1// SPDX-License-Identifier: GPL-2.0
2/*
3 *  Kernel timekeeping code and accessor functions. Based on code from
4 *  timer.c, moved in commit 8524070b7982.
5 */
6#include <linux/timekeeper_internal.h>
7#include <linux/module.h>
8#include <linux/interrupt.h>
9#include <linux/percpu.h>
10#include <linux/init.h>
11#include <linux/mm.h>
12#include <linux/nmi.h>
13#include <linux/sched.h>
14#include <linux/sched/loadavg.h>
15#include <linux/sched/clock.h>
16#include <linux/syscore_ops.h>
17#include <linux/clocksource.h>
18#include <linux/jiffies.h>
19#include <linux/time.h>
20#include <linux/timex.h>
21#include <linux/tick.h>
22#include <linux/stop_machine.h>
23#include <linux/pvclock_gtod.h>
24#include <linux/compiler.h>
25#include <linux/audit.h>
26#include <linux/random.h>
27
28#include "tick-internal.h"
29#include "ntp_internal.h"
30#include "timekeeping_internal.h"
31
32#define TK_CLEAR_NTP		(1 << 0)
33#define TK_MIRROR		(1 << 1)
34#define TK_CLOCK_WAS_SET	(1 << 2)
35
36enum timekeeping_adv_mode {
37	/* Update timekeeper when a tick has passed */
38	TK_ADV_TICK,
39
40	/* Update timekeeper on a direct frequency change */
41	TK_ADV_FREQ
42};
43
44DEFINE_RAW_SPINLOCK(timekeeper_lock);
45
46/*
47 * The most important data for readout fits into a single 64 byte
48 * cache line.
49 */
50static struct {
51	seqcount_raw_spinlock_t	seq;
52	struct timekeeper	timekeeper;
53} tk_core ____cacheline_aligned = {
54	.seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
55};
56
57static struct timekeeper shadow_timekeeper;
58
59/* flag for if timekeeping is suspended */
60int __read_mostly timekeeping_suspended;
61
62/**
63 * struct tk_fast - NMI safe timekeeper
64 * @seq:	Sequence counter for protecting updates. The lowest bit
65 *		is the index for the tk_read_base array
66 * @base:	tk_read_base array. Access is indexed by the lowest bit of
67 *		@seq.
68 *
69 * See @update_fast_timekeeper() below.
70 */
71struct tk_fast {
72	seqcount_latch_t	seq;
73	struct tk_read_base	base[2];
74};
75
76/* Suspend-time cycles value for halted fast timekeeper. */
77static u64 cycles_at_suspend;
78
79static u64 dummy_clock_read(struct clocksource *cs)
80{
81	if (timekeeping_suspended)
82		return cycles_at_suspend;
83	return local_clock();
84}
85
86static struct clocksource dummy_clock = {
87	.read = dummy_clock_read,
88};
89
90/*
91 * Boot time initialization which allows local_clock() to be utilized
92 * during early boot when clocksources are not available. local_clock()
93 * returns nanoseconds already so no conversion is required, hence mult=1
94 * and shift=0. When the first proper clocksource is installed then
95 * the fast time keepers are updated with the correct values.
96 */
97#define FAST_TK_INIT						\
98	{							\
99		.clock		= &dummy_clock,			\
100		.mask		= CLOCKSOURCE_MASK(64),		\
101		.mult		= 1,				\
102		.shift		= 0,				\
103	}
104
105static struct tk_fast tk_fast_mono ____cacheline_aligned = {
106	.seq     = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
107	.base[0] = FAST_TK_INIT,
108	.base[1] = FAST_TK_INIT,
109};
110
111static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
112	.seq     = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
113	.base[0] = FAST_TK_INIT,
114	.base[1] = FAST_TK_INIT,
115};
116
117static inline void tk_normalize_xtime(struct timekeeper *tk)
118{
119	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
120		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
121		tk->xtime_sec++;
122	}
123	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
124		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
125		tk->raw_sec++;
126	}
127}
128
129static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
130{
131	struct timespec64 ts;
132
133	ts.tv_sec = tk->xtime_sec;
134	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
135	return ts;
136}
137
138static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
139{
140	tk->xtime_sec = ts->tv_sec;
141	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
142}
143
144static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
145{
146	tk->xtime_sec += ts->tv_sec;
147	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
148	tk_normalize_xtime(tk);
149}
150
151static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
152{
153	struct timespec64 tmp;
154
155	/*
156	 * Verify consistency of: offset_real = -wall_to_monotonic
157	 * before modifying anything
158	 */
159	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
160					-tk->wall_to_monotonic.tv_nsec);
161	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
162	tk->wall_to_monotonic = wtm;
163	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
164	tk->offs_real = timespec64_to_ktime(tmp);
165	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
166}
167
168static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
169{
170	tk->offs_boot = ktime_add(tk->offs_boot, delta);
171	/*
172	 * Timespec representation for VDSO update to avoid 64bit division
173	 * on every update.
174	 */
175	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
176}
177
178/*
179 * tk_clock_read - atomic clocksource read() helper
180 *
181 * This helper is necessary to use in the read paths because, while the
182 * seqcount ensures we don't return a bad value while structures are updated,
183 * it doesn't protect from potential crashes. There is the possibility that
184 * the tkr's clocksource may change between the read reference, and the
185 * clock reference passed to the read function.  This can cause crashes if
186 * the wrong clocksource is passed to the wrong read function.
187 * This isn't necessary to use when holding the timekeeper_lock or doing
188 * a read of the fast-timekeeper tkrs (which is protected by its own locking
189 * and update logic).
190 */
191static inline u64 tk_clock_read(const struct tk_read_base *tkr)
192{
193	struct clocksource *clock = READ_ONCE(tkr->clock);
194
195	return clock->read(clock);
196}
197
198#ifdef CONFIG_DEBUG_TIMEKEEPING
199#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
200
201static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
202{
203
204	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
205	const char *name = tk->tkr_mono.clock->name;
206
207	if (offset > max_cycles) {
208		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
209				offset, name, max_cycles);
210		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
211	} else {
212		if (offset > (max_cycles >> 1)) {
213			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
214					offset, name, max_cycles >> 1);
215			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
216		}
217	}
218
219	if (tk->underflow_seen) {
220		if (jiffies - tk->last_warning > WARNING_FREQ) {
221			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
222			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
223			printk_deferred("         Your kernel is probably still fine.\n");
224			tk->last_warning = jiffies;
225		}
226		tk->underflow_seen = 0;
227	}
228
229	if (tk->overflow_seen) {
230		if (jiffies - tk->last_warning > WARNING_FREQ) {
231			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
232			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
233			printk_deferred("         Your kernel is probably still fine.\n");
234			tk->last_warning = jiffies;
235		}
236		tk->overflow_seen = 0;
237	}
238}
239
240static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles);
241
242static inline u64 timekeeping_debug_get_ns(const struct tk_read_base *tkr)
243{
244	struct timekeeper *tk = &tk_core.timekeeper;
245	u64 now, last, mask, max, delta;
246	unsigned int seq;
247
248	/*
249	 * Since we're called holding a seqcount, the data may shift
250	 * under us while we're doing the calculation. This can cause
251	 * false positives, since we'd note a problem but throw the
252	 * results away. So nest another seqcount here to atomically
253	 * grab the points we are checking with.
254	 */
255	do {
256		seq = read_seqcount_begin(&tk_core.seq);
257		now = tk_clock_read(tkr);
258		last = tkr->cycle_last;
259		mask = tkr->mask;
260		max = tkr->clock->max_cycles;
261	} while (read_seqcount_retry(&tk_core.seq, seq));
262
263	delta = clocksource_delta(now, last, mask);
264
265	/*
266	 * Try to catch underflows by checking if we are seeing small
267	 * mask-relative negative values.
268	 */
269	if (unlikely((~delta & mask) < (mask >> 3)))
270		tk->underflow_seen = 1;
271
272	/* Check for multiplication overflows */
273	if (unlikely(delta > max))
274		tk->overflow_seen = 1;
275
276	/* timekeeping_cycles_to_ns() handles both under and overflow */
277	return timekeeping_cycles_to_ns(tkr, now);
278}
279#else
280static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
281{
282}
283static inline u64 timekeeping_debug_get_ns(const struct tk_read_base *tkr)
284{
285	BUG();
286}
287#endif
288
289/**
290 * tk_setup_internals - Set up internals to use clocksource clock.
291 *
292 * @tk:		The target timekeeper to setup.
293 * @clock:		Pointer to clocksource.
294 *
295 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
296 * pair and interval request.
297 *
298 * Unless you're the timekeeping code, you should not be using this!
299 */
300static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
301{
302	u64 interval;
303	u64 tmp, ntpinterval;
304	struct clocksource *old_clock;
305
306	++tk->cs_was_changed_seq;
307	old_clock = tk->tkr_mono.clock;
308	tk->tkr_mono.clock = clock;
309	tk->tkr_mono.mask = clock->mask;
310	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
311
312	tk->tkr_raw.clock = clock;
313	tk->tkr_raw.mask = clock->mask;
314	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
315
316	/* Do the ns -> cycle conversion first, using original mult */
317	tmp = NTP_INTERVAL_LENGTH;
318	tmp <<= clock->shift;
319	ntpinterval = tmp;
320	tmp += clock->mult/2;
321	do_div(tmp, clock->mult);
322	if (tmp == 0)
323		tmp = 1;
324
325	interval = (u64) tmp;
326	tk->cycle_interval = interval;
327
328	/* Go back from cycles -> shifted ns */
329	tk->xtime_interval = interval * clock->mult;
330	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
331	tk->raw_interval = interval * clock->mult;
332
333	 /* if changing clocks, convert xtime_nsec shift units */
334	if (old_clock) {
335		int shift_change = clock->shift - old_clock->shift;
336		if (shift_change < 0) {
337			tk->tkr_mono.xtime_nsec >>= -shift_change;
338			tk->tkr_raw.xtime_nsec >>= -shift_change;
339		} else {
340			tk->tkr_mono.xtime_nsec <<= shift_change;
341			tk->tkr_raw.xtime_nsec <<= shift_change;
342		}
343	}
344
345	tk->tkr_mono.shift = clock->shift;
346	tk->tkr_raw.shift = clock->shift;
347
348	tk->ntp_error = 0;
349	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
350	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
351
352	/*
353	 * The timekeeper keeps its own mult values for the currently
354	 * active clocksource. These value will be adjusted via NTP
355	 * to counteract clock drifting.
