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 #ifdef TC_COUNTERS
 107 #define TC_STATS(foo) \
 108         static u_int foo; \
 109         SYSCTL_UINT(_kern_timecounter, OID_AUTO, foo, CTLFLAG_RD, &foo, 0, "");\
 110         struct __hack
 111
 112 TC_STATS(nbinuptime);    TC_STATS(nnanouptime);    TC_STATS(nmicrouptime);
 113 TC_STATS(nbintime);      TC_STATS(nnanotime);      TC_STATS(nmicrotime);
 114 TC_STATS(ngetbinuptime); TC_STATS(ngetnanouptime); TC_STATS(ngetmicrouptime);
 115 TC_STATS(ngetbintime);   TC_STATS(ngetnanotime);   TC_STATS(ngetmicrotime);
 116 TC_STATS(nsetclock);
 117
 118 #define TC_COUNT(var)   var++
 119 #undef TC_STATS
 120 #else
 121 #define TC_COUNT(var)   /* nothing */
 122 #endif /* TC_COUNTERS */
 123
 124 static void tc_windup(void);
 125 static void cpu_tick_calibrate(int);
 126
 127 static int
 128 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 129 {
 130 #ifdef SCTL_MASK32
 131         int tv[2];
 132
 133         if (req->flags & SCTL_MASK32) {
 134                 tv[0] = boottime.tv_sec;
 135                 tv[1] = boottime.tv_usec;
 136                 return SYSCTL_OUT(req, tv, sizeof(tv));
 137         } else
 138 #endif
 139                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
 140 }
 141
 142 static int
 143 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 144 {
 145         u_int ncount;
 146         struct timecounter *tc = arg1;
 147
 148         ncount = tc->tc_get_timecount(tc);
 149         return sysctl_handle_int(oidp, &ncount, 0, req);
 150 }
 151
 152 static int
 153 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 154 {
 155         u_int64_t freq;
 156         struct timecounter *tc = arg1;
 157
 158         freq = tc->tc_frequency;
 159         return sysctl_handle_quad(oidp, &freq, 0, req);
 160 }
 161
 162 /*
 163  * Return the difference between the timehands' counter value now and what
 164  * was when we copied it to the timehands' offset_count.
 165  */
 166 static __inline u_int
 167 tc_delta(struct timehands *th)
 168 {
 169         struct timecounter *tc;
 170
 171         tc = th->th_counter;
 172         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 173             tc->tc_counter_mask);
 174 }
 175
 176 /*
 177  * Functions for reading the time.  We have to loop until we are sure that
 178  * the timehands that we operated on was not updated under our feet.  See
 179  * the comment in <sys/time.h> for a description of these 12 functions.
 180  */
 181
 182 void
 183 binuptime(struct bintime *bt)
 184 {
 185         struct timehands *th;
 186         u_int gen;
 187
 188         TC_COUNT(nbinuptime);
 189         do {
 190                 th = timehands;
 191                 gen = th->th_generation;
 192                 *bt = th->th_offset;
 193                 bintime_addx(bt, th->th_scale * tc_delta(th));
 194         } while (gen == 0 || gen != th->th_generation);
 195 }
 196
 197 void
 198 nanouptime(struct timespec *tsp)
 199 {
 200         struct bintime bt;
 201
 202         TC_COUNT(nnanouptime);
 203         binuptime(&bt);
 204         bintime2timespec(&bt, tsp);
 205 }
 206
 207 void
 208 microuptime(struct timeval *tvp)
 209 {
 210         struct bintime bt;
 211
 212         TC_COUNT(nmicrouptime);
 213         binuptime(&bt);
 214         bintime2timeval(&bt, tvp);
 215 }
 216
 217 void
 218 bintime(struct bintime *bt)
 219 {
 220
 221         TC_COUNT(nbintime);
 222         binuptime(bt);
 223         bintime_add(bt, &boottimebin);
 224 }
 225
 226 void
 227 nanotime(struct timespec *tsp)
 228 {
 229         struct bintime bt;
 230
 231         TC_COUNT(nnanotime);
 232         bintime(&bt);
 233         bintime2timespec(&bt, tsp);
 234 }
 235
 236 void
 237 microtime(struct timeval *tvp)
 238 {
 239         struct bintime bt;
 240
 241         TC_COUNT(nmicrotime);
 242         bintime(&bt);
 243         bintime2timeval(&bt, tvp);
 244 }
 245
 246 void
 247 getbinuptime(struct bintime *bt)
 248 {
 249         struct timehands *th;
 250         u_int gen;
 251
 252         TC_COUNT(ngetbinuptime);
 253         do {
 254                 th = timehands;
 255                 gen = th->th_generation;
 256                 *bt = th->th_offset;
