kernel/time/timekeeping.c

   1 /*
   2  *  linux/kernel/time/timekeeping.c
   3  *
   4  *  Kernel timekeeping code and accessor functions
   5  *
   6  *  This code was moved from linux/kernel/timer.c.
   7  *  Please see that file for copyright and history logs.
   8  *
   9  */
  10
  11 #include <linux/timekeeper_internal.h>
  12 #include <linux/module.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/percpu.h>
  15 #include <linux/init.h>
  16 #include <linux/mm.h>
  17 #include <linux/sched.h>
  18 #include <linux/syscore_ops.h>
  19 #include <linux/clocksource.h>
  20 #include <linux/jiffies.h>
  21 #include <linux/time.h>
  22 #include <linux/tick.h>
  23 #include <linux/stop_machine.h>
  24 #include <linux/pvclock_gtod.h>
  25 #include <linux/compiler.h>
  26
  27 #include "tick-internal.h"
  28 #include "ntp_internal.h"
  29 #include "timekeeping_internal.h"
  30
  31 #define TK_CLEAR_NTP            (1 << 0)
  32 #define TK_MIRROR               (1 << 1)
  33 #define TK_CLOCK_WAS_SET        (1 << 2)
  34
  35 /*
  36  * The most important data for readout fits into a single 64 byte
  37  * cache line.
  38  */
  39 static struct {
  40         seqcount_t              seq;
  41         struct timekeeper       timekeeper;
  42 } tk_core ____cacheline_aligned;
  43
  44 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
  45 static struct timekeeper shadow_timekeeper;
  46
  47 /**
  48  * struct tk_fast - NMI safe timekeeper
  49  * @seq:        Sequence counter for protecting updates. The lowest bit
  50  *              is the index for the tk_read_base array
  51  * @base:       tk_read_base array. Access is indexed by the lowest bit of
  52  *              @seq.
  53  *
  54  * See @update_fast_timekeeper() below.
  55  */
  56 struct tk_fast {
  57         seqcount_t              seq;
  58         struct tk_read_base     base[2];
  59 };
  60
  61 static struct tk_fast tk_fast_mono ____cacheline_aligned;
  62 static struct tk_fast tk_fast_raw  ____cacheline_aligned;
  63
  64 /* flag for if timekeeping is suspended */
  65 int __read_mostly timekeeping_suspended;
  66
  67 static inline void tk_normalize_xtime(struct timekeeper *tk)
  68 {
  69         while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
  70                 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
  71                 tk->xtime_sec++;
  72         }
  73 }
  74
  75 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
  76 {
  77         struct timespec64 ts;
  78
  79         ts.tv_sec = tk->xtime_sec;
  80         ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
  81         return ts;
  82 }
  83
  84 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
  85 {
  86         tk->xtime_sec = ts->tv_sec;
  87         tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
  88 }
  89
  90 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
  91 {
  92         tk->xtime_sec += ts->tv_sec;
  93         tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
  94         tk_normalize_xtime(tk);
  95 }
  96
  97 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
  98 {
  99         struct timespec64 tmp;
 100
 101         /*
 102          * Verify consistency of: offset_real = -wall_to_monotonic
 103          * before modifying anything
 104          */
 105         set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
 106                                         -tk->wall_to_monotonic.tv_nsec);
 107         WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
 108         tk->wall_to_monotonic = wtm;
 109         set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
 110         tk->offs_real = timespec64_to_ktime(tmp);
 111         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
 112 }
 113
 114 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
 115 {
 116         tk->offs_boot = ktime_add(tk->offs_boot, delta);
 117 }
 118
 119 #ifdef CONFIG_DEBUG_TIMEKEEPING
 120 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
 121
 122 static void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
 123 {
 124
 125         cycle_t max_cycles = tk->tkr_mono.clock->max_cycles;
 126         const char *name = tk->tkr_mono.clock->name;
 127
 128         if (offset > max_cycles) {
 129                 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
 130                                 offset, name, max_cycles);
 131                 printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
 132         } else {
 133                 if (offset > (max_cycles >> 1)) {
 134                         printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
 135                                         offset, name, max_cycles >> 1);
 136                         printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
 137                 }
 138         }
 139
 140         if (tk->underflow_seen) {
 141                 if (jiffies - tk->last_warning > WARNING_FREQ) {
 142                         printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
 143                         printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 144                         printk_deferred("         Your kernel is probably still fine.\n");
 145                         tk->last_warning = jiffies;
 146                 }
 147                 tk->underflow_seen = 0;
 148         }
 149
 150         if (tk->overflow_seen) {
 151                 if (jiffies - tk->last_warning > WARNING_FREQ) {
 152                         printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
 153                         printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
 154                         printk_deferred("         Your kernel is probably still fine.\n");
 155                         tk->last_warning = jiffies;
 156                 }
 157                 tk->overflow_seen = 0;
 158         }
 159 }
 160
 161 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
 162 {
 163         struct timekeeper *tk = &tk_core.timekeeper;
 164         cycle_t now, last, mask, max, delta;
 165         unsigned int seq;
 166
 167         /*
 168          * Since we're called holding a seqlock, the data may shift
 169          * under us while we're doing the calculation. This can cause
 170          * false positives, since we'd note a problem but throw the
 171          * results away. So nest another seqlock here to atomically
 172          * grab the points we are checking with.
 173          */
 174         do {
 175                 seq = read_seqcount_begin(&tk_core.seq);
 176                 now = tkr->read(tkr->clock);
 177                 last = tkr->cycle_last;
 178                 mask = tkr->mask;
 179                 max = tkr->clock->max_cycles;
 180         } while (read_seqcount_retry(&tk_core.seq, seq));
 181
 182         delta = clocksource_delta(now, last, mask);
 183
 184         /*
 185          * Try to catch underflows by checking if we are seeing small
 186          * mask-relative negative values.
 187          */
 188         if (unlikely((~delta & mask) < (mask >> 3))) {
 189                 tk->underflow_seen = 1;
 190                 delta = 0;
 191         }
 192
 193         /* Cap delta value to the max_cycles values to avoid mult overflows */
 194         if (unlikely(delta > max)) {
 195                 tk->overflow_seen = 1;
 196                 delta = tkr->clock->max_cycles;
 197         }
 198
 199         return delta;
 200 }
 201 #else
 202 static inline void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
 203 {
 204 }
 205 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
 206 {
 207         cycle_t cycle_now, delta;
 208
 209         /* read clocksource */
 210         cycle_now = tkr->read(tkr->clock);
 211
 212         /* calculate the delta since the last update_wall_time */
 213         delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 214
 215         return delta;
 216 }
 217 #endif
 218
 219 /**
 220  * tk_setup_internals - Set up internals to use clocksource clock.
 221  *
 222  * @tk:         The target timekeeper to setup.
 223  * @clock:              Pointer to clocksource.
 224  *
 225  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
 226  * pair and interval request.
 227  *
 228  * Unless you're the timekeeping code, you should not be using this!
 229  */
 230 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
 231 {
 232         cycle_t interval;
 233         u64 tmp, ntpinterval;
 234         struct clocksource *old_clock;
 235
 236         ++tk->cs_was_changed_seq;
 237         old_clock = tk->tkr_mono.clock;
 238         tk->tkr_mono.clock = clock;
 239         tk->tkr_mono.read = clock->read;
 240         tk->tkr_mono.mask = clock->mask;
 241         tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
 242
 243         tk->tkr_raw.clock = clock;
 244         tk->tkr_raw.read = clock->read;
 245         tk->tkr_raw.mask = clock->mask;
 246         tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
 247
 248         /* Do the ns -> cycle conversion first, using original mult */
 249         tmp = NTP_INTERVAL_LENGTH;
 250         tmp <<= clock->shift;
 251         ntpinterval = tmp;
 252         tmp += clock->mult/2;
 253         do_div(tmp, clock->mult);
 254         if (tmp == 0)
 255                 tmp = 1;
 256
 257         interval = (cycle_t) tmp;
 258         tk->cycle_interval = interval;
 259
 260         /* Go back from cycles -> shifted ns */
 261         tk->xtime_interval = (u64) interval * clock->mult;
 262         tk->xtime_remainder = ntpinterval - tk->xtime_interval;
 263         tk->raw_interval =
 264                 ((u64) interval * clock->mult) >> clock->shift;
 265
 266          /* if changing clocks, convert xtime_nsec shift units */
 267         if (old_clock) {
 268                 int shift_change = clock->shift - old_clock->shift;
 269                 if (shift_change < 0)
 270                         tk->tkr_mono.xtime_nsec >>= -shift_change;
 271                 else
 272                         tk->tkr_mono.xtime_nsec <<= shift_change;
 273         }
 274         tk->tkr_raw.xtime_nsec = 0;
 275
 276         tk->tkr_mono.shift = clock->shift;
 277         tk->tkr_raw.shift = clock->shift;
 278
 279         tk->ntp_error = 0;
 280         tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
 281         tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
 282
 283         /*
 284          * The timekeeper keeps its own mult values for the currently
 285          * active clocksource. These value will be adjusted via NTP
 286          * to counteract clock drifting.
