kern_tc.c revision 155534
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: head/sys/kern/kern_tc.c 155534 2006-02-11 09:33:07Z phk $"); 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 38static u_int 39dummy_get_timecount(struct timecounter *tc) 40{ 41 static u_int now; 42 43 return (++now); 44} 45 46static struct timecounter dummy_timecounter = { 47 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000 48}; 49 50struct 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 64static struct timehands th0; 65static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0}; 66static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9}; 67static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8}; 68static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7}; 69static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6}; 70static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5}; 71static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4}; 72static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3}; 73static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2}; 74static 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 86static struct timehands *volatile timehands = &th0; 87struct timecounter *timecounter = &dummy_timecounter; 88static struct timecounter *timecounters = &dummy_timecounter; 89 90time_t time_second = 1; 91time_t time_uptime = 1; 92 93static struct bintime boottimebin; 94struct timeval boottime; 95static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS); 96SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD, 97 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime"); 98 99SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, ""); 100 101static int timestepwarnings; 102SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW, 103 ×tepwarnings, 0, ""); 104 105#define TC_STATS(foo) \ 106 static u_int foo; \ 107 SYSCTL_UINT(_kern_timecounter, OID_AUTO, foo, CTLFLAG_RD, &foo, 0, "");\ 108 struct __hack 109 110TC_STATS(nbinuptime); TC_STATS(nnanouptime); TC_STATS(nmicrouptime); 111TC_STATS(nbintime); TC_STATS(nnanotime); TC_STATS(nmicrotime); 112TC_STATS(ngetbinuptime); TC_STATS(ngetnanouptime); TC_STATS(ngetmicrouptime); 113TC_STATS(ngetbintime); TC_STATS(ngetnanotime); TC_STATS(ngetmicrotime); 114TC_STATS(nsetclock); 115 116#undef TC_STATS 117 118static void tc_windup(void); 119static void cpu_tick_calibrate(int); 120 121static int 122sysctl_kern_boottime(SYSCTL_HANDLER_ARGS) 123{ 124#ifdef SCTL_MASK32 125 int tv[2]; 126 127 if (req->flags & SCTL_MASK32) { 128 tv[0] = boottime.tv_sec; 129 tv[1] = boottime.tv_usec; 130 return SYSCTL_OUT(req, tv, sizeof(tv)); 131 } else 132#endif 133 return SYSCTL_OUT(req, &boottime, sizeof(boottime)); 134} 135 136/* 137 * Return the difference between the timehands' counter value now and what 138 * was when we copied it to the timehands' offset_count. 139 */ 140static __inline u_int 141tc_delta(struct timehands *th) 142{ 143 struct timecounter *tc; 144 145 tc = th->th_counter; 146 return ((tc->tc_get_timecount(tc) - th->th_offset_count) & 147 tc->tc_counter_mask); 148} 149 150/* 151 * Functions for reading the time. We have to loop until we are sure that 152 * the timehands that we operated on was not updated under our feet. See 153 * the comment in <sys/time.h> for a description of these 12 functions. 154 */ 155 156void 157binuptime(struct bintime *bt) 158{ 159 struct timehands *th; 160 u_int gen; 161 162 nbinuptime++; 163 do { 164 th = timehands; 165 gen = th->th_generation; 166 *bt = th->th_offset; 167 bintime_addx(bt, th->th_scale * tc_delta(th)); 168 } while (gen == 0 || gen != th->th_generation); 169} 170 171void 172nanouptime(struct timespec *tsp) 173{ 174 struct bintime bt; 175 176 nnanouptime++; 177 binuptime(&bt); 178 bintime2timespec(&bt, tsp); 179} 180 181void 182microuptime(struct timeval *tvp) 183{ 184 struct bintime bt; 185 186 nmicrouptime++; 187 binuptime(&bt); 188 bintime2timeval(&bt, tvp); 189} 190 191void 192bintime(struct bintime *bt) 193{ 194 195 nbintime++; 196 binuptime(bt); 197 bintime_add(bt, &boottimebin); 198} 199 200void 201nanotime(struct timespec *tsp) 202{ 203 struct bintime bt; 204 205 nnanotime++; 206 bintime(&bt); 207 bintime2timespec(&bt, tsp); 208} 209 210void 211microtime(struct timeval *tvp) 212{ 213 struct bintime bt; 214 215 nmicrotime++; 216 bintime(&bt); 217 bintime2timeval(&bt, tvp); 218} 219 220void 221getbinuptime(struct bintime *bt) 222{ 223 struct timehands *th; 224 u_int gen; 225 226 ngetbinuptime++; 227 do { 228 th = timehands; 229 gen = th->th_generation; 230 *bt = th->th_offset; 231 } while (gen == 0 || gen != th->th_generation); 232} 233 234void 235getnanouptime(struct timespec *tsp) 236{ 237 struct timehands *th; 238 u_int gen; 239 240 ngetnanouptime++; 241 do { 242 th = timehands; 243 gen = th->th_generation; 244 bintime2timespec(&th->th_offset, tsp); 245 } while (gen == 0 || gen != th->th_generation); 246} 247 248void 249getmicrouptime(struct timeval *tvp) 250{ 251 struct timehands *th; 252 u_int gen; 253 254 ngetmicrouptime++; 255 do { 256 th = timehands; 257 gen = th->th_generation; 258 bintime2timeval(&th->th_offset, tvp); 259 } while (gen == 0 || gen != th->th_generation); 260} 261 262void 263getbintime(struct bintime *bt) 264{ 265 struct timehands *th; 266 u_int gen; 267 268 ngetbintime++; 269 do { 270 th = timehands; 271 gen = th->th_generation; 272 *bt = th->th_offset; 273 } while (gen == 0 || gen != th->th_generation); 274 bintime_add(bt, &boottimebin); 275} 276 277void 278getnanotime(struct timespec *tsp) 279{ 280 struct timehands *th; 281 u_int gen; 282 283 ngetnanotime++; 284 do { 285 th = timehands; 286 gen = th->th_generation; 287 *tsp = th->th_nanotime; 288 } while (gen == 0 || gen != th->th_generation); 289} 290 291void 292getmicrotime(struct timeval *tvp) 293{ 294 struct timehands *th; 295 u_int gen; 296 297 ngetmicrotime++; 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 */ 308void 309tc_init(struct timecounter *tc) 310{ 311 u_int u; 312 313 u = tc->tc_frequency / tc->tc_counter_mask; 314 /* XXX: We need some margin here, 10% is a guess */ 315 u *= 11; 316 u /= 10; 317 if (u > hz && tc->tc_quality >= 0) { 318 tc->tc_quality = -2000; 319 if (bootverbose) { 320 printf("Timecounter \"%s\" frequency %ju Hz", 321 tc->tc_name, (uintmax_t)tc->tc_frequency); 322 printf(" -- Insufficient hz, needs at least %u\n", u); 323 } 324 } else if (tc->tc_quality >= 0 || bootverbose) { 325 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n", 326 tc->tc_name, (uintmax_t)tc->tc_frequency, 327 tc->tc_quality); 328 } 329 330 tc->tc_next = timecounters; 331 timecounters = tc; 332 /* 333 * Never automatically use a timecounter with negative quality. 334 * Even though we run on the dummy counter, switching here may be 335 * worse since this timecounter may not be monotonous. 336 */ 337 if (tc->tc_quality < 0) 338 return; 339 if (tc->tc_quality < timecounter->tc_quality) 340 return; 341 if (tc->tc_quality == timecounter->tc_quality && 342 tc->tc_frequency < timecounter->tc_frequency) 343 return; 344 (void)tc->tc_get_timecount(tc); 345 (void)tc->tc_get_timecount(tc); 346 timecounter = tc; 347} 348 349/* Report the frequency of the current timecounter. */ 350u_int64_t 351tc_getfrequency(void) 352{ 353 354 return (timehands->th_counter->tc_frequency); 355} 356 357/* 358 * Step our concept of UTC. This is done by modifying our estimate of 359 * when we booted. 360 * XXX: not locked. 361 */ 362void 363tc_setclock(struct timespec *ts) 364{ 365 struct timespec ts2; 366 struct bintime bt, bt2; 367 368 cpu_tick_calibrate(1); 369 nsetclock++; 370 binuptime(&bt2); 371 timespec2bintime(ts, &bt); 372 bintime_sub(&bt, &bt2); 373 bintime_add(&bt2, &boottimebin); 374 boottimebin = bt; 375 bintime2timeval(&bt, &boottime); 376 377 /* XXX fiddle all the little crinkly bits around the fiords... */ 378 tc_windup(); 379 if (timestepwarnings) { 380 bintime2timespec(&bt2, &ts2); 381 log(LOG_INFO, "Time stepped from %jd.%09ld to %jd.