 257         } while (gen == 0 || gen != th->th_generation);
 258 }
 259
 260 void
 261 getnanouptime(struct timespec *tsp)
 262 {
 263         struct timehands *th;
 264         u_int gen;
 265
 266         TC_COUNT(ngetnanouptime);
 267         do {
 268                 th = timehands;
 269                 gen = th->th_generation;
 270                 bintime2timespec(&th->th_offset, tsp);
 271         } while (gen == 0 || gen != th->th_generation);
 272 }
 273
 274 void
 275 getmicrouptime(struct timeval *tvp)
 276 {
 277         struct timehands *th;
 278         u_int gen;
 279
 280         TC_COUNT(ngetmicrouptime);
 281         do {
 282                 th = timehands;
 283                 gen = th->th_generation;
 284                 bintime2timeval(&th->th_offset, tvp);
 285         } while (gen == 0 || gen != th->th_generation);
 286 }
 287
 288 void
 289 getbintime(struct bintime *bt)
 290 {
 291         struct timehands *th;
 292         u_int gen;
 293
 294         TC_COUNT(ngetbintime);
 295         do {
 296                 th = timehands;
 297                 gen = th->th_generation;
 298                 *bt = th->th_offset;
 299         } while (gen == 0 || gen != th->th_generation);
 300         bintime_add(bt, &boottimebin);
 301 }
 302
 303 void
 304 getnanotime(struct timespec *tsp)
 305 {
 306         struct timehands *th;
 307         u_int gen;
 308
 309         TC_COUNT(ngetnanotime);
 310         do {
 311                 th = timehands;
 312                 gen = th->th_generation;
 313                 *tsp = th->th_nanotime;
 314         } while (gen == 0 || gen != th->th_generation);
 315 }
 316
 317 void
 318 getmicrotime(struct timeval *tvp)
 319 {
 320         struct timehands *th;
 321         u_int gen;
 322
 323         TC_COUNT(ngetmicrotime);
 324         do {
 325                 th = timehands;
 326                 gen = th->th_generation;
 327                 *tvp = th->th_microtime;
 328         } while (gen == 0 || gen != th->th_generation);
 329 }
 330
 331 /*
 332  * Initialize a new timecounter and possibly use it.
 333  */
 334 void
 335 tc_init(struct timecounter *tc)
 336 {
 337         u_int u;
 338         struct sysctl_oid *tc_root;
 339
 340         u = tc->tc_frequency / tc->tc_counter_mask;
 341         /* XXX: We need some margin here, 10% is a guess */
 342         u *= 11;
 343         u /= 10;
 344         if (u > hz && tc->tc_quality >= 0) {
 345                 tc->tc_quality = -2000;
 346                 if (bootverbose) {
 347                         printf("Timecounter \"%s\" frequency %ju Hz",
 348                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 349                         printf(" -- Insufficient hz, needs at least %u\n", u);
 350                 }
 351         } else if (tc->tc_quality >= 0 || bootverbose) {
 352                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 353                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 354                     tc->tc_quality);
 355         }
 356
 357         tc->tc_next = timecounters;
 358         timecounters = tc;
 359         /*
 360          * Set up sysctl tree for this counter.
 361          */
 362         tc_root = SYSCTL_ADD_NODE(NULL,
 363             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 364             CTLFLAG_RW, 0, "timecounter description");
 365         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 366             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 367             "mask for implemented bits");
 368         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 369             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 370             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 371         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 372             "frequency", CTLTYPE_QUAD | CTLFLAG_RD, tc, sizeof(*tc),
 373              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 374         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 375             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 376             "goodness of time counter");
 377         /*
 378          * Never automatically use a timecounter with negative quality.
