1 /*- 2 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> 3 * Copyright (c) 1982, 1986, 1991, 1993 4 * The Regents of the University of California. All rights reserved. 5 * (c) UNIX System Laboratories, Inc. 6 * All or some portions of this file are derived from material licensed 7 * to the University of California by American Telephone and Telegraph 8 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 9 * the permission of UNIX System Laboratories, Inc. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. All advertising materials mentioning features or use of this software 20 * must display the following acknowledgement: 21 * This product includes software developed by the University of 22 * California, Berkeley and its contributors. 23 * 4. Neither the name of the University nor the names of its contributors 24 * may be used to endorse or promote products derived from this software 25 * without specific prior written permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 * 39 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 40 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ 41 * $DragonFly: src/sys/kern/kern_clock.c,v 1.2 2003/06/17 04:28:41 dillon Exp $ 42 */ 43 44 #include "opt_ntp.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/dkstat.h> 49 #include <sys/callout.h> 50 #include <sys/kernel.h> 51 #include <sys/proc.h> 52 #include <sys/malloc.h> 53 #include <sys/resourcevar.h> 54 #include <sys/signalvar.h> 55 #include <sys/timex.h> 56 #include <sys/timepps.h> 57 #include <vm/vm.h> 58 #include <sys/lock.h> 59 #include <vm/pmap.h> 60 #include <vm/vm_map.h> 61 #include <sys/sysctl.h> 62 63 #include <machine/cpu.h> 64 #include <machine/limits.h> 65 #include <machine/smp.h> 66 67 #ifdef GPROF 68 #include <sys/gmon.h> 69 #endif 70 71 #ifdef DEVICE_POLLING 72 extern void init_device_poll(void); 73 extern void hardclock_device_poll(void); 74 #endif /* DEVICE_POLLING */ 75 76 /* 77 * Number of timecounters used to implement stable storage 78 */ 79 #ifndef NTIMECOUNTER 80 #define NTIMECOUNTER 5 81 #endif 82 83 static MALLOC_DEFINE(M_TIMECOUNTER, "timecounter", 84 "Timecounter stable storage"); 85 86 static void initclocks __P((void *dummy)); 87 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 88 89 static void tco_forward __P((int force)); 90 static void tco_setscales __P((struct timecounter *tc)); 91 static __inline unsigned tco_delta __P((struct timecounter *tc)); 92 93 /* Some of these don't belong here, but it's easiest to concentrate them. */ 94 long cp_time[CPUSTATES]; 95 96 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time), 97 "LU", "CPU time statistics"); 98 99 long tk_cancc; 100 long tk_nin; 101 long tk_nout; 102 long tk_rawcc; 103 104 time_t time_second; 105 106 struct timeval boottime; 107 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 108 &boottime, timeval, "System boottime"); 109 110 /* 111 * Which update policy to use. 112 * 0 - every tick, bad hardware may fail with "calcru negative..." 113 * 1 - more resistent to the above hardware, but less efficient. 114 */ 115 static int tco_method; 116 117 /* 118 * Implement a dummy timecounter which we can use until we get a real one 119 * in the air. This allows the console and other early stuff to use 120 * timeservices. 