Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Kernel timekeeping code and accessor functions. Based on code from
   4 *  timer.c, moved in commit 8524070b7982.
   5 */
   6#include <linux/timekeeper_internal.h>
   7#include <linux/module.h>
   8#include <linux/interrupt.h>
   9#include <linux/percpu.h>
  10#include <linux/init.h>
  11#include <linux/mm.h>
  12#include <linux/nmi.h>
  13#include <linux/sched.h>
  14#include <linux/sched/loadavg.h>
  15#include <linux/sched/clock.h>
  16#include <linux/syscore_ops.h>
  17#include <linux/clocksource.h>
  18#include <linux/jiffies.h>
  19#include <linux/time.h>
  20#include <linux/timex.h>
  21#include <linux/tick.h>
  22#include <linux/stop_machine.h>
  23#include <linux/pvclock_gtod.h>
  24#include <linux/compiler.h>
  25#include <linux/audit.h>
  26#include <linux/random.h>
  27
  28#include "tick-internal.h"
  29#include "ntp_internal.h"
  30#include "timekeeping_internal.h"
  31
  32#define TK_CLEAR_NTP		(1 << 0)
  33#define TK_MIRROR		(1 << 1)
  34#define TK_CLOCK_WAS_SET	(1 << 2)
  35
  36enum timekeeping_adv_mode {
  37	/* Update timekeeper when a tick has passed */
  38	TK_ADV_TICK,
  39
  40	/* Update timekeeper on a direct frequency change */
  41	TK_ADV_FREQ
  42};
  43
  44DEFINE_RAW_SPINLOCK(timekeeper_lock);
  45
  46/*
  47 * The most important data for readout fits into a single 64 byte
  48 * cache line.
  49 */
  50static struct {
  51	seqcount_raw_spinlock_t	seq;
  52	struct timekeeper	timekeeper;
  53} tk_core ____cacheline_aligned = {
  54	.seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
  55};
  56
  57static struct timekeeper shadow_timekeeper;
  58
  59/* flag for if timekeeping is suspended */
  60int __read_mostly timekeeping_suspended;
  61
  62/**
  63 * struct tk_fast - NMI safe timekeeper
  64 * @seq:	Sequence counter for protecting updates. The lowest bit
  65 *		is the index for the tk_read_base array
  66 * @base:	tk_read_base array. Access is indexed by the lowest bit of
  67 *		@seq.
  68 *
  69 * See @update_fast_timekeeper() below.
  70 */
  71struct tk_fast {
  72	seqcount_latch_t	seq;
  73	struct tk_read_base	base[2];
  74};
  75
  76/* Suspend-time cycles value for halted fast timekeeper. */
  77static u64 cycles_at_suspend;
  78
  79static u64 dummy_clock_read(struct clocksource *cs)
  80{
  81	if (timekeeping_suspended)
  82		return cycles_at_suspend;
  83	return local_clock();
  84}
  85
  86static struct clocksource dummy_clock = {
  87	.read = dummy_clock_read,
  88};
  89
  90/*
  91 * Boot time initialization which allows local_clock() to be utilized
  92 * during early boot when clocksources are not available. local_clock()
  93 * returns nanoseconds already so no conversion is required, hence mult=1
  94 * and shift=0. When the first proper clocksource is installed then
  95 * the fast time keepers are updated with the correct values.
  96 */
  97#define FAST_TK_INIT						\
  98	{							\
  99		.clock		= &dummy_clock,			\
 100		.mask		= CLOCKSOURCE_MASK(64),		\
 101		.mult		= 1,				\
 102		.shift		= 0,				\
 103	}
 104
 105static struct tk_fast tk_fast_mono ____cacheline_aligned = {
 106	.seq     = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
 107	.base[0] = FAST_TK_INIT,
 108	.base[1] = FAST_TK_INIT,
 109};
 110
 111static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
 112	.seq     = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
 113	.base[0] = FAST_TK_INIT,
 114	.base[1] = FAST_TK_INIT,
 115};
 116
 
 
 
 117static inline void tk_normalize_xtime(struct timekeeper *tk)
 118{
 119	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
 120		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
 121		tk->xtime_sec++;
 122	}
 123	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
 124		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
 125		tk->raw_sec++;
 126	}
 127}
 128
 129static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
 130{
 131	struct timespec64 ts;
 132
 133	ts.tv_sec = tk->xtime_sec;
 134	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 135	return ts;
 136}
 137
 138static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
 139{
 140	tk->xtime_sec = ts->tv_sec;
 141	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
 142}
 143
 144static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
 145{
 146	tk->xtime_sec += ts->tv_sec;
 147	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
 148	tk_normalize_xtime(tk);
 149}
 150
 151static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
 152{
 153	struct timespec64 tmp;
 154
 155	/*
 156	 * Verify consistency of: offset_real = -wall_to_monotonic
 157	 * before modifying anything
 158	 */
 159	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
 160					-tk->wall_to_monotonic.tv_nsec);
 161	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
 162	tk->wall_to_monotonic = wtm;
 163	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
 164	tk->offs_real = timespec64_to_ktime(tmp);
 165	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
 166}
 167
 168static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
 169{
 170	tk->offs_boot = ktime_add(tk->offs_boot, delta);
 171	/*
 172	 * Timespec representation for VDSO update to avoid 64bit division
 173	 * on every update.
 174	 */
 175	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
 176}
 177
 178/*
 179 * tk_clock_read - atomic clocksource read() helper
 180 *
 181 * This helper is necessary to use in the read paths because, while the
 182 * seqcount ensures we don't return a bad value while structures are updated,
 183 * it doesn't protect from potential crashes. There is the possibility that
 184 * the tkr's clocksource may change between the read reference, and the
 185 * clock reference passed to the read function.  This can cause crashes if
 186 * the wrong clocksource is passed to the wrong read function.
 187 * This isn't necessary to use when holding the timekeeper_lock or doing
 188 * a read of the fast-timekeeper tkrs (which is protected by its own locking
 189 * and update logic).
 190 */
 191static inline u64 tk_clock_read(const struct tk_read_base *tkr)
 192{
 193	struct clocksource *clock = READ_ONCE(tkr->clock);
 194
 195	return clock->read(clock);
 196}
 197
 198#ifdef CONFIG_DEBUG_TIMEKEEPING
 199#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
 200
 201static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 202{
 203
 204	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
 205	const char *name = tk->tkr_mono.clock->name;
 206
 207	if (offset > max_cycles) {
 208		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
 209				offset, name, max_cycles);
 210		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
 211	} else {
 212		if (offset > (max_cycles >> 1)) {
 213			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
 214					offset, name, max_cycles >> 1);
 215			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
 216		}
 217	}
 218
 219	if (tk->underflow_seen) {
 220		if (jiffies - tk->last_warning > WARNING_FREQ) {
 221			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
 222			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 223			printk_deferred("         Your kernel is probably still fine.\n");
 224			tk->last_warning = jiffies;
 225		}
 226		tk->underflow_seen = 0;
 227	}
 228
 229	if (tk->overflow_seen) {
 230		if (jiffies - tk->last_warning > WARNING_FREQ) {
 231			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
 232			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 233			printk_deferred("         Your kernel is probably still fine.\n");
 234			tk->last_warning = jiffies;
 235		}
 236		tk->overflow_seen = 0;
 237	}
 238}
 239
 240static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 241{
 242	struct timekeeper *tk = &tk_core.timekeeper;
 243	u64 now, last, mask, max, delta;
 244	unsigned int seq;
 245
 246	/*
 247	 * Since we're called holding a seqcount, the data may shift
 248	 * under us while we're doing the calculation. This can cause
 249	 * false positives, since we'd note a problem but throw the
 250	 * results away. So nest another seqcount here to atomically
 251	 * grab the points we are checking with.
 252	 */
 253	do {
 254		seq = read_seqcount_begin(&tk_core.seq);
 255		now = tk_clock_read(tkr);
 256		last = tkr->cycle_last;
 257		mask = tkr->mask;
 258		max = tkr->clock->max_cycles;
 259	} while (read_seqcount_retry(&tk_core.seq, seq));
 260
 261	delta = clocksource_delta(now, last, mask);
 262
 263	/*
 264	 * Try to catch underflows by checking if we are seeing small
 265	 * mask-relative negative values.
 266	 */
 267	if (unlikely((~delta & mask) < (mask >> 3))) {
 268		tk->underflow_seen = 1;
 269		delta = 0;
 270	}
 271
 272	/* Cap delta value to the max_cycles values to avoid mult overflows */
 273	if (unlikely(delta > max)) {
 274		tk->overflow_seen = 1;
 275		delta = tkr->clock->max_cycles;
 276	}
 277
 278	return delta;
 279}
 280#else
 281static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 282{
 283}
 284static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 285{
 286	u64 cycle_now, delta;
 287
 288	/* read clocksource */
 289	cycle_now = tk_clock_read(tkr);
 290
 291	/* calculate the delta since the last update_wall_time */
 292	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 293
 294	return delta;
 295}
 296#endif
 297
 298/**
 299 * tk_setup_internals - Set up internals to use clocksource clock.
 300 *
 301 * @tk:		The target timekeeper to setup.
 302 * @clock:		Pointer to clocksource.
 303 *
 304 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
 305 * pair and interval request.
 306 *
 307 * Unless you're the timekeeping code, you should not be using this!
 308 */
 309static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
 310{
 311	u64 interval;
 312	u64 tmp, ntpinterval;
 313	struct clocksource *old_clock;
 314
 315	++tk->cs_was_changed_seq;
 316	old_clock = tk->tkr_mono.clock;
 317	tk->tkr_mono.clock = clock;
 318	tk->tkr_mono.mask = clock->mask;
 319	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
 320
 321	tk->tkr_raw.clock = clock;
 322	tk->tkr_raw.mask = clock->mask;
 323	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
 324
 325	/* Do the ns -> cycle conversion first, using original mult */
 326	tmp = NTP_INTERVAL_LENGTH;
 327	tmp <<= clock->shift;
 328	ntpinterval = tmp;
 329	tmp += clock->mult/2;
 330	do_div(tmp, clock->mult);
 331	if (tmp == 0)
 332		tmp = 1;
 333
 334	interval = (u64) tmp;
 335	tk->cycle_interval = interval;
 336
 337	/* Go back from cycles -> shifted ns */
 338	tk->xtime_interval = interval * clock->mult;
 339	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
 340	tk->raw_interval = interval * clock->mult;
 341
 342	 /* if changing clocks, convert xtime_nsec shift units */
 343	if (old_clock) {
 344		int shift_change = clock->shift - old_clock->shift;
 345		if (shift_change < 0) {
 346			tk->tkr_mono.xtime_nsec >>= -shift_change;
 347			tk->tkr_raw.xtime_nsec >>= -shift_change;
 348		} else {
 349			tk->tkr_mono.xtime_nsec <<= shift_change;
 350			tk->tkr_raw.xtime_nsec <<= shift_change;
 351		}
 352	}
 353
 354	tk->tkr_mono.shift = clock->shift;
 355	tk->tkr_raw.shift = clock->shift;
 356
 357	tk->ntp_error = 0;
 358	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 359	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
 360
 361	/*
 362	 * The timekeeper keeps its own mult values for the currently
 363	 * active clocksource. These value will be adjusted via NTP
 364	 * to counteract clock drifting.
 365	 */
 366	tk->tkr_mono.mult = clock->mult;
 367	tk->tkr_raw.mult = clock->mult;
 368	tk->ntp_err_mult = 0;
 369	tk->skip_second_overflow = 0;
 370}
 371
 372/* Timekeeper helper functions. */
 373
 
 
 
 
 
 
 
 374static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
 375{
 376	u64 nsec;
 377
 378	nsec = delta * tkr->mult + tkr->xtime_nsec;
 379	nsec >>= tkr->shift;
 380
 381	return nsec;
 
 382}
 383
 384static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
 385{
 386	u64 delta;
 387
 388	delta = timekeeping_get_delta(tkr);
 389	return timekeeping_delta_to_ns(tkr, delta);
 390}
 391
 392static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
 393{
 394	u64 delta;
 395
 396	/* calculate the delta since the last update_wall_time */
 397	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 398	return timekeeping_delta_to_ns(tkr, delta);
 399}
 400
 401/**
 402 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
 403 * @tkr: Timekeeping readout base from which we take the update
 404 * @tkf: Pointer to NMI safe timekeeper
 405 *
 406 * We want to use this from any context including NMI and tracing /
 407 * instrumenting the timekeeping code itself.
 408 *
 409 * Employ the latch technique; see @raw_write_seqcount_latch.
 410 *
 411 * So if a NMI hits the update of base[0] then it will use base[1]
 412 * which is still consistent. In the worst case this can result is a
 413 * slightly wrong timestamp (a few nanoseconds). See
 414 * @ktime_get_mono_fast_ns.
 415 */
 416static void update_fast_timekeeper(const struct tk_read_base *tkr,
 417				   struct tk_fast *tkf)
 418{
 419	struct tk_read_base *base = tkf->base;
 420
 421	/* Force readers off to base[1] */
 422	raw_write_seqcount_latch(&tkf->seq);
 423
 424	/* Update base[0] */
 425	memcpy(base, tkr, sizeof(*base));
 426
 427	/* Force readers back to base[0] */
 428	raw_write_seqcount_latch(&tkf->seq);
 429
 430	/* Update base[1] */
 431	memcpy(base + 1, base, sizeof(*base));
 432}
 433
 434static __always_inline u64 fast_tk_get_delta_ns(struct tk_read_base *tkr)
 435{
 436	u64 delta, cycles = tk_clock_read(tkr);
 437
 438	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 439	return timekeeping_delta_to_ns(tkr, delta);
 440}
 441
 442static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
 443{
 444	struct tk_read_base *tkr;
 445	unsigned int seq;
 446	u64 now;
 447
 448	do {
 449		seq = raw_read_seqcount_latch(&tkf->seq);
 450		tkr = tkf->base + (seq & 0x01);
 451		now = ktime_to_ns(tkr->base);
 452		now += fast_tk_get_delta_ns(tkr);
 453	} while (read_seqcount_latch_retry(&tkf->seq, seq));
 454
 455	return now;
 456}
 457
 458/**
 459 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
 460 *
 461 * This timestamp is not guaranteed to be monotonic across an update.
 462 * The timestamp is calculated by:
 463 *
 464 *	now = base_mono + clock_delta * slope
 465 *
 466 * So if the update lowers the slope, readers who are forced to the
 467 * not yet updated second array are still using the old steeper slope.
 468 *
 469 * tmono
 470 * ^
 471 * |    o  n
 472 * |   o n
 473 * |  u
 474 * | o
 475 * |o
 476 * |12345678---> reader order
 477 *
 478 * o = old slope
 479 * u = update
 480 * n = new slope
 481 *
 482 * So reader 6 will observe time going backwards versus reader 5.
 483 *
 484 * While other CPUs are likely to be able to observe that, the only way
 485 * for a CPU local observation is when an NMI hits in the middle of
 486 * the update. Timestamps taken from that NMI context might be ahead
 487 * of the following timestamps. Callers need to be aware of that and
 488 * deal with it.
 489 */
 490u64 notrace ktime_get_mono_fast_ns(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 491{
 492	return __ktime_get_fast_ns(&tk_fast_mono);
 493}
 494EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 495
 496/**
 497 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
 498 *
 499 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
 500 * conversion factor is not affected by NTP/PTP correction.
 501 */
 502u64 notrace ktime_get_raw_fast_ns(void)
 503{
 504	return __ktime_get_fast_ns(&tk_fast_raw);
 505}
 506EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
 507
 508/**
 509 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
 510 *
 511 * To keep it NMI safe since we're accessing from tracing, we're not using a
 512 * separate timekeeper with updates to monotonic clock and boot offset
 513 * protected with seqcounts. This has the following minor side effects:
 514 *
 515 * (1) Its possible that a timestamp be taken after the boot offset is updated
 516 * but before the timekeeper is updated. If this happens, the new boot offset
 517 * is added to the old timekeeping making the clock appear to update slightly
 518 * earlier:
 519 *    CPU 0                                        CPU 1
 520 *    timekeeping_inject_sleeptime64()
 521 *    __timekeeping_inject_sleeptime(tk, delta);
 522 *                                                 timestamp();
 523 *    timekeeping_update(tk, TK_CLEAR_NTP...);
 524 *
 525 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
 526 * partially updated.  Since the tk->offs_boot update is a rare event, this
 527 * should be a rare occurrence which postprocessing should be able to handle.
 528 *
 529 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns()
 530 * apply as well.
 531 */
 532u64 notrace ktime_get_boot_fast_ns(void)
 533{
 534	struct timekeeper *tk = &tk_core.timekeeper;
 535
 536	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
 537}
 538EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
 539
 540/**
 541 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
 542 *
 543 * The same limitations as described for ktime_get_boot_fast_ns() apply. The
 544 * mono time and the TAI offset are not read atomically which may yield wrong
 545 * readouts. However, an update of the TAI offset is an rare event e.g., caused
 546 * by settime or adjtimex with an offset. The user of this function has to deal
 547 * with the possibility of wrong timestamps in post processing.
 548 */
 549u64 notrace ktime_get_tai_fast_ns(void)
 550{
 551	struct timekeeper *tk = &tk_core.timekeeper;
 552
 553	return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
 554}
 555EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
 556
 557static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
 
 
 
 558{
 559	struct tk_read_base *tkr;
 560	u64 basem, baser, delta;
 561	unsigned int seq;
 
 562
 563	do {
 564		seq = raw_read_seqcount_latch(&tkf->seq);
 565		tkr = tkf->base + (seq & 0x01);
 566		basem = ktime_to_ns(tkr->base);
 567		baser = ktime_to_ns(tkr->base_real);
 568		delta = fast_tk_get_delta_ns(tkr);
 569	} while (read_seqcount_latch_retry(&tkf->seq, seq));
 570
 571	if (mono)
 572		*mono = basem + delta;
 573	return baser + delta;
 
