Linux Audio

Check our new training course

Loading...
v5.14.15
   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/* flag for if timekeeping is suspended */
  58int __read_mostly timekeeping_suspended;
  59
  60/**
  61 * struct tk_fast - NMI safe timekeeper
  62 * @seq:	Sequence counter for protecting updates. The lowest bit
  63 *		is the index for the tk_read_base array
  64 * @base:	tk_read_base array. Access is indexed by the lowest bit of
  65 *		@seq.
  66 *
  67 * See @update_fast_timekeeper() below.
  68 */
  69struct tk_fast {
  70	seqcount_latch_t	seq;
  71	struct tk_read_base	base[2];
  72};
  73
  74/* Suspend-time cycles value for halted fast timekeeper. */
  75static u64 cycles_at_suspend;
  76
  77static u64 dummy_clock_read(struct clocksource *cs)
  78{
  79	if (timekeeping_suspended)
  80		return cycles_at_suspend;
  81	return local_clock();
  82}
  83
  84static struct clocksource dummy_clock = {
  85	.read = dummy_clock_read,
  86};
  87
  88/*
  89 * Boot time initialization which allows local_clock() to be utilized
  90 * during early boot when clocksources are not available. local_clock()
  91 * returns nanoseconds already so no conversion is required, hence mult=1
  92 * and shift=0. When the first proper clocksource is installed then
  93 * the fast time keepers are updated with the correct values.
  94 */
  95#define FAST_TK_INIT						\
  96	{							\
  97		.clock		= &dummy_clock,			\
  98		.mask		= CLOCKSOURCE_MASK(64),		\
  99		.mult		= 1,				\
 100		.shift		= 0,				\
 101	}
 102
 103static struct tk_fast tk_fast_mono ____cacheline_aligned = {
 104	.seq     = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
 105	.base[0] = FAST_TK_INIT,
 106	.base[1] = FAST_TK_INIT,
 107};
 108
 109static struct tk_fast tk_fast_raw  ____cacheline_aligned = {
 110	.seq     = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
 111	.base[0] = FAST_TK_INIT,
 112	.base[1] = FAST_TK_INIT,
 113};
 114
 
 
 
 115static inline void tk_normalize_xtime(struct timekeeper *tk)
 116{
 117	while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
 118		tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
 119		tk->xtime_sec++;
 120	}
 121	while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
 122		tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
 123		tk->raw_sec++;
 124	}
 125}
 126
 127static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
 128{
 129	struct timespec64 ts;
 130
 131	ts.tv_sec = tk->xtime_sec;
 132	ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 133	return ts;
 134}
 135
 136static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
 137{
 138	tk->xtime_sec = ts->tv_sec;
 139	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
 140}
 141
 142static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
 143{
 144	tk->xtime_sec += ts->tv_sec;
 145	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
 146	tk_normalize_xtime(tk);
 147}
 148
 149static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
 150{
 151	struct timespec64 tmp;
 152
 153	/*
 154	 * Verify consistency of: offset_real = -wall_to_monotonic
 155	 * before modifying anything
 156	 */
 157	set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
 158					-tk->wall_to_monotonic.tv_nsec);
 159	WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
 160	tk->wall_to_monotonic = wtm;
 161	set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
 162	tk->offs_real = timespec64_to_ktime(tmp);
 163	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
 164}
 165
 166static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
 167{
 168	tk->offs_boot = ktime_add(tk->offs_boot, delta);
 169	/*
 170	 * Timespec representation for VDSO update to avoid 64bit division
 171	 * on every update.
 172	 */
 173	tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
 174}
 175
 176/*
 177 * tk_clock_read - atomic clocksource read() helper
 178 *
 179 * This helper is necessary to use in the read paths because, while the
 180 * seqcount ensures we don't return a bad value while structures are updated,
 181 * it doesn't protect from potential crashes. There is the possibility that
 182 * the tkr's clocksource may change between the read reference, and the
 183 * clock reference passed to the read function.  This can cause crashes if
 184 * the wrong clocksource is passed to the wrong read function.
 185 * This isn't necessary to use when holding the timekeeper_lock or doing
 186 * a read of the fast-timekeeper tkrs (which is protected by its own locking
 187 * and update logic).
 188 */
 189static inline u64 tk_clock_read(const struct tk_read_base *tkr)
 190{
 191	struct clocksource *clock = READ_ONCE(tkr->clock);
 192
 193	return clock->read(clock);
 194}
 195
 196#ifdef CONFIG_DEBUG_TIMEKEEPING
 197#define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
 198
 199static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 200{
 201
 202	u64 max_cycles = tk->tkr_mono.clock->max_cycles;
 203	const char *name = tk->tkr_mono.clock->name;
 204
 205	if (offset > max_cycles) {
 206		printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
 207				offset, name, max_cycles);
 208		printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
 209	} else {
 210		if (offset > (max_cycles >> 1)) {
 211			printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
 212					offset, name, max_cycles >> 1);
 213			printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
 214		}
 215	}
 216
 217	if (tk->underflow_seen) {
 218		if (jiffies - tk->last_warning > WARNING_FREQ) {
 219			printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
 220			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 221			printk_deferred("         Your kernel is probably still fine.\n");
 222			tk->last_warning = jiffies;
 223		}
 224		tk->underflow_seen = 0;
 225	}
 226
 227	if (tk->overflow_seen) {
 228		if (jiffies - tk->last_warning > WARNING_FREQ) {
 229			printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
 230			printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 231			printk_deferred("         Your kernel is probably still fine.\n");
 232			tk->last_warning = jiffies;
 233		}
 234		tk->overflow_seen = 0;
 235	}
 236}
 237
 238static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 239{
 240	struct timekeeper *tk = &tk_core.timekeeper;
 241	u64 now, last, mask, max, delta;
 242	unsigned int seq;
 243
 244	/*
 245	 * Since we're called holding a seqcount, the data may shift
 246	 * under us while we're doing the calculation. This can cause
 247	 * false positives, since we'd note a problem but throw the
 248	 * results away. So nest another seqcount here to atomically
 249	 * grab the points we are checking with.
 250	 */
 251	do {
 252		seq = read_seqcount_begin(&tk_core.seq);
 253		now = tk_clock_read(tkr);
 254		last = tkr->cycle_last;
 255		mask = tkr->mask;
 256		max = tkr->clock->max_cycles;
 257	} while (read_seqcount_retry(&tk_core.seq, seq));
 258
 259	delta = clocksource_delta(now, last, mask);
 260
 261	/*
 262	 * Try to catch underflows by checking if we are seeing small
 263	 * mask-relative negative values.
 264	 */
 265	if (unlikely((~delta & mask) < (mask >> 3))) {
 266		tk->underflow_seen = 1;
 267		delta = 0;
 268	}
 269
 270	/* Cap delta value to the max_cycles values to avoid mult overflows */
 271	if (unlikely(delta > max)) {
 272		tk->overflow_seen = 1;
 273		delta = tkr->clock->max_cycles;
 274	}
 275
 276	return delta;
 277}
 278#else
 279static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
 280{
 281}
 282static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
 283{
 284	u64 cycle_now, delta;
 285
 286	/* read clocksource */
 287	cycle_now = tk_clock_read(tkr);
 288
 289	/* calculate the delta since the last update_wall_time */
 290	delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 291
 292	return delta;
 293}
 294#endif
 295
 296/**
 297 * tk_setup_internals - Set up internals to use clocksource clock.
 298 *
 299 * @tk:		The target timekeeper to setup.
 300 * @clock:		Pointer to clocksource.
 301 *
 302 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
 303 * pair and interval request.
 304 *
 305 * Unless you're the timekeeping code, you should not be using this!
 306 */
 307static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
 308{
 309	u64 interval;
 310	u64 tmp, ntpinterval;
 311	struct clocksource *old_clock;
 312
 313	++tk->cs_was_changed_seq;
 314	old_clock = tk->tkr_mono.clock;
 315	tk->tkr_mono.clock = clock;
 316	tk->tkr_mono.mask = clock->mask;
 317	tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
 318
 319	tk->tkr_raw.clock = clock;
 320	tk->tkr_raw.mask = clock->mask;
 321	tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
 322
 323	/* Do the ns -> cycle conversion first, using original mult */
 324	tmp = NTP_INTERVAL_LENGTH;
 325	tmp <<= clock->shift;
 326	ntpinterval = tmp;
 327	tmp += clock->mult/2;
 328	do_div(tmp, clock->mult);
 329	if (tmp == 0)
 330		tmp = 1;
 331
 332	interval = (u64) tmp;
 333	tk->cycle_interval = interval;
 334
 335	/* Go back from cycles -> shifted ns */
 336	tk->xtime_interval = interval * clock->mult;
 337	tk->xtime_remainder = ntpinterval - tk->xtime_interval;
 338	tk->raw_interval = interval * clock->mult;
 339
 340	 /* if changing clocks, convert xtime_nsec shift units */
 341	if (old_clock) {
 342		int shift_change = clock->shift - old_clock->shift;
 343		if (shift_change < 0) {
 344			tk->tkr_mono.xtime_nsec >>= -shift_change;
 345			tk->tkr_raw.xtime_nsec >>= -shift_change;
 346		} else {
 347			tk->tkr_mono.xtime_nsec <<= shift_change;
 348			tk->tkr_raw.xtime_nsec <<= shift_change;
 349		}
 350	}
 351
 352	tk->tkr_mono.shift = clock->shift;
 353	tk->tkr_raw.shift = clock->shift;
 354
 355	tk->ntp_error = 0;
 356	tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 357	tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
 358
 359	/*
 360	 * The timekeeper keeps its own mult values for the currently
 361	 * active clocksource. These value will be adjusted via NTP
 362	 * to counteract clock drifting.
 363	 */
 364	tk->tkr_mono.mult = clock->mult;
 365	tk->tkr_raw.mult = clock->mult;
 366	tk->ntp_err_mult = 0;
 367	tk->skip_second_overflow = 0;
 368}
 369
 370/* Timekeeper helper functions. */
 371
 
