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/timex.h>
21#include <linux/tick.h>
22#include <linux/stop_machine.h>
23#include <linux/pvclock_gtod.h>
24#include <linux/compiler.h>
25#include <linux/audit.h>
26#include <linux/random.h>
27
28#include "tick-internal.h"
29#include "ntp_internal.h"
30#include "timekeeping_internal.h"
31
32#define TK_CLEAR_NTP (1 << 0)
33#define TK_MIRROR (1 << 1)
34#define TK_CLOCK_WAS_SET (1 << 2)
35
36enum timekeeping_adv_mode {
37 /* Update timekeeper when a tick has passed */
38 TK_ADV_TICK,
39
40 /* Update timekeeper on a direct frequency change */
41 TK_ADV_FREQ
42};
43
44DEFINE_RAW_SPINLOCK(timekeeper_lock);
45
46/*
47 * The most important data for readout fits into a single 64 byte
48 * cache line.
49 */
50static struct {
51 seqcount_raw_spinlock_t seq;
52 struct timekeeper timekeeper;
53} tk_core ____cacheline_aligned = {
54 .seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
55};
56
57static struct timekeeper shadow_timekeeper;
58
59/* flag for if timekeeping is suspended */
60int __read_mostly timekeeping_suspended;
61
62/**
63 * struct tk_fast - NMI safe timekeeper
64 * @seq: Sequence counter for protecting updates. The lowest bit
65 * is the index for the tk_read_base array
66 * @base: tk_read_base array. Access is indexed by the lowest bit of
67 * @seq.
68 *
69 * See @update_fast_timekeeper() below.
70 */
71struct tk_fast {
72 seqcount_latch_t seq;
73 struct tk_read_base base[2];
74};
75
76/* Suspend-time cycles value for halted fast timekeeper. */
77static u64 cycles_at_suspend;
78
79static u64 dummy_clock_read(struct clocksource *cs)
80{
81 if (timekeeping_suspended)
82 return cycles_at_suspend;
83 return local_clock();
84}
85
86static struct clocksource dummy_clock = {
87 .read = dummy_clock_read,
88};
89
90/*
91 * Boot time initialization which allows local_clock() to be utilized
92 * during early boot when clocksources are not available. local_clock()
93 * returns nanoseconds already so no conversion is required, hence mult=1
94 * and shift=0. When the first proper clocksource is installed then
95 * the fast time keepers are updated with the correct values.
96 */
97#define FAST_TK_INIT \
98 { \
99 .clock = &dummy_clock, \
100 .mask = CLOCKSOURCE_MASK(64), \
101 .mult = 1, \
102 .shift = 0, \
103 }
104
105static struct tk_fast tk_fast_mono ____cacheline_aligned = {
106 .seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
107 .base[0] = FAST_TK_INIT,
108 .base[1] = FAST_TK_INIT,
109};
110
111static struct tk_fast tk_fast_raw ____cacheline_aligned = {
112 .seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
113 .base[0] = FAST_TK_INIT,
114 .base[1] = FAST_TK_INIT,
115};
116
117static inline void tk_normalize_xtime(struct timekeeper *tk)
118{
119 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
120 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
121 tk->xtime_sec++;
122 }
123 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
124 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
125 tk->raw_sec++;
126 }
127}
128
129static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
130{
131 struct timespec64 ts;
132
133 ts.tv_sec = tk->xtime_sec;
134 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
135 return ts;
136}
137
138static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
139{
140 tk->xtime_sec = ts->tv_sec;
141 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
142}
143
144static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
145{
146 tk->xtime_sec += ts->tv_sec;
147 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
148 tk_normalize_xtime(tk);
149}
150
151static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
152{
153 struct timespec64 tmp;
154
155 /*
156 * Verify consistency of: offset_real = -wall_to_monotonic
157 * before modifying anything
158 */
159 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
160 -tk->wall_to_monotonic.tv_nsec);
161 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
162 tk->wall_to_monotonic = wtm;
163 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
164 tk->offs_real = timespec64_to_ktime(tmp);
165 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
166}
167
168static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
169{
170 tk->offs_boot = ktime_add(tk->offs_boot, delta);
171 /*
172 * Timespec representation for VDSO update to avoid 64bit division
173 * on every update.
174 */
175 tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
176}
177
178/*
179 * tk_clock_read - atomic clocksource read() helper
180 *
181 * This helper is necessary to use in the read paths because, while the
182 * seqcount ensures we don't return a bad value while structures are updated,
183 * it doesn't protect from potential crashes. There is the possibility that
184 * the tkr's clocksource may change between the read reference, and the
185 * clock reference passed to the read function. This can cause crashes if
186 * the wrong clocksource is passed to the wrong read function.
187 * This isn't necessary to use when holding the timekeeper_lock or doing
188 * a read of the fast-timekeeper tkrs (which is protected by its own locking
189 * and update logic).
190 */
191static inline u64 tk_clock_read(const struct tk_read_base *tkr)
192{
193 struct clocksource *clock = READ_ONCE(tkr->clock);
194
195 return clock->read(clock);
196}
197
198#ifdef CONFIG_DEBUG_TIMEKEEPING
199#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
200
201static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
202{
203
204 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
205 const char *name = tk->tkr_mono.clock->name;
206
207 if (offset > max_cycles) {
208 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
209 offset, name, max_cycles);
210 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
211 } else {
212 if (offset > (max_cycles >> 1)) {
213 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
214 offset, name, max_cycles >> 1);
215 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
216 }
217 }
218
219 if (tk->underflow_seen) {
220 if (jiffies - tk->last_warning > WARNING_FREQ) {
221 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
222 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
223 printk_deferred(" Your kernel is probably still fine.\n");
224 tk->last_warning = jiffies;
225 }
226 tk->underflow_seen = 0;
227 }
228
229 if (tk->overflow_seen) {
230 if (jiffies - tk->last_warning > WARNING_FREQ) {
231 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
232 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
233 printk_deferred(" Your kernel is probably still fine.\n");
234 tk->last_warning = jiffies;
235 }
236 tk->overflow_seen = 0;
237 }
238}
239
240static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
241{
242 struct timekeeper *tk = &tk_core.timekeeper;
243 u64 now, last, mask, max, delta;
244 unsigned int seq;
245
246 /*
247 * Since we're called holding a seqcount, the data may shift
248 * under us while we're doing the calculation. This can cause
249 * false positives, since we'd note a problem but throw the
250 * results away. So nest another seqcount here to atomically
251 * grab the points we are checking with.
252 */
253 do {
254 seq = read_seqcount_begin(&tk_core.seq);
255 now = tk_clock_read(tkr);
256 last = tkr->cycle_last;
257 mask = tkr->mask;
258 max = tkr->clock->max_cycles;
259 } while (read_seqcount_retry(&tk_core.seq, seq));
260
261 delta = clocksource_delta(now, last, mask);
262
263 /*
264 * Try to catch underflows by checking if we are seeing small
265 * mask-relative negative values.
266 */
267 if (unlikely((~delta & mask) < (mask >> 3))) {
268 tk->underflow_seen = 1;
269 delta = 0;
270 }
271
272 /* Cap delta value to the max_cycles values to avoid mult overflows */
273 if (unlikely(delta > max)) {
274 tk->overflow_seen = 1;
275 delta = tkr->clock->max_cycles;
276 }
277
278 return delta;
279}
280#else
281static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
282{
283}
284static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
285{
286 u64 cycle_now, delta;
287
288 /* read clocksource */
289 cycle_now = tk_clock_read(tkr);
290
291 /* calculate the delta since the last update_wall_time */
292 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
293
294 return delta;
295}
296#endif
297
298/**
299 * tk_setup_internals - Set up internals to use clocksource clock.
300 *
301 * @tk: The target timekeeper to setup.
302 * @clock: Pointer to clocksource.
303 *
304 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
305 * pair and interval request.
306 *
307 * Unless you're the timekeeping code, you should not be using this!
308 */
309static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
310{
311 u64 interval;
312 u64 tmp, ntpinterval;
313 struct clocksource *old_clock;
314
315 ++tk->cs_was_changed_seq;
316 old_clock = tk->tkr_mono.clock;
317 tk->tkr_mono.clock = clock;
318 tk->tkr_mono.mask = clock->mask;
319 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
320
321 tk->tkr_raw.clock = clock;
322 tk->tkr_raw.mask = clock->mask;
323 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
324
325 /* Do the ns -> cycle conversion first, using original mult */
326 tmp = NTP_INTERVAL_LENGTH;
327 tmp <<= clock->shift;
328 ntpinterval = tmp;
329 tmp += clock->mult/2;
330 do_div(tmp, clock->mult);
331 if (tmp == 0)
332 tmp = 1;
333
334 interval = (u64) tmp;
335 tk->cycle_interval = interval;
336
337 /* Go back from cycles -> shifted ns */
338 tk->xtime_interval = interval * clock->mult;
339 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
340 tk->raw_interval = interval * clock->mult;
341
342 /* if changing clocks, convert xtime_nsec shift units */
343 if (old_clock) {
344 int shift_change = clock->shift - old_clock->shift;
345 if (shift_change < 0) {
346 tk->tkr_mono.xtime_nsec >>= -shift_change;
347 tk->tkr_raw.xtime_nsec >>= -shift_change;
348 } else {
349 tk->tkr_mono.xtime_nsec <<= shift_change;
350 tk->tkr_raw.xtime_nsec <<= shift_change;
351 }
352 }
353
354 tk->tkr_mono.shift = clock->shift;
355 tk->tkr_raw.shift = clock->shift;
356
357 tk->ntp_error = 0;
358 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
359 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
360
361 /*
362 * The timekeeper keeps its own mult values for the currently
363 * active clocksource. These value will be adjusted via NTP
364 * to counteract clock drifting.