 379          * Even though we run on the dummy counter, switching here may be
 380          * worse since this timecounter may not be monotonous.
 381          */
 382         if (tc->tc_quality < 0)
 383                 return;
 384         if (tc->tc_quality < timecounter->tc_quality)
 385                 return;
 386         if (tc->tc_quality == timecounter->tc_quality &&
 387             tc->tc_frequency < timecounter->tc_frequency)
 388                 return;
 389         (void)tc->tc_get_timecount(tc);
 390         (void)tc->tc_get_timecount(tc);
 391         timecounter = tc;
 392 }
 393
 394 /* Report the frequency of the current timecounter. */
 395 u_int64_t
 396 tc_getfrequency(void)
 397 {
 398
 399         return (timehands->th_counter->tc_frequency);
 400 }
 401
 402 /*
 403  * Step our concept of UTC.  This is done by modifying our estimate of
 404  * when we booted.
 405  * XXX: not locked.
 406  */
 407 void
 408 tc_setclock(struct timespec *ts)
 409 {
 410         struct timespec tbef, taft;
 411         struct bintime bt, bt2;
 412
 413         cpu_tick_calibrate(1);
 414         TC_COUNT(nsetclock);
 415         nanotime(&tbef);
 416         timespec2bintime(ts, &bt);
 417         binuptime(&bt2);
 418         bintime_sub(&bt, &bt2);
 419         bintime_add(&bt2, &boottimebin);
 420         boottimebin = bt;
 421         bintime2timeval(&bt, &boottime);
 422
 423         /* XXX fiddle all the little crinkly bits around the fiords... */
 424         tc_windup();
 425         nanotime(&taft);
 426         if (timestepwarnings) {
 427                 log(LOG_INFO,
 428                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 429                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 430                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 431                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 432         }
 433         cpu_tick_calibrate(1);
 434 }
 435
 436 /*
 437  * Initialize the next struct timehands in the ring and make
 438  * it the active timehands.  Along the way we might switch to a different
 439  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 440  */
 441 static void
 442 tc_windup(void)
 443 {
 444         struct bintime bt;
 445         struct timehands *th, *tho;
 446         u_int64_t scale;
 447         u_int delta, ncount, ogen;
 448         int i;
 449         time_t t;
 450
 451         /*
 452          * Make the next timehands a copy of the current one, but do not
 453          * overwrite the generation or next pointer.  While we update
 454          * the contents, the generation must be zero.
 455          */
 456         tho = timehands;
 457         th = tho->th_next;
 458         ogen = th->th_generation;
 459         th->th_generation = 0;
 460         bcopy(tho, th, offsetof(struct timehands, th_generation));
 461
 462         /*
 463          * Capture a timecounter delta on the current timecounter and if
 464          * changing timecounters, a counter value from the new timecounter.
 465          * Update the offset fields accordingly.
 466          */
 467         delta = tc_delta(th);
 468         if (th->th_counter != timecounter)
 469                 ncount = timecounter->tc_get_timecount(timecounter);
 470         else
 471                 ncount = 0;
 472         th->th_offset_count += delta;
 473         th->th_offset_count &= th->th_counter->tc_counter_mask;
 474         bintime_addx(&th->th_offset, th->th_scale * delta);
 475
 476         /*
 477          * Hardware latching timecounters may not generate interrupts on
 478          * PPS events, so instead we poll them.  There is a finite risk that
 479          * the hardware might capture a count which is later than the one we
 480          * got above, and therefore possibly in the next NTP second which might
 481          * have a different rate than the current NTP second.  It doesn't
 482          * matter in practice.
 483          */
 484         if (tho->th_counter->tc_poll_pps)
 485                 tho->th_counter->tc_poll_pps(tho->th_counter);
 486
 487         /*
 488          * Deal with NTP second processing.  The for loop normally
 489          * iterates at most once, but in extreme situations it might
 490          * keep NTP sane if timeouts are not run for several seconds.
 491          * At boot, the time step can be large when the TOD hardware
 492          * has been read, so on really large steps, we call
 493          * ntp_update_second only twice.  We need to call it twice in
 494          * case we missed a leap second.
