sys/kern/kern_tc.c

   1 /*-
   2  * SPDX-License-Identifier: Beerware
   3  *
   4  * ----------------------------------------------------------------------------
   5  * "THE BEER-WARE LICENSE" (Revision 42):
   6  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
   7  * can do whatever you want with this stuff. If we meet some day, and you think
   8  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
   9  * ----------------------------------------------------------------------------
  10  *
  11  * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
  12  * All rights reserved.
  13  *
  14  * Portions of this software were developed by Julien Ridoux at the University
  15  * of Melbourne under sponsorship from the FreeBSD Foundation.
  16  *
  17  * Portions of this software were developed by Konstantin Belousov
  18  * under sponsorship from the FreeBSD Foundation.
  19  */
  20
  21 #include <sys/cdefs.h>
  22 __FBSDID("$FreeBSD$");
  23
  24 #include "opt_ntp.h"
  25 #include "opt_ffclock.h"
  26
  27 #include <sys/param.h>
  28 #include <sys/kernel.h>
  29 #include <sys/limits.h>
  30 #include <sys/lock.h>
  31 #include <sys/mutex.h>
  32 #include <sys/proc.h>
  33 #include <sys/sbuf.h>
  34 #include <sys/sleepqueue.h>
  35 #include <sys/sysctl.h>
  36 #include <sys/syslog.h>
  37 #include <sys/systm.h>
  38 #include <sys/timeffc.h>
  39 #include <sys/timepps.h>
  40 #include <sys/timetc.h>
  41 #include <sys/timex.h>
  42 #include <sys/vdso.h>
  43
  44 /*
  45  * A large step happens on boot.  This constant detects such steps.
  46  * It is relatively small so that ntp_update_second gets called enough
  47  * in the typical 'missed a couple of seconds' case, but doesn't loop
  48  * forever when the time step is large.
  49  */
  50 #define LARGE_STEP      200
  51
  52 /*
  53  * Implement a dummy timecounter which we can use until we get a real one
  54  * in the air.  This allows the console and other early stuff to use
  55  * time services.
  56  */
  57
  58 static u_int
  59 dummy_get_timecount(struct timecounter *tc)
  60 {
  61         static u_int now;
  62
  63         return (++now);
  64 }
  65
  66 static struct timecounter dummy_timecounter = {
  67         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
  68 };
  69
  70 struct timehands {
  71         /* These fields must be initialized by the driver. */
  72         struct timecounter      *th_counter;
  73         int64_t                 th_adjustment;
  74         uint64_t                th_scale;
  75         u_int                   th_large_delta;
  76         u_int                   th_offset_count;
  77         struct bintime          th_offset;
  78         struct bintime          th_bintime;
  79         struct timeval          th_microtime;
  80         struct timespec         th_nanotime;
  81         struct bintime          th_boottime;
  82         /* Fields not to be copied in tc_windup start with th_generation. */
  83         u_int                   th_generation;
  84         struct timehands        *th_next;
  85 };
  86
  87 static struct timehands ths[16] = {
  88     [0] =  {
  89         .th_counter = &dummy_timecounter,
  90         .th_scale = (uint64_t)-1 / 1000000,
  91         .th_large_delta = 1000000,
  92         .th_offset = { .sec = 1 },
  93         .th_generation = 1,
  94     },
  95 };
  96
  97 static struct timehands *volatile timehands = &ths[0];
  98 struct timecounter *timecounter = &dummy_timecounter;
  99 static struct timecounter *timecounters = &dummy_timecounter;
 100
 101 /* Mutex to protect the timecounter list. */
 102 static struct mtx tc_lock;
 103
 104 int tc_min_ticktock_freq = 1;
 105
 106 volatile time_t time_second = 1;
 107 volatile time_t time_uptime = 1;
 108
 109 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
 110 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
 111     CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 112     sysctl_kern_boottime, "S,timeval",
 113     "System boottime");
 114
 115 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 116     "");
 117 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
 118     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 119     "");
 120
 121 static int timestepwarnings;
 122 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
 123     &timestepwarnings, 0, "Log time steps");
 124
 125 static int timehands_count = 2;
 126 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
 127     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 128     &timehands_count, 0, "Count of timehands in rotation");
 129
 130 struct bintime bt_timethreshold;
 131 struct bintime bt_tickthreshold;
 132 sbintime_t sbt_timethreshold;
 133 sbintime_t sbt_tickthreshold;
 134 struct bintime tc_tick_bt;
 135 sbintime_t tc_tick_sbt;
 136 int tc_precexp;
 137 int tc_timepercentage = TC_DEFAULTPERC;
 138 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
 139 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
 140     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
 141     sysctl_kern_timecounter_adjprecision, "I",
 142     "Allowed time interval deviation in percents");
 143
 144 volatile int rtc_generation = 1;
 145
 146 static int tc_chosen;   /* Non-zero if a specific tc was chosen via sysctl. */
 147 static char tc_from_tunable[16];
 148
 149 static void tc_windup(struct bintime *new_boottimebin);
 150 static void cpu_tick_calibrate(int);
 151
 152 void dtrace_getnanotime(struct timespec *tsp);
 153 void dtrace_getnanouptime(struct timespec *tsp);
 154
 155 static int
 156 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 157 {
 158         struct timeval boottime;
 159
 160         getboottime(&boottime);
 161
 162 /* i386 is the only arch which uses a 32bits time_t */
 163 #ifdef __amd64__
 164 #ifdef SCTL_MASK32
 165         int tv[2];
 166
 167         if (req->flags & SCTL_MASK32) {
 168                 tv[0] = boottime.tv_sec;
 169                 tv[1] = boottime.tv_usec;
 170                 return (SYSCTL_OUT(req, tv, sizeof(tv)));
 171         }
 172 #endif
 173 #endif
 174         return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
 175 }
 176
 177 static int
 178 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 179 {
 180         u_int ncount;
 181         struct timecounter *tc = arg1;
 182
 183         ncount = tc->tc_get_timecount(tc);
 184         return (sysctl_handle_int(oidp, &ncount, 0, req));
 185 }
 186
 187 static int
 188 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 189 {
 190         uint64_t freq;
 191         struct timecounter *tc = arg1;
 192
 193         freq = tc->tc_frequency;
 194         return (sysctl_handle_64(oidp, &freq, 0, req));
 195 }
 196
 197 /*
 198  * Return the difference between the timehands' counter value now and what
 199  * was when we copied it to the timehands' offset_count.
 200  */
 201 static __inline u_int
 202 tc_delta(struct timehands *th)
 203 {
 204         struct timecounter *tc;
 205
 206         tc = th->th_counter;
 207         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 208             tc->tc_counter_mask);
 209 }
 210
 211 static __inline void
 212 bintime_add_tc_delta(struct bintime *bt, uint64_t scale,
 213     uint64_t large_delta, uint64_t delta)
 214 {
 215         uint64_t x;
 216
 217         if (__predict_false(delta >= large_delta)) {
 218                 /* Avoid overflow for scale * delta. */
 219                 x = (scale >> 32) * delta;
 220                 bt->sec += x >> 32;
 221                 bintime_addx(bt, x << 32);
 222                 bintime_addx(bt, (scale & 0xffffffff) * delta);
 223         } else {
 224                 bintime_addx(bt, scale * delta);
 225         }
 226 }
 227
 228 /*
 229  * Functions for reading the time.  We have to loop until we are sure that
 230  * the timehands that we operated on was not updated under our feet.  See
 231  * the comment in <sys/time.h> for a description of these 12 functions.
