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         uint64_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 int tc_min_ticktock_freq = 1;
  91
  92 time_t time_second = 1;
  93 time_t time_uptime = 1;
  94
  95 struct bintime boottimebin;
  96 struct timeval boottime;
  97 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
  98 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
  99     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
 100
 101 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
 102 SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
 103
 104 static int timestepwarnings;
 105 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
 106     &timestepwarnings, 0, "Log time steps");
 107
 108 static void tc_windup(void);
 109 static void cpu_tick_calibrate(int);
 110
 111 static int
 112 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 113 {
 114 #ifdef SCTL_MASK32
 115         int tv[2];
 116
 117         if (req->flags & SCTL_MASK32) {
 118                 tv[0] = boottime.tv_sec;
 119                 tv[1] = boottime.tv_usec;
 120                 return SYSCTL_OUT(req, tv, sizeof(tv));
 121         } else
 122 #endif
 123                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
 124 }
 125
 126 static int
 127 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 128 {
 129         u_int ncount;
 130         struct timecounter *tc = arg1;
 131
 132         ncount = tc->tc_get_timecount(tc);
 133         return sysctl_handle_int(oidp, &ncount, 0, req);
 134 }
 135
 136 static int
 137 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 138 {
 139         uint64_t freq;
 140         struct timecounter *tc = arg1;
 141
 142         freq = tc->tc_frequency;
 143         return sysctl_handle_64(oidp, &freq, 0, req);
 144 }
 145
 146 /*
 147  * Return the difference between the timehands' counter value now and what
 148  * was when we copied it to the timehands' offset_count.
 149  */
 150 static __inline u_int
 151 tc_delta(struct timehands *th)
 152 {
 153         struct timecounter *tc;
 154
 155         tc = th->th_counter;
 156         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 157             tc->tc_counter_mask);
 158 }
 159
 160 /*
 161  * Functions for reading the time.  We have to loop until we are sure that
 162  * the timehands that we operated on was not updated under our feet.  See
 163  * the comment in <sys/time.h> for a description of these 12 functions.
 164  */
 165
 166 void
 167 binuptime(struct bintime *bt)
 168 {
 169         struct timehands *th;
 170         u_int gen;
 171
 172         do {
 173                 th = timehands;
 174                 gen = th->th_generation;
 175                 *bt = th->th_offset;
 176                 bintime_addx(bt, th->th_scale * tc_delta(th));
 177         } while (gen == 0 || gen != th->th_generation);
 178 }
 179
 180 void
 181 nanouptime(struct timespec *tsp)
 182 {
 183         struct bintime bt;
 184
 185         binuptime(&bt);
 186         bintime2timespec(&bt, tsp);
 187 }
 188
 189 void
 190 microuptime(struct timeval *tvp)
 191 {
 192         struct bintime bt;
 193
 194         binuptime(&bt);
 195         bintime2timeval(&bt, tvp);
 196 }
 197
 198 void
 199 bintime(struct bintime *bt)
 200 {
 201
 202         binuptime(bt);
 203         bintime_add(bt, &boottimebin);
 204 }
 205
 206 void
 207 nanotime(struct timespec *tsp)
 208 {
 209         struct bintime bt;
 210
 211         bintime(&bt);
 212         bintime2timespec(&bt, tsp);
 213 }
 214
 215 void
 216 microtime(struct timeval *tvp)
 217 {
 218         struct bintime bt;
 219
 220         bintime(&bt);
 221         bintime2timeval(&bt, tvp);
 222 }
 223
 224 void
 225 getbinuptime(struct bintime *bt)
 226 {
 227         struct timehands *th;
 228         u_int gen;
 229
 230         do {
 231                 th = timehands;
 232                 gen = th->th_generation;
 233                 *bt = th->th_offset;
 234         } while (gen == 0 || gen != th->th_generation);
 235 }
 236
 237 void
 238 getnanouptime(struct timespec *tsp)
 239 {
 240         struct timehands *th;
 241         u_int gen;
 242
 243         do {
 244                 th = timehands;
 245                 gen = th->th_generation;
 246                 bintime2timespec(&th->th_offset, tsp);
 247         } while (gen == 0 || gen != th->th_generation);
 248 }
 249
 250 void
 251 getmicrouptime(struct timeval *tvp)
 252 {
