xref: /openbmc/linux/kernel/time/timekeeping.c (revision 98ddec80)
1 /*
2  *  linux/kernel/time/timekeeping.c
3  *
4  *  Kernel timekeeping code and accessor functions
5  *
6  *  This code was moved from linux/kernel/timer.c.
7  *  Please see that file for copyright and history logs.
8  *
9  */
10 
11 #include <linux/timekeeper_internal.h>
12 #include <linux/module.h>
13 #include <linux/interrupt.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
16 #include <linux/mm.h>
17 #include <linux/nmi.h>
18 #include <linux/sched.h>
19 #include <linux/sched/loadavg.h>
20 #include <linux/syscore_ops.h>
21 #include <linux/clocksource.h>
22 #include <linux/jiffies.h>
23 #include <linux/time.h>
24 #include <linux/tick.h>
25 #include <linux/stop_machine.h>
26 #include <linux/pvclock_gtod.h>
27 #include <linux/compiler.h>
28 
29 #include "tick-internal.h"
30 #include "ntp_internal.h"
31 #include "timekeeping_internal.h"
32 
33 #define TK_CLEAR_NTP		(1 << 0)
34 #define TK_MIRROR		(1 << 1)
35 #define TK_CLOCK_WAS_SET	(1 << 2)
36 
37 /*
38  * The most important data for readout fits into a single 64 byte
39  * cache line.
40  */
41 static struct {
42 	seqcount_t		seq;
43 	struct timekeeper	timekeeper;
44 } tk_core ____cacheline_aligned;
45 
46 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
47 static struct timekeeper shadow_timekeeper;
48 
49 /**
50  * struct tk_fast - NMI safe timekeeper
51  * @seq:	Sequence counter for protecting updates. The lowest bit
52  *		is the index for the tk_read_base array
53  * @base:	tk_read_base array. Access is indexed by the lowest bit of
54  *		@seq.
55  *
56  * See @update_fast_timekeeper() below.
57  */
58 struct tk_fast {
59 	seqcount_t		seq;
60 	struct tk_read_base	base[2];
61 };
62 
63 /* Suspend-time cycles value for halted fast timekeeper. */
64 static u64 cycles_at_suspend;
65 
66 static u64 dummy_clock_read(struct clocksource *cs)
67 {
68 	return cycles_at_suspend;
69 }
70 
71 static struct clocksource dummy_clock = {
72 	.read = dummy_clock_read,
73 };
74 
75 static struct tk_fast tk_fast_mono ____cacheline_aligned = {
76 	.base[0] = { .clock = &dummy_clock, },
77 	.base[1] = { .clock = &dummy_clock, },
78 };
79 
80 static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
81 	.base[0] = { .clock = &dummy_clock, },
82 	.base[1] = { .clock = &dummy_clock, },
83 };
84 
85 /* flag for if timekeeping is suspended */
86 int __read_mostly timekeeping_suspended;
87 
88 static inline void tk_normalize_xtime(struct timekeeper *tk)
89 {
90 	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
91 		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
92 		tk->xtime_sec++;
93 	}
94 	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
95 		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
96 		tk->raw_sec++;
97 	}
98 }
99 
100 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
101 {
102 	struct timespec64 ts;
103 
104 	ts.tv_sec = tk->xtime_sec;
105 	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
106 	return ts;
107 }
108 
109 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
110 {
111 	tk->xtime_sec = ts->tv_sec;
112 	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
113 }
114 
115 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
116 {
117 	tk->xtime_sec += ts->tv_sec;
118 	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
119 	tk_normalize_xtime(tk);
120 }
121 
122 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
123 {
124 	struct timespec64 tmp;
125 
126 	/*
127 	 * Verify consistency of: offset_real = -wall_to_monotonic
128 	 * before modifying anything
129 	 */
130 	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
131 					-tk->wall_to_monotonic.tv_nsec);
132 	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
133 	tk->wall_to_monotonic = wtm;
134 	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
135 	tk->offs_real = timespec64_to_ktime(tmp);
136 	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
137 }
138 
139 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
140 {
141 	tk->offs_boot = ktime_add(tk->offs_boot, delta);
142 }
143 
144 /*
145  * tk_clock_read - atomic clocksource read() helper
146  *
147  * This helper is necessary to use in the read paths because, while the
148  * seqlock ensures we don't return a bad value while structures are updated,
149  * it doesn't protect from potential crashes. There is the possibility that
150  * the tkr's clocksource may change between the read reference, and the
151  * clock reference passed to the read function.  This can cause crashes if
152  * the wrong clocksource is passed to the wrong read function.
153  * This isn't necessary to use when holding the timekeeper_lock or doing
154  * a read of the fast-timekeeper tkrs (which is protected by its own locking
155  * and update logic).
156  */
157 static inline u64 tk_clock_read(struct tk_read_base *tkr)
158 {
159 	struct clocksource *clock = READ_ONCE(tkr->clock);
160 
161 	return clock->read(clock);
162 }
163 
164 #ifdef CONFIG_DEBUG_TIMEKEEPING
165 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
166 
167 static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
168 {
169 
170 	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
171 	const char *name = tk->tkr_mono.clock->name;
172 
173 	if (offset > max_cycles) {
174 		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
175 				offset, name, max_cycles);
176 		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
177 	} else {
178 		if (offset > (max_cycles >> 1)) {
179 			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
180 					offset, name, max_cycles >> 1);
181 			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
182 		}
183 	}
184 
185 	if (tk->underflow_seen) {
186 		if (jiffies - tk->last_warning > WARNING_FREQ) {
187 			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
188 			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
189 			printk_deferred("         Your kernel is probably still fine.\n");
190 			tk->last_warning = jiffies;
191 		}
192 		tk->underflow_seen = 0;
193 	}
194 
195 	if (tk->overflow_seen) {
196 		if (jiffies - tk->last_warning > WARNING_FREQ) {
197 			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
198 			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
199 			printk_deferred("         Your kernel is probably still fine.\n");
200 			tk->last_warning = jiffies;
201 		}
202 		tk->overflow_seen = 0;
203 	}
204 }
205 
206 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
207 {
208 	struct timekeeper *tk = &tk_core.timekeeper;
209 	u64 now, last, mask, max, delta;
210 	unsigned int seq;
211 
212 	/*
213 	 * Since we're called holding a seqlock, the data may shift
214 	 * under us while we're doing the calculation. This can cause
215 	 * false positives, since we'd note a problem but throw the
216 	 * results away. So nest another seqlock here to atomically
217 	 * grab the points we are checking with.
218 	 */
219 	do {
220 		seq = read_seqcount_begin(&tk_core.seq);
221 		now = tk_clock_read(tkr);
222 		last = tkr->cycle_last;
223 		mask = tkr->mask;
224 		max = tkr->clock->max_cycles;
225 	} while (read_seqcount_retry(&tk_core.seq, seq));
226 
227 	delta = clocksource_delta(now, last, mask);
228 
229 	/*
230 	 * Try to catch underflows by checking if we are seeing small
231 	 * mask-relative negative values.
232 	 */
233 	if (unlikely((~delta & mask) < (mask >> 3))) {
234 		tk->underflow_seen = 1;
235 		delta = 0;
236 	}
237 
238 	/* Cap delta value to the max_cycles values to avoid mult overflows */
239 	if (unlikely(delta > max)) {
240 		tk->overflow_seen = 1;
241 		delta = tkr->clock->max_cycles;
242 	}
243 
244 	return delta;
245 }
246 #else
247 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
248 {
249 }
250 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
251 {
252 	u64 cycle_now, delta;
253 
254 	/* read clocksource */
255 	cycle_now = tk_clock_read(tkr);
256 
257 	/* calculate the delta since the last update_wall_time */
258 	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
259 
260 	return delta;
261 }
262 #endif
263 
264 /**
265  * tk_setup_internals - Set up internals to use clocksource clock.
266  *
267  * @tk:		The target timekeeper to setup.
268  * @clock:		Pointer to clocksource.
269  *
270  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
271  * pair and interval request.
272  *
273  * Unless you're the timekeeping code, you should not be using this!
274  */
275 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
276 {
277 	u64 interval;
278 	u64 tmp, ntpinterval;
279 	struct clocksource *old_clock;
280 
281 	++tk->cs_was_changed_seq;
282 	old_clock = tk->tkr_mono.clock;
283 	tk->tkr_mono.clock = clock;
284 	tk->tkr_mono.mask = clock->mask;
285 	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
286 
287 	tk->tkr_raw.clock = clock;
288 	tk->tkr_raw.mask = clock->mask;
289 	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
290 
291 	/* Do the ns -> cycle conversion first, using original mult */
292 	tmp = NTP_INTERVAL_LENGTH;
293 	tmp <<= clock->shift;
294 	ntpinterval = tmp;
295 	tmp += clock->mult/2;
296 	do_div(tmp, clock->mult);
297 	if (tmp == 0)
298 		tmp = 1;
299 
300 	interval = (u64) tmp;
301 	tk->cycle_interval = interval;
302 
303 	/* Go back from cycles -> shifted ns */
304 	tk->xtime_interval = interval * clock->mult;
305 	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
306 	tk->raw_interval = interval * clock->mult;
307 
308 	 /* if changing clocks, convert xtime_nsec shift units */
309 	if (old_clock) {
310 		int shift_change = clock->shift - old_clock->shift;
311 		if (shift_change < 0) {
312 			tk->tkr_mono.xtime_nsec >>= -shift_change;
313 			tk->tkr_raw.xtime_nsec >>= -shift_change;
314 		} else {
315 			tk->tkr_mono.xtime_nsec <<= shift_change;
316 			tk->tkr_raw.xtime_nsec <<= shift_change;
317 		}
318 	}
319 
320 	tk->tkr_mono.shift = clock->shift;
321 	tk->tkr_raw.shift = clock->shift;
322 
323 	tk->ntp_error = 0;
324 	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
325 	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
326 
327 	/*
328 	 * The timekeeper keeps its own mult values for the currently
329 	 * active clocksource. These value will be adjusted via NTP
330 	 * to counteract clock drifting.
