sys/kern/kern_tc.c

   1 /*-
   2  * ----------------------------------------------------------------------------
   3  * "THE BEER-WARE LICENSE" (Revision 42):
   4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
   5  * can do whatever you want with this stuff. If we meet some day, and you think
   6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
   7  * ----------------------------------------------------------------------------
   8  */
   9
  10 #include <sys/cdefs.h>
  11 __FBSDID("$FreeBSD$");
  12
  13 #include "opt_ntp.h"
  14
  15 #include <sys/param.h>
  16 #include <sys/kernel.h>
  17 #include <sys/sysctl.h>
  18 #include <sys/syslog.h>
  19 #include <sys/systm.h>
  20 #include <sys/timepps.h>
  21 #include <sys/timetc.h>
  22 #include <sys/timex.h>
  23
  24 /*
  25  * A large step happens on boot.  This constant detects such steps.
  26  * It is relatively small so that ntp_update_second gets called enough
  27  * in the typical 'missed a couple of seconds' case, but doesn't loop
  28  * forever when the time step is large.
  29  */
  30 #define LARGE_STEP      200
  31
  32 /*
  33  * Implement a dummy timecounter which we can use until we get a real one
  34  * in the air.  This allows the console and other early stuff to use
  35  * time services.
  36  */
  37
  38 static u_int
  39 dummy_get_timecount(struct timecounter *tc)
  40 {
  41         static u_int now;
  42
  43         return (++now);
  44 }
  45
  46 static struct timecounter dummy_timecounter = {
  47         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
  48 };
  49
  50 struct timehands {
  51         /* These fields must be initialized by the driver. */
  52         struct timecounter      *th_counter;
  53         int64_t                 th_adjustment;
  54         u_int64_t               th_scale;
  55         u_int                   th_offset_count;
  56         struct bintime          th_offset;
  57         struct timeval          th_microtime;
  58         struct timespec         th_nanotime;
  59         /* Fields not to be copied in tc_windup start with th_generation. */
  60         volatile u_int          th_generation;
  61         struct timehands        *th_next;
  62 };
  63
  64 static struct timehands th0;
  65 static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0};
  66 static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9};
  67 static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8};
  68 static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7};
  69 static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6};
  70 static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5};
  71 static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4};
  72 static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3};
  73 static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2};
  74 static struct timehands th0 = {
  75         &dummy_timecounter,
  76         0,
  77         (uint64_t)-1 / 1000000,
  78         0,
  79         {1, 0},
  80         {0, 0},
  81         {0, 0},
  82         1,
  83         &th1
  84 };
  85
  86 static struct timehands *volatile timehands = &th0;
  87 struct timecounter *timecounter = &dummy_timecounter;
  88 static struct timecounter *timecounters = &dummy_timecounter;
  89
  90 time_t time_second = 1;
  91 time_t time_uptime = 1;
  92
  93 static struct bintime boottimebin;
  94 struct timeval boottime;
  95 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
  96 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
  97     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
  98
  99 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
 100 SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
 101
 102 static int timestepwarnings;
 103 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
 104     &timestepwarnings, 0, "");
 105
 106 static void tc_windup(void);
 107 static void cpu_tick_calibrate(int);
 108
 109 static int
 110 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 111 {
 112 #ifdef SCTL_MASK32
 113         int tv[2];
 114
 115         if (req->flags & SCTL_MASK32) {
 116                 tv[0] = boottime.tv_sec;
 117                 tv[1] = boottime.tv_usec;
 118                 return SYSCTL_OUT(req, tv, sizeof(tv));
 119         } else
 120 #endif
 121                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
 122 }
 123
 124 static int
 125 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 126 {
 127         u_int ncount;
 128         struct timecounter *tc = arg1;
 129
 130         ncount = tc->tc_get_timecount(tc);
 131         return sysctl_handle_int(oidp, &ncount, 0, req);
 132 }
 133
 134 static int
 135 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 136 {
 137         u_int64_t freq;
 138         struct timecounter *tc = arg1;
 139
 140         freq = tc->tc_frequency;
 141         return sysctl_handle_quad(oidp, &freq, 0, req);
 142 }
 143
 144 /*
 145  * Return the difference between the timehands' counter value now and what
 146  * was when we copied it to the timehands' offset_count.