 
 
 
 
 
 
 372static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
 373{
 374	u64 nsec;
 375
 376	nsec = delta * tkr->mult + tkr->xtime_nsec;
 377	nsec >>= tkr->shift;
 378
 379	return nsec;
 
 380}
 381
 382static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
 383{
 384	u64 delta;
 385
 386	delta = timekeeping_get_delta(tkr);
 387	return timekeeping_delta_to_ns(tkr, delta);
 388}
 389
 390static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
 391{
 392	u64 delta;
 393
 394	/* calculate the delta since the last update_wall_time */
 395	delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 396	return timekeeping_delta_to_ns(tkr, delta);
 397}
 398
 399/**
 400 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
 401 * @tkr: Timekeeping readout base from which we take the update
 402 * @tkf: Pointer to NMI safe timekeeper
 403 *
 404 * We want to use this from any context including NMI and tracing /
 405 * instrumenting the timekeeping code itself.
 406 *
 407 * Employ the latch technique; see @raw_write_seqcount_latch.
 408 *
 409 * So if a NMI hits the update of base[0] then it will use base[1]
 410 * which is still consistent. In the worst case this can result is a
 411 * slightly wrong timestamp (a few nanoseconds). See
 412 * @ktime_get_mono_fast_ns.
 413 */
 414static void update_fast_timekeeper(const struct tk_read_base *tkr,
 415				   struct tk_fast *tkf)
 416{
 417	struct tk_read_base *base = tkf->base;
 418
 419	/* Force readers off to base[1] */
 420	raw_write_seqcount_latch(&tkf->seq);
 421
 422	/* Update base[0] */
 423	memcpy(base, tkr, sizeof(*base));
 424
 425	/* Force readers back to base[0] */
 426	raw_write_seqcount_latch(&tkf->seq);
 427
 428	/* Update base[1] */
 429	memcpy(base + 1, base, sizeof(*base));
 430}
 431
 432static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
 433{
 434	struct tk_read_base *tkr;
 435	unsigned int seq;
 436	u64 now;
 437
 438	do {
 439		seq = raw_read_seqcount_latch(&tkf->seq);
 440		tkr = tkf->base + (seq & 0x01);
 441		now = ktime_to_ns(tkr->base);
 442
 443		now += timekeeping_delta_to_ns(tkr,
 444				clocksource_delta(
 445					tk_clock_read(tkr),
 446					tkr->cycle_last,
 447					tkr->mask));
 448	} while (read_seqcount_latch_retry(&tkf->seq, seq));
 449
 450	return now;
 451}
 452
 453/**
 454 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
 455 *
 456 * This timestamp is not guaranteed to be monotonic across an update.
 457 * The timestamp is calculated by:
 458 *
 459 *	now = base_mono + clock_delta * slope
 460 *
 461 * So if the update lowers the slope, readers who are forced to the
 462 * not yet updated second array are still using the old steeper slope.
 463 *
 464 * tmono
 465 * ^
 466 * |    o  n
 467 * |   o n
 468 * |  u
 469 * | o
 470 * |o
 471 * |12345678---> reader order
 472 *
 473 * o = old slope
 474 * u = update
 475 * n = new slope
 476 *
 477 * So reader 6 will observe time going backwards versus reader 5.
 478 *
 479 * While other CPUs are likely to be able to observe that, the only way
 480 * for a CPU local observation is when an NMI hits in the middle of
 481 * the update. Timestamps taken from that NMI context might be ahead
 482 * of the following timestamps. Callers need to be aware of that and
 483 * deal with it.
 484 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 485u64 ktime_get_mono_fast_ns(void)
 486{
 487	return __ktime_get_fast_ns(&tk_fast_mono);
 488}
 489EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 490
 491/**
 492 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
 493 *
 494 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
 495 * conversion factor is not affected by NTP/PTP correction.
 496 */
 497u64 ktime_get_raw_fast_ns(void)
 498{
 499	return __ktime_get_fast_ns(&tk_fast_raw);
 500}
 501EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
 502
 503/**
 504 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
 505 *
 506 * To keep it NMI safe since we're accessing from tracing, we're not using a
 507 * separate timekeeper with updates to monotonic clock and boot offset
 508 * protected with seqcounts. This has the following minor side effects:
 509 *
 510 * (1) Its possible that a timestamp be taken after the boot offset is updated
 511 * but before the timekeeper is updated. If this happens, the new boot offset
 512 * is added to the old timekeeping making the clock appear to update slightly
 513 * earlier:
 514 *    CPU 0                                        CPU 1
 515 *    timekeeping_inject_sleeptime64()
 516 *    __timekeeping_inject_sleeptime(tk, delta);
 517 *                                                 timestamp();
 518 *    timekeeping_update(tk, TK_CLEAR_NTP...);
 519 *
 520 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
 521 * partially updated.  Since the tk->offs_boot update is a rare event, this
 522 * should be a rare occurrence which postprocessing should be able to handle.
 523 *
 524 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns()
 525 * apply as well.
 526 */
 527u64 notrace ktime_get_boot_fast_ns(void)
 528{
 529	struct timekeeper *tk = &tk_core.timekeeper;
 530
 531	return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
 532}
 533EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
 534
 535static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
 
 
 
 
 536{
 537	struct tk_read_base *tkr;
 538	u64 basem, baser, delta;
 539	unsigned int seq;
 
 540
 541	do {
 542		seq = raw_read_seqcount_latch(&tkf->seq);
 543		tkr = tkf->base + (seq & 0x01);
 544		basem = ktime_to_ns(tkr->base);
 545		baser = ktime_to_ns(tkr->base_real);
 546
 547		delta = timekeeping_delta_to_ns(tkr,
 548				clocksource_delta(tk_clock_read(tkr),
 549				tkr->cycle_last, tkr->mask));
 550	} while (read_seqcount_latch_retry(&tkf->seq, seq));
 551
 552	if (mono)
 553		*mono = basem + delta;
 554	return baser + delta;
 555}
 556
 557/**
 558 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
 559 *
 560 * See ktime_get_fast_ns() for documentation of the time stamp ordering.
 561 */
 562u64 ktime_get_real_fast_ns(void)
 563{
 564	return __ktime_get_real_fast(&tk_fast_mono, NULL);
 565}
 566EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
 567
 568/**
 569 * ktime_get_fast_timestamps: - NMI safe timestamps
 570 * @snapshot:	Pointer to timestamp storage
 571 *
 572 * Stores clock monotonic, boottime and realtime timestamps.
 573 *
 574 * Boot time is a racy access on 32bit systems if the sleep time injection
 575 * happens late during resume and not in timekeeping_resume(). That could
 576 * be avoided by expanding struct tk_read_base with boot offset for 32bit
 577 * and adding more overhead to the update. As this is a hard to observe
 578 * once per resume event which can be filtered with reasonable effort using
 579 * the accurate mono/real timestamps, it's probably not worth the trouble.
 580 *
 581 * Aside of that it might be possible on 32 and 64 bit to observe the
 582 * following when the sleep time injection happens late:
 583 *
 584 * CPU 0				CPU 1
 585 * timekeeping_resume()
 586 * ktime_get_fast_timestamps()
 587 *	mono, real = __ktime_get_real_fast()
 588 *					inject_sleep_time()
 589 *					   update boot offset
 590 *	boot = mono + bootoffset;
 591 *
 592 * That means that boot time already has the sleep time adjustment, but
 593 * real time does not. On the next readout both are in sync again.
 594 *
 595 * Preventing this for 64bit is not really feasible without destroying the
 596 * careful cache layout of the timekeeper because the sequence count and
 597 * struct tk_read_base would then need two cache lines instead of one.
 598 *
 599 * Access to the time keeper clock source is disabled across the innermost
 600 * steps of suspend/resume. The accessors still work, but the timestamps
 601 * are frozen until time keeping is resumed which happens very early.
 602 *
 603 * For regular suspend/resume there is no observable difference vs. sched
 604 * clock, but it might affect some of the nasty low level debug printks.
 605 *
 606 * OTOH, access to sched clock is not guaranteed across suspend/resume on
 607 * all systems either so it depends on the hardware in use.
 608 *
 609 * If that turns out to be a real problem then this could be mitigated by
 610 * using sched clock in a similar way as during early boot. But it's not as
 611 * trivial as on early boot because it needs some careful protection
 612 * against the clock monotonic timestamp jumping backwards on resume.
 613 */
 614void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
 615{
 616	struct timekeeper *tk = &tk_core.timekeeper;
 617
 618	snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
 619	snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
 620}
 621
 622/**
 623 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
 624 * @tk: Timekeeper to snapshot.
 625 *
 626 * It generally is unsafe to access the clocksource after timekeeping has been
 627 * suspended, so take a snapshot of the readout base of @tk and use it as the
 628 * fast timekeeper's readout base while suspended.  It will return the same
 629 * number of cycles every time until timekeeping is resumed at which time the
 630 * proper readout base for the fast timekeeper will be restored automatically.
 631 */
 632static void halt_fast_timekeeper(const struct timekeeper *tk)
 633{
 634	static struct tk_read_base tkr_dummy;
 635	const struct tk_read_base *tkr = &tk->tkr_mono;
 636
 637	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 638	cycles_at_suspend = tk_clock_read(tkr);
 639	tkr_dummy.clock = &dummy_clock;
 640	tkr_dummy.base_real = tkr->base + tk->offs_real;
 641	update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
 642
 643	tkr = &tk->tkr_raw;
 644	memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 645	tkr_dummy.clock = &dummy_clock;
 646	update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
 647}
 648
 649static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 650
 651static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 652{
 653	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 654}
 655
 656/**
 657 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 658 * @nb: Pointer to the notifier block to register
 659 */
 660int pvclock_gtod_register_notifier(struct notifier_block *nb)
 661{
 662	struct timekeeper *tk = &tk_core.timekeeper;
 663	unsigned long flags;
 664	int ret;
 665
 666	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 667	ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 668	update_pvclock_gtod(tk, true);
 669	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 670
 671	return ret;
 672}
 673EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 674
 675/**
 676 * pvclock_gtod_unregister_notifier - unregister a pvclock
 677 * timedata update listener
 678 * @nb: Pointer to the notifier block to unregister
 679 */
 680int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 681{
 682	unsigned long flags;
 683	int ret;
 684
 685	raw_spin_lock_irqsave(&timekeeper_lock, flags);
 686	ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 687	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 688
 689	return ret;
 690}
 691EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 692
 693/*
 694 * tk_update_leap_state - helper to update the next_leap_ktime
 695 */
 696static inline void tk_update_leap_state(struct timekeeper *tk)
 697{
 698	tk->next_leap_ktime = ntp_get_next_leap();
 699	if (tk->next_leap_ktime != KTIME_MAX)
 700		/* Convert to monotonic time */
 701		tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 702}
 703
 704/*
 705 * Update the ktime_t based scalar nsec members of the timekeeper
 706 */
 707static inline void tk_update_ktime_data(struct timekeeper *tk)
 708{
 709	u64 seconds;
 710	u32 nsec;
 711
 712	/*
 713	 * The xtime based monotonic readout is:
 714	 *	nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 715	 * The ktime based monotonic readout is:
 716	 *	nsec = base_mono + now();
 717	 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 718	 */
 719	seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 720	nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 721	tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 722
 723	/*
 724	 * The sum of the nanoseconds portions of xtime and
 725	 * wall_to_monotonic can be greater/equal one second. Take
 726	 * this into account before updating tk->ktime_sec.
 727	 */
 728	nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 729	if (nsec >= NSEC_PER_SEC)
 730		seconds++;
 731	tk->ktime_sec = seconds;
 732
 733	/* Update the monotonic raw base */
 734	tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
 735}
 736
 737/* must hold timekeeper_lock */
 738static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 739{
 740	if (action & TK_CLEAR_NTP) {
 741		tk->ntp_error = 0;
 742		ntp_clear();
 743	}
 744
 745	tk_update_leap_state(tk);
 746	tk_update_ktime_data(tk);
 747
 748	update_vsyscall(tk);
 749	update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 750
 751	tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
 752	update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 753	update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 754
 755	if (action & TK_CLOCK_WAS_SET)
 756		tk->clock_was_set_seq++;
 757	/*
 758	 * The mirroring of the data to the shadow-timekeeper needs
 759	 * to happen last here to ensure we don't over-write the
 760	 * timekeeper structure on the next update with stale data
 761	 */
 762	if (action & TK_MIRROR)
 763		memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 764		       sizeof(tk_core.timekeeper));
 765}
 766
 767/**
 768 * timekeeping_forward_now - update clock to the current time
 769 * @tk:		Pointer to the timekeeper to update
 770 *
 771 * Forward the current clock to update its state since the last call to
 772 * update_wall_time(). This is useful before significant clock changes,
 773 * as it avoids having to deal with this time offset explicitly.
 774 */
 775static void timekeeping_forward_now(struct timekeeper *tk)
 776{
 777	u64 cycle_now, delta;
 778
 779	cycle_now = tk_clock_read(&tk->tkr_mono);
 780	delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 781	tk->tkr_mono.cycle_last = cycle_now;
 782	tk->tkr_raw.cycle_last  = cycle_now;
 783
 784	tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 
 
