kern_tc.c revision 231341
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 * Copyright (c) 2011 The FreeBSD Foundation 10 * All rights reserved. 11 * 12 * Portions of this software were developed by Julien Ridoux at the University 13 * of Melbourne under sponsorship from the FreeBSD Foundation. 14 */ 15 16#include <sys/cdefs.h> 17__FBSDID("$FreeBSD: head/sys/kern/kern_tc.c 231341 2012-02-10 06:30:52Z kevlo $"); 18 19#include "opt_ntp.h" 20#include "opt_ffclock.h" 21 22#include <sys/param.h> 23#include <sys/kernel.h> 24#ifdef FFCLOCK 25#include <sys/lock.h> 26#include <sys/mutex.h> 27#endif 28#include <sys/sysctl.h> 29#include <sys/syslog.h> 30#include <sys/systm.h> 31#include <sys/timeffc.h> 32#include <sys/timepps.h> 33#include <sys/timetc.h> 34#include <sys/timex.h> 35 36/* 37 * A large step happens on boot. This constant detects such steps. 38 * It is relatively small so that ntp_update_second gets called enough 39 * in the typical 'missed a couple of seconds' case, but doesn't loop 40 * forever when the time step is large. 41 */ 42#define LARGE_STEP 200 43 44/* 45 * Implement a dummy timecounter which we can use until we get a real one 46 * in the air. This allows the console and other early stuff to use 47 * time services. 48 */ 49 50static u_int 51dummy_get_timecount(struct timecounter *tc) 52{ 53 static u_int now; 54 55 return (++now); 56} 57 58static struct timecounter dummy_timecounter = { 59 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000 60}; 61 62struct timehands { 63 /* These fields must be initialized by the driver. */ 64 struct timecounter *th_counter; 65 int64_t th_adjustment; 66 uint64_t th_scale; 67 u_int th_offset_count; 68 struct bintime th_offset; 69 struct timeval th_microtime; 70 struct timespec th_nanotime; 71 /* Fields not to be copied in tc_windup start with th_generation. */ 72 volatile u_int th_generation; 73 struct timehands *th_next; 74}; 75 76static struct timehands th0; 77static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0}; 78static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9}; 79static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8}; 80static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7}; 81static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6}; 82static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5}; 83static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4}; 84static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3}; 85static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2}; 86static struct timehands th0 = { 87 &dummy_timecounter, 88 0, 89 (uint64_t)-1 / 1000000, 90 0, 91 {1, 0}, 92 {0, 0}, 93 {0, 0}, 94 1, 95 &th1 96}; 97 98static struct timehands *volatile timehands = &th0; 99struct timecounter *timecounter = &dummy_timecounter; 100static struct timecounter *timecounters = &dummy_timecounter; 101 102int tc_min_ticktock_freq = 1; 103 104time_t time_second = 1; 105time_t time_uptime = 1; 106 107struct bintime boottimebin; 108struct timeval boottime; 109static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS); 110SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD, 111 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime"); 112 113SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, ""); 114static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, ""); 115 116static int timestepwarnings; 117SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW, 118 ×tepwarnings, 0, "Log time steps"); 119 120static void tc_windup(void); 121static void cpu_tick_calibrate(int); 122 123static int 124sysctl_kern_boottime(SYSCTL_HANDLER_ARGS) 125{ 126#ifdef SCTL_MASK32 127 int tv[2]; 128 129 if (req->flags & SCTL_MASK32) { 130 tv[0] = boottime.tv_sec; 131 tv[1] = boottime.tv_usec; 132 return SYSCTL_OUT(req, tv, sizeof(tv)); 133 } else 134#endif 135 return SYSCTL_OUT(req, &boottime, sizeof(boottime)); 136} 137 138static int 139sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS) 140{ 141 u_int ncount; 142 struct timecounter *tc = arg1; 143 144 ncount = tc->tc_get_timecount(tc); 145 return sysctl_handle_int(oidp, &ncount, 0, req); 146} 147 148static int 149sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS) 150{ 151 uint64_t freq; 152 struct timecounter *tc = arg1; 153 154 freq = tc->tc_frequency; 155 return sysctl_handle_64(oidp, &freq, 0, req); 156} 157 158/* 159 * Return the difference between the timehands' counter value now and what 160 * was when we copied it to the timehands' offset_count. 161 */ 162static __inline u_int 163tc_delta(struct timehands *th) 164{ 165 struct timecounter *tc; 166 167 tc = th->th_counter; 168 return ((tc->tc_get_timecount(tc) - th->th_offset_count) & 169 tc->tc_counter_mask); 170} 171 172/* 173 * Functions for reading the time. We have to loop until we are sure that 174 * the timehands that we operated on was not updated under our feet. See 175 * the comment in <sys/time.h> for a description of these 12 functions. 176 */ 177 178#ifdef FFCLOCK 179void 180fbclock_binuptime(struct bintime *bt) 181{ 182 struct timehands *th; 183 unsigned int gen; 184 185 do { 186 th = timehands; 187 gen = th->th_generation; 188 *bt = th->th_offset; 189 bintime_addx(bt, th->th_scale * tc_delta(th)); 190 } while (gen == 0 || gen != th->th_generation); 191} 192 193void 194fbclock_nanouptime(struct timespec *tsp) 195{ 196 struct bintime bt; 197 198 fbclock_binuptime(&bt); 199 bintime2timespec(&bt, tsp); 200} 201 202void 203fbclock_microuptime(struct timeval *tvp) 204{ 205 struct bintime bt; 206 207 fbclock_binuptime(&bt); 208 bintime2timeval(&bt, tvp); 209} 210 211void 212fbclock_bintime(struct bintime *bt) 213{ 214 215 fbclock_binuptime(bt); 216 bintime_add(bt, &boottimebin); 217} 218 219void 220fbclock_nanotime(struct timespec *tsp) 221{ 222 struct bintime bt; 223 224 fbclock_bintime(&bt); 225 bintime2timespec(&bt, tsp); 226} 227 228void 229fbclock_microtime(struct timeval *tvp) 230{ 231 struct bintime bt; 232 233 fbclock_bintime(&bt); 234 bintime2timeval(&bt, tvp); 235} 236 237void 238fbclock_getbinuptime(struct bintime *bt) 239{ 240 struct timehands *th; 241 unsigned int gen; 242 243 do { 244 th = timehands; 245 gen = th->th_generation; 246 *bt = th->th_offset; 247 } while (gen == 0 || gen != th->th_generation); 248} 249 250void 251fbclock_getnanouptime(struct timespec *tsp) 252{ 253 struct timehands *th; 254 unsigned int gen; 255 256 do { 257 th = timehands; 258 gen = th->th_generation; 