 232  */
 233
 234 static __inline void
 235 bintime_off(struct bintime *bt, u_int off)
 236 {
 237         struct timehands *th;
 238         struct bintime *btp;
 239         uint64_t scale;
 240         u_int delta, gen, large_delta;
 241
 242         do {
 243                 th = timehands;
 244                 gen = atomic_load_acq_int(&th->th_generation);
 245                 btp = (struct bintime *)((vm_offset_t)th + off);
 246                 *bt = *btp;
 247                 scale = th->th_scale;
 248                 delta = tc_delta(th);
 249                 large_delta = th->th_large_delta;
 250                 atomic_thread_fence_acq();
 251         } while (gen == 0 || gen != th->th_generation);
 252
 253         bintime_add_tc_delta(bt, scale, large_delta, delta);
 254 }
 255 #define GETTHBINTIME(dst, member)                                       \
 256 do {                                                                    \
 257         _Static_assert(_Generic(((struct timehands *)NULL)->member,     \
 258             struct bintime: 1, default: 0) == 1,                        \
 259             "struct timehands member is not of struct bintime type");   \
 260         bintime_off(dst, __offsetof(struct timehands, member));         \
 261 } while (0)
 262
 263 static __inline void
 264 getthmember(void *out, size_t out_size, u_int off)
 265 {
 266         struct timehands *th;
 267         u_int gen;
 268
 269         do {
 270                 th = timehands;
 271                 gen = atomic_load_acq_int(&th->th_generation);
 272                 memcpy(out, (char *)th + off, out_size);
 273                 atomic_thread_fence_acq();
 274         } while (gen == 0 || gen != th->th_generation);
 275 }
 276 #define GETTHMEMBER(dst, member)                                        \
 277 do {                                                                    \
 278         _Static_assert(_Generic(*dst,                                   \
 279             __typeof(((struct timehands *)NULL)->member): 1,            \
 280             default: 0) == 1,                                           \
 281             "*dst and struct timehands member have different types");   \
 282         getthmember(dst, sizeof(*dst), __offsetof(struct timehands,     \
 283             member));                                                   \
 284 } while (0)
 285
 286 #ifdef FFCLOCK
 287 void
 288 fbclock_binuptime(struct bintime *bt)
 289 {
 290
 291         GETTHBINTIME(bt, th_offset);
 292 }
 293
 294 void
 295 fbclock_nanouptime(struct timespec *tsp)
 296 {
 297         struct bintime bt;
 298
 299         fbclock_binuptime(&bt);
 300         bintime2timespec(&bt, tsp);
 301 }
 302
 303 void
 304 fbclock_microuptime(struct timeval *tvp)
 305 {
 306         struct bintime bt;
 307
 308         fbclock_binuptime(&bt);
 309         bintime2timeval(&bt, tvp);
 310 }
 311
 312 void
 313 fbclock_bintime(struct bintime *bt)
 314 {
 315
 316         GETTHBINTIME(bt, th_bintime);
 317 }
 318
 319 void
 320 fbclock_nanotime(struct timespec *tsp)
 321 {
 322         struct bintime bt;
 323
 324         fbclock_bintime(&bt);
 325         bintime2timespec(&bt, tsp);
 326 }
 327
 328 void
 329 fbclock_microtime(struct timeval *tvp)
 330 {
 331         struct bintime bt;
 332
 333         fbclock_bintime(&bt);
 334         bintime2timeval(&bt, tvp);
 335 }
 336
 337 void
 338 fbclock_getbinuptime(struct bintime *bt)
 339 {
 340
 341         GETTHMEMBER(bt, th_offset);
 342 }
 343
 344 void
 345 fbclock_getnanouptime(struct timespec *tsp)
 346 {
 347         struct bintime bt;
 348
 349         GETTHMEMBER(&bt, th_offset);
 350         bintime2timespec(&bt, tsp);
 351 }
 352
 353 void
 354 fbclock_getmicrouptime(struct timeval *tvp)
 355 {
 356         struct bintime bt;
 357
 358         GETTHMEMBER(&bt, th_offset);
 359         bintime2timeval(&bt, tvp);
 360 }
 361
 362 void
 363 fbclock_getbintime(struct bintime *bt)
 364 {
 365
 366         GETTHMEMBER(bt, th_bintime);
 367 }
 368
 369 void
 370 fbclock_getnanotime(struct timespec *tsp)
 371 {
 372
 373         GETTHMEMBER(tsp, th_nanotime);
 374 }
 375
 376 void
 377 fbclock_getmicrotime(struct timeval *tvp)
 378 {
 379
 380         GETTHMEMBER(tvp, th_microtime);
 381 }
 382 #else /* !FFCLOCK */
 383
 384 void
 385 binuptime(struct bintime *bt)
 386 {
 387
 388         GETTHBINTIME(bt, th_offset);
 389 }
 390
 391 void
 392 nanouptime(struct timespec *tsp)
 393 {
 394         struct bintime bt;
 395
 396         binuptime(&bt);
 397         bintime2timespec(&bt, tsp);
 398 }
 399
 400 void
 401 microuptime(struct timeval *tvp)
 402 {
 403         struct bintime bt;
 404
 405         binuptime(&bt);
 406         bintime2timeval(&bt, tvp);
 407 }
 408
 409 void
 410 bintime(struct bintime *bt)
 411 {
 412
 413         GETTHBINTIME(bt, th_bintime);
 414 }
 415
 416 void
 417 nanotime(struct timespec *tsp)
 418 {
 419         struct bintime bt;
 420
 421         bintime(&bt);
 422         bintime2timespec(&bt, tsp);
 423 }
 424
 425 void
 426 microtime(struct timeval *tvp)
 427 {
 428         struct bintime bt;
 429
 430         bintime(&bt);
 431         bintime2timeval(&bt, tvp);
 432 }
 433
 434 void
 435 getbinuptime(struct bintime *bt)
 436 {
 437
 438         GETTHMEMBER(bt, th_offset);
 439 }
 440
 441 void
 442 getnanouptime(struct timespec *tsp)
 443 {
 444         struct bintime bt;
 445
 446         GETTHMEMBER(&bt, th_offset);
 447         bintime2timespec(&bt, tsp);
 448 }
 449
 450 void
 451 getmicrouptime(struct timeval *tvp)
 452 {
 453         struct bintime bt;
 454
 455         GETTHMEMBER(&bt, th_offset);
 456         bintime2timeval(&bt, tvp);
 457 }
 458
 459 void
 460 getbintime(struct bintime *bt)
 461 {
 462
 463         GETTHMEMBER(bt, th_bintime);
 464 }
 465
 466 void
 467 getnanotime(struct timespec *tsp)
 468 {
 469
 470         GETTHMEMBER(tsp, th_nanotime);
 471 }
 472
 473 void
 474 getmicrotime(struct timeval *tvp)
 475 {
 476
 477         GETTHMEMBER(tvp, th_microtime);
 478 }
 479 #endif /* FFCLOCK */
 480
 481 void
 482 getboottime(struct timeval *boottime)
 483 {
 484         struct bintime boottimebin;
 485
 486         getboottimebin(&boottimebin);
 487         bintime2timeval(&boottimebin, boottime);
 488 }
 489
 490 void
 491 getboottimebin(struct bintime *boottimebin)
 492 {
 493
 494         GETTHMEMBER(boottimebin, th_boottime);
 495 }
 496
 497 #ifdef FFCLOCK
 498 /*
 499  * Support for feed-forward synchronization algorithms. This is heavily inspired
 500  * by the timehands mechanism but kept independent from it. *_windup() functions
 501  * have some connection to avoid accessing the timecounter hardware more than
 502  * necessary.
 503  */
 504
 505 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
 506 struct ffclock_estimate ffclock_estimate;
 507 struct bintime ffclock_boottime;        /* Feed-forward boot time estimate. */
 508 uint32_t ffclock_status;                /* Feed-forward clock status. */
 509 int8_t ffclock_updated;                 /* New estimates are available. */
 510 struct mtx ffclock_mtx;                 /* Mutex on ffclock_estimate. */
 511
 512 struct fftimehands {
 513         struct ffclock_estimate cest;
 514         struct bintime          tick_time;
 515         struct bintime          tick_time_lerp;
 516         ffcounter               tick_ffcount;
 517         uint64_t                period_lerp;
 518         volatile uint8_t        gen;
 519         struct fftimehands      *next;
 520 };
 521
 522 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
 523
 524 static struct fftimehands ffth[10];
 525 static struct fftimehands *volatile fftimehands = ffth;
 526
 527 static void
 528 ffclock_init(void)
 529 {
 530         struct fftimehands *cur;
 531         struct fftimehands *last;
 532
 533         memset(ffth, 0, sizeof(ffth));
 534
 535         last = ffth + NUM_ELEMENTS(ffth) - 1;
 536         for (cur = ffth; cur < last; cur++)
 537                 cur->next = cur + 1;
 538         last->next = ffth;
 539
 540         ffclock_updated = 0;
 541         ffclock_status = FFCLOCK_STA_UNSYNC;
 542         mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
 543 }
 544
 545 /*
 546  * Reset the feed-forward clock estimates. Called from inittodr() to get things
 547  * kick started and uses the timecounter nominal frequency as a first period
 548  * estimate. Note: this function may be called several time just after boot.
 549  * Note: this is the only function that sets the value of boot time for the
 550  * monotonic (i.e. uptime) version of the feed-forward clock.
 551  */
 552 void
 553 ffclock_reset_clock(struct timespec *ts)
 554 {
 555         struct timecounter *tc;
 556         struct ffclock_estimate cest;
 557
 558         tc = timehands->th_counter;
 559         memset(&cest, 0, sizeof(struct ffclock_estimate));
 560
 561         timespec2bintime(ts, &ffclock_boottime);
 562         timespec2bintime(ts, &(cest.update_time));
 563         ffclock_read_counter(&cest.update_ffcount);
 564         cest.leapsec_next = 0;
 565         cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
 566         cest.errb_abs = 0;
 567         cest.errb_rate = 0;
 568         cest.status = FFCLOCK_STA_UNSYNC;
 569         cest.leapsec_total = 0;
 570         cest.leapsec = 0;
 571
 572         mtx_lock(&ffclock_mtx);
 573         bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
 574         ffclock_updated = INT8_MAX;
 575         mtx_unlock(&ffclock_mtx);
 576
 577         printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
 578             (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
 579             (unsigned long)ts->tv_nsec);
 580 }
 581
 582 /*
 583  * Sub-routine to convert a time interval measured in RAW counter units to time
 584  * in seconds stored in bintime format.
 585  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
 586  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
 587  * extra cycles.