%09ld\n", 382 (intmax_t)ts2.tv_sec, ts2.tv_nsec, 383 (intmax_t)ts->tv_sec, ts->tv_nsec); 384 } 385 cpu_tick_calibrate(1); 386} 387 388/* 389 * Initialize the next struct timehands in the ring and make 390 * it the active timehands. Along the way we might switch to a different 391 * timecounter and/or do seconds processing in NTP. Slightly magic. 392 */ 393static void 394tc_windup(void) 395{ 396 struct bintime bt; 397 struct timehands *th, *tho; 398 u_int64_t scale; 399 u_int delta, ncount, ogen; 400 int i; 401 time_t t; 402 403 /* 404 * Make the next timehands a copy of the current one, but do not 405 * overwrite the generation or next pointer. While we update 406 * the contents, the generation must be zero. 407 */ 408 tho = timehands; 409 th = tho->th_next; 410 ogen = th->th_generation; 411 th->th_generation = 0; 412 bcopy(tho, th, offsetof(struct timehands, th_generation)); 413 414 /* 415 * Capture a timecounter delta on the current timecounter and if 416 * changing timecounters, a counter value from the new timecounter. 417 * Update the offset fields accordingly. 418 */ 419 delta = tc_delta(th); 420 if (th->th_counter != timecounter) 421 ncount = timecounter->tc_get_timecount(timecounter); 422 else 423 ncount = 0; 424 th->th_offset_count += delta; 425 th->th_offset_count &= th->th_counter->tc_counter_mask; 426 bintime_addx(&th->th_offset, th->th_scale * delta); 427 428 /* 429 * Hardware latching timecounters may not generate interrupts on 430 * PPS events, so instead we poll them. There is a finite risk that 431 * the hardware might capture a count which is later than the one we 432 * got above, and therefore possibly in the next NTP second which might 433 * have a different rate than the current NTP second. It doesn't 434 * matter in practice. 435 */ 436 if (tho->th_counter->tc_poll_pps) 437 tho->th_counter->tc_poll_pps(tho->th_counter); 438 439 /* 440 * Deal with NTP second processing. The for loop normally 441 * iterates at most once, but in extreme situations it might 442 * keep NTP sane if timeouts are not run for several seconds. 443 * At boot, the time step can be large when the TOD hardware 444 * has been read, so on really large steps, we call 445 * ntp_update_second only twice. We need to call it twice in 446 * case we missed a leap second. 447 */ 448 bt = th->th_offset; 449 bintime_add(&bt, &boottimebin); 450 i = bt.sec - tho->th_microtime.tv_sec; 451 if (i > LARGE_STEP) 452 i = 2; 453 for (; i > 0; i--) { 454 t = bt.sec; 455 ntp_update_second(&th->th_adjustment, &bt.sec); 456 if (bt.sec != t) 457 boottimebin.sec += bt.sec - t; 458 } 459 /* Update the UTC timestamps used by the get*() functions. */ 460 /* XXX shouldn't do this here. Should force non-`get' versions. */ 461 bintime2timeval(&bt, &th->th_microtime); 462 bintime2timespec(&bt, &th->th_nanotime); 463 464 /* Now is a good time to change timecounters. */ 465 if (th->th_counter != timecounter) { 466 th->th_counter = timecounter; 467 th->th_offset_count = ncount; 468 } 469 470 /*- 471 * Recalculate the scaling factor. We want the number of 1/2^64 472 * fractions of a second per period of the hardware counter, taking 473 * into account the th_adjustment factor which the NTP PLL/adjtime(2) 474 * processing provides us with. 475 * 476 * The th_adjustment is nanoseconds per second with 32 bit binary 477 * fraction and we want 64 bit binary fraction of second: 478 * 479 * x = a * 2^32 / 10^9 = a * 4.294967296 480 * 481 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int 482 * we can only multiply by about 850 without overflowing, that 483 * leaves no suitably precise fractions for multiply before divide. 484 * 485 * Divide before multiply with a fraction of 2199/512 results in a 486 * systematic undercompensation of 10PPM of th_adjustment. On a 487 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable. 488 * 489 * We happily sacrifice the lowest of the 64 bits of our result 490 * to the goddess of code clarity. 