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 /*
 109  * The system time is always computed by summing the estimated boot time and the
 110  * system uptime. The timehands track boot time, but it changes when the system
 111  * time is set by the user, stepped by ntpd or adjusted when resuming. It
 112  * is set to new_time - uptime.
 113  */
 114 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
 115 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
 116     CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 117     sysctl_kern_boottime, "S,timeval",
 118     "Estimated system boottime");
 119
 120 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 121     "");
 122 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
 123     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 124     "");
 125
 126 static int timestepwarnings;
 127 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
 128     &timestepwarnings, 0, "Log time steps");
 129
 130 static int timehands_count = 2;
 131 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
 132     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 133     &timehands_count, 0, "Count of timehands in rotation");
 134
 135 struct bintime bt_timethreshold;
 136 struct bintime bt_tickthreshold;
 137 sbintime_t sbt_timethreshold;
 138 sbintime_t sbt_tickthreshold;
 139 struct bintime tc_tick_bt;
 140 sbintime_t tc_tick_sbt;
 141 int tc_precexp;
 142 int tc_timepercentage = TC_DEFAULTPERC;
 143 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
 144 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
 145     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
 146     sysctl_kern_timecounter_adjprecision, "I",
 147     "Allowed time interval deviation in percents");
 148
 149 volatile int rtc_generation = 1;
 150
 151 static int tc_chosen;   /* Non-zero if a specific tc was chosen via sysctl. */
 152 static char tc_from_tunable[16];
 153
 154 static void tc_windup(struct bintime *new_boottimebin);
 155 static void cpu_tick_calibrate(int);
 156
 157 void dtrace_getnanotime(struct timespec *tsp);
 158 void dtrace_getnanouptime(struct timespec *tsp);
 159
 160 static int
 161 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 162 {
 163         struct timeval boottime;
 164
 165         getboottime(&boottime);
 166
 167 /* i386 is the only arch which uses a 32bits time_t */
 168 #ifdef __amd64__
 169 #ifdef SCTL_MASK32
 170         int tv[2];
 171
 172         if (req->flags & SCTL_MASK32) {
 173                 tv[0] = boottime.tv_sec;
 174                 tv[1] = boottime.tv_usec;
 175                 return (SYSCTL_OUT(req, tv, sizeof(tv)));
 176         }
 177 #endif
 178 #endif
 179         return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
 180 }
 181
 182 static int
 183 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 184 {
 185         u_int ncount;
 186         struct timecounter *tc = arg1;
 187
 188         ncount = tc->tc_get_timecount(tc);
 189         return (sysctl_handle_int(oidp, &ncount, 0, req));
 190 }
 191
 192 static int
 193 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 194 {
 195         uint64_t freq;
 196         struct timecounter *tc = arg1;
 197
 198         freq = tc->tc_frequency;
 199         return (sysctl_handle_64(oidp, &freq, 0, req));
 200 }
 201
 202 /*
 203  * Return the difference between the timehands' counter value now and what
 204  * was when we copied it to the timehands' offset_count.
 205  */
 206 static __inline u_int
 207 tc_delta(struct timehands *th)
 208 {
 209         struct timecounter *tc;
 210
 211         tc = th->th_counter;
 212         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 213             tc->tc_counter_mask);
 214 }
 215
 216 static __inline void
 217 bintime_add_tc_delta(struct bintime *bt, uint64_t scale,
 218     uint64_t large_delta, uint64_t delta)
 219 {
 220         uint64_t x;
 221
 222         if (__predict_false(delta >= large_delta)) {
 223                 /* Avoid overflow for scale * delta. */
 224                 x = (scale >> 32) * delta;
 225                 bt->sec += x >> 32;
 226                 bintime_addx(bt, x << 32);
 227                 bintime_addx(bt, (scale & 0xffffffff) * delta);
 228         } else {
 229                 bintime_addx(bt, scale * delta);
 230         }
 231 }
 232
 233 /*
 234  * Functions for reading the time.  We have to loop until we are sure that
 235  * the timehands that we operated on was not updated under our feet.  See
 236  * the comment in <sys/time.h> for a description of these 12 functions.
 237  */
 238
 239 static __inline void
 240 bintime_off(struct bintime *bt, u_int off)
 241 {
 242         struct timehands *th;
 243         struct bintime *btp;
 244         uint64_t scale;
 245         u_int delta, gen, large_delta;
 246
 247         do {
 248                 th = timehands;
 249                 gen = atomic_load_acq_int(&th->th_generation);
 250                 btp = (struct bintime *)((vm_offset_t)th + off);
 251                 *bt = *btp;
 252                 scale = th->th_scale;
 253                 delta = tc_delta(th);
 254                 large_delta = th->th_large_delta;
 255                 atomic_thread_fence_acq();
 256         } while (gen == 0 || gen != th->th_generation);
 257
 258         bintime_add_tc_delta(bt, scale, large_delta, delta);
 259 }
 260 #define GETTHBINTIME(dst, member)                                       \
 261 do {                                                                    \
 262         _Static_assert(_Generic(((struct timehands *)NULL)->member,     \
 263             struct bintime: 1, default: 0) == 1,                        \
 264             "struct timehands member is not of struct bintime type");   \
 265         bintime_off(dst, __offsetof(struct timehands, member));         \
 266 } while (0)
 267
 268 static __inline void
 269 getthmember(void *out, size_t out_size, u_int off)
 270 {
 271         struct timehands *th;
 272         u_int gen;
 273
 274         do {
 275                 th = timehands;
 276                 gen = atomic_load_acq_int(&th->th_generation);
 277                 memcpy(out, (char *)th + off, out_size);
 278                 atomic_thread_fence_acq();
 279         } while (gen == 0 || gen != th->th_generation);
 280 }
 281 #define GETTHMEMBER(dst, member)                                        \
 282 do {                                                                    \
 283         _Static_assert(_Generic(*dst,                                   \
 284             __typeof(((struct timehands *)NULL)->member): 1,            \
 285             default: 0) == 1,                                           \
 286             "*dst and struct timehands member have different types");   \
 287         getthmember(dst, sizeof(*dst), __offsetof(struct timehands,     \
 288             member));                                                   \
 289 } while (0)
 290
 291 #ifdef FFCLOCK
 292 void
 293 fbclock_binuptime(struct bintime *bt)
 294 {
 295
 296         GETTHBINTIME(bt, th_offset);
 297 }
 298
 299 void
 300 fbclock_nanouptime(struct timespec *tsp)
 301 {
 302         struct bintime bt;
 303
 304         fbclock_binuptime(&bt);
 305         bintime2timespec(&bt, tsp);
 306 }
 307
 308 void
 309 fbclock_microuptime(struct timeval *tvp)
 310 {
 311         struct bintime bt;
 312
 313         fbclock_binuptime(&bt);
 314         bintime2timeval(&bt, tvp);
 315 }
 316
 317 void
 318 fbclock_bintime(struct bintime *bt)
 319 {
 320
 321         GETTHBINTIME(bt, th_bintime);
 322 }
 323
 324 void
 325 fbclock_nanotime(struct timespec *tsp)
 326 {
 327         struct bintime bt;
 328
 329         fbclock_bintime(&bt);
 330         bintime2timespec(&bt, tsp);
 331 }
 332
 333 void
 334 fbclock_microtime(struct timeval *tvp)
 335 {
 336         struct bintime bt;
 337
 338         fbclock_bintime(&bt);
 339         bintime2timeval(&bt, tvp);
 340 }
 341
 342 void
 343 fbclock_getbinuptime(struct bintime *bt)
 344 {
 345
 346         GETTHMEMBER(bt, th_offset);
 347 }
 348
 349 void
 350 fbclock_getnanouptime(struct timespec *tsp)
 351 {
 352         struct bintime bt;
 353
 354         GETTHMEMBER(&bt, th_offset);
 355         bintime2timespec(&bt, tsp);
 356 }
 357
 358 void
 359 fbclock_getmicrouptime(struct timeval *tvp)
 360 {
 361         struct bintime bt;
 362
 363         GETTHMEMBER(&bt, th_offset);
 364         bintime2timeval(&bt, tvp);
 365 }
 366
 367 void
 368 fbclock_getbintime(struct bintime *bt)
 369 {
 370
 371         GETTHMEMBER(bt, th_bintime);
 372 }
 373
 374 void
 375 fbclock_getnanotime(struct timespec *tsp)
 376 {
 377
 378         GETTHMEMBER(tsp, th_nanotime);
 379 }
 380
 381 void
 382 fbclock_getmicrotime(struct timeval *tvp)
 383 {
 384
 385         GETTHMEMBER(tvp, th_microtime);
 386 }
 387 #else /* !FFCLOCK */
 388
 389 void
 390 binuptime(struct bintime *bt)
 391 {
 392
 393         GETTHBINTIME(bt, th_offset);
 394 }
 395
 396 void
 397 nanouptime(struct timespec *tsp)
 398 {
 399         struct bintime bt;