 
 574}
 575
 576/**
 577 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
 578 *
 579 * See ktime_get_fast_ns() for documentation of the time stamp ordering.
 580 */
 581u64 ktime_get_real_fast_ns(void)
 582{
 583	return __ktime_get_real_fast(&tk_fast_mono, NULL);
 584}
 585EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
 586
 587/**
 588 * ktime_get_fast_timestamps: - NMI safe timestamps
 589 * @snapshot:	Pointer to timestamp storage
 590 *
 591 * Stores clock monotonic, boottime and realtime timestamps.
 592 *
 593 * Boot time is a racy access on 32bit systems if the sleep time injection
 594 * happens late during resume and not in timekeeping_resume(). That could
 595 * be avoided by expanding struct tk_read_base with boot offset for 32bit
 596 * and adding more overhead to the update. As this is a hard to observe
 597 * once per resume event which can be filtered with reasonable effort using
 598 * the accurate mono/real timestamps, it's probably not worth the trouble.
 599 *
 600 * Aside of that it might be possible on 32 and 64 bit to observe the
 601 * following when the sleep time injection happens late:
 602 *
 603 * CPU 0				CPU 1
 604 * timekeeping_resume()
 605 * ktime_get_fast_timestamps()
 606 *	mono, real = __ktime_get_real_fast()
 607 *					inject_sleep_time()
 608 *					   update boot offset
 609 *	boot = mono + bootoffset;
 610 *
 611 * That means that boot time already has the sleep time adjustment, but
 612 * real time does not. On the next readout both are in sync again.
 613 *
 614 * Preventing this for 64bit is not really feasible without destroying the
 615 * careful cache layout of the timekeeper because the sequence count and
 616 * struct tk_read_base would then need two cache lines instead of one.
 617 *
 618 * Access to the time keeper clock source is disabled across the innermost
 619 * steps of suspend/resume. The accessors still work, but the timestamps
 620 * are frozen until time keeping is resumed which happens very early.
 621 *
 622 * For regular suspend/resume there is no observable difference vs. sched
 623 * clock, but it might affect some of the nasty low level debug printks.
 624 *
 625 * OTOH, access to sched clock is not guaranteed across suspend/resume on
 626 * all systems either so it depends on the hardware in use.
 627 *
 628 * If that turns out to be a real problem then this could be mitigated by
 629 * using sched clock in a similar way as during early boot. But it's not as
 630 * trivial as on early boot because it needs some careful protection
 631 * against the clock monotonic timestamp jumping backwards on resume.
 632 */
 633void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
 634{
 635	struct timekeeper *tk = &tk_core.timekeeper;
 636
 637	snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
 638	snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
 639}
 640
 641/**
 642 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
 643 * @tk: Timekeeper to snapshot.
 644 *
 645 * It generally is unsafe to access the clocksource after timekeeping has been
 646 * suspended, so take a snapshot of the readout base of @tk and use it as the
 647 * fast timekeeper's readout base while suspended.  It will return the same
 648 * number of cycles every time until timekeeping is resumed at which time the
 649 * proper readout base for the fast timekeeper will be restored automatically.
 650 */
 651static void halt_fast_timekeeper(const struct timekeeper *tk)
 652{
 653	static struct tk_read_base tkr_dummy;
 654	const struct tk_read_base *tkr = &tk->tkr_mono;
 655
 656	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 657	cycles_at_suspend = tk_clock_read(tkr);
 658	tkr_dummy.clock = &dummy_clock;
 659	tkr_dummy.base_real = tkr->base + tk->offs_real;
 660	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
 661
 662	tkr = &tk->tkr_raw;
 663	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 664	tkr_dummy.clock = &dummy_clock;
 665	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
 666}
 667
 668static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 669
 670static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 671{
 672	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 673}
 674
 675/**
 676 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 677 * @nb: Pointer to the notifier block to register
 678 */
 679int pvclock_gtod_register_notifier(struct notifier_block *nb)
 680{
 681	struct timekeeper *tk = &tk_core.timekeeper;
 682	unsigned long flags;
 683	int ret;
 684
 685	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 686	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 687	update_pvclock_gtod(tk, true);
 688	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 689
 690	return ret;
 691}
 692EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 693
 694/**
 695 * pvclock_gtod_unregister_notifier - unregister a pvclock
 696 * timedata update listener
 697 * @nb: Pointer to the notifier block to unregister
 698 */
 699int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 700{
 701	unsigned long flags;
 702	int ret;
 703
 704	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 705	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 706	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 707
 708	return ret;
 709}
 710EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 711
 712/*
 713 * tk_update_leap_state - helper to update the next_leap_ktime
 714 */
 715static inline void tk_update_leap_state(struct timekeeper *tk)
 716{
 717	tk->next_leap_ktime = ntp_get_next_leap();
 718	if (tk->next_leap_ktime != KTIME_MAX)
 719		/* Convert to monotonic time */
 720		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 721}
 722
 723/*
 724 * Update the ktime_t based scalar nsec members of the timekeeper
 725 */
 726static inline void tk_update_ktime_data(struct timekeeper *tk)
 727{
 728	u64 seconds;
 729	u32 nsec;
 730
 731	/*
 732	 * The xtime based monotonic readout is:
 733	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 734	 * The ktime based monotonic readout is:
 735	 *	nsec = base_mono + now();
 736	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 737	 */
 738	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 739	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 740	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 741
 742	/*
 743	 * The sum of the nanoseconds portions of xtime and
 744	 * wall_to_monotonic can be greater/equal one second. Take
 745	 * this into account before updating tk->ktime_sec.
 746	 */
 747	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 748	if (nsec >= NSEC_PER_SEC)
 749		seconds++;
 750	tk->ktime_sec = seconds;
 751
 752	/* Update the monotonic raw base */
 753	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
 754}
 755
 756/* must hold timekeeper_lock */
 757static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 758{
 759	if (action & TK_CLEAR_NTP) {
 760		tk->ntp_error = 0;
 761		ntp_clear();
 762	}
 763
 764	tk_update_leap_state(tk);
 765	tk_update_ktime_data(tk);
 766
 767	update_vsyscall(tk);
 768	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 769
 770	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
 771	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 772	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 773
 774	if (action & TK_CLOCK_WAS_SET)
 775		tk->clock_was_set_seq++;
 776	/*
 777	 * The mirroring of the data to the shadow-timekeeper needs
 778	 * to happen last here to ensure we don't over-write the
 779	 * timekeeper structure on the next update with stale data
 780	 */
 781	if (action & TK_MIRROR)
 782		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 783		       sizeof(tk_core.timekeeper));
 784}
 785
 786/**
 787 * timekeeping_forward_now - update clock to the current time
 788 * @tk:		Pointer to the timekeeper to update
 789 *
 790 * Forward the current clock to update its state since the last call to
 791 * update_wall_time(). This is useful before significant clock changes,
 792 * as it avoids having to deal with this time offset explicitly.
 793 */
 794static void timekeeping_forward_now(struct timekeeper *tk)
 795{
 796	u64 cycle_now, delta;
 797
 798	cycle_now = tk_clock_read(&tk->tkr_mono);
 799	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 800	tk->tkr_mono.cycle_last = cycle_now;
 801	tk->tkr_raw.cycle_last  = cycle_now;
 802
 803	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 
 
 
 
 
 804	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
 805
 
 
 
 806	tk_normalize_xtime(tk);
 807}
 808
 809/**
 810 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
 811 * @ts:		pointer to the timespec to be set
 812 *
 813 * Returns the time of day in a timespec64 (WARN if suspended).
 814 */
 815void ktime_get_real_ts64(struct timespec64 *ts)
 816{
 817	struct timekeeper *tk = &tk_core.timekeeper;
 818	unsigned int seq;
 819	u64 nsecs;
 820
 821	WARN_ON(timekeeping_suspended);
 822
 823	do {
 824		seq = read_seqcount_begin(&tk_core.seq);
 825
 826		ts->tv_sec = tk->xtime_sec;
 827		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 828
 829	} while (read_seqcount_retry(&tk_core.seq, seq));
 830
 831	ts->tv_nsec = 0;
 832	timespec64_add_ns(ts, nsecs);
 833}
 834EXPORT_SYMBOL(ktime_get_real_ts64);
 835
 836ktime_t ktime_get(void)
 837{
 838	struct timekeeper *tk = &tk_core.timekeeper;
 839	unsigned int seq;
 840	ktime_t base;
 841	u64 nsecs;
 842
 843	WARN_ON(timekeeping_suspended);
 844
 845	do {
 846		seq = read_seqcount_begin(&tk_core.seq);
 847		base = tk->tkr_mono.base;
 848		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 849
 850	} while (read_seqcount_retry(&tk_core.seq, seq));
 851
 852	return ktime_add_ns(base, nsecs);
 853}
 854EXPORT_SYMBOL_GPL(ktime_get);
 855
 856u32 ktime_get_resolution_ns(void)
 857{
 858	struct timekeeper *tk = &tk_core.timekeeper;
 859	unsigned int seq;
 860	u32 nsecs;
 861
 862	WARN_ON(timekeeping_suspended);
 863
 864	do {
 865		seq = read_seqcount_begin(&tk_core.seq);
 866		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 867	} while (read_seqcount_retry(&tk_core.seq, seq));
 868
 869	return nsecs;
 870}
 871EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 872
 873static ktime_t *offsets[TK_OFFS_MAX] = {
 874	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
 875	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
 876	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
 877};
 878
 879ktime_t ktime_get_with_offset(enum tk_offsets offs)
 880{
 881	struct timekeeper *tk = &tk_core.timekeeper;
 882	unsigned int seq;
 883	ktime_t base, *offset = offsets[offs];
 884	u64 nsecs;
 885
 886	WARN_ON(timekeeping_suspended);
 887
 888	do {
 889		seq = read_seqcount_begin(&tk_core.seq);
 890		base = ktime_add(tk->tkr_mono.base, *offset);
 891		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 892
 893	} while (read_seqcount_retry(&tk_core.seq, seq));
 894
 895	return ktime_add_ns(base, nsecs);
 896
 897}
 898EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 899
 900ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
 901{
 902	struct timekeeper *tk = &tk_core.timekeeper;
 903	unsigned int seq;
 904	ktime_t base, *offset = offsets[offs];
 905	u64 nsecs;
 906
 907	WARN_ON(timekeeping_suspended);
 908
 909	do {
 910		seq = read_seqcount_begin(&tk_core.seq);
 911		base = ktime_add(tk->tkr_mono.base, *offset);
 912		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 913
 914	} while (read_seqcount_retry(&tk_core.seq, seq));
 915
 916	return ktime_add_ns(base, nsecs);
 917}
 918EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
 919
 920/**
 921 * ktime_mono_to_any() - convert monotonic time to any other time
 922 * @tmono:	time to convert.
 923 * @offs:	which offset to use
 924 */
 925ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 926{
 927	ktime_t *offset = offsets[offs];
 928	unsigned int seq;
 929	ktime_t tconv;
 930
 931	do {
 932		seq = read_seqcount_begin(&tk_core.seq);
 933		tconv = ktime_add(tmono, *offset);
 934	} while (read_seqcount_retry(&tk_core.seq, seq));
 935
 936	return tconv;
 937}
 938EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 939
 940/**
 941 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 942 */
 943ktime_t ktime_get_raw(void)
 944{
 945	struct timekeeper *tk = &tk_core.timekeeper;
 946	unsigned int seq;
 947	ktime_t base;
 948	u64 nsecs;
 949
 950	do {
 951		seq = read_seqcount_begin(&tk_core.seq);
 952		base = tk->tkr_raw.base;
 953		nsecs = timekeeping_get_ns(&tk->tkr_raw);
 954
 955	} while (read_seqcount_retry(&tk_core.seq, seq));
 956
 957	return ktime_add_ns(base, nsecs);
 958}
 959EXPORT_SYMBOL_GPL(ktime_get_raw);
 960
 961/**
 962 * ktime_get_ts64 - get the monotonic clock in timespec64 format
 963 * @ts:		pointer to timespec variable
 964 *
 965 * The function calculates the monotonic clock from the realtime
 966 * clock and the wall_to_monotonic offset and stores the result
 967 * in normalized timespec64 format in the variable pointed to by @ts.
 968 */
 969void ktime_get_ts64(struct timespec64 *ts)
 970{
 971	struct timekeeper *tk = &tk_core.timekeeper;
 972	struct timespec64 tomono;
 973	unsigned int seq;
 974	u64 nsec;
 975
 976	WARN_ON(timekeeping_suspended);
 977
 978	do {
 979		seq = read_seqcount_begin(&tk_core.seq);
 980		ts->tv_sec = tk->xtime_sec;
 981		nsec = timekeeping_get_ns(&tk->tkr_mono);
 982		tomono = tk->wall_to_monotonic;
 983
 984	} while (read_seqcount_retry(&tk_core.seq, seq));
 985
 986	ts->tv_sec += tomono.tv_sec;
 987	ts->tv_nsec = 0;
 988	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 989}
 990EXPORT_SYMBOL_GPL(ktime_get_ts64);
 991
 992/**
 993 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 994 *
 995 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 996 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 997 * works on both 32 and 64 bit systems. On 32 bit systems the readout
 998 * covers ~136 years of uptime which should be enough to prevent
 999 * premature wrap arounds.
1000 */
1001time64_t ktime_get_seconds(void)
1002{
1003	struct timekeeper *tk = &tk_core.timekeeper;
1004
1005	WARN_ON(timekeeping_suspended);
1006	return tk->ktime_sec;
1007}
1008EXPORT_SYMBOL_GPL(ktime_get_seconds);
1009
1010/**
1011 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1012 *
1013 * Returns the wall clock seconds since 1970.
 
1014 *
1015 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1016 * 32bit systems the access must be protected with the sequence
1017 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1018 * value.
1019 */
1020time64_t ktime_get_real_seconds(void)
1021{
1022	struct timekeeper *tk = &tk_core.timekeeper;
1023	time64_t seconds;
1024	unsigned int seq;
1025
1026	if (IS_ENABLED(CONFIG_64BIT))
1027		return tk->xtime_sec;
1028
1029	do {
1030		seq = read_seqcount_begin(&tk_core.seq);
1031		seconds = tk->xtime_sec;
1032
1033	} while (read_seqcount_retry(&tk_core.seq, seq));
1034
1035	return seconds;
1036}
1037EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1038
1039/**
1040 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1041 * but without the sequence counter protect. This internal function
1042 * is called just when timekeeping lock is already held.
1043 */
1044noinstr time64_t __ktime_get_real_seconds(void)
1045{
1046	struct timekeeper *tk = &tk_core.timekeeper;
1047
1048	return tk->xtime_sec;
1049}
1050
1051/**
1052 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1053 * @systime_snapshot:	pointer to struct receiving the system time snapshot
1054 */
1055void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1056{
1057	struct timekeeper *tk = &tk_core.timekeeper;
1058	unsigned int seq;
1059	ktime_t base_raw;
1060	ktime_t base_real;
1061	u64 nsec_raw;
1062	u64 nsec_real;
1063	u64 now;
1064
1065	WARN_ON_ONCE(timekeeping_suspended);
1066
1067	do {
1068		seq = read_seqcount_begin(&tk_core.seq);
1069		now = tk_clock_read(&tk->tkr_mono);
1070		systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1071		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1072		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1073		base_real = ktime_add(tk->tkr_mono.base,
1074				      tk_core.timekeeper.offs_real);
1075		base_raw = tk->tkr_raw.base;
1076		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1077		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1078	} while (read_seqcount_retry(&tk_core.seq, seq));
1079
1080	systime_snapshot->cycles = now;
1081	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1082	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1083}
1084EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1085
1086/* Scale base by mult/div checking for overflow */
1087static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1088{
1089	u64 tmp, rem;
1090
1091	tmp = div64_u64_rem(*base, div, &rem);
1092
1093	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1094	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1095		return -EOVERFLOW;
1096	tmp *= mult;
1097
1098	rem = div64_u64(rem * mult, div);
1099	*base = tmp + rem;
1100	return 0;
1101}
1102
1103/**
1104 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1105 * @history:			Snapshot representing start of history
1106 * @partial_history_cycles:	Cycle offset into history (fractional part)
1107 * @total_history_cycles:	Total history length in cycles
1108 * @discontinuity:		True indicates clock was set on history period
1109 * @ts:				Cross timestamp that should be adjusted using
1110 *	partial/total ratio
1111 *
1112 * Helper function used by get_device_system_crosststamp() to correct the
1113 * crosstimestamp corresponding to the start of the current interval to the
1114 * system counter value (timestamp point) provided by the driver. The
1115 * total_history_* quantities are the total history starting at the provided
1116 * reference point and ending at the start of the current interval. The cycle
1117 * count between the driver timestamp point and the start of the current
1118 * interval is partial_history_cycles.
1119 */
1120static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1121					 u64 partial_history_cycles,
1122					 u64 total_history_cycles,
1123					 bool discontinuity,
1124					 struct system_device_crosststamp *ts)
1125{
1126	struct timekeeper *tk = &tk_core.timekeeper;
1127	u64 corr_raw, corr_real;
1128	bool interp_forward;
1129	int ret;
1130
1131	if (total_history_cycles == 0 || partial_history_cycles == 0)
1132		return 0;
1133
1134	/* Interpolate shortest distance from beginning or end of history */
1135	interp_forward = partial_history_cycles > total_history_cycles / 2;
1136	partial_history_cycles = interp_forward ?
1137		total_history_cycles - partial_history_cycles :
1138		partial_history_cycles;
1139
1140	/*
1141	 * Scale the monotonic raw time delta by:
1142	 *	partial_history_cycles / total_history_cycles
1143	 */
1144	corr_raw = (u64)ktime_to_ns(
1145		ktime_sub(ts->sys_monoraw, history->raw));
1146	ret = scale64_check_overflow(partial_history_cycles,
1147				     total_history_cycles, &corr_raw);
1148	if (ret)
1149		return ret;
1150
1151	/*
1152	 * If there is a discontinuity in the history, scale monotonic raw
1153	 *	correction by:
1154	 *	mult(real)/mult(raw) yielding the realtime correction
1155	 * Otherwise, calculate the realtime correction similar to monotonic
1156	 *	raw calculation
1157	 */
1158	if (discontinuity) {
1159		corr_real = mul_u64_u32_div
1160			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1161	} else {
1162		corr_real = (u64)ktime_to_ns(
1163			ktime_sub(ts->sys_realtime, history->real));
1164		ret = scale64_check_overflow(partial_history_cycles,
1165					     total_history_cycles, &corr_real);
1166		if (ret)
1167			return ret;
1168	}
1169
1170	/* Fixup monotonic raw and real time time values */
1171	if (interp_forward) {
1172		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1173		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1174	} else {
1175		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1176		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1177	}
1178
1179	return 0;
1180}
1181
1182/*
1183 * cycle_between - true if test occurs chronologically between before and after
1184 */
1185static bool cycle_between(u64 before, u64 test, u64 after)
1186{
1187	if (test > before && test < after)
1188		return true;
1189	if (test < before && before > after)
1190		return true;
1191	return false;
1192}
1193
1194/**
1195 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1196 * @get_time_fn:	Callback to get simultaneous device time and
1197 *	system counter from the device driver
1198 * @ctx:		Context passed to get_time_fn()
1199 * @history_begin:	Historical reference point used to interpolate system
1200 *	time when counter provided by the driver is before the current interval
1201 * @xtstamp:		Receives simultaneously captured system and device time
1202 *
1203 * Reads a timestamp from a device and correlates it to system time
1204 */
1205int get_device_system_crosststamp(int (*get_time_fn)
1206				  (ktime_t *device_time,
1207				   struct system_counterval_t *sys_counterval,
1208				   void *ctx),
1209				  void *ctx,
1210				  struct system_time_snapshot *history_begin,
1211				  struct system_device_crosststamp *xtstamp)
1212{
1213	struct system_counterval_t system_counterval;
1214	struct timekeeper *tk = &tk_core.timekeeper;
1215	u64 cycles, now, interval_start;
1216	unsigned int clock_was_set_seq = 0;
1217	ktime_t base_real, base_raw;
1218	u64 nsec_real, nsec_raw;
1219	u8 cs_was_changed_seq;
1220	unsigned int seq;
1221	bool do_interp;
1222	int ret;
1223
1224	do {
1225		seq = read_seqcount_begin(&tk_core.seq);
1226		/*
1227		 * Try to synchronously capture device time and a system
1228		 * counter value calling back into the device driver
1229		 */
1230		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1231		if (ret)
1232			return ret;
1233
1234		/*
1235		 * Verify that the clocksource associated with the captured
1236		 * system counter value is the same as the currently installed
1237		 * timekeeper clocksource
1238		 */
1239		if (tk->tkr_mono.clock != system_counterval.cs)
1240			return -ENODEV;
1241		cycles = system_counterval.cycles;
1242
1243		/*
1244		 * Check whether the system counter value provided by the
1245		 * device driver is on the current timekeeping interval.
1246		 */
1247		now = tk_clock_read(&tk->tkr_mono);
1248		interval_start = tk->tkr_mono.cycle_last;
1249		if (!cycle_between(interval_start, cycles, now)) {
1250			clock_was_set_seq = tk->clock_was_set_seq;
1251			cs_was_changed_seq = tk->cs_was_changed_seq;
1252			cycles = interval_start;
1253			do_interp = true;
1254		} else {
1255			do_interp = false;
1256		}
1257
1258		base_real = ktime_add(tk->tkr_mono.base,
1259				      tk_core.timekeeper.offs_real);
1260		base_raw = tk->tkr_raw.base;
1261
1262		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1263						     system_counterval.cycles);
1264		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1265						    system_counterval.cycles);
1266	} while (read_seqcount_retry(&tk_core.seq, seq));
1267
1268	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1269	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1270
1271	/*
1272	 * Interpolate if necessary, adjusting back from the start of the
1273	 * current interval
1274	 */
1275	if (do_interp) {
1276		u64 partial_history_cycles, total_history_cycles;
1277		bool discontinuity;
1278
1279		/*
1280		 * Check that the counter value occurs after the provided
1281		 * history reference and that the history doesn't cross a
1282		 * clocksource change
1283		 */
1284		if (!history_begin ||
1285		    !cycle_between(history_begin->cycles,
1286				   system_counterval.cycles, cycles) ||
1287		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1288			return -EINVAL;
1289		partial_history_cycles = cycles - system_counterval.cycles;
1290		total_history_cycles = cycles - history_begin->cycles;
1291		discontinuity =
1292			history_begin->clock_was_set_seq != clock_was_set_seq;
1293
1294		ret = adjust_historical_crosststamp(history_begin,
1295						    partial_history_cycles,
1296						    total_history_cycles,
1297						    discontinuity, xtstamp);
1298		if (ret)
1299			return ret;
1300	}
1301
1302	return 0;
1303}
1304EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1305
1306/**
1307 * do_settimeofday64 - Sets the time of day.
1308 * @ts:     pointer to the timespec64 variable containing the new time
1309 *
1310 * Sets the time of day to the new time and update NTP and notify hrtimers
1311 */
1312int do_settimeofday64(const struct timespec64 *ts)
1313{
1314	struct timekeeper *tk = &tk_core.timekeeper;
1315	struct timespec64 ts_delta, xt;
1316	unsigned long flags;
1317	int ret = 0;
1318
1319	if (!timespec64_valid_settod(ts))
1320		return -EINVAL;
1321
1322	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1323	write_seqcount_begin(&tk_core.seq);
1324
1325	timekeeping_forward_now(tk);
1326
1327	xt = tk_xtime(tk);
1328	ts_delta = timespec64_sub(*ts, xt);
 
