1 /*- 2 * Copyright (c) 2004, Matthew Dillon <dillon@backplane.com> 3 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> 4 * Copyright (c) 1982, 1986, 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 41 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ 42 * $DragonFly: src/sys/kern/kern_clock.c,v 1.16 2004/02/11 21:47:51 dillon Exp $ 43 */ 44 45 #include "opt_ntp.h" 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/dkstat.h> 50 #include <sys/callout.h> 51 #include <sys/kernel.h> 52 #include <sys/proc.h> 53 #include <sys/malloc.h> 54 #include <sys/resourcevar.h> 55 #include <sys/signalvar.h> 56 #include <sys/timex.h> 57 #include <sys/timepps.h> 58 #include <vm/vm.h> 59 #include <sys/lock.h> 60 #include <vm/pmap.h> 61 #include <vm/vm_map.h> 62 #include <sys/sysctl.h> 63 #include <sys/thread2.h> 64 65 #include <machine/cpu.h> 66 #include <machine/limits.h> 67 #include <machine/smp.h> 68 69 #ifdef GPROF 70 #include <sys/gmon.h> 71 #endif 72 73 #ifdef DEVICE_POLLING 74 extern void init_device_poll(void); 75 extern void hardclock_device_poll(void); 76 #endif /* DEVICE_POLLING */ 77 78 static void initclocks (void *dummy); 79 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 80 81 /* 82 * Some of these don't belong here, but it's easiest to concentrate them. 83 * Note that cp_time[] counts in microseconds, but most userland programs 84 * just compare relative times against the total by delta. 85 */ 86 long cp_time[CPUSTATES]; 87 88 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time), 89 "LU", "CPU time statistics"); 90 91 long tk_cancc; 92 long tk_nin; 93 long tk_nout; 94 long tk_rawcc; 95 96 /* 97 * boottime is used to calculate the 'real' uptime. Do not confuse this with 98 * microuptime(). microtime() is not drift compensated. The real uptime 99 * with compensation is nanotime() - bootime. 100 * 101 * basetime is used to calculate the compensated real time of day. Chunky 102 * changes to the time, aka settimeofday(), are made by modifying basetime. 103 * 104 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 105 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 106 * the real time. 107 */ 108 struct timespec boottime; /* boot time (realtime) for reference only */ 109 struct timespec basetime; /* base time adjusts uptime -> realtime */ 110 time_t time_second; /* read-only 'passive' uptime in seconds */ 111 112 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 113 &boottime, timeval, "System boottime"); 114 SYSCTL_STRUCT(_kern, OID_AUTO, basetime, CTLFLAG_RD, 115 &basetime, timeval, "System basetime"); 116 117 static void hardclock(systimer_t info, struct intrframe *frame); 118 static void statclock(systimer_t info, struct intrframe *frame); 119 static void schedclock(systimer_t info, struct intrframe *frame); 120 121 int ticks; /* system master ticks at hz */ 122 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 123 int64_t nsec_acc; /* accumulator */ 124 125 /* 126 * Finish initializing clock frequencies and start all clocks running. 