 partial_history_cycles;
1037
1038 /*
1039 * Scale the monotonic raw time delta by:
1040 * partial_history_cycles / total_history_cycles
1041 */
1042 corr_raw = (u64)ktime_to_ns(
1043 ktime_sub(ts->sys_monoraw, history->raw));
1044 ret = scale64_check_overflow(partial_history_cycles,
1045 total_history_cycles, &corr_raw);
1046 if (ret)
1047 return ret;
1048
1049 /*
1050 * If there is a discontinuity in the history, scale monotonic raw
1051 * correction by:
1052 * mult(real)/mult(raw) yielding the realtime correction
1053 * Otherwise, calculate the realtime correction similar to monotonic
1054 * raw calculation
1055 */
1056 if (discontinuity) {
1057 corr_real = mul_u64_u32_div
1058 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1059 } else {
1060 corr_real = (u64)ktime_to_ns(
1061 ktime_sub(ts->sys_realtime, history->real));
1062 ret = scale64_check_overflow(partial_history_cycles,
1063 total_history_cycles, &corr_real);
1064 if (ret)
1065 return ret;
1066 }
1067
1068 /* Fixup monotonic raw and real time time values */
1069 if (interp_forward) {
1070 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1071 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1072 } else {
1073 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1074 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1075 }
1076
1077 return 0;
1078}
1079
1080/*
1081 * cycle_between - true if test occurs chronologically between before and after
1082 */
1083static bool cycle_between(u64 before, u64 test, u64 after)
1084{
1085 if (test > before && test < after)
1086 return true;
1087 if (test < before && before > after)
1088 return true;
1089 return false;
1090}
1091
1092/**
1093 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1094 * @get_time_fn: Callback to get simultaneous device time and
1095 * system counter from the device driver
1096 * @ctx: Context passed to get_time_fn()
1097 * @history_begin: Historical reference point used to interpolate system
1098 * time when counter provided by the driver is before the current interval
1099 * @xtstamp: Receives simultaneously captured system and device time
1100 *
1101 * Reads a timestamp from a device and correlates it to system time
1102 */
1103int get_device_system_crosststamp(int (*get_time_fn)
1104 (ktime_t *device_time,
1105 struct system_counterval_t *sys_counterval,
1106 void *ctx),
1107 void *ctx,
1108 struct system_time_snapshot *history_begin,
1109 struct system_device_crosststamp *xtstamp)
1110{
1111 struct system_counterval_t system_counterval;
1112 struct timekeeper *tk = &tk_core.timekeeper;
1113 u64 cycles, now, interval_start;
1114 unsigned int clock_was_set_seq = 0;
1115 ktime_t base_real, base_raw;
1116 u64 nsec_real, nsec_raw;
1117 u8 cs_was_changed_seq;
1118 unsigned long seq;
1119 bool do_interp;
1120 int ret;
1121
1122 do {
1123 seq = read_seqcount_begin(&tk_core.seq);
1124 /*
1125 * Try to synchronously capture device time and a system
1126 * counter value calling back into the device driver
1127 */
1128 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1129 if (ret)
1130 return ret;
1131
1132 /*
1133 * Verify that the clocksource associated with the captured
1134 * system counter value is the same as the currently installed
1135 * timekeeper clocksource
1136 */
1137 if (tk->tkr_mono.clock != system_counterval.cs)
1138 return -ENODEV;
1139 cycles = system_counterval.cycles;
1140
1141 /*
1142 * Check whether the system counter value provided by the
1143 * device driver is on the current timekeeping interval.
1144 */
1145 now = tk_clock_read(&tk->tkr_mono);
1146 interval_start = tk->tkr_mono.cycle_last;
1147 if (!cycle_between(interval_start, cycles, now)) {
1148 clock_was_set_seq = tk->clock_was_set_seq;
1149 cs_was_changed_seq = tk->cs_was_changed_seq;
1150 cycles = interval_start;
1151 do_interp = true;
1152 } else {
1153 do_interp = false;
1154 }
1155
1156 base_real = ktime_add(tk->tkr_mono.base,
1157 tk_core.timekeeper.offs_real);
1158 base_raw = tk->tkr_raw.base;
1159
1160 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1161 system_counterval.cycles);
1162 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1163 system_counterval.cycles);
1164 } while (read_seqcount_retry(&tk_core.seq, seq));
1165
1166 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1167 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1168
1169 /*
1170 * Interpolate if necessary, adjusting back from the start of the
1171 * current interval
1172 */
1173 if (do_interp) {
1174 u64 partial_history_cycles, total_history_cycles;
1175 bool discontinuity;
1176
1177 /*
1178 * Check that the counter value occurs after the provided
1179 * history reference and that the history doesn't cross a
1180 * clocksource change
1181 */
1182 if (!history_begin ||
1183 !cycle_between(history_begin->cycles,
1184 system_counterval.cycles, cycles) ||
1185 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1186 return -EINVAL;
1187 partial_history_cycles = cycles - system_counterval.cycles;
1188 total_history_cycles = cycles - history_begin->cycles;
1189 discontinuity =
1190 history_begin->clock_was_set_seq != clock_was_set_seq;
1191
1192 ret = adjust_historical_crosststamp(history_begin,
1193 partial_history_cycles,
1194 total_history_cycles,
1195 discontinuity, xtstamp);
1196 if (ret)
1197 return ret;
1198 }
1199
1200 return 0;
1201}
1202EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1203
1204/**
1205 * do_gettimeofday - Returns the time of day in a timeval
1206 * @tv: pointer to the timeval to be set
1207 *
1208 * NOTE: Users should be converted to using getnstimeofday()
1209 */
1210void do_gettimeofday(struct timeval *tv)
1211{
1212 struct timespec64 now;
1213
1214 getnstimeofday64(&now);
1215 tv->tv_sec = now.tv_sec;
1216 tv->tv_usec = now.tv_nsec/1000;
1217}
1218EXPORT_SYMBOL(do_gettimeofday);
1219
1220/**
1221 * do_settimeofday64 - Sets the time of day.