1329
1330	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1331		ret = -EINVAL;
1332		goto out;
1333	}
1334
1335	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1336
1337	tk_set_xtime(tk, ts);
1338out:
1339	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1340
1341	write_seqcount_end(&tk_core.seq);
1342	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1343
1344	/* Signal hrtimers about time change */
1345	clock_was_set(CLOCK_SET_WALL);
1346
1347	if (!ret) {
1348		audit_tk_injoffset(ts_delta);
1349		add_device_randomness(ts, sizeof(*ts));
1350	}
1351
1352	return ret;
1353}
1354EXPORT_SYMBOL(do_settimeofday64);
1355
1356/**
1357 * timekeeping_inject_offset - Adds or subtracts from the current time.
1358 * @ts:		Pointer to the timespec variable containing the offset
1359 *
1360 * Adds or subtracts an offset value from the current time.
1361 */
1362static int timekeeping_inject_offset(const struct timespec64 *ts)
1363{
1364	struct timekeeper *tk = &tk_core.timekeeper;
1365	unsigned long flags;
1366	struct timespec64 tmp;
1367	int ret = 0;
1368
1369	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1370		return -EINVAL;
1371
1372	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1373	write_seqcount_begin(&tk_core.seq);
1374
1375	timekeeping_forward_now(tk);
1376
1377	/* Make sure the proposed value is valid */
1378	tmp = timespec64_add(tk_xtime(tk), *ts);
1379	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1380	    !timespec64_valid_settod(&tmp)) {
1381		ret = -EINVAL;
1382		goto error;
1383	}
1384
1385	tk_xtime_add(tk, ts);
1386	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1387
1388error: /* even if we error out, we forwarded the time, so call update */
1389	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1390
1391	write_seqcount_end(&tk_core.seq);
1392	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1393
1394	/* Signal hrtimers about time change */
1395	clock_was_set(CLOCK_SET_WALL);
1396
1397	return ret;
1398}
1399
1400/*
1401 * Indicates if there is an offset between the system clock and the hardware
1402 * clock/persistent clock/rtc.
1403 */
1404int persistent_clock_is_local;
1405
1406/*
1407 * Adjust the time obtained from the CMOS to be UTC time instead of
1408 * local time.
1409 *
1410 * This is ugly, but preferable to the alternatives.  Otherwise we
1411 * would either need to write a program to do it in /etc/rc (and risk
1412 * confusion if the program gets run more than once; it would also be
1413 * hard to make the program warp the clock precisely n hours)  or
1414 * compile in the timezone information into the kernel.  Bad, bad....
1415 *
1416 *						- TYT, 1992-01-01
1417 *
1418 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1419 * as real UNIX machines always do it. This avoids all headaches about
1420 * daylight saving times and warping kernel clocks.
1421 */
1422void timekeeping_warp_clock(void)
1423{
1424	if (sys_tz.tz_minuteswest != 0) {
1425		struct timespec64 adjust;
1426
1427		persistent_clock_is_local = 1;
1428		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1429		adjust.tv_nsec = 0;
1430		timekeeping_inject_offset(&adjust);
1431	}
1432}
1433
1434/*
1435 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
 
1436 */
1437static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1438{
1439	tk->tai_offset = tai_offset;
1440	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1441}
1442
1443/*
1444 * change_clocksource - Swaps clocksources if a new one is available
1445 *
1446 * Accumulates current time interval and initializes new clocksource
1447 */
1448static int change_clocksource(void *data)
1449{
1450	struct timekeeper *tk = &tk_core.timekeeper;
1451	struct clocksource *new, *old = NULL;
1452	unsigned long flags;
1453	bool change = false;
1454
1455	new = (struct clocksource *) data;
1456
 
 
 
 
1457	/*
1458	 * If the cs is in module, get a module reference. Succeeds
1459	 * for built-in code (owner == NULL) as well.
1460	 */
1461	if (try_module_get(new->owner)) {
1462		if (!new->enable || new->enable(new) == 0)
1463			change = true;
1464		else
 
 
 
 
1465			module_put(new->owner);
 
1466	}
1467
1468	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1469	write_seqcount_begin(&tk_core.seq);
1470
1471	timekeeping_forward_now(tk);
1472
1473	if (change) {
1474		old = tk->tkr_mono.clock;
1475		tk_setup_internals(tk, new);
1476	}
1477
1478	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1479
1480	write_seqcount_end(&tk_core.seq);
1481	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1482
1483	if (old) {
1484		if (old->disable)
1485			old->disable(old);
1486
1487		module_put(old->owner);
1488	}
1489
1490	return 0;
1491}
1492
1493/**
1494 * timekeeping_notify - Install a new clock source
1495 * @clock:		pointer to the clock source
1496 *
1497 * This function is called from clocksource.c after a new, better clock
1498 * source has been registered. The caller holds the clocksource_mutex.
1499 */
1500int timekeeping_notify(struct clocksource *clock)
1501{
1502	struct timekeeper *tk = &tk_core.timekeeper;
1503
1504	if (tk->tkr_mono.clock == clock)
1505		return 0;
1506	stop_machine(change_clocksource, clock, NULL);
1507	tick_clock_notify();
1508	return tk->tkr_mono.clock == clock ? 0 : -1;
1509}
1510
1511/**
1512 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1513 * @ts:		pointer to the timespec64 to be set
1514 *
1515 * Returns the raw monotonic time (completely un-modified by ntp)
1516 */
1517void ktime_get_raw_ts64(struct timespec64 *ts)
1518{
1519	struct timekeeper *tk = &tk_core.timekeeper;
1520	unsigned int seq;
1521	u64 nsecs;
1522
1523	do {
1524		seq = read_seqcount_begin(&tk_core.seq);
1525		ts->tv_sec = tk->raw_sec;
1526		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1527
1528	} while (read_seqcount_retry(&tk_core.seq, seq));
1529
1530	ts->tv_nsec = 0;
1531	timespec64_add_ns(ts, nsecs);
1532}
1533EXPORT_SYMBOL(ktime_get_raw_ts64);
1534
1535
1536/**
1537 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1538 */
1539int timekeeping_valid_for_hres(void)
1540{
1541	struct timekeeper *tk = &tk_core.timekeeper;
1542	unsigned int seq;
1543	int ret;
1544
1545	do {
1546		seq = read_seqcount_begin(&tk_core.seq);
1547
1548		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1549
1550	} while (read_seqcount_retry(&tk_core.seq, seq));
1551
1552	return ret;
1553}
1554
1555/**
1556 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1557 */
1558u64 timekeeping_max_deferment(void)
1559{
1560	struct timekeeper *tk = &tk_core.timekeeper;
1561	unsigned int seq;
1562	u64 ret;
1563
1564	do {
1565		seq = read_seqcount_begin(&tk_core.seq);
1566
1567		ret = tk->tkr_mono.clock->max_idle_ns;
1568
1569	} while (read_seqcount_retry(&tk_core.seq, seq));
1570
1571	return ret;
1572}
1573
1574/**
1575 * read_persistent_clock64 -  Return time from the persistent clock.
1576 * @ts: Pointer to the storage for the readout value
1577 *
1578 * Weak dummy function for arches that do not yet support it.
1579 * Reads the time from the battery backed persistent clock.
1580 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1581 *
1582 *  XXX - Do be sure to remove it once all arches implement it.
1583 */
1584void __weak read_persistent_clock64(struct timespec64 *ts)
1585{
1586	ts->tv_sec = 0;
1587	ts->tv_nsec = 0;
1588}
1589
1590/**
1591 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1592 *                                        from the boot.
1593 * @wall_time:	  current time as returned by persistent clock
1594 * @boot_offset:  offset that is defined as wall_time - boot_time
1595 *
1596 * Weak dummy function for arches that do not yet support it.
1597 *
 
1598 * The default function calculates offset based on the current value of
1599 * local_clock(). This way architectures that support sched_clock() but don't
1600 * support dedicated boot time clock will provide the best estimate of the
1601 * boot time.
1602 */
1603void __weak __init
1604read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1605				     struct timespec64 *boot_offset)
1606{
1607	read_persistent_clock64(wall_time);
1608	*boot_offset = ns_to_timespec64(local_clock());
1609}
1610
1611/*
1612 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1613 *
1614 * The flag starts of false and is only set when a suspend reaches
1615 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1616 * timekeeper clocksource is not stopping across suspend and has been
1617 * used to update sleep time. If the timekeeper clocksource has stopped
1618 * then the flag stays true and is used by the RTC resume code to decide
1619 * whether sleeptime must be injected and if so the flag gets false then.
1620 *
1621 * If a suspend fails before reaching timekeeping_resume() then the flag
1622 * stays false and prevents erroneous sleeptime injection.
1623 */
1624static bool suspend_timing_needed;
1625
1626/* Flag for if there is a persistent clock on this platform */
1627static bool persistent_clock_exists;
1628
1629/*
1630 * timekeeping_init - Initializes the clocksource and common timekeeping values
1631 */
1632void __init timekeeping_init(void)
1633{
1634	struct timespec64 wall_time, boot_offset, wall_to_mono;
1635	struct timekeeper *tk = &tk_core.timekeeper;
1636	struct clocksource *clock;
1637	unsigned long flags;
1638
1639	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1640	if (timespec64_valid_settod(&wall_time) &&
1641	    timespec64_to_ns(&wall_time) > 0) {
1642		persistent_clock_exists = true;
1643	} else if (timespec64_to_ns(&wall_time) != 0) {
1644		pr_warn("Persistent clock returned invalid value");
1645		wall_time = (struct timespec64){0};
1646	}
1647
1648	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1649		boot_offset = (struct timespec64){0};
1650
1651	/*
1652	 * We want set wall_to_mono, so the following is true:
1653	 * wall time + wall_to_mono = boot time
1654	 */
1655	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1656
1657	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1658	write_seqcount_begin(&tk_core.seq);
1659	ntp_init();
1660
1661	clock = clocksource_default_clock();
1662	if (clock->enable)
1663		clock->enable(clock);
1664	tk_setup_internals(tk, clock);
1665
1666	tk_set_xtime(tk, &wall_time);
1667	tk->raw_sec = 0;
1668
1669	tk_set_wall_to_mono(tk, wall_to_mono);
1670
1671	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1672
1673	write_seqcount_end(&tk_core.seq);
1674	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1675}
1676
1677/* time in seconds when suspend began for persistent clock */
1678static struct timespec64 timekeeping_suspend_time;
1679
1680/**
1681 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1682 * @tk:		Pointer to the timekeeper to be updated
1683 * @delta:	Pointer to the delta value in timespec64 format
1684 *
1685 * Takes a timespec offset measuring a suspend interval and properly
1686 * adds the sleep offset to the timekeeping variables.
1687 */
1688static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1689					   const struct timespec64 *delta)
1690{
1691	if (!timespec64_valid_strict(delta)) {
1692		printk_deferred(KERN_WARNING
1693				"__timekeeping_inject_sleeptime: Invalid "
1694				"sleep delta value!\n");
1695		return;
1696	}
1697	tk_xtime_add(tk, delta);
1698	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1699	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1700	tk_debug_account_sleep_time(delta);
1701}
1702
1703#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1704/*
1705 * We have three kinds of time sources to use for sleep time
1706 * injection, the preference order is:
1707 * 1) non-stop clocksource
1708 * 2) persistent clock (ie: RTC accessible when irqs are off)
1709 * 3) RTC
1710 *
1711 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1712 * If system has neither 1) nor 2), 3) will be used finally.
1713 *
1714 *
1715 * If timekeeping has injected sleeptime via either 1) or 2),
1716 * 3) becomes needless, so in this case we don't need to call
1717 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1718 * means.
1719 */
1720bool timekeeping_rtc_skipresume(void)
1721{
1722	return !suspend_timing_needed;
1723}
1724
1725/*
1726 * 1) can be determined whether to use or not only when doing
1727 * timekeeping_resume() which is invoked after rtc_suspend(),
1728 * so we can't skip rtc_suspend() surely if system has 1).
1729 *
1730 * But if system has 2), 2) will definitely be used, so in this
1731 * case we don't need to call rtc_suspend(), and this is what
1732 * timekeeping_rtc_skipsuspend() means.
1733 */
1734bool timekeeping_rtc_skipsuspend(void)
1735{
1736	return persistent_clock_exists;
1737}
1738
1739/**
1740 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1741 * @delta: pointer to a timespec64 delta value
1742 *
1743 * This hook is for architectures that cannot support read_persistent_clock64
1744 * because their RTC/persistent clock is only accessible when irqs are enabled.
1745 * and also don't have an effective nonstop clocksource.
1746 *
1747 * This function should only be called by rtc_resume(), and allows
1748 * a suspend offset to be injected into the timekeeping values.
1749 */
1750void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1751{
1752	struct timekeeper *tk = &tk_core.timekeeper;
1753	unsigned long flags;
1754
1755	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1756	write_seqcount_begin(&tk_core.seq);
1757
1758	suspend_timing_needed = false;
1759
1760	timekeeping_forward_now(tk);
1761
1762	__timekeeping_inject_sleeptime(tk, delta);
1763
1764	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1765
1766	write_seqcount_end(&tk_core.seq);
1767	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1768
1769	/* Signal hrtimers about time change */
1770	clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1771}
1772#endif
1773
1774/**
1775 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1776 */
1777void timekeeping_resume(void)
1778{
1779	struct timekeeper *tk = &tk_core.timekeeper;
1780	struct clocksource *clock = tk->tkr_mono.clock;
1781	unsigned long flags;
1782	struct timespec64 ts_new, ts_delta;
1783	u64 cycle_now, nsec;
1784	bool inject_sleeptime = false;
1785
1786	read_persistent_clock64(&ts_new);
1787
1788	clockevents_resume();
1789	clocksource_resume();
1790
1791	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1792	write_seqcount_begin(&tk_core.seq);
1793
1794	/*
1795	 * After system resumes, we need to calculate the suspended time and
1796	 * compensate it for the OS time. There are 3 sources that could be
1797	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1798	 * device.
1799	 *
1800	 * One specific platform may have 1 or 2 or all of them, and the
1801	 * preference will be:
1802	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1803	 * The less preferred source will only be tried if there is no better
1804	 * usable source. The rtc part is handled separately in rtc core code.
1805	 */
1806	cycle_now = tk_clock_read(&tk->tkr_mono);
1807	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1808	if (nsec > 0) {
1809		ts_delta = ns_to_timespec64(nsec);
1810		inject_sleeptime = true;
1811	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1812		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1813		inject_sleeptime = true;
1814	}
1815
1816	if (inject_sleeptime) {
1817		suspend_timing_needed = false;
1818		__timekeeping_inject_sleeptime(tk, &ts_delta);
1819	}
1820
1821	/* Re-base the last cycle value */
1822	tk->tkr_mono.cycle_last = cycle_now;
1823	tk->tkr_raw.cycle_last  = cycle_now;
1824
1825	tk->ntp_error = 0;
1826	timekeeping_suspended = 0;
1827	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1828	write_seqcount_end(&tk_core.seq);
1829	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1830
1831	touch_softlockup_watchdog();
1832
1833	/* Resume the clockevent device(s) and hrtimers */
1834	tick_resume();
1835	/* Notify timerfd as resume is equivalent to clock_was_set() */
1836	timerfd_resume();
1837}
1838
1839int timekeeping_suspend(void)
1840{
1841	struct timekeeper *tk = &tk_core.timekeeper;
1842	unsigned long flags;
1843	struct timespec64		delta, delta_delta;
1844	static struct timespec64	old_delta;
1845	struct clocksource *curr_clock;
1846	u64 cycle_now;
1847
1848	read_persistent_clock64(&timekeeping_suspend_time);
1849
1850	/*
1851	 * On some systems the persistent_clock can not be detected at
1852	 * timekeeping_init by its return value, so if we see a valid
1853	 * value returned, update the persistent_clock_exists flag.
1854	 */
1855	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1856		persistent_clock_exists = true;
1857
1858	suspend_timing_needed = true;
1859
1860	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1861	write_seqcount_begin(&tk_core.seq);
1862	timekeeping_forward_now(tk);
1863	timekeeping_suspended = 1;
1864
1865	/*
1866	 * Since we've called forward_now, cycle_last stores the value
1867	 * just read from the current clocksource. Save this to potentially
1868	 * use in suspend timing.
1869	 */
1870	curr_clock = tk->tkr_mono.clock;
1871	cycle_now = tk->tkr_mono.cycle_last;
1872	clocksource_start_suspend_timing(curr_clock, cycle_now);
1873
1874	if (persistent_clock_exists) {
1875		/*
1876		 * To avoid drift caused by repeated suspend/resumes,
1877		 * which each can add ~1 second drift error,
1878		 * try to compensate so the difference in system time
1879		 * and persistent_clock time stays close to constant.
1880		 */
1881		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1882		delta_delta = timespec64_sub(delta, old_delta);
1883		if (abs(delta_delta.tv_sec) >= 2) {
1884			/*
1885			 * if delta_delta is too large, assume time correction
1886			 * has occurred and set old_delta to the current delta.
1887			 */
1888			old_delta = delta;
1889		} else {
1890			/* Otherwise try to adjust old_system to compensate */
1891			timekeeping_suspend_time =
1892				timespec64_add(timekeeping_suspend_time, delta_delta);
1893		}
1894	}
1895
1896	timekeeping_update(tk, TK_MIRROR);
1897	halt_fast_timekeeper(tk);
1898	write_seqcount_end(&tk_core.seq);
1899	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1900
1901	tick_suspend();
1902	clocksource_suspend();
1903	clockevents_suspend();
1904
1905	return 0;
1906}
1907
1908/* sysfs resume/suspend bits for timekeeping */
1909static struct syscore_ops timekeeping_syscore_ops = {
1910	.resume		= timekeeping_resume,
1911	.suspend	= timekeeping_suspend,
1912};
1913
1914static int __init timekeeping_init_ops(void)
1915{
1916	register_syscore_ops(&timekeeping_syscore_ops);
1917	return 0;
1918}
1919device_initcall(timekeeping_init_ops);
1920
1921/*
1922 * Apply a multiplier adjustment to the timekeeper
1923 */
1924static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1925							 s64 offset,
1926							 s32 mult_adj)
1927{
1928	s64 interval = tk->cycle_interval;
1929
1930	if (mult_adj == 0) {
1931		return;
1932	} else if (mult_adj == -1) {
1933		interval = -interval;
1934		offset = -offset;
1935	} else if (mult_adj != 1) {
1936		interval *= mult_adj;
1937		offset *= mult_adj;
1938	}
1939
1940	/*
1941	 * So the following can be confusing.
1942	 *
1943	 * To keep things simple, lets assume mult_adj == 1 for now.
1944	 *
1945	 * When mult_adj != 1, remember that the interval and offset values
1946	 * have been appropriately scaled so the math is the same.
1947	 *
1948	 * The basic idea here is that we're increasing the multiplier
1949	 * by one, this causes the xtime_interval to be incremented by
1950	 * one cycle_interval. This is because:
1951	 *	xtime_interval = cycle_interval * mult
1952	 * So if mult is being incremented by one:
1953	 *	xtime_interval = cycle_interval * (mult + 1)
1954	 * Its the same as:
1955	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1956	 * Which can be shortened to:
1957	 *	xtime_interval += cycle_interval
1958	 *
1959	 * So offset stores the non-accumulated cycles. Thus the current
1960	 * time (in shifted nanoseconds) is:
1961	 *	now = (offset * adj) + xtime_nsec
1962	 * Now, even though we're adjusting the clock frequency, we have
1963	 * to keep time consistent. In other words, we can't jump back
1964	 * in time, and we also want to avoid jumping forward in time.
1965	 *
1966	 * So given the same offset value, we need the time to be the same
1967	 * both before and after the freq adjustment.
1968	 *	now = (offset * adj_1) + xtime_nsec_1
1969	 *	now = (offset * adj_2) + xtime_nsec_2
1970	 * So:
1971	 *	(offset * adj_1) + xtime_nsec_1 =
1972	 *		(offset * adj_2) + xtime_nsec_2
1973	 * And we know:
1974	 *	adj_2 = adj_1 + 1
1975	 * So:
1976	 *	(offset * adj_1) + xtime_nsec_1 =
1977	 *		(offset * (adj_1+1)) + xtime_nsec_2
1978	 *	(offset * adj_1) + xtime_nsec_1 =
1979	 *		(offset * adj_1) + offset + xtime_nsec_2
1980	 * Canceling the sides:
1981	 *	xtime_nsec_1 = offset + xtime_nsec_2
1982	 * Which gives us:
1983	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1984	 * Which simplifies to:
1985	 *	xtime_nsec -= offset
1986	 */
1987	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1988		/* NTP adjustment caused clocksource mult overflow */
1989		WARN_ON_ONCE(1);
1990		return;
1991	}
1992
1993	tk->tkr_mono.mult += mult_adj;
1994	tk->xtime_interval += interval;
1995	tk->tkr_mono.xtime_nsec -= offset;
1996}
1997
1998/*
1999 * Adjust the timekeeper's multiplier to the correct frequency
2000 * and also to reduce the accumulated error value.
2001 */
2002static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2003{
2004	u32 mult;
2005
2006	/*
2007	 * Determine the multiplier from the current NTP tick length.
2008	 * Avoid expensive division when the tick length doesn't change.
2009	 */
2010	if (likely(tk->ntp_tick == ntp_tick_length())) {
2011		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2012	} else {
2013		tk->ntp_tick = ntp_tick_length();
2014		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2015				 tk->xtime_remainder, tk->cycle_interval);
2016	}
2017
2018	/*
2019	 * If the clock is behind the NTP time, increase the multiplier by 1
2020	 * to catch up with it. If it's ahead and there was a remainder in the
2021	 * tick division, the clock will slow down. Otherwise it will stay
2022	 * ahead until the tick length changes to a non-divisible value.
2023	 */
2024	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2025	mult += tk->ntp_err_mult;
2026
2027	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2028
2029	if (unlikely(tk->tkr_mono.clock->maxadj &&
2030		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2031			> tk->tkr_mono.clock->maxadj))) {
2032		printk_once(KERN_WARNING
2033			"Adjusting %s more than 11%% (%ld vs %ld)\n",
2034			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2035			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2036	}
2037
2038	/*
2039	 * It may be possible that when we entered this function, xtime_nsec
2040	 * was very small.  Further, if we're slightly speeding the clocksource
2041	 * in the code above, its possible the required corrective factor to
2042	 * xtime_nsec could cause it to underflow.
2043	 *
2044	 * Now, since we have already accumulated the second and the NTP
2045	 * subsystem has been notified via second_overflow(), we need to skip
2046	 * the next update.
2047	 */
2048	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2049		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2050							tk->tkr_mono.shift;
2051		tk->xtime_sec--;
2052		tk->skip_second_overflow = 1;
2053	}
2054}
2055
2056/*
2057 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2058 *
2059 * Helper function that accumulates the nsecs greater than a second
2060 * from the xtime_nsec field to the xtime_secs field.
2061 * It also calls into the NTP code to handle leapsecond processing.
 
