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_compat.h"
  14 #include "opt_ntp.h"
  15
  16 #include <sys/param.h>
  17 #include <sys/kernel.h>
  18 #include <sys/sysctl.h>
  19 #include <sys/syslog.h>
  20 #include <sys/systm.h>
  21 #include <sys/timepps.h>
  22 #include <sys/timetc.h>
  23 #include <sys/timex.h>
  24 #include <sys/vdso.h>
  25
  26 /*
  27  * A large step happens on boot.  This constant detects such steps.
  28  * It is relatively small so that ntp_update_second gets called enough
  29  * in the typical 'missed a couple of seconds' case, but doesn't loop
  30  * forever when the time step is large.
  31  */
  32 #define LARGE_STEP      200
  33
  34 /*
  35  * Implement a dummy timecounter which we can use until we get a real one
  36  * in the air.  This allows the console and other early stuff to use
  37  * time services.
  38  */
  39
  40 static u_int
  41 dummy_get_timecount(struct timecounter *tc)
  42 {
  43         static u_int now;
  44
  45         return (++now);
  46 }
  47
  48 static struct timecounter dummy_timecounter = {
  49         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
  50 };
  51
  52 struct timehands {
  53         /* These fields must be initialized by the driver. */
  54         struct timecounter      *th_counter;
  55         int64_t                 th_adjustment;
  56         uint64_t                th_scale;
  57         u_int                   th_offset_count;
  58         struct bintime          th_offset;
  59         struct timeval          th_microtime;
  60         struct timespec         th_nanotime;
  61         /* Fields not to be copied in tc_windup start with th_generation. */
  62         volatile u_int          th_generation;
  63         struct timehands        *th_next;
  64 };
  65
  66 static struct timehands th0;
  67 static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0};
  68 static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9};
  69 static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8};
  70 static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7};
  71 static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6};
  72 static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5};
  73 static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4};
  74 static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3};
  75 static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2};
  76 static struct timehands th0 = {
  77         &dummy_timecounter,
  78         0,
  79         (uint64_t)-1 / 1000000,
  80         0,
  81         {1, 0},
  82         {0, 0},
  83         {0, 0},
  84         1,
  85         &th1
  86 };
  87
  88 static struct timehands *volatile timehands = &th0;
  89 struct timecounter *timecounter = &dummy_timecounter;
  90 static struct timecounter *timecounters = &dummy_timecounter;
  91
  92 int tc_min_ticktock_freq = 1;
  93
  94 time_t time_second = 1;
  95 time_t time_uptime = 1;
  96
  97 struct bintime boottimebin;
  98 struct timeval boottime;
  99 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
 100 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
 101     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
 102
 103 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
 104 SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
 105
 106 static int timestepwarnings;
 107 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
 108     &timestepwarnings, 0, "Log time steps");
 109
 110 static void tc_windup(void);
 111 static void cpu_tick_calibrate(int);
 112
 113 static int
 114 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 115 {
 116 #ifdef SCTL_MASK32
 117         int tv[2];
 118
 119         if (req->flags & SCTL_MASK32) {
 120                 tv[0] = boottime.tv_sec;
 121                 tv[1] = boottime.tv_usec;
 122                 return SYSCTL_OUT(req, tv, sizeof(tv));
 123         } else
 124 #endif
 125                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
 126 }
 127
 128 static int
 129 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 130 {
 131         u_int ncount;
 132         struct timecounter *tc = arg1;
 133
 134         ncount = tc->tc_get_timecount(tc);
 135         return sysctl_handle_int(oidp, &ncount, 0, req);
 136 }
 137
 138 static int
 139 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 140 {
 141         uint64_t freq;
 142         struct timecounter *tc = arg1;
 143
 144         freq = tc->tc_frequency;
 145         return sysctl_handle_64(oidp, &freq, 0, req);
 146 }
 147
 148 /*
 149  * Return the difference between the timehands' counter value now and what
 150  * was when we copied it to the timehands' offset_count.