127 */ 128 /* ARGSUSED*/ 129 static void 130 initclocks(void *dummy) 131 { 132 cpu_initclocks(); 133 #ifdef DEVICE_POLLING 134 init_device_poll(); 135 #endif 136 /*psratio = profhz / stathz;*/ 137 initclocks_pcpu(); 138 } 139 140 /* 141 * Called on a per-cpu basis 142 */ 143 void 144 initclocks_pcpu(void) 145 { 146 struct globaldata *gd = mycpu; 147 148 crit_enter(); 149 if (gd->gd_cpuid == 0) { 150 gd->gd_time_seconds = 1; 151 gd->gd_cpuclock_base = cputimer_count(); 152 } else { 153 /* XXX */ 154 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 155 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 156 } 157 systimer_init_periodic(&gd->gd_hardclock, hardclock, NULL, hz); 158 systimer_init_periodic(&gd->gd_statclock, statclock, NULL, stathz); 159 /* XXX correct the frequency for scheduler / estcpu tests */ 160 systimer_init_periodic(&gd->gd_schedclock, schedclock, NULL, 10); 161 crit_exit(); 162 } 163 164 /* 165 * This sets the current real time of day. Timespecs are in seconds and 166 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 167 * instead we adjust basetime so basetime + gd_* results in the current 168 * time of day. This way the gd_* fields are guarenteed to represent 169 * a monotonically increasing 'uptime' value. 170 */ 171 void 172 set_timeofday(struct timespec *ts) 173 { 174 struct timespec ts2; 175 176 /* 177 * XXX SMP / non-atomic basetime updates 178 */ 179 crit_enter(); 180 nanouptime(&ts2); 181 basetime.tv_sec = ts->tv_sec - ts2.tv_sec; 182 basetime.tv_nsec = ts->tv_nsec - ts2.tv_nsec; 183 if (basetime.tv_nsec < 0) { 184 basetime.tv_nsec += 1000000000; 185 --basetime.tv_sec; 186 } 187 if (boottime.tv_sec == 0) 188 boottime = basetime; 189 timedelta = 0; 190 crit_exit(); 191 } 192 193 /* 194 * Each cpu has its own hardclock, but we only increments ticks and softticks 195 * on cpu #0. 196 * 197 * NOTE! systimer! the MP lock might not be held here. We can only safely 198 * manipulate objects owned by the current cpu. 199 */ 200 static void 201 hardclock(systimer_t info, struct intrframe *frame) 202 { 203 sysclock_t cputicks; 204 struct proc *p; 205 struct pstats *pstats; 206 struct globaldata *gd = mycpu; 207 208 /* 209 * Realtime updates are per-cpu. Note that timer corrections as 210 * returned by microtime() and friends make an additional adjustment 211 * using a system-wise 'basetime', but the running time is always 212 * taken from the per-cpu globaldata area. Since the same clock 213 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 214 * stay in synch. 215 * 216 * Note that we never allow info->time (aka gd->gd_hardclock.time) 217 * to reverse index gd_cpuclock_base. 218 */ 219 cputicks = info->time - gd->gd_cpuclock_base; 220 if (cputicks > cputimer_freq) { 221 ++gd->gd_time_seconds; 222 gd->gd_cpuclock_base += cputimer_freq; 223 } 224 225 /* 226 * The system-wide ticks and softticks are only updated by cpu #0. 227 * Callwheel actions are also (at the moment) only handled by cpu #0. 228 * Finally, we also do NTP related timedelta/tickdelta adjustments 229 * by adjusting basetime. 230 */ 231 if (gd->gd_cpuid == 0) { 232 struct timespec nts; 233 int leap; 234 235 ++ticks; 236 237 #ifdef DEVICE_POLLING 238 hardclock_device_poll(); /* mpsafe, short and quick */ 239 #endif /* DEVICE_POLLING */ 240 241 if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) { 242 setsoftclock(); 243 } else if (softticks + 1 == ticks) { 244 ++softticks; 245 } 246 247 #if 0 248 if (tco->tc_poll_pps) 249 tco->tc_poll_pps(tco); 250 #endif 251 /* 252 * Apply adjtime corrections. At the moment only do this if 253 * we can get the MP lock to interlock with adjtime's modification 254 * of these variables. Note that basetime adjustments are not 255 * MP safe either XXX. 256 */ 257 if (timedelta != 0 && try_mplock()) { 258 basetime.tv_nsec += tickdelta * 1000; 259 if (basetime.tv_nsec >= 1000000000) { 260 basetime.tv_nsec -= 1000000000; 261 ++basetime.tv_sec; 262 } else if (basetime.tv_nsec < 0) { 263 basetime.tv_nsec += 1000000000; 264 --basetime.tv_sec; 265 } 266 timedelta -= tickdelta; 267 rel_mplock(); 268 } 269 270 /* 271 * Apply per-tick compensation. ticks_adj adjusts for both 272 * offset and frequency, and could be negative. 