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 volatile time_t time_second = 1;
  95 volatile 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 static 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 void dtrace_getnanotime(struct timespec *tsp);
 114
 115 static int
 116 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
 117 {
 118 #ifdef SCTL_MASK32
 119         int tv[2];
 120
 121         if (req->flags & SCTL_MASK32) {
 122                 tv[0] = boottime.tv_sec;
 123                 tv[1] = boottime.tv_usec;
 124                 return SYSCTL_OUT(req, tv, sizeof(tv));
 125         } else
 126 #endif
 127                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
 128 }
 129
 130 static int
 131 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
 132 {
 133         u_int ncount;
 134         struct timecounter *tc = arg1;
 135
 136         ncount = tc->tc_get_timecount(tc);
 137         return sysctl_handle_int(oidp, &ncount, 0, req);
 138 }
 139
 140 static int
 141 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
 142 {
 143         uint64_t freq;
 144         struct timecounter *tc = arg1;
 145
 146         freq = tc->tc_frequency;
 147         return sysctl_handle_64(oidp, &freq, 0, req);
 148 }
 149
 150 /*
 151  * Return the difference between the timehands' counter value now and what
 152  * was when we copied it to the timehands' offset_count.
 153  */
 154 static __inline u_int
 155 tc_delta(struct timehands *th)
 156 {
 157         struct timecounter *tc;
 158
 159         tc = th->th_counter;
 160         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
 161             tc->tc_counter_mask);
 162 }
 163
 164 /*
 165  * Functions for reading the time.  We have to loop until we are sure that
 166  * the timehands that we operated on was not updated under our feet.  See
 167  * the comment in <sys/time.h> for a description of these 12 functions.
 168  */
 169
 170 void
 171 binuptime(struct bintime *bt)
 172 {
 173         struct timehands *th;
 174         u_int gen;
 175
 176         do {
 177                 th = timehands;
 178                 gen = th->th_generation;
 179                 *bt = th->th_offset;
 180                 bintime_addx(bt, th->th_scale * tc_delta(th));
 181         } while (gen == 0 || gen != th->th_generation);
 182 }
 183
 184 void
 185 nanouptime(struct timespec *tsp)
 186 {
 187         struct bintime bt;
 188
 189         binuptime(&bt);
 190         bintime2timespec(&bt, tsp);
 191 }
 192
 193 void
 194 microuptime(struct timeval *tvp)
 195 {
 196         struct bintime bt;
 197
 198         binuptime(&bt);
 199         bintime2timeval(&bt, tvp);
 200 }
 201
 202 void
 203 bintime(struct bintime *bt)
 204 {
 205
 206         binuptime(bt);
 207         bintime_add(bt, &boottimebin);
 208 }
 209
 210 void
 211 nanotime(struct timespec *tsp)
 212 {
 213         struct bintime bt;
 214
 215         bintime(&bt);
 216         bintime2timespec(&bt, tsp);
 217 }
 218
 219 void
 220 microtime(struct timeval *tvp)
 221 {
 222         struct bintime bt;
 223
 224         bintime(&bt);
 225         bintime2timeval(&bt, tvp);
 226 }
 227
 228 void
 229 getbinuptime(struct bintime *bt)
 230 {
 231         struct timehands *th;
 232         u_int gen;
 233
 234         do {
 235                 th = timehands;
 236                 gen = th->th_generation;
 237                 *bt = th->th_offset;
 238         } while (gen == 0 || gen != th->th_generation);
 239 }
 240
 241 void
 242 getnanouptime(struct timespec *tsp)
 243 {
 244         struct timehands *th;
 245         u_int gen;
 246
 247         do {
 248                 th = timehands;
 249                 gen = th->th_generation;
 250                 bintime2timespec(&th->th_offset, tsp);
 251         } while (gen == 0 || gen != th->th_generation);
 252 }
 253
 254 void
 255 getmicrouptime(struct timeval *tvp)
 256 {
 257         struct timehands *th;
 258         u_int gen;
 259
 260         do {
 261                 th = timehands;
 262                 gen = th->th_generation;
 263                 bintime2timeval(&th->th_offset, tvp);
 264         } while (gen == 0 || gen != th->th_generation);
 265 }
 266
 267 void
 268 getbintime(struct bintime *bt)
 269 {
 270         struct timehands *th;
 271         u_int gen;
 272
 273         do {
 274                 th = timehands;
 275                 gen = th->th_generation;
 276                 *bt = th->th_offset;
 277         } while (gen == 0 || gen != th->th_generation);
 278         bintime_add(bt, &boottimebin);
 279 }
 280
 281 void
 282 getnanotime(struct timespec *tsp)
 283 {
 284         struct timehands *th;
 285         u_int gen;
 286
 287         do {
 288                 th = timehands;
 289                 gen = th->th_generation;
 290                 *tsp = th->th_nanotime;
 291         } while (gen == 0 || gen != th->th_generation);
 292 }
 293
 294 void
 295 getmicrotime(struct timeval *tvp)
 296 {
 297         struct timehands *th;
 298         u_int gen;
 299
 300         do {
 301                 th = timehands;
 302                 gen = th->th_generation;
 303                 *tvp = th->th_microtime;
 304         } while (gen == 0 || gen != th->th_generation);
 305 }
 306
 307 /*
 308  * This is a clone of getnanotime and used for walltimestamps.
