xref: /linux/kernel/time/timekeeping.c (revision bd628c1bed7902ec1f24ba0fe70758949146abbe)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Kernel timekeeping code and accessor functions. Based on code from
4  *  timer.c, moved in commit 8524070b7982.
5  */
6 #include <linux/timekeeper_internal.h>
7 #include <linux/module.h>
8 #include <linux/interrupt.h>
9 #include <linux/percpu.h>
10 #include <linux/init.h>
11 #include <linux/mm.h>
12 #include <linux/nmi.h>
13 #include <linux/sched.h>
14 #include <linux/sched/loadavg.h>
15 #include <linux/sched/clock.h>
16 #include <linux/syscore_ops.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/tick.h>
21 #include <linux/stop_machine.h>
22 #include <linux/pvclock_gtod.h>
23 #include <linux/compiler.h>
24 
25 #include "tick-internal.h"
26 #include "ntp_internal.h"
27 #include "timekeeping_internal.h"
28 
29 #define TK_CLEAR_NTP		(1 << 0)
30 #define TK_MIRROR		(1 << 1)
31 #define TK_CLOCK_WAS_SET	(1 << 2)
32 
33 enum timekeeping_adv_mode {
34 	/* Update timekeeper when a tick has passed */
35 	TK_ADV_TICK,
36 
37 	/* Update timekeeper on a direct frequency change */
38 	TK_ADV_FREQ
39 };
40 
41 /*
42  * The most important data for readout fits into a single 64 byte
43  * cache line.
44  */
45 static struct {
46 	seqcount_t		seq;
47 	struct timekeeper	timekeeper;
48 } tk_core ____cacheline_aligned = {
49 	.seq = SEQCNT_ZERO(tk_core.seq),
50 };
51 
52 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
53 static struct timekeeper shadow_timekeeper;
54 
55 /**
56  * struct tk_fast - NMI safe timekeeper
57  * @seq:	Sequence counter for protecting updates. The lowest bit
58  *		is the index for the tk_read_base array
59  * @base:	tk_read_base array. Access is indexed by the lowest bit of
60  *		@seq.
61  *
62  * See @update_fast_timekeeper() below.
63  */
64 struct tk_fast {
65 	seqcount_t		seq;
66 	struct tk_read_base	base[2];
67 };
68 
69 /* Suspend-time cycles value for halted fast timekeeper. */
70 static u64 cycles_at_suspend;
71 
72 static u64 dummy_clock_read(struct clocksource *cs)
73 {
74 	return cycles_at_suspend;
75 }
76 
77 static struct clocksource dummy_clock = {
78 	.read = dummy_clock_read,
79 };
80 
81 static struct tk_fast tk_fast_mono ____cacheline_aligned = {
82 	.base[0] = { .clock = &dummy_clock, },
83 	.base[1] = { .clock = &dummy_clock, },
84 };
85 
86 static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
87 	.base[0] = { .clock = &dummy_clock, },
88 	.base[1] = { .clock = &dummy_clock, },
89 };
90 
91 /* flag for if timekeeping is suspended */
92 int __read_mostly timekeeping_suspended;
93 
94 static inline void tk_normalize_xtime(struct timekeeper *tk)
95 {
96 	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
97 		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
98 		tk->xtime_sec++;
99 	}
100 	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
101 		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
102 		tk->raw_sec++;
103 	}
104 }
105 
106 static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
107 {
108 	struct timespec64 ts;
109 
110 	ts.tv_sec = tk->xtime_sec;
111 	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
112 	return ts;
113 }
114 
115 static void tk_set_xtime(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 }
120 
121 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
122 {
123 	tk->xtime_sec += ts->tv_sec;
124 	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
125 	tk_normalize_xtime(tk);
126 }
127 
128 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
129 {
130 	struct timespec64 tmp;
131 
132 	/*
133 	 * Verify consistency of: offset_real = -wall_to_monotonic
134 	 * before modifying anything
135 	 */
136 	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
137 					-tk->wall_to_monotonic.tv_nsec);
138 	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
139 	tk->wall_to_monotonic = wtm;
140 	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
141 	tk->offs_real = timespec64_to_ktime(tmp);
142 	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
143 }
144 
145 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
146 {
147 	tk->offs_boot = ktime_add(tk->offs_boot, delta);
148 }
149 
150 /*
151  * tk_clock_read - atomic clocksource read() helper
152  *
153  * This helper is necessary to use in the read paths because, while the
154  * seqlock ensures we don't return a bad value while structures are updated,
155  * it doesn't protect from potential crashes. There is the possibility that
156  * the tkr's clocksource may change between the read reference, and the
157  * clock reference passed to the read function.  This can cause crashes if
158  * the wrong clocksource is passed to the wrong read function.
159  * This isn't necessary to use when holding the timekeeper_lock or doing
160  * a read of the fast-timekeeper tkrs (which is protected by its own locking
161  * and update logic).
162  */
163 static inline u64 tk_clock_read(const struct tk_read_base *tkr)
164 {
165 	struct clocksource *clock = READ_ONCE(tkr->clock);
166 
167 	return clock->read(clock);
168 }
169 
170 #ifdef CONFIG_DEBUG_TIMEKEEPING
171 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
172 
173 static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
174 {
175 
176 	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
177 	const char *name = tk->tkr_mono.clock->name;
178 
179 	if (offset > max_cycles) {
180 		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
181 				offset, name, max_cycles);
182 		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
183 	} else {
184 		if (offset > (max_cycles >> 1)) {
185 			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
186 					offset, name, max_cycles >> 1);
187 			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
188 		}
189 	}
190 
191 	if (tk->underflow_seen) {
192 		if (jiffies - tk->last_warning > WARNING_FREQ) {
193 			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
194 			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
195 			printk_deferred("         Your kernel is probably still fine.\n");
196 			tk->last_warning = jiffies;
197 		}
198 		tk->underflow_seen = 0;
199 	}
200 
201 	if (tk->overflow_seen) {
202 		if (jiffies - tk->last_warning > WARNING_FREQ) {
203 			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
204 			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
205 			printk_deferred("         Your kernel is probably still fine.\n");
206 			tk->last_warning = jiffies;
207 		}
208 		tk->overflow_seen = 0;
209 	}
210 }
211 
212 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
213 {
214 	struct timekeeper *tk = &tk_core.timekeeper;
215 	u64 now, last, mask, max, delta;
216 	unsigned int seq;
217 
218 	/*
219 	 * Since we're called holding a seqlock, the data may shift
220 	 * under us while we're doing the calculation. This can cause
221 	 * false positives, since we'd note a problem but throw the
222 	 * results away. So nest another seqlock here to atomically
223 	 * grab the points we are checking with.
224 	 */
225 	do {
226 		seq = read_seqcount_begin(&tk_core.seq);
227 		now = tk_clock_read(tkr);
228 		last = tkr->cycle_last;
229 		mask = tkr->mask;
230 		max = tkr->clock->max_cycles;
231 	} while (read_seqcount_retry(&tk_core.seq, seq));
232 
233 	delta = clocksource_delta(now, last, mask);
234 
235 	/*
236 	 * Try to catch underflows by checking if we are seeing small
237 	 * mask-relative negative values.
238 	 */
239 	if (unlikely((~delta & mask) < (mask >> 3))) {
240 		tk->underflow_seen = 1;
241 		delta = 0;
242 	}
243 
244 	/* Cap delta value to the max_cycles values to avoid mult overflows */
245 	if (unlikely(delta > max)) {
246 		tk->overflow_seen = 1;
247 		delta = tkr->clock->max_cycles;
248 	}
249 
250 	return delta;
251 }
252 #else
253 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
254 {
255 }
256 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
257 {
258 	u64 cycle_now, delta;
259 
260 	/* read clocksource */
261 	cycle_now = tk_clock_read(tkr);
262 
263 	/* calculate the delta since the last update_wall_time */
264 	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
265 
266 	return delta;
267 }
268 #endif
269 
270 /**
271  * tk_setup_internals - Set up internals to use clocksource clock.
272  *
273  * @tk:		The target timekeeper to setup.
274  * @clock:		Pointer to clocksource.
275  *
276  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
277  * pair and interval request.
278  *
279  * Unless you're the timekeeping code, you should not be using this!
280  */
281 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
282 {
283 	u64 interval;
284 	u64 tmp, ntpinterval;
285 	struct clocksource *old_clock;
286 
287 	++tk->cs_was_changed_seq;
288 	old_clock = tk->tkr_mono.clock;
289 	tk->tkr_mono.clock = clock;
290 	tk->tkr_mono.mask = clock->mask;
291 	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
292 
293 	tk->tkr_raw.clock = clock;
294 	tk->tkr_raw.mask = clock->mask;
295 	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
296 
297 	/* Do the ns -> cycle conversion first, using original mult */
298 	tmp = NTP_INTERVAL_LENGTH;
299 	tmp <<= clock->shift;
300 	ntpinterval = tmp;
301 	tmp += clock->mult/2;
302 	do_div(tmp, clock->mult);
303 	if (tmp == 0)
304 		tmp = 1;
305 
306 	interval = (u64) tmp;
307 	tk->cycle_interval = interval;
308 
309 	/* Go back from cycles -> shifted ns */
310 	tk->xtime_interval = interval * clock->mult;
311 	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
312 	tk->raw_interval = interval * clock->mult;
313 
314 	 /* if changing clocks, convert xtime_nsec shift units */
315 	if (old_clock) {
316 		int shift_change = clock->shift - old_clock->shift;
317 		if (shift_change < 0) {
318 			tk->tkr_mono.xtime_nsec >>= -shift_change;
319 			tk->tkr_raw.xtime_nsec >>= -shift_change;
320 		} else {
321 			tk->tkr_mono.xtime_nsec <<= shift_change;
322 			tk->tkr_raw.xtime_nsec <<= shift_change;
323 		}
324 	}
325 
326 	tk->tkr_mono.shift = clock->shift;
327 	tk->tkr_raw.shift = clock->shift;
328 
329 	tk->ntp_error = 0;
330 	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
331 	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
332 
333 	/*
334 	 * The timekeeper keeps its own mult values for the currently
335 	 * active clocksource. These value will be adjusted via NTP
336 	 * to counteract clock drifting.