273 */ 274 if (nsec_adj != 0 && try_mplock()) { 275 nsec_acc += nsec_adj; 276 if (nsec_acc >= 0x100000000LL) { 277 basetime.tv_nsec += nsec_acc >> 32; 278 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 279 } else if (nsec_acc <= -0x100000000LL) { 280 basetime.tv_nsec -= -nsec_acc >> 32; 281 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 282 } 283 if (basetime.tv_nsec >= 1000000000) { 284 basetime.tv_nsec -= 1000000000; 285 ++basetime.tv_sec; 286 } else if (basetime.tv_nsec < 0) { 287 basetime.tv_nsec += 1000000000; 288 --basetime.tv_sec; 289 } 290 rel_mplock(); 291 } 292 293 /* 294 * If the realtime-adjusted seconds hand rolls over then tell 295 * ntp_update_second() what we did in the last second so it can 296 * calculate what to do in the next second. It may also add 297 * or subtract a leap second. 298 */ 299 getnanotime(&nts); 300 if (time_second != nts.tv_sec) { 301 leap = ntp_update_second(time_second, &nsec_adj); 302 basetime.tv_sec += leap; 303 time_second = nts.tv_sec + leap; 304 nsec_adj /= hz; 305 } 306 } 307 308 /* 309 * ITimer handling is per-tick, per-cpu. I don't think psignal() 310 * is mpsafe on curproc, so XXX get the mplock. 311 */ 312 if ((p = curproc) != NULL && try_mplock()) { 313 pstats = p->p_stats; 314 if (frame && CLKF_USERMODE(frame) && 315 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 316 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 317 psignal(p, SIGVTALRM); 318 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 319 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 320 psignal(p, SIGPROF); 321 rel_mplock(); 322 } 323 setdelayed(); 324 } 325 326 /* 327 * The statistics clock typically runs at a 125Hz rate, and is intended 328 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 329 * 330 * NOTE! systimer! the MP lock might not be held here. We can only safely 331 * manipulate objects owned by the current cpu. 332 * 333 * The stats clock is responsible for grabbing a profiling sample. 334 * Most of the statistics are only used by user-level statistics programs. 335 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 336 * p->p_estcpu. 337 * 338 * Like the other clocks, the stat clock is called from what is effectively 339 * a fast interrupt, so the context should be the thread/process that got 340 * interrupted. 341 */ 342 static void 343 statclock(systimer_t info, struct intrframe *frame) 344 { 345 #ifdef GPROF 346 struct gmonparam *g; 347 int i; 348 #endif 349 thread_t td; 350 struct proc *p; 351 int bump; 352 struct timeval tv; 353 struct timeval *stv; 354 355 /* 356 * How big was our timeslice relative to the last time? 357 */ 358 microuptime(&tv); /* mpsafe */ 359 stv = &mycpu->gd_stattv; 360 if (stv->tv_sec == 0) { 361 bump = 1; 362 } else { 363 bump = tv.tv_usec - stv->tv_usec + 364 (tv.tv_sec - stv->tv_sec) * 1000000; 365 if (bump < 0) 366 bump = 0; 367 if (bump > 1000000) 368 bump = 1000000; 369 } 370 *stv = tv; 371 372 td = curthread; 373 p = td->td_proc; 374 375 if (frame && CLKF_USERMODE(frame)) { 376 /* 377 * Came from userland, handle user time and deal with 378 * possible process. 