 309  * The dtrace_ prefix prevents fbt from creating probes for
 310  * it so walltimestamp can be safely used in all fbt probes.
 311  */
 312 void
 313 dtrace_getnanotime(struct timespec *tsp)
 314 {
 315         struct timehands *th;
 316         u_int gen;
 317
 318         do {
 319                 th = timehands;
 320                 gen = th->th_generation;
 321                 *tsp = th->th_nanotime;
 322         } while (gen == 0 || gen != th->th_generation);
 323 }
 324
 325 /*
 326  * Initialize a new timecounter and possibly use it.
 327  */
 328 void
 329 tc_init(struct timecounter *tc)
 330 {
 331         u_int u;
 332         struct sysctl_oid *tc_root;
 333
 334         u = tc->tc_frequency / tc->tc_counter_mask;
 335         /* XXX: We need some margin here, 10% is a guess */
 336         u *= 11;
 337         u /= 10;
 338         if (u > hz && tc->tc_quality >= 0) {
 339                 tc->tc_quality = -2000;
 340                 if (bootverbose) {
 341                         printf("Timecounter \"%s\" frequency %ju Hz",
 342                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 343                         printf(" -- Insufficient hz, needs at least %u\n", u);
 344                 }
 345         } else if (tc->tc_quality >= 0 || bootverbose) {
 346                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 347                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 348                     tc->tc_quality);
 349         }
 350
 351         tc->tc_next = timecounters;
 352         timecounters = tc;
 353         /*
 354          * Set up sysctl tree for this counter.
 355          */
 356         tc_root = SYSCTL_ADD_NODE(NULL,
 357             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 358             CTLFLAG_RW, 0, "timecounter description");
 359         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 360             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 361             "mask for implemented bits");
 362         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 363             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 364             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 365         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 366             "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
 367              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 368         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 369             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 370             "goodness of time counter");
 371         /*
 372          * Never automatically use a timecounter with negative quality.
 373          * Even though we run on the dummy counter, switching here may be
 374          * worse since this timecounter may not be monotonous.
 375          */
 376         if (tc->tc_quality < 0)
 377                 return;
 378         if (tc->tc_quality < timecounter->tc_quality)
 379                 return;
 380         if (tc->tc_quality == timecounter->tc_quality &&
 381             tc->tc_frequency < timecounter->tc_frequency)
 382                 return;
 383         (void)tc->tc_get_timecount(tc);
 384         (void)tc->tc_get_timecount(tc);
 385         timecounter = tc;
 386 }
 387
 388 /* Report the frequency of the current timecounter. */
 389 uint64_t
 390 tc_getfrequency(void)
 391 {
 392
 393         return (timehands->th_counter->tc_frequency);
 394 }
 395
 396 /*
 397  * Step our concept of UTC.  This is done by modifying our estimate of
 398  * when we booted.
 399  * XXX: not locked.