331 	 */
332 	tk->tkr_mono.mult = clock->mult;
333 	tk->tkr_raw.mult = clock->mult;
334 	tk->ntp_err_mult = 0;
335 	tk->skip_second_overflow = 0;
336 }
337 
338 /* Timekeeper helper functions. */
339 
340 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
341 static u32 default_arch_gettimeoffset(void) { return 0; }
342 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
343 #else
344 static inline u32 arch_gettimeoffset(void) { return 0; }
345 #endif
346 
347 static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta)
348 {
349 	u64 nsec;
350 
351 	nsec = delta * tkr->mult + tkr->xtime_nsec;
352 	nsec >>= tkr->shift;
353 
354 	/* If arch requires, add in get_arch_timeoffset() */
355 	return nsec + arch_gettimeoffset();
356 }
357 
358 static inline u64 timekeeping_get_ns(struct tk_read_base *tkr)
359 {
360 	u64 delta;
361 
362 	delta = timekeeping_get_delta(tkr);
363 	return timekeeping_delta_to_ns(tkr, delta);
364 }
365 
366 static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles)
367 {
368 	u64 delta;
369 
370 	/* calculate the delta since the last update_wall_time */
371 	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
372 	return timekeeping_delta_to_ns(tkr, delta);
373 }
374 
375 /**
376  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
377  * @tkr: Timekeeping readout base from which we take the update
378  *
379  * We want to use this from any context including NMI and tracing /
380  * instrumenting the timekeeping code itself.
381  *
382  * Employ the latch technique; see @raw_write_seqcount_latch.
383  *
384  * So if a NMI hits the update of base[0] then it will use base[1]
385  * which is still consistent. In the worst case this can result is a
386  * slightly wrong timestamp (a few nanoseconds). See
387  * @ktime_get_mono_fast_ns.
388  */
389 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
390 {
391 	struct tk_read_base *base = tkf->base;
392 
393 	/* Force readers off to base[1] */
394 	raw_write_seqcount_latch(&tkf->seq);
395 
396 	/* Update base[0] */
397 	memcpy(base, tkr, sizeof(*base));
398 
399 	/* Force readers back to base[0] */
400 	raw_write_seqcount_latch(&tkf->seq);
401 
402 	/* Update base[1] */
403 	memcpy(base + 1, base, sizeof(*base));
404 }
405 
406 /**
407  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
408  *
409  * This timestamp is not guaranteed to be monotonic across an update.
410  * The timestamp is calculated by:
411  *
412  *	now = base_mono + clock_delta * slope
413  *
414  * So if the update lowers the slope, readers who are forced to the
415  * not yet updated second array are still using the old steeper slope.
416  *
417  * tmono
418  * ^
419  * |    o  n
420  * |   o n
421  * |  u
422  * | o
423  * |o
424  * |12345678---> reader order
425  *
426  * o = old slope
427  * u = update
428  * n = new slope
429  *
430  * So reader 6 will observe time going backwards versus reader 5.
431  *
432  * While other CPUs are likely to be able observe that, the only way
433  * for a CPU local observation is when an NMI hits in the middle of
434  * the update. Timestamps taken from that NMI context might be ahead
435  * of the following timestamps. Callers need to be aware of that and
436  * deal with it.
437  */
438 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
439 {
440 	struct tk_read_base *tkr;
441 	unsigned int seq;
442 	u64 now;
443 
444 	do {
445 		seq = raw_read_seqcount_latch(&tkf->seq);
446 		tkr = tkf->base + (seq & 0x01);
447 		now = ktime_to_ns(tkr->base);
448 
449 		now += timekeeping_delta_to_ns(tkr,
450 				clocksource_delta(
451 					tk_clock_read(tkr),
452 					tkr->cycle_last,
453 					tkr->mask));
454 	} while (read_seqcount_retry(&tkf->seq, seq));
455 
456 	return now;
457 }
458 
459 u64 ktime_get_mono_fast_ns(void)
460 {
461 	return __ktime_get_fast_ns(&tk_fast_mono);
462 }
463 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
464 
465 u64 ktime_get_raw_fast_ns(void)
466 {
467 	return __ktime_get_fast_ns(&tk_fast_raw);
468 }
469 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
470 
471 /**
472  * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
473  *
474  * To keep it NMI safe since we're accessing from tracing, we're not using a
475  * separate timekeeper with updates to monotonic clock and boot offset
476  * protected with seqlocks. This has the following minor side effects:
477  *
478  * (1) Its possible that a timestamp be taken after the boot offset is updated
479  * but before the timekeeper is updated. If this happens, the new boot offset
480  * is added to the old timekeeping making the clock appear to update slightly
481  * earlier:
482  *    CPU 0                                        CPU 1
483  *    timekeeping_inject_sleeptime64()
484  *    __timekeeping_inject_sleeptime(tk, delta);
485  *                                                 timestamp();
486  *    timekeeping_update(tk, TK_CLEAR_NTP...);
487  *
488  * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
489  * partially updated.  Since the tk->offs_boot update is a rare event, this
490  * should be a rare occurrence which postprocessing should be able to handle.
491  */
492 u64 notrace ktime_get_boot_fast_ns(void)
493 {
494 	struct timekeeper *tk = &tk_core.timekeeper;
495 
496 	return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
497 }
498 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
499 
500 
501 /*
502  * See comment for __ktime_get_fast_ns() vs. timestamp ordering
503  */
504 static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf)
505 {
506 	struct tk_read_base *tkr;
507 	unsigned int seq;
508 	u64 now;
509 
510 	do {
511 		seq = raw_read_seqcount_latch(&tkf->seq);
512 		tkr = tkf->base + (seq & 0x01);
513 		now = ktime_to_ns(tkr->base_real);
514 
515 		now += timekeeping_delta_to_ns(tkr,
516 				clocksource_delta(
517 					tk_clock_read(tkr),
518 					tkr->cycle_last,
519 					tkr->mask));
520 	} while (read_seqcount_retry(&tkf->seq, seq));
521 
522 	return now;
523 }
524 
525 /**
526  * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
527  */
528 u64 ktime_get_real_fast_ns(void)
529 {
530 	return __ktime_get_real_fast_ns(&tk_fast_mono);
531 }
532 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
533 
534 /**
535  * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
536  * @tk: Timekeeper to snapshot.
537  *
538  * It generally is unsafe to access the clocksource after timekeeping has been
539  * suspended, so take a snapshot of the readout base of @tk and use it as the
540  * fast timekeeper's readout base while suspended.  It will return the same
541  * number of cycles every time until timekeeping is resumed at which time the
542  * proper readout base for the fast timekeeper will be restored automatically.
543  */
544 static void halt_fast_timekeeper(struct timekeeper *tk)
545 {
546 	static struct tk_read_base tkr_dummy;
547 	struct tk_read_base *tkr = &tk->tkr_mono;
548 
549 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
550 	cycles_at_suspend = tk_clock_read(tkr);
551 	tkr_dummy.clock = &dummy_clock;
552 	tkr_dummy.base_real = tkr->base + tk->offs_real;
553 	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
554 
555 	tkr = &tk->tkr_raw;
556 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
557 	tkr_dummy.clock = &dummy_clock;
558 	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
559 }
560 
561 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
562 
563 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
564 {
565 	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
566 }
567 
568 /**
569  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
570  */
571 int pvclock_gtod_register_notifier(struct notifier_block *nb)
572 {
573 	struct timekeeper *tk = &tk_core.timekeeper;
574 	unsigned long flags;
575 	int ret;
576 
577 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
578 	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
579 	update_pvclock_gtod(tk, true);
580 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
581 
582 	return ret;
583 }
584 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
585 
586 /**
587  * pvclock_gtod_unregister_notifier - unregister a pvclock
588  * timedata update listener
589  */
590 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
591 {
592 	unsigned long flags;
593 	int ret;
594 
595 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
596 	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
597 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
598 
599 	return ret;
600 }
601 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
602 
603 /*
604  * tk_update_leap_state - helper to update the next_leap_ktime
605  */
606 static inline void tk_update_leap_state(struct timekeeper *tk)
607 {
608 	tk->next_leap_ktime = ntp_get_next_leap();
609 	if (tk->next_leap_ktime != KTIME_MAX)
610 		/* Convert to monotonic time */
611 		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
612 }
613 
614 /*
615  * Update the ktime_t based scalar nsec members of the timekeeper
616  */
617 static inline void tk_update_ktime_data(struct timekeeper *tk)
618 {
619 	u64 seconds;
620 	u32 nsec;
621 
622 	/*
623 	 * The xtime based monotonic readout is:
624 	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
625 	 * The ktime based monotonic readout is:
626 	 *	nsec = base_mono + now();
627 	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
628 	 */
629 	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
630 	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
631 	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
632 
633 	/*
634 	 * The sum of the nanoseconds portions of xtime and
635 	 * wall_to_monotonic can be greater/equal one second. Take
636 	 * this into account before updating tk->ktime_sec.