 588  */
 589 static void
 590 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
 591 {
 592         struct bintime bt2;
 593         ffcounter delta, delta_max;
 594
 595         delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
 596         bintime_clear(bt);
 597         do {
 598                 if (ffdelta > delta_max)
 599                         delta = delta_max;
 600                 else
 601                         delta = ffdelta;
 602                 bt2.sec = 0;
 603                 bt2.frac = period;
 604                 bintime_mul(&bt2, (unsigned int)delta);
 605                 bintime_add(bt, &bt2);
 606                 ffdelta -= delta;
 607         } while (ffdelta > 0);
 608 }
 609
 610 /*
 611  * Update the fftimehands.
 612  * Push the tick ffcount and time(s) forward based on current clock estimate.
 613  * The conversion from ffcounter to bintime relies on the difference clock
 614  * principle, whose accuracy relies on computing small time intervals. If a new
 615  * clock estimate has been passed by the synchronisation daemon, make it
 616  * current, and compute the linear interpolation for monotonic time if needed.
 617  */
 618 static void
 619 ffclock_windup(unsigned int delta)
 620 {
 621         struct ffclock_estimate *cest;
 622         struct fftimehands *ffth;
 623         struct bintime bt, gap_lerp;
 624         ffcounter ffdelta;
 625         uint64_t frac;
 626         unsigned int polling;
 627         uint8_t forward_jump, ogen;
 628
 629         /*
 630          * Pick the next timehand, copy current ffclock estimates and move tick
 631          * times and counter forward.
 632          */
 633         forward_jump = 0;
 634         ffth = fftimehands->next;
 635         ogen = ffth->gen;
 636         ffth->gen = 0;
 637         cest = &ffth->cest;
 638         bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
 639         ffdelta = (ffcounter)delta;
 640         ffth->period_lerp = fftimehands->period_lerp;
 641
 642         ffth->tick_time = fftimehands->tick_time;
 643         ffclock_convert_delta(ffdelta, cest->period, &bt);
 644         bintime_add(&ffth->tick_time, &bt);
 645
 646         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
 647         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
 648         bintime_add(&ffth->tick_time_lerp, &bt);
 649
 650         ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
 651
 652         /*
 653          * Assess the status of the clock, if the last update is too old, it is
 654          * likely the synchronisation daemon is dead and the clock is free
 655          * running.
 656          */
 657         if (ffclock_updated == 0) {
 658                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
 659                 ffclock_convert_delta(ffdelta, cest->period, &bt);
 660                 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
 661                         ffclock_status |= FFCLOCK_STA_UNSYNC;
 662         }
 663
 664         /*
 665          * If available, grab updated clock estimates and make them current.
 666          * Recompute time at this tick using the updated estimates. The clock
 667          * estimates passed the feed-forward synchronisation daemon may result
 668          * in time conversion that is not monotonically increasing (just after
 669          * the update). time_lerp is a particular linear interpolation over the
 670          * synchronisation algo polling period that ensures monotonicity for the
 671          * clock ids requesting it.
 672          */
 673         if (ffclock_updated > 0) {
 674                 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
 675                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
 676                 ffth->tick_time = cest->update_time;
 677                 ffclock_convert_delta(ffdelta, cest->period, &bt);
 678                 bintime_add(&ffth->tick_time, &bt);
 679
 680                 /* ffclock_reset sets ffclock_updated to INT8_MAX */
 681                 if (ffclock_updated == INT8_MAX)
 682                         ffth->tick_time_lerp = ffth->tick_time;
 683
 684                 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
 685                         forward_jump = 1;
 686                 else
 687                         forward_jump = 0;
 688
 689                 bintime_clear(&gap_lerp);
 690                 if (forward_jump) {
 691                         gap_lerp = ffth->tick_time;
 692                         bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
 693                 } else {
 694                         gap_lerp = ffth->tick_time_lerp;
 695                         bintime_sub(&gap_lerp, &ffth->tick_time);
 696                 }
 697
 698                 /*
 699                  * The reset from the RTC clock may be far from accurate, and
 700                  * reducing the gap between real time and interpolated time
 701                  * could take a very long time if the interpolated clock insists
 702                  * on strict monotonicity. The clock is reset under very strict
 703                  * conditions (kernel time is known to be wrong and
 704                  * synchronization daemon has been restarted recently.
 705                  * ffclock_boottime absorbs the jump to ensure boot time is
 706                  * correct and uptime functions stay consistent.
 707                  */
 708                 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
 709                     ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
 710                     ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
 711                         if (forward_jump)
 712                                 bintime_add(&ffclock_boottime, &gap_lerp);
 713                         else
 714                                 bintime_sub(&ffclock_boottime, &gap_lerp);
 715                         ffth->tick_time_lerp = ffth->tick_time;
 716                         bintime_clear(&gap_lerp);
 717                 }
 718
 719                 ffclock_status = cest->status;
 720                 ffth->period_lerp = cest->period;
 721
 722                 /*
 723                  * Compute corrected period used for the linear interpolation of
 724                  * time. The rate of linear interpolation is capped to 5000PPM
 725                  * (5ms/s).
 726                  */
 727                 if (bintime_isset(&gap_lerp)) {
 728                         ffdelta = cest->update_ffcount;
 729                         ffdelta -= fftimehands->cest.update_ffcount;
 730                         ffclock_convert_delta(ffdelta, cest->period, &bt);
 731                         polling = bt.sec;
 732                         bt.sec = 0;
 733                         bt.frac = 5000000 * (uint64_t)18446744073LL;
 734                         bintime_mul(&bt, polling);
 735                         if (bintime_cmp(&gap_lerp, &bt, >))
 736                                 gap_lerp = bt;
 737
 738                         /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
 739                         frac = 0;
 740                         if (gap_lerp.sec > 0) {
 741                                 frac -= 1;
 742                                 frac /= ffdelta / gap_lerp.sec;
 743                         }
 744                         frac += gap_lerp.frac / ffdelta;
 745
 746                         if (forward_jump)
 747                                 ffth->period_lerp += frac;
 748                         else
 749                                 ffth->period_lerp -= frac;
 750                 }
 751
 752                 ffclock_updated = 0;
 753         }
 754         if (++ogen == 0)
 755                 ogen = 1;
 756         ffth->gen = ogen;
 757         fftimehands = ffth;
 758 }
 759
 760 /*
 761  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
 762  * the old and new hardware counter cannot be read simultaneously. tc_windup()
 763  * does read the two counters 'back to back', but a few cycles are effectively
 764  * lost, and not accumulated in tick_ffcount. This is a fairly radical
 765  * operation for a feed-forward synchronization daemon, and it is its job to not
 766  * pushing irrelevant data to the kernel. Because there is no locking here,
 767  * simply force to ignore pending or next update to give daemon a chance to
 768  * realize the counter has changed.
 769  */
 770 static void
 771 ffclock_change_tc(struct timehands *th)
 772 {
 773         struct fftimehands *ffth;
 774         struct ffclock_estimate *cest;
 775         struct timecounter *tc;
 776         uint8_t ogen;
 777
 778         tc = th->th_counter;
 779         ffth = fftimehands->next;
 780         ogen = ffth->gen;
 781         ffth->gen = 0;
 782
 783         cest = &ffth->cest;
 784         bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
 785         cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
 786         cest->errb_abs = 0;
 787         cest->errb_rate = 0;
 788         cest->status |= FFCLOCK_STA_UNSYNC;
 789
 790         ffth->tick_ffcount = fftimehands->tick_ffcount;
 791         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
 792         ffth->tick_time = fftimehands->tick_time;
 793         ffth->period_lerp = cest->period;
 794
 795         /* Do not lock but ignore next update from synchronization daemon. */
 796         ffclock_updated--;
 797
 798         if (++ogen == 0)
 799                 ogen = 1;
 800         ffth->gen = ogen;
 801         fftimehands = ffth;
 802 }
 803
 804 /*
 805  * Retrieve feed-forward counter and time of last kernel tick.
 806  */
 807 void
 808 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
 809 {
 810         struct fftimehands *ffth;
 811         uint8_t gen;
 812
 813         /*
 814          * No locking but check generation has not changed. Also need to make
 815          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
 816          */
 817         do {
 818                 ffth = fftimehands;
 819                 gen = ffth->gen;
 820                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
 821                         *bt = ffth->tick_time_lerp;
 822                 else
 823                         *bt = ffth->tick_time;
 824                 *ffcount = ffth->tick_ffcount;
 825         } while (gen == 0 || gen != ffth->gen);
 826 }
 827
 828 /*
 829  * Absolute clock conversion. Low level function to convert ffcounter to
 830  * bintime. The ffcounter is converted using the current ffclock period estimate
 831  * or the "interpolated period" to ensure monotonicity.
 832  * NOTE: this conversion may have been deferred, and the clock updated since the
 833  * hardware counter has been read.