 400  */
 401 void
 402 tc_setclock(struct timespec *ts)
 403 {
 404         struct timespec tbef, taft;
 405         struct bintime bt, bt2;
 406
 407         cpu_tick_calibrate(1);
 408         nanotime(&tbef);
 409         timespec2bintime(ts, &bt);
 410         binuptime(&bt2);
 411         bintime_sub(&bt, &bt2);
 412         bintime_add(&bt2, &boottimebin);
 413         boottimebin = bt;
 414         bintime2timeval(&bt, &boottime);
 415
 416         /* XXX fiddle all the little crinkly bits around the fiords... */
 417         tc_windup();
 418         nanotime(&taft);
 419         if (timestepwarnings) {
 420                 log(LOG_INFO,
 421                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 422                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 423                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 424                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 425         }
 426         cpu_tick_calibrate(1);
 427 }
 428
 429 /*
 430  * Initialize the next struct timehands in the ring and make
 431  * it the active timehands.  Along the way we might switch to a different
 432  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 433  */
 434 static void
 435 tc_windup(void)
 436 {
 437         struct bintime bt;
 438         struct timehands *th, *tho;
 439         uint64_t scale;
 440         u_int delta, ncount, ogen;
 441         int i;
 442         time_t t;
 443
 444         /*
 445          * Make the next timehands a copy of the current one, but do not
 446          * overwrite the generation or next pointer.  While we update
 447          * the contents, the generation must be zero.
 448          */
 449         tho = timehands;
 450         th = tho->th_next;
 451         ogen = th->th_generation;
 452         th->th_generation = 0;
 453         bcopy(tho, th, offsetof(struct timehands, th_generation));
 454
 455         /*
 456          * Capture a timecounter delta on the current timecounter and if
 457          * changing timecounters, a counter value from the new timecounter.
 458          * Update the offset fields accordingly.
 459          */
 460         delta = tc_delta(th);
 461         if (th->th_counter != timecounter)
 462                 ncount = timecounter->tc_get_timecount(timecounter);
 463         else
 464                 ncount = 0;
 465         th->th_offset_count += delta;
 466         th->th_offset_count &= th->th_counter->tc_counter_mask;
 467         while (delta > th->th_counter->tc_frequency) {
 468                 /* Eat complete unadjusted seconds. */
 469                 delta -= th->th_counter->tc_frequency;
 470                 th->th_offset.sec++;
 471         }
 472         if ((delta > th->th_counter->tc_frequency / 2) &&
 473             (th->th_scale * delta < ((uint64_t)1 << 63))) {
 474                 /* The product th_scale * delta just barely overflows. */
 475                 th->th_offset.sec++;
 476         }
 477         bintime_addx(&th->th_offset, th->th_scale * delta);
 478
 479         /*
 480          * Hardware latching timecounters may not generate interrupts on
 481          * PPS events, so instead we poll them.  There is a finite risk that
 482          * the hardware might capture a count which is later than the one we
 483          * got above, and therefore possibly in the next NTP second which might
 484          * have a different rate than the current NTP second.  It doesn't
 485          * matter in practice.
 486          */
 487         if (tho->th_counter->tc_poll_pps)
 488                 tho->th_counter->tc_poll_pps(tho->th_counter);
 489
 490         /*
 491          * Deal with NTP second processing.  The for loop normally
 492          * iterates at most once, but in extreme situations it might
 493          * keep NTP sane if timeouts are not run for several seconds.
 494          * At boot, the time step can be large when the TOD hardware
 495          * has been read, so on really large steps, we call
 496          * ntp_update_second only twice.  We need to call it twice in
 497          * case we missed a leap second.
 498          */
 499         bt = th->th_offset;
 500         bintime_add(&bt, &boottimebin);
 501         i = bt.sec - tho->th_microtime.tv_sec;
 502         if (i > LARGE_STEP)
 503                 i = 2;
 504         for (; i > 0; i--) {
 505                 t = bt.sec;
 506                 ntp_update_second(&th->th_adjustment, &bt.sec);
 507                 if (bt.sec != t)
 508                         boottimebin.sec += bt.sec - t;
 509         }
 510         /* Update the UTC timestamps used by the get*() functions. */
 511         /* XXX shouldn't do this here.  Should force non-`get' versions. */
 512         bintime2timeval(&bt, &th->th_microtime);
 513         bintime2timespec(&bt, &th->th_nanotime);
 514
 515         /* Now is a good time to change timecounters. */
 516         if (th->th_counter != timecounter) {
 517 #ifndef __arm__
 518                 if ((timecounter->tc_flags & TC_FLAGS_C3STOP) != 0)
 519                         cpu_disable_deep_sleep++;
 520                 if ((th->th_counter->tc_flags & TC_FLAGS_C3STOP) != 0)
 521                         cpu_disable_deep_sleep--;
 522 #endif
 523                 th->th_counter = timecounter;
 524                 th->th_offset_count = ncount;
 525                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
 526                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
 527         }
 528
 529         /*-
 530          * Recalculate the scaling factor.  We want the number of 1/2^64
 531          * fractions of a second per period of the hardware counter, taking
 532          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 533          * processing provides us with.
