1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Kernel timekeeping code and accessor functions. Based on code from
   4 *  timer.c, moved in commit 8524070b7982.
   5 */
   6#include <linux/timekeeper_internal.h>
   7#include <linux/module.h>
   8#include <linux/interrupt.h>
   9#include <linux/percpu.h>
  10#include <linux/init.h>
  11#include <linux/mm.h>
  12#include <linux/nmi.h>
  13#include <linux/sched.h>
  14#include <linux/sched/loadavg.h>
  15#include <linux/sched/clock.h>
  16#include <linux/syscore_ops.h>
  17#include <linux/clocksource.h>
  18#include <linux/jiffies.h>
  19#include <linux/time.h>
 
  20#include <linux/tick.h>
  21#include <linux/stop_machine.h>
  22#include <linux/pvclock_gtod.h>
  23#include <linux/compiler.h>
  24#include <linux/audit.h>
 
  25
  26#include "tick-internal.h"
  27#include "ntp_internal.h"
  28#include "timekeeping_internal.h"
  29
  30#define TK_CLEAR_NTP		(1 << 0)
  31#define TK_MIRROR		(1 << 1)
  32#define TK_CLOCK_WAS_SET	(1 << 2)
  33
  34enum timekeeping_adv_mode {
  35	/* Update timekeeper when a tick has passed */
  36	TK_ADV_TICK,
  37
  38	/* Update timekeeper on a direct frequency change */
  39	TK_ADV_FREQ
  40};
  41
 
 
  42/*
  43 * The most important data for readout fits into a single 64 byte
  44 * cache line.
  45 */
  46static struct {
  47	seqcount_t		seq;
  48	struct timekeeper	timekeeper;
  49} tk_core ____cacheline_aligned = {
  50	.seq = SEQCNT_ZERO(tk_core.seq),
  51};
  52
  53static DEFINE_RAW_SPINLOCK(timekeeper_lock);
  54static struct timekeeper shadow_timekeeper;
  55
 
 
 
  56/**
  57 * struct tk_fast - NMI safe timekeeper
  58 * @seq:	Sequence counter for protecting updates. The lowest bit
  59 *		is the index for the tk_read_base array
  60 * @base:	tk_read_base array. Access is indexed by the lowest bit of
  61 *		@seq.
  62 *
  63 * See @update_fast_timekeeper() below.
  64 */
  65struct tk_fast {
  66	seqcount_t		seq;
  67	struct tk_read_base	base[2];
  68};
  69
  70/* Suspend-time cycles value for halted fast timekeeper. */
  71static u64 cycles_at_suspend;
  72
  73static u64 dummy_clock_read(struct clocksource *cs)
  74{
  75	return cycles_at_suspend;
 
 
  76}
  77
  78static struct clocksource dummy_clock = {
  79	.read = dummy_clock_read,
  80};
  81
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  82static struct tk_fast tk_fast_mono ____cacheline_aligned = {
  83	.base[0] = { .clock = &dummy_clock, },
  84	.base[1] = { .clock = &dummy_clock, },
 
  85};
  86
  87static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
  88	.base[0] = { .clock = &dummy_clock, },
  89	.base[1] = { .clock = &dummy_clock, },
 
  90};
  91
  92/* flag for if timekeeping is suspended */
  93int __read_mostly timekeeping_suspended;
  94
  95static inline void tk_normalize_xtime(struct timekeeper *tk)
  96{
  97	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
  98		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
  99		tk->xtime_sec++;
 100	}
 101	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
 102		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
 103		tk->raw_sec++;
 104	}
 105}
 106
 107static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
 108{
 109	struct timespec64 ts;
 110
 111	ts.tv_sec = tk->xtime_sec;
 112	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 113	return ts;
 114}
 115
 116static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
 117{
 118	tk->xtime_sec = ts->tv_sec;
 119	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
 120}
 121
 122static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
 123{
 124	tk->xtime_sec += ts->tv_sec;
 125	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
 126	tk_normalize_xtime(tk);
 127}
 128
 129static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
 130{
 131	struct timespec64 tmp;
 132
 133	/*
 134	 * Verify consistency of: offset_real = -wall_to_monotonic
 135	 * before modifying anything
 136	 */
 137	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
 138					-tk->wall_to_monotonic.tv_nsec);
 139	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
 140	tk->wall_to_monotonic = wtm;
 141	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
 142	tk->offs_real = timespec64_to_ktime(tmp);
 143	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
 144}
 145
 146static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
 147{
 148	tk->offs_boot = ktime_add(tk->offs_boot, delta);
 149	/*
 150	 * Timespec representation for VDSO update to avoid 64bit division
 151	 * on every update.
 152	 */
 153	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
 154}
 155
 156/*
 157 * tk_clock_read - atomic clocksource read() helper
 158 *
 159 * This helper is necessary to use in the read paths because, while the
 160 * seqlock ensures we don't return a bad value while structures are updated,
 161 * it doesn't protect from potential crashes. There is the possibility that
 162 * the tkr's clocksource may change between the read reference, and the
 163 * clock reference passed to the read function.  This can cause crashes if
 164 * the wrong clocksource is passed to the wrong read function.
 165 * This isn't necessary to use when holding the timekeeper_lock or doing
 166 * a read of the fast-timekeeper tkrs (which is protected by its own locking
 167 * and update logic).
 168 */
 169static inline u64 tk_clock_read(const struct tk_read_base *tkr)
 170{
 171	struct clocksource *clock = READ_ONCE(tkr->clock);
 172
 173	return clock->read(clock);
 174}
 175
 176#ifdef CONFIG_DEBUG_TIMEKEEPING
 177#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
 178
 179static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 180{
 181
 182	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
 183	const char *name = tk->tkr_mono.clock->name;
 184
 185	if (offset > max_cycles) {
 186		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
 187				offset, name, max_cycles);
 188		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
 189	} else {
 190		if (offset > (max_cycles >> 1)) {
 191			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
 192					offset, name, max_cycles >> 1);
 193			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
 194		}
 195	}
 196
 197	if (tk->underflow_seen) {
 198		if (jiffies - tk->last_warning > WARNING_FREQ) {
 199			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
 200			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 201			printk_deferred("         Your kernel is probably still fine.\n");
 202			tk->last_warning = jiffies;
 203		}
 204		tk->underflow_seen = 0;
 205	}
 206
 207	if (tk->overflow_seen) {
 208		if (jiffies - tk->last_warning > WARNING_FREQ) {
 209			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
 210			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 211			printk_deferred("         Your kernel is probably still fine.\n");
 212			tk->last_warning = jiffies;
 213		}
 214		tk->overflow_seen = 0;
 215	}
 216}
 217
 218static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 219{
 220	struct timekeeper *tk = &tk_core.timekeeper;
 221	u64 now, last, mask, max, delta;
 222	unsigned int seq;
 223
 224	/*
 225	 * Since we're called holding a seqlock, the data may shift
 226	 * under us while we're doing the calculation. This can cause
 227	 * false positives, since we'd note a problem but throw the
 228	 * results away. So nest another seqlock here to atomically
 229	 * grab the points we are checking with.
 230	 */
 231	do {
 232		seq = read_seqcount_begin(&tk_core.seq);
 233		now = tk_clock_read(tkr);
 234		last = tkr->cycle_last;
 235		mask = tkr->mask;
 236		max = tkr->clock->max_cycles;
 237	} while (read_seqcount_retry(&tk_core.seq, seq));
 238
 239	delta = clocksource_delta(now, last, mask);
 240
 241	/*
 242	 * Try to catch underflows by checking if we are seeing small
 243	 * mask-relative negative values.
 244	 */
 245	if (unlikely((~delta & mask) < (mask >> 3))) {
 246		tk->underflow_seen = 1;
 247		delta = 0;
 248	}
 249
 250	/* Cap delta value to the max_cycles values to avoid mult overflows */
 251	if (unlikely(delta > max)) {
 252		tk->overflow_seen = 1;
 253		delta = tkr->clock->max_cycles;
 254	}
 255
 256	return delta;
 257}
 258#else
 259static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 260{
 261}
 262static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 263{
 264	u64 cycle_now, delta;
 265
 266	/* read clocksource */
 267	cycle_now = tk_clock_read(tkr);
 268
 269	/* calculate the delta since the last update_wall_time */
 270	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 271
 272	return delta;
 273}
 274#endif
 275
 276/**
 277 * tk_setup_internals - Set up internals to use clocksource clock.
 278 *
 279 * @tk:		The target timekeeper to setup.
 280 * @clock:		Pointer to clocksource.
 281 *
 282 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
 283 * pair and interval request.
 284 *
 285 * Unless you're the timekeeping code, you should not be using this!
 286 */
 287static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
 288{
 289	u64 interval;
 290	u64 tmp, ntpinterval;
 291	struct clocksource *old_clock;
 292
 293	++tk->cs_was_changed_seq;
 294	old_clock = tk->tkr_mono.clock;
 295	tk->tkr_mono.clock = clock;
 296	tk->tkr_mono.mask = clock->mask;
 297	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
 298
 299	tk->tkr_raw.clock = clock;
 300	tk->tkr_raw.mask = clock->mask;
 301	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
 302
 303	/* Do the ns -> cycle conversion first, using original mult */
 304	tmp = NTP_INTERVAL_LENGTH;
 305	tmp <<= clock->shift;
 306	ntpinterval = tmp;
 307	tmp += clock->mult/2;
 308	do_div(tmp, clock->mult);
 309	if (tmp == 0)
 310		tmp = 1;
 311
 312	interval = (u64) tmp;
 313	tk->cycle_interval = interval;
 314
 315	/* Go back from cycles -> shifted ns */
 316	tk->xtime_interval = interval * clock->mult;
 317	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
 318	tk->raw_interval = interval * clock->mult;
 319
 320	 /* if changing clocks, convert xtime_nsec shift units */
 321	if (old_clock) {
 322		int shift_change = clock->shift - old_clock->shift;
 323		if (shift_change < 0) {
 324			tk->tkr_mono.xtime_nsec >>= -shift_change;
 325			tk->tkr_raw.xtime_nsec >>= -shift_change;
 326		} else {
 327			tk->tkr_mono.xtime_nsec <<= shift_change;
 328			tk->tkr_raw.xtime_nsec <<= shift_change;
 329		}
 330	}
 331
 332	tk->tkr_mono.shift = clock->shift;
 333	tk->tkr_raw.shift = clock->shift;
 334
 335	tk->ntp_error = 0;
 336	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 337	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
 338
 339	/*
 340	 * The timekeeper keeps its own mult values for the currently
 341	 * active clocksource. These value will be adjusted via NTP
 342	 * to counteract clock drifting.
 343	 */
 344	tk->tkr_mono.mult = clock->mult;
 345	tk->tkr_raw.mult = clock->mult;
 346	tk->ntp_err_mult = 0;
 347	tk->skip_second_overflow = 0;
 348}
 349
 350/* Timekeeper helper functions. */
 351
 352#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 353static u32 default_arch_gettimeoffset(void) { return 0; }
 354u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
 355#else
 356static inline u32 arch_gettimeoffset(void) { return 0; }
 357#endif
 358
 359static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
 360{
 361	u64 nsec;
 362
 363	nsec = delta * tkr->mult + tkr->xtime_nsec;
 364	nsec >>= tkr->shift;
 365
 366	/* If arch requires, add in get_arch_timeoffset() */
 367	return nsec + arch_gettimeoffset();
 368}
 369
 370static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
 371{
 372	u64 delta;
 373
 374	delta = timekeeping_get_delta(tkr);
 375	return timekeeping_delta_to_ns(tkr, delta);
 376}
 377
 378static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
 379{
 380	u64 delta;
 381
 382	/* calculate the delta since the last update_wall_time */
 383	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 384	return timekeeping_delta_to_ns(tkr, delta);
 385}
 386
 387/**
 388 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
 389 * @tkr: Timekeeping readout base from which we take the update
 