 147  */
 148 static __inline u_int
 149 tc_delta(struct timehands *th)
 150 {
 151         struct timecounter *tc;
 152
 153         tc = th->th_counter;
 154         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 155             tc->tc_counter_mask);
 156 }
 157
 158 /*
 159  * Functions for reading the time.  We have to loop until we are sure that
 160  * the timehands that we operated on was not updated under our feet.  See
 161  * the comment in <sys/time.h> for a description of these 12 functions.
 162  */
 163
 164 void
 165 binuptime(struct bintime *bt)
 166 {
 167         struct timehands *th;
 168         u_int gen;
 169
 170         do {
 171                 th = timehands;
 172                 gen = th->th_generation;
 173                 *bt = th->th_offset;
 174                 bintime_addx(bt, th->th_scale * tc_delta(th));
 175         } while (gen == 0 || gen != th->th_generation);
 176 }
 177
 178 void
 179 nanouptime(struct timespec *tsp)
 180 {
 181         struct bintime bt;
 182
 183         binuptime(&bt);
 184         bintime2timespec(&bt, tsp);
 185 }
 186
 187 void
 188 microuptime(struct timeval *tvp)
 189 {
 190         struct bintime bt;
 191
 192         binuptime(&bt);
 193         bintime2timeval(&bt, tvp);
 194 }
 195
 196 void
 197 bintime(struct bintime *bt)
 198 {
 199
 200         binuptime(bt);
 201         bintime_add(bt, &boottimebin);
 202 }
 203
 204 void
 205 nanotime(struct timespec *tsp)
 206 {
 207         struct bintime bt;
 208
 209         bintime(&bt);
 210         bintime2timespec(&bt, tsp);
 211 }
 212
 213 void
 214 microtime(struct timeval *tvp)
 215 {
 216         struct bintime bt;
 217
 218         bintime(&bt);
 219         bintime2timeval(&bt, tvp);
 220 }
 221
 222 void
 223 getbinuptime(struct bintime *bt)
 224 {
 225         struct timehands *th;
 226         u_int gen;
 227
 228         do {
 229                 th = timehands;
 230                 gen = th->th_generation;
 231                 *bt = th->th_offset;
 232         } while (gen == 0 || gen != th->th_generation);
 233 }
 234
 235 void
 236 getnanouptime(struct timespec *tsp)
 237 {
 238         struct timehands *th;
 239         u_int gen;
 240
 241         do {
 242                 th = timehands;
 243                 gen = th->th_generation;
 244                 bintime2timespec(&th->th_offset, tsp);
 245         } while (gen == 0 || gen != th->th_generation);
 246 }
 247
 248 void
 249 getmicrouptime(struct timeval *tvp)
 250 {
 251         struct timehands *th;
 252         u_int gen;
 253
 254         do {
 255                 th = timehands;
 256                 gen = th->th_generation;
 257                 bintime2timeval(&th->th_offset, tvp);
 258         } while (gen == 0 || gen != th->th_generation);
 259 }
 260
 261 void
 262 getbintime(struct bintime *bt)
 263 {
 264         struct timehands *th;
 265         u_int gen;
 266
 267         do {
 268                 th = timehands;
 269                 gen = th->th_generation;
 270                 *bt = th->th_offset;
 271         } while (gen == 0 || gen != th->th_generation);
 272         bintime_add(bt, &boottimebin);
 273 }
 274
 275 void
 276 getnanotime(struct timespec *tsp)
 277 {
 278         struct timehands *th;
 279         u_int gen;
 280
 281         do {
 282                 th = timehands;
 283                 gen = th->th_generation;
 284                 *tsp = th->th_nanotime;
 285         } while (gen == 0 || gen != th->th_generation);
 286 }
 287
 288 void
 289 getmicrotime(struct timeval *tvp)
 290 {
 291         struct timehands *th;
 292         u_int gen;
 293
 294         do {
 295                 th = timehands;
 296                 gen = th->th_generation;
 297                 *tvp = th->th_microtime;
 298         } while (gen == 0 || gen != th->th_generation);
 299 }
 300
 301 /*
 302  * Initialize a new timecounter and possibly use it.
