xref: /linux/kernel/time/timekeeping.c (revision c25ca0c2e42c77e0241411d374d44c41e253b3f5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Kernel timekeeping code and accessor functions. Based on code from
4  *  timer.c, moved in commit 8524070b7982.
5  */
6 #include <linux/timekeeper_internal.h>
7 #include <linux/module.h>
8 #include <linux/interrupt.h>
9 #include <linux/percpu.h>
10 #include <linux/init.h>
11 #include <linux/mm.h>
12 #include <linux/nmi.h>
13 #include <linux/sched.h>
14 #include <linux/sched/loadavg.h>
15 #include <linux/sched/clock.h>
16 #include <linux/syscore_ops.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/timex.h>
21 #include <linux/tick.h>
22 #include <linux/stop_machine.h>
23 #include <linux/pvclock_gtod.h>
24 #include <linux/compiler.h>
25 #include <linux/audit.h>
26 #include <linux/random.h>
27 
28 #include "tick-internal.h"
29 #include "ntp_internal.h"
30 #include "timekeeping_internal.h"
31 
32 #define TK_CLEAR_NTP		(1 << 0)
33 #define TK_CLOCK_WAS_SET	(1 << 1)
34 
35 #define TK_UPDATE_ALL		(TK_CLEAR_NTP | TK_CLOCK_WAS_SET)
36 
37 enum timekeeping_adv_mode {
38 	/* Update timekeeper when a tick has passed */
39 	TK_ADV_TICK,
40 
41 	/* Update timekeeper on a direct frequency change */
42 	TK_ADV_FREQ
43 };
44 
45 /*
46  * The most important data for readout fits into a single 64 byte
47  * cache line.
48  */
49 struct tk_data {
50 	seqcount_raw_spinlock_t	seq;
51 	struct timekeeper	timekeeper;
52 	struct timekeeper	shadow_timekeeper;
53 	raw_spinlock_t		lock;
54 } ____cacheline_aligned;
55 
56 static struct tk_data tk_core;
57 
58 /* flag for if timekeeping is suspended */
59 int __read_mostly timekeeping_suspended;
60 
61 /**
62  * struct tk_fast - NMI safe timekeeper
63  * @seq:	Sequence counter for protecting updates. The lowest bit
64  *		is the index for the tk_read_base array
65  * @base:	tk_read_base array. Access is indexed by the lowest bit of
66  *		@seq.
67  *
68  * See @update_fast_timekeeper() below.
69  */
70 struct tk_fast {
71 	seqcount_latch_t	seq;
72 	struct tk_read_base	base[2];
73 };
74 
75 /* Suspend-time cycles value for halted fast timekeeper. */
76 static u64 cycles_at_suspend;
77 
dummy_clock_read(struct clocksource * cs)78 static u64 dummy_clock_read(struct clocksource *cs)
79 {
80 	if (timekeeping_suspended)
81 		return cycles_at_suspend;
82 	return local_clock();
83 }
84 
85 static struct clocksource dummy_clock = {
86 	.read = dummy_clock_read,
87 };
88 
89 /*
90  * Boot time initialization which allows local_clock() to be utilized
91  * during early boot when clocksources are not available. local_clock()
92  * returns nanoseconds already so no conversion is required, hence mult=1
93  * and shift=0. When the first proper clocksource is installed then
94  * the fast time keepers are updated with the correct values.
95  */
96 #define FAST_TK_INIT						\
97 	{							\
98 		.clock		= &dummy_clock,			\
99 		.mask		= CLOCKSOURCE_MASK(64),		\
100 		.mult		= 1,				\
101 		.shift		= 0,				\
102 	}
103 
104 static struct tk_fast tk_fast_mono ____cacheline_aligned = {
105 	.seq     = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
106 	.base[0] = FAST_TK_INIT,
107 	.base[1] = FAST_TK_INIT,
108 };
109 
110 static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
111 	.seq     = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
112 	.base[0] = FAST_TK_INIT,
113 	.base[1] = FAST_TK_INIT,
114 };
115 
timekeeper_lock_irqsave(void)116 unsigned long timekeeper_lock_irqsave(void)
117 {
118 	unsigned long flags;
119 
120 	raw_spin_lock_irqsave(&tk_core.lock, flags);
121 	return flags;
122 }
123 
timekeeper_unlock_irqrestore(unsigned long flags)124 void timekeeper_unlock_irqrestore(unsigned long flags)
125 {
126 	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
127 }
128 
129 /*
130  * Multigrain timestamps require tracking the latest fine-grained timestamp
131  * that has been issued, and never returning a coarse-grained timestamp that is
132  * earlier than that value.
133  *
134  * mg_floor represents the latest fine-grained time that has been handed out as
135  * a file timestamp on the system. This is tracked as a monotonic ktime_t, and
136  * converted to a realtime clock value on an as-needed basis.
137  *
138  * Maintaining mg_floor ensures the multigrain interfaces never issue a
139  * timestamp earlier than one that has been previously issued.
140  *
141  * The exception to this rule is when there is a backward realtime clock jump. If
142  * such an event occurs, a timestamp can appear to be earlier than a previous one.
143  */
144 static __cacheline_aligned_in_smp atomic64_t mg_floor;
145 
tk_normalize_xtime(struct timekeeper * tk)146 static inline void tk_normalize_xtime(struct timekeeper *tk)
147 {
148 	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
149 		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
150 		tk->xtime_sec++;
151 	}
152 	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
153 		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
154 		tk->raw_sec++;
155 	}
156 }
157 
tk_xtime(const struct timekeeper * tk)158 static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
159 {
160 	struct timespec64 ts;
161 
162 	ts.tv_sec = tk->xtime_sec;
163 	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
164 	return ts;
165 }
166 
tk_set_xtime(struct timekeeper * tk,const struct timespec64 * ts)167 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
168 {
169 	tk->xtime_sec = ts->tv_sec;
170 	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
171 }
172 
tk_xtime_add(struct timekeeper * tk,const struct timespec64 * ts)173 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
174 {
175 	tk->xtime_sec += ts->tv_sec;
176 	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
177 	tk_normalize_xtime(tk);
178 }
179 
tk_set_wall_to_mono(struct timekeeper * tk,struct timespec64 wtm)180 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
181 {
182 	struct timespec64 tmp;
183 
184 	/*
185 	 * Verify consistency of: offset_real = -wall_to_monotonic
186 	 * before modifying anything
187 	 */
188 	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
189 					-tk->wall_to_monotonic.tv_nsec);
190 	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
191 	tk->wall_to_monotonic = wtm;
192 	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
193 	/* Paired with READ_ONCE() in ktime_mono_to_any() */
194 	WRITE_ONCE(tk->offs_real, timespec64_to_ktime(tmp));
195 	WRITE_ONCE(tk->offs_tai, ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0)));
196 }
197 
tk_update_sleep_time(struct timekeeper * tk,ktime_t delta)198 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
199 {
200 	/* Paired with READ_ONCE() in ktime_mono_to_any() */
201 	WRITE_ONCE(tk->offs_boot, ktime_add(tk->offs_boot, delta));
202 	/*
203 	 * Timespec representation for VDSO update to avoid 64bit division
204 	 * on every update.
205 	 */
206 	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
207 }
208 
209 /*
210  * tk_clock_read - atomic clocksource read() helper
211  *
212  * This helper is necessary to use in the read paths because, while the
213  * seqcount ensures we don't return a bad value while structures are updated,
214  * it doesn't protect from potential crashes. There is the possibility that
215  * the tkr's clocksource may change between the read reference, and the
216  * clock reference passed to the read function.  This can cause crashes if
217  * the wrong clocksource is passed to the wrong read function.
218  * This isn't necessary to use when holding the tk_core.lock or doing
219  * a read of the fast-timekeeper tkrs (which is protected by its own locking
220  * and update logic).
221  */
tk_clock_read(const struct tk_read_base * tkr)222 static inline u64 tk_clock_read(const struct tk_read_base *tkr)
223 {
224 	struct clocksource *clock = READ_ONCE(tkr->clock);
225 
226 	return clock->read(clock);
227 }
228 
229 /**
230  * tk_setup_internals - Set up internals to use clocksource clock.
231  *
232  * @tk:		The target timekeeper to setup.
233  * @clock:		Pointer to clocksource.
234  *
235  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
236  * pair and interval request.
237  *
238  * Unless you're the timekeeping code, you should not be using this!
239  */
tk_setup_internals(struct timekeeper * tk,struct clocksource * clock)240 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
241 {
242 	u64 interval;
243 	u64 tmp, ntpinterval;
244 	struct clocksource *old_clock;
245 
246 	++tk->cs_was_changed_seq;
247 	old_clock = tk->tkr_mono.clock;
248 	tk->tkr_mono.clock = clock;
249 	tk->tkr_mono.mask = clock->mask;
250 	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
251 
252 	tk->tkr_raw.clock = clock;
253 	tk->tkr_raw.mask = clock->mask;
254 	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
255 
256 	/* Do the ns -> cycle conversion first, using original mult */
257 	tmp = NTP_INTERVAL_LENGTH;
258 	tmp <<= clock->shift;
259 	ntpinterval = tmp;
260 	tmp += clock->mult/2;
261 	do_div(tmp, clock->mult);
262 	if (tmp == 0)
263 		tmp = 1;
264 
265 	interval = (u64) tmp;
266 	tk->cycle_interval = interval;
267 
268 	/* Go back from cycles -> shifted ns */
269 	tk->xtime_interval = interval * clock->mult;
270 	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
271 	tk->raw_interval = interval * clock->mult;
272 
273 	 /* if changing clocks, convert xtime_nsec shift units */
274 	if (old_clock) {
275 		int shift_change = clock->shift - old_clock->shift;
276 		if (shift_change < 0) {
277 			tk->tkr_mono.xtime_nsec >>= -shift_change;
278 			tk->tkr_raw.xtime_nsec >>= -shift_change;
279 		} else {
280 			tk->tkr_mono.xtime_nsec <<= shift_change;
281 			tk->tkr_raw.xtime_nsec <<= shift_change;
282 		}
283 	}
284 
285 	tk->tkr_mono.shift = clock->shift;
286 	tk->tkr_raw.shift = clock->shift;
287 
288 	tk->ntp_error = 0;
289 	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
290 	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
291 
292 	/*
293 	 * The timekeeper keeps its own mult values for the currently
294 	 * active clocksource. These value will be adjusted via NTP
295 	 * to counteract clock drifting.
296 	 */
297 	tk->tkr_mono.mult = clock->mult;
298 	tk->tkr_raw.mult = clock->mult;
299 	tk->ntp_err_mult = 0;
300 	tk->skip_second_overflow = 0;
301 }
302 
303 /* Timekeeper helper functions. */
delta_to_ns_safe(const struct tk_read_base * tkr,u64 delta)304 static noinline u64 delta_to_ns_safe(const struct tk_read_base *tkr, u64 delta)
305 {
306 	return mul_u64_u32_add_u64_shr(delta, tkr->mult, tkr->xtime_nsec, tkr->shift);
307 }
308 
timekeeping_cycles_to_ns(const struct tk_read_base * tkr,u64 cycles)309 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
310 {
311 	/* Calculate the delta since the last update_wall_time() */
312 	u64 mask = tkr->mask, delta = (cycles - tkr->cycle_last) & mask;
313 
314 	/*
315 	 * This detects both negative motion and the case where the delta
316 	 * overflows the multiplication with tkr->mult.
317 	 */
318 	if (unlikely(delta > tkr->clock->max_cycles)) {
319 		/*
320 		 * Handle clocksource inconsistency between CPUs to prevent
321 		 * time from going backwards by checking for the MSB of the
322 		 * mask being set in the delta.
323 		 */
324 		if (delta & ~(mask >> 1))
325 			return tkr->xtime_nsec >> tkr->shift;
326 
327 		return delta_to_ns_safe(tkr, delta);
328 	}
329 
330 	return ((delta * tkr->mult) + tkr->xtime_nsec) >> tkr->shift;
331 }
332 
timekeeping_get_ns(const struct tk_read_base * tkr)333 static __always_inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
334 {
335 	return timekeeping_cycles_to_ns(tkr, tk_clock_read(tkr));
336 }
337 
338 /**
339  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
340  * @tkr: Timekeeping readout base from which we take the update
341  * @tkf: Pointer to NMI safe timekeeper
342  *
343  * We want to use this from any context including NMI and tracing /
344  * instrumenting the timekeeping code itself.