 287          */
 288         tk->tkr_mono.mult = clock->mult;
 289         tk->tkr_raw.mult = clock->mult;
 290         tk->ntp_err_mult = 0;
 291 }
 292
 293 /* Timekeeper helper functions. */
 294
 295 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 296 static u32 default_arch_gettimeoffset(void) { return 0; }
 297 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
 298 #else
 299 static inline u32 arch_gettimeoffset(void) { return 0; }
 300 #endif
 301
 302 static inline s64 timekeeping_delta_to_ns(struct tk_read_base *tkr,
 303                                           cycle_t delta)
 304 {
 305         s64 nsec;
 306
 307         nsec = delta * tkr->mult + tkr->xtime_nsec;
 308         nsec >>= tkr->shift;
 309
 310         /* If arch requires, add in get_arch_timeoffset() */
 311         return nsec + arch_gettimeoffset();
 312 }
 313
 314 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
 315 {
 316         cycle_t delta;
 317
 318         delta = timekeeping_get_delta(tkr);
 319         return timekeeping_delta_to_ns(tkr, delta);
 320 }
 321
 322 static inline s64 timekeeping_cycles_to_ns(struct tk_read_base *tkr,
 323                                             cycle_t cycles)
 324 {
 325         cycle_t delta;
 326
 327         /* calculate the delta since the last update_wall_time */
 328         delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
 329         return timekeeping_delta_to_ns(tkr, delta);
 330 }
 331
 332 /**
 333  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
 334  * @tkr: Timekeeping readout base from which we take the update
 335  *
 336  * We want to use this from any context including NMI and tracing /
 337  * instrumenting the timekeeping code itself.
 338  *
 339  * Employ the latch technique; see @raw_write_seqcount_latch.
 340  *
 341  * So if a NMI hits the update of base[0] then it will use base[1]
 342  * which is still consistent. In the worst case this can result is a
 343  * slightly wrong timestamp (a few nanoseconds). See
 344  * @ktime_get_mono_fast_ns.
 345  */
 346 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
 347 {
 348         struct tk_read_base *base = tkf->base;
 349
 350         /* Force readers off to base[1] */
 351         raw_write_seqcount_latch(&tkf->seq);
 352
 353         /* Update base[0] */
 354         memcpy(base, tkr, sizeof(*base));
 355
 356         /* Force readers back to base[0] */
 357         raw_write_seqcount_latch(&tkf->seq);
 358
 359         /* Update base[1] */
 360         memcpy(base + 1, base, sizeof(*base));
 361 }
 362
 363 /**
 364  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
 365  *
 366  * This timestamp is not guaranteed to be monotonic across an update.
 367  * The timestamp is calculated by:
 368  *
 369  *      now = base_mono + clock_delta * slope
 370  *
 371  * So if the update lowers the slope, readers who are forced to the
 372  * not yet updated second array are still using the old steeper slope.
 373  *
 374  * tmono
 375  * ^
 376  * |    o  n
 377  * |   o n
 378  * |  u
 379  * | o
 380  * |o
 381  * |12345678---> reader order
 382  *
 383  * o = old slope
 384  * u = update
 385  * n = new slope
 386  *
 387  * So reader 6 will observe time going backwards versus reader 5.
 388  *
 389  * While other CPUs are likely to be able observe that, the only way
 390  * for a CPU local observation is when an NMI hits in the middle of
 391  * the update. Timestamps taken from that NMI context might be ahead
 392  * of the following timestamps. Callers need to be aware of that and
 393  * deal with it.
 394  */
 395 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
 396 {
 397         struct tk_read_base *tkr;
 398         unsigned int seq;
 399         u64 now;
 400
 401         do {
 402                 seq = raw_read_seqcount_latch(&tkf->seq);
 403                 tkr = tkf->base + (seq & 0x01);
 404                 now = ktime_to_ns(tkr->base) + timekeeping_get_ns(tkr);
 405         } while (read_seqcount_retry(&tkf->seq, seq));
 406
 407         return now;
 408 }
 409
 410 u64 ktime_get_mono_fast_ns(void)
 411 {
 412         return __ktime_get_fast_ns(&tk_fast_mono);
 413 }
 414 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 415
 416 u64 ktime_get_raw_fast_ns(void)
 417 {
 418         return __ktime_get_fast_ns(&tk_fast_raw);
 419 }
 420 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
 421
 422 /* Suspend-time cycles value for halted fast timekeeper. */
 423 static cycle_t cycles_at_suspend;
 424
 425 static cycle_t dummy_clock_read(struct clocksource *cs)
 426 {
 427         return cycles_at_suspend;
 428 }
 429
 430 /**
 431  * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
 432  * @tk: Timekeeper to snapshot.
 433  *
 434  * It generally is unsafe to access the clocksource after timekeeping has been
 435  * suspended, so take a snapshot of the readout base of @tk and use it as the
 436  * fast timekeeper's readout base while suspended.  It will return the same
 437  * number of cycles every time until timekeeping is resumed at which time the
 438  * proper readout base for the fast timekeeper will be restored automatically.
 439  */
 440 static void halt_fast_timekeeper(struct timekeeper *tk)
 441 {
 442         static struct tk_read_base tkr_dummy;
 443         struct tk_read_base *tkr = &tk->tkr_mono;
 444
 445         memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 446         cycles_at_suspend = tkr->read(tkr->clock);
 447         tkr_dummy.read = dummy_clock_read;
 448         update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
 449
 450         tkr = &tk->tkr_raw;
 451         memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
 452         tkr_dummy.read = dummy_clock_read;
 453         update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
 454 }
 455
 456 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
 457
 458 static inline void update_vsyscall(struct timekeeper *tk)
 459 {
 460         struct timespec xt, wm;
 461
 462         xt = timespec64_to_timespec(tk_xtime(tk));
 463         wm = timespec64_to_timespec(tk->wall_to_monotonic);
 464         update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
 465                             tk->tkr_mono.cycle_last);
 466 }
 467
 468 static inline void old_vsyscall_fixup(struct timekeeper *tk)
 469 {
 470         s64 remainder;
 471
 472         /*
 473         * Store only full nanoseconds into xtime_nsec after rounding
 474         * it up and add the remainder to the error difference.
 475         * XXX - This is necessary to avoid small 1ns inconsistnecies caused
 476         * by truncating the remainder in vsyscalls. However, it causes
 477         * additional work to be done in timekeeping_adjust(). Once
 478         * the vsyscall implementations are converted to use xtime_nsec
 479         * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
 480         * users are removed, this can be killed.
 481         */
 482         remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
 483         if (remainder != 0) {
 484                 tk->tkr_mono.xtime_nsec -= remainder;
 485                 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
 486                 tk->ntp_error += remainder << tk->ntp_error_shift;
 487                 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
 488         }
 489 }
 490 #else
 491 #define old_vsyscall_fixup(tk)
 492 #endif
 493
 494 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 495
 496 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 497 {
 498         raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 499 }
 500
 501 /**
 502  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 503  */
 504 int pvclock_gtod_register_notifier(struct notifier_block *nb)
 505 {
 506         struct timekeeper *tk = &tk_core.timekeeper;
 507         unsigned long flags;
 508         int ret;
 509
 510         raw_spin_lock_irqsave(&timekeeper_lock, flags);
 511         ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 512         update_pvclock_gtod(tk, true);
 513         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 514
 515         return ret;
 516 }
 517 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 518
 519 /**
 520  * pvclock_gtod_unregister_notifier - unregister a pvclock
 521  * timedata update listener
 522  */
 523 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 524 {
 525         unsigned long flags;
 526         int ret;
 527
 528         raw_spin_lock_irqsave(&timekeeper_lock, flags);
 529         ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 530         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 531
 532         return ret;
 533 }
 534 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 535
 536 /*
 537  * tk_update_leap_state - helper to update the next_leap_ktime
 538  */
 539 static inline void tk_update_leap_state(struct timekeeper *tk)
 540 {
 541         tk->next_leap_ktime = ntp_get_next_leap();
 542         if (tk->next_leap_ktime.tv64 != KTIME_MAX)
 543                 /* Convert to monotonic time */
 544                 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 545 }
 546
 547 /*
 548  * Update the ktime_t based scalar nsec members of the timekeeper
 549  */
 550 static inline void tk_update_ktime_data(struct timekeeper *tk)
 551 {
 552         u64 seconds;
 553         u32 nsec;
 554
 555         /*
 556          * The xtime based monotonic readout is:
 557          *      nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 558          * The ktime based monotonic readout is:
 559          *      nsec = base_mono + now();
 560          * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 561          */
 562         seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 563         nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 564         tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 565
 566         /* Update the monotonic raw base */
 567         tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
 568
 569         /*
 570          * The sum of the nanoseconds portions of xtime and
 571          * wall_to_monotonic can be greater/equal one second. Take
 572          * this into account before updating tk->ktime_sec.
 573          */
 574         nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 575         if (nsec >= NSEC_PER_SEC)
 576                 seconds++;
 577         tk->ktime_sec = seconds;
 578 }
 579
 580 /* must hold timekeeper_lock */
 581 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 582 {
 583         if (action & TK_CLEAR_NTP) {
 584                 tk->ntp_error = 0;
 585                 ntp_clear();
 586         }
 587
 588         tk_update_leap_state(tk);
 589         tk_update_ktime_data(tk);
 590
 591         update_vsyscall(tk);
 592         update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 593
 594         update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 595         update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 596
 597         if (action & TK_CLOCK_WAS_SET)
 598                 tk->clock_was_set_seq++;
 599         /*
 600          * The mirroring of the data to the shadow-timekeeper needs
 601          * to happen last here to ensure we don't over-write the
 602          * timekeeper structure on the next update with stale data
 603          */
 604         if (action & TK_MIRROR)
 605                 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 606                        sizeof(tk_core.timekeeper));
 607 }
 608
 609 /**
 610  * timekeeping_forward_now - update clock to the current time
 611  *
 612  * Forward the current clock to update its state since the last call to
 613  * update_wall_time(). This is useful before significant clock changes,
 614  * as it avoids having to deal with this time offset explicitly.