2062 */
2063static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2064{
2065	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2066	unsigned int clock_set = 0;
2067
2068	while (tk->tkr_mono.xtime_nsec >= nsecps) {
2069		int leap;
2070
2071		tk->tkr_mono.xtime_nsec -= nsecps;
2072		tk->xtime_sec++;
2073
2074		/*
2075		 * Skip NTP update if this second was accumulated before,
2076		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2077		 */
2078		if (unlikely(tk->skip_second_overflow)) {
2079			tk->skip_second_overflow = 0;
2080			continue;
2081		}
2082
2083		/* Figure out if its a leap sec and apply if needed */
2084		leap = second_overflow(tk->xtime_sec);
2085		if (unlikely(leap)) {
2086			struct timespec64 ts;
2087
2088			tk->xtime_sec += leap;
2089
2090			ts.tv_sec = leap;
2091			ts.tv_nsec = 0;
2092			tk_set_wall_to_mono(tk,
2093				timespec64_sub(tk->wall_to_monotonic, ts));
2094
2095			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2096
2097			clock_set = TK_CLOCK_WAS_SET;
2098		}
2099	}
2100	return clock_set;
2101}
2102
2103/*
2104 * logarithmic_accumulation - shifted accumulation of cycles
2105 *
2106 * This functions accumulates a shifted interval of cycles into
2107 * a shifted interval nanoseconds. Allows for O(log) accumulation
2108 * loop.
2109 *
2110 * Returns the unconsumed cycles.
2111 */
2112static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2113				    u32 shift, unsigned int *clock_set)
2114{
2115	u64 interval = tk->cycle_interval << shift;
2116	u64 snsec_per_sec;
2117
2118	/* If the offset is smaller than a shifted interval, do nothing */
2119	if (offset < interval)
2120		return offset;
2121
2122	/* Accumulate one shifted interval */
2123	offset -= interval;
2124	tk->tkr_mono.cycle_last += interval;
2125	tk->tkr_raw.cycle_last  += interval;
2126
2127	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2128	*clock_set |= accumulate_nsecs_to_secs(tk);
2129
2130	/* Accumulate raw time */
2131	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2132	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2133	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2134		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2135		tk->raw_sec++;
2136	}
2137
2138	/* Accumulate error between NTP and clock interval */
2139	tk->ntp_error += tk->ntp_tick << shift;
2140	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2141						(tk->ntp_error_shift + shift);
2142
2143	return offset;
2144}
2145
2146/*
2147 * timekeeping_advance - Updates the timekeeper to the current time and
2148 * current NTP tick length
2149 */
2150static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2151{
2152	struct timekeeper *real_tk = &tk_core.timekeeper;
2153	struct timekeeper *tk = &shadow_timekeeper;
2154	u64 offset;
2155	int shift = 0, maxshift;
2156	unsigned int clock_set = 0;
2157	unsigned long flags;
2158
2159	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2160
2161	/* Make sure we're fully resumed: */
2162	if (unlikely(timekeeping_suspended))
2163		goto out;
2164
 
 
 
 
 
 
2165	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2166				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2167
2168	/* Check if there's really nothing to do */
2169	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2170		goto out;
 
2171
2172	/* Do some additional sanity checking */
2173	timekeeping_check_update(tk, offset);
2174
2175	/*
2176	 * With NO_HZ we may have to accumulate many cycle_intervals
2177	 * (think "ticks") worth of time at once. To do this efficiently,
2178	 * we calculate the largest doubling multiple of cycle_intervals
2179	 * that is smaller than the offset.  We then accumulate that
2180	 * chunk in one go, and then try to consume the next smaller
2181	 * doubled multiple.
2182	 */
2183	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2184	shift = max(0, shift);
2185	/* Bound shift to one less than what overflows tick_length */
2186	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2187	shift = min(shift, maxshift);
2188	while (offset >= tk->cycle_interval) {
2189		offset = logarithmic_accumulation(tk, offset, shift,
2190							&clock_set);
2191		if (offset < tk->cycle_interval<<shift)
2192			shift--;
2193	}
2194
2195	/* Adjust the multiplier to correct NTP error */
2196	timekeeping_adjust(tk, offset);
2197
2198	/*
2199	 * Finally, make sure that after the rounding
2200	 * xtime_nsec isn't larger than NSEC_PER_SEC
2201	 */
2202	clock_set |= accumulate_nsecs_to_secs(tk);
2203
2204	write_seqcount_begin(&tk_core.seq);
2205	/*
2206	 * Update the real timekeeper.
2207	 *
2208	 * We could avoid this memcpy by switching pointers, but that
2209	 * requires changes to all other timekeeper usage sites as
2210	 * well, i.e. move the timekeeper pointer getter into the
2211	 * spinlocked/seqcount protected sections. And we trade this
2212	 * memcpy under the tk_core.seq against one before we start
2213	 * updating.
2214	 */
2215	timekeeping_update(tk, clock_set);
2216	memcpy(real_tk, tk, sizeof(*tk));
2217	/* The memcpy must come last. Do not put anything here! */
2218	write_seqcount_end(&tk_core.seq);
2219out:
2220	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2221
2222	return !!clock_set;
 
2223}
2224
2225/**
2226 * update_wall_time - Uses the current clocksource to increment the wall time
2227 *
2228 */
2229void update_wall_time(void)
2230{
2231	if (timekeeping_advance(TK_ADV_TICK))
2232		clock_was_set_delayed();
2233}
2234
2235/**
2236 * getboottime64 - Return the real time of system boot.
2237 * @ts:		pointer to the timespec64 to be set
2238 *
2239 * Returns the wall-time of boot in a timespec64.
2240 *
2241 * This is based on the wall_to_monotonic offset and the total suspend
2242 * time. Calls to settimeofday will affect the value returned (which
2243 * basically means that however wrong your real time clock is at boot time,
2244 * you get the right time here).
2245 */
2246void getboottime64(struct timespec64 *ts)
2247{
2248	struct timekeeper *tk = &tk_core.timekeeper;
2249	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2250
2251	*ts = ktime_to_timespec64(t);
2252}
2253EXPORT_SYMBOL_GPL(getboottime64);
2254
2255void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2256{
2257	struct timekeeper *tk = &tk_core.timekeeper;
2258	unsigned int seq;
2259
2260	do {
2261		seq = read_seqcount_begin(&tk_core.seq);
2262
2263		*ts = tk_xtime(tk);
2264	} while (read_seqcount_retry(&tk_core.seq, seq));
2265}
2266EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2267
2268void ktime_get_coarse_ts64(struct timespec64 *ts)
2269{
2270	struct timekeeper *tk = &tk_core.timekeeper;
2271	struct timespec64 now, mono;
2272	unsigned int seq;
2273
2274	do {
2275		seq = read_seqcount_begin(&tk_core.seq);
2276
2277		now = tk_xtime(tk);
2278		mono = tk->wall_to_monotonic;
2279	} while (read_seqcount_retry(&tk_core.seq, seq));
2280
2281	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2282				now.tv_nsec + mono.tv_nsec);
2283}
2284EXPORT_SYMBOL(ktime_get_coarse_ts64);
2285
2286/*
2287 * Must hold jiffies_lock
2288 */
2289void do_timer(unsigned long ticks)
2290{
2291	jiffies_64 += ticks;
2292	calc_global_load();
2293}
2294
2295/**
2296 * ktime_get_update_offsets_now - hrtimer helper
2297 * @cwsseq:	pointer to check and store the clock was set sequence number
2298 * @offs_real:	pointer to storage for monotonic -> realtime offset
2299 * @offs_boot:	pointer to storage for monotonic -> boottime offset
2300 * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2301 *
2302 * Returns current monotonic time and updates the offsets if the
2303 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2304 * different.
2305 *
2306 * Called from hrtimer_interrupt() or retrigger_next_event()
2307 */
2308ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2309				     ktime_t *offs_boot, ktime_t *offs_tai)
2310{
2311	struct timekeeper *tk = &tk_core.timekeeper;
2312	unsigned int seq;
2313	ktime_t base;
2314	u64 nsecs;
2315
2316	do {
2317		seq = read_seqcount_begin(&tk_core.seq);
2318
2319		base = tk->tkr_mono.base;
2320		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2321		base = ktime_add_ns(base, nsecs);
2322
2323		if (*cwsseq != tk->clock_was_set_seq) {
2324			*cwsseq = tk->clock_was_set_seq;
2325			*offs_real = tk->offs_real;
2326			*offs_boot = tk->offs_boot;
2327			*offs_tai = tk->offs_tai;
2328		}
2329
2330		/* Handle leapsecond insertion adjustments */
2331		if (unlikely(base >= tk->next_leap_ktime))
2332			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2333
2334	} while (read_seqcount_retry(&tk_core.seq, seq));
2335
2336	return base;
2337}
2338
2339/*
2340 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2341 */
2342static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2343{
2344	if (txc->modes & ADJ_ADJTIME) {
2345		/* singleshot must not be used with any other mode bits */
2346		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2347			return -EINVAL;
2348		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2349		    !capable(CAP_SYS_TIME))
2350			return -EPERM;
2351	} else {
2352		/* In order to modify anything, you gotta be super-user! */
2353		if (txc->modes && !capable(CAP_SYS_TIME))
2354			return -EPERM;
2355		/*
2356		 * if the quartz is off by more than 10% then
2357		 * something is VERY wrong!
2358		 */
2359		if (txc->modes & ADJ_TICK &&
2360		    (txc->tick <  900000/USER_HZ ||
2361		     txc->tick > 1100000/USER_HZ))
2362			return -EINVAL;
2363	}
2364
2365	if (txc->modes & ADJ_SETOFFSET) {
2366		/* In order to inject time, you gotta be super-user! */
2367		if (!capable(CAP_SYS_TIME))
2368			return -EPERM;
2369
2370		/*
2371		 * Validate if a timespec/timeval used to inject a time
2372		 * offset is valid.  Offsets can be positive or negative, so
2373		 * we don't check tv_sec. The value of the timeval/timespec
2374		 * is the sum of its fields,but *NOTE*:
2375		 * The field tv_usec/tv_nsec must always be non-negative and
2376		 * we can't have more nanoseconds/microseconds than a second.
2377		 */
2378		if (txc->time.tv_usec < 0)
2379			return -EINVAL;
2380
2381		if (txc->modes & ADJ_NANO) {
2382			if (txc->time.tv_usec >= NSEC_PER_SEC)
2383				return -EINVAL;
2384		} else {
2385			if (txc->time.tv_usec >= USEC_PER_SEC)
2386				return -EINVAL;
2387		}
2388	}
2389
2390	/*
2391	 * Check for potential multiplication overflows that can
2392	 * only happen on 64-bit systems:
2393	 */
2394	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2395		if (LLONG_MIN / PPM_SCALE > txc->freq)
2396			return -EINVAL;
2397		if (LLONG_MAX / PPM_SCALE < txc->freq)
2398			return -EINVAL;
2399	}
2400
2401	return 0;
2402}
2403
2404/**
2405 * random_get_entropy_fallback - Returns the raw clock source value,
2406 * used by random.c for platforms with no valid random_get_entropy().
2407 */
2408unsigned long random_get_entropy_fallback(void)
2409{
2410	struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2411	struct clocksource *clock = READ_ONCE(tkr->clock);
2412
2413	if (unlikely(timekeeping_suspended || !clock))
2414		return 0;
2415	return clock->read(clock);
2416}
2417EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2418
2419/**
2420 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2421 */
2422int do_adjtimex(struct __kernel_timex *txc)
2423{
2424	struct timekeeper *tk = &tk_core.timekeeper;
2425	struct audit_ntp_data ad;
2426	bool clock_set = false;
2427	struct timespec64 ts;
2428	unsigned long flags;
 
