projects/linux·blob v6.3 · refs

2503 lines · 71.1 KiB

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 (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_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 (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_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 * cycle_between - true if test occurs chronologically between before and after
1184 */
1185static bool cycle_between(u64 before, u64 test, u64 after)
1186{
1187 if (test > before && test < after)
1188 return true;
1189 if (test < before && before > after)
1190 return true;
1191 return false;
1192}
1193
1194/**
1195 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1196 * @get_time_fn: Callback to get simultaneous device time and
1197 * system counter from the device driver
1198 * @ctx: Context passed to get_time_fn()
1199 * @history_begin: Historical reference point used to interpolate system
1200 * time when counter provided by the driver is before the current interval
1201 * @xtstamp: Receives simultaneously captured system and device time
1202 *
1203 * Reads a timestamp from a device and correlates it to system time
1204 */
1205int get_device_system_crosststamp(int (*get_time_fn)
1206 (ktime_t *device_time,
1207 struct system_counterval_t *sys_counterval,
1208 void *ctx),
1209 void *ctx,
1210 struct system_time_snapshot *history_begin,
1211 struct system_device_crosststamp *xtstamp)
1212{
1213 struct system_counterval_t system_counterval;
1214 struct timekeeper *tk = &tk_core.timekeeper;
1215 u64 cycles, now, interval_start;
1216 unsigned int clock_was_set_seq = 0;
1217 ktime_t base_real, base_raw;
1218 u64 nsec_real, nsec_raw;
1219 u8 cs_was_changed_seq;
1220 unsigned int seq;
1221 bool do_interp;
1222 int ret;
1223
1224 do {
1225 seq = read_seqcount_begin(&tk_core.seq);
1226 /*
1227 * Try to synchronously capture device time and a system
1228 * counter value calling back into the device driver
1229 */
1230 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1231 if (ret)
1232 return ret;
1233
1234 /*
1235 * Verify that the clocksource associated with the captured
1236 * system counter value is the same as the currently installed
1237 * timekeeper clocksource
1238 */
1239 if (tk->tkr_mono.clock != system_counterval.cs)
1240 return -ENODEV;
1241 cycles = system_counterval.cycles;
1242
1243 /*
1244 * Check whether the system counter value provided by the
1245 * device driver is on the current timekeeping interval.
1246 */
1247 now = tk_clock_read(&tk->tkr_mono);
1248 interval_start = tk->tkr_mono.cycle_last;
1249 if (!cycle_between(interval_start, cycles, now)) {
1250 clock_was_set_seq = tk->clock_was_set_seq;
1251 cs_was_changed_seq = tk->cs_was_changed_seq;
1252 cycles = interval_start;
1253 do_interp = true;
1254 } else {
1255 do_interp = false;
1256 }
1257
1258 base_real = ktime_add(tk->tkr_mono.base,
1259 tk_core.timekeeper.offs_real);
1260 base_raw = tk->tkr_raw.base;
1261
1262 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1263 system_counterval.cycles);
1264 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1265 system_counterval.cycles);
1266 } while (read_seqcount_retry(&tk_core.seq, seq));
1267
1268 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1269 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1270
1271 /*
1272 * Interpolate if necessary, adjusting back from the start of the
1273 * current interval
1274 */
1275 if (do_interp) {
1276 u64 partial_history_cycles, total_history_cycles;
1277 bool discontinuity;
1278
1279 /*
1280 * Check that the counter value occurs after the provided
1281 * history reference and that the history doesn't cross a
1282 * clocksource change
1283 */
1284 if (!history_begin ||
1285 !cycle_between(history_begin->cycles,
1286 system_counterval.cycles, cycles) ||
1287 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1288 return -EINVAL;
1289 partial_history_cycles = cycles - system_counterval.cycles;
1290 total_history_cycles = cycles - history_begin->cycles;
1291 discontinuity =
1292 history_begin->clock_was_set_seq != clock_was_set_seq;
1293
1294 ret = adjust_historical_crosststamp(history_begin,
1295 partial_history_cycles,
1296 total_history_cycles,
1297 discontinuity, xtstamp);
1298 if (ret)
1299 return ret;
1300 }
1301
1302 return 0;
1303}
1304EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1305
1306/**
1307 * do_settimeofday64 - Sets the time of day.
1308 * @ts: pointer to the timespec64 variable containing the new time
1309 *
1310 * Sets the time of day to the new time and update NTP and notify hrtimers
1311 */
1312int do_settimeofday64(const struct timespec64 *ts)
1313{
1314 struct timekeeper *tk = &tk_core.timekeeper;
1315 struct timespec64 ts_delta, xt;
1316 unsigned long flags;
1317 int ret = 0;
1318
1319 if (!timespec64_valid_settod(ts))
1320 return -EINVAL;
1321
1322 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1323 write_seqcount_begin(&tk_core.seq);
1324
1325 timekeeping_forward_now(tk);
1326
1327 xt = tk_xtime(tk);
1328 ts_delta = timespec64_sub(*ts, xt);
1329
1330 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1331 ret = -EINVAL;
1332 goto out;
1333 }
1334
1335 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1336
1337 tk_set_xtime(tk, ts);
1338out:
1339 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1340
1341 write_seqcount_end(&tk_core.seq);
1342 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1343
1344 /* Signal hrtimers about time change */
1345 clock_was_set(CLOCK_SET_WALL);
1346
1347 if (!ret) {
1348 audit_tk_injoffset(ts_delta);
1349 add_device_randomness(ts, sizeof(*ts));
1350 }
1351
1352 return ret;
1353}
1354EXPORT_SYMBOL(do_settimeofday64);
1355
1356/**
1357 * timekeeping_inject_offset - Adds or subtracts from the current time.