365 */
366 tk->tkr_mono.mult = clock->mult;
367 tk->tkr_raw.mult = clock->mult;
368 tk->ntp_err_mult = 0;
369 tk->skip_second_overflow = 0;
370}
371
372/* Timekeeper helper functions. */
373
374static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
375{
376 u64 nsec;
377
378 nsec = delta * tkr->mult + tkr->xtime_nsec;
379 nsec >>= tkr->shift;
380
381 return nsec;
382}
383
384static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
385{
386 u64 delta;
387
388 delta = timekeeping_get_delta(tkr);
389 return timekeeping_delta_to_ns(tkr, delta);
390}
391
392static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
393{
394 u64 delta;
395
396 /* calculate the delta since the last update_wall_time */
397 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
398 return timekeeping_delta_to_ns(tkr, delta);
399}
400
401/**
402 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
403 * @tkr: Timekeeping readout base from which we take the update
404 * @tkf: Pointer to NMI safe timekeeper
405 *
406 * We want to use this from any context including NMI and tracing /
407 * instrumenting the timekeeping code itself.
408 *
409 * Employ the latch technique; see @raw_write_seqcount_latch.
410 *
411 * So if a NMI hits the update of base[0] then it will use base[1]
412 * which is still consistent. In the worst case this can result is a
413 * slightly wrong timestamp (a few nanoseconds). See
414 * @ktime_get_mono_fast_ns.
415 */
416static void update_fast_timekeeper(const struct tk_read_base *tkr,
417 struct tk_fast *tkf)
418{
419 struct tk_read_base *base = tkf->base;
420
421 /* Force readers off to base[1] */
422 raw_write_seqcount_latch(&tkf->seq);
423
424 /* Update base[0] */
425 memcpy(base, tkr, sizeof(*base));
426
427 /* Force readers back to base[0] */
428 raw_write_seqcount_latch(&tkf->seq);
429
430 /* Update base[1] */
431 memcpy(base + 1, base, sizeof(*base));
432}
433
434static __always_inline u64 fast_tk_get_delta_ns(struct tk_read_base *tkr)
435{
436 u64 delta, cycles = tk_clock_read(tkr);
437
438 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
439 return timekeeping_delta_to_ns(tkr, delta);
440}
441
442static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
443{
444 struct tk_read_base *tkr;
445 unsigned int seq;
446 u64 now;
447
448 do {
449 seq = raw_read_seqcount_latch(&tkf->seq);
450 tkr = tkf->base + (seq & 0x01);
451 now = ktime_to_ns(tkr->base);
452 now += fast_tk_get_delta_ns(tkr);
453 } while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
454
455 return now;
456}
457
458/**
459 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
460 *
461 * This timestamp is not guaranteed to be monotonic across an update.
462 * The timestamp is calculated by:
463 *
464 * now = base_mono + clock_delta * slope
465 *
466 * So if the update lowers the slope, readers who are forced to the
467 * not yet updated second array are still using the old steeper slope.
468 *
469 * tmono
470 * ^
471 * | o n
472 * | o n
473 * | u
474 * | o
475 * |o
476 * |12345678---> reader order
477 *
478 * o = old slope
479 * u = update
480 * n = new slope
481 *
482 * So reader 6 will observe time going backwards versus reader 5.
483 *
484 * While other CPUs are likely to be able to observe that, the only way
485 * for a CPU local observation is when an NMI hits in the middle of
486 * the update. Timestamps taken from that NMI context might be ahead
487 * of the following timestamps. Callers need to be aware of that and
488 * deal with it.
489 */
490u64 notrace ktime_get_mono_fast_ns(void)
491{
492 return __ktime_get_fast_ns(&tk_fast_mono);
493}
494EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
495
496/**
497 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
498 *
499 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
500 * conversion factor is not affected by NTP/PTP correction.
501 */
502u64 notrace ktime_get_raw_fast_ns(void)
503{
504 return __ktime_get_fast_ns(&tk_fast_raw);
505}
506EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
507
508/**
509 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
510 *
511 * To keep it NMI safe since we're accessing from tracing, we're not using a
512 * separate timekeeper with updates to monotonic clock and boot offset
513 * protected with seqcounts. This has the following minor side effects:
514 *
515 * (1) Its possible that a timestamp be taken after the boot offset is updated
516 * but before the timekeeper is updated. If this happens, the new boot offset
517 * is added to the old timekeeping making the clock appear to update slightly
518 * earlier:
519 * CPU 0 CPU 1
520 * timekeeping_inject_sleeptime64()
521 * __timekeeping_inject_sleeptime(tk, delta);
522 * timestamp();
523 * timekeeping_update(tk, TK_CLEAR_NTP...);
524 *
525 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
526 * partially updated. Since the tk->offs_boot update is a rare event, this
527 * should be a rare occurrence which postprocessing should be able to handle.
528 *
529 * The caveats vs. timestamp ordering as documented for ktime_get_mono_fast_ns()
530 * apply as well.
531 */
532u64 notrace ktime_get_boot_fast_ns(void)
533{
534 struct timekeeper *tk = &tk_core.timekeeper;
535
536 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
537}
538EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
539
540/**
541 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
542 *
543 * The same limitations as described for ktime_get_boot_fast_ns() apply. The
544 * mono time and the TAI offset are not read atomically which may yield wrong
545 * readouts. However, an update of the TAI offset is an rare event e.g., caused
546 * by settime or adjtimex with an offset. The user of this function has to deal
547 * with the possibility of wrong timestamps in post processing.
548 */
549u64 notrace ktime_get_tai_fast_ns(void)
550{
551 struct timekeeper *tk = &tk_core.timekeeper;
552
553 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
554}
555EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
556
557static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
558{
559 struct tk_read_base *tkr;
560 u64 basem, baser, delta;
561 unsigned int seq;
562
563 do {
564 seq = raw_read_seqcount_latch(&tkf->seq);
565 tkr = tkf->base + (seq & 0x01);
566 basem = ktime_to_ns(tkr->base);
567 baser = ktime_to_ns(tkr->base_real);
568 delta = fast_tk_get_delta_ns(tkr);
569 } while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
570
571 if (mono)
572 *mono = basem + delta;
573 return baser + delta;
574}
575
576/**
577 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
578 *
579 * See ktime_get_mono_fast_ns() for documentation of the time stamp ordering.
580 */
581u64 ktime_get_real_fast_ns(void)
582{
583 return __ktime_get_real_fast(&tk_fast_mono, NULL);
584}
585EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
586
587/**
588 * ktime_get_fast_timestamps: - NMI safe timestamps
589 * @snapshot: Pointer to timestamp storage
590 *
591 * Stores clock monotonic, boottime and realtime timestamps.
592 *
593 * Boot time is a racy access on 32bit systems if the sleep time injection
594 * happens late during resume and not in timekeeping_resume(). That could
595 * be avoided by expanding struct tk_read_base with boot offset for 32bit
596 * and adding more overhead to the update. As this is a hard to observe
597 * once per resume event which can be filtered with reasonable effort using
598 * the accurate mono/real timestamps, it's probably not worth the trouble.
599 *
600 * Aside of that it might be possible on 32 and 64 bit to observe the
601 * following when the sleep time injection happens late:
602 *
603 * CPU 0 CPU 1
604 * timekeeping_resume()
605 * ktime_get_fast_timestamps()
606 * mono, real = __ktime_get_real_fast()
607 * inject_sleep_time()
608 * update boot offset
609 * boot = mono + bootoffset;
610 *
611 * That means that boot time already has the sleep time adjustment, but
612 * real time does not. On the next readout both are in sync again.
613 *
614 * Preventing this for 64bit is not really feasible without destroying the
615 * careful cache layout of the timekeeper because the sequence count and
616 * struct tk_read_base would then need two cache lines instead of one.
617 *
618 * Access to the time keeper clock source is disabled across the innermost
619 * steps of suspend/resume. The accessors still work, but the timestamps
620 * are frozen until time keeping is resumed which happens very early.
621 *
622 * For regular suspend/resume there is no observable difference vs. sched
623 * clock, but it might affect some of the nasty low level debug printks.
624 *
625 * OTOH, access to sched clock is not guaranteed across suspend/resume on
626 * all systems either so it depends on the hardware in use.
627 *
628 * If that turns out to be a real problem then this could be mitigated by
629 * using sched clock in a similar way as during early boot. But it's not as
630 * trivial as on early boot because it needs some careful protection
631 * against the clock monotonic timestamp jumping backwards on resume.
632 */
633void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
634{
635 struct timekeeper *tk = &tk_core.timekeeper;
636
637 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
638 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
639}
640
641/**
642 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
643 * @tk: Timekeeper to snapshot.
644 *
645 * It generally is unsafe to access the clocksource after timekeeping has been
646 * suspended, so take a snapshot of the readout base of @tk and use it as the
647 * fast timekeeper's readout base while suspended. It will return the same
648 * number of cycles every time until timekeeping is resumed at which time the
649 * proper readout base for the fast timekeeper will be restored automatically.