345  *
346  * Employ the latch technique; see @write_seqcount_latch.
347  *
348  * So if a NMI hits the update of base[0] then it will use base[1]
349  * which is still consistent. In the worst case this can result is a
350  * slightly wrong timestamp (a few nanoseconds). See
351  * @ktime_get_mono_fast_ns.
352  */
update_fast_timekeeper(const struct tk_read_base * tkr,struct tk_fast * tkf)353 static void update_fast_timekeeper(const struct tk_read_base *tkr,
354 				   struct tk_fast *tkf)
355 {
356 	struct tk_read_base *base = tkf->base;
357 
358 	/* Force readers off to base[1] */
359 	write_seqcount_latch_begin(&tkf->seq);
360 
361 	/* Update base[0] */
362 	memcpy(base, tkr, sizeof(*base));
363 
364 	/* Force readers back to base[0] */
365 	write_seqcount_latch(&tkf->seq);
366 
367 	/* Update base[1] */
368 	memcpy(base + 1, base, sizeof(*base));
369 
370 	write_seqcount_latch_end(&tkf->seq);
371 }
372 
__ktime_get_fast_ns(struct tk_fast * tkf)373 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
374 {
375 	struct tk_read_base *tkr;
376 	unsigned int seq;
377 	u64 now;
378 
379 	do {
380 		seq = read_seqcount_latch(&tkf->seq);
381 		tkr = tkf->base + (seq & 0x01);
382 		now = ktime_to_ns(tkr->base);
383 		now += timekeeping_get_ns(tkr);
384 	} while (read_seqcount_latch_retry(&tkf->seq, seq));
385 
386 	return now;
387 }
388 
389 /**
390  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
391  *
392  * This timestamp is not guaranteed to be monotonic across an update.
393  * The timestamp is calculated by:
394  *
395  *	now = base_mono + clock_delta * slope
396  *
397  * So if the update lowers the slope, readers who are forced to the
398  * not yet updated second array are still using the old steeper slope.
399  *
400  * tmono
401  * ^
402  * |    o  n
403  * |   o n
404  * |  u
405  * | o
406  * |o
407  * |12345678---> reader order
408  *
409  * o = old slope
410  * u = update
411  * n = new slope
412  *
413  * So reader 6 will observe time going backwards versus reader 5.
414  *
415  * While other CPUs are likely to be able to observe that, the only way
416  * for a CPU local observation is when an NMI hits in the middle of
417  * the update. Timestamps taken from that NMI context might be ahead
418  * of the following timestamps. Callers need to be aware of that and
419  * deal with it.
420  */
ktime_get_mono_fast_ns(void)421 u64 notrace ktime_get_mono_fast_ns(void)
422 {
423 	return __ktime_get_fast_ns(&tk_fast_mono);
424 }
425 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
426 
427 /**
428  * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
429  *
430  * Contrary to ktime_get_mono_fast_ns() this is always correct because the
431  * conversion factor is not affected by NTP/PTP correction.
432  */
ktime_get_raw_fast_ns(void)433 u64 notrace ktime_get_raw_fast_ns(void)
434 {
435 	return __ktime_get_fast_ns(&tk_fast_raw);
436 }
437 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
438 
439 /**
440  * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
441  *
442  * To keep it NMI safe since we're accessing from tracing, we're not using a
443  * separate timekeeper with updates to monotonic clock and boot offset
444  * protected with seqcounts. This has the following minor side effects:
445  *
446  * (1) Its possible that a timestamp be taken after the boot offset is updated
447  * but before the timekeeper is updated. If this happens, the new boot offset
448  * is added to the old timekeeping making the clock appear to update slightly
449  * earlier:
450  *    CPU 0                                        CPU 1
451  *    timekeeping_inject_sleeptime64()
452  *    __timekeeping_inject_sleeptime(tk, delta);
453  *                                                 timestamp();
454  *    timekeeping_update_staged(tkd, TK_CLEAR_NTP...);
455  *
456  * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
457  * partially updated.  Since the tk->offs_boot update is a rare event, this
458  * should be a rare occurrence which postprocessing should be able to handle.
459  *
460  * The caveats vs. timestamp ordering as documented for ktime_get_mono_fast_ns()
461  * apply as well.
462  */
ktime_get_boot_fast_ns(void)463 u64 notrace ktime_get_boot_fast_ns(void)
464 {
465 	struct timekeeper *tk = &tk_core.timekeeper;
466 
467 	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
468 }
469 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
470 
471 /**
472  * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
473  *
474  * The same limitations as described for ktime_get_boot_fast_ns() apply. The
475  * mono time and the TAI offset are not read atomically which may yield wrong
476  * readouts. However, an update of the TAI offset is an rare event e.g., caused
477  * by settime or adjtimex with an offset. The user of this function has to deal
478  * with the possibility of wrong timestamps in post processing.
479  */
ktime_get_tai_fast_ns(void)480 u64 notrace ktime_get_tai_fast_ns(void)
481 {
482 	struct timekeeper *tk = &tk_core.timekeeper;
483 
484 	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
485 }
486 EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
487 
__ktime_get_real_fast(struct tk_fast * tkf,u64 * mono)488 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
489 {
490 	struct tk_read_base *tkr;
491 	u64 basem, baser, delta;
492 	unsigned int seq;
493 
494 	do {
495 		seq = raw_read_seqcount_latch(&tkf->seq);
496 		tkr = tkf->base + (seq & 0x01);
497 		basem = ktime_to_ns(tkr->base);
498 		baser = ktime_to_ns(tkr->base_real);
499 		delta = timekeeping_get_ns(tkr);
500 	} while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
501 
502 	if (mono)
503 		*mono = basem + delta;
504 	return baser + delta;
505 }
506 
507 /**
508  * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
509  *
510  * See ktime_get_mono_fast_ns() for documentation of the time stamp ordering.
511  */
ktime_get_real_fast_ns(void)512 u64 ktime_get_real_fast_ns(void)
513 {
514 	return __ktime_get_real_fast(&tk_fast_mono, NULL);
515 }
516 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
517 
518 /**
519  * ktime_get_fast_timestamps: - NMI safe timestamps
520  * @snapshot:	Pointer to timestamp storage
521  *
522  * Stores clock monotonic, boottime and realtime timestamps.
523  *
524  * Boot time is a racy access on 32bit systems if the sleep time injection
525  * happens late during resume and not in timekeeping_resume(). That could
526  * be avoided by expanding struct tk_read_base with boot offset for 32bit
527  * and adding more overhead to the update. As this is a hard to observe
528  * once per resume event which can be filtered with reasonable effort using
529  * the accurate mono/real timestamps, it's probably not worth the trouble.
530  *
531  * Aside of that it might be possible on 32 and 64 bit to observe the
532  * following when the sleep time injection happens late:
533  *
534  * CPU 0				CPU 1
535  * timekeeping_resume()
536  * ktime_get_fast_timestamps()
537  *	mono, real = __ktime_get_real_fast()
538  *					inject_sleep_time()
539  *					   update boot offset
540  *	boot = mono + bootoffset;
541  *
542  * That means that boot time already has the sleep time adjustment, but
543  * real time does not. On the next readout both are in sync again.
544  *
545  * Preventing this for 64bit is not really feasible without destroying the
546  * careful cache layout of the timekeeper because the sequence count and
547  * struct tk_read_base would then need two cache lines instead of one.
548  *
549  * Access to the time keeper clock source is disabled across the innermost
550  * steps of suspend/resume. The accessors still work, but the timestamps
551  * are frozen until time keeping is resumed which happens very early.
552  *
553  * For regular suspend/resume there is no observable difference vs. sched
554  * clock, but it might affect some of the nasty low level debug printks.
555  *
556  * OTOH, access to sched clock is not guaranteed across suspend/resume on
557  * all systems either so it depends on the hardware in use.
558  *
559  * If that turns out to be a real problem then this could be mitigated by
560  * using sched clock in a similar way as during early boot. But it's not as
561  * trivial as on early boot because it needs some careful protection
562  * against the clock monotonic timestamp jumping backwards on resume.
563  */
ktime_get_fast_timestamps(struct ktime_timestamps * snapshot)564 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
565 {
566 	struct timekeeper *tk = &tk_core.timekeeper;
567 
568 	snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
569 	snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
570 }
571 
572 /**
573  * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
574  * @tk: Timekeeper to snapshot.
575  *
576  * It generally is unsafe to access the clocksource after timekeeping has been
577  * suspended, so take a snapshot of the readout base of @tk and use it as the
578  * fast timekeeper's readout base while suspended.  It will return the same
579  * number of cycles every time until timekeeping is resumed at which time the
580  * proper readout base for the fast timekeeper will be restored automatically.
581  */
halt_fast_timekeeper(const struct timekeeper * tk)582 static void halt_fast_timekeeper(const struct timekeeper *tk)
583 {
584 	static struct tk_read_base tkr_dummy;
585 	const struct tk_read_base *tkr = &tk->tkr_mono;
586 
587 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
588 	cycles_at_suspend = tk_clock_read(tkr);
589 	tkr_dummy.clock = &dummy_clock;
590 	tkr_dummy.base_real = tkr->base + tk->offs_real;
591 	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
592 
593 	tkr = &tk->tkr_raw;
594 	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
595 	tkr_dummy.clock = &dummy_clock;
596 	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
597 }
598 
599 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
600 
update_pvclock_gtod(struct timekeeper * tk,bool was_set)601 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
602 {
603 	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
604 }
605 
606 /**
607  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
608  * @nb: Pointer to the notifier block to register
609  */
pvclock_gtod_register_notifier(struct notifier_block * nb)610 int pvclock_gtod_register_notifier(struct notifier_block *nb)
611 {
612 	struct timekeeper *tk = &tk_core.timekeeper;
613 	int ret;
614 
615 	guard(raw_spinlock_irqsave)(&tk_core.lock);
616 	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
617 	update_pvclock_gtod(tk, true);
618 
619 	return ret;
620 }
621 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
622 
623 /**
624  * pvclock_gtod_unregister_notifier - unregister a pvclock
625  * timedata update listener
626  * @nb: Pointer to the notifier block to unregister
627  */
pvclock_gtod_unregister_notifier(struct notifier_block * nb)628 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
629 {
630 	guard(raw_spinlock_irqsave)(&tk_core.lock);
631 	return raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
632 }
633 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
634 
635 /*
636  * tk_update_leap_state - helper to update the next_leap_ktime
637  */
tk_update_leap_state(struct timekeeper * tk)638 static inline void tk_update_leap_state(struct timekeeper *tk)
639 {
640 	tk->next_leap_ktime = ntp_get_next_leap();
641 	if (tk->next_leap_ktime != KTIME_MAX)
642 		/* Convert to monotonic time */
643 		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
644 }
645 
646 /*
647  * Leap state update for both shadow and the real timekeeper
648  * Separate to spare a full memcpy() of the timekeeper.
649  */
tk_update_leap_state_all(struct tk_data * tkd)650 static void tk_update_leap_state_all(struct tk_data *tkd)
651 {
652 	write_seqcount_begin(&tkd->seq);
653 	tk_update_leap_state(&tkd->shadow_timekeeper);
654 	tkd->timekeeper.next_leap_ktime = tkd->shadow_timekeeper.next_leap_ktime;
655 	write_seqcount_end(&tkd->seq);
656 }
657 
658 /*
659  * Update the ktime_t based scalar nsec members of the timekeeper
660  */
tk_update_ktime_data(struct timekeeper * tk)661 static inline void tk_update_ktime_data(struct timekeeper *tk)
662 {
663 	u64 seconds;
664 	u32 nsec;
665 
666 	/*
667 	 * The xtime based monotonic readout is:
668 	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
669 	 * The ktime based monotonic readout is:
670 	 *	nsec = base_mono + now();
671 	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
672 	 */
673 	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
674 	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
675 	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
676 
677 	/*
678 	 * The sum of the nanoseconds portions of xtime and
679 	 * wall_to_monotonic can be greater/equal one second. Take
680 	 * this into account before updating tk->ktime_sec.