1358 * @ts: Pointer to the timespec variable containing the offset
1359 *
1360 * Adds or subtracts an offset value from the current time.
1361 */
1362static int timekeeping_inject_offset(const struct timespec64 *ts)
1363{
1364 struct timekeeper *tk = &tk_core.timekeeper;
1365 unsigned long flags;
1366 struct timespec64 tmp;
1367 int ret = 0;
1368
1369 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1370 return -EINVAL;
1371
1372 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1373 write_seqcount_begin(&tk_core.seq);
1374
1375 timekeeping_forward_now(tk);
1376
1377 /* Make sure the proposed value is valid */
1378 tmp = timespec64_add(tk_xtime(tk), *ts);
1379 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1380 !timespec64_valid_settod(&tmp)) {
1381 ret = -EINVAL;
1382 goto error;
1383 }
1384
1385 tk_xtime_add(tk, ts);
1386 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1387
1388error: /* even if we error out, we forwarded the time, so call update */
1389 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1390
1391 write_seqcount_end(&tk_core.seq);
1392 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1393
1394 /* Signal hrtimers about time change */
1395 clock_was_set(CLOCK_SET_WALL);
1396
1397 return ret;
1398}
1399
1400/*
1401 * Indicates if there is an offset between the system clock and the hardware
1402 * clock/persistent clock/rtc.
1403 */
1404int persistent_clock_is_local;
1405
1406/*
1407 * Adjust the time obtained from the CMOS to be UTC time instead of
1408 * local time.
1409 *
1410 * This is ugly, but preferable to the alternatives. Otherwise we
1411 * would either need to write a program to do it in /etc/rc (and risk
1412 * confusion if the program gets run more than once; it would also be
1413 * hard to make the program warp the clock precisely n hours) or
1414 * compile in the timezone information into the kernel. Bad, bad....
1415 *
1416 * - TYT, 1992-01-01
1417 *
1418 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1419 * as real UNIX machines always do it. This avoids all headaches about
1420 * daylight saving times and warping kernel clocks.
1421 */
1422void timekeeping_warp_clock(void)
1423{
1424 if (sys_tz.tz_minuteswest != 0) {
1425 struct timespec64 adjust;
1426
1427 persistent_clock_is_local = 1;
1428 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1429 adjust.tv_nsec = 0;
1430 timekeeping_inject_offset(&adjust);
1431 }
1432}
1433
1434/*
1435 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1436 */
1437static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1438{
1439 tk->tai_offset = tai_offset;
1440 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1441}
1442
1443/*
1444 * change_clocksource - Swaps clocksources if a new one is available
1445 *
1446 * Accumulates current time interval and initializes new clocksource
1447 */
1448static int change_clocksource(void *data)
1449{
1450 struct timekeeper *tk = &tk_core.timekeeper;
1451 struct clocksource *new, *old = NULL;
1452 unsigned long flags;
1453 bool change = false;
1454
1455 new = (struct clocksource *) data;
1456
1457 /*
1458 * If the cs is in module, get a module reference. Succeeds
1459 * for built-in code (owner == NULL) as well.
1460 */
1461 if (try_module_get(new->owner)) {
1462 if (!new->enable || new->enable(new) == 0)
1463 change = true;
1464 else
1465 module_put(new->owner);
1466 }
1467
1468 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1469 write_seqcount_begin(&tk_core.seq);
1470
1471 timekeeping_forward_now(tk);
1472
1473 if (change) {
1474 old = tk->tkr_mono.clock;
1475 tk_setup_internals(tk, new);
1476 }
1477
1478 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1479
1480 write_seqcount_end(&tk_core.seq);
1481 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1482
1483 if (old) {
1484 if (old->disable)
1485 old->disable(old);
1486
1487 module_put(old->owner);
1488 }
1489
1490 return 0;
1491}
1492
1493/**
1494 * timekeeping_notify - Install a new clock source
1495 * @clock: pointer to the clock source
1496 *
1497 * This function is called from clocksource.c after a new, better clock
1498 * source has been registered. The caller holds the clocksource_mutex.