 253         struct timehands *th;
 254         u_int gen;
 255
 256         do {
 257                 th = timehands;
 258                 gen = th->th_generation;
 259                 bintime2timeval(&th->th_offset, tvp);
 260         } while (gen == 0 || gen != th->th_generation);
 261 }
 262
 263 void
 264 getbintime(struct bintime *bt)
 265 {
 266         struct timehands *th;
 267         u_int gen;
 268
 269         do {
 270                 th = timehands;
 271                 gen = th->th_generation;
 272                 *bt = th->th_offset;
 273         } while (gen == 0 || gen != th->th_generation);
 274         bintime_add(bt, &boottimebin);
 275 }
 276
 277 void
 278 getnanotime(struct timespec *tsp)
 279 {
 280         struct timehands *th;
 281         u_int gen;
 282
 283         do {
 284                 th = timehands;
 285                 gen = th->th_generation;
 286                 *tsp = th->th_nanotime;
 287         } while (gen == 0 || gen != th->th_generation);
 288 }
 289
 290 void
 291 getmicrotime(struct timeval *tvp)
 292 {
 293         struct timehands *th;
 294         u_int gen;
 295
 296         do {
 297                 th = timehands;
 298                 gen = th->th_generation;
 299                 *tvp = th->th_microtime;
 300         } while (gen == 0 || gen != th->th_generation);
 301 }
 302
 303 /*
 304  * Initialize a new timecounter and possibly use it.
 305  */
 306 void
 307 tc_init(struct timecounter *tc)
 308 {
 309         u_int u;
 310         struct sysctl_oid *tc_root;
 311
 312         u = tc->tc_frequency / tc->tc_counter_mask;
 313         /* XXX: We need some margin here, 10% is a guess */
 314         u *= 11;
 315         u /= 10;
 316         if (u > hz && tc->tc_quality >= 0) {
 317                 tc->tc_quality = -2000;
 318                 if (bootverbose) {
 319                         printf("Timecounter \"%s\" frequency %ju Hz",
 320                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 321                         printf(" -- Insufficient hz, needs at least %u\n", u);
 322                 }
 323         } else if (tc->tc_quality >= 0 || bootverbose) {
 324                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 325                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 326                     tc->tc_quality);
 327         }
 328
 329         tc->tc_next = timecounters;
 330         timecounters = tc;
 331         /*
 332          * Set up sysctl tree for this counter.
 333          */
 334         tc_root = SYSCTL_ADD_NODE(NULL,
 335             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 336             CTLFLAG_RW, 0, "timecounter description");
 337         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 338             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 339             "mask for implemented bits");
 340         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 341             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 342             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 343         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 344             "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
 345              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 346         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 347             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 348             "goodness of time counter");
 349         /*
 350          * Never automatically use a timecounter with negative quality.
 351          * Even though we run on the dummy counter, switching here may be
 352          * worse since this timecounter may not be monotonous.
 353          */
 354         if (tc->tc_quality < 0)
 355                 return;
 356         if (tc->tc_quality < timecounter->tc_quality)
 357                 return;
 358         if (tc->tc_quality == timecounter->tc_quality &&
 359             tc->tc_frequency < timecounter->tc_frequency)
 360                 return;
 361         (void)tc->tc_get_timecount(tc);
 362         (void)tc->tc_get_timecount(tc);
 363         timecounter = tc;
 364 }
 365
 366 /* Report the frequency of the current timecounter. */
 367 uint64_t
 368 tc_getfrequency(void)
 369 {
 370
 371         return (timehands->th_counter->tc_frequency);
 372 }
 373
 374 /*
 375  * Step our concept of UTC.  This is done by modifying our estimate of
 376  * when we booted.