 
 
 
 785	tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
 786
 
 
 
 787	tk_normalize_xtime(tk);
 788}
 789
 790/**
 791 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
 792 * @ts:		pointer to the timespec to be set
 793 *
 794 * Returns the time of day in a timespec64 (WARN if suspended).
 795 */
 796void ktime_get_real_ts64(struct timespec64 *ts)
 797{
 798	struct timekeeper *tk = &tk_core.timekeeper;
 799	unsigned int seq;
 800	u64 nsecs;
 801
 802	WARN_ON(timekeeping_suspended);
 803
 804	do {
 805		seq = read_seqcount_begin(&tk_core.seq);
 806
 807		ts->tv_sec = tk->xtime_sec;
 808		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 809
 810	} while (read_seqcount_retry(&tk_core.seq, seq));
 811
 812	ts->tv_nsec = 0;
 813	timespec64_add_ns(ts, nsecs);
 814}
 815EXPORT_SYMBOL(ktime_get_real_ts64);
 816
 817ktime_t ktime_get(void)
 818{
 819	struct timekeeper *tk = &tk_core.timekeeper;
 820	unsigned int seq;
 821	ktime_t base;
 822	u64 nsecs;
 823
 824	WARN_ON(timekeeping_suspended);
 825
 826	do {
 827		seq = read_seqcount_begin(&tk_core.seq);
 828		base = tk->tkr_mono.base;
 829		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 830
 831	} while (read_seqcount_retry(&tk_core.seq, seq));
 832
 833	return ktime_add_ns(base, nsecs);
 834}
 835EXPORT_SYMBOL_GPL(ktime_get);
 836
 837u32 ktime_get_resolution_ns(void)
 838{
 839	struct timekeeper *tk = &tk_core.timekeeper;
 840	unsigned int seq;
 841	u32 nsecs;
 842
 843	WARN_ON(timekeeping_suspended);
 844
 845	do {
 846		seq = read_seqcount_begin(&tk_core.seq);
 847		nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 848	} while (read_seqcount_retry(&tk_core.seq, seq));
 849
 850	return nsecs;
 851}
 852EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 853
 854static ktime_t *offsets[TK_OFFS_MAX] = {
 855	[TK_OFFS_REAL]	= &tk_core.timekeeper.offs_real,
 856	[TK_OFFS_BOOT]	= &tk_core.timekeeper.offs_boot,
 857	[TK_OFFS_TAI]	= &tk_core.timekeeper.offs_tai,
 858};
 859
 860ktime_t ktime_get_with_offset(enum tk_offsets offs)
 861{
 862	struct timekeeper *tk = &tk_core.timekeeper;
 863	unsigned int seq;
 864	ktime_t base, *offset = offsets[offs];
 865	u64 nsecs;
 866
 867	WARN_ON(timekeeping_suspended);
 868
 869	do {
 870		seq = read_seqcount_begin(&tk_core.seq);
 871		base = ktime_add(tk->tkr_mono.base, *offset);
 872		nsecs = timekeeping_get_ns(&tk->tkr_mono);
 873
 874	} while (read_seqcount_retry(&tk_core.seq, seq));
 875
 876	return ktime_add_ns(base, nsecs);
 877
 878}
 879EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 880
 881ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
 882{
 883	struct timekeeper *tk = &tk_core.timekeeper;
 884	unsigned int seq;
 885	ktime_t base, *offset = offsets[offs];
 886	u64 nsecs;
 887
 888	WARN_ON(timekeeping_suspended);
 889
 890	do {
 891		seq = read_seqcount_begin(&tk_core.seq);
 892		base = ktime_add(tk->tkr_mono.base, *offset);
 893		nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 894
 895	} while (read_seqcount_retry(&tk_core.seq, seq));
 896
 897	return ktime_add_ns(base, nsecs);
 898}
 899EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
 900
 901/**
 902 * ktime_mono_to_any() - convert monotonic time to any other time
 903 * @tmono:	time to convert.
 904 * @offs:	which offset to use
 905 */
 906ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 907{
 908	ktime_t *offset = offsets[offs];
 909	unsigned int seq;
 910	ktime_t tconv;
 911
 912	do {
 913		seq = read_seqcount_begin(&tk_core.seq);
 914		tconv = ktime_add(tmono, *offset);
 915	} while (read_seqcount_retry(&tk_core.seq, seq));
 916
 917	return tconv;
 918}
 919EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 920
 921/**
 922 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 923 */
 924ktime_t ktime_get_raw(void)
 925{
 926	struct timekeeper *tk = &tk_core.timekeeper;
 927	unsigned int seq;
 928	ktime_t base;
 929	u64 nsecs;
 930
 931	do {
 932		seq = read_seqcount_begin(&tk_core.seq);
 933		base = tk->tkr_raw.base;
 934		nsecs = timekeeping_get_ns(&tk->tkr_raw);
 935
 936	} while (read_seqcount_retry(&tk_core.seq, seq));
 937
 938	return ktime_add_ns(base, nsecs);
 939}
 940EXPORT_SYMBOL_GPL(ktime_get_raw);
 941
 942/**
 943 * ktime_get_ts64 - get the monotonic clock in timespec64 format
 944 * @ts:		pointer to timespec variable
 945 *
 946 * The function calculates the monotonic clock from the realtime
 947 * clock and the wall_to_monotonic offset and stores the result
 948 * in normalized timespec64 format in the variable pointed to by @ts.
 949 */
 950void ktime_get_ts64(struct timespec64 *ts)
 951{
 952	struct timekeeper *tk = &tk_core.timekeeper;
 953	struct timespec64 tomono;
 954	unsigned int seq;
 955	u64 nsec;
 956
 957	WARN_ON(timekeeping_suspended);
 958
 959	do {
 960		seq = read_seqcount_begin(&tk_core.seq);
 961		ts->tv_sec = tk->xtime_sec;
 962		nsec = timekeeping_get_ns(&tk->tkr_mono);
 963		tomono = tk->wall_to_monotonic;
 964
 965	} while (read_seqcount_retry(&tk_core.seq, seq));
 966
 967	ts->tv_sec += tomono.tv_sec;
 968	ts->tv_nsec = 0;
 969	timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 970}
 971EXPORT_SYMBOL_GPL(ktime_get_ts64);
 972
 973/**
 974 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 975 *
 976 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 977 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 978 * works on both 32 and 64 bit systems. On 32 bit systems the readout
 979 * covers ~136 years of uptime which should be enough to prevent
 980 * premature wrap arounds.
 981 */
 982time64_t ktime_get_seconds(void)
 983{
 984	struct timekeeper *tk = &tk_core.timekeeper;
 985
 986	WARN_ON(timekeeping_suspended);
 987	return tk->ktime_sec;
 988}
 989EXPORT_SYMBOL_GPL(ktime_get_seconds);
 990
 991/**
 992 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
 993 *
 994 * Returns the wall clock seconds since 1970.
 