650 */
651static void halt_fast_timekeeper(const struct timekeeper *tk)
652{
653 static struct tk_read_base tkr_dummy;
654 const struct tk_read_base *tkr = &tk->tkr_mono;
655
656 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
657 cycles_at_suspend = tk_clock_read(tkr);
658 tkr_dummy.clock = &dummy_clock;
659 tkr_dummy.base_real = tkr->base + tk->offs_real;
660 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
661
662 tkr = &tk->tkr_raw;
663 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
664 tkr_dummy.clock = &dummy_clock;
665 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
666}
667
668static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
669
670static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
671{
672 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
673}
674
675/**
676 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
677 * @nb: Pointer to the notifier block to register
678 */
679int pvclock_gtod_register_notifier(struct notifier_block *nb)
680{
681 struct timekeeper *tk = &tk_core.timekeeper;
682 unsigned long flags;
683 int ret;
684
685 raw_spin_lock_irqsave(&timekeeper_lock, flags);
686 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
687 update_pvclock_gtod(tk, true);
688 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
689
690 return ret;
691}
692EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
693
694/**
695 * pvclock_gtod_unregister_notifier - unregister a pvclock
696 * timedata update listener
697 * @nb: Pointer to the notifier block to unregister
698 */
699int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
700{
701 unsigned long flags;
702 int ret;
703
704 raw_spin_lock_irqsave(&timekeeper_lock, flags);
705 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
706 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
707
708 return ret;
709}
710EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
711
712/*
713 * tk_update_leap_state - helper to update the next_leap_ktime
714 */
715static inline void tk_update_leap_state(struct timekeeper *tk)
716{
717 tk->next_leap_ktime = ntp_get_next_leap();
718 if (tk->next_leap_ktime != KTIME_MAX)
719 /* Convert to monotonic time */
720 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
721}
722
723/*
724 * Update the ktime_t based scalar nsec members of the timekeeper
725 */
726static inline void tk_update_ktime_data(struct timekeeper *tk)
727{
728 u64 seconds;
729 u32 nsec;
730
731 /*
732 * The xtime based monotonic readout is:
733 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
734 * The ktime based monotonic readout is:
735 * nsec = base_mono + now();
736 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
737 */
738 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
739 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
740 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
741
742 /*
743 * The sum of the nanoseconds portions of xtime and
744 * wall_to_monotonic can be greater/equal one second. Take
745 * this into account before updating tk->ktime_sec.
746 */
747 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
748 if (nsec >= NSEC_PER_SEC)
749 seconds++;
750 tk->ktime_sec = seconds;
751
752 /* Update the monotonic raw base */
753 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
754}
755
756/* must hold timekeeper_lock */
757static void timekeeping_update(struct timekeeper *tk, unsigned int action)
758{
759 if (action & TK_CLEAR_NTP) {
760 tk->ntp_error = 0;
761 ntp_clear();
762 }
763
764 tk_update_leap_state(tk);
765 tk_update_ktime_data(tk);
766
767 update_vsyscall(tk);
768 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
769
770 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
771 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
772 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
773
774 if (action & TK_CLOCK_WAS_SET)
775 tk->clock_was_set_seq++;
776 /*
777 * The mirroring of the data to the shadow-timekeeper needs
778 * to happen last here to ensure we don't over-write the
779 * timekeeper structure on the next update with stale data
780 */
781 if (action & TK_MIRROR)
782 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
783 sizeof(tk_core.timekeeper));
784}
785
786/**
787 * timekeeping_forward_now - update clock to the current time
788 * @tk: Pointer to the timekeeper to update
789 *
790 * Forward the current clock to update its state since the last call to
791 * update_wall_time(). This is useful before significant clock changes,
792 * as it avoids having to deal with this time offset explicitly.
793 */
794static void timekeeping_forward_now(struct timekeeper *tk)
795{
796 u64 cycle_now, delta;
797
798 cycle_now = tk_clock_read(&tk->tkr_mono);
799 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
800 tk->tkr_mono.cycle_last = cycle_now;
801 tk->tkr_raw.cycle_last = cycle_now;
802
803 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
804 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
805
806 tk_normalize_xtime(tk);
807}
808
809/**
810 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
811 * @ts: pointer to the timespec to be set
812 *
813 * Returns the time of day in a timespec64 (WARN if suspended).
814 */
815void ktime_get_real_ts64(struct timespec64 *ts)
816{
817 struct timekeeper *tk = &tk_core.timekeeper;
818 unsigned int seq;
819 u64 nsecs;
820
821 WARN_ON(timekeeping_suspended);
822
823 do {
824 seq = read_seqcount_begin(&tk_core.seq);
825
826 ts->tv_sec = tk->xtime_sec;
827 nsecs = timekeeping_get_ns(&tk->tkr_mono);
828
829 } while (read_seqcount_retry(&tk_core.seq, seq));
830
831 ts->tv_nsec = 0;
832 timespec64_add_ns(ts, nsecs);
833}
834EXPORT_SYMBOL(ktime_get_real_ts64);
835
836ktime_t ktime_get(void)
837{
838 struct timekeeper *tk = &tk_core.timekeeper;
839 unsigned int seq;
840 ktime_t base;
841 u64 nsecs;
842
843 WARN_ON(timekeeping_suspended);
844
845 do {
846 seq = read_seqcount_begin(&tk_core.seq);
847 base = tk->tkr_mono.base;
848 nsecs = timekeeping_get_ns(&tk->tkr_mono);
849
850 } while (read_seqcount_retry(&tk_core.seq, seq));
851
852 return ktime_add_ns(base, nsecs);
853}
854EXPORT_SYMBOL_GPL(ktime_get);
855
856u32 ktime_get_resolution_ns(void)
857{
858 struct timekeeper *tk = &tk_core.timekeeper;
859 unsigned int seq;
860 u32 nsecs;
861
862 WARN_ON(timekeeping_suspended);
863
864 do {
865 seq = read_seqcount_begin(&tk_core.seq);
866 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
867 } while (read_seqcount_retry(&tk_core.seq, seq));
868
869 return nsecs;
870}
871EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
872
873static ktime_t *offsets[TK_OFFS_MAX] = {
874 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
875 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
876 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
877};
878
879ktime_t ktime_get_with_offset(enum tk_offsets offs)
880{
881 struct timekeeper *tk = &tk_core.timekeeper;
882 unsigned int seq;
883 ktime_t base, *offset = offsets[offs];
884 u64 nsecs;
885
886 WARN_ON(timekeeping_suspended);
887
888 do {
889 seq = read_seqcount_begin(&tk_core.seq);
890 base = ktime_add(tk->tkr_mono.base, *offset);
891 nsecs = timekeeping_get_ns(&tk->tkr_mono);
892
893 } while (read_seqcount_retry(&tk_core.seq, seq));
894
895 return ktime_add_ns(base, nsecs);
896
897}
898EXPORT_SYMBOL_GPL(ktime_get_with_offset);
899
900ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
901{
902 struct timekeeper *tk = &tk_core.timekeeper;
903 unsigned int seq;
904 ktime_t base, *offset = offsets[offs];
905 u64 nsecs;
906
907 WARN_ON(timekeeping_suspended);
908
909 do {
910 seq = read_seqcount_begin(&tk_core.seq);
911 base = ktime_add(tk->tkr_mono.base, *offset);
912 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
913
914 } while (read_seqcount_retry(&tk_core.seq, seq));
915
916 return ktime_add_ns(base, nsecs);
917}
918EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
919
920/**
921 * ktime_mono_to_any() - convert monotonic time to any other time
922 * @tmono: time to convert.
923 * @offs: which offset to use
924 */
925ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
926{
927 ktime_t *offset = offsets[offs];
928 unsigned int seq;
929 ktime_t tconv;
930
931 do {
932 seq = read_seqcount_begin(&tk_core.seq);
933 tconv = ktime_add(tmono, *offset);
934 } while (read_seqcount_retry(&tk_core.seq, seq));
935
936 return tconv;
937}
938EXPORT_SYMBOL_GPL(ktime_mono_to_any);
939
940/**
941 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
942 */
943ktime_t ktime_get_raw(void)
944{
945 struct timekeeper *tk = &tk_core.timekeeper;
946 unsigned int seq;
947 ktime_t base;
948 u64 nsecs;
949
950 do {
951 seq = read_seqcount_begin(&tk_core.seq);
952 base = tk->tkr_raw.base;
953 nsecs = timekeeping_get_ns(&tk->tkr_raw);
954
955 } while (read_seqcount_retry(&tk_core.seq, seq));
956
957 return ktime_add_ns(base, nsecs);
958}
959EXPORT_SYMBOL_GPL(ktime_get_raw);
960
961/**
962 * ktime_get_ts64 - get the monotonic clock in timespec64 format
963 * @ts: pointer to timespec variable
964 *
965 * The function calculates the monotonic clock from the realtime
966 * clock and the wall_to_monotonic offset and stores the result
967 * in normalized timespec64 format in the variable pointed to by @ts.
968 */
969void ktime_get_ts64(struct timespec64 *ts)
970{
971 struct timekeeper *tk = &tk_core.timekeeper;
972 struct timespec64 tomono;
973 unsigned int seq;
974 u64 nsec;
975
976 WARN_ON(timekeeping_suspended);
977
978 do {
979 seq = read_seqcount_begin(&tk_core.seq);
980 ts->tv_sec = tk->xtime_sec;
981 nsec = timekeeping_get_ns(&tk->tkr_mono);
982 tomono = tk->wall_to_monotonic;
983
984 } while (read_seqcount_retry(&tk_core.seq, seq));
985
986 ts->tv_sec += tomono.tv_sec;
987 ts->tv_nsec = 0;
988 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
989}
990EXPORT_SYMBOL_GPL(ktime_get_ts64);
991
992/**
993 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
994 *
995 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
996 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
997 * works on both 32 and 64 bit systems. On 32 bit systems the readout
998 * covers ~136 years of uptime which should be enough to prevent
999 * premature wrap arounds.