 377  * XXX: not locked.
 378  */
 379 void
 380 tc_setclock(struct timespec *ts)
 381 {
 382         struct timespec tbef, taft;
 383         struct bintime bt, bt2;
 384
 385         cpu_tick_calibrate(1);
 386         nanotime(&tbef);
 387         timespec2bintime(ts, &bt);
 388         binuptime(&bt2);
 389         bintime_sub(&bt, &bt2);
 390         bintime_add(&bt2, &boottimebin);
 391         boottimebin = bt;
 392         bintime2timeval(&bt, &boottime);
 393
 394         /* XXX fiddle all the little crinkly bits around the fiords... */
 395         tc_windup();
 396         nanotime(&taft);
 397         if (timestepwarnings) {
 398                 log(LOG_INFO,
 399                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 400                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 401                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 402                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 403         }
 404         cpu_tick_calibrate(1);
 405 }
 406
 407 /*
 408  * Initialize the next struct timehands in the ring and make
 409  * it the active timehands.  Along the way we might switch to a different
 410  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 411  */
 412 static void
 413 tc_windup(void)
 414 {
 415         struct bintime bt;
 416         struct timehands *th, *tho;
 417         uint64_t scale;
 418         u_int delta, ncount, ogen;
 419         int i;
 420         time_t t;
 421
 422         /*
 423          * Make the next timehands a copy of the current one, but do not
 424          * overwrite the generation or next pointer.  While we update
 425          * the contents, the generation must be zero.
 426          */
 427         tho = timehands;
 428         th = tho->th_next;
 429         ogen = th->th_generation;
 430         th->th_generation = 0;
 431         bcopy(tho, th, offsetof(struct timehands, th_generation));
 432
 433         /*
 434          * Capture a timecounter delta on the current timecounter and if
 435          * changing timecounters, a counter value from the new timecounter.
 436          * Update the offset fields accordingly.
 437          */
 438         delta = tc_delta(th);
 439         if (th->th_counter != timecounter)
 440                 ncount = timecounter->tc_get_timecount(timecounter);
 441         else
 442                 ncount = 0;
 443         th->th_offset_count += delta;
 444         th->th_offset_count &= th->th_counter->tc_counter_mask;
 445         while (delta > th->th_counter->tc_frequency) {
 446                 /* Eat complete unadjusted seconds. */
 447                 delta -= th->th_counter->tc_frequency;
 448                 th->th_offset.sec++;
 449         }
 450         if ((delta > th->th_counter->tc_frequency / 2) &&
 451             (th->th_scale * delta < ((uint64_t)1 << 63))) {
 452                 /* The product th_scale * delta just barely overflows. */
 453                 th->th_offset.sec++;
 454         }
 455         bintime_addx(&th->th_offset, th->th_scale * delta);
 456
 457         /*
 458          * Hardware latching timecounters may not generate interrupts on
 459          * PPS events, so instead we poll them.  There is a finite risk that
 460          * the hardware might capture a count which is later than the one we
 461          * got above, and therefore possibly in the next NTP second which might
 462          * have a different rate than the current NTP second.  It doesn't
 463          * matter in practice.
 464          */
 465         if (tho->th_counter->tc_poll_pps)
 466                 tho->th_counter->tc_poll_pps(tho->th_counter);
 467
 468         /*
 469          * Deal with NTP second processing.  The for loop normally
 470          * iterates at most once, but in extreme situations it might
 471          * keep NTP sane if timeouts are not run for several seconds.
 472          * At boot, the time step can be large when the TOD hardware
 473          * has been read, so on really large steps, we call
 474          * ntp_update_second only twice.  We need to call it twice in
 475          * case we missed a leap second.