259 bintime2timespec(&th->th_offset, tsp); 260 } while (gen == 0 || gen != th->th_generation); 261} 262 263void 264fbclock_getmicrouptime(struct timeval *tvp) 265{ 266 struct timehands *th; 267 unsigned int gen; 268 269 do { 270 th = timehands; 271 gen = th->th_generation; 272 bintime2timeval(&th->th_offset, tvp); 273 } while (gen == 0 || gen != th->th_generation); 274} 275 276void 277fbclock_getbintime(struct bintime *bt) 278{ 279 struct timehands *th; 280 unsigned int gen; 281 282 do { 283 th = timehands; 284 gen = th->th_generation; 285 *bt = th->th_offset; 286 } while (gen == 0 || gen != th->th_generation); 287 bintime_add(bt, &boottimebin); 288} 289 290void 291fbclock_getnanotime(struct timespec *tsp) 292{ 293 struct timehands *th; 294 unsigned int gen; 295 296 do { 297 th = timehands; 298 gen = th->th_generation; 299 *tsp = th->th_nanotime; 300 } while (gen == 0 || gen != th->th_generation); 301} 302 303void 304fbclock_getmicrotime(struct timeval *tvp) 305{ 306 struct timehands *th; 307 unsigned int gen; 308 309 do { 310 th = timehands; 311 gen = th->th_generation; 312 *tvp = th->th_microtime; 313 } while (gen == 0 || gen != th->th_generation); 314} 315#else /* !FFCLOCK */ 316void 317binuptime(struct bintime *bt) 318{ 319 struct timehands *th; 320 u_int gen; 321 322 do { 323 th = timehands; 324 gen = th->th_generation; 325 *bt = th->th_offset; 326 bintime_addx(bt, th->th_scale * tc_delta(th)); 327 } while (gen == 0 || gen != th->th_generation); 328} 329 330void 331nanouptime(struct timespec *tsp) 332{ 333 struct bintime bt; 334 335 binuptime(&bt); 336 bintime2timespec(&bt, tsp); 337} 338 339void 340microuptime(struct timeval *tvp) 341{ 342 struct bintime bt; 343 344 binuptime(&bt); 345 bintime2timeval(&bt, tvp); 346} 347 348void 349bintime(struct bintime *bt) 350{ 351 352 binuptime(bt); 353 bintime_add(bt, &boottimebin); 354} 355 356void 357nanotime(struct timespec *tsp) 358{ 359 struct bintime bt; 360 361 bintime(&bt); 362 bintime2timespec(&bt, tsp); 363} 364 365void 366microtime(struct timeval *tvp) 367{ 368 struct bintime bt; 369 370 bintime(&bt); 371 bintime2timeval(&bt, tvp); 372} 373 374void 375getbinuptime(struct bintime *bt) 376{ 377 struct timehands *th; 378 u_int gen; 379 380 do { 381 th = timehands; 382 gen = th->th_generation; 383 *bt = th->th_offset; 384 } while (gen == 0 || gen != th->th_generation); 385} 386 387void 388getnanouptime(struct timespec *tsp) 389{ 390 struct timehands *th; 391 u_int gen; 392 393 do { 394 th = timehands; 395 gen = th->th_generation; 396 bintime2timespec(&th->th_offset, tsp); 397 } while (gen == 0 || gen != th->th_generation); 398} 399 400void 401getmicrouptime(struct timeval *tvp) 402{ 403 struct timehands *th; 404 u_int gen; 405 406 do { 407 th = timehands; 408 gen = th->th_generation; 409 bintime2timeval(&th->th_offset, tvp); 410 } while (gen == 0 || gen != th->th_generation); 411} 412 413void 414getbintime(struct bintime *bt) 415{ 416 struct timehands *th; 417 u_int gen; 418 419 do { 420 th = timehands; 421 gen = th->th_generation; 422 *bt = th->th_offset; 423 } while (gen == 0 || gen != th->th_generation); 424 bintime_add(bt, &boottimebin); 425} 426 427void 428getnanotime(struct timespec *tsp) 429{ 430 struct timehands *th; 431 u_int gen; 432 433 do { 434 th = timehands; 435 gen = th->th_generation; 436 *tsp = th->th_nanotime; 437 } while (gen == 0 || gen != th->th_generation); 438} 439 440void 441getmicrotime(struct timeval *tvp) 442{ 443 struct timehands *th; 444 u_int gen; 445 446 do { 447 th = timehands; 448 gen = th->th_generation; 449 *tvp = th->th_microtime; 450 } while (gen == 0 || gen != th->th_generation); 451} 452#endif /* FFCLOCK */ 453 454#ifdef FFCLOCK 455/* 456 * Support for feed-forward synchronization algorithms. This is heavily inspired 457 * by the timehands mechanism but kept independent from it. *_windup() functions 458 * have some connection to avoid accessing the timecounter hardware more than 459 * necessary. 460 */ 461 462/* Feed-forward clock estimates kept updated by the synchronization daemon. */ 463struct ffclock_estimate ffclock_estimate; 464struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */ 465uint32_t ffclock_status; /* Feed-forward clock status. */ 466int8_t ffclock_updated; /* New estimates are available. */ 467struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */ 468 469struct fftimehands { 470 struct ffclock_estimate cest; 471 struct bintime tick_time; 472 struct bintime tick_time_lerp; 473 ffcounter tick_ffcount; 474 uint64_t period_lerp; 475 volatile uint8_t gen; 476 struct fftimehands *next; 477}; 478 479#define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x)) 480 481static struct fftimehands ffth[10]; 482static struct fftimehands *volatile fftimehands = ffth; 483 484static void 485ffclock_init(void) 486{ 487 struct fftimehands *cur; 488 struct fftimehands *last; 489 490 memset(ffth, 0, sizeof(ffth)); 491 492 last = ffth + NUM_ELEMENTS(ffth) - 1; 493 for (cur = ffth; cur < last; cur++) 494 cur->next = cur + 1; 495 last->next = ffth; 496 497 ffclock_updated = 0; 498 ffclock_status = FFCLOCK_STA_UNSYNC; 499 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF); 500} 501 502/* 503 * Reset the feed-forward clock estimates. Called from inittodr() to get things 504 * kick started and uses the timecounter nominal frequency as a first period 505 * estimate. Note: this function may be called several time just after boot. 506 * Note: this is the only function that sets the value of boot time for the 507 * monotonic (i.e. uptime) version of the feed-forward clock. 508 */ 509void 510ffclock_reset_clock(struct timespec *ts) 511{ 512 struct timecounter *tc; 513 struct ffclock_estimate cest; 514 515 tc = timehands->th_counter; 516 memset(&cest, 0, sizeof(struct ffclock_estimate)); 517 518 timespec2bintime(ts, &ffclock_boottime); 519 timespec2bintime(ts, &(cest.update_time)); 520 ffclock_read_counter(&cest.update_ffcount); 521 cest.leapsec_next = 0; 522 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1; 523 cest.errb_abs = 0; 524 cest.errb_rate = 0; 525 cest.status = FFCLOCK_STA_UNSYNC; 526 cest.leapsec_total = 0; 527 cest.leapsec = 0; 528 529 mtx_lock(&ffclock_mtx); 530 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate)); 531 ffclock_updated = INT8_MAX; 532 mtx_unlock(&ffclock_mtx); 533 534 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name, 535 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec, 536 (unsigned long)ts->tv_nsec); 537} 538 539/* 540 * Sub-routine to convert a time interval measured in RAW counter units to time 541 * in seconds stored in bintime format. 