681 	 */
682 	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
683 	if (nsec >= NSEC_PER_SEC)
684 		seconds++;
685 	tk->ktime_sec = seconds;
686 
687 	/* Update the monotonic raw base */
688 	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
689 }
690 
691 /*
692  * Restore the shadow timekeeper from the real timekeeper.
693  */
timekeeping_restore_shadow(struct tk_data * tkd)694 static void timekeeping_restore_shadow(struct tk_data *tkd)
695 {
696 	lockdep_assert_held(&tkd->lock);
697 	memcpy(&tkd->shadow_timekeeper, &tkd->timekeeper, sizeof(tkd->timekeeper));
698 }
699 
timekeeping_update_from_shadow(struct tk_data * tkd,unsigned int action)700 static void timekeeping_update_from_shadow(struct tk_data *tkd, unsigned int action)
701 {
702 	struct timekeeper *tk = &tk_core.shadow_timekeeper;
703 
704 	lockdep_assert_held(&tkd->lock);
705 
706 	/*
707 	 * Block out readers before running the updates below because that
708 	 * updates VDSO and other time related infrastructure. Not blocking
709 	 * the readers might let a reader see time going backwards when
710 	 * reading from the VDSO after the VDSO update and then reading in
711 	 * the kernel from the timekeeper before that got updated.
712 	 */
713 	write_seqcount_begin(&tkd->seq);
714 
715 	if (action & TK_CLEAR_NTP) {
716 		tk->ntp_error = 0;
717 		ntp_clear();
718 	}
719 
720 	tk_update_leap_state(tk);
721 	tk_update_ktime_data(tk);
722 
723 	update_vsyscall(tk);
724 	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
725 
726 	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
727 	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
728 	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
729 
730 	if (action & TK_CLOCK_WAS_SET)
731 		tk->clock_was_set_seq++;
732 
733 	/*
734 	 * Update the real timekeeper.
735 	 *
736 	 * We could avoid this memcpy() by switching pointers, but that has
737 	 * the downside that the reader side does not longer benefit from
738 	 * the cacheline optimized data layout of the timekeeper and requires
739 	 * another indirection.
740 	 */
741 	memcpy(&tkd->timekeeper, tk, sizeof(*tk));
742 	write_seqcount_end(&tkd->seq);
743 }
744 
745 /**
746  * timekeeping_forward_now - update clock to the current time
747  * @tk:		Pointer to the timekeeper to update
748  *
749  * Forward the current clock to update its state since the last call to
750  * update_wall_time(). This is useful before significant clock changes,
751  * as it avoids having to deal with this time offset explicitly.
752  */
timekeeping_forward_now(struct timekeeper * tk)753 static void timekeeping_forward_now(struct timekeeper *tk)
754 {
755 	u64 cycle_now, delta;
756 
757 	cycle_now = tk_clock_read(&tk->tkr_mono);
758 	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask,
759 				  tk->tkr_mono.clock->max_raw_delta);
760 	tk->tkr_mono.cycle_last = cycle_now;
761 	tk->tkr_raw.cycle_last  = cycle_now;
762 
763 	while (delta > 0) {
764 		u64 max = tk->tkr_mono.clock->max_cycles;
765 		u64 incr = delta < max ? delta : max;
766 
767 		tk->tkr_mono.xtime_nsec += incr * tk->tkr_mono.mult;
768 		tk->tkr_raw.xtime_nsec += incr * tk->tkr_raw.mult;
769 		tk_normalize_xtime(tk);
770 		delta -= incr;
771 	}
772 }
773 
774 /**
775  * ktime_get_real_ts64 - Returns the time of day in a timespec64.
776  * @ts:		pointer to the timespec to be set
777  *
778  * Returns the time of day in a timespec64 (WARN if suspended).
779  */
ktime_get_real_ts64(struct timespec64 * ts)780 void ktime_get_real_ts64(struct timespec64 *ts)
781 {
782 	struct timekeeper *tk = &tk_core.timekeeper;
783 	unsigned int seq;
784 	u64 nsecs;
785 
786 	WARN_ON(timekeeping_suspended);
787 
788 	do {
789 		seq = read_seqcount_begin(&tk_core.seq);
790 
791 		ts->tv_sec = tk->xtime_sec;
792 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
793 
794 	} while (read_seqcount_retry(&tk_core.seq, seq));
795 
796 	ts->tv_nsec = 0;
797 	timespec64_add_ns(ts, nsecs);
798 }
799 EXPORT_SYMBOL(ktime_get_real_ts64);
800 
ktime_get(void)801 ktime_t ktime_get(void)
802 {
803 	struct timekeeper *tk = &tk_core.timekeeper;
804 	unsigned int seq;
805 	ktime_t base;
806 	u64 nsecs;
807 
808 	WARN_ON(timekeeping_suspended);
809 
810 	do {
811 		seq = read_seqcount_begin(&tk_core.seq);
812 		base = tk->tkr_mono.base;
813 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
814 
815 	} while (read_seqcount_retry(&tk_core.seq, seq));
816 
817 	return ktime_add_ns(base, nsecs);
818 }
819 EXPORT_SYMBOL_GPL(ktime_get);
820 
ktime_get_resolution_ns(void)821 u32 ktime_get_resolution_ns(void)
822 {
823 	struct timekeeper *tk = &tk_core.timekeeper;
824 	unsigned int seq;
825 	u32 nsecs;
826 
827 	WARN_ON(timekeeping_suspended);
828 
829 	do {
830 		seq = read_seqcount_begin(&tk_core.seq);
831 		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
832 	} while (read_seqcount_retry(&tk_core.seq, seq));
833 
834 	return nsecs;
835 }
836 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
837 
838 static ktime_t *offsets[TK_OFFS_MAX] = {
839 	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
840 	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
841 	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
842 };
843 
ktime_get_with_offset(enum tk_offsets offs)844 ktime_t ktime_get_with_offset(enum tk_offsets offs)
845 {
846 	struct timekeeper *tk = &tk_core.timekeeper;
847 	unsigned int seq;
848 	ktime_t base, *offset = offsets[offs];
849 	u64 nsecs;
850 
851 	WARN_ON(timekeeping_suspended);
852 
853 	do {
854 		seq = read_seqcount_begin(&tk_core.seq);
855 		base = ktime_add(tk->tkr_mono.base, *offset);
856 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
857 
858 	} while (read_seqcount_retry(&tk_core.seq, seq));
859 
860 	return ktime_add_ns(base, nsecs);
861 
862 }
863 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
864 
ktime_get_coarse_with_offset(enum tk_offsets offs)865 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
866 {
867 	struct timekeeper *tk = &tk_core.timekeeper;
868 	unsigned int seq;
869 	ktime_t base, *offset = offsets[offs];
870 	u64 nsecs;
871 
872 	WARN_ON(timekeeping_suspended);
873 
874 	do {
875 		seq = read_seqcount_begin(&tk_core.seq);
876 		base = ktime_add(tk->tkr_mono.base, *offset);
877 		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
878 
879 	} while (read_seqcount_retry(&tk_core.seq, seq));
880 
881 	return ktime_add_ns(base, nsecs);
882 }
883 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
884 
885 /**
886  * ktime_mono_to_any() - convert monotonic time to any other time
887  * @tmono:	time to convert.
888  * @offs:	which offset to use
889  */
ktime_mono_to_any(ktime_t tmono,enum tk_offsets offs)890 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
891 {
892 	ktime_t *offset = offsets[offs];
893 	unsigned int seq;
894 	ktime_t tconv;
895 
896 	if (IS_ENABLED(CONFIG_64BIT)) {
897 		/*
898 		 * Paired with WRITE_ONCE()s in tk_set_wall_to_mono() and
899 		 * tk_update_sleep_time().
900 		 */
901 		return ktime_add(tmono, READ_ONCE(*offset));
902 	}
903 
904 	do {
905 		seq = read_seqcount_begin(&tk_core.seq);
906 		tconv = ktime_add(tmono, *offset);
907 	} while (read_seqcount_retry(&tk_core.seq, seq));
908 
909 	return tconv;
910 }
911 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
912 
913 /**
914  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
915  */
ktime_get_raw(void)916 ktime_t ktime_get_raw(void)
917 {
918 	struct timekeeper *tk = &tk_core.timekeeper;
919 	unsigned int seq;
920 	ktime_t base;
921 	u64 nsecs;
922 
923 	do {
924 		seq = read_seqcount_begin(&tk_core.seq);
925 		base = tk->tkr_raw.base;
926 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
927 
928 	} while (read_seqcount_retry(&tk_core.seq, seq));
929 
930 	return ktime_add_ns(base, nsecs);
931 }
932 EXPORT_SYMBOL_GPL(ktime_get_raw);
933 
934 /**
935  * ktime_get_ts64 - get the monotonic clock in timespec64 format
936  * @ts:		pointer to timespec variable
937  *
938  * The function calculates the monotonic clock from the realtime
939  * clock and the wall_to_monotonic offset and stores the result
940  * in normalized timespec64 format in the variable pointed to by @ts.
941  */
ktime_get_ts64(struct timespec64 * ts)942 void ktime_get_ts64(struct timespec64 *ts)
943 {
944 	struct timekeeper *tk = &tk_core.timekeeper;
945 	struct timespec64 tomono;
946 	unsigned int seq;
947 	u64 nsec;
948 
949 	WARN_ON(timekeeping_suspended);
950 
951 	do {
952 		seq = read_seqcount_begin(&tk_core.seq);
953 		ts->tv_sec = tk->xtime_sec;
954 		nsec = timekeeping_get_ns(&tk->tkr_mono);
955 		tomono = tk->wall_to_monotonic;
956 
957 	} while (read_seqcount_retry(&tk_core.seq, seq));
958 
959 	ts->tv_sec += tomono.tv_sec;
960 	ts->tv_nsec = 0;
961 	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
962 }
963 EXPORT_SYMBOL_GPL(ktime_get_ts64);
964 
965 /**
966  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
967  *
968  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
969  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
970  * works on both 32 and 64 bit systems. On 32 bit systems the readout
971  * covers ~136 years of uptime which should be enough to prevent
972  * premature wrap arounds.
973  */
ktime_get_seconds(void)974 time64_t ktime_get_seconds(void)
975 {
976 	struct timekeeper *tk = &tk_core.timekeeper;
977 
978 	WARN_ON(timekeeping_suspended);
979 	return tk->ktime_sec;
980 }
981 EXPORT_SYMBOL_GPL(ktime_get_seconds);
982 
983 /**
984  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
985  *
986  * Returns the wall clock seconds since 1970.
987  *
988  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
989  * 32bit systems the access must be protected with the sequence
990  * counter to provide "atomic" access to the 64bit tk->xtime_sec
991  * value.
992  */
ktime_get_real_seconds(void)993 time64_t ktime_get_real_seconds(void)
994 {
995 	struct timekeeper *tk = &tk_core.timekeeper;
996 	time64_t seconds;
997 	unsigned int seq;
998 
999 	if (IS_ENABLED(CONFIG_64BIT))
1000 		return tk->xtime_sec;
1001 
1002 	do {
1003 		seq = read_seqcount_begin(&tk_core.seq);
1004 		seconds = tk->xtime_sec;
1005 
1006 	} while (read_seqcount_retry(&tk_core.seq, seq));
1007 
1008 	return seconds;
1009 }
1010 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1011 
1012 /**
1013  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1014  * but without the sequence counter protect. This internal function
1015  * is called just when timekeeping lock is already held.