 390 *
 391 * We want to use this from any context including NMI and tracing /
 392 * instrumenting the timekeeping code itself.
 393 *
 394 * Employ the latch technique; see @raw_write_seqcount_latch.
 395 *
 396 * So if a NMI hits the update of base[0] then it will use base[1]
 397 * which is still consistent. In the worst case this can result is a
 398 * slightly wrong timestamp (a few nanoseconds). See
 399 * @ktime_get_mono_fast_ns.
 400 */
 401static void update_fast_timekeeper(const struct tk_read_base *tkr,
 402				   struct tk_fast *tkf)
 403{
 404	struct tk_read_base *base = tkf->base;
 405
 406	/* Force readers off to base[1] */
 407	raw_write_seqcount_latch(&tkf->seq);
 408
 409	/* Update base[0] */
 410	memcpy(base, tkr, sizeof(*base));
 411
 412	/* Force readers back to base[0] */
 413	raw_write_seqcount_latch(&tkf->seq);
 414
 415	/* Update base[1] */
 416	memcpy(base + 1, base, sizeof(*base));
 417}
 418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 419/**
 420 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
 421 *
 422 * This timestamp is not guaranteed to be monotonic across an update.
 423 * The timestamp is calculated by:
 424 *
 425 *	now = base_mono + clock_delta * slope
 426 *
 427 * So if the update lowers the slope, readers who are forced to the
 428 * not yet updated second array are still using the old steeper slope.
 429 *
 430 * tmono
 431 * ^
 432 * |    o  n
 433 * |   o n
 434 * |  u
 435 * | o
 436 * |o
 437 * |12345678---> reader order
 438 *
 439 * o = old slope
 440 * u = update
 441 * n = new slope
 442 *
 443 * So reader 6 will observe time going backwards versus reader 5.
 444 *
 445 * While other CPUs are likely to be able observe that, the only way
 446 * for a CPU local observation is when an NMI hits in the middle of
 447 * the update. Timestamps taken from that NMI context might be ahead
 448 * of the following timestamps. Callers need to be aware of that and
 449 * deal with it.
 450 */
 451static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
 452{
 453	struct tk_read_base *tkr;
 454	unsigned int seq;
 455	u64 now;
 456
 457	do {
 458		seq = raw_read_seqcount_latch(&tkf->seq);
 459		tkr = tkf->base + (seq & 0x01);
 460		now = ktime_to_ns(tkr->base);
 461
 462		now += timekeeping_delta_to_ns(tkr,
 463				clocksource_delta(
 464					tk_clock_read(tkr),
 465					tkr->cycle_last,
 466					tkr->mask));
 467	} while (read_seqcount_retry(&tkf->seq, seq));
 468
 469	return now;
 470}
 471
 472u64 ktime_get_mono_fast_ns(void)
 473{
 474	return __ktime_get_fast_ns(&tk_fast_mono);
 475}
 476EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 477
 478u64 ktime_get_raw_fast_ns(void)
 
 
 
 
 
 
 479{
 480	return __ktime_get_fast_ns(&tk_fast_raw);
 481}
 482EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
 483
 484/**
 485 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
 486 *
 487 * To keep it NMI safe since we're accessing from tracing, we're not using a
 488 * separate timekeeper with updates to monotonic clock and boot offset
 489 * protected with seqlocks. This has the following minor side effects:
 490 *
 491 * (1) Its possible that a timestamp be taken after the boot offset is updated
 492 * but before the timekeeper is updated. If this happens, the new boot offset
 493 * is added to the old timekeeping making the clock appear to update slightly
 494 * earlier:
 495 *    CPU 0                                        CPU 1
 496 *    timekeeping_inject_sleeptime64()
 497 *    __timekeeping_inject_sleeptime(tk, delta);
 498 *                                                 timestamp();
 499 *    timekeeping_update(tk, TK_CLEAR_NTP...);
 500 *
 501 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
 502 * partially updated.  Since the tk->offs_boot update is a rare event, this
 503 * should be a rare occurrence which postprocessing should be able to handle.
 
 
 
 504 */
 505u64 notrace ktime_get_boot_fast_ns(void)
 506{
 507	struct timekeeper *tk = &tk_core.timekeeper;
 508
 509	return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
 510}
 511EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
 512
 513
 514/*
 515 * See comment for __ktime_get_fast_ns() vs. timestamp ordering
 
 
 
 
 
 516 */
 517static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf)
 
 
 
 
 
 
 
 
 518{
 519	struct tk_read_base *tkr;
 
 520	unsigned int seq;
 521	u64 now;
 522
 523	do {
 524		seq = raw_read_seqcount_latch(&tkf->seq);
 525		tkr = tkf->base + (seq & 0x01);
 526		now = ktime_to_ns(tkr->base_real);
 527
 528		now += timekeeping_delta_to_ns(tkr,
 529				clocksource_delta(
 530					tk_clock_read(tkr),
 531					tkr->cycle_last,
 532					tkr->mask));
 533	} while (read_seqcount_retry(&tkf->seq, seq));
 534
 535	return now;
 536}
 537
 538/**
 539 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
 
 
 540 */
 541u64 ktime_get_real_fast_ns(void)
 542{
 543	return __ktime_get_real_fast_ns(&tk_fast_mono);
 544}
 545EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
 546
 547/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 548 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
 549 * @tk: Timekeeper to snapshot.
 550 *
 551 * It generally is unsafe to access the clocksource after timekeeping has been
 552 * suspended, so take a snapshot of the readout base of @tk and use it as the
 553 * fast timekeeper's readout base while suspended.  It will return the same
 554 * number of cycles every time until timekeeping is resumed at which time the
 555 * proper readout base for the fast timekeeper will be restored automatically.
 556 */
 557static void halt_fast_timekeeper(const struct timekeeper *tk)
 558{
 559	static struct tk_read_base tkr_dummy;
 560	const struct tk_read_base *tkr = &tk->tkr_mono;
 561
 562	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 563	cycles_at_suspend = tk_clock_read(tkr);
 564	tkr_dummy.clock = &dummy_clock;
 565	tkr_dummy.base_real = tkr->base + tk->offs_real;
 566	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
 567
 568	tkr = &tk->tkr_raw;
 569	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 570	tkr_dummy.clock = &dummy_clock;
 571	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
 572}
 573
 574static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 575
 576static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 577{
 578	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 579}
 580
 581/**
 582 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 
 583 */
 584int pvclock_gtod_register_notifier(struct notifier_block *nb)
 585{
 586	struct timekeeper *tk = &tk_core.timekeeper;
 587	unsigned long flags;
 588	int ret;
 589
 590	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 591	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 592	update_pvclock_gtod(tk, true);
 593	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 594
 595	return ret;
 596}
 597EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 598
 599/**
 600 * pvclock_gtod_unregister_notifier - unregister a pvclock
 601 * timedata update listener
 