 303  */
 304 void
 305 tc_init(struct timecounter *tc)
 306 {
 307         u_int u;
 308         struct sysctl_oid *tc_root;
 309
 310         u = tc->tc_frequency / tc->tc_counter_mask;
 311         /* XXX: We need some margin here, 10% is a guess */
 312         u *= 11;
 313         u /= 10;
 314         if (u > hz && tc->tc_quality >= 0) {
 315                 tc->tc_quality = -2000;
 316                 if (bootverbose) {
 317                         printf("Timecounter \"%s\" frequency %ju Hz",
 318                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 319                         printf(" -- Insufficient hz, needs at least %u\n", u);
 320                 }
 321         } else if (tc->tc_quality >= 0 || bootverbose) {
 322                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 323                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 324                     tc->tc_quality);
 325         }
 326
 327         tc->tc_next = timecounters;
 328         timecounters = tc;
 329         /*
 330          * Set up sysctl tree for this counter.
 331          */
 332         tc_root = SYSCTL_ADD_NODE(NULL,
 333             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 334             CTLFLAG_RW, 0, "timecounter description");
 335         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 336             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 337             "mask for implemented bits");
 338         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 339             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 340             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 341         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 342             "frequency", CTLTYPE_QUAD | CTLFLAG_RD, tc, sizeof(*tc),
 343              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 344         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 345             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 346             "goodness of time counter");
 347         /*
 348          * Never automatically use a timecounter with negative quality.
 349          * Even though we run on the dummy counter, switching here may be
 350          * worse since this timecounter may not be monotonous.
 351          */
 352         if (tc->tc_quality < 0)
 353                 return;
 354         if (tc->tc_quality < timecounter->tc_quality)
 355                 return;
 356         if (tc->tc_quality == timecounter->tc_quality &&
 357             tc->tc_frequency < timecounter->tc_frequency)
 358                 return;
 359         (void)tc->tc_get_timecount(tc);
 360         (void)tc->tc_get_timecount(tc);
 361         timecounter = tc;
 362 }
 363
 364 /* Report the frequency of the current timecounter. */
 365 u_int64_t
 366 tc_getfrequency(void)
 367 {
 368
 369         return (timehands->th_counter->tc_frequency);
 370 }
 371
 372 /*
 373  * Step our concept of UTC.  This is done by modifying our estimate of
 374  * when we booted.
 375  * XXX: not locked.
 376  */
 377 void
 378 tc_setclock(struct timespec *ts)
 379 {
 380         struct timespec tbef, taft;
 381         struct bintime bt, bt2;
 382
 383         cpu_tick_calibrate(1);
 384         nanotime(&tbef);
 385         timespec2bintime(ts, &bt);
 386         binuptime(&bt2);
 387         bintime_sub(&bt, &bt2);
 388         bintime_add(&bt2, &boottimebin);
 389         boottimebin = bt;
 390         bintime2timeval(&bt, &boottime);
 391
 392         /* XXX fiddle all the little crinkly bits around the fiords... */
 393         tc_windup();
 394         nanotime(&taft);
 395         if (timestepwarnings) {
 396                 log(LOG_INFO,
 397                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 398                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 399                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 400                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 401         }
 402         cpu_tick_calibrate(1);
 403 }
 404
 405 /*
 406  * Initialize the next struct timehands in the ring and make
 407  * it the active timehands.  Along the way we might switch to a different
 408  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 409  */
 410 static void
 411 tc_windup(void)
 412 {
 413         struct bintime bt;
 414         struct timehands *th, *tho;
 415         u_int64_t scale;
 416         u_int delta, ncount, ogen;
 417         int i;
 418         time_t t;
 419
 420         /*
 421          * Make the next timehands a copy of the current one, but do not
 422          * overwrite the generation or next pointer.  While we update
 423          * the contents, the generation must be zero.
 424          */
 425         tho = timehands;
 426         th = tho->th_next;
 427         ogen = th->th_generation;
 428         th->th_generation = 0;
 429         bcopy(tho, th, offsetof(struct timehands, th_generation));
 430
 431         /*
 432          * Capture a timecounter delta on the current timecounter and if
 433          * changing timecounters, a counter value from the new timecounter.
 434          * Update the offset fields accordingly.
 435          */
 436         delta = tc_delta(th);
 437         if (th->th_counter != timecounter)
 438                 ncount = timecounter->tc_get_timecount(timecounter);
 439         else
 440                 ncount = 0;
 441         th->th_offset_count += delta;
 442         th->th_offset_count &= th->th_counter->tc_counter_mask;
 443         bintime_addx(&th->th_offset, th->th_scale * delta);
 444
 445         /*
 446          * Hardware latching timecounters may not generate interrupts on
 447          * PPS events, so instead we poll them.  There is a finite risk that
 448          * the hardware might capture a count which is later than the one we
 449          * got above, and therefore possibly in the next NTP second which might
 450          * have a different rate than the current NTP second.  It doesn't
 451          * matter in practice.