637 	 */
638 	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
639 	if (nsec >= NSEC_PER_SEC)
640 		seconds++;
641 	tk->ktime_sec = seconds;
642 
643 	/* Update the monotonic raw base */
644 	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
645 }
646 
647 /* must hold timekeeper_lock */
648 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
649 {
650 	if (action & TK_CLEAR_NTP) {
651 		tk->ntp_error = 0;
652 		ntp_clear();
653 	}
654 
655 	tk_update_leap_state(tk);
656 	tk_update_ktime_data(tk);
657 
658 	update_vsyscall(tk);
659 	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
660 
661 	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
662 	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
663 	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
664 
665 	if (action & TK_CLOCK_WAS_SET)
666 		tk->clock_was_set_seq++;
667 	/*
668 	 * The mirroring of the data to the shadow-timekeeper needs
669 	 * to happen last here to ensure we don't over-write the
670 	 * timekeeper structure on the next update with stale data
671 	 */
672 	if (action & TK_MIRROR)
673 		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
674 		       sizeof(tk_core.timekeeper));
675 }
676 
677 /**
678  * timekeeping_forward_now - update clock to the current time
679  *
680  * Forward the current clock to update its state since the last call to
681  * update_wall_time(). This is useful before significant clock changes,
682  * as it avoids having to deal with this time offset explicitly.
683  */
684 static void timekeeping_forward_now(struct timekeeper *tk)
685 {
686 	u64 cycle_now, delta;
687 
688 	cycle_now = tk_clock_read(&tk->tkr_mono);
689 	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
690 	tk->tkr_mono.cycle_last = cycle_now;
691 	tk->tkr_raw.cycle_last  = cycle_now;
692 
693 	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
694 
695 	/* If arch requires, add in get_arch_timeoffset() */
696 	tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
697 
698 
699 	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
700 
701 	/* If arch requires, add in get_arch_timeoffset() */
702 	tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
703 
704 	tk_normalize_xtime(tk);
705 }
706 
707 /**
708  * ktime_get_real_ts64 - Returns the time of day in a timespec64.
709  * @ts:		pointer to the timespec to be set
710  *
711  * Returns the time of day in a timespec64 (WARN if suspended).
712  */
713 void ktime_get_real_ts64(struct timespec64 *ts)
714 {
715 	struct timekeeper *tk = &tk_core.timekeeper;
716 	unsigned long seq;
717 	u64 nsecs;
718 
719 	WARN_ON(timekeeping_suspended);
720 
721 	do {
722 		seq = read_seqcount_begin(&tk_core.seq);
723 
724 		ts->tv_sec = tk->xtime_sec;
725 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
726 
727 	} while (read_seqcount_retry(&tk_core.seq, seq));
728 
729 	ts->tv_nsec = 0;
730 	timespec64_add_ns(ts, nsecs);
731 }
732 EXPORT_SYMBOL(ktime_get_real_ts64);
733 
734 ktime_t ktime_get(void)
735 {
736 	struct timekeeper *tk = &tk_core.timekeeper;
737 	unsigned int seq;
738 	ktime_t base;
739 	u64 nsecs;
740 
741 	WARN_ON(timekeeping_suspended);
742 
743 	do {
744 		seq = read_seqcount_begin(&tk_core.seq);
745 		base = tk->tkr_mono.base;
746 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
747 
748 	} while (read_seqcount_retry(&tk_core.seq, seq));
749 
750 	return ktime_add_ns(base, nsecs);
751 }
752 EXPORT_SYMBOL_GPL(ktime_get);
753 
754 u32 ktime_get_resolution_ns(void)
755 {
756 	struct timekeeper *tk = &tk_core.timekeeper;
757 	unsigned int seq;
758 	u32 nsecs;
759 
760 	WARN_ON(timekeeping_suspended);
761 
762 	do {
763 		seq = read_seqcount_begin(&tk_core.seq);
764 		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
765 	} while (read_seqcount_retry(&tk_core.seq, seq));
766 
767 	return nsecs;
768 }
769 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
770 
771 static ktime_t *offsets[TK_OFFS_MAX] = {
772 	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
773 	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
774 	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
775 };
776 
777 ktime_t ktime_get_with_offset(enum tk_offsets offs)
778 {
779 	struct timekeeper *tk = &tk_core.timekeeper;
780 	unsigned int seq;
781 	ktime_t base, *offset = offsets[offs];
782 	u64 nsecs;
783 
784 	WARN_ON(timekeeping_suspended);
785 
786 	do {
787 		seq = read_seqcount_begin(&tk_core.seq);
788 		base = ktime_add(tk->tkr_mono.base, *offset);
789 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
790 
791 	} while (read_seqcount_retry(&tk_core.seq, seq));
792 
793 	return ktime_add_ns(base, nsecs);
794 
795 }
796 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
797 
798 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
799 {
800 	struct timekeeper *tk = &tk_core.timekeeper;
801 	unsigned int seq;
802 	ktime_t base, *offset = offsets[offs];
803 
804 	WARN_ON(timekeeping_suspended);
805 
806 	do {
807 		seq = read_seqcount_begin(&tk_core.seq);
808 		base = ktime_add(tk->tkr_mono.base, *offset);
809 
810 	} while (read_seqcount_retry(&tk_core.seq, seq));
811 
812 	return base;
813 
814 }
815 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
816 
817 /**
818  * ktime_mono_to_any() - convert mononotic time to any other time
819  * @tmono:	time to convert.
820  * @offs:	which offset to use
821  */
822 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
823 {
824 	ktime_t *offset = offsets[offs];
825 	unsigned long seq;
826 	ktime_t tconv;
827 
828 	do {
829 		seq = read_seqcount_begin(&tk_core.seq);
830 		tconv = ktime_add(tmono, *offset);
831 	} while (read_seqcount_retry(&tk_core.seq, seq));
832 
833 	return tconv;
834 }
835 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
836 
837 /**
838  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
839  */
840 ktime_t ktime_get_raw(void)
841 {
842 	struct timekeeper *tk = &tk_core.timekeeper;
843 	unsigned int seq;
844 	ktime_t base;
845 	u64 nsecs;
846 
847 	do {
848 		seq = read_seqcount_begin(&tk_core.seq);
849 		base = tk->tkr_raw.base;
850 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
851 
852 	} while (read_seqcount_retry(&tk_core.seq, seq));
853 
854 	return ktime_add_ns(base, nsecs);
855 }
856 EXPORT_SYMBOL_GPL(ktime_get_raw);
857 
858 /**
859  * ktime_get_ts64 - get the monotonic clock in timespec64 format
860  * @ts:		pointer to timespec variable
861  *
862  * The function calculates the monotonic clock from the realtime
863  * clock and the wall_to_monotonic offset and stores the result
864  * in normalized timespec64 format in the variable pointed to by @ts.
865  */
866 void ktime_get_ts64(struct timespec64 *ts)
867 {
868 	struct timekeeper *tk = &tk_core.timekeeper;
869 	struct timespec64 tomono;
870 	unsigned int seq;
871 	u64 nsec;
872 
873 	WARN_ON(timekeeping_suspended);
874 
875 	do {
876 		seq = read_seqcount_begin(&tk_core.seq);
877 		ts->tv_sec = tk->xtime_sec;
878 		nsec = timekeeping_get_ns(&tk->tkr_mono);
879 		tomono = tk->wall_to_monotonic;
880 
881 	} while (read_seqcount_retry(&tk_core.seq, seq));
882 
883 	ts->tv_sec += tomono.tv_sec;
884 	ts->tv_nsec = 0;
885 	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
886 }
887 EXPORT_SYMBOL_GPL(ktime_get_ts64);
888 
889 /**
890  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
891  *
892  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
893  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
894  * works on both 32 and 64 bit systems. On 32 bit systems the readout
895  * covers ~136 years of uptime which should be enough to prevent
896  * premature wrap arounds.
897  */
898 time64_t ktime_get_seconds(void)
899 {
900 	struct timekeeper *tk = &tk_core.timekeeper;
901 
902 	WARN_ON(timekeeping_suspended);
903 	return tk->ktime_sec;
904 }
905 EXPORT_SYMBOL_GPL(ktime_get_seconds);
906 
907 /**
908  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
909  *
910  * Returns the wall clock seconds since 1970. This replaces the
911  * get_seconds() interface which is not y2038 safe on 32bit systems.
912  *
913  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
914  * 32bit systems the access must be protected with the sequence
915  * counter to provide "atomic" access to the 64bit tk->xtime_sec
916  * value.
917  */
918 time64_t ktime_get_real_seconds(void)
919 {
920 	struct timekeeper *tk = &tk_core.timekeeper;
921 	time64_t seconds;
922 	unsigned int seq;
923 
924 	if (IS_ENABLED(CONFIG_64BIT))
925 		return tk->xtime_sec;
926 
927 	do {
928 		seq = read_seqcount_begin(&tk_core.seq);
929 		seconds = tk->xtime_sec;
930 
931 	} while (read_seqcount_retry(&tk_core.seq, seq));
932 
933 	return seconds;
934 }
935 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
936 
937 /**
938  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
939  * but without the sequence counter protect. This internal function
940  * is called just when timekeeping lock is already held.