 615  */
 616 static void timekeeping_forward_now(struct timekeeper *tk)
 617 {
 618         struct clocksource *clock = tk->tkr_mono.clock;
 619         cycle_t cycle_now, delta;
 620         s64 nsec;
 621
 622         cycle_now = tk->tkr_mono.read(clock);
 623         delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 624         tk->tkr_mono.cycle_last = cycle_now;
 625         tk->tkr_raw.cycle_last  = cycle_now;
 626
 627         tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 628
 629         /* If arch requires, add in get_arch_timeoffset() */
 630         tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
 631
 632         tk_normalize_xtime(tk);
 633
 634         nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
 635         timespec64_add_ns(&tk->raw_time, nsec);
 636 }
 637
 638 /**
 639  * __getnstimeofday64 - Returns the time of day in a timespec64.
 640  * @ts:         pointer to the timespec to be set
 641  *
 642  * Updates the time of day in the timespec.
 643  * Returns 0 on success, or -ve when suspended (timespec will be undefined).
 644  */
 645 int __getnstimeofday64(struct timespec64 *ts)
 646 {
 647         struct timekeeper *tk = &tk_core.timekeeper;
 648         unsigned long seq;
 649         s64 nsecs = 0;
 650
 651         do {
 652                 seq = read_seqcount_begin(&tk_core.seq);
 653
 654                 ts->tv_sec = tk->xtime_sec;
 655                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 656
 657         } while (read_seqcount_retry(&tk_core.seq, seq));
 658
 659         ts->tv_nsec = 0;
 660         timespec64_add_ns(ts, nsecs);
 661
 662         /*
 663          * Do not bail out early, in case there were callers still using
 664          * the value, even in the face of the WARN_ON.
 665          */
 666         if (unlikely(timekeeping_suspended))
 667                 return -EAGAIN;
 668         return 0;
 669 }
 670 EXPORT_SYMBOL(__getnstimeofday64);
 671
 672 /**
 673  * getnstimeofday64 - Returns the time of day in a timespec64.
 674  * @ts:         pointer to the timespec64 to be set
 675  *
 676  * Returns the time of day in a timespec64 (WARN if suspended).
 677  */
 678 void getnstimeofday64(struct timespec64 *ts)
 679 {
 680         WARN_ON(__getnstimeofday64(ts));
 681 }
 682 EXPORT_SYMBOL(getnstimeofday64);
 683
 684 ktime_t ktime_get(void)
 685 {
 686         struct timekeeper *tk = &tk_core.timekeeper;
 687         unsigned int seq;
 688         ktime_t base;
 689         s64 nsecs;
 690
 691         WARN_ON(timekeeping_suspended);
 692
 693         do {
 694                 seq = read_seqcount_begin(&tk_core.seq);
 695                 base = tk->tkr_mono.base;
 696                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 697
 698         } while (read_seqcount_retry(&tk_core.seq, seq));
 699
 700         return ktime_add_ns(base, nsecs);
 701 }
 702 EXPORT_SYMBOL_GPL(ktime_get);
 703
 704 u32 ktime_get_resolution_ns(void)
 705 {
 706         struct timekeeper *tk = &tk_core.timekeeper;
 707         unsigned int seq;
 708         u32 nsecs;
 709
 710         WARN_ON(timekeeping_suspended);
 711
 712         do {
 713                 seq = read_seqcount_begin(&tk_core.seq);
 714                 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 715         } while (read_seqcount_retry(&tk_core.seq, seq));
 716
 717         return nsecs;
 718 }
 719 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 720
 721 static ktime_t *offsets[TK_OFFS_MAX] = {
 722         [TK_OFFS_REAL]  = &tk_core.timekeeper.offs_real,
 723         [TK_OFFS_BOOT]  = &tk_core.timekeeper.offs_boot,
 724         [TK_OFFS_TAI]   = &tk_core.timekeeper.offs_tai,
 725 };
 726
 727 ktime_t ktime_get_with_offset(enum tk_offsets offs)
 728 {
 729         struct timekeeper *tk = &tk_core.timekeeper;
 730         unsigned int seq;
 731         ktime_t base, *offset = offsets[offs];
 732         s64 nsecs;
 733
 734         WARN_ON(timekeeping_suspended);
 735
 736         do {
 737                 seq = read_seqcount_begin(&tk_core.seq);
 738                 base = ktime_add(tk->tkr_mono.base, *offset);
 739                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 740
 741         } while (read_seqcount_retry(&tk_core.seq, seq));
 742
 743         return ktime_add_ns(base, nsecs);
 744
 745 }
 746 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 747
 748 /**
 749  * ktime_mono_to_any() - convert mononotic time to any other time
 750  * @tmono:      time to convert.
 751  * @offs:       which offset to use
 752  */
 753 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 754 {
 755         ktime_t *offset = offsets[offs];
 756         unsigned long seq;
 757         ktime_t tconv;
 758
 759         do {
 760                 seq = read_seqcount_begin(&tk_core.seq);
 761                 tconv = ktime_add(tmono, *offset);
 762         } while (read_seqcount_retry(&tk_core.seq, seq));
 763
 764         return tconv;
 765 }
 766 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 767
 768 /**
 769  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 770  */
 771 ktime_t ktime_get_raw(void)
 772 {
 773         struct timekeeper *tk = &tk_core.timekeeper;
 774         unsigned int seq;
 775         ktime_t base;
 776         s64 nsecs;
 777
 778         do {
 779                 seq = read_seqcount_begin(&tk_core.seq);
 780                 base = tk->tkr_raw.base;
 781                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
 782
 783         } while (read_seqcount_retry(&tk_core.seq, seq));
 784
 785         return ktime_add_ns(base, nsecs);
 786 }
 787 EXPORT_SYMBOL_GPL(ktime_get_raw);
 788
 789 /**
 790  * ktime_get_ts64 - get the monotonic clock in timespec64 format
 791  * @ts:         pointer to timespec variable
 792  *
 793  * The function calculates the monotonic clock from the realtime
 794  * clock and the wall_to_monotonic offset and stores the result
 795  * in normalized timespec64 format in the variable pointed to by @ts.
 796  */
 797 void ktime_get_ts64(struct timespec64 *ts)
 798 {
 799         struct timekeeper *tk = &tk_core.timekeeper;
 800         struct timespec64 tomono;
 801         s64 nsec;
 802         unsigned int seq;
 803
 804         WARN_ON(timekeeping_suspended);
 805
 806         do {
 807                 seq = read_seqcount_begin(&tk_core.seq);
 808                 ts->tv_sec = tk->xtime_sec;
 809                 nsec = timekeeping_get_ns(&tk->tkr_mono);
 810                 tomono = tk->wall_to_monotonic;
 811
 812         } while (read_seqcount_retry(&tk_core.seq, seq));
 813
 814         ts->tv_sec += tomono.tv_sec;
 815         ts->tv_nsec = 0;
 816         timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 817 }
 818 EXPORT_SYMBOL_GPL(ktime_get_ts64);
 819
 820 /**
 821  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 822  *
 823  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 824  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 825  * works on both 32 and 64 bit systems. On 32 bit systems the readout
 826  * covers ~136 years of uptime which should be enough to prevent
 827  * premature wrap arounds.
 828  */
 829 time64_t ktime_get_seconds(void)
 830 {
 831         struct timekeeper *tk = &tk_core.timekeeper;
 832
 833         WARN_ON(timekeeping_suspended);
 834         return tk->ktime_sec;
 835 }
 836 EXPORT_SYMBOL_GPL(ktime_get_seconds);
 837
 838 /**
 839  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
 840  *
 841  * Returns the wall clock seconds since 1970. This replaces the
 842  * get_seconds() interface which is not y2038 safe on 32bit systems.
 843  *
 844  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
 845  * 32bit systems the access must be protected with the sequence
 846  * counter to provide "atomic" access to the 64bit tk->xtime_sec
 847  * value.
 848  */
 849 time64_t ktime_get_real_seconds(void)
 850 {
 851         struct timekeeper *tk = &tk_core.timekeeper;
 852         time64_t seconds;
 853         unsigned int seq;
 854
 855         if (IS_ENABLED(CONFIG_64BIT))
 856                 return tk->xtime_sec;
 857
 858         do {
 859                 seq = read_seqcount_begin(&tk_core.seq);
 860                 seconds = tk->xtime_sec;
 861
 862         } while (read_seqcount_retry(&tk_core.seq, seq));
 863
 864         return seconds;
 865 }
 866 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
 867
 868 /**
 869  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
 870  * but without the sequence counter protect. This internal function
 871  * is called just when timekeeping lock is already held.