 400
 401         binuptime(&bt);
 402         bintime2timespec(&bt, tsp);
 403 }
 404
 405 void
 406 microuptime(struct timeval *tvp)
 407 {
 408         struct bintime bt;
 409
 410         binuptime(&bt);
 411         bintime2timeval(&bt, tvp);
 412 }
 413
 414 void
 415 bintime(struct bintime *bt)
 416 {
 417
 418         GETTHBINTIME(bt, th_bintime);
 419 }
 420
 421 void
 422 nanotime(struct timespec *tsp)
 423 {
 424         struct bintime bt;
 425
 426         bintime(&bt);
 427         bintime2timespec(&bt, tsp);
 428 }
 429
 430 void
 431 microtime(struct timeval *tvp)
 432 {
 433         struct bintime bt;
 434
 435         bintime(&bt);
 436         bintime2timeval(&bt, tvp);
 437 }
 438
 439 void
 440 getbinuptime(struct bintime *bt)
 441 {
 442
 443         GETTHMEMBER(bt, th_offset);
 444 }
 445
 446 void
 447 getnanouptime(struct timespec *tsp)
 448 {
 449         struct bintime bt;
 450
 451         GETTHMEMBER(&bt, th_offset);
 452         bintime2timespec(&bt, tsp);
 453 }
 454
 455 void
 456 getmicrouptime(struct timeval *tvp)
 457 {
 458         struct bintime bt;
 459
 460         GETTHMEMBER(&bt, th_offset);
 461         bintime2timeval(&bt, tvp);
 462 }
 463
 464 void
 465 getbintime(struct bintime *bt)
 466 {
 467
 468         GETTHMEMBER(bt, th_bintime);
 469 }
 470
 471 void
 472 getnanotime(struct timespec *tsp)
 473 {
 474
 475         GETTHMEMBER(tsp, th_nanotime);
 476 }
 477
 478 void
 479 getmicrotime(struct timeval *tvp)
 480 {
 481
 482         GETTHMEMBER(tvp, th_microtime);
 483 }
 484 #endif /* FFCLOCK */
 485
 486 void
 487 getboottime(struct timeval *boottime)
 488 {
 489         struct bintime boottimebin;
 490
 491         getboottimebin(&boottimebin);
 492         bintime2timeval(&boottimebin, boottime);
 493 }
 494
 495 void
 496 getboottimebin(struct bintime *boottimebin)
 497 {
 498
 499         GETTHMEMBER(boottimebin, th_boottime);
 500 }
 501
 502 #ifdef FFCLOCK
 503 /*
 504  * Support for feed-forward synchronization algorithms. This is heavily inspired
 505  * by the timehands mechanism but kept independent from it. *_windup() functions
 506  * have some connection to avoid accessing the timecounter hardware more than
 507  * necessary.
 508  */
 509
 510 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
 511 struct ffclock_estimate ffclock_estimate;
 512 struct bintime ffclock_boottime;        /* Feed-forward boot time estimate. */
 513 uint32_t ffclock_status;                /* Feed-forward clock status. */
 514 int8_t ffclock_updated;                 /* New estimates are available. */
 515 struct mtx ffclock_mtx;                 /* Mutex on ffclock_estimate. */
 516
 517 struct fftimehands {
 518         struct ffclock_estimate cest;
 519         struct bintime          tick_time;
 520         struct bintime          tick_time_lerp;
 521         ffcounter               tick_ffcount;
 522         uint64_t                period_lerp;
 523         volatile uint8_t        gen;
 524         struct fftimehands      *next;
 525 };
 526
 527 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
 528
 529 static struct fftimehands ffth[10];
 530 static struct fftimehands *volatile fftimehands = ffth;
 531
 532 static void
 533 ffclock_init(void)
 534 {
 535         struct fftimehands *cur;
 536         struct fftimehands *last;
 537
 538         memset(ffth, 0, sizeof(ffth));
 539
 540         last = ffth + NUM_ELEMENTS(ffth) - 1;
 541         for (cur = ffth; cur < last; cur++)
 542                 cur->next = cur + 1;
 543         last->next = ffth;
 544
 545         ffclock_updated = 0;
 546         ffclock_status = FFCLOCK_STA_UNSYNC;
 547         mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
 548 }
 549
 550 /*
 551  * Reset the feed-forward clock estimates. Called from inittodr() to get things
 552  * kick started and uses the timecounter nominal frequency as a first period
 553  * estimate. Note: this function may be called several time just after boot.
 554  * Note: this is the only function that sets the value of boot time for the
 555  * monotonic (i.e. uptime) version of the feed-forward clock.
 556  */
 557 void
 558 ffclock_reset_clock(struct timespec *ts)
 559 {
 560         struct timecounter *tc;
 561         struct ffclock_estimate cest;
 562
 563         tc = timehands->th_counter;
 564         memset(&cest, 0, sizeof(struct ffclock_estimate));
 565
 566         timespec2bintime(ts, &ffclock_boottime);
 567         timespec2bintime(ts, &(cest.update_time));
 568         ffclock_read_counter(&cest.update_ffcount);
 569         cest.leapsec_next = 0;
 570         cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
 571         cest.errb_abs = 0;
 572         cest.errb_rate = 0;
 573         cest.status = FFCLOCK_STA_UNSYNC;
 574         cest.leapsec_total = 0;
 575         cest.leapsec = 0;
 576
 577         mtx_lock(&ffclock_mtx);
 578         bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
 579         ffclock_updated = INT8_MAX;
 580         mtx_unlock(&ffclock_mtx);
 581
 582         printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
 583             (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
 584             (unsigned long)ts->tv_nsec);
 585 }
 586
 587 /*
 588  * Sub-routine to convert a time interval measured in RAW counter units to time
 589  * in seconds stored in bintime format.
 590  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
 591  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
 592  * extra cycles.
 593  */
 594 static void
 595 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
 596 {
 597         struct bintime bt2;
 598         ffcounter delta, delta_max;
 599
 600         delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
 601         bintime_clear(bt);
 602         do {
 603                 if (ffdelta > delta_max)
 604                         delta = delta_max;
 605                 else
 606                         delta = ffdelta;
 607                 bt2.sec = 0;
 608                 bt2.frac = period;
 609                 bintime_mul(&bt2, (unsigned int)delta);
 610                 bintime_add(bt, &bt2);
 611                 ffdelta -= delta;
 612         } while (ffdelta > 0);
 613 }
 614
 615 /*
 616  * Update the fftimehands.
 617  * Push the tick ffcount and time(s) forward based on current clock estimate.
 618  * The conversion from ffcounter to bintime relies on the difference clock
 619  * principle, whose accuracy relies on computing small time intervals. If a new
 620  * clock estimate has been passed by the synchronisation daemon, make it
 621  * current, and compute the linear interpolation for monotonic time if needed.
 622  */
 623 static void
 624 ffclock_windup(unsigned int delta)
 625 {
 626         struct ffclock_estimate *cest;
 627         struct fftimehands *ffth;
 628         struct bintime bt, gap_lerp;
 629         ffcounter ffdelta;
 630         uint64_t frac;
 631         unsigned int polling;
 632         uint8_t forward_jump, ogen;
 633
 634         /*
 635          * Pick the next timehand, copy current ffclock estimates and move tick
 636          * times and counter forward.
 637          */
 638         forward_jump = 0;
 639         ffth = fftimehands->next;
 640         ogen = ffth->gen;
 641         ffth->gen = 0;
 642         cest = &ffth->cest;
 643         bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
 644         ffdelta = (ffcounter)delta;
 645         ffth->period_lerp = fftimehands->period_lerp;
 646
 647         ffth->tick_time = fftimehands->tick_time;
 648         ffclock_convert_delta(ffdelta, cest->period, &bt);
 649         bintime_add(&ffth->tick_time, &bt);
 650
 651         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
 652         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
 653         bintime_add(&ffth->tick_time_lerp, &bt);
 654
 655         ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
 656
 657         /*
 658          * Assess the status of the clock, if the last update is too old, it is
 659          * likely the synchronisation daemon is dead and the clock is free
 660          * running.
 661          */
 662         if (ffclock_updated == 0) {
 663                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
 664                 ffclock_convert_delta(ffdelta, cest->period, &bt);
 665                 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
 666                         ffclock_status |= FFCLOCK_STA_UNSYNC;
 667         }
 668
 669         /*
 670          * If available, grab updated clock estimates and make them current.
 671          * Recompute time at this tick using the updated estimates. The clock
 672          * estimates passed the feed-forward synchronisation daemon may result
 673          * in time conversion that is not monotonically increasing (just after
 674          * the update). time_lerp is a particular linear interpolation over the
 675          * synchronisation algo polling period that ensures monotonicity for the
 676          * clock ids requesting it.