 834  */
 835 void
 836 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
 837 {
 838         struct fftimehands *ffth;
 839         struct bintime bt2;
 840         ffcounter ffdelta;
 841         uint8_t gen;
 842
 843         /*
 844          * No locking but check generation has not changed. Also need to make
 845          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
 846          */
 847         do {
 848                 ffth = fftimehands;
 849                 gen = ffth->gen;
 850                 if (ffcount > ffth->tick_ffcount)
 851                         ffdelta = ffcount - ffth->tick_ffcount;
 852                 else
 853                         ffdelta = ffth->tick_ffcount - ffcount;
 854
 855                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
 856                         *bt = ffth->tick_time_lerp;
 857                         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
 858                 } else {
 859                         *bt = ffth->tick_time;
 860                         ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
 861                 }
 862
 863                 if (ffcount > ffth->tick_ffcount)
 864                         bintime_add(bt, &bt2);
 865                 else
 866                         bintime_sub(bt, &bt2);
 867         } while (gen == 0 || gen != ffth->gen);
 868 }
 869
 870 /*
 871  * Difference clock conversion.
 872  * Low level function to Convert a time interval measured in RAW counter units
 873  * into bintime. The difference clock allows measuring small intervals much more
 874  * reliably than the absolute clock.
 875  */
 876 void
 877 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
 878 {
 879         struct fftimehands *ffth;
 880         uint8_t gen;
 881
 882         /* No locking but check generation has not changed. */
 883         do {
 884                 ffth = fftimehands;
 885                 gen = ffth->gen;
 886                 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
 887         } while (gen == 0 || gen != ffth->gen);
 888 }
 889
 890 /*
 891  * Access to current ffcounter value.
 892  */
 893 void
 894 ffclock_read_counter(ffcounter *ffcount)
 895 {
 896         struct timehands *th;
 897         struct fftimehands *ffth;
 898         unsigned int gen, delta;
 899
 900         /*
 901          * ffclock_windup() called from tc_windup(), safe to rely on
 902          * th->th_generation only, for correct delta and ffcounter.
 903          */
 904         do {
 905                 th = timehands;
 906                 gen = atomic_load_acq_int(&th->th_generation);
 907                 ffth = fftimehands;
 908                 delta = tc_delta(th);
 909                 *ffcount = ffth->tick_ffcount;
 910                 atomic_thread_fence_acq();
 911         } while (gen == 0 || gen != th->th_generation);
 912
 913         *ffcount += delta;
 914 }
 915
 916 void
 917 binuptime(struct bintime *bt)
 918 {
 919
 920         binuptime_fromclock(bt, sysclock_active);
 921 }
 922
 923 void
 924 nanouptime(struct timespec *tsp)
 925 {
 926
 927         nanouptime_fromclock(tsp, sysclock_active);
 928 }
 929
 930 void
 931 microuptime(struct timeval *tvp)
 932 {
 933
 934         microuptime_fromclock(tvp, sysclock_active);
 935 }
 936
 937 void
 938 bintime(struct bintime *bt)
 939 {
 940
 941         bintime_fromclock(bt, sysclock_active);
 942 }
 943
 944 void
 945 nanotime(struct timespec *tsp)
 946 {
 947
 948         nanotime_fromclock(tsp, sysclock_active);
 949 }
 950
 951 void
 952 microtime(struct timeval *tvp)
 953 {
 954
 955         microtime_fromclock(tvp, sysclock_active);
 956 }
 957
 958 void
 959 getbinuptime(struct bintime *bt)
 960 {
 961
 962         getbinuptime_fromclock(bt, sysclock_active);
 963 }
 964
 965 void
 966 getnanouptime(struct timespec *tsp)
 967 {
 968
 969         getnanouptime_fromclock(tsp, sysclock_active);
 970 }
 971
 972 void
 973 getmicrouptime(struct timeval *tvp)
 974 {
 975
 976         getmicrouptime_fromclock(tvp, sysclock_active);
 977 }
 978
 979 void
 980 getbintime(struct bintime *bt)
 981 {
 982
 983         getbintime_fromclock(bt, sysclock_active);
 984 }
 985
 986 void
 987 getnanotime(struct timespec *tsp)
 988 {
 989
 990         getnanotime_fromclock(tsp, sysclock_active);
 991 }
 992
 993 void
 994 getmicrotime(struct timeval *tvp)
 995 {
 996
 997         getmicrouptime_fromclock(tvp, sysclock_active);
 998 }
 999
1000 #endif /* FFCLOCK */
1001
1002 /*
1003  * This is a clone of getnanotime and used for walltimestamps.
1004  * The dtrace_ prefix prevents fbt from creating probes for
1005  * it so walltimestamp can be safely used in all fbt probes.
1006  */
1007 void
1008 dtrace_getnanotime(struct timespec *tsp)
1009 {
1010
1011         GETTHMEMBER(tsp, th_nanotime);
1012 }
1013
1014 /*
1015  * This is a clone of getnanouptime used for time since boot.
1016  * The dtrace_ prefix prevents fbt from creating probes for
1017  * it so an uptime that can be safely used in all fbt probes.
1018  */
1019 void
1020 dtrace_getnanouptime(struct timespec *tsp)
1021 {
1022         struct bintime bt;
1023
1024         GETTHMEMBER(&bt, th_offset);
1025         bintime2timespec(&bt, tsp);
1026 }
1027
1028 /*
1029  * System clock currently providing time to the system. Modifiable via sysctl
1030  * when the FFCLOCK option is defined.
1031  */
1032 int sysclock_active = SYSCLOCK_FBCK;
1033
1034 /* Internal NTP status and error estimates. */
1035 extern int time_status;
1036 extern long time_esterror;
1037
1038 /*
1039  * Take a snapshot of sysclock data which can be used to compare system clocks
1040  * and generate timestamps after the fact.
1041  */
1042 void
1043 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1044 {
1045         struct fbclock_info *fbi;
1046         struct timehands *th;
1047         struct bintime bt;
1048         unsigned int delta, gen;
1049 #ifdef FFCLOCK
1050         ffcounter ffcount;
1051         struct fftimehands *ffth;
1052         struct ffclock_info *ffi;
1053         struct ffclock_estimate cest;
1054
1055         ffi = &clock_snap->ff_info;
1056 #endif
1057
1058         fbi = &clock_snap->fb_info;
1059         delta = 0;
1060
1061         do {
1062                 th = timehands;
1063                 gen = atomic_load_acq_int(&th->th_generation);
1064                 fbi->th_scale = th->th_scale;
1065                 fbi->tick_time = th->th_offset;
1066 #ifdef FFCLOCK
1067                 ffth = fftimehands;
1068                 ffi->tick_time = ffth->tick_time_lerp;
1069                 ffi->tick_time_lerp = ffth->tick_time_lerp;
1070                 ffi->period = ffth->cest.period;
1071                 ffi->period_lerp = ffth->period_lerp;
1072                 clock_snap->ffcount = ffth->tick_ffcount;
1073                 cest = ffth->cest;
1074 #endif
1075                 if (!fast)
1076                         delta = tc_delta(th);
1077                 atomic_thread_fence_acq();
1078         } while (gen == 0 || gen != th->th_generation);
1079
1080         clock_snap->delta = delta;
1081         clock_snap->sysclock_active = sysclock_active;
1082
1083         /* Record feedback clock status and error. */
1084         clock_snap->fb_info.status = time_status;
1085         /* XXX: Very crude estimate of feedback clock error. */
1086         bt.sec = time_esterror / 1000000;
1087         bt.frac = ((time_esterror - bt.sec) * 1000000) *
1088             (uint64_t)18446744073709ULL;
1089         clock_snap->fb_info.error = bt;
1090
1091 #ifdef FFCLOCK
1092         if (!fast)
1093                 clock_snap->ffcount += delta;
1094
1095         /* Record feed-forward clock leap second adjustment. */
1096         ffi->leapsec_adjustment = cest.leapsec_total;
1097         if (clock_snap->ffcount > cest.leapsec_next)
1098                 ffi->leapsec_adjustment -= cest.leapsec;
1099
1100         /* Record feed-forward clock status and error. */
1101         clock_snap->ff_info.status = cest.status;
1102         ffcount = clock_snap->ffcount - cest.update_ffcount;
1103         ffclock_convert_delta(ffcount, cest.period, &bt);
1104         /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1105         bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1106         /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1107         bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1108         clock_snap->ff_info.error = bt;
1109 #endif
1110 }
1111
1112 /*
1113  * Convert a sysclock snapshot into a struct bintime based on the specified
1114  * clock source and flags.
1115  */
1116 int
1117 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1118     int whichclock, uint32_t flags)
1119 {
1120         struct bintime boottimebin;
1121 #ifdef FFCLOCK
1122         struct bintime bt2;
1123         uint64_t period;
1124 #endif
1125
1126         switch (whichclock) {
1127         case SYSCLOCK_FBCK:
1128                 *bt = cs->fb_info.tick_time;
1129
1130                 /* If snapshot was created with !fast, delta will be >0. */
1131                 if (cs->delta > 0)
1132                         bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1133
1134                 if ((flags & FBCLOCK_UPTIME) == 0) {
1135                         getboottimebin(&boottimebin);
1136                         bintime_add(bt, &boottimebin);
1137                 }
1138                 break;
1139 #ifdef FFCLOCK
1140         case SYSCLOCK_FFWD:
1141                 if (flags & FFCLOCK_LERP) {
1142                         *bt = cs->ff_info.tick_time_lerp;
1143                         period = cs->ff_info.period_lerp;
1144                 } else {
1145                         *bt = cs->ff_info.tick_time;
1146                         period = cs->ff_info.period;
1147                 }
1148
1149                 /* If snapshot was created with !fast, delta will be >0. */
1150                 if (cs->delta > 0) {
1151                         ffclock_convert_delta(cs->delta, period, &bt2);
1152                         bintime_add(bt, &bt2);
1153                 }
1154
1155                 /* Leap second adjustment. */
1156                 if (flags & FFCLOCK_LEAPSEC)
1157                         bt->sec -= cs->ff_info.leapsec_adjustment;
1158
1159                 /* Boot time adjustment, for uptime/monotonic clocks. */
1160                 if (flags & FFCLOCK_UPTIME)
1161                         bintime_sub(bt, &ffclock_boottime);
1162                 break;
1163 #endif
1164         default:
1165                 return (EINVAL);
1166                 break;
1167         }
1168
1169         return (0);
1170 }
1171
1172 /*
1173  * Initialize a new timecounter and possibly use it.
