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         tk->tkr_mono.xtime_nsec -= remainder;
 484         tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
 485         tk->ntp_error += remainder << tk->ntp_error_shift;
 486         tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
 487 }
 488 #else
 489 #define old_vsyscall_fixup(tk)
 490 #endif
 491
 492 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
 493
 494 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 495 {
 496         raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
 497 }
 498
 499 /**
 500  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
 501  */
 502 int pvclock_gtod_register_notifier(struct notifier_block *nb)
 503 {
 504         struct timekeeper *tk = &tk_core.timekeeper;
 505         unsigned long flags;
 506         int ret;
 507
 508         raw_spin_lock_irqsave(&timekeeper_lock, flags);
 509         ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
 510         update_pvclock_gtod(tk, true);
 511         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 512
 513         return ret;
 514 }
 515 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
 516
 517 /**
 518  * pvclock_gtod_unregister_notifier - unregister a pvclock
 519  * timedata update listener
 520  */
 521 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
 522 {
 523         unsigned long flags;
 524         int ret;
 525
 526         raw_spin_lock_irqsave(&timekeeper_lock, flags);
 527         ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
 528         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 529
 530         return ret;
 531 }
 532 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
 533
 534 /*
 535  * tk_update_leap_state - helper to update the next_leap_ktime
 536  */
 537 static inline void tk_update_leap_state(struct timekeeper *tk)
 538 {
 539         tk->next_leap_ktime = ntp_get_next_leap();
 540         if (tk->next_leap_ktime.tv64 != KTIME_MAX)
 541                 /* Convert to monotonic time */
 542                 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
 543 }
 544
 545 /*
 546  * Update the ktime_t based scalar nsec members of the timekeeper
 547  */
 548 static inline void tk_update_ktime_data(struct timekeeper *tk)
 549 {
 550         u64 seconds;
 551         u32 nsec;
 552
 553         /*
 554          * The xtime based monotonic readout is:
 555          *      nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
 556          * The ktime based monotonic readout is:
 557          *      nsec = base_mono + now();
 558          * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
 559          */
 560         seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
 561         nsec = (u32) tk->wall_to_monotonic.tv_nsec;
 562         tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
 563
 564         /* Update the monotonic raw base */
 565         tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
 566
 567         /*
 568          * The sum of the nanoseconds portions of xtime and
 569          * wall_to_monotonic can be greater/equal one second. Take
 570          * this into account before updating tk->ktime_sec.
 571          */
 572         nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 573         if (nsec >= NSEC_PER_SEC)
 574                 seconds++;
 575         tk->ktime_sec = seconds;
 576 }
 577
 578 /* must hold timekeeper_lock */
 579 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
 580 {
 581         if (action & TK_CLEAR_NTP) {
 582                 tk->ntp_error = 0;
 583                 ntp_clear();
 584         }
 585
 586         tk_update_leap_state(tk);
 587         tk_update_ktime_data(tk);
 588
 589         update_vsyscall(tk);
 590         update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
 591
 592         update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
 593         update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
 594
 595         if (action & TK_CLOCK_WAS_SET)
 596                 tk->clock_was_set_seq++;
 597         /*
 598          * The mirroring of the data to the shadow-timekeeper needs
 599          * to happen last here to ensure we don't over-write the
 600          * timekeeper structure on the next update with stale data
 601          */
 602         if (action & TK_MIRROR)
 603                 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
 604                        sizeof(tk_core.timekeeper));
 605 }
 606
 607 /**
 608  * timekeeping_forward_now - update clock to the current time
 609  *
 610  * Forward the current clock to update its state since the last call to
 611  * update_wall_time(). This is useful before significant clock changes,
 612  * as it avoids having to deal with this time offset explicitly.
 613  */
 614 static void timekeeping_forward_now(struct timekeeper *tk)
 615 {
 616         struct clocksource *clock = tk->tkr_mono.clock;
 617         cycle_t cycle_now, delta;
 618         s64 nsec;
 619
 620         cycle_now = tk->tkr_mono.read(clock);
 621         delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 622         tk->tkr_mono.cycle_last = cycle_now;
 623         tk->tkr_raw.cycle_last  = cycle_now;
 624
 625         tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
 626
 627         /* If arch requires, add in get_arch_timeoffset() */
 628         tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
 629
 630         tk_normalize_xtime(tk);
 631
 632         nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
 633         timespec64_add_ns(&tk->raw_time, nsec);
 634 }
 635
 636 /**
 637  * __getnstimeofday64 - Returns the time of day in a timespec64.
 638  * @ts:         pointer to the timespec to be set
 639  *
 640  * Updates the time of day in the timespec.
 641  * Returns 0 on success, or -ve when suspended (timespec will be undefined).
 642  */
 643 int __getnstimeofday64(struct timespec64 *ts)
 644 {
 645         struct timekeeper *tk = &tk_core.timekeeper;
 646         unsigned long seq;
 647         s64 nsecs = 0;
 648
 649         do {
 650                 seq = read_seqcount_begin(&tk_core.seq);
 651
 652                 ts->tv_sec = tk->xtime_sec;
 653                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 654
 655         } while (read_seqcount_retry(&tk_core.seq, seq));
 656
 657         ts->tv_nsec = 0;
 658         timespec64_add_ns(ts, nsecs);
 659
 660         /*
 661          * Do not bail out early, in case there were callers still using
 662          * the value, even in the face of the WARN_ON.
 663          */
 664         if (unlikely(timekeeping_suspended))
 665                 return -EAGAIN;
 666         return 0;
 667 }
 668 EXPORT_SYMBOL(__getnstimeofday64);
 669
 670 /**
 671  * getnstimeofday64 - Returns the time of day in a timespec64.
 672  * @ts:         pointer to the timespec64 to be set
 673  *
 674  * Returns the time of day in a timespec64 (WARN if suspended).
 675  */
 676 void getnstimeofday64(struct timespec64 *ts)
 677 {
 678         WARN_ON(__getnstimeofday64(ts));
 679 }
 680 EXPORT_SYMBOL(getnstimeofday64);
 681
 682 ktime_t ktime_get(void)
 683 {
 684         struct timekeeper *tk = &tk_core.timekeeper;
 685         unsigned int seq;
 686         ktime_t base;
 687         s64 nsecs;
 688
 689         WARN_ON(timekeeping_suspended);
 690
 691         do {
 692                 seq = read_seqcount_begin(&tk_core.seq);
 693                 base = tk->tkr_mono.base;
 694                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 695
 696         } while (read_seqcount_retry(&tk_core.seq, seq));
 697
 698         return ktime_add_ns(base, nsecs);
 699 }
 700 EXPORT_SYMBOL_GPL(ktime_get);
 701
 702 u32 ktime_get_resolution_ns(void)
 703 {
 704         struct timekeeper *tk = &tk_core.timekeeper;
 705         unsigned int seq;
 706         u32 nsecs;
 707
 708         WARN_ON(timekeeping_suspended);
 709
 710         do {
 711                 seq = read_seqcount_begin(&tk_core.seq);
 712                 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
 713         } while (read_seqcount_retry(&tk_core.seq, seq));
 714
 715         return nsecs;
 716 }
 717 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
 718
 719 static ktime_t *offsets[TK_OFFS_MAX] = {
 720         [TK_OFFS_REAL]  = &tk_core.timekeeper.offs_real,
 721         [TK_OFFS_BOOT]  = &tk_core.timekeeper.offs_boot,
 722         [TK_OFFS_TAI]   = &tk_core.timekeeper.offs_tai,
 723 };
 724
 725 ktime_t ktime_get_with_offset(enum tk_offsets offs)
 726 {
 727         struct timekeeper *tk = &tk_core.timekeeper;
 728         unsigned int seq;
 729         ktime_t base, *offset = offsets[offs];
 730         s64 nsecs;
 731
 732         WARN_ON(timekeeping_suspended);
 733
 734         do {
 735                 seq = read_seqcount_begin(&tk_core.seq);
 736                 base = ktime_add(tk->tkr_mono.base, *offset);
 737                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
 738
 739         } while (read_seqcount_retry(&tk_core.seq, seq));
 740
 741         return ktime_add_ns(base, nsecs);
 742
 743 }
 744 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
 745
 746 /**
 747  * ktime_mono_to_any() - convert mononotic time to any other time
 748  * @tmono:      time to convert.