121 */ 122 123 static unsigned 124 dummy_get_timecount(struct timecounter *tc) 125 { 126 static unsigned now; 127 return (++now); 128 } 129 130 static struct timecounter dummy_timecounter = { 131 dummy_get_timecount, 132 0, 133 ~0u, 134 1000000, 135 "dummy" 136 }; 137 138 struct timecounter *timecounter = &dummy_timecounter; 139 140 /* 141 * Clock handling routines. 142 * 143 * This code is written to operate with two timers that run independently of 144 * each other. 145 * 146 * The main timer, running hz times per second, is used to trigger interval 147 * timers, timeouts and rescheduling as needed. 148 * 149 * The second timer handles kernel and user profiling, 150 * and does resource use estimation. If the second timer is programmable, 151 * it is randomized to avoid aliasing between the two clocks. For example, 152 * the randomization prevents an adversary from always giving up the cpu 153 * just before its quantum expires. Otherwise, it would never accumulate 154 * cpu ticks. The mean frequency of the second timer is stathz. 155 * 156 * If no second timer exists, stathz will be zero; in this case we drive 157 * profiling and statistics off the main clock. This WILL NOT be accurate; 158 * do not do it unless absolutely necessary. 159 * 160 * The statistics clock may (or may not) be run at a higher rate while 161 * profiling. This profile clock runs at profhz. We require that profhz 162 * be an integral multiple of stathz. 163 * 164 * If the statistics clock is running fast, it must be divided by the ratio 165 * profhz/stathz for statistics. (For profiling, every tick counts.) 166 * 167 * Time-of-day is maintained using a "timecounter", which may or may 168 * not be related to the hardware generating the above mentioned 169 * interrupts. 170 */ 171 172 int stathz; 173 int profhz; 174 static int profprocs; 175 int ticks; 176 static int psdiv, pscnt; /* prof => stat divider */ 177 int psratio; /* ratio: prof / stat */ 178 179 /* 180 * Initialize clock frequencies and start both clocks running. 181 */ 182 /* ARGSUSED*/ 183 static void 184 initclocks(dummy) 185 void *dummy; 186 { 187 register int i; 188 189 /* 190 * Set divisors to 1 (normal case) and let the machine-specific 191 * code do its bit. 192 */ 193 psdiv = pscnt = 1; 194 cpu_initclocks(); 195 196 #ifdef DEVICE_POLLING 197 init_device_poll(); 198 #endif 199 200 /* 201 * Compute profhz/stathz, and fix profhz if needed. 202 */ 203 i = stathz ? stathz : hz; 204 if (profhz == 0) 205 profhz = i; 206 psratio = profhz / i; 207 } 208 209 /* 210 * The real-time timer, interrupting hz times per second. 211 */ 212 void 213 hardclock(frame) 214 register struct clockframe *frame; 215 { 216 register struct proc *p; 217 218 p = curproc; 219 if (p) { 220 register struct pstats *pstats; 221 222 /* 223 * Run current process's virtual and profile time, as needed. 224 */ 225 pstats = p->p_stats; 226 if (CLKF_USERMODE(frame) && 227 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 228 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 229 psignal(p, SIGVTALRM); 230 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 231 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 232 psignal(p, SIGPROF); 233 } 234 235 #if defined(SMP) && defined(BETTER_CLOCK) 236 forward_hardclock(pscnt); 237 #endif 238 239 /* 240 * If no separate statistics clock is available, run it from here. 241 */ 242 if (stathz == 0) 243 statclock(frame); 244 245 tco_forward(0); 246 ticks++; 247 248 #ifdef DEVICE_POLLING 249 hardclock_device_poll(); /* this is very short and quick */ 250 #endif /* DEVICE_POLLING */ 251 252 /* 253 * Process callouts at a very low cpu priority, so we don't keep the 254 * relatively high clock interrupt priority any longer than necessary. 