1174  */
1175 void
1176 tc_init(struct timecounter *tc)
1177 {
1178         u_int u;
1179         struct sysctl_oid *tc_root;
1180
1181         u = tc->tc_frequency / tc->tc_counter_mask;
1182         /* XXX: We need some margin here, 10% is a guess */
1183         u *= 11;
1184         u /= 10;
1185         if (u > hz && tc->tc_quality >= 0) {
1186                 tc->tc_quality = -2000;
1187                 if (bootverbose) {
1188                         printf("Timecounter \"%s\" frequency %ju Hz",
1189                             tc->tc_name, (uintmax_t)tc->tc_frequency);
1190                         printf(" -- Insufficient hz, needs at least %u\n", u);
1191                 }
1192         } else if (tc->tc_quality >= 0 || bootverbose) {
1193                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1194                     tc->tc_name, (uintmax_t)tc->tc_frequency,
1195                     tc->tc_quality);
1196         }
1197
1198         /*
1199          * Set up sysctl tree for this counter.
1200          */
1201         tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1202             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1203             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1204             "timecounter description", "timecounter");
1205         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1206             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1207             "mask for implemented bits");
1208         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1209             "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1210             sizeof(*tc), sysctl_kern_timecounter_get, "IU",
1211             "current timecounter value");
1212         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1213             "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1214             sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
1215             "timecounter frequency");
1216         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1217             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1218             "goodness of time counter");
1219
1220         mtx_lock(&tc_lock);
1221         tc->tc_next = timecounters;
1222         timecounters = tc;
1223
1224         /*
1225          * Do not automatically switch if the current tc was specifically
1226          * chosen.  Never automatically use a timecounter with negative quality.
1227          * Even though we run on the dummy counter, switching here may be
1228          * worse since this timecounter may not be monotonic.
1229          */
1230         if (tc_chosen)
1231                 goto unlock;
1232         if (tc->tc_quality < 0)
1233                 goto unlock;
1234         if (tc_from_tunable[0] != '\0' &&
1235             strcmp(tc->tc_name, tc_from_tunable) == 0) {
1236                 tc_chosen = 1;
1237                 tc_from_tunable[0] = '\0';
1238         } else {
1239                 if (tc->tc_quality < timecounter->tc_quality)
1240                         goto unlock;
1241                 if (tc->tc_quality == timecounter->tc_quality &&
1242                     tc->tc_frequency < timecounter->tc_frequency)
1243                         goto unlock;
1244         }
1245         (void)tc->tc_get_timecount(tc);
1246         timecounter = tc;
1247 unlock:
1248         mtx_unlock(&tc_lock);
1249 }
1250
1251 /* Report the frequency of the current timecounter. */
1252 uint64_t
1253 tc_getfrequency(void)
1254 {
1255
1256         return (timehands->th_counter->tc_frequency);
1257 }
1258
1259 static bool
1260 sleeping_on_old_rtc(struct thread *td)
1261 {
1262
1263         /*
1264          * td_rtcgen is modified by curthread when it is running,
1265          * and by other threads in this function.  By finding the thread
1266          * on a sleepqueue and holding the lock on the sleepqueue
1267          * chain, we guarantee that the thread is not running and that
1268          * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
1269          * the thread that it was woken due to a real-time clock adjustment.
1270          * (The declaration of td_rtcgen refers to this comment.)
1271          */
1272         if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1273                 td->td_rtcgen = 0;
1274                 return (true);
1275         }
1276         return (false);
1277 }
1278
1279 static struct mtx tc_setclock_mtx;
1280 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1281
1282 /*
1283  * Step our concept of UTC.  This is done by modifying our estimate of
1284  * when we booted.
1285  */
1286 void
1287 tc_setclock(struct timespec *ts)
1288 {
1289         struct timespec tbef, taft;
1290         struct bintime bt, bt2;
1291
1292         timespec2bintime(ts, &bt);
1293         nanotime(&tbef);
1294         mtx_lock_spin(&tc_setclock_mtx);
1295         cpu_tick_calibrate(1);
1296         binuptime(&bt2);
1297         bintime_sub(&bt, &bt2);
1298
1299         /* XXX fiddle all the little crinkly bits around the fiords... */
1300         tc_windup(&bt);
1301         mtx_unlock_spin(&tc_setclock_mtx);
1302
1303         /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1304         atomic_add_rel_int(&rtc_generation, 2);
1305         sleepq_chains_remove_matching(sleeping_on_old_rtc);
1306         if (timestepwarnings) {
1307                 nanotime(&taft);
1308                 log(LOG_INFO,
1309                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1310                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1311                     (intmax_t)taft.tv_sec, taft.tv_nsec,
1312                     (intmax_t)ts->tv_sec, ts->tv_nsec);
1313         }
1314 }
1315
1316 /*
1317  * Recalculate the scaling factor.  We want the number of 1/2^64
1318  * fractions of a second per period of the hardware counter, taking
1319  * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1320  * processing provides us with.
1321  *
1322  * The th_adjustment is nanoseconds per second with 32 bit binary
1323  * fraction and we want 64 bit binary fraction of second:
1324  *
1325  *       x = a * 2^32 / 10^9 = a * 4.294967296
1326  *
1327  * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1328  * we can only multiply by about 850 without overflowing, that
1329  * leaves no suitably precise fractions for multiply before divide.
1330  *
1331  * Divide before multiply with a fraction of 2199/512 results in a
1332  * systematic undercompensation of 10PPM of th_adjustment.  On a
1333  * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
1334  *
1335  * We happily sacrifice the lowest of the 64 bits of our result
1336  * to the goddess of code clarity.
1337  */
1338 static void
1339 recalculate_scaling_factor_and_large_delta(struct timehands *th)
1340 {
1341         uint64_t scale;
1342
1343         scale = (uint64_t)1 << 63;
1344         scale += (th->th_adjustment / 1024) * 2199;
1345         scale /= th->th_counter->tc_frequency;
1346         th->th_scale = scale * 2;
1347         th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1348 }
1349
1350 /*
1351  * Initialize the next struct timehands in the ring and make
1352  * it the active timehands.  Along the way we might switch to a different
1353  * timecounter and/or do seconds processing in NTP.  Slightly magic.
1354  */
1355 static void
1356 tc_windup(struct bintime *new_boottimebin)
1357 {
1358         struct bintime bt;
1359         struct timehands *th, *tho;
1360         u_int delta, ncount, ogen;
1361         int i;
1362         time_t t;
1363
1364         /*
1365          * Make the next timehands a copy of the current one, but do
1366          * not overwrite the generation or next pointer.  While we
1367          * update the contents, the generation must be zero.  We need
1368          * to ensure that the zero generation is visible before the
1369          * data updates become visible, which requires release fence.
1370          * For similar reasons, re-reading of the generation after the
1371          * data is read should use acquire fence.
1372          */
1373         tho = timehands;
1374         th = tho->th_next;
1375         ogen = th->th_generation;
1376         th->th_generation = 0;
1377         atomic_thread_fence_rel();
1378         memcpy(th, tho, offsetof(struct timehands, th_generation));
1379         if (new_boottimebin != NULL)
1380                 th->th_boottime = *new_boottimebin;
1381
1382         /*
1383          * Capture a timecounter delta on the current timecounter and if
1384          * changing timecounters, a counter value from the new timecounter.
1385          * Update the offset fields accordingly.
1386          */
1387         delta = tc_delta(th);
1388         if (th->th_counter != timecounter)
1389                 ncount = timecounter->tc_get_timecount(timecounter);
1390         else
1391                 ncount = 0;
1392 #ifdef FFCLOCK
1393         ffclock_windup(delta);
1394 #endif
1395         th->th_offset_count += delta;
1396         th->th_offset_count &= th->th_counter->tc_counter_mask;
1397         bintime_add_tc_delta(&th->th_offset, th->th_scale,
1398             th->th_large_delta, delta);
1399
1400         /*
1401          * Hardware latching timecounters may not generate interrupts on
1402          * PPS events, so instead we poll them.  There is a finite risk that
1403          * the hardware might capture a count which is later than the one we
1404          * got above, and therefore possibly in the next NTP second which might
1405          * have a different rate than the current NTP second.  It doesn't
1406          * matter in practice.