 677          */
 678         if (ffclock_updated > 0) {
 679                 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
 680                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
 681                 ffth->tick_time = cest->update_time;
 682                 ffclock_convert_delta(ffdelta, cest->period, &bt);
 683                 bintime_add(&ffth->tick_time, &bt);
 684
 685                 /* ffclock_reset sets ffclock_updated to INT8_MAX */
 686                 if (ffclock_updated == INT8_MAX)
 687                         ffth->tick_time_lerp = ffth->tick_time;
 688
 689                 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
 690                         forward_jump = 1;
 691                 else
 692                         forward_jump = 0;
 693
 694                 bintime_clear(&gap_lerp);
 695                 if (forward_jump) {
 696                         gap_lerp = ffth->tick_time;
 697                         bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
 698                 } else {
 699                         gap_lerp = ffth->tick_time_lerp;
 700                         bintime_sub(&gap_lerp, &ffth->tick_time);
 701                 }
 702
 703                 /*
 704                  * The reset from the RTC clock may be far from accurate, and
 705                  * reducing the gap between real time and interpolated time
 706                  * could take a very long time if the interpolated clock insists
 707                  * on strict monotonicity. The clock is reset under very strict
 708                  * conditions (kernel time is known to be wrong and
 709                  * synchronization daemon has been restarted recently.
 710                  * ffclock_boottime absorbs the jump to ensure boot time is
 711                  * correct and uptime functions stay consistent.
 712                  */
 713                 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
 714                     ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
 715                     ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
 716                         if (forward_jump)
 717                                 bintime_add(&ffclock_boottime, &gap_lerp);
 718                         else
 719                                 bintime_sub(&ffclock_boottime, &gap_lerp);
 720                         ffth->tick_time_lerp = ffth->tick_time;
 721                         bintime_clear(&gap_lerp);
 722                 }
 723
 724                 ffclock_status = cest->status;
 725                 ffth->period_lerp = cest->period;
 726
 727                 /*
 728                  * Compute corrected period used for the linear interpolation of
 729                  * time. The rate of linear interpolation is capped to 5000PPM
 730                  * (5ms/s).
 731                  */
 732                 if (bintime_isset(&gap_lerp)) {
 733                         ffdelta = cest->update_ffcount;
 734                         ffdelta -= fftimehands->cest.update_ffcount;
 735                         ffclock_convert_delta(ffdelta, cest->period, &bt);
 736                         polling = bt.sec;
 737                         bt.sec = 0;
 738                         bt.frac = 5000000 * (uint64_t)18446744073LL;
 739                         bintime_mul(&bt, polling);
 740                         if (bintime_cmp(&gap_lerp, &bt, >))
 741                                 gap_lerp = bt;
 742
 743                         /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
 744                         frac = 0;
 745                         if (gap_lerp.sec > 0) {
 746                                 frac -= 1;
 747                                 frac /= ffdelta / gap_lerp.sec;
 748                         }
 749                         frac += gap_lerp.frac / ffdelta;
 750
 751                         if (forward_jump)
 752                                 ffth->period_lerp += frac;
 753                         else
 754                                 ffth->period_lerp -= frac;
 755                 }
 756
 757                 ffclock_updated = 0;
 758         }
 759         if (++ogen == 0)
 760                 ogen = 1;
 761         ffth->gen = ogen;
 762         fftimehands = ffth;
 763 }
 764
 765 /*
 766  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
 767  * the old and new hardware counter cannot be read simultaneously. tc_windup()
 768  * does read the two counters 'back to back', but a few cycles are effectively
 769  * lost, and not accumulated in tick_ffcount. This is a fairly radical
 770  * operation for a feed-forward synchronization daemon, and it is its job to not
 771  * pushing irrelevant data to the kernel. Because there is no locking here,
 772  * simply force to ignore pending or next update to give daemon a chance to
 773  * realize the counter has changed.
 774  */
 775 static void
 776 ffclock_change_tc(struct timehands *th)
 777 {
 778         struct fftimehands *ffth;
 779         struct ffclock_estimate *cest;
 780         struct timecounter *tc;
 781         uint8_t ogen;
 782
 783         tc = th->th_counter;
 784         ffth = fftimehands->next;
 785         ogen = ffth->gen;
 786         ffth->gen = 0;
 787
 788         cest = &ffth->cest;
 789         bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
 790         cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
 791         cest->errb_abs = 0;
 792         cest->errb_rate = 0;
 793         cest->status |= FFCLOCK_STA_UNSYNC;
 794
 795         ffth->tick_ffcount = fftimehands->tick_ffcount;
 796         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
 797         ffth->tick_time = fftimehands->tick_time;
 798         ffth->period_lerp = cest->period;
 799
 800         /* Do not lock but ignore next update from synchronization daemon. */
 801         ffclock_updated--;
 802
 803         if (++ogen == 0)
 804                 ogen = 1;
 805         ffth->gen = ogen;
 806         fftimehands = ffth;
 807 }
 808
 809 /*
 810  * Retrieve feed-forward counter and time of last kernel tick.
 811  */
 812 void
 813 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
 814 {
 815         struct fftimehands *ffth;
 816         uint8_t gen;
 817
 818         /*
 819          * No locking but check generation has not changed. Also need to make
 820          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
 821          */
 822         do {
 823                 ffth = fftimehands;
 824                 gen = ffth->gen;
 825                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
 826                         *bt = ffth->tick_time_lerp;
 827                 else
 828                         *bt = ffth->tick_time;
 829                 *ffcount = ffth->tick_ffcount;
 830         } while (gen == 0 || gen != ffth->gen);
 831 }
 832
 833 /*
 834  * Absolute clock conversion. Low level function to convert ffcounter to
 835  * bintime. The ffcounter is converted using the current ffclock period estimate
 836  * or the "interpolated period" to ensure monotonicity.
 837  * NOTE: this conversion may have been deferred, and the clock updated since the
 838  * hardware counter has been read.
 839  */
 840 void
 841 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
 842 {
 843         struct fftimehands *ffth;
 844         struct bintime bt2;
 845         ffcounter ffdelta;
 846         uint8_t gen;
 847
 848         /*
 849          * No locking but check generation has not changed. Also need to make
 850          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
 851          */
 852         do {
 853                 ffth = fftimehands;
 854                 gen = ffth->gen;
 855                 if (ffcount > ffth->tick_ffcount)
 856                         ffdelta = ffcount - ffth->tick_ffcount;
 857                 else
 858                         ffdelta = ffth->tick_ffcount - ffcount;
 859
 860                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
 861                         *bt = ffth->tick_time_lerp;
 862                         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
 863                 } else {
 864                         *bt = ffth->tick_time;
 865                         ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
 866                 }
 867
 868                 if (ffcount > ffth->tick_ffcount)
 869                         bintime_add(bt, &bt2);
 870                 else
 871                         bintime_sub(bt, &bt2);
 872         } while (gen == 0 || gen != ffth->gen);
 873 }
 874
 875 /*
 876  * Difference clock conversion.
 877  * Low level function to Convert a time interval measured in RAW counter units
 878  * into bintime. The difference clock allows measuring small intervals much more
 879  * reliably than the absolute clock.
 880  */
 881 void
 882 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
 883 {
 884         struct fftimehands *ffth;
 885         uint8_t gen;
 886
 887         /* No locking but check generation has not changed. */
 888         do {
 889                 ffth = fftimehands;
 890                 gen = ffth->gen;
 891                 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
 892         } while (gen == 0 || gen != ffth->gen);
 893 }
 894
 895 /*
 896  * Access to current ffcounter value.
 897  */
 898 void
 899 ffclock_read_counter(ffcounter *ffcount)
 900 {
 901         struct timehands *th;
 902         struct fftimehands *ffth;
 903         unsigned int gen, delta;
 904
 905         /*
 906          * ffclock_windup() called from tc_windup(), safe to rely on
 907          * th->th_generation only, for correct delta and ffcounter.