1499 */
1500int timekeeping_notify(struct clocksource *clock)
1501{
1502 struct timekeeper *tk = &tk_core.timekeeper;
1503
1504 if (tk->tkr_mono.clock == clock)
1505 return 0;
1506 stop_machine(change_clocksource, clock, NULL);
1507 tick_clock_notify();
1508 return tk->tkr_mono.clock == clock ? 0 : -1;
1509}
1510
1511/**
1512 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1513 * @ts: pointer to the timespec64 to be set
1514 *
1515 * Returns the raw monotonic time (completely un-modified by ntp)
1516 */
1517void ktime_get_raw_ts64(struct timespec64 *ts)
1518{
1519 struct timekeeper *tk = &tk_core.timekeeper;
1520 unsigned int seq;
1521 u64 nsecs;
1522
1523 do {
1524 seq = read_seqcount_begin(&tk_core.seq);
1525 ts->tv_sec = tk->raw_sec;
1526 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1527
1528 } while (read_seqcount_retry(&tk_core.seq, seq));
1529
1530 ts->tv_nsec = 0;
1531 timespec64_add_ns(ts, nsecs);
1532}
1533EXPORT_SYMBOL(ktime_get_raw_ts64);
1534
1535
1536/**
1537 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1538 */
1539int timekeeping_valid_for_hres(void)
1540{
1541 struct timekeeper *tk = &tk_core.timekeeper;
1542 unsigned int seq;
1543 int ret;
1544
1545 do {
1546 seq = read_seqcount_begin(&tk_core.seq);
1547
1548 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1549
1550 } while (read_seqcount_retry(&tk_core.seq, seq));
1551
1552 return ret;
1553}
1554
1555/**
1556 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1557 */
1558u64 timekeeping_max_deferment(void)
1559{
1560 struct timekeeper *tk = &tk_core.timekeeper;
1561 unsigned int seq;
1562 u64 ret;
1563
1564 do {
1565 seq = read_seqcount_begin(&tk_core.seq);
1566
1567 ret = tk->tkr_mono.clock->max_idle_ns;
1568
1569 } while (read_seqcount_retry(&tk_core.seq, seq));
1570
1571 return ret;
1572}
1573
1574/**
1575 * read_persistent_clock64 - Return time from the persistent clock.
1576 * @ts: Pointer to the storage for the readout value
1577 *
1578 * Weak dummy function for arches that do not yet support it.
1579 * Reads the time from the battery backed persistent clock.
1580 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1581 *
1582 * XXX - Do be sure to remove it once all arches implement it.
1583 */
1584void __weak read_persistent_clock64(struct timespec64 *ts)
1585{
1586 ts->tv_sec = 0;
1587 ts->tv_nsec = 0;
1588}
1589
1590/**
1591 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1592 * from the boot.
1593 * @wall_time: current time as returned by persistent clock
1594 * @boot_offset: offset that is defined as wall_time - boot_time
1595 *
1596 * Weak dummy function for arches that do not yet support it.
1597 *
1598 * The default function calculates offset based on the current value of
1599 * local_clock(). This way architectures that support sched_clock() but don't
1600 * support dedicated boot time clock will provide the best estimate of the
1601 * boot time.
1602 */
1603void __weak __init
1604read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1605 struct timespec64 *boot_offset)
1606{
1607 read_persistent_clock64(wall_time);
1608 *boot_offset = ns_to_timespec64(local_clock());
1609}
1610
1611/*
1612 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1613 *
1614 * The flag starts of false and is only set when a suspend reaches
1615 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1616 * timekeeper clocksource is not stopping across suspend and has been
1617 * used to update sleep time. If the timekeeper clocksource has stopped
1618 * then the flag stays true and is used by the RTC resume code to decide
1619 * whether sleeptime must be injected and if so the flag gets false then.
1620 *
1621 * If a suspend fails before reaching timekeeping_resume() then the flag
1622 * stays false and prevents erroneous sleeptime injection.
1623 */
1624static bool suspend_timing_needed;
1625
1626/* Flag for if there is a persistent clock on this platform */
1627static bool persistent_clock_exists;
1628
1629/*
1630 * timekeeping_init - Initializes the clocksource and common timekeeping values
1631 */
1632void __init timekeeping_init(void)
1633{
1634 struct timespec64 wall_time, boot_offset, wall_to_mono;
1635 struct timekeeper *tk = &tk_core.timekeeper;
1636 struct clocksource *clock;
1637 unsigned long flags;
1638
1639 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1640 if (timespec64_valid_settod(&wall_time) &&
1641 timespec64_to_ns(&wall_time) > 0) {
1642 persistent_clock_exists = true;
1643 } else if (timespec64_to_ns(&wall_time) != 0) {
1644 pr_warn("Persistent clock returned invalid value");
1645 wall_time = (struct timespec64){0};
1646 }
1647
1648 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1649 boot_offset = (struct timespec64){0};
1650
1651 /*
1652 * We want set wall_to_mono, so the following is true:
1653 * wall time + wall_to_mono = boot time
1654 */
1655 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1656
1657 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1658 write_seqcount_begin(&tk_core.seq);
1659 ntp_init();
1660
1661 clock = clocksource_default_clock();
1662 if (clock->enable)
1663 clock->enable(clock);
1664 tk_setup_internals(tk, clock);
1665
1666 tk_set_xtime(tk, &wall_time);
1667 tk->raw_sec = 0;
1668
1669 tk_set_wall_to_mono(tk, wall_to_mono);
1670
1671 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1672
1673 write_seqcount_end(&tk_core.seq);
1674 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1675}
1676
1677/* time in seconds when suspend began for persistent clock */
1678static struct timespec64 timekeeping_suspend_time;
1679
1680/**
1681 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1682 * @tk: Pointer to the timekeeper to be updated
1683 * @delta: Pointer to the delta value in timespec64 format
1684 *
1685 * Takes a timespec offset measuring a suspend interval and properly
1686 * adds the sleep offset to the timekeeping variables.