255 */ 256 if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) { 257 if (CLKF_BASEPRI(frame)) { 258 /* 259 * Save the overhead of a software interrupt; 260 * it will happen as soon as we return, so do it now. 261 */ 262 (void)splsoftclock(); 263 softclock(); 264 } else 265 setsoftclock(); 266 } else if (softticks + 1 == ticks) 267 ++softticks; 268 } 269 270 /* 271 * Compute number of ticks in the specified amount of time. 272 */ 273 int 274 tvtohz(tv) 275 struct timeval *tv; 276 { 277 register unsigned long ticks; 278 register long sec, usec; 279 280 /* 281 * If the number of usecs in the whole seconds part of the time 282 * difference fits in a long, then the total number of usecs will 283 * fit in an unsigned long. Compute the total and convert it to 284 * ticks, rounding up and adding 1 to allow for the current tick 285 * to expire. Rounding also depends on unsigned long arithmetic 286 * to avoid overflow. 287 * 288 * Otherwise, if the number of ticks in the whole seconds part of 289 * the time difference fits in a long, then convert the parts to 290 * ticks separately and add, using similar rounding methods and 291 * overflow avoidance. This method would work in the previous 292 * case but it is slightly slower and assumes that hz is integral. 293 * 294 * Otherwise, round the time difference down to the maximum 295 * representable value. 296 * 297 * If ints have 32 bits, then the maximum value for any timeout in 298 * 10ms ticks is 248 days. 299 */ 300 sec = tv->tv_sec; 301 usec = tv->tv_usec; 302 if (usec < 0) { 303 sec--; 304 usec += 1000000; 305 } 306 if (sec < 0) { 307 #ifdef DIAGNOSTIC 308 if (usec > 0) { 309 sec++; 310 usec -= 1000000; 311 } 312 printf("tvotohz: negative time difference %ld sec %ld usec\n", 313 sec, usec); 314 #endif 315 ticks = 1; 316 } else if (sec <= LONG_MAX / 1000000) 317 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 318 / tick + 1; 319 else if (sec <= LONG_MAX / hz) 320 ticks = sec * hz 321 + ((unsigned long)usec + (tick - 1)) / tick + 1; 322 else 323 ticks = LONG_MAX; 324 if (ticks > INT_MAX) 325 ticks = INT_MAX; 326 return ((int)ticks); 327 } 328 329 /* 330 * Start profiling on a process. 331 * 332 * Kernel profiling passes proc0 which never exits and hence 333 * keeps the profile clock running constantly. 334 */ 335 void 336 startprofclock(p) 337 register struct proc *p; 338 { 339 int s; 340 341 if ((p->p_flag & P_PROFIL) == 0) { 342 p->p_flag |= P_PROFIL; 343 if (++profprocs == 1 && stathz != 0) { 344 s = splstatclock(); 345 psdiv = pscnt = psratio; 346 setstatclockrate(profhz); 347 splx(s); 348 } 349 } 350 } 351 352 /* 353 * Stop profiling on a process. 354 */ 355 void 356 stopprofclock(p) 357 register struct proc *p; 358 { 359 int s; 360 361 if (p->p_flag & P_PROFIL) { 362 p->p_flag &= ~P_PROFIL; 363 if (--profprocs == 0 && stathz != 0) { 364 s = splstatclock(); 365 psdiv = pscnt = 1; 366 setstatclockrate(stathz); 367 splx(s); 368 } 369 } 370 } 371 372 /* 373 * Statistics clock. Grab profile sample, and if divider reaches 0, 374 * do process and kernel statistics. Most of the statistics are only 375 * used by user-level statistics programs. The main exceptions are 376 * p->p_uticks, p->p_sticks, p->p_iticks, and p->p_estcpu. 