 602 */
 603int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 604{
 605	unsigned long flags;
 606	int ret;
 607
 608	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 609	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 610	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 611
 612	return ret;
 613}
 614EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 615
 616/*
 617 * tk_update_leap_state - helper to update the next_leap_ktime
 618 */
 619static inline void tk_update_leap_state(struct timekeeper *tk)
 620{
 621	tk->next_leap_ktime = ntp_get_next_leap();
 622	if (tk->next_leap_ktime != KTIME_MAX)
 623		/* Convert to monotonic time */
 624		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 625}
 626
 627/*
 628 * Update the ktime_t based scalar nsec members of the timekeeper
 629 */
 630static inline void tk_update_ktime_data(struct timekeeper *tk)
 631{
 632	u64 seconds;
 633	u32 nsec;
 634
 635	/*
 636	 * The xtime based monotonic readout is:
 637	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 638	 * The ktime based monotonic readout is:
 639	 *	nsec = base_mono + now();
 640	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 641	 */
 642	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 643	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 644	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 645
 646	/*
 647	 * The sum of the nanoseconds portions of xtime and
 648	 * wall_to_monotonic can be greater/equal one second. Take
 649	 * this into account before updating tk->ktime_sec.
 650	 */
 651	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 652	if (nsec >= NSEC_PER_SEC)
 653		seconds++;
 654	tk->ktime_sec = seconds;
 655
 656	/* Update the monotonic raw base */
 657	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
 658}
 659
 660/* must hold timekeeper_lock */
 661static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 662{
 663	if (action & TK_CLEAR_NTP) {
 664		tk->ntp_error = 0;
 665		ntp_clear();
 666	}
 667
 668	tk_update_leap_state(tk);
 669	tk_update_ktime_data(tk);
 670
 671	update_vsyscall(tk);
 672	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 673
 674	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
 675	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 676	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 677
 678	if (action & TK_CLOCK_WAS_SET)
 679		tk->clock_was_set_seq++;
 680	/*
 681	 * The mirroring of the data to the shadow-timekeeper needs
 682	 * to happen last here to ensure we don't over-write the
 683	 * timekeeper structure on the next update with stale data
 684	 */
 685	if (action & TK_MIRROR)
 686		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 687		       sizeof(tk_core.timekeeper));
 688}
 689
 690/**
 691 * timekeeping_forward_now - update clock to the current time
 
 692 *
 693 * Forward the current clock to update its state since the last call to
 694 * update_wall_time(). This is useful before significant clock changes,
 695 * as it avoids having to deal with this time offset explicitly.
 696 */
 697static void timekeeping_forward_now(struct timekeeper *tk)
 698{
 699	u64 cycle_now, delta;
 700
 701	cycle_now = tk_clock_read(&tk->tkr_mono);
 702	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 703	tk->tkr_mono.cycle_last = cycle_now;
 704	tk->tkr_raw.cycle_last  = cycle_now;
 705
 706	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 707
 708	/* If arch requires, add in get_arch_timeoffset() */
 709	tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
 710
 711
 712	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
 713
 714	/* If arch requires, add in get_arch_timeoffset() */
 715	tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
 716
 717	tk_normalize_xtime(tk);
 718}
 719
 720/**
 721 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
 722 * @ts:		pointer to the timespec to be set
 723 *
 724 * Returns the time of day in a timespec64 (WARN if suspended).
 725 */
 726void ktime_get_real_ts64(struct timespec64 *ts)
 727{
 728	struct timekeeper *tk = &tk_core.timekeeper;
 729	unsigned int seq;
 730	u64 nsecs;
 731
 732	WARN_ON(timekeeping_suspended);
 733
 734	do {
 735		seq = read_seqcount_begin(&tk_core.seq);
 736
 737		ts->tv_sec = tk->xtime_sec;
 738		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 739
 740	} while (read_seqcount_retry(&tk_core.seq, seq));
 741
 742	ts->tv_nsec = 0;
 743	timespec64_add_ns(ts, nsecs);
 744}
 745EXPORT_SYMBOL(ktime_get_real_ts64);
 746
 747ktime_t ktime_get(void)
 748{
 749	struct timekeeper *tk = &tk_core.timekeeper;
 750	unsigned int seq;
 751	ktime_t base;
 752	u64 nsecs;
 753
 754	WARN_ON(timekeeping_suspended);
 755
 756	do {
 757		seq = read_seqcount_begin(&tk_core.seq);
 758		base = tk->tkr_mono.base;
 759		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 760
 761	} while (read_seqcount_retry(&tk_core.seq, seq));
 762
 763	return ktime_add_ns(base, nsecs);
 764}
 765EXPORT_SYMBOL_GPL(ktime_get);
 766
 767u32 ktime_get_resolution_ns(void)
 768{
 769	struct timekeeper *tk = &tk_core.timekeeper;
 770	unsigned int seq;
 771	u32 nsecs;
 772
 773	WARN_ON(timekeeping_suspended);
 774
 775	do {
 776		seq = read_seqcount_begin(&tk_core.seq);
 777		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 778	} while (read_seqcount_retry(&tk_core.seq, seq));
 779
 780	return nsecs;
 781}
 782EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 783
 784static ktime_t *offsets[TK_OFFS_MAX] = {
 785	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
 786	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
 787	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
 788};
 789
 790ktime_t ktime_get_with_offset(enum tk_offsets offs)
 791{
 792	struct timekeeper *tk = &tk_core.timekeeper;
 793	unsigned int seq;
 794	ktime_t base, *offset = offsets[offs];
 795	u64 nsecs;
 796
 797	WARN_ON(timekeeping_suspended);
 798
 799	do {
 800		seq = read_seqcount_begin(&tk_core.seq);
 801		base = ktime_add(tk->tkr_mono.base, *offset);
 802		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 803
 804	} while (read_seqcount_retry(&tk_core.seq, seq));
 805
 806	return ktime_add_ns(base, nsecs);
 807
 808}
 809EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 810
 811ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
 812{
 813	struct timekeeper *tk = &tk_core.timekeeper;
 814	unsigned int seq;
 815	ktime_t base, *offset = offsets[offs];
 816	u64 nsecs;
 817
 818	WARN_ON(timekeeping_suspended);
 819
 820	do {
 821		seq = read_seqcount_begin(&tk_core.seq);
 822		base = ktime_add(tk->tkr_mono.base, *offset);
 823		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 824
 825	} while (read_seqcount_retry(&tk_core.seq, seq));
 826
 827	return ktime_add_ns(base, nsecs);
 828}
 829EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
 830
 831/**
 832 * ktime_mono_to_any() - convert mononotic time to any other time
 833 * @tmono:	time to convert.
 834 * @offs:	which offset to use
 835 */
 836ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 837{
 838	ktime_t *offset = offsets[offs];
 839	unsigned int seq;
 840	ktime_t tconv;
 841
 842	do {
 843		seq = read_seqcount_begin(&tk_core.seq);
 844		tconv = ktime_add(tmono, *offset);
 845	} while (read_seqcount_retry(&tk_core.seq, seq));
 846
 847	return tconv;
 848}
 849EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 850
 851/**
 852 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 853 */
 854ktime_t ktime_get_raw(void)
 855{
 856	struct timekeeper *tk = &tk_core.timekeeper;
 857	unsigned int seq;
 858	ktime_t base;
 859	u64 nsecs;
 860
 861	do {
 862		seq = read_seqcount_begin(&tk_core.seq);
 863		base = tk->tkr_raw.base;
 864		nsecs = timekeeping_get_ns(&tk->tkr_raw);
 865
 866	} while (read_seqcount_retry(&tk_core.seq, seq));
 867
 868	return ktime_add_ns(base, nsecs);
 869}
 870EXPORT_SYMBOL_GPL(ktime_get_raw);
 871
 872/**
 873 * ktime_get_ts64 - get the monotonic clock in timespec64 format
 874 * @ts:		pointer to timespec variable
 875 *
 876 * The function calculates the monotonic clock from the realtime
 877 * clock and the wall_to_monotonic offset and stores the result
 878 * in normalized timespec64 format in the variable pointed to by @ts.
 879 */
 880void ktime_get_ts64(struct timespec64 *ts)
 881{
 882	struct timekeeper *tk = &tk_core.timekeeper;
 883	struct timespec64 tomono;
 884	unsigned int seq;
 885	u64 nsec;
 886
 887	WARN_ON(timekeeping_suspended);
 888
 889	do {
 890		seq = read_seqcount_begin(&tk_core.seq);
 891		ts->tv_sec = tk->xtime_sec;
 892		nsec = timekeeping_get_ns(&tk->tkr_mono);
 893		tomono = tk->wall_to_monotonic;
 894
 895	} while (read_seqcount_retry(&tk_core.seq, seq));
 896
 897	ts->tv_sec += tomono.tv_sec;
 898	ts->tv_nsec = 0;
 899	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 900}
 901EXPORT_SYMBOL_GPL(ktime_get_ts64);
 902
 903/**
 904 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 905 *
 906 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 907 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 908 * works on both 32 and 64 bit systems. On 32 bit systems the readout
 909 * covers ~136 years of uptime which should be enough to prevent
 910 * premature wrap arounds.
 911 */
 912time64_t ktime_get_seconds(void)
 913{
 914	struct timekeeper *tk = &tk_core.timekeeper;
 915
 916	WARN_ON(timekeeping_suspended);
 917	return tk->ktime_sec;
 918}
 919EXPORT_SYMBOL_GPL(ktime_get_seconds);
 920
 921/**
 922 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
 923 *
 924 * Returns the wall clock seconds since 1970. This replaces the
 925 * get_seconds() interface which is not y2038 safe on 32bit systems.
 926 *
 927 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
 928 * 32bit systems the access must be protected with the sequence
 929 * counter to provide "atomic" access to the 64bit tk->xtime_sec
 930 * value.
 931 */
 932time64_t ktime_get_real_seconds(void)
 933{
 934	struct timekeeper *tk = &tk_core.timekeeper;
 935	time64_t seconds;
 936	unsigned int seq;
 937
 938	if (IS_ENABLED(CONFIG_64BIT))
 939		return tk->xtime_sec;
 940
 941	do {
 942		seq = read_seqcount_begin(&tk_core.seq);
 943		seconds = tk->xtime_sec;
 944
 945	} while (read_seqcount_retry(&tk_core.seq, seq));
 946
 947	return seconds;
 948}
 949EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
 950
 951/**
 952 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
 953 * but without the sequence counter protect. This internal function
 954 * is called just when timekeeping lock is already held.
 955 */
 956time64_t __ktime_get_real_seconds(void)
 957{
 958	struct timekeeper *tk = &tk_core.timekeeper;
 959
 960	return tk->xtime_sec;
 961}
 962
 963/**
 964 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
 965 * @systime_snapshot:	pointer to struct receiving the system time snapshot
 966 */
 967void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
 968{
 969	struct timekeeper *tk = &tk_core.timekeeper;
 970	unsigned int seq;
 971	ktime_t base_raw;
 972	ktime_t base_real;
 973	u64 nsec_raw;
 974	u64 nsec_real;
 975	u64 now;
 976
 977	WARN_ON_ONCE(timekeeping_suspended);
 978
 979	do {
 980		seq = read_seqcount_begin(&tk_core.seq);
 981		now = tk_clock_read(&tk->tkr_mono);
 