379 */ 380 if (p && (p->p_flag & P_PROFIL)) 381 addupc_intr(p, CLKF_PC(frame), 1); 382 td->td_uticks += bump; 383 384 /* 385 * Charge the time as appropriate 386 */ 387 if (p && p->p_nice > NZERO) 388 cp_time[CP_NICE] += bump; 389 else 390 cp_time[CP_USER] += bump; 391 } else { 392 #ifdef GPROF 393 /* 394 * Kernel statistics are just like addupc_intr, only easier. 395 */ 396 g = &_gmonparam; 397 if (g->state == GMON_PROF_ON && frame) { 398 i = CLKF_PC(frame) - g->lowpc; 399 if (i < g->textsize) { 400 i /= HISTFRACTION * sizeof(*g->kcount); 401 g->kcount[i]++; 402 } 403 } 404 #endif 405 /* 406 * Came from kernel mode, so we were: 407 * - handling an interrupt, 408 * - doing syscall or trap work on behalf of the current 409 * user process, or 410 * - spinning in the idle loop. 411 * Whichever it is, charge the time as appropriate. 412 * Note that we charge interrupts to the current process, 413 * regardless of whether they are ``for'' that process, 414 * so that we know how much of its real time was spent 415 * in ``non-process'' (i.e., interrupt) work. 416 * 417 * XXX assume system if frame is NULL. A NULL frame 418 * can occur if ipi processing is done from an splx(). 419 */ 420 if (frame && CLKF_INTR(frame)) 421 td->td_iticks += bump; 422 else 423 td->td_sticks += bump; 424 425 if (frame && CLKF_INTR(frame)) { 426 cp_time[CP_INTR] += bump; 427 } else { 428 if (td == &mycpu->gd_idlethread) 429 cp_time[CP_IDLE] += bump; 430 else 431 cp_time[CP_SYS] += bump; 432 } 433 } 434 } 435 436 /* 437 * The scheduler clock typically runs at a 10Hz rate. NOTE! systimer, 438 * the MP lock might not be held. We can safely manipulate parts of curproc 439 * but that's about it. 440 */ 441 static void 442 schedclock(systimer_t info, struct intrframe *frame) 443 { 444 struct proc *p; 445 struct pstats *pstats; 446 struct rusage *ru; 447 struct vmspace *vm; 448 long rss; 449 450 schedulerclock(NULL); /* mpsafe */ 451 if ((p = curproc) != NULL) { 452 /* Update resource usage integrals and maximums. */ 453 if ((pstats = p->p_stats) != NULL && 454 (ru = &pstats->p_ru) != NULL && 455 (vm = p->p_vmspace) != NULL) { 456 ru->ru_ixrss += pgtok(vm->vm_tsize); 457 ru->ru_idrss += pgtok(vm->vm_dsize); 458 ru->ru_isrss += pgtok(vm->vm_ssize); 459 rss = pgtok(vmspace_resident_count(vm)); 460 if (ru->ru_maxrss < rss) 461 ru->ru_maxrss = rss; 462 } 463 } 464 } 465 466 /* 467 * Compute number of ticks for the specified amount of time. The 468 * return value is intended to be used in a clock interrupt timed 469 * operation and guarenteed to meet or exceed the requested time. 470 * If the representation overflows, return INT_MAX. The minimum return 471 * value is 1 ticks and the function will average the calculation up. 472 * If any value greater then 0 microseconds is supplied, a value 473 * of at least 2 will be returned to ensure that a near-term clock 474 * interrupt does not cause the timeout to occur (degenerately) early. 475 * 476 * Note that limit checks must take into account microseconds, which is 477 * done simply by using the smaller signed long maximum instead of 478 * the unsigned long maximum. 479 * 480 * If ints have 32 bits, then the maximum value for any timeout in 481 * 10ms ticks is 248 days. 482 */ 483 int 484 tvtohz_high(struct timeval *tv) 485 { 486 int ticks; 487 long sec, usec; 488 489 sec = tv->tv_sec; 490 usec = tv->tv_usec; 491 if (usec < 0) { 492 sec--; 493 usec += 1000000; 494 } 495 if (sec < 0) { 496 #ifdef DIAGNOSTIC 497 if (usec > 0) { 498 sec++; 499 usec -= 1000000; 500 } 501 printf("tvotohz: negative time difference %ld sec %ld usec\n", 502 sec, usec); 503 #endif 504 ticks = 1; 505 } else if (sec <= INT_MAX / hz) { 506 ticks = (int)(sec * hz + 507 ((u_long)usec + (tick - 1)) / tick) + 1; 508 } else { 509 ticks = INT_MAX; 510 } 511 return (ticks); 512 } 513 514 /* 515 * Compute number of ticks for the specified amount of time, erroring on 516 * the side of it being too low to ensure that sleeping the returned number 517 * of ticks will not result in a late return. 