941  */
942 time64_t __ktime_get_real_seconds(void)
943 {
944 	struct timekeeper *tk = &tk_core.timekeeper;
945 
946 	return tk->xtime_sec;
947 }
948 
949 /**
950  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
951  * @systime_snapshot:	pointer to struct receiving the system time snapshot
952  */
953 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
954 {
955 	struct timekeeper *tk = &tk_core.timekeeper;
956 	unsigned long seq;
957 	ktime_t base_raw;
958 	ktime_t base_real;
959 	u64 nsec_raw;
960 	u64 nsec_real;
961 	u64 now;
962 
963 	WARN_ON_ONCE(timekeeping_suspended);
964 
965 	do {
966 		seq = read_seqcount_begin(&tk_core.seq);
967 		now = tk_clock_read(&tk->tkr_mono);
968 		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
969 		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
970 		base_real = ktime_add(tk->tkr_mono.base,
971 				      tk_core.timekeeper.offs_real);
972 		base_raw = tk->tkr_raw.base;
973 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
974 		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
975 	} while (read_seqcount_retry(&tk_core.seq, seq));
976 
977 	systime_snapshot->cycles = now;
978 	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
979 	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
980 }
981 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
982 
983 /* Scale base by mult/div checking for overflow */
984 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
985 {
986 	u64 tmp, rem;
987 
988 	tmp = div64_u64_rem(*base, div, &rem);
989 
990 	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
991 	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
992 		return -EOVERFLOW;
993 	tmp *= mult;
994 	rem *= mult;
995 
996 	do_div(rem, div);
997 	*base = tmp + rem;
998 	return 0;
999 }
1000 
1001 /**
1002  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1003  * @history:			Snapshot representing start of history
1004  * @partial_history_cycles:	Cycle offset into history (fractional part)
1005  * @total_history_cycles:	Total history length in cycles
1006  * @discontinuity:		True indicates clock was set on history period
1007  * @ts:				Cross timestamp that should be adjusted using
1008  *	partial/total ratio
1009  *
1010  * Helper function used by get_device_system_crosststamp() to correct the
1011  * crosstimestamp corresponding to the start of the current interval to the
1012  * system counter value (timestamp point) provided by the driver. The
1013  * total_history_* quantities are the total history starting at the provided
1014  * reference point and ending at the start of the current interval. The cycle
1015  * count between the driver timestamp point and the start of the current
1016  * interval is partial_history_cycles.
1017  */
1018 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1019 					 u64 partial_history_cycles,
1020 					 u64 total_history_cycles,
1021 					 bool discontinuity,
1022 					 struct system_device_crosststamp *ts)
1023 {
1024 	struct timekeeper *tk = &tk_core.timekeeper;
1025 	u64 corr_raw, corr_real;
1026 	bool interp_forward;
1027 	int ret;
1028 
1029 	if (total_history_cycles == 0 || partial_history_cycles == 0)
1030 		return 0;
1031 
1032 	/* Interpolate shortest distance from beginning or end of history */
1033 	interp_forward = partial_history_cycles > total_history_cycles / 2;
1034 	partial_history_cycles = interp_forward ?
1035 		total_history_cycles - partial_history_cycles :
1036 		partial_history_cycles;
1037 
1038 	/*
1039 	 * Scale the monotonic raw time delta by:
1040 	 *	partial_history_cycles / total_history_cycles
1041 	 */
1042 	corr_raw = (u64)ktime_to_ns(
1043 		ktime_sub(ts->sys_monoraw, history->raw));
1044 	ret = scale64_check_overflow(partial_history_cycles,
1045 				     total_history_cycles, &corr_raw);
1046 	if (ret)
1047 		return ret;
1048 
1049 	/*
1050 	 * If there is a discontinuity in the history, scale monotonic raw
1051 	 *	correction by:
1052 	 *	mult(real)/mult(raw) yielding the realtime correction
1053 	 * Otherwise, calculate the realtime correction similar to monotonic
1054 	 *	raw calculation
1055 	 */
1056 	if (discontinuity) {
1057 		corr_real = mul_u64_u32_div
1058 			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1059 	} else {
1060 		corr_real = (u64)ktime_to_ns(
1061 			ktime_sub(ts->sys_realtime, history->real));
1062 		ret = scale64_check_overflow(partial_history_cycles,
1063 					     total_history_cycles, &corr_real);
1064 		if (ret)
1065 			return ret;
1066 	}
1067 
1068 	/* Fixup monotonic raw and real time time values */
1069 	if (interp_forward) {
1070 		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1071 		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1072 	} else {
1073 		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1074 		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1075 	}
1076 
1077 	return 0;
1078 }
1079 
1080 /*
1081  * cycle_between - true if test occurs chronologically between before and after
1082  */
1083 static bool cycle_between(u64 before, u64 test, u64 after)
1084 {
1085 	if (test > before && test < after)
1086 		return true;
1087 	if (test < before && before > after)
1088 		return true;
1089 	return false;
1090 }
1091 
1092 /**
1093  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1094  * @get_time_fn:	Callback to get simultaneous device time and
1095  *	system counter from the device driver
1096  * @ctx:		Context passed to get_time_fn()
1097  * @history_begin:	Historical reference point used to interpolate system
1098  *	time when counter provided by the driver is before the current interval
1099  * @xtstamp:		Receives simultaneously captured system and device time
1100  *
1101  * Reads a timestamp from a device and correlates it to system time
1102  */
1103 int get_device_system_crosststamp(int (*get_time_fn)
1104 				  (ktime_t *device_time,
1105 				   struct system_counterval_t *sys_counterval,
1106 				   void *ctx),
1107 				  void *ctx,
1108 				  struct system_time_snapshot *history_begin,
1109 				  struct system_device_crosststamp *xtstamp)
1110 {
1111 	struct system_counterval_t system_counterval;
1112 	struct timekeeper *tk = &tk_core.timekeeper;
1113 	u64 cycles, now, interval_start;
1114 	unsigned int clock_was_set_seq = 0;
1115 	ktime_t base_real, base_raw;
1116 	u64 nsec_real, nsec_raw;
1117 	u8 cs_was_changed_seq;
1118 	unsigned long seq;
1119 	bool do_interp;
1120 	int ret;
1121 
1122 	do {
1123 		seq = read_seqcount_begin(&tk_core.seq);
1124 		/*
1125 		 * Try to synchronously capture device time and a system
1126 		 * counter value calling back into the device driver
1127 		 */
1128 		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1129 		if (ret)
1130 			return ret;
1131 
1132 		/*
1133 		 * Verify that the clocksource associated with the captured
1134 		 * system counter value is the same as the currently installed
1135 		 * timekeeper clocksource
1136 		 */
1137 		if (tk->tkr_mono.clock != system_counterval.cs)
1138 			return -ENODEV;
1139 		cycles = system_counterval.cycles;
1140 
1141 		/*
1142 		 * Check whether the system counter value provided by the
1143 		 * device driver is on the current timekeeping interval.
1144 		 */
1145 		now = tk_clock_read(&tk->tkr_mono);
1146 		interval_start = tk->tkr_mono.cycle_last;
1147 		if (!cycle_between(interval_start, cycles, now)) {
1148 			clock_was_set_seq = tk->clock_was_set_seq;
1149 			cs_was_changed_seq = tk->cs_was_changed_seq;
1150 			cycles = interval_start;
1151 			do_interp = true;
1152 		} else {
1153 			do_interp = false;
1154 		}
1155 
1156 		base_real = ktime_add(tk->tkr_mono.base,
1157 				      tk_core.timekeeper.offs_real);
1158 		base_raw = tk->tkr_raw.base;
1159 
1160 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1161 						     system_counterval.cycles);
1162 		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1163 						    system_counterval.cycles);
1164 	} while (read_seqcount_retry(&tk_core.seq, seq));
1165 
1166 	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1167 	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1168 
1169 	/*
1170 	 * Interpolate if necessary, adjusting back from the start of the
1171 	 * current interval
1172 	 */
1173 	if (do_interp) {
1174 		u64 partial_history_cycles, total_history_cycles;
1175 		bool discontinuity;
1176 
1177 		/*
1178 		 * Check that the counter value occurs after the provided
1179 		 * history reference and that the history doesn't cross a
1180 		 * clocksource change
1181 		 */
1182 		if (!history_begin ||
1183 		    !cycle_between(history_begin->cycles,
1184 				   system_counterval.cycles, cycles) ||
1185 		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1186 			return -EINVAL;
1187 		partial_history_cycles = cycles - system_counterval.cycles;
1188 		total_history_cycles = cycles - history_begin->cycles;
1189 		discontinuity =
1190 			history_begin->clock_was_set_seq != clock_was_set_seq;
1191 
1192 		ret = adjust_historical_crosststamp(history_begin,
1193 						    partial_history_cycles,
1194 						    total_history_cycles,
1195 						    discontinuity, xtstamp);
1196 		if (ret)
1197 			return ret;
1198 	}
1199 
1200 	return 0;
1201 }
1202 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1203 
1204 /**
1205  * do_gettimeofday - Returns the time of day in a timeval
1206  * @tv:		pointer to the timeval to be set
1207  *
1208  * NOTE: Users should be converted to using getnstimeofday()
1209  */
1210 void do_gettimeofday(struct timeval *tv)
1211 {
1212 	struct timespec64 now;
1213 
1214 	getnstimeofday64(&now);
1215 	tv->tv_sec = now.tv_sec;
1216 	tv->tv_usec = now.tv_nsec/1000;
1217 }
1218 EXPORT_SYMBOL(do_gettimeofday);
1219 
1220 /**
1221  * do_settimeofday64 - Sets the time of day.