 151  */
 152 static __inline u_int
 153 tc_delta(struct timehands *th)
 154 {
 155         struct timecounter *tc;
 156
 157         tc = th->th_counter;
 158         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 159             tc->tc_counter_mask);
 160 }
 161
 162 /*
 163  * Functions for reading the time.  We have to loop until we are sure that
 164  * the timehands that we operated on was not updated under our feet.  See
 165  * the comment in <sys/time.h> for a description of these 12 functions.
 166  */
 167
 168 void
 169 binuptime(struct bintime *bt)
 170 {
 171         struct timehands *th;
 172         u_int gen;
 173
 174         do {
 175                 th = timehands;
 176                 gen = th->th_generation;
 177                 *bt = th->th_offset;
 178                 bintime_addx(bt, th->th_scale * tc_delta(th));
 179         } while (gen == 0 || gen != th->th_generation);
 180 }
 181
 182 void
 183 nanouptime(struct timespec *tsp)
 184 {
 185         struct bintime bt;
 186
 187         binuptime(&bt);
 188         bintime2timespec(&bt, tsp);
 189 }
 190
 191 void
 192 microuptime(struct timeval *tvp)
 193 {
 194         struct bintime bt;
 195
 196         binuptime(&bt);
 197         bintime2timeval(&bt, tvp);
 198 }
 199
 200 void
 201 bintime(struct bintime *bt)
 202 {
 203
 204         binuptime(bt);
 205         bintime_add(bt, &boottimebin);
 206 }
 207
 208 void
 209 nanotime(struct timespec *tsp)
 210 {
 211         struct bintime bt;
 212
 213         bintime(&bt);
 214         bintime2timespec(&bt, tsp);
 215 }
 216
 217 void
 218 microtime(struct timeval *tvp)
 219 {
 220         struct bintime bt;
 221
 222         bintime(&bt);
 223         bintime2timeval(&bt, tvp);
 224 }
 225
 226 void
 227 getbinuptime(struct bintime *bt)
 228 {
 229         struct timehands *th;
 230         u_int gen;
 231
 232         do {
 233                 th = timehands;
 234                 gen = th->th_generation;
 235                 *bt = th->th_offset;
 236         } while (gen == 0 || gen != th->th_generation);
 237 }
 238
 239 void
 240 getnanouptime(struct timespec *tsp)
 241 {
 242         struct timehands *th;
 243         u_int gen;
 244
 245         do {
 246                 th = timehands;
 247                 gen = th->th_generation;
 248                 bintime2timespec(&th->th_offset, tsp);
 249         } while (gen == 0 || gen != th->th_generation);
 250 }
 251
 252 void
 253 getmicrouptime(struct timeval *tvp)
 254 {
 255         struct timehands *th;
 256         u_int gen;
 257
 258         do {
 259                 th = timehands;
 260                 gen = th->th_generation;
 261                 bintime2timeval(&th->th_offset, tvp);
 262         } while (gen == 0 || gen != th->th_generation);
 263 }
 264
 265 void
 266 getbintime(struct bintime *bt)
 267 {
 268         struct timehands *th;
 269         u_int gen;
 270
 271         do {
 272                 th = timehands;
 273                 gen = th->th_generation;
 274                 *bt = th->th_offset;
 275         } while (gen == 0 || gen != th->th_generation);
 276         bintime_add(bt, &boottimebin);
 277 }
 278
 279 void
 280 getnanotime(struct timespec *tsp)
 281 {
 282         struct timehands *th;
 283         u_int gen;
 284
 285         do {
 286                 th = timehands;
 287                 gen = th->th_generation;
 288                 *tsp = th->th_nanotime;
 289         } while (gen == 0 || gen != th->th_generation);
 290 }
 291
 292 void
 293 getmicrotime(struct timeval *tvp)
 294 {
 295         struct timehands *th;
 296         u_int gen;
 297
 298         do {
 299                 th = timehands;
 300                 gen = th->th_generation;
 301                 *tvp = th->th_microtime;
 302         } while (gen == 0 || gen != th->th_generation);
 303 }
 304
 305 /*
 306  * Initialize a new timecounter and possibly use it.