 995 *
 996 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
 997 * 32bit systems the access must be protected with the sequence
 998 * counter to provide "atomic" access to the 64bit tk->xtime_sec
 999 * value.
1000 */
1001time64_t ktime_get_real_seconds(void)
1002{
1003	struct timekeeper *tk = &tk_core.timekeeper;
1004	time64_t seconds;
1005	unsigned int seq;
1006
1007	if (IS_ENABLED(CONFIG_64BIT))
1008		return tk->xtime_sec;
1009
1010	do {
1011		seq = read_seqcount_begin(&tk_core.seq);
1012		seconds = tk->xtime_sec;
1013
1014	} while (read_seqcount_retry(&tk_core.seq, seq));
1015
1016	return seconds;
1017}
1018EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1019
1020/**
1021 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1022 * but without the sequence counter protect. This internal function
1023 * is called just when timekeeping lock is already held.
1024 */
1025noinstr time64_t __ktime_get_real_seconds(void)
1026{
1027	struct timekeeper *tk = &tk_core.timekeeper;
1028
1029	return tk->xtime_sec;
1030}
1031
1032/**
1033 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1034 * @systime_snapshot:	pointer to struct receiving the system time snapshot
1035 */
1036void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1037{
1038	struct timekeeper *tk = &tk_core.timekeeper;
1039	unsigned int seq;
1040	ktime_t base_raw;
1041	ktime_t base_real;
1042	u64 nsec_raw;
1043	u64 nsec_real;
1044	u64 now;
1045
1046	WARN_ON_ONCE(timekeeping_suspended);
1047
1048	do {
1049		seq = read_seqcount_begin(&tk_core.seq);
1050		now = tk_clock_read(&tk->tkr_mono);
1051		systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1052		systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1053		systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1054		base_real = ktime_add(tk->tkr_mono.base,
1055				      tk_core.timekeeper.offs_real);
1056		base_raw = tk->tkr_raw.base;
1057		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1058		nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1059	} while (read_seqcount_retry(&tk_core.seq, seq));
1060
1061	systime_snapshot->cycles = now;
1062	systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1063	systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1064}
1065EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1066
1067/* Scale base by mult/div checking for overflow */
1068static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1069{
1070	u64 tmp, rem;
1071
1072	tmp = div64_u64_rem(*base, div, &rem);
1073
1074	if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1075	    ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1076		return -EOVERFLOW;
1077	tmp *= mult;
1078
1079	rem = div64_u64(rem * mult, div);
1080	*base = tmp + rem;
1081	return 0;
1082}
1083
1084/**
1085 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1086 * @history:			Snapshot representing start of history
1087 * @partial_history_cycles:	Cycle offset into history (fractional part)
1088 * @total_history_cycles:	Total history length in cycles
1089 * @discontinuity:		True indicates clock was set on history period
1090 * @ts:				Cross timestamp that should be adjusted using
1091 *	partial/total ratio
1092 *
1093 * Helper function used by get_device_system_crosststamp() to correct the
1094 * crosstimestamp corresponding to the start of the current interval to the
1095 * system counter value (timestamp point) provided by the driver. The
1096 * total_history_* quantities are the total history starting at the provided
1097 * reference point and ending at the start of the current interval. The cycle
1098 * count between the driver timestamp point and the start of the current
1099 * interval is partial_history_cycles.
1100 */
1101static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1102					 u64 partial_history_cycles,
1103					 u64 total_history_cycles,
1104					 bool discontinuity,
1105					 struct system_device_crosststamp *ts)
1106{
1107	struct timekeeper *tk = &tk_core.timekeeper;
1108	u64 corr_raw, corr_real;
1109	bool interp_forward;
1110	int ret;
1111
1112	if (total_history_cycles == 0 || partial_history_cycles == 0)
1113		return 0;
1114
1115	/* Interpolate shortest distance from beginning or end of history */
1116	interp_forward = partial_history_cycles > total_history_cycles / 2;
1117	partial_history_cycles = interp_forward ?
1118		total_history_cycles - partial_history_cycles :
1119		partial_history_cycles;
1120
1121	/*
1122	 * Scale the monotonic raw time delta by:
1123	 *	partial_history_cycles / total_history_cycles
1124	 */
1125	corr_raw = (u64)ktime_to_ns(
1126		ktime_sub(ts->sys_monoraw, history->raw));
1127	ret = scale64_check_overflow(partial_history_cycles,
1128				     total_history_cycles, &corr_raw);
1129	if (ret)
1130		return ret;
1131
1132	/*
1133	 * If there is a discontinuity in the history, scale monotonic raw
1134	 *	correction by:
1135	 *	mult(real)/mult(raw) yielding the realtime correction
1136	 * Otherwise, calculate the realtime correction similar to monotonic
1137	 *	raw calculation
1138	 */
1139	if (discontinuity) {
1140		corr_real = mul_u64_u32_div
1141			(corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1142	} else {
1143		corr_real = (u64)ktime_to_ns(
1144			ktime_sub(ts->sys_realtime, history->real));
1145		ret = scale64_check_overflow(partial_history_cycles,
1146					     total_history_cycles, &corr_real);
1147		if (ret)
1148			return ret;
1149	}
1150
1151	/* Fixup monotonic raw and real time time values */
1152	if (interp_forward) {
1153		ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1154		ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1155	} else {
1156		ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1157		ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1158	}
1159
1160	return 0;
1161}
1162
1163/*
1164 * cycle_between - true if test occurs chronologically between before and after
1165 */
1166static bool cycle_between(u64 before, u64 test, u64 after)
1167{
1168	if (test > before && test < after)
1169		return true;
1170	if (test < before && before > after)
1171		return true;
1172	return false;
1173}
1174
1175/**
1176 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1177 * @get_time_fn:	Callback to get simultaneous device time and
1178 *	system counter from the device driver
1179 * @ctx:		Context passed to get_time_fn()
1180 * @history_begin:	Historical reference point used to interpolate system
1181 *	time when counter provided by the driver is before the current interval
1182 * @xtstamp:		Receives simultaneously captured system and device time
1183 *
1184 * Reads a timestamp from a device and correlates it to system time
1185 */
1186int get_device_system_crosststamp(int (*get_time_fn)
1187				  (ktime_t *device_time,
1188				   struct system_counterval_t *sys_counterval,
1189				   void *ctx),
1190				  void *ctx,
1191				  struct system_time_snapshot *history_begin,
1192				  struct system_device_crosststamp *xtstamp)
1193{
1194	struct system_counterval_t system_counterval;
1195	struct timekeeper *tk = &tk_core.timekeeper;
1196	u64 cycles, now, interval_start;
1197	unsigned int clock_was_set_seq = 0;
1198	ktime_t base_real, base_raw;
1199	u64 nsec_real, nsec_raw;
1200	u8 cs_was_changed_seq;
1201	unsigned int seq;
1202	bool do_interp;
1203	int ret;
1204
1205	do {
1206		seq = read_seqcount_begin(&tk_core.seq);
1207		/*
1208		 * Try to synchronously capture device time and a system
1209		 * counter value calling back into the device driver
1210		 */
1211		ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1212		if (ret)
1213			return ret;
1214
1215		/*
1216		 * Verify that the clocksource associated with the captured
1217		 * system counter value is the same as the currently installed
1218		 * timekeeper clocksource
1219		 */
1220		if (tk->tkr_mono.clock != system_counterval.cs)
1221			return -ENODEV;
1222		cycles = system_counterval.cycles;
1223
1224		/*
1225		 * Check whether the system counter value provided by the
1226		 * device driver is on the current timekeeping interval.
1227		 */
1228		now = tk_clock_read(&tk->tkr_mono);
1229		interval_start = tk->tkr_mono.cycle_last;
1230		if (!cycle_between(interval_start, cycles, now)) {
1231			clock_was_set_seq = tk->clock_was_set_seq;
1232			cs_was_changed_seq = tk->cs_was_changed_seq;
1233			cycles = interval_start;
1234			do_interp = true;
1235		} else {
1236			do_interp = false;
1237		}
1238
1239		base_real = ktime_add(tk->tkr_mono.base,
1240				      tk_core.timekeeper.offs_real);
1241		base_raw = tk->tkr_raw.base;
1242
1243		nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1244						     system_counterval.cycles);
1245		nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1246						    system_counterval.cycles);
1247	} while (read_seqcount_retry(&tk_core.seq, seq));
1248
1249	xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1250	xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1251
1252	/*
1253	 * Interpolate if necessary, adjusting back from the start of the
1254	 * current interval
1255	 */
1256	if (do_interp) {
1257		u64 partial_history_cycles, total_history_cycles;
1258		bool discontinuity;
1259
1260		/*
1261		 * Check that the counter value occurs after the provided
1262		 * history reference and that the history doesn't cross a
1263		 * clocksource change
1264		 */
1265		if (!history_begin ||
1266		    !cycle_between(history_begin->cycles,
1267				   system_counterval.cycles, cycles) ||
1268		    history_begin->cs_was_changed_seq != cs_was_changed_seq)
1269			return -EINVAL;
1270		partial_history_cycles = cycles - system_counterval.cycles;
1271		total_history_cycles = cycles - history_begin->cycles;
1272		discontinuity =
1273			history_begin->clock_was_set_seq != clock_was_set_seq;
1274
1275		ret = adjust_historical_crosststamp(history_begin,
1276						    partial_history_cycles,
1277						    total_history_cycles,
1278						    discontinuity, xtstamp);
1279		if (ret)
1280			return ret;
1281	}
1282
1283	return 0;
1284}
1285EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1286
1287/**
1288 * do_settimeofday64 - Sets the time of day.
1289 * @ts:     pointer to the timespec64 variable containing the new time
1290 *
1291 * Sets the time of day to the new time and update NTP and notify hrtimers
1292 */
1293int do_settimeofday64(const struct timespec64 *ts)
1294{
1295	struct timekeeper *tk = &tk_core.timekeeper;
1296	struct timespec64 ts_delta, xt;
1297	unsigned long flags;
1298	int ret = 0;
1299
1300	if (!timespec64_valid_settod(ts))
1301		return -EINVAL;
1302
1303	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1304	write_seqcount_begin(&tk_core.seq);
1305
1306	timekeeping_forward_now(tk);
1307
1308	xt = tk_xtime(tk);
1309	ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1310	ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1311
1312	if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1313		ret = -EINVAL;
1314		goto out;
1315	}
1316
1317	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1318
1319	tk_set_xtime(tk, ts);
1320out:
1321	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1322
1323	write_seqcount_end(&tk_core.seq);
1324	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1325
1326	/* signal hrtimers about time change */
1327	clock_was_set();
1328
1329	if (!ret)
1330		audit_tk_injoffset(ts_delta);
1331
1332	return ret;
1333}
1334EXPORT_SYMBOL(do_settimeofday64);
1335
1336/**
1337 * timekeeping_inject_offset - Adds or subtracts from the current time.
1338 * @ts:		Pointer to the timespec variable containing the offset
1339 *
1340 * Adds or subtracts an offset value from the current time.
1341 */
1342static int timekeeping_inject_offset(const struct timespec64 *ts)
1343{
1344	struct timekeeper *tk = &tk_core.timekeeper;
1345	unsigned long flags;
1346	struct timespec64 tmp;
1347	int ret = 0;
1348
1349	if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1350		return -EINVAL;
1351
1352	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1353	write_seqcount_begin(&tk_core.seq);
1354
1355	timekeeping_forward_now(tk);
1356
1357	/* Make sure the proposed value is valid */
1358	tmp = timespec64_add(tk_xtime(tk), *ts);
1359	if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1360	    !timespec64_valid_settod(&tmp)) {
1361		ret = -EINVAL;
1362		goto error;
1363	}
1364
1365	tk_xtime_add(tk, ts);
1366	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1367
1368error: /* even if we error out, we forwarded the time, so call update */
1369	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1370
1371	write_seqcount_end(&tk_core.seq);
1372	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1373
1374	/* signal hrtimers about time change */
1375	clock_was_set();
1376
1377	return ret;
1378}
1379
1380/*
1381 * Indicates if there is an offset between the system clock and the hardware
1382 * clock/persistent clock/rtc.
1383 */
1384int persistent_clock_is_local;
1385
1386/*
1387 * Adjust the time obtained from the CMOS to be UTC time instead of
1388 * local time.
1389 *
1390 * This is ugly, but preferable to the alternatives.  Otherwise we
1391 * would either need to write a program to do it in /etc/rc (and risk
1392 * confusion if the program gets run more than once; it would also be
1393 * hard to make the program warp the clock precisely n hours)  or
1394 * compile in the timezone information into the kernel.  Bad, bad....
1395 *
1396 *						- TYT, 1992-01-01
1397 *
1398 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1399 * as real UNIX machines always do it. This avoids all headaches about
1400 * daylight saving times and warping kernel clocks.
1401 */
1402void timekeeping_warp_clock(void)
1403{
1404	if (sys_tz.tz_minuteswest != 0) {
1405		struct timespec64 adjust;
1406
1407		persistent_clock_is_local = 1;
1408		adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1409		adjust.tv_nsec = 0;
1410		timekeeping_inject_offset(&adjust);
1411	}
1412}
1413
1414/*
1415 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
 