2429	s32 orig_tai, tai;
2430	int ret;
2431
2432	/* Validate the data before disabling interrupts */
2433	ret = timekeeping_validate_timex(txc);
2434	if (ret)
2435		return ret;
2436	add_device_randomness(txc, sizeof(*txc));
2437
2438	if (txc->modes & ADJ_SETOFFSET) {
2439		struct timespec64 delta;
2440		delta.tv_sec  = txc->time.tv_sec;
2441		delta.tv_nsec = txc->time.tv_usec;
2442		if (!(txc->modes & ADJ_NANO))
2443			delta.tv_nsec *= 1000;
2444		ret = timekeeping_inject_offset(&delta);
2445		if (ret)
2446			return ret;
2447
2448		audit_tk_injoffset(delta);
2449	}
2450
2451	audit_ntp_init(&ad);
2452
2453	ktime_get_real_ts64(&ts);
2454	add_device_randomness(&ts, sizeof(ts));
2455
2456	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2457	write_seqcount_begin(&tk_core.seq);
2458
2459	orig_tai = tai = tk->tai_offset;
2460	ret = __do_adjtimex(txc, &ts, &tai, &ad);
2461
2462	if (tai != orig_tai) {
2463		__timekeeping_set_tai_offset(tk, tai);
2464		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2465		clock_set = true;
2466	}
2467	tk_update_leap_state(tk);
2468
2469	write_seqcount_end(&tk_core.seq);
2470	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2471
2472	audit_ntp_log(&ad);
2473
2474	/* Update the multiplier immediately if frequency was set directly */
2475	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2476		clock_set |= timekeeping_advance(TK_ADV_FREQ);
2477
2478	if (clock_set)
2479		clock_was_set(CLOCK_REALTIME);
2480
2481	ntp_notify_cmos_timer();
2482
2483	return ret;
2484}
2485
2486#ifdef CONFIG_NTP_PPS
2487/**
2488 * hardpps() - Accessor function to NTP __hardpps function
2489 */
2490void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2491{
2492	unsigned long flags;
2493
2494	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2495	write_seqcount_begin(&tk_core.seq);
2496
2497	__hardpps(phase_ts, raw_ts);
2498
2499	write_seqcount_end(&tk_core.seq);
2500	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2501}
2502EXPORT_SYMBOL(hardpps);
2503#endif /* CONFIG_NTP_PPS */
v5.9
   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
  42DEFINE_RAW_SPINLOCK(timekeeper_lock);
  43
  44/*
  45 * The most important data for readout fits into a single 64 byte
  46 * cache line.
  47 */
  48static struct {
  49	seqcount_raw_spinlock_t	seq;
  50	struct timekeeper	timekeeper;
  51} tk_core ____cacheline_aligned = {
  52	.seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
  53};
  54
  55static struct timekeeper shadow_timekeeper;
  56
 
 
 
  57/**
  58 * struct tk_fast - NMI safe timekeeper
  59 * @seq:	Sequence counter for protecting updates. The lowest bit
  60 *		is the index for the tk_read_base array
  61 * @base:	tk_read_base array. Access is indexed by the lowest bit of
  62 *		@seq.
  63 *
  64 * See @update_fast_timekeeper() below.
  65 */
  66struct tk_fast {
  67	seqcount_raw_spinlock_t	seq;
  68	struct tk_read_base	base[2];
  69};
  70
  71/* Suspend-time cycles value for halted fast timekeeper. */
  72static u64 cycles_at_suspend;
  73
  74static u64 dummy_clock_read(struct clocksource *cs)
  75{
  76	return cycles_at_suspend;
 
 
  77}
  78
  79static struct clocksource dummy_clock = {
  80	.read = dummy_clock_read,
  81};
  82
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  83static struct tk_fast tk_fast_mono ____cacheline_aligned = {
  84	.seq     = SEQCNT_RAW_SPINLOCK_ZERO(tk_fast_mono.seq, &timekeeper_lock),
  85	.base[0] = { .clock = &dummy_clock, },
  86	.base[1] = { .clock = &dummy_clock, },
  87};
  88
  89static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
  90	.seq     = SEQCNT_RAW_SPINLOCK_ZERO(tk_fast_raw.seq, &timekeeper_lock),
  91	.base[0] = { .clock = &dummy_clock, },
  92	.base[1] = { .clock = &dummy_clock, },
  93};
  94
  95/* flag for if timekeeping is suspended */
  96int __read_mostly timekeeping_suspended;
  97
  98static inline void tk_normalize_xtime(struct timekeeper *tk)
  99{
 100	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
 101		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
 102		tk->xtime_sec++;
 103	}
 104	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
 105		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
 106		tk->raw_sec++;
 107	}
 108}
 109
 110static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
 111{
 112	struct timespec64 ts;
 113
 114	ts.tv_sec = tk->xtime_sec;
 115	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 116	return ts;
 117}
 118
 119static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
 120{
 121	tk->xtime_sec = ts->tv_sec;
 122	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
 123}
 124
 125static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
 126{
 127	tk->xtime_sec += ts->tv_sec;
 128	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
 129	tk_normalize_xtime(tk);
 130}
 131
 132static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
 133{
 134	struct timespec64 tmp;
 135
 136	/*
 137	 * Verify consistency of: offset_real = -wall_to_monotonic
 138	 * before modifying anything
 139	 */
 140	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
 141					-tk->wall_to_monotonic.tv_nsec);
 142	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
 143	tk->wall_to_monotonic = wtm;
 144	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
 145	tk->offs_real = timespec64_to_ktime(tmp);
 146	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
 147}
 148
 149static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
 150{
 151	tk->offs_boot = ktime_add(tk->offs_boot, delta);
 152	/*
 153	 * Timespec representation for VDSO update to avoid 64bit division
 154	 * on every update.
 155	 */
 156	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
 157}
 158
 159/*
 160 * tk_clock_read - atomic clocksource read() helper
 161 *
 162 * This helper is necessary to use in the read paths because, while the
 163 * seqcount ensures we don't return a bad value while structures are updated,
 164 * it doesn't protect from potential crashes. There is the possibility that
 165 * the tkr's clocksource may change between the read reference, and the
 166 * clock reference passed to the read function.  This can cause crashes if
 167 * the wrong clocksource is passed to the wrong read function.
 168 * This isn't necessary to use when holding the timekeeper_lock or doing
 169 * a read of the fast-timekeeper tkrs (which is protected by its own locking
 170 * and update logic).
 171 */
 172static inline u64 tk_clock_read(const struct tk_read_base *tkr)
 173{
 174	struct clocksource *clock = READ_ONCE(tkr->clock);
 175
 176	return clock->read(clock);
 177}
 178
 179#ifdef CONFIG_DEBUG_TIMEKEEPING
 180#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
 181
 182static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 183{
 184
 185	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
 186	const char *name = tk->tkr_mono.clock->name;
 187
 188	if (offset > max_cycles) {
 189		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
 190				offset, name, max_cycles);
 191		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
 192	} else {
 193		if (offset > (max_cycles >> 1)) {
 194			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
 195					offset, name, max_cycles >> 1);
 196			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
 197		}
 198	}
 199
 200	if (tk->underflow_seen) {
 201		if (jiffies - tk->last_warning > WARNING_FREQ) {
 202			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
 203			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 204			printk_deferred("         Your kernel is probably still fine.\n");
 205			tk->last_warning = jiffies;
 206		}
 207		tk->underflow_seen = 0;
 208	}
 209
 210	if (tk->overflow_seen) {
 211		if (jiffies - tk->last_warning > WARNING_FREQ) {
 212			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
 213			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 214			printk_deferred("         Your kernel is probably still fine.\n");
 215			tk->last_warning = jiffies;
 216		}
 217		tk->overflow_seen = 0;
 218	}
 219}
 220
 221static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 222{
 223	struct timekeeper *tk = &tk_core.timekeeper;
 224	u64 now, last, mask, max, delta;
 225	unsigned int seq;
 226
 227	/*
 228	 * Since we're called holding a seqcount, the data may shift
 229	 * under us while we're doing the calculation. This can cause
 230	 * false positives, since we'd note a problem but throw the
 231	 * results away. So nest another seqcount here to atomically
 232	 * grab the points we are checking with.
 233	 */
 234	do {
 235		seq = read_seqcount_begin(&tk_core.seq);
 236		now = tk_clock_read(tkr);
 237		last = tkr->cycle_last;
 238		mask = tkr->mask;
 239		max = tkr->clock->max_cycles;
 240	} while (read_seqcount_retry(&tk_core.seq, seq));
 241
 242	delta = clocksource_delta(now, last, mask);
 243
 244	/*
 245	 * Try to catch underflows by checking if we are seeing small
 246	 * mask-relative negative values.
 247	 */
 248	if (unlikely((~delta & mask) < (mask >> 3))) {
 249		tk->underflow_seen = 1;
 250		delta = 0;
 251	}
 252
 253	/* Cap delta value to the max_cycles values to avoid mult overflows */
 254	if (unlikely(delta > max)) {
 255		tk->overflow_seen = 1;
 256		delta = tkr->clock->max_cycles;
 257	}
 258
 259	return delta;
 260}
 261#else
 262static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 263{
 264}
 265static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 266{
 267	u64 cycle_now, delta;
 268
 269	/* read clocksource */
 270	cycle_now = tk_clock_read(tkr);
 271
 272	/* calculate the delta since the last update_wall_time */
 273	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 274
 275	return delta;
 276}
 277#endif
 278
 279/**
 280 * tk_setup_internals - Set up internals to use clocksource clock.
 281 *
 282 * @tk:		The target timekeeper to setup.
 283 * @clock:		Pointer to clocksource.
 284 *
 285 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
 286 * pair and interval request.
 287 *
 288 * Unless you're the timekeeping code, you should not be using this!
 289 */
 290static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
 291{
 292	u64 interval;
 293	u64 tmp, ntpinterval;
 294	struct clocksource *old_clock;
 295
 296	++tk->cs_was_changed_seq;
 297	old_clock = tk->tkr_mono.clock;
 298	tk->tkr_mono.clock = clock;
 299	tk->tkr_mono.mask = clock->mask;
 300	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
 301
 302	tk->tkr_raw.clock = clock;
 303	tk->tkr_raw.mask = clock->mask;
 304	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
 305
 306	/* Do the ns -> cycle conversion first, using original mult */
 307	tmp = NTP_INTERVAL_LENGTH;
 308	tmp <<= clock->shift;
 309	ntpinterval = tmp;
 310	tmp += clock->mult/2;
 311	do_div(tmp, clock->mult);
 312	if (tmp == 0)
 313		tmp = 1;
 314
 315	interval = (u64) tmp;
 316	tk->cycle_interval = interval;
 317
 318	/* Go back from cycles -> shifted ns */
 319	tk->xtime_interval = interval * clock->mult;
 320	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
 321	tk->raw_interval = interval * clock->mult;
 322
 323	 /* if changing clocks, convert xtime_nsec shift units */
 324	if (old_clock) {
 325		int shift_change = clock->shift - old_clock->shift;
 326		if (shift_change < 0) {
 327			tk->tkr_mono.xtime_nsec >>= -shift_change;
 328			tk->tkr_raw.xtime_nsec >>= -shift_change;
 329		} else {
 330			tk->tkr_mono.xtime_nsec <<= shift_change;
 331			tk->tkr_raw.xtime_nsec <<= shift_change;
 332		}
 333	}
 334
 335	tk->tkr_mono.shift = clock->shift;
 336	tk->tkr_raw.shift = clock->shift;
 337
 338	tk->ntp_error = 0;
 339	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 340	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
 341
 342	/*
 343	 * The timekeeper keeps its own mult values for the currently
 344	 * active clocksource. These value will be adjusted via NTP
 345	 * to counteract clock drifting.
 346	 */
 347	tk->tkr_mono.mult = clock->mult;
 348	tk->tkr_raw.mult = clock->mult;
 349	tk->ntp_err_mult = 0;
 350	tk->skip_second_overflow = 0;
 351}
 352
 353/* Timekeeper helper functions. */
 354
 355#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 356static u32 default_arch_gettimeoffset(void) { return 0; }
 357u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
 358#else
 359static inline u32 arch_gettimeoffset(void) { return 0; }
 360#endif
 361
 362static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
 363{
 364	u64 nsec;
 365
 366	nsec = delta * tkr->mult + tkr->xtime_nsec;
 367	nsec >>= tkr->shift;
 368
 369	/* If arch requires, add in get_arch_timeoffset() */
 370	return nsec + arch_gettimeoffset();
 371}
 372
 373static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
 374{
 375	u64 delta;
 376
 377	delta = timekeeping_get_delta(tkr);
 378	return timekeeping_delta_to_ns(tkr, delta);
 379}
 380
 381static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
 382{
 383	u64 delta;
 384
 385	/* calculate the delta since the last update_wall_time */
 386	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 387	return timekeeping_delta_to_ns(tkr, delta);
 388}
 389
 390/**
 391 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
 392 * @tkr: Timekeeping readout base from which we take the update
 
 393 *
 394 * We want to use this from any context including NMI and tracing /
 395 * instrumenting the timekeeping code itself.
 396 *
 397 * Employ the latch technique; see @raw_write_seqcount_latch.
 398 *
 399 * So if a NMI hits the update of base[0] then it will use base[1]
 400 * which is still consistent. In the worst case this can result is a
 401 * slightly wrong timestamp (a few nanoseconds). See
 402 * @ktime_get_mono_fast_ns.
 403 */
 404static void update_fast_timekeeper(const struct tk_read_base *tkr,
 405				   struct tk_fast *tkf)
 406{
 407	struct tk_read_base *base = tkf->base;
 408
 409	/* Force readers off to base[1] */
 410	raw_write_seqcount_latch(&tkf->seq);
 411
 412	/* Update base[0] */
 413	memcpy(base, tkr, sizeof(*base));
 414
 415	/* Force readers back to base[0] */
 416	raw_write_seqcount_latch(&tkf->seq);
 417
 418	/* Update base[1] */
 419	memcpy(base + 1, base, sizeof(*base));
 420}
 421
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 422/**
 423 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
 424 *
 425 * This timestamp is not guaranteed to be monotonic across an update.
 426 * The timestamp is calculated by:
 427 *
 428 *	now = base_mono + clock_delta * slope
 429 *
 430 * So if the update lowers the slope, readers who are forced to the
 431 * not yet updated second array are still using the old steeper slope.
 432 *
 433 * tmono
 434 * ^
 435 * |    o  n
 436 * |   o n
 437 * |  u
 438 * | o
 439 * |o
 440 * |12345678---> reader order
 441 *
 442 * o = old slope
 443 * u = update
 444 * n = new slope
 445 *
 446 * So reader 6 will observe time going backwards versus reader 5.
 447 *
 448 * While other CPUs are likely to be able observe that, the only way
 449 * for a CPU local observation is when an NMI hits in the middle of
 450 * the update. Timestamps taken from that NMI context might be ahead
 451 * of the following timestamps. Callers need to be aware of that and
 452 * deal with it.
 453 */
 454static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
 455{
 456	struct tk_read_base *tkr;
 457	unsigned int seq;
 458	u64 now;
 459
 460	do {
 461		seq = raw_read_seqcount_latch(&tkf->seq);
 462		tkr = tkf->base + (seq & 0x01);
 463		now = ktime_to_ns(tkr->base);
 464
 465		now += timekeeping_delta_to_ns(tkr,
 466				clocksource_delta(
 467					tk_clock_read(tkr),
 468					tkr->cycle_last,
 469					tkr->mask));
 470	} while (read_seqcount_retry(&tkf->seq, seq));
 471
 472	return now;
 473}
 474
 475u64 ktime_get_mono_fast_ns(void)
 476{
 477	return __ktime_get_fast_ns(&tk_fast_mono);
 478}
 479EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 480
 481u64 ktime_get_raw_fast_ns(void)
 
 
 
 
 
 
 482{
 483	return __ktime_get_fast_ns(&tk_fast_raw);
 484}
 485EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
 486
 487/**
 488 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
 489 *
 490 * To keep it NMI safe since we're accessing from tracing, we're not using a
 491 * separate timekeeper with updates to monotonic clock and boot offset
 492 * protected with seqcounts. This has the following minor side effects:
 493 *
 494 * (1) Its possible that a timestamp be taken after the boot offset is updated
 495 * but before the timekeeper is updated. If this happens, the new boot offset
 496 * is added to the old timekeeping making the clock appear to update slightly
 497 * earlier:
 498 *    CPU 0                                        CPU 1
 499 *    timekeeping_inject_sleeptime64()
 500 *    __timekeeping_inject_sleeptime(tk, delta);
 501 *                                                 timestamp();
 502 *    timekeeping_update(tk, TK_CLEAR_NTP...);
 503 *
 504 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
 505 * partially updated.  Since the tk->offs_boot update is a rare event, this
 506 * should be a rare occurrence which postprocessing should be able to handle.
 