 749  * @offs:       which offset to use
 750  */
 751 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
 752 {
 753         ktime_t *offset = offsets[offs];
 754         unsigned long seq;
 755         ktime_t tconv;
 756
 757         do {
 758                 seq = read_seqcount_begin(&tk_core.seq);
 759                 tconv = ktime_add(tmono, *offset);
 760         } while (read_seqcount_retry(&tk_core.seq, seq));
 761
 762         return tconv;
 763 }
 764 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
 765
 766 /**
 767  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
 768  */
 769 ktime_t ktime_get_raw(void)
 770 {
 771         struct timekeeper *tk = &tk_core.timekeeper;
 772         unsigned int seq;
 773         ktime_t base;
 774         s64 nsecs;
 775
 776         do {
 777                 seq = read_seqcount_begin(&tk_core.seq);
 778                 base = tk->tkr_raw.base;
 779                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
 780
 781         } while (read_seqcount_retry(&tk_core.seq, seq));
 782
 783         return ktime_add_ns(base, nsecs);
 784 }
 785 EXPORT_SYMBOL_GPL(ktime_get_raw);
 786
 787 /**
 788  * ktime_get_ts64 - get the monotonic clock in timespec64 format
 789  * @ts:         pointer to timespec variable
 790  *
 791  * The function calculates the monotonic clock from the realtime
 792  * clock and the wall_to_monotonic offset and stores the result
 793  * in normalized timespec64 format in the variable pointed to by @ts.
 794  */
 795 void ktime_get_ts64(struct timespec64 *ts)
 796 {
 797         struct timekeeper *tk = &tk_core.timekeeper;
 798         struct timespec64 tomono;
 799         s64 nsec;
 800         unsigned int seq;
 801
 802         WARN_ON(timekeeping_suspended);
 803
 804         do {
 805                 seq = read_seqcount_begin(&tk_core.seq);
 806                 ts->tv_sec = tk->xtime_sec;
 807                 nsec = timekeeping_get_ns(&tk->tkr_mono);
 808                 tomono = tk->wall_to_monotonic;
 809
 810         } while (read_seqcount_retry(&tk_core.seq, seq));
 811
 812         ts->tv_sec += tomono.tv_sec;
 813         ts->tv_nsec = 0;
 814         timespec64_add_ns(ts, nsec + tomono.tv_nsec);
 815 }
 816 EXPORT_SYMBOL_GPL(ktime_get_ts64);
 817
 818 /**
 819  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
 820  *
 821  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
 822  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
 823  * works on both 32 and 64 bit systems. On 32 bit systems the readout
 824  * covers ~136 years of uptime which should be enough to prevent
 825  * premature wrap arounds.
 826  */
 827 time64_t ktime_get_seconds(void)
 828 {
 829         struct timekeeper *tk = &tk_core.timekeeper;
 830
 831         WARN_ON(timekeeping_suspended);
 832         return tk->ktime_sec;
 833 }
 834 EXPORT_SYMBOL_GPL(ktime_get_seconds);
 835
 836 /**
 837  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
 838  *
 839  * Returns the wall clock seconds since 1970. This replaces the
 840  * get_seconds() interface which is not y2038 safe on 32bit systems.
 841  *
 842  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
 843  * 32bit systems the access must be protected with the sequence
 844  * counter to provide "atomic" access to the 64bit tk->xtime_sec
 845  * value.
 846  */
 847 time64_t ktime_get_real_seconds(void)
 848 {
 849         struct timekeeper *tk = &tk_core.timekeeper;
 850         time64_t seconds;
 851         unsigned int seq;
 852
 853         if (IS_ENABLED(CONFIG_64BIT))
 854                 return tk->xtime_sec;
 855
 856         do {
 857                 seq = read_seqcount_begin(&tk_core.seq);
 858                 seconds = tk->xtime_sec;
 859
 860         } while (read_seqcount_retry(&tk_core.seq, seq));
 861
 862         return seconds;
 863 }
 864 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
 865
 866 /**
 867  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
 868  * but without the sequence counter protect. This internal function
 869  * is called just when timekeeping lock is already held.
 870  */
 871 time64_t __ktime_get_real_seconds(void)
 872 {
 873         struct timekeeper *tk = &tk_core.timekeeper;
 874
 875         return tk->xtime_sec;
 876 }
 877
 878 /**
 879  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
 880  * @systime_snapshot:   pointer to struct receiving the system time snapshot
 881  */
 882 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
 883 {
 884         struct timekeeper *tk = &tk_core.timekeeper;
 885         unsigned long seq;
 886         ktime_t base_raw;
 887         ktime_t base_real;
 888         s64 nsec_raw;
 889         s64 nsec_real;
 890         cycle_t now;
 891
 892         WARN_ON_ONCE(timekeeping_suspended);
 893
 894         do {
 895                 seq = read_seqcount_begin(&tk_core.seq);
 896
 897                 now = tk->tkr_mono.read(tk->tkr_mono.clock);
 898                 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
 899                 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
 900                 base_real = ktime_add(tk->tkr_mono.base,
 901                                       tk_core.timekeeper.offs_real);
 902                 base_raw = tk->tkr_raw.base;
 903                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
 904                 nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
 905         } while (read_seqcount_retry(&tk_core.seq, seq));
 906
 907         systime_snapshot->cycles = now;
 908         systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
 909         systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
 910 }
 911 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
 912
 913 /* Scale base by mult/div checking for overflow */
 914 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
 915 {
 916         u64 tmp, rem;
 917
 918         tmp = div64_u64_rem(*base, div, &rem);
 919
 920         if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
 921             ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
 922                 return -EOVERFLOW;
 923         tmp *= mult;
 924         rem *= mult;
 925
 926         do_div(rem, div);
 927         *base = tmp + rem;
 928         return 0;
 929 }
 930
 931 /**
 932  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
 933  * @history:                    Snapshot representing start of history
 934  * @partial_history_cycles:     Cycle offset into history (fractional part)
 935  * @total_history_cycles:       Total history length in cycles
 936  * @discontinuity:              True indicates clock was set on history period
 937  * @ts:                         Cross timestamp that should be adjusted using
 938  *      partial/total ratio
 939  *
 940  * Helper function used by get_device_system_crosststamp() to correct the
 941  * crosstimestamp corresponding to the start of the current interval to the
 942  * system counter value (timestamp point) provided by the driver. The
 943  * total_history_* quantities are the total history starting at the provided
 944  * reference point and ending at the start of the current interval. The cycle
 945  * count between the driver timestamp point and the start of the current
 946  * interval is partial_history_cycles.
