linux - kernel/time/timekeeping.c

3412 lines · 97.7 KiB

1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Kernel timekeeping code and accessor functions. Based on code from
4	* timer.c, moved in commit 8524070b7982.
5	*/
6	#include <linux/audit.h>
7	#include <linux/clocksource.h>
8	#include <linux/compiler.h>
9	#include <linux/jiffies.h>
10	#include <linux/kobject.h>
11	#include <linux/module.h>
12	#include <linux/nmi.h>
13	#include <linux/pvclock_gtod.h>
14	#include <linux/random.h>
15	#include <linux/sched/clock.h>
16	#include <linux/sched/loadavg.h>
17	#include <linux/static_key.h>
18	#include <linux/stop_machine.h>
19	#include <linux/syscore_ops.h>
20	#include <linux/tick.h>
21	#include <linux/time.h>
22	#include <linux/timex.h>
23	#include <linux/timekeeper_internal.h>
24
25	#include <vdso/auxclock.h>
26
27	#include "tick-internal.h"
28	#include "timekeeping_internal.h"
29	#include "ntp_internal.h"
30
31	#define TK_CLEAR_NTP (1 << 0)
32	#define TK_CLOCK_WAS_SET (1 << 1)
33
34	#define TK_UPDATE_ALL (TK_CLEAR_NTP \| TK_CLOCK_WAS_SET)
35
36	enum 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
44	/*
45	* The most important data for readout fits into a single 64 byte
46	* cache line.
47	*/
48	struct tk_data {
49	seqcount_raw_spinlock_t seq;
50	struct timekeeper timekeeper;
51	struct timekeeper shadow_timekeeper;
52	raw_spinlock_t lock;
53	} ____cacheline_aligned;
54
55	static struct tk_data timekeeper_data[TIMEKEEPERS_MAX];
56
57	/* The core timekeeper */
58	#define tk_core (timekeeper_data[TIMEKEEPER_CORE])
59
60	#ifdef CONFIG_POSIX_AUX_CLOCKS
61	static inline bool tk_get_aux_ts64(unsigned int tkid, struct timespec64 *ts)
62	{
63	return ktime_get_aux_ts64(CLOCK_AUX + tkid - TIMEKEEPER_AUX_FIRST, ts);
64	}
65
66	static inline bool tk_is_aux(const struct timekeeper *tk)
67	{
68	return tk->id >= TIMEKEEPER_AUX_FIRST && tk->id <= TIMEKEEPER_AUX_LAST;
69	}
70	static inline struct tk_data *aux_get_tk_data(clockid_t id);
71	#else
72	static inline bool tk_get_aux_ts64(unsigned int tkid, struct timespec64 *ts)
73	{
74	return false;
75	}
76
77	static inline bool tk_is_aux(const struct timekeeper *tk)
78	{
79	return false;
80	}
81	static inline struct tk_data *aux_get_tk_data(clockid_t id)
82	{
83	return NULL;
84	}
85	#endif
86
87	static inline void tk_update_aux_offs(struct timekeeper *tk, ktime_t offs)
88	{
89	tk->offs_aux = offs;
90	tk->monotonic_to_aux = ktime_to_timespec64(offs);
91	}
92
93	/* flag for if timekeeping is suspended */
94	int __read_mostly timekeeping_suspended;
95
96	/**
97	* struct tk_fast - NMI safe timekeeper
98	* @seq: Sequence counter for protecting updates. The lowest bit
99	* is the index for the tk_read_base array
100	* @base: tk_read_base array. Access is indexed by the lowest bit of
101	* @seq.
102	*
103	* See @update_fast_timekeeper() below.
104	*/
105	struct tk_fast {
106	seqcount_latch_t seq;
107	struct tk_read_base base[2];
108	};
109
110	/* Suspend-time cycles value for halted fast timekeeper. */
111	static u64 cycles_at_suspend;
112
113	static u64 dummy_clock_read(struct clocksource *cs)
114	{
115	if (timekeeping_suspended)
116	return cycles_at_suspend;
117	return local_clock();
118	}
119
120	static struct clocksource dummy_clock = {
121	.read = dummy_clock_read,
122	};
123
124	/*
125	* Boot time initialization which allows local_clock() to be utilized
126	* during early boot when clocksources are not available. local_clock()
127	* returns nanoseconds already so no conversion is required, hence mult=1
128	* and shift=0. When the first proper clocksource is installed then
129	* the fast time keepers are updated with the correct values.
130	*/
131	#define FAST_TK_INIT \
132	{ \
133	.clock = &dummy_clock, \
134	.mask = CLOCKSOURCE_MASK(64), \
135	.mult = 1, \
136	.shift = 0, \
137	}
138
139	static struct tk_fast tk_fast_mono ____cacheline_aligned = {
140	.seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
141	.base[0] = FAST_TK_INIT,
142	.base[1] = FAST_TK_INIT,
143	};
144
145	static struct tk_fast tk_fast_raw ____cacheline_aligned = {
146	.seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
147	.base[0] = FAST_TK_INIT,
148	.base[1] = FAST_TK_INIT,
149	};
150
151	#ifdef CONFIG_POSIX_AUX_CLOCKS
152	static __init void tk_aux_setup(void);
153	static void tk_aux_update_clocksource(void);
154	static void tk_aux_advance(void);
155	#else
156	static inline void tk_aux_setup(void) { }
157	static inline void tk_aux_update_clocksource(void) { }
158	static inline void tk_aux_advance(void) { }
159	#endif
160
161	unsigned long timekeeper_lock_irqsave(void)
162	{
163	unsigned long flags;
164
165	raw_spin_lock_irqsave(&tk_core.lock, flags);
166	return flags;
167	}
168
169	void timekeeper_unlock_irqrestore(unsigned long flags)
170	{
171	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
172	}
173
174	/*
175	* Multigrain timestamps require tracking the latest fine-grained timestamp
176	* that has been issued, and never returning a coarse-grained timestamp that is
177	* earlier than that value.
178	*
179	* mg_floor represents the latest fine-grained time that has been handed out as
180	* a file timestamp on the system. This is tracked as a monotonic ktime_t, and
181	* converted to a realtime clock value on an as-needed basis.
182	*
183	* Maintaining mg_floor ensures the multigrain interfaces never issue a
184	* timestamp earlier than one that has been previously issued.
185	*
186	* The exception to this rule is when there is a backward realtime clock jump. If
187	* such an event occurs, a timestamp can appear to be earlier than a previous one.
188	*/
189	static __cacheline_aligned_in_smp atomic64_t mg_floor;
190
191	static inline void tk_normalize_xtime(struct timekeeper *tk)
192	{
193	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
194	tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
195	tk->xtime_sec++;
196	}
197	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
198	tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
199	tk->raw_sec++;
200	}
201	}
202
203	static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
204	{
205	struct timespec64 ts;
206
207	ts.tv_sec = tk->xtime_sec;
208	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
209	return ts;
210	}
211
212	static inline struct timespec64 tk_xtime_coarse(const struct timekeeper *tk)
213	{
214	struct timespec64 ts;
215
216	ts.tv_sec = tk->xtime_sec;
217	ts.tv_nsec = tk->coarse_nsec;
218	return ts;
219	}
220
221	/*
222	* Update the nanoseconds part for the coarse time keepers. They can't rely
223	* on xtime_nsec because xtime_nsec could be adjusted by a small negative
224	* amount when the multiplication factor of the clock is adjusted, which
225	* could cause the coarse clocks to go slightly backwards. See
226	* timekeeping_apply_adjustment(). Thus we keep a separate copy for the coarse
227	* clockids which only is updated when the clock has been set or we have
228	* accumulated time.
229	*/
230	static inline void tk_update_coarse_nsecs(struct timekeeper *tk)
231	{
232	tk->coarse_nsec = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
233	}
234
235	static void tk_set_xtime(struct timekeeper tk, const struct timespec64 ts)
236	{
237	tk->xtime_sec = ts->tv_sec;
238	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
239	tk_update_coarse_nsecs(tk);
240	}
241
242	static void tk_xtime_add(struct timekeeper tk, const struct timespec64 ts)
243	{
244	tk->xtime_sec += ts->tv_sec;
245	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
246	tk_normalize_xtime(tk);
247	tk_update_coarse_nsecs(tk);
248	}
249
250	static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
251	{
252	struct timespec64 tmp;
253
254	/*
255	* Verify consistency of: offset_real = -wall_to_monotonic
256	* before modifying anything
257	*/
258	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
259	-tk->wall_to_monotonic.tv_nsec);
260	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
261	tk->wall_to_monotonic = wtm;
262	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
263	/* Paired with READ_ONCE() in ktime_mono_to_any() */
264	WRITE_ONCE(tk->offs_real, timespec64_to_ktime(tmp));
265	WRITE_ONCE(tk->offs_tai, ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0)));
266	}
267
268	static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
269	{
270	/* Paired with READ_ONCE() in ktime_mono_to_any() */
271	WRITE_ONCE(tk->offs_boot, ktime_add(tk->offs_boot, delta));
272	/*
273	* Timespec representation for VDSO update to avoid 64bit division
274	* on every update.
275	*/
276	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
277	}
278
279	#ifdef CONFIG_ARCH_WANTS_CLOCKSOURCE_READ_INLINE
280	#include <asm/clock_inlined.h>
281
282	static DEFINE_STATIC_KEY_FALSE(clocksource_read_inlined);
283
284	/*
285	* tk_clock_read - atomic clocksource read() helper
286	*
287	* This helper is necessary to use in the read paths because, while the
288	* seqcount ensures we don't return a bad value while structures are updated,
289	* it doesn't protect from potential crashes. There is the possibility that
290	* the tkr's clocksource may change between the read reference, and the
291	* clock reference passed to the read function. This can cause crashes if
292	* the wrong clocksource is passed to the wrong read function.
293	* This isn't necessary to use when holding the tk_core.lock or doing
294	* a read of the fast-timekeeper tkrs (which is protected by its own locking
295	* and update logic).
296	*/
297	static __always_inline u64 tk_clock_read(const struct tk_read_base *tkr)
298	{
299	struct clocksource *clock = READ_ONCE(tkr->clock);
300
301	if (static_branch_likely(&clocksource_read_inlined))
302	return arch_inlined_clocksource_read(clock);
303
304	return clock->read(clock);
305	}
306
307	static inline void clocksource_disable_inline_read(void)
308	{
309	static_branch_disable(&clocksource_read_inlined);
310	}
311
312	static inline void clocksource_enable_inline_read(void)
313	{
314	static_branch_enable(&clocksource_read_inlined);
315	}
316	#else
317	static __always_inline u64 tk_clock_read(const struct tk_read_base *tkr)
318	{
319	struct clocksource *clock = READ_ONCE(tkr->clock);
320
321	return clock->read(clock);
322	}
323
324	static inline void clocksource_disable_inline_read(void) { }
325	static inline void clocksource_enable_inline_read(void) { }
326	#endif
327
328	/**
329	* tk_setup_internals - Set up internals to use clocksource clock.
330	*
331	* @tk: The target timekeeper to setup.
332	* @clock: Pointer to clocksource.
333	*
334	* Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
335	* pair and interval request.
336	*
337	* Unless you're the timekeeping code, you should not be using this!
338	*/
339	static void tk_setup_internals(struct timekeeper tk, struct clocksource clock)
340	{
341	u64 interval;
342	struct clocksource *old_clock;
343
344	++tk->cs_was_changed_seq;
345	old_clock = tk->tkr_mono.clock;
346	tk->tkr_mono.clock = clock;
347	tk->tkr_mono.mask = clock->mask;
348	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
349
350	tk->tkr_raw.clock = clock;
351	tk->tkr_raw.mask = clock->mask;
352	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
353
354	/* Do the ns -> cycle conversion first, using original mult */
355	interval = (u64)NTP_INTERVAL_LENGTH << clock->shift;
356	interval += clock->mult / 2;
357	do_div(interval, clock->mult);
358	if (interval == 0)
359	interval = 1;
360
361	tk->cycle_interval = interval;
362
363	/* Go back from cycles -> shifted ns */
364	tk->xtime_interval = interval * clock->mult;
365	tk->raw_interval = interval * clock->mult;
366
367	/* if changing clocks, convert xtime_nsec shift units */
368	if (old_clock) {
369	int shift_change = clock->shift - old_clock->shift;
370	if (shift_change < 0) {
371	tk->tkr_mono.xtime_nsec >>= -shift_change;
372	tk->tkr_raw.xtime_nsec >>= -shift_change;
373	} else {
374	tk->tkr_mono.xtime_nsec <<= shift_change;
375	tk->tkr_raw.xtime_nsec <<= shift_change;
376	}
377	}
378
379	tk->tkr_mono.shift = clock->shift;
380	tk->tkr_raw.shift = clock->shift;
381
382	tk->ntp_error = 0;
383	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
384
385	/*
386	* ntp_tick is the tick length that NTP disciplines (its ±500 PPM
387	* scales only this part), in NTP-shifted ns: the real interval of
388	* a whole number of counter cycles. Because cycle_interval is
389	* rounded to an integer number of cycles, this ntp_tick differs
390	* from the true intended 1/HZ tick length by up to half a cycle
391	* period.
392	*/
393	tk->ntp_tick = (u64)tk->xtime_interval << tk->ntp_error_shift;
394
395	/*
396	* cs_tick_adj is the constant difference between the disciplined
397	* ntp_tick above and the true 1/HZ tick, expressed per-second to
398	* match the ntp_update_frequency() addends and handed to NTP via
399	* ntp_clear() to be explicitly included in its tick_length.
400	*
401	* Worked example: HZ=1000, ACPI PM timer at 3.579545 MHz, which
402	* has 3579.545 cycles in 1ms, rounded to cycle_interval = 3580.
403	*
404	* So ntp_tick is actually 1.000127ms, as that is the amount of
405	* time that 3580 cycles will take at the nominal frequency. This
406	* is the part that NTP disciplines, causing each 3580 counts to
407	* advance the clock by up to NTP's ±500PPM of that amount.
408	*
409	* The "extra" 127ns/tick is what's stored in cs_tick_adj and
410	* applied as a constant correction by ntp_update_frequency() so
411	* that NTP believes it's disciplining a 1ms tick.
412	*/
413	tk->cs_tick_adj = (s64)tk->ntp_tick -
414	((s64)NTP_INTERVAL_LENGTH << NTP_SCALE_SHIFT);
415	tk->cs_tick_adj *= NTP_INTERVAL_FREQ;
416
417	/*
418	* The timekeeper keeps its own mult values for the currently
419	* active clocksource. These value will be adjusted via NTP
420	* to counteract clock drifting.
421	*/
422	tk->tkr_mono.mult = clock->mult;
423	tk->tkr_raw.mult = clock->mult;
424	tk->ntp_err_mult = 0;
425	tk->skip_second_overflow = 0;
426	tk->skew_delta = 0;
427
428	tk->cs_id = clock->id;
429
430	/* Coupled clockevent data */
431	if (IS_ENABLED(CONFIG_GENERIC_CLOCKEVENTS_COUPLED) &&
432	clock->flags & CLOCK_SOURCE_HAS_COUPLED_CLOCK_EVENT) {
433	/*
434	* Aim for an one hour maximum delta and use KHz to handle
435	* clocksources with a frequency above 4GHz correctly as
436	* the frequency argument of clocks_calc_mult_shift() is u32.
437	*/
438	clocks_calc_mult_shift(&tk->cs_ns_to_cyc_mult, &tk->cs_ns_to_cyc_shift,
439	NSEC_PER_MSEC, clock->freq_khz, 3600 * 1000);
440	/*
441	* Initialize the conversion limit as the previous clocksource
442	* might have the same shift/mult pair so the quick check in
443	* tk_update_ns_to_cyc() fails to update it after a clocksource
444	* change leaving it effectivly zero.