 982		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
 983		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
 984		base_real = ktime_add(tk->tkr_mono.base,
 985				      tk_core.timekeeper.offs_real);
 986		base_raw = tk->tkr_raw.base;
 987		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
 988		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
 989	} while (read_seqcount_retry(&tk_core.seq, seq));
 990
 991	systime_snapshot->cycles = now;
 992	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
 993	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
 994}
 995EXPORT_SYMBOL_GPL(ktime_get_snapshot);
 996
 997/* Scale base by mult/div checking for overflow */
 998static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
 999{
1000	u64 tmp, rem;
1001
1002	tmp = div64_u64_rem(*base, div, &rem);
1003
1004	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1005	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1006		return -EOVERFLOW;
1007	tmp *= mult;
1008	rem *= mult;
1009
1010	do_div(rem, div);
1011	*base = tmp + rem;
1012	return 0;
1013}
1014
1015/**
1016 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1017 * @history:			Snapshot representing start of history
1018 * @partial_history_cycles:	Cycle offset into history (fractional part)
1019 * @total_history_cycles:	Total history length in cycles
1020 * @discontinuity:		True indicates clock was set on history period
1021 * @ts:				Cross timestamp that should be adjusted using
1022 *	partial/total ratio
1023 *
1024 * Helper function used by get_device_system_crosststamp() to correct the
1025 * crosstimestamp corresponding to the start of the current interval to the
1026 * system counter value (timestamp point) provided by the driver. The
1027 * total_history_* quantities are the total history starting at the provided
1028 * reference point and ending at the start of the current interval. The cycle
1029 * count between the driver timestamp point and the start of the current
1030 * interval is partial_history_cycles.
1031 */
1032static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1033					 u64 partial_history_cycles,
1034					 u64 total_history_cycles,
1035					 bool discontinuity,
1036					 struct system_device_crosststamp *ts)
1037{
1038	struct timekeeper *tk = &tk_core.timekeeper;
1039	u64 corr_raw, corr_real;
1040	bool interp_forward;
1041	int ret;
1042
1043	if (total_history_cycles == 0 || partial_history_cycles == 0)
1044		return 0;
1045
1046	/* Interpolate shortest distance from beginning or end of history */
1047	interp_forward = partial_history_cycles > total_history_cycles / 2;
1048	partial_history_cycles = interp_forward ?
1049		total_history_cycles - partial_history_cycles :
1050		partial_history_cycles;
1051
1052	/*
1053	 * Scale the monotonic raw time delta by:
1054	 *	partial_history_cycles / total_history_cycles
1055	 */
1056	corr_raw = (u64)ktime_to_ns(
1057		ktime_sub(ts->sys_monoraw, history->raw));
1058	ret = scale64_check_overflow(partial_history_cycles,
1059				     total_history_cycles, &corr_raw);
1060	if (ret)
1061		return ret;
1062
1063	/*
1064	 * If there is a discontinuity in the history, scale monotonic raw
1065	 *	correction by:
1066	 *	mult(real)/mult(raw) yielding the realtime correction
1067	 * Otherwise, calculate the realtime correction similar to monotonic
1068	 *	raw calculation
1069	 */
1070	if (discontinuity) {
1071		corr_real = mul_u64_u32_div
1072			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1073	} else {
1074		corr_real = (u64)ktime_to_ns(
1075			ktime_sub(ts->sys_realtime, history->real));
1076		ret = scale64_check_overflow(partial_history_cycles,
1077					     total_history_cycles, &corr_real);
1078		if (ret)
1079			return ret;
1080	}
1081
1082	/* Fixup monotonic raw and real time time values */
1083	if (interp_forward) {
1084		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1085		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1086	} else {
1087		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1088		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1089	}
1090
1091	return 0;
1092}
1093
1094/*
1095 * cycle_between - true if test occurs chronologically between before and after
1096 */
1097static bool cycle_between(u64 before, u64 test, u64 after)
1098{
1099	if (test > before && test < after)
1100		return true;
1101	if (test < before && before > after)
1102		return true;
1103	return false;
1104}
1105
1106/**
1107 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1108 * @get_time_fn:	Callback to get simultaneous device time and
1109 *	system counter from the device driver
1110 * @ctx:		Context passed to get_time_fn()
1111 * @history_begin:	Historical reference point used to interpolate system
1112 *	time when counter provided by the driver is before the current interval
1113 * @xtstamp:		Receives simultaneously captured system and device time
1114 *
1115 * Reads a timestamp from a device and correlates it to system time
1116 */
1117int get_device_system_crosststamp(int (*get_time_fn)
1118				  (ktime_t *device_time,
1119				   struct system_counterval_t *sys_counterval,
1120				   void *ctx),
1121				  void *ctx,
1122				  struct system_time_snapshot *history_begin,
1123				  struct system_device_crosststamp *xtstamp)
1124{
1125	struct system_counterval_t system_counterval;
1126	struct timekeeper *tk = &tk_core.timekeeper;
1127	u64 cycles, now, interval_start;
1128	unsigned int clock_was_set_seq = 0;
1129	ktime_t base_real, base_raw;
1130	u64 nsec_real, nsec_raw;
1131	u8 cs_was_changed_seq;
1132	unsigned int seq;
1133	bool do_interp;
1134	int ret;
1135
1136	do {
1137		seq = read_seqcount_begin(&tk_core.seq);
1138		/*
1139		 * Try to synchronously capture device time and a system
1140		 * counter value calling back into the device driver
1141		 */
1142		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1143		if (ret)
1144			return ret;
1145
1146		/*
1147		 * Verify that the clocksource associated with the captured
1148		 * system counter value is the same as the currently installed
1149		 * timekeeper clocksource
1150		 */
1151		if (tk->tkr_mono.clock != system_counterval.cs)
1152			return -ENODEV;
1153		cycles = system_counterval.cycles;
1154
1155		/*
1156		 * Check whether the system counter value provided by the
1157		 * device driver is on the current timekeeping interval.
1158		 */
1159		now = tk_clock_read(&tk->tkr_mono);
1160		interval_start = tk->tkr_mono.cycle_last;
1161		if (!cycle_between(interval_start, cycles, now)) {
1162			clock_was_set_seq = tk->clock_was_set_seq;
1163			cs_was_changed_seq = tk->cs_was_changed_seq;
1164			cycles = interval_start;
1165			do_interp = true;
1166		} else {
1167			do_interp = false;
1168		}
1169
1170		base_real = ktime_add(tk->tkr_mono.base,
1171				      tk_core.timekeeper.offs_real);
1172		base_raw = tk->tkr_raw.base;
1173
1174		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1175						     system_counterval.cycles);
1176		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1177						    system_counterval.cycles);
1178	} while (read_seqcount_retry(&tk_core.seq, seq));
1179
1180	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1181	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1182
1183	/*
1184	 * Interpolate if necessary, adjusting back from the start of the
1185	 * current interval
1186	 */
1187	if (do_interp) {
1188		u64 partial_history_cycles, total_history_cycles;
1189		bool discontinuity;
1190
1191		/*
1192		 * Check that the counter value occurs after the provided
1193		 * history reference and that the history doesn't cross a
1194		 * clocksource change
1195		 */
1196		if (!history_begin ||
1197		    !cycle_between(history_begin->cycles,
1198				   system_counterval.cycles, cycles) ||
1199		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1200			return -EINVAL;
1201		partial_history_cycles = cycles - system_counterval.cycles;
1202		total_history_cycles = cycles - history_begin->cycles;
1203		discontinuity =
1204			history_begin->clock_was_set_seq != clock_was_set_seq;
1205
1206		ret = adjust_historical_crosststamp(history_begin,
1207						    partial_history_cycles,
1208						    total_history_cycles,
1209						    discontinuity, xtstamp);
1210		if (ret)
1211			return ret;
1212	}
1213
1214	return 0;
1215}
1216EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1217
1218/**
1219 * do_settimeofday64 - Sets the time of day.
1220 * @ts:     pointer to the timespec64 variable containing the new time
1221 *
1222 * Sets the time of day to the new time and update NTP and notify hrtimers
1223 */
1224int do_settimeofday64(const struct timespec64 *ts)
1225{
1226	struct timekeeper *tk = &tk_core.timekeeper;
1227	struct timespec64 ts_delta, xt;
1228	unsigned long flags;
1229	int ret = 0;
1230
1231	if (!timespec64_valid_settod(ts))
1232		return -EINVAL;
1233
1234	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1235	write_seqcount_begin(&tk_core.seq);
1236
1237	timekeeping_forward_now(tk);
1238
1239	xt = tk_xtime(tk);
1240	ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1241	ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1242
1243	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1244		ret = -EINVAL;
1245		goto out;
1246	}
1247
1248	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1249
1250	tk_set_xtime(tk, ts);
1251out:
1252	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1253
1254	write_seqcount_end(&tk_core.seq);
1255	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1256
1257	/* signal hrtimers about time change */
1258	clock_was_set();
1259
1260	if (!ret)
1261		audit_tk_injoffset(ts_delta);
 