 307  */
 308 void
 309 tc_init(struct timecounter *tc)
 310 {
 311         u_int u;
 312         struct sysctl_oid *tc_root;
 313
 314         u = tc->tc_frequency / tc->tc_counter_mask;
 315         /* XXX: We need some margin here, 10% is a guess */
 316         u *= 11;
 317         u /= 10;
 318         if (u > hz && tc->tc_quality >= 0) {
 319                 tc->tc_quality = -2000;
 320                 if (bootverbose) {
 321                         printf("Timecounter \"%s\" frequency %ju Hz",
 322                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 323                         printf(" -- Insufficient hz, needs at least %u\n", u);
 324                 }
 325         } else if (tc->tc_quality >= 0 || bootverbose) {
 326                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 327                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 328                     tc->tc_quality);
 329         }
 330
 331         tc->tc_next = timecounters;
 332         timecounters = tc;
 333         /*
 334          * Set up sysctl tree for this counter.
 335          */
 336         tc_root = SYSCTL_ADD_NODE(NULL,
 337             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 338             CTLFLAG_RW, 0, "timecounter description");
 339         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 340             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 341             "mask for implemented bits");
 342         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 343             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 344             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 345         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 346             "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
 347              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 348         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 349             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 350             "goodness of time counter");
 351         /*
 352          * Never automatically use a timecounter with negative quality.
 353          * Even though we run on the dummy counter, switching here may be
 354          * worse since this timecounter may not be monotonous.
 355          */
 356         if (tc->tc_quality < 0)
 357                 return;
 358         if (tc->tc_quality < timecounter->tc_quality)
 359                 return;
 360         if (tc->tc_quality == timecounter->tc_quality &&
 361             tc->tc_frequency < timecounter->tc_frequency)
 362                 return;
 363         (void)tc->tc_get_timecount(tc);
 364         (void)tc->tc_get_timecount(tc);
 365         timecounter = tc;
 366 }
 367
 368 /* Report the frequency of the current timecounter. */
 369 uint64_t
 370 tc_getfrequency(void)
 371 {
 372
 373         return (timehands->th_counter->tc_frequency);
 374 }
 375
 376 /*
 377  * Step our concept of UTC.  This is done by modifying our estimate of
 378  * when we booted.
 379  * XXX: not locked.
 380  */
 381 void
 382 tc_setclock(struct timespec *ts)
 383 {
 384         struct timespec tbef, taft;
 385         struct bintime bt, bt2;
 386
 387         cpu_tick_calibrate(1);
 388         nanotime(&tbef);
 389         timespec2bintime(ts, &bt);
 390         binuptime(&bt2);
 391         bintime_sub(&bt, &bt2);
 392         bintime_add(&bt2, &boottimebin);
 393         boottimebin = bt;
 394         bintime2timeval(&bt, &boottime);
 395
 396         /* XXX fiddle all the little crinkly bits around the fiords... */
 397         tc_windup();
 398         nanotime(&taft);
 399         if (timestepwarnings) {
 400                 log(LOG_INFO,
 401                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 402                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 403                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 404                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 405         }
 406         cpu_tick_calibrate(1);
 407 }
 408
 409 /*
 410  * Initialize the next struct timehands in the ring and make
 411  * it the active timehands.  Along the way we might switch to a different
 412  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 413  */
 414 static void
 415 tc_windup(void)
 416 {
 417         struct bintime bt;
 418         struct timehands *th, *tho;
 419         uint64_t scale;
 420         u_int delta, ncount, ogen;
 421         int i;
 422         time_t t;
 423
 424         /*
 425          * Make the next timehands a copy of the current one, but do not
 426          * overwrite the generation or next pointer.  While we update
 427          * the contents, the generation must be zero.
 428          */
 429         tho = timehands;
 430         th = tho->th_next;
 431         ogen = th->th_generation;
 432         th->th_generation = 0;
 433         bcopy(tho, th, offsetof(struct timehands, th_generation));
 434
 435         /*
 436          * Capture a timecounter delta on the current timecounter and if
 437          * changing timecounters, a counter value from the new timecounter.