1416 */
1417static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1418{
1419	tk->tai_offset = tai_offset;
1420	tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1421}
1422
1423/*
1424 * change_clocksource - Swaps clocksources if a new one is available
1425 *
1426 * Accumulates current time interval and initializes new clocksource
1427 */
1428static int change_clocksource(void *data)
1429{
1430	struct timekeeper *tk = &tk_core.timekeeper;
1431	struct clocksource *new, *old = NULL;
1432	unsigned long flags;
1433	bool change = false;
1434
1435	new = (struct clocksource *) data;
1436
 
 
 
 
1437	/*
1438	 * If the cs is in module, get a module reference. Succeeds
1439	 * for built-in code (owner == NULL) as well.
1440	 */
1441	if (try_module_get(new->owner)) {
1442		if (!new->enable || new->enable(new) == 0)
1443			change = true;
1444		else
 
 
 
 
1445			module_put(new->owner);
 
1446	}
1447
1448	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1449	write_seqcount_begin(&tk_core.seq);
1450
1451	timekeeping_forward_now(tk);
1452
1453	if (change) {
1454		old = tk->tkr_mono.clock;
1455		tk_setup_internals(tk, new);
1456	}
1457
1458	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1459
1460	write_seqcount_end(&tk_core.seq);
1461	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1462
1463	if (old) {
1464		if (old->disable)
1465			old->disable(old);
1466
1467		module_put(old->owner);
1468	}
1469
1470	return 0;
1471}
1472
1473/**
1474 * timekeeping_notify - Install a new clock source
1475 * @clock:		pointer to the clock source
1476 *
1477 * This function is called from clocksource.c after a new, better clock
1478 * source has been registered. The caller holds the clocksource_mutex.
1479 */
1480int timekeeping_notify(struct clocksource *clock)
1481{
1482	struct timekeeper *tk = &tk_core.timekeeper;
1483
1484	if (tk->tkr_mono.clock == clock)
1485		return 0;
1486	stop_machine(change_clocksource, clock, NULL);
1487	tick_clock_notify();
1488	return tk->tkr_mono.clock == clock ? 0 : -1;
1489}
1490
1491/**
1492 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1493 * @ts:		pointer to the timespec64 to be set
1494 *
1495 * Returns the raw monotonic time (completely un-modified by ntp)
1496 */
1497void ktime_get_raw_ts64(struct timespec64 *ts)
1498{
1499	struct timekeeper *tk = &tk_core.timekeeper;
1500	unsigned int seq;
1501	u64 nsecs;
1502
1503	do {
1504		seq = read_seqcount_begin(&tk_core.seq);
1505		ts->tv_sec = tk->raw_sec;
1506		nsecs = timekeeping_get_ns(&tk->tkr_raw);
1507
1508	} while (read_seqcount_retry(&tk_core.seq, seq));
1509
1510	ts->tv_nsec = 0;
1511	timespec64_add_ns(ts, nsecs);
1512}
1513EXPORT_SYMBOL(ktime_get_raw_ts64);
1514
1515
1516/**
1517 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1518 */
1519int timekeeping_valid_for_hres(void)
1520{
1521	struct timekeeper *tk = &tk_core.timekeeper;
1522	unsigned int seq;
1523	int ret;
1524
1525	do {
1526		seq = read_seqcount_begin(&tk_core.seq);
1527
1528		ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1529
1530	} while (read_seqcount_retry(&tk_core.seq, seq));
1531
1532	return ret;
1533}
1534
1535/**
1536 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1537 */
1538u64 timekeeping_max_deferment(void)
1539{
1540	struct timekeeper *tk = &tk_core.timekeeper;
1541	unsigned int seq;
1542	u64 ret;
1543
1544	do {
1545		seq = read_seqcount_begin(&tk_core.seq);
1546
1547		ret = tk->tkr_mono.clock->max_idle_ns;
1548
1549	} while (read_seqcount_retry(&tk_core.seq, seq));
1550
1551	return ret;
1552}
1553
1554/**
1555 * read_persistent_clock64 -  Return time from the persistent clock.
1556 * @ts: Pointer to the storage for the readout value
1557 *
1558 * Weak dummy function for arches that do not yet support it.
1559 * Reads the time from the battery backed persistent clock.
1560 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1561 *
1562 *  XXX - Do be sure to remove it once all arches implement it.
1563 */
1564void __weak read_persistent_clock64(struct timespec64 *ts)
1565{
1566	ts->tv_sec = 0;
1567	ts->tv_nsec = 0;
1568}
1569
1570/**
1571 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1572 *                                        from the boot.
1573 *
1574 * Weak dummy function for arches that do not yet support it.
1575 * @wall_time:	- current time as returned by persistent clock
1576 * @boot_offset: - offset that is defined as wall_time - boot_time
1577 *
1578 * The default function calculates offset based on the current value of
1579 * local_clock(). This way architectures that support sched_clock() but don't
1580 * support dedicated boot time clock will provide the best estimate of the
1581 * boot time.
1582 */
1583void __weak __init
1584read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1585				     struct timespec64 *boot_offset)
1586{
1587	read_persistent_clock64(wall_time);
1588	*boot_offset = ns_to_timespec64(local_clock());
1589}
1590
1591/*
1592 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1593 *
1594 * The flag starts of false and is only set when a suspend reaches
1595 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1596 * timekeeper clocksource is not stopping across suspend and has been
1597 * used to update sleep time. If the timekeeper clocksource has stopped
1598 * then the flag stays true and is used by the RTC resume code to decide
1599 * whether sleeptime must be injected and if so the flag gets false then.
1600 *
1601 * If a suspend fails before reaching timekeeping_resume() then the flag
1602 * stays false and prevents erroneous sleeptime injection.
1603 */
1604static bool suspend_timing_needed;
1605
1606/* Flag for if there is a persistent clock on this platform */
1607static bool persistent_clock_exists;
1608
1609/*
1610 * timekeeping_init - Initializes the clocksource and common timekeeping values
1611 */
1612void __init timekeeping_init(void)
1613{
1614	struct timespec64 wall_time, boot_offset, wall_to_mono;
1615	struct timekeeper *tk = &tk_core.timekeeper;
1616	struct clocksource *clock;
1617	unsigned long flags;
1618
1619	read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1620	if (timespec64_valid_settod(&wall_time) &&
1621	    timespec64_to_ns(&wall_time) > 0) {
1622		persistent_clock_exists = true;
1623	} else if (timespec64_to_ns(&wall_time) != 0) {
1624		pr_warn("Persistent clock returned invalid value");
1625		wall_time = (struct timespec64){0};
1626	}
1627
1628	if (timespec64_compare(&wall_time, &boot_offset) < 0)
1629		boot_offset = (struct timespec64){0};
1630
1631	/*
1632	 * We want set wall_to_mono, so the following is true:
1633	 * wall time + wall_to_mono = boot time
1634	 */
1635	wall_to_mono = timespec64_sub(boot_offset, wall_time);
1636
1637	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1638	write_seqcount_begin(&tk_core.seq);
1639	ntp_init();
1640
1641	clock = clocksource_default_clock();
1642	if (clock->enable)
1643		clock->enable(clock);
1644	tk_setup_internals(tk, clock);
1645
1646	tk_set_xtime(tk, &wall_time);
1647	tk->raw_sec = 0;
1648
1649	tk_set_wall_to_mono(tk, wall_to_mono);
1650
1651	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1652
1653	write_seqcount_end(&tk_core.seq);
1654	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1655}
1656
1657/* time in seconds when suspend began for persistent clock */
1658static struct timespec64 timekeeping_suspend_time;
1659
1660/**
1661 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1662 * @tk:		Pointer to the timekeeper to be updated
1663 * @delta:	Pointer to the delta value in timespec64 format
1664 *
1665 * Takes a timespec offset measuring a suspend interval and properly
1666 * adds the sleep offset to the timekeeping variables.
1667 */
1668static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1669					   const struct timespec64 *delta)
1670{
1671	if (!timespec64_valid_strict(delta)) {
1672		printk_deferred(KERN_WARNING
1673				"__timekeeping_inject_sleeptime: Invalid "
1674				"sleep delta value!\n");
1675		return;
1676	}
1677	tk_xtime_add(tk, delta);
1678	tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1679	tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1680	tk_debug_account_sleep_time(delta);
1681}
1682
1683#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1684/**
1685 * We have three kinds of time sources to use for sleep time
1686 * injection, the preference order is:
1687 * 1) non-stop clocksource
1688 * 2) persistent clock (ie: RTC accessible when irqs are off)
1689 * 3) RTC
1690 *
1691 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1692 * If system has neither 1) nor 2), 3) will be used finally.
1693 *
1694 *
1695 * If timekeeping has injected sleeptime via either 1) or 2),
1696 * 3) becomes needless, so in this case we don't need to call
1697 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1698 * means.
1699 */
1700bool timekeeping_rtc_skipresume(void)
1701{
1702	return !suspend_timing_needed;
1703}
1704
1705/**
1706 * 1) can be determined whether to use or not only when doing
1707 * timekeeping_resume() which is invoked after rtc_suspend(),
1708 * so we can't skip rtc_suspend() surely if system has 1).
1709 *
1710 * But if system has 2), 2) will definitely be used, so in this
1711 * case we don't need to call rtc_suspend(), and this is what
1712 * timekeeping_rtc_skipsuspend() means.
1713 */
1714bool timekeeping_rtc_skipsuspend(void)
1715{
1716	return persistent_clock_exists;
1717}
1718
1719/**
1720 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1721 * @delta: pointer to a timespec64 delta value
1722 *
1723 * This hook is for architectures that cannot support read_persistent_clock64
1724 * because their RTC/persistent clock is only accessible when irqs are enabled.
1725 * and also don't have an effective nonstop clocksource.
1726 *
1727 * This function should only be called by rtc_resume(), and allows
1728 * a suspend offset to be injected into the timekeeping values.
1729 */
1730void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1731{
1732	struct timekeeper *tk = &tk_core.timekeeper;
1733	unsigned long flags;
1734
1735	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1736	write_seqcount_begin(&tk_core.seq);
1737
1738	suspend_timing_needed = false;