 476          */
 477         bt = th->th_offset;
 478         bintime_add(&bt, &boottimebin);
 479         i = bt.sec - tho->th_microtime.tv_sec;
 480         if (i > LARGE_STEP)
 481                 i = 2;
 482         for (; i > 0; i--) {
 483                 t = bt.sec;
 484                 ntp_update_second(&th->th_adjustment, &bt.sec);
 485                 if (bt.sec != t)
 486                         boottimebin.sec += bt.sec - t;
 487         }
 488         /* Update the UTC timestamps used by the get*() functions. */
 489         /* XXX shouldn't do this here.  Should force non-`get' versions. */
 490         bintime2timeval(&bt, &th->th_microtime);
 491         bintime2timespec(&bt, &th->th_nanotime);
 492
 493         /* Now is a good time to change timecounters. */
 494         if (th->th_counter != timecounter) {
 495                 th->th_counter = timecounter;
 496                 th->th_offset_count = ncount;
 497                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
 498                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
 499         }
 500
 501         /*-
 502          * Recalculate the scaling factor.  We want the number of 1/2^64
 503          * fractions of a second per period of the hardware counter, taking
 504          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 505          * processing provides us with.
 506          *
 507          * The th_adjustment is nanoseconds per second with 32 bit binary
 508          * fraction and we want 64 bit binary fraction of second:
 509          *
 510          *       x = a * 2^32 / 10^9 = a * 4.294967296
 511          *
 512          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 513          * we can only multiply by about 850 without overflowing, that
 514          * leaves no suitably precise fractions for multiply before divide.
 515          *
 516          * Divide before multiply with a fraction of 2199/512 results in a
 517          * systematic undercompensation of 10PPM of th_adjustment.  On a
 518          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 519          *
 520          * We happily sacrifice the lowest of the 64 bits of our result
 521          * to the goddess of code clarity.
 522          *
 523          */
 524         scale = (uint64_t)1 << 63;
 525         scale += (th->th_adjustment / 1024) * 2199;
 526         scale /= th->th_counter->tc_frequency;
 527         th->th_scale = scale * 2;
 528
 529         /*
 530          * Now that the struct timehands is again consistent, set the new
 531          * generation number, making sure to not make it zero.
 532          */
 533         if (++ogen == 0)
 534                 ogen = 1;
 535         th->th_generation = ogen;
 536
 537         /* Go live with the new struct timehands. */
 538         time_second = th->th_microtime.tv_sec;
 539         time_uptime = th->th_offset.sec;
 540         timehands = th;
 541 }
 542
 543 /* Report or change the active timecounter hardware. */
 544 static int
 545 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 546 {
 547         char newname[32];
 548         struct timecounter *newtc, *tc;
 549         int error;
 550
 551         tc = timecounter;
 552         strlcpy(newname, tc->tc_name, sizeof(newname));
 553
 554         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 555         if (error != 0 || req->newptr == NULL ||
 556             strcmp(newname, tc->tc_name) == 0)
 557                 return (error);
 558         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 559                 if (strcmp(newname, newtc->tc_name) != 0)
 560                         continue;
 561
 562                 /* Warm up new timecounter. */
 563                 (void)newtc->tc_get_timecount(newtc);
 564                 (void)newtc->tc_get_timecount(newtc);
 565
 566                 timecounter = newtc;
 567                 return (0);
 568         }
 569         return (EINVAL);
 570 }
 571
 572 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 573     0, 0, sysctl_kern_timecounter_hardware, "A",
 574     "Timecounter hardware selected");
 575
 576
 577 /* Report or change the active timecounter hardware. */
 578 static int
 579 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 580 {
 581         char buf[32], *spc;
 582         struct timecounter *tc;
 583         int error;
 584
 585         spc = "";
 586         error = 0;
 587         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
 588                 sprintf(buf, "%s%s(%d)",
 589                     spc, tc->tc_name, tc->tc_quality);
 590                 error = SYSCTL_OUT(req, buf, strlen(buf));
 591                 spc = " ";
 592         }
 593         return (error);
 594 }
 595
 596 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 597     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
 598
 599 /*
 600  * RFC 2783 PPS-API implementation.