1222 * @ts: pointer to the timespec64 variable containing the new time
1223 *
1224 * Sets the time of day to the new time and update NTP and notify hrtimers
1225 */
1226int do_settimeofday64(const struct timespec64 *ts)
1227{
1228 struct timekeeper *tk = &tk_core.timekeeper;
1229 struct timespec64 ts_delta, xt;
1230 unsigned long flags;
1231 int ret = 0;
1232
1233 if (!timespec64_valid_strict(ts))
1234 return -EINVAL;
1235
1236 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1237 write_seqcount_begin(&tk_core.seq);
1238
1239 timekeeping_forward_now(tk);
1240
1241 xt = tk_xtime(tk);
1242 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1243 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1244
1245 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1246 ret = -EINVAL;
1247 goto out;
1248 }
1249
1250 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1251
1252 tk_set_xtime(tk, ts);
1253out:
1254 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1255
1256 write_seqcount_end(&tk_core.seq);
1257 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1258
1259 /* signal hrtimers about time change */
1260 clock_was_set();
1261
1262 return ret;
1263}
1264EXPORT_SYMBOL(do_settimeofday64);
1265
1266/**
1267 * timekeeping_inject_offset - Adds or subtracts from the current time.
1268 * @tv: pointer to the timespec variable containing the offset
1269 *
1270 * Adds or subtracts an offset value from the current time.
1271 */
1272static int timekeeping_inject_offset(struct timespec64 *ts)
1273{
1274 struct timekeeper *tk = &tk_core.timekeeper;
1275 unsigned long flags;
1276 struct timespec64 tmp;
1277 int ret = 0;
1278
1279 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1280 return -EINVAL;
1281
1282 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1283 write_seqcount_begin(&tk_core.seq);
1284
1285 timekeeping_forward_now(tk);
1286
1287 /* Make sure the proposed value is valid */
1288 tmp = timespec64_add(tk_xtime(tk), *ts);
1289 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1290 !timespec64_valid_strict(&tmp)) {
1291 ret = -EINVAL;
1292 goto error;
1293 }
1294
1295 tk_xtime_add(tk, ts);
1296 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1297
1298error: /* even if we error out, we forwarded the time, so call update */
1299 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1300
1301 write_seqcount_end(&tk_core.seq);
1302 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1303
1304 /* signal hrtimers about time change */
1305 clock_was_set();
1306
1307 return ret;
1308}
1309
1310/*
1311 * Indicates if there is an offset between the system clock and the hardware
1312 * clock/persistent clock/rtc.
1313 */
1314int persistent_clock_is_local;
1315
1316/*
1317 * Adjust the time obtained from the CMOS to be UTC time instead of
1318 * local time.
1319 *
1320 * This is ugly, but preferable to the alternatives. Otherwise we
1321 * would either need to write a program to do it in /etc/rc (and risk
1322 * confusion if the program gets run more than once; it would also be
1323 * hard to make the program warp the clock precisely n hours) or
1324 * compile in the timezone information into the kernel. Bad, bad....
1325 *
1326 * - TYT, 1992-01-01
1327 *
1328 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1329 * as real UNIX machines always do it. This avoids all headaches about
1330 * daylight saving times and warping kernel clocks.
1331 */
1332void timekeeping_warp_clock(void)
1333{
1334 if (sys_tz.tz_minuteswest != 0) {
1335 struct timespec64 adjust;
1336
1337 persistent_clock_is_local = 1;
1338 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1339 adjust.tv_nsec = 0;
1340 timekeeping_inject_offset(&adjust);
1341 }
1342}
1343
1344/**
1345 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1346 *
1347 */
1348static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1349{
1350 tk->tai_offset = tai_offset;
1351 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1352}
1353
1354/**
1355 * change_clocksource - Swaps clocksources if a new one is available
1356 *
1357 * Accumulates current time interval and initializes new clocksource
1358 */
1359static int change_clocksource(void *data)
1360{
1361 struct timekeeper *tk = &tk_core.timekeeper;
1362 struct clocksource *new, *old;
1363 unsigned long flags;
1364
1365 new = (struct clocksource *) data;
1366
1367 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1368 write_seqcount_begin(&tk_core.seq);
1369
1370 timekeeping_forward_now(tk);
1371 /*
1372 * If the cs is in module, get a module reference. Succeeds
1373 * for built-in code (owner == NULL) as well.