1016  */
__ktime_get_real_seconds(void)1017 noinstr time64_t __ktime_get_real_seconds(void)
1018 {
1019 	struct timekeeper *tk = &tk_core.timekeeper;
1020 
1021 	return tk->xtime_sec;
1022 }
1023 
1024 /**
1025  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1026  * @systime_snapshot:	pointer to struct receiving the system time snapshot
1027  */
ktime_get_snapshot(struct system_time_snapshot * systime_snapshot)1028 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1029 {
1030 	struct timekeeper *tk = &tk_core.timekeeper;
1031 	unsigned int seq;
1032 	ktime_t base_raw;
1033 	ktime_t base_real;
1034 	ktime_t base_boot;
1035 	u64 nsec_raw;
1036 	u64 nsec_real;
1037 	u64 now;
1038 
1039 	WARN_ON_ONCE(timekeeping_suspended);
1040 
1041 	do {
1042 		seq = read_seqcount_begin(&tk_core.seq);
1043 		now = tk_clock_read(&tk->tkr_mono);
1044 		systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1045 		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1046 		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1047 		base_real = ktime_add(tk->tkr_mono.base,
1048 				      tk_core.timekeeper.offs_real);
1049 		base_boot = ktime_add(tk->tkr_mono.base,
1050 				      tk_core.timekeeper.offs_boot);
1051 		base_raw = tk->tkr_raw.base;
1052 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1053 		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1054 	} while (read_seqcount_retry(&tk_core.seq, seq));
1055 
1056 	systime_snapshot->cycles = now;
1057 	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1058 	systime_snapshot->boot = ktime_add_ns(base_boot, nsec_real);
1059 	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1060 }
1061 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1062 
1063 /* Scale base by mult/div checking for overflow */
scale64_check_overflow(u64 mult,u64 div,u64 * base)1064 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1065 {
1066 	u64 tmp, rem;
1067 
1068 	tmp = div64_u64_rem(*base, div, &rem);
1069 
1070 	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1071 	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1072 		return -EOVERFLOW;
1073 	tmp *= mult;
1074 
1075 	rem = div64_u64(rem * mult, div);
1076 	*base = tmp + rem;
1077 	return 0;
1078 }
1079 
1080 /**
1081  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1082  * @history:			Snapshot representing start of history
1083  * @partial_history_cycles:	Cycle offset into history (fractional part)
1084  * @total_history_cycles:	Total history length in cycles
1085  * @discontinuity:		True indicates clock was set on history period
1086  * @ts:				Cross timestamp that should be adjusted using
1087  *	partial/total ratio
1088  *
1089  * Helper function used by get_device_system_crosststamp() to correct the
1090  * crosstimestamp corresponding to the start of the current interval to the
1091  * system counter value (timestamp point) provided by the driver. The
1092  * total_history_* quantities are the total history starting at the provided
1093  * reference point and ending at the start of the current interval. The cycle
1094  * count between the driver timestamp point and the start of the current
1095  * interval is partial_history_cycles.
1096  */
adjust_historical_crosststamp(struct system_time_snapshot * history,u64 partial_history_cycles,u64 total_history_cycles,bool discontinuity,struct system_device_crosststamp * ts)1097 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1098 					 u64 partial_history_cycles,
1099 					 u64 total_history_cycles,
1100 					 bool discontinuity,
1101 					 struct system_device_crosststamp *ts)
1102 {
1103 	struct timekeeper *tk = &tk_core.timekeeper;
1104 	u64 corr_raw, corr_real;
1105 	bool interp_forward;
1106 	int ret;
1107 
1108 	if (total_history_cycles == 0 || partial_history_cycles == 0)
1109 		return 0;
1110 
1111 	/* Interpolate shortest distance from beginning or end of history */
1112 	interp_forward = partial_history_cycles > total_history_cycles / 2;
1113 	partial_history_cycles = interp_forward ?
1114 		total_history_cycles - partial_history_cycles :
1115 		partial_history_cycles;
1116 
1117 	/*
1118 	 * Scale the monotonic raw time delta by:
1119 	 *	partial_history_cycles / total_history_cycles
1120 	 */
1121 	corr_raw = (u64)ktime_to_ns(
1122 		ktime_sub(ts->sys_monoraw, history->raw));
1123 	ret = scale64_check_overflow(partial_history_cycles,
1124 				     total_history_cycles, &corr_raw);
1125 	if (ret)
1126 		return ret;
1127 
1128 	/*
1129 	 * If there is a discontinuity in the history, scale monotonic raw
1130 	 *	correction by:
1131 	 *	mult(real)/mult(raw) yielding the realtime correction
1132 	 * Otherwise, calculate the realtime correction similar to monotonic
1133 	 *	raw calculation
1134 	 */
1135 	if (discontinuity) {
1136 		corr_real = mul_u64_u32_div
1137 			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1138 	} else {
1139 		corr_real = (u64)ktime_to_ns(
1140 			ktime_sub(ts->sys_realtime, history->real));
1141 		ret = scale64_check_overflow(partial_history_cycles,
1142 					     total_history_cycles, &corr_real);
1143 		if (ret)
1144 			return ret;
1145 	}
1146 
1147 	/* Fixup monotonic raw and real time time values */
1148 	if (interp_forward) {
1149 		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1150 		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1151 	} else {
1152 		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1153 		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1154 	}
1155 
1156 	return 0;
1157 }
1158 
1159 /*
1160  * timestamp_in_interval - true if ts is chronologically in [start, end]
1161  *
1162  * True if ts occurs chronologically at or after start, and before or at end.
1163  */
timestamp_in_interval(u64 start,u64 end,u64 ts)1164 static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1165 {
1166 	if (ts >= start && ts <= end)
1167 		return true;
1168 	if (start > end && (ts >= start || ts <= end))
1169 		return true;
1170 	return false;
1171 }
1172 
convert_clock(u64 * val,u32 numerator,u32 denominator)1173 static bool convert_clock(u64 *val, u32 numerator, u32 denominator)
1174 {
1175 	u64 rem, res;
1176 
1177 	if (!numerator || !denominator)
1178 		return false;
1179 
1180 	res = div64_u64_rem(*val, denominator, &rem) * numerator;
1181 	*val = res + div_u64(rem * numerator, denominator);
1182 	return true;
1183 }
1184 
convert_base_to_cs(struct system_counterval_t * scv)1185 static bool convert_base_to_cs(struct system_counterval_t *scv)
1186 {
1187 	struct clocksource *cs = tk_core.timekeeper.tkr_mono.clock;
1188 	struct clocksource_base *base;
1189 	u32 num, den;
1190 
1191 	/* The timestamp was taken from the time keeper clock source */
1192 	if (cs->id == scv->cs_id)
1193 		return true;
1194 
1195 	/*
1196 	 * Check whether cs_id matches the base clock. Prevent the compiler from
1197 	 * re-evaluating @base as the clocksource might change concurrently.
1198 	 */
1199 	base = READ_ONCE(cs->base);
1200 	if (!base || base->id != scv->cs_id)
1201 		return false;
1202 
1203 	num = scv->use_nsecs ? cs->freq_khz : base->numerator;
1204 	den = scv->use_nsecs ? USEC_PER_SEC : base->denominator;
1205 
1206 	if (!convert_clock(&scv->cycles, num, den))
1207 		return false;
1208 
1209 	scv->cycles += base->offset;
1210 	return true;
1211 }
1212 
convert_cs_to_base(u64 * cycles,enum clocksource_ids base_id)1213 static bool convert_cs_to_base(u64 *cycles, enum clocksource_ids base_id)
1214 {
1215 	struct clocksource *cs = tk_core.timekeeper.tkr_mono.clock;
1216 	struct clocksource_base *base;
1217 
1218 	/*
1219 	 * Check whether base_id matches the base clock. Prevent the compiler from
1220 	 * re-evaluating @base as the clocksource might change concurrently.
1221 	 */
1222 	base = READ_ONCE(cs->base);
1223 	if (!base || base->id != base_id)
1224 		return false;
1225 
1226 	*cycles -= base->offset;
1227 	if (!convert_clock(cycles, base->denominator, base->numerator))
1228 		return false;
1229 	return true;
1230 }
1231 
convert_ns_to_cs(u64 * delta)1232 static bool convert_ns_to_cs(u64 *delta)
1233 {
1234 	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
1235 
1236 	if (BITS_TO_BYTES(fls64(*delta) + tkr->shift) >= sizeof(*delta))
1237 		return false;
1238 
1239 	*delta = div_u64((*delta << tkr->shift) - tkr->xtime_nsec, tkr->mult);
1240 	return true;
1241 }
1242 
1243 /**
1244  * ktime_real_to_base_clock() - Convert CLOCK_REALTIME timestamp to a base clock timestamp
1245  * @treal:	CLOCK_REALTIME timestamp to convert
1246  * @base_id:	base clocksource id
1247  * @cycles:	pointer to store the converted base clock timestamp
1248  *
1249  * Converts a supplied, future realtime clock value to the corresponding base clock value.
1250  *
1251  * Return:  true if the conversion is successful, false otherwise.
1252  */
ktime_real_to_base_clock(ktime_t treal,enum clocksource_ids base_id,u64 * cycles)1253 bool ktime_real_to_base_clock(ktime_t treal, enum clocksource_ids base_id, u64 *cycles)
1254 {
1255 	struct timekeeper *tk = &tk_core.timekeeper;
1256 	unsigned int seq;
1257 	u64 delta;
1258 
1259 	do {
1260 		seq = read_seqcount_begin(&tk_core.seq);
1261 		if ((u64)treal < tk->tkr_mono.base_real)
1262 			return false;
1263 		delta = (u64)treal - tk->tkr_mono.base_real;
1264 		if (!convert_ns_to_cs(&delta))
1265 			return false;
1266 		*cycles = tk->tkr_mono.cycle_last + delta;
1267 		if (!convert_cs_to_base(cycles, base_id))
1268 			return false;
1269 	} while (read_seqcount_retry(&tk_core.seq, seq));
1270 
1271 	return true;
1272 }
1273 EXPORT_SYMBOL_GPL(ktime_real_to_base_clock);
1274 
1275 /**
1276  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1277  * @get_time_fn:	Callback to get simultaneous device time and
1278  *	system counter from the device driver
1279  * @ctx:		Context passed to get_time_fn()
1280  * @history_begin:	Historical reference point used to interpolate system
1281  *	time when counter provided by the driver is before the current interval
1282  * @xtstamp:		Receives simultaneously captured system and device time
1283  *
1284  * Reads a timestamp from a device and correlates it to system time
1285  */
get_device_system_crosststamp(int (* get_time_fn)(ktime_t * device_time,struct system_counterval_t * sys_counterval,void * ctx),void * ctx,struct system_time_snapshot * history_begin,struct system_device_crosststamp * xtstamp)1286 int get_device_system_crosststamp(int (*get_time_fn)
1287 				  (ktime_t *device_time,
1288 				   struct system_counterval_t *sys_counterval,
1289 				   void *ctx),
1290 				  void *ctx,
1291 				  struct system_time_snapshot *history_begin,
1292 				  struct system_device_crosststamp *xtstamp)
1293 {
1294 	struct system_counterval_t system_counterval;
1295 	struct timekeeper *tk = &tk_core.timekeeper;
1296 	u64 cycles, now, interval_start;
1297 	unsigned int clock_was_set_seq = 0;
1298 	ktime_t base_real, base_raw;
1299 	u64 nsec_real, nsec_raw;
1300 	u8 cs_was_changed_seq;
1301 	unsigned int seq;
1302 	bool do_interp;
1303 	int ret;
1304 
1305 	do {
1306 		seq = read_seqcount_begin(&tk_core.seq);
1307 		/*
1308 		 * Try to synchronously capture device time and a system
1309 		 * counter value calling back into the device driver
1310 		 */
1311 		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1312 		if (ret)
1313 			return ret;
1314 
1315 		/*
1316 		 * Verify that the clocksource ID associated with the captured
1317 		 * system counter value is the same as for the currently
1318 		 * installed timekeeper clocksource
1319 		 */
1320 		if (system_counterval.cs_id == CSID_GENERIC ||
1321 		    !convert_base_to_cs(&system_counterval))
1322 			return -ENODEV;
1323 		cycles = system_counterval.cycles;
1324 
1325 		/*
1326 		 * Check whether the system counter value provided by the
1327 		 * device driver is on the current timekeeping interval.