445	*/
446	tk->cs_ns_to_cyc_maxns = div_u64(clock->mask, tk->cs_ns_to_cyc_mult);
447	}
448	}
449
450	/* Timekeeper helper functions. */
451	static noinline u64 delta_to_ns_safe(const struct tk_read_base *tkr, u64 delta)
452	{
453	return mul_u64_u32_add_u64_shr(delta, tkr->mult, tkr->xtime_nsec, tkr->shift);
454	}
455
456	static __always_inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
457	{
458	/* Calculate the delta since the last update_wall_time() */
459	u64 mask = tkr->mask, delta = (cycles - tkr->cycle_last) & mask;
460
461	/*
462	* This detects both negative motion and the case where the delta
463	* overflows the multiplication with tkr->mult.
464	*/
465	if (unlikely(delta > tkr->clock->max_cycles)) {
466	/*
467	* Handle clocksource inconsistency between CPUs to prevent
468	* time from going backwards by checking for the MSB of the
469	* mask being set in the delta.
470	*/
471	if (delta & ~(mask >> 1))
472	return tkr->xtime_nsec >> tkr->shift;
473
474	return delta_to_ns_safe(tkr, delta);
475	}
476
477	return ((delta * tkr->mult) + tkr->xtime_nsec) >> tkr->shift;
478	}
479
480	static __always_inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
481	{
482	return timekeeping_cycles_to_ns(tkr, tk_clock_read(tkr));
483	}
484
485	/**
486	* update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
487	* @tkr: Timekeeping readout base from which we take the update
488	* @tkf: Pointer to NMI safe timekeeper
489	*
490	* We want to use this from any context including NMI and tracing /
491	* instrumenting the timekeeping code itself.
492	*
493	* Employ the latch technique; see @write_seqcount_latch.
494	*
495	* So if a NMI hits the update of base[0] then it will use base[1]
496	* which is still consistent. In the worst case this can result is a
497	* slightly wrong timestamp (a few nanoseconds). See
498	* @ktime_get_mono_fast_ns.
499	*/
500	static void update_fast_timekeeper(const struct tk_read_base *tkr,
501	struct tk_fast *tkf)
502	{
503	struct tk_read_base *base = tkf->base;
504
505	/* Force readers off to base[1] */
506	write_seqcount_latch_begin(&tkf->seq);
507
508	/* Update base[0] */
509	memcpy(base, tkr, sizeof(*base));
510
511	/* Force readers back to base[0] */
512	write_seqcount_latch(&tkf->seq);
513
514	/* Update base[1] */
515	memcpy(base + 1, base, sizeof(*base));
516
517	write_seqcount_latch_end(&tkf->seq);
518	}
519
520	static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
521	{
522	struct tk_read_base *tkr;
523	unsigned int seq;
524	u64 now;
525
526	do {
527	seq = read_seqcount_latch(&tkf->seq);
528	tkr = tkf->base + (seq & 0x01);
529	now = ktime_to_ns(tkr->base);
530	now += timekeeping_get_ns(tkr);
531	} while (read_seqcount_latch_retry(&tkf->seq, seq));
532
533	return now;
534	}
535
536	/**
537	* ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
538	*
539	* This timestamp is not guaranteed to be monotonic across an update.
540	* The timestamp is calculated by:
541	*
542	* now = base_mono + clock_delta * slope
543	*
544	* So if the update lowers the slope, readers who are forced to the
545	* not yet updated second array are still using the old steeper slope.
546	*
547	* tmono
548	* ^
549	* \| o n
550	* \| o n
551	* \| u
552	* \| o
553	* \|o
554	* \|12345678---> reader order
555	*
556	* o = old slope
557	* u = update
558	* n = new slope
559	*
560	* So reader 6 will observe time going backwards versus reader 5.
561	*
562	* While other CPUs are likely to be able to observe that, the only way
563	* for a CPU local observation is when an NMI hits in the middle of
564	* the update. Timestamps taken from that NMI context might be ahead
565	* of the following timestamps. Callers need to be aware of that and
566	* deal with it.
567	*/
568	u64 notrace ktime_get_mono_fast_ns(void)
569	{
570	return __ktime_get_fast_ns(&tk_fast_mono);
571	}
572	EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
573
574	/**
575	* ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
576	*
577	* Contrary to ktime_get_mono_fast_ns() this is always correct because the
578	* conversion factor is not affected by NTP/PTP correction.
579	*/
580	u64 notrace ktime_get_raw_fast_ns(void)
581	{
582	return __ktime_get_fast_ns(&tk_fast_raw);
583	}
584	EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
585
586	/**
587	* ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
588	*
589	* To keep it NMI safe since we're accessing from tracing, we're not using a
590	* separate timekeeper with updates to monotonic clock and boot offset
591	* protected with seqcounts. This has the following minor side effects:
592	*
593	* (1) Its possible that a timestamp be taken after the boot offset is updated
594	* but before the timekeeper is updated. If this happens, the new boot offset
595	* is added to the old timekeeping making the clock appear to update slightly
596	* earlier:
597	* CPU 0 CPU 1
598	* timekeeping_inject_sleeptime64()
599	* __timekeeping_inject_sleeptime(tk, delta);
600	* timestamp();
601	* timekeeping_update_staged(tkd, TK_CLEAR_NTP...);
602	*
603	* (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
604	* partially updated. Since the tk->offs_boot update is a rare event, this
605	* should be a rare occurrence which postprocessing should be able to handle.
606	*
607	* The caveats vs. timestamp ordering as documented for ktime_get_mono_fast_ns()
608	* apply as well.
609	*/
610	u64 notrace ktime_get_boot_fast_ns(void)
611	{
612	struct timekeeper *tk = &tk_core.timekeeper;
613
614	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
615	}
616	EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
617
618	/**
619	* ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
620	*
621	* The same limitations as described for ktime_get_boot_fast_ns() apply. The
622	* mono time and the TAI offset are not read atomically which may yield wrong
623	* readouts. However, an update of the TAI offset is an rare event e.g., caused
624	* by settime or adjtimex with an offset. The user of this function has to deal
625	* with the possibility of wrong timestamps in post processing.
626	*/
627	u64 notrace ktime_get_tai_fast_ns(void)
628	{
629	struct timekeeper *tk = &tk_core.timekeeper;
630
631	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
632	}
633	EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
634
635	/**
636	* ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
637	*
638	* See ktime_get_mono_fast_ns() for documentation of the time stamp ordering.
639	*/
640	u64 ktime_get_real_fast_ns(void)
641	{
642	struct tk_fast *tkf = &tk_fast_mono;
643	struct tk_read_base *tkr;
644	u64 baser, delta;
645	unsigned int seq;
646
647	do {
648	seq = raw_read_seqcount_latch(&tkf->seq);
649	tkr = tkf->base + (seq & 0x01);
650	baser = ktime_to_ns(tkr->base_real);
651	delta = timekeeping_get_ns(tkr);
652	} while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
653
654	return baser + delta;
655	}
656	EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
657
658	/**
659	* halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
660	* @tk: Timekeeper to snapshot.
661	*
662	* It generally is unsafe to access the clocksource after timekeeping has been
663	* suspended, so take a snapshot of the readout base of @tk and use it as the
664	* fast timekeeper's readout base while suspended. It will return the same
665	* number of cycles every time until timekeeping is resumed at which time the
666	* proper readout base for the fast timekeeper will be restored automatically.
667	*/
668	static void halt_fast_timekeeper(const struct timekeeper *tk)
669	{
670	static struct tk_read_base tkr_dummy;
671	const struct tk_read_base *tkr = &tk->tkr_mono;
672
673	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
674	cycles_at_suspend = tk_clock_read(tkr);
675	tkr_dummy.clock = &dummy_clock;
676	tkr_dummy.base_real = tkr->base + tk->offs_real;
677	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
678
679	tkr = &tk->tkr_raw;
680	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
681	tkr_dummy.clock = &dummy_clock;
682	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
683	}
684
685	static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
686
687	static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
688	{
689	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
690	}
691
692	/**
693	* pvclock_gtod_register_notifier - register a pvclock timedata update listener
694	* @nb: Pointer to the notifier block to register
695	*/
696	int pvclock_gtod_register_notifier(struct notifier_block *nb)
697	{
698	struct timekeeper *tk = &tk_core.timekeeper;
699	int ret;
700
701	guard(raw_spinlock_irqsave)(&tk_core.lock);
702	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
703	update_pvclock_gtod(tk, true);
704
705	return ret;
706	}
707	EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
708
709	/**
710	* pvclock_gtod_unregister_notifier - unregister a pvclock
711	* timedata update listener
712	* @nb: Pointer to the notifier block to unregister
713	*/
714	int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
715	{
716	guard(raw_spinlock_irqsave)(&tk_core.lock);
717	return raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
718	}
719	EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
720
721	/*
722	* tk_update_leap_state - helper to update the next_leap_ktime
723	*/
724	static inline void tk_update_leap_state(struct timekeeper *tk)
725	{
726	tk->next_leap_ktime = ntp_get_next_leap(tk->id);
727	if (tk->next_leap_ktime != KTIME_MAX)
728	/* Convert to monotonic time */
729	tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
730	}
731
732	/*
733	* Leap state update for both shadow and the real timekeeper
734	* Separate to spare a full memcpy() of the timekeeper.
735	*/
736	static void tk_update_leap_state_all(struct tk_data *tkd)
737	{
738	write_seqcount_begin(&tkd->seq);
739	tk_update_leap_state(&tkd->shadow_timekeeper);
740	tkd->timekeeper.next_leap_ktime = tkd->shadow_timekeeper.next_leap_ktime;
741	write_seqcount_end(&tkd->seq);
742	}
743
744	/*
745	* Update the ktime_t based scalar nsec members of the timekeeper
746	*/
747	static inline void tk_update_ktime_data(struct timekeeper *tk)
748	{
749	u64 seconds;
750	u32 nsec;
751
752	/*
753	* The xtime based monotonic readout is:
754	* nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
755	* The ktime based monotonic readout is:
756	* nsec = base_mono + now();
757	* ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
758	*/
759	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
760	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
761	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
762
763	/*
764	* The sum of the nanoseconds portions of xtime and
765	* wall_to_monotonic can be greater/equal one second. Take
766	* this into account before updating tk->ktime_sec.
767	*/
768	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
769	if (nsec >= NSEC_PER_SEC)
770	seconds++;
771	tk->ktime_sec = seconds;
772
773	/* Update the monotonic raw base */
774	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
775	}
776
777	static inline void tk_update_ns_to_cyc(struct timekeeper tks, struct timekeeper tkc)
778	{
779	struct tk_read_base *tkrs = &tks->tkr_mono;
780	struct tk_read_base *tkrc = &tkc->tkr_mono;
781	unsigned int shift;
782
783	if (!IS_ENABLED(CONFIG_GENERIC_CLOCKEVENTS_COUPLED) \|\|
784	!(tkrs->clock->flags & CLOCK_SOURCE_HAS_COUPLED_CLOCK_EVENT))
785	return;
786
787	if (tkrs->mult == tkrc->mult && tkrs->shift == tkrc->shift)
788	return;
789	/*
790	* The conversion math is simple:
791	*
792	* CS::MULT (1 << NS_TO_CYC_SHIFT)
793	* --------------- = ----------------------
794	* (1 << CS:SHIFT) NS_TO_CYC_MULT
795	*
796	* Ergo:
797	*
798	* NS_TO_CYC_MULT = (1 << (CS::SHIFT + NS_TO_CYC_SHIFT)) / CS::MULT
799	*
800	* NS_TO_CYC_SHIFT has been set up in tk_setup_internals()
801	*/
802	shift = tkrs->shift + tks->cs_ns_to_cyc_shift;
803	tks->cs_ns_to_cyc_mult = (u32)div_u64(1ULL << shift, tkrs->mult);
804	tks->cs_ns_to_cyc_maxns = div_u64(tkrs->clock->mask, tks->cs_ns_to_cyc_mult);
805	}
806
807	/*
808	* Restore the shadow timekeeper from the real timekeeper.
809	*/
810	static void timekeeping_restore_shadow(struct tk_data *tkd)
811	{
812	lockdep_assert_held(&tkd->lock);
813	memcpy(&tkd->shadow_timekeeper, &tkd->timekeeper, sizeof(tkd->timekeeper));
814	}
815
816	static void timekeeping_update_from_shadow(struct tk_data *tkd, unsigned int action)
817	{
818	struct timekeeper *tk = &tkd->shadow_timekeeper;
819
820	lockdep_assert_held(&tkd->lock);
821
822	/*
823	* Block out readers before running the updates below because that
824	* updates VDSO and other time related infrastructure. Not blocking
825	* the readers might let a reader see time going backwards when
826	* reading from the VDSO after the VDSO update and then reading in
827	* the kernel from the timekeeper before that got updated.
828	*/
829	write_seqcount_begin(&tkd->seq);
830
831	if (action & TK_CLEAR_NTP) {
832	tk->ntp_error = 0;
833	ntp_clear(tk->id, tk->cs_tick_adj);
834	}
835
836	tk_update_leap_state(tk);
837	tk_update_ktime_data(tk);
838	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
839
840	if (tk->id == TIMEKEEPER_CORE) {
841	tk_update_ns_to_cyc(tk, &tkd->timekeeper);
842	update_vsyscall(tk);
843	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
844
845	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
846	update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
847	} else if (tk_is_aux(tk)) {
848	vdso_time_update_aux(tk);
849	}
850
851	if (action & TK_CLOCK_WAS_SET)
852	tk->clock_was_set_seq++;
853
854	/*
855	* Update the real timekeeper.
856	*
857	* We could avoid this memcpy() by switching pointers, but that has
858	* the downside that the reader side does not longer benefit from
859	* the cacheline optimized data layout of the timekeeper and requires
860	* another indirection.
861	*
862	* Write xtime_sec first so that even if the memcpy() tears the store
863	* data integrity is provided for ktime_get_real_seconds().
864	*/
865	WRITE_ONCE(tkd->timekeeper.xtime_sec, tk->xtime_sec);
866	memcpy(&tkd->timekeeper, tk, sizeof(*tk));
867	write_seqcount_end(&tkd->seq);
868	}
869
870	/**
871	* timekeeping_forward_now - update clock to the current time
872	* @tk: Pointer to the timekeeper to update
873	*
874	* Forward the current clock to update its state since the last call to
875	* update_wall_time(). This is useful before significant clock changes,
876	* as it avoids having to deal with this time offset explicitly.
877	*/
878	static void timekeeping_forward_now(struct timekeeper *tk)
879	{
880	u64 cycle_now, delta;
881
882	cycle_now = tk_clock_read(&tk->tkr_mono);
883	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask,
884	tk->tkr_mono.clock->max_raw_delta);
885	tk->tkr_mono.cycle_last = cycle_now;
886	tk->tkr_raw.cycle_last = cycle_now;
887
888	while (delta > 0) {
889	u64 max = tk->tkr_mono.clock->max_cycles;
890	u64 incr = delta < max ? delta : max;
891
892	tk->tkr_mono.xtime_nsec += incr * tk->tkr_mono.mult;
893	tk->tkr_raw.xtime_nsec += incr * tk->tkr_raw.mult;
894	tk_normalize_xtime(tk);
895	delta -= incr;
896	}
897	tk_update_coarse_nsecs(tk);
898	}
899
900	/*
901	* ktime_expiry_to_cycles - Convert a expiry time to clocksource cycles
902	* @id: Clocksource ID which is required for validity
903	* @expires_ns: Absolute CLOCK_MONOTONIC expiry time (nsecs) to be converted
904	* @cycles: Pointer to storage for corresponding absolute cycles value
905	*
906	* Convert a CLOCK_MONOTONIC based absolute expiry time to a cycles value
907	* based on the correlated clocksource of the clockevent device by using
908	* the base nanoseconds and cycles values of the last timekeeper update and
909	* converting the delta between @expires_ns and base nanoseconds to cycles.