1000 */
1001time64_t ktime_get_seconds(void)
1002{
1003 struct timekeeper *tk = &tk_core.timekeeper;
1004
1005 WARN_ON(timekeeping_suspended);
1006 return tk->ktime_sec;
1007}
1008EXPORT_SYMBOL_GPL(ktime_get_seconds);
1009
1010/**
1011 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1012 *
1013 * Returns the wall clock seconds since 1970.
1014 *
1015 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1016 * 32bit systems the access must be protected with the sequence
1017 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1018 * value.
1019 */
1020time64_t ktime_get_real_seconds(void)
1021{
1022 struct timekeeper *tk = &tk_core.timekeeper;
1023 time64_t seconds;
1024 unsigned int seq;
1025
1026 if (IS_ENABLED(CONFIG_64BIT))
1027 return tk->xtime_sec;
1028
1029 do {
1030 seq = read_seqcount_begin(&tk_core.seq);
1031 seconds = tk->xtime_sec;
1032
1033 } while (read_seqcount_retry(&tk_core.seq, seq));
1034
1035 return seconds;
1036}
1037EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1038
1039/**
1040 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1041 * but without the sequence counter protect. This internal function
1042 * is called just when timekeeping lock is already held.
1043 */
1044noinstr time64_t __ktime_get_real_seconds(void)
1045{
1046 struct timekeeper *tk = &tk_core.timekeeper;
1047
1048 return tk->xtime_sec;
1049}
1050
1051/**
1052 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1053 * @systime_snapshot: pointer to struct receiving the system time snapshot
1054 */
1055void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1056{
1057 struct timekeeper *tk = &tk_core.timekeeper;
1058 unsigned int seq;
1059 ktime_t base_raw;
1060 ktime_t base_real;
1061 u64 nsec_raw;
1062 u64 nsec_real;
1063 u64 now;
1064
1065 WARN_ON_ONCE(timekeeping_suspended);
1066
1067 do {
1068 seq = read_seqcount_begin(&tk_core.seq);
1069 now = tk_clock_read(&tk->tkr_mono);
1070 systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1071 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1072 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1073 base_real = ktime_add(tk->tkr_mono.base,
1074 tk_core.timekeeper.offs_real);
1075 base_raw = tk->tkr_raw.base;
1076 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1077 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1078 } while (read_seqcount_retry(&tk_core.seq, seq));
1079
1080 systime_snapshot->cycles = now;
1081 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1082 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1083}
1084EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1085
1086/* Scale base by mult/div checking for overflow */
1087static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1088{
1089 u64 tmp, rem;
1090
1091 tmp = div64_u64_rem(*base, div, &rem);
1092
1093 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1094 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1095 return -EOVERFLOW;
1096 tmp *= mult;
1097
1098 rem = div64_u64(rem * mult, div);
1099 *base = tmp + rem;
1100 return 0;
1101}
1102
1103/**
1104 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1105 * @history: Snapshot representing start of history
1106 * @partial_history_cycles: Cycle offset into history (fractional part)
1107 * @total_history_cycles: Total history length in cycles
1108 * @discontinuity: True indicates clock was set on history period
1109 * @ts: Cross timestamp that should be adjusted using
1110 * partial/total ratio
1111 *
1112 * Helper function used by get_device_system_crosststamp() to correct the
1113 * crosstimestamp corresponding to the start of the current interval to the
1114 * system counter value (timestamp point) provided by the driver. The
1115 * total_history_* quantities are the total history starting at the provided
1116 * reference point and ending at the start of the current interval. The cycle
1117 * count between the driver timestamp point and the start of the current
1118 * interval is partial_history_cycles.
1119 */
1120static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1121 u64 partial_history_cycles,
1122 u64 total_history_cycles,
1123 bool discontinuity,
1124 struct system_device_crosststamp *ts)
1125{
1126 struct timekeeper *tk = &tk_core.timekeeper;
1127 u64 corr_raw, corr_real;
1128 bool interp_forward;
1129 int ret;
1130
1131 if (total_history_cycles == 0 || partial_history_cycles == 0)
1132 return 0;
1133
1134 /* Interpolate shortest distance from beginning or end of history */
1135 interp_forward = partial_history_cycles > total_history_cycles / 2;
1136 partial_history_cycles = interp_forward ?
1137 total_history_cycles - partial_history_cycles :
1138 partial_history_cycles;
1139
1140 /*
1141 * Scale the monotonic raw time delta by:
1142 * partial_history_cycles / total_history_cycles
1143 */
1144 corr_raw = (u64)ktime_to_ns(
1145 ktime_sub(ts->sys_monoraw, history->raw));
1146 ret = scale64_check_overflow(partial_history_cycles,
1147 total_history_cycles, &corr_raw);
1148 if (ret)
1149 return ret;
1150
1151 /*
1152 * If there is a discontinuity in the history, scale monotonic raw
1153 * correction by:
1154 * mult(real)/mult(raw) yielding the realtime correction
1155 * Otherwise, calculate the realtime correction similar to monotonic
1156 * raw calculation
1157 */
1158 if (discontinuity) {
1159 corr_real = mul_u64_u32_div
1160 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1161 } else {
1162 corr_real = (u64)ktime_to_ns(
1163 ktime_sub(ts->sys_realtime, history->real));
1164 ret = scale64_check_overflow(partial_history_cycles,
1165 total_history_cycles, &corr_real);
1166 if (ret)
1167 return ret;
1168 }
1169
1170 /* Fixup monotonic raw and real time time values */
1171 if (interp_forward) {
1172 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1173 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1174 } else {
1175 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1176 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1177 }
1178
1179 return 0;
1180}
1181
1182/*
1183 * timestamp_in_interval - true if ts is chronologically in [start, end]
1184 *
1185 * True if ts occurs chronologically at or after start, and before or at end.
1186 */
1187static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1188{
1189 if (ts >= start && ts <= end)
1190 return true;
1191 if (start > end && (ts >= start || ts <= end))
1192 return true;
1193 return false;
1194}
1195
1196/**
1197 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1198 * @get_time_fn: Callback to get simultaneous device time and
1199 * system counter from the device driver
1200 * @ctx: Context passed to get_time_fn()
1201 * @history_begin: Historical reference point used to interpolate system
1202 * time when counter provided by the driver is before the current interval
1203 * @xtstamp: Receives simultaneously captured system and device time
1204 *
1205 * Reads a timestamp from a device and correlates it to system time
1206 */
1207int get_device_system_crosststamp(int (*get_time_fn)
1208 (ktime_t *device_time,
1209 struct system_counterval_t *sys_counterval,
1210 void *ctx),
1211 void *ctx,
1212 struct system_time_snapshot *history_begin,
1213 struct system_device_crosststamp *xtstamp)
1214{
1215 struct system_counterval_t system_counterval;
1216 struct timekeeper *tk = &tk_core.timekeeper;
1217 u64 cycles, now, interval_start;
1218 unsigned int clock_was_set_seq = 0;
1219 ktime_t base_real, base_raw;
1220 u64 nsec_real, nsec_raw;
1221 u8 cs_was_changed_seq;
1222 unsigned int seq;
1223 bool do_interp;
1224 int ret;
1225
1226 do {
1227 seq = read_seqcount_begin(&tk_core.seq);
1228 /*
1229 * Try to synchronously capture device time and a system
1230 * counter value calling back into the device driver
1231 */
1232 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1233 if (ret)
1234 return ret;
1235
1236 /*
1237 * Verify that the clocksource ID associated with the captured
1238 * system counter value is the same as for the currently
1239 * installed timekeeper clocksource
1240 */
1241 if (system_counterval.cs_id == CSID_GENERIC ||
1242 tk->tkr_mono.clock->id != system_counterval.cs_id)
1243 return -ENODEV;
1244 cycles = system_counterval.cycles;
1245
1246 /*
1247 * Check whether the system counter value provided by the
1248 * device driver is on the current timekeeping interval.
1249 */
1250 now = tk_clock_read(&tk->tkr_mono);
1251 interval_start = tk->tkr_mono.cycle_last;
1252 if (!timestamp_in_interval(interval_start, now, cycles)) {
1253 clock_was_set_seq = tk->clock_was_set_seq;
1254 cs_was_changed_seq = tk->cs_was_changed_seq;
1255 cycles = interval_start;
1256 do_interp = true;
1257 } else {
1258 do_interp = false;
1259 }
1260
1261 base_real = ktime_add(tk->tkr_mono.base,
1262 tk_core.timekeeper.offs_real);
1263 base_raw = tk->tkr_raw.base;
1264
1265 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1266 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1267 } while (read_seqcount_retry(&tk_core.seq, seq));
1268
1269 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1270 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1271
1272 /*
1273 * Interpolate if necessary, adjusting back from the start of the
1274 * current interval
1275 */
1276 if (do_interp) {
1277 u64 partial_history_cycles, total_history_cycles;
1278 bool discontinuity;
1279
1280 /*
1281 * Check that the counter value is not before the provided
1282 * history reference and that the history doesn't cross a
1283 * clocksource change
1284 */
1285 if (!history_begin ||
1286 !timestamp_in_interval(history_begin->cycles,
1287 cycles, system_counterval.cycles) ||
1288 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1289 return -EINVAL;
1290 partial_history_cycles = cycles - system_counterval.cycles;
1291 total_history_cycles = cycles - history_begin->cycles;
1292 discontinuity =
1293 history_begin->clock_was_set_seq != clock_was_set_seq;
1294
1295 ret = adjust_historical_crosststamp(history_begin,
1296 partial_history_cycles,
1297 total_history_cycles,
1298 discontinuity, xtstamp);
1299 if (ret)
1300 return ret;
1301 }
1302
1303 return 0;
1304}
1305EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1306
1307/**
1308 * do_settimeofday64 - Sets the time of day.