 947  */
 948 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
 949                                          cycle_t partial_history_cycles,
 950                                          cycle_t total_history_cycles,
 951                                          bool discontinuity,
 952                                          struct system_device_crosststamp *ts)
 953 {
 954         struct timekeeper *tk = &tk_core.timekeeper;
 955         u64 corr_raw, corr_real;
 956         bool interp_forward;
 957         int ret;
 958
 959         if (total_history_cycles == 0 || partial_history_cycles == 0)
 960                 return 0;
 961
 962         /* Interpolate shortest distance from beginning or end of history */
 963         interp_forward = partial_history_cycles > total_history_cycles/2 ?
 964                 true : false;
 965         partial_history_cycles = interp_forward ?
 966                 total_history_cycles - partial_history_cycles :
 967                 partial_history_cycles;
 968
 969         /*
 970          * Scale the monotonic raw time delta by:
 971          *      partial_history_cycles / total_history_cycles
 972          */
 973         corr_raw = (u64)ktime_to_ns(
 974                 ktime_sub(ts->sys_monoraw, history->raw));
 975         ret = scale64_check_overflow(partial_history_cycles,
 976                                      total_history_cycles, &corr_raw);
 977         if (ret)
 978                 return ret;
 979
 980         /*
 981          * If there is a discontinuity in the history, scale monotonic raw
 982          *      correction by:
 983          *      mult(real)/mult(raw) yielding the realtime correction
 984          * Otherwise, calculate the realtime correction similar to monotonic
 985          *      raw calculation
 986          */
 987         if (discontinuity) {
 988                 corr_real = mul_u64_u32_div
 989                         (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
 990         } else {
 991                 corr_real = (u64)ktime_to_ns(
 992                         ktime_sub(ts->sys_realtime, history->real));
 993                 ret = scale64_check_overflow(partial_history_cycles,
 994                                              total_history_cycles, &corr_real);
 995                 if (ret)
 996                         return ret;
 997         }
 998
 999         /* Fixup monotonic raw and real time time values */
1000         if (interp_forward) {
1001                 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1002                 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1003         } else {
1004                 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1005                 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1006         }
1007
1008         return 0;
1009 }
1010
1011 /*
1012  * cycle_between - true if test occurs chronologically between before and after
1013  */
1014 static bool cycle_between(cycle_t before, cycle_t test, cycle_t after)
1015 {
1016         if (test > before && test < after)
1017                 return true;
1018         if (test < before && before > after)
1019                 return true;
1020         return false;
1021 }
1022
1023 /**
1024  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1025  * @get_time_fn:        Callback to get simultaneous device time and
1026  *      system counter from the device driver
1027  * @ctx:                Context passed to get_time_fn()
1028  * @history_begin:      Historical reference point used to interpolate system
1029  *      time when counter provided by the driver is before the current interval
1030  * @xtstamp:            Receives simultaneously captured system and device time
1031  *
1032  * Reads a timestamp from a device and correlates it to system time
1033  */
1034 int get_device_system_crosststamp(int (*get_time_fn)
1035                                   (ktime_t *device_time,
1036                                    struct system_counterval_t *sys_counterval,
1037                                    void *ctx),
1038                                   void *ctx,
1039                                   struct system_time_snapshot *history_begin,
1040                                   struct system_device_crosststamp *xtstamp)
1041 {
1042         struct system_counterval_t system_counterval;
1043         struct timekeeper *tk = &tk_core.timekeeper;
1044         cycle_t cycles, now, interval_start;
1045         unsigned int clock_was_set_seq = 0;
1046         ktime_t base_real, base_raw;
1047         s64 nsec_real, nsec_raw;
1048         u8 cs_was_changed_seq;
1049         unsigned long seq;
1050         bool do_interp;
1051         int ret;
1052
1053         do {
1054                 seq = read_seqcount_begin(&tk_core.seq);
1055                 /*
1056                  * Try to synchronously capture device time and a system
1057                  * counter value calling back into the device driver
1058                  */
1059                 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1060                 if (ret)
1061                         return ret;
1062
1063                 /*
1064                  * Verify that the clocksource associated with the captured
1065                  * system counter value is the same as the currently installed
1066                  * timekeeper clocksource
1067                  */
1068                 if (tk->tkr_mono.clock != system_counterval.cs)
1069                         return -ENODEV;
1070                 cycles = system_counterval.cycles;
1071
1072                 /*
1073                  * Check whether the system counter value provided by the
1074                  * device driver is on the current timekeeping interval.
1075                  */
1076                 now = tk->tkr_mono.read(tk->tkr_mono.clock);
1077                 interval_start = tk->tkr_mono.cycle_last;
1078                 if (!cycle_between(interval_start, cycles, now)) {
1079                         clock_was_set_seq = tk->clock_was_set_seq;
1080                         cs_was_changed_seq = tk->cs_was_changed_seq;
1081                         cycles = interval_start;
1082                         do_interp = true;
1083                 } else {
1084                         do_interp = false;
1085                 }
1086
1087                 base_real = ktime_add(tk->tkr_mono.base,
1088                                       tk_core.timekeeper.offs_real);
1089                 base_raw = tk->tkr_raw.base;
1090
1091                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1092                                                      system_counterval.cycles);
1093                 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1094                                                     system_counterval.cycles);
1095         } while (read_seqcount_retry(&tk_core.seq, seq));
1096
1097         xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1098         xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1099
1100         /*
1101          * Interpolate if necessary, adjusting back from the start of the
1102          * current interval
1103          */
1104         if (do_interp) {
1105                 cycle_t partial_history_cycles, total_history_cycles;
1106                 bool discontinuity;
1107
1108                 /*
1109                  * Check that the counter value occurs after the provided
1110                  * history reference and that the history doesn't cross a
1111                  * clocksource change
1112                  */
1113                 if (!history_begin ||
1114                     !cycle_between(history_begin->cycles,
1115                                    system_counterval.cycles, cycles) ||
1116                     history_begin->cs_was_changed_seq != cs_was_changed_seq)
1117                         return -EINVAL;
1118                 partial_history_cycles = cycles - system_counterval.cycles;
1119                 total_history_cycles = cycles - history_begin->cycles;
1120                 discontinuity =
1121                         history_begin->clock_was_set_seq != clock_was_set_seq;
1122
1123                 ret = adjust_historical_crosststamp(history_begin,
1124                                                     partial_history_cycles,
1125                                                     total_history_cycles,
1126                                                     discontinuity, xtstamp);
1127                 if (ret)
1128                         return ret;
1129         }
1130
1131         return 0;
1132 }
1133 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1134
1135 /**
1136  * do_gettimeofday - Returns the time of day in a timeval
1137  * @tv:         pointer to the timeval to be set
1138  *
1139  * NOTE: Users should be converted to using getnstimeofday()
1140  */
1141 void do_gettimeofday(struct timeval *tv)
1142 {
1143         struct timespec64 now;
1144
1145         getnstimeofday64(&now);
1146         tv->tv_sec = now.tv_sec;
1147         tv->tv_usec = now.tv_nsec/1000;
1148 }
1149 EXPORT_SYMBOL(do_gettimeofday);
1150
1151 /**
1152  * do_settimeofday64 - Sets the time of day.
1153  * @ts:     pointer to the timespec64 variable containing the new time
1154  *
1155  * Sets the time of day to the new time and update NTP and notify hrtimers
1156  */
1157 int do_settimeofday64(const struct timespec64 *ts)
1158 {
1159         struct timekeeper *tk = &tk_core.timekeeper;
1160         struct timespec64 ts_delta, xt;
1161         unsigned long flags;
1162         int ret = 0;
1163
1164         if (!timespec64_valid_strict(ts))
1165                 return -EINVAL;
1166
1167         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1168         write_seqcount_begin(&tk_core.seq);
1169
1170         timekeeping_forward_now(tk);
1171
1172         xt = tk_xtime(tk);
1173         ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1174         ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1175
1176         if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1177                 ret = -EINVAL;
1178                 goto out;
1179         }
1180
1181         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1182
1183         tk_set_xtime(tk, ts);
1184 out:
1185         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1186
1187         write_seqcount_end(&tk_core.seq);
1188         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1189
1190         /* signal hrtimers about time change */
1191         clock_was_set();
1192
1193         return ret;
1194 }
1195 EXPORT_SYMBOL(do_settimeofday64);
1196
1197 /**
1198  * timekeeping_inject_offset - Adds or subtracts from the current time.