1222  * @ts:     pointer to the timespec64 variable containing the new time
1223  *
1224  * Sets the time of day to the new time and update NTP and notify hrtimers
1225  */
1226 int do_settimeofday64(const struct timespec64 *ts)
1227 {
1228 	struct timekeeper *tk = &tk_core.timekeeper;
1229 	struct timespec64 ts_delta, xt;
1230 	unsigned long flags;
1231 	int ret = 0;
1232 
1233 	if (!timespec64_valid_strict(ts))
1234 		return -EINVAL;
1235 
1236 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1237 	write_seqcount_begin(&tk_core.seq);
1238 
1239 	timekeeping_forward_now(tk);
1240 
1241 	xt = tk_xtime(tk);
1242 	ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1243 	ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1244 
1245 	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1246 		ret = -EINVAL;
1247 		goto out;
1248 	}
1249 
1250 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1251 
1252 	tk_set_xtime(tk, ts);
1253 out:
1254 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1255 
1256 	write_seqcount_end(&tk_core.seq);
1257 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1258 
1259 	/* signal hrtimers about time change */
1260 	clock_was_set();
1261 
1262 	return ret;
1263 }
1264 EXPORT_SYMBOL(do_settimeofday64);
1265 
1266 /**
1267  * timekeeping_inject_offset - Adds or subtracts from the current time.
1268  * @tv:		pointer to the timespec variable containing the offset
1269  *
1270  * Adds or subtracts an offset value from the current time.
1271  */
1272 static int timekeeping_inject_offset(struct timespec64 *ts)
1273 {
1274 	struct timekeeper *tk = &tk_core.timekeeper;
1275 	unsigned long flags;
1276 	struct timespec64 tmp;
1277 	int ret = 0;
1278 
1279 	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1280 		return -EINVAL;
1281 
1282 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1283 	write_seqcount_begin(&tk_core.seq);
1284 
1285 	timekeeping_forward_now(tk);
1286 
1287 	/* Make sure the proposed value is valid */
1288 	tmp = timespec64_add(tk_xtime(tk), *ts);
1289 	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1290 	    !timespec64_valid_strict(&tmp)) {
1291 		ret = -EINVAL;
1292 		goto error;
1293 	}
1294 
1295 	tk_xtime_add(tk, ts);
1296 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1297 
1298 error: /* even if we error out, we forwarded the time, so call update */
1299 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1300 
1301 	write_seqcount_end(&tk_core.seq);
1302 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1303 
1304 	/* signal hrtimers about time change */
1305 	clock_was_set();
1306 
1307 	return ret;
1308 }
1309 
1310 /*
1311  * Indicates if there is an offset between the system clock and the hardware
1312  * clock/persistent clock/rtc.
1313  */
1314 int persistent_clock_is_local;
1315 
1316 /*
1317  * Adjust the time obtained from the CMOS to be UTC time instead of
1318  * local time.
1319  *
1320  * This is ugly, but preferable to the alternatives.  Otherwise we
1321  * would either need to write a program to do it in /etc/rc (and risk
1322  * confusion if the program gets run more than once; it would also be
1323  * hard to make the program warp the clock precisely n hours)  or
1324  * compile in the timezone information into the kernel.  Bad, bad....
1325  *
1326  *						- TYT, 1992-01-01
1327  *
1328  * The best thing to do is to keep the CMOS clock in universal time (UTC)
1329  * as real UNIX machines always do it. This avoids all headaches about
1330  * daylight saving times and warping kernel clocks.
1331  */
1332 void timekeeping_warp_clock(void)
1333 {
1334 	if (sys_tz.tz_minuteswest != 0) {
1335 		struct timespec64 adjust;
1336 
1337 		persistent_clock_is_local = 1;
1338 		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1339 		adjust.tv_nsec = 0;
1340 		timekeeping_inject_offset(&adjust);
1341 	}
1342 }
1343 
1344 /**
1345  * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1346  *
1347  */
1348 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1349 {
1350 	tk->tai_offset = tai_offset;
1351 	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1352 }
1353 
1354 /**
1355  * change_clocksource - Swaps clocksources if a new one is available
1356  *
1357  * Accumulates current time interval and initializes new clocksource
1358  */
1359 static int change_clocksource(void *data)
1360 {
1361 	struct timekeeper *tk = &tk_core.timekeeper;
1362 	struct clocksource *new, *old;
1363 	unsigned long flags;
1364 
1365 	new = (struct clocksource *) data;
1366 
1367 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1368 	write_seqcount_begin(&tk_core.seq);
1369 
1370 	timekeeping_forward_now(tk);
1371 	/*
1372 	 * If the cs is in module, get a module reference. Succeeds
1373 	 * for built-in code (owner == NULL) as well.
1374 	 */
1375 	if (try_module_get(new->owner)) {
1376 		if (!new->enable || new->enable(new) == 0) {
1377 			old = tk->tkr_mono.clock;
1378 			tk_setup_internals(tk, new);
1379 			if (old->disable)
1380 				old->disable(old);
1381 			module_put(old->owner);
1382 		} else {
1383 			module_put(new->owner);
1384 		}
1385 	}
1386 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1387 
1388 	write_seqcount_end(&tk_core.seq);
1389 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1390 
1391 	return 0;
1392 }
1393 
1394 /**
1395  * timekeeping_notify - Install a new clock source
1396  * @clock:		pointer to the clock source
1397  *
1398  * This function is called from clocksource.c after a new, better clock
1399  * source has been registered. The caller holds the clocksource_mutex.
1400  */
1401 int timekeeping_notify(struct clocksource *clock)
1402 {
1403 	struct timekeeper *tk = &tk_core.timekeeper;
1404 
1405 	if (tk->tkr_mono.clock == clock)
1406 		return 0;
1407 	stop_machine(change_clocksource, clock, NULL);
1408 	tick_clock_notify();
1409 	return tk->tkr_mono.clock == clock ? 0 : -1;
1410 }
1411 
1412 /**
1413  * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1414  * @ts:		pointer to the timespec64 to be set
1415  *
1416  * Returns the raw monotonic time (completely un-modified by ntp)
1417  */
1418 void ktime_get_raw_ts64(struct timespec64 *ts)
1419 {
1420 	struct timekeeper *tk = &tk_core.timekeeper;
1421 	unsigned long seq;
1422 	u64 nsecs;
1423 
1424 	do {
1425 		seq = read_seqcount_begin(&tk_core.seq);
1426 		ts->tv_sec = tk->raw_sec;
1427 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1428 
1429 	} while (read_seqcount_retry(&tk_core.seq, seq));
1430 
1431 	ts->tv_nsec = 0;
1432 	timespec64_add_ns(ts, nsecs);
1433 }
1434 EXPORT_SYMBOL(ktime_get_raw_ts64);
1435 
1436 
1437 /**
1438  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1439  */
1440 int timekeeping_valid_for_hres(void)
1441 {
1442 	struct timekeeper *tk = &tk_core.timekeeper;
1443 	unsigned long seq;
1444 	int ret;
1445 
1446 	do {
1447 		seq = read_seqcount_begin(&tk_core.seq);
1448 
1449 		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1450 
1451 	} while (read_seqcount_retry(&tk_core.seq, seq));
1452 
1453 	return ret;
1454 }
1455 
1456 /**
1457  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1458  */
1459 u64 timekeeping_max_deferment(void)
1460 {
1461 	struct timekeeper *tk = &tk_core.timekeeper;
1462 	unsigned long seq;
1463 	u64 ret;
1464 
1465 	do {
1466 		seq = read_seqcount_begin(&tk_core.seq);
1467 
1468 		ret = tk->tkr_mono.clock->max_idle_ns;
1469 
1470 	} while (read_seqcount_retry(&tk_core.seq, seq));
1471 
1472 	return ret;
1473 }
1474 
1475 /**
1476  * read_persistent_clock -  Return time from the persistent clock.
1477  *
1478  * Weak dummy function for arches that do not yet support it.
1479  * Reads the time from the battery backed persistent clock.
1480  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1481  *
1482  *  XXX - Do be sure to remove it once all arches implement it.
1483  */
1484 void __weak read_persistent_clock(struct timespec *ts)
1485 {
1486 	ts->tv_sec = 0;
1487 	ts->tv_nsec = 0;
1488 }
1489 
1490 void __weak read_persistent_clock64(struct timespec64 *ts64)
1491 {
1492 	struct timespec ts;
1493 
1494 	read_persistent_clock(&ts);
1495 	*ts64 = timespec_to_timespec64(ts);
1496 }
1497 
1498 /**
1499  * read_boot_clock64 -  Return time of the system start.
1500  *
1501  * Weak dummy function for arches that do not yet support it.
1502  * Function to read the exact time the system has been started.
1503  * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1504  *
1505  *  XXX - Do be sure to remove it once all arches implement it.