 908          */
 909         do {
 910                 th = timehands;
 911                 gen = atomic_load_acq_int(&th->th_generation);
 912                 ffth = fftimehands;
 913                 delta = tc_delta(th);
 914                 *ffcount = ffth->tick_ffcount;
 915                 atomic_thread_fence_acq();
 916         } while (gen == 0 || gen != th->th_generation);
 917
 918         *ffcount += delta;
 919 }
 920
 921 void
 922 binuptime(struct bintime *bt)
 923 {
 924
 925         binuptime_fromclock(bt, sysclock_active);
 926 }
 927
 928 void
 929 nanouptime(struct timespec *tsp)
 930 {
 931
 932         nanouptime_fromclock(tsp, sysclock_active);
 933 }
 934
 935 void
 936 microuptime(struct timeval *tvp)
 937 {
 938
 939         microuptime_fromclock(tvp, sysclock_active);
 940 }
 941
 942 void
 943 bintime(struct bintime *bt)
 944 {
 945
 946         bintime_fromclock(bt, sysclock_active);
 947 }
 948
 949 void
 950 nanotime(struct timespec *tsp)
 951 {
 952
 953         nanotime_fromclock(tsp, sysclock_active);
 954 }
 955
 956 void
 957 microtime(struct timeval *tvp)
 958 {
 959
 960         microtime_fromclock(tvp, sysclock_active);
 961 }
 962
 963 void
 964 getbinuptime(struct bintime *bt)
 965 {
 966
 967         getbinuptime_fromclock(bt, sysclock_active);
 968 }
 969
 970 void
 971 getnanouptime(struct timespec *tsp)
 972 {
 973
 974         getnanouptime_fromclock(tsp, sysclock_active);
 975 }
 976
 977 void
 978 getmicrouptime(struct timeval *tvp)
 979 {
 980
 981         getmicrouptime_fromclock(tvp, sysclock_active);
 982 }
 983
 984 void
 985 getbintime(struct bintime *bt)
 986 {
 987
 988         getbintime_fromclock(bt, sysclock_active);
 989 }
 990
 991 void
 992 getnanotime(struct timespec *tsp)
 993 {
 994
 995         getnanotime_fromclock(tsp, sysclock_active);
 996 }
 997
 998 void
 999 getmicrotime(struct timeval *tvp)
1000 {
1001
1002         getmicrouptime_fromclock(tvp, sysclock_active);
1003 }
1004
1005 #endif /* FFCLOCK */
1006
1007 /*
1008  * This is a clone of getnanotime and used for walltimestamps.
1009  * The dtrace_ prefix prevents fbt from creating probes for
1010  * it so walltimestamp can be safely used in all fbt probes.
1011  */
1012 void
1013 dtrace_getnanotime(struct timespec *tsp)
1014 {
1015
1016         GETTHMEMBER(tsp, th_nanotime);
1017 }
1018
1019 /*
1020  * This is a clone of getnanouptime used for time since boot.
1021  * The dtrace_ prefix prevents fbt from creating probes for
1022  * it so an uptime that can be safely used in all fbt probes.
1023  */
1024 void
1025 dtrace_getnanouptime(struct timespec *tsp)
1026 {
1027         struct bintime bt;
1028
1029         GETTHMEMBER(&bt, th_offset);
1030         bintime2timespec(&bt, tsp);
1031 }
1032
1033 /*
1034  * System clock currently providing time to the system. Modifiable via sysctl
1035  * when the FFCLOCK option is defined.
1036  */
1037 int sysclock_active = SYSCLOCK_FBCK;
1038
1039 /* Internal NTP status and error estimates. */
1040 extern int time_status;
1041 extern long time_esterror;
1042
1043 /*
1044  * Take a snapshot of sysclock data which can be used to compare system clocks
1045  * and generate timestamps after the fact.
1046  */
1047 void
1048 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1049 {
1050         struct fbclock_info *fbi;
1051         struct timehands *th;
1052         struct bintime bt;
1053         unsigned int delta, gen;
1054 #ifdef FFCLOCK
1055         ffcounter ffcount;
1056         struct fftimehands *ffth;
1057         struct ffclock_info *ffi;
1058         struct ffclock_estimate cest;
1059
1060         ffi = &clock_snap->ff_info;
1061 #endif
1062
1063         fbi = &clock_snap->fb_info;
1064         delta = 0;
1065
1066         do {
1067                 th = timehands;
1068                 gen = atomic_load_acq_int(&th->th_generation);
1069                 fbi->th_scale = th->th_scale;
1070                 fbi->tick_time = th->th_offset;
1071 #ifdef FFCLOCK
1072                 ffth = fftimehands;
1073                 ffi->tick_time = ffth->tick_time_lerp;
1074                 ffi->tick_time_lerp = ffth->tick_time_lerp;
1075                 ffi->period = ffth->cest.period;
1076                 ffi->period_lerp = ffth->period_lerp;
1077                 clock_snap->ffcount = ffth->tick_ffcount;
1078                 cest = ffth->cest;
1079 #endif
1080                 if (!fast)
1081                         delta = tc_delta(th);
1082                 atomic_thread_fence_acq();
1083         } while (gen == 0 || gen != th->th_generation);
1084
1085         clock_snap->delta = delta;
1086         clock_snap->sysclock_active = sysclock_active;
1087
1088         /* Record feedback clock status and error. */
1089         clock_snap->fb_info.status = time_status;
1090         /* XXX: Very crude estimate of feedback clock error. */
1091         bt.sec = time_esterror / 1000000;
1092         bt.frac = ((time_esterror - bt.sec) * 1000000) *
1093             (uint64_t)18446744073709ULL;
1094         clock_snap->fb_info.error = bt;
1095
1096 #ifdef FFCLOCK
1097         if (!fast)
1098                 clock_snap->ffcount += delta;
1099
1100         /* Record feed-forward clock leap second adjustment. */
1101         ffi->leapsec_adjustment = cest.leapsec_total;
1102         if (clock_snap->ffcount > cest.leapsec_next)
1103                 ffi->leapsec_adjustment -= cest.leapsec;
1104
1105         /* Record feed-forward clock status and error. */
1106         clock_snap->ff_info.status = cest.status;
1107         ffcount = clock_snap->ffcount - cest.update_ffcount;
1108         ffclock_convert_delta(ffcount, cest.period, &bt);
1109         /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1110         bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1111         /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1112         bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1113         clock_snap->ff_info.error = bt;
1114 #endif
1115 }
1116
1117 /*
1118  * Convert a sysclock snapshot into a struct bintime based on the specified
1119  * clock source and flags.
1120  */
1121 int
1122 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1123     int whichclock, uint32_t flags)
1124 {
1125         struct bintime boottimebin;
1126 #ifdef FFCLOCK
1127         struct bintime bt2;
1128         uint64_t period;
1129 #endif
1130
1131         switch (whichclock) {
1132         case SYSCLOCK_FBCK:
1133                 *bt = cs->fb_info.tick_time;
1134
1135                 /* If snapshot was created with !fast, delta will be >0. */
1136                 if (cs->delta > 0)
1137                         bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1138
1139                 if ((flags & FBCLOCK_UPTIME) == 0) {
1140                         getboottimebin(&boottimebin);
1141                         bintime_add(bt, &boottimebin);
1142                 }
1143                 break;
1144 #ifdef FFCLOCK
1145         case SYSCLOCK_FFWD:
1146                 if (flags & FFCLOCK_LERP) {
1147                         *bt = cs->ff_info.tick_time_lerp;
1148                         period = cs->ff_info.period_lerp;
1149                 } else {
1150                         *bt = cs->ff_info.tick_time;
1151                         period = cs->ff_info.period;
1152                 }
1153
1154                 /* If snapshot was created with !fast, delta will be >0. */
1155                 if (cs->delta > 0) {
1156                         ffclock_convert_delta(cs->delta, period, &bt2);
1157                         bintime_add(bt, &bt2);
1158                 }
1159
1160                 /* Leap second adjustment. */
1161                 if (flags & FFCLOCK_LEAPSEC)
1162                         bt->sec -= cs->ff_info.leapsec_adjustment;
1163
1164                 /* Boot time adjustment, for uptime/monotonic clocks. */
1165                 if (flags & FFCLOCK_UPTIME)
1166                         bintime_sub(bt, &ffclock_boottime);
1167                 break;
1168 #endif
1169         default:
1170                 return (EINVAL);
1171                 break;
1172         }
1173
1174         return (0);
1175 }
1176
1177 /*
1178  * Initialize a new timecounter and possibly use it.
1179  */
1180 void
1181 tc_init(struct timecounter *tc)
1182 {
1183         u_int u;
1184         struct sysctl_oid *tc_root;
1185
1186         u = tc->tc_frequency / tc->tc_counter_mask;
1187         /* XXX: We need some margin here, 10% is a guess */
1188         u *= 11;
1189         u /= 10;
1190         if (u > hz && tc->tc_quality >= 0) {
1191                 tc->tc_quality = -2000;
1192                 if (bootverbose) {
1193                         printf("Timecounter \"%s\" frequency %ju Hz",
1194                             tc->tc_name, (uintmax_t)tc->tc_frequency);
1195                         printf(" -- Insufficient hz, needs at least %u\n", u);
1196                 }
1197         } else if (tc->tc_quality >= 0 || bootverbose) {
1198                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1199                     tc->tc_name, (uintmax_t)tc->tc_frequency,
1200                     tc->tc_quality);
1201         }
1202
1203         /*
1204          * Set up sysctl tree for this counter.