910	*
911	* This only works for clockevent devices which are using a less than or
912	* equal comparator against the clocksource.
913	*
914	* Utilizing this avoids two clocksource reads for such devices, the
915	* ktime_get() in clockevents_program_event() to calculate the delta expiry
916	* value and the readout in the device::set_next_event() callback to
917	* convert the delta back to a absolute comparator value.
918	*
919	* Returns: True if @id matches the current clocksource ID, false otherwise
920	*/
921	bool ktime_expiry_to_cycles(enum clocksource_ids id, ktime_t expires_ns, u64 *cycles)
922	{
923	struct timekeeper *tk = &tk_core.timekeeper;
924	struct tk_read_base *tkrm = &tk->tkr_mono;
925	ktime_t base_ns, delta_ns, max_ns;
926	u64 base_cycles, delta_cycles;
927	unsigned int seq;
928	u32 mult, shift;
929
930	/*
931	* Racy check to avoid the seqcount overhead when ID does not match. If
932	* the relevant clocksource is installed concurrently, then this will
933	* just delay the switch over to this mechanism until the next event is
934	* programmed. If the ID is not matching the clock events code will use
935	* the regular relative set_next_event() callback as before.
936	*/
937	if (data_race(tk->cs_id) != id)
938	return false;
939
940	do {
941	seq = read_seqcount_begin(&tk_core.seq);
942
943	if (tk->cs_id != id)
944	return false;
945
946	base_cycles = tkrm->cycle_last;
947	base_ns = tkrm->base + (tkrm->xtime_nsec >> tkrm->shift);
948
949	mult = tk->cs_ns_to_cyc_mult;
950	shift = tk->cs_ns_to_cyc_shift;
951	max_ns = tk->cs_ns_to_cyc_maxns;
952
953	} while (read_seqcount_retry(&tk_core.seq, seq));
954
955	/* Prevent negative deltas and multiplication overflows */
956	delta_ns = min(expires_ns - base_ns, max_ns);
957	delta_ns = max(delta_ns, 0);
958
959	/* Convert to cycles */
960	delta_cycles = ((u64)delta_ns * mult) >> shift;
961	*cycles = base_cycles + delta_cycles;
962	return true;
963	}
964
965	/**
966	* ktime_get_real_ts64 - Returns the time of day in a timespec64.
967	* @ts: pointer to the timespec to be set
968	*
969	* Returns the time of day in a timespec64 (WARN if suspended).
970	*/
971	void ktime_get_real_ts64(struct timespec64 *ts)
972	{
973	struct timekeeper *tk = &tk_core.timekeeper;
974	unsigned int seq;
975	u64 nsecs;
976
977	WARN_ON(timekeeping_suspended);
978
979	do {
980	seq = read_seqcount_begin(&tk_core.seq);
981
982	ts->tv_sec = tk->xtime_sec;
983	nsecs = timekeeping_get_ns(&tk->tkr_mono);
984
985	} while (read_seqcount_retry(&tk_core.seq, seq));
986
987	ts->tv_nsec = 0;
988	timespec64_add_ns(ts, nsecs);
989	}
990	EXPORT_SYMBOL(ktime_get_real_ts64);
991
992	ktime_t ktime_get(void)
993	{
994	struct timekeeper *tk = &tk_core.timekeeper;
995	unsigned int seq;
996	ktime_t base;
997	u64 nsecs;
998
999	WARN_ON(timekeeping_suspended);
1000
1001	do {
1002	seq = read_seqcount_begin(&tk_core.seq);
1003	base = tk->tkr_mono.base;
1004	nsecs = timekeeping_get_ns(&tk->tkr_mono);
1005
1006	} while (read_seqcount_retry(&tk_core.seq, seq));
1007
1008	return ktime_add_ns(base, nsecs);
1009	}
1010	EXPORT_SYMBOL_GPL(ktime_get);
1011
1012	u32 ktime_get_resolution_ns(void)
1013	{
1014	struct timekeeper *tk = &tk_core.timekeeper;
1015	unsigned int seq;
1016	u32 nsecs;
1017
1018	WARN_ON(timekeeping_suspended);
1019
1020	do {
1021	seq = read_seqcount_begin(&tk_core.seq);
1022	nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
1023	} while (read_seqcount_retry(&tk_core.seq, seq));
1024
1025	return nsecs;
1026	}
1027	EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
1028
1029	static const ktime_t *const offsets[TK_OFFS_MAX] = {
1030	[TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
1031	[TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
1032	[TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
1033	};
1034
1035	ktime_t ktime_get_with_offset(enum tk_offsets offs)
1036	{
1037	struct timekeeper *tk = &tk_core.timekeeper;
1038	const ktime_t *offset = offsets[offs];
1039	unsigned int seq;
1040	ktime_t base;
1041	u64 nsecs;
1042
1043	WARN_ON(timekeeping_suspended);
1044
1045	do {
1046	seq = read_seqcount_begin(&tk_core.seq);
1047	base = ktime_add(tk->tkr_mono.base, *offset);
1048	nsecs = timekeeping_get_ns(&tk->tkr_mono);
1049
1050	} while (read_seqcount_retry(&tk_core.seq, seq));
1051
1052	return ktime_add_ns(base, nsecs);
1053
1054	}
1055	EXPORT_SYMBOL_GPL(ktime_get_with_offset);
1056
1057	ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
1058	{
1059	struct timekeeper *tk = &tk_core.timekeeper;
1060	const ktime_t *offset = offsets[offs];
1061	unsigned int seq;
1062	ktime_t base;
1063	u64 nsecs;
1064
1065	WARN_ON(timekeeping_suspended);
1066
1067	do {
1068	seq = read_seqcount_begin(&tk_core.seq);
1069	base = ktime_add(tk->tkr_mono.base, *offset);
1070	nsecs = tk->coarse_nsec;
1071
1072	} while (read_seqcount_retry(&tk_core.seq, seq));
1073
1074	return ktime_add_ns(base, nsecs);
1075	}
1076	EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
1077
1078	/**
1079	* ktime_mono_to_any() - convert monotonic time to any other time
1080	* @tmono: time to convert.
1081	* @offs: which offset to use
1082	*/
1083	ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
1084	{
1085	const ktime_t *offset = offsets[offs];
1086	unsigned int seq;
1087	ktime_t tconv;
1088
1089	if (IS_ENABLED(CONFIG_64BIT)) {
1090	/*
1091	* Paired with WRITE_ONCE()s in tk_set_wall_to_mono() and
1092	* tk_update_sleep_time().
1093	*/
1094	return ktime_add(tmono, READ_ONCE(*offset));
1095	}
1096
1097	do {
1098	seq = read_seqcount_begin(&tk_core.seq);
1099	tconv = ktime_add(tmono, *offset);
1100	} while (read_seqcount_retry(&tk_core.seq, seq));
1101
1102	return tconv;
1103	}
1104	EXPORT_SYMBOL_GPL(ktime_mono_to_any);
1105
1106	/**
1107	* ktime_get_raw - Returns the raw monotonic time in ktime_t format
1108	*/
1109	ktime_t ktime_get_raw(void)
1110	{
1111	struct timekeeper *tk = &tk_core.timekeeper;
1112	unsigned int seq;
1113	ktime_t base;
1114	u64 nsecs;
1115
1116	do {
1117	seq = read_seqcount_begin(&tk_core.seq);
1118	base = tk->tkr_raw.base;
1119	nsecs = timekeeping_get_ns(&tk->tkr_raw);
1120
1121	} while (read_seqcount_retry(&tk_core.seq, seq));
1122
1123	return ktime_add_ns(base, nsecs);
1124	}
1125	EXPORT_SYMBOL_GPL(ktime_get_raw);
1126
1127	/**
1128	* ktime_get_ts64 - get the monotonic clock in timespec64 format
1129	* @ts: pointer to timespec variable
1130	*
1131	* The function calculates the monotonic clock from the realtime
1132	* clock and the wall_to_monotonic offset and stores the result
1133	* in normalized timespec64 format in the variable pointed to by @ts.
1134	*/
1135	void ktime_get_ts64(struct timespec64 *ts)
1136	{
1137	struct timekeeper *tk = &tk_core.timekeeper;
1138	struct timespec64 tomono;
1139	unsigned int seq;
1140	u64 nsec;
1141
1142	WARN_ON(timekeeping_suspended);
1143
1144	do {
1145	seq = read_seqcount_begin(&tk_core.seq);
1146	ts->tv_sec = tk->xtime_sec;
1147	nsec = timekeeping_get_ns(&tk->tkr_mono);
1148	tomono = tk->wall_to_monotonic;
1149
1150	} while (read_seqcount_retry(&tk_core.seq, seq));
1151
1152	ts->tv_sec += tomono.tv_sec;
1153	ts->tv_nsec = 0;
1154	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
1155	}
1156	EXPORT_SYMBOL_GPL(ktime_get_ts64);
1157
1158	/**
1159	* ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
1160	*
1161	* Returns the seconds portion of CLOCK_MONOTONIC with a single non
1162	* serialized read. tk->ktime_sec is of type 'unsigned long' so this
1163	* works on both 32 and 64 bit systems. On 32 bit systems the readout
1164	* covers ~136 years of uptime which should be enough to prevent
1165	* premature wrap arounds.
1166	*/
1167	time64_t ktime_get_seconds(void)
1168	{
1169	struct timekeeper *tk = &tk_core.timekeeper;
1170
1171	WARN_ON(timekeeping_suspended);
1172	return tk->ktime_sec;
1173	}
1174	EXPORT_SYMBOL_GPL(ktime_get_seconds);
1175
1176	/**
1177	* ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1178	*
1179	* Returns the wall clock seconds since 1970.
1180	*
1181	* For 64bit systems the fast access to tk->xtime_sec is preserved. On
1182	* 32bit systems the access must be protected with the sequence
1183	* counter to provide "atomic" access to the 64bit tk->xtime_sec
1184	* value.
1185	*/
1186	time64_t ktime_get_real_seconds(void)
1187	{
1188	struct timekeeper *tk = &tk_core.timekeeper;
1189	time64_t seconds;
1190	unsigned int seq;
1191
1192	if (IS_ENABLED(CONFIG_64BIT))
1193	return READ_ONCE(tk->xtime_sec);
1194
1195	do {
1196	seq = read_seqcount_begin(&tk_core.seq);
1197	seconds = READ_ONCE(tk->xtime_sec);
1198
1199	} while (read_seqcount_retry(&tk_core.seq, seq));
1200
1201	return seconds;
1202	}
1203	EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1204
1205	/**
1206	* __ktime_get_real_seconds - Unprotected access to CLOCK_REALTIME seconds
1207	*
1208	* The same as ktime_get_real_seconds() but without the sequence counter
1209	* protection. This function is used in restricted contexts like the x86 MCE
1210	* handler and in KGDB. It's unprotected on 32-bit vs. concurrent half
1211	* completed modification and only to be used for such critical contexts.
1212	*
1213	* Returns: Racy snapshot of the CLOCK_REALTIME seconds value
1214	*/
1215	noinstr time64_t __ktime_get_real_seconds(void)
1216	{
1217	struct timekeeper *tk = &tk_core.timekeeper;
1218
1219	return READ_ONCE(tk->xtime_sec);
1220	}
1221
1222	static inline u64 tk_clock_read_snapshot(const struct tk_read_base *tkr,
1223	struct clocksource_hw_snapshot *chs)
1224	{
1225	struct clocksource *clock = READ_ONCE(tkr->clock);
1226
1227	if (unlikely(clock->read_snapshot))
1228	return clock->read_snapshot(clock, chs);
1229
1230	return clock->read(clock);
1231	}
1232
1233
1234	/**
1235	* ktime_get_snapshot_id - Simultaneously snapshot a given clock ID with
1236	* the corresponding monotonic raw and the underlying
1237	* clocksource counter value.
1238	* @clock_id: The clock ID to snapshot
1239	* @systime_snapshot: Pointer to struct receiving the system time snapshot
1240	*
1241	* For the system time keeping clocks (REALTIME, MONOTONIC and BOOTTIME) the
1242	* monotonic raw clock is CLOCK_MONOTONIC_RAW. For AUX clocks this is the
1243	* monotonic raw clock related to the AUX clock. These AUX clock related
1244	* monotonic raw clocks have a strict linear offset to the system time
1245	* CLOCK_MONOTONIC_RAW:
1246	*
1247	* MONOTONIC_RAW(AUX$N) = CLOCK_MONOTONIC_RAW(system) + offset(AUX$N)
1248	*
1249	* The offset is established when a AUX clock is initialized, but it is
1250	* currently not accessible.
1251	*/
1252	void ktime_get_snapshot_id(clockid_t clock_id, struct system_time_snapshot *systime_snapshot)
1253	{
1254	ktime_t base_raw, base_sys, offs_sys, *offs, offs_zero = 0;
1255	u64 nsec_raw, nsec_sys, now;
1256	struct timekeeper *tk;
1257	struct tk_data *tkd;
1258	unsigned int seq;
1259
1260	/* Invalidate the snapshot for all failure cases */
1261	systime_snapshot->valid = false;
1262
1263	if (WARN_ON_ONCE(timekeeping_suspended))
1264	return;
1265
1266	switch (clock_id) {
1267	case CLOCK_REALTIME:
1268	tkd = &tk_core;
1269	offs = &tk_core.timekeeper.offs_real;
1270	break;
1271	/* Map RAW to MONOTONIC so the loop below is trivial */
1272	case CLOCK_MONOTONIC_RAW:
1273	case CLOCK_MONOTONIC:
1274	tkd = &tk_core;
1275	offs = &offs_zero;
1276	break;
1277	case CLOCK_BOOTTIME:
1278	tkd = &tk_core;
1279	offs = &tk_core.timekeeper.offs_boot;
1280	break;
1281	case CLOCK_AUX ... CLOCK_AUX_LAST:
1282	tkd = aux_get_tk_data(clock_id);
1283	if (!tkd)
1284	return;
1285	offs = &tkd->timekeeper.offs_aux;
1286	break;
1287	default:
1288	WARN_ON_ONCE(1);
1289	return;
1290	}
1291
1292	tk = &tkd->timekeeper;
1293
1294	do {
1295	struct clocksource_hw_snapshot chs = { };
1296
1297	seq = read_seqcount_begin(&tkd->seq);
1298
1299	/* Aux clocks can be invalid */
1300	if (!tk->clock_valid)
1301	return;
1302
1303	now = tk_clock_read_snapshot(&tk->tkr_mono, &chs);
1304	systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1305
1306	systime_snapshot->hw_cycles = chs.hw_cycles;
1307	systime_snapshot->hw_csid = chs.hw_csid;
1308
1309	systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1310	systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1311
1312	base_sys = tk->tkr_mono.base;
1313	offs_sys = *offs;
1314	base_raw = tk->tkr_raw.base;
1315
1316	nsec_sys = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1317	nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1318	} while (read_seqcount_retry(&tkd->seq, seq));
1319
1320	systime_snapshot->cycles = now;
1321	systime_snapshot->systime = ktime_add_ns(base_sys, offs_sys + nsec_sys);
1322	systime_snapshot->monoraw = ktime_add_ns(base_raw, nsec_raw);
1323
1324	/*
1325	* Special case for PTP. Just transfer the raw time into sys,
1326	* so the call sites can consistently use snap::systime.