 438          * Update the offset fields accordingly.
 439          */
 440         delta = tc_delta(th);
 441         if (th->th_counter != timecounter)
 442                 ncount = timecounter->tc_get_timecount(timecounter);
 443         else
 444                 ncount = 0;
 445         th->th_offset_count += delta;
 446         th->th_offset_count &= th->th_counter->tc_counter_mask;
 447         while (delta > th->th_counter->tc_frequency) {
 448                 /* Eat complete unadjusted seconds. */
 449                 delta -= th->th_counter->tc_frequency;
 450                 th->th_offset.sec++;
 451         }
 452         if ((delta > th->th_counter->tc_frequency / 2) &&
 453             (th->th_scale * delta < ((uint64_t)1 << 63))) {
 454                 /* The product th_scale * delta just barely overflows. */
 455                 th->th_offset.sec++;
 456         }
 457         bintime_addx(&th->th_offset, th->th_scale * delta);
 458
 459         /*
 460          * Hardware latching timecounters may not generate interrupts on
 461          * PPS events, so instead we poll them.  There is a finite risk that
 462          * the hardware might capture a count which is later than the one we
 463          * got above, and therefore possibly in the next NTP second which might
 464          * have a different rate than the current NTP second.  It doesn't
 465          * matter in practice.
 466          */
 467         if (tho->th_counter->tc_poll_pps)
 468                 tho->th_counter->tc_poll_pps(tho->th_counter);
 469
 470         /*
 471          * Deal with NTP second processing.  The for loop normally
 472          * iterates at most once, but in extreme situations it might
 473          * keep NTP sane if timeouts are not run for several seconds.
 474          * At boot, the time step can be large when the TOD hardware
 475          * has been read, so on really large steps, we call
 476          * ntp_update_second only twice.  We need to call it twice in
 477          * case we missed a leap second.
 478          */
 479         bt = th->th_offset;
 480         bintime_add(&bt, &boottimebin);
 481         i = bt.sec - tho->th_microtime.tv_sec;
 482         if (i > LARGE_STEP)
 483                 i = 2;
 484         for (; i > 0; i--) {
 485                 t = bt.sec;
 486                 ntp_update_second(&th->th_adjustment, &bt.sec);
 487                 if (bt.sec != t)
 488                         boottimebin.sec += bt.sec - t;
 489         }
 490         /* Update the UTC timestamps used by the get*() functions. */
 491         /* XXX shouldn't do this here.  Should force non-`get' versions. */
 492         bintime2timeval(&bt, &th->th_microtime);
 493         bintime2timespec(&bt, &th->th_nanotime);
 494
 495         /* Now is a good time to change timecounters. */
 496         if (th->th_counter != timecounter) {
 497 #ifndef __arm__
 498                 if ((timecounter->tc_flags & TC_FLAGS_C3STOP) != 0)
 499                         cpu_disable_deep_sleep++;
 500                 if ((th->th_counter->tc_flags & TC_FLAGS_C3STOP) != 0)
 501                         cpu_disable_deep_sleep--;
 502 #endif
 503                 th->th_counter = timecounter;
 504                 th->th_offset_count = ncount;
 505                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
 506                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
 507         }
 508
 509         /*-
 510          * Recalculate the scaling factor.  We want the number of 1/2^64
 511          * fractions of a second per period of the hardware counter, taking
 512          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 513          * processing provides us with.
 514          *
 515          * The th_adjustment is nanoseconds per second with 32 bit binary
 516          * fraction and we want 64 bit binary fraction of second:
 517          *
 518          *       x = a * 2^32 / 10^9 = a * 4.294967296
 519          *
 520          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 521          * we can only multiply by about 850 without overflowing, that
 522          * leaves no suitably precise fractions for multiply before divide.