 534          *
 535          * The th_adjustment is nanoseconds per second with 32 bit binary
 536          * fraction and we want 64 bit binary fraction of second:
 537          *
 538          *       x = a * 2^32 / 10^9 = a * 4.294967296
 539          *
 540          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 541          * we can only multiply by about 850 without overflowing, that
 542          * leaves no suitably precise fractions for multiply before divide.
 543          *
 544          * Divide before multiply with a fraction of 2199/512 results in a
 545          * systematic undercompensation of 10PPM of th_adjustment.  On a
 546          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 547          *
 548          * We happily sacrifice the lowest of the 64 bits of our result
 549          * to the goddess of code clarity.
 550          *
 551          */
 552         scale = (uint64_t)1 << 63;
 553         scale += (th->th_adjustment / 1024) * 2199;
 554         scale /= th->th_counter->tc_frequency;
 555         th->th_scale = scale * 2;
 556
 557         /*
 558          * Now that the struct timehands is again consistent, set the new
 559          * generation number, making sure to not make it zero.
 560          */
 561         if (++ogen == 0)
 562                 ogen = 1;
 563         th->th_generation = ogen;
 564
 565         /* Go live with the new struct timehands. */
 566         time_second = th->th_microtime.tv_sec;
 567         time_uptime = th->th_offset.sec;
 568         timehands = th;
 569         timekeep_push_vdso();
 570 }
 571
 572 /* Report or change the active timecounter hardware. */
 573 static int
 574 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 575 {
 576         char newname[32];
 577         struct timecounter *newtc, *tc;
 578         int error;
 579
 580         tc = timecounter;
 581         strlcpy(newname, tc->tc_name, sizeof(newname));
 582
 583         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 584         if (error != 0 || req->newptr == NULL ||
 585             strcmp(newname, tc->tc_name) == 0)
 586                 return (error);
 587         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 588                 if (strcmp(newname, newtc->tc_name) != 0)
 589                         continue;
 590
 591                 /* Warm up new timecounter. */
 592                 (void)newtc->tc_get_timecount(newtc);
 593                 (void)newtc->tc_get_timecount(newtc);
 594
 595                 timecounter = newtc;
 596                 timekeep_push_vdso();
 597                 return (0);
 598         }
 599         return (EINVAL);
 600 }
 601
 602 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 603     0, 0, sysctl_kern_timecounter_hardware, "A",
 604     "Timecounter hardware selected");
 605
 606
 607 /* Report or change the active timecounter hardware. */
 608 static int
 609 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 610 {
 611         char buf[32], *spc;
 612         struct timecounter *tc;
 613         int error;
 614
 615         spc = "";
 616         error = 0;
 617         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
 618                 sprintf(buf, "%s%s(%d)",
 619                     spc, tc->tc_name, tc->tc_quality);
 620                 error = SYSCTL_OUT(req, buf, strlen(buf));
 621                 spc = " ";
 622         }
 623         return (error);
 624 }
 625
 626 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 627     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
 628
 629 /*
 630  * RFC 2783 PPS-API implementation.