1374 */
1375 if (try_module_get(new->owner)) {
1376 if (!new->enable || new->enable(new) == 0) {
1377 old = tk->tkr_mono.clock;
1378 tk_setup_internals(tk, new);
1379 if (old->disable)
1380 old->disable(old);
1381 module_put(old->owner);
1382 } else {
1383 module_put(new->owner);
1384 }
1385 }
1386 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1387
1388 write_seqcount_end(&tk_core.seq);
1389 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1390
1391 return 0;
1392}
1393
1394/**
1395 * timekeeping_notify - Install a new clock source
1396 * @clock: pointer to the clock source
1397 *
1398 * This function is called from clocksource.c after a new, better clock
1399 * source has been registered. The caller holds the clocksource_mutex.
1400 */
1401int timekeeping_notify(struct clocksource *clock)
1402{
1403 struct timekeeper *tk = &tk_core.timekeeper;
1404
1405 if (tk->tkr_mono.clock == clock)
1406 return 0;
1407 stop_machine(change_clocksource, clock, NULL);
1408 tick_clock_notify();
1409 return tk->tkr_mono.clock == clock ? 0 : -1;
1410}
1411
1412/**
1413 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1414 * @ts: pointer to the timespec64 to be set
1415 *
1416 * Returns the raw monotonic time (completely un-modified by ntp)
1417 */
1418void getrawmonotonic64(struct timespec64 *ts)
1419{
1420 struct timekeeper *tk = &tk_core.timekeeper;
1421 unsigned long seq;
1422 u64 nsecs;
1423
1424 do {
1425 seq = read_seqcount_begin(&tk_core.seq);
1426 ts->tv_sec = tk->raw_sec;
1427 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1428
1429 } while (read_seqcount_retry(&tk_core.seq, seq));
1430
1431 ts->tv_nsec = 0;
1432 timespec64_add_ns(ts, nsecs);
1433}
1434EXPORT_SYMBOL(getrawmonotonic64);
1435
1436
1437/**
1438 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1439 */
1440int timekeeping_valid_for_hres(void)
1441{
1442 struct timekeeper *tk = &tk_core.timekeeper;
1443 unsigned long seq;
1444 int ret;
1445
1446 do {
1447 seq = read_seqcount_begin(&tk_core.seq);
1448
1449 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1450
1451 } while (read_seqcount_retry(&tk_core.seq, seq));
1452
1453 return ret;
1454}
1455
1456/**
1457 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1458 */
1459u64 timekeeping_max_deferment(void)
1460{
1461 struct timekeeper *tk = &tk_core.timekeeper;
1462 unsigned long seq;
1463 u64 ret;
1464
1465 do {
1466 seq = read_seqcount_begin(&tk_core.seq);
1467
1468 ret = tk->tkr_mono.clock->max_idle_ns;
1469
1470 } while (read_seqcount_retry(&tk_core.seq, seq));
1471
1472 return ret;
1473}
1474
1475/**
1476 * read_persistent_clock - Return time from the persistent clock.
1477 *
1478 * Weak dummy function for arches that do not yet support it.
1479 * Reads the time from the battery backed persistent clock.
1480 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1481 *
1482 * XXX - Do be sure to remove it once all arches implement it.
1483 */
1484void __weak read_persistent_clock(struct timespec *ts)
1485{
1486 ts->tv_sec = 0;
1487 ts->tv_nsec = 0;
1488}
1489
1490void __weak read_persistent_clock64(struct timespec64 *ts64)
1491{
1492 struct timespec ts;
1493
1494 read_persistent_clock(&ts);
1495 *ts64 = timespec_to_timespec64(ts);
1496}
1497
1498/**
1499 * read_boot_clock64 - Return time of the system start.
1500 *
1501 * Weak dummy function for arches that do not yet support it.
1502 * Function to read the exact time the system has been started.
1503 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1504 *
1505 * XXX - Do be sure to remove it once all arches implement it.
1506 */
1507void __weak read_boot_clock64(struct timespec64 *ts)
1508{
1509 ts->tv_sec = 0;
1510 ts->tv_nsec = 0;
1511}
1512
1513/* Flag for if timekeeping_resume() has injected sleeptime */
1514static bool sleeptime_injected;
1515
1516/* Flag for if there is a persistent clock on this platform */
1517static bool persistent_clock_exists;
1518
1519/*
1520 * timekeeping_init - Initializes the clocksource and common timekeeping values
1521 */
1522void __init timekeeping_init(void)
1523{
1524 struct timekeeper *tk = &tk_core.timekeeper;
1525 struct clocksource *clock;
1526 unsigned long flags;
1527 struct timespec64 now, boot, tmp;
1528
1529 read_persistent_clock64(&now);
1530 if (!timespec64_valid_strict(&now)) {
1531 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1532 " Check your CMOS/BIOS settings.\n");
1533 now.tv_sec = 0;
1534 now.tv_nsec = 0;
1535 } else if (now.tv_sec || now.tv_nsec)
1536 persistent_clock_exists = true;
1537
1538 read_boot_clock64(&boot);
1539 if (!timespec64_valid_strict(&boot)) {
1540 pr_warn("WARNING: Boot clock returned invalid value!\n"
1541 " Check your CMOS/BIOS settings.\n");
1542 boot.tv_sec = 0;
1543 boot.tv_nsec = 0;
1544 }
1545
1546 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1547 write_seqcount_begin(&tk_core.seq);
1548 ntp_init();
1549
1550 clock = clocksource_default_clock();
1551 if (clock->enable)
1552 clock->enable(clock);
1553 tk_setup_internals(tk, clock);
1554
1555 tk_set_xtime(tk, &now);
1556 tk->raw_sec = 0;
1557 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1558 boot = tk_xtime(tk);
1559
1560 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1561 tk_set_wall_to_mono(tk, tmp);
1562
1563 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1564
1565 write_seqcount_end(&tk_core.seq);
1566 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1567}
1568
1569/* time in seconds when suspend began for persistent clock */
1570static struct timespec64 timekeeping_suspend_time;
1571
1572/**
1573 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1574 * @delta: pointer to a timespec delta value
1575 *
1576 * Takes a timespec offset measuring a suspend interval and properly
1577 * adds the sleep offset to the timekeeping variables.