1328 		 */
1329 		now = tk_clock_read(&tk->tkr_mono);
1330 		interval_start = tk->tkr_mono.cycle_last;
1331 		if (!timestamp_in_interval(interval_start, now, cycles)) {
1332 			clock_was_set_seq = tk->clock_was_set_seq;
1333 			cs_was_changed_seq = tk->cs_was_changed_seq;
1334 			cycles = interval_start;
1335 			do_interp = true;
1336 		} else {
1337 			do_interp = false;
1338 		}
1339 
1340 		base_real = ktime_add(tk->tkr_mono.base,
1341 				      tk_core.timekeeper.offs_real);
1342 		base_raw = tk->tkr_raw.base;
1343 
1344 		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1345 		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1346 	} while (read_seqcount_retry(&tk_core.seq, seq));
1347 
1348 	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1349 	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1350 
1351 	/*
1352 	 * Interpolate if necessary, adjusting back from the start of the
1353 	 * current interval
1354 	 */
1355 	if (do_interp) {
1356 		u64 partial_history_cycles, total_history_cycles;
1357 		bool discontinuity;
1358 
1359 		/*
1360 		 * Check that the counter value is not before the provided
1361 		 * history reference and that the history doesn't cross a
1362 		 * clocksource change
1363 		 */
1364 		if (!history_begin ||
1365 		    !timestamp_in_interval(history_begin->cycles,
1366 					   cycles, system_counterval.cycles) ||
1367 		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1368 			return -EINVAL;
1369 		partial_history_cycles = cycles - system_counterval.cycles;
1370 		total_history_cycles = cycles - history_begin->cycles;
1371 		discontinuity =
1372 			history_begin->clock_was_set_seq != clock_was_set_seq;
1373 
1374 		ret = adjust_historical_crosststamp(history_begin,
1375 						    partial_history_cycles,
1376 						    total_history_cycles,
1377 						    discontinuity, xtstamp);
1378 		if (ret)
1379 			return ret;
1380 	}
1381 
1382 	return 0;
1383 }
1384 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1385 
1386 /**
1387  * timekeeping_clocksource_has_base - Check whether the current clocksource
1388  *				      is based on given a base clock
1389  * @id:		base clocksource ID
1390  *
1391  * Note:	The return value is a snapshot which can become invalid right
1392  *		after the function returns.
1393  *
1394  * Return:	true if the timekeeper clocksource has a base clock with @id,
1395  *		false otherwise
1396  */
timekeeping_clocksource_has_base(enum clocksource_ids id)1397 bool timekeeping_clocksource_has_base(enum clocksource_ids id)
1398 {
1399 	/*
1400 	 * This is a snapshot, so no point in using the sequence
1401 	 * count. Just prevent the compiler from re-evaluating @base as the
1402 	 * clocksource might change concurrently.
1403 	 */
1404 	struct clocksource_base *base = READ_ONCE(tk_core.timekeeper.tkr_mono.clock->base);
1405 
1406 	return base ? base->id == id : false;
1407 }
1408 EXPORT_SYMBOL_GPL(timekeeping_clocksource_has_base);
1409 
1410 /**
1411  * do_settimeofday64 - Sets the time of day.
1412  * @ts:     pointer to the timespec64 variable containing the new time
1413  *
1414  * Sets the time of day to the new time and update NTP and notify hrtimers
1415  */
do_settimeofday64(const struct timespec64 * ts)1416 int do_settimeofday64(const struct timespec64 *ts)
1417 {
1418 	struct timespec64 ts_delta, xt;
1419 
1420 	if (!timespec64_valid_settod(ts))
1421 		return -EINVAL;
1422 
1423 	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
1424 		struct timekeeper *tks = &tk_core.shadow_timekeeper;
1425 
1426 		timekeeping_forward_now(tks);
1427 
1428 		xt = tk_xtime(tks);
1429 		ts_delta = timespec64_sub(*ts, xt);
1430 
1431 		if (timespec64_compare(&tks->wall_to_monotonic, &ts_delta) > 0) {
1432 			timekeeping_restore_shadow(&tk_core);
1433 			return -EINVAL;
1434 		}
1435 
1436 		tk_set_wall_to_mono(tks, timespec64_sub(tks->wall_to_monotonic, ts_delta));
1437 		tk_set_xtime(tks, ts);
1438 		timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1439 	}
1440 
1441 	/* Signal hrtimers about time change */
1442 	clock_was_set(CLOCK_SET_WALL);
1443 
1444 	audit_tk_injoffset(ts_delta);
1445 	add_device_randomness(ts, sizeof(*ts));
1446 	return 0;
1447 }
1448 EXPORT_SYMBOL(do_settimeofday64);
1449 
1450 /**
1451  * timekeeping_inject_offset - Adds or subtracts from the current time.
1452  * @ts:		Pointer to the timespec variable containing the offset
1453  *
1454  * Adds or subtracts an offset value from the current time.
1455  */
timekeeping_inject_offset(const struct timespec64 * ts)1456 static int timekeeping_inject_offset(const struct timespec64 *ts)
1457 {
1458 	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1459 		return -EINVAL;
1460 
1461 	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
1462 		struct timekeeper *tks = &tk_core.shadow_timekeeper;
1463 		struct timespec64 tmp;
1464 
1465 		timekeeping_forward_now(tks);
1466 
1467 		/* Make sure the proposed value is valid */
1468 		tmp = timespec64_add(tk_xtime(tks), *ts);
1469 		if (timespec64_compare(&tks->wall_to_monotonic, ts) > 0 ||
1470 		    !timespec64_valid_settod(&tmp)) {
1471 			timekeeping_restore_shadow(&tk_core);
1472 			return -EINVAL;
1473 		}
1474 
1475 		tk_xtime_add(tks, ts);
1476 		tk_set_wall_to_mono(tks, timespec64_sub(tks->wall_to_monotonic, *ts));
1477 		timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1478 	}
1479 
1480 	/* Signal hrtimers about time change */
1481 	clock_was_set(CLOCK_SET_WALL);
1482 	return 0;
1483 }
1484 
1485 /*
1486  * Indicates if there is an offset between the system clock and the hardware
1487  * clock/persistent clock/rtc.
1488  */
1489 int persistent_clock_is_local;
1490 
1491 /*
1492  * Adjust the time obtained from the CMOS to be UTC time instead of
1493  * local time.
1494  *
1495  * This is ugly, but preferable to the alternatives.  Otherwise we
1496  * would either need to write a program to do it in /etc/rc (and risk
1497  * confusion if the program gets run more than once; it would also be
1498  * hard to make the program warp the clock precisely n hours)  or
1499  * compile in the timezone information into the kernel.  Bad, bad....
1500  *
1501  *						- TYT, 1992-01-01
1502  *
1503  * The best thing to do is to keep the CMOS clock in universal time (UTC)
1504  * as real UNIX machines always do it. This avoids all headaches about
1505  * daylight saving times and warping kernel clocks.
1506  */
timekeeping_warp_clock(void)1507 void timekeeping_warp_clock(void)
1508 {
1509 	if (sys_tz.tz_minuteswest != 0) {
1510 		struct timespec64 adjust;
1511 
1512 		persistent_clock_is_local = 1;
1513 		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1514 		adjust.tv_nsec = 0;
1515 		timekeeping_inject_offset(&adjust);
1516 	}
1517 }
1518 
1519 /*
1520  * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1521  */
__timekeeping_set_tai_offset(struct timekeeper * tk,s32 tai_offset)1522 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1523 {
1524 	tk->tai_offset = tai_offset;
1525 	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1526 }
1527 
1528 /*
1529  * change_clocksource - Swaps clocksources if a new one is available
1530  *
1531  * Accumulates current time interval and initializes new clocksource
1532  */
change_clocksource(void * data)1533 static int change_clocksource(void *data)
1534 {
1535 	struct clocksource *new = data, *old = NULL;
1536 
1537 	/*
1538 	 * If the clocksource is in a module, get a module reference.
1539 	 * Succeeds for built-in code (owner == NULL) as well. Abort if the
1540 	 * reference can't be acquired.
1541 	 */
1542 	if (!try_module_get(new->owner))
1543 		return 0;
1544 
1545 	/* Abort if the device can't be enabled */
1546 	if (new->enable && new->enable(new) != 0) {
1547 		module_put(new->owner);
1548 		return 0;
1549 	}
1550 
1551 	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
1552 		struct timekeeper *tks = &tk_core.shadow_timekeeper;
1553 
1554 		timekeeping_forward_now(tks);
1555 		old = tks->tkr_mono.clock;
1556 		tk_setup_internals(tks, new);
1557 		timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1558 	}
1559 
1560 	if (old) {
1561 		if (old->disable)
1562 			old->disable(old);
1563 		module_put(old->owner);
1564 	}
1565 
1566 	return 0;
1567 }
1568 
1569 /**
1570  * timekeeping_notify - Install a new clock source
1571  * @clock:		pointer to the clock source
1572  *
1573  * This function is called from clocksource.c after a new, better clock
1574  * source has been registered. The caller holds the clocksource_mutex.
1575  */
timekeeping_notify(struct clocksource * clock)1576 int timekeeping_notify(struct clocksource *clock)
1577 {
1578 	struct timekeeper *tk = &tk_core.timekeeper;
1579 
1580 	if (tk->tkr_mono.clock == clock)
1581 		return 0;
1582 	stop_machine(change_clocksource, clock, NULL);
1583 	tick_clock_notify();
1584 	return tk->tkr_mono.clock == clock ? 0 : -1;
1585 }
1586 
1587 /**
1588  * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1589  * @ts:		pointer to the timespec64 to be set
1590  *
1591  * Returns the raw monotonic time (completely un-modified by ntp)
1592  */
ktime_get_raw_ts64(struct timespec64 * ts)1593 void ktime_get_raw_ts64(struct timespec64 *ts)
1594 {
1595 	struct timekeeper *tk = &tk_core.timekeeper;
1596 	unsigned int seq;
1597 	u64 nsecs;
1598 
1599 	do {
1600 		seq = read_seqcount_begin(&tk_core.seq);
1601 		ts->tv_sec = tk->raw_sec;
1602 		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1603 
1604 	} while (read_seqcount_retry(&tk_core.seq, seq));
1605 
1606 	ts->tv_nsec = 0;
1607 	timespec64_add_ns(ts, nsecs);
1608 }
1609 EXPORT_SYMBOL(ktime_get_raw_ts64);
1610 
1611 
1612 /**
1613  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1614  */
timekeeping_valid_for_hres(void)1615 int timekeeping_valid_for_hres(void)
1616 {
1617 	struct timekeeper *tk = &tk_core.timekeeper;
1618 	unsigned int seq;
1619 	int ret;
1620 
1621 	do {
1622 		seq = read_seqcount_begin(&tk_core.seq);
1623 
1624 		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1625 
1626 	} while (read_seqcount_retry(&tk_core.seq, seq));
1627 
1628 	return ret;
1629 }
1630 
1631 /**
1632  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1633  */
timekeeping_max_deferment(void)1634 u64 timekeeping_max_deferment(void)
1635 {
1636 	struct timekeeper *tk = &tk_core.timekeeper;
1637 	unsigned int seq;
1638 	u64 ret;
1639 
1640 	do {
1641 		seq = read_seqcount_begin(&tk_core.seq);
1642 
1643 		ret = tk->tkr_mono.clock->max_idle_ns;
1644 
1645 	} while (read_seqcount_retry(&tk_core.seq, seq));
1646 
1647 	return ret;
1648 }
1649 
1650 /**
1651  * read_persistent_clock64 -  Return time from the persistent clock.
1652  * @ts: Pointer to the storage for the readout value
1653  *
1654  * Weak dummy function for arches that do not yet support it.
1655  * Reads the time from the battery backed persistent clock.
1656  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1657  *
1658  *  XXX - Do be sure to remove it once all arches implement it.
1659  */
read_persistent_clock64(struct timespec64 * ts)1660 void __weak read_persistent_clock64(struct timespec64 *ts)
1661 {
1662 	ts->tv_sec = 0;
1663 	ts->tv_nsec = 0;
1664 }
1665 
1666 /**
1667  * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1668  *                                        from the boot.
1669  * @wall_time:	  current time as returned by persistent clock
1670  * @boot_offset:  offset that is defined as wall_time - boot_time
1671  *
1672  * Weak dummy function for arches that do not yet support it.
1673  *
1674  * The default function calculates offset based on the current value of
1675  * local_clock(). This way architectures that support sched_clock() but don't
1676  * support dedicated boot time clock will provide the best estimate of the
1677  * boot time.