 872  */
 873 time64_t __ktime_get_real_seconds(void)
 874 {
 875         struct timekeeper *tk = &tk_core.timekeeper;
 876
 877         return tk->xtime_sec;
 878 }
 879
 880 /**
 881  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
 882  * @systime_snapshot:   pointer to struct receiving the system time snapshot
 883  */
 884 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
 885 {
 886         struct timekeeper *tk = &tk_core.timekeeper;
 887         unsigned long seq;
 888         ktime_t base_raw;
 889         ktime_t base_real;
 890         s64 nsec_raw;
 891         s64 nsec_real;
 892         cycle_t now;
 893
 894         WARN_ON_ONCE(timekeeping_suspended);
 895
 896         do {
 897                 seq = read_seqcount_begin(&tk_core.seq);
 898
 899                 now = tk->tkr_mono.read(tk->tkr_mono.clock);
 900                 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
 901                 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
 902                 base_real = ktime_add(tk->tkr_mono.base,
 903                                       tk_core.timekeeper.offs_real);
 904                 base_raw = tk->tkr_raw.base;
 905                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
 906                 nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
 907         } while (read_seqcount_retry(&tk_core.seq, seq));
 908
 909         systime_snapshot->cycles = now;
 910         systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
 911         systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
 912 }
 913 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
 914
 915 /* Scale base by mult/div checking for overflow */
 916 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
 917 {
 918         u64 tmp, rem;
 919
 920         tmp = div64_u64_rem(*base, div, &rem);
 921
 922         if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
 923             ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
 924                 return -EOVERFLOW;
 925         tmp *= mult;
 926         rem *= mult;
 927
 928         do_div(rem, div);
 929         *base = tmp + rem;
 930         return 0;
 931 }
 932
 933 /**
 934  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
 935  * @history:                    Snapshot representing start of history
 936  * @partial_history_cycles:     Cycle offset into history (fractional part)
 937  * @total_history_cycles:       Total history length in cycles
 938  * @discontinuity:              True indicates clock was set on history period
 939  * @ts:                         Cross timestamp that should be adjusted using
 940  *      partial/total ratio
 941  *
 942  * Helper function used by get_device_system_crosststamp() to correct the
 943  * crosstimestamp corresponding to the start of the current interval to the
 944  * system counter value (timestamp point) provided by the driver. The
 945  * total_history_* quantities are the total history starting at the provided
 946  * reference point and ending at the start of the current interval. The cycle
 947  * count between the driver timestamp point and the start of the current
 948  * interval is partial_history_cycles.
 949  */
 950 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
 951                                          cycle_t partial_history_cycles,
 952                                          cycle_t total_history_cycles,
 953                                          bool discontinuity,
 954                                          struct system_device_crosststamp *ts)
 955 {
 956         struct timekeeper *tk = &tk_core.timekeeper;
 957         u64 corr_raw, corr_real;
 958         bool interp_forward;
 959         int ret;
 960
 961         if (total_history_cycles == 0 || partial_history_cycles == 0)
 962                 return 0;
 963
 964         /* Interpolate shortest distance from beginning or end of history */
 965         interp_forward = partial_history_cycles > total_history_cycles/2 ?
 966                 true : false;
 967         partial_history_cycles = interp_forward ?
 968                 total_history_cycles - partial_history_cycles :
 969                 partial_history_cycles;
 970
 971         /*
 972          * Scale the monotonic raw time delta by:
 973          *      partial_history_cycles / total_history_cycles
 974          */
 975         corr_raw = (u64)ktime_to_ns(
 976                 ktime_sub(ts->sys_monoraw, history->raw));
 977         ret = scale64_check_overflow(partial_history_cycles,
 978                                      total_history_cycles, &corr_raw);
 979         if (ret)
 980                 return ret;
 981
 982         /*
 983          * If there is a discontinuity in the history, scale monotonic raw
 984          *      correction by:
 985          *      mult(real)/mult(raw) yielding the realtime correction
 986          * Otherwise, calculate the realtime correction similar to monotonic
 987          *      raw calculation
 988          */
 989         if (discontinuity) {
 990                 corr_real = mul_u64_u32_div
 991                         (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
 992         } else {
 993                 corr_real = (u64)ktime_to_ns(
 994                         ktime_sub(ts->sys_realtime, history->real));
 995                 ret = scale64_check_overflow(partial_history_cycles,
 996                                              total_history_cycles, &corr_real);
 997                 if (ret)
 998                         return ret;
 999         }
1000
1001         /* Fixup monotonic raw and real time time values */
1002         if (interp_forward) {
1003                 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1004                 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1005         } else {
1006                 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1007                 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1008         }
1009
1010         return 0;
1011 }
1012
1013 /*
1014  * cycle_between - true if test occurs chronologically between before and after
1015  */
1016 static bool cycle_between(cycle_t before, cycle_t test, cycle_t after)
1017 {
1018         if (test > before && test < after)
1019                 return true;
1020         if (test < before && before > after)
1021                 return true;
1022         return false;
1023 }
1024
1025 /**
1026  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1027  * @get_time_fn:        Callback to get simultaneous device time and
1028  *      system counter from the device driver
1029  * @ctx:                Context passed to get_time_fn()
1030  * @history_begin:      Historical reference point used to interpolate system
1031  *      time when counter provided by the driver is before the current interval
1032  * @xtstamp:            Receives simultaneously captured system and device time
1033  *
1034  * Reads a timestamp from a device and correlates it to system time
1035  */
1036 int get_device_system_crosststamp(int (*get_time_fn)
1037                                   (ktime_t *device_time,
1038                                    struct system_counterval_t *sys_counterval,
1039                                    void *ctx),
1040                                   void *ctx,
1041                                   struct system_time_snapshot *history_begin,
1042                                   struct system_device_crosststamp *xtstamp)
1043 {
1044         struct system_counterval_t system_counterval;
1045         struct timekeeper *tk = &tk_core.timekeeper;
1046         cycle_t cycles, now, interval_start;
1047         unsigned int clock_was_set_seq = 0;
1048         ktime_t base_real, base_raw;
1049         s64 nsec_real, nsec_raw;
1050         u8 cs_was_changed_seq;
1051         unsigned long seq;
1052         bool do_interp;
1053         int ret;
1054
1055         do {
1056                 seq = read_seqcount_begin(&tk_core.seq);
1057                 /*
1058                  * Try to synchronously capture device time and a system
1059                  * counter value calling back into the device driver
1060                  */
1061                 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1062                 if (ret)
1063                         return ret;
1064
1065                 /*
1066                  * Verify that the clocksource associated with the captured
1067                  * system counter value is the same as the currently installed
1068                  * timekeeper clocksource
1069                  */
1070                 if (tk->tkr_mono.clock != system_counterval.cs)
1071                         return -ENODEV;
1072                 cycles = system_counterval.cycles;
1073
1074                 /*
1075                  * Check whether the system counter value provided by the
1076                  * device driver is on the current timekeeping interval.
1077                  */
1078                 now = tk->tkr_mono.read(tk->tkr_mono.clock);
1079                 interval_start = tk->tkr_mono.cycle_last;
1080                 if (!cycle_between(interval_start, cycles, now)) {
1081                         clock_was_set_seq = tk->clock_was_set_seq;
1082                         cs_was_changed_seq = tk->cs_was_changed_seq;
1083                         cycles = interval_start;
1084                         do_interp = true;
1085                 } else {
1086                         do_interp = false;
1087                 }
1088
1089                 base_real = ktime_add(tk->tkr_mono.base,
1090                                       tk_core.timekeeper.offs_real);
1091                 base_raw = tk->tkr_raw.base;
1092
1093                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1094                                                      system_counterval.cycles);
1095                 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1096                                                     system_counterval.cycles);
1097         } while (read_seqcount_retry(&tk_core.seq, seq));
1098
1099         xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1100         xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1101
1102         /*
1103          * Interpolate if necessary, adjusting back from the start of the
1104          * current interval
1105          */
1106         if (do_interp) {
1107                 cycle_t partial_history_cycles, total_history_cycles;
1108                 bool discontinuity;
1109
1110                 /*
1111                  * Check that the counter value occurs after the provided
1112                  * history reference and that the history doesn't cross a
1113                  * clocksource change
1114                  */
1115                 if (!history_begin ||
1116                     !cycle_between(history_begin->cycles,
1117                                    system_counterval.cycles, cycles) ||
1118                     history_begin->cs_was_changed_seq != cs_was_changed_seq)
1119                         return -EINVAL;
1120                 partial_history_cycles = cycles - system_counterval.cycles;
1121                 total_history_cycles = cycles - history_begin->cycles;
1122                 discontinuity =
1123                         history_begin->clock_was_set_seq != clock_was_set_seq;
1124
1125                 ret = adjust_historical_crosststamp(history_begin,
1126                                                     partial_history_cycles,
1127                                                     total_history_cycles,
1128                                                     discontinuity, xtstamp);
1129                 if (ret)
1130                         return ret;
1131         }
1132
1133         return 0;
1134 }
1135 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1136
1137 /**
1138  * do_gettimeofday - Returns the time of day in a timeval
1139  * @tv:         pointer to the timeval to be set
1140  *
1141  * NOTE: Users should be converted to using getnstimeofday()
1142  */
1143 void do_gettimeofday(struct timeval *tv)
1144 {
1145         struct timespec64 now;
1146
1147         getnstimeofday64(&now);
1148         tv->tv_sec = now.tv_sec;
1149         tv->tv_usec = now.tv_nsec/1000;
1150 }
1151 EXPORT_SYMBOL(do_gettimeofday);
1152
1153 /**
1154  * do_settimeofday64 - Sets the time of day.