377 */ 378 void 379 statclock(frame) 380 register struct clockframe *frame; 381 { 382 #ifdef GPROF 383 register struct gmonparam *g; 384 int i; 385 #endif 386 register struct proc *p; 387 struct pstats *pstats; 388 long rss; 389 struct rusage *ru; 390 struct vmspace *vm; 391 392 if (curproc != NULL && CLKF_USERMODE(frame)) { 393 /* 394 * Came from user mode; CPU was in user state. 395 * If this process is being profiled, record the tick. 396 */ 397 p = curproc; 398 if (p->p_flag & P_PROFIL) 399 addupc_intr(p, CLKF_PC(frame), 1); 400 #if defined(SMP) && defined(BETTER_CLOCK) 401 if (stathz != 0) 402 forward_statclock(pscnt); 403 #endif 404 if (--pscnt > 0) 405 return; 406 /* 407 * Charge the time as appropriate. 408 */ 409 p->p_uticks++; 410 if (p->p_nice > NZERO) 411 cp_time[CP_NICE]++; 412 else 413 cp_time[CP_USER]++; 414 } else { 415 #ifdef GPROF 416 /* 417 * Kernel statistics are just like addupc_intr, only easier. 418 */ 419 g = &_gmonparam; 420 if (g->state == GMON_PROF_ON) { 421 i = CLKF_PC(frame) - g->lowpc; 422 if (i < g->textsize) { 423 i /= HISTFRACTION * sizeof(*g->kcount); 424 g->kcount[i]++; 425 } 426 } 427 #endif 428 #if defined(SMP) && defined(BETTER_CLOCK) 429 if (stathz != 0) 430 forward_statclock(pscnt); 431 #endif 432 if (--pscnt > 0) 433 return; 434 /* 435 * Came from kernel mode, so we were: 436 * - handling an interrupt, 437 * - doing syscall or trap work on behalf of the current 438 * user process, or 439 * - spinning in the idle loop. 440 * Whichever it is, charge the time as appropriate. 441 * Note that we charge interrupts to the current process, 442 * regardless of whether they are ``for'' that process, 443 * so that we know how much of its real time was spent 444 * in ``non-process'' (i.e., interrupt) work. 445 */ 446 p = curproc; 447 if (CLKF_INTR(frame)) { 448 if (p != NULL) 449 p->p_iticks++; 450 cp_time[CP_INTR]++; 451 } else if (p != NULL) { 452 p->p_sticks++; 453 cp_time[CP_SYS]++; 454 } else 455 cp_time[CP_IDLE]++; 456 } 457 pscnt = psdiv; 458 459 if (p != NULL) { 460 schedclock(p); 461 462 /* Update resource usage integrals and maximums. */ 463 if ((pstats = p->p_stats) != NULL && 464 (ru = &pstats->p_ru) != NULL && 465 (vm = p->p_vmspace) != NULL) { 466 ru->ru_ixrss += pgtok(vm->vm_tsize); 467 ru->ru_idrss += pgtok(vm->vm_dsize); 468 ru->ru_isrss += pgtok(vm->vm_ssize); 469 rss = pgtok(vmspace_resident_count(vm)); 470 if (ru->ru_maxrss < rss) 471 ru->ru_maxrss = rss; 472 } 473 } 474 } 475 476 /* 477 * Return information about system clocks. 478 */ 479 static int 480 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 481 { 482 struct clockinfo clkinfo; 483 /* 484 * Construct clockinfo structure. 485 */ 486 clkinfo.hz = hz; 487 clkinfo.tick = tick; 488 clkinfo.tickadj = tickadj; 489 clkinfo.profhz = profhz; 490 clkinfo.stathz = stathz ? stathz : hz; 491 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 492 } 493 494 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 495 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 496 497 static __inline unsigned 498 tco_delta(struct timecounter *tc) 499 { 500 501 return ((tc->tc_get_timecount(tc) - tc->tc_offset_count) & 502 tc->tc_counter_mask); 503 } 504 505 /* 506 * We have eight functions for looking at the clock, four for 507 * microseconds and four for nanoseconds. For each there is fast 508 * but less precise version "get{nano|micro}[up]time" which will 509 * return a time which is up to 1/HZ previous to the call, whereas 510 * the raw version "{nano|micro}[up]time" will return a timestamp 511 * which is as precise as possible. The "up" variants return the 512 * time relative to system boot, these are well suited for time 513 * interval measurements. 