1199  * @tv:         pointer to the timespec variable containing the offset
1200  *
1201  * Adds or subtracts an offset value from the current time.
1202  */
1203 int timekeeping_inject_offset(struct timespec *ts)
1204 {
1205         struct timekeeper *tk = &tk_core.timekeeper;
1206         unsigned long flags;
1207         struct timespec64 ts64, tmp;
1208         int ret = 0;
1209
1210         if (!timespec_inject_offset_valid(ts))
1211                 return -EINVAL;
1212
1213         ts64 = timespec_to_timespec64(*ts);
1214
1215         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1216         write_seqcount_begin(&tk_core.seq);
1217
1218         timekeeping_forward_now(tk);
1219
1220         /* Make sure the proposed value is valid */
1221         tmp = timespec64_add(tk_xtime(tk),  ts64);
1222         if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1223             !timespec64_valid_strict(&tmp)) {
1224                 ret = -EINVAL;
1225                 goto error;
1226         }
1227
1228         tk_xtime_add(tk, &ts64);
1229         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1230
1231 error: /* even if we error out, we forwarded the time, so call update */
1232         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1233
1234         write_seqcount_end(&tk_core.seq);
1235         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1236
1237         /* signal hrtimers about time change */
1238         clock_was_set();
1239
1240         return ret;
1241 }
1242 EXPORT_SYMBOL(timekeeping_inject_offset);
1243
1244
1245 /**
1246  * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1247  *
1248  */
1249 s32 timekeeping_get_tai_offset(void)
1250 {
1251         struct timekeeper *tk = &tk_core.timekeeper;
1252         unsigned int seq;
1253         s32 ret;
1254
1255         do {
1256                 seq = read_seqcount_begin(&tk_core.seq);
1257                 ret = tk->tai_offset;
1258         } while (read_seqcount_retry(&tk_core.seq, seq));
1259
1260         return ret;
1261 }
1262
1263 /**
1264  * __timekeeping_set_tai_offset - Lock free worker function
1265  *
1266  */
1267 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1268 {
1269         tk->tai_offset = tai_offset;
1270         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1271 }
1272
1273 /**
1274  * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1275  *
1276  */
1277 void timekeeping_set_tai_offset(s32 tai_offset)
1278 {
1279         struct timekeeper *tk = &tk_core.timekeeper;
1280         unsigned long flags;
1281
1282         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1283         write_seqcount_begin(&tk_core.seq);
1284         __timekeeping_set_tai_offset(tk, tai_offset);
1285         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1286         write_seqcount_end(&tk_core.seq);
1287         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1288         clock_was_set();
1289 }
1290
1291 /**
1292  * change_clocksource - Swaps clocksources if a new one is available
1293  *
1294  * Accumulates current time interval and initializes new clocksource
1295  */
1296 static int change_clocksource(void *data)
1297 {
1298         struct timekeeper *tk = &tk_core.timekeeper;
1299         struct clocksource *new, *old;
1300         unsigned long flags;
1301
1302         new = (struct clocksource *) data;
1303
1304         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1305         write_seqcount_begin(&tk_core.seq);
1306
1307         timekeeping_forward_now(tk);
1308         /*
1309          * If the cs is in module, get a module reference. Succeeds
1310          * for built-in code (owner == NULL) as well.
1311          */
1312         if (try_module_get(new->owner)) {
1313                 if (!new->enable || new->enable(new) == 0) {
1314                         old = tk->tkr_mono.clock;
1315                         tk_setup_internals(tk, new);
1316                         if (old->disable)
1317                                 old->disable(old);
1318                         module_put(old->owner);
1319                 } else {
1320                         module_put(new->owner);
1321                 }
1322         }
1323         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1324
1325         write_seqcount_end(&tk_core.seq);
1326         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1327
1328         return 0;
1329 }
1330
1331 /**
1332  * timekeeping_notify - Install a new clock source
1333  * @clock:              pointer to the clock source
1334  *
1335  * This function is called from clocksource.c after a new, better clock
1336  * source has been registered. The caller holds the clocksource_mutex.
1337  */
1338 int timekeeping_notify(struct clocksource *clock)
1339 {
1340         struct timekeeper *tk = &tk_core.timekeeper;
1341
1342         if (tk->tkr_mono.clock == clock)
1343                 return 0;
1344         stop_machine(change_clocksource, clock, NULL);
1345         tick_clock_notify();
1346         return tk->tkr_mono.clock == clock ? 0 : -1;
1347 }
1348
1349 /**
1350  * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1351  * @ts:         pointer to the timespec64 to be set
1352  *
1353  * Returns the raw monotonic time (completely un-modified by ntp)
1354  */
1355 void getrawmonotonic64(struct timespec64 *ts)
1356 {
1357         struct timekeeper *tk = &tk_core.timekeeper;
1358         struct timespec64 ts64;
1359         unsigned long seq;
1360         s64 nsecs;
1361
1362         do {
1363                 seq = read_seqcount_begin(&tk_core.seq);
1364                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1365                 ts64 = tk->raw_time;
1366
1367         } while (read_seqcount_retry(&tk_core.seq, seq));
1368
1369         timespec64_add_ns(&ts64, nsecs);
1370         *ts = ts64;
1371 }
1372 EXPORT_SYMBOL(getrawmonotonic64);
1373
1374
1375 /**
1376  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1377  */
1378 int timekeeping_valid_for_hres(void)
1379 {
1380         struct timekeeper *tk = &tk_core.timekeeper;
1381         unsigned long seq;
1382         int ret;
1383
1384         do {
1385                 seq = read_seqcount_begin(&tk_core.seq);
1386
1387                 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1388
1389         } while (read_seqcount_retry(&tk_core.seq, seq));
1390
1391         return ret;
1392 }
1393
1394 /**
1395  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1396  */
1397 u64 timekeeping_max_deferment(void)
1398 {
1399         struct timekeeper *tk = &tk_core.timekeeper;
1400         unsigned long seq;
1401         u64 ret;
1402
1403         do {
1404                 seq = read_seqcount_begin(&tk_core.seq);
1405
1406                 ret = tk->tkr_mono.clock->max_idle_ns;
1407
1408         } while (read_seqcount_retry(&tk_core.seq, seq));
1409
1410         return ret;
1411 }
1412
1413 /**
1414  * read_persistent_clock -  Return time from the persistent clock.
1415  *
1416  * Weak dummy function for arches that do not yet support it.
1417  * Reads the time from the battery backed persistent clock.
1418  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1419  *
1420  *  XXX - Do be sure to remove it once all arches implement it.
1421  */
1422 void __weak read_persistent_clock(struct timespec *ts)
1423 {
1424         ts->tv_sec = 0;
1425         ts->tv_nsec = 0;
1426 }
1427
1428 void __weak read_persistent_clock64(struct timespec64 *ts64)
1429 {
1430         struct timespec ts;
1431
1432         read_persistent_clock(&ts);
1433         *ts64 = timespec_to_timespec64(ts);
1434 }
1435
1436 /**
1437  * read_boot_clock64 -  Return time of the system start.