1739
1740	timekeeping_forward_now(tk);
1741
1742	__timekeeping_inject_sleeptime(tk, delta);
1743
1744	timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1745
1746	write_seqcount_end(&tk_core.seq);
1747	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1748
1749	/* signal hrtimers about time change */
1750	clock_was_set();
1751}
1752#endif
1753
1754/**
1755 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1756 */
1757void timekeeping_resume(void)
1758{
1759	struct timekeeper *tk = &tk_core.timekeeper;
1760	struct clocksource *clock = tk->tkr_mono.clock;
1761	unsigned long flags;
1762	struct timespec64 ts_new, ts_delta;
1763	u64 cycle_now, nsec;
1764	bool inject_sleeptime = false;
1765
1766	read_persistent_clock64(&ts_new);
1767
1768	clockevents_resume();
1769	clocksource_resume();
1770
1771	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1772	write_seqcount_begin(&tk_core.seq);
1773
1774	/*
1775	 * After system resumes, we need to calculate the suspended time and
1776	 * compensate it for the OS time. There are 3 sources that could be
1777	 * used: Nonstop clocksource during suspend, persistent clock and rtc
1778	 * device.
1779	 *
1780	 * One specific platform may have 1 or 2 or all of them, and the
1781	 * preference will be:
1782	 *	suspend-nonstop clocksource -> persistent clock -> rtc
1783	 * The less preferred source will only be tried if there is no better
1784	 * usable source. The rtc part is handled separately in rtc core code.
1785	 */
1786	cycle_now = tk_clock_read(&tk->tkr_mono);
1787	nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1788	if (nsec > 0) {
1789		ts_delta = ns_to_timespec64(nsec);
1790		inject_sleeptime = true;
1791	} else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1792		ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1793		inject_sleeptime = true;
1794	}
1795
1796	if (inject_sleeptime) {
1797		suspend_timing_needed = false;
1798		__timekeeping_inject_sleeptime(tk, &ts_delta);
1799	}
1800
1801	/* Re-base the last cycle value */
1802	tk->tkr_mono.cycle_last = cycle_now;
1803	tk->tkr_raw.cycle_last  = cycle_now;
1804
1805	tk->ntp_error = 0;
1806	timekeeping_suspended = 0;
1807	timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1808	write_seqcount_end(&tk_core.seq);
1809	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1810
1811	touch_softlockup_watchdog();
1812
1813	tick_resume();
1814	hrtimers_resume();
1815}
1816
1817int timekeeping_suspend(void)
1818{
1819	struct timekeeper *tk = &tk_core.timekeeper;
1820	unsigned long flags;
1821	struct timespec64		delta, delta_delta;
1822	static struct timespec64	old_delta;
1823	struct clocksource *curr_clock;
1824	u64 cycle_now;
1825
1826	read_persistent_clock64(&timekeeping_suspend_time);
1827
1828	/*
1829	 * On some systems the persistent_clock can not be detected at
1830	 * timekeeping_init by its return value, so if we see a valid
1831	 * value returned, update the persistent_clock_exists flag.
1832	 */
1833	if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1834		persistent_clock_exists = true;
1835
1836	suspend_timing_needed = true;
1837
1838	raw_spin_lock_irqsave(&timekeeper_lock, flags);
1839	write_seqcount_begin(&tk_core.seq);
1840	timekeeping_forward_now(tk);
1841	timekeeping_suspended = 1;
1842
1843	/*
1844	 * Since we've called forward_now, cycle_last stores the value
1845	 * just read from the current clocksource. Save this to potentially
1846	 * use in suspend timing.
1847	 */
1848	curr_clock = tk->tkr_mono.clock;
1849	cycle_now = tk->tkr_mono.cycle_last;
1850	clocksource_start_suspend_timing(curr_clock, cycle_now);
1851
1852	if (persistent_clock_exists) {
1853		/*
1854		 * To avoid drift caused by repeated suspend/resumes,
1855		 * which each can add ~1 second drift error,
1856		 * try to compensate so the difference in system time
1857		 * and persistent_clock time stays close to constant.
1858		 */
1859		delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1860		delta_delta = timespec64_sub(delta, old_delta);
1861		if (abs(delta_delta.tv_sec) >= 2) {
1862			/*
1863			 * if delta_delta is too large, assume time correction
1864			 * has occurred and set old_delta to the current delta.
1865			 */
1866			old_delta = delta;
1867		} else {
1868			/* Otherwise try to adjust old_system to compensate */
1869			timekeeping_suspend_time =
1870				timespec64_add(timekeeping_suspend_time, delta_delta);
1871		}
1872	}
1873
1874	timekeeping_update(tk, TK_MIRROR);
1875	halt_fast_timekeeper(tk);
1876	write_seqcount_end(&tk_core.seq);
1877	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1878
1879	tick_suspend();
1880	clocksource_suspend();
1881	clockevents_suspend();
1882
1883	return 0;
1884}
1885
1886/* sysfs resume/suspend bits for timekeeping */
1887static struct syscore_ops timekeeping_syscore_ops = {
1888	.resume		= timekeeping_resume,
1889	.suspend	= timekeeping_suspend,
1890};
1891
1892static int __init timekeeping_init_ops(void)
1893{
1894	register_syscore_ops(&timekeeping_syscore_ops);
1895	return 0;
1896}
1897device_initcall(timekeeping_init_ops);
1898
1899/*
1900 * Apply a multiplier adjustment to the timekeeper
1901 */
1902static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1903							 s64 offset,
1904							 s32 mult_adj)
1905{
1906	s64 interval = tk->cycle_interval;
1907
1908	if (mult_adj == 0) {
1909		return;
1910	} else if (mult_adj == -1) {
1911		interval = -interval;
1912		offset = -offset;
1913	} else if (mult_adj != 1) {
1914		interval *= mult_adj;
1915		offset *= mult_adj;
1916	}
1917
1918	/*
1919	 * So the following can be confusing.
1920	 *
1921	 * To keep things simple, lets assume mult_adj == 1 for now.
1922	 *
1923	 * When mult_adj != 1, remember that the interval and offset values
1924	 * have been appropriately scaled so the math is the same.
1925	 *
1926	 * The basic idea here is that we're increasing the multiplier
1927	 * by one, this causes the xtime_interval to be incremented by
1928	 * one cycle_interval. This is because:
1929	 *	xtime_interval = cycle_interval * mult
1930	 * So if mult is being incremented by one:
1931	 *	xtime_interval = cycle_interval * (mult + 1)
1932	 * Its the same as:
1933	 *	xtime_interval = (cycle_interval * mult) + cycle_interval
1934	 * Which can be shortened to:
1935	 *	xtime_interval += cycle_interval
1936	 *
1937	 * So offset stores the non-accumulated cycles. Thus the current
1938	 * time (in shifted nanoseconds) is:
1939	 *	now = (offset * adj) + xtime_nsec
1940	 * Now, even though we're adjusting the clock frequency, we have
1941	 * to keep time consistent. In other words, we can't jump back
1942	 * in time, and we also want to avoid jumping forward in time.
1943	 *
1944	 * So given the same offset value, we need the time to be the same
1945	 * both before and after the freq adjustment.
1946	 *	now = (offset * adj_1) + xtime_nsec_1
1947	 *	now = (offset * adj_2) + xtime_nsec_2
1948	 * So:
1949	 *	(offset * adj_1) + xtime_nsec_1 =
1950	 *		(offset * adj_2) + xtime_nsec_2
1951	 * And we know:
1952	 *	adj_2 = adj_1 + 1
1953	 * So:
1954	 *	(offset * adj_1) + xtime_nsec_1 =
1955	 *		(offset * (adj_1+1)) + xtime_nsec_2
1956	 *	(offset * adj_1) + xtime_nsec_1 =
1957	 *		(offset * adj_1) + offset + xtime_nsec_2
1958	 * Canceling the sides:
1959	 *	xtime_nsec_1 = offset + xtime_nsec_2
1960	 * Which gives us:
1961	 *	xtime_nsec_2 = xtime_nsec_1 - offset
1962	 * Which simplifies to:
1963	 *	xtime_nsec -= offset
1964	 */
1965	if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1966		/* NTP adjustment caused clocksource mult overflow */
1967		WARN_ON_ONCE(1);
1968		return;
1969	}
1970
1971	tk->tkr_mono.mult += mult_adj;
1972	tk->xtime_interval += interval;
1973	tk->tkr_mono.xtime_nsec -= offset;
1974}
1975
1976/*
1977 * Adjust the timekeeper's multiplier to the correct frequency
1978 * and also to reduce the accumulated error value.
1979 */
1980static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1981{
1982	u32 mult;
1983
1984	/*
1985	 * Determine the multiplier from the current NTP tick length.
1986	 * Avoid expensive division when the tick length doesn't change.
1987	 */
1988	if (likely(tk->ntp_tick == ntp_tick_length())) {
1989		mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1990	} else {
1991		tk->ntp_tick = ntp_tick_length();
1992		mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1993				 tk->xtime_remainder, tk->cycle_interval);
1994	}
1995
1996	/*
1997	 * If the clock is behind the NTP time, increase the multiplier by 1
1998	 * to catch up with it. If it's ahead and there was a remainder in the
1999	 * tick division, the clock will slow down. Otherwise it will stay
2000	 * ahead until the tick length changes to a non-divisible value.
2001	 */
2002	tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2003	mult += tk->ntp_err_mult;
2004
2005	timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2006
2007	if (unlikely(tk->tkr_mono.clock->maxadj &&
2008		(abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2009			> tk->tkr_mono.clock->maxadj))) {
2010		printk_once(KERN_WARNING
2011			"Adjusting %s more than 11%% (%ld vs %ld)\n",
2012			tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2013			(long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2014	}
2015
2016	/*
2017	 * It may be possible that when we entered this function, xtime_nsec
2018	 * was very small.  Further, if we're slightly speeding the clocksource
2019	 * in the code above, its possible the required corrective factor to
2020	 * xtime_nsec could cause it to underflow.
2021	 *
2022	 * Now, since we have already accumulated the second and the NTP
2023	 * subsystem has been notified via second_overflow(), we need to skip
2024	 * the next update.
2025	 */
2026	if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2027		tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2028							tk->tkr_mono.shift;
2029		tk->xtime_sec--;
2030		tk->skip_second_overflow = 1;
2031	}
2032}
2033
2034/*
2035 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2036 *
2037 * Helper function that accumulates the nsecs greater than a second
2038 * from the xtime_nsec field to the xtime_secs field.
2039 * It also calls into the NTP code to handle leapsecond processing.
 