356	 */
357	tk->tkr_mono.mult = clock->mult;
358	tk->tkr_raw.mult = clock->mult;
359	tk->ntp_err_mult = 0;
360	tk->skip_second_overflow = 0;
361}
362
363/* Timekeeper helper functions. */
364static noinline u64 delta_to_ns_safe(const struct tk_read_base *tkr, u64 delta)
365{
366	return mul_u64_u32_add_u64_shr(delta, tkr->mult, tkr->xtime_nsec, tkr->shift);
367}
368
369static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
370{
371	/* Calculate the delta since the last update_wall_time() */
372	u64 mask = tkr->mask, delta = (cycles - tkr->cycle_last) & mask;
373
374	/*
375	 * This detects both negative motion and the case where the delta
376	 * overflows the multiplication with tkr->mult.
377	 */
378	if (unlikely(delta > tkr->clock->max_cycles)) {
379		/*
380		 * Handle clocksource inconsistency between CPUs to prevent
381		 * time from going backwards by checking for the MSB of the
382		 * mask being set in the delta.
383		 */
384		if (delta & ~(mask >> 1))
385			return tkr->xtime_nsec >> tkr->shift;
386
387		return delta_to_ns_safe(tkr, delta);
388	}
389
390	return ((delta * tkr->mult) + tkr->xtime_nsec) >> tkr->shift;
391}
392
393static __always_inline u64 __timekeeping_get_ns(const struct tk_read_base *tkr)
394{
395	return timekeeping_cycles_to_ns(tkr, tk_clock_read(tkr));
396}
397
398static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
399{
400	if (IS_ENABLED(CONFIG_DEBUG_TIMEKEEPING))
401		return timekeeping_debug_get_ns(tkr);
402
403	return __timekeeping_get_ns(tkr);
404}
405
406/**
407 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
408 * @tkr: Timekeeping readout base from which we take the update
409 * @tkf: Pointer to NMI safe timekeeper
410 *
411 * We want to use this from any context including NMI and tracing /
412 * instrumenting the timekeeping code itself.
413 *
414 * Employ the latch technique; see @raw_write_seqcount_latch.
415 *
416 * So if a NMI hits the update of base[0] then it will use base[1]
417 * which is still consistent. In the worst case this can result is a
418 * slightly wrong timestamp (a few nanoseconds). See
419 * @ktime_get_mono_fast_ns.
420 */
421static void update_fast_timekeeper(const struct tk_read_base *tkr,
422				   struct tk_fast *tkf)
423{
424	struct tk_read_base *base = tkf->base;
425
426	/* Force readers off to base[1] */
427	raw_write_seqcount_latch(&tkf->seq);
428
429	/* Update base[0] */
430	memcpy(base, tkr, sizeof(*base));
431
432	/* Force readers back to base[0] */
433	raw_write_seqcount_latch(&tkf->seq);
434
435	/* Update base[1] */
436	memcpy(base + 1, base, sizeof(*base));
437}
438
439static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
440{
441	struct tk_read_base *tkr;
442	unsigned int seq;
443	u64 now;
444
445	do {
446		seq = raw_read_seqcount_latch(&tkf->seq);
447		tkr = tkf->base + (seq & 0x01);
448		now = ktime_to_ns(tkr->base);
449		now += __timekeeping_get_ns(tkr);
450	} while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
451
452	return now;
453}
454
455/**
456 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
457 *
458 * This timestamp is not guaranteed to be monotonic across an update.
459 * The timestamp is calculated by:
460 *
461 *	now = base_mono + clock_delta * slope
462 *
463 * So if the update lowers the slope, readers who are forced to the
464 * not yet updated second array are still using the old steeper slope.
465 *
466 * tmono
467 * ^
468 * |    o  n
469 * |   o n
470 * |  u
471 * | o
472 * |o
473 * |12345678---> reader order
474 *
475 * o = old slope
476 * u = update
477 * n = new slope
478 *
479 * So reader 6 will observe time going backwards versus reader 5.
480 *
481 * While other CPUs are likely to be able to observe that, the only way
482 * for a CPU local observation is when an NMI hits in the middle of
483 * the update. Timestamps taken from that NMI context might be ahead
484 * of the following timestamps. Callers need to be aware of that and
485 * deal with it.
486 */
487u64 notrace ktime_get_mono_fast_ns(void)
488{
489	return __ktime_get_fast_ns(&tk_fast_mono);
490}
491EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
492
493/**
494 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
495 *
496 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
497 * conversion factor is not affected by NTP/PTP correction.
498 */
499u64 notrace ktime_get_raw_fast_ns(void)
500{
501	return __ktime_get_fast_ns(&tk_fast_raw);
502}
503EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
504
505/**
506 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
507 *
508 * To keep it NMI safe since we're accessing from tracing, we're not using a
509 * separate timekeeper with updates to monotonic clock and boot offset
510 * protected with seqcounts. This has the following minor side effects:
511 *
512 * (1) Its possible that a timestamp be taken after the boot offset is updated
513 * but before the timekeeper is updated. If this happens, the new boot offset
514 * is added to the old timekeeping making the clock appear to update slightly
515 * earlier:
516 *    CPU 0                                        CPU 1
517 *    timekeeping_inject_sleeptime64()
518 *    __timekeeping_inject_sleeptime(tk, delta);
519 *                                                 timestamp();
520 *    timekeeping_update(tk, TK_CLEAR_NTP...);
521 *
522 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
523 * partially updated.  Since the tk->offs_boot update is a rare event, this
524 * should be a rare occurrence which postprocessing should be able to handle.
525 *
526 * The caveats vs. timestamp ordering as documented for ktime_get_mono_fast_ns()
527 * apply as well.
528 */
529u64 notrace ktime_get_boot_fast_ns(void)
530{
531	struct timekeeper *tk = &tk_core.timekeeper;
532
533	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
534}
535EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
536
537/**
538 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
539 *
540 * The same limitations as described for ktime_get_boot_fast_ns() apply. The
541 * mono time and the TAI offset are not read atomically which may yield wrong
542 * readouts. However, an update of the TAI offset is an rare event e.g., caused
543 * by settime or adjtimex with an offset. The user of this function has to deal
544 * with the possibility of wrong timestamps in post processing.
545 */
546u64 notrace ktime_get_tai_fast_ns(void)
547{
548	struct timekeeper *tk = &tk_core.timekeeper;
549
550	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
551}
552EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
553
554static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
555{
556	struct tk_read_base *tkr;
557	u64 basem, baser, delta;
558	unsigned int seq;
559
560	do {
561		seq = raw_read_seqcount_latch(&tkf->seq);
562		tkr = tkf->base + (seq & 0x01);
563		basem = ktime_to_ns(tkr->base);
564		baser = ktime_to_ns(tkr->base_real);
565		delta = __timekeeping_get_ns(tkr);
566	} while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
567
568	if (mono)
569		*mono = basem + delta;
570	return baser + delta;
571}
572
573/**
574 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
575 *
576 * See ktime_get_mono_fast_ns() for documentation of the time stamp ordering.
577 */
578u64 ktime_get_real_fast_ns(void)
579{
580	return __ktime_get_real_fast(&tk_fast_mono, NULL);
581}
582EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
583
584/**
585 * ktime_get_fast_timestamps: - NMI safe timestamps
586 * @snapshot:	Pointer to timestamp storage
587 *
588 * Stores clock monotonic, boottime and realtime timestamps.
589 *
590 * Boot time is a racy access on 32bit systems if the sleep time injection
591 * happens late during resume and not in timekeeping_resume(). That could
592 * be avoided by expanding struct tk_read_base with boot offset for 32bit
593 * and adding more overhead to the update. As this is a hard to observe
594 * once per resume event which can be filtered with reasonable effort using
595 * the accurate mono/real timestamps, it's probably not worth the trouble.
596 *
597 * Aside of that it might be possible on 32 and 64 bit to observe the
598 * following when the sleep time injection happens late:
599 *
600 * CPU 0				CPU 1
601 * timekeeping_resume()
602 * ktime_get_fast_timestamps()
603 *	mono, real = __ktime_get_real_fast()
604 *					inject_sleep_time()
605 *					   update boot offset
606 *	boot = mono + bootoffset;
607 *
608 * That means that boot time already has the sleep time adjustment, but
609 * real time does not. On the next readout both are in sync again.
610 *
611 * Preventing this for 64bit is not really feasible without destroying the
612 * careful cache layout of the timekeeper because the sequence count and
613 * struct tk_read_base would then need two cache lines instead of one.
614 *
615 * Access to the time keeper clock source is disabled across the innermost
616 * steps of suspend/resume. The accessors still work, but the timestamps
617 * are frozen until time keeping is resumed which happens very early.
618 *
619 * For regular suspend/resume there is no observable difference vs. sched
620 * clock, but it might affect some of the nasty low level debug printks.
621 *
622 * OTOH, access to sched clock is not guaranteed across suspend/resume on
623 * all systems either so it depends on the hardware in use.
624 *
625 * If that turns out to be a real problem then this could be mitigated by
626 * using sched clock in a similar way as during early boot. But it's not as
627 * trivial as on early boot because it needs some careful protection
628 * against the clock monotonic timestamp jumping backwards on resume.