1687 */
1688static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1689 const struct timespec64 *delta)
1690{
1691 if (!timespec64_valid_strict(delta)) {
1692 printk_deferred(KERN_WARNING
1693 "__timekeeping_inject_sleeptime: Invalid "
1694 "sleep delta value!\n");
1695 return;
1696 }
1697 tk_xtime_add(tk, delta);
1698 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1699 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1700 tk_debug_account_sleep_time(delta);
1701}
1702
1703#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1704/*
1705 * We have three kinds of time sources to use for sleep time
1706 * injection, the preference order is:
1707 * 1) non-stop clocksource
1708 * 2) persistent clock (ie: RTC accessible when irqs are off)
1709 * 3) RTC
1710 *
1711 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1712 * If system has neither 1) nor 2), 3) will be used finally.
1713 *
1714 *
1715 * If timekeeping has injected sleeptime via either 1) or 2),
1716 * 3) becomes needless, so in this case we don't need to call
1717 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1718 * means.
1719 */
1720bool timekeeping_rtc_skipresume(void)
1721{
1722 return !suspend_timing_needed;
1723}
1724
1725/*
1726 * 1) can be determined whether to use or not only when doing
1727 * timekeeping_resume() which is invoked after rtc_suspend(),
1728 * so we can't skip rtc_suspend() surely if system has 1).
1729 *
1730 * But if system has 2), 2) will definitely be used, so in this
1731 * case we don't need to call rtc_suspend(), and this is what
1732 * timekeeping_rtc_skipsuspend() means.
1733 */
1734bool timekeeping_rtc_skipsuspend(void)
1735{
1736 return persistent_clock_exists;
1737}
1738
1739/**
1740 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1741 * @delta: pointer to a timespec64 delta value
1742 *
1743 * This hook is for architectures that cannot support read_persistent_clock64
1744 * because their RTC/persistent clock is only accessible when irqs are enabled.
1745 * and also don't have an effective nonstop clocksource.
1746 *
1747 * This function should only be called by rtc_resume(), and allows
1748 * a suspend offset to be injected into the timekeeping values.
1749 */
1750void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1751{
1752 struct timekeeper *tk = &tk_core.timekeeper;
1753 unsigned long flags;
1754
1755 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1756 write_seqcount_begin(&tk_core.seq);
1757
1758 suspend_timing_needed = false;
1759
1760 timekeeping_forward_now(tk);
1761
1762 __timekeeping_inject_sleeptime(tk, delta);
1763
1764 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1765
1766 write_seqcount_end(&tk_core.seq);
1767 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1768
1769 /* Signal hrtimers about time change */
1770 clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1771}
1772#endif
1773
1774/**
1775 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1776 */
1777void timekeeping_resume(void)
1778{
1779 struct timekeeper *tk = &tk_core.timekeeper;
1780 struct clocksource *clock = tk->tkr_mono.clock;
1781 unsigned long flags;
1782 struct timespec64 ts_new, ts_delta;
1783 u64 cycle_now, nsec;
1784 bool inject_sleeptime = false;
1785
1786 read_persistent_clock64(&ts_new);
1787
1788 clockevents_resume();
1789 clocksource_resume();
1790
1791 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1792 write_seqcount_begin(&tk_core.seq);
1793
1794 /*
1795 * After system resumes, we need to calculate the suspended time and
1796 * compensate it for the OS time. There are 3 sources that could be
1797 * used: Nonstop clocksource during suspend, persistent clock and rtc
1798 * device.
1799 *
1800 * One specific platform may have 1 or 2 or all of them, and the
1801 * preference will be:
1802 * suspend-nonstop clocksource -> persistent clock -> rtc
1803 * The less preferred source will only be tried if there is no better
1804 * usable source. The rtc part is handled separately in rtc core code.
1805 */
1806 cycle_now = tk_clock_read(&tk->tkr_mono);
1807 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1808 if (nsec > 0) {
1809 ts_delta = ns_to_timespec64(nsec);
1810 inject_sleeptime = true;
1811 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1812 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1813 inject_sleeptime = true;
1814 }
1815
1816 if (inject_sleeptime) {
1817 suspend_timing_needed = false;
1818 __timekeeping_inject_sleeptime(tk, &ts_delta);
1819 }
1820
1821 /* Re-base the last cycle value */
1822 tk->tkr_mono.cycle_last = cycle_now;
1823 tk->tkr_raw.cycle_last = cycle_now;
1824
1825 tk->ntp_error = 0;
1826 timekeeping_suspended = 0;
1827 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1828 write_seqcount_end(&tk_core.seq);
1829 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1830
1831 touch_softlockup_watchdog();
1832
1833 /* Resume the clockevent device(s) and hrtimers */
1834 tick_resume();
1835 /* Notify timerfd as resume is equivalent to clock_was_set() */
1836 timerfd_resume();
1837}
1838
1839int timekeeping_suspend(void)
1840{
1841 struct timekeeper *tk = &tk_core.timekeeper;
1842 unsigned long flags;
1843 struct timespec64 delta, delta_delta;
1844 static struct timespec64 old_delta;
1845 struct clocksource *curr_clock;
1846 u64 cycle_now;
1847
1848 read_persistent_clock64(&timekeeping_suspend_time);
1849
1850 /*
1851 * On some systems the persistent_clock can not be detected at
1852 * timekeeping_init by its return value, so if we see a valid
1853 * value returned, update the persistent_clock_exists flag.