 
 
 507 */
 508u64 notrace ktime_get_boot_fast_ns(void)
 509{
 510	struct timekeeper *tk = &tk_core.timekeeper;
 511
 512	return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
 513}
 514EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
 515
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 516
 517/*
 518 * See comment for __ktime_get_fast_ns() vs. timestamp ordering
 519 */
 520static __always_inline u64 __ktime_get_real_fast_ns(struct tk_fast *tkf)
 521{
 522	struct tk_read_base *tkr;
 
 523	unsigned int seq;
 524	u64 now;
 525
 526	do {
 527		seq = raw_read_seqcount_latch(&tkf->seq);
 528		tkr = tkf->base + (seq & 0x01);
 529		now = ktime_to_ns(tkr->base_real);
 530
 531		now += timekeeping_delta_to_ns(tkr,
 532				clocksource_delta(
 533					tk_clock_read(tkr),
 534					tkr->cycle_last,
 535					tkr->mask));
 536	} while (read_seqcount_retry(&tkf->seq, seq));
 537
 538	return now;
 539}
 540
 541/**
 542 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
 
 
 543 */
 544u64 ktime_get_real_fast_ns(void)
 545{
 546	return __ktime_get_real_fast_ns(&tk_fast_mono);
 547}
 548EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
 549
 550/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 551 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
 552 * @tk: Timekeeper to snapshot.
 553 *
 554 * It generally is unsafe to access the clocksource after timekeeping has been
 555 * suspended, so take a snapshot of the readout base of @tk and use it as the
 556 * fast timekeeper's readout base while suspended.  It will return the same
 557 * number of cycles every time until timekeeping is resumed at which time the
 558 * proper readout base for the fast timekeeper will be restored automatically.
 559 */
 560static void halt_fast_timekeeper(const struct timekeeper *tk)
 561{
 562	static struct tk_read_base tkr_dummy;
 563	const struct tk_read_base *tkr = &tk->tkr_mono;
 564
 565	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 566	cycles_at_suspend = tk_clock_read(tkr);
 567	tkr_dummy.clock = &dummy_clock;
 568	tkr_dummy.base_real = tkr->base + tk->offs_real;
 569	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
 570
 571	tkr = &tk->tkr_raw;
 572	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 573	tkr_dummy.clock = &dummy_clock;
 574	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
 575}
 576
 577static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 578
 579static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 580{
 581	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 582}
 583
 584/**
 585 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 
 586 */
 587int pvclock_gtod_register_notifier(struct notifier_block *nb)
 588{
 589	struct timekeeper *tk = &tk_core.timekeeper;
 590	unsigned long flags;
 591	int ret;
 592
 593	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 594	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 595	update_pvclock_gtod(tk, true);
 596	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 597
 598	return ret;
 599}
 600EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 601
 602/**
 603 * pvclock_gtod_unregister_notifier - unregister a pvclock
 604 * timedata update listener
 
 605 */
 606int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 607{
 608	unsigned long flags;
 609	int ret;
 610
 611	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 612	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 613	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 614
 615	return ret;
 616}
 617EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 618
 619/*
 620 * tk_update_leap_state - helper to update the next_leap_ktime
 621 */
 622static inline void tk_update_leap_state(struct timekeeper *tk)
 623{
 624	tk->next_leap_ktime = ntp_get_next_leap();
 625	if (tk->next_leap_ktime != KTIME_MAX)
 626		/* Convert to monotonic time */
 627		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 628}
 629
 630/*
 631 * Update the ktime_t based scalar nsec members of the timekeeper
 632 */
 633static inline void tk_update_ktime_data(struct timekeeper *tk)
 634{
 635	u64 seconds;
 636	u32 nsec;
 637
 638	/*
 639	 * The xtime based monotonic readout is:
 640	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 641	 * The ktime based monotonic readout is:
 642	 *	nsec = base_mono + now();
 643	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 644	 */
 645	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 646	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 647	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 648
 649	/*
 650	 * The sum of the nanoseconds portions of xtime and
 651	 * wall_to_monotonic can be greater/equal one second. Take
 652	 * this into account before updating tk->ktime_sec.
 653	 */
 654	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 655	if (nsec >= NSEC_PER_SEC)
 656		seconds++;
 657	tk->ktime_sec = seconds;
 658
 659	/* Update the monotonic raw base */
 660	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
 661}
 662
 663/* must hold timekeeper_lock */
 664static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 665{
 666	if (action & TK_CLEAR_NTP) {
 667		tk->ntp_error = 0;
 668		ntp_clear();
 669	}
 670
 671	tk_update_leap_state(tk);
 672	tk_update_ktime_data(tk);
 673
 674	update_vsyscall(tk);
 675	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 676
 677	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
 678	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 679	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 680
 681	if (action & TK_CLOCK_WAS_SET)
 682		tk->clock_was_set_seq++;
 683	/*
 684	 * The mirroring of the data to the shadow-timekeeper needs
 685	 * to happen last here to ensure we don't over-write the
 686	 * timekeeper structure on the next update with stale data
 687	 */
 688	if (action & TK_MIRROR)
 689		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 690		       sizeof(tk_core.timekeeper));
 691}
 692
 693/**
 694 * timekeeping_forward_now - update clock to the current time
 
 695 *
 696 * Forward the current clock to update its state since the last call to
 697 * update_wall_time(). This is useful before significant clock changes,
 698 * as it avoids having to deal with this time offset explicitly.
 699 */
 700static void timekeeping_forward_now(struct timekeeper *tk)
 701{
 702	u64 cycle_now, delta;
 703
 704	cycle_now = tk_clock_read(&tk->tkr_mono);
 705	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 706	tk->tkr_mono.cycle_last = cycle_now;
 707	tk->tkr_raw.cycle_last  = cycle_now;
 708
 709	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 710
 711	/* If arch requires, add in get_arch_timeoffset() */
 712	tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
 713
 714
 715	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
 716
 717	/* If arch requires, add in get_arch_timeoffset() */
 718	tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
 719
 720	tk_normalize_xtime(tk);
 721}
 722
 723/**
 724 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
 725 * @ts:		pointer to the timespec to be set
 726 *
 727 * Returns the time of day in a timespec64 (WARN if suspended).
 728 */
 729void ktime_get_real_ts64(struct timespec64 *ts)
 730{
 731	struct timekeeper *tk = &tk_core.timekeeper;
 732	unsigned int seq;
 733	u64 nsecs;
 734
 735	WARN_ON(timekeeping_suspended);
 736
 737	do {
 738		seq = read_seqcount_begin(&tk_core.seq);
 739
 740		ts->tv_sec = tk->xtime_sec;
 741		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 742
 743	} while (read_seqcount_retry(&tk_core.seq, seq));
 744
 745	ts->tv_nsec = 0;
 746	timespec64_add_ns(ts, nsecs);
 747}
 748EXPORT_SYMBOL(ktime_get_real_ts64);
 749
 750ktime_t ktime_get(void)
 751{
 752	struct timekeeper *tk = &tk_core.timekeeper;
 753	unsigned int seq;
 754	ktime_t base;
 755	u64 nsecs;
 756
 757	WARN_ON(timekeeping_suspended);
 758
 759	do {
 760		seq = read_seqcount_begin(&tk_core.seq);
 761		base = tk->tkr_mono.base;
 762		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 763
 764	} while (read_seqcount_retry(&tk_core.seq, seq));
 765
 766	return ktime_add_ns(base, nsecs);
 767}
 768EXPORT_SYMBOL_GPL(ktime_get);
 769
 770u32 ktime_get_resolution_ns(void)
 771{
 772	struct timekeeper *tk = &tk_core.timekeeper;
 773	unsigned int seq;
 774	u32 nsecs;
 775
 776	WARN_ON(timekeeping_suspended);
 777
 778	do {
 779		seq = read_seqcount_begin(&tk_core.seq);
 780		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 781	} while (read_seqcount_retry(&tk_core.seq, seq));
 782
 783	return nsecs;
 784}
 785EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 786
 787static ktime_t *offsets[TK_OFFS_MAX] = {
 788	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
 789	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
 790	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
 791};
 792
 793ktime_t ktime_get_with_offset(enum tk_offsets offs)
 794{
 795	struct timekeeper *tk = &tk_core.timekeeper;
 796	unsigned int seq;
 797	ktime_t base, *offset = offsets[offs];
 798	u64 nsecs;
 799
 800	WARN_ON(timekeeping_suspended);
 801
 802	do {
 803		seq = read_seqcount_begin(&tk_core.seq);
 804		base = ktime_add(tk->tkr_mono.base, *offset);
 805		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 806
 807	} while (read_seqcount_retry(&tk_core.seq, seq));
 808
 809	return ktime_add_ns(base, nsecs);
 810
 811}
 812EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 813
 814ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
 815{
 816	struct timekeeper *tk = &tk_core.timekeeper;
 817	unsigned int seq;
 818	ktime_t base, *offset = offsets[offs];
 819	u64 nsecs;
 820
 821	WARN_ON(timekeeping_suspended);
 822
 823	do {
 824		seq = read_seqcount_begin(&tk_core.seq);
 825		base = ktime_add(tk->tkr_mono.base, *offset);
 826		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 827
 828	} while (read_seqcount_retry(&tk_core.seq, seq));
 829
 830	return ktime_add_ns(base, nsecs);
 831}
 832EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
 833
 834/**
 835 * ktime_mono_to_any() - convert mononotic time to any other time
 836 * @tmono:	time to convert.
 837 * @offs:	which offset to use
 838 */
 839ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 840{
 841	ktime_t *offset = offsets[offs];
 842	unsigned int seq;
 843	ktime_t tconv;
 844
 845	do {
 846		seq = read_seqcount_begin(&tk_core.seq);
 847		tconv = ktime_add(tmono, *offset);
 848	} while (read_seqcount_retry(&tk_core.seq, seq));
 849
 850	return tconv;
 851}
 852EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 853
 854/**
 855 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 856 */
 857ktime_t ktime_get_raw(void)
 858{
 859	struct timekeeper *tk = &tk_core.timekeeper;
 860	unsigned int seq;
 861	ktime_t base;
 862	u64 nsecs;
 863
 864	do {
 865		seq = read_seqcount_begin(&tk_core.seq);
 866		base = tk->tkr_raw.base;
 867		nsecs = timekeeping_get_ns(&tk->tkr_raw);
 868
 869	} while (read_seqcount_retry(&tk_core.seq, seq));
 870
 871	return ktime_add_ns(base, nsecs);
 872}
 873EXPORT_SYMBOL_GPL(ktime_get_raw);
 874
 875/**
 876 * ktime_get_ts64 - get the monotonic clock in timespec64 format
 877 * @ts:		pointer to timespec variable
 878 *
 879 * The function calculates the monotonic clock from the realtime
 880 * clock and the wall_to_monotonic offset and stores the result
 881 * in normalized timespec64 format in the variable pointed to by @ts.
 882 */
 883void ktime_get_ts64(struct timespec64 *ts)
 884{
 885	struct timekeeper *tk = &tk_core.timekeeper;
 886	struct timespec64 tomono;
 887	unsigned int seq;
 888	u64 nsec;
 889
 890	WARN_ON(timekeeping_suspended);
 891
 892	do {
 893		seq = read_seqcount_begin(&tk_core.seq);
 894		ts->tv_sec = tk->xtime_sec;
 895		nsec = timekeeping_get_ns(&tk->tkr_mono);
 896		tomono = tk->wall_to_monotonic;
 897
 898	} while (read_seqcount_retry(&tk_core.seq, seq));
 899
 900	ts->tv_sec += tomono.tv_sec;
 901	ts->tv_nsec = 0;
 902	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 903}
 904EXPORT_SYMBOL_GPL(ktime_get_ts64);
 905
 906/**
 907 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 908 *
 909 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 910 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 911 * works on both 32 and 64 bit systems. On 32 bit systems the readout
 912 * covers ~136 years of uptime which should be enough to prevent
 913 * premature wrap arounds.
 914 */
 915time64_t ktime_get_seconds(void)
 916{
 917	struct timekeeper *tk = &tk_core.timekeeper;
 918
 919	WARN_ON(timekeeping_suspended);
 920	return tk->ktime_sec;
 921}
 922EXPORT_SYMBOL_GPL(ktime_get_seconds);
 923
 924/**
 925 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
 926 *
 927 * Returns the wall clock seconds since 1970. This replaces the
 928 * get_seconds() interface which is not y2038 safe on 32bit systems.
 929 *
 930 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
 931 * 32bit systems the access must be protected with the sequence
 932 * counter to provide "atomic" access to the 64bit tk->xtime_sec
 933 * value.
 934 */
 935time64_t ktime_get_real_seconds(void)
 936{
 937	struct timekeeper *tk = &tk_core.timekeeper;
 938	time64_t seconds;
 939	unsigned int seq;
 940
 941	if (IS_ENABLED(CONFIG_64BIT))
 942		return tk->xtime_sec;
 943
 944	do {
 945		seq = read_seqcount_begin(&tk_core.seq);
 946		seconds = tk->xtime_sec;
 947
 948	} while (read_seqcount_retry(&tk_core.seq, seq));
 949
 950	return seconds;
 951}
 952EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
 953
 954/**
 955 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
 956 * but without the sequence counter protect. This internal function
 957 * is called just when timekeeping lock is already held.
 958 */
 959noinstr time64_t __ktime_get_real_seconds(void)
 960{
 961	struct timekeeper *tk = &tk_core.timekeeper;
 962
 963	return tk->xtime_sec;
 964}
 965
 966/**
 967 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
 968 * @systime_snapshot:	pointer to struct receiving the system time snapshot
 969 */
 970void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
 971{
 972	struct timekeeper *tk = &tk_core.timekeeper;
 973	unsigned int seq;
 974	ktime_t base_raw;
 975	ktime_t base_real;
 976	u64 nsec_raw;
 977	u64 nsec_real;
 978	u64 now;
 979
 980	WARN_ON_ONCE(timekeeping_suspended);
 981
 982	do {
 983		seq = read_seqcount_begin(&tk_core.seq);
 984		now = tk_clock_read(&tk->tkr_mono);
 
 985		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
 986		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
 987		base_real = ktime_add(tk->tkr_mono.base,
 988				      tk_core.timekeeper.offs_real);
 989		base_raw = tk->tkr_raw.base;
 990		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
 991		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
 992	} while (read_seqcount_retry(&tk_core.seq, seq));
 993
 994	systime_snapshot->cycles = now;
 995	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
 996	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
 997}
 998EXPORT_SYMBOL_GPL(ktime_get_snapshot);
 999
1000/* Scale base by mult/div checking for overflow */
1001static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1002{
1003	u64 tmp, rem;
1004
1005	tmp = div64_u64_rem(*base, div, &rem);
1006
1007	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1008	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1009		return -EOVERFLOW;
1010	tmp *= mult;
1011
1012	rem = div64_u64(rem * mult, div);
1013	*base = tmp + rem;
1014	return 0;
1015}
1016
1017/**
1018 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1019 * @history:			Snapshot representing start of history
1020 * @partial_history_cycles:	Cycle offset into history (fractional part)
1021 * @total_history_cycles:	Total history length in cycles
1022 * @discontinuity:		True indicates clock was set on history period
1023 * @ts:				Cross timestamp that should be adjusted using
1024 *	partial/total ratio
1025 *
1026 * Helper function used by get_device_system_crosststamp() to correct the
1027 * crosstimestamp corresponding to the start of the current interval to the
1028 * system counter value (timestamp point) provided by the driver. The
1029 * total_history_* quantities are the total history starting at the provided
1030 * reference point and ending at the start of the current interval. The cycle
1031 * count between the driver timestamp point and the start of the current
1032 * interval is partial_history_cycles.
1033 */
1034static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1035					 u64 partial_history_cycles,
1036					 u64 total_history_cycles,
1037					 bool discontinuity,
1038					 struct system_device_crosststamp *ts)
1039{
1040	struct timekeeper *tk = &tk_core.timekeeper;
1041	u64 corr_raw, corr_real;
1042	bool interp_forward;
1043	int ret;
1044
1045	if (total_history_cycles == 0 || partial_history_cycles == 0)
1046		return 0;
1047
1048	/* Interpolate shortest distance from beginning or end of history */
1049	interp_forward = partial_history_cycles > total_history_cycles / 2;
1050	partial_history_cycles = interp_forward ?
1051		total_history_cycles - partial_history_cycles :
1052		partial_history_cycles;
1053
1054	/*
1055	 * Scale the monotonic raw time delta by:
1056	 *	partial_history_cycles / total_history_cycles
1057	 */
1058	corr_raw = (u64)ktime_to_ns(
1059		ktime_sub(ts->sys_monoraw, history->raw));
1060	ret = scale64_check_overflow(partial_history_cycles,
1061				     total_history_cycles, &corr_raw);
1062	if (ret)
1063		return ret;
1064
1065	/*
1066	 * If there is a discontinuity in the history, scale monotonic raw
1067	 *	correction by:
1068	 *	mult(real)/mult(raw) yielding the realtime correction
1069	 * Otherwise, calculate the realtime correction similar to monotonic
1070	 *	raw calculation
1071	 */
1072	if (discontinuity) {
1073		corr_real = mul_u64_u32_div
1074			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1075	} else {
1076		corr_real = (u64)ktime_to_ns(
1077			ktime_sub(ts->sys_realtime, history->real));
1078		ret = scale64_check_overflow(partial_history_cycles,
1079					     total_history_cycles, &corr_real);
1080		if (ret)
1081			return ret;
1082	}
1083
1084	/* Fixup monotonic raw and real time time values */
1085	if (interp_forward) {
1086		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1087		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1088	} else {
1089		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1090		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1091	}
1092
1093	return 0;
1094}
1095
1096/*
1097 * cycle_between - true if test occurs chronologically between before and after
1098 */
1099static bool cycle_between(u64 before, u64 test, u64 after)
1100{
1101	if (test > before && test < after)
1102		return true;
1103	if (test < before && before > after)
1104		return true;
1105	return false;
1106}
1107
1108/**
1109 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1110 * @get_time_fn:	Callback to get simultaneous device time and
1111 *	system counter from the device driver
1112 * @ctx:		Context passed to get_time_fn()
1113 * @history_begin:	Historical reference point used to interpolate system
1114 *	time when counter provided by the driver is before the current interval
1115 * @xtstamp:		Receives simultaneously captured system and device time
1116 *
1117 * Reads a timestamp from a device and correlates it to system time
1118 */
1119int get_device_system_crosststamp(int (*get_time_fn)
1120				  (ktime_t *device_time,
1121				   struct system_counterval_t *sys_counterval,
1122				   void *ctx),
1123				  void *ctx,
1124				  struct system_time_snapshot *history_begin,
1125				  struct system_device_crosststamp *xtstamp)
1126{
1127	struct system_counterval_t system_counterval;
1128	struct timekeeper *tk = &tk_core.timekeeper;
1129	u64 cycles, now, interval_start;
1130	unsigned int clock_was_set_seq = 0;
1131	ktime_t base_real, base_raw;
1132	u64 nsec_real, nsec_raw;
1133	u8 cs_was_changed_seq;
1134	unsigned int seq;
1135	bool do_interp;
1136	int ret;
1137
1138	do {
1139		seq = read_seqcount_begin(&tk_core.seq);
1140		/*
1141		 * Try to synchronously capture device time and a system
1142		 * counter value calling back into the device driver
1143		 */
1144		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1145		if (ret)
1146			return ret;
1147
1148		/*
1149		 * Verify that the clocksource associated with the captured
1150		 * system counter value is the same as the currently installed
1151		 * timekeeper clocksource
1152		 */
1153		if (tk->tkr_mono.clock != system_counterval.cs)
1154			return -ENODEV;
1155		cycles = system_counterval.cycles;
1156
1157		/*
1158		 * Check whether the system counter value provided by the
1159		 * device driver is on the current timekeeping interval.
1160		 */
1161		now = tk_clock_read(&tk->tkr_mono);
1162		interval_start = tk->tkr_mono.cycle_last;
1163		if (!cycle_between(interval_start, cycles, now)) {
1164			clock_was_set_seq = tk->clock_was_set_seq;
1165			cs_was_changed_seq = tk->cs_was_changed_seq;
1166			cycles = interval_start;
1167			do_interp = true;
1168		} else {
1169			do_interp = false;
1170		}
1171
1172		base_real = ktime_add(tk->tkr_mono.base,
1173				      tk_core.timekeeper.offs_real);
1174		base_raw = tk->tkr_raw.base;
1175
1176		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1177						     system_counterval.cycles);
1178		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1179						    system_counterval.cycles);
1180	} while (read_seqcount_retry(&tk_core.seq, seq));
1181
1182	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1183	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1184
1185	/*
1186	 * Interpolate if necessary, adjusting back from the start of the
1187	 * current interval
1188	 */
1189	if (do_interp) {
1190		u64 partial_history_cycles, total_history_cycles;
1191		bool discontinuity;
1192
1193		/*
1194		 * Check that the counter value occurs after the provided
1195		 * history reference and that the history doesn't cross a
1196		 * clocksource change
1197		 */
1198		if (!history_begin ||
1199		    !cycle_between(history_begin->cycles,
1200				   system_counterval.cycles, cycles) ||
1201		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1202			return -EINVAL;
1203		partial_history_cycles = cycles - system_counterval.cycles;
1204		total_history_cycles = cycles - history_begin->cycles;
1205		discontinuity =
1206			history_begin->clock_was_set_seq != clock_was_set_seq;
1207
1208		ret = adjust_historical_crosststamp(history_begin,
1209						    partial_history_cycles,
1210						    total_history_cycles,
1211						    discontinuity, xtstamp);
1212		if (ret)
1213			return ret;
1214	}
1215
1216	return 0;
1217}
1218EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1219
1220/**
1221 * do_settimeofday64 - Sets the time of day.
1222 * @ts:     pointer to the timespec64 variable containing the new time
1223 *
1224 * Sets the time of day to the new time and update NTP and notify hrtimers
1225 */
1226int do_settimeofday64(const struct timespec64 *ts)
1227{
1228	struct timekeeper *tk = &tk_core.timekeeper;
1229	struct timespec64 ts_delta, xt;
1230	unsigned long flags;
1231	int ret = 0;
1232
1233	if (!timespec64_valid_settod(ts))
1234		return -EINVAL;
1235
1236	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1237	write_seqcount_begin(&tk_core.seq);
1238
1239	timekeeping_forward_now(tk);
1240
1241	xt = tk_xtime(tk);
1242	ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1243	ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1244
1245	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1246		ret = -EINVAL;
1247		goto out;
1248	}
1249
1250	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1251
1252	tk_set_xtime(tk, ts);
1253out:
1254	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1255
1256	write_seqcount_end(&tk_core.seq);
1257	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1258
1259	/* signal hrtimers about time change */
1260	clock_was_set();
1261
1262	if (!ret)
1263		audit_tk_injoffset(ts_delta);
 