337 	 */
338 	tk->tkr_mono.mult = clock->mult;
339 	tk->tkr_raw.mult = clock->mult;
340 	tk->ntp_err_mult = 0;
341 	tk->skip_second_overflow = 0;
342 }
343 
344 /* Timekeeper helper functions. */
345 
346 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
347 static u32 default_arch_gettimeoffset(void) { return 0; }
348 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
349 #else
350 static inline u32 arch_gettimeoffset(void) { return 0; }
351 #endif
352 
353 static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
354 {
355 	u64 nsec;
356 
357 	nsec = delta * tkr->mult + tkr->xtime_nsec;
358 	nsec >>= tkr->shift;
359 
360 	/* If arch requires, add in get_arch_timeoffset() */
361 	return nsec + arch_gettimeoffset();
362 }
363 
364 static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
365 {
366 	u64 delta;
367 
368 	delta = timekeeping_get_delta(tkr);
369 	return timekeeping_delta_to_ns(tkr, delta);
370 }
371 
372 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
373 {
374 	u64 delta;
375 
376 	/* calculate the delta since the last update_wall_time */
377 	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
378 	return timekeeping_delta_to_ns(tkr, delta);
379 }
380 
381 /**
382  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
383  * @tkr: Timekeeping readout base from which we take the update
384  *
385  * We want to use this from any context including NMI and tracing /
386  * instrumenting the timekeeping code itself.
387  *
388  * Employ the latch technique; see @raw_write_seqcount_latch.
389  *
390  * So if a NMI hits the update of base[0] then it will use base[1]
391  * which is still consistent. In the worst case this can result is a
392  * slightly wrong timestamp (a few nanoseconds). See
393  * @ktime_get_mono_fast_ns.
394  */
395 static void update_fast_timekeeper(const struct tk_read_base *tkr,
396 				   struct tk_fast *tkf)
397 {
398 	struct tk_read_base *base = tkf->base;
399 
400 	/* Force readers off to base[1] */
401 	raw_write_seqcount_latch(&tkf->seq);
402 
403 	/* Update base[0] */
404 	memcpy(base, tkr, sizeof(*base));
405 
406 	/* Force readers back to base[0] */
407 	raw_write_seqcount_latch(&tkf->seq);
408 
409 	/* Update base[1] */
410 	memcpy(base + 1, base, sizeof(*base));
411 }
412 
413 /**
414  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
415  *
416  * This timestamp is not guaranteed to be monotonic across an update.
417  * The timestamp is calculated by:
418  *
419  *	now = base_mono + clock_delta * slope
420  *
421  * So if the update lowers the slope, readers who are forced to the
422  * not yet updated second array are still using the old steeper slope.
423  *
424  * tmono
425  * ^
426  * |    o  n
427  * |   o n
428  * |  u
429  * | o
430  * |o
431  * |12345678---> reader order
432  *
433  * o = old slope
434  * u = update
435  * n = new slope
436  *
437  * So reader 6 will observe time going backwards versus reader 5.
438  *
439  * While other CPUs are likely to be able observe that, the only way
440  * for a CPU local observation is when an NMI hits in the middle of
441  * the update. Timestamps taken from that NMI context might be ahead
442  * of the following timestamps. Callers need to be aware of that and
443  * deal with it.
444  */
445 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
446 {
447 	struct tk_read_base *tkr;
448 	unsigned int seq;
449 	u64 now;
450 
451 	do {
452 		seq = raw_read_seqcount_latch(&tkf->seq);
453 		tkr = tkf->base + (seq & 0x01);
454 		now = ktime_to_ns(tkr->base);
455 
456 		now += timekeeping_delta_to_ns(tkr,
457 				clocksource_delta(
458 					tk_clock_read(tkr),
459 					tkr->cycle_last,
460 					tkr->mask));
461 	} while (read_seqcount_retry(&tkf->seq, seq));
462 
463 	return now;
464 }
465 
466 u64 ktime_get_mono_fast_ns(void)
467 {
468 	return __ktime_get_fast_ns(&tk_fast_mono);
469 }
470 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
471 
472 u64 ktime_get_raw_fast_ns(void)
473 {
474 	return __ktime_get_fast_ns(&tk_fast_raw);
475 }
476 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
477 
478 /**
479  * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
480  *
481  * To keep it NMI safe since we're accessing from tracing, we're not using a
482  * separate timekeeper with updates to monotonic clock and boot offset
483  * protected with seqlocks. This has the following minor side effects:
484  *
485  * (1) Its possible that a timestamp be taken after the boot offset is updated
486  * but before the timekeeper is updated. If this happens, the new boot offset
487  * is added to the old timekeeping making the clock appear to update slightly
488  * earlier:
489  *    CPU 0                                        CPU 1
490  *    timekeeping_inject_sleeptime64()
491  *    __timekeeping_inject_sleeptime(tk, delta);
492  *                                                 timestamp();
493  *    timekeeping_update(tk, TK_CLEAR_NTP...);
494  *
495  * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
496  * partially updated.  Since the tk->offs_boot update is a rare event, this
497  * should be a rare occurrence which postprocessing should be able to handle.
498  */
499 u64 notrace ktime_get_boot_fast_ns(void)
500 {
501 	struct timekeeper *tk = &tk_core.timekeeper;
502 
503 	return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
504 }
505 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
506 
507 
508 /*
509  * See comment for __ktime_get_fast_ns() vs. timestamp ordering
510  */
511 static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf)
512 {
513 	struct tk_read_base *tkr;
514 	unsigned int seq;
515 	u64 now;
516 
517 	do {
518 		seq = raw_read_seqcount_latch(&tkf->seq);
519 		tkr = tkf->base + (seq & 0x01);
520 		now = ktime_to_ns(tkr->base_real);
521 
522 		now += timekeeping_delta_to_ns(tkr,
523 				clocksource_delta(
524 					tk_clock_read(tkr),
525 					tkr->cycle_last,
526 					tkr->mask));
527 	} while (read_seqcount_retry(&tkf->seq, seq));
528 
529 	return now;
530 }
531 
532 /**
533  * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
534  */
535 u64 ktime_get_real_fast_ns(void)
536 {
537 	return __ktime_get_real_fast_ns(&tk_fast_mono);
538 }
539 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
540 
541 /**
542  * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
543  * @tk: Timekeeper to snapshot.
544  *
545  * It generally is unsafe to access the clocksource after timekeeping has been
546  * suspended, so take a snapshot of the readout base of @tk and use it as the
547  * fast timekeeper's readout base while suspended.  It will return the same
548  * number of cycles every time until timekeeping is resumed at which time the
549  * proper readout base for the fast timekeeper will be restored automatically.
550  */
551 static void halt_fast_timekeeper(const struct timekeeper *tk)
552 {
553 	static struct tk_read_base tkr_dummy;
554 	const struct tk_read_base *tkr = &tk->tkr_mono;
555 
556 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
557 	cycles_at_suspend = tk_clock_read(tkr);
558 	tkr_dummy.clock = &dummy_clock;
559 	tkr_dummy.base_real = tkr->base + tk->offs_real;
560 	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
561 
562 	tkr = &tk->tkr_raw;
563 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
564 	tkr_dummy.clock = &dummy_clock;
565 	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
566 }
567 
568 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
569 
570 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
571 {
572 	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
573 }
574 
575 /**
576  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
577  */
578 int pvclock_gtod_register_notifier(struct notifier_block *nb)
579 {
580 	struct timekeeper *tk = &tk_core.timekeeper;
581 	unsigned long flags;
582 	int ret;
583 
584 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
585 	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
586 	update_pvclock_gtod(tk, true);
587 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
588 
589 	return ret;
590 }
591 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
592 
593 /**
594  * pvclock_gtod_unregister_notifier - unregister a pvclock
595  * timedata update listener
596  */
597 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
598 {
599 	unsigned long flags;
600 	int ret;
601 
602 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
603 	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
604 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
605 
606 	return ret;
607 }
608 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
609 
610 /*
611  * tk_update_leap_state - helper to update the next_leap_ktime
612  */
613 static inline void tk_update_leap_state(struct timekeeper *tk)
614 {
615 	tk->next_leap_ktime = ntp_get_next_leap();
616 	if (tk->next_leap_ktime != KTIME_MAX)
617 		/* Convert to monotonic time */
618 		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
619 }
620 
621 /*
622  * Update the ktime_t based scalar nsec members of the timekeeper
623  */
624 static inline void tk_update_ktime_data(struct timekeeper *tk)
625 {
626 	u64 seconds;
627 	u32 nsec;
628 
629 	/*
630 	 * The xtime based monotonic readout is:
631 	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
632 	 * The ktime based monotonic readout is:
633 	 *	nsec = base_mono + now();
634 	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
635 	 */
636 	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
637 	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
638 	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
639 
640 	/*
641 	 * The sum of the nanoseconds portions of xtime and
642 	 * wall_to_monotonic can be greater/equal one second. Take
643 	 * this into account before updating tk->ktime_sec.