1309 * @ts: pointer to the timespec64 variable containing the new time
1310 *
1311 * Sets the time of day to the new time and update NTP and notify hrtimers
1312 */
1313int do_settimeofday64(const struct timespec64 *ts)
1314{
1315 struct timekeeper *tk = &tk_core.timekeeper;
1316 struct timespec64 ts_delta, xt;
1317 unsigned long flags;
1318 int ret = 0;
1319
1320 if (!timespec64_valid_settod(ts))
1321 return -EINVAL;
1322
1323 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1324 write_seqcount_begin(&tk_core.seq);
1325
1326 timekeeping_forward_now(tk);
1327
1328 xt = tk_xtime(tk);
1329 ts_delta = timespec64_sub(*ts, xt);
1330
1331 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1332 ret = -EINVAL;
1333 goto out;
1334 }
1335
1336 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1337
1338 tk_set_xtime(tk, ts);
1339out:
1340 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1341
1342 write_seqcount_end(&tk_core.seq);
1343 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1344
1345 /* Signal hrtimers about time change */
1346 clock_was_set(CLOCK_SET_WALL);
1347
1348 if (!ret) {
1349 audit_tk_injoffset(ts_delta);
1350 add_device_randomness(ts, sizeof(*ts));
1351 }
1352
1353 return ret;
1354}
1355EXPORT_SYMBOL(do_settimeofday64);
1356
1357/**
1358 * timekeeping_inject_offset - Adds or subtracts from the current time.
1359 * @ts: Pointer to the timespec variable containing the offset
1360 *
1361 * Adds or subtracts an offset value from the current time.
1362 */
1363static int timekeeping_inject_offset(const struct timespec64 *ts)
1364{
1365 struct timekeeper *tk = &tk_core.timekeeper;
1366 unsigned long flags;
1367 struct timespec64 tmp;
1368 int ret = 0;
1369
1370 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1371 return -EINVAL;
1372
1373 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1374 write_seqcount_begin(&tk_core.seq);
1375
1376 timekeeping_forward_now(tk);
1377
1378 /* Make sure the proposed value is valid */
1379 tmp = timespec64_add(tk_xtime(tk), *ts);
1380 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1381 !timespec64_valid_settod(&tmp)) {
1382 ret = -EINVAL;
1383 goto error;
1384 }
1385
1386 tk_xtime_add(tk, ts);
1387 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1388
1389error: /* even if we error out, we forwarded the time, so call update */
1390 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1391
1392 write_seqcount_end(&tk_core.seq);
1393 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1394
1395 /* Signal hrtimers about time change */
1396 clock_was_set(CLOCK_SET_WALL);
1397
1398 return ret;
1399}
1400
1401/*
1402 * Indicates if there is an offset between the system clock and the hardware
1403 * clock/persistent clock/rtc.
1404 */
1405int persistent_clock_is_local;
1406
1407/*
1408 * Adjust the time obtained from the CMOS to be UTC time instead of
1409 * local time.
1410 *
1411 * This is ugly, but preferable to the alternatives. Otherwise we
1412 * would either need to write a program to do it in /etc/rc (and risk
1413 * confusion if the program gets run more than once; it would also be
1414 * hard to make the program warp the clock precisely n hours) or
1415 * compile in the timezone information into the kernel. Bad, bad....
1416 *
1417 * - TYT, 1992-01-01
1418 *
1419 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1420 * as real UNIX machines always do it. This avoids all headaches about
1421 * daylight saving times and warping kernel clocks.
1422 */
1423void timekeeping_warp_clock(void)
1424{
1425 if (sys_tz.tz_minuteswest != 0) {
1426 struct timespec64 adjust;
1427
1428 persistent_clock_is_local = 1;
1429 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1430 adjust.tv_nsec = 0;
1431 timekeeping_inject_offset(&adjust);
1432 }
1433}
1434
1435/*
1436 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1437 */
1438static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1439{
1440 tk->tai_offset = tai_offset;
1441 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1442}
1443
1444/*
1445 * change_clocksource - Swaps clocksources if a new one is available
1446 *
1447 * Accumulates current time interval and initializes new clocksource
1448 */
1449static int change_clocksource(void *data)
1450{
1451 struct timekeeper *tk = &tk_core.timekeeper;
1452 struct clocksource *new, *old = NULL;
1453 unsigned long flags;
1454 bool change = false;
1455
1456 new = (struct clocksource *) data;
1457
1458 /*
1459 * If the cs is in module, get a module reference. Succeeds
1460 * for built-in code (owner == NULL) as well.
1461 */
1462 if (try_module_get(new->owner)) {
1463 if (!new->enable || new->enable(new) == 0)
1464 change = true;
1465 else
1466 module_put(new->owner);
1467 }
1468
1469 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1470 write_seqcount_begin(&tk_core.seq);
1471
1472 timekeeping_forward_now(tk);
1473
1474 if (change) {
1475 old = tk->tkr_mono.clock;
1476 tk_setup_internals(tk, new);
1477 }
1478
1479 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1480
1481 write_seqcount_end(&tk_core.seq);
1482 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1483
1484 if (old) {
1485 if (old->disable)
1486 old->disable(old);
1487
1488 module_put(old->owner);
1489 }
1490
1491 return 0;
1492}
1493
1494/**
1495 * timekeeping_notify - Install a new clock source
1496 * @clock: pointer to the clock source
1497 *
1498 * This function is called from clocksource.c after a new, better clock
1499 * source has been registered. The caller holds the clocksource_mutex.
1500 */
1501int timekeeping_notify(struct clocksource *clock)
1502{
1503 struct timekeeper *tk = &tk_core.timekeeper;
1504
1505 if (tk->tkr_mono.clock == clock)
1506 return 0;
1507 stop_machine(change_clocksource, clock, NULL);
1508 tick_clock_notify();
1509 return tk->tkr_mono.clock == clock ? 0 : -1;
1510}
1511
1512/**
1513 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1514 * @ts: pointer to the timespec64 to be set
1515 *
1516 * Returns the raw monotonic time (completely un-modified by ntp)
1517 */
1518void ktime_get_raw_ts64(struct timespec64 *ts)
1519{
1520 struct timekeeper *tk = &tk_core.timekeeper;
1521 unsigned int seq;
1522 u64 nsecs;
1523
1524 do {
1525 seq = read_seqcount_begin(&tk_core.seq);
1526 ts->tv_sec = tk->raw_sec;
1527 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1528
1529 } while (read_seqcount_retry(&tk_core.seq, seq));
1530
1531 ts->tv_nsec = 0;
1532 timespec64_add_ns(ts, nsecs);
1533}
1534EXPORT_SYMBOL(ktime_get_raw_ts64);
1535
1536
1537/**
1538 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1539 */
1540int timekeeping_valid_for_hres(void)
1541{
1542 struct timekeeper *tk = &tk_core.timekeeper;
1543 unsigned int seq;
1544 int ret;
1545
1546 do {
1547 seq = read_seqcount_begin(&tk_core.seq);
1548
1549 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1550
1551 } while (read_seqcount_retry(&tk_core.seq, seq));
1552
1553 return ret;
1554}
1555
1556/**
1557 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1558 */
1559u64 timekeeping_max_deferment(void)
1560{
1561 struct timekeeper *tk = &tk_core.timekeeper;
1562 unsigned int seq;
1563 u64 ret;
1564
1565 do {
1566 seq = read_seqcount_begin(&tk_core.seq);
1567
1568 ret = tk->tkr_mono.clock->max_idle_ns;
1569
1570 } while (read_seqcount_retry(&tk_core.seq, seq));
1571
1572 return ret;
1573}
1574
1575/**
1576 * read_persistent_clock64 - Return time from the persistent clock.
1577 * @ts: Pointer to the storage for the readout value
1578 *
1579 * Weak dummy function for arches that do not yet support it.
1580 * Reads the time from the battery backed persistent clock.
1581 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1582 *
1583 * XXX - Do be sure to remove it once all arches implement it.
1584 */
1585void __weak read_persistent_clock64(struct timespec64 *ts)
1586{
1587 ts->tv_sec = 0;
1588 ts->tv_nsec = 0;
1589}
1590
1591/**
1592 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1593 * from the boot.
1594 * @wall_time: current time as returned by persistent clock
1595 * @boot_offset: offset that is defined as wall_time - boot_time
1596 *
1597 * Weak dummy function for arches that do not yet support it.
1598 *
1599 * The default function calculates offset based on the current value of
1600 * local_clock(). This way architectures that support sched_clock() but don't
1601 * support dedicated boot time clock will provide the best estimate of the
1602 * boot time.
1603 */
1604void __weak __init
1605read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1606 struct timespec64 *boot_offset)
1607{
1608 read_persistent_clock64(wall_time);
1609 *boot_offset = ns_to_timespec64(local_clock());
1610}
1611
1612/*
1613 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1614 *
1615 * The flag starts of false and is only set when a suspend reaches
1616 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1617 * timekeeper clocksource is not stopping across suspend and has been
1618 * used to update sleep time. If the timekeeper clocksource has stopped
1619 * then the flag stays true and is used by the RTC resume code to decide
1620 * whether sleeptime must be injected and if so the flag gets false then.