542 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be 543 * larger than the max value of u_int (on 32 bit architecture). Loop to consume 544 * extra cycles. 545 */ 546static void 547ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt) 548{ 549 struct bintime bt2; 550 ffcounter delta, delta_max; 551 552 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1; 553 bintime_clear(bt); 554 do { 555 if (ffdelta > delta_max) 556 delta = delta_max; 557 else 558 delta = ffdelta; 559 bt2.sec = 0; 560 bt2.frac = period; 561 bintime_mul(&bt2, (unsigned int)delta); 562 bintime_add(bt, &bt2); 563 ffdelta -= delta; 564 } while (ffdelta > 0); 565} 566 567/* 568 * Update the fftimehands. 569 * Push the tick ffcount and time(s) forward based on current clock estimate. 570 * The conversion from ffcounter to bintime relies on the difference clock 571 * principle, whose accuracy relies on computing small time intervals. If a new 572 * clock estimate has been passed by the synchronisation daemon, make it 573 * current, and compute the linear interpolation for monotonic time if needed. 574 */ 575static void 576ffclock_windup(unsigned int delta) 577{ 578 struct ffclock_estimate *cest; 579 struct fftimehands *ffth; 580 struct bintime bt, gap_lerp; 581 ffcounter ffdelta; 582 uint64_t frac; 583 unsigned int polling; 584 uint8_t forward_jump, ogen; 585 586 /* 587 * Pick the next timehand, copy current ffclock estimates and move tick 588 * times and counter forward. 589 */ 590 forward_jump = 0; 591 ffth = fftimehands->next; 592 ogen = ffth->gen; 593 ffth->gen = 0; 594 cest = &ffth->cest; 595 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate)); 596 ffdelta = (ffcounter)delta; 597 ffth->period_lerp = fftimehands->period_lerp; 598 599 ffth->tick_time = fftimehands->tick_time; 600 ffclock_convert_delta(ffdelta, cest->period, &bt); 601 bintime_add(&ffth->tick_time, &bt); 602 603 ffth->tick_time_lerp = fftimehands->tick_time_lerp; 604 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt); 605 bintime_add(&ffth->tick_time_lerp, &bt); 606 607 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta; 608 609 /* 610 * Assess the status of the clock, if the last update is too old, it is 611 * likely the synchronisation daemon is dead and the clock is free 612 * running. 613 */ 614 if (ffclock_updated == 0) { 615 ffdelta = ffth->tick_ffcount - cest->update_ffcount; 616 ffclock_convert_delta(ffdelta, cest->period, &bt); 617 if (bt.sec > 2 * FFCLOCK_SKM_SCALE) 618 ffclock_status |= FFCLOCK_STA_UNSYNC; 619 } 620 621 /* 622 * If available, grab updated clock estimates and make them current. 623 * Recompute time at this tick using the updated estimates. The clock 624 * estimates passed the feed-forward synchronisation daemon may result 625 * in time conversion that is not monotonically increasing (just after 626 * the update). time_lerp is a particular linear interpolation over the 627 * synchronisation algo polling period that ensures monotonicity for the 628 * clock ids requesting it. 629 */ 630 if (ffclock_updated > 0) { 631 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate)); 632 ffdelta = ffth->tick_ffcount - cest->update_ffcount; 633 ffth->tick_time = cest->update_time; 634 ffclock_convert_delta(ffdelta, cest->period, &bt); 635 bintime_add(&ffth->tick_time, &bt); 636 637 /* ffclock_reset sets ffclock_updated to INT8_MAX */ 638 if (ffclock_updated == INT8_MAX) 639 ffth->tick_time_lerp = ffth->tick_time; 640 641 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >)) 642 forward_jump = 1; 643 else 644 forward_jump = 0; 645 646 bintime_clear(&gap_lerp); 647 if (forward_jump) { 648 gap_lerp = ffth->tick_time; 649 bintime_sub(&gap_lerp, &ffth->tick_time_lerp); 650 } else { 651 gap_lerp = ffth->tick_time_lerp; 652 bintime_sub(&gap_lerp, &ffth->tick_time); 653 } 654 655 /* 656 * The reset from the RTC clock may be far from accurate, and 657 * reducing the gap between real time and interpolated time 658 * could take a very long time if the interpolated clock insists 659 * on strict monotonicity. The clock is reset under very strict 660 * conditions (kernel time is known to be wrong and 661 * synchronization daemon has been restarted recently. 662 * ffclock_boottime absorbs the jump to ensure boot time is 663 * correct and uptime functions stay consistent. 664 */ 665 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) && 666 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) && 667 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) { 668 if (forward_jump) 669 bintime_add(&ffclock_boottime, &gap_lerp); 670 else 671 bintime_sub(&ffclock_boottime, &gap_lerp); 672 ffth->tick_time_lerp = ffth->tick_time; 673 bintime_clear(&gap_lerp); 674 } 675 676 ffclock_status = cest->status; 677 ffth->period_lerp = cest->period; 678 679 /* 680 * Compute corrected period used for the linear interpolation of 681 * time. The rate of linear interpolation is capped to 5000PPM 682 * (5ms/s). 683 */ 684 if (bintime_isset(&gap_lerp)) { 685 ffdelta = cest->update_ffcount; 686 ffdelta -= fftimehands->cest.update_ffcount; 687 ffclock_convert_delta(ffdelta, cest->period, &bt); 688 polling = bt.sec; 689 bt.sec = 0; 690 bt.frac = 5000000 * (uint64_t)18446744073LL; 691 bintime_mul(&bt, polling); 692 if (bintime_cmp(&gap_lerp, &bt, >)) 693 gap_lerp = bt; 694 695 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */ 696 frac = 0; 697 if (gap_lerp.sec > 0) { 698 frac -= 1; 699 frac /= ffdelta / gap_lerp.sec; 700 } 701 frac += gap_lerp.frac / ffdelta; 702 703 if (forward_jump) 704 ffth->period_lerp += frac; 705 else 706 ffth->period_lerp -= frac; 707 } 708 709 ffclock_updated = 0; 710 } 711 if (++ogen == 0) 712 ogen = 1; 713 ffth->gen = ogen; 714 fftimehands = ffth; 715} 716 717/* 718 * Adjust the fftimehands when the timecounter is changed. Stating the obvious, 719 * the old and new hardware counter cannot be read simultaneously. tc_windup() 720 * does read the two counters 'back to back', but a few cycles are effectively 721 * lost, and not accumulated in tick_ffcount. This is a fairly radical 722 * operation for a feed-forward synchronization daemon, and it is its job to not 723 * pushing irrelevant data to the kernel. Because there is no locking here, 724 * simply force to ignore pending or next update to give daemon a chance to 725 * realize the counter has changed. 