 523          *
 524          * Divide before multiply with a fraction of 2199/512 results in a
 525          * systematic undercompensation of 10PPM of th_adjustment.  On a
 526          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 527          *
 528          * We happily sacrifice the lowest of the 64 bits of our result
 529          * to the goddess of code clarity.
 530          *
 531          */
 532         scale = (uint64_t)1 << 63;
 533         scale += (th->th_adjustment / 1024) * 2199;
 534         scale /= th->th_counter->tc_frequency;
 535         th->th_scale = scale * 2;
 536
 537         /*
 538          * Now that the struct timehands is again consistent, set the new
 539          * generation number, making sure to not make it zero.
 540          */
 541         if (++ogen == 0)
 542                 ogen = 1;
 543         th->th_generation = ogen;
 544
 545         /* Go live with the new struct timehands. */
 546         time_second = th->th_microtime.tv_sec;
 547         time_uptime = th->th_offset.sec;
 548         timehands = th;
 549         timekeep_push_vdso();
 550 }
 551
 552 /* Report or change the active timecounter hardware. */
 553 static int
 554 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 555 {
 556         char newname[32];
 557         struct timecounter *newtc, *tc;
 558         int error;
 559
 560         tc = timecounter;
 561         strlcpy(newname, tc->tc_name, sizeof(newname));
 562
 563         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 564         if (error != 0 || req->newptr == NULL ||
 565             strcmp(newname, tc->tc_name) == 0)
 566                 return (error);
 567         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 568                 if (strcmp(newname, newtc->tc_name) != 0)
 569                         continue;
 570
 571                 /* Warm up new timecounter. */
 572                 (void)newtc->tc_get_timecount(newtc);
 573                 (void)newtc->tc_get_timecount(newtc);
 574
 575                 timecounter = newtc;
 576                 timekeep_push_vdso();
 577                 return (0);
 578         }
 579         return (EINVAL);
 580 }
 581
 582 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 583     0, 0, sysctl_kern_timecounter_hardware, "A",
 584     "Timecounter hardware selected");
 585
 586
 587 /* Report or change the active timecounter hardware. */
 588 static int
 589 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 590 {
 591         char buf[32], *spc;
 592         struct timecounter *tc;
 593         int error;
 594
 595         spc = "";
 596         error = 0;
 597         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
 598                 sprintf(buf, "%s%s(%d)",
 599                     spc, tc->tc_name, tc->tc_quality);
 600                 error = SYSCTL_OUT(req, buf, strlen(buf));
 601                 spc = " ";
 602         }
 603         return (error);
 604 }
 605
 606 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 607     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
 608
 609 /*
 610  * RFC 2783 PPS-API implementation.
 611  */
 612
 613 int
 614 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 615 {
 616         pps_params_t *app;
 617         struct pps_fetch_args *fapi;
 618 #ifdef PPS_SYNC
 619         struct pps_kcbind_args *kapi;
 620 #endif
 621
 622         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 623         switch (cmd) {
 624         case PPS_IOC_CREATE:
 625                 return (0);
 626         case PPS_IOC_DESTROY:
 627                 return (0);
 628         case PPS_IOC_SETPARAMS:
 629                 app = (pps_params_t *)data;
 630                 if (app->mode & ~pps->ppscap)
 631                         return (EINVAL);
 632                 pps->ppsparam = *app;
 633                 return (0);
 634         case PPS_IOC_GETPARAMS:
 635                 app = (pps_params_t *)data;
 636                 *app = pps->ppsparam;
 637                 app->api_version = PPS_API_VERS_1;
 638                 return (0);
 639         case PPS_IOC_GETCAP:
 640                 *(int*)data = pps->ppscap;
 641                 return (0);
 642         case PPS_IOC_FETCH:
 643                 fapi = (struct pps_fetch_args *)data;
 644                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 645                         return (EINVAL);
 646                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
 647                         return (EOPNOTSUPP);
 648                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