1578 */
1579static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1580 struct timespec64 *delta)
1581{
1582 if (!timespec64_valid_strict(delta)) {
1583 printk_deferred(KERN_WARNING
1584 "__timekeeping_inject_sleeptime: Invalid "
1585 "sleep delta value!\n");
1586 return;
1587 }
1588 tk_xtime_add(tk, delta);
1589 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1590 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1591 tk_debug_account_sleep_time(delta);
1592}
1593
1594#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1595/**
1596 * We have three kinds of time sources to use for sleep time
1597 * injection, the preference order is:
1598 * 1) non-stop clocksource
1599 * 2) persistent clock (ie: RTC accessible when irqs are off)
1600 * 3) RTC
1601 *
1602 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1603 * If system has neither 1) nor 2), 3) will be used finally.
1604 *
1605 *
1606 * If timekeeping has injected sleeptime via either 1) or 2),
1607 * 3) becomes needless, so in this case we don't need to call
1608 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1609 * means.
1610 */
1611bool timekeeping_rtc_skipresume(void)
1612{
1613 return sleeptime_injected;
1614}
1615
1616/**
1617 * 1) can be determined whether to use or not only when doing
1618 * timekeeping_resume() which is invoked after rtc_suspend(),
1619 * so we can't skip rtc_suspend() surely if system has 1).
1620 *
1621 * But if system has 2), 2) will definitely be used, so in this
1622 * case we don't need to call rtc_suspend(), and this is what
1623 * timekeeping_rtc_skipsuspend() means.
1624 */
1625bool timekeeping_rtc_skipsuspend(void)
1626{
1627 return persistent_clock_exists;
1628}
1629
1630/**
1631 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1632 * @delta: pointer to a timespec64 delta value
1633 *
1634 * This hook is for architectures that cannot support read_persistent_clock64
1635 * because their RTC/persistent clock is only accessible when irqs are enabled.
1636 * and also don't have an effective nonstop clocksource.
1637 *
1638 * This function should only be called by rtc_resume(), and allows
1639 * a suspend offset to be injected into the timekeeping values.
1640 */
1641void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1642{
1643 struct timekeeper *tk = &tk_core.timekeeper;
1644 unsigned long flags;
1645
1646 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1647 write_seqcount_begin(&tk_core.seq);
1648
1649 timekeeping_forward_now(tk);
1650
1651 __timekeeping_inject_sleeptime(tk, delta);
1652
1653 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1654
1655 write_seqcount_end(&tk_core.seq);
1656 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1657
1658 /* signal hrtimers about time change */
1659 clock_was_set();
1660}
1661#endif
1662
1663/**
1664 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1665 */
1666void timekeeping_resume(void)
1667{
1668 struct timekeeper *tk = &tk_core.timekeeper;
1669 struct clocksource *clock = tk->tkr_mono.clock;
1670 unsigned long flags;
1671 struct timespec64 ts_new, ts_delta;
1672 u64 cycle_now;
1673
1674 sleeptime_injected = false;
1675 read_persistent_clock64(&ts_new);
1676
1677 clockevents_resume();
1678 clocksource_resume();
1679
1680 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1681 write_seqcount_begin(&tk_core.seq);
1682
1683 /*
1684 * After system resumes, we need to calculate the suspended time and
1685 * compensate it for the OS time. There are 3 sources that could be
1686 * used: Nonstop clocksource during suspend, persistent clock and rtc
1687 * device.
1688 *
1689 * One specific platform may have 1 or 2 or all of them, and the
1690 * preference will be:
1691 * suspend-nonstop clocksource -> persistent clock -> rtc
1692 * The less preferred source will only be tried if there is no better
1693 * usable source. The rtc part is handled separately in rtc core code.
1694 */
1695 cycle_now = tk_clock_read(&tk->tkr_mono);
1696 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1697 cycle_now > tk->tkr_mono.cycle_last) {
1698 u64 nsec, cyc_delta;
1699
1700 cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1701 tk->tkr_mono.mask);
1702 nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift);
1703 ts_delta = ns_to_timespec64(nsec);
1704 sleeptime_injected = true;
1705 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1706 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1707 sleeptime_injected = true;
1708 }
1709
1710 if (sleeptime_injected)
1711 __timekeeping_inject_sleeptime(tk, &ts_delta);
1712
1713 /* Re-base the last cycle value */
1714 tk->tkr_mono.cycle_last = cycle_now;
1715 tk->tkr_raw.cycle_last = cycle_now;
1716
1717 tk->ntp_error = 0;
1718 timekeeping_suspended = 0;
1719 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1720 write_seqcount_end(&tk_core.seq);
1721 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1722
1723 touch_softlockup_watchdog();
1724
1725 tick_resume();
1726 hrtimers_resume();
1727}
1728
1729int timekeeping_suspend(void)
1730{
1731 struct timekeeper *tk = &tk_core.timekeeper;
1732 unsigned long flags;
1733 struct timespec64 delta, delta_delta;
1734 static struct timespec64 old_delta;
1735
1736 read_persistent_clock64(&timekeeping_suspend_time);
1737
1738 /*
1739 * On some systems the persistent_clock can not be detected at
1740 * timekeeping_init by its return value, so if we see a valid
1741 * value returned, update the persistent_clock_exists flag.