1621 *
1622 * If a suspend fails before reaching timekeeping_resume() then the flag
1623 * stays false and prevents erroneous sleeptime injection.
1624 */
1625static bool suspend_timing_needed;
1626
1627/* Flag for if there is a persistent clock on this platform */
1628static bool persistent_clock_exists;
1629
1630/*
1631 * timekeeping_init - Initializes the clocksource and common timekeeping values
1632 */
1633void __init timekeeping_init(void)
1634{
1635 struct timespec64 wall_time, boot_offset, wall_to_mono;
1636 struct timekeeper *tk = &tk_core.timekeeper;
1637 struct clocksource *clock;
1638 unsigned long flags;
1639
1640 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1641 if (timespec64_valid_settod(&wall_time) &&
1642 timespec64_to_ns(&wall_time) > 0) {
1643 persistent_clock_exists = true;
1644 } else if (timespec64_to_ns(&wall_time) != 0) {
1645 pr_warn("Persistent clock returned invalid value");
1646 wall_time = (struct timespec64){0};
1647 }
1648
1649 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1650 boot_offset = (struct timespec64){0};
1651
1652 /*
1653 * We want set wall_to_mono, so the following is true:
1654 * wall time + wall_to_mono = boot time
1655 */
1656 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1657
1658 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1659 write_seqcount_begin(&tk_core.seq);
1660 ntp_init();
1661
1662 clock = clocksource_default_clock();
1663 if (clock->enable)
1664 clock->enable(clock);
1665 tk_setup_internals(tk, clock);
1666
1667 tk_set_xtime(tk, &wall_time);
1668 tk->raw_sec = 0;
1669
1670 tk_set_wall_to_mono(tk, wall_to_mono);
1671
1672 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1673
1674 write_seqcount_end(&tk_core.seq);
1675 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1676}
1677
1678/* time in seconds when suspend began for persistent clock */
1679static struct timespec64 timekeeping_suspend_time;
1680
1681/**
1682 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1683 * @tk: Pointer to the timekeeper to be updated
1684 * @delta: Pointer to the delta value in timespec64 format
1685 *
1686 * Takes a timespec offset measuring a suspend interval and properly
1687 * adds the sleep offset to the timekeeping variables.
1688 */
1689static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1690 const struct timespec64 *delta)
1691{
1692 if (!timespec64_valid_strict(delta)) {
1693 printk_deferred(KERN_WARNING
1694 "__timekeeping_inject_sleeptime: Invalid "
1695 "sleep delta value!\n");
1696 return;
1697 }
1698 tk_xtime_add(tk, delta);
1699 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1700 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1701 tk_debug_account_sleep_time(delta);
1702}
1703
1704#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1705/*
1706 * We have three kinds of time sources to use for sleep time
1707 * injection, the preference order is:
1708 * 1) non-stop clocksource
1709 * 2) persistent clock (ie: RTC accessible when irqs are off)
1710 * 3) RTC
1711 *
1712 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1713 * If system has neither 1) nor 2), 3) will be used finally.
1714 *
1715 *
1716 * If timekeeping has injected sleeptime via either 1) or 2),
1717 * 3) becomes needless, so in this case we don't need to call
1718 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1719 * means.
1720 */
1721bool timekeeping_rtc_skipresume(void)
1722{
1723 return !suspend_timing_needed;
1724}
1725
1726/*
1727 * 1) can be determined whether to use or not only when doing
1728 * timekeeping_resume() which is invoked after rtc_suspend(),
1729 * so we can't skip rtc_suspend() surely if system has 1).
1730 *
1731 * But if system has 2), 2) will definitely be used, so in this
1732 * case we don't need to call rtc_suspend(), and this is what
1733 * timekeeping_rtc_skipsuspend() means.
1734 */
1735bool timekeeping_rtc_skipsuspend(void)
1736{
1737 return persistent_clock_exists;
1738}
1739
1740/**
1741 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1742 * @delta: pointer to a timespec64 delta value
1743 *
1744 * This hook is for architectures that cannot support read_persistent_clock64
1745 * because their RTC/persistent clock is only accessible when irqs are enabled.
1746 * and also don't have an effective nonstop clocksource.
1747 *
1748 * This function should only be called by rtc_resume(), and allows
1749 * a suspend offset to be injected into the timekeeping values.
1750 */
1751void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1752{
1753 struct timekeeper *tk = &tk_core.timekeeper;
1754 unsigned long flags;
1755
1756 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1757 write_seqcount_begin(&tk_core.seq);
1758
1759 suspend_timing_needed = false;
1760
1761 timekeeping_forward_now(tk);
1762
1763 __timekeeping_inject_sleeptime(tk, delta);
1764
1765 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1766
1767 write_seqcount_end(&tk_core.seq);
1768 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1769
1770 /* Signal hrtimers about time change */
1771 clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1772}
1773#endif
1774
1775/**
1776 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1777 */
1778void timekeeping_resume(void)
1779{
1780 struct timekeeper *tk = &tk_core.timekeeper;
1781 struct clocksource *clock = tk->tkr_mono.clock;
1782 unsigned long flags;
1783 struct timespec64 ts_new, ts_delta;
1784 u64 cycle_now, nsec;
1785 bool inject_sleeptime = false;
1786
1787 read_persistent_clock64(&ts_new);
1788
1789 clockevents_resume();
1790 clocksource_resume();
1791
1792 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1793 write_seqcount_begin(&tk_core.seq);
1794
1795 /*
1796 * After system resumes, we need to calculate the suspended time and
1797 * compensate it for the OS time. There are 3 sources that could be
1798 * used: Nonstop clocksource during suspend, persistent clock and rtc
1799 * device.
1800 *
1801 * One specific platform may have 1 or 2 or all of them, and the
1802 * preference will be:
1803 * suspend-nonstop clocksource -> persistent clock -> rtc
1804 * The less preferred source will only be tried if there is no better
1805 * usable source. The rtc part is handled separately in rtc core code.
1806 */
1807 cycle_now = tk_clock_read(&tk->tkr_mono);
1808 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1809 if (nsec > 0) {
1810 ts_delta = ns_to_timespec64(nsec);
1811 inject_sleeptime = true;
1812 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1813 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1814 inject_sleeptime = true;
1815 }
1816
1817 if (inject_sleeptime) {
1818 suspend_timing_needed = false;
1819 __timekeeping_inject_sleeptime(tk, &ts_delta);
1820 }
1821
1822 /* Re-base the last cycle value */
1823 tk->tkr_mono.cycle_last = cycle_now;
1824 tk->tkr_raw.cycle_last = cycle_now;
1825
1826 tk->ntp_error = 0;
1827 timekeeping_suspended = 0;
1828 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1829 write_seqcount_end(&tk_core.seq);
1830 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1831
1832 touch_softlockup_watchdog();
1833
1834 /* Resume the clockevent device(s) and hrtimers */
1835 tick_resume();
1836 /* Notify timerfd as resume is equivalent to clock_was_set() */
1837 timerfd_resume();
1838}
1839
1840int timekeeping_suspend(void)
1841{
1842 struct timekeeper *tk = &tk_core.timekeeper;
1843 unsigned long flags;
1844 struct timespec64 delta, delta_delta;
1845 static struct timespec64 old_delta;
1846 struct clocksource *curr_clock;
1847 u64 cycle_now;
1848
1849 read_persistent_clock64(&timekeeping_suspend_time);
1850
1851 /*
1852 * On some systems the persistent_clock can not be detected at
1853 * timekeeping_init by its return value, so if we see a valid
1854 * value returned, update the persistent_clock_exists flag.
1855 */
1856 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1857 persistent_clock_exists = true;
1858
1859 suspend_timing_needed = true;
1860
1861 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1862 write_seqcount_begin(&tk_core.seq);
1863 timekeeping_forward_now(tk);
1864 timekeeping_suspended = 1;
1865
1866 /*
1867 * Since we've called forward_now, cycle_last stores the value
1868 * just read from the current clocksource. Save this to potentially
1869 * use in suspend timing.
1870 */
1871 curr_clock = tk->tkr_mono.clock;
1872 cycle_now = tk->tkr_mono.cycle_last;
1873 clocksource_start_suspend_timing(curr_clock, cycle_now);
1874
1875 if (persistent_clock_exists) {
1876 /*
1877 * To avoid drift caused by repeated suspend/resumes,
1878 * which each can add ~1 second drift error,
1879 * try to compensate so the difference in system time
1880 * and persistent_clock time stays close to constant.
1881 */
1882 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1883 delta_delta = timespec64_sub(delta, old_delta);
1884 if (abs(delta_delta.tv_sec) >= 2) {
1885 /*
1886 * if delta_delta is too large, assume time correction
1887 * has occurred and set old_delta to the current delta.