726 */ 727static void 728ffclock_change_tc(struct timehands *th) 729{ 730 struct fftimehands *ffth; 731 struct ffclock_estimate *cest; 732 struct timecounter *tc; 733 uint8_t ogen; 734 735 tc = th->th_counter; 736 ffth = fftimehands->next; 737 ogen = ffth->gen; 738 ffth->gen = 0; 739 740 cest = &ffth->cest; 741 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate)); 742 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1; 743 cest->errb_abs = 0; 744 cest->errb_rate = 0; 745 cest->status |= FFCLOCK_STA_UNSYNC; 746 747 ffth->tick_ffcount = fftimehands->tick_ffcount; 748 ffth->tick_time_lerp = fftimehands->tick_time_lerp; 749 ffth->tick_time = fftimehands->tick_time; 750 ffth->period_lerp = cest->period; 751 752 /* Do not lock but ignore next update from synchronization daemon. */ 753 ffclock_updated--; 754 755 if (++ogen == 0) 756 ogen = 1; 757 ffth->gen = ogen; 758 fftimehands = ffth; 759} 760 761/* 762 * Retrieve feed-forward counter and time of last kernel tick. 763 */ 764void 765ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags) 766{ 767 struct fftimehands *ffth; 768 uint8_t gen; 769 770 /* 771 * No locking but check generation has not changed. Also need to make 772 * sure ffdelta is positive, i.e. ffcount > tick_ffcount. 773 */ 774 do { 775 ffth = fftimehands; 776 gen = ffth->gen; 777 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) 778 *bt = ffth->tick_time_lerp; 779 else 780 *bt = ffth->tick_time; 781 *ffcount = ffth->tick_ffcount; 782 } while (gen == 0 || gen != ffth->gen); 783} 784 785/* 786 * Absolute clock conversion. Low level function to convert ffcounter to 787 * bintime. The ffcounter is converted using the current ffclock period estimate 788 * or the "interpolated period" to ensure monotonicity. 789 * NOTE: this conversion may have been deferred, and the clock updated since the 790 * hardware counter has been read. 791 */ 792void 793ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags) 794{ 795 struct fftimehands *ffth; 796 struct bintime bt2; 797 ffcounter ffdelta; 798 uint8_t gen; 799 800 /* 801 * No locking but check generation has not changed. Also need to make 802 * sure ffdelta is positive, i.e. ffcount > tick_ffcount. 803 */ 804 do { 805 ffth = fftimehands; 806 gen = ffth->gen; 807 if (ffcount > ffth->tick_ffcount) 808 ffdelta = ffcount - ffth->tick_ffcount; 809 else 810 ffdelta = ffth->tick_ffcount - ffcount; 811 812 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) { 813 *bt = ffth->tick_time_lerp; 814 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2); 815 } else { 816 *bt = ffth->tick_time; 817 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2); 818 } 819 820 if (ffcount > ffth->tick_ffcount) 821 bintime_add(bt, &bt2); 822 else 823 bintime_sub(bt, &bt2); 824 } while (gen == 0 || gen != ffth->gen); 825} 826 827/* 828 * Difference clock conversion. 829 * Low level function to Convert a time interval measured in RAW counter units 830 * into bintime. The difference clock allows measuring small intervals much more 831 * reliably than the absolute clock. 832 */ 833void 834ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt) 835{ 836 struct fftimehands *ffth; 837 uint8_t gen; 838 839 /* No locking but check generation has not changed. */ 840 do { 841 ffth = fftimehands; 842 gen = ffth->gen; 843 ffclock_convert_delta(ffdelta, ffth->cest.period, bt); 844 } while (gen == 0 || gen != ffth->gen); 845} 846 847/* 848 * Access to current ffcounter value. 849 */ 850void 851ffclock_read_counter(ffcounter *ffcount) 852{ 853 struct timehands *th; 854 struct fftimehands *ffth; 855 unsigned int gen, delta; 856 857 /* 858 * ffclock_windup() called from tc_windup(), safe to rely on 859 * th->th_generation only, for correct delta and ffcounter. 860 */ 861 do { 862 th = timehands; 863 gen = th->th_generation; 864 ffth = fftimehands; 865 delta = tc_delta(th); 866 *ffcount = ffth->tick_ffcount; 867 } while (gen == 0 || gen != th->th_generation); 868 869 *ffcount += delta; 870} 871 872void 873binuptime(struct bintime *bt) 874{ 875 876 binuptime_fromclock(bt, sysclock_active); 877} 878 879void 880nanouptime(struct timespec *tsp) 881{ 882 883 nanouptime_fromclock(tsp, sysclock_active); 884} 885 886void 887microuptime(struct timeval *tvp) 888{ 889 890 microuptime_fromclock(tvp, sysclock_active); 891} 892 893void 894bintime(struct bintime *bt) 895{ 896 897 bintime_fromclock(bt, sysclock_active); 898} 899 900void 901nanotime(struct timespec *tsp) 902{ 903 904 nanotime_fromclock(tsp, sysclock_active); 905} 906 907void 908microtime(struct timeval *tvp) 909{ 910 911 microtime_fromclock(tvp, sysclock_active); 912} 913 914void 915getbinuptime(struct bintime *bt) 916{ 917 918 getbinuptime_fromclock(bt, sysclock_active); 919} 920 921void 922getnanouptime(struct timespec *tsp) 923{ 924 925 getnanouptime_fromclock(tsp, sysclock_active); 926} 927 928void 929getmicrouptime(struct timeval *tvp) 930{ 931 932 getmicrouptime_fromclock(tvp, sysclock_active); 933} 934 935void 936getbintime(struct bintime *bt) 937{ 938 939 getbintime_fromclock(bt, sysclock_active); 940} 941 942void 943getnanotime(struct timespec *tsp) 944{ 945 946 getnanotime_fromclock(tsp, sysclock_active); 947} 948 949void 950getmicrotime(struct timeval *tvp) 951{ 952 953 getmicrouptime_fromclock(tvp, sysclock_active); 954} 955 956#endif /* FFCLOCK */ 957 958/* 959 * System clock currently providing time to the system. Modifiable via sysctl 960 * when the FFCLOCK option is defined. 