 495          */
 496         bt = th->th_offset;
 497         bintime_add(&bt, &boottimebin);
 498         i = bt.sec - tho->th_microtime.tv_sec;
 499         if (i > LARGE_STEP)
 500                 i = 2;
 501         for (; i > 0; i--) {
 502                 t = bt.sec;
 503                 ntp_update_second(&th->th_adjustment, &bt.sec);
 504                 if (bt.sec != t)
 505                         boottimebin.sec += bt.sec - t;
 506         }
 507         /* Update the UTC timestamps used by the get*() functions. */
 508         /* XXX shouldn't do this here.  Should force non-`get' versions. */
 509         bintime2timeval(&bt, &th->th_microtime);
 510         bintime2timespec(&bt, &th->th_nanotime);
 511
 512         /* Now is a good time to change timecounters. */
 513         if (th->th_counter != timecounter) {
 514                 th->th_counter = timecounter;
 515                 th->th_offset_count = ncount;
 516         }
 517
 518         /*-
 519          * Recalculate the scaling factor.  We want the number of 1/2^64
 520          * fractions of a second per period of the hardware counter, taking
 521          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 522          * processing provides us with.
 523          *
 524          * The th_adjustment is nanoseconds per second with 32 bit binary
 525          * fraction and we want 64 bit binary fraction of second:
 526          *
 527          *       x = a * 2^32 / 10^9 = a * 4.294967296
 528          *
 529          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 530          * we can only multiply by about 850 without overflowing, that
 531          * leaves no suitably precise fractions for multiply before divide.
 532          *
 533          * Divide before multiply with a fraction of 2199/512 results in a
 534          * systematic undercompensation of 10PPM of th_adjustment.  On a
 535          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 536          *
 537          * We happily sacrifice the lowest of the 64 bits of our result
 538          * to the goddess of code clarity.
 539          *
 540          */
 541         scale = (u_int64_t)1 << 63;
 542         scale += (th->th_adjustment / 1024) * 2199;
 543         scale /= th->th_counter->tc_frequency;
 544         th->th_scale = scale * 2;
 545
 546         /*
 547          * Now that the struct timehands is again consistent, set the new
 548          * generation number, making sure to not make it zero.
 549          */
 550         if (++ogen == 0)
 551                 ogen = 1;
 552         th->th_generation = ogen;
 553
 554         /* Go live with the new struct timehands. */
 555         time_second = th->th_microtime.tv_sec;
 556         time_uptime = th->th_offset.sec;
 557         timehands = th;
 558 }
 559
 560 /* Report or change the active timecounter hardware. */
 561 static int
 562 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 563 {
 564         char newname[32];
 565         struct timecounter *newtc, *tc;
 566         int error;
 567
 568         tc = timecounter;
 569         strlcpy(newname, tc->tc_name, sizeof(newname));
 570
 571         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 572         if (error != 0 || req->newptr == NULL ||
 573             strcmp(newname, tc->tc_name) == 0)
 574                 return (error);
 575         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 576                 if (strcmp(newname, newtc->tc_name) != 0)
 577                         continue;
 578
 579                 /* Warm up new timecounter. */
 580                 (void)newtc->tc_get_timecount(newtc);
 581                 (void)newtc->tc_get_timecount(newtc);
 582
 583                 timecounter = newtc;
 584                 return (0);
 585         }
 586         return (EINVAL);
 587 }
 588
 589 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 590     0, 0, sysctl_kern_timecounter_hardware, "A", "");
 591
 592
 593 /* Report or change the active timecounter hardware. */
 594 static int
 595 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 596 {
 597         char buf[32], *spc;
 598         struct timecounter *tc;
 599         int error;
 600
 601         spc = "";
 602         error = 0;
 603         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
 604                 sprintf(buf, "%s%s(%d)",
 605                     spc, tc->tc_name, tc->tc_quality);
 606                 error = SYSCTL_OUT(req, buf, strlen(buf));
 607                 spc = " ";
 608         }
 609         return (error);
 610 }
 611
 612 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 613     0, 0, sysctl_kern_timecounter_choice, "A", "");
 614
 615 /*
 616  * RFC 2783 PPS-API implementation.