1438  *
1439  * Weak dummy function for arches that do not yet support it.
1440  * Function to read the exact time the system has been started.
1441  * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1442  *
1443  *  XXX - Do be sure to remove it once all arches implement it.
1444  */
1445 void __weak read_boot_clock64(struct timespec64 *ts)
1446 {
1447         ts->tv_sec = 0;
1448         ts->tv_nsec = 0;
1449 }
1450
1451 /* Flag for if timekeeping_resume() has injected sleeptime */
1452 static bool sleeptime_injected;
1453
1454 /* Flag for if there is a persistent clock on this platform */
1455 static bool persistent_clock_exists;
1456
1457 /*
1458  * timekeeping_init - Initializes the clocksource and common timekeeping values
1459  */
1460 void __init timekeeping_init(void)
1461 {
1462         struct timekeeper *tk = &tk_core.timekeeper;
1463         struct clocksource *clock;
1464         unsigned long flags;
1465         struct timespec64 now, boot, tmp;
1466
1467         read_persistent_clock64(&now);
1468         if (!timespec64_valid_strict(&now)) {
1469                 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1470                         "         Check your CMOS/BIOS settings.\n");
1471                 now.tv_sec = 0;
1472                 now.tv_nsec = 0;
1473         } else if (now.tv_sec || now.tv_nsec)
1474                 persistent_clock_exists = true;
1475
1476         read_boot_clock64(&boot);
1477         if (!timespec64_valid_strict(&boot)) {
1478                 pr_warn("WARNING: Boot clock returned invalid value!\n"
1479                         "         Check your CMOS/BIOS settings.\n");
1480                 boot.tv_sec = 0;
1481                 boot.tv_nsec = 0;
1482         }
1483
1484         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1485         write_seqcount_begin(&tk_core.seq);
1486         ntp_init();
1487
1488         clock = clocksource_default_clock();
1489         if (clock->enable)
1490                 clock->enable(clock);
1491         tk_setup_internals(tk, clock);
1492
1493         tk_set_xtime(tk, &now);
1494         tk->raw_time.tv_sec = 0;
1495         tk->raw_time.tv_nsec = 0;
1496         if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1497                 boot = tk_xtime(tk);
1498
1499         set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1500         tk_set_wall_to_mono(tk, tmp);
1501
1502         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1503
1504         write_seqcount_end(&tk_core.seq);
1505         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1506 }
1507
1508 /* time in seconds when suspend began for persistent clock */
1509 static struct timespec64 timekeeping_suspend_time;
1510
1511 /**
1512  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1513  * @delta: pointer to a timespec delta value
1514  *
1515  * Takes a timespec offset measuring a suspend interval and properly
1516  * adds the sleep offset to the timekeeping variables.
1517  */
1518 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1519                                            struct timespec64 *delta)
1520 {
1521         if (!timespec64_valid_strict(delta)) {
1522                 printk_deferred(KERN_WARNING
1523                                 "__timekeeping_inject_sleeptime: Invalid "
1524                                 "sleep delta value!\n");
1525                 return;
1526         }
1527         tk_xtime_add(tk, delta);
1528         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1529         tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1530         tk_debug_account_sleep_time(delta);
1531 }
1532
1533 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1534 /**
1535  * We have three kinds of time sources to use for sleep time
1536  * injection, the preference order is:
1537  * 1) non-stop clocksource
1538  * 2) persistent clock (ie: RTC accessible when irqs are off)
1539  * 3) RTC
1540  *
1541  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1542  * If system has neither 1) nor 2), 3) will be used finally.
1543  *
1544  *
1545  * If timekeeping has injected sleeptime via either 1) or 2),
1546  * 3) becomes needless, so in this case we don't need to call
1547  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1548  * means.
1549  */
1550 bool timekeeping_rtc_skipresume(void)
1551 {
1552         return sleeptime_injected;
1553 }
1554
1555 /**
1556  * 1) can be determined whether to use or not only when doing
1557  * timekeeping_resume() which is invoked after rtc_suspend(),
1558  * so we can't skip rtc_suspend() surely if system has 1).
1559  *
1560  * But if system has 2), 2) will definitely be used, so in this
1561  * case we don't need to call rtc_suspend(), and this is what
1562  * timekeeping_rtc_skipsuspend() means.
1563  */
1564 bool timekeeping_rtc_skipsuspend(void)
1565 {
1566         return persistent_clock_exists;
1567 }
1568
1569 /**
1570  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1571  * @delta: pointer to a timespec64 delta value
1572  *
1573  * This hook is for architectures that cannot support read_persistent_clock64
1574  * because their RTC/persistent clock is only accessible when irqs are enabled.
1575  * and also don't have an effective nonstop clocksource.
1576  *
1577  * This function should only be called by rtc_resume(), and allows
1578  * a suspend offset to be injected into the timekeeping values.
1579  */
1580 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1581 {
1582         struct timekeeper *tk = &tk_core.timekeeper;
1583         unsigned long flags;
1584
1585         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1586         write_seqcount_begin(&tk_core.seq);
1587
1588         timekeeping_forward_now(tk);
1589
1590         __timekeeping_inject_sleeptime(tk, delta);
1591
1592         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1593
1594         write_seqcount_end(&tk_core.seq);
1595         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1596
1597         /* signal hrtimers about time change */
1598         clock_was_set();
1599 }
1600 #endif
1601
1602 /**
1603  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1604  */
1605 void timekeeping_resume(void)
1606 {
1607         struct timekeeper *tk = &tk_core.timekeeper;
1608         struct clocksource *clock = tk->tkr_mono.clock;
1609         unsigned long flags;
1610         struct timespec64 ts_new, ts_delta;
1611         cycle_t cycle_now, cycle_delta;
1612
1613         sleeptime_injected = false;
1614         read_persistent_clock64(&ts_new);
1615
1616         clockevents_resume();
1617         clocksource_resume();
1618
1619         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1620         write_seqcount_begin(&tk_core.seq);
1621
1622         /*
1623          * After system resumes, we need to calculate the suspended time and
1624          * compensate it for the OS time. There are 3 sources that could be
1625          * used: Nonstop clocksource during suspend, persistent clock and rtc
1626          * device.
1627          *
1628          * One specific platform may have 1 or 2 or all of them, and the
1629          * preference will be:
1630          *      suspend-nonstop clocksource -> persistent clock -> rtc
1631          * The less preferred source will only be tried if there is no better
1632          * usable source. The rtc part is handled separately in rtc core code.
1633          */
1634         cycle_now = tk->tkr_mono.read(clock);
1635         if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1636                 cycle_now > tk->tkr_mono.cycle_last) {
1637                 u64 num, max = ULLONG_MAX;
1638                 u32 mult = clock->mult;
1639                 u32 shift = clock->shift;
1640                 s64 nsec = 0;
1641
1642                 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1643                                                 tk->tkr_mono.mask);
1644
1645                 /*
1646                  * "cycle_delta * mutl" may cause 64 bits overflow, if the
1647                  * suspended time is too long. In that case we need do the
1648                  * 64 bits math carefully
1649                  */
1650                 do_div(max, mult);
1651                 if (cycle_delta > max) {
1652                         num = div64_u64(cycle_delta, max);
1653                         nsec = (((u64) max * mult) >> shift) * num;
1654                         cycle_delta -= num * max;
1655                 }
1656                 nsec += ((u64) cycle_delta * mult) >> shift;
1657
1658                 ts_delta = ns_to_timespec64(nsec);
1659                 sleeptime_injected = true;
1660         } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1661                 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1662                 sleeptime_injected = true;
1663         }
1664
1665         if (sleeptime_injected)
1666                 __timekeeping_inject_sleeptime(tk, &ts_delta);
1667
1668         /* Re-base the last cycle value */
1669         tk->tkr_mono.cycle_last = cycle_now;
1670         tk->tkr_raw.cycle_last  = cycle_now;
1671
1672         tk->ntp_error = 0;
1673         timekeeping_suspended = 0;
1674         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1675         write_seqcount_end(&tk_core.seq);
1676         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1677
1678         touch_softlockup_watchdog();
1679
1680         tick_resume();
1681         hrtimers_resume();
1682 }
1683
1684 int timekeeping_suspend(void)
1685 {
1686         struct timekeeper *tk = &tk_core.timekeeper;
1687         unsigned long flags;
1688         struct timespec64               delta, delta_delta;
1689         static struct timespec64        old_delta;
1690
1691         read_persistent_clock64(&timekeeping_suspend_time);
1692
1693         /*
1694          * On some systems the persistent_clock can not be detected at
1695          * timekeeping_init by its return value, so if we see a valid
1696          * value returned, update the persistent_clock_exists flag.