518 * 519 * The supplied timeval may not be negative and should be normalized. A 520 * return value of 0 is possible if the timeval converts to less then 521 * 1 tick. 522 * 523 * If ints have 32 bits, then the maximum value for any timeout in 524 * 10ms ticks is 248 days. 525 */ 526 int 527 tvtohz_low(struct timeval *tv) 528 { 529 int ticks; 530 long sec; 531 532 sec = tv->tv_sec; 533 if (sec <= INT_MAX / hz) 534 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 535 else 536 ticks = INT_MAX; 537 return (ticks); 538 } 539 540 541 /* 542 * Start profiling on a process. 543 * 544 * Kernel profiling passes proc0 which never exits and hence 545 * keeps the profile clock running constantly. 546 */ 547 void 548 startprofclock(struct proc *p) 549 { 550 if ((p->p_flag & P_PROFIL) == 0) { 551 p->p_flag |= P_PROFIL; 552 #if 0 /* XXX */ 553 if (++profprocs == 1 && stathz != 0) { 554 s = splstatclock(); 555 psdiv = psratio; 556 setstatclockrate(profhz); 557 splx(s); 558 } 559 #endif 560 } 561 } 562 563 /* 564 * Stop profiling on a process. 565 */ 566 void 567 stopprofclock(struct proc *p) 568 { 569 if (p->p_flag & P_PROFIL) { 570 p->p_flag &= ~P_PROFIL; 571 #if 0 /* XXX */ 572 if (--profprocs == 0 && stathz != 0) { 573 s = splstatclock(); 574 psdiv = 1; 575 setstatclockrate(stathz); 576 splx(s); 577 } 578 #endif 579 } 580 } 581 582 /* 583 * Return information about system clocks. 584 */ 585 static int 586 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 587 { 588 struct clockinfo clkinfo; 589 /* 590 * Construct clockinfo structure. 591 */ 592 clkinfo.hz = hz; 593 clkinfo.tick = tick; 594 clkinfo.tickadj = tickadj; 595 clkinfo.profhz = profhz; 596 clkinfo.stathz = stathz ? stathz : hz; 597 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 598 } 599 600 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 601 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 602 603 /* 604 * We have eight functions for looking at the clock, four for 605 * microseconds and four for nanoseconds. For each there is fast 606 * but less precise version "get{nano|micro}[up]time" which will 607 * return a time which is up to 1/HZ previous to the call, whereas 608 * the raw version "{nano|micro}[up]time" will return a timestamp 609 * which is as precise as possible. The "up" variants return the 610 * time relative to system boot, these are well suited for time 611 * interval measurements. 612 * 613 * Each cpu independantly maintains the current time of day, so all 614 * we need to do to protect ourselves from changes is to do a loop 615 * check on the seconds field changing out from under us. 616 */ 617 void 618 getmicrouptime(struct timeval *tvp) 619 { 620 struct globaldata *gd = mycpu; 621 sysclock_t delta; 622 623 do { 624 tvp->tv_sec = gd->gd_time_seconds; 625 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 626 } while (tvp->tv_sec != gd->gd_time_seconds); 627 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 628 if (tvp->tv_usec >= 1000000) { 629 tvp->tv_usec -= 1000000; 630 ++tvp->tv_sec; 631 } 632 } 633 634 void 635 getnanouptime(struct timespec *tsp) 636 { 637 struct