1155  * @ts:     pointer to the timespec64 variable containing the new time
1156  *
1157  * Sets the time of day to the new time and update NTP and notify hrtimers
1158  */
1159 int do_settimeofday64(const struct timespec64 *ts)
1160 {
1161         struct timekeeper *tk = &tk_core.timekeeper;
1162         struct timespec64 ts_delta, xt;
1163         unsigned long flags;
1164         int ret = 0;
1165
1166         if (!timespec64_valid_strict(ts))
1167                 return -EINVAL;
1168
1169         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1170         write_seqcount_begin(&tk_core.seq);
1171
1172         timekeeping_forward_now(tk);
1173
1174         xt = tk_xtime(tk);
1175         ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1176         ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1177
1178         if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1179                 ret = -EINVAL;
1180                 goto out;
1181         }
1182
1183         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1184
1185         tk_set_xtime(tk, ts);
1186 out:
1187         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1188
1189         write_seqcount_end(&tk_core.seq);
1190         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1191
1192         /* signal hrtimers about time change */
1193         clock_was_set();
1194
1195         return ret;
1196 }
1197 EXPORT_SYMBOL(do_settimeofday64);
1198
1199 /**
1200  * timekeeping_inject_offset - Adds or subtracts from the current time.
1201  * @tv:         pointer to the timespec variable containing the offset
1202  *
1203  * Adds or subtracts an offset value from the current time.
1204  */
1205 int timekeeping_inject_offset(struct timespec *ts)
1206 {
1207         struct timekeeper *tk = &tk_core.timekeeper;
1208         unsigned long flags;
1209         struct timespec64 ts64, tmp;
1210         int ret = 0;
1211
1212         if (!timespec_inject_offset_valid(ts))
1213                 return -EINVAL;
1214
1215         ts64 = timespec_to_timespec64(*ts);
1216
1217         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1218         write_seqcount_begin(&tk_core.seq);
1219
1220         timekeeping_forward_now(tk);
1221
1222         /* Make sure the proposed value is valid */
1223         tmp = timespec64_add(tk_xtime(tk),  ts64);
1224         if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1225             !timespec64_valid_strict(&tmp)) {
1226                 ret = -EINVAL;
1227                 goto error;
1228         }
1229
1230         tk_xtime_add(tk, &ts64);
1231         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1232
1233 error: /* even if we error out, we forwarded the time, so call update */
1234         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1235
1236         write_seqcount_end(&tk_core.seq);
1237         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1238
1239         /* signal hrtimers about time change */
1240         clock_was_set();
1241
1242         return ret;
1243 }
1244 EXPORT_SYMBOL(timekeeping_inject_offset);
1245
1246
1247 /**
1248  * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1249  *
1250  */
1251 s32 timekeeping_get_tai_offset(void)
1252 {
1253         struct timekeeper *tk = &tk_core.timekeeper;
1254         unsigned int seq;
1255         s32 ret;
1256
1257         do {
1258                 seq = read_seqcount_begin(&tk_core.seq);
1259                 ret = tk->tai_offset;
1260         } while (read_seqcount_retry(&tk_core.seq, seq));
1261
1262         return ret;
1263 }
1264
1265 /**
1266  * __timekeeping_set_tai_offset - Lock free worker function
1267  *
1268  */
1269 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1270 {
1271         tk->tai_offset = tai_offset;
1272         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1273 }
1274
1275 /**
1276  * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1277  *
1278  */
1279 void timekeeping_set_tai_offset(s32 tai_offset)
1280 {
1281         struct timekeeper *tk = &tk_core.timekeeper;
1282         unsigned long flags;
1283
1284         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1285         write_seqcount_begin(&tk_core.seq);
1286         __timekeeping_set_tai_offset(tk, tai_offset);
1287         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1288         write_seqcount_end(&tk_core.seq);
1289         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1290         clock_was_set();
1291 }
1292
1293 /**
1294  * change_clocksource - Swaps clocksources if a new one is available
1295  *
1296  * Accumulates current time interval and initializes new clocksource
1297  */
1298 static int change_clocksource(void *data)
1299 {
1300         struct timekeeper *tk = &tk_core.timekeeper;
1301         struct clocksource *new, *old;
1302         unsigned long flags;
1303
1304         new = (struct clocksource *) data;
1305
1306         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1307         write_seqcount_begin(&tk_core.seq);
1308
1309         timekeeping_forward_now(tk);
1310         /*
1311          * If the cs is in module, get a module reference. Succeeds
1312          * for built-in code (owner == NULL) as well.
1313          */
1314         if (try_module_get(new->owner)) {
1315                 if (!new->enable || new->enable(new) == 0) {
1316                         old = tk->tkr_mono.clock;
1317                         tk_setup_internals(tk, new);
1318                         if (old->disable)
1319                                 old->disable(old);
1320                         module_put(old->owner);
1321                 } else {
1322                         module_put(new->owner);
1323                 }
1324         }
1325         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1326
1327         write_seqcount_end(&tk_core.seq);
1328         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1329
1330         return 0;
1331 }
1332
1333 /**
1334  * timekeeping_notify - Install a new clock source
1335  * @clock:              pointer to the clock source
1336  *
1337  * This function is called from clocksource.c after a new, better clock
1338  * source has been registered. The caller holds the clocksource_mutex.
1339  */
1340 int timekeeping_notify(struct clocksource *clock)
1341 {
1342         struct timekeeper *tk = &tk_core.timekeeper;
1343
1344         if (tk->tkr_mono.clock == clock)
1345                 return 0;
1346         stop_machine(change_clocksource, clock, NULL);
1347         tick_clock_notify();
1348         return tk->tkr_mono.clock == clock ? 0 : -1;
1349 }
1350
1351 /**
1352  * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1353  * @ts:         pointer to the timespec64 to be set
1354  *
1355  * Returns the raw monotonic time (completely un-modified by ntp)
1356  */
1357 void getrawmonotonic64(struct timespec64 *ts)
1358 {
1359         struct timekeeper *tk = &tk_core.timekeeper;
1360         struct timespec64 ts64;
1361         unsigned long seq;
1362         s64 nsecs;
1363
1364         do {
1365                 seq = read_seqcount_begin(&tk_core.seq);
1366                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1367                 ts64 = tk->raw_time;
1368
1369         } while (read_seqcount_retry(&tk_core.seq, seq));
1370
1371         timespec64_add_ns(&ts64, nsecs);
1372         *ts = ts64;
1373 }
1374 EXPORT_SYMBOL(getrawmonotonic64);
1375
1376
1377 /**
1378  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1379  */
1380 int timekeeping_valid_for_hres(void)
1381 {
1382         struct timekeeper *tk = &tk_core.timekeeper;
1383         unsigned long seq;
1384         int ret;
1385
1386         do {
1387                 seq = read_seqcount_begin(&tk_core.seq);
1388
1389                 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1390
1391         } while (read_seqcount_retry(&tk_core.seq, seq));
1392
1393         return ret;
1394 }
1395
1396 /**
1397  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1398  */
1399 u64 timekeeping_max_deferment(void)
1400 {
1401         struct timekeeper *tk = &tk_core.timekeeper;
1402         unsigned long seq;
1403         u64 ret;
1404
1405         do {
1406                 seq = read_seqcount_begin(&tk_core.seq);
1407
1408                 ret = tk->tkr_mono.clock->max_idle_ns;
1409
1410         } while (read_seqcount_retry(&tk_core.seq, seq));
1411
1412         return ret;
1413 }
1414
1415 /**
1416  * read_persistent_clock -  Return time from the persistent clock.
1417  *
1418  * Weak dummy function for arches that do not yet support it.
1419  * Reads the time from the battery backed persistent clock.
1420  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1421  *
1422  *  XXX - Do be sure to remove it once all arches implement it.
1423  */
1424 void __weak read_persistent_clock(struct timespec *ts)
1425 {
1426         ts->tv_sec = 0;
1427         ts->tv_nsec = 0;
1428 }
1429
1430 void __weak read_persistent_clock64(struct timespec64 *ts64)
1431 {
1432         struct timespec ts;
1433
1434         read_persistent_clock(&ts);
1435         *ts64 = timespec_to_timespec64(ts);
1436 }
1437
1438 /**
1439  * read_boot_clock64 -  Return time of the system start.
1440  *
1441  * Weak dummy function for arches that do not yet support it.
1442  * Function to read the exact time the system has been started.
1443  * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1444  *
1445  *  XXX - Do be sure to remove it once all arches implement it.