 649                 fapi->pps_info_buf = pps->ppsinfo;
 650                 return (0);
 651         case PPS_IOC_KCBIND:
 652 #ifdef PPS_SYNC
 653                 kapi = (struct pps_kcbind_args *)data;
 654                 /* XXX Only root should be able to do this */
 655                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 656                         return (EINVAL);
 657                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 658                         return (EINVAL);
 659                 if (kapi->edge & ~pps->ppscap)
 660                         return (EINVAL);
 661                 pps->kcmode = kapi->edge;
 662                 return (0);
 663 #else
 664                 return (EOPNOTSUPP);
 665 #endif
 666         default:
 667                 return (ENOIOCTL);
 668         }
 669 }
 670
 671 void
 672 pps_init(struct pps_state *pps)
 673 {
 674         pps->ppscap |= PPS_TSFMT_TSPEC;
 675         if (pps->ppscap & PPS_CAPTUREASSERT)
 676                 pps->ppscap |= PPS_OFFSETASSERT;
 677         if (pps->ppscap & PPS_CAPTURECLEAR)
 678                 pps->ppscap |= PPS_OFFSETCLEAR;
 679 }
 680
 681 void
 682 pps_capture(struct pps_state *pps)
 683 {
 684         struct timehands *th;
 685
 686         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 687         th = timehands;
 688         pps->capgen = th->th_generation;
 689         pps->capth = th;
 690         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 691         if (pps->capgen != th->th_generation)
 692                 pps->capgen = 0;
 693 }
 694
 695 void
 696 pps_event(struct pps_state *pps, int event)
 697 {
 698         struct bintime bt;
 699         struct timespec ts, *tsp, *osp;
 700         u_int tcount, *pcount;
 701         int foff, fhard;
 702         pps_seq_t *pseq;
 703
 704         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 705         /* If the timecounter was wound up underneath us, bail out. */
 706         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
 707                 return;
 708
 709         /* Things would be easier with arrays. */
 710         if (event == PPS_CAPTUREASSERT) {
 711                 tsp = &pps->ppsinfo.assert_timestamp;
 712                 osp = &pps->ppsparam.assert_offset;
 713                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 714                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 715                 pcount = &pps->ppscount[0];
 716                 pseq = &pps->ppsinfo.assert_sequence;
 717         } else {
 718                 tsp = &pps->ppsinfo.clear_timestamp;
 719                 osp = &pps->ppsparam.clear_offset;
 720                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 721                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 722                 pcount = &pps->ppscount[1];
 723                 pseq = &pps->ppsinfo.clear_sequence;
 724         }
 725
 726         /*
 727          * If the timecounter changed, we cannot compare the count values, so
 728          * we have to drop the rest of the PPS-stuff until the next event.
 729          */
 730         if (pps->ppstc != pps->capth->th_counter) {
 731                 pps->ppstc = pps->capth->th_counter;
 732                 *pcount = pps->capcount;
 733                 pps->ppscount[2] = pps->capcount;
 734                 return;
 735         }
 736
 737         /* Convert the count to a timespec. */
 738         tcount = pps->capcount - pps->capth->th_offset_count;
 739         tcount &= pps->capth->th_counter->tc_counter_mask;
 740         bt = pps->capth->th_offset;
 741         bintime_addx(&bt, pps->capth->th_scale * tcount);
 742         bintime_add(&bt, &boottimebin);
 743         bintime2timespec(&bt, &ts);
 744
 745         /* If the timecounter was wound up underneath us, bail out. */
 746         if (pps->capgen != pps->capth->th_generation)
 747                 return;
 748
 749         *pcount = pps->capcount;
 750         (*pseq)++;
 751         *tsp = ts;
 752
 753         if (foff) {
 754                 timespecadd(tsp, osp);
 755                 if (tsp->tv_nsec < 0) {
 756                         tsp->tv_nsec += 1000000000;
 757                         tsp->tv_sec -= 1;
 758                 }
 759         }
 760 #ifdef PPS_SYNC
 761         if (fhard) {
 762                 uint64_t scale;
 763
 764                 /*
 765                  * Feed the NTP PLL/FLL.
 766                  * The FLL wants to know how many (hardware) nanoseconds
 767                  * elapsed since the previous event.