514 */ 515 516 void 517 getmicrotime(struct timeval *tvp) 518 { 519 struct timecounter *tc; 520 521 if (!tco_method) { 522 tc = timecounter; 523 *tvp = tc->tc_microtime; 524 } else { 525 microtime(tvp); 526 } 527 } 528 529 void 530 getnanotime(struct timespec *tsp) 531 { 532 struct timecounter *tc; 533 534 if (!tco_method) { 535 tc = timecounter; 536 *tsp = tc->tc_nanotime; 537 } else { 538 nanotime(tsp); 539 } 540 } 541 542 void 543 microtime(struct timeval *tv) 544 { 545 struct timecounter *tc; 546 547 tc = timecounter; 548 tv->tv_sec = tc->tc_offset_sec; 549 tv->tv_usec = tc->tc_offset_micro; 550 tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; 551 tv->tv_usec += boottime.tv_usec; 552 tv->tv_sec += boottime.tv_sec; 553 while (tv->tv_usec < 0) { 554 tv->tv_usec += 1000000; 555 if (tv->tv_sec > 0) 556 tv->tv_sec--; 557 } 558 while (tv->tv_usec >= 1000000) { 559 tv->tv_usec -= 1000000; 560 tv->tv_sec++; 561 } 562 } 563 564 void 565 nanotime(struct timespec *ts) 566 { 567 unsigned count; 568 u_int64_t delta; 569 struct timecounter *tc; 570 571 tc = timecounter; 572 ts->tv_sec = tc->tc_offset_sec; 573 count = tco_delta(tc); 574 delta = tc->tc_offset_nano; 575 delta += ((u_int64_t)count * tc->tc_scale_nano_f); 576 delta >>= 32; 577 delta += ((u_int64_t)count * tc->tc_scale_nano_i); 578 delta += boottime.tv_usec * 1000; 579 ts->tv_sec += boottime.tv_sec; 580 while (delta < 0) { 581 delta += 1000000000; 582 if (ts->tv_sec > 0) 583 ts->tv_sec--; 584 } 585 while (delta >= 1000000000) { 586 delta -= 1000000000; 587 ts->tv_sec++; 588 } 589 ts->tv_nsec = delta; 590 } 591 592 void 593 getmicrouptime(struct timeval *tvp) 594 { 595 struct timecounter *tc; 596 597 if (!tco_method) { 598 tc = timecounter; 599 tvp->tv_sec = tc->tc_offset_sec; 600 tvp->tv_usec = tc->tc_offset_micro; 601 } else { 602 microuptime(tvp); 603 } 604 } 605 606 void 607 getnanouptime(struct timespec *tsp) 608 { 609 struct timecounter *tc; 610 611 if (!tco_method) { 612 tc = timecounter; 613 tsp->tv_sec = tc->tc_offset_sec; 614 tsp->tv_nsec = tc->tc_offset_nano >> 32; 615 } else { 616 nanouptime(tsp); 617 } 618 } 619 620 void 621 microuptime(struct timeval *tv) 622 { 623 struct timecounter *tc; 624 625 tc = timecounter; 626 tv->tv_sec = tc->tc_offset_sec; 627 tv->tv_usec = tc->tc_offset_micro; 628 tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; 629 while (tv->tv_usec < 0) { 630 tv->tv_usec += 1000000; 631 if (tv->tv_sec > 0) 632 tv->tv_sec--; 633 } 634 while (tv->tv_usec >= 1000000) { 635 tv->tv_usec -= 1000000; 636 tv->tv_sec++; 637 } 638 } 639 640 void 641 nanouptime(struct timespec *ts) 642 { 643 unsigned count; 644 u_int64_t delta; 645 struct timecounter *tc; 646 647 tc = timecounter; 648 ts->tv_sec = tc->tc_offset_sec; 649 count = tco_delta(tc); 650 delta = tc->tc_offset_nano; 651 delta += ((u_int64_t)count * tc->tc_scale_nano_f); 652 delta >>= 32; 653 delta += ((u_int64_t)count * tc->tc_scale_nano_i); 654 while (delta < 0) { 655 delta += 1000000000; 656 if (ts->tv_sec > 0) 657 ts->tv_sec--; 658 } 659 while (delta >= 1000000000) { 660 delta -= 1000000000; 661 ts->tv_sec++; 662 } 663 ts->tv_nsec = delta; 664 } 665 666 static void 667 tco_setscales(struct timecounter *tc) 668 { 669 u_int64_t scale; 670 671 scale = 1000000000LL << 32; 672 scale += tc->tc_adjustment; 673 scale /= tc->tc_tweak->tc_frequency; 674 tc->tc_scale_micro = scale / 1000; 675 tc->tc_scale_nano_f = scale & 0xffffffff; 676 tc->tc_scale_nano_i = scale >> 32; 677 } 678 679 void 680 update_timecounter(struct timecounter *tc) 681 { 682 tco_setscales(tc); 683 } 684 685 void 686 init_timecounter(struct timecounter *tc) 687 { 688 struct timespec ts1; 689 struct timecounter *t1, *t2, *t3; 690 unsigned u; 691 int i; 692 693 u = tc->tc_frequency / tc->tc_counter_mask; 694 if (u > hz) { 695 printf("Timecounter \"%s\" frequency %lu Hz" 696 " -- Insufficient hz, needs at least %u\n", 697 tc->tc_name, (u_long) tc->tc_frequency, u); 698 return; 699 } 700 701 tc->tc_adjustment = 0; 702 tc->tc_tweak = tc; 703 tco_setscales(tc); 704 tc->tc_offset_count = tc->tc_get_timecount(tc); 705 if (timecounter == &dummy_timecounter) 706 tc->tc_avail = tc; 707 else { 708 tc->tc_avail = timecounter->tc_tweak->tc_avail; 709 timecounter->tc_tweak->tc_avail = tc; 710 } 711 MALLOC(t1, struct timecounter *, sizeof *t1, M_TIMECOUNTER, M_WAITOK); 712 tc->tc_other = t1; 713 *t1 = *tc; 714 t2 = t1; 715 for (i = 1; i < NTIMECOUNTER; i++) { 716 MALLOC(t3, struct timecounter *, sizeof *t3, 717 M_TIMECOUNTER, M_WAITOK); 718 *t3 = *tc; 719 t3->tc_other = t2; 720 t2 = t3; 721 } 722 t1->tc_other = t3; 723 tc = t1; 724 725 printf("Timecounter \"%s\" frequency %lu Hz\n", 726 tc->tc_name, (u_long)tc->tc_frequency); 727 728 /* XXX: For now always start using the counter. */ 729 tc->tc_offset_count = tc->tc_get_timecount(tc); 730 nanouptime(&ts1); 731 tc->tc_offset_nano = (u_int64_t)ts1.tv_nsec << 32; 732 tc->tc_offset_micro = ts1.tv_nsec / 1000; 733 tc->tc_offset_sec = ts1.tv_sec; 734 timecounter = tc; 735 } 736 737 void 738 set_timecounter(struct timespec *ts) 739 { 740 struct timespec ts2; 741 742 nanouptime(&ts2); 743 boottime.tv_sec = ts->tv_sec - ts2.tv_sec; 744 boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000; 745 if (boottime.tv_usec < 0) { 746 boottime.tv_usec += 1000000; 747 boottime.tv_sec--; 748 } 749 /* fiddle all the little crinkly bits around the fiords... */ 750 tco_forward(1); 751 } 752 753 static void 754 switch_timecounter(struct timecounter *newtc) 755 { 756 int s; 757 struct timecounter *tc; 758 struct timespec ts; 759 760 s = splclock(); 761 tc = timecounter; 762 if (newtc->tc_tweak == tc->tc_tweak) { 763 splx(s); 764 return; 765 } 766 newtc = newtc->tc_tweak->tc_other; 767 nanouptime(&ts); 768 newtc->tc_offset_sec = ts.tv_sec; 769 newtc->tc_offset_nano = (u_int64_t)ts.tv_nsec << 32; 770 newtc->tc_offset_micro = ts.tv_nsec / 1000; 771 newtc->tc_offset_count = newtc->tc_get_timecount(newtc); 772 tco_setscales(newtc); 773 timecounter = newtc; 774 splx(s); 775 } 776 777 static struct timecounter * 778 sync_other_counter(void) 779 { 780 struct timecounter *tc, *tcn, *tco; 781 unsigned delta; 782 783 tco = timecounter; 784 tc = tco->tc_other; 785 tcn = tc->tc_other; 786 *tc = *tco; 787 tc->tc_other = tcn; 788 delta = tco_delta(tc); 789 tc->tc_offset_count += delta; 790 tc->tc_offset_count &= tc->tc_counter_mask; 791 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_f; 792 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_i << 32; 793 return (tc); 794 } 795 796 static void 797 tco_forward(int force) 798 { 799 struct timecounter *tc, *tco; 800 struct timeval tvt; 801 802 tco = timecounter; 803 tc = sync_other_counter(); 804 /* 805 * We may be inducing a tiny error here, the tc_poll_pps() may 806 * process a latched count which happens after the tco_delta() 807 * in sync_other_counter(), which would extend the previous 808 * counters parameters into the domain of this new one. 809 * Since the timewindow is very small for this, the error is 810 * going to be only a few weenieseconds (as Dave Mills would 811 * say), so lets just not talk more about it, OK ? 