644 	 */
645 	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
646 	if (nsec >= NSEC_PER_SEC)
647 		seconds++;
648 	tk->ktime_sec = seconds;
649 
650 	/* Update the monotonic raw base */
651 	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
652 }
653 
654 /* must hold timekeeper_lock */
655 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
656 {
657 	if (action & TK_CLEAR_NTP) {
658 		tk->ntp_error = 0;
659 		ntp_clear();
660 	}
661 
662 	tk_update_leap_state(tk);
663 	tk_update_ktime_data(tk);
664 
665 	update_vsyscall(tk);
666 	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
667 
668 	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
669 	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
670 	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
671 
672 	if (action & TK_CLOCK_WAS_SET)
673 		tk->clock_was_set_seq++;
674 	/*
675 	 * The mirroring of the data to the shadow-timekeeper needs
676 	 * to happen last here to ensure we don't over-write the
677 	 * timekeeper structure on the next update with stale data
678 	 */
679 	if (action & TK_MIRROR)
680 		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
681 		       sizeof(tk_core.timekeeper));
682 }
683 
684 /**
685  * timekeeping_forward_now - update clock to the current time
686  *
687  * Forward the current clock to update its state since the last call to
688  * update_wall_time(). This is useful before significant clock changes,
689  * as it avoids having to deal with this time offset explicitly.
690  */
691 static void timekeeping_forward_now(struct timekeeper *tk)
692 {
693 	u64 cycle_now, delta;
694 
695 	cycle_now = tk_clock_read(&tk->tkr_mono);
696 	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
697 	tk->tkr_mono.cycle_last = cycle_now;
698 	tk->tkr_raw.cycle_last  = cycle_now;
699 
700 	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
701 
702 	/* If arch requires, add in get_arch_timeoffset() */
703 	tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
704 
705 
706 	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
707 
708 	/* If arch requires, add in get_arch_timeoffset() */
709 	tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
710 
711 	tk_normalize_xtime(tk);
712 }
713 
714 /**
715  * ktime_get_real_ts64 - Returns the time of day in a timespec64.
716  * @ts:		pointer to the timespec to be set
717  *
718  * Returns the time of day in a timespec64 (WARN if suspended).
719  */
720 void ktime_get_real_ts64(struct timespec64 *ts)
721 {
722 	struct timekeeper *tk = &tk_core.timekeeper;
723 	unsigned long seq;
724 	u64 nsecs;
725 
726 	WARN_ON(timekeeping_suspended);
727 
728 	do {
729 		seq = read_seqcount_begin(&tk_core.seq);
730 
731 		ts->tv_sec = tk->xtime_sec;
732 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
733 
734 	} while (read_seqcount_retry(&tk_core.seq, seq));
735 
736 	ts->tv_nsec = 0;
737 	timespec64_add_ns(ts, nsecs);
738 }
739 EXPORT_SYMBOL(ktime_get_real_ts64);
740 
741 ktime_t ktime_get(void)
742 {
743 	struct timekeeper *tk = &tk_core.timekeeper;
744 	unsigned int seq;
745 	ktime_t base;
746 	u64 nsecs;
747 
748 	WARN_ON(timekeeping_suspended);
749 
750 	do {
751 		seq = read_seqcount_begin(&tk_core.seq);
752 		base = tk->tkr_mono.base;
753 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
754 
755 	} while (read_seqcount_retry(&tk_core.seq, seq));
756 
757 	return ktime_add_ns(base, nsecs);
758 }
759 EXPORT_SYMBOL_GPL(ktime_get);
760 
761 u32 ktime_get_resolution_ns(void)
762 {
763 	struct timekeeper *tk = &tk_core.timekeeper;
764 	unsigned int seq;
765 	u32 nsecs;
766 
767 	WARN_ON(timekeeping_suspended);
768 
769 	do {
770 		seq = read_seqcount_begin(&tk_core.seq);
771 		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
772 	} while (read_seqcount_retry(&tk_core.seq, seq));
773 
774 	return nsecs;
775 }
776 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
777 
778 static ktime_t *offsets[TK_OFFS_MAX] = {
779 	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
780 	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
781 	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
782 };
783 
784 ktime_t ktime_get_with_offset(enum tk_offsets offs)
785 {
786 	struct timekeeper *tk = &tk_core.timekeeper;
787 	unsigned int seq;
788 	ktime_t base, *offset = offsets[offs];
789 	u64 nsecs;
790 
791 	WARN_ON(timekeeping_suspended);
792 
793 	do {
794 		seq = read_seqcount_begin(&tk_core.seq);
795 		base = ktime_add(tk->tkr_mono.base, *offset);
796 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
797 
798 	} while (read_seqcount_retry(&tk_core.seq, seq));
799 
800 	return ktime_add_ns(base, nsecs);
801 
802 }
803 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
804 
805 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
806 {
807 	struct timekeeper *tk = &tk_core.timekeeper;
808 	unsigned int seq;
809 	ktime_t base, *offset = offsets[offs];
810 
811 	WARN_ON(timekeeping_suspended);
812 
813 	do {
814 		seq = read_seqcount_begin(&tk_core.seq);
815 		base = ktime_add(tk->tkr_mono.base, *offset);
816 
817 	} while (read_seqcount_retry(&tk_core.seq, seq));
818 
819 	return base;
820 
821 }
822 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
823 
824 /**
825  * ktime_mono_to_any() - convert mononotic time to any other time
826  * @tmono:	time to convert.
827  * @offs:	which offset to use
828  */
829 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
830 {
831 	ktime_t *offset = offsets[offs];
832 	unsigned long seq;
833 	ktime_t tconv;
834 
835 	do {
836 		seq = read_seqcount_begin(&tk_core.seq);
837 		tconv = ktime_add(tmono, *offset);
838 	} while (read_seqcount_retry(&tk_core.seq, seq));
839 
840 	return tconv;
841 }
842 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
843 
844 /**
845  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
846  */
847 ktime_t ktime_get_raw(void)
848 {
849 	struct timekeeper *tk = &tk_core.timekeeper;
850 	unsigned int seq;
851 	ktime_t base;
852 	u64 nsecs;
853 
854 	do {
855 		seq = read_seqcount_begin(&tk_core.seq);
856 		base = tk->tkr_raw.base;
857 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
858 
859 	} while (read_seqcount_retry(&tk_core.seq, seq));
860 
861 	return ktime_add_ns(base, nsecs);
862 }
863 EXPORT_SYMBOL_GPL(ktime_get_raw);
864 
865 /**
866  * ktime_get_ts64 - get the monotonic clock in timespec64 format
867  * @ts:		pointer to timespec variable
868  *
869  * The function calculates the monotonic clock from the realtime
870  * clock and the wall_to_monotonic offset and stores the result
871  * in normalized timespec64 format in the variable pointed to by @ts.
872  */
873 void ktime_get_ts64(struct timespec64 *ts)
874 {
875 	struct timekeeper *tk = &tk_core.timekeeper;
876 	struct timespec64 tomono;
877 	unsigned int seq;
878 	u64 nsec;
879 
880 	WARN_ON(timekeeping_suspended);
881 
882 	do {
883 		seq = read_seqcount_begin(&tk_core.seq);
884 		ts->tv_sec = tk->xtime_sec;
885 		nsec = timekeeping_get_ns(&tk->tkr_mono);
886 		tomono = tk->wall_to_monotonic;
887 
888 	} while (read_seqcount_retry(&tk_core.seq, seq));
889 
890 	ts->tv_sec += tomono.tv_sec;
891 	ts->tv_nsec = 0;
892 	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
893 }
894 EXPORT_SYMBOL_GPL(ktime_get_ts64);
895 
896 /**
897  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
898  *
899  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
900  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
901  * works on both 32 and 64 bit systems. On 32 bit systems the readout
902  * covers ~136 years of uptime which should be enough to prevent
903  * premature wrap arounds.
904  */
905 time64_t ktime_get_seconds(void)
906 {
907 	struct timekeeper *tk = &tk_core.timekeeper;
908 
909 	WARN_ON(timekeeping_suspended);
910 	return tk->ktime_sec;
911 }
912 EXPORT_SYMBOL_GPL(ktime_get_seconds);
913 
914 /**
915  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
916  *
917  * Returns the wall clock seconds since 1970. This replaces the
918  * get_seconds() interface which is not y2038 safe on 32bit systems.
919  *
920  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
921  * 32bit systems the access must be protected with the sequence
922  * counter to provide "atomic" access to the 64bit tk->xtime_sec
923  * value.
924  */
925 time64_t ktime_get_real_seconds(void)
926 {
927 	struct timekeeper *tk = &tk_core.timekeeper;
928 	time64_t seconds;
929 	unsigned int seq;
930 
931 	if (IS_ENABLED(CONFIG_64BIT))
932 		return tk->xtime_sec;
933 
934 	do {
935 		seq = read_seqcount_begin(&tk_core.seq);
936 		seconds = tk->xtime_sec;
937 
938 	} while (read_seqcount_retry(&tk_core.seq, seq));
939 
940 	return seconds;
941 }
942 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
943 
944 /**
945  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
946  * but without the sequence counter protect. This internal function
947  * is called just when timekeeping lock is already held.