1407          */
1408         if (tho->th_counter->tc_poll_pps)
1409                 tho->th_counter->tc_poll_pps(tho->th_counter);
1410
1411         /*
1412          * Deal with NTP second processing.  The loop normally
1413          * iterates at most once, but in extreme situations it might
1414          * keep NTP sane if timeouts are not run for several seconds.
1415          * At boot, the time step can be large when the TOD hardware
1416          * has been read, so on really large steps, we call
1417          * ntp_update_second only twice.  We need to call it twice in
1418          * case we missed a leap second.
1419          */
1420         bt = th->th_offset;
1421         bintime_add(&bt, &th->th_boottime);
1422         i = bt.sec - tho->th_microtime.tv_sec;
1423         if (i > 0) {
1424                 if (i > LARGE_STEP)
1425                         i = 2;
1426
1427                 do {
1428                         t = bt.sec;
1429                         ntp_update_second(&th->th_adjustment, &bt.sec);
1430                         if (bt.sec != t)
1431                                 th->th_boottime.sec += bt.sec - t;
1432                         --i;
1433                 } while (i > 0);
1434
1435                 recalculate_scaling_factor_and_large_delta(th);
1436         }
1437
1438         /* Update the UTC timestamps used by the get*() functions. */
1439         th->th_bintime = bt;
1440         bintime2timeval(&bt, &th->th_microtime);
1441         bintime2timespec(&bt, &th->th_nanotime);
1442
1443         /* Now is a good time to change timecounters. */
1444         if (th->th_counter != timecounter) {
1445 #ifndef __arm__
1446                 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1447                         cpu_disable_c2_sleep++;
1448                 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1449                         cpu_disable_c2_sleep--;
1450 #endif
1451                 th->th_counter = timecounter;
1452                 th->th_offset_count = ncount;
1453                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1454                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1455                 recalculate_scaling_factor_and_large_delta(th);
1456 #ifdef FFCLOCK
1457                 ffclock_change_tc(th);
1458 #endif
1459         }
1460
1461         /*
1462          * Now that the struct timehands is again consistent, set the new
1463          * generation number, making sure to not make it zero.
1464          */
1465         if (++ogen == 0)
1466                 ogen = 1;
1467         atomic_store_rel_int(&th->th_generation, ogen);
1468
1469         /* Go live with the new struct timehands. */
1470 #ifdef FFCLOCK
1471         switch (sysclock_active) {
1472         case SYSCLOCK_FBCK:
1473 #endif
1474                 time_second = th->th_microtime.tv_sec;
1475                 time_uptime = th->th_offset.sec;
1476 #ifdef FFCLOCK
1477                 break;
1478         case SYSCLOCK_FFWD:
1479                 time_second = fftimehands->tick_time_lerp.sec;
1480                 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1481                 break;
1482         }
1483 #endif
1484
1485         timehands = th;
1486         timekeep_push_vdso();
1487 }
1488
1489 /* Report or change the active timecounter hardware. */
1490 static int
1491 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1492 {
1493         char newname[32];
1494         struct timecounter *newtc, *tc;
1495         int error;
1496
1497         mtx_lock(&tc_lock);
1498         tc = timecounter;
1499         strlcpy(newname, tc->tc_name, sizeof(newname));
1500         mtx_unlock(&tc_lock);
1501
1502         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1503         if (error != 0 || req->newptr == NULL)
1504                 return (error);
1505
1506         mtx_lock(&tc_lock);
1507         /* Record that the tc in use now was specifically chosen. */
1508         tc_chosen = 1;
1509         if (strcmp(newname, tc->tc_name) == 0) {
1510                 mtx_unlock(&tc_lock);
1511                 return (0);
1512         }
1513         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1514                 if (strcmp(newname, newtc->tc_name) != 0)
1515                         continue;
1516
1517                 /* Warm up new timecounter. */
1518                 (void)newtc->tc_get_timecount(newtc);
1519
1520                 timecounter = newtc;
1521
1522                 /*
1523                  * The vdso timehands update is deferred until the next
1524                  * 'tc_windup()'.
1525                  *
1526                  * This is prudent given that 'timekeep_push_vdso()' does not
1527                  * use any locking and that it can be called in hard interrupt
1528                  * context via 'tc_windup()'.
1529                  */
1530                 break;
1531         }
1532         mtx_unlock(&tc_lock);
1533         return (newtc != NULL ? 0 : EINVAL);
1534 }
1535 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1536     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1537     sysctl_kern_timecounter_hardware, "A",
1538     "Timecounter hardware selected");
1539
1540 /* Report the available timecounter hardware. */
1541 static int
1542 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1543 {
1544         struct sbuf sb;
1545         struct timecounter *tc;
1546         int error;
1547
1548         error = sysctl_wire_old_buffer(req, 0);
1549         if (error != 0)
1550                 return (error);
1551         sbuf_new_for_sysctl(&sb, NULL, 0, req);
1552         mtx_lock(&tc_lock);
1553         for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1554                 if (tc != timecounters)
1555                         sbuf_putc(&sb, ' ');
1556                 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1557         }
1558         mtx_unlock(&tc_lock);
1559         error = sbuf_finish(&sb);
1560         sbuf_delete(&sb);
1561         return (error);
1562 }
1563
1564 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
1565     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
1566     sysctl_kern_timecounter_choice, "A",
1567     "Timecounter hardware detected");
1568
1569 /*
1570  * RFC 2783 PPS-API implementation.
1571  */
1572
1573 /*
1574  *  Return true if the driver is aware of the abi version extensions in the
1575  *  pps_state structure, and it supports at least the given abi version number.
1576  */
1577 static inline int
1578 abi_aware(struct pps_state *pps, int vers)
1579 {
1580
1581         return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1582 }
1583
1584 static int
1585 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1586 {
1587         int err, timo;
1588         pps_seq_t aseq, cseq;
1589         struct timeval tv;
1590
1591         if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1592                 return (EINVAL);
1593
1594         /*
1595          * If no timeout is requested, immediately return whatever values were
1596          * most recently captured.  If timeout seconds is -1, that's a request
1597          * to block without a timeout.  WITNESS won't let us sleep forever
1598          * without a lock (we really don't need a lock), so just repeatedly
1599          * sleep a long time.
1600          */
1601         if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1602                 if (fapi->timeout.tv_sec == -1)
1603                         timo = 0x7fffffff;
1604                 else {
1605                         tv.tv_sec = fapi->timeout.tv_sec;
1606                         tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1607                         timo = tvtohz(&tv);
1608                 }
1609                 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1610                 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1611                 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1612                     cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1613                         if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1614                                 if (pps->flags & PPSFLAG_MTX_SPIN) {
1615                                         err = msleep_spin(pps, pps->driver_mtx,
1616                                             "ppsfch", timo);
1617                                 } else {
1618                                         err = msleep(pps, pps->driver_mtx, PCATCH,
1619                                             "ppsfch", timo);
1620                                 }
1621                         } else {
1622                                 err = tsleep(pps, PCATCH, "ppsfch", timo);
1623                         }
1624                         if (err == EWOULDBLOCK) {
1625                                 if (fapi->timeout.tv_sec == -1) {
1626                                         continue;
1627                                 } else {
1628                                         return (ETIMEDOUT);
1629                                 }
1630                         } else if (err != 0) {
1631                                 return (err);
1632                         }
1633                 }
1634         }
1635
1636         pps->ppsinfo.current_mode = pps->ppsparam.mode;
1637         fapi->pps_info_buf = pps->ppsinfo;
1638
1639         return (0);
1640 }
1641
1642 int
1643 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1644 {
1645         pps_params_t *app;
1646         struct pps_fetch_args *fapi;
1647 #ifdef FFCLOCK
1648         struct pps_fetch_ffc_args *fapi_ffc;
1649 #endif
1650 #ifdef PPS_SYNC
1651         struct pps_kcbind_args *kapi;
1652 #endif
1653
1654         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1655         switch (cmd) {
1656         case PPS_IOC_CREATE:
1657                 return (0);
1658         case PPS_IOC_DESTROY:
1659                 return (0);
1660         case PPS_IOC_SETPARAMS:
1661                 app = (pps_params_t *)data;
1662                 if (app->mode & ~pps->ppscap)
1663                         return (EINVAL);
1664 #ifdef FFCLOCK
1665                 /* Ensure only a single clock is selected for ffc timestamp. */
1666                 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1667                         return (EINVAL);
1668 #endif
1669                 pps->ppsparam = *app;
1670                 return (0);
1671         case PPS_IOC_GETPARAMS:
1672                 app = (pps_params_t *)data;
1673                 *app = pps->ppsparam;
1674                 app->api_version = PPS_API_VERS_1;
1675                 return (0);
1676         case PPS_IOC_GETCAP:
1677                 *(int*)data = pps->ppscap;
1678                 return (0);
1679         case PPS_IOC_FETCH:
1680                 fapi = (struct pps_fetch_args *)data;
1681                 return (pps_fetch(fapi, pps));
1682 #ifdef FFCLOCK
1683         case PPS_IOC_FETCH_FFCOUNTER:
1684                 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1685                 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1686                     PPS_TSFMT_TSPEC)
1687                         return (EINVAL);
1688                 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1689                         return (EOPNOTSUPP);