961 */ 962int sysclock_active = SYSCLOCK_FBCK; 963 964/* Internal NTP status and error estimates. */ 965extern int time_status; 966extern long time_esterror; 967 968/* 969 * Take a snapshot of sysclock data which can be used to compare system clocks 970 * and generate timestamps after the fact. 971 */ 972void 973sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast) 974{ 975 struct fbclock_info *fbi; 976 struct timehands *th; 977 struct bintime bt; 978 unsigned int delta, gen; 979#ifdef FFCLOCK 980 ffcounter ffcount; 981 struct fftimehands *ffth; 982 struct ffclock_info *ffi; 983 struct ffclock_estimate cest; 984 985 ffi = &clock_snap->ff_info; 986#endif 987 988 fbi = &clock_snap->fb_info; 989 delta = 0; 990 991 do { 992 th = timehands; 993 gen = th->th_generation; 994 fbi->th_scale = th->th_scale; 995 fbi->tick_time = th->th_offset; 996#ifdef FFCLOCK 997 ffth = fftimehands; 998 ffi->tick_time = ffth->tick_time_lerp; 999 ffi->tick_time_lerp = ffth->tick_time_lerp; 1000 ffi->period = ffth->cest.period; 1001 ffi->period_lerp = ffth->period_lerp; 1002 clock_snap->ffcount = ffth->tick_ffcount; 1003 cest = ffth->cest; 1004#endif 1005 if (!fast) 1006 delta = tc_delta(th); 1007 } while (gen == 0 || gen != th->th_generation); 1008 1009 clock_snap->delta = delta; 1010 clock_snap->sysclock_active = sysclock_active; 1011 1012 /* Record feedback clock status and error. */ 1013 clock_snap->fb_info.status = time_status; 1014 /* XXX: Very crude estimate of feedback clock error. */ 1015 bt.sec = time_esterror / 1000000; 1016 bt.frac = ((time_esterror - bt.sec) * 1000000) * 1017 (uint64_t)18446744073709ULL; 1018 clock_snap->fb_info.error = bt; 1019 1020#ifdef FFCLOCK 1021 if (!fast) 1022 clock_snap->ffcount += delta; 1023 1024 /* Record feed-forward clock leap second adjustment. */ 1025 ffi->leapsec_adjustment = cest.leapsec_total; 1026 if (clock_snap->ffcount > cest.leapsec_next) 1027 ffi->leapsec_adjustment -= cest.leapsec; 1028 1029 /* Record feed-forward clock status and error. */ 1030 clock_snap->ff_info.status = cest.status; 1031 ffcount = clock_snap->ffcount - cest.update_ffcount; 1032 ffclock_convert_delta(ffcount, cest.period, &bt); 1033 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */ 1034 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL); 1035 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */ 1036 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL); 1037 clock_snap->ff_info.error = bt; 1038#endif 1039} 1040 1041/* 1042 * Convert a sysclock snapshot into a struct bintime based on the specified 1043 * clock source and flags. 1044 */ 1045int 1046sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt, 1047 int whichclock, uint32_t flags) 1048{ 1049#ifdef FFCLOCK 1050 struct bintime bt2; 1051 uint64_t period; 1052#endif 1053 1054 switch (whichclock) { 1055 case SYSCLOCK_FBCK: 1056 *bt = cs->fb_info.tick_time; 1057 1058 /* If snapshot was created with !fast, delta will be >0. */ 1059 if (cs->delta > 0) 1060 bintime_addx(bt, cs->fb_info.th_scale * cs->delta); 1061 1062 if ((flags & FBCLOCK_UPTIME) == 0) 1063 bintime_add(bt, &boottimebin); 1064 break; 1065#ifdef FFCLOCK 1066 case SYSCLOCK_FFWD: 1067 if (flags & FFCLOCK_LERP) { 1068 *bt = cs->ff_info.tick_time_lerp; 1069 period = cs->ff_info.period_lerp; 1070 } else { 1071 *bt = cs->ff_info.tick_time; 1072 period = cs->ff_info.period; 1073 } 1074 1075 /* If snapshot was created with !fast, delta will be >0. */ 1076 if (cs->delta > 0) { 1077 ffclock_convert_delta(cs->delta, period, &bt2); 1078 bintime_add(bt, &bt2); 1079 } 1080 1081 /* Leap second adjustment. */ 1082 if (flags & FFCLOCK_LEAPSEC) 1083 bt->sec -= cs->ff_info.leapsec_adjustment; 1084 1085 /* Boot time adjustment, for uptime/monotonic clocks. */ 1086 if (flags & FFCLOCK_UPTIME) 1087 bintime_sub(bt, &ffclock_boottime); 1088 break; 1089#endif 1090 default: 1091 return (EINVAL); 1092 break; 1093 } 1094 1095 return (0); 1096} 1097 1098/* 1099 * Initialize a new timecounter and possibly use it. 1100 */ 1101void 1102tc_init(struct timecounter *tc) 1103{ 1104 u_int u; 1105 struct sysctl_oid *tc_root; 1106 1107 u = tc->tc_frequency / tc->tc_counter_mask; 1108 /* XXX: We need some margin here, 10% is a guess */ 1109 u *= 11; 1110 u /= 10; 1111 if (u > hz && tc->tc_quality >= 0) { 1112 tc->tc_quality = -2000; 1113 if (bootverbose) { 1114 printf("Timecounter \"%s\" frequency %ju Hz", 1115 tc->tc_name, (uintmax_t)tc->tc_frequency); 1116 printf(" -- Insufficient hz, needs at least %u\n", u); 1117 } 1118 } else if (tc->tc_quality >= 0 || bootverbose) { 1119 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n", 1120 tc->tc_name, (uintmax_t)tc->tc_frequency, 1121 tc->tc_quality); 1122 } 1123 1124 tc->tc_next = timecounters; 1125 timecounters = tc; 1126 /* 1127 * Set up sysctl tree for this counter. 1128 */ 1129 tc_root = SYSCTL_ADD_NODE(NULL, 1130 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name, 1131 CTLFLAG_RW, 0, "timecounter description"); 1132 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO, 1133 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0, 1134 "mask for implemented bits"); 1135 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO, 1136 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc), 1137 sysctl_kern_timecounter_get, "IU", "current timecounter value"); 1138 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO, 1139 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc), 1140 sysctl_kern_timecounter_freq, "QU", "timecounter frequency"); 1141 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO, 1142 "quality", CTLFLAG_RD, &(tc->tc_quality), 0, 1143 "goodness of time counter"); 1144 /* 1145 * Never automatically use a timecounter with negative quality. 1146 * Even though we run on the dummy counter, switching here may be 1147 * worse since this timecounter may not be monotonous. 1148 */ 1149 if (tc->tc_quality < 0) 1150 return; 1151 if (tc->tc_quality < timecounter->tc_quality) 1152 return; 1153 if (tc->tc_quality == timecounter->tc_quality && 1154 tc->tc_frequency < timecounter->tc_frequency) 1155 return; 1156 (void)tc->tc_get_timecount(tc); 1157 (void)tc->tc_get_timecount(tc); 1158 timecounter = tc; 1159} 1160 1161/* Report the frequency of the current timecounter. */ 1162uint64_t 1163tc_getfrequency(void) 1164{ 1165 1166 return (timehands->th_counter->tc_frequency); 1167} 1168 1169/* 1170 * Step our concept of UTC. This is done by modifying our estimate of 1171 * when we booted. 1172 * XXX: not locked. 1173 */ 1174void 1175tc_setclock(struct timespec *ts) 1176{ 1177 struct timespec tbef, taft; 1178 struct bintime bt, bt2; 1179 1180 cpu_tick_calibrate(1); 1181 nanotime(&tbef); 1182 timespec2bintime(ts, &bt); 1183 binuptime(&bt2); 1184 bintime_sub(&bt, &bt2); 1185 bintime_add(&bt2, &boottimebin); 1186 boottimebin = bt; 1187 bintime2timeval(&bt, &boottime); 1188 1189 /* XXX fiddle all the little crinkly bits around the fiords... */ 1190 tc_windup(); 1191 nanotime(&taft); 1192 if (timestepwarnings) { 1193 log(LOG_INFO, 1194 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n", 1195 (intmax_t)tbef.tv_sec, tbef.tv_nsec, 1196 (intmax_t)taft.tv_sec, taft.tv_nsec, 1197 (intmax_t)ts->tv_sec, ts->tv_nsec); 1198 } 1199 cpu_tick_calibrate(1); 1200} 1201 1202/* 1203 * Initialize the next struct timehands in the ring and make 1204 * it the active timehands. Along the way we might switch to a different 1205 * timecounter and/or do seconds processing in NTP. Slightly magic. 