2040 */
2041static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2042{
2043	u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2044	unsigned int clock_set = 0;
2045
2046	while (tk->tkr_mono.xtime_nsec >= nsecps) {
2047		int leap;
2048
2049		tk->tkr_mono.xtime_nsec -= nsecps;
2050		tk->xtime_sec++;
2051
2052		/*
2053		 * Skip NTP update if this second was accumulated before,
2054		 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2055		 */
2056		if (unlikely(tk->skip_second_overflow)) {
2057			tk->skip_second_overflow = 0;
2058			continue;
2059		}
2060
2061		/* Figure out if its a leap sec and apply if needed */
2062		leap = second_overflow(tk->xtime_sec);
2063		if (unlikely(leap)) {
2064			struct timespec64 ts;
2065
2066			tk->xtime_sec += leap;
2067
2068			ts.tv_sec = leap;
2069			ts.tv_nsec = 0;
2070			tk_set_wall_to_mono(tk,
2071				timespec64_sub(tk->wall_to_monotonic, ts));
2072
2073			__timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2074
2075			clock_set = TK_CLOCK_WAS_SET;
2076		}
2077	}
2078	return clock_set;
2079}
2080
2081/*
2082 * logarithmic_accumulation - shifted accumulation of cycles
2083 *
2084 * This functions accumulates a shifted interval of cycles into
2085 * a shifted interval nanoseconds. Allows for O(log) accumulation
2086 * loop.
2087 *
2088 * Returns the unconsumed cycles.
2089 */
2090static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2091				    u32 shift, unsigned int *clock_set)
2092{
2093	u64 interval = tk->cycle_interval << shift;
2094	u64 snsec_per_sec;
2095
2096	/* If the offset is smaller than a shifted interval, do nothing */
2097	if (offset < interval)
2098		return offset;
2099
2100	/* Accumulate one shifted interval */
2101	offset -= interval;
2102	tk->tkr_mono.cycle_last += interval;
2103	tk->tkr_raw.cycle_last  += interval;
2104
2105	tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2106	*clock_set |= accumulate_nsecs_to_secs(tk);
2107
2108	/* Accumulate raw time */
2109	tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2110	snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2111	while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2112		tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2113		tk->raw_sec++;
2114	}
2115
2116	/* Accumulate error between NTP and clock interval */
2117	tk->ntp_error += tk->ntp_tick << shift;
2118	tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2119						(tk->ntp_error_shift + shift);
2120
2121	return offset;
2122}
2123
2124/*
2125 * timekeeping_advance - Updates the timekeeper to the current time and
2126 * current NTP tick length
2127 */
2128static void timekeeping_advance(enum timekeeping_adv_mode mode)
2129{
2130	struct timekeeper *real_tk = &tk_core.timekeeper;
2131	struct timekeeper *tk = &shadow_timekeeper;
2132	u64 offset;
2133	int shift = 0, maxshift;
2134	unsigned int clock_set = 0;
2135	unsigned long flags;
2136
2137	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2138
2139	/* Make sure we're fully resumed: */
2140	if (unlikely(timekeeping_suspended))
2141		goto out;
2142
 
 
 
 
 
 
2143	offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2144				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2145
2146	/* Check if there's really nothing to do */
2147	if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2148		goto out;
 