1742 */
1743 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1744 persistent_clock_exists = true;
1745
1746 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1747 write_seqcount_begin(&tk_core.seq);
1748 timekeeping_forward_now(tk);
1749 timekeeping_suspended = 1;
1750
1751 if (persistent_clock_exists) {
1752 /*
1753 * To avoid drift caused by repeated suspend/resumes,
1754 * which each can add ~1 second drift error,
1755 * try to compensate so the difference in system time
1756 * and persistent_clock time stays close to constant.
1757 */
1758 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1759 delta_delta = timespec64_sub(delta, old_delta);
1760 if (abs(delta_delta.tv_sec) >= 2) {
1761 /*
1762 * if delta_delta is too large, assume time correction
1763 * has occurred and set old_delta to the current delta.
1764 */
1765 old_delta = delta;
1766 } else {
1767 /* Otherwise try to adjust old_system to compensate */
1768 timekeeping_suspend_time =
1769 timespec64_add(timekeeping_suspend_time, delta_delta);
1770 }
1771 }
1772
1773 timekeeping_update(tk, TK_MIRROR);
1774 halt_fast_timekeeper(tk);
1775 write_seqcount_end(&tk_core.seq);
1776 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1777
1778 tick_suspend();
1779 clocksource_suspend();
1780 clockevents_suspend();
1781
1782 return 0;
1783}
1784
1785/* sysfs resume/suspend bits for timekeeping */
1786static struct syscore_ops timekeeping_syscore_ops = {
1787 .resume = timekeeping_resume,
1788 .suspend = timekeeping_suspend,
1789};
1790
1791static int __init timekeeping_init_ops(void)
1792{
1793 register_syscore_ops(&timekeeping_syscore_ops);
1794 return 0;
1795}
1796device_initcall(timekeeping_init_ops);
1797
1798/*
1799 * Apply a multiplier adjustment to the timekeeper
1800 */
1801static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1802 s64 offset,
1803 s32 mult_adj)
1804{
1805 s64 interval = tk->cycle_interval;
1806
1807 if (mult_adj == 0) {
1808 return;
1809 } else if (mult_adj == -1) {
1810 interval = -interval;
1811 offset = -offset;
1812 } else if (mult_adj != 1) {
1813 interval *= mult_adj;
1814 offset *= mult_adj;
1815 }
1816
1817 /*
1818 * So the following can be confusing.
1819 *
1820 * To keep things simple, lets assume mult_adj == 1 for now.
1821 *
1822 * When mult_adj != 1, remember that the interval and offset values
1823 * have been appropriately scaled so the math is the same.
1824 *
1825 * The basic idea here is that we're increasing the multiplier
1826 * by one, this causes the xtime_interval to be incremented by
1827 * one cycle_interval. This is because:
1828 * xtime_interval = cycle_interval * mult
1829 * So if mult is being incremented by one:
1830 * xtime_interval = cycle_interval * (mult + 1)
1831 * Its the same as:
1832 * xtime_interval = (cycle_interval * mult) + cycle_interval
1833 * Which can be shortened to:
1834 * xtime_interval += cycle_interval
1835 *
1836 * So offset stores the non-accumulated cycles. Thus the current
1837 * time (in shifted nanoseconds) is:
1838 * now = (offset * adj) + xtime_nsec
1839 * Now, even though we're adjusting the clock frequency, we have
1840 * to keep time consistent. In other words, we can't jump back
1841 * in time, and we also want to avoid jumping forward in time.
1842 *
1843 * So given the same offset value, we need the time to be the same
1844 * both before and after the freq adjustment.
1845 * now = (offset * adj_1) + xtime_nsec_1
1846 * now = (offset * adj_2) + xtime_nsec_2
1847 * So:
1848 * (offset * adj_1) + xtime_nsec_1 =
1849 * (offset * adj_2) + xtime_nsec_2
1850 * And we know:
1851 * adj_2 = adj_1 + 1
1852 * So:
1853 * (offset * adj_1) + xtime_nsec_1 =
1854 * (offset * (adj_1+1)) + xtime_nsec_2
1855 * (offset * adj_1) + xtime_nsec_1 =
1856 * (offset * adj_1) + offset + xtime_nsec_2
1857 * Canceling the sides:
1858 * xtime_nsec_1 = offset + xtime_nsec_2
1859 * Which gives us:
1860 * xtime_nsec_2 = xtime_nsec_1 - offset
1861 * Which simplfies to:
1862 * xtime_nsec -= offset
1863 */
1864 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1865 /* NTP adjustment caused clocksource mult overflow */
1866 WARN_ON_ONCE(1);
1867 return;
1868 }
1869
1870 tk->tkr_mono.mult += mult_adj;
1871 tk->xtime_interval += interval;
1872 tk->tkr_mono.xtime_nsec -= offset;
1873}
1874
1875/*
1876 * Adjust the timekeeper's multiplier to the correct frequency
1877 * and also to reduce the accumulated error value.