 
1262
1263	return ret;
1264}
1265EXPORT_SYMBOL(do_settimeofday64);
1266
1267/**
1268 * timekeeping_inject_offset - Adds or subtracts from the current time.
1269 * @tv:		pointer to the timespec variable containing the offset
1270 *
1271 * Adds or subtracts an offset value from the current time.
1272 */
1273static int timekeeping_inject_offset(const struct timespec64 *ts)
1274{
1275	struct timekeeper *tk = &tk_core.timekeeper;
1276	unsigned long flags;
1277	struct timespec64 tmp;
1278	int ret = 0;
1279
1280	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1281		return -EINVAL;
1282
1283	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1284	write_seqcount_begin(&tk_core.seq);
1285
1286	timekeeping_forward_now(tk);
1287
1288	/* Make sure the proposed value is valid */
1289	tmp = timespec64_add(tk_xtime(tk), *ts);
1290	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1291	    !timespec64_valid_settod(&tmp)) {
1292		ret = -EINVAL;
1293		goto error;
1294	}
1295
1296	tk_xtime_add(tk, ts);
1297	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1298
1299error: /* even if we error out, we forwarded the time, so call update */
1300	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1301
1302	write_seqcount_end(&tk_core.seq);
1303	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1304
1305	/* signal hrtimers about time change */
1306	clock_was_set();
1307
1308	return ret;
1309}
1310
1311/*
1312 * Indicates if there is an offset between the system clock and the hardware
1313 * clock/persistent clock/rtc.
1314 */
1315int persistent_clock_is_local;
1316
1317/*
1318 * Adjust the time obtained from the CMOS to be UTC time instead of
1319 * local time.
1320 *
1321 * This is ugly, but preferable to the alternatives.  Otherwise we
1322 * would either need to write a program to do it in /etc/rc (and risk
1323 * confusion if the program gets run more than once; it would also be
1324 * hard to make the program warp the clock precisely n hours)  or
1325 * compile in the timezone information into the kernel.  Bad, bad....
1326 *
1327 *						- TYT, 1992-01-01
1328 *
1329 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1330 * as real UNIX machines always do it. This avoids all headaches about
1331 * daylight saving times and warping kernel clocks.
1332 */
1333void timekeeping_warp_clock(void)
1334{
1335	if (sys_tz.tz_minuteswest != 0) {
1336		struct timespec64 adjust;
1337
1338		persistent_clock_is_local = 1;
1339		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1340		adjust.tv_nsec = 0;
1341		timekeeping_inject_offset(&adjust);
1342	}
1343}
1344
1345/**
1346 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1347 *
1348 */
1349static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1350{
1351	tk->tai_offset = tai_offset;
1352	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1353}
1354
1355/**
1356 * change_clocksource - Swaps clocksources if a new one is available
1357 *
1358 * Accumulates current time interval and initializes new clocksource
1359 */
1360static int change_clocksource(void *data)
1361{
1362	struct timekeeper *tk = &tk_core.timekeeper;
1363	struct clocksource *new, *old;
1364	unsigned long flags;
 
1365
1366	new = (struct clocksource *) data;
1367
1368	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1369	write_seqcount_begin(&tk_core.seq);
1370
1371	timekeeping_forward_now(tk);
1372	/*
1373	 * If the cs is in module, get a module reference. Succeeds
1374	 * for built-in code (owner == NULL) as well.
1375	 */
1376	if (try_module_get(new->owner)) {
1377		if (!new->enable || new->enable(new) == 0) {
1378			old = tk->tkr_mono.clock;
1379			tk_setup_internals(tk, new);
1380			if (old->disable)
1381				old->disable(old);
1382			module_put(old->owner);
1383		} else {
1384			module_put(new->owner);
1385		}
1386	}
 
 
 
 
 
 
 
 
 
 
 
1387	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1388
1389	write_seqcount_end(&tk_core.seq);
1390	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1391
 
 
 
 
 
 
 
1392	return 0;
1393}
1394
1395/**
1396 * timekeeping_notify - Install a new clock source
1397 * @clock:		pointer to the clock source
1398 *
1399 * This function is called from clocksource.c after a new, better clock
1400 * source has been registered. The caller holds the clocksource_mutex.
1401 */
1402int timekeeping_notify(struct clocksource *clock)
1403{
1404	struct timekeeper *tk = &tk_core.timekeeper;
1405
1406	if (tk->tkr_mono.clock == clock)
1407		return 0;
1408	stop_machine(change_clocksource, clock, NULL);
1409	tick_clock_notify();
1410	return tk->tkr_mono.clock == clock ? 0 : -1;
1411}
1412
1413/**
1414 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1415 * @ts:		pointer to the timespec64 to be set
1416 *
1417 * Returns the raw monotonic time (completely un-modified by ntp)
1418 */
1419void ktime_get_raw_ts64(struct timespec64 *ts)
1420{
1421	struct timekeeper *tk = &tk_core.timekeeper;
1422	unsigned int seq;
1423	u64 nsecs;
1424
1425	do {
1426		seq = read_seqcount_begin(&tk_core.seq);
1427		ts->tv_sec = tk->raw_sec;
1428		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1429
1430	} while (read_seqcount_retry(&tk_core.seq, seq));
1431
1432	ts->tv_nsec = 0;
1433	timespec64_add_ns(ts, nsecs);
1434}
1435EXPORT_SYMBOL(ktime_get_raw_ts64);
1436
1437
1438/**
1439 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1440 */
1441int timekeeping_valid_for_hres(void)
1442{
1443	struct timekeeper *tk = &tk_core.timekeeper;
1444	unsigned int seq;
1445	int ret;
1446
1447	do {
1448		seq = read_seqcount_begin(&tk_core.seq);
1449
1450		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1451
1452	} while (read_seqcount_retry(&tk_core.seq, seq));
1453
1454	return ret;
1455}
1456
1457/**
1458 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1459 */
1460u64 timekeeping_max_deferment(void)
1461{
1462	struct timekeeper *tk = &tk_core.timekeeper;
1463	unsigned int seq;
1464	u64 ret;
1465
1466	do {
1467		seq = read_seqcount_begin(&tk_core.seq);
1468
1469		ret = tk->tkr_mono.clock->max_idle_ns;
1470
1471	} while (read_seqcount_retry(&tk_core.seq, seq));
1472
1473	return ret;
1474}
1475
1476/**
1477 * read_persistent_clock64 -  Return time from the persistent clock.
 
1478 *
1479 * Weak dummy function for arches that do not yet support it.
1480 * Reads the time from the battery backed persistent clock.
1481 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1482 *
1483 *  XXX - Do be sure to remove it once all arches implement it.
1484 */
1485void __weak read_persistent_clock64(struct timespec64 *ts)
1486{
1487	ts->tv_sec = 0;
1488	ts->tv_nsec = 0;
1489}
1490
1491/**
1492 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1493 *                                        from the boot.
 