1888 */
1889 old_delta = delta;
1890 } else {
1891 /* Otherwise try to adjust old_system to compensate */
1892 timekeeping_suspend_time =
1893 timespec64_add(timekeeping_suspend_time, delta_delta);
1894 }
1895 }
1896
1897 timekeeping_update(tk, TK_MIRROR);
1898 halt_fast_timekeeper(tk);
1899 write_seqcount_end(&tk_core.seq);
1900 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1901
1902 tick_suspend();
1903 clocksource_suspend();
1904 clockevents_suspend();
1905
1906 return 0;
1907}
1908
1909/* sysfs resume/suspend bits for timekeeping */
1910static struct syscore_ops timekeeping_syscore_ops = {
1911 .resume = timekeeping_resume,
1912 .suspend = timekeeping_suspend,
1913};
1914
1915static int __init timekeeping_init_ops(void)
1916{
1917 register_syscore_ops(&timekeeping_syscore_ops);
1918 return 0;
1919}
1920device_initcall(timekeeping_init_ops);
1921
1922/*
1923 * Apply a multiplier adjustment to the timekeeper
1924 */
1925static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1926 s64 offset,
1927 s32 mult_adj)
1928{
1929 s64 interval = tk->cycle_interval;
1930
1931 if (mult_adj == 0) {
1932 return;
1933 } else if (mult_adj == -1) {
1934 interval = -interval;
1935 offset = -offset;
1936 } else if (mult_adj != 1) {
1937 interval *= mult_adj;
1938 offset *= mult_adj;
1939 }
1940
1941 /*
1942 * So the following can be confusing.
1943 *
1944 * To keep things simple, lets assume mult_adj == 1 for now.
1945 *
1946 * When mult_adj != 1, remember that the interval and offset values
1947 * have been appropriately scaled so the math is the same.
1948 *
1949 * The basic idea here is that we're increasing the multiplier
1950 * by one, this causes the xtime_interval to be incremented by
1951 * one cycle_interval. This is because:
1952 * xtime_interval = cycle_interval * mult
1953 * So if mult is being incremented by one:
1954 * xtime_interval = cycle_interval * (mult + 1)
1955 * Its the same as:
1956 * xtime_interval = (cycle_interval * mult) + cycle_interval
1957 * Which can be shortened to:
1958 * xtime_interval += cycle_interval
1959 *
1960 * So offset stores the non-accumulated cycles. Thus the current
1961 * time (in shifted nanoseconds) is:
1962 * now = (offset * adj) + xtime_nsec
1963 * Now, even though we're adjusting the clock frequency, we have
1964 * to keep time consistent. In other words, we can't jump back
1965 * in time, and we also want to avoid jumping forward in time.
1966 *
1967 * So given the same offset value, we need the time to be the same
1968 * both before and after the freq adjustment.
1969 * now = (offset * adj_1) + xtime_nsec_1
1970 * now = (offset * adj_2) + xtime_nsec_2
1971 * So:
1972 * (offset * adj_1) + xtime_nsec_1 =
1973 * (offset * adj_2) + xtime_nsec_2
1974 * And we know:
1975 * adj_2 = adj_1 + 1
1976 * So:
1977 * (offset * adj_1) + xtime_nsec_1 =
1978 * (offset * (adj_1+1)) + xtime_nsec_2
1979 * (offset * adj_1) + xtime_nsec_1 =
1980 * (offset * adj_1) + offset + xtime_nsec_2
1981 * Canceling the sides:
1982 * xtime_nsec_1 = offset + xtime_nsec_2
1983 * Which gives us:
1984 * xtime_nsec_2 = xtime_nsec_1 - offset
1985 * Which simplifies to:
1986 * xtime_nsec -= offset
1987 */
1988 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1989 /* NTP adjustment caused clocksource mult overflow */
1990 WARN_ON_ONCE(1);
1991 return;
1992 }
1993
1994 tk->tkr_mono.mult += mult_adj;
1995 tk->xtime_interval += interval;
1996 tk->tkr_mono.xtime_nsec -= offset;
1997}
1998
1999/*
2000 * Adjust the timekeeper's multiplier to the correct frequency
2001 * and also to reduce the accumulated error value.
2002 */
2003static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2004{
2005 u32 mult;
2006
2007 /*
2008 * Determine the multiplier from the current NTP tick length.
2009 * Avoid expensive division when the tick length doesn't change.
2010 */
2011 if (likely(tk->ntp_tick == ntp_tick_length())) {
2012 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2013 } else {
2014 tk->ntp_tick = ntp_tick_length();
2015 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2016 tk->xtime_remainder, tk->cycle_interval);
2017 }
2018
2019 /*
2020 * If the clock is behind the NTP time, increase the multiplier by 1
2021 * to catch up with it. If it's ahead and there was a remainder in the
2022 * tick division, the clock will slow down. Otherwise it will stay
2023 * ahead until the tick length changes to a non-divisible value.
2024 */
2025 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2026 mult += tk->ntp_err_mult;
2027
2028 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2029
2030 if (unlikely(tk->tkr_mono.clock->maxadj &&
2031 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2032 > tk->tkr_mono.clock->maxadj))) {
2033 printk_once(KERN_WARNING
2034 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2035 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2036 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2037 }
2038
2039 /*
2040 * It may be possible that when we entered this function, xtime_nsec
2041 * was very small. Further, if we're slightly speeding the clocksource
2042 * in the code above, its possible the required corrective factor to
2043 * xtime_nsec could cause it to underflow.
2044 *
2045 * Now, since we have already accumulated the second and the NTP
2046 * subsystem has been notified via second_overflow(), we need to skip
2047 * the next update.
2048 */
2049 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2050 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2051 tk->tkr_mono.shift;
2052 tk->xtime_sec--;
2053 tk->skip_second_overflow = 1;
2054 }
2055}
2056
2057/*
2058 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2059 *
2060 * Helper function that accumulates the nsecs greater than a second
2061 * from the xtime_nsec field to the xtime_secs field.
2062 * It also calls into the NTP code to handle leapsecond processing.
2063 */
2064static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2065{
2066 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2067 unsigned int clock_set = 0;
2068
2069 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2070 int leap;
2071
2072 tk->tkr_mono.xtime_nsec -= nsecps;
2073 tk->xtime_sec++;
2074
2075 /*
2076 * Skip NTP update if this second was accumulated before,
2077 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2078 */
2079 if (unlikely(tk->skip_second_overflow)) {
2080 tk->skip_second_overflow = 0;
2081 continue;
2082 }
2083
2084 /* Figure out if its a leap sec and apply if needed */
2085 leap = second_overflow(tk->xtime_sec);
2086 if (unlikely(leap)) {
2087 struct timespec64 ts;
2088
2089 tk->xtime_sec += leap;
2090
2091 ts.tv_sec = leap;
2092 ts.tv_nsec = 0;
2093 tk_set_wall_to_mono(tk,
2094 timespec64_sub(tk->wall_to_monotonic, ts));
2095
2096 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2097
2098 clock_set = TK_CLOCK_WAS_SET;
2099 }
2100 }
2101 return clock_set;
2102}
2103
2104/*
2105 * logarithmic_accumulation - shifted accumulation of cycles
2106 *
2107 * This functions accumulates a shifted interval of cycles into
2108 * a shifted interval nanoseconds. Allows for O(log) accumulation
2109 * loop.
2110 *
2111 * Returns the unconsumed cycles.
2112 */
2113static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2114 u32 shift, unsigned int *clock_set)
2115{
2116 u64 interval = tk->cycle_interval << shift;
2117 u64 snsec_per_sec;
2118
2119 /* If the offset is smaller than a shifted interval, do nothing */
2120 if (offset < interval)
2121 return offset;
2122
2123 /* Accumulate one shifted interval */
2124 offset -= interval;
2125 tk->tkr_mono.cycle_last += interval;
2126 tk->tkr_raw.cycle_last += interval;
2127
2128 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2129 *clock_set |= accumulate_nsecs_to_secs(tk);
2130
2131 /* Accumulate raw time */
2132 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2133 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2134 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2135 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2136 tk->raw_sec++;
2137 }
2138
2139 /* Accumulate error between NTP and clock interval */
2140 tk->ntp_error += tk->ntp_tick << shift;
2141 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2142 (tk->ntp_error_shift + shift);
2143
2144 return offset;
2145}
2146
2147/*
2148 * timekeeping_advance - Updates the timekeeper to the current time and
2149 * current NTP tick length
2150 */
2151static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2152{
2153 struct timekeeper *real_tk = &tk_core.timekeeper;
2154 struct timekeeper *tk = &shadow_timekeeper;
2155 u64 offset;
2156 int shift = 0, maxshift;
2157 unsigned int clock_set = 0;
2158 unsigned long flags;
2159
2160 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2161
2162 /* Make sure we're fully resumed: */
2163 if (unlikely(timekeeping_suspended))
2164 goto out;
2165
2166 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2167 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2168
2169 /* Check if there's really nothing to do */
2170 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2171 goto out;
2172
2173 /* Do some additional sanity checking */
2174 timekeeping_check_update(tk, offset);
2175
2176 /*
2177 * With NO_HZ we may have to accumulate many cycle_intervals
2178 * (think "ticks") worth of time at once. To do this efficiently,
2179 * we calculate the largest doubling multiple of cycle_intervals
2180 * that is smaller than the offset. We then accumulate that
2181 * chunk in one go, and then try to consume the next smaller
2182 * doubled multiple.
2183 */
2184 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2185 shift = max(0, shift);
2186 /* Bound shift to one less than what overflows tick_length */
2187 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2188 shift = min(shift, maxshift);
2189 while (offset >= tk->cycle_interval) {
2190 offset = logarithmic_accumulation(tk, offset, shift,
2191 &clock_set);
2192 if (offset < tk->cycle_interval<<shift)
2193 shift--;
2194 }
2195
2196 /* Adjust the multiplier to correct NTP error */
2197 timekeeping_adjust(tk, offset);
2198
2199 /*
2200 * Finally, make sure that after the rounding
2201 * xtime_nsec isn't larger than NSEC_PER_SEC
2202 */
2203 clock_set |= accumulate_nsecs_to_secs(tk);
2204
2205 write_seqcount_begin(&tk_core.seq);
2206 /*
2207 * Update the real timekeeper.