1327	*/
1328	if (clock_id == CLOCK_MONOTONIC_RAW)
1329	systime_snapshot->systime = systime_snapshot->monoraw;
1330	/* Tell the consumer that this snapshot is valid */
1331	systime_snapshot->valid = true;
1332	}
1333	EXPORT_SYMBOL_GPL(ktime_get_snapshot_id);
1334
1335	/* Scale base by mult/div checking for overflow */
1336	static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1337	{
1338	u64 tmp, rem;
1339
1340	tmp = div64_u64_rem(*base, div, &rem);
1341
1342	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) \|\|
1343	((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1344	return -EOVERFLOW;
1345	tmp *= mult;
1346
1347	rem = div64_u64(rem * mult, div);
1348	*base = tmp + rem;
1349	return 0;
1350	}
1351
1352	/**
1353	* adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1354	* @history: Snapshot representing start of history
1355	* @partial_history_cycles: Cycle offset into history (fractional part)
1356	* @total_history_cycles: Total history length in cycles
1357	* @discontinuity: True indicates clock was set on history period
1358	* @ts: Cross timestamp that should be adjusted using
1359	* partial/total ratio
1360	*
1361	* Helper function used by get_device_system_crosststamp() to correct the
1362	* crosstimestamp corresponding to the start of the current interval to the
1363	* system counter value (timestamp point) provided by the driver. The
1364	* total_history_* quantities are the total history starting at the provided
1365	* reference point and ending at the start of the current interval. The cycle
1366	* count between the driver timestamp point and the start of the current
1367	* interval is partial_history_cycles.
1368	*/
1369	static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1370	u64 partial_history_cycles,
1371	u64 total_history_cycles,
1372	bool discontinuity,
1373	struct system_device_crosststamp *ts)
1374	{
1375	struct timekeeper *tk = &tk_core.timekeeper;
1376	u64 corr_raw, corr_sys;
1377	bool interp_forward;
1378	int ret;
1379
1380	if (total_history_cycles == 0 \|\| partial_history_cycles == 0)
1381	return 0;
1382
1383	/* Interpolate shortest distance from beginning or end of history */
1384	interp_forward = partial_history_cycles > total_history_cycles / 2;
1385	partial_history_cycles = interp_forward ?
1386	total_history_cycles - partial_history_cycles :
1387	partial_history_cycles;
1388
1389	/*
1390	* Scale the monotonic raw time delta by:
1391	* partial_history_cycles / total_history_cycles
1392	*/
1393	corr_raw = (u64)ktime_to_ns(ktime_sub(ts->sys_monoraw, history->monoraw));
1394	ret = scale64_check_overflow(partial_history_cycles,
1395	total_history_cycles, &corr_raw);
1396	if (ret)
1397	return ret;
1398
1399	/*
1400	* If there is a discontinuity in the history, scale monotonic raw
1401	* correction by:
1402	* mult(sys)/mult(raw) yielding the system time correction
1403	*
1404	* Otherwise, calculate the system time correction similar to monotonic
1405	* raw calculation
1406	*/
1407	if (discontinuity) {
1408	corr_sys = mul_u64_u32_div(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1409	} else {
1410	corr_sys = (u64)ktime_to_ns(ktime_sub(ts->sys_systime, history->systime));
1411	ret = scale64_check_overflow(partial_history_cycles, total_history_cycles,
1412	&corr_sys);
1413	if (ret)
1414	return ret;
1415	}
1416
1417	/* Fixup monotonic raw and system time time values */
1418	if (interp_forward) {
1419	ts->sys_monoraw = ktime_add_ns(history->monoraw, corr_raw);
1420	ts->sys_systime = ktime_add_ns(history->systime, corr_sys);
1421	} else {
1422	ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1423	ts->sys_systime = ktime_sub_ns(ts->sys_systime, corr_sys);
1424	}
1425
1426	return 0;
1427	}
1428
1429	/*
1430	* timestamp_in_interval - true if ts is chronologically in [start, end]
1431	*
1432	* True if ts occurs chronologically at or after start, and before or at end.
1433	*/
1434	static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1435	{
1436	if (ts >= start && ts <= end)
1437	return true;
1438	if (start > end && (ts >= start \|\| ts <= end))
1439	return true;
1440	return false;
1441	}
1442
1443	static bool convert_clock(u64 *val, u32 numerator, u32 denominator)
1444	{
1445	u64 rem, res;
1446
1447	if (!numerator \|\| !denominator)
1448	return false;
1449
1450	res = div64_u64_rem(val, denominator, &rem) numerator;
1451	val = res + div_u64(rem numerator, denominator);
1452	return true;
1453	}
1454
1455	static bool convert_base_to_cs(struct system_counterval_t *scv)
1456	{
1457	struct clocksource *cs = tk_core.timekeeper.tkr_mono.clock;
1458	struct clocksource_base *base;
1459	u32 num, den;
1460
1461	/* The timestamp was taken from the time keeper clock source */
1462	if (cs->id == scv->cs_id)
1463	return true;
1464
1465	/*
1466	* Check whether cs_id matches the base clock. Prevent the compiler from
1467	* re-evaluating @base as the clocksource might change concurrently.
1468	*/
1469	base = READ_ONCE(cs->base);
1470	if (!base \|\| base->id != scv->cs_id)
1471	return false;
1472
1473	num = scv->use_nsecs ? cs->freq_khz : base->numerator;
1474	den = scv->use_nsecs ? USEC_PER_SEC : base->denominator;
1475
1476	if (!convert_clock(&scv->cycles, num, den))
1477	return false;
1478
1479	scv->cycles += base->offset;
1480	/* Set the clocksource ID as scv::cycles is now clocksource based */
1481	scv->cs_id = cs->id;
1482	return true;
1483	}
1484
1485	static bool convert_cs_to_base(u64 *cycles, enum clocksource_ids base_id)
1486	{
1487	struct clocksource *cs = tk_core.timekeeper.tkr_mono.clock;
1488	struct clocksource_base *base;
1489
1490	/*
1491	* Check whether base_id matches the base clock. Prevent the compiler from
1492	* re-evaluating @base as the clocksource might change concurrently.
1493	*/
1494	base = READ_ONCE(cs->base);
1495	if (!base \|\| base->id != base_id)
1496	return false;
1497
1498	*cycles -= base->offset;
1499	if (!convert_clock(cycles, base->denominator, base->numerator))
1500	return false;
1501	return true;
1502	}
1503
1504	static bool convert_ns_to_cs(u64 *delta)
1505	{
1506	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
1507
1508	if (BITS_TO_BYTES(fls64(delta) + tkr->shift) >= sizeof(delta))
1509	return false;
1510
1511	delta = div_u64((delta << tkr->shift) - tkr->xtime_nsec, tkr->mult);
1512	return true;
1513	}
1514
1515	/**
1516	* ktime_real_to_base_clock() - Convert CLOCK_REALTIME timestamp to a base clock timestamp
1517	* @treal: CLOCK_REALTIME timestamp to convert
1518	* @base_id: base clocksource id
1519	* @cycles: pointer to store the converted base clock timestamp
1520	*
1521	* Converts a supplied, future realtime clock value to the corresponding base clock value.
1522	*
1523	* Return: true if the conversion is successful, false otherwise.
1524	*/
1525	bool ktime_real_to_base_clock(ktime_t treal, enum clocksource_ids base_id, u64 *cycles)
1526	{
1527	struct timekeeper *tk = &tk_core.timekeeper;
1528	unsigned int seq;
1529	u64 delta;
1530
1531	do {
1532	seq = read_seqcount_begin(&tk_core.seq);
1533	if ((u64)treal < tk->tkr_mono.base_real)
1534	return false;
1535	delta = (u64)treal - tk->tkr_mono.base_real;
1536	if (!convert_ns_to_cs(&delta))
1537	return false;
1538	*cycles = tk->tkr_mono.cycle_last + delta;
1539	if (!convert_cs_to_base(cycles, base_id))
1540	return false;
1541	} while (read_seqcount_retry(&tk_core.seq, seq));
1542
1543	return true;
1544	}
1545	EXPORT_SYMBOL_GPL(ktime_real_to_base_clock);
1546
1547	/**
1548	* get_device_system_crosststamp - Synchronously capture system/device timestamp
1549	* @get_time_fn: Callback to get simultaneous device time and system counter
1550	* from the device driver
1551	* @ctx: Context passed to get_time_fn()
1552	* @history_begin: Historical reference point used to interpolate system time when
1553	* the counter value provided by the driver is before the current interval
1554	* @xtstamp: Receives simultaneously captured system and device time
1555	*
1556	* Reads a timestamp from a device and correlates it to system time
1557	*
1558	* See documentation for ktime_get_snapshot_id() for information about the raw
1559	* monotonic time stamp which is used here.
1560	*/
1561	int get_device_system_crosststamp(int (*get_time_fn)
1562	(ktime_t *device_time,
1563	struct system_counterval_t *sys_counterval,
1564	void *ctx),
1565	void *ctx,
1566	struct system_time_snapshot *history_begin,
1567	struct system_device_crosststamp *xtstamp)
1568	{
1569	u64 syscnt_cycles, cycles, now, interval_start;
1570	ktime_t base_sys, base_raw, *offs;
1571	u32 clock_was_set_seq = 0;
1572	u64 nsec_sys, nsec_raw;
1573	u8 cs_was_changed_seq;
1574	unsigned int seq;
1575	bool do_interp;
1576	struct timekeeper *tk;
1577	struct tk_data *tkd;
1578	int ret;
1579
1580	switch (xtstamp->clock_id) {
1581	case CLOCK_REALTIME:
1582	tkd = &tk_core;
1583	offs = &tk_core.timekeeper.offs_real;
1584	break;
1585	case CLOCK_AUX ... CLOCK_AUX_LAST:
1586	tkd = aux_get_tk_data(xtstamp->clock_id);
1587	if (!tkd)
1588	return -ENODEV;
1589	offs = &tkd->timekeeper.offs_aux;
1590	break;
1591	default:
1592	WARN_ON_ONCE(1);
1593	return -ENODEV;
1594	}
1595
1596	tk = &tkd->timekeeper;
1597
1598	do {
1599	seq = read_seqcount_begin(&tkd->seq);
1600	/*
1601	* Try to synchronously capture device time and a system
1602	* counter value calling back into the device driver
1603	*/
1604	ret = get_time_fn(&xtstamp->device, &xtstamp->sys_counter, ctx);
1605	if (ret)
1606	return ret;
1607
1608	/*
1609	* Verify that the clocksource ID associated with the captured
1610	* system counter value is the same as for the currently
1611	* installed timekeeper clocksource and convert to it.
1612	*/
1613	if (xtstamp->sys_counter.cs_id == CSID_GENERIC \|\|
1614	!convert_base_to_cs(&xtstamp->sys_counter))
1615	return -ENODEV;
1616
1617	cycles = syscnt_cycles = xtstamp->sys_counter.cycles;
1618
1619	/*
1620	* Check whether the system counter value provided by the
1621	* device driver is on the current timekeeping interval.
1622	*/
1623	now = tk_clock_read(&tk->tkr_mono);
1624	interval_start = tk->tkr_mono.cycle_last;
1625	if (!timestamp_in_interval(interval_start, now, cycles)) {
1626	clock_was_set_seq = tk->clock_was_set_seq;
1627	cs_was_changed_seq = tk->cs_was_changed_seq;
1628	cycles = interval_start;
1629	do_interp = true;
1630	} else {
1631	do_interp = false;
1632	}
1633
1634	base_sys = ktime_add(tk->tkr_mono.base, *offs);
1635	base_raw = tk->tkr_raw.base;
1636
1637	nsec_sys = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1638	nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1639	} while (read_seqcount_retry(&tkd->seq, seq));
1640
1641	xtstamp->sys_systime = ktime_add_ns(base_sys, nsec_sys);
1642	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1643
1644	/*
1645	* Interpolate if necessary, adjusting back from the start of the
1646	* current interval
1647	*/
1648	if (do_interp) {
1649	u64 partial_history_cycles, total_history_cycles;
1650	bool discontinuity;
1651
1652	/*
1653	* Check that the counter value is not before the provided
1654	* history reference and that the history doesn't cross a
1655	* clocksource change
1656	*/
1657	if (!history_begin \|\|
1658	!timestamp_in_interval(history_begin->cycles, cycles, syscnt_cycles) \|\|
1659	history_begin->cs_was_changed_seq != cs_was_changed_seq)
1660	return -EINVAL;
1661
1662	partial_history_cycles = cycles - syscnt_cycles;
1663	total_history_cycles = cycles - history_begin->cycles;
1664	discontinuity = history_begin->clock_was_set_seq != clock_was_set_seq;
1665
1666	ret = adjust_historical_crosststamp(history_begin, partial_history_cycles,
1667	total_history_cycles, discontinuity, xtstamp);
1668	}
1669
1670	return ret;
1671	}
1672	EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1673
1674	/**
1675	* timekeeping_clocksource_has_base - Check whether the current clocksource
1676	* is based on given a base clock
1677	* @id: base clocksource ID
1678	*
1679	* Note: The return value is a snapshot which can become invalid right
1680	* after the function returns.
1681	*
1682	* Return: true if the timekeeper clocksource has a base clock with @id,
1683	* false otherwise
1684	*/
1685	bool timekeeping_clocksource_has_base(enum clocksource_ids id)
1686	{
1687	/*
1688	* This is a snapshot, so no point in using the sequence
1689	* count. Just prevent the compiler from re-evaluating @base as the
1690	* clocksource might change concurrently.
1691	*/
1692	struct clocksource_base *base = READ_ONCE(tk_core.timekeeper.tkr_mono.clock->base);
1693
1694	return base ? base->id == id : false;
1695	}
1696	EXPORT_SYMBOL_GPL(timekeeping_clocksource_has_base);
1697
1698	/**
1699	* do_settimeofday64 - Sets the time of day.
1700	* @ts: pointer to the timespec64 variable containing the new time
1701	*
1702	* Sets the time of day to the new time and update NTP and notify hrtimers
1703	*/
1704	int do_settimeofday64(const struct timespec64 *ts)
1705	{
1706	struct timespec64 ts_delta, xt;
1707
1708	if (!timespec64_valid_settod(ts))
1709	return -EINVAL;
1710
1711	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
1712	struct timekeeper *tks = &tk_core.shadow_timekeeper;
1713
1714	timekeeping_forward_now(tks);
1715
1716	xt = tk_xtime(tks);
1717	ts_delta = timespec64_sub(*ts, xt);
1718
1719	if (timespec64_compare(&tks->wall_to_monotonic, &ts_delta) > 0) {
1720	timekeeping_restore_shadow(&tk_core);
1721	return -EINVAL;
1722	}
1723
1724	tk_set_wall_to_mono(tks, timespec64_sub(tks->wall_to_monotonic, ts_delta));
1725	tk_set_xtime(tks, ts);
1726	timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1727	}
1728
1729	/* Signal hrtimers about time change */
1730	clock_was_set(CLOCK_SET_WALL);
1731
1732	audit_tk_injoffset(ts_delta);
1733	add_device_randomness(ts, sizeof(*ts));
1734	return 0;
1735	}
1736	EXPORT_SYMBOL(do_settimeofday64);
1737
1738	static inline bool timekeeper_is_core_tk(struct timekeeper *tk)
1739	{
1740	return !IS_ENABLED(CONFIG_POSIX_AUX_CLOCKS) \|\| tk->id == TIMEKEEPER_CORE;
1741	}
1742
1743	/**
1744	* __timekeeping_inject_offset - Adds or subtracts from the current time.