 601  */
 602
 603 int
 604 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 605 {
 606         pps_params_t *app;
 607         struct pps_fetch_args *fapi;
 608 #ifdef PPS_SYNC
 609         struct pps_kcbind_args *kapi;
 610 #endif
 611
 612         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 613         switch (cmd) {
 614         case PPS_IOC_CREATE:
 615                 return (0);
 616         case PPS_IOC_DESTROY:
 617                 return (0);
 618         case PPS_IOC_SETPARAMS:
 619                 app = (pps_params_t *)data;
 620                 if (app->mode & ~pps->ppscap)
 621                         return (EINVAL);
 622                 pps->ppsparam = *app;
 623                 return (0);
 624         case PPS_IOC_GETPARAMS:
 625                 app = (pps_params_t *)data;
 626                 *app = pps->ppsparam;
 627                 app->api_version = PPS_API_VERS_1;
 628                 return (0);
 629         case PPS_IOC_GETCAP:
 630                 *(int*)data = pps->ppscap;
 631                 return (0);
 632         case PPS_IOC_FETCH:
 633                 fapi = (struct pps_fetch_args *)data;
 634                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 635                         return (EINVAL);
 636                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
 637                         return (EOPNOTSUPP);
 638                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
 639                 fapi->pps_info_buf = pps->ppsinfo;
 640                 return (0);
 641         case PPS_IOC_KCBIND:
 642 #ifdef PPS_SYNC
 643                 kapi = (struct pps_kcbind_args *)data;
 644                 /* XXX Only root should be able to do this */
 645                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 646                         return (EINVAL);
 647                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 648                         return (EINVAL);
 649                 if (kapi->edge & ~pps->ppscap)
 650                         return (EINVAL);
 651                 pps->kcmode = kapi->edge;
 652                 return (0);
 653 #else
 654                 return (EOPNOTSUPP);
 655 #endif
 656         default:
 657                 return (ENOIOCTL);
 658         }
 659 }
 660
 661 void
 662 pps_init(struct pps_state *pps)
 663 {
 664         pps->ppscap |= PPS_TSFMT_TSPEC;
 665         if (pps->ppscap & PPS_CAPTUREASSERT)
 666                 pps->ppscap |= PPS_OFFSETASSERT;
 667         if (pps->ppscap & PPS_CAPTURECLEAR)
 668                 pps->ppscap |= PPS_OFFSETCLEAR;
 669 }
 670
 671 void
 672 pps_capture(struct pps_state *pps)
 673 {
 674         struct timehands *th;
 675
 676         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 677         th = timehands;
 678         pps->capgen = th->th_generation;
 679         pps->capth = th;
 680         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 681         if (pps->capgen != th->th_generation)
 682                 pps->capgen = 0;
 683 }
 684
 685 void
 686 pps_event(struct pps_state *pps, int event)
 687 {
 688         struct bintime bt;
 689         struct timespec ts, *tsp, *osp;
 690         u_int tcount, *pcount;
 691         int foff, fhard;
 692         pps_seq_t *pseq;
 693
 694         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 695         /* If the timecounter was wound up underneath us, bail out. */
 696         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
 697                 return;
 698
 699         /* Things would be easier with arrays. */
 700         if (event == PPS_CAPTUREASSERT) {
 701                 tsp = &pps->ppsinfo.assert_timestamp;
 702                 osp = &pps->ppsparam.assert_offset;
 703                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 704                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 705                 pcount = &pps->ppscount[0];
 706                 pseq = &pps->ppsinfo.assert_sequence;
 707         } else {
 708                 tsp = &pps->ppsinfo.clear_timestamp;
 709                 osp = &pps->ppsparam.clear_offset;
 710                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 711                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 712                 pcount = &pps->ppscount[1];
 713                 pseq = &pps->ppsinfo.clear_sequence;
 714         }
 715
 716         /*
 717          * If the timecounter changed, we cannot compare the count values, so
 718          * we have to drop the rest of the PPS-stuff until the next event.