2208 *
2209 * We could avoid this memcpy by switching pointers, but that
2210 * requires changes to all other timekeeper usage sites as
2211 * well, i.e. move the timekeeper pointer getter into the
2212 * spinlocked/seqcount protected sections. And we trade this
2213 * memcpy under the tk_core.seq against one before we start
2214 * updating.
2215 */
2216 timekeeping_update(tk, clock_set);
2217 memcpy(real_tk, tk, sizeof(*tk));
2218 /* The memcpy must come last. Do not put anything here! */
2219 write_seqcount_end(&tk_core.seq);
2220out:
2221 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2222
2223 return !!clock_set;
2224}
2225
2226/**
2227 * update_wall_time - Uses the current clocksource to increment the wall time
2228 *
2229 */
2230void update_wall_time(void)
2231{
2232 if (timekeeping_advance(TK_ADV_TICK))
2233 clock_was_set_delayed();
2234}
2235
2236/**
2237 * getboottime64 - Return the real time of system boot.
2238 * @ts: pointer to the timespec64 to be set
2239 *
2240 * Returns the wall-time of boot in a timespec64.
2241 *
2242 * This is based on the wall_to_monotonic offset and the total suspend
2243 * time. Calls to settimeofday will affect the value returned (which
2244 * basically means that however wrong your real time clock is at boot time,
2245 * you get the right time here).
2246 */
2247void getboottime64(struct timespec64 *ts)
2248{
2249 struct timekeeper *tk = &tk_core.timekeeper;
2250 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2251
2252 *ts = ktime_to_timespec64(t);
2253}
2254EXPORT_SYMBOL_GPL(getboottime64);
2255
2256void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2257{
2258 struct timekeeper *tk = &tk_core.timekeeper;
2259 unsigned int seq;
2260
2261 do {
2262 seq = read_seqcount_begin(&tk_core.seq);
2263
2264 *ts = tk_xtime(tk);
2265 } while (read_seqcount_retry(&tk_core.seq, seq));
2266}
2267EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2268
2269void ktime_get_coarse_ts64(struct timespec64 *ts)
2270{
2271 struct timekeeper *tk = &tk_core.timekeeper;
2272 struct timespec64 now, mono;
2273 unsigned int seq;
2274
2275 do {
2276 seq = read_seqcount_begin(&tk_core.seq);
2277
2278 now = tk_xtime(tk);
2279 mono = tk->wall_to_monotonic;
2280 } while (read_seqcount_retry(&tk_core.seq, seq));
2281
2282 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2283 now.tv_nsec + mono.tv_nsec);
2284}
2285EXPORT_SYMBOL(ktime_get_coarse_ts64);
2286
2287/*
2288 * Must hold jiffies_lock
2289 */
2290void do_timer(unsigned long ticks)
2291{
2292 jiffies_64 += ticks;
2293 calc_global_load();
2294}
2295
2296/**
2297 * ktime_get_update_offsets_now - hrtimer helper
2298 * @cwsseq: pointer to check and store the clock was set sequence number
2299 * @offs_real: pointer to storage for monotonic -> realtime offset
2300 * @offs_boot: pointer to storage for monotonic -> boottime offset
2301 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2302 *
2303 * Returns current monotonic time and updates the offsets if the
2304 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2305 * different.
2306 *
2307 * Called from hrtimer_interrupt() or retrigger_next_event()
2308 */
2309ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2310 ktime_t *offs_boot, ktime_t *offs_tai)
2311{
2312 struct timekeeper *tk = &tk_core.timekeeper;
2313 unsigned int seq;
2314 ktime_t base;
2315 u64 nsecs;
2316
2317 do {
2318 seq = read_seqcount_begin(&tk_core.seq);
2319
2320 base = tk->tkr_mono.base;
2321 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2322 base = ktime_add_ns(base, nsecs);
2323
2324 if (*cwsseq != tk->clock_was_set_seq) {
2325 *cwsseq = tk->clock_was_set_seq;
2326 *offs_real = tk->offs_real;
2327 *offs_boot = tk->offs_boot;
2328 *offs_tai = tk->offs_tai;
2329 }
2330
2331 /* Handle leapsecond insertion adjustments */
2332 if (unlikely(base >= tk->next_leap_ktime))
2333 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2334
2335 } while (read_seqcount_retry(&tk_core.seq, seq));
2336
2337 return base;
2338}
2339
2340/*
2341 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2342 */
2343static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2344{
2345 if (txc->modes & ADJ_ADJTIME) {
2346 /* singleshot must not be used with any other mode bits */
2347 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2348 return -EINVAL;
2349 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2350 !capable(CAP_SYS_TIME))
2351 return -EPERM;
2352 } else {
2353 /* In order to modify anything, you gotta be super-user! */
2354 if (txc->modes && !capable(CAP_SYS_TIME))
2355 return -EPERM;
2356 /*
2357 * if the quartz is off by more than 10% then
2358 * something is VERY wrong!
2359 */
2360 if (txc->modes & ADJ_TICK &&
2361 (txc->tick < 900000/USER_HZ ||
2362 txc->tick > 1100000/USER_HZ))
2363 return -EINVAL;
2364 }
2365
2366 if (txc->modes & ADJ_SETOFFSET) {
2367 /* In order to inject time, you gotta be super-user! */
2368 if (!capable(CAP_SYS_TIME))
2369 return -EPERM;
2370
2371 /*
2372 * Validate if a timespec/timeval used to inject a time
2373 * offset is valid. Offsets can be positive or negative, so
2374 * we don't check tv_sec. The value of the timeval/timespec
2375 * is the sum of its fields,but *NOTE*:
2376 * The field tv_usec/tv_nsec must always be non-negative and
2377 * we can't have more nanoseconds/microseconds than a second.
2378 */
2379 if (txc->time.tv_usec < 0)
2380 return -EINVAL;
2381
2382 if (txc->modes & ADJ_NANO) {
2383 if (txc->time.tv_usec >= NSEC_PER_SEC)
2384 return -EINVAL;
2385 } else {
2386 if (txc->time.tv_usec >= USEC_PER_SEC)
2387 return -EINVAL;
2388 }
2389 }
2390
2391 /*
2392 * Check for potential multiplication overflows that can
2393 * only happen on 64-bit systems:
2394 */
2395 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2396 if (LLONG_MIN / PPM_SCALE > txc->freq)
2397 return -EINVAL;
2398 if (LLONG_MAX / PPM_SCALE < txc->freq)
2399 return -EINVAL;
2400 }
2401
2402 return 0;
2403}
2404
2405/**
2406 * random_get_entropy_fallback - Returns the raw clock source value,
2407 * used by random.c for platforms with no valid random_get_entropy().
2408 */
2409unsigned long random_get_entropy_fallback(void)
2410{
2411 struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2412 struct clocksource *clock = READ_ONCE(tkr->clock);
2413
2414 if (unlikely(timekeeping_suspended || !clock))
2415 return 0;
2416 return clock->read(clock);
2417}
2418EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2419
2420/**
2421 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2422 */
2423int do_adjtimex(struct __kernel_timex *txc)
2424{
2425 struct timekeeper *tk = &tk_core.timekeeper;
2426 struct audit_ntp_data ad;
2427 bool clock_set = false;
2428 struct timespec64 ts;
2429 unsigned long flags;
2430 s32 orig_tai, tai;
2431 int ret;
2432
2433 /* Validate the data before disabling interrupts */
2434 ret = timekeeping_validate_timex(txc);
2435 if (ret)
2436 return ret;
2437 add_device_randomness(txc, sizeof(*txc));
2438
2439 if (txc->modes & ADJ_SETOFFSET) {
2440 struct timespec64 delta;
2441 delta.tv_sec = txc->time.tv_sec;
2442 delta.tv_nsec = txc->time.tv_usec;
2443 if (!(txc->modes & ADJ_NANO))
2444 delta.tv_nsec *= 1000;
2445 ret = timekeeping_inject_offset(&delta);
2446 if (ret)
2447 return ret;
2448
2449 audit_tk_injoffset(delta);
2450 }
2451
2452 audit_ntp_init(&ad);
2453
2454 ktime_get_real_ts64(&ts);
2455 add_device_randomness(&ts, sizeof(ts));
2456
2457 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2458 write_seqcount_begin(&tk_core.seq);
2459
2460 orig_tai = tai = tk->tai_offset;
2461 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2462
2463 if (tai != orig_tai) {
2464 __timekeeping_set_tai_offset(tk, tai);
2465 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2466 clock_set = true;
2467 }
2468 tk_update_leap_state(tk);
2469
2470 write_seqcount_end(&tk_core.seq);
2471 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2472
2473 audit_ntp_log(&ad);
2474
2475 /* Update the multiplier immediately if frequency was set directly */
2476 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2477 clock_set |= timekeeping_advance(TK_ADV_FREQ);
2478
2479 if (clock_set)
2480 clock_was_set(CLOCK_REALTIME);
2481
2482 ntp_notify_cmos_timer();
2483
2484 return ret;
2485}
2486
2487#ifdef CONFIG_NTP_PPS
2488/**
2489 * hardpps() - Accessor function to NTP __hardpps function
2490 */
2491void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2492{
2493 unsigned long flags;
2494
2495 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2496 write_seqcount_begin(&tk_core.seq);
2497
2498 __hardpps(phase_ts, raw_ts);
2499
2500 write_seqcount_end(&tk_core.seq);
2501 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2502}
2503EXPORT_SYMBOL(hardpps);
2504#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}