1697          */
1698         if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1699                 persistent_clock_exists = true;
1700
1701         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1702         write_seqcount_begin(&tk_core.seq);
1703         timekeeping_forward_now(tk);
1704         timekeeping_suspended = 1;
1705
1706         if (persistent_clock_exists) {
1707                 /*
1708                  * To avoid drift caused by repeated suspend/resumes,
1709                  * which each can add ~1 second drift error,
1710                  * try to compensate so the difference in system time
1711                  * and persistent_clock time stays close to constant.
1712                  */
1713                 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1714                 delta_delta = timespec64_sub(delta, old_delta);
1715                 if (abs(delta_delta.tv_sec) >= 2) {
1716                         /*
1717                          * if delta_delta is too large, assume time correction
1718                          * has occurred and set old_delta to the current delta.
1719                          */
1720                         old_delta = delta;
1721                 } else {
1722                         /* Otherwise try to adjust old_system to compensate */
1723                         timekeeping_suspend_time =
1724                                 timespec64_add(timekeeping_suspend_time, delta_delta);
1725                 }
1726         }
1727
1728         timekeeping_update(tk, TK_MIRROR);
1729         halt_fast_timekeeper(tk);
1730         write_seqcount_end(&tk_core.seq);
1731         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1732
1733         tick_suspend();
1734         clocksource_suspend();
1735         clockevents_suspend();
1736
1737         return 0;
1738 }
1739
1740 /* sysfs resume/suspend bits for timekeeping */
1741 static struct syscore_ops timekeeping_syscore_ops = {
1742         .resume         = timekeeping_resume,
1743         .suspend        = timekeeping_suspend,
1744 };
1745
1746 static int __init timekeeping_init_ops(void)
1747 {
1748         register_syscore_ops(&timekeeping_syscore_ops);
1749         return 0;
1750 }
1751 device_initcall(timekeeping_init_ops);
1752
1753 /*
1754  * Apply a multiplier adjustment to the timekeeper
1755  */
1756 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1757                                                          s64 offset,
1758                                                          bool negative,
1759                                                          int adj_scale)
1760 {
1761         s64 interval = tk->cycle_interval;
1762         s32 mult_adj = 1;
1763
1764         if (negative) {
1765                 mult_adj = -mult_adj;
1766                 interval = -interval;
1767                 offset  = -offset;
1768         }
1769         mult_adj <<= adj_scale;
1770         interval <<= adj_scale;
1771         offset <<= adj_scale;
1772
1773         /*
1774          * So the following can be confusing.
1775          *
1776          * To keep things simple, lets assume mult_adj == 1 for now.
1777          *
1778          * When mult_adj != 1, remember that the interval and offset values
1779          * have been appropriately scaled so the math is the same.
1780          *
1781          * The basic idea here is that we're increasing the multiplier
1782          * by one, this causes the xtime_interval to be incremented by
1783          * one cycle_interval. This is because:
1784          *      xtime_interval = cycle_interval * mult
1785          * So if mult is being incremented by one:
1786          *      xtime_interval = cycle_interval * (mult + 1)
1787          * Its the same as:
1788          *      xtime_interval = (cycle_interval * mult) + cycle_interval
1789          * Which can be shortened to:
1790          *      xtime_interval += cycle_interval
1791          *
1792          * So offset stores the non-accumulated cycles. Thus the current
1793          * time (in shifted nanoseconds) is:
1794          *      now = (offset * adj) + xtime_nsec
1795          * Now, even though we're adjusting the clock frequency, we have
1796          * to keep time consistent. In other words, we can't jump back
1797          * in time, and we also want to avoid jumping forward in time.
1798          *
1799          * So given the same offset value, we need the time to be the same
1800          * both before and after the freq adjustment.
1801          *      now = (offset * adj_1) + xtime_nsec_1
1802          *      now = (offset * adj_2) + xtime_nsec_2
1803          * So:
1804          *      (offset * adj_1) + xtime_nsec_1 =
1805          *              (offset * adj_2) + xtime_nsec_2
1806          * And we know:
1807          *      adj_2 = adj_1 + 1
1808          * So:
1809          *      (offset * adj_1) + xtime_nsec_1 =
1810          *              (offset * (adj_1+1)) + xtime_nsec_2
1811          *      (offset * adj_1) + xtime_nsec_1 =
1812          *              (offset * adj_1) + offset + xtime_nsec_2
1813          * Canceling the sides:
1814          *      xtime_nsec_1 = offset + xtime_nsec_2
1815          * Which gives us:
1816          *      xtime_nsec_2 = xtime_nsec_1 - offset
1817          * Which simplfies to:
1818          *      xtime_nsec -= offset
1819          *
1820          * XXX - TODO: Doc ntp_error calculation.
1821          */
1822         if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1823                 /* NTP adjustment caused clocksource mult overflow */
1824                 WARN_ON_ONCE(1);
1825                 return;
1826         }
1827
1828         tk->tkr_mono.mult += mult_adj;
1829         tk->xtime_interval += interval;
1830         tk->tkr_mono.xtime_nsec -= offset;
1831         tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1832 }
1833
1834 /*
1835  * Calculate the multiplier adjustment needed to match the frequency
1836  * specified by NTP
1837  */
1838 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1839                                                         s64 offset)
1840 {
1841         s64 interval = tk->cycle_interval;
1842         s64 xinterval = tk->xtime_interval;
1843         u32 base = tk->tkr_mono.clock->mult;
1844         u32 max = tk->tkr_mono.clock->maxadj;
1845         u32 cur_adj = tk->tkr_mono.mult;
1846         s64 tick_error;
1847         bool negative;
1848         u32 adj_scale;
1849
1850         /* Remove any current error adj from freq calculation */
1851         if (tk->ntp_err_mult)
1852                 xinterval -= tk->cycle_interval;
1853
1854         tk->ntp_tick = ntp_tick_length();
1855
1856         /* Calculate current error per tick */
1857         tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1858         tick_error -= (xinterval + tk->xtime_remainder);
1859
1860         /* Don't worry about correcting it if its small */
1861         if (likely((tick_error >= 0) && (tick_error <= interval)))
1862                 return;
1863
1864         /* preserve the direction of correction */
1865         negative = (tick_error < 0);
1866
1867         /* If any adjustment would pass the max, just return */
1868         if (negative && (cur_adj - 1) <= (base - max))
1869                 return;
1870         if (!negative && (cur_adj + 1) >= (base + max))
1871                 return;
1872         /*
1873          * Sort out the magnitude of the correction, but
1874          * avoid making so large a correction that we go
1875          * over the max adjustment.