1205          */
1206         tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1207             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1208             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1209             "timecounter description", "timecounter");
1210         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1211             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1212             "mask for implemented bits");
1213         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1214             "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1215             sizeof(*tc), sysctl_kern_timecounter_get, "IU",
1216             "current timecounter value");
1217         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1218             "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1219             sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
1220             "timecounter frequency");
1221         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1222             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1223             "goodness of time counter");
1224
1225         mtx_lock(&tc_lock);
1226         tc->tc_next = timecounters;
1227         timecounters = tc;
1228
1229         /*
1230          * Do not automatically switch if the current tc was specifically
1231          * chosen.  Never automatically use a timecounter with negative quality.
1232          * Even though we run on the dummy counter, switching here may be
1233          * worse since this timecounter may not be monotonic.
1234          */
1235         if (tc_chosen)
1236                 goto unlock;
1237         if (tc->tc_quality < 0)
1238                 goto unlock;
1239         if (tc_from_tunable[0] != '\0' &&
1240             strcmp(tc->tc_name, tc_from_tunable) == 0) {
1241                 tc_chosen = 1;
1242                 tc_from_tunable[0] = '\0';
1243         } else {
1244                 if (tc->tc_quality < timecounter->tc_quality)
1245                         goto unlock;
1246                 if (tc->tc_quality == timecounter->tc_quality &&
1247                     tc->tc_frequency < timecounter->tc_frequency)
1248                         goto unlock;
1249         }
1250         (void)tc->tc_get_timecount(tc);
1251         timecounter = tc;
1252 unlock:
1253         mtx_unlock(&tc_lock);
1254 }
1255
1256 /* Report the frequency of the current timecounter. */
1257 uint64_t
1258 tc_getfrequency(void)
1259 {
1260
1261         return (timehands->th_counter->tc_frequency);
1262 }
1263
1264 static bool
1265 sleeping_on_old_rtc(struct thread *td)
1266 {
1267
1268         /*
1269          * td_rtcgen is modified by curthread when it is running,
1270          * and by other threads in this function.  By finding the thread
1271          * on a sleepqueue and holding the lock on the sleepqueue
1272          * chain, we guarantee that the thread is not running and that
1273          * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
1274          * the thread that it was woken due to a real-time clock adjustment.
1275          * (The declaration of td_rtcgen refers to this comment.)
1276          */
1277         if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1278                 td->td_rtcgen = 0;
1279                 return (true);
1280         }
1281         return (false);
1282 }
1283
1284 static struct mtx tc_setclock_mtx;
1285 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1286
1287 /*
1288  * Step our concept of UTC.  This is done by modifying our estimate of
1289  * when we booted.
1290  */
1291 void
1292 tc_setclock(struct timespec *ts)
1293 {
1294         struct timespec tbef, taft;
1295         struct bintime bt, bt2;
1296
1297         timespec2bintime(ts, &bt);
1298         nanotime(&tbef);
1299         mtx_lock_spin(&tc_setclock_mtx);
1300         cpu_tick_calibrate(1);
1301         binuptime(&bt2);
1302         bintime_sub(&bt, &bt2);
1303
1304         /* XXX fiddle all the little crinkly bits around the fiords... */
1305         tc_windup(&bt);
1306         mtx_unlock_spin(&tc_setclock_mtx);
1307
1308         /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1309         atomic_add_rel_int(&rtc_generation, 2);
1310         sleepq_chains_remove_matching(sleeping_on_old_rtc);
1311         if (timestepwarnings) {
1312                 nanotime(&taft);
1313                 log(LOG_INFO,
1314                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1315                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1316                     (intmax_t)taft.tv_sec, taft.tv_nsec,
1317                     (intmax_t)ts->tv_sec, ts->tv_nsec);
1318         }
1319 }
1320
1321 /*
1322  * Recalculate the scaling factor.  We want the number of 1/2^64
1323  * fractions of a second per period of the hardware counter, taking
1324  * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1325  * processing provides us with.
1326  *
1327  * The th_adjustment is nanoseconds per second with 32 bit binary
1328  * fraction and we want 64 bit binary fraction of second:
1329  *
1330  *       x = a * 2^32 / 10^9 = a * 4.294967296
1331  *
1332  * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1333  * we can only multiply by about 850 without overflowing, that
1334  * leaves no suitably precise fractions for multiply before divide.
1335  *
1336  * Divide before multiply with a fraction of 2199/512 results in a
1337  * systematic undercompensation of 10PPM of th_adjustment.  On a
1338  * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
1339  *
1340  * We happily sacrifice the lowest of the 64 bits of our result
1341  * to the goddess of code clarity.
1342  */
1343 static void
1344 recalculate_scaling_factor_and_large_delta(struct timehands *th)
1345 {
1346         uint64_t scale;
1347
1348         scale = (uint64_t)1 << 63;
1349         scale += (th->th_adjustment / 1024) * 2199;
1350         scale /= th->th_counter->tc_frequency;
1351         th->th_scale = scale * 2;
1352         th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1353 }
1354
1355 /*
1356  * Initialize the next struct timehands in the ring and make
1357  * it the active timehands.  Along the way we might switch to a different
1358  * timecounter and/or do seconds processing in NTP.  Slightly magic.
1359  */
1360 static void
1361 tc_windup(struct bintime *new_boottimebin)
1362 {
1363         struct bintime bt;
1364         struct timecounter *tc;
1365         struct timehands *th, *tho;
1366         u_int delta, ncount, ogen;
1367         int i;
1368         time_t t;
1369
1370         /*
1371          * Make the next timehands a copy of the current one, but do
1372          * not overwrite the generation or next pointer.  While we
1373          * update the contents, the generation must be zero.  We need
1374          * to ensure that the zero generation is visible before the
1375          * data updates become visible, which requires release fence.
1376          * For similar reasons, re-reading of the generation after the
1377          * data is read should use acquire fence.
1378          */
1379         tho = timehands;
1380         th = tho->th_next;
1381         ogen = th->th_generation;
1382         th->th_generation = 0;
1383         atomic_thread_fence_rel();
1384         memcpy(th, tho, offsetof(struct timehands, th_generation));
1385         if (new_boottimebin != NULL)
1386                 th->th_boottime = *new_boottimebin;
1387
1388         /*
1389          * Capture a timecounter delta on the current timecounter and if
1390          * changing timecounters, a counter value from the new timecounter.
1391          * Update the offset fields accordingly.
1392          */
1393         tc = atomic_load_ptr(&timecounter);
1394         delta = tc_delta(th);
1395         if (th->th_counter != tc)
1396                 ncount = tc->tc_get_timecount(tc);
1397         else
1398                 ncount = 0;
1399 #ifdef FFCLOCK
1400         ffclock_windup(delta);
1401 #endif
1402         th->th_offset_count += delta;
1403         th->th_offset_count &= th->th_counter->tc_counter_mask;
1404         bintime_add_tc_delta(&th->th_offset, th->th_scale,
1405             th->th_large_delta, delta);
1406
1407         /*
1408          * Hardware latching timecounters may not generate interrupts on
1409          * PPS events, so instead we poll them.  There is a finite risk that
1410          * the hardware might capture a count which is later than the one we
1411          * got above, and therefore possibly in the next NTP second which might
1412          * have a different rate than the current NTP second.  It doesn't
1413          * matter in practice.
1414          */
1415         if (tho->th_counter->tc_poll_pps)
1416                 tho->th_counter->tc_poll_pps(tho->th_counter);
1417
1418         /*
1419          * Deal with NTP second processing.  The loop normally
1420          * iterates at most once, but in extreme situations it might
1421          * keep NTP sane if timeouts are not run for several seconds.
1422          * At boot, the time step can be large when the TOD hardware
1423          * has been read, so on really large steps, we call
1424          * ntp_update_second only twice.  We need to call it twice in
1425          * case we missed a leap second.
1426          */
1427         bt = th->th_offset;
1428         bintime_add(&bt, &th->th_boottime);
1429         i = bt.sec - tho->th_microtime.tv_sec;
1430         if (i > 0) {
1431                 if (i > LARGE_STEP)
1432                         i = 2;
1433
1434                 do {
1435                         t = bt.sec;
1436                         ntp_update_second(&th->th_adjustment, &bt.sec);
1437                         if (bt.sec != t)
1438                                 th->th_boottime.sec += bt.sec - t;
1439                         --i;
1440                 } while (i > 0);
1441
1442                 recalculate_scaling_factor_and_large_delta(th);
1443         }
1444
1445         /* Update the UTC timestamps used by the get*() functions. */
1446         th->th_bintime = bt;
1447         bintime2timeval(&bt, &th->th_microtime);
1448         bintime2timespec(&bt, &th->th_nanotime);
1449
1450         /* Now is a good time to change timecounters. */
1451         if (th->th_counter != tc) {
1452 #ifndef __arm__
1453                 if ((tc->tc_flags & TC_FLAGS_C2STOP) != 0)
1454                         cpu_disable_c2_sleep++;
1455                 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1456                         cpu_disable_c2_sleep--;
1457 #endif
1458                 th->th_counter = tc;
1459                 th->th_offset_count = ncount;
1460                 tc_min_ticktock_freq = max(1, tc->tc_frequency /
1461                     (((uint64_t)tc->tc_counter_mask + 1) / 3));
1462                 recalculate_scaling_factor_and_large_delta(th);
1463 #ifdef FFCLOCK
1464                 ffclock_change_tc(th);
1465 #endif
1466         }
1467
1468         /*
1469          * Now that the struct timehands is again consistent, set the new
1470          * generation number, making sure to not make it zero.