1678  */
1679 void __weak __init
read_persistent_wall_and_boot_offset(struct timespec64 * wall_time,struct timespec64 * boot_offset)1680 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1681 				     struct timespec64 *boot_offset)
1682 {
1683 	read_persistent_clock64(wall_time);
1684 	*boot_offset = ns_to_timespec64(local_clock());
1685 }
1686 
tkd_basic_setup(struct tk_data * tkd)1687 static __init void tkd_basic_setup(struct tk_data *tkd)
1688 {
1689 	raw_spin_lock_init(&tkd->lock);
1690 	seqcount_raw_spinlock_init(&tkd->seq, &tkd->lock);
1691 }
1692 
1693 /*
1694  * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1695  *
1696  * The flag starts of false and is only set when a suspend reaches
1697  * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1698  * timekeeper clocksource is not stopping across suspend and has been
1699  * used to update sleep time. If the timekeeper clocksource has stopped
1700  * then the flag stays true and is used by the RTC resume code to decide
1701  * whether sleeptime must be injected and if so the flag gets false then.
1702  *
1703  * If a suspend fails before reaching timekeeping_resume() then the flag
1704  * stays false and prevents erroneous sleeptime injection.
1705  */
1706 static bool suspend_timing_needed;
1707 
1708 /* Flag for if there is a persistent clock on this platform */
1709 static bool persistent_clock_exists;
1710 
1711 /*
1712  * timekeeping_init - Initializes the clocksource and common timekeeping values
1713  */
timekeeping_init(void)1714 void __init timekeeping_init(void)
1715 {
1716 	struct timespec64 wall_time, boot_offset, wall_to_mono;
1717 	struct timekeeper *tks = &tk_core.shadow_timekeeper;
1718 	struct clocksource *clock;
1719 
1720 	tkd_basic_setup(&tk_core);
1721 
1722 	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1723 	if (timespec64_valid_settod(&wall_time) &&
1724 	    timespec64_to_ns(&wall_time) > 0) {
1725 		persistent_clock_exists = true;
1726 	} else if (timespec64_to_ns(&wall_time) != 0) {
1727 		pr_warn("Persistent clock returned invalid value");
1728 		wall_time = (struct timespec64){0};
1729 	}
1730 
1731 	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1732 		boot_offset = (struct timespec64){0};
1733 
1734 	/*
1735 	 * We want set wall_to_mono, so the following is true:
1736 	 * wall time + wall_to_mono = boot time
1737 	 */
1738 	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1739 
1740 	guard(raw_spinlock_irqsave)(&tk_core.lock);
1741 
1742 	ntp_init();
1743 
1744 	clock = clocksource_default_clock();
1745 	if (clock->enable)
1746 		clock->enable(clock);
1747 	tk_setup_internals(tks, clock);
1748 
1749 	tk_set_xtime(tks, &wall_time);
1750 	tks->raw_sec = 0;
1751 
1752 	tk_set_wall_to_mono(tks, wall_to_mono);
1753 
1754 	timekeeping_update_from_shadow(&tk_core, TK_CLOCK_WAS_SET);
1755 }
1756 
1757 /* time in seconds when suspend began for persistent clock */
1758 static struct timespec64 timekeeping_suspend_time;
1759 
1760 /**
1761  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1762  * @tk:		Pointer to the timekeeper to be updated
1763  * @delta:	Pointer to the delta value in timespec64 format
1764  *
1765  * Takes a timespec offset measuring a suspend interval and properly
1766  * adds the sleep offset to the timekeeping variables.
1767  */
__timekeeping_inject_sleeptime(struct timekeeper * tk,const struct timespec64 * delta)1768 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1769 					   const struct timespec64 *delta)
1770 {
1771 	if (!timespec64_valid_strict(delta)) {
1772 		printk_deferred(KERN_WARNING
1773 				"__timekeeping_inject_sleeptime: Invalid "
1774 				"sleep delta value!\n");
1775 		return;
1776 	}
1777 	tk_xtime_add(tk, delta);
1778 	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1779 	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1780 	tk_debug_account_sleep_time(delta);
1781 }
1782 
1783 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1784 /*
1785  * We have three kinds of time sources to use for sleep time
1786  * injection, the preference order is:
1787  * 1) non-stop clocksource
1788  * 2) persistent clock (ie: RTC accessible when irqs are off)
1789  * 3) RTC
1790  *
1791  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1792  * If system has neither 1) nor 2), 3) will be used finally.
1793  *
1794  *
1795  * If timekeeping has injected sleeptime via either 1) or 2),
1796  * 3) becomes needless, so in this case we don't need to call
1797  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1798  * means.
1799  */
timekeeping_rtc_skipresume(void)1800 bool timekeeping_rtc_skipresume(void)
1801 {
1802 	return !suspend_timing_needed;
1803 }
1804 
1805 /*
1806  * 1) can be determined whether to use or not only when doing
1807  * timekeeping_resume() which is invoked after rtc_suspend(),
1808  * so we can't skip rtc_suspend() surely if system has 1).
1809  *
1810  * But if system has 2), 2) will definitely be used, so in this
1811  * case we don't need to call rtc_suspend(), and this is what
1812  * timekeeping_rtc_skipsuspend() means.
1813  */
timekeeping_rtc_skipsuspend(void)1814 bool timekeeping_rtc_skipsuspend(void)
1815 {
1816 	return persistent_clock_exists;
1817 }
1818 
1819 /**
1820  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1821  * @delta: pointer to a timespec64 delta value
1822  *
1823  * This hook is for architectures that cannot support read_persistent_clock64
1824  * because their RTC/persistent clock is only accessible when irqs are enabled.
1825  * and also don't have an effective nonstop clocksource.
1826  *
1827  * This function should only be called by rtc_resume(), and allows
1828  * a suspend offset to be injected into the timekeeping values.
1829  */
timekeeping_inject_sleeptime64(const struct timespec64 * delta)1830 void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1831 {
1832 	scoped_guard(raw_spinlock_irqsave, &tk_core.lock) {
1833 		struct timekeeper *tks = &tk_core.shadow_timekeeper;
1834 
1835 		suspend_timing_needed = false;
1836 		timekeeping_forward_now(tks);
1837 		__timekeeping_inject_sleeptime(tks, delta);
1838 		timekeeping_update_from_shadow(&tk_core, TK_UPDATE_ALL);
1839 	}
1840 
1841 	/* Signal hrtimers about time change */
1842 	clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1843 }
1844 #endif
1845 
1846 /**
1847  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1848  */
timekeeping_resume(void)1849 void timekeeping_resume(void)
1850 {
1851 	struct timekeeper *tks = &tk_core.shadow_timekeeper;
1852 	struct clocksource *clock = tks->tkr_mono.clock;
1853 	struct timespec64 ts_new, ts_delta;
1854 	bool inject_sleeptime = false;
1855 	u64 cycle_now, nsec;
1856 	unsigned long flags;
1857 
1858 	read_persistent_clock64(&ts_new);
1859 
1860 	clockevents_resume();
1861 	clocksource_resume();
1862 
1863 	raw_spin_lock_irqsave(&tk_core.lock, flags);
1864 
1865 	/*
1866 	 * After system resumes, we need to calculate the suspended time and
1867 	 * compensate it for the OS time. There are 3 sources that could be
1868 	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1869 	 * device.
1870 	 *
1871 	 * One specific platform may have 1 or 2 or all of them, and the
1872 	 * preference will be:
1873 	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1874 	 * The less preferred source will only be tried if there is no better
1875 	 * usable source. The rtc part is handled separately in rtc core code.
1876 	 */
1877 	cycle_now = tk_clock_read(&tks->tkr_mono);
1878 	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1879 	if (nsec > 0) {
1880 		ts_delta = ns_to_timespec64(nsec);
1881 		inject_sleeptime = true;
1882 	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1883 		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1884 		inject_sleeptime = true;
1885 	}
1886 
1887 	if (inject_sleeptime) {
1888 		suspend_timing_needed = false;
1889 		__timekeeping_inject_sleeptime(tks, &ts_delta);
1890 	}
1891 
1892 	/* Re-base the last cycle value */
1893 	tks->tkr_mono.cycle_last = cycle_now;
1894 	tks->tkr_raw.cycle_last  = cycle_now;
1895 
1896 	tks->ntp_error = 0;
1897 	timekeeping_suspended = 0;
1898 	timekeeping_update_from_shadow(&tk_core, TK_CLOCK_WAS_SET);
1899 	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
1900 
1901 	touch_softlockup_watchdog();
1902 
1903 	/* Resume the clockevent device(s) and hrtimers */
1904 	tick_resume();
1905 	/* Notify timerfd as resume is equivalent to clock_was_set() */
1906 	timerfd_resume();
1907 }
1908 
timekeeping_suspend(void)1909 int timekeeping_suspend(void)
1910 {
1911 	struct timekeeper *tks = &tk_core.shadow_timekeeper;
1912 	struct timespec64 delta, delta_delta;
1913 	static struct timespec64 old_delta;
1914 	struct clocksource *curr_clock;
1915 	unsigned long flags;
1916 	u64 cycle_now;
1917 
1918 	read_persistent_clock64(&timekeeping_suspend_time);
1919 
1920 	/*
1921 	 * On some systems the persistent_clock can not be detected at
1922 	 * timekeeping_init by its return value, so if we see a valid
1923 	 * value returned, update the persistent_clock_exists flag.
1924 	 */
1925 	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1926 		persistent_clock_exists = true;
1927 
1928 	suspend_timing_needed = true;
1929 
1930 	raw_spin_lock_irqsave(&tk_core.lock, flags);
1931 	timekeeping_forward_now(tks);
1932 	timekeeping_suspended = 1;
1933 
1934 	/*
1935 	 * Since we've called forward_now, cycle_last stores the value
1936 	 * just read from the current clocksource. Save this to potentially
1937 	 * use in suspend timing.
1938 	 */
1939 	curr_clock = tks->tkr_mono.clock;
1940 	cycle_now = tks->tkr_mono.cycle_last;
1941 	clocksource_start_suspend_timing(curr_clock, cycle_now);
1942 
1943 	if (persistent_clock_exists) {
1944 		/*
1945 		 * To avoid drift caused by repeated suspend/resumes,
1946 		 * which each can add ~1 second drift error,
1947 		 * try to compensate so the difference in system time
1948 		 * and persistent_clock time stays close to constant.
1949 		 */
1950 		delta = timespec64_sub(tk_xtime(tks), timekeeping_suspend_time);
1951 		delta_delta = timespec64_sub(delta, old_delta);
1952 		if (abs(delta_delta.tv_sec) >= 2) {
1953 			/*
1954 			 * if delta_delta is too large, assume time correction
1955 			 * has occurred and set old_delta to the current delta.
1956 			 */
1957 			old_delta = delta;
1958 		} else {
1959 			/* Otherwise try to adjust old_system to compensate */
1960 			timekeeping_suspend_time =
1961 				timespec64_add(timekeeping_suspend_time, delta_delta);
1962 		}
1963 	}
1964 
1965 	timekeeping_update_from_shadow(&tk_core, 0);
1966 	halt_fast_timekeeper(tks);
1967 	raw_spin_unlock_irqrestore(&tk_core.lock, flags);
1968 
1969 	tick_suspend();
1970 	clocksource_suspend();
1971 	clockevents_suspend();
1972 
1973 	return 0;
1974 }
1975 
1976 /* sysfs resume/suspend bits for timekeeping */
1977 static struct syscore_ops timekeeping_syscore_ops = {
1978 	.resume		= timekeeping_resume,
1979 	.suspend	= timekeeping_suspend,
1980 };
1981 
timekeeping_init_ops(void)1982 static int __init timekeeping_init_ops(void)
1983 {
1984 	register_syscore_ops(&timekeeping_syscore_ops);
1985 	return 0;
1986 }
1987 device_initcall(timekeeping_init_ops);
1988 
1989 /*
1990  * Apply a multiplier adjustment to the timekeeper
1991  */
timekeeping_apply_adjustment(struct timekeeper * tk,s64 offset,s32 mult_adj)1992 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1993 							 s64 offset,
1994 							 s32 mult_adj)
1995 {
1996 	s64 interval = tk->cycle_interval;
1997 
1998 	if (mult_adj == 0) {
1999 		return;
2000 	} else if (mult_adj == -1) {
2001 		interval = -interval;
2002 		offset = -offset;
2003 	} else if (mult_adj != 1) {
2004 		interval *= mult_adj;
2005 		offset *= mult_adj;
2006 	}
2007 
2008 	/*
2009 	 * So the following can be confusing.