 617  */
 618
 619 int
 620 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 621 {
 622         pps_params_t *app;
 623         struct pps_fetch_args *fapi;
 624 #ifdef PPS_SYNC
 625         struct pps_kcbind_args *kapi;
 626 #endif
 627
 628         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 629         switch (cmd) {
 630         case PPS_IOC_CREATE:
 631                 return (0);
 632         case PPS_IOC_DESTROY:
 633                 return (0);
 634         case PPS_IOC_SETPARAMS:
 635                 app = (pps_params_t *)data;
 636                 if (app->mode & ~pps->ppscap)
 637                         return (EINVAL);
 638                 pps->ppsparam = *app;
 639                 return (0);
 640         case PPS_IOC_GETPARAMS:
 641                 app = (pps_params_t *)data;
 642                 *app = pps->ppsparam;
 643                 app->api_version = PPS_API_VERS_1;
 644                 return (0);
 645         case PPS_IOC_GETCAP:
 646                 *(int*)data = pps->ppscap;
 647                 return (0);
 648         case PPS_IOC_FETCH:
 649                 fapi = (struct pps_fetch_args *)data;
 650                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 651                         return (EINVAL);
 652                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
 653                         return (EOPNOTSUPP);
 654                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
 655                 fapi->pps_info_buf = pps->ppsinfo;
 656                 return (0);
 657         case PPS_IOC_KCBIND:
 658 #ifdef PPS_SYNC
 659                 kapi = (struct pps_kcbind_args *)data;
 660                 /* XXX Only root should be able to do this */
 661                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 662                         return (EINVAL);
 663                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 664                         return (EINVAL);
 665                 if (kapi->edge & ~pps->ppscap)
 666                         return (EINVAL);
 667                 pps->kcmode = kapi->edge;
 668                 return (0);
 669 #else
 670                 return (EOPNOTSUPP);
 671 #endif
 672         default:
 673                 return (ENOIOCTL);
 674         }
 675 }
 676
 677 void
 678 pps_init(struct pps_state *pps)
 679 {
 680         pps->ppscap |= PPS_TSFMT_TSPEC;
 681         if (pps->ppscap & PPS_CAPTUREASSERT)
 682                 pps->ppscap |= PPS_OFFSETASSERT;
 683         if (pps->ppscap & PPS_CAPTURECLEAR)
 684                 pps->ppscap |= PPS_OFFSETCLEAR;
 685 }
 686
 687 void
 688 pps_capture(struct pps_state *pps)
 689 {
 690         struct timehands *th;
 691
 692         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 693         th = timehands;
 694         pps->capgen = th->th_generation;
 695         pps->capth = th;
 696         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 697         if (pps->capgen != th->th_generation)
 698                 pps->capgen = 0;
 699 }
 700
 701 void
 702 pps_event(struct pps_state *pps, int event)
 703 {
 704         struct bintime bt;
 705         struct timespec ts, *tsp, *osp;
 706         u_int tcount, *pcount;
 707         int foff, fhard;
 708         pps_seq_t *pseq;
 709
 710         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 711         /* If the timecounter was wound up underneath us, bail out. */
 712         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
 713                 return;
 714
 715         /* Things would be easier with arrays. */
 716         if (event == PPS_CAPTUREASSERT) {
 717                 tsp = &pps->ppsinfo.assert_timestamp;
 718                 osp = &pps->ppsparam.assert_offset;
 719                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 720                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 721                 pcount = &pps->ppscount[0];
 722                 pseq = &pps->ppsinfo.assert_sequence;
 723         } else {
 724                 tsp = &pps->ppsinfo.clear_timestamp;
 725                 osp = &pps->ppsparam.clear_offset;
 726                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 727                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 728                 pcount = &pps->ppscount[1];
 729                 pseq = &pps->ppsinfo.clear_sequence;
 730         }
 731
 732         /*
 733          * If the timecounter changed, we cannot compare the count values, so
 734          * we have to drop the rest of the PPS-stuff until the next event.