2149
2150	/* Do some additional sanity checking */
2151	timekeeping_check_update(tk, offset);
2152
2153	/*
2154	 * With NO_HZ we may have to accumulate many cycle_intervals
2155	 * (think "ticks") worth of time at once. To do this efficiently,
2156	 * we calculate the largest doubling multiple of cycle_intervals
2157	 * that is smaller than the offset.  We then accumulate that
2158	 * chunk in one go, and then try to consume the next smaller
2159	 * doubled multiple.
2160	 */
2161	shift = ilog2(offset) - ilog2(tk->cycle_interval);
2162	shift = max(0, shift);
2163	/* Bound shift to one less than what overflows tick_length */
2164	maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2165	shift = min(shift, maxshift);
2166	while (offset >= tk->cycle_interval) {
2167		offset = logarithmic_accumulation(tk, offset, shift,
2168							&clock_set);
2169		if (offset < tk->cycle_interval<<shift)
2170			shift--;
2171	}
2172
2173	/* Adjust the multiplier to correct NTP error */
2174	timekeeping_adjust(tk, offset);
2175
2176	/*
2177	 * Finally, make sure that after the rounding
2178	 * xtime_nsec isn't larger than NSEC_PER_SEC
2179	 */
2180	clock_set |= accumulate_nsecs_to_secs(tk);
2181
2182	write_seqcount_begin(&tk_core.seq);
2183	/*
2184	 * Update the real timekeeper.
2185	 *
2186	 * We could avoid this memcpy by switching pointers, but that
2187	 * requires changes to all other timekeeper usage sites as
2188	 * well, i.e. move the timekeeper pointer getter into the
2189	 * spinlocked/seqcount protected sections. And we trade this
2190	 * memcpy under the tk_core.seq against one before we start
2191	 * updating.
2192	 */
2193	timekeeping_update(tk, clock_set);
2194	memcpy(real_tk, tk, sizeof(*tk));
2195	/* The memcpy must come last. Do not put anything here! */
2196	write_seqcount_end(&tk_core.seq);
2197out:
2198	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2199	if (clock_set)
2200		/* Have to call _delayed version, since in irq context*/
2201		clock_was_set_delayed();
2202}
2203
2204/**
2205 * update_wall_time - Uses the current clocksource to increment the wall time
2206 *
2207 */
2208void update_wall_time(void)
2209{
2210	timekeeping_advance(TK_ADV_TICK);
2211}
2212
2213/**
2214 * getboottime64 - Return the real time of system boot.
2215 * @ts:		pointer to the timespec64 to be set
2216 *
2217 * Returns the wall-time of boot in a timespec64.
2218 *
2219 * This is based on the wall_to_monotonic offset and the total suspend
2220 * time. Calls to settimeofday will affect the value returned (which
2221 * basically means that however wrong your real time clock is at boot time,
2222 * you get the right time here).
2223 */
2224void getboottime64(struct timespec64 *ts)
2225{
2226	struct timekeeper *tk = &tk_core.timekeeper;
2227	ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2228
2229	*ts = ktime_to_timespec64(t);
2230}
2231EXPORT_SYMBOL_GPL(getboottime64);
2232
2233void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2234{
2235	struct timekeeper *tk = &tk_core.timekeeper;
2236	unsigned int seq;
2237
2238	do {
2239		seq = read_seqcount_begin(&tk_core.seq);
2240
2241		*ts = tk_xtime(tk);
2242	} while (read_seqcount_retry(&tk_core.seq, seq));
2243}
2244EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2245
2246void ktime_get_coarse_ts64(struct timespec64 *ts)
2247{
2248	struct timekeeper *tk = &tk_core.timekeeper;
2249	struct timespec64 now, mono;
2250	unsigned int seq;
2251
2252	do {
2253		seq = read_seqcount_begin(&tk_core.seq);
2254
2255		now = tk_xtime(tk);
2256		mono = tk->wall_to_monotonic;
2257	} while (read_seqcount_retry(&tk_core.seq, seq));
2258
2259	set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2260				now.tv_nsec + mono.tv_nsec);
2261}
2262EXPORT_SYMBOL(ktime_get_coarse_ts64);
2263
2264/*
2265 * Must hold jiffies_lock
2266 */
2267void do_timer(unsigned long ticks)
2268{
2269	jiffies_64 += ticks;
2270	calc_global_load();
2271}
2272
2273/**
2274 * ktime_get_update_offsets_now - hrtimer helper
2275 * @cwsseq:	pointer to check and store the clock was set sequence number
2276 * @offs_real:	pointer to storage for monotonic -> realtime offset
2277 * @offs_boot:	pointer to storage for monotonic -> boottime offset
2278 * @offs_tai:	pointer to storage for monotonic -> clock tai offset
2279 *
2280 * Returns current monotonic time and updates the offsets if the
2281 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2282 * different.
2283 *
2284 * Called from hrtimer_interrupt() or retrigger_next_event()
2285 */
2286ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2287				     ktime_t *offs_boot, ktime_t *offs_tai)
2288{
2289	struct timekeeper *tk = &tk_core.timekeeper;
2290	unsigned int seq;
2291	ktime_t base;
2292	u64 nsecs;
2293
2294	do {
2295		seq = read_seqcount_begin(&tk_core.seq);
2296
2297		base = tk->tkr_mono.base;
2298		nsecs = timekeeping_get_ns(&tk->tkr_mono);
2299		base = ktime_add_ns(base, nsecs);
2300
2301		if (*cwsseq != tk->clock_was_set_seq) {
2302			*cwsseq = tk->clock_was_set_seq;
2303			*offs_real = tk->offs_real;
2304			*offs_boot = tk->offs_boot;
2305			*offs_tai = tk->offs_tai;
2306		}
2307
2308		/* Handle leapsecond insertion adjustments */
2309		if (unlikely(base >= tk->next_leap_ktime))
2310			*offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2311
2312	} while (read_seqcount_retry(&tk_core.seq, seq));
2313
2314	return base;
2315}
2316
2317/*
2318 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2319 */
2320static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2321{
2322	if (txc->modes & ADJ_ADJTIME) {
2323		/* singleshot must not be used with any other mode bits */
2324		if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2325			return -EINVAL;
2326		if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2327		    !capable(CAP_SYS_TIME))
2328			return -EPERM;
2329	} else {
2330		/* In order to modify anything, you gotta be super-user! */
2331		if (txc->modes && !capable(CAP_SYS_TIME))
2332			return -EPERM;
2333		/*
2334		 * if the quartz is off by more than 10% then
2335		 * something is VERY wrong!
2336		 */
2337		if (txc->modes & ADJ_TICK &&
2338		    (txc->tick <  900000/USER_HZ ||
2339		     txc->tick > 1100000/USER_HZ))
2340			return -EINVAL;
2341	}
2342
2343	if (txc->modes & ADJ_SETOFFSET) {
2344		/* In order to inject time, you gotta be super-user! */
2345		if (!capable(CAP_SYS_TIME))
2346			return -EPERM;
2347
2348		/*
2349		 * Validate if a timespec/timeval used to inject a time
2350		 * offset is valid.  Offsets can be positive or negative, so
2351		 * we don't check tv_sec. The value of the timeval/timespec
2352		 * is the sum of its fields,but *NOTE*:
2353		 * The field tv_usec/tv_nsec must always be non-negative and
2354		 * we can't have more nanoseconds/microseconds than a second.
2355		 */
2356		if (txc->time.tv_usec < 0)
2357			return -EINVAL;
2358
2359		if (txc->modes & ADJ_NANO) {
2360			if (txc->time.tv_usec >= NSEC_PER_SEC)
2361				return -EINVAL;
2362		} else {
2363			if (txc->time.tv_usec >= USEC_PER_SEC)
2364				return -EINVAL;
2365		}
2366	}
2367
2368	/*
2369	 * Check for potential multiplication overflows that can
2370	 * only happen on 64-bit systems:
2371	 */
2372	if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2373		if (LLONG_MIN / PPM_SCALE > txc->freq)
2374			return -EINVAL;
2375		if (LLONG_MAX / PPM_SCALE < txc->freq)
2376			return -EINVAL;
2377	}
2378
2379	return 0;
2380}
2381
2382
2383/**
2384 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2385 */
2386int do_adjtimex(struct __kernel_timex *txc)
2387{
2388	struct timekeeper *tk = &tk_core.timekeeper;
2389	struct audit_ntp_data ad;
2390	unsigned long flags;
2391	struct timespec64 ts;
2392	s32 orig_tai, tai;
2393	int ret;
2394
2395	/* Validate the data before disabling interrupts */
2396	ret = timekeeping_validate_timex(txc);
2397	if (ret)
2398		return ret;
2399
2400	if (txc->modes & ADJ_SETOFFSET) {
2401		struct timespec64 delta;
2402		delta.tv_sec  = txc->time.tv_sec;
2403		delta.tv_nsec = txc->time.tv_usec;
2404		if (!(txc->modes & ADJ_NANO))
2405			delta.tv_nsec *= 1000;
2406		ret = timekeeping_inject_offset(&delta);
2407		if (ret)
2408			return ret;
2409
2410		audit_tk_injoffset(delta);
2411	}
2412
2413	audit_ntp_init(&ad);
2414
2415	ktime_get_real_ts64(&ts);
2416
2417	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2418	write_seqcount_begin(&tk_core.seq);
2419
2420	orig_tai = tai = tk->tai_offset;
2421	ret = __do_adjtimex(txc, &ts, &tai, &ad);
2422
2423	if (tai != orig_tai) {
2424		__timekeeping_set_tai_offset(tk, tai);
2425		timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2426	}
2427	tk_update_leap_state(tk);
2428
2429	write_seqcount_end(&tk_core.seq);
2430	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2431
2432	audit_ntp_log(&ad);
2433
2434	/* Update the multiplier immediately if frequency was set directly */
2435	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2436		timekeeping_advance(TK_ADV_FREQ);
2437
2438	if (tai != orig_tai)
2439		clock_was_set();
2440
2441	ntp_notify_cmos_timer();
2442
2443	return ret;
2444}
2445
2446#ifdef CONFIG_NTP_PPS
2447/**
2448 * hardpps() - Accessor function to NTP __hardpps function
2449 */
2450void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2451{
2452	unsigned long flags;
2453
2454	raw_spin_lock_irqsave(&timekeeper_lock, flags);
2455	write_seqcount_begin(&tk_core.seq);
2456
2457	__hardpps(phase_ts, raw_ts);
2458
2459	write_seqcount_end(&tk_core.seq);
2460	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2461}
2462EXPORT_SYMBOL(hardpps);
2463#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}