1854 */
1855 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1856 persistent_clock_exists = true;
1857
1858 suspend_timing_needed = true;
1859
1860 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1861 write_seqcount_begin(&tk_core.seq);
1862 timekeeping_forward_now(tk);
1863 timekeeping_suspended = 1;
1864
1865 /*
1866 * Since we've called forward_now, cycle_last stores the value
1867 * just read from the current clocksource. Save this to potentially
1868 * use in suspend timing.
1869 */
1870 curr_clock = tk->tkr_mono.clock;
1871 cycle_now = tk->tkr_mono.cycle_last;
1872 clocksource_start_suspend_timing(curr_clock, cycle_now);
1873
1874 if (persistent_clock_exists) {
1875 /*
1876 * To avoid drift caused by repeated suspend/resumes,
1877 * which each can add ~1 second drift error,
1878 * try to compensate so the difference in system time
1879 * and persistent_clock time stays close to constant.
1880 */
1881 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1882 delta_delta = timespec64_sub(delta, old_delta);
1883 if (abs(delta_delta.tv_sec) >= 2) {
1884 /*
1885 * if delta_delta is too large, assume time correction
1886 * has occurred and set old_delta to the current delta.
1887 */
1888 old_delta = delta;
1889 } else {
1890 /* Otherwise try to adjust old_system to compensate */
1891 timekeeping_suspend_time =
1892 timespec64_add(timekeeping_suspend_time, delta_delta);
1893 }
1894 }
1895
1896 timekeeping_update(tk, TK_MIRROR);
1897 halt_fast_timekeeper(tk);
1898 write_seqcount_end(&tk_core.seq);
1899 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1900
1901 tick_suspend();
1902 clocksource_suspend();
1903 clockevents_suspend();
1904
1905 return 0;
1906}
1907
1908/* sysfs resume/suspend bits for timekeeping */
1909static struct syscore_ops timekeeping_syscore_ops = {
1910 .resume = timekeeping_resume,
1911 .suspend = timekeeping_suspend,
1912};
1913
1914static int __init timekeeping_init_ops(void)
1915{
1916 register_syscore_ops(&timekeeping_syscore_ops);
1917 return 0;
1918}
1919device_initcall(timekeeping_init_ops);
1920
1921/*
1922 * Apply a multiplier adjustment to the timekeeper
1923 */
1924static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1925 s64 offset,
1926 s32 mult_adj)
1927{
1928 s64 interval = tk->cycle_interval;
1929
1930 if (mult_adj == 0) {
1931 return;
1932 } else if (mult_adj == -1) {
1933 interval = -interval;
1934 offset = -offset;
1935 } else if (mult_adj != 1) {
1936 interval *= mult_adj;
1937 offset *= mult_adj;
1938 }
1939
1940 /*
1941 * So the following can be confusing.
1942 *
1943 * To keep things simple, lets assume mult_adj == 1 for now.
1944 *
1945 * When mult_adj != 1, remember that the interval and offset values
1946 * have been appropriately scaled so the math is the same.
1947 *
1948 * The basic idea here is that we're increasing the multiplier
1949 * by one, this causes the xtime_interval to be incremented by
1950 * one cycle_interval. This is because:
1951 * xtime_interval = cycle_interval * mult
1952 * So if mult is being incremented by one:
1953 * xtime_interval = cycle_interval * (mult + 1)
1954 * Its the same as:
1955 * xtime_interval = (cycle_interval * mult) + cycle_interval
1956 * Which can be shortened to:
1957 * xtime_interval += cycle_interval
1958 *
1959 * So offset stores the non-accumulated cycles. Thus the current
1960 * time (in shifted nanoseconds) is:
1961 * now = (offset * adj) + xtime_nsec
1962 * Now, even though we're adjusting the clock frequency, we have
1963 * to keep time consistent. In other words, we can't jump back
1964 * in time, and we also want to avoid jumping forward in time.
1965 *
1966 * So given the same offset value, we need the time to be the same
1967 * both before and after the freq adjustment.
1968 * now = (offset * adj_1) + xtime_nsec_1
1969 * now = (offset * adj_2) + xtime_nsec_2
1970 * So:
1971 * (offset * adj_1) + xtime_nsec_1 =
1972 * (offset * adj_2) + xtime_nsec_2
1973 * And we know:
1974 * adj_2 = adj_1 + 1
1975 * So:
1976 * (offset * adj_1) + xtime_nsec_1 =
1977 * (offset * (adj_1+1)) + xtime_nsec_2
1978 * (offset * adj_1) + xtime_nsec_1 =
1979 * (offset * adj_1) + offset + xtime_nsec_2
1980 * Canceling the sides:
1981 * xtime_nsec_1 = offset + xtime_nsec_2
1982 * Which gives us:
1983 * xtime_nsec_2 = xtime_nsec_1 - offset
1984 * Which simplifies to:
1985 * xtime_nsec -= offset
1986 */
1987 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1988 /* NTP adjustment caused clocksource mult overflow */
1989 WARN_ON_ONCE(1);
1990 return;
1991 }
1992
1993 tk->tkr_mono.mult += mult_adj;
1994 tk->xtime_interval += interval;
1995 tk->tkr_mono.xtime_nsec -= offset;
1996}
1997
1998/*
1999 * Adjust the timekeeper's multiplier to the correct frequency
2000 * and also to reduce the accumulated error value.