 631  */
 632
 633 int
 634 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 635 {
 636         pps_params_t *app;
 637         struct pps_fetch_args *fapi;
 638 #ifdef PPS_SYNC
 639         struct pps_kcbind_args *kapi;
 640 #endif
 641
 642         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 643         switch (cmd) {
 644         case PPS_IOC_CREATE:
 645                 return (0);
 646         case PPS_IOC_DESTROY:
 647                 return (0);
 648         case PPS_IOC_SETPARAMS:
 649                 app = (pps_params_t *)data;
 650                 if (app->mode & ~pps->ppscap)
 651                         return (EINVAL);
 652                 pps->ppsparam = *app;
 653                 return (0);
 654         case PPS_IOC_GETPARAMS:
 655                 app = (pps_params_t *)data;
 656                 *app = pps->ppsparam;
 657                 app->api_version = PPS_API_VERS_1;
 658                 return (0);
 659         case PPS_IOC_GETCAP:
 660                 *(int*)data = pps->ppscap;
 661                 return (0);
 662         case PPS_IOC_FETCH:
 663                 fapi = (struct pps_fetch_args *)data;
 664                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 665                         return (EINVAL);
 666                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
 667                         return (EOPNOTSUPP);
 668                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
 669                 fapi->pps_info_buf = pps->ppsinfo;
 670                 return (0);
 671         case PPS_IOC_KCBIND:
 672 #ifdef PPS_SYNC
 673                 kapi = (struct pps_kcbind_args *)data;
 674                 /* XXX Only root should be able to do this */
 675                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 676                         return (EINVAL);
 677                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 678                         return (EINVAL);
 679                 if (kapi->edge & ~pps->ppscap)
 680                         return (EINVAL);
 681                 pps->kcmode = kapi->edge;
 682                 return (0);
 683 #else
 684                 return (EOPNOTSUPP);
 685 #endif
 686         default:
 687                 return (ENOIOCTL);
 688         }
 689 }
 690
 691 void
 692 pps_init(struct pps_state *pps)
 693 {
 694         pps->ppscap |= PPS_TSFMT_TSPEC;
 695         if (pps->ppscap & PPS_CAPTUREASSERT)
 696                 pps->ppscap |= PPS_OFFSETASSERT;
 697         if (pps->ppscap & PPS_CAPTURECLEAR)
 698                 pps->ppscap |= PPS_OFFSETCLEAR;
 699 }
 700
 701 void
 702 pps_capture(struct pps_state *pps)
 703 {
 704         struct timehands *th;
 705
 706         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 707         th = timehands;
 708         pps->capgen = th->th_generation;
 709         pps->capth = th;
 710         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 711         if (pps->capgen != th->th_generation)
 712                 pps->capgen = 0;
 713 }
 714
 715 void
 716 pps_event(struct pps_state *pps, int event)
 717 {
 718         struct bintime bt;
 719         struct timespec ts, *tsp, *osp;
 720         u_int tcount, *pcount;
 721         int foff, fhard;
 722         pps_seq_t *pseq;
 723
 724         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 725         /* If the timecounter was wound up underneath us, bail out. */
 726         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
 727                 return;
 728
 729         /* Things would be easier with arrays. */
 730         if (event == PPS_CAPTUREASSERT) {
 731                 tsp = &pps->ppsinfo.assert_timestamp;
 732                 osp = &pps->ppsparam.assert_offset;
 733                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 734                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 735                 pcount = &pps->ppscount[0];
 736                 pseq = &pps->ppsinfo.assert_sequence;
 737         } else {
 738                 tsp = &pps->ppsinfo.clear_timestamp;
 739                 osp = &pps->ppsparam.clear_offset;
 740                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 741                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 742                 pcount = &pps->ppscount[1];
 743                 pseq = &pps->ppsinfo.clear_sequence;
 744         }
 745
 746         /*
 747          * If the timecounter changed, we cannot compare the count values, so
 748          * we have to drop the rest of the PPS-stuff until the next event.
 749          */
 750         if (pps->ppstc != pps->capth->th_counter) {
 751                 pps->ppstc = pps->capth->th_counter;
 752                 *pcount = pps->capcount;
 753                 pps->ppscount[2] = pps->capcount;
 754                 return;
 755         }
 756
 757         /* Convert the count to a timespec. */
 758         tcount = pps->capcount - pps->capth->th_offset_count;
 759         tcount &= pps->capth->th_counter->tc_counter_mask;
 760         bt = pps->capth->th_offset;
 761         bintime_addx(&bt, pps->capth->th_scale * tcount);
 762         bintime_add(&bt, &boottimebin);
 763         bintime2timespec(&bt, &ts);
 764
 765         /* If the timecounter was wound up underneath us, bail out. */
 766         if (pps->capgen != pps->capth->th_generation)
 767                 return;
 768
 769         *pcount = pps->capcount;
 770         (*pseq)++;
 771         *tsp = ts;
 772
 773         if (foff) {
 774                 timespecadd(tsp, osp);
 775                 if (tsp->tv_nsec < 0) {
 776                         tsp->tv_nsec += 1000000000;
 777                         tsp->tv_sec -= 1;
 778                 }
 779         }
 780 #ifdef PPS_SYNC
 781         if (fhard) {
 782                 uint64_t scale;
 783
 784                 /*
 785                  * Feed the NTP PLL/FLL.