 
1494 *
1495 * Weak dummy function for arches that do not yet support it.
1496 * wall_time	- current time as returned by persistent clock
1497 * boot_offset	- offset that is defined as wall_time - boot_time
1498 * The default function calculates offset based on the current value of
1499 * local_clock(). This way architectures that support sched_clock() but don't
1500 * support dedicated boot time clock will provide the best estimate of the
1501 * boot time.
1502 */
1503void __weak __init
1504read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1505				     struct timespec64 *boot_offset)
1506{
1507	read_persistent_clock64(wall_time);
1508	*boot_offset = ns_to_timespec64(local_clock());
1509}
1510
1511/*
1512 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1513 *
1514 * The flag starts of false and is only set when a suspend reaches
1515 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1516 * timekeeper clocksource is not stopping across suspend and has been
1517 * used to update sleep time. If the timekeeper clocksource has stopped
1518 * then the flag stays true and is used by the RTC resume code to decide
1519 * whether sleeptime must be injected and if so the flag gets false then.
1520 *
1521 * If a suspend fails before reaching timekeeping_resume() then the flag
1522 * stays false and prevents erroneous sleeptime injection.
1523 */
1524static bool suspend_timing_needed;
1525
1526/* Flag for if there is a persistent clock on this platform */
1527static bool persistent_clock_exists;
1528
1529/*
1530 * timekeeping_init - Initializes the clocksource and common timekeeping values
1531 */
1532void __init timekeeping_init(void)
1533{
1534	struct timespec64 wall_time, boot_offset, wall_to_mono;
1535	struct timekeeper *tk = &tk_core.timekeeper;
1536	struct clocksource *clock;
1537	unsigned long flags;
1538
1539	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1540	if (timespec64_valid_settod(&wall_time) &&
1541	    timespec64_to_ns(&wall_time) > 0) {
1542		persistent_clock_exists = true;
1543	} else if (timespec64_to_ns(&wall_time) != 0) {
1544		pr_warn("Persistent clock returned invalid value");
1545		wall_time = (struct timespec64){0};
1546	}
1547
1548	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1549		boot_offset = (struct timespec64){0};
1550
1551	/*
1552	 * We want set wall_to_mono, so the following is true:
1553	 * wall time + wall_to_mono = boot time
1554	 */
1555	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1556
1557	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1558	write_seqcount_begin(&tk_core.seq);
1559	ntp_init();
1560
1561	clock = clocksource_default_clock();
1562	if (clock->enable)
1563		clock->enable(clock);
1564	tk_setup_internals(tk, clock);
1565
1566	tk_set_xtime(tk, &wall_time);
1567	tk->raw_sec = 0;
1568
1569	tk_set_wall_to_mono(tk, wall_to_mono);
1570
1571	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1572
1573	write_seqcount_end(&tk_core.seq);
1574	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1575}
1576
1577/* time in seconds when suspend began for persistent clock */
1578static struct timespec64 timekeeping_suspend_time;
1579
1580/**
1581 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1582 * @delta: pointer to a timespec delta value
 
1583 *
1584 * Takes a timespec offset measuring a suspend interval and properly
1585 * adds the sleep offset to the timekeeping variables.
1586 */
1587static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1588					   const struct timespec64 *delta)
1589{
1590	if (!timespec64_valid_strict(delta)) {
1591		printk_deferred(KERN_WARNING
1592				"__timekeeping_inject_sleeptime: Invalid "
1593				"sleep delta value!\n");
1594		return;
1595	}
1596	tk_xtime_add(tk, delta);
1597	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1598	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1599	tk_debug_account_sleep_time(delta);
1600}
1601
1602#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1603/**
1604 * We have three kinds of time sources to use for sleep time
1605 * injection, the preference order is:
1606 * 1) non-stop clocksource
1607 * 2) persistent clock (ie: RTC accessible when irqs are off)
1608 * 3) RTC
1609 *
1610 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1611 * If system has neither 1) nor 2), 3) will be used finally.
1612 *
1613 *
1614 * If timekeeping has injected sleeptime via either 1) or 2),
1615 * 3) becomes needless, so in this case we don't need to call
1616 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1617 * means.
1618 */
1619bool timekeeping_rtc_skipresume(void)
1620{
1621	return !suspend_timing_needed;
1622}
1623
1624/**
1625 * 1) can be determined whether to use or not only when doing
1626 * timekeeping_resume() which is invoked after rtc_suspend(),
1627 * so we can't skip rtc_suspend() surely if system has 1).
1628 *
1629 * But if system has 2), 2) will definitely be used, so in this
1630 * case we don't need to call rtc_suspend(), and this is what
1631 * timekeeping_rtc_skipsuspend() means.
1632 */
1633bool timekeeping_rtc_skipsuspend(void)
1634{
1635	return persistent_clock_exists;
1636}
1637
1638/**
1639 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1640 * @delta: pointer to a timespec64 delta value
1641 *
1642 * This hook is for architectures that cannot support read_persistent_clock64
1643 * because their RTC/persistent clock is only accessible when irqs are enabled.
1644 * and also don't have an effective nonstop clocksource.
1645 *
1646 * This function should only be called by rtc_resume(), and allows
1647 * a suspend offset to be injected into the timekeeping values.
1648 */
1649void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1650{
1651	struct timekeeper *tk = &tk_core.timekeeper;
1652	unsigned long flags;
1653
1654	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1655	write_seqcount_begin(&tk_core.seq);
1656
1657	suspend_timing_needed = false;
1658
1659	timekeeping_forward_now(tk);
1660
1661	__timekeeping_inject_sleeptime(tk, delta);
1662
1663	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1664
1665	write_seqcount_end(&tk_core.seq);
1666	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1667
1668	/* signal hrtimers about time change */
1669	clock_was_set();
1670}
1671#endif
1672
1673/**
1674 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1675 */
1676void timekeeping_resume(void)
1677{
1678	struct timekeeper *tk = &tk_core.timekeeper;
1679	struct clocksource *clock = tk->tkr_mono.clock;
1680	unsigned long flags;
1681	struct timespec64 ts_new, ts_delta;
1682	u64 cycle_now, nsec;
1683	bool inject_sleeptime = false;
1684
1685	read_persistent_clock64(&ts_new);
1686
1687	clockevents_resume();
1688	clocksource_resume();
1689
1690	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1691	write_seqcount_begin(&tk_core.seq);
1692
1693	/*
1694	 * After system resumes, we need to calculate the suspended time and
1695	 * compensate it for the OS time. There are 3 sources that could be
1696	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1697	 * device.
1698	 *
1699	 * One specific platform may have 1 or 2 or all of them, and the
1700	 * preference will be:
1701	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1702	 * The less preferred source will only be tried if there is no better
1703	 * usable source. The rtc part is handled separately in rtc core code.
1704	 */
1705	cycle_now = tk_clock_read(&tk->tkr_mono);
1706	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1707	if (nsec > 0) {
1708		ts_delta = ns_to_timespec64(nsec);
1709		inject_sleeptime = true;
1710	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1711		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1712		inject_sleeptime = true;
1713	}
1714
1715	if (inject_sleeptime) {
1716		suspend_timing_needed = false;
1717		__timekeeping_inject_sleeptime(tk, &ts_delta);
1718	}
1719
1720	/* Re-base the last cycle value */
1721	tk->tkr_mono.cycle_last = cycle_now;
1722	tk->tkr_raw.cycle_last  = cycle_now;
1723
1724	tk->ntp_error = 0;
1725	timekeeping_suspended = 0;
1726	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1727	write_seqcount_end(&tk_core.seq);
1728	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1729
1730	touch_softlockup_watchdog();
1731
 
1732	tick_resume();
1733	hrtimers_resume();
 