1446  */
1447 void __weak read_boot_clock64(struct timespec64 *ts)
1448 {
1449         ts->tv_sec = 0;
1450         ts->tv_nsec = 0;
1451 }
1452
1453 /* Flag for if timekeeping_resume() has injected sleeptime */
1454 static bool sleeptime_injected;
1455
1456 /* Flag for if there is a persistent clock on this platform */
1457 static bool persistent_clock_exists;
1458
1459 /*
1460  * timekeeping_init - Initializes the clocksource and common timekeeping values
1461  */
1462 void __init timekeeping_init(void)
1463 {
1464         struct timekeeper *tk = &tk_core.timekeeper;
1465         struct clocksource *clock;
1466         unsigned long flags;
1467         struct timespec64 now, boot, tmp;
1468
1469         read_persistent_clock64(&now);
1470         if (!timespec64_valid_strict(&now)) {
1471                 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1472                         "         Check your CMOS/BIOS settings.\n");
1473                 now.tv_sec = 0;
1474                 now.tv_nsec = 0;
1475         } else if (now.tv_sec || now.tv_nsec)
1476                 persistent_clock_exists = true;
1477
1478         read_boot_clock64(&boot);
1479         if (!timespec64_valid_strict(&boot)) {
1480                 pr_warn("WARNING: Boot clock returned invalid value!\n"
1481                         "         Check your CMOS/BIOS settings.\n");
1482                 boot.tv_sec = 0;
1483                 boot.tv_nsec = 0;
1484         }
1485
1486         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1487         write_seqcount_begin(&tk_core.seq);
1488         ntp_init();
1489
1490         clock = clocksource_default_clock();
1491         if (clock->enable)
1492                 clock->enable(clock);
1493         tk_setup_internals(tk, clock);
1494
1495         tk_set_xtime(tk, &now);
1496         tk->raw_time.tv_sec = 0;
1497         tk->raw_time.tv_nsec = 0;
1498         if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1499                 boot = tk_xtime(tk);
1500
1501         set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1502         tk_set_wall_to_mono(tk, tmp);
1503
1504         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1505
1506         write_seqcount_end(&tk_core.seq);
1507         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1508 }
1509
1510 /* time in seconds when suspend began for persistent clock */
1511 static struct timespec64 timekeeping_suspend_time;
1512
1513 /**
1514  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1515  * @delta: pointer to a timespec delta value
1516  *
1517  * Takes a timespec offset measuring a suspend interval and properly
1518  * adds the sleep offset to the timekeeping variables.
1519  */
1520 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1521                                            struct timespec64 *delta)
1522 {
1523         if (!timespec64_valid_strict(delta)) {
1524                 printk_deferred(KERN_WARNING
1525                                 "__timekeeping_inject_sleeptime: Invalid "
1526                                 "sleep delta value!\n");
1527                 return;
1528         }
1529         tk_xtime_add(tk, delta);
1530         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1531         tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1532         tk_debug_account_sleep_time(delta);
1533 }
1534
1535 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1536 /**
1537  * We have three kinds of time sources to use for sleep time
1538  * injection, the preference order is:
1539  * 1) non-stop clocksource
1540  * 2) persistent clock (ie: RTC accessible when irqs are off)
1541  * 3) RTC
1542  *
1543  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1544  * If system has neither 1) nor 2), 3) will be used finally.
1545  *
1546  *
1547  * If timekeeping has injected sleeptime via either 1) or 2),
1548  * 3) becomes needless, so in this case we don't need to call
1549  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1550  * means.
1551  */
1552 bool timekeeping_rtc_skipresume(void)
1553 {
1554         return sleeptime_injected;
1555 }
1556
1557 /**
1558  * 1) can be determined whether to use or not only when doing
1559  * timekeeping_resume() which is invoked after rtc_suspend(),
1560  * so we can't skip rtc_suspend() surely if system has 1).
1561  *
1562  * But if system has 2), 2) will definitely be used, so in this
1563  * case we don't need to call rtc_suspend(), and this is what
1564  * timekeeping_rtc_skipsuspend() means.
1565  */
1566 bool timekeeping_rtc_skipsuspend(void)
1567 {
1568         return persistent_clock_exists;
1569 }
1570
1571 /**
1572  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1573  * @delta: pointer to a timespec64 delta value
1574  *
1575  * This hook is for architectures that cannot support read_persistent_clock64
1576  * because their RTC/persistent clock is only accessible when irqs are enabled.
1577  * and also don't have an effective nonstop clocksource.
1578  *
1579  * This function should only be called by rtc_resume(), and allows
1580  * a suspend offset to be injected into the timekeeping values.
1581  */
1582 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1583 {
1584         struct timekeeper *tk = &tk_core.timekeeper;
1585         unsigned long flags;
1586
1587         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1588         write_seqcount_begin(&tk_core.seq);
1589
1590         timekeeping_forward_now(tk);
1591
1592         __timekeeping_inject_sleeptime(tk, delta);
1593
1594         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1595
1596         write_seqcount_end(&tk_core.seq);
1597         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1598
1599         /* signal hrtimers about time change */
1600         clock_was_set();
1601 }
1602 #endif
1603
1604 /**
1605  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1606  */
1607 void timekeeping_resume(void)
1608 {
1609         struct timekeeper *tk = &tk_core.timekeeper;
1610         struct clocksource *clock = tk->tkr_mono.clock;
1611         unsigned long flags;
1612         struct timespec64 ts_new, ts_delta;
1613         cycle_t cycle_now, cycle_delta;
1614
1615         sleeptime_injected = false;
1616         read_persistent_clock64(&ts_new);
1617
1618         clockevents_resume();
1619         clocksource_resume();
1620
1621         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1622         write_seqcount_begin(&tk_core.seq);
1623
1624         /*
1625          * After system resumes, we need to calculate the suspended time and
1626          * compensate it for the OS time. There are 3 sources that could be
1627          * used: Nonstop clocksource during suspend, persistent clock and rtc
1628          * device.
1629          *
1630          * One specific platform may have 1 or 2 or all of them, and the
1631          * preference will be:
1632          *      suspend-nonstop clocksource -> persistent clock -> rtc
1633          * The less preferred source will only be tried if there is no better
1634          * usable source. The rtc part is handled separately in rtc core code.
1635          */
1636         cycle_now = tk->tkr_mono.read(clock);
1637         if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1638                 cycle_now > tk->tkr_mono.cycle_last) {
1639                 u64 num, max = ULLONG_MAX;
1640                 u32 mult = clock->mult;
1641                 u32 shift = clock->shift;
1642                 s64 nsec = 0;
1643
1644                 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1645                                                 tk->tkr_mono.mask);
1646
1647                 /*
1648                  * "cycle_delta * mutl" may cause 64 bits overflow, if the
1649                  * suspended time is too long. In that case we need do the
1650                  * 64 bits math carefully
1651                  */
1652                 do_div(max, mult);
1653                 if (cycle_delta > max) {
1654                         num = div64_u64(cycle_delta, max);
1655                         nsec = (((u64) max * mult) >> shift) * num;
1656                         cycle_delta -= num * max;
1657                 }
1658                 nsec += ((u64) cycle_delta * mult) >> shift;
1659
1660                 ts_delta = ns_to_timespec64(nsec);
1661                 sleeptime_injected = true;
1662         } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1663                 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1664                 sleeptime_injected = true;
1665         }
1666
1667         if (sleeptime_injected)
1668                 __timekeeping_inject_sleeptime(tk, &ts_delta);
1669
1670         /* Re-base the last cycle value */
1671         tk->tkr_mono.cycle_last = cycle_now;
1672         tk->tkr_raw.cycle_last  = cycle_now;
1673
1674         tk->ntp_error = 0;
1675         timekeeping_suspended = 0;
1676         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1677         write_seqcount_end(&tk_core.seq);
1678         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1679
1680         touch_softlockup_watchdog();
1681
1682         tick_resume();
1683         hrtimers_resume();
1684 }
1685
1686 int timekeeping_suspend(void)
1687 {
1688         struct timekeeper *tk = &tk_core.timekeeper;
1689         unsigned long flags;
1690         struct timespec64               delta, delta_delta;
1691         static struct timespec64        old_delta;
1692
1693         read_persistent_clock64(&timekeeping_suspend_time);
1694
1695         /*
1696          * On some systems the persistent_clock can not be detected at
1697          * timekeeping_init by its return value, so if we see a valid
1698          * value returned, update the persistent_clock_exists flag.
1699          */
1700         if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1701                 persistent_clock_exists = true;
1702
1703         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1704         write_seqcount_begin(&tk_core.seq);
1705         timekeeping_forward_now(tk);
1706         timekeeping_suspended = 1;
1707
1708         if (persistent_clock_exists) {
1709                 /*
1710                  * To avoid drift caused by repeated suspend/resumes,
1711                  * which each can add ~1 second drift error,
1712                  * try to compensate so the difference in system time
1713                  * and persistent_clock time stays close to constant.
1714                  */
1715                 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1716                 delta_delta = timespec64_sub(delta, old_delta);
1717                 if (abs(delta_delta.tv_sec) >= 2) {
1718                         /*
1719                          * if delta_delta is too large, assume time correction
1720                          * has occurred and set old_delta to the current delta.