 719          */
 720         if (pps->ppstc != pps->capth->th_counter) {
 721                 pps->ppstc = pps->capth->th_counter;
 722                 *pcount = pps->capcount;
 723                 pps->ppscount[2] = pps->capcount;
 724                 return;
 725         }
 726
 727         /* Convert the count to a timespec. */
 728         tcount = pps->capcount - pps->capth->th_offset_count;
 729         tcount &= pps->capth->th_counter->tc_counter_mask;
 730         bt = pps->capth->th_offset;
 731         bintime_addx(&bt, pps->capth->th_scale * tcount);
 732         bintime_add(&bt, &boottimebin);
 733         bintime2timespec(&bt, &ts);
 734
 735         /* If the timecounter was wound up underneath us, bail out. */
 736         if (pps->capgen != pps->capth->th_generation)
 737                 return;
 738
 739         *pcount = pps->capcount;
 740         (*pseq)++;
 741         *tsp = ts;
 742
 743         if (foff) {
 744                 timespecadd(tsp, osp);
 745                 if (tsp->tv_nsec < 0) {
 746                         tsp->tv_nsec += 1000000000;
 747                         tsp->tv_sec -= 1;
 748                 }
 749         }
 750 #ifdef PPS_SYNC
 751         if (fhard) {
 752                 uint64_t scale;
 753
 754                 /*
 755                  * Feed the NTP PLL/FLL.
 756                  * The FLL wants to know how many (hardware) nanoseconds
 757                  * elapsed since the previous event.
 758                  */
 759                 tcount = pps->capcount - pps->ppscount[2];
 760                 pps->ppscount[2] = pps->capcount;
 761                 tcount &= pps->capth->th_counter->tc_counter_mask;
 762                 scale = (uint64_t)1 << 63;
 763                 scale /= pps->capth->th_counter->tc_frequency;
 764                 scale *= 2;
 765                 bt.sec = 0;
 766                 bt.frac = 0;
 767                 bintime_addx(&bt, scale * tcount);
 768                 bintime2timespec(&bt, &ts);
 769                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 770         }
 771 #endif
 772 }
 773
 774 /*
 775  * Timecounters need to be updated every so often to prevent the hardware
 776  * counter from overflowing.  Updating also recalculates the cached values
 777  * used by the get*() family of functions, so their precision depends on
 778  * the update frequency.
 779  */
 780
 781 static int tc_tick;
 782 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
 783     "Approximate number of hardclock ticks in a millisecond");
 784
 785 void
 786 tc_ticktock(int cnt)
 787 {
 788         static int count;
 789
 790         count += cnt;
 791         if (count < tc_tick)
 792                 return;
 793         count = 0;
 794         tc_windup();
 795 }
 796
 797 static void
 798 inittimecounter(void *dummy)
 799 {
 800         u_int p;
 801
 802         /*
 803          * Set the initial timeout to
 804          * max(1, <approx. number of hardclock ticks in a millisecond>).
 805          * People should probably not use the sysctl to set the timeout
 806          * to smaller than its inital value, since that value is the
 807          * smallest reasonable one.  If they want better timestamps they
 808          * should use the non-"get"* functions.