1745	* @tkd: Pointer to the timekeeper to modify
1746	* @ts: Pointer to the timespec variable containing the offset
1747	*
1748	* Adds or subtracts an offset value from the current time.
1749	*/
1750	static int __timekeeping_inject_offset(struct tk_data tkd, const struct timespec64 ts)
1751	{
1752	struct timekeeper *tks = &tkd->shadow_timekeeper;
1753	struct timespec64 tmp;
1754
1755	if (ts->tv_nsec < 0 \|\| ts->tv_nsec >= NSEC_PER_SEC)
1756	return -EINVAL;
1757
1758	timekeeping_forward_now(tks);
1759
1760	if (timekeeper_is_core_tk(tks)) {
1761	/* Make sure the proposed value is valid */
1762	tmp = timespec64_add(tk_xtime(tks), *ts);
1763	if (timespec64_compare(&tks->wall_to_monotonic, ts) > 0 \|\|
1764	!timespec64_valid_settod(&tmp)) {
1765	timekeeping_restore_shadow(tkd);
1766	return -EINVAL;
1767	}
1768
1769	tk_xtime_add(tks, ts);
1770	tk_set_wall_to_mono(tks, timespec64_sub(tks->wall_to_monotonic, *ts));
1771	} else {
1772	struct tk_read_base *tkr_mono = &tks->tkr_mono;
1773	ktime_t now, offs;
1774
1775	/* Get the current time */
1776	now = ktime_add_ns(tkr_mono->base, timekeeping_get_ns(tkr_mono));
1777	/* Add the relative offset change */
1778	offs = ktime_add(tks->offs_aux, timespec64_to_ktime(*ts));
1779
1780	/* Prevent that the resulting time becomes negative */
1781	if (ktime_add(now, offs) < 0) {
1782	timekeeping_restore_shadow(tkd);
1783	return -EINVAL;
1784	}
1785	tk_update_aux_offs(tks, offs);
1786	}
1787
1788	timekeeping_update_from_shadow(tkd, TK_UPDATE_ALL);
1789	return 0;
1790	}
1791
1792	static int timekeeping_inject_offset(const struct timespec64 *ts)
1793	{
1794	int ret;
1795
1796	scoped_guard (raw_spinlock_irqsave, &tk_core.lock)
1797	ret = __timekeeping_inject_offset(&tk_core, ts);
1798
1799	/* Signal hrtimers about time change */
1800	if (!ret)
1801	clock_was_set(CLOCK_SET_WALL);
1802	return ret;
1803	}
1804
1805	/*
1806	* Indicates if there is an offset between the system clock and the hardware
1807	* clock/persistent clock/rtc.
1808	*/
1809	int persistent_clock_is_local;
1810
1811	/*
1812	* Adjust the time obtained from the CMOS to be UTC time instead of
1813	* local time.
1814	*
1815	* This is ugly, but preferable to the alternatives. Otherwise we
1816	* would either need to write a program to do it in /etc/rc (and risk
1817	* confusion if the program gets run more than once; it would also be
1818	* hard to make the program warp the clock precisely n hours) or
1819	* compile in the timezone information into the kernel. Bad, bad....
1820	*
1821	* - TYT, 1992-01-01
1822	*
1823	* The best thing to do is to keep the CMOS clock in universal time (UTC)
1824	* as real UNIX machines always do it. This avoids all headaches about
1825	* daylight saving times and warping kernel clocks.
1826	*/
1827	void timekeeping_warp_clock(void)
1828	{
1829	if (sys_tz.tz_minuteswest != 0) {
1830	struct timespec64 adjust;
1831
1832	persistent_clock_is_local = 1;
1833	adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1834	adjust.tv_nsec = 0;
1835	timekeeping_inject_offset(&adjust);
1836	}
1837	}
1838
1839	/*
1840	* __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1841	*/
1842	static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1843	{
1844	tk->tai_offset = tai_offset;
1845	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1846	}
1847
1848	/*
1849	* change_clocksource - Swaps clocksources if a new one is available
1850	*
1851	* Accumulates current time interval and initializes new clocksource
1852	*/
1853	static int change_clocksource(void *data)
1854	{
1855	struct clocksource new = data, old = NULL;
1856
1857	/*
1858	* If the clocksource is in a module, get a module reference.
1859	* Succeeds for built-in code (owner == NULL) as well. Abort if the
1860	* reference can't be acquired.
1861	*/
1862	if (!try_module_get(new->owner))
1863	return 0;
1864
1865	/* Abort if the device can't be enabled */
1866	if (new->enable && new->enable(new) != 0) {
1867	module_put(new->owner);
1868	return 0;
1869	}
1870
1871	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
1872	struct timekeeper *tks = &tk_core.shadow_timekeeper;
1873
1874	timekeeping_forward_now(tks);
1875	old = tks->tkr_mono.clock;
1876	tk_setup_internals(tks, new);
1877	timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1878	}
1879
1880	tk_aux_update_clocksource();
1881
1882	if (old) {
1883	if (old->disable)
1884	old->disable(old);
1885	module_put(old->owner);
1886	}
1887
1888	return 0;
1889	}
1890
1891	/**
1892	* timekeeping_notify - Install a new clock source
1893	* @clock: pointer to the clock source
1894	*
1895	* This function is called from clocksource.c after a new, better clock
1896	* source has been registered. The caller holds the clocksource_mutex.
1897	*/
1898	int timekeeping_notify(struct clocksource *clock)
1899	{
1900	struct timekeeper *tk = &tk_core.timekeeper;
1901
1902	if (tk->tkr_mono.clock == clock)
1903	return 0;
1904
1905	/* Disable inlined reads accross the clocksource switch */
1906	clocksource_disable_inline_read();
1907
1908	stop_machine(change_clocksource, clock, NULL);
1909
1910	/*
1911	* If the clocksource has been selected and supports inlined reads
1912	* enable the branch.
1913	*/
1914	if (tk->tkr_mono.clock == clock && clock->flags & CLOCK_SOURCE_CAN_INLINE_READ)
1915	clocksource_enable_inline_read();
1916
1917	tick_clock_notify();
1918	return tk->tkr_mono.clock == clock ? 0 : -1;
1919	}
1920
1921	/**
1922	* ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1923	* @ts: pointer to the timespec64 to be set
1924	*
1925	* Returns the raw monotonic time (completely un-modified by ntp)
1926	*/
1927	void ktime_get_raw_ts64(struct timespec64 *ts)
1928	{
1929	struct timekeeper *tk = &tk_core.timekeeper;
1930	unsigned int seq;
1931	u64 nsecs;
1932
1933	do {
1934	seq = read_seqcount_begin(&tk_core.seq);
1935	ts->tv_sec = tk->raw_sec;
1936	nsecs = timekeeping_get_ns(&tk->tkr_raw);
1937
1938	} while (read_seqcount_retry(&tk_core.seq, seq));
1939
1940	ts->tv_nsec = 0;
1941	timespec64_add_ns(ts, nsecs);
1942	}
1943	EXPORT_SYMBOL(ktime_get_raw_ts64);
1944
1945	/**
1946	* timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1947	*/
1948	int timekeeping_valid_for_hres(void)
1949	{
1950	struct timekeeper *tk = &tk_core.timekeeper;
1951	unsigned int seq;
1952	int ret;
1953
1954	do {
1955	seq = read_seqcount_begin(&tk_core.seq);
1956
1957	ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1958
1959	} while (read_seqcount_retry(&tk_core.seq, seq));
1960
1961	return ret;
1962	}
1963
1964	/**
1965	* timekeeping_max_deferment - Returns max time the clocksource can be deferred
1966	*/
1967	u64 timekeeping_max_deferment(void)
1968	{
1969	struct timekeeper *tk = &tk_core.timekeeper;
1970	unsigned int seq;
1971	u64 ret;
1972
1973	do {
1974	seq = read_seqcount_begin(&tk_core.seq);
1975
1976	ret = tk->tkr_mono.clock->max_idle_ns;
1977
1978	} while (read_seqcount_retry(&tk_core.seq, seq));
1979
1980	return ret;
1981	}
1982
1983	/**
1984	* read_persistent_clock64 - Return time from the persistent clock.
1985	* @ts: Pointer to the storage for the readout value
1986	*
1987	* Weak dummy function for arches that do not yet support it.
1988	* Reads the time from the battery backed persistent clock.
1989	* Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1990	*
1991	* XXX - Do be sure to remove it once all arches implement it.
1992	*/
1993	void __weak read_persistent_clock64(struct timespec64 *ts)
1994	{
1995	ts->tv_sec = 0;
1996	ts->tv_nsec = 0;
1997	}
1998
1999	/**
2000	* read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
2001	* from the boot.
2002	* @wall_time: current time as returned by persistent clock
2003	* @boot_offset: offset that is defined as wall_time - boot_time
2004	*
2005	* Weak dummy function for arches that do not yet support it.
2006	*
2007	* The default function calculates offset based on the current value of
2008	* local_clock(). This way architectures that support sched_clock() but don't
2009	* support dedicated boot time clock will provide the best estimate of the
2010	* boot time.
2011	*/
2012	void __weak __init
2013	read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
2014	struct timespec64 *boot_offset)
2015	{
2016	read_persistent_clock64(wall_time);
2017	*boot_offset = ns_to_timespec64(local_clock());
2018	}
2019
2020	static __init void tkd_basic_setup(struct tk_data *tkd, enum timekeeper_ids tk_id, bool valid)
2021	{
2022	raw_spin_lock_init(&tkd->lock);
2023	seqcount_raw_spinlock_init(&tkd->seq, &tkd->lock);
2024	tkd->timekeeper.id = tkd->shadow_timekeeper.id = tk_id;
2025	tkd->timekeeper.clock_valid = tkd->shadow_timekeeper.clock_valid = valid;
2026	}
2027
2028	/*
2029	* Flag reflecting whether timekeeping_resume() has injected sleeptime.
2030	*
2031	* The flag starts of false and is only set when a suspend reaches
2032	* timekeeping_suspend(), timekeeping_resume() sets it to false when the
2033	* timekeeper clocksource is not stopping across suspend and has been
2034	* used to update sleep time. If the timekeeper clocksource has stopped
2035	* then the flag stays true and is used by the RTC resume code to decide
2036	* whether sleeptime must be injected and if so the flag gets false then.
2037	*
2038	* If a suspend fails before reaching timekeeping_resume() then the flag
2039	* stays false and prevents erroneous sleeptime injection.
2040	*/
2041	static bool suspend_timing_needed;
2042
2043	/* Flag for if there is a persistent clock on this platform */
2044	static bool persistent_clock_exists;
2045
2046	/*
2047	* timekeeping_init - Initializes the clocksource and common timekeeping values
2048	*/
2049	void __init timekeeping_init(void)
2050	{
2051	struct timespec64 wall_time, boot_offset, wall_to_mono;
2052	struct timekeeper *tks = &tk_core.shadow_timekeeper;
2053	struct clocksource *clock;
2054
2055	tkd_basic_setup(&tk_core, TIMEKEEPER_CORE, true);
2056	tk_aux_setup();
2057
2058	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
2059	if (timespec64_valid_settod(&wall_time) &&
2060	timespec64_to_ns(&wall_time) > 0) {
2061	persistent_clock_exists = true;
2062	} else if (timespec64_to_ns(&wall_time) != 0) {
2063	pr_warn("Persistent clock returned invalid value");
2064	wall_time = (struct timespec64){0};
2065	}
2066
2067	if (timespec64_compare(&wall_time, &boot_offset) < 0)
2068	boot_offset = (struct timespec64){0};
2069
2070	/*
2071	* We want set wall_to_mono, so the following is true:
2072	* wall time + wall_to_mono = boot time
2073	*/
2074	wall_to_mono = timespec64_sub(boot_offset, wall_time);
2075
2076	clock = clocksource_default_clock();
2077	if (clock->enable)
2078	clock->enable(clock);
2079
2080	guard(raw_spinlock_irqsave)(&tk_core.lock);
2081
2082	ntp_init();
2083
2084	tk_setup_internals(tks, clock);
2085
2086	tk_set_xtime(tks, &wall_time);
2087	tks->raw_sec = 0;
2088
2089	tk_set_wall_to_mono(tks, wall_to_mono);
2090
2091	/*
2092	* Use TK_UPDATE_ALL so the NTP layer picks up the clocksource's
2093	* cs_tick_adj via ntp_clear(). Clearing NTP here is otherwise
2094	* redundant as ntp_init() already initialised it above.
2095	*/
2096	timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
2097	}
2098
2099	/* time in seconds when suspend began for persistent clock */
2100	static struct timespec64 timekeeping_suspend_time;
2101
2102	/**
2103	* __timekeeping_inject_sleeptime - Internal function to add sleep interval
2104	* @tk: Pointer to the timekeeper to be updated
2105	* @delta: Pointer to the delta value in timespec64 format
2106	*
2107	* Takes a timespec offset measuring a suspend interval and properly
2108	* adds the sleep offset to the timekeeping variables.
2109	*/
2110	static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
2111	const struct timespec64 *delta)
2112	{
2113	if (!timespec64_valid_strict(delta)) {
2114	printk_deferred(KERN_WARNING
2115	"__timekeeping_inject_sleeptime: Invalid "
2116	"sleep delta value!\n");
2117	return;
2118	}
2119	tk_xtime_add(tk, delta);
2120	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
2121	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
2122	tk_debug_account_sleep_time(delta);
2123	}
2124
2125	#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
2126	/*
2127	* We have three kinds of time sources to use for sleep time
2128	* injection, the preference order is:
2129	* 1) non-stop clocksource
2130	* 2) persistent clock (ie: RTC accessible when irqs are off)
2131	* 3) RTC
2132	*
2133	* 1) and 2) are used by timekeeping, 3) by RTC subsystem.
2134	* If system has neither 1) nor 2), 3) will be used finally.
2135	*
2136	*
2137	* If timekeeping has injected sleeptime via either 1) or 2),
2138	* 3) becomes needless, so in this case we don't need to call
2139	* rtc_resume(), and this is what timekeeping_rtc_skipresume()
2140	* means.
2141	*/
2142	bool timekeeping_rtc_skipresume(void)
2143	{
2144	return !suspend_timing_needed;
2145	}
2146
2147	/*
2148	* 1) can be determined whether to use or not only when doing
2149	* timekeeping_resume() which is invoked after rtc_suspend(),
2150	* so we can't skip rtc_suspend() surely if system has 1).
2151	*
2152	* But if system has 2), 2) will definitely be used, so in this
2153	* case we don't need to call rtc_suspend(), and this is what
2154	* timekeeping_rtc_skipsuspend() means.
2155	*/
2156	bool timekeeping_rtc_skipsuspend(void)
2157	{
2158	return persistent_clock_exists;
2159	}
2160
2161	/**
2162	* timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
2163	* @delta: pointer to a timespec64 delta value
2164	*
2165	* This hook is for architectures that cannot support read_persistent_clock64
2166	* because their RTC/persistent clock is only accessible when irqs are enabled.
2167	* and also don't have an effective nonstop clocksource.
2168	*
2169	* This function should only be called by rtc_resume(), and allows
2170	* a suspend offset to be injected into the timekeeping values.