948  */
949 time64_t __ktime_get_real_seconds(void)
950 {
951 	struct timekeeper *tk = &tk_core.timekeeper;
952 
953 	return tk->xtime_sec;
954 }
955 
956 /**
957  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
958  * @systime_snapshot:	pointer to struct receiving the system time snapshot
959  */
960 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
961 {
962 	struct timekeeper *tk = &tk_core.timekeeper;
963 	unsigned long seq;
964 	ktime_t base_raw;
965 	ktime_t base_real;
966 	u64 nsec_raw;
967 	u64 nsec_real;
968 	u64 now;
969 
970 	WARN_ON_ONCE(timekeeping_suspended);
971 
972 	do {
973 		seq = read_seqcount_begin(&tk_core.seq);
974 		now = tk_clock_read(&tk->tkr_mono);
975 		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
976 		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
977 		base_real = ktime_add(tk->tkr_mono.base,
978 				      tk_core.timekeeper.offs_real);
979 		base_raw = tk->tkr_raw.base;
980 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
981 		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
982 	} while (read_seqcount_retry(&tk_core.seq, seq));
983 
984 	systime_snapshot->cycles = now;
985 	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
986 	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
987 }
988 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
989 
990 /* Scale base by mult/div checking for overflow */
991 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
992 {
993 	u64 tmp, rem;
994 
995 	tmp = div64_u64_rem(*base, div, &rem);
996 
997 	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
998 	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
999 		return -EOVERFLOW;
1000 	tmp *= mult;
1001 	rem *= mult;
1002 
1003 	do_div(rem, div);
1004 	*base = tmp + rem;
1005 	return 0;
1006 }
1007 
1008 /**
1009  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1010  * @history:			Snapshot representing start of history
1011  * @partial_history_cycles:	Cycle offset into history (fractional part)
1012  * @total_history_cycles:	Total history length in cycles
1013  * @discontinuity:		True indicates clock was set on history period
1014  * @ts:				Cross timestamp that should be adjusted using
1015  *	partial/total ratio
1016  *
1017  * Helper function used by get_device_system_crosststamp() to correct the
1018  * crosstimestamp corresponding to the start of the current interval to the
1019  * system counter value (timestamp point) provided by the driver. The
1020  * total_history_* quantities are the total history starting at the provided
1021  * reference point and ending at the start of the current interval. The cycle
1022  * count between the driver timestamp point and the start of the current
1023  * interval is partial_history_cycles.
1024  */
1025 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1026 					 u64 partial_history_cycles,
1027 					 u64 total_history_cycles,
1028 					 bool discontinuity,
1029 					 struct system_device_crosststamp *ts)
1030 {
1031 	struct timekeeper *tk = &tk_core.timekeeper;
1032 	u64 corr_raw, corr_real;
1033 	bool interp_forward;
1034 	int ret;
1035 
1036 	if (total_history_cycles == 0 || partial_history_cycles == 0)
1037 		return 0;
1038 
1039 	/* Interpolate shortest distance from beginning or end of history */
1040 	interp_forward = partial_history_cycles > total_history_cycles / 2;
1041 	partial_history_cycles = interp_forward ?
1042 		total_history_cycles - partial_history_cycles :
1043 		partial_history_cycles;
1044 
1045 	/*
1046 	 * Scale the monotonic raw time delta by:
1047 	 *	partial_history_cycles / total_history_cycles
1048 	 */
1049 	corr_raw = (u64)ktime_to_ns(
1050 		ktime_sub(ts->sys_monoraw, history->raw));
1051 	ret = scale64_check_overflow(partial_history_cycles,
1052 				     total_history_cycles, &corr_raw);
1053 	if (ret)
1054 		return ret;
1055 
1056 	/*
1057 	 * If there is a discontinuity in the history, scale monotonic raw
1058 	 *	correction by:
1059 	 *	mult(real)/mult(raw) yielding the realtime correction
1060 	 * Otherwise, calculate the realtime correction similar to monotonic
1061 	 *	raw calculation
1062 	 */
1063 	if (discontinuity) {
1064 		corr_real = mul_u64_u32_div
1065 			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1066 	} else {
1067 		corr_real = (u64)ktime_to_ns(
1068 			ktime_sub(ts->sys_realtime, history->real));
1069 		ret = scale64_check_overflow(partial_history_cycles,
1070 					     total_history_cycles, &corr_real);
1071 		if (ret)
1072 			return ret;
1073 	}
1074 
1075 	/* Fixup monotonic raw and real time time values */
1076 	if (interp_forward) {
1077 		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1078 		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1079 	} else {
1080 		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1081 		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1082 	}
1083 
1084 	return 0;
1085 }
1086 
1087 /*
1088  * cycle_between - true if test occurs chronologically between before and after
1089  */
1090 static bool cycle_between(u64 before, u64 test, u64 after)
1091 {
1092 	if (test > before && test < after)
1093 		return true;
1094 	if (test < before && before > after)
1095 		return true;
1096 	return false;
1097 }
1098 
1099 /**
1100  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1101  * @get_time_fn:	Callback to get simultaneous device time and
1102  *	system counter from the device driver
1103  * @ctx:		Context passed to get_time_fn()
1104  * @history_begin:	Historical reference point used to interpolate system
1105  *	time when counter provided by the driver is before the current interval
1106  * @xtstamp:		Receives simultaneously captured system and device time
1107  *
1108  * Reads a timestamp from a device and correlates it to system time
1109  */
1110 int get_device_system_crosststamp(int (*get_time_fn)
1111 				  (ktime_t *device_time,
1112 				   struct system_counterval_t *sys_counterval,
1113 				   void *ctx),
1114 				  void *ctx,
1115 				  struct system_time_snapshot *history_begin,
1116 				  struct system_device_crosststamp *xtstamp)
1117 {
1118 	struct system_counterval_t system_counterval;
1119 	struct timekeeper *tk = &tk_core.timekeeper;
1120 	u64 cycles, now, interval_start;
1121 	unsigned int clock_was_set_seq = 0;
1122 	ktime_t base_real, base_raw;
1123 	u64 nsec_real, nsec_raw;
1124 	u8 cs_was_changed_seq;
1125 	unsigned long seq;
1126 	bool do_interp;
1127 	int ret;
1128 
1129 	do {
1130 		seq = read_seqcount_begin(&tk_core.seq);
1131 		/*
1132 		 * Try to synchronously capture device time and a system
1133 		 * counter value calling back into the device driver
1134 		 */
1135 		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1136 		if (ret)
1137 			return ret;
1138 
1139 		/*
1140 		 * Verify that the clocksource associated with the captured
1141 		 * system counter value is the same as the currently installed
1142 		 * timekeeper clocksource
1143 		 */
1144 		if (tk->tkr_mono.clock != system_counterval.cs)
1145 			return -ENODEV;
1146 		cycles = system_counterval.cycles;
1147 
1148 		/*
1149 		 * Check whether the system counter value provided by the
1150 		 * device driver is on the current timekeeping interval.
1151 		 */
1152 		now = tk_clock_read(&tk->tkr_mono);
1153 		interval_start = tk->tkr_mono.cycle_last;
1154 		if (!cycle_between(interval_start, cycles, now)) {
1155 			clock_was_set_seq = tk->clock_was_set_seq;
1156 			cs_was_changed_seq = tk->cs_was_changed_seq;
1157 			cycles = interval_start;
1158 			do_interp = true;
1159 		} else {
1160 			do_interp = false;
1161 		}
1162 
1163 		base_real = ktime_add(tk->tkr_mono.base,
1164 				      tk_core.timekeeper.offs_real);
1165 		base_raw = tk->tkr_raw.base;
1166 
1167 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1168 						     system_counterval.cycles);
1169 		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1170 						    system_counterval.cycles);
1171 	} while (read_seqcount_retry(&tk_core.seq, seq));
1172 
1173 	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1174 	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1175 
1176 	/*
1177 	 * Interpolate if necessary, adjusting back from the start of the
1178 	 * current interval
1179 	 */
1180 	if (do_interp) {
1181 		u64 partial_history_cycles, total_history_cycles;
1182 		bool discontinuity;
1183 
1184 		/*
1185 		 * Check that the counter value occurs after the provided
1186 		 * history reference and that the history doesn't cross a
1187 		 * clocksource change
1188 		 */
1189 		if (!history_begin ||
1190 		    !cycle_between(history_begin->cycles,
1191 				   system_counterval.cycles, cycles) ||
1192 		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1193 			return -EINVAL;
1194 		partial_history_cycles = cycles - system_counterval.cycles;
1195 		total_history_cycles = cycles - history_begin->cycles;
1196 		discontinuity =
1197 			history_begin->clock_was_set_seq != clock_was_set_seq;
1198 
1199 		ret = adjust_historical_crosststamp(history_begin,
1200 						    partial_history_cycles,
1201 						    total_history_cycles,
1202 						    discontinuity, xtstamp);
1203 		if (ret)
1204 			return ret;
1205 	}
1206 
1207 	return 0;
1208 }
1209 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1210 
1211 /**
1212  * do_settimeofday64 - Sets the time of day.