2010 	 *
2011 	 * To keep things simple, lets assume mult_adj == 1 for now.
2012 	 *
2013 	 * When mult_adj != 1, remember that the interval and offset values
2014 	 * have been appropriately scaled so the math is the same.
2015 	 *
2016 	 * The basic idea here is that we're increasing the multiplier
2017 	 * by one, this causes the xtime_interval to be incremented by
2018 	 * one cycle_interval. This is because:
2019 	 *	xtime_interval = cycle_interval * mult
2020 	 * So if mult is being incremented by one:
2021 	 *	xtime_interval = cycle_interval * (mult + 1)
2022 	 * Its the same as:
2023 	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
2024 	 * Which can be shortened to:
2025 	 *	xtime_interval += cycle_interval
2026 	 *
2027 	 * So offset stores the non-accumulated cycles. Thus the current
2028 	 * time (in shifted nanoseconds) is:
2029 	 *	now = (offset * adj) + xtime_nsec
2030 	 * Now, even though we're adjusting the clock frequency, we have
2031 	 * to keep time consistent. In other words, we can't jump back
2032 	 * in time, and we also want to avoid jumping forward in time.
2033 	 *
2034 	 * So given the same offset value, we need the time to be the same
2035 	 * both before and after the freq adjustment.
2036 	 *	now = (offset * adj_1) + xtime_nsec_1
2037 	 *	now = (offset * adj_2) + xtime_nsec_2
2038 	 * So:
2039 	 *	(offset * adj_1) + xtime_nsec_1 =
2040 	 *		(offset * adj_2) + xtime_nsec_2
2041 	 * And we know:
2042 	 *	adj_2 = adj_1 + 1
2043 	 * So:
2044 	 *	(offset * adj_1) + xtime_nsec_1 =
2045 	 *		(offset * (adj_1+1)) + xtime_nsec_2
2046 	 *	(offset * adj_1) + xtime_nsec_1 =
2047 	 *		(offset * adj_1) + offset + xtime_nsec_2
2048 	 * Canceling the sides:
2049 	 *	xtime_nsec_1 = offset + xtime_nsec_2
2050 	 * Which gives us:
2051 	 *	xtime_nsec_2 = xtime_nsec_1 - offset
2052 	 * Which simplifies to:
2053 	 *	xtime_nsec -= offset
2054 	 */
2055 	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
2056 		/* NTP adjustment caused clocksource mult overflow */
2057 		WARN_ON_ONCE(1);
2058 		return;
2059 	}
2060 
2061 	tk->tkr_mono.mult += mult_adj;
2062 	tk->xtime_interval += interval;
2063 	tk->tkr_mono.xtime_nsec -= offset;
2064 }
2065 
2066 /*
2067  * Adjust the timekeeper's multiplier to the correct frequency
2068  * and also to reduce the accumulated error value.
2069  */
timekeeping_adjust(struct timekeeper * tk,s64 offset)2070 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2071 {
2072 	u64 ntp_tl = ntp_tick_length();
2073 	u32 mult;
2074 
2075 	/*
2076 	 * Determine the multiplier from the current NTP tick length.
2077 	 * Avoid expensive division when the tick length doesn't change.
2078 	 */
2079 	if (likely(tk->ntp_tick == ntp_tl)) {
2080 		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2081 	} else {
2082 		tk->ntp_tick = ntp_tl;
2083 		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2084 				 tk->xtime_remainder, tk->cycle_interval);
2085 	}
2086 
2087 	/*
2088 	 * If the clock is behind the NTP time, increase the multiplier by 1
2089 	 * to catch up with it. If it's ahead and there was a remainder in the
2090 	 * tick division, the clock will slow down. Otherwise it will stay
2091 	 * ahead until the tick length changes to a non-divisible value.
2092 	 */
2093 	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2094 	mult += tk->ntp_err_mult;
2095 
2096 	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2097 
2098 	if (unlikely(tk->tkr_mono.clock->maxadj &&
2099 		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2100 			> tk->tkr_mono.clock->maxadj))) {
2101 		printk_once(KERN_WARNING
2102 			"Adjusting %s more than 11%% (%ld vs %ld)\n",
2103 			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2104 			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2105 	}
2106 
2107 	/*
2108 	 * It may be possible that when we entered this function, xtime_nsec
2109 	 * was very small.  Further, if we're slightly speeding the clocksource
2110 	 * in the code above, its possible the required corrective factor to
2111 	 * xtime_nsec could cause it to underflow.
2112 	 *
2113 	 * Now, since we have already accumulated the second and the NTP
2114 	 * subsystem has been notified via second_overflow(), we need to skip
2115 	 * the next update.
2116 	 */
2117 	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2118 		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2119 							tk->tkr_mono.shift;
2120 		tk->xtime_sec--;
2121 		tk->skip_second_overflow = 1;
2122 	}
2123 }
2124 
2125 /*
2126  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2127  *
2128  * Helper function that accumulates the nsecs greater than a second
2129  * from the xtime_nsec field to the xtime_secs field.
2130  * It also calls into the NTP code to handle leapsecond processing.
2131  */
accumulate_nsecs_to_secs(struct timekeeper * tk)2132 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2133 {
2134 	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2135 	unsigned int clock_set = 0;
2136 
2137 	while (tk->tkr_mono.xtime_nsec >= nsecps) {
2138 		int leap;
2139 
2140 		tk->tkr_mono.xtime_nsec -= nsecps;
2141 		tk->xtime_sec++;
2142 
2143 		/*
2144 		 * Skip NTP update if this second was accumulated before,
2145 		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2146 		 */
2147 		if (unlikely(tk->skip_second_overflow)) {
2148 			tk->skip_second_overflow = 0;
2149 			continue;
2150 		}
2151 
2152 		/* Figure out if its a leap sec and apply if needed */
2153 		leap = second_overflow(tk->xtime_sec);
2154 		if (unlikely(leap)) {
2155 			struct timespec64 ts;
2156 
2157 			tk->xtime_sec += leap;
2158 
2159 			ts.tv_sec = leap;
2160 			ts.tv_nsec = 0;
2161 			tk_set_wall_to_mono(tk,
2162 				timespec64_sub(tk->wall_to_monotonic, ts));
2163 
2164 			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2165 
2166 			clock_set = TK_CLOCK_WAS_SET;
2167 		}
2168 	}
2169 	return clock_set;
2170 }
2171 
2172 /*
2173  * logarithmic_accumulation - shifted accumulation of cycles
2174  *
2175  * This functions accumulates a shifted interval of cycles into
2176  * a shifted interval nanoseconds. Allows for O(log) accumulation
2177  * loop.
2178  *
2179  * Returns the unconsumed cycles.
2180  */
logarithmic_accumulation(struct timekeeper * tk,u64 offset,u32 shift,unsigned int * clock_set)2181 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2182 				    u32 shift, unsigned int *clock_set)
2183 {
2184 	u64 interval = tk->cycle_interval << shift;
2185 	u64 snsec_per_sec;
2186 
2187 	/* If the offset is smaller than a shifted interval, do nothing */
2188 	if (offset < interval)
2189 		return offset;
2190 
2191 	/* Accumulate one shifted interval */
2192 	offset -= interval;
2193 	tk->tkr_mono.cycle_last += interval;
2194 	tk->tkr_raw.cycle_last  += interval;
2195 
2196 	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2197 	*clock_set |= accumulate_nsecs_to_secs(tk);
2198 
2199 	/* Accumulate raw time */
2200 	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2201 	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2202 	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2203 		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2204 		tk->raw_sec++;
2205 	}
2206 
2207 	/* Accumulate error between NTP and clock interval */
2208 	tk->ntp_error += tk->ntp_tick << shift;
2209 	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2210 						(tk->ntp_error_shift + shift);
2211 
2212 	return offset;
2213 }
2214 
2215 /*
2216  * timekeeping_advance - Updates the timekeeper to the current time and
2217  * current NTP tick length
2218  */
timekeeping_advance(enum timekeeping_adv_mode mode)2219 static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2220 {
2221 	struct timekeeper *tk = &tk_core.shadow_timekeeper;
2222 	struct timekeeper *real_tk = &tk_core.timekeeper;
2223 	unsigned int clock_set = 0;
2224 	int shift = 0, maxshift;
2225 	u64 offset;
2226 
2227 	guard(raw_spinlock_irqsave)(&tk_core.lock);
2228 
2229 	/* Make sure we're fully resumed: */
2230 	if (unlikely(timekeeping_suspended))
2231 		return false;
2232 
2233 	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2234 				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask,
2235 				   tk->tkr_mono.clock->max_raw_delta);
2236 
2237 	/* Check if there's really nothing to do */
2238 	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2239 		return false;
2240 
2241 	/*
2242 	 * With NO_HZ we may have to accumulate many cycle_intervals
2243 	 * (think "ticks") worth of time at once. To do this efficiently,
2244 	 * we calculate the largest doubling multiple of cycle_intervals
2245 	 * that is smaller than the offset.  We then accumulate that
2246 	 * chunk in one go, and then try to consume the next smaller
2247 	 * doubled multiple.
2248 	 */
2249 	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2250 	shift = max(0, shift);
2251 	/* Bound shift to one less than what overflows tick_length */
2252 	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2253 	shift = min(shift, maxshift);
2254 	while (offset >= tk->cycle_interval) {
2255 		offset = logarithmic_accumulation(tk, offset, shift, &clock_set);
2256 		if (offset < tk->cycle_interval<<shift)
2257 			shift--;
2258 	}
2259 
2260 	/* Adjust the multiplier to correct NTP error */
2261 	timekeeping_adjust(tk, offset);
2262 
2263 	/*
2264 	 * Finally, make sure that after the rounding
2265 	 * xtime_nsec isn't larger than NSEC_PER_SEC
2266 	 */
2267 	clock_set |= accumulate_nsecs_to_secs(tk);
2268 
2269 	timekeeping_update_from_shadow(&tk_core, clock_set);
2270 
2271 	return !!clock_set;
2272 }
2273 
2274 /**
2275  * update_wall_time - Uses the current clocksource to increment the wall time
2276  *
2277  */
update_wall_time(void)2278 void update_wall_time(void)
2279 {
2280 	if (timekeeping_advance(TK_ADV_TICK))
2281 		clock_was_set_delayed();
2282 }
2283 
2284 /**
2285  * getboottime64 - Return the real time of system boot.
2286  * @ts:		pointer to the timespec64 to be set
2287  *
2288  * Returns the wall-time of boot in a timespec64.
2289  *
2290  * This is based on the wall_to_monotonic offset and the total suspend
2291  * time. Calls to settimeofday will affect the value returned (which
2292  * basically means that however wrong your real time clock is at boot time,
2293  * you get the right time here).
2294  */
getboottime64(struct timespec64 * ts)2295 void getboottime64(struct timespec64 *ts)
2296 {
2297 	struct timekeeper *tk = &tk_core.timekeeper;
2298 	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2299 
2300 	*ts = ktime_to_timespec64(t);
2301 }
2302 EXPORT_SYMBOL_GPL(getboottime64);
2303 
ktime_get_coarse_real_ts64(struct timespec64 * ts)2304 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2305 {
2306 	struct timekeeper *tk = &tk_core.timekeeper;
2307 	unsigned int seq;
2308 
2309 	do {
2310 		seq = read_seqcount_begin(&tk_core.seq);
2311 
2312 		*ts = tk_xtime(tk);
2313 	} while (read_seqcount_retry(&tk_core.seq, seq));
2314 }
2315 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2316 
2317 /**
2318  * ktime_get_coarse_real_ts64_mg - return latter of coarse grained time or floor
2319  * @ts:		timespec64 to be filled
2320  *
2321  * Fetch the global mg_floor value, convert it to realtime and compare it
2322  * to the current coarse-grained time. Fill @ts with whichever is
2323  * latest. Note that this is a filesystem-specific interface and should be
2324  * avoided outside of that context.