2171	*/
2172	void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
2173	{
2174	scoped_guard(raw_spinlock_irqsave, &tk_core.lock) {
2175	struct timekeeper *tks = &tk_core.shadow_timekeeper;
2176
2177	suspend_timing_needed = false;
2178	timekeeping_forward_now(tks);
2179	__timekeeping_inject_sleeptime(tks, delta);
2180	timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
2181	}
2182
2183	/* Signal hrtimers about time change */
2184	clock_was_set(CLOCK_SET_WALL \| CLOCK_SET_BOOT);
2185	}
2186	#endif
2187
2188	/**
2189	* timekeeping_resume - Resumes the generic timekeeping subsystem.
2190	*/
2191	void timekeeping_resume(void)
2192	{
2193	struct timekeeper *tks = &tk_core.shadow_timekeeper;
2194	struct clocksource *clock = tks->tkr_mono.clock;
2195	struct timespec64 ts_new, ts_delta;
2196	bool inject_sleeptime = false;
2197	u64 cycle_now, nsec;
2198	unsigned long flags;
2199
2200	read_persistent_clock64(&ts_new);
2201
2202	clockevents_resume();
2203	clocksource_resume();
2204
2205	raw_spin_lock_irqsave(&tk_core.lock, flags);
2206
2207	/*
2208	* After system resumes, we need to calculate the suspended time and
2209	* compensate it for the OS time. There are 3 sources that could be
2210	* used: Nonstop clocksource during suspend, persistent clock and rtc
2211	* device.
2212	*
2213	* One specific platform may have 1 or 2 or all of them, and the
2214	* preference will be:
2215	* suspend-nonstop clocksource -> persistent clock -> rtc
2216	* The less preferred source will only be tried if there is no better
2217	* usable source. The rtc part is handled separately in rtc core code.
2218	*/
2219	cycle_now = tk_clock_read(&tks->tkr_mono);
2220	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
2221	if (nsec > 0) {
2222	ts_delta = ns_to_timespec64(nsec);
2223	inject_sleeptime = true;
2224	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
2225	ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
2226	inject_sleeptime = true;
2227	}
2228
2229	if (inject_sleeptime) {
2230	suspend_timing_needed = false;
2231	__timekeeping_inject_sleeptime(tks, &ts_delta);
2232	}
2233
2234	/* Re-base the last cycle value */
2235	tks->tkr_mono.cycle_last = cycle_now;
2236	tks->tkr_raw.cycle_last = cycle_now;
2237
2238	tks->ntp_error = 0;
2239	timekeeping_suspended = 0;
2240	timekeeping_update_from_shadow(&tk_core, TK_CLOCK_WAS_SET);
2241	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
2242
2243	touch_softlockup_watchdog();
2244
2245	/* Resume the clockevent device(s) and hrtimers */
2246	tick_resume();
2247	/* Notify timerfd as resume is equivalent to clock_was_set() */
2248	timerfd_resume();
2249	}
2250
2251	static void timekeeping_syscore_resume(void *data)
2252	{
2253	timekeeping_resume();
2254	}
2255
2256	int timekeeping_suspend(void)
2257	{
2258	struct timekeeper *tks = &tk_core.shadow_timekeeper;
2259	struct timespec64 delta, delta_delta;
2260	static struct timespec64 old_delta;
2261	struct clocksource *curr_clock;
2262	unsigned long flags;
2263	u64 cycle_now;
2264
2265	read_persistent_clock64(&timekeeping_suspend_time);
2266
2267	/*
2268	* On some systems the persistent_clock can not be detected at
2269	* timekeeping_init by its return value, so if we see a valid
2270	* value returned, update the persistent_clock_exists flag.
2271	*/
2272	if (timekeeping_suspend_time.tv_sec \|\| timekeeping_suspend_time.tv_nsec)
2273	persistent_clock_exists = true;
2274
2275	suspend_timing_needed = true;
2276
2277	raw_spin_lock_irqsave(&tk_core.lock, flags);
2278	timekeeping_forward_now(tks);
2279	timekeeping_suspended = 1;
2280
2281	/*
2282	* Since we've called forward_now, cycle_last stores the value
2283	* just read from the current clocksource. Save this to potentially
2284	* use in suspend timing.
2285	*/
2286	curr_clock = tks->tkr_mono.clock;
2287	cycle_now = tks->tkr_mono.cycle_last;
2288	clocksource_start_suspend_timing(curr_clock, cycle_now);
2289
2290	if (persistent_clock_exists) {
2291	/*
2292	* To avoid drift caused by repeated suspend/resumes,
2293	* which each can add ~1 second drift error,
2294	* try to compensate so the difference in system time
2295	* and persistent_clock time stays close to constant.
2296	*/
2297	delta = timespec64_sub(tk_xtime(tks), timekeeping_suspend_time);
2298	delta_delta = timespec64_sub(delta, old_delta);
2299	if (abs(delta_delta.tv_sec) >= 2) {
2300	/*
2301	* if delta_delta is too large, assume time correction
2302	* has occurred and set old_delta to the current delta.
2303	*/
2304	old_delta = delta;
2305	} else {
2306	/* Otherwise try to adjust old_system to compensate */
2307	timekeeping_suspend_time =
2308	timespec64_add(timekeeping_suspend_time, delta_delta);
2309	}
2310	}
2311
2312	timekeeping_update_from_shadow(&tk_core, 0);
2313	halt_fast_timekeeper(tks);
2314	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
2315
2316	tick_suspend();
2317	clocksource_suspend();
2318	clockevents_suspend();
2319
2320	return 0;
2321	}
2322
2323	static int timekeeping_syscore_suspend(void *data)
2324	{
2325	return timekeeping_suspend();
2326	}
2327
2328	/* sysfs resume/suspend bits for timekeeping */
2329	static const struct syscore_ops timekeeping_syscore_ops = {
2330	.resume = timekeeping_syscore_resume,
2331	.suspend = timekeeping_syscore_suspend,
2332	};
2333
2334	static struct syscore timekeeping_syscore = {
2335	.ops = &timekeeping_syscore_ops,
2336	};
2337
2338	static int __init timekeeping_init_ops(void)
2339	{
2340	register_syscore(&timekeeping_syscore);
2341	return 0;
2342	}
2343	device_initcall(timekeeping_init_ops);
2344
2345	/*
2346	* Apply a multiplier adjustment to the timekeeper
2347	*/
2348	static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
2349	s64 offset,
2350	s32 mult_adj)
2351	{
2352	s64 interval = tk->cycle_interval;
2353
2354	if (mult_adj == 0) {
2355	return;
2356	} else if (mult_adj == -1) {
2357	interval = -interval;
2358	offset = -offset;
2359	} else if (mult_adj != 1) {
2360	interval *= mult_adj;
2361	offset *= mult_adj;
2362	}
2363
2364	/*
2365	* So the following can be confusing.
2366	*
2367	* To keep things simple, lets assume mult_adj == 1 for now.
2368	*
2369	* When mult_adj != 1, remember that the interval and offset values
2370	* have been appropriately scaled so the math is the same.
2371	*
2372	* The basic idea here is that we're increasing the multiplier
2373	* by one, this causes the xtime_interval to be incremented by
2374	* one cycle_interval. This is because:
2375	* xtime_interval = cycle_interval * mult
2376	* So if mult is being incremented by one:
2377	* xtime_interval = cycle_interval * (mult + 1)
2378	* Its the same as:
2379	* xtime_interval = (cycle_interval * mult) + cycle_interval
2380	* Which can be shortened to:
2381	* xtime_interval += cycle_interval
2382	*
2383	* So offset stores the non-accumulated cycles. Thus the current
2384	* time (in shifted nanoseconds) is:
2385	* now = (offset * adj) + xtime_nsec
2386	* Now, even though we're adjusting the clock frequency, we have
2387	* to keep time consistent. In other words, we can't jump back
2388	* in time, and we also want to avoid jumping forward in time.
2389	*
2390	* So given the same offset value, we need the time to be the same
2391	* both before and after the freq adjustment.
2392	* now = (offset * adj_1) + xtime_nsec_1
2393	* now = (offset * adj_2) + xtime_nsec_2
2394	* So:
2395	* (offset * adj_1) + xtime_nsec_1 =
2396	* (offset * adj_2) + xtime_nsec_2
2397	* And we know:
2398	* adj_2 = adj_1 + 1
2399	* So:
2400	* (offset * adj_1) + xtime_nsec_1 =
2401	* (offset * (adj_1+1)) + xtime_nsec_2
2402	* (offset * adj_1) + xtime_nsec_1 =
2403	* (offset * adj_1) + offset + xtime_nsec_2
2404	* Canceling the sides:
2405	* xtime_nsec_1 = offset + xtime_nsec_2
2406	* Which gives us:
2407	* xtime_nsec_2 = xtime_nsec_1 - offset
2408	* Which simplifies to:
2409	* xtime_nsec -= offset
2410	*
2411	* When subtracting offset from xtime_nsec, the same amount
2412	* (in appropriate units) has to be added to ntp_error, in
2413	* order to correctly track the delta between the time
2414	* reported in xtime_nsec, and the intended time.
2415	*/
2416	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
2417	/* NTP adjustment caused clocksource mult overflow */
2418	WARN_ON_ONCE(1);
2419	return;
2420	}
2421
2422	tk->tkr_mono.mult += mult_adj;
2423	tk->xtime_interval += interval;
2424	tk->tkr_mono.xtime_nsec -= offset;
2425	tk->ntp_error += offset << tk->ntp_error_shift;
2426	}
2427
2428	/*
2429	* Adjust the timekeeper's multiplier to the correct frequency
2430	* and also to reduce the accumulated error value.
2431	*/
2432	static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2433	{
2434	u64 ntp_tl = ntp_tick_length(tk->id);
2435	s64 skew = ntp_get_skew_delta(tk->id);
2436	u32 mult;
2437
2438	/*
2439	* Determine the multiplier from the current NTP tick length plus
2440	* skew_delta. The skew biases mult so that ±1 dithering can deliver
2441	* the time_offset slew rate. Recompute when either changes.
2442	*/
2443	if (likely(tk->ntp_tick == ntp_tl && tk->skew_delta == skew)) {
2444	/* Revert to the base mult rate. */
2445	mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2446	} else {
2447	tk->ntp_tick = ntp_tl;
2448	tk->skew_delta = skew;
2449	/*
2450	* skew_delta is stored pre-divided by HZ (matching time_offset);
2451	* scale it back up to the full per-tick rate for the mult bias.
2452	*/
2453	skew *= NTP_INTERVAL_FREQ;
2454	mult = div64_u64((tk->ntp_tick + skew) >> tk->ntp_error_shift,
2455	tk->cycle_interval);
2456	}
2457
2458	/*
2459	* If the clock is behind the NTP time, increase the multiplier by 1
2460	* to catch up with it. If it's ahead and there was a remainder in the
2461	* tick division, the clock will slow down. Otherwise it will stay
2462	* ahead until the tick length changes to a non-divisible value.
2463	*/
2464	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2465	mult += tk->ntp_err_mult;
2466
2467	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2468
2469	if (unlikely(tk->tkr_mono.clock->maxadj &&
2470	(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2471	> tk->tkr_mono.clock->maxadj))) {
2472	printk_once(KERN_WARNING
2473	"Adjusting %s more than 11%% (%ld vs %ld)\n",
2474	tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2475	(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2476	}
2477
2478	/*
2479	* It may be possible that when we entered this function, xtime_nsec
2480	* was very small. Further, if we're slightly speeding the clocksource
2481	* in the code above, its possible the required corrective factor to
2482	* xtime_nsec could cause it to underflow.
2483	*
2484	* Now, since we have already accumulated the second and the NTP
2485	* subsystem has been notified via second_overflow(), we need to skip
2486	* the next update.
2487	*/
2488	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2489	tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2490	tk->tkr_mono.shift;
2491	tk->xtime_sec--;
2492	tk->skip_second_overflow = 1;
2493	}
2494	}
2495
2496	/*
2497	* accumulate_nsecs_to_secs - Accumulates nsecs into secs
2498	*
2499	* Helper function that accumulates the nsecs greater than a second
2500	* from the xtime_nsec field to the xtime_secs field.
2501	* It also calls into the NTP code to handle leapsecond processing.
2502	*/
2503	static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2504	{
2505	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2506	unsigned int clock_set = 0;
2507
2508	while (tk->tkr_mono.xtime_nsec >= nsecps) {
2509	int leap;
2510
2511	tk->tkr_mono.xtime_nsec -= nsecps;
2512	tk->xtime_sec++;
2513
2514	/*
2515	* Skip NTP update if this second was accumulated before,
2516	* i.e. xtime_nsec underflowed in timekeeping_adjust()
2517	*/
2518	if (unlikely(tk->skip_second_overflow)) {
2519	tk->skip_second_overflow = 0;
2520	continue;
2521	}
2522
2523	/* Figure out if its a leap sec and apply if needed */
2524	leap = second_overflow(tk->id, tk->xtime_sec);
2525	if (unlikely(leap)) {
2526	struct timespec64 ts;
2527
2528	tk->xtime_sec += leap;
2529
2530	ts.tv_sec = leap;
2531	ts.tv_nsec = 0;
2532	tk_set_wall_to_mono(tk,
2533	timespec64_sub(tk->wall_to_monotonic, ts));
2534
2535	__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2536
2537	clock_set = TK_CLOCK_WAS_SET;
2538	}
2539	}
2540	return clock_set;
2541	}
2542
2543	/*
2544	* logarithmic_accumulation - shifted accumulation of cycles
2545	*
2546	* This functions accumulates a shifted interval of cycles into
2547	* a shifted interval nanoseconds. Allows for O(log) accumulation
2548	* loop.
2549	*
2550	* Returns the unconsumed cycles.
2551	*/
2552	static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2553	u32 shift, unsigned int *clock_set)
2554	{
2555	u64 interval = tk->cycle_interval << shift;
2556	u64 snsec_per_sec;
2557
2558	/* If the offset is smaller than a shifted interval, do nothing */
2559	if (offset < interval)
2560	return offset;
2561
2562	/* Accumulate one shifted interval */
2563	offset -= interval;
2564	tk->tkr_mono.cycle_last += interval;
2565	tk->tkr_raw.cycle_last += interval;
2566
2567	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2568	*clock_set \|= accumulate_nsecs_to_secs(tk);
2569
2570	/* Accumulate raw time */
2571	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2572	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2573	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2574	tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2575	tk->raw_sec++;
2576	}
2577
2578	/* Accumulate error between NTP and clock interval */
2579	tk->ntp_error += tk->ntp_tick << shift;
2580	tk->ntp_error -= tk->xtime_interval << (tk->ntp_error_shift + shift);
2581
2582	/*
2583	* When skewing, do so by adjusting ntp_error to impart an extra
2584	* target delta into ntp_error per tick, limited to what can be
2585	* drained from time_offset / time_adjust to avoid overshoot.
2586	*
2587	* The base 'mult' value was calculated with the skew taken into
2588	* account, such that the per-tick choice of 'mult' vs. 'mult+1'
2589	* allows for the desired effective rate and ntp_error does not
2590	* grow unbounded.
2591	*
2592	* Once the full desired phase offset is delivered, any remaining
2593	* skew imparted by the adjusted 'mult', accounted above, remains
2594	* in ntp_error and will be compensated by the dithering over time.