2001 */
2002static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2003{
2004 u32 mult;
2005
2006 /*
2007 * Determine the multiplier from the current NTP tick length.
2008 * Avoid expensive division when the tick length doesn't change.
2009 */
2010 if (likely(tk->ntp_tick == ntp_tick_length())) {
2011 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2012 } else {
2013 tk->ntp_tick = ntp_tick_length();
2014 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2015 tk->xtime_remainder, tk->cycle_interval);
2016 }
2017
2018 /*
2019 * If the clock is behind the NTP time, increase the multiplier by 1
2020 * to catch up with it. If it's ahead and there was a remainder in the
2021 * tick division, the clock will slow down. Otherwise it will stay
2022 * ahead until the tick length changes to a non-divisible value.
2023 */
2024 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2025 mult += tk->ntp_err_mult;
2026
2027 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2028
2029 if (unlikely(tk->tkr_mono.clock->maxadj &&
2030 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2031 > tk->tkr_mono.clock->maxadj))) {
2032 printk_once(KERN_WARNING
2033 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2034 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2035 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2036 }
2037
2038 /*
2039 * It may be possible that when we entered this function, xtime_nsec
2040 * was very small. Further, if we're slightly speeding the clocksource
2041 * in the code above, its possible the required corrective factor to
2042 * xtime_nsec could cause it to underflow.
2043 *
2044 * Now, since we have already accumulated the second and the NTP
2045 * subsystem has been notified via second_overflow(), we need to skip
2046 * the next update.
2047 */
2048 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2049 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2050 tk->tkr_mono.shift;
2051 tk->xtime_sec--;
2052 tk->skip_second_overflow = 1;
2053 }
2054}
2055
2056/*
2057 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2058 *
2059 * Helper function that accumulates the nsecs greater than a second
2060 * from the xtime_nsec field to the xtime_secs field.
2061 * It also calls into the NTP code to handle leapsecond processing.
2062 */
2063static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2064{
2065 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2066 unsigned int clock_set = 0;
2067
2068 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2069 int leap;
2070
2071 tk->tkr_mono.xtime_nsec -= nsecps;
2072 tk->xtime_sec++;
2073
2074 /*
2075 * Skip NTP update if this second was accumulated before,
2076 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2077 */
2078 if (unlikely(tk->skip_second_overflow)) {
2079 tk->skip_second_overflow = 0;
2080 continue;
2081 }
2082
2083 /* Figure out if its a leap sec and apply if needed */
2084 leap = second_overflow(tk->xtime_sec);
2085 if (unlikely(leap)) {
2086 struct timespec64 ts;
2087
2088 tk->xtime_sec += leap;
2089
2090 ts.tv_sec = leap;
2091 ts.tv_nsec = 0;
2092 tk_set_wall_to_mono(tk,
2093 timespec64_sub(tk->wall_to_monotonic, ts));
2094
2095 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2096
2097 clock_set = TK_CLOCK_WAS_SET;
2098 }
2099 }
2100 return clock_set;
2101}
2102
2103/*
2104 * logarithmic_accumulation - shifted accumulation of cycles
2105 *
2106 * This functions accumulates a shifted interval of cycles into
2107 * a shifted interval nanoseconds. Allows for O(log) accumulation
2108 * loop.
2109 *
2110 * Returns the unconsumed cycles.
2111 */
2112static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2113 u32 shift, unsigned int *clock_set)
2114{
2115 u64 interval = tk->cycle_interval << shift;
2116 u64 snsec_per_sec;
2117
2118 /* If the offset is smaller than a shifted interval, do nothing */
2119 if (offset < interval)
2120 return offset;
2121
2122 /* Accumulate one shifted interval */
2123 offset -= interval;
2124 tk->tkr_mono.cycle_last += interval;
2125 tk->tkr_raw.cycle_last += interval;
2126
2127 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2128 *clock_set |= accumulate_nsecs_to_secs(tk);
2129
2130 /* Accumulate raw time */
2131 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2132 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2133 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2134 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2135 tk->raw_sec++;
2136 }
2137
2138 /* Accumulate error between NTP and clock interval */
2139 tk->ntp_error += tk->ntp_tick << shift;
2140 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2141 (tk->ntp_error_shift + shift);
2142
2143 return offset;
2144}
2145
2146/*
2147 * timekeeping_advance - Updates the timekeeper to the current time and
2148 * current NTP tick length
2149 */
2150static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2151{
2152 struct timekeeper *real_tk = &tk_core.timekeeper;
2153 struct timekeeper *tk = &shadow_timekeeper;
2154 u64 offset;
2155 int shift = 0, maxshift;
2156 unsigned int clock_set = 0;
2157 unsigned long flags;
2158
2159 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2160
2161 /* Make sure we're fully resumed: */
2162 if (unlikely(timekeeping_suspended))
2163 goto out;
2164
2165 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2166 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2167
2168 /* Check if there's really nothing to do */
2169 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2170 goto out;
2171
2172 /* Do some additional sanity checking */
2173 timekeeping_check_update(tk, offset);
2174
2175 /*
2176 * With NO_HZ we may have to accumulate many cycle_intervals
2177 * (think "ticks") worth of time at once. To do this efficiently,
2178 * we calculate the largest doubling multiple of cycle_intervals
2179 * that is smaller than the offset. We then accumulate that
2180 * chunk in one go, and then try to consume the next smaller
2181 * doubled multiple.