1213  * @ts:     pointer to the timespec64 variable containing the new time
1214  *
1215  * Sets the time of day to the new time and update NTP and notify hrtimers
1216  */
1217 int do_settimeofday64(const struct timespec64 *ts)
1218 {
1219 	struct timekeeper *tk = &tk_core.timekeeper;
1220 	struct timespec64 ts_delta, xt;
1221 	unsigned long flags;
1222 	int ret = 0;
1223 
1224 	if (!timespec64_valid_strict(ts))
1225 		return -EINVAL;
1226 
1227 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1228 	write_seqcount_begin(&tk_core.seq);
1229 
1230 	timekeeping_forward_now(tk);
1231 
1232 	xt = tk_xtime(tk);
1233 	ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1234 	ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1235 
1236 	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1237 		ret = -EINVAL;
1238 		goto out;
1239 	}
1240 
1241 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1242 
1243 	tk_set_xtime(tk, ts);
1244 out:
1245 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1246 
1247 	write_seqcount_end(&tk_core.seq);
1248 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1249 
1250 	/* signal hrtimers about time change */
1251 	clock_was_set();
1252 
1253 	return ret;
1254 }
1255 EXPORT_SYMBOL(do_settimeofday64);
1256 
1257 /**
1258  * timekeeping_inject_offset - Adds or subtracts from the current time.
1259  * @tv:		pointer to the timespec variable containing the offset
1260  *
1261  * Adds or subtracts an offset value from the current time.
1262  */
1263 static int timekeeping_inject_offset(const struct timespec64 *ts)
1264 {
1265 	struct timekeeper *tk = &tk_core.timekeeper;
1266 	unsigned long flags;
1267 	struct timespec64 tmp;
1268 	int ret = 0;
1269 
1270 	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1271 		return -EINVAL;
1272 
1273 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1274 	write_seqcount_begin(&tk_core.seq);
1275 
1276 	timekeeping_forward_now(tk);
1277 
1278 	/* Make sure the proposed value is valid */
1279 	tmp = timespec64_add(tk_xtime(tk), *ts);
1280 	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1281 	    !timespec64_valid_strict(&tmp)) {
1282 		ret = -EINVAL;
1283 		goto error;
1284 	}
1285 
1286 	tk_xtime_add(tk, ts);
1287 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1288 
1289 error: /* even if we error out, we forwarded the time, so call update */
1290 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1291 
1292 	write_seqcount_end(&tk_core.seq);
1293 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1294 
1295 	/* signal hrtimers about time change */
1296 	clock_was_set();
1297 
1298 	return ret;
1299 }
1300 
1301 /*
1302  * Indicates if there is an offset between the system clock and the hardware
1303  * clock/persistent clock/rtc.
1304  */
1305 int persistent_clock_is_local;
1306 
1307 /*
1308  * Adjust the time obtained from the CMOS to be UTC time instead of
1309  * local time.
1310  *
1311  * This is ugly, but preferable to the alternatives.  Otherwise we
1312  * would either need to write a program to do it in /etc/rc (and risk
1313  * confusion if the program gets run more than once; it would also be
1314  * hard to make the program warp the clock precisely n hours)  or
1315  * compile in the timezone information into the kernel.  Bad, bad....
1316  *
1317  *						- TYT, 1992-01-01
1318  *
1319  * The best thing to do is to keep the CMOS clock in universal time (UTC)
1320  * as real UNIX machines always do it. This avoids all headaches about
1321  * daylight saving times and warping kernel clocks.
1322  */
1323 void timekeeping_warp_clock(void)
1324 {
1325 	if (sys_tz.tz_minuteswest != 0) {
1326 		struct timespec64 adjust;
1327 
1328 		persistent_clock_is_local = 1;
1329 		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1330 		adjust.tv_nsec = 0;
1331 		timekeeping_inject_offset(&adjust);
1332 	}
1333 }
1334 
1335 /**
1336  * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1337  *
1338  */
1339 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1340 {
1341 	tk->tai_offset = tai_offset;
1342 	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1343 }
1344 
1345 /**
1346  * change_clocksource - Swaps clocksources if a new one is available
1347  *
1348  * Accumulates current time interval and initializes new clocksource
1349  */
1350 static int change_clocksource(void *data)
1351 {
1352 	struct timekeeper *tk = &tk_core.timekeeper;
1353 	struct clocksource *new, *old;
1354 	unsigned long flags;
1355 
1356 	new = (struct clocksource *) data;
1357 
1358 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1359 	write_seqcount_begin(&tk_core.seq);
1360 
1361 	timekeeping_forward_now(tk);
1362 	/*
1363 	 * If the cs is in module, get a module reference. Succeeds
1364 	 * for built-in code (owner == NULL) as well.
1365 	 */
1366 	if (try_module_get(new->owner)) {
1367 		if (!new->enable || new->enable(new) == 0) {
1368 			old = tk->tkr_mono.clock;
1369 			tk_setup_internals(tk, new);
1370 			if (old->disable)
1371 				old->disable(old);
1372 			module_put(old->owner);
1373 		} else {
1374 			module_put(new->owner);
1375 		}
1376 	}
1377 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1378 
1379 	write_seqcount_end(&tk_core.seq);
1380 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1381 
1382 	return 0;
1383 }
1384 
1385 /**
1386  * timekeeping_notify - Install a new clock source
1387  * @clock:		pointer to the clock source
1388  *
1389  * This function is called from clocksource.c after a new, better clock
1390  * source has been registered. The caller holds the clocksource_mutex.
1391  */
1392 int timekeeping_notify(struct clocksource *clock)
1393 {
1394 	struct timekeeper *tk = &tk_core.timekeeper;
1395 
1396 	if (tk->tkr_mono.clock == clock)
1397 		return 0;
1398 	stop_machine(change_clocksource, clock, NULL);
1399 	tick_clock_notify();
1400 	return tk->tkr_mono.clock == clock ? 0 : -1;
1401 }
1402 
1403 /**
1404  * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1405  * @ts:		pointer to the timespec64 to be set
1406  *
1407  * Returns the raw monotonic time (completely un-modified by ntp)
1408  */
1409 void ktime_get_raw_ts64(struct timespec64 *ts)
1410 {
1411 	struct timekeeper *tk = &tk_core.timekeeper;
1412 	unsigned long seq;
1413 	u64 nsecs;
1414 
1415 	do {
1416 		seq = read_seqcount_begin(&tk_core.seq);
1417 		ts->tv_sec = tk->raw_sec;
1418 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1419 
1420 	} while (read_seqcount_retry(&tk_core.seq, seq));
1421 
1422 	ts->tv_nsec = 0;
1423 	timespec64_add_ns(ts, nsecs);
1424 }
1425 EXPORT_SYMBOL(ktime_get_raw_ts64);
1426 
1427 
1428 /**
1429  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1430  */
1431 int timekeeping_valid_for_hres(void)
1432 {
1433 	struct timekeeper *tk = &tk_core.timekeeper;
1434 	unsigned long seq;
1435 	int ret;
1436 
1437 	do {
1438 		seq = read_seqcount_begin(&tk_core.seq);
1439 
1440 		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1441 
1442 	} while (read_seqcount_retry(&tk_core.seq, seq));
1443 
1444 	return ret;
1445 }
1446 
1447 /**
1448  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1449  */
1450 u64 timekeeping_max_deferment(void)
1451 {
1452 	struct timekeeper *tk = &tk_core.timekeeper;
1453 	unsigned long seq;
1454 	u64 ret;
1455 
1456 	do {
1457 		seq = read_seqcount_begin(&tk_core.seq);
1458 
1459 		ret = tk->tkr_mono.clock->max_idle_ns;
1460 
1461 	} while (read_seqcount_retry(&tk_core.seq, seq));
1462 
1463 	return ret;
1464 }
1465 
1466 /**
1467  * read_persistent_clock64 -  Return time from the persistent clock.
1468  *
1469  * Weak dummy function for arches that do not yet support it.
1470  * Reads the time from the battery backed persistent clock.
1471  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1472  *
1473  *  XXX - Do be sure to remove it once all arches implement it.
1474  */
1475 void __weak read_persistent_clock64(struct timespec64 *ts)
1476 {
1477 	ts->tv_sec = 0;
1478 	ts->tv_nsec = 0;
1479 }
1480 
1481 /**
1482  * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1483  *                                        from the boot.
1484  *
1485  * Weak dummy function for arches that do not yet support it.
1486  * wall_time	- current time as returned by persistent clock
1487  * boot_offset	- offset that is defined as wall_time - boot_time
1488  * The default function calculates offset based on the current value of
1489  * local_clock(). This way architectures that support sched_clock() but don't
1490  * support dedicated boot time clock will provide the best estimate of the
1491  * boot time.
1492  */
1493 void __weak __init
1494 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1495 				     struct timespec64 *boot_offset)
1496 {
1497 	read_persistent_clock64(wall_time);
1498 	*boot_offset = ns_to_timespec64(local_clock());
1499 }
1500 
1501 /*
1502  * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1503  *
1504  * The flag starts of false and is only set when a suspend reaches
1505  * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1506  * timekeeper clocksource is not stopping across suspend and has been
1507  * used to update sleep time. If the timekeeper clocksource has stopped
1508  * then the flag stays true and is used by the RTC resume code to decide
1509  * whether sleeptime must be injected and if so the flag gets false then.
1510  *
1511  * If a suspend fails before reaching timekeeping_resume() then the flag
1512  * stays false and prevents erroneous sleeptime injection.