1206 */ 1207static void 1208tc_windup(void) 1209{ 1210 struct bintime bt; 1211 struct timehands *th, *tho; 1212 uint64_t scale; 1213 u_int delta, ncount, ogen; 1214 int i; 1215 time_t t; 1216 1217 /* 1218 * Make the next timehands a copy of the current one, but do not 1219 * overwrite the generation or next pointer. While we update 1220 * the contents, the generation must be zero. 1221 */ 1222 tho = timehands; 1223 th = tho->th_next; 1224 ogen = th->th_generation; 1225 th->th_generation = 0; 1226 bcopy(tho, th, offsetof(struct timehands, th_generation)); 1227 1228 /* 1229 * Capture a timecounter delta on the current timecounter and if 1230 * changing timecounters, a counter value from the new timecounter. 1231 * Update the offset fields accordingly. 1232 */ 1233 delta = tc_delta(th); 1234 if (th->th_counter != timecounter) 1235 ncount = timecounter->tc_get_timecount(timecounter); 1236 else 1237 ncount = 0; 1238#ifdef FFCLOCK 1239 ffclock_windup(delta); 1240#endif 1241 th->th_offset_count += delta; 1242 th->th_offset_count &= th->th_counter->tc_counter_mask; 1243 while (delta > th->th_counter->tc_frequency) { 1244 /* Eat complete unadjusted seconds. */ 1245 delta -= th->th_counter->tc_frequency; 1246 th->th_offset.sec++; 1247 } 1248 if ((delta > th->th_counter->tc_frequency / 2) && 1249 (th->th_scale * delta < ((uint64_t)1 << 63))) { 1250 /* The product th_scale * delta just barely overflows. */ 1251 th->th_offset.sec++; 1252 } 1253 bintime_addx(&th->th_offset, th->th_scale * delta); 1254 1255 /* 1256 * Hardware latching timecounters may not generate interrupts on 1257 * PPS events, so instead we poll them. There is a finite risk that 1258 * the hardware might capture a count which is later than the one we 1259 * got above, and therefore possibly in the next NTP second which might 1260 * have a different rate than the current NTP second. It doesn't 1261 * matter in practice. 1262 */ 1263 if (tho->th_counter->tc_poll_pps) 1264 tho->th_counter->tc_poll_pps(tho->th_counter); 1265 1266 /* 1267 * Deal with NTP second processing. The for loop normally 1268 * iterates at most once, but in extreme situations it might 1269 * keep NTP sane if timeouts are not run for several seconds. 1270 * At boot, the time step can be large when the TOD hardware 1271 * has been read, so on really large steps, we call 1272 * ntp_update_second only twice. We need to call it twice in 1273 * case we missed a leap second. 1274 */ 1275 bt = th->th_offset; 1276 bintime_add(&bt, &boottimebin); 1277 i = bt.sec - tho->th_microtime.tv_sec; 1278 if (i > LARGE_STEP) 1279 i = 2; 1280 for (; i > 0; i--) { 1281 t = bt.sec; 1282 ntp_update_second(&th->th_adjustment, &bt.sec); 1283 if (bt.sec != t) 1284 boottimebin.sec += bt.sec - t; 1285 } 1286 /* Update the UTC timestamps used by the get*() functions. */ 1287 /* XXX shouldn't do this here. Should force non-`get' versions. */ 1288 bintime2timeval(&bt, &th->th_microtime); 1289 bintime2timespec(&bt, &th->th_nanotime); 1290 1291 /* Now is a good time to change timecounters. */ 1292 if (th->th_counter != timecounter) { 1293#ifndef __arm__ 1294 if ((timecounter->tc_flags & TC_FLAGS_C3STOP) != 0) 1295 cpu_disable_deep_sleep++; 1296 if ((th->th_counter->tc_flags & TC_FLAGS_C3STOP) != 0) 1297 cpu_disable_deep_sleep--; 1298#endif 1299 th->th_counter = timecounter; 1300 th->th_offset_count = ncount; 1301 tc_min_ticktock_freq = max(1, timecounter->tc_frequency / 1302 (((uint64_t)timecounter->tc_counter_mask + 1) / 3)); 1303#ifdef FFCLOCK 1304 ffclock_change_tc(th); 1305#endif 1306 } 1307 1308 /*- 1309 * Recalculate the scaling factor. We want the number of 1/2^64 1310 * fractions of a second per period of the hardware counter, taking 1311 * into account the th_adjustment factor which the NTP PLL/adjtime(2) 1312 * processing provides us with. 1313 * 1314 * The th_adjustment is nanoseconds per second with 32 bit binary 1315 * fraction and we want 64 bit binary fraction of second: 1316 * 1317 * x = a * 2^32 / 10^9 = a * 4.294967296 1318 * 1319 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int 1320 * we can only multiply by about 850 without overflowing, that 1321 * leaves no suitably precise fractions for multiply before divide. 1322 * 1323 * Divide before multiply with a fraction of 2199/512 results in a 1324 * systematic undercompensation of 10PPM of th_adjustment. On a 1325 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable. 1326 * 1327 * We happily sacrifice the lowest of the 64 bits of our result 1328 * to the goddess of code clarity. 1329 * 1330 */ 1331 scale = (uint64_t)1 << 63; 1332 scale += (th->th_adjustment / 1024) * 2199; 1333 scale /= th->th_counter->tc_frequency; 1334 th->th_scale = scale * 2; 1335 1336 /* 1337 * Now that the struct timehands is again consistent, set the new 1338 * generation number, making sure to not make it zero. 1339 */ 1340 if (++ogen == 0) 1341 ogen = 1; 1342 th->th_generation = ogen; 1343 1344 /* Go live with the new struct timehands. */ 1345#ifdef FFCLOCK 1346 switch (sysclock_active) { 1347 case SYSCLOCK_FBCK: 1348#endif 1349 time_second = th->th_microtime.tv_sec; 1350 time_uptime = th->th_offset.sec; 1351#ifdef FFCLOCK 1352 break; 1353 case SYSCLOCK_FFWD: 1354 time_second = fftimehands->tick_time_lerp.sec; 1355 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec; 1356 break; 1357 } 1358#endif 1359 1360 timehands = th; 1361} 1362 1363/* Report or change the active timecounter hardware. */ 1364static int 1365sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS) 1366{ 1367 char newname[32]; 1368 struct timecounter *newtc, *tc; 1369 int error; 1370 1371 tc = timecounter; 1372 strlcpy(newname, tc->tc_name, sizeof(newname)); 1373 1374 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req); 1375 if (error != 0 || req->newptr == NULL || 1376 strcmp(newname, tc->tc_name) == 0) 1377 return (error); 1378 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) { 1379 if (strcmp(newname, newtc->tc_name) != 0) 1380 continue; 1381 1382 /* Warm up new timecounter. */ 1383 (void)newtc->tc_get_timecount(newtc); 1384 (void)newtc->tc_get_timecount(newtc); 1385 1386 timecounter = newtc; 1387 return (0); 1388 } 1389 return (EINVAL); 1390} 1391 1392SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW, 1393 0, 0, sysctl_kern_timecounter_hardware, "A", 1394 "Timecounter hardware selected"); 1395 1396 1397/* Report or change the active timecounter hardware. */ 1398static int 1399sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS) 1400{ 1401 char buf[32], *spc; 1402 struct timecounter *tc; 1403 int error; 1404 1405 spc = ""; 1406 error = 0; 1407 for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) { 1408 sprintf(buf, "%s%s(%d)", 1409 spc, tc->tc_name, tc->tc_quality); 1410 error = SYSCTL_OUT(req, buf, strlen(buf)); 1411 spc = " "; 1412 } 1413 return (error); 1414} 1415 1416SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD, 1417 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected"); 1418 1419/* 1420 * RFC 2783 PPS-API implementation. 