 
1264
1265	return ret;
1266}
1267EXPORT_SYMBOL(do_settimeofday64);
1268
1269/**
1270 * timekeeping_inject_offset - Adds or subtracts from the current time.
1271 * @tv:		pointer to the timespec variable containing the offset
1272 *
1273 * Adds or subtracts an offset value from the current time.
1274 */
1275static int timekeeping_inject_offset(const struct timespec64 *ts)
1276{
1277	struct timekeeper *tk = &tk_core.timekeeper;
1278	unsigned long flags;
1279	struct timespec64 tmp;
1280	int ret = 0;
1281
1282	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1283		return -EINVAL;
1284
1285	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1286	write_seqcount_begin(&tk_core.seq);
1287
1288	timekeeping_forward_now(tk);
1289
1290	/* Make sure the proposed value is valid */
1291	tmp = timespec64_add(tk_xtime(tk), *ts);
1292	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1293	    !timespec64_valid_settod(&tmp)) {
1294		ret = -EINVAL;
1295		goto error;
1296	}
1297
1298	tk_xtime_add(tk, ts);
1299	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1300
1301error: /* even if we error out, we forwarded the time, so call update */
1302	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1303
1304	write_seqcount_end(&tk_core.seq);
1305	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1306
1307	/* signal hrtimers about time change */
1308	clock_was_set();
1309
1310	return ret;
1311}
1312
1313/*
1314 * Indicates if there is an offset between the system clock and the hardware
1315 * clock/persistent clock/rtc.
1316 */
1317int persistent_clock_is_local;
1318
1319/*
1320 * Adjust the time obtained from the CMOS to be UTC time instead of
1321 * local time.
1322 *
1323 * This is ugly, but preferable to the alternatives.  Otherwise we
1324 * would either need to write a program to do it in /etc/rc (and risk
1325 * confusion if the program gets run more than once; it would also be
1326 * hard to make the program warp the clock precisely n hours)  or
1327 * compile in the timezone information into the kernel.  Bad, bad....
1328 *
1329 *						- TYT, 1992-01-01
1330 *
1331 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1332 * as real UNIX machines always do it. This avoids all headaches about
1333 * daylight saving times and warping kernel clocks.
1334 */
1335void timekeeping_warp_clock(void)
1336{
1337	if (sys_tz.tz_minuteswest != 0) {
1338		struct timespec64 adjust;
1339
1340		persistent_clock_is_local = 1;
1341		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1342		adjust.tv_nsec = 0;
1343		timekeeping_inject_offset(&adjust);
1344	}
1345}
1346
1347/**
1348 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1349 *
1350 */
1351static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1352{
1353	tk->tai_offset = tai_offset;
1354	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1355}
1356
1357/**
1358 * change_clocksource - Swaps clocksources if a new one is available
1359 *
1360 * Accumulates current time interval and initializes new clocksource
1361 */
1362static int change_clocksource(void *data)
1363{
1364	struct timekeeper *tk = &tk_core.timekeeper;
1365	struct clocksource *new, *old;
1366	unsigned long flags;
 
1367
1368	new = (struct clocksource *) data;
1369
1370	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1371	write_seqcount_begin(&tk_core.seq);
1372
1373	timekeeping_forward_now(tk);
1374	/*
1375	 * If the cs is in module, get a module reference. Succeeds
1376	 * for built-in code (owner == NULL) as well.
1377	 */
1378	if (try_module_get(new->owner)) {
1379		if (!new->enable || new->enable(new) == 0) {
1380			old = tk->tkr_mono.clock;
1381			tk_setup_internals(tk, new);
1382			if (old->disable)
1383				old->disable(old);
1384			module_put(old->owner);
1385		} else {
1386			module_put(new->owner);
1387		}
1388	}
 
 
 
 
 
 
 
 
 
 
 
1389	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1390
1391	write_seqcount_end(&tk_core.seq);
1392	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1393
 
 
 
 
 
 
 
1394	return 0;
1395}
1396
1397/**
1398 * timekeeping_notify - Install a new clock source
1399 * @clock:		pointer to the clock source
1400 *
1401 * This function is called from clocksource.c after a new, better clock
1402 * source has been registered. The caller holds the clocksource_mutex.
1403 */
1404int timekeeping_notify(struct clocksource *clock)
1405{
1406	struct timekeeper *tk = &tk_core.timekeeper;
1407
1408	if (tk->tkr_mono.clock == clock)
1409		return 0;
1410	stop_machine(change_clocksource, clock, NULL);
1411	tick_clock_notify();
1412	return tk->tkr_mono.clock == clock ? 0 : -1;
1413}
1414
1415/**
1416 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1417 * @ts:		pointer to the timespec64 to be set
1418 *
1419 * Returns the raw monotonic time (completely un-modified by ntp)
1420 */
1421void ktime_get_raw_ts64(struct timespec64 *ts)
1422{
1423	struct timekeeper *tk = &tk_core.timekeeper;
1424	unsigned int seq;
1425	u64 nsecs;
1426
1427	do {
1428		seq = read_seqcount_begin(&tk_core.seq);
1429		ts->tv_sec = tk->raw_sec;
1430		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1431
1432	} while (read_seqcount_retry(&tk_core.seq, seq));
1433
1434	ts->tv_nsec = 0;
1435	timespec64_add_ns(ts, nsecs);
1436}
1437EXPORT_SYMBOL(ktime_get_raw_ts64);
1438
1439
1440/**
1441 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1442 */
1443int timekeeping_valid_for_hres(void)
1444{
1445	struct timekeeper *tk = &tk_core.timekeeper;
1446	unsigned int seq;
1447	int ret;
1448
1449	do {
1450		seq = read_seqcount_begin(&tk_core.seq);
1451
1452		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1453
1454	} while (read_seqcount_retry(&tk_core.seq, seq));
1455
1456	return ret;
1457}
1458
1459/**
1460 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1461 */
1462u64 timekeeping_max_deferment(void)
1463{
1464	struct timekeeper *tk = &tk_core.timekeeper;
1465	unsigned int seq;
1466	u64 ret;
1467
1468	do {
1469		seq = read_seqcount_begin(&tk_core.seq);
1470
1471		ret = tk->tkr_mono.clock->max_idle_ns;
1472
1473	} while (read_seqcount_retry(&tk_core.seq, seq));
1474
1475	return ret;
1476}
1477
1478/**
1479 * read_persistent_clock64 -  Return time from the persistent clock.
 
1480 *
1481 * Weak dummy function for arches that do not yet support it.
1482 * Reads the time from the battery backed persistent clock.
1483 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1484 *
1485 *  XXX - Do be sure to remove it once all arches implement it.
1486 */
1487void __weak read_persistent_clock64(struct timespec64 *ts)
1488{
1489	ts->tv_sec = 0;
1490	ts->tv_nsec = 0;
1491}
1492
1493/**
1494 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1495 *                                        from the boot.
 
 
1496 *
1497 * Weak dummy function for arches that do not yet support it.
1498 * wall_time	- current time as returned by persistent clock
1499 * boot_offset	- offset that is defined as wall_time - boot_time
1500 * The default function calculates offset based on the current value of
1501 * local_clock(). This way architectures that support sched_clock() but don't
1502 * support dedicated boot time clock will provide the best estimate of the
1503 * boot time.
1504 */
1505void __weak __init
1506read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1507				     struct timespec64 *boot_offset)
1508{
1509	read_persistent_clock64(wall_time);
1510	*boot_offset = ns_to_timespec64(local_clock());
1511}
1512
1513/*
1514 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1515 *
1516 * The flag starts of false and is only set when a suspend reaches
1517 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1518 * timekeeper clocksource is not stopping across suspend and has been
1519 * used to update sleep time. If the timekeeper clocksource has stopped
1520 * then the flag stays true and is used by the RTC resume code to decide
1521 * whether sleeptime must be injected and if so the flag gets false then.
1522 *
1523 * If a suspend fails before reaching timekeeping_resume() then the flag
1524 * stays false and prevents erroneous sleeptime injection.
1525 */
1526static bool suspend_timing_needed;
1527
1528/* Flag for if there is a persistent clock on this platform */
1529static bool persistent_clock_exists;
1530
1531/*
1532 * timekeeping_init - Initializes the clocksource and common timekeeping values
1533 */
1534void __init timekeeping_init(void)
1535{
1536	struct timespec64 wall_time, boot_offset, wall_to_mono;
1537	struct timekeeper *tk = &tk_core.timekeeper;
1538	struct clocksource *clock;
1539	unsigned long flags;
1540
1541	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1542	if (timespec64_valid_settod(&wall_time) &&
1543	    timespec64_to_ns(&wall_time) > 0) {
1544		persistent_clock_exists = true;
1545	} else if (timespec64_to_ns(&wall_time) != 0) {
1546		pr_warn("Persistent clock returned invalid value");
1547		wall_time = (struct timespec64){0};
1548	}
1549
1550	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1551		boot_offset = (struct timespec64){0};
1552
1553	/*
1554	 * We want set wall_to_mono, so the following is true:
1555	 * wall time + wall_to_mono = boot time
1556	 */
1557	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1558
1559	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1560	write_seqcount_begin(&tk_core.seq);
1561	ntp_init();
1562
1563	clock = clocksource_default_clock();
1564	if (clock->enable)
1565		clock->enable(clock);
1566	tk_setup_internals(tk, clock);
1567
1568	tk_set_xtime(tk, &wall_time);
1569	tk->raw_sec = 0;
1570
1571	tk_set_wall_to_mono(tk, wall_to_mono);
1572
1573	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1574
1575	write_seqcount_end(&tk_core.seq);
1576	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1577}
1578
1579/* time in seconds when suspend began for persistent clock */
1580static struct timespec64 timekeeping_suspend_time;
1581
1582/**
1583 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1584 * @delta: pointer to a timespec delta value
 
1585 *
1586 * Takes a timespec offset measuring a suspend interval and properly
1587 * adds the sleep offset to the timekeeping variables.
1588 */
1589static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1590					   const struct timespec64 *delta)
1591{
1592	if (!timespec64_valid_strict(delta)) {
1593		printk_deferred(KERN_WARNING
1594				"__timekeeping_inject_sleeptime: Invalid "
1595				"sleep delta value!\n");
1596		return;
1597	}
1598	tk_xtime_add(tk, delta);
1599	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1600	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1601	tk_debug_account_sleep_time(delta);
1602}
1603
1604#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1605/**
1606 * We have three kinds of time sources to use for sleep time
1607 * injection, the preference order is:
1608 * 1) non-stop clocksource
1609 * 2) persistent clock (ie: RTC accessible when irqs are off)
1610 * 3) RTC
1611 *
1612 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1613 * If system has neither 1) nor 2), 3) will be used finally.
1614 *
1615 *
1616 * If timekeeping has injected sleeptime via either 1) or 2),
1617 * 3) becomes needless, so in this case we don't need to call
1618 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1619 * means.
1620 */
1621bool timekeeping_rtc_skipresume(void)
1622{
1623	return !suspend_timing_needed;
1624}
1625
1626/**
1627 * 1) can be determined whether to use or not only when doing
1628 * timekeeping_resume() which is invoked after rtc_suspend(),
1629 * so we can't skip rtc_suspend() surely if system has 1).
1630 *
1631 * But if system has 2), 2) will definitely be used, so in this
1632 * case we don't need to call rtc_suspend(), and this is what
1633 * timekeeping_rtc_skipsuspend() means.
1634 */
1635bool timekeeping_rtc_skipsuspend(void)
1636{
1637	return persistent_clock_exists;
1638}
1639
1640/**
1641 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1642 * @delta: pointer to a timespec64 delta value
1643 *
1644 * This hook is for architectures that cannot support read_persistent_clock64
1645 * because their RTC/persistent clock is only accessible when irqs are enabled.
1646 * and also don't have an effective nonstop clocksource.
1647 *
1648 * This function should only be called by rtc_resume(), and allows
1649 * a suspend offset to be injected into the timekeeping values.
1650 */
1651void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1652{
1653	struct timekeeper *tk = &tk_core.timekeeper;
1654	unsigned long flags;
1655
1656	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1657	write_seqcount_begin(&tk_core.seq);
1658
1659	suspend_timing_needed = false;
1660
1661	timekeeping_forward_now(tk);
1662
1663	__timekeeping_inject_sleeptime(tk, delta);
1664
1665	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1666
1667	write_seqcount_end(&tk_core.seq);
1668	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1669
1670	/* signal hrtimers about time change */
1671	clock_was_set();
1672}
1673#endif
1674
1675/**
1676 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1677 */
1678void timekeeping_resume(void)
1679{
1680	struct timekeeper *tk = &tk_core.timekeeper;
1681	struct clocksource *clock = tk->tkr_mono.clock;
1682	unsigned long flags;
1683	struct timespec64 ts_new, ts_delta;
1684	u64 cycle_now, nsec;
1685	bool inject_sleeptime = false;
1686
1687	read_persistent_clock64(&ts_new);
1688
1689	clockevents_resume();
1690	clocksource_resume();
1691
1692	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1693	write_seqcount_begin(&tk_core.seq);
1694
1695	/*
1696	 * After system resumes, we need to calculate the suspended time and
1697	 * compensate it for the OS time. There are 3 sources that could be
1698	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1699	 * device.
1700	 *
1701	 * One specific platform may have 1 or 2 or all of them, and the
1702	 * preference will be:
1703	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1704	 * The less preferred source will only be tried if there is no better
1705	 * usable source. The rtc part is handled separately in rtc core code.
1706	 */
1707	cycle_now = tk_clock_read(&tk->tkr_mono);
1708	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1709	if (nsec > 0) {
1710		ts_delta = ns_to_timespec64(nsec);
1711		inject_sleeptime = true;
1712	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1713		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1714		inject_sleeptime = true;
1715	}
1716
1717	if (inject_sleeptime) {
1718		suspend_timing_needed = false;
1719		__timekeeping_inject_sleeptime(tk, &ts_delta);
1720	}
1721
1722	/* Re-base the last cycle value */
1723	tk->tkr_mono.cycle_last = cycle_now;
1724	tk->tkr_raw.cycle_last  = cycle_now;
1725
1726	tk->ntp_error = 0;
1727	timekeeping_suspended = 0;
1728	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1729	write_seqcount_end(&tk_core.seq);
1730	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1731
1732	touch_softlockup_watchdog();
1733
 
1734	tick_resume();
1735	hrtimers_resume();
 