1876          */
1877         adj_scale = 0;
1878         tick_error = abs(tick_error);
1879         while (tick_error > interval) {
1880                 u32 adj = 1 << (adj_scale + 1);
1881
1882                 /* Check if adjustment gets us within 1 unit from the max */
1883                 if (negative && (cur_adj - adj) <= (base - max))
1884                         break;
1885                 if (!negative && (cur_adj + adj) >= (base + max))
1886                         break;
1887
1888                 adj_scale++;
1889                 tick_error >>= 1;
1890         }
1891
1892         /* scale the corrections */
1893         timekeeping_apply_adjustment(tk, offset, negative, adj_scale);
1894 }
1895
1896 /*
1897  * Adjust the timekeeper's multiplier to the correct frequency
1898  * and also to reduce the accumulated error value.
1899  */
1900 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1901 {
1902         /* Correct for the current frequency error */
1903         timekeeping_freqadjust(tk, offset);
1904
1905         /* Next make a small adjustment to fix any cumulative error */
1906         if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1907                 tk->ntp_err_mult = 1;
1908                 timekeeping_apply_adjustment(tk, offset, 0, 0);
1909         } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1910                 /* Undo any existing error adjustment */
1911                 timekeeping_apply_adjustment(tk, offset, 1, 0);
1912                 tk->ntp_err_mult = 0;
1913         }
1914
1915         if (unlikely(tk->tkr_mono.clock->maxadj &&
1916                 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1917                         > tk->tkr_mono.clock->maxadj))) {
1918                 printk_once(KERN_WARNING
1919                         "Adjusting %s more than 11%% (%ld vs %ld)\n",
1920                         tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1921                         (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1922         }
1923
1924         /*
1925          * It may be possible that when we entered this function, xtime_nsec
1926          * was very small.  Further, if we're slightly speeding the clocksource
1927          * in the code above, its possible the required corrective factor to
1928          * xtime_nsec could cause it to underflow.
1929          *
1930          * Now, since we already accumulated the second, cannot simply roll
1931          * the accumulated second back, since the NTP subsystem has been
1932          * notified via second_overflow. So instead we push xtime_nsec forward
1933          * by the amount we underflowed, and add that amount into the error.
1934          *
1935          * We'll correct this error next time through this function, when
1936          * xtime_nsec is not as small.
1937          */
1938         if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1939                 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1940                 tk->tkr_mono.xtime_nsec = 0;
1941                 tk->ntp_error += neg << tk->ntp_error_shift;
1942         }
1943 }
1944
1945 /**
1946  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1947  *
1948  * Helper function that accumulates the nsecs greater than a second
1949  * from the xtime_nsec field to the xtime_secs field.
1950  * It also calls into the NTP code to handle leapsecond processing.
1951  *
1952  */
1953 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1954 {
1955         u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1956         unsigned int clock_set = 0;
1957
1958         while (tk->tkr_mono.xtime_nsec >= nsecps) {
1959                 int leap;
1960
1961                 tk->tkr_mono.xtime_nsec -= nsecps;
1962                 tk->xtime_sec++;
1963
1964                 /* Figure out if its a leap sec and apply if needed */
1965                 leap = second_overflow(tk->xtime_sec);
1966                 if (unlikely(leap)) {
1967                         struct timespec64 ts;
1968
1969                         tk->xtime_sec += leap;
1970
1971                         ts.tv_sec = leap;
1972                         ts.tv_nsec = 0;
1973                         tk_set_wall_to_mono(tk,
1974                                 timespec64_sub(tk->wall_to_monotonic, ts));
1975
1976                         __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1977
1978                         clock_set = TK_CLOCK_WAS_SET;
1979                 }
1980         }
1981         return clock_set;
1982 }
1983
1984 /**
1985  * logarithmic_accumulation - shifted accumulation of cycles
1986  *
1987  * This functions accumulates a shifted interval of cycles into
1988  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1989  * loop.
1990  *
1991  * Returns the unconsumed cycles.
1992  */
1993 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1994                                                 u32 shift,
1995                                                 unsigned int *clock_set)
1996 {
1997         cycle_t interval = tk->cycle_interval << shift;
1998         u64 raw_nsecs;
1999
2000         /* If the offset is smaller than a shifted interval, do nothing */
2001         if (offset < interval)
2002                 return offset;
2003
2004         /* Accumulate one shifted interval */
2005         offset -= interval;
2006         tk->tkr_mono.cycle_last += interval;
2007         tk->tkr_raw.cycle_last  += interval;
2008
2009         tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2010         *clock_set |= accumulate_nsecs_to_secs(tk);
2011
2012         /* Accumulate raw time */
2013         raw_nsecs = (u64)tk->raw_interval << shift;
2014         raw_nsecs += tk->raw_time.tv_nsec;
2015         if (raw_nsecs >= NSEC_PER_SEC) {
2016                 u64 raw_secs = raw_nsecs;
2017                 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
2018                 tk->raw_time.tv_sec += raw_secs;
2019         }
2020         tk->raw_time.tv_nsec = raw_nsecs;
2021
2022         /* Accumulate error between NTP and clock interval */
2023         tk->ntp_error += tk->ntp_tick << shift;
2024         tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2025                                                 (tk->ntp_error_shift + shift);
2026
2027         return offset;
2028 }
2029
2030 /**
2031  * update_wall_time - Uses the current clocksource to increment the wall time
2032  *
2033  */
2034 void update_wall_time(void)
2035 {
2036         struct timekeeper *real_tk = &tk_core.timekeeper;
2037         struct timekeeper *tk = &shadow_timekeeper;
2038         cycle_t offset;
2039         int shift = 0, maxshift;
2040         unsigned int clock_set = 0;
2041         unsigned long flags;
2042
2043         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2044
2045         /* Make sure we're fully resumed: */
2046         if (unlikely(timekeeping_suspended))
2047                 goto out;
2048
2049 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2050         offset = real_tk->cycle_interval;
2051 #else
2052         offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
2053                                    tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2054 #endif
2055
2056         /* Check if there's really nothing to do */
2057         if (offset < real_tk->cycle_interval)
2058                 goto out;
2059
2060         /* Do some additional sanity checking */
2061         timekeeping_check_update(real_tk, offset);
2062
2063         /*
2064          * With NO_HZ we may have to accumulate many cycle_intervals
2065          * (think "ticks") worth of time at once. To do this efficiently,
2066          * we calculate the largest doubling multiple of cycle_intervals
2067          * that is smaller than the offset.  We then accumulate that
2068          * chunk in one go, and then try to consume the next smaller
2069          * doubled multiple.
2070          */
2071         shift = ilog2(offset) - ilog2(tk->cycle_interval);
2072         shift = max(0, shift);
2073         /* Bound shift to one less than what overflows tick_length */
2074         maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2075         shift = min(shift, maxshift);
2076         while (offset >= tk->cycle_interval) {
2077                 offset = logarithmic_accumulation(tk, offset, shift,
2078                                                         &clock_set);
2079                 if (offset < tk->cycle_interval<<shift)
2080                         shift--;
2081         }
2082
2083         /* correct the clock when NTP error is too big */
2084         timekeeping_adjust(tk, offset);
2085
2086         /*
2087          * XXX This can be killed once everyone converts
2088          * to the new update_vsyscall.
2089          */
2090         old_vsyscall_fixup(tk);
2091
2092         /*
2093          * Finally, make sure that after the rounding
2094          * xtime_nsec isn't larger than NSEC_PER_SEC
2095          */
2096         clock_set |= accumulate_nsecs_to_secs(tk);
2097
2098         write_seqcount_begin(&tk_core.seq);
2099         /*
2100          * Update the real timekeeper.