globaldata *gd = mycpu; 638 sysclock_t delta; 639 640 do { 641 tsp->tv_sec = gd->gd_time_seconds; 642 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 643 } while (tsp->tv_sec != gd->gd_time_seconds); 644 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 645 if (tsp->tv_nsec >= 1000000000) { 646 tsp->tv_nsec -= 1000000000; 647 ++tsp->tv_sec; 648 } 649 } 650 651 void 652 microuptime(struct timeval *tvp) 653 { 654 struct globaldata *gd = mycpu; 655 sysclock_t delta; 656 657 do { 658 tvp->tv_sec = gd->gd_time_seconds; 659 delta = cputimer_count() - gd->gd_cpuclock_base; 660 } while (tvp->tv_sec != gd->gd_time_seconds); 661 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 662 if (tvp->tv_usec >= 1000000) { 663 tvp->tv_usec -= 1000000; 664 ++tvp->tv_sec; 665 } 666 } 667 668 void 669 nanouptime(struct timespec *tsp) 670 { 671 struct globaldata *gd = mycpu; 672 sysclock_t delta; 673 674 do { 675 tsp->tv_sec = gd->gd_time_seconds; 676 delta = cputimer_count() - gd->gd_cpuclock_base; 677 } while (tsp->tv_sec != gd->gd_time_seconds); 678 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 679 if (tsp->tv_nsec >= 1000000000) { 680 tsp->tv_nsec -= 1000000000; 681 ++tsp->tv_sec; 682 } 683 } 684 685 /* 686 * realtime routines 687 */ 688 689 void 690 getmicrotime(struct timeval *tvp) 691 { 692 struct globaldata *gd = mycpu; 693 sysclock_t delta; 694 695 do { 696 tvp->tv_sec = gd->gd_time_seconds; 697 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 698 } while (tvp->tv_sec != gd->gd_time_seconds); 699 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 700 701 tvp->tv_sec += basetime.tv_sec; 702 tvp->tv_usec += basetime.tv_nsec / 1000; 703 while (tvp->tv_usec >= 1000000) { 704 tvp->tv_usec -= 1000000; 705 ++tvp->tv_sec; 706 } 707 } 708 709 void 710 getnanotime(struct timespec *tsp) 711 { 712 struct globaldata *gd = mycpu; 713 sysclock_t delta; 714 715 do { 716 tsp->tv_sec = gd->gd_time_seconds; 717 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 718 } while (tsp->tv_sec != gd->gd_time_seconds); 719 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 720 721 tsp->tv_sec += basetime.tv_sec; 722 tsp->tv_nsec += basetime.tv_nsec; 723 while (tsp->tv_nsec >= 1000000000) { 724 tsp->tv_nsec -= 1000000000; 725 ++tsp->tv_sec; 726 } 727 } 728 729 void 730 microtime(struct timeval *tvp) 731 { 732 struct globaldata *gd = mycpu; 733 sysclock_t delta; 734 735 do { 736 tvp->tv_sec = gd->gd_time_seconds; 737 delta = cputimer_count() - gd->gd_cpuclock_base; 738 } while (tvp->tv_sec != gd->gd_time_seconds); 739 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 740 741 tvp->tv_sec += basetime.tv_sec; 742 tvp->tv_usec += basetime.tv_nsec / 1000; 743 while (tvp->tv_usec >= 1000000) { 744 tvp->tv_usec -= 1000000; 745 ++tvp->tv_sec; 746 } 747 } 748 749 void 750 nanotime(struct timespec *tsp) 751 { 752 struct globaldata *gd = mycpu; 753 sysclock_t delta; 754 755 do { 756 tsp->tv_sec = gd->gd_time_seconds; 757 delta = cputimer_count() - gd->gd_cpuclock_base; 758 } while (tsp->tv_sec != gd->gd_time_seconds); 759 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 760 761 tsp->tv_sec += basetime.tv_sec; 762 tsp->tv_nsec += basetime.tv_nsec; 763 while (tsp->tv_nsec >= 1000000000) { 764 tsp->tv_nsec -= 1000000000; 765 ++tsp->tv_sec; 766 } 767 } 768 769 int 770 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 771 { 772 pps_params_t *app; 773 struct pps_fetch_args *fapi; 774 #ifdef PPS_SYNC 775 struct pps_kcbind_args *kapi; 776 #endif 777 778 switch (cmd) { 779 case PPS_IOC_CREATE: 780 return (0); 781 case PPS_IOC_DESTROY: 782 return (0); 783 case PPS_IOC_SETPARAMS: 784 app = (pps_params_t *)data; 785 if (app->mode & ~pps->ppscap) 786 return (EINVAL); 787 pps->ppsparam = *app; 788 return (0); 789 case PPS_IOC_GETPARAMS: 790 app = (pps_params_t *)data; 791 *app = pps->ppsparam; 792 app->api_version = PPS_API_VERS_1; 793 return (0); 794 case PPS_IOC_GETCAP: 795 *(int*)data = pps->ppscap; 796 return (0); 797 case PPS_IOC_FETCH: 798 fapi = (struct pps_fetch_args *)data; 799 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 800 return (EINVAL); 801 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 802 return (EOPNOTSUPP); 803 pps->ppsinfo.current_mode = pps->ppsparam.mode; 804 fapi->pps_info_buf = pps->ppsinfo; 805 return (0); 806 case PPS_IOC_KCBIND: 807 #ifdef PPS_SYNC 808 kapi = (struct pps_kcbind_args *)data; 809 /* XXX Only root should be able to do this */ 810 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 811 return (EINVAL); 812 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 813 return (EINVAL); 814 if (kapi->edge & ~pps->ppscap) 815 return (EINVAL); 816 pps->kcmode = kapi->edge; 817 return (0); 818 #else 819 return (EOPNOTSUPP); 820 #endif 821 default: 822 return (ENOTTY); 823 } 824 } 825 826 void 827 pps_init(struct pps_state *pps) 828 { 829 pps->ppscap |= PPS_TSFMT_TSPEC; 830 if (pps->ppscap & PPS_CAPTUREASSERT) 831 pps->ppscap |= PPS_OFFSETASSERT; 832 if (pps->ppscap & PPS_CAPTURECLEAR) 833 pps->ppscap |= PPS_OFFSETCLEAR; 834 } 835 836 void 837 pps_event(struct pps_state *pps, sysclock_t count, int event) 838 { 839 struct globaldata *gd; 840 struct timespec *tsp; 841 struct timespec *osp; 842 struct timespec ts; 843 sysclock_t *pcount; 844 #ifdef PPS_SYNC 845 sysclock_t tcount; 846 #endif 847 sysclock_t delta; 848 pps_seq_t *pseq; 849 int foff; 850 int fhard; 851 852 gd = mycpu; 853 854 /* Things would be easier with arrays... */ 855 if (event == PPS_CAPTUREASSERT) { 856 tsp = &pps->ppsinfo.assert_timestamp; 857 osp = &pps->ppsparam.assert_offset; 858 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 859 fhard = pps->kcmode & PPS_CAPTUREASSERT; 860 pcount = &pps->ppscount[0]; 861 pseq = &pps->ppsinfo.assert_sequence; 862 } else { 863 tsp = &pps->ppsinfo.clear_timestamp; 864 osp = &pps->ppsparam.clear_offset; 865 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 866 fhard = pps->kcmode & PPS_CAPTURECLEAR; 867 pcount = &pps->ppscount[1]; 868 pseq = &pps->ppsinfo.clear_sequence; 869 } 870 871 /* Nothing really happened */ 872 if (*pcount == count) 873 return; 874 875 *pcount = count; 876 877 do { 878 ts.tv_sec = gd->gd_time_seconds; 879 delta = count - gd->gd_cpuclock_base; 880 } while (ts.tv_sec != gd->gd_time_seconds); 881 if (delta > cputimer_freq) { 882 ts.tv_sec += delta / cputimer_freq; 883 delta %= cputimer_freq; 884 } 885 ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 886 ts.tv_sec += basetime.tv_sec; 887 ts.tv_nsec += basetime.tv_nsec; 888 while (ts.tv_nsec >= 1000000000) { 889 ts.tv_nsec -= 1000000000; 890 ++ts.tv_sec; 891 } 892 893 (*pseq)++; 894 *tsp = ts; 895 896 if (foff) { 897 timespecadd(tsp, osp); 898 if (tsp->tv_nsec < 0) { 899 tsp->tv_nsec += 1000000000; 900 tsp->tv_sec -= 1; 901 } 902 } 903 #ifdef PPS_SYNC 904 if (fhard) { 905 /* magic, at its best... */ 906 tcount = count - pps->ppscount[2]; 907 pps->ppscount[2] = count; 908 delta = (cputimer_freq64_nsec * tcount) >> 32; 909 hardpps(tsp, delta); 910 } 911 #endif 912 } 913 914