1513  */
1514 static bool suspend_timing_needed;
1515 
1516 /* Flag for if there is a persistent clock on this platform */
1517 static bool persistent_clock_exists;
1518 
1519 /*
1520  * timekeeping_init - Initializes the clocksource and common timekeeping values
1521  */
1522 void __init timekeeping_init(void)
1523 {
1524 	struct timespec64 wall_time, boot_offset, wall_to_mono;
1525 	struct timekeeper *tk = &tk_core.timekeeper;
1526 	struct clocksource *clock;
1527 	unsigned long flags;
1528 
1529 	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1530 	if (timespec64_valid_strict(&wall_time) &&
1531 	    timespec64_to_ns(&wall_time) > 0) {
1532 		persistent_clock_exists = true;
1533 	} else if (timespec64_to_ns(&wall_time) != 0) {
1534 		pr_warn("Persistent clock returned invalid value");
1535 		wall_time = (struct timespec64){0};
1536 	}
1537 
1538 	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1539 		boot_offset = (struct timespec64){0};
1540 
1541 	/*
1542 	 * We want set wall_to_mono, so the following is true:
1543 	 * wall time + wall_to_mono = boot time
1544 	 */
1545 	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1546 
1547 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1548 	write_seqcount_begin(&tk_core.seq);
1549 	ntp_init();
1550 
1551 	clock = clocksource_default_clock();
1552 	if (clock->enable)
1553 		clock->enable(clock);
1554 	tk_setup_internals(tk, clock);
1555 
1556 	tk_set_xtime(tk, &wall_time);
1557 	tk->raw_sec = 0;
1558 
1559 	tk_set_wall_to_mono(tk, wall_to_mono);
1560 
1561 	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1562 
1563 	write_seqcount_end(&tk_core.seq);
1564 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1565 }
1566 
1567 /* time in seconds when suspend began for persistent clock */
1568 static struct timespec64 timekeeping_suspend_time;
1569 
1570 /**
1571  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1572  * @delta: pointer to a timespec delta value
1573  *
1574  * Takes a timespec offset measuring a suspend interval and properly
1575  * adds the sleep offset to the timekeeping variables.
1576  */
1577 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1578 					   const struct timespec64 *delta)
1579 {
1580 	if (!timespec64_valid_strict(delta)) {
1581 		printk_deferred(KERN_WARNING
1582 				"__timekeeping_inject_sleeptime: Invalid "
1583 				"sleep delta value!\n");
1584 		return;
1585 	}
1586 	tk_xtime_add(tk, delta);
1587 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1588 	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1589 	tk_debug_account_sleep_time(delta);
1590 }
1591 
1592 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1593 /**
1594  * We have three kinds of time sources to use for sleep time
1595  * injection, the preference order is:
1596  * 1) non-stop clocksource
1597  * 2) persistent clock (ie: RTC accessible when irqs are off)
1598  * 3) RTC
1599  *
1600  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1601  * If system has neither 1) nor 2), 3) will be used finally.
1602  *
1603  *
1604  * If timekeeping has injected sleeptime via either 1) or 2),
1605  * 3) becomes needless, so in this case we don't need to call
1606  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1607  * means.
1608  */
1609 bool timekeeping_rtc_skipresume(void)
1610 {
1611 	return !suspend_timing_needed;
1612 }
1613 
1614 /**
1615  * 1) can be determined whether to use or not only when doing
1616  * timekeeping_resume() which is invoked after rtc_suspend(),
1617  * so we can't skip rtc_suspend() surely if system has 1).
1618  *
1619  * But if system has 2), 2) will definitely be used, so in this
1620  * case we don't need to call rtc_suspend(), and this is what
1621  * timekeeping_rtc_skipsuspend() means.
1622  */
1623 bool timekeeping_rtc_skipsuspend(void)
1624 {
1625 	return persistent_clock_exists;
1626 }
1627 
1628 /**
1629  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1630  * @delta: pointer to a timespec64 delta value
1631  *
1632  * This hook is for architectures that cannot support read_persistent_clock64
1633  * because their RTC/persistent clock is only accessible when irqs are enabled.
1634  * and also don't have an effective nonstop clocksource.
1635  *
1636  * This function should only be called by rtc_resume(), and allows
1637  * a suspend offset to be injected into the timekeeping values.
1638  */
1639 void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1640 {
1641 	struct timekeeper *tk = &tk_core.timekeeper;
1642 	unsigned long flags;
1643 
1644 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1645 	write_seqcount_begin(&tk_core.seq);
1646 
1647 	suspend_timing_needed = false;
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  */
1666 void 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, nsec;
1673 	bool inject_sleeptime = false;
1674 
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 	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1697 	if (nsec > 0) {
1698 		ts_delta = ns_to_timespec64(nsec);
1699 		inject_sleeptime = true;
1700 	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1701 		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1702 		inject_sleeptime = true;
1703 	}
1704 
1705 	if (inject_sleeptime) {
1706 		suspend_timing_needed = false;
1707 		__timekeeping_inject_sleeptime(tk, &ts_delta);
1708 	}
1709 
1710 	/* Re-base the last cycle value */
1711 	tk->tkr_mono.cycle_last = cycle_now;
1712 	tk->tkr_raw.cycle_last  = cycle_now;
1713 
1714 	tk->ntp_error = 0;
1715 	timekeeping_suspended = 0;
1716 	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1717 	write_seqcount_end(&tk_core.seq);
1718 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1719 
1720 	touch_softlockup_watchdog();
1721 
1722 	tick_resume();
1723 	hrtimers_resume();
1724 }
1725 
1726 int timekeeping_suspend(void)
1727 {
1728 	struct timekeeper *tk = &tk_core.timekeeper;
1729 	unsigned long flags;
1730 	struct timespec64		delta, delta_delta;
1731 	static struct timespec64	old_delta;
1732 	struct clocksource *curr_clock;
1733 	u64 cycle_now;
1734 
1735 	read_persistent_clock64(&timekeeping_suspend_time);
1736 
1737 	/*
1738 	 * On some systems the persistent_clock can not be detected at
1739 	 * timekeeping_init by its return value, so if we see a valid
1740 	 * value returned, update the persistent_clock_exists flag.
1741 	 */
1742 	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1743 		persistent_clock_exists = true;
1744 
1745 	suspend_timing_needed = true;
1746 
1747 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1748 	write_seqcount_begin(&tk_core.seq);
1749 	timekeeping_forward_now(tk);
1750 	timekeeping_suspended = 1;
1751 
1752 	/*
1753 	 * Since we've called forward_now, cycle_last stores the value
1754 	 * just read from the current clocksource. Save this to potentially
1755 	 * use in suspend timing.
1756 	 */
1757 	curr_clock = tk->tkr_mono.clock;
1758 	cycle_now = tk->tkr_mono.cycle_last;
1759 	clocksource_start_suspend_timing(curr_clock, cycle_now);
1760 
1761 	if (persistent_clock_exists) {
1762 		/*
1763 		 * To avoid drift caused by repeated suspend/resumes,
1764 		 * which each can add ~1 second drift error,
1765 		 * try to compensate so the difference in system time
1766 		 * and persistent_clock time stays close to constant.
1767 		 */
1768 		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1769 		delta_delta = timespec64_sub(delta, old_delta);
1770 		if (abs(delta_delta.tv_sec) >= 2) {
1771 			/*
1772 			 * if delta_delta is too large, assume time correction
1773 			 * has occurred and set old_delta to the current delta.
1774 			 */
1775 			old_delta = delta;
1776 		} else {
1777 			/* Otherwise try to adjust old_system to compensate */
1778 			timekeeping_suspend_time =
1779 				timespec64_add(timekeeping_suspend_time, delta_delta);
1780 		}
1781 	}
1782 
1783 	timekeeping_update(tk, TK_MIRROR);
1784 	halt_fast_timekeeper(tk);
1785 	write_seqcount_end(&tk_core.seq);
1786 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1787 
1788 	tick_suspend();
1789 	clocksource_suspend();
1790 	clockevents_suspend();
1791 
1792 	return 0;
1793 }
1794 
1795 /* sysfs resume/suspend bits for timekeeping */
1796 static struct syscore_ops timekeeping_syscore_ops = {
1797 	.resume		= timekeeping_resume,
1798 	.suspend	= timekeeping_suspend,
1799 };
1800 
1801 static int __init timekeeping_init_ops(void)
1802 {
1803 	register_syscore_ops(&timekeeping_syscore_ops);
1804 	return 0;
1805 }
1806 device_initcall(timekeeping_init_ops);
1807 
1808 /*
1809  * Apply a multiplier adjustment to the timekeeper
1810  */
1811 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1812 							 s64 offset,
1813 							 s32 mult_adj)
1814 {
1815 	s64 interval = tk->cycle_interval;
1816 
1817 	if (mult_adj == 0) {
1818 		return;
1819 	} else if (mult_adj == -1) {
1820 		interval = -interval;
1821 		offset = -offset;
1822 	} else if (mult_adj != 1) {
1823 		interval *= mult_adj;
1824 		offset *= mult_adj;
1825 	}
1826 
1827 	/*
1828 	 * So the following can be confusing.
1829 	 *
1830 	 * To keep things simple, lets assume mult_adj == 1 for now.
1831 	 *
1832 	 * When mult_adj != 1, remember that the interval and offset values
1833 	 * have been appropriately scaled so the math is the same.
1834 	 *
1835 	 * The basic idea here is that we're increasing the multiplier
1836 	 * by one, this causes the xtime_interval to be incremented by
1837 	 * one cycle_interval. This is because:
1838 	 *	xtime_interval = cycle_interval * mult
1839 	 * So if mult is being incremented by one:
1840 	 *	xtime_interval = cycle_interval * (mult + 1)
1841 	 * Its the same as:
1842 	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1843 	 * Which can be shortened to:
1844 	 *	xtime_interval += cycle_interval
1845 	 *
1846 	 * So offset stores the non-accumulated cycles. Thus the current
1847 	 * time (in shifted nanoseconds) is:
1848 	 *	now = (offset * adj) + xtime_nsec
1849 	 * Now, even though we're adjusting the clock frequency, we have
1850 	 * to keep time consistent. In other words, we can't jump back
1851 	 * in time, and we also want to avoid jumping forward in time.