1878 */
1879static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1880{
1881 u32 mult;
1882
1883 /*
1884 * Determine the multiplier from the current NTP tick length.
1885 * Avoid expensive division when the tick length doesn't change.
1886 */
1887 if (likely(tk->ntp_tick == ntp_tick_length())) {
1888 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1889 } else {
1890 tk->ntp_tick = ntp_tick_length();
1891 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1892 tk->xtime_remainder, tk->cycle_interval);
1893 }
1894
1895 /*
1896 * If the clock is behind the NTP time, increase the multiplier by 1
1897 * to catch up with it. If it's ahead and there was a remainder in the
1898 * tick division, the clock will slow down. Otherwise it will stay
1899 * ahead until the tick length changes to a non-divisible value.
1900 */
1901 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1902 mult += tk->ntp_err_mult;
1903
1904 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
1905
1906 if (unlikely(tk->tkr_mono.clock->maxadj &&
1907 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1908 > tk->tkr_mono.clock->maxadj))) {
1909 printk_once(KERN_WARNING
1910 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1911 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1912 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1913 }
1914
1915 /*
1916 * It may be possible that when we entered this function, xtime_nsec
1917 * was very small. Further, if we're slightly speeding the clocksource
1918 * in the code above, its possible the required corrective factor to
1919 * xtime_nsec could cause it to underflow.
1920 *
1921 * Now, since we have already accumulated the second and the NTP
1922 * subsystem has been notified via second_overflow(), we need to skip
1923 * the next update.
1924 */
1925 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1926 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
1927 tk->tkr_mono.shift;
1928 tk->xtime_sec--;
1929 tk->skip_second_overflow = 1;
1930 }
1931}
1932
1933/**
1934 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1935 *
1936 * Helper function that accumulates the nsecs greater than a second
1937 * from the xtime_nsec field to the xtime_secs field.
1938 * It also calls into the NTP code to handle leapsecond processing.
1939 *
1940 */
1941static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1942{
1943 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1944 unsigned int clock_set = 0;
1945
1946 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1947 int leap;
1948
1949 tk->tkr_mono.xtime_nsec -= nsecps;
1950 tk->xtime_sec++;
1951
1952 /*
1953 * Skip NTP update if this second was accumulated before,
1954 * i.e. xtime_nsec underflowed in timekeeping_adjust()
1955 */
1956 if (unlikely(tk->skip_second_overflow)) {
1957 tk->skip_second_overflow = 0;
1958 continue;
1959 }
1960
1961 /* Figure out if its a leap sec and apply if needed */
1962 leap = second_overflow(tk->xtime_sec);
1963 if (unlikely(leap)) {
1964 struct timespec64 ts;
1965
1966 tk->xtime_sec += leap;
1967
1968 ts.tv_sec = leap;
1969 ts.tv_nsec = 0;
1970 tk_set_wall_to_mono(tk,
1971 timespec64_sub(tk->wall_to_monotonic, ts));
1972
1973 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1974
1975 clock_set = TK_CLOCK_WAS_SET;
1976 }
1977 }
1978 return clock_set;
1979}
1980
1981/**
1982 * logarithmic_accumulation - shifted accumulation of cycles
1983 *
1984 * This functions accumulates a shifted interval of cycles into
1985 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1986 * loop.
1987 *
1988 * Returns the unconsumed cycles.
1989 */
1990static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
1991 u32 shift, unsigned int *clock_set)
1992{
1993 u64 interval = tk->cycle_interval << shift;
1994 u64 snsec_per_sec;
1995
1996 /* If the offset is smaller than a shifted interval, do nothing */
1997 if (offset < interval)
1998 return offset;
1999
2000 /* Accumulate one shifted interval */
2001 offset -= interval;
2002 tk->tkr_mono.cycle_last += interval;
2003 tk->tkr_raw.cycle_last += interval;
2004
2005 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2006 *clock_set |= accumulate_nsecs_to_secs(tk);
2007
2008 /* Accumulate raw time */
2009 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2010 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2011 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2012 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2013 tk->raw_sec++;
2014 }
2015
2016 /* Accumulate error between NTP and clock interval */
2017 tk->ntp_error += tk->ntp_tick << shift;
2018 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2019 (tk->ntp_error_shift + shift);
2020
2021 return offset;
2022}
2023
2024/**
2025 * update_wall_time - Uses the current clocksource to increment the wall time
2026 *
2027 */
2028void update_wall_time(void)
2029{
2030 struct timekeeper *real_tk = &tk_core.timekeeper;
2031 struct timekeeper *tk = &shadow_timekeeper;
2032 u64 offset;
2033 int shift = 0, maxshift;
2034 unsigned int clock_set = 0;
2035 unsigned long flags;
2036
2037 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2038
2039 /* Make sure we're fully resumed: */
2040 if (unlikely(timekeeping_suspended))
2041 goto out;
2042
2043#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2044 offset = real_tk->cycle_interval;
2045#else
2046 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2047 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2048#endif
2049
2050 /* Check if there's really nothing to do */
2051 if (offset < real_tk->cycle_interval)
2052 goto out;
2053
2054 /* Do some additional sanity checking */
2055 timekeeping_check_update(tk, offset);
2056
2057 /*
2058 * With NO_HZ we may have to accumulate many cycle_intervals
2059 * (think "ticks") worth of time at once. To do this efficiently,
2060 * we calculate the largest doubling multiple of cycle_intervals
2061 * that is smaller than the offset. We then accumulate that
2062 * chunk in one go, and then try to consume the next smaller
2063 * doubled multiple.