 735          */
 736         if (pps->ppstc != pps->capth->th_counter) {
 737                 pps->ppstc = pps->capth->th_counter;
 738                 *pcount = pps->capcount;
 739                 pps->ppscount[2] = pps->capcount;
 740                 return;
 741         }
 742
 743         /* Convert the count to a timespec. */
 744         tcount = pps->capcount - pps->capth->th_offset_count;
 745         tcount &= pps->capth->th_counter->tc_counter_mask;
 746         bt = pps->capth->th_offset;
 747         bintime_addx(&bt, pps->capth->th_scale * tcount);
 748         bintime_add(&bt, &boottimebin);
 749         bintime2timespec(&bt, &ts);
 750
 751         /* If the timecounter was wound up underneath us, bail out. */
 752         if (pps->capgen != pps->capth->th_generation)
 753                 return;
 754
 755         *pcount = pps->capcount;
 756         (*pseq)++;
 757         *tsp = ts;
 758
 759         if (foff) {
 760                 timespecadd(tsp, osp);
 761                 if (tsp->tv_nsec < 0) {
 762                         tsp->tv_nsec += 1000000000;
 763                         tsp->tv_sec -= 1;
 764                 }
 765         }
 766 #ifdef PPS_SYNC
 767         if (fhard) {
 768                 u_int64_t scale;
 769
 770                 /*
 771                  * Feed the NTP PLL/FLL.
 772                  * The FLL wants to know how many (hardware) nanoseconds
 773                  * elapsed since the previous event.
 774                  */
 775                 tcount = pps->capcount - pps->ppscount[2];
 776                 pps->ppscount[2] = pps->capcount;
 777                 tcount &= pps->capth->th_counter->tc_counter_mask;
 778                 scale = (u_int64_t)1 << 63;
 779                 scale /= pps->capth->th_counter->tc_frequency;
 780                 scale *= 2;
 781                 bt.sec = 0;
 782                 bt.frac = 0;
 783                 bintime_addx(&bt, scale * tcount);
 784                 bintime2timespec(&bt, &ts);
 785                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 786         }
 787 #endif
 788 }
 789
 790 /*
 791  * Timecounters need to be updated every so often to prevent the hardware
 792  * counter from overflowing.  Updating also recalculates the cached values
 793  * used by the get*() family of functions, so their precision depends on
 794  * the update frequency.
 795  */
 796
 797 static int tc_tick;
 798 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
 799
 800 void
 801 tc_ticktock(void)
 802 {
 803         static int count;
 804         static time_t last_calib;
 805
 806         if (++count < tc_tick)
 807                 return;
 808         count = 0;
 809         tc_windup();
 810         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 811                 cpu_tick_calibrate(0);
 812                 last_calib = time_uptime;
 813         }
 814 }
 815
 816 static void
 817 inittimecounter(void *dummy)
 818 {
 819         u_int p;
 820
 821         /*
 822          * Set the initial timeout to
 823          * max(1, <approx. number of hardclock ticks in a millisecond>).
 824          * People should probably not use the sysctl to set the timeout
 825          * to smaller than its inital value, since that value is the
 826          * smallest reasonable one.  If they want better timestamps they
 827          * should use the non-"get"* functions.
 828          */
 829         if (hz > 1000)
 830                 tc_tick = (hz + 500) / 1000;
 831         else
 832                 tc_tick = 1;
 833         p = (tc_tick * 1000000) / hz;
 834         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 835
 836         /* warm up new timecounter (again) and get rolling. */
 837         (void)timecounter->tc_get_timecount(timecounter);
 838         (void)timecounter->tc_get_timecount(timecounter);
 839 }
 840
 841 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 842
 843 /* Cpu tick handling -------------------------------------------------*/
 844
 845 static int cpu_tick_variable;
 846 static uint64_t cpu_tick_frequency;
 847
 848 static uint64_t
 849 tc_cpu_ticks(void)
 850 {
 851         static uint64_t base;
 852         static unsigned last;
 853         unsigned u;
 854         struct timecounter *tc;
 855
 856         tc = timehands->th_counter;
 857         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 858         if (u < last)
 859                 base += (uint64_t)tc->tc_counter_mask + 1;
 860         last = u;
 861         return (u + base);
 862 }
 863
 864 /*
 865  * This function gets called ever 16 seconds on only one designated
 866  * CPU in the system from hardclock() via tc_ticktock().