2182 */
2183 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2184 shift = max(0, shift);
2185 /* Bound shift to one less than what overflows tick_length */
2186 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2187 shift = min(shift, maxshift);
2188 while (offset >= tk->cycle_interval) {
2189 offset = logarithmic_accumulation(tk, offset, shift,
2190 &clock_set);
2191 if (offset < tk->cycle_interval<<shift)
2192 shift--;
2193 }
2194
2195 /* Adjust the multiplier to correct NTP error */
2196 timekeeping_adjust(tk, offset);
2197
2198 /*
2199 * Finally, make sure that after the rounding
2200 * xtime_nsec isn't larger than NSEC_PER_SEC
2201 */
2202 clock_set |= accumulate_nsecs_to_secs(tk);
2203
2204 write_seqcount_begin(&tk_core.seq);
2205 /*
2206 * Update the real timekeeper.
2207 *
2208 * We could avoid this memcpy by switching pointers, but that
2209 * requires changes to all other timekeeper usage sites as
2210 * well, i.e. move the timekeeper pointer getter into the
2211 * spinlocked/seqcount protected sections. And we trade this
2212 * memcpy under the tk_core.seq against one before we start
2213 * updating.
2214 */
2215 timekeeping_update(tk, clock_set);
2216 memcpy(real_tk, tk, sizeof(*tk));
2217 /* The memcpy must come last. Do not put anything here! */
2218 write_seqcount_end(&tk_core.seq);
2219out:
2220 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2221
2222 return !!clock_set;
2223}
2224
2225/**
2226 * update_wall_time - Uses the current clocksource to increment the wall time
2227 *
2228 */
2229void update_wall_time(void)
2230{
2231 if (timekeeping_advance(TK_ADV_TICK))
2232 clock_was_set_delayed();
2233}
2234
2235/**
2236 * getboottime64 - Return the real time of system boot.
2237 * @ts: pointer to the timespec64 to be set
2238 *
2239 * Returns the wall-time of boot in a timespec64.
2240 *
2241 * This is based on the wall_to_monotonic offset and the total suspend
2242 * time. Calls to settimeofday will affect the value returned (which
2243 * basically means that however wrong your real time clock is at boot time,
2244 * you get the right time here).
2245 */
2246void getboottime64(struct timespec64 *ts)
2247{
2248 struct timekeeper *tk = &tk_core.timekeeper;
2249 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2250
2251 *ts = ktime_to_timespec64(t);
2252}
2253EXPORT_SYMBOL_GPL(getboottime64);
2254
2255void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2256{
2257 struct timekeeper *tk = &tk_core.timekeeper;
2258 unsigned int seq;
2259
2260 do {
2261 seq = read_seqcount_begin(&tk_core.seq);
2262
2263 *ts = tk_xtime(tk);
2264 } while (read_seqcount_retry(&tk_core.seq, seq));
2265}
2266EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2267
2268void ktime_get_coarse_ts64(struct timespec64 *ts)
2269{
2270 struct timekeeper *tk = &tk_core.timekeeper;
2271 struct timespec64 now, mono;
2272 unsigned int seq;
2273
2274 do {
2275 seq = read_seqcount_begin(&tk_core.seq);
2276
2277 now = tk_xtime(tk);
2278 mono = tk->wall_to_monotonic;
2279 } while (read_seqcount_retry(&tk_core.seq, seq));
2280
2281 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2282 now.tv_nsec + mono.tv_nsec);
2283}
2284EXPORT_SYMBOL(ktime_get_coarse_ts64);
2285
2286/*
2287 * Must hold jiffies_lock
2288 */
2289void do_timer(unsigned long ticks)
2290{
2291 jiffies_64 += ticks;
2292 calc_global_load();
2293}
2294
2295/**
2296 * ktime_get_update_offsets_now - hrtimer helper
2297 * @cwsseq: pointer to check and store the clock was set sequence number
2298 * @offs_real: pointer to storage for monotonic -> realtime offset
2299 * @offs_boot: pointer to storage for monotonic -> boottime offset
2300 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2301 *
2302 * Returns current monotonic time and updates the offsets if the
2303 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2304 * different.
2305 *
2306 * Called from hrtimer_interrupt() or retrigger_next_event()
2307 */
2308ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2309 ktime_t *offs_boot, ktime_t *offs_tai)
2310{
2311 struct timekeeper *tk = &tk_core.timekeeper;
2312 unsigned int seq;
2313 ktime_t base;
2314 u64 nsecs;
2315
2316 do {
2317 seq = read_seqcount_begin(&tk_core.seq);
2318
2319 base = tk->tkr_mono.base;
2320 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2321 base = ktime_add_ns(base, nsecs);
2322
2323 if (*cwsseq != tk->clock_was_set_seq) {
2324 *cwsseq = tk->clock_was_set_seq;
2325 *offs_real = tk->offs_real;
2326 *offs_boot = tk->offs_boot;
2327 *offs_tai = tk->offs_tai;
2328 }
2329
2330 /* Handle leapsecond insertion adjustments */
2331 if (unlikely(base >= tk->next_leap_ktime))
2332 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2333
2334 } while (read_seqcount_retry(&tk_core.seq, seq));
2335
2336 return base;
2337}
2338
2339/*
2340 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2341 */
2342static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2343{
2344 if (txc->modes & ADJ_ADJTIME) {
2345 /* singleshot must not be used with any other mode bits */
2346 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2347 return -EINVAL;
2348 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2349 !capable(CAP_SYS_TIME))
2350 return -EPERM;
2351 } else {
2352 /* In order to modify anything, you gotta be super-user! */
2353 if (txc->modes && !capable(CAP_SYS_TIME))
2354 return -EPERM;
2355 /*
2356 * if the quartz is off by more than 10% then
2357 * something is VERY wrong!