 452          */
 453         if (tho->th_counter->tc_poll_pps)
 454                 tho->th_counter->tc_poll_pps(tho->th_counter);
 455
 456         /*
 457          * Deal with NTP second processing.  The for loop normally
 458          * iterates at most once, but in extreme situations it might
 459          * keep NTP sane if timeouts are not run for several seconds.
 460          * At boot, the time step can be large when the TOD hardware
 461          * has been read, so on really large steps, we call
 462          * ntp_update_second only twice.  We need to call it twice in
 463          * case we missed a leap second.
 464          */
 465         bt = th->th_offset;
 466         bintime_add(&bt, &boottimebin);
 467         i = bt.sec - tho->th_microtime.tv_sec;
 468         if (i > LARGE_STEP)
 469                 i = 2;
 470         for (; i > 0; i--) {
 471                 t = bt.sec;
 472                 ntp_update_second(&th->th_adjustment, &bt.sec);
 473                 if (bt.sec != t)
 474                         boottimebin.sec += bt.sec - t;
 475         }
 476         /* Update the UTC timestamps used by the get*() functions. */
 477         /* XXX shouldn't do this here.  Should force non-`get' versions. */
 478         bintime2timeval(&bt, &th->th_microtime);
 479         bintime2timespec(&bt, &th->th_nanotime);
 480
 481         /* Now is a good time to change timecounters. */
 482         if (th->th_counter != timecounter) {
 483                 th->th_counter = timecounter;
 484                 th->th_offset_count = ncount;
 485         }
 486
 487         /*-
 488          * Recalculate the scaling factor.  We want the number of 1/2^64
 489          * fractions of a second per period of the hardware counter, taking
 490          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 491          * processing provides us with.
 492          *
 493          * The th_adjustment is nanoseconds per second with 32 bit binary
 494          * fraction and we want 64 bit binary fraction of second:
 495          *
 496          *       x = a * 2^32 / 10^9 = a * 4.294967296
 497          *
 498          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 499          * we can only multiply by about 850 without overflowing, that
 500          * leaves no suitably precise fractions for multiply before divide.
 501          *
 502          * Divide before multiply with a fraction of 2199/512 results in a
 503          * systematic undercompensation of 10PPM of th_adjustment.  On a
 504          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 505          *
 506          * We happily sacrifice the lowest of the 64 bits of our result
 507          * to the goddess of code clarity.
 508          *
 509          */
 510         scale = (u_int64_t)1 << 63;
 511         scale += (th->th_adjustment / 1024) * 2199;
 512         scale /= th->th_counter->tc_frequency;
 513         th->th_scale = scale * 2;
 514
 515         /*
 516          * Now that the struct timehands is again consistent, set the new
 517          * generation number, making sure to not make it zero.
 518          */
 519         if (++ogen == 0)
 520                 ogen = 1;
 521         th->th_generation = ogen;
 522
 523         /* Go live with the new struct timehands. */
 524         time_second = th->th_microtime.tv_sec;
 525         time_uptime = th->th_offset.sec;
 526         timehands = th;
 527 }
 528
 529 /* Report or change the active timecounter hardware. */
 530 static int
 531 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 532 {
 533         char newname[32];
 534         struct timecounter *newtc, *tc;
 535         int error;
 536
 537         tc = timecounter;
 538         strlcpy(newname, tc->tc_name, sizeof(newname));
 539
 540         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 541         if (error != 0 || req->newptr == NULL ||
 542             strcmp(newname, tc->tc_name) == 0)
 543                 return (error);
 544         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 545                 if (strcmp(newname, newtc->tc_name) != 0)
 546                         continue;
 547
 548                 /* Warm up new timecounter. */
 549                 (void)newtc->tc_get_timecount(newtc);
 550                 (void)newtc->tc_get_timecount(newtc);
 551
 552                 timecounter = newtc;
 553                 return (0);
 554         }
 555         return (EINVAL);
 556 }
 557
 558 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 559     0, 0, sysctl_kern_timecounter_hardware, "A", "");
 560
 561
 562 /* Report or change the active timecounter hardware. */
 563 static int
 564 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 565 {
 566         char buf[32], *spc;
 567         struct timecounter *tc;
 568         int error;
 569
 570         spc = "";
 571         error = 0;
 572         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
 573                 sprintf(buf, "%s%s(%d)",
 574                     spc, tc->tc_name, tc->tc_quality);
 575                 error = SYSCTL_OUT(req, buf, strlen(buf));
 576                 spc = " ";
 577         }
 578         return (error);
 579 }
 580
 581 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 582     0, 0, sysctl_kern_timecounter_choice, "A", "");
 583
 584 /*
 585  * RFC 2783 PPS-API implementation.