 867  *
 868  * Whenever the real time clock is stepped we get called with reset=1
 869  * to make sure we handle suspend/resume and similar events correctly.
 870  */
 871
 872 static void
 873 cpu_tick_calibrate(int reset)
 874 {
 875         static uint64_t c_last;
 876         uint64_t c_this, c_delta;
 877         static struct bintime  t_last;
 878         struct bintime t_this, t_delta;
 879         uint32_t divi;
 880
 881         if (reset) {
 882                 /* The clock was stepped, abort & reset */
 883                 t_last.sec = 0;
 884                 return;
 885         }
 886
 887         /* we don't calibrate fixed rate cputicks */
 888         if (!cpu_tick_variable)
 889                 return;
 890
 891         getbinuptime(&t_this);
 892         c_this = cpu_ticks();
 893         if (t_last.sec != 0) {
 894                 c_delta = c_this - c_last;
 895                 t_delta = t_this;
 896                 bintime_sub(&t_delta, &t_last);
 897                 /*
 898                  * Validate that 16 +/- 1/256 seconds passed.
 899                  * After division by 16 this gives us a precision of
 900                  * roughly 250PPM which is sufficient
 901                  */
 902                 if (t_delta.sec > 16 || (
 903                     t_delta.sec == 16 && t_delta.frac >= (0x01LL << 56))) {
 904                         /* too long */
 905                         if (bootverbose)
 906                                 printf("t_delta %ju.%016jx too long\n",
 907                                     (uintmax_t)t_delta.sec,
 908                                     (uintmax_t)t_delta.frac);
 909                 } else if (t_delta.sec < 15 ||
 910                     (t_delta.sec == 15 && t_delta.frac <= (0xffLL << 56))) {
 911                         /* too short */
 912                         if (bootverbose)
 913                                 printf("t_delta %ju.%016jx too short\n",
 914                                     (uintmax_t)t_delta.sec,
 915                                     (uintmax_t)t_delta.frac);
 916                 } else {
 917                         /* just right */
 918                         /*
 919                          * Headroom:
 920                          *      2^(64-20) / 16[s] =
 921                          *      2^(44) / 16[s] =
 922                          *      17.592.186.044.416 / 16 =
 923                          *      1.099.511.627.776 [Hz]
 924                          */
 925                         divi = t_delta.sec << 20;
 926                         divi |= t_delta.frac >> (64 - 20);
 927                         c_delta <<= 20;
 928                         c_delta /= divi;
 929                         if (c_delta  > cpu_tick_frequency) {
 930                                 if (0 && bootverbose)
 931                                         printf("cpu_tick increased to %ju Hz\n",
 932                                             c_delta);
 933                                 cpu_tick_frequency = c_delta;
 934                         }
 935                 }
 936         }
 937         c_last = c_this;
 938         t_last = t_this;
 939 }
 940
 941 void
 942 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 943 {
 944
 945         if (func == NULL) {
 946                 cpu_ticks = tc_cpu_ticks;
 947         } else {
 948                 cpu_tick_frequency = freq;
 949                 cpu_tick_variable = var;
 950                 cpu_ticks = func;
 951         }
 952 }
 953
 954 uint64_t
 955 cpu_tickrate(void)
 956 {
 957
 958         if (cpu_ticks == tc_cpu_ticks)
 959                 return (tc_getfrequency());
 960         return (cpu_tick_frequency);
 961 }
 962
 963 /*
 964  * We need to be slightly careful converting cputicks to microseconds.
 965  * There is plenty of margin in 64 bits of microseconds (half a million
 966  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 967  * before divide conversion (to retain precision) we find that the
 968  * margin shrinks to 1.5 hours (one millionth of 146y).
 969  * With a three prong approach we never lose significant bits, no
 970  * matter what the cputick rate and length of timeinterval is.
 971  */
 972
 973 uint64_t
 974 cputick2usec(uint64_t tick)
 975 {
 976
 977         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 978                 return (tick / (cpu_tickrate() / 1000000LL));
 979         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 980                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 981         else
 982                 return ((tick * 1000000LL) / cpu_tickrate());
 983 }
 984
 985 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;