1421 */ 1422 1423int 1424pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1425{ 1426 pps_params_t *app; 1427 struct pps_fetch_args *fapi; 1428#ifdef FFCLOCK 1429 struct pps_fetch_ffc_args *fapi_ffc; 1430#endif 1431#ifdef PPS_SYNC 1432 struct pps_kcbind_args *kapi; 1433#endif 1434 1435 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl")); 1436 switch (cmd) { 1437 case PPS_IOC_CREATE: 1438 return (0); 1439 case PPS_IOC_DESTROY: 1440 return (0); 1441 case PPS_IOC_SETPARAMS: 1442 app = (pps_params_t *)data; 1443 if (app->mode & ~pps->ppscap) 1444 return (EINVAL); 1445#ifdef FFCLOCK 1446 /* Ensure only a single clock is selected for ffc timestamp. */ 1447 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK) 1448 return (EINVAL); 1449#endif 1450 pps->ppsparam = *app; 1451 return (0); 1452 case PPS_IOC_GETPARAMS: 1453 app = (pps_params_t *)data; 1454 *app = pps->ppsparam; 1455 app->api_version = PPS_API_VERS_1; 1456 return (0); 1457 case PPS_IOC_GETCAP: 1458 *(int*)data = pps->ppscap; 1459 return (0); 1460 case PPS_IOC_FETCH: 1461 fapi = (struct pps_fetch_args *)data; 1462 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1463 return (EINVAL); 1464 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1465 return (EOPNOTSUPP); 1466 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1467 fapi->pps_info_buf = pps->ppsinfo; 1468 return (0); 1469#ifdef FFCLOCK 1470 case PPS_IOC_FETCH_FFCOUNTER: 1471 fapi_ffc = (struct pps_fetch_ffc_args *)data; 1472 if (fapi_ffc->tsformat && fapi_ffc->tsformat != 1473 PPS_TSFMT_TSPEC) 1474 return (EINVAL); 1475 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec) 1476 return (EOPNOTSUPP); 1477 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode; 1478 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc; 1479 /* Overwrite timestamps if feedback clock selected. */ 1480 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) { 1481 case PPS_TSCLK_FBCK: 1482 fapi_ffc->pps_info_buf_ffc.assert_timestamp = 1483 pps->ppsinfo.assert_timestamp; 1484 fapi_ffc->pps_info_buf_ffc.clear_timestamp = 1485 pps->ppsinfo.clear_timestamp; 1486 break; 1487 case PPS_TSCLK_FFWD: 1488 break; 1489 default: 1490 break; 1491 } 1492 return (0); 1493#endif /* FFCLOCK */ 1494 case PPS_IOC_KCBIND: 1495#ifdef PPS_SYNC 1496 kapi = (struct pps_kcbind_args *)data; 1497 /* XXX Only root should be able to do this */ 1498 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1499 return (EINVAL); 1500 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1501 return (EINVAL); 1502 if (kapi->edge & ~pps->ppscap) 1503 return (EINVAL); 1504 pps->kcmode = kapi->edge; 1505 return (0); 1506#else 1507 return (EOPNOTSUPP); 1508#endif 1509 default: 1510 return (ENOIOCTL); 1511 } 1512} 1513 1514void 1515pps_init(struct pps_state *pps) 1516{ 1517 pps->ppscap |= PPS_TSFMT_TSPEC; 1518 if (pps->ppscap & PPS_CAPTUREASSERT) 1519 pps->ppscap |= PPS_OFFSETASSERT; 1520 if (pps->ppscap & PPS_CAPTURECLEAR) 1521 pps->ppscap |= PPS_OFFSETCLEAR; 1522#ifdef FFCLOCK 1523 pps->ppscap |= PPS_TSCLK_MASK; 1524#endif 1525} 1526 1527void 1528pps_capture(struct pps_state *pps) 1529{ 1530 struct timehands *th; 1531 1532 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture")); 1533 th = timehands; 1534 pps->capgen = th->th_generation; 1535 pps->capth = th; 1536#ifdef FFCLOCK 1537 pps->capffth = fftimehands; 1538#endif 1539 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter); 1540 if (pps->capgen != th->th_generation) 1541 pps->capgen = 0; 1542} 1543 1544void 1545pps_event(struct pps_state *pps, int event) 1546{ 1547 struct bintime bt; 1548 struct timespec ts, *tsp, *osp; 1549 u_int tcount, *pcount; 1550 int foff, fhard; 1551 pps_seq_t *pseq; 1552#ifdef FFCLOCK 1553 struct timespec *tsp_ffc; 1554 pps_seq_t *pseq_ffc; 1555 ffcounter *ffcount; 1556#endif 1557 1558 KASSERT(pps != NULL, ("NULL pps pointer in pps_event")); 1559 /* If the timecounter was wound up underneath us, bail out. */ 1560 if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation) 1561 return; 1562 1563 /* Things would be easier with arrays. */ 1564 if (event == PPS_CAPTUREASSERT) { 1565 tsp = &pps->ppsinfo.assert_timestamp; 1566 osp = &pps->ppsparam.assert_offset; 1567 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1568 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1569 pcount = &pps->ppscount[0]; 1570 pseq = &pps->ppsinfo.assert_sequence; 1571#ifdef FFCLOCK 1572 ffcount = &pps->ppsinfo_ffc.assert_ffcount; 1573 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp; 1574 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence; 1575#endif 1576 } else { 1577 tsp = &pps->ppsinfo.clear_timestamp; 1578 osp = &pps->ppsparam.clear_offset; 1579 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1580 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1581 pcount = &pps->ppscount[1]; 1582 pseq = &pps->ppsinfo.clear_sequence; 1583#ifdef FFCLOCK 1584 ffcount = &pps->ppsinfo_ffc.clear_ffcount; 1585 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp; 1586 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence; 1587#endif 1588 } 1589 1590 /* 1591 * If the timecounter changed, we cannot compare the count values, so 1592 * we have to drop the rest of the PPS-stuff until the next event. 