2358 */
2359 if (txc->modes & ADJ_TICK &&
2360 (txc->tick < 900000/USER_HZ ||
2361 txc->tick > 1100000/USER_HZ))
2362 return -EINVAL;
2363 }
2364
2365 if (txc->modes & ADJ_SETOFFSET) {
2366 /* In order to inject time, you gotta be super-user! */
2367 if (!capable(CAP_SYS_TIME))
2368 return -EPERM;
2369
2370 /*
2371 * Validate if a timespec/timeval used to inject a time
2372 * offset is valid. Offsets can be positive or negative, so
2373 * we don't check tv_sec. The value of the timeval/timespec
2374 * is the sum of its fields,but *NOTE*:
2375 * The field tv_usec/tv_nsec must always be non-negative and
2376 * we can't have more nanoseconds/microseconds than a second.
2377 */
2378 if (txc->time.tv_usec < 0)
2379 return -EINVAL;
2380
2381 if (txc->modes & ADJ_NANO) {
2382 if (txc->time.tv_usec >= NSEC_PER_SEC)
2383 return -EINVAL;
2384 } else {
2385 if (txc->time.tv_usec >= USEC_PER_SEC)
2386 return -EINVAL;
2387 }
2388 }
2389
2390 /*
2391 * Check for potential multiplication overflows that can
2392 * only happen on 64-bit systems:
2393 */
2394 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2395 if (LLONG_MIN / PPM_SCALE > txc->freq)
2396 return -EINVAL;
2397 if (LLONG_MAX / PPM_SCALE < txc->freq)
2398 return -EINVAL;
2399 }
2400
2401 return 0;
2402}
2403
2404/**
2405 * random_get_entropy_fallback - Returns the raw clock source value,
2406 * used by random.c for platforms with no valid random_get_entropy().
2407 */
2408unsigned long random_get_entropy_fallback(void)
2409{
2410 struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2411 struct clocksource *clock = READ_ONCE(tkr->clock);
2412
2413 if (unlikely(timekeeping_suspended || !clock))
2414 return 0;
2415 return clock->read(clock);
2416}
2417EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2418
2419/**
2420 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2421 */
2422int do_adjtimex(struct __kernel_timex *txc)
2423{
2424 struct timekeeper *tk = &tk_core.timekeeper;
2425 struct audit_ntp_data ad;
2426 bool clock_set = false;
2427 struct timespec64 ts;
2428 unsigned long flags;
2429 s32 orig_tai, tai;
2430 int ret;
2431
2432 /* Validate the data before disabling interrupts */
2433 ret = timekeeping_validate_timex(txc);
2434 if (ret)
2435 return ret;
2436 add_device_randomness(txc, sizeof(*txc));
2437
2438 if (txc->modes & ADJ_SETOFFSET) {
2439 struct timespec64 delta;
2440 delta.tv_sec = txc->time.tv_sec;
2441 delta.tv_nsec = txc->time.tv_usec;
2442 if (!(txc->modes & ADJ_NANO))
2443 delta.tv_nsec *= 1000;
2444 ret = timekeeping_inject_offset(&delta);
2445 if (ret)
2446 return ret;
2447
2448 audit_tk_injoffset(delta);
2449 }
2450
2451 audit_ntp_init(&ad);
2452
2453 ktime_get_real_ts64(&ts);
2454 add_device_randomness(&ts, sizeof(ts));
2455
2456 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2457 write_seqcount_begin(&tk_core.seq);
2458
2459 orig_tai = tai = tk->tai_offset;
2460 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2461
2462 if (tai != orig_tai) {
2463 __timekeeping_set_tai_offset(tk, tai);
2464 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2465 clock_set = true;
2466 }
2467 tk_update_leap_state(tk);
2468
2469 write_seqcount_end(&tk_core.seq);
2470 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2471
2472 audit_ntp_log(&ad);
2473
2474 /* Update the multiplier immediately if frequency was set directly */
2475 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2476 clock_set |= timekeeping_advance(TK_ADV_FREQ);
2477
2478 if (clock_set)
2479 clock_was_set(CLOCK_REALTIME);
2480
2481 ntp_notify_cmos_timer();
2482
2483 return ret;
2484}
2485
2486#ifdef CONFIG_NTP_PPS
2487/**
2488 * hardpps() - Accessor function to NTP __hardpps function
2489 */
2490void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2491{
2492 unsigned long flags;
2493
2494 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2495 write_seqcount_begin(&tk_core.seq);
2496
2497 __hardpps(phase_ts, raw_ts);
2498
2499 write_seqcount_end(&tk_core.seq);
2500 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2501}
2502EXPORT_SYMBOL(hardpps);
2503#endif /* CONFIG_NTP_PPS */