2595	*/
2596	if (tk->skew_delta)
2597	tk->ntp_error += ntp_drain_skew(tk->id, tk->skew_delta << shift,
2598	shift) * NTP_INTERVAL_FREQ;
2599
2600	return offset;
2601	}
2602
2603	/*
2604	* timekeeping_advance - Updates the timekeeper to the current time and
2605	* current NTP tick length
2606	*/
2607	static bool __timekeeping_advance(struct tk_data *tkd, enum timekeeping_adv_mode mode)
2608	{
2609	struct timekeeper *tk = &tkd->shadow_timekeeper;
2610	struct timekeeper *real_tk = &tkd->timekeeper;
2611	unsigned int clock_set = 0;
2612	int shift = 0, maxshift;
2613	u64 offset, orig_offset;
2614
2615	/* Make sure we're fully resumed: */
2616	if (unlikely(timekeeping_suspended))
2617	return false;
2618
2619	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2620	tk->tkr_mono.cycle_last, tk->tkr_mono.mask,
2621	tk->tkr_mono.clock->max_raw_delta);
2622	orig_offset = offset;
2623	/* Check if there's really nothing to do */
2624	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2625	return false;
2626
2627	/*
2628	* With NO_HZ we may have to accumulate many cycle_intervals
2629	* (think "ticks") worth of time at once. To do this efficiently,
2630	* we calculate the largest doubling multiple of cycle_intervals
2631	* that is smaller than the offset. We then accumulate that
2632	* chunk in one go, and then try to consume the next smaller
2633	* doubled multiple.
2634	*/
2635	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2636	shift = max(0, shift);
2637	/* Bound shift to one less than what overflows tick_length */
2638	maxshift = (64 - (ilog2(ntp_tick_length(tk->id)) + 1)) - 1;
2639	shift = min(shift, maxshift);
2640	while (offset >= tk->cycle_interval) {
2641	offset = logarithmic_accumulation(tk, offset, shift, &clock_set);
2642	if (offset < tk->cycle_interval<<shift)
2643	shift--;
2644	}
2645
2646	/* Adjust the multiplier to correct NTP error */
2647	timekeeping_adjust(tk, offset);
2648
2649	/*
2650	* Finally, make sure that after the rounding
2651	* xtime_nsec isn't larger than NSEC_PER_SEC
2652	*/
2653	clock_set \|= accumulate_nsecs_to_secs(tk);
2654
2655	/*
2656	* To avoid inconsistencies caused adjtimex TK_ADV_FREQ calls
2657	* making small negative adjustments to the base xtime_nsec
2658	* value, only update the coarse clocks if we accumulated time
2659	*/
2660	if (orig_offset != offset)
2661	tk_update_coarse_nsecs(tk);
2662
2663	timekeeping_update_from_shadow(tkd, clock_set);
2664
2665	return !!clock_set;
2666	}
2667
2668	static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2669	{
2670	guard(raw_spinlock_irqsave)(&tk_core.lock);
2671	return __timekeeping_advance(&tk_core, mode);
2672	}
2673
2674	/**
2675	* update_wall_time - Uses the current clocksource to increment the wall time
2676	*
2677	* It also updates the enabled auxiliary clock timekeepers
2678	*/
2679	void update_wall_time(void)
2680	{
2681	if (timekeeping_advance(TK_ADV_TICK))
2682	clock_was_set_delayed();
2683	tk_aux_advance();
2684	}
2685
2686	/**
2687	* getboottime64 - Return the real time of system boot.
2688	* @ts: pointer to the timespec64 to be set
2689	*
2690	* Returns the wall-time of boot in a timespec64.
2691	*
2692	* This is based on the wall_to_monotonic offset and the total suspend
2693	* time. Calls to settimeofday will affect the value returned (which
2694	* basically means that however wrong your real time clock is at boot time,
2695	* you get the right time here).
2696	*/
2697	void getboottime64(struct timespec64 *ts)
2698	{
2699	struct timekeeper *tk = &tk_core.timekeeper;
2700	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2701
2702	*ts = ktime_to_timespec64(t);
2703	}
2704	EXPORT_SYMBOL_GPL(getboottime64);
2705
2706	void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2707	{
2708	struct timekeeper *tk = &tk_core.timekeeper;
2709	unsigned int seq;
2710
2711	do {
2712	seq = read_seqcount_begin(&tk_core.seq);
2713
2714	*ts = tk_xtime_coarse(tk);
2715	} while (read_seqcount_retry(&tk_core.seq, seq));
2716	}
2717	EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2718
2719	/**
2720	* ktime_get_coarse_real_ts64_mg - return latter of coarse grained time or floor
2721	* @ts: timespec64 to be filled
2722	*
2723	* Fetch the global mg_floor value, convert it to realtime and compare it
2724	* to the current coarse-grained time. Fill @ts with whichever is
2725	* latest. Note that this is a filesystem-specific interface and should be
2726	* avoided outside of that context.
2727	*/
2728	void ktime_get_coarse_real_ts64_mg(struct timespec64 *ts)
2729	{
2730	struct timekeeper *tk = &tk_core.timekeeper;
2731	u64 floor = atomic64_read(&mg_floor);
2732	ktime_t f_real, offset, coarse;
2733	unsigned int seq;
2734
2735	do {
2736	seq = read_seqcount_begin(&tk_core.seq);
2737	*ts = tk_xtime_coarse(tk);
2738	offset = tk_core.timekeeper.offs_real;
2739	} while (read_seqcount_retry(&tk_core.seq, seq));
2740
2741	coarse = timespec64_to_ktime(*ts);
2742	f_real = ktime_add(floor, offset);
2743	if (ktime_after(f_real, coarse))
2744	*ts = ktime_to_timespec64(f_real);
2745	}
2746
2747	/**
2748	* ktime_get_real_ts64_mg - attempt to update floor value and return result
2749	* @ts: pointer to the timespec to be set
2750	*
2751	* Get a monotonic fine-grained time value and attempt to swap it into
2752	* mg_floor. If that succeeds then accept the new floor value. If it fails
2753	* then another task raced in during the interim time and updated the
2754	* floor. Since any update to the floor must be later than the previous
2755	* floor, either outcome is acceptable.
2756	*
2757	* Typically this will be called after calling ktime_get_coarse_real_ts64_mg(),
2758	* and determining that the resulting coarse-grained timestamp did not effect
2759	* a change in ctime. Any more recent floor value would effect a change to
2760	* ctime, so there is no need to retry the atomic64_try_cmpxchg() on failure.
2761	*
2762	* @ts will be filled with the latest floor value, regardless of the outcome of
2763	* the cmpxchg. Note that this is a filesystem specific interface and should be
2764	* avoided outside of that context.
2765	*/
2766	void ktime_get_real_ts64_mg(struct timespec64 *ts)
2767	{
2768	struct timekeeper *tk = &tk_core.timekeeper;
2769	ktime_t old = atomic64_read(&mg_floor);
2770	ktime_t offset, mono;
2771	unsigned int seq;
2772	u64 nsecs;
2773
2774	do {
2775	seq = read_seqcount_begin(&tk_core.seq);
2776
2777	ts->tv_sec = tk->xtime_sec;
2778	mono = tk->tkr_mono.base;
2779	nsecs = timekeeping_get_ns(&tk->tkr_mono);
2780	offset = tk_core.timekeeper.offs_real;
2781	} while (read_seqcount_retry(&tk_core.seq, seq));
2782
2783	mono = ktime_add_ns(mono, nsecs);
2784
2785	/*
2786	* Attempt to update the floor with the new time value. As any
2787	* update must be later then the existing floor, and would effect
2788	* a change to ctime from the perspective of the current task,
2789	* accept the resulting floor value regardless of the outcome of
2790	* the swap.
2791	*/
2792	if (atomic64_try_cmpxchg(&mg_floor, &old, mono)) {
2793	ts->tv_nsec = 0;
2794	timespec64_add_ns(ts, nsecs);
2795	timekeeping_inc_mg_floor_swaps();
2796	} else {
2797	/*
2798	* Another task changed mg_floor since "old" was fetched.
2799	* "old" has been updated with the latest value of "mg_floor".
2800	* That value is newer than the previous floor value, which
2801	* is enough to effect a change to ctime. Accept it.
2802	*/
2803	*ts = ktime_to_timespec64(ktime_add(old, offset));
2804	}
2805	}
2806
2807	void ktime_get_coarse_ts64(struct timespec64 *ts)
2808	{
2809	struct timekeeper *tk = &tk_core.timekeeper;
2810	struct timespec64 now, mono;
2811	unsigned int seq;
2812
2813	do {
2814	seq = read_seqcount_begin(&tk_core.seq);
2815
2816	now = tk_xtime_coarse(tk);
2817	mono = tk->wall_to_monotonic;
2818	} while (read_seqcount_retry(&tk_core.seq, seq));
2819
2820	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2821	now.tv_nsec + mono.tv_nsec);
2822	}
2823	EXPORT_SYMBOL(ktime_get_coarse_ts64);
2824
2825	/*
2826	* Must hold jiffies_lock
2827	*/
2828	void do_timer(unsigned long ticks)
2829	{
2830	jiffies_64 += ticks;
2831	calc_global_load();
2832	}
2833
2834	/**
2835	* ktime_get_update_offsets_now - hrtimer helper
2836	* @cwsseq: pointer to check and store the clock was set sequence number
2837	* @offs_real: pointer to storage for monotonic -> realtime offset
2838	* @offs_boot: pointer to storage for monotonic -> boottime offset
2839	* @offs_tai: pointer to storage for monotonic -> clock tai offset
2840	*
2841	* Returns current monotonic time and updates the offsets if the
2842	* sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2843	* different.
2844	*
2845	* Called from hrtimer_interrupt() or retrigger_next_event()
2846	*/
2847	ktime_t ktime_get_update_offsets_now(u32 cwsseq, ktime_t offs_real,
2848	ktime_t offs_boot, ktime_t offs_tai)
2849	{
2850	struct timekeeper *tk = &tk_core.timekeeper;
2851	unsigned int seq;
2852	ktime_t base;
2853	u64 nsecs;
2854
2855	do {
2856	seq = read_seqcount_begin(&tk_core.seq);
2857
2858	base = tk->tkr_mono.base;
2859	nsecs = timekeeping_get_ns(&tk->tkr_mono);
2860	base = ktime_add_ns(base, nsecs);
2861
2862	if (*cwsseq != tk->clock_was_set_seq) {
2863	*cwsseq = tk->clock_was_set_seq;
2864	*offs_real = tk->offs_real;
2865	*offs_boot = tk->offs_boot;
2866	*offs_tai = tk->offs_tai;
2867	}
2868
2869	/* Handle leapsecond insertion adjustments */
2870	if (unlikely(base >= tk->next_leap_ktime))
2871	*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2872
2873	} while (read_seqcount_retry(&tk_core.seq, seq));
2874
2875	return base;
2876	}
2877
2878	/*
2879	* timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2880	*/
2881	static int timekeeping_validate_timex(const struct __kernel_timex *txc, bool aux_clock)
2882	{
2883	if (txc->modes & ADJ_ADJTIME) {
2884	/* singleshot must not be used with any other mode bits */
2885	if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2886	return -EINVAL;
2887	if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2888	!capable(CAP_SYS_TIME))
2889	return -EPERM;
2890	} else {
2891	/* In order to modify anything, you gotta be super-user! */
2892	if (txc->modes && !capable(CAP_SYS_TIME))
2893	return -EPERM;
2894	/*
2895	* if the quartz is off by more than 10% then
2896	* something is VERY wrong!
2897	*/
2898	if (txc->modes & ADJ_TICK &&
2899	(txc->tick < 900000/USER_HZ \|\|
2900	txc->tick > 1100000/USER_HZ))
2901	return -EINVAL;
2902	}
2903
2904	if (txc->modes & ADJ_SETOFFSET) {
2905	/* In order to inject time, you gotta be super-user! */
2906	if (!capable(CAP_SYS_TIME))
2907	return -EPERM;
2908
2909	/*
2910	* Validate if a timespec/timeval used to inject a time
2911	* offset is valid. Offsets can be positive or negative, so
2912	* we don't check tv_sec. The value of the timeval/timespec
2913	* is the sum of its fields,but NOTE:
2914	* The field tv_usec/tv_nsec must always be non-negative and
2915	* we can't have more nanoseconds/microseconds than a second.
2916	*/
2917	if (txc->time.tv_usec < 0)
2918	return -EINVAL;
2919
2920	if (txc->modes & ADJ_NANO) {
2921	if (txc->time.tv_usec >= NSEC_PER_SEC)
2922	return -EINVAL;
2923	} else {
2924	if (txc->time.tv_usec >= USEC_PER_SEC)
2925	return -EINVAL;
2926	}
2927	}
2928
2929	/*
2930	* Check for potential multiplication overflows that can
2931	* only happen on 64-bit systems:
2932	*/
2933	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2934	if (LLONG_MIN / PPM_SCALE > txc->freq)
2935	return -EINVAL;
2936	if (LLONG_MAX / PPM_SCALE < txc->freq)
2937	return -EINVAL;
2938	}
2939
2940	if (aux_clock) {
2941	/* Auxiliary clocks are similar to TAI and do not have leap seconds */
2942	if (txc->modes & ADJ_STATUS &&
2943	txc->status & (STA_INS \| STA_DEL))
2944	return -EINVAL;
2945
2946	/* No TAI offset setting */
2947	if (txc->modes & ADJ_TAI)
2948	return -EINVAL;
2949
2950	/* No PPS support either */
2951	if (txc->modes & ADJ_STATUS &&
2952	txc->status & (STA_PPSFREQ \| STA_PPSTIME))
2953	return -EINVAL;
2954	}
2955
2956	return 0;
2957	}
2958
2959	/**
2960	* random_get_entropy_fallback - Returns the raw clock source value,
2961	* used by random.c for platforms with no valid random_get_entropy().
2962	*/
2963	unsigned long random_get_entropy_fallback(void)
2964	{
2965	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2966	struct clocksource *clock = READ_ONCE(tkr->clock);
2967
2968	if (unlikely(timekeeping_suspended \|\| !clock))
2969	return 0;
2970	return clock->read(clock);
2971	}
2972	EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2973
2974	struct adjtimex_result {
2975	struct audit_ntp_data ad;
2976	struct timespec64 delta;
2977	bool clock_set;
2978	};
2979
2980	static int __do_adjtimex(struct tk_data tkd, struct __kernel_timex txc,
2981	struct adjtimex_result *result)
2982	{
2983	struct timekeeper *tks = &tkd->shadow_timekeeper;
2984	bool aux_clock = !timekeeper_is_core_tk(tks);
2985	struct timespec64 ts;
2986	s32 orig_tai, tai;
2987	int ret;
2988
2989	/* Validate the data before disabling interrupts */
2990	ret = timekeeping_validate_timex(txc, aux_clock);
2991	if (ret)
2992	return ret;
2993	add_device_randomness(txc, sizeof(*txc));
2994
2995	if (!aux_clock) {
2996	ktime_get_real_ts64(&ts);
2997	} else {
2998	if (!tk_get_aux_ts64(tkd->timekeeper.id, &ts))
2999	return -ENODEV;
3000	}
3001
3002	add_device_randomness(&ts, sizeof(ts));
3003
3004	guard(raw_spinlock_irqsave)(&tkd->lock);
3005
3006	if (!tks->clock_valid)
3007	return -ENODEV;
3008
3009	if (txc->modes & ADJ_SETOFFSET) {
3010	result->delta.tv_sec = txc->time.tv_sec;
3011	result->delta.tv_nsec = txc->time.tv_usec;
3012	if (!(txc->modes & ADJ_NANO))
3013	result->delta.tv_nsec *= 1000;
3014	ret = __timekeeping_inject_offset(tkd, &result->delta);
3015	if (ret)
3016	return ret;
3017	result->clock_set = true;
3018	}
3019
3020	orig_tai = tai = tks->tai_offset;
3021	ret = ntp_adjtimex(tks->id, txc, &ts, &tai, &result->ad);
3022
3023	if (tai != orig_tai) {
3024	__timekeeping_set_tai_offset(tks, tai);
3025	timekeeping_update_from_shadow(tkd, TK_CLOCK_WAS_SET);
3026	result->clock_set = true;
3027	} else {
3028	tk_update_leap_state_all(tkd);
3029	}
3030
3031	/* Update the multiplier immediately if frequency was set directly */
3032	if (txc->modes & (ADJ_FREQUENCY \| ADJ_TICK))
3033	result->clock_set \|= __timekeeping_advance(tkd, TK_ADV_FREQ);
3034
3035	return ret;
3036	}
3037
3038	/**
3039	* do_adjtimex() - Accessor function to NTP __do_adjtimex function
3040	* @txc: Pointer to kernel_timex structure containing NTP parameters
3041	*/
3042	int do_adjtimex(struct __kernel_timex *txc)
3043	{
3044	struct adjtimex_result result = { };
3045	int ret;
3046
3047	ret = __do_adjtimex(&tk_core, txc, &result);
3048	if (ret < 0)
3049	return ret;
3050
3051	if (txc->modes & ADJ_SETOFFSET)
3052	audit_tk_injoffset(result.delta);
3053
3054	audit_ntp_log(&result.ad);
3055
3056	if (result.clock_set)
3057	clock_was_set(CLOCK_SET_WALL);
3058
3059	ntp_notify_cmos_timer(result.delta.tv_sec != 0);
3060
3061	return ret;
3062	}
3063
3064	/*
3065	* Invoked from NTP with the time keeper lock held, so lockless access is
3066	* fine.