1593 */ 1594 if (pps->ppstc != pps->capth->th_counter) { 1595 pps->ppstc = pps->capth->th_counter; 1596 *pcount = pps->capcount; 1597 pps->ppscount[2] = pps->capcount; 1598 return; 1599 } 1600 1601 /* Convert the count to a timespec. */ 1602 tcount = pps->capcount - pps->capth->th_offset_count; 1603 tcount &= pps->capth->th_counter->tc_counter_mask; 1604 bt = pps->capth->th_offset; 1605 bintime_addx(&bt, pps->capth->th_scale * tcount); 1606 bintime_add(&bt, &boottimebin); 1607 bintime2timespec(&bt, &ts); 1608 1609 /* If the timecounter was wound up underneath us, bail out. */ 1610 if (pps->capgen != pps->capth->th_generation) 1611 return; 1612 1613 *pcount = pps->capcount; 1614 (*pseq)++; 1615 *tsp = ts; 1616 1617 if (foff) { 1618 timespecadd(tsp, osp); 1619 if (tsp->tv_nsec < 0) { 1620 tsp->tv_nsec += 1000000000; 1621 tsp->tv_sec -= 1; 1622 } 1623 } 1624 1625#ifdef FFCLOCK 1626 *ffcount = pps->capffth->tick_ffcount + tcount; 1627 bt = pps->capffth->tick_time; 1628 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt); 1629 bintime_add(&bt, &pps->capffth->tick_time); 1630 bintime2timespec(&bt, &ts); 1631 (*pseq_ffc)++; 1632 *tsp_ffc = ts; 1633#endif 1634 1635#ifdef PPS_SYNC 1636 if (fhard) { 1637 uint64_t scale; 1638 1639 /* 1640 * Feed the NTP PLL/FLL. 1641 * The FLL wants to know how many (hardware) nanoseconds 1642 * elapsed since the previous event. 1643 */ 1644 tcount = pps->capcount - pps->ppscount[2]; 1645 pps->ppscount[2] = pps->capcount; 1646 tcount &= pps->capth->th_counter->tc_counter_mask; 1647 scale = (uint64_t)1 << 63; 1648 scale /= pps->capth->th_counter->tc_frequency; 1649 scale *= 2; 1650 bt.sec = 0; 1651 bt.frac = 0; 1652 bintime_addx(&bt, scale * tcount); 1653 bintime2timespec(&bt, &ts); 1654 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec); 1655 } 1656#endif 1657} 1658 1659/* 1660 * Timecounters need to be updated every so often to prevent the hardware 1661 * counter from overflowing. Updating also recalculates the cached values 1662 * used by the get*() family of functions, so their precision depends on 1663 * the update frequency. 1664 */ 1665 1666static int tc_tick; 1667SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, 1668 "Approximate number of hardclock ticks in a millisecond"); 1669 1670void 1671tc_ticktock(int cnt) 1672{ 1673 static int count; 1674 1675 count += cnt; 1676 if (count < tc_tick) 1677 return; 1678 count = 0; 1679 tc_windup(); 1680} 1681 1682static void 1683inittimecounter(void *dummy) 1684{ 1685 u_int p; 1686 1687 /* 1688 * Set the initial timeout to 1689 * max(1, <approx. number of hardclock ticks in a millisecond>). 1690 * People should probably not use the sysctl to set the timeout 1691 * to smaller than its inital value, since that value is the 1692 * smallest reasonable one. If they want better timestamps they 1693 * should use the non-"get"* functions. 1694 */ 1695 if (hz > 1000) 1696 tc_tick = (hz + 500) / 1000; 1697 else 1698 tc_tick = 1; 1699 p = (tc_tick * 1000000) / hz; 1700 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000); 1701 1702#ifdef FFCLOCK 1703 ffclock_init(); 1704#endif 1705 /* warm up new timecounter (again) and get rolling. */ 1706 (void)timecounter->tc_get_timecount(timecounter); 1707 (void)timecounter->tc_get_timecount(timecounter); 1708 tc_windup(); 1709} 1710 1711SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL); 1712 1713/* Cpu tick handling -------------------------------------------------*/ 1714 1715static int cpu_tick_variable; 1716static uint64_t cpu_tick_frequency; 1717 1718static uint64_t 1719tc_cpu_ticks(void) 1720{ 1721 static uint64_t base; 1722 static unsigned last; 1723 unsigned u; 1724 struct timecounter *tc; 1725 1726 tc = timehands->th_counter; 1727 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask; 1728 if (u < last) 1729 base += (uint64_t)tc->tc_counter_mask + 1; 1730 last = u; 1731 return (u + base); 1732} 1733 1734void 1735cpu_tick_calibration(void) 1736{ 1737 static time_t last_calib; 1738 1739 if (time_uptime != last_calib && !(time_uptime & 0xf)) { 1740 cpu_tick_calibrate(0); 1741 last_calib = time_uptime; 1742 } 1743} 1744 1745/* 1746 * This function gets called every 16 seconds on only one designated 1747 * CPU in the system from hardclock() via cpu_tick_calibration()(). 1748 * 1749 * Whenever the real time clock is stepped we get called with reset=1 1750 * to make sure we handle suspend/resume and similar events correctly. 1751 */ 1752 1753static void 1754cpu_tick_calibrate(int reset) 1755{ 1756 static uint64_t c_last; 1757 uint64_t c_this, c_delta; 1758 static struct bintime t_last; 1759 struct bintime t_this, t_delta; 1760 uint32_t divi; 1761 1762 if (reset) { 1763 /* The clock was stepped, abort & reset */ 1764 t_last.sec = 0; 1765 return; 1766 } 1767 1768 /* we don't calibrate fixed rate cputicks */ 1769 if (!cpu_tick_variable) 1770 return; 1771 1772 getbinuptime(&t_this); 1773 c_this = cpu_ticks(); 1774 if (t_last.sec != 0) { 1775 c_delta = c_this - c_last; 1776 t_delta = t_this; 1777 bintime_sub(&t_delta, &t_last); 1778 /* 1779 * Headroom: 1780 * 2^(64-20) / 16[s] = 1781 * 2^(44) / 16[s] = 1782 * 17.592.186.044.416 / 16 = 1783 * 1.099.511.627.776 [Hz] 1784 */ 1785 divi = t_delta.sec << 20; 1786 divi |= t_delta.frac >> (64 - 20); 1787 c_delta <<= 20; 1788 c_delta /= divi; 1789 if (c_delta > cpu_tick_frequency) { 1790 if (0 && bootverbose) 1791 printf("cpu_tick increased to %ju Hz\n", 1792 c_delta); 1793 cpu_tick_frequency = c_delta; 1794 } 1795 } 1796 c_last = c_this; 1797 t_last = t_this; 1798} 1799 1800void 1801set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var) 1802{ 1803 1804 if (func == NULL) { 1805 cpu_ticks = tc_cpu_ticks; 1806 } else { 1807 cpu_tick_frequency = freq; 1808 cpu_tick_variable = var; 1809 cpu_ticks = func; 1810 } 1811} 1812 1813uint64_t 1814cpu_tickrate(void) 1815{ 1816 1817 if (cpu_ticks == tc_cpu_ticks) 1818 return (tc_getfrequency()); 1819 return (cpu_tick_frequency); 1820} 1821 1822/* 1823 * We need to be slightly careful converting cputicks to microseconds. 1824 * There is plenty of margin in 64 bits of microseconds (half a million 1825 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply 1826 * before divide conversion (to retain precision) we find that the 1827 * margin shrinks to 1.5 hours (one millionth of 146y). 1828 * With a three prong approach we never lose significant bits, no 1829 * matter what the cputick rate and length of timeinterval is. 1830 */ 1831 1832uint64_t 1833cputick2usec(uint64_t tick) 1834{ 1835 1836 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */ 1837 return (tick / (cpu_tickrate() / 1000000LL)); 1838 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */ 1839 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL)); 1840 else 1841 return ((tick * 1000000LL) / cpu_tickrate()); 1842} 1843 1844cpu_tick_f *cpu_ticks = tc_cpu_ticks; 1845