1690                 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1691                 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1692                 /* Overwrite timestamps if feedback clock selected. */
1693                 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1694                 case PPS_TSCLK_FBCK:
1695                         fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1696                             pps->ppsinfo.assert_timestamp;
1697                         fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1698                             pps->ppsinfo.clear_timestamp;
1699                         break;
1700                 case PPS_TSCLK_FFWD:
1701                         break;
1702                 default:
1703                         break;
1704                 }
1705                 return (0);
1706 #endif /* FFCLOCK */
1707         case PPS_IOC_KCBIND:
1708 #ifdef PPS_SYNC
1709                 kapi = (struct pps_kcbind_args *)data;
1710                 /* XXX Only root should be able to do this */
1711                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1712                         return (EINVAL);
1713                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1714                         return (EINVAL);
1715                 if (kapi->edge & ~pps->ppscap)
1716                         return (EINVAL);
1717                 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1718                     (pps->kcmode & KCMODE_ABIFLAG);
1719                 return (0);
1720 #else
1721                 return (EOPNOTSUPP);
1722 #endif
1723         default:
1724                 return (ENOIOCTL);
1725         }
1726 }
1727
1728 void
1729 pps_init(struct pps_state *pps)
1730 {
1731         pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1732         if (pps->ppscap & PPS_CAPTUREASSERT)
1733                 pps->ppscap |= PPS_OFFSETASSERT;
1734         if (pps->ppscap & PPS_CAPTURECLEAR)
1735                 pps->ppscap |= PPS_OFFSETCLEAR;
1736 #ifdef FFCLOCK
1737         pps->ppscap |= PPS_TSCLK_MASK;
1738 #endif
1739         pps->kcmode &= ~KCMODE_ABIFLAG;
1740 }
1741
1742 void
1743 pps_init_abi(struct pps_state *pps)
1744 {
1745
1746         pps_init(pps);
1747         if (pps->driver_abi > 0) {
1748                 pps->kcmode |= KCMODE_ABIFLAG;
1749                 pps->kernel_abi = PPS_ABI_VERSION;
1750         }
1751 }
1752
1753 void
1754 pps_capture(struct pps_state *pps)
1755 {
1756         struct timehands *th;
1757
1758         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1759         th = timehands;
1760         pps->capgen = atomic_load_acq_int(&th->th_generation);
1761         pps->capth = th;
1762 #ifdef FFCLOCK
1763         pps->capffth = fftimehands;
1764 #endif
1765         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1766         atomic_thread_fence_acq();
1767         if (pps->capgen != th->th_generation)
1768                 pps->capgen = 0;
1769 }
1770
1771 void
1772 pps_event(struct pps_state *pps, int event)
1773 {
1774         struct bintime bt;
1775         struct timespec ts, *tsp, *osp;
1776         u_int tcount, *pcount;
1777         int foff;
1778         pps_seq_t *pseq;
1779 #ifdef FFCLOCK
1780         struct timespec *tsp_ffc;
1781         pps_seq_t *pseq_ffc;
1782         ffcounter *ffcount;
1783 #endif
1784 #ifdef PPS_SYNC
1785         int fhard;
1786 #endif
1787
1788         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1789         /* Nothing to do if not currently set to capture this event type. */
1790         if ((event & pps->ppsparam.mode) == 0)
1791                 return;
1792         /* If the timecounter was wound up underneath us, bail out. */
1793         if (pps->capgen == 0 || pps->capgen !=
1794             atomic_load_acq_int(&pps->capth->th_generation))
1795                 return;
1796
1797         /* Things would be easier with arrays. */
1798         if (event == PPS_CAPTUREASSERT) {
1799                 tsp = &pps->ppsinfo.assert_timestamp;
1800                 osp = &pps->ppsparam.assert_offset;
1801                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1802 #ifdef PPS_SYNC
1803                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1804 #endif
1805                 pcount = &pps->ppscount[0];
1806                 pseq = &pps->ppsinfo.assert_sequence;
1807 #ifdef FFCLOCK
1808                 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1809                 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1810                 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1811 #endif
1812         } else {
1813                 tsp = &pps->ppsinfo.clear_timestamp;
1814                 osp = &pps->ppsparam.clear_offset;
1815                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1816 #ifdef PPS_SYNC
1817                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1818 #endif
1819                 pcount = &pps->ppscount[1];
1820                 pseq = &pps->ppsinfo.clear_sequence;
1821 #ifdef FFCLOCK
1822                 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1823                 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1824                 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1825 #endif
1826         }
1827
1828         /*
1829          * If the timecounter changed, we cannot compare the count values, so
1830          * we have to drop the rest of the PPS-stuff until the next event.
1831          */
1832         if (pps->ppstc != pps->capth->th_counter) {
1833                 pps->ppstc = pps->capth->th_counter;
1834                 *pcount = pps->capcount;
1835                 pps->ppscount[2] = pps->capcount;
1836                 return;
1837         }
1838
1839         /* Convert the count to a timespec. */
1840         tcount = pps->capcount - pps->capth->th_offset_count;
1841         tcount &= pps->capth->th_counter->tc_counter_mask;
1842         bt = pps->capth->th_bintime;
1843         bintime_addx(&bt, pps->capth->th_scale * tcount);
1844         bintime2timespec(&bt, &ts);
1845
1846         /* If the timecounter was wound up underneath us, bail out. */
1847         atomic_thread_fence_acq();
1848         if (pps->capgen != pps->capth->th_generation)
1849                 return;
1850
1851         *pcount = pps->capcount;
1852         (*pseq)++;
1853         *tsp = ts;
1854
1855         if (foff) {
1856                 timespecadd(tsp, osp, tsp);
1857                 if (tsp->tv_nsec < 0) {
1858                         tsp->tv_nsec += 1000000000;
1859                         tsp->tv_sec -= 1;
1860                 }
1861         }
1862
1863 #ifdef FFCLOCK
1864         *ffcount = pps->capffth->tick_ffcount + tcount;
1865         bt = pps->capffth->tick_time;
1866         ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1867         bintime_add(&bt, &pps->capffth->tick_time);
1868         bintime2timespec(&bt, &ts);
1869         (*pseq_ffc)++;
1870         *tsp_ffc = ts;
1871 #endif
1872
1873 #ifdef PPS_SYNC
1874         if (fhard) {
1875                 uint64_t scale;
1876
1877                 /*
1878                  * Feed the NTP PLL/FLL.
1879                  * The FLL wants to know how many (hardware) nanoseconds
1880                  * elapsed since the previous event.
1881                  */
1882                 tcount = pps->capcount - pps->ppscount[2];
1883                 pps->ppscount[2] = pps->capcount;
1884                 tcount &= pps->capth->th_counter->tc_counter_mask;
1885                 scale = (uint64_t)1 << 63;
1886                 scale /= pps->capth->th_counter->tc_frequency;
1887                 scale *= 2;
1888                 bt.sec = 0;
1889                 bt.frac = 0;
1890                 bintime_addx(&bt, scale * tcount);
1891                 bintime2timespec(&bt, &ts);
1892                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1893         }
1894 #endif
1895
1896         /* Wakeup anyone sleeping in pps_fetch().  */
1897         wakeup(pps);
1898 }
1899
1900 /*
1901  * Timecounters need to be updated every so often to prevent the hardware
1902  * counter from overflowing.  Updating also recalculates the cached values
1903  * used by the get*() family of functions, so their precision depends on
1904  * the update frequency.
1905  */
1906
1907 static int tc_tick;
1908 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1909     "Approximate number of hardclock ticks in a millisecond");
1910
1911 void
1912 tc_ticktock(int cnt)
1913 {
1914         static int count;
1915
1916         if (mtx_trylock_spin(&tc_setclock_mtx)) {
1917                 count += cnt;
1918                 if (count >= tc_tick) {
1919                         count = 0;
1920                         tc_windup(NULL);
1921                 }
1922                 mtx_unlock_spin(&tc_setclock_mtx);
1923         }
1924 }
1925
1926 static void __inline
1927 tc_adjprecision(void)
1928 {
1929         int t;
1930
1931         if (tc_timepercentage > 0) {
1932                 t = (99 + tc_timepercentage) / tc_timepercentage;
1933                 tc_precexp = fls(t + (t >> 1)) - 1;
1934                 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1935                 FREQ2BT(hz, &bt_tickthreshold);
1936                 bintime_shift(&bt_timethreshold, tc_precexp);
1937                 bintime_shift(&bt_tickthreshold, tc_precexp);
1938         } else {
1939                 tc_precexp = 31;
1940                 bt_timethreshold.sec = INT_MAX;
1941                 bt_timethreshold.frac = ~(uint64_t)0;
1942                 bt_tickthreshold = bt_timethreshold;
1943         }
1944         sbt_timethreshold = bttosbt(bt_timethreshold);
1945         sbt_tickthreshold = bttosbt(bt_tickthreshold);
1946 }
1947
1948 static int
1949 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1950 {
1951         int error, val;
1952
1953         val = tc_timepercentage;
1954         error = sysctl_handle_int(oidp, &val, 0, req);
1955         if (error != 0 || req->newptr == NULL)
1956                 return (error);
1957         tc_timepercentage = val;
1958         if (cold)
1959                 goto done;
1960         tc_adjprecision();
1961 done:
1962         return (0);
1963 }
1964
1965 /* Set up the requested number of timehands. */
1966 static void
1967 inittimehands(void *dummy)
1968 {
1969         struct timehands *thp;
1970         int i;
1971
1972         TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1973             &timehands_count);
1974         if (timehands_count < 1)
1975                 timehands_count = 1;
1976         if (timehands_count > nitems(ths))
1977                 timehands_count = nitems(ths);
1978         for (i = 1, thp = &ths[0]; i < timehands_count;  thp = &ths[i++])
1979                 thp->th_next = &ths[i];
1980         thp->th_next = &ths[0];
1981
1982         TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1983             sizeof(tc_from_tunable));
1984
1985         mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
1986 }
1987 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1988
1989 static void
1990 inittimecounter(void *dummy)
1991 {
1992         u_int p;
1993         int tick_rate;
1994
1995         /*
1996          * Set the initial timeout to
1997          * max(1, <approx. number of hardclock ticks in a millisecond>).