1471          */
1472         if (++ogen == 0)
1473                 ogen = 1;
1474         atomic_store_rel_int(&th->th_generation, ogen);
1475
1476         /* Go live with the new struct timehands. */
1477 #ifdef FFCLOCK
1478         switch (sysclock_active) {
1479         case SYSCLOCK_FBCK:
1480 #endif
1481                 time_second = th->th_microtime.tv_sec;
1482                 time_uptime = th->th_offset.sec;
1483 #ifdef FFCLOCK
1484                 break;
1485         case SYSCLOCK_FFWD:
1486                 time_second = fftimehands->tick_time_lerp.sec;
1487                 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1488                 break;
1489         }
1490 #endif
1491
1492         timehands = th;
1493         timekeep_push_vdso();
1494 }
1495
1496 /* Report or change the active timecounter hardware. */
1497 static int
1498 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1499 {
1500         char newname[32];
1501         struct timecounter *newtc, *tc;
1502         int error;
1503
1504         mtx_lock(&tc_lock);
1505         tc = timecounter;
1506         strlcpy(newname, tc->tc_name, sizeof(newname));
1507         mtx_unlock(&tc_lock);
1508
1509         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1510         if (error != 0 || req->newptr == NULL)
1511                 return (error);
1512
1513         mtx_lock(&tc_lock);
1514         /* Record that the tc in use now was specifically chosen. */
1515         tc_chosen = 1;
1516         if (strcmp(newname, tc->tc_name) == 0) {
1517                 mtx_unlock(&tc_lock);
1518                 return (0);
1519         }
1520         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1521                 if (strcmp(newname, newtc->tc_name) != 0)
1522                         continue;
1523
1524                 /* Warm up new timecounter. */
1525                 (void)newtc->tc_get_timecount(newtc);
1526
1527                 timecounter = newtc;
1528
1529                 /*
1530                  * The vdso timehands update is deferred until the next
1531                  * 'tc_windup()'.
1532                  *
1533                  * This is prudent given that 'timekeep_push_vdso()' does not
1534                  * use any locking and that it can be called in hard interrupt
1535                  * context via 'tc_windup()'.
1536                  */
1537                 break;
1538         }
1539         mtx_unlock(&tc_lock);
1540         return (newtc != NULL ? 0 : EINVAL);
1541 }
1542 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1543     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1544     sysctl_kern_timecounter_hardware, "A",
1545     "Timecounter hardware selected");
1546
1547 /* Report the available timecounter hardware. */
1548 static int
1549 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1550 {
1551         struct sbuf sb;
1552         struct timecounter *tc;
1553         int error;
1554
1555         error = sysctl_wire_old_buffer(req, 0);
1556         if (error != 0)
1557                 return (error);
1558         sbuf_new_for_sysctl(&sb, NULL, 0, req);
1559         mtx_lock(&tc_lock);
1560         for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1561                 if (tc != timecounters)
1562                         sbuf_putc(&sb, ' ');
1563                 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1564         }
1565         mtx_unlock(&tc_lock);
1566         error = sbuf_finish(&sb);
1567         sbuf_delete(&sb);
1568         return (error);
1569 }
1570
1571 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
1572     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
1573     sysctl_kern_timecounter_choice, "A",
1574     "Timecounter hardware detected");
1575
1576 /*
1577  * RFC 2783 PPS-API implementation.
1578  */
1579
1580 /*
1581  *  Return true if the driver is aware of the abi version extensions in the
1582  *  pps_state structure, and it supports at least the given abi version number.
1583  */
1584 static inline int
1585 abi_aware(struct pps_state *pps, int vers)
1586 {
1587
1588         return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1589 }
1590
1591 static int
1592 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1593 {
1594         int err, timo;
1595         pps_seq_t aseq, cseq;
1596         struct timeval tv;
1597
1598         if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1599                 return (EINVAL);
1600
1601         /*
1602          * If no timeout is requested, immediately return whatever values were
1603          * most recently captured.  If timeout seconds is -1, that's a request
1604          * to block without a timeout.  WITNESS won't let us sleep forever
1605          * without a lock (we really don't need a lock), so just repeatedly
1606          * sleep a long time.
1607          */
1608         if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1609                 if (fapi->timeout.tv_sec == -1)
1610                         timo = 0x7fffffff;
1611                 else {
1612                         tv.tv_sec = fapi->timeout.tv_sec;
1613                         tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1614                         timo = tvtohz(&tv);
1615                 }
1616                 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1617                 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1618                 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1619                     cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1620                         if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1621                                 if (pps->flags & PPSFLAG_MTX_SPIN) {
1622                                         err = msleep_spin(pps, pps->driver_mtx,
1623                                             "ppsfch", timo);
1624                                 } else {
1625                                         err = msleep(pps, pps->driver_mtx, PCATCH,
1626                                             "ppsfch", timo);
1627                                 }
1628                         } else {
1629                                 err = tsleep(pps, PCATCH, "ppsfch", timo);
1630                         }
1631                         if (err == EWOULDBLOCK) {
1632                                 if (fapi->timeout.tv_sec == -1) {
1633                                         continue;
1634                                 } else {
1635                                         return (ETIMEDOUT);
1636                                 }
1637                         } else if (err != 0) {
1638                                 return (err);
1639                         }
1640                 }
1641         }
1642
1643         pps->ppsinfo.current_mode = pps->ppsparam.mode;
1644         fapi->pps_info_buf = pps->ppsinfo;
1645
1646         return (0);
1647 }
1648
1649 int
1650 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1651 {
1652         pps_params_t *app;
1653         struct pps_fetch_args *fapi;
1654 #ifdef FFCLOCK
1655         struct pps_fetch_ffc_args *fapi_ffc;
1656 #endif
1657 #ifdef PPS_SYNC
1658         struct pps_kcbind_args *kapi;
1659 #endif
1660
1661         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1662         switch (cmd) {
1663         case PPS_IOC_CREATE:
1664                 return (0);
1665         case PPS_IOC_DESTROY:
1666                 return (0);
1667         case PPS_IOC_SETPARAMS:
1668                 app = (pps_params_t *)data;
1669                 if (app->mode & ~pps->ppscap)
1670                         return (EINVAL);
1671 #ifdef FFCLOCK
1672                 /* Ensure only a single clock is selected for ffc timestamp. */
1673                 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1674                         return (EINVAL);
1675 #endif
1676                 pps->ppsparam = *app;
1677                 return (0);
1678         case PPS_IOC_GETPARAMS:
1679                 app = (pps_params_t *)data;
1680                 *app = pps->ppsparam;
1681                 app->api_version = PPS_API_VERS_1;
1682                 return (0);
1683         case PPS_IOC_GETCAP:
1684                 *(int*)data = pps->ppscap;
1685                 return (0);
1686         case PPS_IOC_FETCH:
1687                 fapi = (struct pps_fetch_args *)data;
1688                 return (pps_fetch(fapi, pps));
1689 #ifdef FFCLOCK
1690         case PPS_IOC_FETCH_FFCOUNTER:
1691                 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1692                 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1693                     PPS_TSFMT_TSPEC)
1694                         return (EINVAL);
1695                 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1696                         return (EOPNOTSUPP);
1697                 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1698                 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1699                 /* Overwrite timestamps if feedback clock selected. */
1700                 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1701                 case PPS_TSCLK_FBCK:
1702                         fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1703                             pps->ppsinfo.assert_timestamp;
1704                         fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1705                             pps->ppsinfo.clear_timestamp;
1706                         break;
1707                 case PPS_TSCLK_FFWD:
1708                         break;
1709                 default:
1710                         break;
1711                 }
1712                 return (0);
1713 #endif /* FFCLOCK */
1714         case PPS_IOC_KCBIND:
1715 #ifdef PPS_SYNC
1716                 kapi = (struct pps_kcbind_args *)data;
1717                 /* XXX Only root should be able to do this */
1718                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1719                         return (EINVAL);
1720                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1721                         return (EINVAL);