812 */ 813 if (tco->tc_poll_pps) 814 tco->tc_poll_pps(tco); 815 if (timedelta != 0) { 816 tvt = boottime; 817 tvt.tv_usec += tickdelta; 818 if (tvt.tv_usec >= 1000000) { 819 tvt.tv_sec++; 820 tvt.tv_usec -= 1000000; 821 } else if (tvt.tv_usec < 0) { 822 tvt.tv_sec--; 823 tvt.tv_usec += 1000000; 824 } 825 boottime = tvt; 826 timedelta -= tickdelta; 827 } 828 829 while (tc->tc_offset_nano >= 1000000000ULL << 32) { 830 tc->tc_offset_nano -= 1000000000ULL << 32; 831 tc->tc_offset_sec++; 832 ntp_update_second(tc); /* XXX only needed if xntpd runs */ 833 tco_setscales(tc); 834 force++; 835 } 836 837 if (tco_method && !force) 838 return; 839 840 tc->tc_offset_micro = (tc->tc_offset_nano / 1000) >> 32; 841 842 /* Figure out the wall-clock time */ 843 tc->tc_nanotime.tv_sec = tc->tc_offset_sec + boottime.tv_sec; 844 tc->tc_nanotime.tv_nsec = 845 (tc->tc_offset_nano >> 32) + boottime.tv_usec * 1000; 846 tc->tc_microtime.tv_usec = tc->tc_offset_micro + boottime.tv_usec; 847 while (tc->tc_nanotime.tv_nsec >= 1000000000) { 848 tc->tc_nanotime.tv_nsec -= 1000000000; 849 tc->tc_microtime.tv_usec -= 1000000; 850 tc->tc_nanotime.tv_sec++; 851 } 852 time_second = tc->tc_microtime.tv_sec = tc->tc_nanotime.tv_sec; 853 854 timecounter = tc; 855 } 856 857 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, ""); 858 859 SYSCTL_INT(_kern_timecounter, OID_AUTO, method, CTLFLAG_RW, &tco_method, 0, 860 "This variable determines the method used for updating timecounters. " 861 "If the default algorithm (0) fails with \"calcru negative...\" messages " 862 "try the alternate algorithm (1) which handles bad hardware better." 863 864 ); 865 866 static int 867 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS) 868 { 869 char newname[32]; 870 struct timecounter *newtc, *tc; 871 int error; 872 873 tc = timecounter->tc_tweak; 874 strncpy(newname, tc->tc_name, sizeof(newname)); 875 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req); 876 if (error == 0 && req->newptr != NULL && 877 strcmp(newname, tc->tc_name) != 0) { 878 for (newtc = tc->tc_avail; newtc != tc; 879 newtc = newtc->tc_avail) { 880 if (strcmp(newname, newtc->tc_name) == 0) { 881 /* Warm up new timecounter. */ 882 (void)newtc->tc_get_timecount(newtc); 883 884 switch_timecounter(newtc); 885 return (0); 886 } 887 } 888 return (EINVAL); 889 } 890 return (error); 891 } 892 893 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW, 894 0, 0, sysctl_kern_timecounter_hardware, "A", ""); 895 896 897 int 898 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 899 { 900 pps_params_t *app; 901 struct pps_fetch_args *fapi; 902 #ifdef PPS_SYNC 903 struct pps_kcbind_args *kapi; 904 #endif 905 906 switch (cmd) { 907 case PPS_IOC_CREATE: 908 return (0); 909 case PPS_IOC_DESTROY: 910 return (0); 911 case PPS_IOC_SETPARAMS: 912 app = (pps_params_t *)data; 913 if (app->mode & ~pps->ppscap) 914 return (EINVAL); 915 pps->ppsparam = *app; 916 return (0); 917 case PPS_IOC_GETPARAMS: 918 app = (pps_params_t *)data; 919 *app = pps->ppsparam; 920 app->api_version = PPS_API_VERS_1; 921 return (0); 922 case PPS_IOC_GETCAP: 923 *(int*)data = pps->ppscap; 924 return (0); 925 case PPS_IOC_FETCH: 926 fapi = (struct pps_fetch_args *)data; 927 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 928 return (EINVAL); 929 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 930 return (EOPNOTSUPP); 931 pps->ppsinfo.current_mode = pps->ppsparam.mode; 932 fapi->pps_info_buf = pps->ppsinfo; 933 return (0); 934 case PPS_IOC_KCBIND: 935 #ifdef PPS_SYNC 936 kapi = (struct pps_kcbind_args *)data; 937 /* XXX Only root should be able to do this */ 938 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 939 return (EINVAL); 940 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 941 return (EINVAL); 942 if (kapi->edge & ~pps->ppscap) 943 return (EINVAL); 944 pps->kcmode = kapi->edge; 945 return (0); 946 #else 947 return (EOPNOTSUPP); 948 #endif 949 default: 950 return (ENOTTY); 951 } 952 } 953 954 void 955 pps_init(struct pps_state *pps) 956 { 957 pps->ppscap |= PPS_TSFMT_TSPEC; 958 if (pps->ppscap & PPS_CAPTUREASSERT) 959 pps->ppscap |= PPS_OFFSETASSERT; 960 if (pps->ppscap & PPS_CAPTURECLEAR) 961 pps->ppscap |= PPS_OFFSETCLEAR; 962 } 963 964 void 965 pps_event(struct pps_state *pps, struct timecounter *tc, unsigned count, int event) 966 { 967 struct timespec ts, *tsp, *osp; 968 u_int64_t delta; 969 unsigned tcount, *pcount; 970 int foff, fhard; 971 pps_seq_t *pseq; 972 973 /* Things would be easier with arrays... */ 974 if (event == PPS_CAPTUREASSERT) { 975 tsp = &pps->ppsinfo.assert_timestamp; 976 osp = &pps->ppsparam.assert_offset; 977 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 978 fhard = pps->kcmode & PPS_CAPTUREASSERT; 979 pcount = &pps->ppscount[0]; 980 pseq = &pps->ppsinfo.assert_sequence; 981 } else { 982 tsp = &pps->ppsinfo.clear_timestamp; 983 osp = &pps->ppsparam.clear_offset; 984 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 985 fhard = pps->kcmode & PPS_CAPTURECLEAR; 986 pcount = &pps->ppscount[1]; 987 pseq = &pps->ppsinfo.clear_sequence; 988 } 989 990 /* The timecounter changed: bail */ 991 if (!pps->ppstc || 992 pps->ppstc->tc_name != tc->tc_name || 993 tc->tc_name != timecounter->tc_name) { 994 pps->ppstc = tc; 995 *pcount = count; 996 return; 997 } 998 999 /* Nothing really happened */ 1000 if (*pcount == count) 1001 return; 1002 1003 *pcount = count; 1004 1005 /* Convert the count to timespec */ 1006 ts.tv_sec = tc->tc_offset_sec; 1007 tcount = count - tc->tc_offset_count; 1008 tcount &= tc->tc_counter_mask; 1009 delta = tc->tc_offset_nano; 1010 delta += ((u_int64_t)tcount * tc->tc_scale_nano_f); 1011 delta >>= 32; 1012 delta += ((u_int64_t)tcount * tc->tc_scale_nano_i); 1013 delta += boottime.tv_usec * 1000; 1014 ts.tv_sec += boottime.tv_sec; 1015 while (delta >= 1000000000) { 1016 delta -= 1000000000; 1017 ts.tv_sec++; 1018 } 1019 ts.tv_nsec = delta; 1020 1021 (*pseq)++; 1022 *tsp = ts; 1023 1024 if (foff) { 1025 timespecadd(tsp, osp); 1026 if (tsp->tv_nsec < 0) { 1027 tsp->tv_nsec += 1000000000; 1028 tsp->tv_sec -= 1; 1029 } 1030 } 1031 #ifdef PPS_SYNC 1032 if (fhard) { 1033 /* magic, at its best... */ 1034 tcount = count - pps->ppscount[2]; 1035 pps->ppscount[2] = count; 1036 tcount &= tc->tc_counter_mask; 1037 delta = ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_f); 1038 delta >>= 32; 1039 delta += ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_i); 1040 hardpps(tsp, delta); 1041 } 1042 #endif 1043 } 1044