2064 */
2065 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2066 shift = max(0, shift);
2067 /* Bound shift to one less than what overflows tick_length */
2068 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2069 shift = min(shift, maxshift);
2070 while (offset >= tk->cycle_interval) {
2071 offset = logarithmic_accumulation(tk, offset, shift,
2072 &clock_set);
2073 if (offset < tk->cycle_interval<<shift)
2074 shift--;
2075 }
2076
2077 /* Adjust the multiplier to correct NTP error */
2078 timekeeping_adjust(tk, offset);
2079
2080 /*
2081 * Finally, make sure that after the rounding
2082 * xtime_nsec isn't larger than NSEC_PER_SEC
2083 */
2084 clock_set |= accumulate_nsecs_to_secs(tk);
2085
2086 write_seqcount_begin(&tk_core.seq);
2087 /*
2088 * Update the real timekeeper.
2089 *
2090 * We could avoid this memcpy by switching pointers, but that
2091 * requires changes to all other timekeeper usage sites as
2092 * well, i.e. move the timekeeper pointer getter into the
2093 * spinlocked/seqcount protected sections. And we trade this
2094 * memcpy under the tk_core.seq against one before we start
2095 * updating.
2096 */
2097 timekeeping_update(tk, clock_set);
2098 memcpy(real_tk, tk, sizeof(*tk));
2099 /* The memcpy must come last. Do not put anything here! */
2100 write_seqcount_end(&tk_core.seq);
2101out:
2102 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2103 if (clock_set)
2104 /* Have to call _delayed version, since in irq context*/
2105 clock_was_set_delayed();
2106}
2107
2108/**
2109 * getboottime64 - Return the real time of system boot.
2110 * @ts: pointer to the timespec64 to be set
2111 *
2112 * Returns the wall-time of boot in a timespec64.
2113 *
2114 * This is based on the wall_to_monotonic offset and the total suspend
2115 * time. Calls to settimeofday will affect the value returned (which
2116 * basically means that however wrong your real time clock is at boot time,
2117 * you get the right time here).
2118 */
2119void getboottime64(struct timespec64 *ts)
2120{
2121 struct timekeeper *tk = &tk_core.timekeeper;
2122 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2123
2124 *ts = ktime_to_timespec64(t);
2125}
2126EXPORT_SYMBOL_GPL(getboottime64);
2127
2128unsigned long get_seconds(void)
2129{
2130 struct timekeeper *tk = &tk_core.timekeeper;
2131
2132 return tk->xtime_sec;
2133}
2134EXPORT_SYMBOL(get_seconds);
2135
2136struct timespec64 current_kernel_time64(void)
2137{
2138 struct timekeeper *tk = &tk_core.timekeeper;
2139 struct timespec64 now;
2140 unsigned long seq;
2141
2142 do {
2143 seq = read_seqcount_begin(&tk_core.seq);
2144
2145 now = tk_xtime(tk);
2146 } while (read_seqcount_retry(&tk_core.seq, seq));
2147
2148 return now;
2149}
2150EXPORT_SYMBOL(current_kernel_time64);
2151
2152struct timespec64 get_monotonic_coarse64(void)
2153{
2154 struct timekeeper *tk = &tk_core.timekeeper;
2155 struct timespec64 now, mono;
2156 unsigned long seq;
2157
2158 do {
2159 seq = read_seqcount_begin(&tk_core.seq);
2160
2161 now = tk_xtime(tk);
2162 mono = tk->wall_to_monotonic;
2163 } while (read_seqcount_retry(&tk_core.seq, seq));
2164
2165 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2166 now.tv_nsec + mono.tv_nsec);
2167
2168 return now;
2169}
2170EXPORT_SYMBOL(get_monotonic_coarse64);
2171
2172/*
2173 * Must hold jiffies_lock
2174 */
2175void do_timer(unsigned long ticks)
2176{
2177 jiffies_64 += ticks;
2178 calc_global_load(ticks);
2179}
2180
2181/**
2182 * ktime_get_update_offsets_now - hrtimer helper
2183 * @cwsseq: pointer to check and store the clock was set sequence number
2184 * @offs_real: pointer to storage for monotonic -> realtime offset
2185 * @offs_boot: pointer to storage for monotonic -> boottime offset
2186 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2187 *
2188 * Returns current monotonic time and updates the offsets if the
2189 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2190 * different.