 586  */
 587
 588 int
 589 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 590 {
 591         pps_params_t *app;
 592         struct pps_fetch_args *fapi;
 593 #ifdef PPS_SYNC
 594         struct pps_kcbind_args *kapi;
 595 #endif
 596
 597         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 598         switch (cmd) {
 599         case PPS_IOC_CREATE:
 600                 return (0);
 601         case PPS_IOC_DESTROY:
 602                 return (0);
 603         case PPS_IOC_SETPARAMS:
 604                 app = (pps_params_t *)data;
 605                 if (app->mode & ~pps->ppscap)
 606                         return (EINVAL);
 607                 pps->ppsparam = *app;
 608                 return (0);
 609         case PPS_IOC_GETPARAMS:
 610                 app = (pps_params_t *)data;
 611                 *app = pps->ppsparam;
 612                 app->api_version = PPS_API_VERS_1;
 613                 return (0);
 614         case PPS_IOC_GETCAP:
 615                 *(int*)data = pps->ppscap;
 616                 return (0);
 617         case PPS_IOC_FETCH:
 618                 fapi = (struct pps_fetch_args *)data;
 619                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 620                         return (EINVAL);
 621                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
 622                         return (EOPNOTSUPP);
 623                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
 624                 fapi->pps_info_buf = pps->ppsinfo;
 625                 return (0);
 626         case PPS_IOC_KCBIND:
 627 #ifdef PPS_SYNC
 628                 kapi = (struct pps_kcbind_args *)data;
 629                 /* XXX Only root should be able to do this */
 630                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 631                         return (EINVAL);
 632                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 633                         return (EINVAL);
 634                 if (kapi->edge & ~pps->ppscap)
 635                         return (EINVAL);
 636                 pps->kcmode = kapi->edge;
 637                 return (0);
 638 #else
 639                 return (EOPNOTSUPP);
 640 #endif
 641         default:
 642                 return (ENOIOCTL);
 643         }
 644 }
 645
 646 void
 647 pps_init(struct pps_state *pps)
 648 {
 649         pps->ppscap |= PPS_TSFMT_TSPEC;
 650         if (pps->ppscap & PPS_CAPTUREASSERT)
 651                 pps->ppscap |= PPS_OFFSETASSERT;
 652         if (pps->ppscap & PPS_CAPTURECLEAR)
 653                 pps->ppscap |= PPS_OFFSETCLEAR;
 654 }
 655
 656 void
 657 pps_capture(struct pps_state *pps)
 658 {
 659         struct timehands *th;
 660
 661         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 662         th = timehands;
 663         pps->capgen = th->th_generation;
 664         pps->capth = th;
 665         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 666         if (pps->capgen != th->th_generation)
 667                 pps->capgen = 0;
 668 }
 669
 670 void
 671 pps_event(struct pps_state *pps, int event)
 672 {
 673         struct bintime bt;
 674         struct timespec ts, *tsp, *osp;
 675         u_int tcount, *pcount;
 676         int foff, fhard;
 677         pps_seq_t *pseq;
 678
 679         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 680         /* If the timecounter was wound up underneath us, bail out. */
 681         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
 682                 return;
 683
 684         /* Things would be easier with arrays. */
 685         if (event == PPS_CAPTUREASSERT) {
 686                 tsp = &pps->ppsinfo.assert_timestamp;
 687                 osp = &pps->ppsparam.assert_offset;
 688                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 689                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 690                 pcount = &pps->ppscount[0];
 691                 pseq = &pps->ppsinfo.assert_sequence;
 692         } else {
 693                 tsp = &pps->ppsinfo.clear_timestamp;
 694                 osp = &pps->ppsparam.clear_offset;
 695                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 696                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 697                 pcount = &pps->ppscount[1];
 698                 pseq = &pps->ppsinfo.clear_sequence;
 699         }
 700
 701         /*
 702          * If the timecounter changed, we cannot compare the count values, so
 703          * we have to drop the rest of the PPS-stuff until the next event.