 786                  * The FLL wants to know how many (hardware) nanoseconds
 787                  * elapsed since the previous event.
 788                  */
 789                 tcount = pps->capcount - pps->ppscount[2];
 790                 pps->ppscount[2] = pps->capcount;
 791                 tcount &= pps->capth->th_counter->tc_counter_mask;
 792                 scale = (uint64_t)1 << 63;
 793                 scale /= pps->capth->th_counter->tc_frequency;
 794                 scale *= 2;
 795                 bt.sec = 0;
 796                 bt.frac = 0;
 797                 bintime_addx(&bt, scale * tcount);
 798                 bintime2timespec(&bt, &ts);
 799                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 800         }
 801 #endif
 802 }
 803
 804 /*
 805  * Timecounters need to be updated every so often to prevent the hardware
 806  * counter from overflowing.  Updating also recalculates the cached values
 807  * used by the get*() family of functions, so their precision depends on
 808  * the update frequency.
 809  */
 810
 811 static int tc_tick;
 812 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
 813     "Approximate number of hardclock ticks in a millisecond");
 814
 815 void
 816 tc_ticktock(int cnt)
 817 {
 818         static int count;
 819
 820         count += cnt;
 821         if (count < tc_tick)
 822                 return;
 823         count = 0;
 824         tc_windup();
 825 }
 826
 827 static void
 828 inittimecounter(void *dummy)
 829 {
 830         u_int p;
 831
 832         /*
 833          * Set the initial timeout to
 834          * max(1, <approx. number of hardclock ticks in a millisecond>).
 835          * People should probably not use the sysctl to set the timeout
 836          * to smaller than its inital value, since that value is the
 837          * smallest reasonable one.  If they want better timestamps they
 838          * should use the non-"get"* functions.
 839          */
 840         if (hz > 1000)
 841                 tc_tick = (hz + 500) / 1000;
 842         else
 843                 tc_tick = 1;
 844         p = (tc_tick * 1000000) / hz;
 845         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 846
 847         /* warm up new timecounter (again) and get rolling. */
 848         (void)timecounter->tc_get_timecount(timecounter);
 849         (void)timecounter->tc_get_timecount(timecounter);
 850         tc_windup();
 851 }
 852
 853 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 854
 855 /* Cpu tick handling -------------------------------------------------*/
 856
 857 static int cpu_tick_variable;
 858 static uint64_t cpu_tick_frequency;
 859
 860 static uint64_t
 861 tc_cpu_ticks(void)
 862 {
 863         static uint64_t base;
 864         static unsigned last;
 865         unsigned u;
 866         struct timecounter *tc;
 867
 868         tc = timehands->th_counter;
 869         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 870         if (u < last)
 871                 base += (uint64_t)tc->tc_counter_mask + 1;
 872         last = u;
 873         return (u + base);
 874 }
 875
 876 void
 877 cpu_tick_calibration(void)
 878 {
 879         static time_t last_calib;
 880
 881         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 882                 cpu_tick_calibrate(0);
 883                 last_calib = time_uptime;
 884         }
 885 }
 886
 887 /*
 888  * This function gets called every 16 seconds on only one designated
 889  * CPU in the system from hardclock() via cpu_tick_calibration()().
 890  *
 891  * Whenever the real time clock is stepped we get called with reset=1
 892  * to make sure we handle suspend/resume and similar events correctly.