1852 	 *
1853 	 * So given the same offset value, we need the time to be the same
1854 	 * both before and after the freq adjustment.
1855 	 *	now = (offset * adj_1) + xtime_nsec_1
1856 	 *	now = (offset * adj_2) + xtime_nsec_2
1857 	 * So:
1858 	 *	(offset * adj_1) + xtime_nsec_1 =
1859 	 *		(offset * adj_2) + xtime_nsec_2
1860 	 * And we know:
1861 	 *	adj_2 = adj_1 + 1
1862 	 * So:
1863 	 *	(offset * adj_1) + xtime_nsec_1 =
1864 	 *		(offset * (adj_1+1)) + xtime_nsec_2
1865 	 *	(offset * adj_1) + xtime_nsec_1 =
1866 	 *		(offset * adj_1) + offset + xtime_nsec_2
1867 	 * Canceling the sides:
1868 	 *	xtime_nsec_1 = offset + xtime_nsec_2
1869 	 * Which gives us:
1870 	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1871 	 * Which simplfies to:
1872 	 *	xtime_nsec -= offset
1873 	 */
1874 	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1875 		/* NTP adjustment caused clocksource mult overflow */
1876 		WARN_ON_ONCE(1);
1877 		return;
1878 	}
1879 
1880 	tk->tkr_mono.mult += mult_adj;
1881 	tk->xtime_interval += interval;
1882 	tk->tkr_mono.xtime_nsec -= offset;
1883 }
1884 
1885 /*
1886  * Adjust the timekeeper's multiplier to the correct frequency
1887  * and also to reduce the accumulated error value.
1888  */
1889 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1890 {
1891 	u32 mult;
1892 
1893 	/*
1894 	 * Determine the multiplier from the current NTP tick length.
1895 	 * Avoid expensive division when the tick length doesn't change.
1896 	 */
1897 	if (likely(tk->ntp_tick == ntp_tick_length())) {
1898 		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1899 	} else {
1900 		tk->ntp_tick = ntp_tick_length();
1901 		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1902 				 tk->xtime_remainder, tk->cycle_interval);
1903 	}
1904 
1905 	/*
1906 	 * If the clock is behind the NTP time, increase the multiplier by 1
1907 	 * to catch up with it. If it's ahead and there was a remainder in the
1908 	 * tick division, the clock will slow down. Otherwise it will stay
1909 	 * ahead until the tick length changes to a non-divisible value.
1910 	 */
1911 	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1912 	mult += tk->ntp_err_mult;
1913 
1914 	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
1915 
1916 	if (unlikely(tk->tkr_mono.clock->maxadj &&
1917 		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1918 			> tk->tkr_mono.clock->maxadj))) {
1919 		printk_once(KERN_WARNING
1920 			"Adjusting %s more than 11%% (%ld vs %ld)\n",
1921 			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1922 			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1923 	}
1924 
1925 	/*
1926 	 * It may be possible that when we entered this function, xtime_nsec
1927 	 * was very small.  Further, if we're slightly speeding the clocksource
1928 	 * in the code above, its possible the required corrective factor to
1929 	 * xtime_nsec could cause it to underflow.
1930 	 *
1931 	 * Now, since we have already accumulated the second and the NTP
1932 	 * subsystem has been notified via second_overflow(), we need to skip
1933 	 * the next update.
1934 	 */
1935 	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1936 		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
1937 							tk->tkr_mono.shift;
1938 		tk->xtime_sec--;
1939 		tk->skip_second_overflow = 1;
1940 	}
1941 }
1942 
1943 /**
1944  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1945  *
1946  * Helper function that accumulates the nsecs greater than a second
1947  * from the xtime_nsec field to the xtime_secs field.
1948  * It also calls into the NTP code to handle leapsecond processing.
1949  *
1950  */
1951 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1952 {
1953 	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1954 	unsigned int clock_set = 0;
1955 
1956 	while (tk->tkr_mono.xtime_nsec >= nsecps) {
1957 		int leap;
1958 
1959 		tk->tkr_mono.xtime_nsec -= nsecps;
1960 		tk->xtime_sec++;
1961 
1962 		/*
1963 		 * Skip NTP update if this second was accumulated before,
1964 		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
1965 		 */
1966 		if (unlikely(tk->skip_second_overflow)) {
1967 			tk->skip_second_overflow = 0;
1968 			continue;
1969 		}
1970 
1971 		/* Figure out if its a leap sec and apply if needed */
1972 		leap = second_overflow(tk->xtime_sec);
1973 		if (unlikely(leap)) {
1974 			struct timespec64 ts;
1975 
1976 			tk->xtime_sec += leap;
1977 
1978 			ts.tv_sec = leap;
1979 			ts.tv_nsec = 0;
1980 			tk_set_wall_to_mono(tk,
1981 				timespec64_sub(tk->wall_to_monotonic, ts));
1982 
1983 			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1984 
1985 			clock_set = TK_CLOCK_WAS_SET;
1986 		}
1987 	}
1988 	return clock_set;
1989 }
1990 
1991 /**
1992  * logarithmic_accumulation - shifted accumulation of cycles
1993  *
1994  * This functions accumulates a shifted interval of cycles into
1995  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1996  * loop.
1997  *
1998  * Returns the unconsumed cycles.
1999  */
2000 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2001 				    u32 shift, unsigned int *clock_set)
2002 {
2003 	u64 interval = tk->cycle_interval << shift;
2004 	u64 snsec_per_sec;
2005 
2006 	/* If the offset is smaller than a shifted interval, do nothing */
2007 	if (offset < interval)
2008 		return offset;
2009 
2010 	/* Accumulate one shifted interval */
2011 	offset -= interval;
2012 	tk->tkr_mono.cycle_last += interval;
2013 	tk->tkr_raw.cycle_last  += interval;
2014 
2015 	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2016 	*clock_set |= accumulate_nsecs_to_secs(tk);
2017 
2018 	/* Accumulate raw time */
2019 	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2020 	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2021 	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2022 		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2023 		tk->raw_sec++;
2024 	}
2025 
2026 	/* Accumulate error between NTP and clock interval */
2027 	tk->ntp_error += tk->ntp_tick << shift;
2028 	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2029 						(tk->ntp_error_shift + shift);
2030 
2031 	return offset;
2032 }
2033 
2034 /*
2035  * timekeeping_advance - Updates the timekeeper to the current time and
2036  * current NTP tick length
2037  */
2038 static void timekeeping_advance(enum timekeeping_adv_mode mode)
2039 {
2040 	struct timekeeper *real_tk = &tk_core.timekeeper;
2041 	struct timekeeper *tk = &shadow_timekeeper;
2042 	u64 offset;
2043 	int shift = 0, maxshift;
2044 	unsigned int clock_set = 0;
2045 	unsigned long flags;
2046 
2047 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2048 
2049 	/* Make sure we're fully resumed: */
2050 	if (unlikely(timekeeping_suspended))
2051 		goto out;
2052 
2053 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2054 	offset = real_tk->cycle_interval;
2055 
2056 	if (mode != TK_ADV_TICK)
2057 		goto out;
2058 #else
2059 	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2060 				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2061 
2062 	/* Check if there's really nothing to do */
2063 	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2064 		goto out;
2065 #endif
2066 
2067 	/* Do some additional sanity checking */
2068 	timekeeping_check_update(tk, offset);
2069 
2070 	/*
2071 	 * With NO_HZ we may have to accumulate many cycle_intervals
2072 	 * (think "ticks") worth of time at once. To do this efficiently,
2073 	 * we calculate the largest doubling multiple of cycle_intervals
2074 	 * that is smaller than the offset.  We then accumulate that
2075 	 * chunk in one go, and then try to consume the next smaller
2076 	 * doubled multiple.
2077 	 */
2078 	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2079 	shift = max(0, shift);
2080 	/* Bound shift to one less than what overflows tick_length */
2081 	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2082 	shift = min(shift, maxshift);
2083 	while (offset >= tk->cycle_interval) {
2084 		offset = logarithmic_accumulation(tk, offset, shift,
2085 							&clock_set);
2086 		if (offset < tk->cycle_interval<<shift)
2087 			shift--;
2088 	}
2089 
2090 	/* Adjust the multiplier to correct NTP error */
2091 	timekeeping_adjust(tk, offset);
2092 
2093 	/*
2094 	 * Finally, make sure that after the rounding
2095 	 * xtime_nsec isn't larger than NSEC_PER_SEC
2096 	 */
2097 	clock_set |= accumulate_nsecs_to_secs(tk);
2098 
2099 	write_seqcount_begin(&tk_core.seq);
2100 	/*
2101 	 * Update the real timekeeper.
2102 	 *
2103 	 * We could avoid this memcpy by switching pointers, but that
2104 	 * requires changes to all other timekeeper usage sites as
2105 	 * well, i.e. move the timekeeper pointer getter into the
2106 	 * spinlocked/seqcount protected sections. And we trade this
2107 	 * memcpy under the tk_core.seq against one before we start
2108 	 * updating.