1734}
1735
1736int timekeeping_suspend(void)
1737{
1738	struct timekeeper *tk = &tk_core.timekeeper;
1739	unsigned long flags;
1740	struct timespec64		delta, delta_delta;
1741	static struct timespec64	old_delta;
1742	struct clocksource *curr_clock;
1743	u64 cycle_now;
1744
1745	read_persistent_clock64(&timekeeping_suspend_time);
1746
1747	/*
1748	 * On some systems the persistent_clock can not be detected at
1749	 * timekeeping_init by its return value, so if we see a valid
1750	 * value returned, update the persistent_clock_exists flag.
1751	 */
1752	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1753		persistent_clock_exists = true;
1754
1755	suspend_timing_needed = true;
1756
1757	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1758	write_seqcount_begin(&tk_core.seq);
1759	timekeeping_forward_now(tk);
1760	timekeeping_suspended = 1;
1761
1762	/*
1763	 * Since we've called forward_now, cycle_last stores the value
1764	 * just read from the current clocksource. Save this to potentially
1765	 * use in suspend timing.
1766	 */
1767	curr_clock = tk->tkr_mono.clock;
1768	cycle_now = tk->tkr_mono.cycle_last;
1769	clocksource_start_suspend_timing(curr_clock, cycle_now);
1770
1771	if (persistent_clock_exists) {
1772		/*
1773		 * To avoid drift caused by repeated suspend/resumes,
1774		 * which each can add ~1 second drift error,
1775		 * try to compensate so the difference in system time
1776		 * and persistent_clock time stays close to constant.
1777		 */
1778		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1779		delta_delta = timespec64_sub(delta, old_delta);
1780		if (abs(delta_delta.tv_sec) >= 2) {
1781			/*
1782			 * if delta_delta is too large, assume time correction
1783			 * has occurred and set old_delta to the current delta.
1784			 */
1785			old_delta = delta;
1786		} else {
1787			/* Otherwise try to adjust old_system to compensate */
1788			timekeeping_suspend_time =
1789				timespec64_add(timekeeping_suspend_time, delta_delta);
1790		}
1791	}
1792
1793	timekeeping_update(tk, TK_MIRROR);
1794	halt_fast_timekeeper(tk);
1795	write_seqcount_end(&tk_core.seq);
1796	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1797
1798	tick_suspend();
1799	clocksource_suspend();
1800	clockevents_suspend();
1801
1802	return 0;
1803}
1804
1805/* sysfs resume/suspend bits for timekeeping */
1806static struct syscore_ops timekeeping_syscore_ops = {
1807	.resume		= timekeeping_resume,
1808	.suspend	= timekeeping_suspend,
1809};
1810
1811static int __init timekeeping_init_ops(void)
1812{
1813	register_syscore_ops(&timekeeping_syscore_ops);
1814	return 0;
1815}
1816device_initcall(timekeeping_init_ops);
1817
1818/*
1819 * Apply a multiplier adjustment to the timekeeper
1820 */
1821static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1822							 s64 offset,
1823							 s32 mult_adj)
1824{
1825	s64 interval = tk->cycle_interval;
1826
1827	if (mult_adj == 0) {
1828		return;
1829	} else if (mult_adj == -1) {
1830		interval = -interval;
1831		offset = -offset;
1832	} else if (mult_adj != 1) {
1833		interval *= mult_adj;
1834		offset *= mult_adj;
1835	}
1836
1837	/*
1838	 * So the following can be confusing.
1839	 *
1840	 * To keep things simple, lets assume mult_adj == 1 for now.
1841	 *
1842	 * When mult_adj != 1, remember that the interval and offset values
1843	 * have been appropriately scaled so the math is the same.
1844	 *
1845	 * The basic idea here is that we're increasing the multiplier
1846	 * by one, this causes the xtime_interval to be incremented by
1847	 * one cycle_interval. This is because:
1848	 *	xtime_interval = cycle_interval * mult
1849	 * So if mult is being incremented by one:
1850	 *	xtime_interval = cycle_interval * (mult + 1)
1851	 * Its the same as:
1852	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1853	 * Which can be shortened to:
1854	 *	xtime_interval += cycle_interval
1855	 *
1856	 * So offset stores the non-accumulated cycles. Thus the current
1857	 * time (in shifted nanoseconds) is:
1858	 *	now = (offset * adj) + xtime_nsec
1859	 * Now, even though we're adjusting the clock frequency, we have
1860	 * to keep time consistent. In other words, we can't jump back
1861	 * in time, and we also want to avoid jumping forward in time.
1862	 *
1863	 * So given the same offset value, we need the time to be the same
1864	 * both before and after the freq adjustment.
1865	 *	now = (offset * adj_1) + xtime_nsec_1
1866	 *	now = (offset * adj_2) + xtime_nsec_2
1867	 * So:
1868	 *	(offset * adj_1) + xtime_nsec_1 =
1869	 *		(offset * adj_2) + xtime_nsec_2
1870	 * And we know:
1871	 *	adj_2 = adj_1 + 1
1872	 * So:
1873	 *	(offset * adj_1) + xtime_nsec_1 =
1874	 *		(offset * (adj_1+1)) + xtime_nsec_2
1875	 *	(offset * adj_1) + xtime_nsec_1 =
1876	 *		(offset * adj_1) + offset + xtime_nsec_2
1877	 * Canceling the sides:
1878	 *	xtime_nsec_1 = offset + xtime_nsec_2
1879	 * Which gives us:
1880	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1881	 * Which simplfies to:
1882	 *	xtime_nsec -= offset
1883	 */
1884	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1885		/* NTP adjustment caused clocksource mult overflow */
1886		WARN_ON_ONCE(1);
1887		return;
1888	}
1889
1890	tk->tkr_mono.mult += mult_adj;
1891	tk->xtime_interval += interval;
1892	tk->tkr_mono.xtime_nsec -= offset;
1893}
1894
1895/*
1896 * Adjust the timekeeper's multiplier to the correct frequency
1897 * and also to reduce the accumulated error value.
1898 */
1899static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1900{
1901	u32 mult;
1902
1903	/*
1904	 * Determine the multiplier from the current NTP tick length.
1905	 * Avoid expensive division when the tick length doesn't change.
1906	 */
1907	if (likely(tk->ntp_tick == ntp_tick_length())) {
1908		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1909	} else {
1910		tk->ntp_tick = ntp_tick_length();
1911		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1912				 tk->xtime_remainder, tk->cycle_interval);
1913	}
1914
1915	/*
1916	 * If the clock is behind the NTP time, increase the multiplier by 1
1917	 * to catch up with it. If it's ahead and there was a remainder in the
1918	 * tick division, the clock will slow down. Otherwise it will stay
1919	 * ahead until the tick length changes to a non-divisible value.
1920	 */
1921	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1922	mult += tk->ntp_err_mult;
1923
1924	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
1925
1926	if (unlikely(tk->tkr_mono.clock->maxadj &&
1927		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1928			> tk->tkr_mono.clock->maxadj))) {
1929		printk_once(KERN_WARNING
1930			"Adjusting %s more than 11%% (%ld vs %ld)\n",
1931			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1932			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1933	}
1934
1935	/*
1936	 * It may be possible that when we entered this function, xtime_nsec
1937	 * was very small.  Further, if we're slightly speeding the clocksource
1938	 * in the code above, its possible the required corrective factor to
1939	 * xtime_nsec could cause it to underflow.
1940	 *
1941	 * Now, since we have already accumulated the second and the NTP
1942	 * subsystem has been notified via second_overflow(), we need to skip
1943	 * the next update.
1944	 */
1945	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1946		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
1947							tk->tkr_mono.shift;
1948		tk->xtime_sec--;
1949		tk->skip_second_overflow = 1;
1950	}
1951}
1952
1953/**
1954 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1955 *
1956 * Helper function that accumulates the nsecs greater than a second
1957 * from the xtime_nsec field to the xtime_secs field.
1958 * It also calls into the NTP code to handle leapsecond processing.
1959 *
1960 */
1961static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1962{
1963	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1964	unsigned int clock_set = 0;
1965
1966	while (tk->tkr_mono.xtime_nsec >= nsecps) {
1967		int leap;
1968
1969		tk->tkr_mono.xtime_nsec -= nsecps;
1970		tk->xtime_sec++;
1971
1972		/*
1973		 * Skip NTP update if this second was accumulated before,
1974		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
1975		 */
1976		if (unlikely(tk->skip_second_overflow)) {
1977			tk->skip_second_overflow = 0;
1978			continue;
1979		}
1980
1981		/* Figure out if its a leap sec and apply if needed */
1982		leap = second_overflow(tk->xtime_sec);
1983		if (unlikely(leap)) {
1984			struct timespec64 ts;
1985
1986			tk->xtime_sec += leap;
1987
1988			ts.tv_sec = leap;
1989			ts.tv_nsec = 0;
1990			tk_set_wall_to_mono(tk,
1991				timespec64_sub(tk->wall_to_monotonic, ts));
1992
1993			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1994
1995			clock_set = TK_CLOCK_WAS_SET;
1996		}
1997	}
1998	return clock_set;
1999}
2000
2001/**
2002 * logarithmic_accumulation - shifted accumulation of cycles
2003 *
2004 * This functions accumulates a shifted interval of cycles into
2005 * into a shifted interval nanoseconds. Allows for O(log) accumulation
2006 * loop.
2007 *
2008 * Returns the unconsumed cycles.
2009 */
2010static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2011				    u32 shift, unsigned int *clock_set)
2012{
2013	u64 interval = tk->cycle_interval << shift;
2014	u64 snsec_per_sec;
2015
2016	/* If the offset is smaller than a shifted interval, do nothing */
2017	if (offset < interval)
2018		return offset;
2019
2020	/* Accumulate one shifted interval */
2021	offset -= interval;
2022	tk->tkr_mono.cycle_last += interval;
2023	tk->tkr_raw.cycle_last  += interval;
2024
2025	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2026	*clock_set |= accumulate_nsecs_to_secs(tk);
2027
2028	/* Accumulate raw time */
2029	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2030	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2031	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2032		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2033		tk->raw_sec++;
2034	}
2035
2036	/* Accumulate error between NTP and clock interval */
2037	tk->ntp_error += tk->ntp_tick << shift;
2038	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2039						(tk->ntp_error_shift + shift);
2040
2041	return offset;
2042}
2043
2044/*
2045 * timekeeping_advance - Updates the timekeeper to the current time and
2046 * current NTP tick length
2047 */
2048static void timekeeping_advance(enum timekeeping_adv_mode mode)
2049{
2050	struct timekeeper *real_tk = &tk_core.timekeeper;
2051	struct timekeeper *tk = &shadow_timekeeper;
2052	u64 offset;
2053	int shift = 0, maxshift;
2054	unsigned int clock_set = 0;
2055	unsigned long flags;
2056
2057	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2058
2059	/* Make sure we're fully resumed: */
2060	if (unlikely(timekeeping_suspended))
2061		goto out;
2062
2063#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2064	offset = real_tk->cycle_interval;
2065
2066	if (mode != TK_ADV_TICK)
2067		goto out;
2068#else
2069	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2070				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2071
2072	/* Check if there's really nothing to do */
2073	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2074		goto out;
2075#endif
2076
2077	/* Do some additional sanity checking */
2078	timekeeping_check_update(tk, offset);
2079
2080	/*
2081	 * With NO_HZ we may have to accumulate many cycle_intervals
2082	 * (think "ticks") worth of time at once. To do this efficiently,
2083	 * we calculate the largest doubling multiple of cycle_intervals
2084	 * that is smaller than the offset.  We then accumulate that
2085	 * chunk in one go, and then try to consume the next smaller
2086	 * doubled multiple.
2087	 */
2088	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2089	shift = max(0, shift);
2090	/* Bound shift to one less than what overflows tick_length */
2091	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2092	shift = min(shift, maxshift);
2093	while (offset >= tk->cycle_interval) {
2094		offset = logarithmic_accumulation(tk, offset, shift,
2095							&clock_set);
2096		if (offset < tk->cycle_interval<<shift)
2097			shift--;
2098	}
2099
2100	/* Adjust the multiplier to correct NTP error */
2101	timekeeping_adjust(tk, offset);
2102
2103	/*
2104	 * Finally, make sure that after the rounding
2105	 * xtime_nsec isn't larger than NSEC_PER_SEC
2106	 */
2107	clock_set |= accumulate_nsecs_to_secs(tk);
2108
2109	write_seqcount_begin(&tk_core.seq);
2110	/*
2111	 * Update the real timekeeper.
2112	 *
2113	 * We could avoid this memcpy by switching pointers, but that
2114	 * requires changes to all other timekeeper usage sites as
2115	 * well, i.e. move the timekeeper pointer getter into the
2116	 * spinlocked/seqcount protected sections. And we trade this
2117	 * memcpy under the tk_core.seq against one before we start
2118	 * updating.
2119	 */
2120	timekeeping_update(tk, clock_set);
2121	memcpy(real_tk, tk, sizeof(*tk));
2122	/* The memcpy must come last. Do not put anything here! */
2123	write_seqcount_end(&tk_core.seq);
2124out:
2125	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2126	if (clock_set)
2127		/* Have to call _delayed version, since in irq context*/
2128		clock_was_set_delayed();
2129}
2130
2131/**
2132 * update_wall_time - Uses the current clocksource to increment the wall time
2133 *
2134 */
2135void update_wall_time(void)
2136{
2137	timekeeping_advance(TK_ADV_TICK);
 