491 * 492 */ 493 scale = (u_int64_t)1 << 63; 494 scale += (th->th_adjustment / 1024) * 2199; 495 scale /= th->th_counter->tc_frequency; 496 th->th_scale = scale * 2; 497 498 /* 499 * Now that the struct timehands is again consistent, set the new 500 * generation number, making sure to not make it zero. 501 */ 502 if (++ogen == 0) 503 ogen = 1; 504 th->th_generation = ogen; 505 506 /* Go live with the new struct timehands. */ 507 time_second = th->th_microtime.tv_sec; 508 time_uptime = th->th_offset.sec; 509 timehands = th; 510} 511 512/* Report or change the active timecounter hardware. */ 513static int 514sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS) 515{ 516 char newname[32]; 517 struct timecounter *newtc, *tc; 518 int error; 519 520 tc = timecounter; 521 strlcpy(newname, tc->tc_name, sizeof(newname)); 522 523 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req); 524 if (error != 0 || req->newptr == NULL || 525 strcmp(newname, tc->tc_name) == 0) 526 return (error); 527 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) { 528 if (strcmp(newname, newtc->tc_name) != 0) 529 continue; 530 531 /* Warm up new timecounter. */ 532 (void)newtc->tc_get_timecount(newtc); 533 (void)newtc->tc_get_timecount(newtc); 534 535 timecounter = newtc; 536 return (0); 537 } 538 return (EINVAL); 539} 540 541SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW, 542 0, 0, sysctl_kern_timecounter_hardware, "A", ""); 543 544 545/* Report or change the active timecounter hardware. */ 546static int 547sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS) 548{ 549 char buf[32], *spc; 550 struct timecounter *tc; 551 int error; 552 553 spc = ""; 554 error = 0; 555 for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) { 556 sprintf(buf, "%s%s(%d)", 557 spc, tc->tc_name, tc->tc_quality); 558 error = SYSCTL_OUT(req, buf, strlen(buf)); 559 spc = " "; 560 } 561 return (error); 562} 563 564SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD, 565 0, 0, sysctl_kern_timecounter_choice, "A", ""); 566 567/* 568 * RFC 2783 PPS-API implementation. 569 */ 570 571int 572pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 573{ 574 pps_params_t *app; 575 struct pps_fetch_args *fapi; 576#ifdef PPS_SYNC 577 struct pps_kcbind_args *kapi; 578#endif 579 580 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl")); 581 switch (cmd) { 582 case PPS_IOC_CREATE: 583 return (0); 584 case PPS_IOC_DESTROY: 585 return (0); 586 case PPS_IOC_SETPARAMS: 587 app = (pps_params_t *)data; 588 if (app->mode & ~pps->ppscap) 589 return (EINVAL); 590 pps->ppsparam = *app; 591 return (0); 592 case PPS_IOC_GETPARAMS: 593 app = (pps_params_t *)data; 594 *app = pps->ppsparam; 595 app->api_version = PPS_API_VERS_1; 596 return (0); 597 case PPS_IOC_GETCAP: 598 *(int*)data = pps->ppscap; 599 return (0); 600 case PPS_IOC_FETCH: 601 fapi = (struct pps_fetch_args *)data; 602 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 603 return (EINVAL); 604 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 605 return (EOPNOTSUPP); 606 pps->ppsinfo.current_mode = pps->ppsparam.mode; 607 fapi->pps_info_buf = pps->ppsinfo; 608 return (0); 609 case PPS_IOC_KCBIND: 610#ifdef PPS_SYNC 611 kapi = (struct pps_kcbind_args *)data; 612 /* XXX Only root should be able to do this */ 613 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 614 return (EINVAL); 615 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 616 return (EINVAL); 617 if (kapi->edge & ~pps->ppscap) 618 return (EINVAL); 619 pps->kcmode = kapi->edge; 620 return (0); 621#else 622 return (EOPNOTSUPP); 623#endif 624 default: 625 return (ENOIOCTL); 626 } 627} 628 629void 630pps_init(struct pps_state *pps) 631{ 632 pps->ppscap |= PPS_TSFMT_TSPEC; 633 if (pps->ppscap & PPS_CAPTUREASSERT) 634 pps->ppscap |= PPS_OFFSETASSERT; 635 if (pps->ppscap & PPS_CAPTURECLEAR) 636 pps->ppscap |= PPS_OFFSETCLEAR; 637} 638 639void 640pps_capture(struct pps_state *pps) 641{ 642 struct timehands *th; 643 644 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture")); 645 th = timehands; 646 pps->capgen = th->th_generation; 647 pps->capth = th; 648 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter); 649 if (pps->capgen != th->th_generation) 650 pps->capgen = 0; 651} 652 653void 654pps_event(struct pps_state *pps, int event) 655{ 656 struct bintime bt; 657 struct timespec ts, *tsp, *osp; 658 u_int tcount, *pcount; 659 int foff, fhard; 660 pps_seq_t *pseq; 661 662 KASSERT(pps != NULL, ("NULL pps pointer in pps_event")); 663 /* If the timecounter was wound up underneath us, bail out. */ 664 if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation) 665 return; 666 667 /* Things would be easier with arrays. */ 668 if (event == PPS_CAPTUREASSERT) { 669 tsp = &pps->ppsinfo.assert_timestamp; 670 osp = &pps->ppsparam.assert_offset; 671 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 672 fhard = pps->kcmode & PPS_CAPTUREASSERT; 673 pcount = &pps->ppscount[0]; 674 pseq = &pps->ppsinfo.assert_sequence; 675 } else { 676 tsp = &pps->ppsinfo.clear_timestamp; 677 osp = &pps->ppsparam.clear_offset; 678 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 679 fhard = pps->kcmode & PPS_CAPTURECLEAR; 680 pcount = &pps->ppscount[1]; 681 pseq = &pps->ppsinfo.clear_sequence; 682 } 683 684 /* 685 * If the timecounter changed, we cannot compare the count values, so 686 * we have to drop the rest of the PPS-stuff until the next event. 687 */ 688 if (pps->ppstc != pps->capth->th_counter) { 689 pps->ppstc = pps->capth->th_counter; 690 *pcount = pps->capcount; 691 pps->ppscount[2] = pps->capcount; 692 return; 693 } 694 695 /* Convert the count to a timespec. */ 696 tcount = pps->capcount - pps->capth->th_offset_count; 697 tcount &= pps->capth->th_counter->tc_counter_mask; 698 bt = pps->capth->th_offset; 699 bintime_addx(&bt, pps->capth->th_scale * tcount); 700 bintime_add(&bt, &boottimebin); 701 bintime2timespec(&bt, &ts); 702 703 /* If the timecounter was wound up underneath us, bail out. */ 704 if (pps->capgen != pps->capth->th_generation) 705 return; 706 707 *pcount = pps->capcount; 708 (*pseq)++; 709 *tsp = ts; 710 711 if (foff) { 712 timespecadd(tsp, osp); 713 if (tsp->tv_nsec < 0) { 714 tsp->tv_nsec += 1000000000; 715 tsp->tv_sec -= 1; 716 } 717 } 718#ifdef PPS_SYNC 719 if (fhard) { 720 u_int64_t scale; 721 722 /* 723 * Feed the NTP PLL/FLL. 724 * The FLL wants to know how many (hardware) nanoseconds 725 * elapsed since the previous event. 726 */ 727 tcount = pps->capcount - pps->ppscount[2]; 728 pps->ppscount[2] = pps->capcount; 729 tcount &= pps->capth->th_counter->tc_counter_mask; 730 scale = (u_int64_t)1 << 63; 731 scale /= pps->capth->th_counter->tc_frequency; 732 scale *= 2; 733 bt.sec = 0; 734 bt.frac = 0; 735 bintime_addx(&bt, scale * tcount); 736 bintime2timespec(&bt, &ts); 737 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec); 738 } 739#endif 740} 741 742/* 743 * Timecounters need to be updated every so often to prevent the hardware 744 * counter from overflowing. Updating also recalculates the cached values 745 * used by the get*() family of functions, so their precision depends on 746 * the update frequency. 747 */ 748 749static int tc_tick; 750SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, ""); 751 752void 753tc_ticktock(void) 754{ 755 static int count; 756 static time_t last_calib; 757 758 if (++count < tc_tick) 759 return; 760 count = 0; 761 tc_windup(); 762 if (time_uptime != last_calib && !(time_uptime & 0xf)) { 763 cpu_tick_calibrate(0); 764 last_calib = time_uptime; 765 } 766} 767 768static void 769inittimecounter(void *dummy) 770{ 771 u_int p; 772 773 /* 774 * Set the initial timeout to 775 * max(1, <approx. number of hardclock ticks in a millisecond>). 776 * People should probably not use the sysctl to set the timeout 777 * to smaller than its inital value, since that value is the 778 * smallest reasonable one. If they want better timestamps they 779 * should use the non-"get"* functions. 