2191 *
2192 * Called from hrtimer_interrupt() or retrigger_next_event()
2193 */
2194ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2195 ktime_t *offs_boot, ktime_t *offs_tai)
2196{
2197 struct timekeeper *tk = &tk_core.timekeeper;
2198 unsigned int seq;
2199 ktime_t base;
2200 u64 nsecs;
2201
2202 do {
2203 seq = read_seqcount_begin(&tk_core.seq);
2204
2205 base = tk->tkr_mono.base;
2206 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2207 base = ktime_add_ns(base, nsecs);
2208
2209 if (*cwsseq != tk->clock_was_set_seq) {
2210 *cwsseq = tk->clock_was_set_seq;
2211 *offs_real = tk->offs_real;
2212 *offs_boot = tk->offs_boot;
2213 *offs_tai = tk->offs_tai;
2214 }
2215
2216 /* Handle leapsecond insertion adjustments */
2217 if (unlikely(base >= tk->next_leap_ktime))
2218 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2219
2220 } while (read_seqcount_retry(&tk_core.seq, seq));
2221
2222 return base;
2223}
2224
2225/**
2226 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2227 */
2228static int timekeeping_validate_timex(struct timex *txc)
2229{
2230 if (txc->modes & ADJ_ADJTIME) {
2231 /* singleshot must not be used with any other mode bits */
2232 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2233 return -EINVAL;
2234 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2235 !capable(CAP_SYS_TIME))
2236 return -EPERM;
2237 } else {
2238 /* In order to modify anything, you gotta be super-user! */
2239 if (txc->modes && !capable(CAP_SYS_TIME))
2240 return -EPERM;
2241 /*
2242 * if the quartz is off by more than 10% then
2243 * something is VERY wrong!
2244 */
2245 if (txc->modes & ADJ_TICK &&
2246 (txc->tick < 900000/USER_HZ ||
2247 txc->tick > 1100000/USER_HZ))
2248 return -EINVAL;
2249 }
2250
2251 if (txc->modes & ADJ_SETOFFSET) {
2252 /* In order to inject time, you gotta be super-user! */
2253 if (!capable(CAP_SYS_TIME))
2254 return -EPERM;
2255
2256 /*
2257 * Validate if a timespec/timeval used to inject a time
2258 * offset is valid. Offsets can be postive or negative, so
2259 * we don't check tv_sec. The value of the timeval/timespec
2260 * is the sum of its fields,but *NOTE*:
2261 * The field tv_usec/tv_nsec must always be non-negative and
2262 * we can't have more nanoseconds/microseconds than a second.
2263 */
2264 if (txc->time.tv_usec < 0)
2265 return -EINVAL;
2266
2267 if (txc->modes & ADJ_NANO) {
2268 if (txc->time.tv_usec >= NSEC_PER_SEC)
2269 return -EINVAL;
2270 } else {
2271 if (txc->time.tv_usec >= USEC_PER_SEC)
2272 return -EINVAL;
2273 }
2274 }
2275
2276 /*
2277 * Check for potential multiplication overflows that can
2278 * only happen on 64-bit systems:
2279 */
2280 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2281 if (LLONG_MIN / PPM_SCALE > txc->freq)
2282 return -EINVAL;
2283 if (LLONG_MAX / PPM_SCALE < txc->freq)
2284 return -EINVAL;
2285 }
2286
2287 return 0;
2288}
2289
2290
2291/**
2292 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2293 */
2294int do_adjtimex(struct timex *txc)
2295{
2296 struct timekeeper *tk = &tk_core.timekeeper;
2297 unsigned long flags;
2298 struct timespec64 ts;
2299 s32 orig_tai, tai;
2300 int ret;
2301
2302 /* Validate the data before disabling interrupts */
2303 ret = timekeeping_validate_timex(txc);
2304 if (ret)
2305 return ret;
2306
2307 if (txc->modes & ADJ_SETOFFSET) {
2308 struct timespec64 delta;
2309 delta.tv_sec = txc->time.tv_sec;
2310 delta.tv_nsec = txc->time.tv_usec;
2311 if (!(txc->modes & ADJ_NANO))
2312 delta.tv_nsec *= 1000;
2313 ret = timekeeping_inject_offset(&delta);
2314 if (ret)
2315 return ret;
2316 }
2317
2318 getnstimeofday64(&ts);
2319
2320 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2321 write_seqcount_begin(&tk_core.seq);
2322
2323 orig_tai = tai = tk->tai_offset;
2324 ret = __do_adjtimex(txc, &ts, &tai);
2325
2326 if (tai != orig_tai) {
2327 __timekeeping_set_tai_offset(tk, tai);
2328 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2329 }
2330 tk_update_leap_state(tk);
2331
2332 write_seqcount_end(&tk_core.seq);
2333 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2334
2335 if (tai != orig_tai)
2336 clock_was_set();
2337
2338 ntp_notify_cmos_timer();
2339
2340 return ret;
2341}
2342
2343#ifdef CONFIG_NTP_PPS
2344/**
2345 * hardpps() - Accessor function to NTP __hardpps function
2346 */
2347void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2348{
2349 unsigned long flags;
2350
2351 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2352 write_seqcount_begin(&tk_core.seq);
2353
2354 __hardpps(phase_ts, raw_ts);
2355
2356 write_seqcount_end(&tk_core.seq);
2357 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2358}
2359EXPORT_SYMBOL(hardpps);
2360#endif /* CONFIG_NTP_PPS */
2361
2362/**
2363 * xtime_update() - advances the timekeeping infrastructure
2364 * @ticks: number of ticks, that have elapsed since the last call.
2365 *
2366 * Must be called with interrupts disabled.
2367 */
2368void xtime_update(unsigned long ticks)
2369{
2370 write_seqlock(&jiffies_lock);
2371 do_timer(ticks);
2372 write_sequnlock(&jiffies_lock);
2373 update_wall_time();
2374}