629 */
630void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
631{
632	struct timekeeper *tk = &tk_core.timekeeper;
633
634	snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
635	snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
636}
637
638/**
639 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
640 * @tk: Timekeeper to snapshot.
641 *
642 * It generally is unsafe to access the clocksource after timekeeping has been
643 * suspended, so take a snapshot of the readout base of @tk and use it as the
644 * fast timekeeper's readout base while suspended.  It will return the same
645 * number of cycles every time until timekeeping is resumed at which time the
646 * proper readout base for the fast timekeeper will be restored automatically.
647 */
648static void halt_fast_timekeeper(const struct timekeeper *tk)
649{
650	static struct tk_read_base tkr_dummy;
651	const struct tk_read_base *tkr = &tk->tkr_mono;
652
653	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
654	cycles_at_suspend = tk_clock_read(tkr);
655	tkr_dummy.clock = &dummy_clock;
656	tkr_dummy.base_real = tkr->base + tk->offs_real;
657	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
658
659	tkr = &tk->tkr_raw;
660	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
661	tkr_dummy.clock = &dummy_clock;
662	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
663}
664
665static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
666
667static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
668{
669	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
670}
671
672/**
673 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
674 * @nb: Pointer to the notifier block to register
675 */
676int pvclock_gtod_register_notifier(struct notifier_block *nb)
677{
678	struct timekeeper *tk = &tk_core.timekeeper;
679	unsigned long flags;
680	int ret;
681
682	raw_spin_lock_irqsave(&timekeeper_lock, flags);
683	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
684	update_pvclock_gtod(tk, true);
685	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
686
687	return ret;
688}
689EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
690
691/**
692 * pvclock_gtod_unregister_notifier - unregister a pvclock
693 * timedata update listener
694 * @nb: Pointer to the notifier block to unregister
695 */
696int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
697{
698	unsigned long flags;
699	int ret;
700
701	raw_spin_lock_irqsave(&timekeeper_lock, flags);
702	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
703	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
704
705	return ret;
706}
707EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
708
709/*
710 * tk_update_leap_state - helper to update the next_leap_ktime
711 */
712static inline void tk_update_leap_state(struct timekeeper *tk)
713{
714	tk->next_leap_ktime = ntp_get_next_leap();
715	if (tk->next_leap_ktime != KTIME_MAX)
716		/* Convert to monotonic time */
717		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
718}
719
720/*
721 * Update the ktime_t based scalar nsec members of the timekeeper
722 */
723static inline void tk_update_ktime_data(struct timekeeper *tk)
724{
725	u64 seconds;
726	u32 nsec;
727
728	/*
729	 * The xtime based monotonic readout is:
730	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
731	 * The ktime based monotonic readout is:
732	 *	nsec = base_mono + now();
733	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
734	 */
735	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
736	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
737	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
738
739	/*
740	 * The sum of the nanoseconds portions of xtime and
741	 * wall_to_monotonic can be greater/equal one second. Take
742	 * this into account before updating tk->ktime_sec.
743	 */
744	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
745	if (nsec >= NSEC_PER_SEC)
746		seconds++;
747	tk->ktime_sec = seconds;
748
749	/* Update the monotonic raw base */
750	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
751}
752
753/* must hold timekeeper_lock */
754static void timekeeping_update(struct timekeeper *tk, unsigned int action)
755{
756	if (action & TK_CLEAR_NTP) {
757		tk->ntp_error = 0;
758		ntp_clear();
759	}
760
761	tk_update_leap_state(tk);
762	tk_update_ktime_data(tk);
763
764	update_vsyscall(tk);
765	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
766
767	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
768	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
769	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
770
771	if (action & TK_CLOCK_WAS_SET)
772		tk->clock_was_set_seq++;
773	/*
774	 * The mirroring of the data to the shadow-timekeeper needs
775	 * to happen last here to ensure we don't over-write the
776	 * timekeeper structure on the next update with stale data
777	 */
778	if (action & TK_MIRROR)
779		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
780		       sizeof(tk_core.timekeeper));
781}
782
783/**
784 * timekeeping_forward_now - update clock to the current time
785 * @tk:		Pointer to the timekeeper to update
786 *
787 * Forward the current clock to update its state since the last call to
788 * update_wall_time(). This is useful before significant clock changes,
789 * as it avoids having to deal with this time offset explicitly.
790 */
791static void timekeeping_forward_now(struct timekeeper *tk)
792{
793	u64 cycle_now, delta;
794
795	cycle_now = tk_clock_read(&tk->tkr_mono);
796	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
797	tk->tkr_mono.cycle_last = cycle_now;
798	tk->tkr_raw.cycle_last  = cycle_now;
799
800	while (delta > 0) {
801		u64 max = tk->tkr_mono.clock->max_cycles;
802		u64 incr = delta < max ? delta : max;
803
804		tk->tkr_mono.xtime_nsec += incr * tk->tkr_mono.mult;
805		tk->tkr_raw.xtime_nsec += incr * tk->tkr_raw.mult;
806		tk_normalize_xtime(tk);
807		delta -= incr;
808	}
809}
810
811/**
812 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
813 * @ts:		pointer to the timespec to be set
814 *
815 * Returns the time of day in a timespec64 (WARN if suspended).
816 */
817void ktime_get_real_ts64(struct timespec64 *ts)
818{
819	struct timekeeper *tk = &tk_core.timekeeper;
820	unsigned int seq;
821	u64 nsecs;
822
823	WARN_ON(timekeeping_suspended);
824
825	do {
826		seq = read_seqcount_begin(&tk_core.seq);
827
828		ts->tv_sec = tk->xtime_sec;
829		nsecs = timekeeping_get_ns(&tk->tkr_mono);
830
831	} while (read_seqcount_retry(&tk_core.seq, seq));
832
833	ts->tv_nsec = 0;
834	timespec64_add_ns(ts, nsecs);
835}
836EXPORT_SYMBOL(ktime_get_real_ts64);
837
838ktime_t ktime_get(void)
839{
840	struct timekeeper *tk = &tk_core.timekeeper;
841	unsigned int seq;
842	ktime_t base;
843	u64 nsecs;
844
845	WARN_ON(timekeeping_suspended);
846
847	do {
848		seq = read_seqcount_begin(&tk_core.seq);
849		base = tk->tkr_mono.base;
850		nsecs = timekeeping_get_ns(&tk->tkr_mono);
851
852	} while (read_seqcount_retry(&tk_core.seq, seq));
853
854	return ktime_add_ns(base, nsecs);
855}
856EXPORT_SYMBOL_GPL(ktime_get);
857
858u32 ktime_get_resolution_ns(void)
859{
860	struct timekeeper *tk = &tk_core.timekeeper;
861	unsigned int seq;
862	u32 nsecs;
863
864	WARN_ON(timekeeping_suspended);
865
866	do {
867		seq = read_seqcount_begin(&tk_core.seq);
868		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
869	} while (read_seqcount_retry(&tk_core.seq, seq));
870
871	return nsecs;
872}
873EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
874
875static ktime_t *offsets[TK_OFFS_MAX] = {
876	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
877	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
878	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
879};
880
881ktime_t ktime_get_with_offset(enum tk_offsets offs)
882{
883	struct timekeeper *tk = &tk_core.timekeeper;
884	unsigned int seq;
885	ktime_t base, *offset = offsets[offs];
886	u64 nsecs;
887
888	WARN_ON(timekeeping_suspended);
889
890	do {
891		seq = read_seqcount_begin(&tk_core.seq);
892		base = ktime_add(tk->tkr_mono.base, *offset);
893		nsecs = timekeeping_get_ns(&tk->tkr_mono);
894
895	} while (read_seqcount_retry(&tk_core.seq, seq));
896
897	return ktime_add_ns(base, nsecs);
898
899}
900EXPORT_SYMBOL_GPL(ktime_get_with_offset);
901
902ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
903{
904	struct timekeeper *tk = &tk_core.timekeeper;
905	unsigned int seq;
906	ktime_t base, *offset = offsets[offs];
907	u64 nsecs;
908
909	WARN_ON(timekeeping_suspended);
910
911	do {
912		seq = read_seqcount_begin(&tk_core.seq);
913		base = ktime_add(tk->tkr_mono.base, *offset);
914		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
915
916	} while (read_seqcount_retry(&tk_core.seq, seq));
917
918	return ktime_add_ns(base, nsecs);
919}
920EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
921
922/**
923 * ktime_mono_to_any() - convert monotonic time to any other time
924 * @tmono:	time to convert.
925 * @offs:	which offset to use
926 */
927ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
928{
929	ktime_t *offset = offsets[offs];
930	unsigned int seq;
931	ktime_t tconv;
932
933	do {
934		seq = read_seqcount_begin(&tk_core.seq);
935		tconv = ktime_add(tmono, *offset);
936	} while (read_seqcount_retry(&tk_core.seq, seq));
937
938	return tconv;
939}
940EXPORT_SYMBOL_GPL(ktime_mono_to_any);
941
942/**
943 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
944 */
945ktime_t ktime_get_raw(void)
946{
947	struct timekeeper *tk = &tk_core.timekeeper;
948	unsigned int seq;
949	ktime_t base;
950	u64 nsecs;
951
952	do {
953		seq = read_seqcount_begin(&tk_core.seq);
954		base = tk->tkr_raw.base;
955		nsecs = timekeeping_get_ns(&tk->tkr_raw);
956
957	} while (read_seqcount_retry(&tk_core.seq, seq));
958
959	return ktime_add_ns(base, nsecs);
960}
961EXPORT_SYMBOL_GPL(ktime_get_raw);
962
963/**
964 * ktime_get_ts64 - get the monotonic clock in timespec64 format
965 * @ts:		pointer to timespec variable
966 *
967 * The function calculates the monotonic clock from the realtime
968 * clock and the wall_to_monotonic offset and stores the result
969 * in normalized timespec64 format in the variable pointed to by @ts.