 809          */
 810         if (hz > 1000)
 811                 tc_tick = (hz + 500) / 1000;
 812         else
 813                 tc_tick = 1;
 814         p = (tc_tick * 1000000) / hz;
 815         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 816
 817         /* warm up new timecounter (again) and get rolling. */
 818         (void)timecounter->tc_get_timecount(timecounter);
 819         (void)timecounter->tc_get_timecount(timecounter);
 820         tc_windup();
 821 }
 822
 823 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 824
 825 /* Cpu tick handling -------------------------------------------------*/
 826
 827 static int cpu_tick_variable;
 828 static uint64_t cpu_tick_frequency;
 829
 830 static uint64_t
 831 tc_cpu_ticks(void)
 832 {
 833         static uint64_t base;
 834         static unsigned last;
 835         unsigned u;
 836         struct timecounter *tc;
 837
 838         tc = timehands->th_counter;
 839         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 840         if (u < last)
 841                 base += (uint64_t)tc->tc_counter_mask + 1;
 842         last = u;
 843         return (u + base);
 844 }
 845
 846 void
 847 cpu_tick_calibration(void)
 848 {
 849         static time_t last_calib;
 850
 851         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 852                 cpu_tick_calibrate(0);
 853                 last_calib = time_uptime;
 854         }
 855 }
 856
 857 /*
 858  * This function gets called every 16 seconds on only one designated
 859  * CPU in the system from hardclock() via cpu_tick_calibration()().
 860  *
 861  * Whenever the real time clock is stepped we get called with reset=1
 862  * to make sure we handle suspend/resume and similar events correctly.
 863  */
 864
 865 static void
 866 cpu_tick_calibrate(int reset)
 867 {
 868         static uint64_t c_last;
 869         uint64_t c_this, c_delta;
 870         static struct bintime  t_last;
 871         struct bintime t_this, t_delta;
 872         uint32_t divi;
 873
 874         if (reset) {
 875                 /* The clock was stepped, abort & reset */
 876                 t_last.sec = 0;
 877                 return;
 878         }
 879
 880         /* we don't calibrate fixed rate cputicks */
 881         if (!cpu_tick_variable)
 882                 return;
 883
 884         getbinuptime(&t_this);
 885         c_this = cpu_ticks();
 886         if (t_last.sec != 0) {
 887                 c_delta = c_this - c_last;
 888                 t_delta = t_this;
 889                 bintime_sub(&t_delta, &t_last);
 890                 /*
 891                  * Headroom:
 892                  *      2^(64-20) / 16[s] =
 893                  *      2^(44) / 16[s] =
 894                  *      17.592.186.044.416 / 16 =
 895                  *      1.099.511.627.776 [Hz]
 896                  */
 897                 divi = t_delta.sec << 20;
 898                 divi |= t_delta.frac >> (64 - 20);
 899                 c_delta <<= 20;
 900                 c_delta /= divi;
 901                 if (c_delta > cpu_tick_frequency) {
 902                         if (0 && bootverbose)
 903                                 printf("cpu_tick increased to %ju Hz\n",
 904                                     c_delta);
 905                         cpu_tick_frequency = c_delta;
 906                 }
 907         }
 908         c_last = c_this;
 909         t_last = t_this;
 910 }
 911
 912 void
 913 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 914 {
 915
 916         if (func == NULL) {
 917                 cpu_ticks = tc_cpu_ticks;
 918         } else {
 919                 cpu_tick_frequency = freq;
 920                 cpu_tick_variable = var;
 921                 cpu_ticks = func;
 922         }
 923 }
 924
 925 uint64_t
 926 cpu_tickrate(void)
 927 {
 928
 929         if (cpu_ticks == tc_cpu_ticks)
 930                 return (tc_getfrequency());
 931         return (cpu_tick_frequency);
 932 }
 933
 934 /*
 935  * We need to be slightly careful converting cputicks to microseconds.
 936  * There is plenty of margin in 64 bits of microseconds (half a million
 937  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 938  * before divide conversion (to retain precision) we find that the
 939  * margin shrinks to 1.5 hours (one millionth of 146y).
 940  * With a three prong approach we never lose significant bits, no
 941  * matter what the cputick rate and length of timeinterval is.
 942  */
 943
 944 uint64_t
 945 cputick2usec(uint64_t tick)
 946 {
 947
 948         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 949                 return (tick / (cpu_tickrate() / 1000000LL));
 950         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 951                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 952         else
 953                 return ((tick * 1000000LL) / cpu_tickrate());
 954 }
 955
 956 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;