1506  */
1507 void __weak read_boot_clock64(struct timespec64 *ts)
1508 {
1509 	ts->tv_sec = 0;
1510 	ts->tv_nsec = 0;
1511 }
1512 
1513 /* Flag for if timekeeping_resume() has injected sleeptime */
1514 static bool sleeptime_injected;
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 timekeeper *tk = &tk_core.timekeeper;
1525 	struct clocksource *clock;
1526 	unsigned long flags;
1527 	struct timespec64 now, boot, tmp;
1528 
1529 	read_persistent_clock64(&now);
1530 	if (!timespec64_valid_strict(&now)) {
1531 		pr_warn("WARNING: Persistent clock returned invalid value!\n"
1532 			"         Check your CMOS/BIOS settings.\n");
1533 		now.tv_sec = 0;
1534 		now.tv_nsec = 0;
1535 	} else if (now.tv_sec || now.tv_nsec)
1536 		persistent_clock_exists = true;
1537 
1538 	read_boot_clock64(&boot);
1539 	if (!timespec64_valid_strict(&boot)) {
1540 		pr_warn("WARNING: Boot clock returned invalid value!\n"
1541 			"         Check your CMOS/BIOS settings.\n");
1542 		boot.tv_sec = 0;
1543 		boot.tv_nsec = 0;
1544 	}
1545 
1546 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1547 	write_seqcount_begin(&tk_core.seq);
1548 	ntp_init();
1549 
1550 	clock = clocksource_default_clock();
1551 	if (clock->enable)
1552 		clock->enable(clock);
1553 	tk_setup_internals(tk, clock);
1554 
1555 	tk_set_xtime(tk, &now);
1556 	tk->raw_sec = 0;
1557 	if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1558 		boot = tk_xtime(tk);
1559 
1560 	set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1561 	tk_set_wall_to_mono(tk, tmp);
1562 
1563 	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1564 
1565 	write_seqcount_end(&tk_core.seq);
1566 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1567 }
1568 
1569 /* time in seconds when suspend began for persistent clock */
1570 static struct timespec64 timekeeping_suspend_time;
1571 
1572 /**
1573  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1574  * @delta: pointer to a timespec delta value
1575  *
1576  * Takes a timespec offset measuring a suspend interval and properly
1577  * adds the sleep offset to the timekeeping variables.
1578  */
1579 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1580 					   struct timespec64 *delta)
1581 {
1582 	if (!timespec64_valid_strict(delta)) {
1583 		printk_deferred(KERN_WARNING
1584 				"__timekeeping_inject_sleeptime: Invalid "
1585 				"sleep delta value!\n");
1586 		return;
1587 	}
1588 	tk_xtime_add(tk, delta);
1589 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1590 	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1591 	tk_debug_account_sleep_time(delta);
1592 }
1593 
1594 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1595 /**
1596  * We have three kinds of time sources to use for sleep time
1597  * injection, the preference order is:
1598  * 1) non-stop clocksource
1599  * 2) persistent clock (ie: RTC accessible when irqs are off)
1600  * 3) RTC
1601  *
1602  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1603  * If system has neither 1) nor 2), 3) will be used finally.
1604  *
1605  *
1606  * If timekeeping has injected sleeptime via either 1) or 2),
1607  * 3) becomes needless, so in this case we don't need to call
1608  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1609  * means.
1610  */
1611 bool timekeeping_rtc_skipresume(void)
1612 {
1613 	return sleeptime_injected;
1614 }
1615 
1616 /**
1617  * 1) can be determined whether to use or not only when doing
1618  * timekeeping_resume() which is invoked after rtc_suspend(),
1619  * so we can't skip rtc_suspend() surely if system has 1).
1620  *
1621  * But if system has 2), 2) will definitely be used, so in this
1622  * case we don't need to call rtc_suspend(), and this is what
1623  * timekeeping_rtc_skipsuspend() means.
1624  */
1625 bool timekeeping_rtc_skipsuspend(void)
1626 {
1627 	return persistent_clock_exists;
1628 }
1629 
1630 /**
1631  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1632  * @delta: pointer to a timespec64 delta value
1633  *
1634  * This hook is for architectures that cannot support read_persistent_clock64
1635  * because their RTC/persistent clock is only accessible when irqs are enabled.
1636  * and also don't have an effective nonstop clocksource.
1637  *
1638  * This function should only be called by rtc_resume(), and allows
1639  * a suspend offset to be injected into the timekeeping values.
1640  */
1641 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1642 {
1643 	struct timekeeper *tk = &tk_core.timekeeper;
1644 	unsigned long flags;
1645 
1646 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1647 	write_seqcount_begin(&tk_core.seq);
1648 
1649 	timekeeping_forward_now(tk);
1650 
1651 	__timekeeping_inject_sleeptime(tk, delta);
1652 
1653 	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1654 
1655 	write_seqcount_end(&tk_core.seq);
1656 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1657 
1658 	/* signal hrtimers about time change */
1659 	clock_was_set();
1660 }
1661 #endif
1662 
1663 /**
1664  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1665  */
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;
1673 
1674 	sleeptime_injected = false;
1675 	read_persistent_clock64(&ts_new);
1676 
1677 	clockevents_resume();
1678 	clocksource_resume();
1679 
1680 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1681 	write_seqcount_begin(&tk_core.seq);
1682 
1683 	/*
1684 	 * After system resumes, we need to calculate the suspended time and
1685 	 * compensate it for the OS time. There are 3 sources that could be
1686 	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1687 	 * device.
1688 	 *
1689 	 * One specific platform may have 1 or 2 or all of them, and the
1690 	 * preference will be:
1691 	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1692 	 * The less preferred source will only be tried if there is no better
1693 	 * usable source. The rtc part is handled separately in rtc core code.
1694 	 */
1695 	cycle_now = tk_clock_read(&tk->tkr_mono);
1696 	if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1697 		cycle_now > tk->tkr_mono.cycle_last) {
1698 		u64 nsec, cyc_delta;
1699 
1700 		cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1701 					      tk->tkr_mono.mask);
1702 		nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift);
1703 		ts_delta = ns_to_timespec64(nsec);
1704 		sleeptime_injected = true;
1705 	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1706 		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1707 		sleeptime_injected = true;
1708 	}
1709 
1710 	if (sleeptime_injected)
1711 		__timekeeping_inject_sleeptime(tk, &ts_delta);
1712 
1713 	/* Re-base the last cycle value */
1714 	tk->tkr_mono.cycle_last = cycle_now;
1715 	tk->tkr_raw.cycle_last  = cycle_now;
1716 
1717 	tk->ntp_error = 0;
1718 	timekeeping_suspended = 0;
1719 	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1720 	write_seqcount_end(&tk_core.seq);
1721 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1722 
1723 	touch_softlockup_watchdog();
1724 
1725 	tick_resume();
1726 	hrtimers_resume();
1727 }
1728 
1729 int timekeeping_suspend(void)
1730 {
1731 	struct timekeeper *tk = &tk_core.timekeeper;
1732 	unsigned long flags;
1733 	struct timespec64		delta, delta_delta;
1734 	static struct timespec64	old_delta;
1735 
1736 	read_persistent_clock64(&timekeeping_suspend_time);
1737 
1738 	/*
1739 	 * On some systems the persistent_clock can not be detected at
1740 	 * timekeeping_init by its return value, so if we see a valid
1741 	 * value returned, update the persistent_clock_exists flag.
1742 	 */
1743 	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1744 		persistent_clock_exists = true;
1745 
1746 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1747 	write_seqcount_begin(&tk_core.seq);
1748 	timekeeping_forward_now(tk);
1749 	timekeeping_suspended = 1;
1750 
1751 	if (persistent_clock_exists) {
1752 		/*
1753 		 * To avoid drift caused by repeated suspend/resumes,
1754 		 * which each can add ~1 second drift error,
1755 		 * try to compensate so the difference in system time
1756 		 * and persistent_clock time stays close to constant.
1757 		 */
1758 		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1759 		delta_delta = timespec64_sub(delta, old_delta);
1760 		if (abs(delta_delta.tv_sec) >= 2) {
1761 			/*
1762 			 * if delta_delta is too large, assume time correction
1763 			 * has occurred and set old_delta to the current delta.
1764 			 */
1765 			old_delta = delta;
1766 		} else {
1767 			/* Otherwise try to adjust old_system to compensate */
1768 			timekeeping_suspend_time =
1769 				timespec64_add(timekeeping_suspend_time, delta_delta);
1770 		}
1771 	}
1772 
1773 	timekeeping_update(tk, TK_MIRROR);
1774 	halt_fast_timekeeper(tk);
1775 	write_seqcount_end(&tk_core.seq);
1776 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1777 
1778 	tick_suspend();
1779 	clocksource_suspend();
1780 	clockevents_suspend();
1781 
1782 	return 0;
1783 }
1784 
1785 /* sysfs resume/suspend bits for timekeeping */
1786 static struct syscore_ops timekeeping_syscore_ops = {
1787 	.resume		= timekeeping_resume,
1788 	.suspend	= timekeeping_suspend,
1789 };
1790 
1791 static int __init timekeeping_init_ops(void)
1792 {
1793 	register_syscore_ops(&timekeeping_syscore_ops);
1794 	return 0;
1795 }
1796 device_initcall(timekeeping_init_ops);
1797 
1798 /*
1799  * Apply a multiplier adjustment to the timekeeper
1800  */
1801 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1802 							 s64 offset,
1803 							 s32 mult_adj)
1804 {
1805 	s64 interval = tk->cycle_interval;
1806 
1807 	if (mult_adj == 0) {
1808 		return;
1809 	} else if (mult_adj == -1) {
1810 		interval = -interval;
1811 		offset = -offset;
1812 	} else if (mult_adj != 1) {
1813 		interval *= mult_adj;
1814 		offset *= mult_adj;
1815 	}
1816 
1817 	/*
1818 	 * So the following can be confusing.
1819 	 *
1820 	 * To keep things simple, lets assume mult_adj == 1 for now.
1821 	 *
1822 	 * When mult_adj != 1, remember that the interval and offset values
1823 	 * have been appropriately scaled so the math is the same.