2138}
2139
2140/**
2141 * getboottime64 - Return the real time of system boot.
2142 * @ts:		pointer to the timespec64 to be set
2143 *
2144 * Returns the wall-time of boot in a timespec64.
2145 *
2146 * This is based on the wall_to_monotonic offset and the total suspend
2147 * time. Calls to settimeofday will affect the value returned (which
2148 * basically means that however wrong your real time clock is at boot time,
2149 * you get the right time here).
2150 */
2151void getboottime64(struct timespec64 *ts)
2152{
2153	struct timekeeper *tk = &tk_core.timekeeper;
2154	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2155
2156	*ts = ktime_to_timespec64(t);
2157}
2158EXPORT_SYMBOL_GPL(getboottime64);
2159
2160void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2161{
2162	struct timekeeper *tk = &tk_core.timekeeper;
2163	unsigned int seq;
2164
2165	do {
2166		seq = read_seqcount_begin(&tk_core.seq);
2167
2168		*ts = tk_xtime(tk);
2169	} while (read_seqcount_retry(&tk_core.seq, seq));
2170}
2171EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2172
2173void ktime_get_coarse_ts64(struct timespec64 *ts)
2174{
2175	struct timekeeper *tk = &tk_core.timekeeper;
2176	struct timespec64 now, mono;
2177	unsigned int seq;
2178
2179	do {
2180		seq = read_seqcount_begin(&tk_core.seq);
2181
2182		now = tk_xtime(tk);
2183		mono = tk->wall_to_monotonic;
2184	} while (read_seqcount_retry(&tk_core.seq, seq));
2185
2186	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2187				now.tv_nsec + mono.tv_nsec);
2188}
2189EXPORT_SYMBOL(ktime_get_coarse_ts64);
2190
2191/*
2192 * Must hold jiffies_lock
2193 */
2194void do_timer(unsigned long ticks)
2195{
2196	jiffies_64 += ticks;
2197	calc_global_load(ticks);
2198}
2199
2200/**
2201 * ktime_get_update_offsets_now - hrtimer helper
2202 * @cwsseq:	pointer to check and store the clock was set sequence number
2203 * @offs_real:	pointer to storage for monotonic -> realtime offset
2204 * @offs_boot:	pointer to storage for monotonic -> boottime offset
2205 * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2206 *
2207 * Returns current monotonic time and updates the offsets if the
2208 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2209 * different.
2210 *
2211 * Called from hrtimer_interrupt() or retrigger_next_event()
2212 */
2213ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2214				     ktime_t *offs_boot, ktime_t *offs_tai)
2215{
2216	struct timekeeper *tk = &tk_core.timekeeper;
2217	unsigned int seq;
2218	ktime_t base;
2219	u64 nsecs;
2220
2221	do {
2222		seq = read_seqcount_begin(&tk_core.seq);
2223
2224		base = tk->tkr_mono.base;
2225		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2226		base = ktime_add_ns(base, nsecs);
2227
2228		if (*cwsseq != tk->clock_was_set_seq) {
2229			*cwsseq = tk->clock_was_set_seq;
2230			*offs_real = tk->offs_real;
2231			*offs_boot = tk->offs_boot;
2232			*offs_tai = tk->offs_tai;
2233		}
2234
2235		/* Handle leapsecond insertion adjustments */
2236		if (unlikely(base >= tk->next_leap_ktime))
2237			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2238
2239	} while (read_seqcount_retry(&tk_core.seq, seq));
2240
2241	return base;
2242}
2243
2244/**
2245 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2246 */
2247static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2248{
2249	if (txc->modes & ADJ_ADJTIME) {
2250		/* singleshot must not be used with any other mode bits */
2251		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2252			return -EINVAL;
2253		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2254		    !capable(CAP_SYS_TIME))
2255			return -EPERM;
2256	} else {
2257		/* In order to modify anything, you gotta be super-user! */
2258		if (txc->modes && !capable(CAP_SYS_TIME))
2259			return -EPERM;
2260		/*
2261		 * if the quartz is off by more than 10% then
2262		 * something is VERY wrong!
2263		 */
2264		if (txc->modes & ADJ_TICK &&
2265		    (txc->tick <  900000/USER_HZ ||
2266		     txc->tick > 1100000/USER_HZ))
2267			return -EINVAL;
2268	}
2269
2270	if (txc->modes & ADJ_SETOFFSET) {
2271		/* In order to inject time, you gotta be super-user! */
2272		if (!capable(CAP_SYS_TIME))
2273			return -EPERM;
2274
2275		/*
2276		 * Validate if a timespec/timeval used to inject a time
2277		 * offset is valid.  Offsets can be postive or negative, so
2278		 * we don't check tv_sec. The value of the timeval/timespec
2279		 * is the sum of its fields,but *NOTE*:
2280		 * The field tv_usec/tv_nsec must always be non-negative and
2281		 * we can't have more nanoseconds/microseconds than a second.
2282		 */
2283		if (txc->time.tv_usec < 0)
2284			return -EINVAL;
2285
2286		if (txc->modes & ADJ_NANO) {
2287			if (txc->time.tv_usec >= NSEC_PER_SEC)
2288				return -EINVAL;
2289		} else {
2290			if (txc->time.tv_usec >= USEC_PER_SEC)
2291				return -EINVAL;
2292		}
2293	}
2294
2295	/*
2296	 * Check for potential multiplication overflows that can
2297	 * only happen on 64-bit systems:
2298	 */
2299	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2300		if (LLONG_MIN / PPM_SCALE > txc->freq)
2301			return -EINVAL;
2302		if (LLONG_MAX / PPM_SCALE < txc->freq)
2303			return -EINVAL;
2304	}
2305
2306	return 0;
2307}
2308
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2309
2310/**
2311 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2312 */
2313int do_adjtimex(struct __kernel_timex *txc)
2314{
2315	struct timekeeper *tk = &tk_core.timekeeper;
2316	struct audit_ntp_data ad;
2317	unsigned long flags;
2318	struct timespec64 ts;
 
2319	s32 orig_tai, tai;
2320	int ret;
2321
2322	/* Validate the data before disabling interrupts */
2323	ret = timekeeping_validate_timex(txc);
2324	if (ret)
2325		return ret;
 
2326
2327	if (txc->modes & ADJ_SETOFFSET) {
2328		struct timespec64 delta;
2329		delta.tv_sec  = txc->time.tv_sec;
2330		delta.tv_nsec = txc->time.tv_usec;
2331		if (!(txc->modes & ADJ_NANO))
2332			delta.tv_nsec *= 1000;
2333		ret = timekeeping_inject_offset(&delta);
2334		if (ret)
2335			return ret;
2336
2337		audit_tk_injoffset(delta);
2338	}
2339
2340	audit_ntp_init(&ad);
2341
2342	ktime_get_real_ts64(&ts);
 
2343
2344	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2345	write_seqcount_begin(&tk_core.seq);
2346
2347	orig_tai = tai = tk->tai_offset;
2348	ret = __do_adjtimex(txc, &ts, &tai, &ad);
2349
2350	if (tai != orig_tai) {
2351		__timekeeping_set_tai_offset(tk, tai);
2352		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
 
2353	}
2354	tk_update_leap_state(tk);
2355
2356	write_seqcount_end(&tk_core.seq);
2357	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2358
2359	audit_ntp_log(&ad);
2360
2361	/* Update the multiplier immediately if frequency was set directly */
2362	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2363		timekeeping_advance(TK_ADV_FREQ);
2364
2365	if (tai != orig_tai)
2366		clock_was_set();
2367
2368	ntp_notify_cmos_timer();
2369
2370	return ret;
2371}
2372
2373#ifdef CONFIG_NTP_PPS
2374/**
2375 * hardpps() - Accessor function to NTP __hardpps function
2376 */
2377void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2378{
2379	unsigned long flags;
2380
2381	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2382	write_seqcount_begin(&tk_core.seq);
2383
2384	__hardpps(phase_ts, raw_ts);
2385
2386	write_seqcount_end(&tk_core.seq);
2387	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2388}
2389EXPORT_SYMBOL(hardpps);
2390#endif /* CONFIG_NTP_PPS */
2391
2392/**
2393 * xtime_update() - advances the timekeeping infrastructure
2394 * @ticks:	number of ticks, that have elapsed since the last call.
2395 *
2396 * Must be called with interrupts disabled.
2397 */
2398void xtime_update(unsigned long ticks)
2399{
2400	write_seqlock(&jiffies_lock);
2401	do_timer(ticks);
2402	write_sequnlock(&jiffies_lock);
2403	update_wall_time();
2404}