1998          * People should probably not use the sysctl to set the timeout
1999          * to smaller than its initial value, since that value is the
2000          * smallest reasonable one.  If they want better timestamps they
2001          * should use the non-"get"* functions.
2002          */
2003         if (hz > 1000)
2004                 tc_tick = (hz + 500) / 1000;
2005         else
2006                 tc_tick = 1;
2007         tc_adjprecision();
2008         FREQ2BT(hz, &tick_bt);
2009         tick_sbt = bttosbt(tick_bt);
2010         tick_rate = hz / tc_tick;
2011         FREQ2BT(tick_rate, &tc_tick_bt);
2012         tc_tick_sbt = bttosbt(tc_tick_bt);
2013         p = (tc_tick * 1000000) / hz;
2014         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
2015
2016 #ifdef FFCLOCK
2017         ffclock_init();
2018 #endif
2019
2020         /* warm up new timecounter (again) and get rolling. */
2021         (void)timecounter->tc_get_timecount(timecounter);
2022         mtx_lock_spin(&tc_setclock_mtx);
2023         tc_windup(NULL);
2024         mtx_unlock_spin(&tc_setclock_mtx);
2025 }
2026
2027 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2028
2029 /* Cpu tick handling -------------------------------------------------*/
2030
2031 static int cpu_tick_variable;
2032 static uint64_t cpu_tick_frequency;
2033
2034 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2035 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2036
2037 static uint64_t
2038 tc_cpu_ticks(void)
2039 {
2040         struct timecounter *tc;
2041         uint64_t res, *base;
2042         unsigned u, *last;
2043
2044         critical_enter();
2045         base = DPCPU_PTR(tc_cpu_ticks_base);
2046         last = DPCPU_PTR(tc_cpu_ticks_last);
2047         tc = timehands->th_counter;
2048         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2049         if (u < *last)
2050                 *base += (uint64_t)tc->tc_counter_mask + 1;
2051         *last = u;
2052         res = u + *base;
2053         critical_exit();
2054         return (res);
2055 }
2056
2057 void
2058 cpu_tick_calibration(void)
2059 {
2060         static time_t last_calib;
2061
2062         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2063                 cpu_tick_calibrate(0);
2064                 last_calib = time_uptime;
2065         }
2066 }
2067
2068 /*
2069  * This function gets called every 16 seconds on only one designated
2070  * CPU in the system from hardclock() via cpu_tick_calibration()().
2071  *
2072  * Whenever the real time clock is stepped we get called with reset=1
2073  * to make sure we handle suspend/resume and similar events correctly.
2074  */
2075
2076 static void
2077 cpu_tick_calibrate(int reset)
2078 {
2079         static uint64_t c_last;
2080         uint64_t c_this, c_delta;
2081         static struct bintime  t_last;
2082         struct bintime t_this, t_delta;
2083         uint32_t divi;
2084
2085         if (reset) {
2086                 /* The clock was stepped, abort & reset */
2087                 t_last.sec = 0;
2088                 return;
2089         }
2090
2091         /* we don't calibrate fixed rate cputicks */
2092         if (!cpu_tick_variable)
2093                 return;
2094
2095         getbinuptime(&t_this);
2096         c_this = cpu_ticks();
2097         if (t_last.sec != 0) {
2098                 c_delta = c_this - c_last;
2099                 t_delta = t_this;
2100                 bintime_sub(&t_delta, &t_last);
2101                 /*
2102                  * Headroom:
2103                  *      2^(64-20) / 16[s] =
2104                  *      2^(44) / 16[s] =
2105                  *      17.592.186.044.416 / 16 =
2106                  *      1.099.511.627.776 [Hz]
2107                  */
2108                 divi = t_delta.sec << 20;
2109                 divi |= t_delta.frac >> (64 - 20);
2110                 c_delta <<= 20;
2111                 c_delta /= divi;
2112                 if (c_delta > cpu_tick_frequency) {
2113                         if (0 && bootverbose)
2114                                 printf("cpu_tick increased to %ju Hz\n",
2115                                     c_delta);
2116                         cpu_tick_frequency = c_delta;
2117                 }
2118         }
2119         c_last = c_this;
2120         t_last = t_this;
2121 }
2122
2123 void
2124 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2125 {
2126
2127         if (func == NULL) {
2128                 cpu_ticks = tc_cpu_ticks;
2129         } else {
2130                 cpu_tick_frequency = freq;
2131                 cpu_tick_variable = var;
2132                 cpu_ticks = func;
2133         }
2134 }
2135
2136 uint64_t
2137 cpu_tickrate(void)
2138 {
2139
2140         if (cpu_ticks == tc_cpu_ticks)
2141                 return (tc_getfrequency());
2142         return (cpu_tick_frequency);
2143 }
2144
2145 /*
2146  * We need to be slightly careful converting cputicks to microseconds.
2147  * There is plenty of margin in 64 bits of microseconds (half a million
2148  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2149  * before divide conversion (to retain precision) we find that the
2150  * margin shrinks to 1.5 hours (one millionth of 146y).
2151  * With a three prong approach we never lose significant bits, no
2152  * matter what the cputick rate and length of timeinterval is.
2153  */
2154
2155 uint64_t
2156 cputick2usec(uint64_t tick)
2157 {
2158
2159         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
2160                 return (tick / (cpu_tickrate() / 1000000LL));
2161         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
2162                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2163         else
2164                 return ((tick * 1000000LL) / cpu_tickrate());
2165 }
2166
2167 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
2168
2169 static int vdso_th_enable = 1;
2170 static int
2171 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2172 {
2173         int old_vdso_th_enable, error;
2174
2175         old_vdso_th_enable = vdso_th_enable;
2176         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2177         if (error != 0)
2178                 return (error);
2179         vdso_th_enable = old_vdso_th_enable;
2180         return (0);
2181 }
2182 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2183     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2184     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2185
2186 uint32_t
2187 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2188 {
2189         struct timehands *th;
2190         uint32_t enabled;
2191
2192         th = timehands;
2193         vdso_th->th_scale = th->th_scale;
2194         vdso_th->th_offset_count = th->th_offset_count;
2195         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2196         vdso_th->th_offset = th->th_offset;
2197         vdso_th->th_boottime = th->th_boottime;
2198         if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2199                 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2200                     th->th_counter);
2201         } else
2202                 enabled = 0;
2203         if (!vdso_th_enable)
2204                 enabled = 0;
2205         return (enabled);
2206 }
2207
2208 #ifdef COMPAT_FREEBSD32
2209 uint32_t
2210 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2211 {
2212         struct timehands *th;
2213         uint32_t enabled;
2214
2215         th = timehands;
2216         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2217         vdso_th32->th_offset_count = th->th_offset_count;
2218         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2219         vdso_th32->th_offset.sec = th->th_offset.sec;
2220         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2221         vdso_th32->th_boottime.sec = th->th_boottime.sec;
2222         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2223         if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2224                 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2225                     th->th_counter);
2226         } else
2227                 enabled = 0;
2228         if (!vdso_th_enable)
2229                 enabled = 0;
2230         return (enabled);
2231 }
2232 #endif
2233
2234 #include "opt_ddb.h"
2235 #ifdef DDB
2236 #include <ddb/ddb.h>
2237
2238 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2239 {
2240         struct timehands *th;
2241         struct timecounter *tc;
2242         u_int val1, val2;
2243
2244         th = timehands;
2245         tc = th->th_counter;
2246         val1 = tc->tc_get_timecount(tc);
2247         __compiler_membar();
2248         val2 = tc->tc_get_timecount(tc);
2249
2250         db_printf("timecounter %p %s\n", tc, tc->tc_name);
2251         db_printf("  mask %#x freq %ju qual %d flags %#x priv %p\n",
2252             tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2253             tc->tc_flags, tc->tc_priv);
2254         db_printf("  val %#x %#x\n", val1, val2);
2255         db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2256             (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2257             th->th_large_delta, th->th_offset_count, th->th_generation);
2258         db_printf("  offset %jd %jd boottime %jd %jd\n",
2259             (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2260             (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);
2261 }
2262 #endif