1824 	 *
1825 	 * The basic idea here is that we're increasing the multiplier
1826 	 * by one, this causes the xtime_interval to be incremented by
1827 	 * one cycle_interval. This is because:
1828 	 *	xtime_interval = cycle_interval * mult
1829 	 * So if mult is being incremented by one:
1830 	 *	xtime_interval = cycle_interval * (mult + 1)
1831 	 * Its the same as:
1832 	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1833 	 * Which can be shortened to:
1834 	 *	xtime_interval += cycle_interval
1835 	 *
1836 	 * So offset stores the non-accumulated cycles. Thus the current
1837 	 * time (in shifted nanoseconds) is:
1838 	 *	now = (offset * adj) + xtime_nsec
1839 	 * Now, even though we're adjusting the clock frequency, we have
1840 	 * to keep time consistent. In other words, we can't jump back
1841 	 * in time, and we also want to avoid jumping forward in time.
1842 	 *
1843 	 * So given the same offset value, we need the time to be the same
1844 	 * both before and after the freq adjustment.
1845 	 *	now = (offset * adj_1) + xtime_nsec_1
1846 	 *	now = (offset * adj_2) + xtime_nsec_2
1847 	 * So:
1848 	 *	(offset * adj_1) + xtime_nsec_1 =
1849 	 *		(offset * adj_2) + xtime_nsec_2
1850 	 * And we know:
1851 	 *	adj_2 = adj_1 + 1
1852 	 * So:
1853 	 *	(offset * adj_1) + xtime_nsec_1 =
1854 	 *		(offset * (adj_1+1)) + xtime_nsec_2
1855 	 *	(offset * adj_1) + xtime_nsec_1 =
1856 	 *		(offset * adj_1) + offset + xtime_nsec_2
1857 	 * Canceling the sides:
1858 	 *	xtime_nsec_1 = offset + xtime_nsec_2
1859 	 * Which gives us:
1860 	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1861 	 * Which simplfies to:
1862 	 *	xtime_nsec -= offset
1863 	 */
1864 	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1865 		/* NTP adjustment caused clocksource mult overflow */
1866 		WARN_ON_ONCE(1);
1867 		return;
1868 	}
1869 
1870 	tk->tkr_mono.mult += mult_adj;
1871 	tk->xtime_interval += interval;
1872 	tk->tkr_mono.xtime_nsec -= offset;
1873 }
1874 
1875 /*
1876  * Adjust the timekeeper's multiplier to the correct frequency
1877  * and also to reduce the accumulated error value.
1878  */
1879 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1880 {
1881 	u32 mult;
1882 
1883 	/*
1884 	 * Determine the multiplier from the current NTP tick length.
1885 	 * Avoid expensive division when the tick length doesn't change.
1886 	 */
1887 	if (likely(tk->ntp_tick == ntp_tick_length())) {
1888 		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1889 	} else {
1890 		tk->ntp_tick = ntp_tick_length();
1891 		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1892 				 tk->xtime_remainder, tk->cycle_interval);
1893 	}
1894 
1895 	/*
1896 	 * If the clock is behind the NTP time, increase the multiplier by 1
1897 	 * to catch up with it. If it's ahead and there was a remainder in the
1898 	 * tick division, the clock will slow down. Otherwise it will stay
1899 	 * ahead until the tick length changes to a non-divisible value.
1900 	 */
1901 	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1902 	mult += tk->ntp_err_mult;
1903 
1904 	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
1905 
1906 	if (unlikely(tk->tkr_mono.clock->maxadj &&
1907 		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1908 			> tk->tkr_mono.clock->maxadj))) {
1909 		printk_once(KERN_WARNING
1910 			"Adjusting %s more than 11%% (%ld vs %ld)\n",
1911 			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1912 			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1913 	}
1914 
1915 	/*
1916 	 * It may be possible that when we entered this function, xtime_nsec
1917 	 * was very small.  Further, if we're slightly speeding the clocksource
1918 	 * in the code above, its possible the required corrective factor to
1919 	 * xtime_nsec could cause it to underflow.
1920 	 *
1921 	 * Now, since we have already accumulated the second and the NTP
1922 	 * subsystem has been notified via second_overflow(), we need to skip
1923 	 * the next update.
1924 	 */
1925 	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1926 		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
1927 							tk->tkr_mono.shift;
1928 		tk->xtime_sec--;
1929 		tk->skip_second_overflow = 1;
1930 	}
1931 }
1932 
1933 /**
1934  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1935  *
1936  * Helper function that accumulates the nsecs greater than a second
1937  * from the xtime_nsec field to the xtime_secs field.
1938  * It also calls into the NTP code to handle leapsecond processing.
1939  *
1940  */
1941 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1942 {
1943 	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1944 	unsigned int clock_set = 0;
1945 
1946 	while (tk->tkr_mono.xtime_nsec >= nsecps) {
1947 		int leap;
1948 
1949 		tk->tkr_mono.xtime_nsec -= nsecps;
1950 		tk->xtime_sec++;
1951 
1952 		/*
1953 		 * Skip NTP update if this second was accumulated before,
1954 		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
1955 		 */
1956 		if (unlikely(tk->skip_second_overflow)) {
1957 			tk->skip_second_overflow = 0;
1958 			continue;
1959 		}
1960 
1961 		/* Figure out if its a leap sec and apply if needed */
1962 		leap = second_overflow(tk->xtime_sec);
1963 		if (unlikely(leap)) {
1964 			struct timespec64 ts;
1965 
1966 			tk->xtime_sec += leap;
1967 
1968 			ts.tv_sec = leap;
1969 			ts.tv_nsec = 0;
1970 			tk_set_wall_to_mono(tk,
1971 				timespec64_sub(tk->wall_to_monotonic, ts));
1972 
1973 			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1974 
1975 			clock_set = TK_CLOCK_WAS_SET;
1976 		}
1977 	}
1978 	return clock_set;
1979 }
1980 
1981 /**
1982  * logarithmic_accumulation - shifted accumulation of cycles
1983  *
1984  * This functions accumulates a shifted interval of cycles into
1985  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1986  * loop.
1987  *
1988  * Returns the unconsumed cycles.
1989  */
1990 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
1991 				    u32 shift, unsigned int *clock_set)
1992 {
1993 	u64 interval = tk->cycle_interval << shift;
1994 	u64 snsec_per_sec;
1995 
1996 	/* If the offset is smaller than a shifted interval, do nothing */
1997 	if (offset < interval)
1998 		return offset;
1999 
2000 	/* Accumulate one shifted interval */
2001 	offset -= interval;
2002 	tk->tkr_mono.cycle_last += interval;
2003 	tk->tkr_raw.cycle_last  += interval;
2004 
2005 	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2006 	*clock_set |= accumulate_nsecs_to_secs(tk);
2007 
2008 	/* Accumulate raw time */
2009 	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2010 	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2011 	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2012 		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2013 		tk->raw_sec++;
2014 	}
2015 
2016 	/* Accumulate error between NTP and clock interval */
2017 	tk->ntp_error += tk->ntp_tick << shift;
2018 	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2019 						(tk->ntp_error_shift + shift);
2020 
2021 	return offset;
2022 }
2023 
2024 /**
2025  * update_wall_time - Uses the current clocksource to increment the wall time
2026  *
2027  */
2028 void update_wall_time(void)
2029 {
2030 	struct timekeeper *real_tk = &tk_core.timekeeper;
2031 	struct timekeeper *tk = &shadow_timekeeper;
2032 	u64 offset;
2033 	int shift = 0, maxshift;
2034 	unsigned int clock_set = 0;
2035 	unsigned long flags;
2036 
2037 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2038 
2039 	/* Make sure we're fully resumed: */
2040 	if (unlikely(timekeeping_suspended))
2041 		goto out;
2042 
2043 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2044 	offset = real_tk->cycle_interval;
2045 #else
2046 	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2047 				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2048 #endif
2049 
2050 	/* Check if there's really nothing to do */
2051 	if (offset < real_tk->cycle_interval)
2052 		goto out;
2053 
2054 	/* Do some additional sanity checking */
2055 	timekeeping_check_update(tk, offset);
2056 
2057 	/*
2058 	 * With NO_HZ we may have to accumulate many cycle_intervals
2059 	 * (think "ticks") worth of time at once. To do this efficiently,
2060 	 * we calculate the largest doubling multiple of cycle_intervals
2061 	 * that is smaller than the offset.  We then accumulate that
2062 	 * chunk in one go, and then try to consume the next smaller
2063 	 * doubled multiple.
2064 	 */
2065 	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2066 	shift = max(0, shift);
2067 	/* Bound shift to one less than what overflows tick_length */
2068 	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2069 	shift = min(shift, maxshift);
2070 	while (offset >= tk->cycle_interval) {
2071 		offset = logarithmic_accumulation(tk, offset, shift,
2072 							&clock_set);
2073 		if (offset < tk->cycle_interval<<shift)
2074 			shift--;
2075 	}
2076 
2077 	/* Adjust the multiplier to correct NTP error */
2078 	timekeeping_adjust(tk, offset);
2079 
2080 	/*
2081 	 * Finally, make sure that after the rounding
2082 	 * xtime_nsec isn't larger than NSEC_PER_SEC
2083 	 */
2084 	clock_set |= accumulate_nsecs_to_secs(tk);
2085 
2086 	write_seqcount_begin(&tk_core.seq);
2087 	/*
2088 	 * Update the real timekeeper.
2089 	 *
2090 	 * We could avoid this memcpy by switching pointers, but that
2091 	 * requires changes to all other timekeeper usage sites as
2092 	 * well, i.e. move the timekeeper pointer getter into the
2093 	 * spinlocked/seqcount protected sections. And we trade this
2094 	 * memcpy under the tk_core.seq against one before we start
2095 	 * updating.