1722                 if (kapi->edge & ~pps->ppscap)
1723                         return (EINVAL);
1724                 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1725                     (pps->kcmode & KCMODE_ABIFLAG);
1726                 return (0);
1727 #else
1728                 return (EOPNOTSUPP);
1729 #endif
1730         default:
1731                 return (ENOIOCTL);
1732         }
1733 }
1734
1735 void
1736 pps_init(struct pps_state *pps)
1737 {
1738         pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1739         if (pps->ppscap & PPS_CAPTUREASSERT)
1740                 pps->ppscap |= PPS_OFFSETASSERT;
1741         if (pps->ppscap & PPS_CAPTURECLEAR)
1742                 pps->ppscap |= PPS_OFFSETCLEAR;
1743 #ifdef FFCLOCK
1744         pps->ppscap |= PPS_TSCLK_MASK;
1745 #endif
1746         pps->kcmode &= ~KCMODE_ABIFLAG;
1747 }
1748
1749 void
1750 pps_init_abi(struct pps_state *pps)
1751 {
1752
1753         pps_init(pps);
1754         if (pps->driver_abi > 0) {
1755                 pps->kcmode |= KCMODE_ABIFLAG;
1756                 pps->kernel_abi = PPS_ABI_VERSION;
1757         }
1758 }
1759
1760 void
1761 pps_capture(struct pps_state *pps)
1762 {
1763         struct timehands *th;
1764
1765         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1766         th = timehands;
1767         pps->capgen = atomic_load_acq_int(&th->th_generation);
1768         pps->capth = th;
1769 #ifdef FFCLOCK
1770         pps->capffth = fftimehands;
1771 #endif
1772         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1773         atomic_thread_fence_acq();
1774         if (pps->capgen != th->th_generation)
1775                 pps->capgen = 0;
1776 }
1777
1778 void
1779 pps_event(struct pps_state *pps, int event)
1780 {
1781         struct bintime bt;
1782         struct timespec ts, *tsp, *osp;
1783         u_int tcount, *pcount;
1784         int foff;
1785         pps_seq_t *pseq;
1786 #ifdef FFCLOCK
1787         struct timespec *tsp_ffc;
1788         pps_seq_t *pseq_ffc;
1789         ffcounter *ffcount;
1790 #endif
1791 #ifdef PPS_SYNC
1792         int fhard;
1793 #endif
1794
1795         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1796         /* Nothing to do if not currently set to capture this event type. */
1797         if ((event & pps->ppsparam.mode) == 0)
1798                 return;
1799         /* If the timecounter was wound up underneath us, bail out. */
1800         if (pps->capgen == 0 || pps->capgen !=
1801             atomic_load_acq_int(&pps->capth->th_generation))
1802                 return;
1803
1804         /* Things would be easier with arrays. */
1805         if (event == PPS_CAPTUREASSERT) {
1806                 tsp = &pps->ppsinfo.assert_timestamp;
1807                 osp = &pps->ppsparam.assert_offset;
1808                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1809 #ifdef PPS_SYNC
1810                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1811 #endif
1812                 pcount = &pps->ppscount[0];
1813                 pseq = &pps->ppsinfo.assert_sequence;
1814 #ifdef FFCLOCK
1815                 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1816                 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1817                 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1818 #endif
1819         } else {
1820                 tsp = &pps->ppsinfo.clear_timestamp;
1821                 osp = &pps->ppsparam.clear_offset;
1822                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1823 #ifdef PPS_SYNC
1824                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1825 #endif
1826                 pcount = &pps->ppscount[1];
1827                 pseq = &pps->ppsinfo.clear_sequence;
1828 #ifdef FFCLOCK
1829                 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1830                 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1831                 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1832 #endif
1833         }
1834
1835         /*
1836          * If the timecounter changed, we cannot compare the count values, so
1837          * we have to drop the rest of the PPS-stuff until the next event.
1838          */
1839         if (pps->ppstc != pps->capth->th_counter) {
1840                 pps->ppstc = pps->capth->th_counter;
1841                 *pcount = pps->capcount;
1842                 pps->ppscount[2] = pps->capcount;
1843                 return;
1844         }
1845
1846         /* Convert the count to a timespec. */
1847         tcount = pps->capcount - pps->capth->th_offset_count;
1848         tcount &= pps->capth->th_counter->tc_counter_mask;
1849         bt = pps->capth->th_bintime;
1850         bintime_addx(&bt, pps->capth->th_scale * tcount);
1851         bintime2timespec(&bt, &ts);
1852
1853         /* If the timecounter was wound up underneath us, bail out. */
1854         atomic_thread_fence_acq();
1855         if (pps->capgen != pps->capth->th_generation)
1856                 return;
1857
1858         *pcount = pps->capcount;
1859         (*pseq)++;
1860         *tsp = ts;
1861
1862         if (foff) {
1863                 timespecadd(tsp, osp, tsp);
1864                 if (tsp->tv_nsec < 0) {
1865                         tsp->tv_nsec += 1000000000;
1866                         tsp->tv_sec -= 1;
1867                 }
1868         }
1869
1870 #ifdef FFCLOCK
1871         *ffcount = pps->capffth->tick_ffcount + tcount;
1872         bt = pps->capffth->tick_time;
1873         ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1874         bintime_add(&bt, &pps->capffth->tick_time);
1875         bintime2timespec(&bt, &ts);
1876         (*pseq_ffc)++;
1877         *tsp_ffc = ts;
1878 #endif
1879
1880 #ifdef PPS_SYNC
1881         if (fhard) {
1882                 uint64_t scale;
1883
1884                 /*
1885                  * Feed the NTP PLL/FLL.
1886                  * The FLL wants to know how many (hardware) nanoseconds
1887                  * elapsed since the previous event.
1888                  */
1889                 tcount = pps->capcount - pps->ppscount[2];
1890                 pps->ppscount[2] = pps->capcount;
1891                 tcount &= pps->capth->th_counter->tc_counter_mask;
1892                 scale = (uint64_t)1 << 63;
1893                 scale /= pps->capth->th_counter->tc_frequency;
1894                 scale *= 2;
1895                 bt.sec = 0;
1896                 bt.frac = 0;
1897                 bintime_addx(&bt, scale * tcount);
1898                 bintime2timespec(&bt, &ts);
1899                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1900         }
1901 #endif
1902
1903         /* Wakeup anyone sleeping in pps_fetch().  */
1904         wakeup(pps);
1905 }
1906
1907 /*
1908  * Timecounters need to be updated every so often to prevent the hardware
1909  * counter from overflowing.  Updating also recalculates the cached values
1910  * used by the get*() family of functions, so their precision depends on
1911  * the update frequency.
1912  */
1913
1914 static int tc_tick;
1915 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1916     "Approximate number of hardclock ticks in a millisecond");
1917
1918 void
1919 tc_ticktock(int cnt)
1920 {
1921         static int count;
1922
1923         if (mtx_trylock_spin(&tc_setclock_mtx)) {
1924                 count += cnt;
1925                 if (count >= tc_tick) {
1926                         count = 0;
1927                         tc_windup(NULL);
1928                 }
1929                 mtx_unlock_spin(&tc_setclock_mtx);
1930         }
1931 }
1932
1933 static void __inline
1934 tc_adjprecision(void)
1935 {
1936         int t;
1937
1938         if (tc_timepercentage > 0) {
1939                 t = (99 + tc_timepercentage) / tc_timepercentage;
1940                 tc_precexp = fls(t + (t >> 1)) - 1;
1941                 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1942                 FREQ2BT(hz, &bt_tickthreshold);
1943                 bintime_shift(&bt_timethreshold, tc_precexp);
1944                 bintime_shift(&bt_tickthreshold, tc_precexp);
1945         } else {
1946                 tc_precexp = 31;
1947                 bt_timethreshold.sec = INT_MAX;
1948                 bt_timethreshold.frac = ~(uint64_t)0;
1949                 bt_tickthreshold = bt_timethreshold;
1950         }
1951         sbt_timethreshold = bttosbt(bt_timethreshold);
1952         sbt_tickthreshold = bttosbt(bt_tickthreshold);
1953 }
1954
1955 static int
1956 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1957 {
1958         int error, val;
1959
1960         val = tc_timepercentage;
1961         error = sysctl_handle_int(oidp, &val, 0, req);
1962         if (error != 0 || req->newptr == NULL)
1963                 return (error);
1964         tc_timepercentage = val;
1965         if (cold)
1966                 goto done;
1967         tc_adjprecision();
1968 done:
1969         return (0);
1970 }
1971
1972 /* Set up the requested number of timehands. */
1973 static void
1974 inittimehands(void *dummy)
1975 {
1976         struct timehands *thp;
1977         int i;
1978
1979         TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1980             &timehands_count);
1981         if (timehands_count < 1)
1982                 timehands_count = 1;
1983         if (timehands_count > nitems(ths))
1984                 timehands_count = nitems(ths);
1985         for (i = 1, thp = &ths[0]; i < timehands_count;  thp = &ths[i++])
1986                 thp->th_next = &ths[i];
1987         thp->th_next = &ths[0];
1988
1989         TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1990             sizeof(tc_from_tunable));
1991
1992         mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
1993 }
1994 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1995
1996 static void
1997 inittimecounter(void *dummy)
1998 {
1999         u_int p;
2000         int tick_rate;
2001
2002         /*
2003          * Set the initial timeout to
2004          * max(1, <approx. number of hardclock ticks in a millisecond>).
