sys/kern/kern_tc.c

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