 768                  */
 769                 tcount = pps->capcount - pps->ppscount[2];
 770                 pps->ppscount[2] = pps->capcount;
 771                 tcount &= pps->capth->th_counter->tc_counter_mask;
 772                 scale = (uint64_t)1 << 63;
 773                 scale /= pps->capth->th_counter->tc_frequency;
 774                 scale *= 2;
 775                 bt.sec = 0;
 776                 bt.frac = 0;
 777                 bintime_addx(&bt, scale * tcount);
 778                 bintime2timespec(&bt, &ts);
 779                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 780         }
 781 #endif
 782 }
 783
 784 /*
 785  * Timecounters need to be updated every so often to prevent the hardware
 786  * counter from overflowing.  Updating also recalculates the cached values
 787  * used by the get*() family of functions, so their precision depends on
 788  * the update frequency.
 789  */
 790
 791 static int tc_tick;
 792 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
 793     "Approximate number of hardclock ticks in a millisecond");
 794
 795 void
 796 tc_ticktock(int cnt)
 797 {
 798         static int count;
 799
 800         count += cnt;
 801         if (count < tc_tick)
 802                 return;
 803         count = 0;
 804         tc_windup();
 805 }
 806
 807 static void
 808 inittimecounter(void *dummy)
 809 {
 810         u_int p;
 811
 812         /*
 813          * Set the initial timeout to
 814          * max(1, <approx. number of hardclock ticks in a millisecond>).
 815          * People should probably not use the sysctl to set the timeout
 816          * to smaller than its inital value, since that value is the
 817          * smallest reasonable one.  If they want better timestamps they
 818          * should use the non-"get"* functions.
 819          */
 820         if (hz > 1000)
 821                 tc_tick = (hz + 500) / 1000;
 822         else
 823                 tc_tick = 1;
 824         p = (tc_tick * 1000000) / hz;
 825         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 826
 827         /* warm up new timecounter (again) and get rolling. */
 828         (void)timecounter->tc_get_timecount(timecounter);
 829         (void)timecounter->tc_get_timecount(timecounter);
 830         tc_windup();
 831 }
 832
 833 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 834
 835 /* Cpu tick handling -------------------------------------------------*/
 836
 837 static int cpu_tick_variable;
 838 static uint64_t cpu_tick_frequency;
 839
 840 static uint64_t
 841 tc_cpu_ticks(void)
 842 {
 843         static uint64_t base;
 844         static unsigned last;
 845         unsigned u;
 846         struct timecounter *tc;
 847
 848         tc = timehands->th_counter;
 849         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 850         if (u < last)
 851                 base += (uint64_t)tc->tc_counter_mask + 1;
 852         last = u;
 853         return (u + base);
 854 }
 855
 856 void
 857 cpu_tick_calibration(void)
 858 {
 859         static time_t last_calib;
 860
 861         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 862                 cpu_tick_calibrate(0);
 863                 last_calib = time_uptime;
 864         }
 865 }
 866
 867 /*
 868  * This function gets called every 16 seconds on only one designated
 869  * CPU in the system from hardclock() via cpu_tick_calibration()().
 870  *
 871  * Whenever the real time clock is stepped we get called with reset=1
 872  * to make sure we handle suspend/resume and similar events correctly.