3067	*/
3068	long ktime_get_ntp_seconds(unsigned int id)
3069	{
3070	return timekeeper_data[id].timekeeper.xtime_sec;
3071	}
3072
3073	#ifdef CONFIG_NTP_PPS
3074	/**
3075	* hardpps() - Accessor function to NTP __hardpps function
3076	* @phase_ts: Pointer to timespec64 structure representing phase timestamp
3077	* @raw_ts: Pointer to timespec64 structure representing raw timestamp
3078	*/
3079	void hardpps(const struct timespec64 phase_ts, const struct timespec64 raw_ts)
3080	{
3081	guard(raw_spinlock_irqsave)(&tk_core.lock);
3082	__hardpps(phase_ts, raw_ts);
3083	}
3084	EXPORT_SYMBOL(hardpps);
3085	#endif /* CONFIG_NTP_PPS */
3086
3087	#ifdef CONFIG_POSIX_AUX_CLOCKS
3088	#include "posix-timers.h"
3089
3090	/*
3091	* Bitmap for the activated auxiliary timekeepers to allow lockless quick
3092	* checks in the hot paths without touching extra cache lines. If set, then
3093	* the state of the corresponding timekeeper has to be re-checked under
3094	* timekeeper::lock.
3095	*/
3096	static unsigned long aux_timekeepers;
3097
3098	static inline unsigned int clockid_to_tkid(unsigned int id)
3099	{
3100	return TIMEKEEPER_AUX_FIRST + id - CLOCK_AUX;
3101	}
3102
3103	static inline struct tk_data *aux_get_tk_data(clockid_t id)
3104	{
3105	if (!clockid_is_aux_clock(id))
3106	return NULL;
3107	return &timekeeper_data[clockid_to_tkid(id)];
3108	}
3109
3110	/* Invoked from timekeeping after a clocksource change */
3111	static void tk_aux_update_clocksource(void)
3112	{
3113	unsigned long active = READ_ONCE(aux_timekeepers);
3114	unsigned int id;
3115
3116	for_each_set_bit(id, &active, BITS_PER_LONG) {
3117	struct tk_data *tkd = &timekeeper_data[id + TIMEKEEPER_AUX_FIRST];
3118	struct timekeeper *tks = &tkd->shadow_timekeeper;
3119
3120	guard(raw_spinlock_irqsave)(&tkd->lock);
3121	if (!tks->clock_valid)
3122	continue;
3123
3124	timekeeping_forward_now(tks);
3125	tk_setup_internals(tks, tk_core.timekeeper.tkr_raw.clock);
3126	timekeeping_update_from_shadow(tkd, TK_UPDATE_ALL);
3127	}
3128	}
3129
3130	static void tk_aux_advance(void)
3131	{
3132	unsigned long active = READ_ONCE(aux_timekeepers);
3133	unsigned int id;
3134
3135	/* Lockless quick check to avoid extra cache lines */
3136	for_each_set_bit(id, &active, BITS_PER_LONG) {
3137	struct tk_data *aux_tkd = &timekeeper_data[id + TIMEKEEPER_AUX_FIRST];
3138
3139	guard(raw_spinlock)(&aux_tkd->lock);
3140	if (aux_tkd->shadow_timekeeper.clock_valid)
3141	__timekeeping_advance(aux_tkd, TK_ADV_TICK);
3142	}
3143	}
3144
3145	/**
3146	* ktime_get_aux - Get time for a AUX clock
3147	* @id: ID of the clock to read (CLOCK_AUX...)
3148	* @kt: Pointer to ktime_t to store the time stamp
3149	*
3150	* Returns: True if the timestamp is valid, false otherwise
3151	*/
3152	bool ktime_get_aux(clockid_t id, ktime_t *kt)
3153	{
3154	struct tk_data *aux_tkd = aux_get_tk_data(id);
3155	struct timekeeper *aux_tk;
3156	unsigned int seq;
3157	ktime_t base;
3158	u64 nsecs;
3159
3160	WARN_ON(timekeeping_suspended);
3161
3162	if (!aux_tkd)
3163	return false;
3164
3165	aux_tk = &aux_tkd->timekeeper;
3166	do {
3167	seq = read_seqcount_begin(&aux_tkd->seq);
3168	if (!aux_tk->clock_valid)
3169	return false;
3170
3171	base = ktime_add(aux_tk->tkr_mono.base, aux_tk->offs_aux);
3172	nsecs = timekeeping_get_ns(&aux_tk->tkr_mono);
3173	} while (read_seqcount_retry(&aux_tkd->seq, seq));
3174
3175	*kt = ktime_add_ns(base, nsecs);
3176	return true;
3177	}
3178	EXPORT_SYMBOL_GPL(ktime_get_aux);
3179
3180	/**
3181	* ktime_get_aux_ts64 - Get time for a AUX clock
3182	* @id: ID of the clock to read (CLOCK_AUX...)
3183	* @ts: Pointer to timespec64 to store the time stamp
3184	*
3185	* Returns: True if the timestamp is valid, false otherwise
3186	*/
3187	bool ktime_get_aux_ts64(clockid_t id, struct timespec64 *ts)
3188	{
3189	ktime_t now;
3190
3191	if (!ktime_get_aux(id, &now))
3192	return false;
3193	*ts = ktime_to_timespec64(now);
3194	return true;
3195	}
3196	EXPORT_SYMBOL_GPL(ktime_get_aux_ts64);
3197
3198	static int aux_get_res(clockid_t id, struct timespec64 *tp)
3199	{
3200	if (!clockid_is_aux_clock(id))
3201	return -ENODEV;
3202
3203	tp->tv_sec = aux_clock_resolution_ns() / NSEC_PER_SEC;
3204	tp->tv_nsec = aux_clock_resolution_ns() % NSEC_PER_SEC;
3205	return 0;
3206	}
3207
3208	static int aux_get_timespec(clockid_t id, struct timespec64 *tp)
3209	{
3210	return ktime_get_aux_ts64(id, tp) ? 0 : -ENODEV;
3211	}
3212
3213	static int aux_clock_set(const clockid_t id, const struct timespec64 *tnew)
3214	{
3215	struct tk_data *aux_tkd = aux_get_tk_data(id);
3216	struct timekeeper *aux_tks;
3217	ktime_t tnow, nsecs;
3218
3219	if (!timespec64_valid_settod(tnew))
3220	return -EINVAL;
3221	if (!aux_tkd)
3222	return -ENODEV;
3223
3224	aux_tks = &aux_tkd->shadow_timekeeper;
3225
3226	guard(raw_spinlock_irq)(&aux_tkd->lock);
3227	if (!aux_tks->clock_valid)
3228	return -ENODEV;
3229
3230	/* Forward the timekeeper base time */
3231	timekeeping_forward_now(aux_tks);
3232	/*
3233	* Get the updated base time. tkr_mono.base has not been
3234	* updated yet, so do that first. That makes the update
3235	* in timekeeping_update_from_shadow() redundant, but
3236	* that's harmless. After that @tnow can be calculated
3237	* by using tkr_mono::cycle_last, which has been set
3238	* by timekeeping_forward_now().
3239	*/
3240	tk_update_ktime_data(aux_tks);
3241	nsecs = timekeeping_cycles_to_ns(&aux_tks->tkr_mono, aux_tks->tkr_mono.cycle_last);
3242	tnow = ktime_add(aux_tks->tkr_mono.base, nsecs);
3243
3244	/*
3245	* Calculate the new AUX offset as delta to @tnow ("monotonic").
3246	* That avoids all the tk::xtime back and forth conversions as
3247	* xtime ("realtime") is not applicable for auxiliary clocks and
3248	* kept in sync with "monotonic".
3249	*/
3250	tk_update_aux_offs(aux_tks, ktime_sub(timespec64_to_ktime(*tnew), tnow));
3251
3252	timekeeping_update_from_shadow(aux_tkd, TK_UPDATE_ALL);
3253	return 0;
3254	}
3255
3256	static int aux_clock_adj(const clockid_t id, struct __kernel_timex *txc)
3257	{
3258	struct tk_data *aux_tkd = aux_get_tk_data(id);
3259	struct adjtimex_result result = { };
3260
3261	if (!aux_tkd)
3262	return -ENODEV;
3263
3264	/*
3265	* @result is ignored for now as there are neither hrtimers nor a
3266	* RTC related to auxiliary clocks for now.
3267	*/
3268	return __do_adjtimex(aux_tkd, txc, &result);
3269	}
3270
3271	const struct k_clock clock_aux = {
3272	.clock_getres = aux_get_res,
3273	.clock_get_timespec = aux_get_timespec,
3274	.clock_set = aux_clock_set,
3275	.clock_adj = aux_clock_adj,
3276	};
3277
3278	static void aux_clock_enable(clockid_t id)
3279	{
3280	struct tk_read_base *tkr_raw = &tk_core.timekeeper.tkr_raw;
3281	struct tk_data *aux_tkd = aux_get_tk_data(id);
3282	struct timekeeper *aux_tks = &aux_tkd->shadow_timekeeper;
3283
3284	/* Prevent the core timekeeper from changing. */
3285	guard(raw_spinlock_irq)(&tk_core.lock);
3286
3287	/*
3288	* Setup the auxiliary clock assuming that the raw core timekeeper
3289	* clock frequency conversion is close enough. Userspace has to
3290	* adjust for the deviation via clock_adjtime(2).
3291	*/
3292	guard(raw_spinlock_nested)(&aux_tkd->lock);
3293
3294	/* Remove leftovers of a previous registration */
3295	memset(aux_tks, 0, sizeof(*aux_tks));
3296	/* Restore the timekeeper id */
3297	aux_tks->id = aux_tkd->timekeeper.id;
3298	/* Setup the timekeeper based on the current system clocksource */
3299	tk_setup_internals(aux_tks, tkr_raw->clock);
3300
3301	/* Mark it valid and set it live */
3302	aux_tks->clock_valid = true;
3303	timekeeping_update_from_shadow(aux_tkd, TK_UPDATE_ALL);
3304	}
3305
3306	static void aux_clock_disable(clockid_t id)
3307	{
3308	struct tk_data *aux_tkd = aux_get_tk_data(id);
3309
3310	guard(raw_spinlock_irq)(&aux_tkd->lock);
3311	aux_tkd->shadow_timekeeper.clock_valid = false;
3312	timekeeping_update_from_shadow(aux_tkd, TK_UPDATE_ALL);
3313	}
3314
3315	static DEFINE_MUTEX(aux_clock_mutex);
3316
3317	static ssize_t aux_clock_enable_store(struct kobject kobj, struct kobj_attribute attr,
3318	const char *buf, size_t count)
3319	{
3320	/* Lazy atoi() as name is "0..7" */
3321	int id = kobj->name[0] & 0x7;
3322	bool enable;
3323
3324	if (!capable(CAP_SYS_TIME))
3325	return -EPERM;
3326
3327	if (kstrtobool(buf, &enable) < 0)
3328	return -EINVAL;
3329
3330	guard(mutex)(&aux_clock_mutex);
3331	if (enable == test_bit(id, &aux_timekeepers))
3332	return count;
3333
3334	if (enable) {
3335	aux_clock_enable(CLOCK_AUX + id);
3336	set_bit(id, &aux_timekeepers);
3337	} else {
3338	aux_clock_disable(CLOCK_AUX + id);
3339	clear_bit(id, &aux_timekeepers);
3340	}
3341	return count;
3342	}
3343
3344	static ssize_t aux_clock_enable_show(struct kobject kobj, struct kobj_attribute attr, char *buf)
3345	{
3346	unsigned long active = READ_ONCE(aux_timekeepers);
3347	/* Lazy atoi() as name is "0..7" */
3348	int id = kobj->name[0] & 0x7;
3349
3350	return sysfs_emit(buf, "%d\n", test_bit(id, &active));
3351	}
3352
3353	static struct kobj_attribute aux_clock_enable_attr = __ATTR_RW(aux_clock_enable);
3354
3355	static struct attribute *aux_clock_enable_attrs[] = {
3356	&aux_clock_enable_attr.attr,
3357	NULL
3358	};
3359
3360	static const struct attribute_group aux_clock_enable_attr_group = {
3361	.attrs = aux_clock_enable_attrs,
3362	};
3363
3364	static int __init tk_aux_sysfs_init(void)
3365	{
3366	struct kobject auxo, tko = kobject_create_and_add("time", kernel_kobj);
3367	struct kobject *clks[MAX_AUX_CLOCKS];
3368	int ret = -ENOMEM;
3369	int i;
3370
3371	if (!tko)
3372	return ret;
3373
3374	auxo = kobject_create_and_add("aux_clocks", tko);
3375	if (!auxo)
3376	goto err_clean;
3377
3378	for (i = 0; i < MAX_AUX_CLOCKS; i++) {
3379	char id[2] = { [0] = '0' + i, };
3380	clks[i] = kobject_create_and_add(id, auxo);
3381
3382	if (!clks[i]) {
3383	ret = -ENOMEM;
3384	goto err_clks;
3385	}
3386
3387	ret = sysfs_create_group(clks[i], &aux_clock_enable_attr_group);
3388	if (ret)
3389	goto err_clk;
3390	}
3391	return 0;
3392
3393	err_clk:
3394	kobject_put(clks[i]);
3395	err_clks:
3396	while (--i >= 0) {
3397	sysfs_remove_group(clks[i], &aux_clock_enable_attr_group);
3398	kobject_put(clks[i]);
3399	}
3400	err_clean:
3401	kobject_put(auxo);
3402	kobject_put(tko);
3403	return ret;
3404	}
3405	late_initcall(tk_aux_sysfs_init);
3406
3407	static __init void tk_aux_setup(void)
3408	{
3409	for (int i = TIMEKEEPER_AUX_FIRST; i <= TIMEKEEPER_AUX_LAST; i++)
3410	tkd_basic_setup(&timekeeper_data[i], i, false);
3411	}
3412	#endif /* CONFIG_POSIX_AUX_CLOCKS */