970 */
971void ktime_get_ts64(struct timespec64 *ts)
972{
973	struct timekeeper *tk = &tk_core.timekeeper;
974	struct timespec64 tomono;
975	unsigned int seq;
976	u64 nsec;
977
978	WARN_ON(timekeeping_suspended);
979
980	do {
981		seq = read_seqcount_begin(&tk_core.seq);
982		ts->tv_sec = tk->xtime_sec;
983		nsec = timekeeping_get_ns(&tk->tkr_mono);
984		tomono = tk->wall_to_monotonic;
985
986	} while (read_seqcount_retry(&tk_core.seq, seq));
987
988	ts->tv_sec += tomono.tv_sec;
989	ts->tv_nsec = 0;
990	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
991}
992EXPORT_SYMBOL_GPL(ktime_get_ts64);
993
994/**
995 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
996 *
997 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
998 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
999 * works on both 32 and 64 bit systems. On 32 bit systems the readout
1000 * covers ~136 years of uptime which should be enough to prevent
1001 * premature wrap arounds.
1002 */
1003time64_t ktime_get_seconds(void)
1004{
1005	struct timekeeper *tk = &tk_core.timekeeper;
1006
1007	WARN_ON(timekeeping_suspended);
1008	return tk->ktime_sec;
1009}
1010EXPORT_SYMBOL_GPL(ktime_get_seconds);
1011
1012/**
1013 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1014 *
1015 * Returns the wall clock seconds since 1970.
1016 *
1017 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1018 * 32bit systems the access must be protected with the sequence
1019 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1020 * value.
1021 */
1022time64_t ktime_get_real_seconds(void)
1023{
1024	struct timekeeper *tk = &tk_core.timekeeper;
1025	time64_t seconds;
1026	unsigned int seq;
1027
1028	if (IS_ENABLED(CONFIG_64BIT))
1029		return tk->xtime_sec;
1030
1031	do {
1032		seq = read_seqcount_begin(&tk_core.seq);
1033		seconds = tk->xtime_sec;
1034
1035	} while (read_seqcount_retry(&tk_core.seq, seq));
1036
1037	return seconds;
1038}
1039EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1040
1041/**
1042 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1043 * but without the sequence counter protect. This internal function
1044 * is called just when timekeeping lock is already held.
1045 */
1046noinstr time64_t __ktime_get_real_seconds(void)
1047{
1048	struct timekeeper *tk = &tk_core.timekeeper;
1049
1050	return tk->xtime_sec;
1051}
1052
1053/**
1054 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1055 * @systime_snapshot:	pointer to struct receiving the system time snapshot
1056 */
1057void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1058{
1059	struct timekeeper *tk = &tk_core.timekeeper;
1060	unsigned int seq;
1061	ktime_t base_raw;
1062	ktime_t base_real;
1063	u64 nsec_raw;
1064	u64 nsec_real;
1065	u64 now;
1066
1067	WARN_ON_ONCE(timekeeping_suspended);
1068
1069	do {
1070		seq = read_seqcount_begin(&tk_core.seq);
1071		now = tk_clock_read(&tk->tkr_mono);
1072		systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1073		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1074		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1075		base_real = ktime_add(tk->tkr_mono.base,
1076				      tk_core.timekeeper.offs_real);
1077		base_raw = tk->tkr_raw.base;
1078		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1079		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1080	} while (read_seqcount_retry(&tk_core.seq, seq));
1081
1082	systime_snapshot->cycles = now;
1083	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1084	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1085}
1086EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1087
1088/* Scale base by mult/div checking for overflow */
1089static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1090{
1091	u64 tmp, rem;
1092
1093	tmp = div64_u64_rem(*base, div, &rem);
1094
1095	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1096	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1097		return -EOVERFLOW;
1098	tmp *= mult;
1099
1100	rem = div64_u64(rem * mult, div);
1101	*base = tmp + rem;
1102	return 0;
1103}
1104
1105/**
1106 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1107 * @history:			Snapshot representing start of history
1108 * @partial_history_cycles:	Cycle offset into history (fractional part)
1109 * @total_history_cycles:	Total history length in cycles
1110 * @discontinuity:		True indicates clock was set on history period
1111 * @ts:				Cross timestamp that should be adjusted using
1112 *	partial/total ratio
1113 *
1114 * Helper function used by get_device_system_crosststamp() to correct the
1115 * crosstimestamp corresponding to the start of the current interval to the
1116 * system counter value (timestamp point) provided by the driver. The
1117 * total_history_* quantities are the total history starting at the provided
1118 * reference point and ending at the start of the current interval. The cycle
1119 * count between the driver timestamp point and the start of the current
1120 * interval is partial_history_cycles.
1121 */
1122static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1123					 u64 partial_history_cycles,
1124					 u64 total_history_cycles,
1125					 bool discontinuity,
1126					 struct system_device_crosststamp *ts)
1127{
1128	struct timekeeper *tk = &tk_core.timekeeper;
1129	u64 corr_raw, corr_real;
1130	bool interp_forward;
1131	int ret;
1132
1133	if (total_history_cycles == 0 || partial_history_cycles == 0)
1134		return 0;
1135
1136	/* Interpolate shortest distance from beginning or end of history */
1137	interp_forward = partial_history_cycles > total_history_cycles / 2;
1138	partial_history_cycles = interp_forward ?
1139		total_history_cycles - partial_history_cycles :
1140		partial_history_cycles;
1141
1142	/*
1143	 * Scale the monotonic raw time delta by:
1144	 *	partial_history_cycles / total_history_cycles
1145	 */
1146	corr_raw = (u64)ktime_to_ns(
1147		ktime_sub(ts->sys_monoraw, history->raw));
1148	ret = scale64_check_overflow(partial_history_cycles,
1149				     total_history_cycles, &corr_raw);
1150	if (ret)
1151		return ret;
1152
1153	/*
1154	 * If there is a discontinuity in the history, scale monotonic raw
1155	 *	correction by:
1156	 *	mult(real)/mult(raw) yielding the realtime correction
1157	 * Otherwise, calculate the realtime correction similar to monotonic
1158	 *	raw calculation
1159	 */
1160	if (discontinuity) {
1161		corr_real = mul_u64_u32_div
1162			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1163	} else {
1164		corr_real = (u64)ktime_to_ns(
1165			ktime_sub(ts->sys_realtime, history->real));
1166		ret = scale64_check_overflow(partial_history_cycles,
1167					     total_history_cycles, &corr_real);
1168		if (ret)
1169			return ret;
1170	}
1171
1172	/* Fixup monotonic raw and real time time values */
1173	if (interp_forward) {
1174		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1175		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1176	} else {
1177		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1178		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1179	}
1180
1181	return 0;
1182}
1183
1184/*
1185 * timestamp_in_interval - true if ts is chronologically in [start, end]
1186 *
1187 * True if ts occurs chronologically at or after start, and before or at end.
1188 */
1189static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1190{
1191	if (ts >= start && ts <= end)
1192		return true;
1193	if (start > end && (ts >= start || ts <= end))
1194		return true;
1195	return false;
1196}
1197
1198/**
1199 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1200 * @get_time_fn:	Callback to get simultaneous device time and
1201 *	system counter from the device driver
1202 * @ctx:		Context passed to get_time_fn()
1203 * @history_begin:	Historical reference point used to interpolate system
1204 *	time when counter provided by the driver is before the current interval
1205 * @xtstamp:		Receives simultaneously captured system and device time
1206 *
1207 * Reads a timestamp from a device and correlates it to system time
1208 */
1209int get_device_system_crosststamp(int (*get_time_fn)
1210				  (ktime_t *device_time,
1211				   struct system_counterval_t *sys_counterval,
1212				   void *ctx),
1213				  void *ctx,
1214				  struct system_time_snapshot *history_begin,
1215				  struct system_device_crosststamp *xtstamp)
1216{
1217	struct system_counterval_t system_counterval;
1218	struct timekeeper *tk = &tk_core.timekeeper;
1219	u64 cycles, now, interval_start;
1220	unsigned int clock_was_set_seq = 0;
1221	ktime_t base_real, base_raw;
1222	u64 nsec_real, nsec_raw;
1223	u8 cs_was_changed_seq;
1224	unsigned int seq;
1225	bool do_interp;
1226	int ret;
1227
1228	do {
1229		seq = read_seqcount_begin(&tk_core.seq);
1230		/*
1231		 * Try to synchronously capture device time and a system
1232		 * counter value calling back into the device driver
1233		 */
1234		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1235		if (ret)
1236			return ret;
1237
1238		/*
1239		 * Verify that the clocksource ID associated with the captured
1240		 * system counter value is the same as for the currently
1241		 * installed timekeeper clocksource
1242		 */
1243		if (system_counterval.cs_id == CSID_GENERIC ||
1244		    tk->tkr_mono.clock->id != system_counterval.cs_id)
1245			return -ENODEV;
1246		cycles = system_counterval.cycles;
1247
1248		/*
1249		 * Check whether the system counter value provided by the
1250		 * device driver is on the current timekeeping interval.