 873  */
 874
 875 static void
 876 cpu_tick_calibrate(int reset)
 877 {
 878         static uint64_t c_last;
 879         uint64_t c_this, c_delta;
 880         static struct bintime  t_last;
 881         struct bintime t_this, t_delta;
 882         uint32_t divi;
 883
 884         if (reset) {
 885                 /* The clock was stepped, abort & reset */
 886                 t_last.sec = 0;
 887                 return;
 888         }
 889
 890         /* we don't calibrate fixed rate cputicks */
 891         if (!cpu_tick_variable)
 892                 return;
 893
 894         getbinuptime(&t_this);
 895         c_this = cpu_ticks();
 896         if (t_last.sec != 0) {
 897                 c_delta = c_this - c_last;
 898                 t_delta = t_this;
 899                 bintime_sub(&t_delta, &t_last);
 900                 /*
 901                  * Headroom:
 902                  *      2^(64-20) / 16[s] =
 903                  *      2^(44) / 16[s] =
 904                  *      17.592.186.044.416 / 16 =
 905                  *      1.099.511.627.776 [Hz]
 906                  */
 907                 divi = t_delta.sec << 20;
 908                 divi |= t_delta.frac >> (64 - 20);
 909                 c_delta <<= 20;
 910                 c_delta /= divi;
 911                 if (c_delta > cpu_tick_frequency) {
 912                         if (0 && bootverbose)
 913                                 printf("cpu_tick increased to %ju Hz\n",
 914                                     c_delta);
 915                         cpu_tick_frequency = c_delta;
 916                 }
 917         }
 918         c_last = c_this;
 919         t_last = t_this;
 920 }
 921
 922 void
 923 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 924 {
 925
 926         if (func == NULL) {
 927                 cpu_ticks = tc_cpu_ticks;
 928         } else {
 929                 cpu_tick_frequency = freq;
 930                 cpu_tick_variable = var;
 931                 cpu_ticks = func;
 932         }
 933 }
 934
 935 uint64_t
 936 cpu_tickrate(void)
 937 {
 938
 939         if (cpu_ticks == tc_cpu_ticks)
 940                 return (tc_getfrequency());
 941         return (cpu_tick_frequency);
 942 }
 943
 944 /*
 945  * We need to be slightly careful converting cputicks to microseconds.
 946  * There is plenty of margin in 64 bits of microseconds (half a million
 947  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 948  * before divide conversion (to retain precision) we find that the
 949  * margin shrinks to 1.5 hours (one millionth of 146y).
 950  * With a three prong approach we never lose significant bits, no
 951  * matter what the cputick rate and length of timeinterval is.
 952  */
 953
 954 uint64_t
 955 cputick2usec(uint64_t tick)
 956 {
 957
 958         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 959                 return (tick / (cpu_tickrate() / 1000000LL));
 960         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 961                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 962         else
 963                 return ((tick * 1000000LL) / cpu_tickrate());
 964 }
 965
 966 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
 967
 968 static int vdso_th_enable = 1;
 969 static int
 970 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
 971 {
 972         int old_vdso_th_enable, error;
 973
 974         old_vdso_th_enable = vdso_th_enable;
 975         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
 976         if (error != 0)
 977                 return (error);
 978         vdso_th_enable = old_vdso_th_enable;
 979         timekeep_push_vdso();
 980         return (0);
 981 }
 982 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
 983     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 984     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
 985
 986 uint32_t
 987 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
 988 {
 989         struct timehands *th;
 990         uint32_t enabled;
 991
 992         th = timehands;
 993         vdso_th->th_algo = VDSO_TH_ALGO_1;
 994         vdso_th->th_scale = th->th_scale;
 995         vdso_th->th_offset_count = th->th_offset_count;
 996         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
 997         vdso_th->th_offset = th->th_offset;
 998         vdso_th->th_boottime = boottimebin;
 999         enabled = cpu_fill_vdso_timehands(vdso_th);
1000         if (!vdso_th_enable)
1001                 enabled = 0;
1002         return (enabled);
1003 }
1004
1005 #ifdef COMPAT_FREEBSD32
1006 uint32_t
1007 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
1008 {
1009         struct timehands *th;
1010         uint32_t enabled;
1011
1012         th = timehands;
1013         vdso_th32->th_algo = VDSO_TH_ALGO_1;
1014         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
1015         vdso_th32->th_offset_count = th->th_offset_count;
1016         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
1017         vdso_th32->th_offset.sec = th->th_offset.sec;
1018         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
1019         vdso_th32->th_boottime.sec = boottimebin.sec;
1020         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = boottimebin.frac;
1021         enabled = cpu_fill_vdso_timehands32(vdso_th32);
1022         if (!vdso_th_enable)
1023                 enabled = 0;
1024         return (enabled);
1025 }
1026 #endif