2325  */
ktime_get_coarse_real_ts64_mg(struct timespec64 * ts)2326 void ktime_get_coarse_real_ts64_mg(struct timespec64 *ts)
2327 {
2328 	struct timekeeper *tk = &tk_core.timekeeper;
2329 	u64 floor = atomic64_read(&mg_floor);
2330 	ktime_t f_real, offset, coarse;
2331 	unsigned int seq;
2332 
2333 	do {
2334 		seq = read_seqcount_begin(&tk_core.seq);
2335 		*ts = tk_xtime(tk);
2336 		offset = tk_core.timekeeper.offs_real;
2337 	} while (read_seqcount_retry(&tk_core.seq, seq));
2338 
2339 	coarse = timespec64_to_ktime(*ts);
2340 	f_real = ktime_add(floor, offset);
2341 	if (ktime_after(f_real, coarse))
2342 		*ts = ktime_to_timespec64(f_real);
2343 }
2344 
2345 /**
2346  * ktime_get_real_ts64_mg - attempt to update floor value and return result
2347  * @ts:		pointer to the timespec to be set
2348  *
2349  * Get a monotonic fine-grained time value and attempt to swap it into
2350  * mg_floor. If that succeeds then accept the new floor value. If it fails
2351  * then another task raced in during the interim time and updated the
2352  * floor.  Since any update to the floor must be later than the previous
2353  * floor, either outcome is acceptable.
2354  *
2355  * Typically this will be called after calling ktime_get_coarse_real_ts64_mg(),
2356  * and determining that the resulting coarse-grained timestamp did not effect
2357  * a change in ctime. Any more recent floor value would effect a change to
2358  * ctime, so there is no need to retry the atomic64_try_cmpxchg() on failure.
2359  *
2360  * @ts will be filled with the latest floor value, regardless of the outcome of
2361  * the cmpxchg. Note that this is a filesystem specific interface and should be
2362  * avoided outside of that context.
2363  */
ktime_get_real_ts64_mg(struct timespec64 * ts)2364 void ktime_get_real_ts64_mg(struct timespec64 *ts)
2365 {
2366 	struct timekeeper *tk = &tk_core.timekeeper;
2367 	ktime_t old = atomic64_read(&mg_floor);
2368 	ktime_t offset, mono;
2369 	unsigned int seq;
2370 	u64 nsecs;
2371 
2372 	do {
2373 		seq = read_seqcount_begin(&tk_core.seq);
2374 
2375 		ts->tv_sec = tk->xtime_sec;
2376 		mono = tk->tkr_mono.base;
2377 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2378 		offset = tk_core.timekeeper.offs_real;
2379 	} while (read_seqcount_retry(&tk_core.seq, seq));
2380 
2381 	mono = ktime_add_ns(mono, nsecs);
2382 
2383 	/*
2384 	 * Attempt to update the floor with the new time value. As any
2385 	 * update must be later then the existing floor, and would effect
2386 	 * a change to ctime from the perspective of the current task,
2387 	 * accept the resulting floor value regardless of the outcome of
2388 	 * the swap.
2389 	 */
2390 	if (atomic64_try_cmpxchg(&mg_floor, &old, mono)) {
2391 		ts->tv_nsec = 0;
2392 		timespec64_add_ns(ts, nsecs);
2393 		timekeeping_inc_mg_floor_swaps();
2394 	} else {
2395 		/*
2396 		 * Another task changed mg_floor since "old" was fetched.
2397 		 * "old" has been updated with the latest value of "mg_floor".
2398 		 * That value is newer than the previous floor value, which
2399 		 * is enough to effect a change to ctime. Accept it.
2400 		 */
2401 		*ts = ktime_to_timespec64(ktime_add(old, offset));
2402 	}
2403 }
2404 
ktime_get_coarse_ts64(struct timespec64 * ts)2405 void ktime_get_coarse_ts64(struct timespec64 *ts)
2406 {
2407 	struct timekeeper *tk = &tk_core.timekeeper;
2408 	struct timespec64 now, mono;
2409 	unsigned int seq;
2410 
2411 	do {
2412 		seq = read_seqcount_begin(&tk_core.seq);
2413 
2414 		now = tk_xtime(tk);
2415 		mono = tk->wall_to_monotonic;
2416 	} while (read_seqcount_retry(&tk_core.seq, seq));
2417 
2418 	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2419 				now.tv_nsec + mono.tv_nsec);
2420 }
2421 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2422 
2423 /*
2424  * Must hold jiffies_lock
2425  */
do_timer(unsigned long ticks)2426 void do_timer(unsigned long ticks)
2427 {
2428 	jiffies_64 += ticks;
2429 	calc_global_load();
2430 }
2431 
2432 /**
2433  * ktime_get_update_offsets_now - hrtimer helper
2434  * @cwsseq:	pointer to check and store the clock was set sequence number
2435  * @offs_real:	pointer to storage for monotonic -> realtime offset
2436  * @offs_boot:	pointer to storage for monotonic -> boottime offset
2437  * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2438  *
2439  * Returns current monotonic time and updates the offsets if the
2440  * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2441  * different.
2442  *
2443  * Called from hrtimer_interrupt() or retrigger_next_event()
2444  */
ktime_get_update_offsets_now(unsigned int * cwsseq,ktime_t * offs_real,ktime_t * offs_boot,ktime_t * offs_tai)2445 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2446 				     ktime_t *offs_boot, ktime_t *offs_tai)
2447 {
2448 	struct timekeeper *tk = &tk_core.timekeeper;
2449 	unsigned int seq;
2450 	ktime_t base;
2451 	u64 nsecs;
2452 
2453 	do {
2454 		seq = read_seqcount_begin(&tk_core.seq);
2455 
2456 		base = tk->tkr_mono.base;
2457 		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2458 		base = ktime_add_ns(base, nsecs);
2459 
2460 		if (*cwsseq != tk->clock_was_set_seq) {
2461 			*cwsseq = tk->clock_was_set_seq;
2462 			*offs_real = tk->offs_real;
2463 			*offs_boot = tk->offs_boot;
2464 			*offs_tai = tk->offs_tai;
2465 		}
2466 
2467 		/* Handle leapsecond insertion adjustments */
2468 		if (unlikely(base >= tk->next_leap_ktime))
2469 			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2470 
2471 	} while (read_seqcount_retry(&tk_core.seq, seq));
2472 
2473 	return base;
2474 }
2475 
2476 /*
2477  * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2478  */
timekeeping_validate_timex(const struct __kernel_timex * txc)2479 static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2480 {
2481 	if (txc->modes & ADJ_ADJTIME) {
2482 		/* singleshot must not be used with any other mode bits */
2483 		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2484 			return -EINVAL;
2485 		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2486 		    !capable(CAP_SYS_TIME))
2487 			return -EPERM;
2488 	} else {
2489 		/* In order to modify anything, you gotta be super-user! */
2490 		if (txc->modes && !capable(CAP_SYS_TIME))
2491 			return -EPERM;
2492 		/*
2493 		 * if the quartz is off by more than 10% then
2494 		 * something is VERY wrong!
2495 		 */
2496 		if (txc->modes & ADJ_TICK &&
2497 		    (txc->tick <  900000/USER_HZ ||
2498 		     txc->tick > 1100000/USER_HZ))
2499 			return -EINVAL;
2500 	}
2501 
2502 	if (txc->modes & ADJ_SETOFFSET) {
2503 		/* In order to inject time, you gotta be super-user! */
2504 		if (!capable(CAP_SYS_TIME))
2505 			return -EPERM;
2506 
2507 		/*
2508 		 * Validate if a timespec/timeval used to inject a time
2509 		 * offset is valid.  Offsets can be positive or negative, so
2510 		 * we don't check tv_sec. The value of the timeval/timespec
2511 		 * is the sum of its fields,but *NOTE*:
2512 		 * The field tv_usec/tv_nsec must always be non-negative and
2513 		 * we can't have more nanoseconds/microseconds than a second.
2514 		 */
2515 		if (txc->time.tv_usec < 0)
2516 			return -EINVAL;
2517 
2518 		if (txc->modes & ADJ_NANO) {
2519 			if (txc->time.tv_usec >= NSEC_PER_SEC)
2520 				return -EINVAL;
2521 		} else {
2522 			if (txc->time.tv_usec >= USEC_PER_SEC)
2523 				return -EINVAL;
2524 		}
2525 	}
2526 
2527 	/*
2528 	 * Check for potential multiplication overflows that can
2529 	 * only happen on 64-bit systems:
2530 	 */
2531 	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2532 		if (LLONG_MIN / PPM_SCALE > txc->freq)
2533 			return -EINVAL;
2534 		if (LLONG_MAX / PPM_SCALE < txc->freq)
2535 			return -EINVAL;
2536 	}
2537 
2538 	return 0;
2539 }
2540 
2541 /**
2542  * random_get_entropy_fallback - Returns the raw clock source value,
2543  * used by random.c for platforms with no valid random_get_entropy().
2544  */
random_get_entropy_fallback(void)2545 unsigned long random_get_entropy_fallback(void)
2546 {
2547 	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2548 	struct clocksource *clock = READ_ONCE(tkr->clock);
2549 
2550 	if (unlikely(timekeeping_suspended || !clock))
2551 		return 0;
2552 	return clock->read(clock);
2553 }
2554 EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2555 
2556 /**
2557  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2558  * @txc:	Pointer to kernel_timex structure containing NTP parameters
2559  */
do_adjtimex(struct __kernel_timex * txc)2560 int do_adjtimex(struct __kernel_timex *txc)
2561 {
2562 	struct audit_ntp_data ad;
2563 	bool offset_set = false;
2564 	bool clock_set = false;
2565 	struct timespec64 ts;
2566 	int ret;
2567 
2568 	/* Validate the data before disabling interrupts */
2569 	ret = timekeeping_validate_timex(txc);
2570 	if (ret)
2571 		return ret;
2572 	add_device_randomness(txc, sizeof(*txc));
2573 
2574 	if (txc->modes & ADJ_SETOFFSET) {
2575 		struct timespec64 delta;
2576 
2577 		delta.tv_sec  = txc->time.tv_sec;
2578 		delta.tv_nsec = txc->time.tv_usec;
2579 		if (!(txc->modes & ADJ_NANO))
2580 			delta.tv_nsec *= 1000;
2581 		ret = timekeeping_inject_offset(&delta);
2582 		if (ret)
2583 			return ret;
2584 
2585 		offset_set = delta.tv_sec != 0;
2586 		audit_tk_injoffset(delta);
2587 	}
2588 
2589 	audit_ntp_init(&ad);
2590 
2591 	ktime_get_real_ts64(&ts);
2592 	add_device_randomness(&ts, sizeof(ts));
2593 
2594 	scoped_guard (raw_spinlock_irqsave, &tk_core.lock) {
2595 		struct timekeeper *tks = &tk_core.shadow_timekeeper;
2596 		s32 orig_tai, tai;
2597 
2598 		orig_tai = tai = tks->tai_offset;
2599 		ret = __do_adjtimex(txc, &ts, &tai, &ad);
2600 
2601 		if (tai != orig_tai) {
2602 			__timekeeping_set_tai_offset(tks, tai);
2603 			timekeeping_update_from_shadow(&tk_core, TK_CLOCK_WAS_SET);
2604 			clock_set = true;
2605 		} else {
2606 			tk_update_leap_state_all(&tk_core);
2607 		}
2608 	}
2609 
2610 	audit_ntp_log(&ad);
2611 
2612 	/* Update the multiplier immediately if frequency was set directly */
2613 	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2614 		clock_set |= timekeeping_advance(TK_ADV_FREQ);
2615 
2616 	if (clock_set)
2617 		clock_was_set(CLOCK_SET_WALL);
2618 
2619 	ntp_notify_cmos_timer(offset_set);
2620 
2621 	return ret;
2622 }
2623 
2624 #ifdef CONFIG_NTP_PPS
2625 /**
2626  * hardpps() - Accessor function to NTP __hardpps function
2627  * @phase_ts:	Pointer to timespec64 structure representing phase timestamp
2628  * @raw_ts:	Pointer to timespec64 structure representing raw timestamp
2629  */
hardpps(const struct timespec64 * phase_ts,const struct timespec64 * raw_ts)2630 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2631 {
2632 	guard(raw_spinlock_irqsave)(&tk_core.lock);
2633 	__hardpps(phase_ts, raw_ts);
2634 }
2635 EXPORT_SYMBOL(hardpps);
2636 #endif /* CONFIG_NTP_PPS */
2637