 704          */
 705         if (pps->ppstc != pps->capth->th_counter) {
 706                 pps->ppstc = pps->capth->th_counter;
 707                 *pcount = pps->capcount;
 708                 pps->ppscount[2] = pps->capcount;
 709                 return;
 710         }
 711
 712         /* Convert the count to a timespec. */
 713         tcount = pps->capcount - pps->capth->th_offset_count;
 714         tcount &= pps->capth->th_counter->tc_counter_mask;
 715         bt = pps->capth->th_offset;
 716         bintime_addx(&bt, pps->capth->th_scale * tcount);
 717         bintime_add(&bt, &boottimebin);
 718         bintime2timespec(&bt, &ts);
 719
 720         /* If the timecounter was wound up underneath us, bail out. */
 721         if (pps->capgen != pps->capth->th_generation)
 722                 return;
 723
 724         *pcount = pps->capcount;
 725         (*pseq)++;
 726         *tsp = ts;
 727
 728         if (foff) {
 729                 timespecadd(tsp, osp);
 730                 if (tsp->tv_nsec < 0) {
 731                         tsp->tv_nsec += 1000000000;
 732                         tsp->tv_sec -= 1;
 733                 }
 734         }
 735 #ifdef PPS_SYNC
 736         if (fhard) {
 737                 u_int64_t scale;
 738
 739                 /*
 740                  * Feed the NTP PLL/FLL.
 741                  * The FLL wants to know how many (hardware) nanoseconds
 742                  * elapsed since the previous event.
 743                  */
 744                 tcount = pps->capcount - pps->ppscount[2];
 745                 pps->ppscount[2] = pps->capcount;
 746                 tcount &= pps->capth->th_counter->tc_counter_mask;
 747                 scale = (u_int64_t)1 << 63;
 748                 scale /= pps->capth->th_counter->tc_frequency;
 749                 scale *= 2;
 750                 bt.sec = 0;
 751                 bt.frac = 0;
 752                 bintime_addx(&bt, scale * tcount);
 753                 bintime2timespec(&bt, &ts);
 754                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 755         }
 756 #endif
 757 }
 758
 759 /*
 760  * Timecounters need to be updated every so often to prevent the hardware
 761  * counter from overflowing.  Updating also recalculates the cached values
 762  * used by the get*() family of functions, so their precision depends on
 763  * the update frequency.
 764  */
 765
 766 static int tc_tick;
 767 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
 768
 769 void
 770 tc_ticktock(void)
 771 {
 772         static int count;
 773         static time_t last_calib;
 774
 775         if (++count < tc_tick)
 776                 return;
 777         count = 0;
 778         tc_windup();
 779         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 780                 cpu_tick_calibrate(0);
 781                 last_calib = time_uptime;
 782         }
 783 }
 784
 785 static void
 786 inittimecounter(void *dummy)
 787 {
 788         u_int p;
 789
 790         /*
 791          * Set the initial timeout to
 792          * max(1, <approx. number of hardclock ticks in a millisecond>).
 793          * People should probably not use the sysctl to set the timeout
 794          * to smaller than its inital value, since that value is the
 795          * smallest reasonable one.  If they want better timestamps they
 796          * should use the non-"get"* functions.
 797          */
 798         if (hz > 1000)
 799                 tc_tick = (hz + 500) / 1000;
 800         else
 801                 tc_tick = 1;
 802         p = (tc_tick * 1000000) / hz;
 803         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 804
 805         /* warm up new timecounter (again) and get rolling. */
 806         (void)timecounter->tc_get_timecount(timecounter);
 807         (void)timecounter->tc_get_timecount(timecounter);
 808 }
 809
 810 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 811
 812 /* Cpu tick handling -------------------------------------------------*/
 813
 814 static int cpu_tick_variable;
 815 static uint64_t cpu_tick_frequency;
 816
 817 static uint64_t
 818 tc_cpu_ticks(void)
 819 {
 820         static uint64_t base;
 821         static unsigned last;
 822         unsigned u;
 823         struct timecounter *tc;
 824
 825         tc = timehands->th_counter;
 826         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 827         if (u < last)
 828                 base += (uint64_t)tc->tc_counter_mask + 1;
 829         last = u;
 830         return (u + base);
 831 }
 832
 833 /*
 834  * This function gets called every 16 seconds on only one designated
 835  * CPU in the system from hardclock() via tc_ticktock().