1721                          */
1722                         old_delta = delta;
1723                 } else {
1724                         /* Otherwise try to adjust old_system to compensate */
1725                         timekeeping_suspend_time =
1726                                 timespec64_add(timekeeping_suspend_time, delta_delta);
1727                 }
1728         }
1729
1730         timekeeping_update(tk, TK_MIRROR);
1731         halt_fast_timekeeper(tk);
1732         write_seqcount_end(&tk_core.seq);
1733         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1734
1735         tick_suspend();
1736         clocksource_suspend();
1737         clockevents_suspend();
1738
1739         return 0;
1740 }
1741
1742 /* sysfs resume/suspend bits for timekeeping */
1743 static struct syscore_ops timekeeping_syscore_ops = {
1744         .resume         = timekeeping_resume,
1745         .suspend        = timekeeping_suspend,
1746 };
1747
1748 static int __init timekeeping_init_ops(void)
1749 {
1750         register_syscore_ops(&timekeeping_syscore_ops);
1751         return 0;
1752 }
1753 device_initcall(timekeeping_init_ops);
1754
1755 /*
1756  * Apply a multiplier adjustment to the timekeeper
1757  */
1758 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1759                                                          s64 offset,
1760                                                          bool negative,
1761                                                          int adj_scale)
1762 {
1763         s64 interval = tk->cycle_interval;
1764         s32 mult_adj = 1;
1765
1766         if (negative) {
1767                 mult_adj = -mult_adj;
1768                 interval = -interval;
1769                 offset  = -offset;
1770         }
1771         mult_adj <<= adj_scale;
1772         interval <<= adj_scale;
1773         offset <<= adj_scale;
1774
1775         /*
1776          * So the following can be confusing.
1777          *
1778          * To keep things simple, lets assume mult_adj == 1 for now.
1779          *
1780          * When mult_adj != 1, remember that the interval and offset values
1781          * have been appropriately scaled so the math is the same.
1782          *
1783          * The basic idea here is that we're increasing the multiplier
1784          * by one, this causes the xtime_interval to be incremented by
1785          * one cycle_interval. This is because:
1786          *      xtime_interval = cycle_interval * mult
1787          * So if mult is being incremented by one:
1788          *      xtime_interval = cycle_interval * (mult + 1)
1789          * Its the same as:
1790          *      xtime_interval = (cycle_interval * mult) + cycle_interval
1791          * Which can be shortened to:
1792          *      xtime_interval += cycle_interval
1793          *
1794          * So offset stores the non-accumulated cycles. Thus the current
1795          * time (in shifted nanoseconds) is:
1796          *      now = (offset * adj) + xtime_nsec
1797          * Now, even though we're adjusting the clock frequency, we have
1798          * to keep time consistent. In other words, we can't jump back
1799          * in time, and we also want to avoid jumping forward in time.
1800          *
1801          * So given the same offset value, we need the time to be the same
1802          * both before and after the freq adjustment.
1803          *      now = (offset * adj_1) + xtime_nsec_1
1804          *      now = (offset * adj_2) + xtime_nsec_2
1805          * So:
1806          *      (offset * adj_1) + xtime_nsec_1 =
1807          *              (offset * adj_2) + xtime_nsec_2
1808          * And we know:
1809          *      adj_2 = adj_1 + 1
1810          * So:
1811          *      (offset * adj_1) + xtime_nsec_1 =
1812          *              (offset * (adj_1+1)) + xtime_nsec_2
1813          *      (offset * adj_1) + xtime_nsec_1 =
1814          *              (offset * adj_1) + offset + xtime_nsec_2
1815          * Canceling the sides:
1816          *      xtime_nsec_1 = offset + xtime_nsec_2
1817          * Which gives us:
1818          *      xtime_nsec_2 = xtime_nsec_1 - offset
1819          * Which simplfies to:
1820          *      xtime_nsec -= offset
1821          *
1822          * XXX - TODO: Doc ntp_error calculation.
1823          */
1824         if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1825                 /* NTP adjustment caused clocksource mult overflow */
1826                 WARN_ON_ONCE(1);
1827                 return;
1828         }
1829
1830         tk->tkr_mono.mult += mult_adj;
1831         tk->xtime_interval += interval;
1832         tk->tkr_mono.xtime_nsec -= offset;
1833         tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1834 }
1835
1836 /*
1837  * Calculate the multiplier adjustment needed to match the frequency
1838  * specified by NTP
1839  */
1840 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1841                                                         s64 offset)
1842 {
1843         s64 interval = tk->cycle_interval;
1844         s64 xinterval = tk->xtime_interval;
1845         u32 base = tk->tkr_mono.clock->mult;
1846         u32 max = tk->tkr_mono.clock->maxadj;
1847         u32 cur_adj = tk->tkr_mono.mult;
1848         s64 tick_error;
1849         bool negative;
1850         u32 adj_scale;
1851
1852         /* Remove any current error adj from freq calculation */
1853         if (tk->ntp_err_mult)
1854                 xinterval -= tk->cycle_interval;
1855
1856         tk->ntp_tick = ntp_tick_length();
1857
1858         /* Calculate current error per tick */
1859         tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1860         tick_error -= (xinterval + tk->xtime_remainder);
1861
1862         /* Don't worry about correcting it if its small */
1863         if (likely((tick_error >= 0) && (tick_error <= interval)))
1864                 return;
1865
1866         /* preserve the direction of correction */
1867         negative = (tick_error < 0);
1868
1869         /* If any adjustment would pass the max, just return */
1870         if (negative && (cur_adj - 1) <= (base - max))
1871                 return;
1872         if (!negative && (cur_adj + 1) >= (base + max))
1873                 return;
1874         /*
1875          * Sort out the magnitude of the correction, but
1876          * avoid making so large a correction that we go
1877          * over the max adjustment.
1878          */
1879         adj_scale = 0;
1880         tick_error = abs(tick_error);
1881         while (tick_error > interval) {
1882                 u32 adj = 1 << (adj_scale + 1);
1883
1884                 /* Check if adjustment gets us within 1 unit from the max */
1885                 if (negative && (cur_adj - adj) <= (base - max))
1886                         break;
1887                 if (!negative && (cur_adj + adj) >= (base + max))
1888                         break;
1889
1890                 adj_scale++;
1891                 tick_error >>= 1;
1892         }
1893
1894         /* scale the corrections */
1895         timekeeping_apply_adjustment(tk, offset, negative, adj_scale);
1896 }
1897
1898 /*
1899  * Adjust the timekeeper's multiplier to the correct frequency
1900  * and also to reduce the accumulated error value.
1901  */
1902 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1903 {
1904         /* Correct for the current frequency error */
1905         timekeeping_freqadjust(tk, offset);
1906
1907         /* Next make a small adjustment to fix any cumulative error */
1908         if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1909                 tk->ntp_err_mult = 1;
1910                 timekeeping_apply_adjustment(tk, offset, 0, 0);
1911         } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1912                 /* Undo any existing error adjustment */
1913                 timekeeping_apply_adjustment(tk, offset, 1, 0);
1914                 tk->ntp_err_mult = 0;
1915         }
1916
1917         if (unlikely(tk->tkr_mono.clock->maxadj &&
1918                 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1919                         > tk->tkr_mono.clock->maxadj))) {
1920                 printk_once(KERN_WARNING
1921                         "Adjusting %s more than 11%% (%ld vs %ld)\n",
1922                         tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1923                         (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1924         }
1925
1926         /*
1927          * It may be possible that when we entered this function, xtime_nsec
1928          * was very small.  Further, if we're slightly speeding the clocksource
1929          * in the code above, its possible the required corrective factor to
1930          * xtime_nsec could cause it to underflow.
1931          *
1932          * Now, since we already accumulated the second, cannot simply roll
1933          * the accumulated second back, since the NTP subsystem has been
1934          * notified via second_overflow. So instead we push xtime_nsec forward
1935          * by the amount we underflowed, and add that amount into the error.
1936          *
1937          * We'll correct this error next time through this function, when
1938          * xtime_nsec is not as small.
1939          */
1940         if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1941                 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1942                 tk->tkr_mono.xtime_nsec = 0;
1943                 tk->ntp_error += neg << tk->ntp_error_shift;
1944         }
1945 }
1946
1947 /**
1948  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1949  *
1950  * Helper function that accumulates the nsecs greater than a second
1951  * from the xtime_nsec field to the xtime_secs field.
1952  * It also calls into the NTP code to handle leapsecond processing.
1953  *
1954  */
1955 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1956 {
1957         u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1958         unsigned int clock_set = 0;
1959
1960         while (tk->tkr_mono.xtime_nsec >= nsecps) {
1961                 int leap;
1962
1963                 tk->tkr_mono.xtime_nsec -= nsecps;
1964                 tk->xtime_sec++;
1965
1966                 /* Figure out if its a leap sec and apply if needed */
1967                 leap = second_overflow(tk->xtime_sec);
1968                 if (unlikely(leap)) {
1969                         struct timespec64 ts;
1970
1971                         tk->xtime_sec += leap;
1972
1973                         ts.tv_sec = leap;
1974                         ts.tv_nsec = 0;
1975                         tk_set_wall_to_mono(tk,
1976                                 timespec64_sub(tk->wall_to_monotonic, ts));
1977
1978                         __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1979
1980                         clock_set = TK_CLOCK_WAS_SET;
1981                 }
1982         }
1983         return clock_set;
1984 }
1985
1986 /**
1987  * logarithmic_accumulation - shifted accumulation of cycles
1988  *
1989  * This functions accumulates a shifted interval of cycles into
1990  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1991  * loop.
1992  *
1993  * Returns the unconsumed cycles.