780 */ 781 if (hz > 1000) 782 tc_tick = (hz + 500) / 1000; 783 else 784 tc_tick = 1; 785 p = (tc_tick * 1000000) / hz; 786 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000); 787 788 /* warm up new timecounter (again) and get rolling. */ 789 (void)timecounter->tc_get_timecount(timecounter); 790 (void)timecounter->tc_get_timecount(timecounter); 791} 792 793SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL) 794 795/* Cpu tick handling -------------------------------------------------*/ 796 797static int cpu_tick_variable; 798static uint64_t cpu_tick_frequency; 799 800static 801uint64_t 802tc_cpu_ticks(void) 803{ 804 static uint64_t base; 805 static unsigned last; 806 unsigned u; 807 struct timecounter *tc; 808 809 tc = timehands->th_counter; 810 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask; 811 if (u < last) 812 base += tc->tc_counter_mask + 1; 813 last = u; 814 return (u + base); 815} 816 817/* 818 * This function gets called ever 16 seconds on only one designated 819 * CPU in the system from hardclock() via tc_ticktock(). 820 * 821 * Whenever the real time clock is stepped we get called with reset=1 822 * to make sure we handle suspend/resume and similar events correctly. 823 */ 824 825static void 826cpu_tick_calibrate(int reset) 827{ 828 static uint64_t c_last; 829 uint64_t c_this, c_delta; 830 static struct bintime t_last; 831 struct bintime t_this, t_delta; 832 833 if (reset) { 834 /* The clock was stepped, abort & reset */ 835 t_last.sec = 0; 836 return; 837 } 838 839 /* we don't calibrate fixed rate cputicks */ 840 if (!cpu_tick_variable) 841 return; 842 843 getbinuptime(&t_this); 844 c_this = cpu_ticks(); 845 if (t_last.sec != 0) { 846 c_delta = c_this - c_last; 847 t_delta = t_this; 848 bintime_sub(&t_delta, &t_last); 849 if (0 && bootverbose) { 850 struct timespec ts; 851 bintime2timespec(&t_delta, &ts); 852 printf("%ju %ju.%016jx %ju.%09ju", 853 (uintmax_t)c_delta >> 4, 854 (uintmax_t)t_delta.sec, (uintmax_t)t_delta.frac, 855 (uintmax_t)ts.tv_sec, (uintmax_t)ts.tv_nsec); 856 } 857 /* 858 * Validate that 16 +/- 1/256 seconds passed. 859 * After division by 16 this gives us a precision of 860 * roughly 250PPM which is sufficient 861 */ 862 if (t_delta.sec > 16 || ( 863 t_delta.sec == 16 && t_delta.frac >= (0x01LL << 56))) { 864 /* too long */ 865 if (0 && bootverbose) 866 printf("\ttoo long\n"); 867 } else if (t_delta.sec < 15 || 868 (t_delta.sec == 15 && t_delta.frac <= (0xffLL << 56))) { 869 /* too short */ 870 if (0 && bootverbose) 871 printf("\ttoo short\n"); 872 } else { 873 /* just right */ 874 c_delta >>= 4; 875 if (c_delta > cpu_tick_frequency) { 876 if (0 && bootverbose) 877 printf("\thigher\n"); 878 cpu_tick_frequency = c_delta; 879 } else { 880 if (0 && bootverbose) 881 printf("\tlower\n"); 882 } 883 } 884 } 885 c_last = c_this; 886 t_last = t_this; 887} 888 889void 890set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var) 891{ 892 893 if (func == NULL) { 894 cpu_ticks = tc_cpu_ticks; 895 } else { 896 cpu_tick_frequency = freq; 897 cpu_tick_variable = var; 898 cpu_ticks = func; 899 } 900} 901 902uint64_t 903cpu_tickrate(void) 904{ 905 906 if (cpu_ticks == tc_cpu_ticks) 907 return (tc_getfrequency()); 908 return (cpu_tick_frequency); 909} 910 911/* 912 * We need to be slightly careful converting cputicks to microseconds. 913 * There is plenty of margin in 64 bits of microseconds (half a million 914 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply 915 * before divide conversion (to retain precision) we find that the 916 * margin shrinks to 1.5 hours (one millionth of 146y). 917 * With a three prong approach we never loose significant bits, no 918 * matter what the cputick rate and length of timeinterval is. 919 */ 920 921uint64_t 922cputick2usec(uint64_t tick) 923{ 924 925 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */ 926 return (tick / (cpu_tickrate() / 1000000LL)); 927 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */ 928 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL)); 929 else 930 return ((tick * 1000000LL) / cpu_tickrate()); 931} 932 933cpu_tick_f *cpu_ticks = tc_cpu_ticks; 934