2096 	 */
2097 	timekeeping_update(tk, clock_set);
2098 	memcpy(real_tk, tk, sizeof(*tk));
2099 	/* The memcpy must come last. Do not put anything here! */
2100 	write_seqcount_end(&tk_core.seq);
2101 out:
2102 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2103 	if (clock_set)
2104 		/* Have to call _delayed version, since in irq context*/
2105 		clock_was_set_delayed();
2106 }
2107 
2108 /**
2109  * getboottime64 - Return the real time of system boot.
2110  * @ts:		pointer to the timespec64 to be set
2111  *
2112  * Returns the wall-time of boot in a timespec64.
2113  *
2114  * This is based on the wall_to_monotonic offset and the total suspend
2115  * time. Calls to settimeofday will affect the value returned (which
2116  * basically means that however wrong your real time clock is at boot time,
2117  * you get the right time here).
2118  */
2119 void getboottime64(struct timespec64 *ts)
2120 {
2121 	struct timekeeper *tk = &tk_core.timekeeper;
2122 	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2123 
2124 	*ts = ktime_to_timespec64(t);
2125 }
2126 EXPORT_SYMBOL_GPL(getboottime64);
2127 
2128 unsigned long get_seconds(void)
2129 {
2130 	struct timekeeper *tk = &tk_core.timekeeper;
2131 
2132 	return tk->xtime_sec;
2133 }
2134 EXPORT_SYMBOL(get_seconds);
2135 
2136 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2137 {
2138 	struct timekeeper *tk = &tk_core.timekeeper;
2139 	unsigned long seq;
2140 
2141 	do {
2142 		seq = read_seqcount_begin(&tk_core.seq);
2143 
2144 		*ts = tk_xtime(tk);
2145 	} while (read_seqcount_retry(&tk_core.seq, seq));
2146 }
2147 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2148 
2149 void ktime_get_coarse_ts64(struct timespec64 *ts)
2150 {
2151 	struct timekeeper *tk = &tk_core.timekeeper;
2152 	struct timespec64 now, mono;
2153 	unsigned long seq;
2154 
2155 	do {
2156 		seq = read_seqcount_begin(&tk_core.seq);
2157 
2158 		now = tk_xtime(tk);
2159 		mono = tk->wall_to_monotonic;
2160 	} while (read_seqcount_retry(&tk_core.seq, seq));
2161 
2162 	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2163 				now.tv_nsec + mono.tv_nsec);
2164 }
2165 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2166 
2167 /*
2168  * Must hold jiffies_lock
2169  */
2170 void do_timer(unsigned long ticks)
2171 {
2172 	jiffies_64 += ticks;
2173 	calc_global_load(ticks);
2174 }
2175 
2176 /**
2177  * ktime_get_update_offsets_now - hrtimer helper
2178  * @cwsseq:	pointer to check and store the clock was set sequence number
2179  * @offs_real:	pointer to storage for monotonic -> realtime offset
2180  * @offs_boot:	pointer to storage for monotonic -> boottime offset
2181  * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2182  *
2183  * Returns current monotonic time and updates the offsets if the
2184  * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2185  * different.
2186  *
2187  * Called from hrtimer_interrupt() or retrigger_next_event()
2188  */
2189 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2190 				     ktime_t *offs_boot, ktime_t *offs_tai)
2191 {
2192 	struct timekeeper *tk = &tk_core.timekeeper;
2193 	unsigned int seq;
2194 	ktime_t base;
2195 	u64 nsecs;
2196 
2197 	do {
2198 		seq = read_seqcount_begin(&tk_core.seq);
2199 
2200 		base = tk->tkr_mono.base;
2201 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2202 		base = ktime_add_ns(base, nsecs);
2203 
2204 		if (*cwsseq != tk->clock_was_set_seq) {
2205 			*cwsseq = tk->clock_was_set_seq;
2206 			*offs_real = tk->offs_real;
2207 			*offs_boot = tk->offs_boot;
2208 			*offs_tai = tk->offs_tai;
2209 		}
2210 
2211 		/* Handle leapsecond insertion adjustments */
2212 		if (unlikely(base >= tk->next_leap_ktime))
2213 			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2214 
2215 	} while (read_seqcount_retry(&tk_core.seq, seq));
2216 
2217 	return base;
2218 }
2219 
2220 /**
2221  * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2222  */
2223 static int timekeeping_validate_timex(struct timex *txc)
2224 {
2225 	if (txc->modes & ADJ_ADJTIME) {
2226 		/* singleshot must not be used with any other mode bits */
2227 		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2228 			return -EINVAL;
2229 		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2230 		    !capable(CAP_SYS_TIME))
2231 			return -EPERM;
2232 	} else {
2233 		/* In order to modify anything, you gotta be super-user! */
2234 		if (txc->modes && !capable(CAP_SYS_TIME))
2235 			return -EPERM;
2236 		/*
2237 		 * if the quartz is off by more than 10% then
2238 		 * something is VERY wrong!
2239 		 */
2240 		if (txc->modes & ADJ_TICK &&
2241 		    (txc->tick <  900000/USER_HZ ||
2242 		     txc->tick > 1100000/USER_HZ))
2243 			return -EINVAL;
2244 	}
2245 
2246 	if (txc->modes & ADJ_SETOFFSET) {
2247 		/* In order to inject time, you gotta be super-user! */
2248 		if (!capable(CAP_SYS_TIME))
2249 			return -EPERM;
2250 
2251 		/*
2252 		 * Validate if a timespec/timeval used to inject a time
2253 		 * offset is valid.  Offsets can be postive or negative, so
2254 		 * we don't check tv_sec. The value of the timeval/timespec
2255 		 * is the sum of its fields,but *NOTE*:
2256 		 * The field tv_usec/tv_nsec must always be non-negative and
2257 		 * we can't have more nanoseconds/microseconds than a second.
2258 		 */
2259 		if (txc->time.tv_usec < 0)
2260 			return -EINVAL;
2261 
2262 		if (txc->modes & ADJ_NANO) {
2263 			if (txc->time.tv_usec >= NSEC_PER_SEC)
2264 				return -EINVAL;
2265 		} else {
2266 			if (txc->time.tv_usec >= USEC_PER_SEC)
2267 				return -EINVAL;
2268 		}
2269 	}
2270 
2271 	/*
2272 	 * Check for potential multiplication overflows that can
2273 	 * only happen on 64-bit systems:
2274 	 */
2275 	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2276 		if (LLONG_MIN / PPM_SCALE > txc->freq)
2277 			return -EINVAL;
2278 		if (LLONG_MAX / PPM_SCALE < txc->freq)
2279 			return -EINVAL;
2280 	}
2281 
2282 	return 0;
2283 }
2284 
2285 
2286 /**
2287  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2288  */
2289 int do_adjtimex(struct timex *txc)
2290 {
2291 	struct timekeeper *tk = &tk_core.timekeeper;
2292 	unsigned long flags;
2293 	struct timespec64 ts;
2294 	s32 orig_tai, tai;
2295 	int ret;
2296 
2297 	/* Validate the data before disabling interrupts */
2298 	ret = timekeeping_validate_timex(txc);
2299 	if (ret)
2300 		return ret;
2301 
2302 	if (txc->modes & ADJ_SETOFFSET) {
2303 		struct timespec64 delta;
2304 		delta.tv_sec  = txc->time.tv_sec;
2305 		delta.tv_nsec = txc->time.tv_usec;
2306 		if (!(txc->modes & ADJ_NANO))
2307 			delta.tv_nsec *= 1000;
2308 		ret = timekeeping_inject_offset(&delta);
2309 		if (ret)
2310 			return ret;
2311 	}
2312 
2313 	getnstimeofday64(&ts);
2314 
2315 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2316 	write_seqcount_begin(&tk_core.seq);
2317 
2318 	orig_tai = tai = tk->tai_offset;
2319 	ret = __do_adjtimex(txc, &ts, &tai);
2320 
2321 	if (tai != orig_tai) {
2322 		__timekeeping_set_tai_offset(tk, tai);
2323 		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2324 	}
2325 	tk_update_leap_state(tk);
2326 
2327 	write_seqcount_end(&tk_core.seq);
2328 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2329 
2330 	if (tai != orig_tai)
2331 		clock_was_set();
2332 
2333 	ntp_notify_cmos_timer();
2334 
2335 	return ret;
2336 }
2337 
2338 #ifdef CONFIG_NTP_PPS
2339 /**
2340  * hardpps() - Accessor function to NTP __hardpps function
2341  */
2342 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2343 {
2344 	unsigned long flags;
2345 
2346 	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2347 	write_seqcount_begin(&tk_core.seq);
2348 
2349 	__hardpps(phase_ts, raw_ts);
2350 
2351 	write_seqcount_end(&tk_core.seq);
2352 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2353 }
2354 EXPORT_SYMBOL(hardpps);
2355 #endif /* CONFIG_NTP_PPS */
2356 
2357 /**
2358  * xtime_update() - advances the timekeeping infrastructure
2359  * @ticks:	number of ticks, that have elapsed since the last call.
2360  *
2361  * Must be called with interrupts disabled.
2362  */
2363 void xtime_update(unsigned long ticks)
2364 {
2365 	write_seqlock(&jiffies_lock);
2366 	do_timer(ticks);
2367 	write_sequnlock(&jiffies_lock);
2368 	update_wall_time();
2369 }
2370