2109 	 */
2110 	timekeeping_update(tk, clock_set);
2111 	memcpy(real_tk, tk, sizeof(*tk));
2112 	/* The memcpy must come last. Do not put anything here! */
2113 	write_seqcount_end(&tk_core.seq);
2114 out:
2115 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2116 	if (clock_set)
2117 		/* Have to call _delayed version, since in irq context*/
2118 		clock_was_set_delayed();
2119 }
2120 
2121 /**
2122  * update_wall_time - Uses the current clocksource to increment the wall time
2123  *
2124  */
2125 void update_wall_time(void)
2126 {
2127 	timekeeping_advance(TK_ADV_TICK);
2128 }
2129 
2130 /**
2131  * getboottime64 - Return the real time of system boot.
2132  * @ts:		pointer to the timespec64 to be set
2133  *
2134  * Returns the wall-time of boot in a timespec64.
2135  *
2136  * This is based on the wall_to_monotonic offset and the total suspend
2137  * time. Calls to settimeofday will affect the value returned (which
2138  * basically means that however wrong your real time clock is at boot time,
2139  * you get the right time here).
2140  */
2141 void getboottime64(struct timespec64 *ts)
2142 {
2143 	struct timekeeper *tk = &tk_core.timekeeper;
2144 	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2145 
2146 	*ts = ktime_to_timespec64(t);
2147 }
2148 EXPORT_SYMBOL_GPL(getboottime64);
2149 
2150 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2151 {
2152 	struct timekeeper *tk = &tk_core.timekeeper;
2153 	unsigned long seq;
2154 
2155 	do {
2156 		seq = read_seqcount_begin(&tk_core.seq);
2157 
2158 		*ts = tk_xtime(tk);
2159 	} while (read_seqcount_retry(&tk_core.seq, seq));
2160 }
2161 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2162 
2163 void ktime_get_coarse_ts64(struct timespec64 *ts)
2164 {
2165 	struct timekeeper *tk = &tk_core.timekeeper;
2166 	struct timespec64 now, mono;
2167 	unsigned long seq;
2168 
2169 	do {
2170 		seq = read_seqcount_begin(&tk_core.seq);
2171 
2172 		now = tk_xtime(tk);
2173 		mono = tk->wall_to_monotonic;
2174 	} while (read_seqcount_retry(&tk_core.seq, seq));
2175 
2176 	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2177 				now.tv_nsec + mono.tv_nsec);
2178 }
2179 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2180 
2181 /*
2182  * Must hold jiffies_lock
2183  */
2184 void do_timer(unsigned long ticks)
2185 {
2186 	jiffies_64 += ticks;
2187 	calc_global_load(ticks);
2188 }
2189 
2190 /**
2191  * ktime_get_update_offsets_now - hrtimer helper
2192  * @cwsseq:	pointer to check and store the clock was set sequence number
2193  * @offs_real:	pointer to storage for monotonic -> realtime offset
2194  * @offs_boot:	pointer to storage for monotonic -> boottime offset
2195  * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2196  *
2197  * Returns current monotonic time and updates the offsets if the
2198  * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2199  * different.
2200  *
2201  * Called from hrtimer_interrupt() or retrigger_next_event()
2202  */
2203 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2204 				     ktime_t *offs_boot, ktime_t *offs_tai)
2205 {
2206 	struct timekeeper *tk = &tk_core.timekeeper;
2207 	unsigned int seq;
2208 	ktime_t base;
2209 	u64 nsecs;
2210 
2211 	do {
2212 		seq = read_seqcount_begin(&tk_core.seq);
2213 
2214 		base = tk->tkr_mono.base;
2215 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2216 		base = ktime_add_ns(base, nsecs);
2217 
2218 		if (*cwsseq != tk->clock_was_set_seq) {
2219 			*cwsseq = tk->clock_was_set_seq;
2220 			*offs_real = tk->offs_real;
2221 			*offs_boot = tk->offs_boot;
2222 			*offs_tai = tk->offs_tai;
2223 		}
2224 
2225 		/* Handle leapsecond insertion adjustments */
2226 		if (unlikely(base >= tk->next_leap_ktime))
2227 			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2228 
2229 	} while (read_seqcount_retry(&tk_core.seq, seq));
2230 
2231 	return base;
2232 }
2233 
2234 /**
2235  * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2236  */
2237 static int timekeeping_validate_timex(const struct timex *txc)
2238 {
2239 	if (txc->modes & ADJ_ADJTIME) {
2240 		/* singleshot must not be used with any other mode bits */
2241 		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2242 			return -EINVAL;
2243 		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2244 		    !capable(CAP_SYS_TIME))
2245 			return -EPERM;
2246 	} else {
2247 		/* In order to modify anything, you gotta be super-user! */
2248 		if (txc->modes && !capable(CAP_SYS_TIME))
2249 			return -EPERM;
2250 		/*
2251 		 * if the quartz is off by more than 10% then
2252 		 * something is VERY wrong!
2253 		 */
2254 		if (txc->modes & ADJ_TICK &&
2255 		    (txc->tick <  900000/USER_HZ ||
2256 		     txc->tick > 1100000/USER_HZ))
2257 			return -EINVAL;
2258 	}
2259 
2260 	if (txc->modes & ADJ_SETOFFSET) {
2261 		/* In order to inject time, you gotta be super-user! */
2262 		if (!capable(CAP_SYS_TIME))
2263 			return -EPERM;
2264 
2265 		/*
2266 		 * Validate if a timespec/timeval used to inject a time
2267 		 * offset is valid.  Offsets can be postive or negative, so
2268 		 * we don't check tv_sec. The value of the timeval/timespec
2269 		 * is the sum of its fields,but *NOTE*:
2270 		 * The field tv_usec/tv_nsec must always be non-negative and
2271 		 * we can't have more nanoseconds/microseconds than a second.
2272 		 */
2273 		if (txc->time.tv_usec < 0)
2274 			return -EINVAL;
2275 
2276 		if (txc->modes & ADJ_NANO) {
2277 			if (txc->time.tv_usec >= NSEC_PER_SEC)
2278 				return -EINVAL;
2279 		} else {
2280 			if (txc->time.tv_usec >= USEC_PER_SEC)
2281 				return -EINVAL;
2282 		}
2283 	}
2284 
2285 	/*
2286 	 * Check for potential multiplication overflows that can
2287 	 * only happen on 64-bit systems:
2288 	 */
2289 	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2290 		if (LLONG_MIN / PPM_SCALE > txc->freq)
2291 			return -EINVAL;
2292 		if (LLONG_MAX / PPM_SCALE < txc->freq)
2293 			return -EINVAL;
2294 	}
2295 
2296 	return 0;
2297 }
2298 
2299 
2300 /**
2301  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2302  */
2303 int do_adjtimex(struct timex *txc)
2304 {
2305 	struct timekeeper *tk = &tk_core.timekeeper;
2306 	unsigned long flags;
2307 	struct timespec64 ts;
2308 	s32 orig_tai, tai;
2309 	int ret;
2310 
2311 	/* Validate the data before disabling interrupts */
2312 	ret = timekeeping_validate_timex(txc);
2313 	if (ret)
2314 		return ret;
2315 
2316 	if (txc->modes & ADJ_SETOFFSET) {
2317 		struct timespec64 delta;
2318 		delta.tv_sec  = txc->time.tv_sec;
2319 		delta.tv_nsec = txc->time.tv_usec;
2320 		if (!(txc->modes & ADJ_NANO))
2321 			delta.tv_nsec *= 1000;
2322 		ret = timekeeping_inject_offset(&delta);
2323 		if (ret)
2324 			return ret;
2325 	}
2326 
2327 	ktime_get_real_ts64(&ts);
2328 
2329 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2330 	write_seqcount_begin(&tk_core.seq);
2331 
2332 	orig_tai = tai = tk->tai_offset;
2333 	ret = __do_adjtimex(txc, &ts, &tai);
2334 
2335 	if (tai != orig_tai) {
2336 		__timekeeping_set_tai_offset(tk, tai);
2337 		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2338 	}
2339 	tk_update_leap_state(tk);
2340 
2341 	write_seqcount_end(&tk_core.seq);
2342 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2343 
2344 	/* Update the multiplier immediately if frequency was set directly */
2345 	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2346 		timekeeping_advance(TK_ADV_FREQ);
2347 
2348 	if (tai != orig_tai)
2349 		clock_was_set();
2350 
2351 	ntp_notify_cmos_timer();
2352 
2353 	return ret;
2354 }
2355 
2356 #ifdef CONFIG_NTP_PPS
2357 /**
2358  * hardpps() - Accessor function to NTP __hardpps function
2359  */
2360 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2361 {
2362 	unsigned long flags;
2363 
2364 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2365 	write_seqcount_begin(&tk_core.seq);
2366 
2367 	__hardpps(phase_ts, raw_ts);
2368 
2369 	write_seqcount_end(&tk_core.seq);
2370 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2371 }
2372 EXPORT_SYMBOL(hardpps);
2373 #endif /* CONFIG_NTP_PPS */
2374 
2375 /**
2376  * xtime_update() - advances the timekeeping infrastructure
2377  * @ticks:	number of ticks, that have elapsed since the last call.
2378  *
2379  * Must be called with interrupts disabled.
2380  */
2381 void xtime_update(unsigned long ticks)
2382 {
2383 	write_seqlock(&jiffies_lock);
2384 	do_timer(ticks);
2385 	write_sequnlock(&jiffies_lock);
2386 	update_wall_time();
2387 }
2388