1251		 */
1252		now = tk_clock_read(&tk->tkr_mono);
1253		interval_start = tk->tkr_mono.cycle_last;
1254		if (!timestamp_in_interval(interval_start, now, cycles)) {
1255			clock_was_set_seq = tk->clock_was_set_seq;
1256			cs_was_changed_seq = tk->cs_was_changed_seq;
1257			cycles = interval_start;
1258			do_interp = true;
1259		} else {
1260			do_interp = false;
1261		}
1262
1263		base_real = ktime_add(tk->tkr_mono.base,
1264				      tk_core.timekeeper.offs_real);
1265		base_raw = tk->tkr_raw.base;
1266
1267		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1268		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1269	} while (read_seqcount_retry(&tk_core.seq, seq));
1270
1271	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1272	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1273
1274	/*
1275	 * Interpolate if necessary, adjusting back from the start of the
1276	 * current interval
1277	 */
1278	if (do_interp) {
1279		u64 partial_history_cycles, total_history_cycles;
1280		bool discontinuity;
1281
1282		/*
1283		 * Check that the counter value is not before the provided
1284		 * history reference and that the history doesn't cross a
1285		 * clocksource change
1286		 */
1287		if (!history_begin ||
1288		    !timestamp_in_interval(history_begin->cycles,
1289					   cycles, system_counterval.cycles) ||
1290		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1291			return -EINVAL;
1292		partial_history_cycles = cycles - system_counterval.cycles;
1293		total_history_cycles = cycles - history_begin->cycles;
1294		discontinuity =
1295			history_begin->clock_was_set_seq != clock_was_set_seq;
1296
1297		ret = adjust_historical_crosststamp(history_begin,
1298						    partial_history_cycles,
1299						    total_history_cycles,
1300						    discontinuity, xtstamp);
1301		if (ret)
1302			return ret;
1303	}
1304
1305	return 0;
1306}
1307EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1308
1309/**
1310 * do_settimeofday64 - Sets the time of day.
1311 * @ts:     pointer to the timespec64 variable containing the new time
1312 *
1313 * Sets the time of day to the new time and update NTP and notify hrtimers
1314 */
1315int do_settimeofday64(const struct timespec64 *ts)
1316{
1317	struct timekeeper *tk = &tk_core.timekeeper;
1318	struct timespec64 ts_delta, xt;
1319	unsigned long flags;
1320	int ret = 0;
1321
1322	if (!timespec64_valid_settod(ts))
1323		return -EINVAL;
1324
1325	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1326	write_seqcount_begin(&tk_core.seq);
1327
1328	timekeeping_forward_now(tk);
1329
1330	xt = tk_xtime(tk);
1331	ts_delta = timespec64_sub(*ts, xt);
1332
1333	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1334		ret = -EINVAL;
1335		goto out;
1336	}
1337
1338	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1339
1340	tk_set_xtime(tk, ts);
1341out:
1342	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1343
1344	write_seqcount_end(&tk_core.seq);
1345	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1346
1347	/* Signal hrtimers about time change */
1348	clock_was_set(CLOCK_SET_WALL);
1349
1350	if (!ret) {
1351		audit_tk_injoffset(ts_delta);
1352		add_device_randomness(ts, sizeof(*ts));
1353	}
1354
1355	return ret;
1356}
1357EXPORT_SYMBOL(do_settimeofday64);
1358
1359/**
1360 * timekeeping_inject_offset - Adds or subtracts from the current time.
1361 * @ts:		Pointer to the timespec variable containing the offset
1362 *
1363 * Adds or subtracts an offset value from the current time.
1364 */
1365static int timekeeping_inject_offset(const struct timespec64 *ts)
1366{
1367	struct timekeeper *tk = &tk_core.timekeeper;
1368	unsigned long flags;
1369	struct timespec64 tmp;
1370	int ret = 0;
1371
1372	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1373		return -EINVAL;
1374
1375	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1376	write_seqcount_begin(&tk_core.seq);
1377
1378	timekeeping_forward_now(tk);
1379
1380	/* Make sure the proposed value is valid */
1381	tmp = timespec64_add(tk_xtime(tk), *ts);
1382	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1383	    !timespec64_valid_settod(&tmp)) {
1384		ret = -EINVAL;
1385		goto error;
1386	}
1387
1388	tk_xtime_add(tk, ts);
1389	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1390
1391error: /* even if we error out, we forwarded the time, so call update */
1392	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1393
1394	write_seqcount_end(&tk_core.seq);
1395	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1396
1397	/* Signal hrtimers about time change */
1398	clock_was_set(CLOCK_SET_WALL);
1399
1400	return ret;
1401}
1402
1403/*
1404 * Indicates if there is an offset between the system clock and the hardware
1405 * clock/persistent clock/rtc.
1406 */
1407int persistent_clock_is_local;
1408
1409/*
1410 * Adjust the time obtained from the CMOS to be UTC time instead of
1411 * local time.
1412 *
1413 * This is ugly, but preferable to the alternatives.  Otherwise we
1414 * would either need to write a program to do it in /etc/rc (and risk
1415 * confusion if the program gets run more than once; it would also be
1416 * hard to make the program warp the clock precisely n hours)  or
1417 * compile in the timezone information into the kernel.  Bad, bad....
1418 *
1419 *						- TYT, 1992-01-01
1420 *
1421 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1422 * as real UNIX machines always do it. This avoids all headaches about
1423 * daylight saving times and warping kernel clocks.
1424 */
1425void timekeeping_warp_clock(void)
1426{
1427	if (sys_tz.tz_minuteswest != 0) {
1428		struct timespec64 adjust;
1429
1430		persistent_clock_is_local = 1;
1431		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1432		adjust.tv_nsec = 0;
1433		timekeeping_inject_offset(&adjust);
1434	}
1435}
1436
1437/*
1438 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1439 */
1440static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1441{
1442	tk->tai_offset = tai_offset;
1443	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1444}
1445
1446/*
1447 * change_clocksource - Swaps clocksources if a new one is available
1448 *
1449 * Accumulates current time interval and initializes new clocksource
1450 */
1451static int change_clocksource(void *data)
1452{
1453	struct timekeeper *tk = &tk_core.timekeeper;
1454	struct clocksource *new, *old = NULL;
1455	unsigned long flags;
1456	bool change = false;
1457
1458	new = (struct clocksource *) data;
1459
1460	/*
1461	 * If the cs is in module, get a module reference. Succeeds
1462	 * for built-in code (owner == NULL) as well.
1463	 */
1464	if (try_module_get(new->owner)) {
1465		if (!new->enable || new->enable(new) == 0)
1466			change = true;
1467		else
1468			module_put(new->owner);
1469	}
1470
1471	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1472	write_seqcount_begin(&tk_core.seq);
1473
1474	timekeeping_forward_now(tk);
1475
1476	if (change) {
1477		old = tk->tkr_mono.clock;
1478		tk_setup_internals(tk, new);
1479	}
1480
1481	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1482
1483	write_seqcount_end(&tk_core.seq);
1484	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1485
1486	if (old) {
1487		if (old->disable)
1488			old->disable(old);
1489
1490		module_put(old->owner);
1491	}
1492
1493	return 0;
1494}
1495
1496/**
1497 * timekeeping_notify - Install a new clock source
1498 * @clock:		pointer to the clock source
1499 *
1500 * This function is called from clocksource.c after a new, better clock
1501 * source has been registered. The caller holds the clocksource_mutex.
1502 */
1503int timekeeping_notify(struct clocksource *clock)
1504{
1505	struct timekeeper *tk = &tk_core.timekeeper;
1506
1507	if (tk->tkr_mono.clock == clock)
1508		return 0;
1509	stop_machine(change_clocksource, clock, NULL);
1510	tick_clock_notify();
1511	return tk->tkr_mono.clock == clock ? 0 : -1;
1512}
1513
1514/**
1515 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1516 * @ts:		pointer to the timespec64 to be set
1517 *
1518 * Returns the raw monotonic time (completely un-modified by ntp)
1519 */
1520void ktime_get_raw_ts64(struct timespec64 *ts)
1521{
1522	struct timekeeper *tk = &tk_core.timekeeper;
1523	unsigned int seq;
1524	u64 nsecs;
1525
1526	do {
1527		seq = read_seqcount_begin(&tk_core.seq);
1528		ts->tv_sec = tk->raw_sec;
1529		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1530
1531	} while (read_seqcount_retry(&tk_core.seq, seq));
1532
1533	ts->tv_nsec = 0;
1534	timespec64_add_ns(ts, nsecs);
1535}
1536EXPORT_SYMBOL(ktime_get_raw_ts64);
1537
1538
1539/**
1540 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1541 */
1542int timekeeping_valid_for_hres(void)
1543{
1544	struct timekeeper *tk = &tk_core.timekeeper;
1545	unsigned int seq;
1546	int ret;
1547
1548	do {
1549		seq = read_seqcount_begin(&tk_core.seq);
1550
1551		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1552
1553	} while (read_seqcount_retry(&tk_core.seq, seq));
1554
1555	return ret;
1556}
1557
1558/**
1559 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1560 */
1561u64 timekeeping_max_deferment(void)
1562{
1563	struct timekeeper *tk = &tk_core.timekeeper;
1564	unsigned int seq;
1565	u64 ret;
1566
1567	do {
1568		seq = read_seqcount_begin(&tk_core.seq);
1569
1570		ret = tk->tkr_mono.clock->max_idle_ns;
1571
1572	} while (read_seqcount_retry(&tk_core.seq, seq));
1573
1574	return ret;
1575}
1576
1577/**
1578 * read_persistent_clock64 -  Return time from the persistent clock.
1579 * @ts: Pointer to the storage for the readout value
1580 *
1581 * Weak dummy function for arches that do not yet support it.
1582 * Reads the time from the battery backed persistent clock.
1583 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1584 *
1585 *  XXX - Do be sure to remove it once all arches implement it.
1586 */
1587void __weak read_persistent_clock64(struct timespec64 *ts)
1588{
1589	ts->tv_sec = 0;
1590	ts->tv_nsec = 0;
1591}
1592
1593/**
1594 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1595 *                                        from the boot.
1596 * @wall_time:	  current time as returned by persistent clock
1597 * @boot_offset:  offset that is defined as wall_time - boot_time
1598 *
1599 * Weak dummy function for arches that do not yet support it.
1600 *
1601 * The default function calculates offset based on the current value of
1602 * local_clock(). This way architectures that support sched_clock() but don't
1603 * support dedicated boot time clock will provide the best estimate of the
1604 * boot time.