 836  *
 837  * Whenever the real time clock is stepped we get called with reset=1
 838  * to make sure we handle suspend/resume and similar events correctly.
 839  */
 840
 841 static void
 842 cpu_tick_calibrate(int reset)
 843 {
 844         static uint64_t c_last;
 845         uint64_t c_this, c_delta;
 846         static struct bintime  t_last;
 847         struct bintime t_this, t_delta;
 848         uint32_t divi;
 849
 850         if (reset) {
 851                 /* The clock was stepped, abort & reset */
 852                 t_last.sec = 0;
 853                 return;
 854         }
 855
 856         /* we don't calibrate fixed rate cputicks */
 857         if (!cpu_tick_variable)
 858                 return;
 859
 860         getbinuptime(&t_this);
 861         c_this = cpu_ticks();
 862         if (t_last.sec != 0) {
 863                 c_delta = c_this - c_last;
 864                 t_delta = t_this;
 865                 bintime_sub(&t_delta, &t_last);
 866                 /*
 867                  * Validate that 16 +/- 1/256 seconds passed.
 868                  * After division by 16 this gives us a precision of
 869                  * roughly 250PPM which is sufficient
 870                  */
 871                 if (t_delta.sec > 16 || (
 872                     t_delta.sec == 16 && t_delta.frac >= (0x01LL << 56))) {
 873                         /* too long */
 874                         if (bootverbose)
 875                                 printf("t_delta %ju.%016jx too long\n",
 876                                     (uintmax_t)t_delta.sec,
 877                                     (uintmax_t)t_delta.frac);
 878                 } else if (t_delta.sec < 15 ||
 879                     (t_delta.sec == 15 && t_delta.frac <= (0xffLL << 56))) {
 880                         /* too short */
 881                         if (bootverbose)
 882                                 printf("t_delta %ju.%016jx too short\n",
 883                                     (uintmax_t)t_delta.sec,
 884                                     (uintmax_t)t_delta.frac);
 885                 } else {
 886                         /* just right */
 887                         /*
 888                          * Headroom:
 889                          *      2^(64-20) / 16[s] =
 890                          *      2^(44) / 16[s] =
 891                          *      17.592.186.044.416 / 16 =
 892                          *      1.099.511.627.776 [Hz]
 893                          */
 894                         divi = t_delta.sec << 20;
 895                         divi |= t_delta.frac >> (64 - 20);
 896                         c_delta <<= 20;
 897                         c_delta /= divi;
 898                         if (c_delta  > cpu_tick_frequency) {
 899                                 if (0 && bootverbose)
 900                                         printf("cpu_tick increased to %ju Hz\n",
 901                                             c_delta);
 902                                 cpu_tick_frequency = c_delta;
 903                         }
 904                 }
 905         }
 906         c_last = c_this;
 907         t_last = t_this;
 908 }
 909
 910 void
 911 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 912 {
 913
 914         if (func == NULL) {
 915                 cpu_ticks = tc_cpu_ticks;
 916         } else {
 917                 cpu_tick_frequency = freq;
 918                 cpu_tick_variable = var;
 919                 cpu_ticks = func;
 920         }
 921 }
 922
 923 uint64_t
 924 cpu_tickrate(void)
 925 {
 926
 927         if (cpu_ticks == tc_cpu_ticks)
 928                 return (tc_getfrequency());
 929         return (cpu_tick_frequency);
 930 }
 931
 932 /*
 933  * We need to be slightly careful converting cputicks to microseconds.
 934  * There is plenty of margin in 64 bits of microseconds (half a million
 935  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 936  * before divide conversion (to retain precision) we find that the
 937  * margin shrinks to 1.5 hours (one millionth of 146y).
 938  * With a three prong approach we never lose significant bits, no
 939  * matter what the cputick rate and length of timeinterval is.
 940  */
 941
 942 uint64_t
 943 cputick2usec(uint64_t tick)
 944 {
 945
 946         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 947                 return (tick / (cpu_tickrate() / 1000000LL));
 948         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 949                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 950         else
 951                 return ((tick * 1000000LL) / cpu_tickrate());
 952 }
 953
 954 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;