 893  */
 894
 895 static void
 896 cpu_tick_calibrate(int reset)
 897 {
 898         static uint64_t c_last;
 899         uint64_t c_this, c_delta;
 900         static struct bintime  t_last;
 901         struct bintime t_this, t_delta;
 902         uint32_t divi;
 903
 904         if (reset) {
 905                 /* The clock was stepped, abort & reset */
 906                 t_last.sec = 0;
 907                 return;
 908         }
 909
 910         /* we don't calibrate fixed rate cputicks */
 911         if (!cpu_tick_variable)
 912                 return;
 913
 914         getbinuptime(&t_this);
 915         c_this = cpu_ticks();
 916         if (t_last.sec != 0) {
 917                 c_delta = c_this - c_last;
 918                 t_delta = t_this;
 919                 bintime_sub(&t_delta, &t_last);
 920                 /*
 921                  * Headroom:
 922                  *      2^(64-20) / 16[s] =
 923                  *      2^(44) / 16[s] =
 924                  *      17.592.186.044.416 / 16 =
 925                  *      1.099.511.627.776 [Hz]
 926                  */
 927                 divi = t_delta.sec << 20;
 928                 divi |= t_delta.frac >> (64 - 20);
 929                 c_delta <<= 20;
 930                 c_delta /= divi;
 931                 if (c_delta > cpu_tick_frequency) {
 932                         if (0 && bootverbose)
 933                                 printf("cpu_tick increased to %ju Hz\n",
 934                                     c_delta);
 935                         cpu_tick_frequency = c_delta;
 936                 }
 937         }
 938         c_last = c_this;
 939         t_last = t_this;
 940 }
 941
 942 void
 943 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 944 {
 945
 946         if (func == NULL) {
 947                 cpu_ticks = tc_cpu_ticks;
 948         } else {
 949                 cpu_tick_frequency = freq;
 950                 cpu_tick_variable = var;
 951                 cpu_ticks = func;
 952         }
 953 }
 954
 955 uint64_t
 956 cpu_tickrate(void)
 957 {
 958
 959         if (cpu_ticks == tc_cpu_ticks)
 960                 return (tc_getfrequency());
 961         return (cpu_tick_frequency);
 962 }
 963
 964 /*
 965  * We need to be slightly careful converting cputicks to microseconds.
 966  * There is plenty of margin in 64 bits of microseconds (half a million
 967  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 968  * before divide conversion (to retain precision) we find that the
 969  * margin shrinks to 1.5 hours (one millionth of 146y).
 970  * With a three prong approach we never lose significant bits, no
 971  * matter what the cputick rate and length of timeinterval is.
 972  */
 973
 974 uint64_t
 975 cputick2usec(uint64_t tick)
 976 {
 977
 978         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 979                 return (tick / (cpu_tickrate() / 1000000LL));
 980         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 981                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 982         else
 983                 return ((tick * 1000000LL) / cpu_tickrate());
 984 }
 985
 986 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
 987
 988 static int vdso_th_enable = 1;
 989 static int
 990 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
 991 {
 992         int old_vdso_th_enable, error;
 993
 994         old_vdso_th_enable = vdso_th_enable;
 995         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
 996         if (error != 0)
 997                 return (error);
 998         vdso_th_enable = old_vdso_th_enable;
 999         timekeep_push_vdso();
1000         return (0);
1001 }
1002 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
1003     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1004     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
1005
1006 uint32_t
1007 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
1008 {
1009         struct timehands *th;
1010         uint32_t enabled;
1011
1012         th = timehands;
1013         vdso_th->th_algo = VDSO_TH_ALGO_1;
1014         vdso_th->th_scale = th->th_scale;
1015         vdso_th->th_offset_count = th->th_offset_count;
1016         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
1017         vdso_th->th_offset = th->th_offset;
1018         vdso_th->th_boottime = boottimebin;
1019         enabled = cpu_fill_vdso_timehands(vdso_th);
1020         if (!vdso_th_enable)
1021                 enabled = 0;
1022         return (enabled);
1023 }
1024
1025 #ifdef COMPAT_FREEBSD32
1026 uint32_t
1027 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
1028 {
1029         struct timehands *th;
1030         uint32_t enabled;
1031
1032         th = timehands;
1033         vdso_th32->th_algo = VDSO_TH_ALGO_1;
1034         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
1035         vdso_th32->th_offset_count = th->th_offset_count;
1036         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
1037         vdso_th32->th_offset.sec = th->th_offset.sec;
1038         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
1039         vdso_th32->th_boottime.sec = boottimebin.sec;
1040         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = boottimebin.frac;
1041         enabled = cpu_fill_vdso_timehands32(vdso_th32);
1042         if (!vdso_th_enable)
1043                 enabled = 0;
1044         return (enabled);
1045 }
1046 #endif