1605 */
1606void __weak __init
1607read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1608				     struct timespec64 *boot_offset)
1609{
1610	read_persistent_clock64(wall_time);
1611	*boot_offset = ns_to_timespec64(local_clock());
1612}
1613
1614/*
1615 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1616 *
1617 * The flag starts of false and is only set when a suspend reaches
1618 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1619 * timekeeper clocksource is not stopping across suspend and has been
1620 * used to update sleep time. If the timekeeper clocksource has stopped
1621 * then the flag stays true and is used by the RTC resume code to decide
1622 * whether sleeptime must be injected and if so the flag gets false then.
1623 *
1624 * If a suspend fails before reaching timekeeping_resume() then the flag
1625 * stays false and prevents erroneous sleeptime injection.
1626 */
1627static bool suspend_timing_needed;
1628
1629/* Flag for if there is a persistent clock on this platform */
1630static bool persistent_clock_exists;
1631
1632/*
1633 * timekeeping_init - Initializes the clocksource and common timekeeping values
1634 */
1635void __init timekeeping_init(void)
1636{
1637	struct timespec64 wall_time, boot_offset, wall_to_mono;
1638	struct timekeeper *tk = &tk_core.timekeeper;
1639	struct clocksource *clock;
1640	unsigned long flags;
1641
1642	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1643	if (timespec64_valid_settod(&wall_time) &&
1644	    timespec64_to_ns(&wall_time) > 0) {
1645		persistent_clock_exists = true;
1646	} else if (timespec64_to_ns(&wall_time) != 0) {
1647		pr_warn("Persistent clock returned invalid value");
1648		wall_time = (struct timespec64){0};
1649	}
1650
1651	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1652		boot_offset = (struct timespec64){0};
1653
1654	/*
1655	 * We want set wall_to_mono, so the following is true:
1656	 * wall time + wall_to_mono = boot time
1657	 */
1658	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1659
1660	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1661	write_seqcount_begin(&tk_core.seq);
1662	ntp_init();
1663
1664	clock = clocksource_default_clock();
1665	if (clock->enable)
1666		clock->enable(clock);
1667	tk_setup_internals(tk, clock);
1668
1669	tk_set_xtime(tk, &wall_time);
1670	tk->raw_sec = 0;
1671
1672	tk_set_wall_to_mono(tk, wall_to_mono);
1673
1674	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1675
1676	write_seqcount_end(&tk_core.seq);
1677	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1678}
1679
1680/* time in seconds when suspend began for persistent clock */
1681static struct timespec64 timekeeping_suspend_time;
1682
1683/**
1684 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1685 * @tk:		Pointer to the timekeeper to be updated
1686 * @delta:	Pointer to the delta value in timespec64 format
1687 *
1688 * Takes a timespec offset measuring a suspend interval and properly
1689 * adds the sleep offset to the timekeeping variables.
1690 */
1691static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1692					   const struct timespec64 *delta)
1693{
1694	if (!timespec64_valid_strict(delta)) {
1695		printk_deferred(KERN_WARNING
1696				"__timekeeping_inject_sleeptime: Invalid "
1697				"sleep delta value!\n");
1698		return;
1699	}
1700	tk_xtime_add(tk, delta);
1701	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1702	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1703	tk_debug_account_sleep_time(delta);
1704}
1705
1706#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1707/*
1708 * We have three kinds of time sources to use for sleep time
1709 * injection, the preference order is:
1710 * 1) non-stop clocksource
1711 * 2) persistent clock (ie: RTC accessible when irqs are off)
1712 * 3) RTC
1713 *
1714 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1715 * If system has neither 1) nor 2), 3) will be used finally.
1716 *
1717 *
1718 * If timekeeping has injected sleeptime via either 1) or 2),
1719 * 3) becomes needless, so in this case we don't need to call
1720 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1721 * means.
1722 */
1723bool timekeeping_rtc_skipresume(void)
1724{
1725	return !suspend_timing_needed;
1726}
1727
1728/*
1729 * 1) can be determined whether to use or not only when doing
1730 * timekeeping_resume() which is invoked after rtc_suspend(),
1731 * so we can't skip rtc_suspend() surely if system has 1).
1732 *
1733 * But if system has 2), 2) will definitely be used, so in this
1734 * case we don't need to call rtc_suspend(), and this is what
1735 * timekeeping_rtc_skipsuspend() means.
1736 */
1737bool timekeeping_rtc_skipsuspend(void)
1738{
1739	return persistent_clock_exists;
1740}
1741
1742/**
1743 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1744 * @delta: pointer to a timespec64 delta value
1745 *
1746 * This hook is for architectures that cannot support read_persistent_clock64
1747 * because their RTC/persistent clock is only accessible when irqs are enabled.
1748 * and also don't have an effective nonstop clocksource.
1749 *
1750 * This function should only be called by rtc_resume(), and allows
1751 * a suspend offset to be injected into the timekeeping values.
1752 */
1753void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1754{
1755	struct timekeeper *tk = &tk_core.timekeeper;
1756	unsigned long flags;
1757
1758	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1759	write_seqcount_begin(&tk_core.seq);
1760
1761	suspend_timing_needed = false;
1762
1763	timekeeping_forward_now(tk);
1764
1765	__timekeeping_inject_sleeptime(tk, delta);
1766
1767	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1768
1769	write_seqcount_end(&tk_core.seq);
1770	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1771
1772	/* Signal hrtimers about time change */
1773	clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1774}
1775#endif
1776
1777/**
1778 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1779 */
1780void timekeeping_resume(void)
1781{
1782	struct timekeeper *tk = &tk_core.timekeeper;
1783	struct clocksource *clock = tk->tkr_mono.clock;
1784	unsigned long flags;
1785	struct timespec64 ts_new, ts_delta;
1786	u64 cycle_now, nsec;
1787	bool inject_sleeptime = false;
1788
1789	read_persistent_clock64(&ts_new);
1790
1791	clockevents_resume();
1792	clocksource_resume();
1793
1794	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1795	write_seqcount_begin(&tk_core.seq);
1796
1797	/*
1798	 * After system resumes, we need to calculate the suspended time and
1799	 * compensate it for the OS time. There are 3 sources that could be
1800	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1801	 * device.
1802	 *
1803	 * One specific platform may have 1 or 2 or all of them, and the
1804	 * preference will be:
1805	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1806	 * The less preferred source will only be tried if there is no better
1807	 * usable source. The rtc part is handled separately in rtc core code.
1808	 */
1809	cycle_now = tk_clock_read(&tk->tkr_mono);
1810	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1811	if (nsec > 0) {
1812		ts_delta = ns_to_timespec64(nsec);
1813		inject_sleeptime = true;
1814	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1815		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1816		inject_sleeptime = true;
1817	}
1818
1819	if (inject_sleeptime) {
1820		suspend_timing_needed = false;
1821		__timekeeping_inject_sleeptime(tk, &ts_delta);
1822	}
1823
1824	/* Re-base the last cycle value */
1825	tk->tkr_mono.cycle_last = cycle_now;
1826	tk->tkr_raw.cycle_last  = cycle_now;
1827
1828	tk->ntp_error = 0;
1829	timekeeping_suspended = 0;
1830	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1831	write_seqcount_end(&tk_core.seq);
1832	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1833
1834	touch_softlockup_watchdog();
1835
1836	/* Resume the clockevent device(s) and hrtimers */
1837	tick_resume();
1838	/* Notify timerfd as resume is equivalent to clock_was_set() */
1839	timerfd_resume();
1840}
1841
1842int timekeeping_suspend(void)
1843{
1844	struct timekeeper *tk = &tk_core.timekeeper;
1845	unsigned long flags;
1846	struct timespec64		delta, delta_delta;
1847	static struct timespec64	old_delta;
1848	struct clocksource *curr_clock;
1849	u64 cycle_now;
1850
1851	read_persistent_clock64(&timekeeping_suspend_time);
1852
1853	/*
1854	 * On some systems the persistent_clock can not be detected at
1855	 * timekeeping_init by its return value, so if we see a valid
1856	 * value returned, update the persistent_clock_exists flag.
1857	 */
1858	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1859		persistent_clock_exists = true;
1860
1861	suspend_timing_needed = true;
1862
1863	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1864	write_seqcount_begin(&tk_core.seq);
1865	timekeeping_forward_now(tk);
1866	timekeeping_suspended = 1;
1867
1868	/*
1869	 * Since we've called forward_now, cycle_last stores the value
1870	 * just read from the current clocksource. Save this to potentially
1871	 * use in suspend timing.
1872	 */
1873	curr_clock = tk->tkr_mono.clock;
1874	cycle_now = tk->tkr_mono.cycle_last;
1875	clocksource_start_suspend_timing(curr_clock, cycle_now);
1876
1877	if (persistent_clock_exists) {
1878		/*
1879		 * To avoid drift caused by repeated suspend/resumes,
1880		 * which each can add ~1 second drift error,
1881		 * try to compensate so the difference in system time
1882		 * and persistent_clock time stays close to constant.
1883		 */
1884		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1885		delta_delta = timespec64_sub(delta, old_delta);
1886		if (abs(delta_delta.tv_sec) >= 2) {
1887			/*
1888			 * if delta_delta is too large, assume time correction
1889			 * has occurred and set old_delta to the current delta.
