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