1736}
1737
1738int timekeeping_suspend(void)
1739{
1740	struct timekeeper *tk = &tk_core.timekeeper;
1741	unsigned long flags;
1742	struct timespec64		delta, delta_delta;
1743	static struct timespec64	old_delta;
1744	struct clocksource *curr_clock;
1745	u64 cycle_now;
1746
1747	read_persistent_clock64(&timekeeping_suspend_time);
1748
1749	/*
1750	 * On some systems the persistent_clock can not be detected at
1751	 * timekeeping_init by its return value, so if we see a valid
1752	 * value returned, update the persistent_clock_exists flag.
1753	 */
1754	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1755		persistent_clock_exists = true;
1756
1757	suspend_timing_needed = true;
1758
1759	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1760	write_seqcount_begin(&tk_core.seq);
1761	timekeeping_forward_now(tk);
1762	timekeeping_suspended = 1;
1763
1764	/*
1765	 * Since we've called forward_now, cycle_last stores the value
1766	 * just read from the current clocksource. Save this to potentially
1767	 * use in suspend timing.
1768	 */
1769	curr_clock = tk->tkr_mono.clock;
1770	cycle_now = tk->tkr_mono.cycle_last;
1771	clocksource_start_suspend_timing(curr_clock, cycle_now);
1772
1773	if (persistent_clock_exists) {
1774		/*
1775		 * To avoid drift caused by repeated suspend/resumes,
1776		 * which each can add ~1 second drift error,
1777		 * try to compensate so the difference in system time
1778		 * and persistent_clock time stays close to constant.
1779		 */
1780		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1781		delta_delta = timespec64_sub(delta, old_delta);
1782		if (abs(delta_delta.tv_sec) >= 2) {
1783			/*
1784			 * if delta_delta is too large, assume time correction
1785			 * has occurred and set old_delta to the current delta.
1786			 */
1787			old_delta = delta;
1788		} else {
1789			/* Otherwise try to adjust old_system to compensate */
1790			timekeeping_suspend_time =
1791				timespec64_add(timekeeping_suspend_time, delta_delta);
1792		}
1793	}
1794
1795	timekeeping_update(tk, TK_MIRROR);
1796	halt_fast_timekeeper(tk);
1797	write_seqcount_end(&tk_core.seq);
1798	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1799
1800	tick_suspend();
1801	clocksource_suspend();
1802	clockevents_suspend();
1803
1804	return 0;
1805}
1806
1807/* sysfs resume/suspend bits for timekeeping */
1808static struct syscore_ops timekeeping_syscore_ops = {
1809	.resume		= timekeeping_resume,
1810	.suspend	= timekeeping_suspend,
1811};
1812
1813static int __init timekeeping_init_ops(void)
1814{
1815	register_syscore_ops(&timekeeping_syscore_ops);
1816	return 0;
1817}
1818device_initcall(timekeeping_init_ops);
1819
1820/*
1821 * Apply a multiplier adjustment to the timekeeper
1822 */
1823static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1824							 s64 offset,
1825							 s32 mult_adj)
1826{
1827	s64 interval = tk->cycle_interval;
1828
1829	if (mult_adj == 0) {
1830		return;
1831	} else if (mult_adj == -1) {
1832		interval = -interval;
1833		offset = -offset;
1834	} else if (mult_adj != 1) {
1835		interval *= mult_adj;
1836		offset *= mult_adj;
1837	}
1838
1839	/*
1840	 * So the following can be confusing.
1841	 *
1842	 * To keep things simple, lets assume mult_adj == 1 for now.
1843	 *
1844	 * When mult_adj != 1, remember that the interval and offset values
1845	 * have been appropriately scaled so the math is the same.
1846	 *
1847	 * The basic idea here is that we're increasing the multiplier
1848	 * by one, this causes the xtime_interval to be incremented by
1849	 * one cycle_interval. This is because:
1850	 *	xtime_interval = cycle_interval * mult
1851	 * So if mult is being incremented by one:
1852	 *	xtime_interval = cycle_interval * (mult + 1)
1853	 * Its the same as:
1854	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1855	 * Which can be shortened to:
1856	 *	xtime_interval += cycle_interval
1857	 *
1858	 * So offset stores the non-accumulated cycles. Thus the current
1859	 * time (in shifted nanoseconds) is:
1860	 *	now = (offset * adj) + xtime_nsec
1861	 * Now, even though we're adjusting the clock frequency, we have
1862	 * to keep time consistent. In other words, we can't jump back
1863	 * in time, and we also want to avoid jumping forward in time.
1864	 *
1865	 * So given the same offset value, we need the time to be the same
1866	 * both before and after the freq adjustment.
1867	 *	now = (offset * adj_1) + xtime_nsec_1
1868	 *	now = (offset * adj_2) + xtime_nsec_2
1869	 * So:
1870	 *	(offset * adj_1) + xtime_nsec_1 =
1871	 *		(offset * adj_2) + xtime_nsec_2
1872	 * And we know:
1873	 *	adj_2 = adj_1 + 1
1874	 * So:
1875	 *	(offset * adj_1) + xtime_nsec_1 =
1876	 *		(offset * (adj_1+1)) + xtime_nsec_2
1877	 *	(offset * adj_1) + xtime_nsec_1 =
1878	 *		(offset * adj_1) + offset + xtime_nsec_2
1879	 * Canceling the sides:
1880	 *	xtime_nsec_1 = offset + xtime_nsec_2
1881	 * Which gives us:
1882	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1883	 * Which simplfies to:
1884	 *	xtime_nsec -= offset
1885	 */
1886	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1887		/* NTP adjustment caused clocksource mult overflow */
1888		WARN_ON_ONCE(1);
1889		return;
1890	}
1891
1892	tk->tkr_mono.mult += mult_adj;
1893	tk->xtime_interval += interval;
1894	tk->tkr_mono.xtime_nsec -= offset;
1895}
1896
1897/*
1898 * Adjust the timekeeper's multiplier to the correct frequency
1899 * and also to reduce the accumulated error value.
1900 */
1901static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1902{
1903	u32 mult;
1904
1905	/*
1906	 * Determine the multiplier from the current NTP tick length.
1907	 * Avoid expensive division when the tick length doesn't change.
1908	 */
1909	if (likely(tk->ntp_tick == ntp_tick_length())) {
1910		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1911	} else {
1912		tk->ntp_tick = ntp_tick_length();
1913		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1914				 tk->xtime_remainder, tk->cycle_interval);
1915	}
1916
1917	/*
1918	 * If the clock is behind the NTP time, increase the multiplier by 1
1919	 * to catch up with it. If it's ahead and there was a remainder in the
1920	 * tick division, the clock will slow down. Otherwise it will stay
1921	 * ahead until the tick length changes to a non-divisible value.
1922	 */
1923	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1924	mult += tk->ntp_err_mult;
1925
1926	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
1927
1928	if (unlikely(tk->tkr_mono.clock->maxadj &&
1929		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1930			> tk->tkr_mono.clock->maxadj))) {
1931		printk_once(KERN_WARNING
1932			"Adjusting %s more than 11%% (%ld vs %ld)\n",
1933			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1934			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1935	}
1936
1937	/*
1938	 * It may be possible that when we entered this function, xtime_nsec
1939	 * was very small.  Further, if we're slightly speeding the clocksource
1940	 * in the code above, its possible the required corrective factor to
1941	 * xtime_nsec could cause it to underflow.
1942	 *
1943	 * Now, since we have already accumulated the second and the NTP
1944	 * subsystem has been notified via second_overflow(), we need to skip
1945	 * the next update.
1946	 */
1947	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1948		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
1949							tk->tkr_mono.shift;
1950		tk->xtime_sec--;
1951		tk->skip_second_overflow = 1;
1952	}
1953}
1954
1955/**
1956 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1957 *
1958 * Helper function that accumulates the nsecs greater than a second
1959 * from the xtime_nsec field to the xtime_secs field.
1960 * It also calls into the NTP code to handle leapsecond processing.
1961 *
1962 */
1963static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1964{
1965	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1966	unsigned int clock_set = 0;
1967
1968	while (tk->tkr_mono.xtime_nsec >= nsecps) {
1969		int leap;
1970
1971		tk->tkr_mono.xtime_nsec -= nsecps;
1972		tk->xtime_sec++;
1973
1974		/*
1975		 * Skip NTP update if this second was accumulated before,
1976		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
1977		 */
1978		if (unlikely(tk->skip_second_overflow)) {
1979			tk->skip_second_overflow = 0;
1980			continue;
1981		}
1982
1983		/* Figure out if its a leap sec and apply if needed */
1984		leap = second_overflow(tk->xtime_sec);
1985		if (unlikely(leap)) {
1986			struct timespec64 ts;
1987
1988			tk->xtime_sec += leap;
1989
1990			ts.tv_sec = leap;
1991			ts.tv_nsec = 0;
1992			tk_set_wall_to_mono(tk,
1993				timespec64_sub(tk->wall_to_monotonic, ts));
1994
1995			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1996
1997			clock_set = TK_CLOCK_WAS_SET;
1998		}
1999	}
2000	return clock_set;
2001}
2002
2003/**
2004 * logarithmic_accumulation - shifted accumulation of cycles
2005 *
2006 * This functions accumulates a shifted interval of cycles into
2007 * a shifted interval nanoseconds. Allows for O(log) accumulation
2008 * loop.
2009 *
2010 * Returns the unconsumed cycles.
2011 */
2012static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2013				    u32 shift, unsigned int *clock_set)
2014{
2015	u64 interval = tk->cycle_interval << shift;
2016	u64 snsec_per_sec;
2017
2018	/* If the offset is smaller than a shifted interval, do nothing */
2019	if (offset < interval)
2020		return offset;
2021
2022	/* Accumulate one shifted interval */
2023	offset -= interval;
2024	tk->tkr_mono.cycle_last += interval;
2025	tk->tkr_raw.cycle_last  += interval;
2026
2027	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2028	*clock_set |= accumulate_nsecs_to_secs(tk);
2029
2030	/* Accumulate raw time */
2031	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2032	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2033	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2034		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2035		tk->raw_sec++;
2036	}
2037
2038	/* Accumulate error between NTP and clock interval */
2039	tk->ntp_error += tk->ntp_tick << shift;
2040	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2041						(tk->ntp_error_shift + shift);
2042
2043	return offset;
2044}
2045
2046/*
2047 * timekeeping_advance - Updates the timekeeper to the current time and
2048 * current NTP tick length
2049 */
2050static void timekeeping_advance(enum timekeeping_adv_mode mode)
2051{
2052	struct timekeeper *real_tk = &tk_core.timekeeper;
2053	struct timekeeper *tk = &shadow_timekeeper;
2054	u64 offset;
2055	int shift = 0, maxshift;
2056	unsigned int clock_set = 0;
2057	unsigned long flags;
2058
2059	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2060
2061	/* Make sure we're fully resumed: */
2062	if (unlikely(timekeeping_suspended))
2063		goto out;
2064
2065#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2066	offset = real_tk->cycle_interval;
2067
2068	if (mode != TK_ADV_TICK)
2069		goto out;
2070#else
2071	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2072				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2073
2074	/* Check if there's really nothing to do */
2075	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2076		goto out;
2077#endif
2078
2079	/* Do some additional sanity checking */
2080	timekeeping_check_update(tk, offset);
2081
2082	/*
2083	 * With NO_HZ we may have to accumulate many cycle_intervals
2084	 * (think "ticks") worth of time at once. To do this efficiently,
2085	 * we calculate the largest doubling multiple of cycle_intervals
2086	 * that is smaller than the offset.  We then accumulate that
2087	 * chunk in one go, and then try to consume the next smaller
2088	 * doubled multiple.
2089	 */
2090	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2091	shift = max(0, shift);
2092	/* Bound shift to one less than what overflows tick_length */
2093	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2094	shift = min(shift, maxshift);
2095	while (offset >= tk->cycle_interval) {
2096		offset = logarithmic_accumulation(tk, offset, shift,
2097							&clock_set);
2098		if (offset < tk->cycle_interval<<shift)
2099			shift--;
2100	}
2101
2102	/* Adjust the multiplier to correct NTP error */
2103	timekeeping_adjust(tk, offset);
2104
2105	/*
2106	 * Finally, make sure that after the rounding
2107	 * xtime_nsec isn't larger than NSEC_PER_SEC
2108	 */
2109	clock_set |= accumulate_nsecs_to_secs(tk);
2110
2111	write_seqcount_begin(&tk_core.seq);
2112	/*
2113	 * Update the real timekeeper.
2114	 *
2115	 * We could avoid this memcpy by switching pointers, but that
2116	 * requires changes to all other timekeeper usage sites as
2117	 * well, i.e. move the timekeeper pointer getter into the
2118	 * spinlocked/seqcount protected sections. And we trade this
2119	 * memcpy under the tk_core.seq against one before we start
2120	 * updating.
2121	 */
2122	timekeeping_update(tk, clock_set);
2123	memcpy(real_tk, tk, sizeof(*tk));
2124	/* The memcpy must come last. Do not put anything here! */
2125	write_seqcount_end(&tk_core.seq);
2126out:
2127	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2128	if (clock_set)
2129		/* Have to call _delayed version, since in irq context*/
2130		clock_was_set_delayed();
2131}
2132
2133/**
2134 * update_wall_time - Uses the current clocksource to increment the wall time
2135 *
2136 */
2137void update_wall_time(void)
2138{
2139	timekeeping_advance(TK_ADV_TICK);
 
2140}
2141
2142/**
2143 * getboottime64 - Return the real time of system boot.
2144 * @ts:		pointer to the timespec64 to be set
2145 *
2146 * Returns the wall-time of boot in a timespec64.
2147 *
2148 * This is based on the wall_to_monotonic offset and the total suspend
2149 * time. Calls to settimeofday will affect the value returned (which
2150 * basically means that however wrong your real time clock is at boot time,
2151 * you get the right time here).
2152 */
2153void getboottime64(struct timespec64 *ts)
2154{
2155	struct timekeeper *tk = &tk_core.timekeeper;
2156	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2157
2158	*ts = ktime_to_timespec64(t);
2159}
2160EXPORT_SYMBOL_GPL(getboottime64);
2161
2162void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2163{
2164	struct timekeeper *tk = &tk_core.timekeeper;
2165	unsigned int seq;
2166
2167	do {
2168		seq = read_seqcount_begin(&tk_core.seq);
2169
2170		*ts = tk_xtime(tk);
2171	} while (read_seqcount_retry(&tk_core.seq, seq));
2172}
2173EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2174
2175void ktime_get_coarse_ts64(struct timespec64 *ts)
2176{
2177	struct timekeeper *tk = &tk_core.timekeeper;
2178	struct timespec64 now, mono;
2179	unsigned int seq;
2180
2181	do {
2182		seq = read_seqcount_begin(&tk_core.seq);
2183
2184		now = tk_xtime(tk);
2185		mono = tk->wall_to_monotonic;
2186	} while (read_seqcount_retry(&tk_core.seq, seq));
2187
2188	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2189				now.tv_nsec + mono.tv_nsec);
2190}
2191EXPORT_SYMBOL(ktime_get_coarse_ts64);
2192
2193/*
2194 * Must hold jiffies_lock
2195 */
2196void do_timer(unsigned long ticks)
2197{
2198	jiffies_64 += ticks;
2199	calc_global_load();
2200}
2201
2202/**
2203 * ktime_get_update_offsets_now - hrtimer helper
2204 * @cwsseq:	pointer to check and store the clock was set sequence number
2205 * @offs_real:	pointer to storage for monotonic -> realtime offset
2206 * @offs_boot:	pointer to storage for monotonic -> boottime offset
2207 * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2208 *
2209 * Returns current monotonic time and updates the offsets if the
2210 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2211 * different.
2212 *
2213 * Called from hrtimer_interrupt() or retrigger_next_event()
2214 */
2215ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2216				     ktime_t *offs_boot, ktime_t *offs_tai)
2217{
2218	struct timekeeper *tk = &tk_core.timekeeper;
2219	unsigned int seq;
2220	ktime_t base;
2221	u64 nsecs;
2222
2223	do {
2224		seq = read_seqcount_begin(&tk_core.seq);
2225
2226		base = tk->tkr_mono.base;
2227		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2228		base = ktime_add_ns(base, nsecs);
2229
2230		if (*cwsseq != tk->clock_was_set_seq) {
2231			*cwsseq = tk->clock_was_set_seq;
2232			*offs_real = tk->offs_real;
2233			*offs_boot = tk->offs_boot;
2234			*offs_tai = tk->offs_tai;
2235		}
2236
2237		/* Handle leapsecond insertion adjustments */
2238		if (unlikely(base >= tk->next_leap_ktime))
2239			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2240
2241	} while (read_seqcount_retry(&tk_core.seq, seq));
2242
2243	return base;
2244}
2245
2246/**
2247 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2248 */
2249static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2250{
2251	if (txc->modes & ADJ_ADJTIME) {
2252		/* singleshot must not be used with any other mode bits */
2253		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2254			return -EINVAL;
2255		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2256		    !capable(CAP_SYS_TIME))
2257			return -EPERM;
2258	} else {
2259		/* In order to modify anything, you gotta be super-user! */
2260		if (txc->modes && !capable(CAP_SYS_TIME))
2261			return -EPERM;
2262		/*
2263		 * if the quartz is off by more than 10% then
2264		 * something is VERY wrong!
2265		 */
2266		if (txc->modes & ADJ_TICK &&
2267		    (txc->tick <  900000/USER_HZ ||
2268		     txc->tick > 1100000/USER_HZ))
2269			return -EINVAL;
2270	}
2271
2272	if (txc->modes & ADJ_SETOFFSET) {
2273		/* In order to inject time, you gotta be super-user! */
2274		if (!capable(CAP_SYS_TIME))
2275			return -EPERM;
2276
2277		/*
2278		 * Validate if a timespec/timeval used to inject a time
2279		 * offset is valid.  Offsets can be postive or negative, so
2280		 * we don't check tv_sec. The value of the timeval/timespec
2281		 * is the sum of its fields,but *NOTE*:
2282		 * The field tv_usec/tv_nsec must always be non-negative and
2283		 * we can't have more nanoseconds/microseconds than a second.
2284		 */
2285		if (txc->time.tv_usec < 0)
2286			return -EINVAL;
2287
2288		if (txc->modes & ADJ_NANO) {
2289			if (txc->time.tv_usec >= NSEC_PER_SEC)
2290				return -EINVAL;
2291		} else {
2292			if (txc->time.tv_usec >= USEC_PER_SEC)
2293				return -EINVAL;
2294		}
2295	}
2296
2297	/*
2298	 * Check for potential multiplication overflows that can
2299	 * only happen on 64-bit systems:
2300	 */
2301	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2302		if (LLONG_MIN / PPM_SCALE > txc->freq)
2303			return -EINVAL;
2304		if (LLONG_MAX / PPM_SCALE < txc->freq)
2305			return -EINVAL;
2306	}
2307
2308	return 0;
2309}
2310
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2311
2312/**
2313 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2314 */
2315int do_adjtimex(struct __kernel_timex *txc)
2316{
2317	struct timekeeper *tk = &tk_core.timekeeper;
2318	struct audit_ntp_data ad;
 
 
2319	unsigned long flags;
2320	struct timespec64 ts;
2321	s32 orig_tai, tai;
2322	int ret;
2323
2324	/* Validate the data before disabling interrupts */
2325	ret = timekeeping_validate_timex(txc);
2326	if (ret)
2327		return ret;
 
2328
2329	if (txc->modes & ADJ_SETOFFSET) {
2330		struct timespec64 delta;
2331		delta.tv_sec  = txc->time.tv_sec;
2332		delta.tv_nsec = txc->time.tv_usec;
2333		if (!(txc->modes & ADJ_NANO))
2334			delta.tv_nsec *= 1000;
2335		ret = timekeeping_inject_offset(&delta);
2336		if (ret)
2337			return ret;
2338
2339		audit_tk_injoffset(delta);
2340	}
2341
2342	audit_ntp_init(&ad);
2343
2344	ktime_get_real_ts64(&ts);
 
2345
2346	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2347	write_seqcount_begin(&tk_core.seq);
2348
2349	orig_tai = tai = tk->tai_offset;
2350	ret = __do_adjtimex(txc, &ts, &tai, &ad);
2351
2352	if (tai != orig_tai) {
2353		__timekeeping_set_tai_offset(tk, tai);
2354		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
 
2355	}
2356	tk_update_leap_state(tk);
2357
2358	write_seqcount_end(&tk_core.seq);
2359	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2360
2361	audit_ntp_log(&ad);
2362
2363	/* Update the multiplier immediately if frequency was set directly */
2364	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2365		timekeeping_advance(TK_ADV_FREQ);
2366
2367	if (tai != orig_tai)
2368		clock_was_set();
2369
2370	ntp_notify_cmos_timer();
2371
2372	return ret;
2373}
2374
2375#ifdef CONFIG_NTP_PPS
2376/**
2377 * hardpps() - Accessor function to NTP __hardpps function
2378 */
2379void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2380{
2381	unsigned long flags;
2382
2383	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2384	write_seqcount_begin(&tk_core.seq);
2385
2386	__hardpps(phase_ts, raw_ts);
2387
2388	write_seqcount_end(&tk_core.seq);
2389	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2390}
2391EXPORT_SYMBOL(hardpps);
2392#endif /* CONFIG_NTP_PPS */
2393
2394/**
2395 * xtime_update() - advances the timekeeping infrastructure
2396 * @ticks:	number of ticks, that have elapsed since the last call.
2397 *
2398 * Must be called with interrupts disabled.
2399 */
2400void xtime_update(unsigned long ticks)
2401{
2402	raw_spin_lock(&jiffies_lock);
2403	write_seqcount_begin(&jiffies_seq);
2404	do_timer(ticks);
2405	write_seqcount_end(&jiffies_seq);
2406	raw_spin_unlock(&jiffies_lock);
2407	update_wall_time();
2408}