1890			 */
1891			old_delta = delta;
1892		} else {
1893			/* Otherwise try to adjust old_system to compensate */
1894			timekeeping_suspend_time =
1895				timespec64_add(timekeeping_suspend_time, delta_delta);
1896		}
1897	}
1898
1899	timekeeping_update(tk, TK_MIRROR);
1900	halt_fast_timekeeper(tk);
1901	write_seqcount_end(&tk_core.seq);
1902	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1903
1904	tick_suspend();
1905	clocksource_suspend();
1906	clockevents_suspend();
1907
1908	return 0;
1909}
1910
1911/* sysfs resume/suspend bits for timekeeping */
1912static struct syscore_ops timekeeping_syscore_ops = {
1913	.resume		= timekeeping_resume,
1914	.suspend	= timekeeping_suspend,
1915};
1916
1917static int __init timekeeping_init_ops(void)
1918{
1919	register_syscore_ops(&timekeeping_syscore_ops);
1920	return 0;
1921}
1922device_initcall(timekeeping_init_ops);
1923
1924/*
1925 * Apply a multiplier adjustment to the timekeeper
1926 */
1927static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1928							 s64 offset,
1929							 s32 mult_adj)
1930{
1931	s64 interval = tk->cycle_interval;
1932
1933	if (mult_adj == 0) {
1934		return;
1935	} else if (mult_adj == -1) {
1936		interval = -interval;
1937		offset = -offset;
1938	} else if (mult_adj != 1) {
1939		interval *= mult_adj;
1940		offset *= mult_adj;
1941	}
1942
1943	/*
1944	 * So the following can be confusing.
1945	 *
1946	 * To keep things simple, lets assume mult_adj == 1 for now.
1947	 *
1948	 * When mult_adj != 1, remember that the interval and offset values
1949	 * have been appropriately scaled so the math is the same.
1950	 *
1951	 * The basic idea here is that we're increasing the multiplier
1952	 * by one, this causes the xtime_interval to be incremented by
1953	 * one cycle_interval. This is because:
1954	 *	xtime_interval = cycle_interval * mult
1955	 * So if mult is being incremented by one:
1956	 *	xtime_interval = cycle_interval * (mult + 1)
1957	 * Its the same as:
1958	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1959	 * Which can be shortened to:
1960	 *	xtime_interval += cycle_interval
1961	 *
1962	 * So offset stores the non-accumulated cycles. Thus the current
1963	 * time (in shifted nanoseconds) is:
1964	 *	now = (offset * adj) + xtime_nsec
1965	 * Now, even though we're adjusting the clock frequency, we have
1966	 * to keep time consistent. In other words, we can't jump back
1967	 * in time, and we also want to avoid jumping forward in time.
1968	 *
1969	 * So given the same offset value, we need the time to be the same
1970	 * both before and after the freq adjustment.
1971	 *	now = (offset * adj_1) + xtime_nsec_1
1972	 *	now = (offset * adj_2) + xtime_nsec_2
1973	 * So:
1974	 *	(offset * adj_1) + xtime_nsec_1 =
1975	 *		(offset * adj_2) + xtime_nsec_2
1976	 * And we know:
1977	 *	adj_2 = adj_1 + 1
1978	 * So:
1979	 *	(offset * adj_1) + xtime_nsec_1 =
1980	 *		(offset * (adj_1+1)) + xtime_nsec_2
1981	 *	(offset * adj_1) + xtime_nsec_1 =
1982	 *		(offset * adj_1) + offset + xtime_nsec_2
1983	 * Canceling the sides:
1984	 *	xtime_nsec_1 = offset + xtime_nsec_2
1985	 * Which gives us:
1986	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1987	 * Which simplifies to:
1988	 *	xtime_nsec -= offset
1989	 */
1990	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1991		/* NTP adjustment caused clocksource mult overflow */
1992		WARN_ON_ONCE(1);
1993		return;
1994	}
1995
1996	tk->tkr_mono.mult += mult_adj;
1997	tk->xtime_interval += interval;
1998	tk->tkr_mono.xtime_nsec -= offset;
1999}
2000
2001/*
2002 * Adjust the timekeeper's multiplier to the correct frequency
2003 * and also to reduce the accumulated error value.
2004 */
2005static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2006{
2007	u32 mult;
2008
2009	/*
2010	 * Determine the multiplier from the current NTP tick length.
2011	 * Avoid expensive division when the tick length doesn't change.
2012	 */
2013	if (likely(tk->ntp_tick == ntp_tick_length())) {
2014		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2015	} else {
2016		tk->ntp_tick = ntp_tick_length();
2017		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2018				 tk->xtime_remainder, tk->cycle_interval);
2019	}
2020
2021	/*
2022	 * If the clock is behind the NTP time, increase the multiplier by 1
2023	 * to catch up with it. If it's ahead and there was a remainder in the
2024	 * tick division, the clock will slow down. Otherwise it will stay
2025	 * ahead until the tick length changes to a non-divisible value.
2026	 */
2027	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2028	mult += tk->ntp_err_mult;
2029
2030	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2031
2032	if (unlikely(tk->tkr_mono.clock->maxadj &&
2033		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2034			> tk->tkr_mono.clock->maxadj))) {
2035		printk_once(KERN_WARNING
2036			"Adjusting %s more than 11%% (%ld vs %ld)\n",
2037			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2038			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2039	}
2040
2041	/*
2042	 * It may be possible that when we entered this function, xtime_nsec
2043	 * was very small.  Further, if we're slightly speeding the clocksource
2044	 * in the code above, its possible the required corrective factor to
2045	 * xtime_nsec could cause it to underflow.
2046	 *
2047	 * Now, since we have already accumulated the second and the NTP
2048	 * subsystem has been notified via second_overflow(), we need to skip
2049	 * the next update.
2050	 */
2051	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2052		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2053							tk->tkr_mono.shift;
2054		tk->xtime_sec--;
2055		tk->skip_second_overflow = 1;
2056	}
2057}
2058
2059/*
2060 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2061 *
2062 * Helper function that accumulates the nsecs greater than a second
2063 * from the xtime_nsec field to the xtime_secs field.
2064 * It also calls into the NTP code to handle leapsecond processing.
2065 */
2066static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2067{
2068	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2069	unsigned int clock_set = 0;
2070
2071	while (tk->tkr_mono.xtime_nsec >= nsecps) {
2072		int leap;
2073
2074		tk->tkr_mono.xtime_nsec -= nsecps;
2075		tk->xtime_sec++;
2076
2077		/*
2078		 * Skip NTP update if this second was accumulated before,
2079		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2080		 */
2081		if (unlikely(tk->skip_second_overflow)) {
2082			tk->skip_second_overflow = 0;
2083			continue;
2084		}
2085
2086		/* Figure out if its a leap sec and apply if needed */
2087		leap = second_overflow(tk->xtime_sec);
2088		if (unlikely(leap)) {
2089			struct timespec64 ts;
2090
2091			tk->xtime_sec += leap;
2092
2093			ts.tv_sec = leap;
2094			ts.tv_nsec = 0;
2095			tk_set_wall_to_mono(tk,
2096				timespec64_sub(tk->wall_to_monotonic, ts));
2097
2098			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2099
2100			clock_set = TK_CLOCK_WAS_SET;
2101		}
2102	}
2103	return clock_set;
2104}
2105
2106/*
2107 * logarithmic_accumulation - shifted accumulation of cycles
2108 *
2109 * This functions accumulates a shifted interval of cycles into
2110 * a shifted interval nanoseconds. Allows for O(log) accumulation
2111 * loop.
2112 *
2113 * Returns the unconsumed cycles.
2114 */
2115static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2116				    u32 shift, unsigned int *clock_set)
2117{
2118	u64 interval = tk->cycle_interval << shift;
2119	u64 snsec_per_sec;
2120
2121	/* If the offset is smaller than a shifted interval, do nothing */
2122	if (offset < interval)
2123		return offset;
2124
2125	/* Accumulate one shifted interval */
2126	offset -= interval;
2127	tk->tkr_mono.cycle_last += interval;
2128	tk->tkr_raw.cycle_last  += interval;
2129
2130	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2131	*clock_set |= accumulate_nsecs_to_secs(tk);
2132
2133	/* Accumulate raw time */
2134	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2135	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2136	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2137		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2138		tk->raw_sec++;
2139	}
2140
2141	/* Accumulate error between NTP and clock interval */
2142	tk->ntp_error += tk->ntp_tick << shift;
2143	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2144						(tk->ntp_error_shift + shift);
2145
2146	return offset;
2147}
2148
2149/*
2150 * timekeeping_advance - Updates the timekeeper to the current time and
2151 * current NTP tick length
2152 */
2153static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2154{
2155	struct timekeeper *real_tk = &tk_core.timekeeper;
2156	struct timekeeper *tk = &shadow_timekeeper;
2157	u64 offset;
2158	int shift = 0, maxshift;
2159	unsigned int clock_set = 0;
2160	unsigned long flags;
2161
2162	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2163
2164	/* Make sure we're fully resumed: */
2165	if (unlikely(timekeeping_suspended))
2166		goto out;
2167
2168	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2169				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2170
2171	/* Check if there's really nothing to do */
2172	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2173		goto out;
2174
2175	/* Do some additional sanity checking */
2176	timekeeping_check_update(tk, offset);
2177
2178	/*
2179	 * With NO_HZ we may have to accumulate many cycle_intervals
2180	 * (think "ticks") worth of time at once. To do this efficiently,
2181	 * we calculate the largest doubling multiple of cycle_intervals
2182	 * that is smaller than the offset.  We then accumulate that
2183	 * chunk in one go, and then try to consume the next smaller
2184	 * doubled multiple.
2185	 */
2186	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2187	shift = max(0, shift);
2188	/* Bound shift to one less than what overflows tick_length */
2189	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2190	shift = min(shift, maxshift);
2191	while (offset >= tk->cycle_interval) {
2192		offset = logarithmic_accumulation(tk, offset, shift,
2193							&clock_set);
2194		if (offset < tk->cycle_interval<<shift)
2195			shift--;
2196	}
2197
2198	/* Adjust the multiplier to correct NTP error */
2199	timekeeping_adjust(tk, offset);
2200
2201	/*
2202	 * Finally, make sure that after the rounding
2203	 * xtime_nsec isn't larger than NSEC_PER_SEC
2204	 */
2205	clock_set |= accumulate_nsecs_to_secs(tk);
2206
2207	write_seqcount_begin(&tk_core.seq);
2208	/*
2209	 * Update the real timekeeper.