1994  */
1995 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1996                                                 u32 shift,
1997                                                 unsigned int *clock_set)
1998 {
1999         cycle_t interval = tk->cycle_interval << shift;
2000         u64 raw_nsecs;
2001
2002         /* If the offset is smaller than a shifted interval, do nothing */
2003         if (offset < interval)
2004                 return offset;
2005
2006         /* Accumulate one shifted interval */
2007         offset -= interval;
2008         tk->tkr_mono.cycle_last += interval;
2009         tk->tkr_raw.cycle_last  += interval;
2010
2011         tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2012         *clock_set |= accumulate_nsecs_to_secs(tk);
2013
2014         /* Accumulate raw time */
2015         raw_nsecs = (u64)tk->raw_interval << shift;
2016         raw_nsecs += tk->raw_time.tv_nsec;
2017         if (raw_nsecs >= NSEC_PER_SEC) {
2018                 u64 raw_secs = raw_nsecs;
2019                 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
2020                 tk->raw_time.tv_sec += raw_secs;
2021         }
2022         tk->raw_time.tv_nsec = raw_nsecs;
2023
2024         /* Accumulate error between NTP and clock interval */
2025         tk->ntp_error += tk->ntp_tick << shift;
2026         tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2027                                                 (tk->ntp_error_shift + shift);
2028
2029         return offset;
2030 }
2031
2032 /**
2033  * update_wall_time - Uses the current clocksource to increment the wall time
2034  *
2035  */
2036 void update_wall_time(void)
2037 {
2038         struct timekeeper *real_tk = &tk_core.timekeeper;
2039         struct timekeeper *tk = &shadow_timekeeper;
2040         cycle_t offset;
2041         int shift = 0, maxshift;
2042         unsigned int clock_set = 0;
2043         unsigned long flags;
2044
2045         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2046
2047         /* Make sure we're fully resumed: */
2048         if (unlikely(timekeeping_suspended))
2049                 goto out;
2050
2051 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2052         offset = real_tk->cycle_interval;
2053 #else
2054         offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
2055                                    tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2056 #endif
2057
2058         /* Check if there's really nothing to do */
2059         if (offset < real_tk->cycle_interval)
2060                 goto out;
2061
2062         /* Do some additional sanity checking */
2063         timekeeping_check_update(real_tk, offset);
2064
2065         /*
2066          * With NO_HZ we may have to accumulate many cycle_intervals
2067          * (think "ticks") worth of time at once. To do this efficiently,
2068          * we calculate the largest doubling multiple of cycle_intervals
2069          * that is smaller than the offset.  We then accumulate that
2070          * chunk in one go, and then try to consume the next smaller
2071          * doubled multiple.
2072          */
2073         shift = ilog2(offset) - ilog2(tk->cycle_interval);
2074         shift = max(0, shift);
2075         /* Bound shift to one less than what overflows tick_length */
2076         maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2077         shift = min(shift, maxshift);
2078         while (offset >= tk->cycle_interval) {
2079                 offset = logarithmic_accumulation(tk, offset, shift,
2080                                                         &clock_set);
2081                 if (offset < tk->cycle_interval<<shift)
2082                         shift--;
2083         }
2084
2085         /* correct the clock when NTP error is too big */
2086         timekeeping_adjust(tk, offset);
2087
2088         /*
2089          * XXX This can be killed once everyone converts
2090          * to the new update_vsyscall.
2091          */
2092         old_vsyscall_fixup(tk);
2093
2094         /*
2095          * Finally, make sure that after the rounding
2096          * xtime_nsec isn't larger than NSEC_PER_SEC
2097          */
2098         clock_set |= accumulate_nsecs_to_secs(tk);
2099
2100         write_seqcount_begin(&tk_core.seq);
2101         /*
2102          * Update the real timekeeper.
2103          *
2104          * We could avoid this memcpy by switching pointers, but that
2105          * requires changes to all other timekeeper usage sites as
2106          * well, i.e. move the timekeeper pointer getter into the
2107          * spinlocked/seqcount protected sections. And we trade this
2108          * memcpy under the tk_core.seq against one before we start
2109          * updating.
2110          */
2111         timekeeping_update(tk, clock_set);
2112         memcpy(real_tk, tk, sizeof(*tk));
2113         /* The memcpy must come last. Do not put anything here! */
2114         write_seqcount_end(&tk_core.seq);
2115 out:
2116         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2117         if (clock_set)
2118                 /* Have to call _delayed version, since in irq context*/
2119                 clock_was_set_delayed();
2120 }
2121
2122 /**
2123  * getboottime64 - Return the real time of system boot.
2124  * @ts:         pointer to the timespec64 to be set
2125  *
2126  * Returns the wall-time of boot in a timespec64.
2127  *
2128  * This is based on the wall_to_monotonic offset and the total suspend
2129  * time. Calls to settimeofday will affect the value returned (which
2130  * basically means that however wrong your real time clock is at boot time,
2131  * you get the right time here).
2132  */
2133 void getboottime64(struct timespec64 *ts)
2134 {
2135         struct timekeeper *tk = &tk_core.timekeeper;
2136         ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2137
2138         *ts = ktime_to_timespec64(t);
2139 }
2140 EXPORT_SYMBOL_GPL(getboottime64);
2141
2142 unsigned long get_seconds(void)
2143 {
2144         struct timekeeper *tk = &tk_core.timekeeper;
2145
2146         return tk->xtime_sec;
2147 }
2148 EXPORT_SYMBOL(get_seconds);
2149
2150 struct timespec __current_kernel_time(void)
2151 {
2152         struct timekeeper *tk = &tk_core.timekeeper;
2153
2154         return timespec64_to_timespec(tk_xtime(tk));
2155 }
2156
2157 struct timespec64 current_kernel_time64(void)
2158 {
2159         struct timekeeper *tk = &tk_core.timekeeper;
2160         struct timespec64 now;
2161         unsigned long seq;
2162
2163         do {
2164                 seq = read_seqcount_begin(&tk_core.seq);
2165
2166                 now = tk_xtime(tk);
2167         } while (read_seqcount_retry(&tk_core.seq, seq));
2168
2169         return now;
2170 }
2171 EXPORT_SYMBOL(current_kernel_time64);
2172
2173 struct timespec64 get_monotonic_coarse64(void)
2174 {
2175         struct timekeeper *tk = &tk_core.timekeeper;
2176         struct timespec64 now, mono;
2177         unsigned long seq;
2178
2179         do {
2180                 seq = read_seqcount_begin(&tk_core.seq);
2181
2182                 now = tk_xtime(tk);
2183                 mono = tk->wall_to_monotonic;
2184         } while (read_seqcount_retry(&tk_core.seq, seq));
2185
2186         set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2187                                 now.tv_nsec + mono.tv_nsec);
2188
2189         return now;
2190 }
2191 EXPORT_SYMBOL(get_monotonic_coarse64);
2192
2193 /*
2194  * Must hold jiffies_lock
2195  */
2196 void do_timer(unsigned long ticks)
2197 {
2198         jiffies_64 += ticks;
2199         calc_global_load(ticks);
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  */
2215 ktime_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.tv64 >= tk->next_leap_ktime.tv64))
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  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2248  */
2249 int do_adjtimex(struct timex *txc)
2250 {
2251         struct timekeeper *tk = &tk_core.timekeeper;
2252         unsigned long flags;
2253         struct timespec64 ts;
2254         s32 orig_tai, tai;
2255         int ret;
2256
2257         /* Validate the data before disabling interrupts */
2258         ret = ntp_validate_timex(txc);
2259         if (ret)
2260                 return ret;
2261
2262         if (txc->modes & ADJ_SETOFFSET) {
2263                 struct timespec delta;
2264                 delta.tv_sec  = txc->time.tv_sec;
2265                 delta.tv_nsec = txc->time.tv_usec;
2266                 if (!(txc->modes & ADJ_NANO))
2267                         delta.tv_nsec *= 1000;
2268                 ret = timekeeping_inject_offset(&delta);
2269                 if (ret)
2270                         return ret;
2271         }
2272
2273         getnstimeofday64(&ts);
2274
2275         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2276         write_seqcount_begin(&tk_core.seq);
2277
2278         orig_tai = tai = tk->tai_offset;
2279         ret = __do_adjtimex(txc, &ts, &tai);
2280
2281         if (tai != orig_tai) {
2282                 __timekeeping_set_tai_offset(tk, tai);
2283                 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2284         }
2285         tk_update_leap_state(tk);
2286
2287         write_seqcount_end(&tk_core.seq);
2288         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2289
2290         if (tai != orig_tai)
2291                 clock_was_set();
2292
2293         ntp_notify_cmos_timer();
2294
2295         return ret;
2296 }
2297
2298 #ifdef CONFIG_NTP_PPS
2299 /**
2300  * hardpps() - Accessor function to NTP __hardpps function
2301  */
2302 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2303 {
2304         unsigned long flags;
2305
2306         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2307         write_seqcount_begin(&tk_core.seq);
2308
2309         __hardpps(phase_ts, raw_ts);
2310
2311         write_seqcount_end(&tk_core.seq);
2312         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2313 }
2314 EXPORT_SYMBOL(hardpps);
2315 #endif
2316
2317 /**
2318  * xtime_update() - advances the timekeeping infrastructure
2319  * @ticks:      number of ticks, that have elapsed since the last call.
2320  *
2321  * Must be called with interrupts disabled.
2322  */
2323 void xtime_update(unsigned long ticks)
2324 {
2325         write_seqlock(&jiffies_lock);
2326         do_timer(ticks);
2327         write_sequnlock(&jiffies_lock);
2328         update_wall_time();
2329 }