2101          *
2102          * We could avoid this memcpy by switching pointers, but that
2103          * requires changes to all other timekeeper usage sites as
2104          * well, i.e. move the timekeeper pointer getter into the
2105          * spinlocked/seqcount protected sections. And we trade this
2106          * memcpy under the tk_core.seq against one before we start
2107          * updating.
2108          */
2109         timekeeping_update(tk, clock_set);
2110         memcpy(real_tk, tk, sizeof(*tk));
2111         /* The memcpy must come last. Do not put anything here! */
2112         write_seqcount_end(&tk_core.seq);
2113 out:
2114         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2115         if (clock_set)
2116                 /* Have to call _delayed version, since in irq context*/
2117                 clock_was_set_delayed();
2118 }
2119
2120 /**
2121  * getboottime64 - Return the real time of system boot.
2122  * @ts:         pointer to the timespec64 to be set
2123  *
2124  * Returns the wall-time of boot in a timespec64.
2125  *
2126  * This is based on the wall_to_monotonic offset and the total suspend
2127  * time. Calls to settimeofday will affect the value returned (which
2128  * basically means that however wrong your real time clock is at boot time,
2129  * you get the right time here).
2130  */
2131 void getboottime64(struct timespec64 *ts)
2132 {
2133         struct timekeeper *tk = &tk_core.timekeeper;
2134         ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2135
2136         *ts = ktime_to_timespec64(t);
2137 }
2138 EXPORT_SYMBOL_GPL(getboottime64);
2139
2140 unsigned long get_seconds(void)
2141 {
2142         struct timekeeper *tk = &tk_core.timekeeper;
2143
2144         return tk->xtime_sec;
2145 }
2146 EXPORT_SYMBOL(get_seconds);
2147
2148 struct timespec __current_kernel_time(void)
2149 {
2150         struct timekeeper *tk = &tk_core.timekeeper;
2151
2152         return timespec64_to_timespec(tk_xtime(tk));
2153 }
2154
2155 struct timespec64 current_kernel_time64(void)
2156 {
2157         struct timekeeper *tk = &tk_core.timekeeper;
2158         struct timespec64 now;
2159         unsigned long seq;
2160
2161         do {
2162                 seq = read_seqcount_begin(&tk_core.seq);
2163
2164                 now = tk_xtime(tk);
2165         } while (read_seqcount_retry(&tk_core.seq, seq));
2166
2167         return now;
2168 }
2169 EXPORT_SYMBOL(current_kernel_time64);
2170
2171 struct timespec64 get_monotonic_coarse64(void)
2172 {
2173         struct timekeeper *tk = &tk_core.timekeeper;
2174         struct timespec64 now, mono;
2175         unsigned long seq;
2176
2177         do {
2178                 seq = read_seqcount_begin(&tk_core.seq);
2179
2180                 now = tk_xtime(tk);
2181                 mono = tk->wall_to_monotonic;
2182         } while (read_seqcount_retry(&tk_core.seq, seq));
2183
2184         set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2185                                 now.tv_nsec + mono.tv_nsec);
2186
2187         return now;
2188 }
2189
2190 /*
2191  * Must hold jiffies_lock
2192  */
2193 void do_timer(unsigned long ticks)
2194 {
2195         jiffies_64 += ticks;
2196         calc_global_load(ticks);
2197 }
2198
2199 /**
2200  * ktime_get_update_offsets_now - hrtimer helper
2201  * @cwsseq:     pointer to check and store the clock was set sequence number
2202  * @offs_real:  pointer to storage for monotonic -> realtime offset
2203  * @offs_boot:  pointer to storage for monotonic -> boottime offset
2204  * @offs_tai:   pointer to storage for monotonic -> clock tai offset
2205  *
2206  * Returns current monotonic time and updates the offsets if the
2207  * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2208  * different.
2209  *
2210  * Called from hrtimer_interrupt() or retrigger_next_event()
2211  */
2212 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2213                                      ktime_t *offs_boot, ktime_t *offs_tai)
2214 {
2215         struct timekeeper *tk = &tk_core.timekeeper;
2216         unsigned int seq;
2217         ktime_t base;
2218         u64 nsecs;
2219
2220         do {
2221                 seq = read_seqcount_begin(&tk_core.seq);
2222
2223                 base = tk->tkr_mono.base;
2224                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2225                 base = ktime_add_ns(base, nsecs);
2226
2227                 if (*cwsseq != tk->clock_was_set_seq) {
2228                         *cwsseq = tk->clock_was_set_seq;
2229                         *offs_real = tk->offs_real;
2230                         *offs_boot = tk->offs_boot;
2231                         *offs_tai = tk->offs_tai;
2232                 }
2233
2234                 /* Handle leapsecond insertion adjustments */
2235                 if (unlikely(base.tv64 >= tk->next_leap_ktime.tv64))
2236                         *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2237
2238         } while (read_seqcount_retry(&tk_core.seq, seq));
2239
2240         return base;
2241 }
2242
2243 /**
2244  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2245  */
2246 int do_adjtimex(struct timex *txc)
2247 {
2248         struct timekeeper *tk = &tk_core.timekeeper;
2249         unsigned long flags;
2250         struct timespec64 ts;
2251         s32 orig_tai, tai;
2252         int ret;
2253
2254         /* Validate the data before disabling interrupts */
2255         ret = ntp_validate_timex(txc);
2256         if (ret)
2257                 return ret;
2258
2259         if (txc->modes & ADJ_SETOFFSET) {
2260                 struct timespec delta;
2261                 delta.tv_sec  = txc->time.tv_sec;
2262                 delta.tv_nsec = txc->time.tv_usec;
2263                 if (!(txc->modes & ADJ_NANO))
2264                         delta.tv_nsec *= 1000;
2265                 ret = timekeeping_inject_offset(&delta);
2266                 if (ret)
2267                         return ret;
2268         }
2269
2270         getnstimeofday64(&ts);
2271
2272         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2273         write_seqcount_begin(&tk_core.seq);
2274
2275         orig_tai = tai = tk->tai_offset;
2276         ret = __do_adjtimex(txc, &ts, &tai);
2277
2278         if (tai != orig_tai) {
2279                 __timekeeping_set_tai_offset(tk, tai);
2280                 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2281         }
2282         tk_update_leap_state(tk);
2283
2284         write_seqcount_end(&tk_core.seq);
2285         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2286
2287         if (tai != orig_tai)
2288                 clock_was_set();
2289
2290         ntp_notify_cmos_timer();
2291
2292         return ret;
2293 }
2294
2295 #ifdef CONFIG_NTP_PPS
2296 /**
2297  * hardpps() - Accessor function to NTP __hardpps function
2298  */
2299 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2300 {
2301         unsigned long flags;
2302
2303         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2304         write_seqcount_begin(&tk_core.seq);
2305
2306         __hardpps(phase_ts, raw_ts);
2307
2308         write_seqcount_end(&tk_core.seq);
2309         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2310 }
2311 EXPORT_SYMBOL(hardpps);
2312 #endif
2313
2314 /**
2315  * xtime_update() - advances the timekeeping infrastructure
2316  * @ticks:      number of ticks, that have elapsed since the last call.
2317  *
2318  * Must be called with interrupts disabled.
2319  */
2320 void xtime_update(unsigned long ticks)
2321 {
2322         write_seqlock(&jiffies_lock);
2323         do_timer(ticks);
2324         write_sequnlock(&jiffies_lock);
2325         update_wall_time();
2326 }