2210	 *
2211	 * We could avoid this memcpy by switching pointers, but that
2212	 * requires changes to all other timekeeper usage sites as
2213	 * well, i.e. move the timekeeper pointer getter into the
2214	 * spinlocked/seqcount protected sections. And we trade this
2215	 * memcpy under the tk_core.seq against one before we start
2216	 * updating.
2217	 */
2218	timekeeping_update(tk, clock_set);
2219	memcpy(real_tk, tk, sizeof(*tk));
2220	/* The memcpy must come last. Do not put anything here! */
2221	write_seqcount_end(&tk_core.seq);
2222out:
2223	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2224
2225	return !!clock_set;
2226}
2227
2228/**
2229 * update_wall_time - Uses the current clocksource to increment the wall time
2230 *
2231 */
2232void update_wall_time(void)
2233{
2234	if (timekeeping_advance(TK_ADV_TICK))
2235		clock_was_set_delayed();
2236}
2237
2238/**
2239 * getboottime64 - Return the real time of system boot.
2240 * @ts:		pointer to the timespec64 to be set
2241 *
2242 * Returns the wall-time of boot in a timespec64.
2243 *
2244 * This is based on the wall_to_monotonic offset and the total suspend
2245 * time. Calls to settimeofday will affect the value returned (which
2246 * basically means that however wrong your real time clock is at boot time,
2247 * you get the right time here).
2248 */
2249void getboottime64(struct timespec64 *ts)
2250{
2251	struct timekeeper *tk = &tk_core.timekeeper;
2252	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2253
2254	*ts = ktime_to_timespec64(t);
2255}
2256EXPORT_SYMBOL_GPL(getboottime64);
2257
2258void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2259{
2260	struct timekeeper *tk = &tk_core.timekeeper;
2261	unsigned int seq;
2262
2263	do {
2264		seq = read_seqcount_begin(&tk_core.seq);
2265
2266		*ts = tk_xtime(tk);
2267	} while (read_seqcount_retry(&tk_core.seq, seq));
2268}
2269EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2270
2271void ktime_get_coarse_ts64(struct timespec64 *ts)
2272{
2273	struct timekeeper *tk = &tk_core.timekeeper;
2274	struct timespec64 now, mono;
2275	unsigned int seq;
2276
2277	do {
2278		seq = read_seqcount_begin(&tk_core.seq);
2279
2280		now = tk_xtime(tk);
2281		mono = tk->wall_to_monotonic;
2282	} while (read_seqcount_retry(&tk_core.seq, seq));
2283
2284	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2285				now.tv_nsec + mono.tv_nsec);
2286}
2287EXPORT_SYMBOL(ktime_get_coarse_ts64);
2288
2289/*
2290 * Must hold jiffies_lock
2291 */
2292void do_timer(unsigned long ticks)
2293{
2294	jiffies_64 += ticks;
2295	calc_global_load();
2296}
2297
2298/**
2299 * ktime_get_update_offsets_now - hrtimer helper
2300 * @cwsseq:	pointer to check and store the clock was set sequence number
2301 * @offs_real:	pointer to storage for monotonic -> realtime offset
2302 * @offs_boot:	pointer to storage for monotonic -> boottime offset
2303 * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2304 *
2305 * Returns current monotonic time and updates the offsets if the
2306 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2307 * different.
2308 *
2309 * Called from hrtimer_interrupt() or retrigger_next_event()
2310 */
2311ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2312				     ktime_t *offs_boot, ktime_t *offs_tai)
2313{
2314	struct timekeeper *tk = &tk_core.timekeeper;
2315	unsigned int seq;
2316	ktime_t base;
2317	u64 nsecs;
2318
2319	do {
2320		seq = read_seqcount_begin(&tk_core.seq);
2321
2322		base = tk->tkr_mono.base;
2323		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2324		base = ktime_add_ns(base, nsecs);
2325
2326		if (*cwsseq != tk->clock_was_set_seq) {
2327			*cwsseq = tk->clock_was_set_seq;
2328			*offs_real = tk->offs_real;
2329			*offs_boot = tk->offs_boot;
2330			*offs_tai = tk->offs_tai;
2331		}
2332
2333		/* Handle leapsecond insertion adjustments */
2334		if (unlikely(base >= tk->next_leap_ktime))
2335			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2336
2337	} while (read_seqcount_retry(&tk_core.seq, seq));
2338
2339	return base;
2340}
2341
2342/*
2343 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2344 */
2345static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2346{
2347	if (txc->modes & ADJ_ADJTIME) {
2348		/* singleshot must not be used with any other mode bits */
2349		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2350			return -EINVAL;
2351		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2352		    !capable(CAP_SYS_TIME))
2353			return -EPERM;
2354	} else {
2355		/* In order to modify anything, you gotta be super-user! */
2356		if (txc->modes && !capable(CAP_SYS_TIME))
2357			return -EPERM;
2358		/*
2359		 * if the quartz is off by more than 10% then
2360		 * something is VERY wrong!
2361		 */
2362		if (txc->modes & ADJ_TICK &&
2363		    (txc->tick <  900000/USER_HZ ||
2364		     txc->tick > 1100000/USER_HZ))
2365			return -EINVAL;
2366	}
2367
2368	if (txc->modes & ADJ_SETOFFSET) {
2369		/* In order to inject time, you gotta be super-user! */
2370		if (!capable(CAP_SYS_TIME))
2371			return -EPERM;
2372
2373		/*
2374		 * Validate if a timespec/timeval used to inject a time
2375		 * offset is valid.  Offsets can be positive or negative, so
2376		 * we don't check tv_sec. The value of the timeval/timespec
2377		 * is the sum of its fields,but *NOTE*:
2378		 * The field tv_usec/tv_nsec must always be non-negative and
2379		 * we can't have more nanoseconds/microseconds than a second.
2380		 */
2381		if (txc->time.tv_usec < 0)
2382			return -EINVAL;
2383
2384		if (txc->modes & ADJ_NANO) {
2385			if (txc->time.tv_usec >= NSEC_PER_SEC)
2386				return -EINVAL;
2387		} else {
2388			if (txc->time.tv_usec >= USEC_PER_SEC)
2389				return -EINVAL;
2390		}
2391	}
2392
2393	/*
2394	 * Check for potential multiplication overflows that can
2395	 * only happen on 64-bit systems:
2396	 */
2397	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2398		if (LLONG_MIN / PPM_SCALE > txc->freq)
2399			return -EINVAL;
2400		if (LLONG_MAX / PPM_SCALE < txc->freq)
2401			return -EINVAL;
2402	}
2403
2404	return 0;
2405}
2406
2407/**
2408 * random_get_entropy_fallback - Returns the raw clock source value,
2409 * used by random.c for platforms with no valid random_get_entropy().
2410 */
2411unsigned long random_get_entropy_fallback(void)
2412{
2413	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2414	struct clocksource *clock = READ_ONCE(tkr->clock);
2415
2416	if (unlikely(timekeeping_suspended || !clock))
2417		return 0;
2418	return clock->read(clock);
2419}
2420EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2421
2422/**
2423 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2424 */
2425int do_adjtimex(struct __kernel_timex *txc)
2426{
2427	struct timekeeper *tk = &tk_core.timekeeper;
2428	struct audit_ntp_data ad;
2429	bool clock_set = false;
2430	struct timespec64 ts;
2431	unsigned long flags;
2432	s32 orig_tai, tai;
2433	int ret;
2434
2435	/* Validate the data before disabling interrupts */
2436	ret = timekeeping_validate_timex(txc);
2437	if (ret)
2438		return ret;
2439	add_device_randomness(txc, sizeof(*txc));
2440
2441	if (txc->modes & ADJ_SETOFFSET) {
2442		struct timespec64 delta;
2443		delta.tv_sec  = txc->time.tv_sec;
2444		delta.tv_nsec = txc->time.tv_usec;
2445		if (!(txc->modes & ADJ_NANO))
2446			delta.tv_nsec *= 1000;
2447		ret = timekeeping_inject_offset(&delta);
2448		if (ret)
2449			return ret;
2450
2451		audit_tk_injoffset(delta);
2452	}
2453
2454	audit_ntp_init(&ad);
2455
2456	ktime_get_real_ts64(&ts);
2457	add_device_randomness(&ts, sizeof(ts));
2458
2459	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2460	write_seqcount_begin(&tk_core.seq);
2461
2462	orig_tai = tai = tk->tai_offset;
2463	ret = __do_adjtimex(txc, &ts, &tai, &ad);
2464
2465	if (tai != orig_tai) {
2466		__timekeeping_set_tai_offset(tk, tai);
2467		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2468		clock_set = true;
2469	}
2470	tk_update_leap_state(tk);
2471
2472	write_seqcount_end(&tk_core.seq);
2473	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2474
2475	audit_ntp_log(&ad);
2476
2477	/* Update the multiplier immediately if frequency was set directly */
2478	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2479		clock_set |= timekeeping_advance(TK_ADV_FREQ);
2480
2481	if (clock_set)
2482		clock_was_set(CLOCK_REALTIME);
2483
2484	ntp_notify_cmos_timer();
2485
2486	return ret;
2487}
2488
2489#ifdef CONFIG_NTP_PPS
2490/**
2491 * hardpps() - Accessor function to NTP __hardpps function
2492 */
2493void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2494{
2495	unsigned long flags;
2496
2497	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2498	write_seqcount_begin(&tk_core.seq);
2499
2500	__hardpps(phase_ts, raw_ts);
2501
2502	write_seqcount_end(&tk_core.seq);
2503	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2504}
2505EXPORT_SYMBOL(hardpps);
2506#endif /* CONFIG_NTP_PPS */
2507