2005          * People should probably not use the sysctl to set the timeout
2006          * to smaller than its initial value, since that value is the
2007          * smallest reasonable one.  If they want better timestamps they
2008          * should use the non-"get"* functions.
2009          */
2010         if (hz > 1000)
2011                 tc_tick = (hz + 500) / 1000;
2012         else
2013                 tc_tick = 1;
2014         tc_adjprecision();
2015         FREQ2BT(hz, &tick_bt);
2016         tick_sbt = bttosbt(tick_bt);
2017         tick_rate = hz / tc_tick;
2018         FREQ2BT(tick_rate, &tc_tick_bt);
2019         tc_tick_sbt = bttosbt(tc_tick_bt);
2020         p = (tc_tick * 1000000) / hz;
2021         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
2022
2023 #ifdef FFCLOCK
2024         ffclock_init();
2025 #endif
2026
2027         /* warm up new timecounter (again) and get rolling. */
2028         (void)timecounter->tc_get_timecount(timecounter);
2029         mtx_lock_spin(&tc_setclock_mtx);
2030         tc_windup(NULL);
2031         mtx_unlock_spin(&tc_setclock_mtx);
2032 }
2033
2034 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2035
2036 /* Cpu tick handling -------------------------------------------------*/
2037
2038 static bool cpu_tick_variable;
2039 static uint64_t cpu_tick_frequency;
2040
2041 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2042 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2043
2044 static uint64_t
2045 tc_cpu_ticks(void)
2046 {
2047         struct timecounter *tc;
2048         uint64_t res, *base;
2049         unsigned u, *last;
2050
2051         critical_enter();
2052         base = DPCPU_PTR(tc_cpu_ticks_base);
2053         last = DPCPU_PTR(tc_cpu_ticks_last);
2054         tc = timehands->th_counter;
2055         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2056         if (u < *last)
2057                 *base += (uint64_t)tc->tc_counter_mask + 1;
2058         *last = u;
2059         res = u + *base;
2060         critical_exit();
2061         return (res);
2062 }
2063
2064 void
2065 cpu_tick_calibration(void)
2066 {
2067         static time_t last_calib;
2068
2069         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2070                 cpu_tick_calibrate(0);
2071                 last_calib = time_uptime;
2072         }
2073 }
2074
2075 /*
2076  * This function gets called every 16 seconds on only one designated
2077  * CPU in the system from hardclock() via cpu_tick_calibration()().
2078  *
2079  * Whenever the real time clock is stepped we get called with reset=1
2080  * to make sure we handle suspend/resume and similar events correctly.
2081  */
2082
2083 static void
2084 cpu_tick_calibrate(int reset)
2085 {
2086         static uint64_t c_last;
2087         uint64_t c_this, c_delta;
2088         static struct bintime  t_last;
2089         struct bintime t_this, t_delta;
2090         uint32_t divi;
2091
2092         if (reset) {
2093                 /* The clock was stepped, abort & reset */
2094                 t_last.sec = 0;
2095                 return;
2096         }
2097
2098         /* we don't calibrate fixed rate cputicks */
2099         if (!cpu_tick_variable)
2100                 return;
2101
2102         getbinuptime(&t_this);
2103         c_this = cpu_ticks();
2104         if (t_last.sec != 0) {
2105                 c_delta = c_this - c_last;
2106                 t_delta = t_this;
2107                 bintime_sub(&t_delta, &t_last);
2108                 /*
2109                  * Headroom:
2110                  *      2^(64-20) / 16[s] =
2111                  *      2^(44) / 16[s] =
2112                  *      17.592.186.044.416 / 16 =
2113                  *      1.099.511.627.776 [Hz]
2114                  */
2115                 divi = t_delta.sec << 20;
2116                 divi |= t_delta.frac >> (64 - 20);
2117                 c_delta <<= 20;
2118                 c_delta /= divi;
2119                 if (c_delta > cpu_tick_frequency) {
2120                         if (0 && bootverbose)
2121                                 printf("cpu_tick increased to %ju Hz\n",
2122                                     c_delta);
2123                         cpu_tick_frequency = c_delta;
2124                 }
2125         }
2126         c_last = c_this;
2127         t_last = t_this;
2128 }
2129
2130 void
2131 set_cputicker(cpu_tick_f *func, uint64_t freq, bool isvariable)
2132 {
2133
2134         if (func == NULL) {
2135                 cpu_ticks = tc_cpu_ticks;
2136         } else {
2137                 cpu_tick_frequency = freq;
2138                 cpu_tick_variable = isvariable;
2139                 cpu_ticks = func;
2140         }
2141 }
2142
2143 uint64_t
2144 cpu_tickrate(void)
2145 {
2146
2147         if (cpu_ticks == tc_cpu_ticks)
2148                 return (tc_getfrequency());
2149         return (cpu_tick_frequency);
2150 }
2151
2152 /*
2153  * We need to be slightly careful converting cputicks to microseconds.
2154  * There is plenty of margin in 64 bits of microseconds (half a million
2155  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2156  * before divide conversion (to retain precision) we find that the
2157  * margin shrinks to 1.5 hours (one millionth of 146y).
2158  */
2159
2160 uint64_t
2161 cputick2usec(uint64_t tick)
2162 {
2163         uint64_t tr;
2164         tr = cpu_tickrate();
2165         return ((tick / tr) * 1000000ULL) + ((tick % tr) * 1000000ULL) / tr;
2166 }
2167
2168 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
2169
2170 static int vdso_th_enable = 1;
2171 static int
2172 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2173 {
2174         int old_vdso_th_enable, error;
2175
2176         old_vdso_th_enable = vdso_th_enable;
2177         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2178         if (error != 0)
2179                 return (error);
2180         vdso_th_enable = old_vdso_th_enable;
2181         return (0);
2182 }
2183 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2184     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2185     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2186
2187 uint32_t
2188 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2189 {
2190         struct timehands *th;
2191         uint32_t enabled;
2192
2193         th = timehands;
2194         vdso_th->th_scale = th->th_scale;
2195         vdso_th->th_offset_count = th->th_offset_count;
2196         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2197         vdso_th->th_offset = th->th_offset;
2198         vdso_th->th_boottime = th->th_boottime;
2199         if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2200                 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2201                     th->th_counter);
2202         } else
2203                 enabled = 0;
2204         if (!vdso_th_enable)
2205                 enabled = 0;
2206         return (enabled);
2207 }
2208
2209 #ifdef COMPAT_FREEBSD32
2210 uint32_t
2211 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2212 {
2213         struct timehands *th;
2214         uint32_t enabled;
2215
2216         th = timehands;
2217         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2218         vdso_th32->th_offset_count = th->th_offset_count;
2219         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2220         vdso_th32->th_offset.sec = th->th_offset.sec;
2221         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2222         vdso_th32->th_boottime.sec = th->th_boottime.sec;
2223         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2224         if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2225                 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2226                     th->th_counter);
2227         } else
2228                 enabled = 0;
2229         if (!vdso_th_enable)
2230                 enabled = 0;
2231         return (enabled);
2232 }
2233 #endif
2234
2235 #include "opt_ddb.h"
2236 #ifdef DDB
2237 #include <ddb/ddb.h>
2238
2239 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2240 {
2241         struct timehands *th;
2242         struct timecounter *tc;
2243         u_int val1, val2;
2244
2245         th = timehands;
2246         tc = th->th_counter;
2247         val1 = tc->tc_get_timecount(tc);
2248         __compiler_membar();
2249         val2 = tc->tc_get_timecount(tc);
2250
2251         db_printf("timecounter %p %s\n", tc, tc->tc_name);
2252         db_printf("  mask %#x freq %ju qual %d flags %#x priv %p\n",
2253             tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2254             tc->tc_flags, tc->tc_priv);
2255         db_printf("  val %#x %#x\n", val1, val2);
2256         db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2257             (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2258             th->th_large_delta, th->th_offset_count, th->th_generation);
2259         db_printf("  offset %jd %jd boottime %jd %jd\n",
2260             (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2261             (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);
2262 }
2263 #endif