1 /* 2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> 35 * Copyright (c) 1982, 1986, 1991, 1993 36 * The Regents of the University of California. All rights reserved. 37 * (c) UNIX System Laboratories, Inc. 38 * All or some portions of this file are derived from material licensed 39 * to the University of California by American Telephone and Telegraph 40 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 41 * the permission of UNIX System Laboratories, Inc. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 3. All advertising materials mentioning features or use of this software 52 * must display the following acknowledgement: 53 * This product includes software developed by the University of 54 * California, Berkeley and its contributors. 55 * 4. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 72 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ 73 * $DragonFly: src/sys/kern/kern_clock.c,v 1.38 2005/04/24 02:01:08 dillon Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 78 #include <sys/param.h> 79 #include <sys/systm.h> 80 #include <sys/dkstat.h> 81 #include <sys/callout.h> 82 #include <sys/kernel.h> 83 #include <sys/kinfo.h> 84 #include <sys/proc.h> 85 #include <sys/malloc.h> 86 #include <sys/resourcevar.h> 87 #include <sys/signalvar.h> 88 #include <sys/timex.h> 89 #include <sys/timepps.h> 90 #include <vm/vm.h> 91 #include <sys/lock.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_map.h> 94 #include <sys/sysctl.h> 95 #include <sys/thread2.h> 96 97 #include <machine/cpu.h> 98 #include <machine/limits.h> 99 #include <machine/smp.h> 100 101 #ifdef GPROF 102 #include <sys/gmon.h> 103 #endif 104 105 #ifdef DEVICE_POLLING 106 extern void init_device_poll(void); 107 extern void hardclock_device_poll(void); 108 #endif /* DEVICE_POLLING */ 109 110 static void initclocks (void *dummy); 111 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 112 113 /* 114 * Some of these don't belong here, but it's easiest to concentrate them. 115 * Note that cp_time counts in microseconds, but most userland programs 116 * just compare relative times against the total by delta. 117 */ 118 struct cp_time cp_time; 119 120 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time), 121 "LU", "CPU time statistics"); 122 123 /* 124 * boottime is used to calculate the 'real' uptime. Do not confuse this with 125 * microuptime(). microtime() is not drift compensated. The real uptime 126 * with compensation is nanotime() - bootime. boottime is recalculated 127 * whenever the real time is set based on the compensated elapsed time 128 * in seconds (gd->gd_time_seconds). 129 * 130 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 131 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 132 * the real time. 133 */ 134 struct timespec boottime; /* boot time (realtime) for reference only */ 135 time_t time_second; /* read-only 'passive' uptime in seconds */ 136 137 /* 138 * basetime is used to calculate the compensated real time of day. The 139 * basetime can be modified on a per-tick basis by the adjtime(), 140 * ntp_adjtime(), and sysctl-based time correction APIs. 141 * 142 * Note that frequency corrections can also be made by adjusting 143 * gd_cpuclock_base. 144 * 145 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 146 * used on both SMP and UP systems to avoid MP races between cpu's and 147 * interrupt races on UP systems. 148 */ 149 #define BASETIME_ARYSIZE 16 150 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 151 static struct timespec basetime[BASETIME_ARYSIZE]; 152 static volatile int basetime_index; 153 154 static int 155 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 156 { 157 struct timespec *bt; 158 int error; 159 160 bt = &basetime[basetime_index]; 161 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 162 return (error); 163 } 164 165 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 166 &boottime, timespec, "System boottime"); 167 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 168 sysctl_get_basetime, "S,timespec", "System basetime"); 169 170 static void hardclock(systimer_t info, struct intrframe *frame); 171 static void statclock(systimer_t info, struct intrframe *frame); 172 static void schedclock(systimer_t info, struct intrframe *frame); 173 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 174 175 int ticks; /* system master ticks at hz */ 176 int clocks_running; /* tsleep/timeout clocks operational */ 177 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 178 int64_t nsec_acc; /* accumulator */ 179 180 /* NTPD time correction fields */ 181 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 182 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 183 int64_t ntp_delta; /* one-time correction in nsec */ 184 int64_t ntp_big_delta = 1000000000; 185 int32_t ntp_tick_delta; /* current adjustment rate */ 186 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 187 time_t ntp_leap_second; /* time of next leap second */ 188 int ntp_leap_insert; /* whether to insert or remove a second */ 189 190 /* 191 * Finish initializing clock frequencies and start all clocks running. 192 */ 193 /* ARGSUSED*/ 194 static void 195 initclocks(void *dummy) 196 { 197 cpu_initclocks(); 198 #ifdef DEVICE_POLLING 199 init_device_poll(); 200 #endif 201 /*psratio = profhz / stathz;*/ 202 initclocks_pcpu(); 203 clocks_running = 1; 204 } 205 206 /* 207 * Called on a per-cpu basis 208 */ 209 void 210 initclocks_pcpu(void) 211 { 212 struct globaldata *gd = mycpu; 213 214 crit_enter(); 215 if (gd->gd_cpuid == 0) { 216 gd->gd_time_seconds = 1; 217 gd->gd_cpuclock_base = cputimer_count(); 218 } else { 219 /* XXX */ 220 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 221 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 222 } 223 224 /* 225 * Use a non-queued periodic systimer to prevent multiple ticks from 226 * building up if the sysclock jumps forward (8254 gets reset). The 227 * sysclock will never jump backwards. Our time sync is based on 228 * the actual sysclock, not the ticks count. 229 */ 230 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 231 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 232 /* XXX correct the frequency for scheduler / estcpu tests */ 233 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 234 NULL, ESTCPUFREQ); 235 crit_exit(); 236 } 237 238 /* 239 * This sets the current real time of day. Timespecs are in seconds and 240 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 241 * instead we adjust basetime so basetime + gd_* results in the current 242 * time of day. This way the gd_* fields are guarenteed to represent 243 * a monotonically increasing 'uptime' value. 244 * 245 * When set_timeofday() is called from userland, the system call forces it 246 * onto cpu #0 since only cpu #0 can update basetime_index. 247 */ 248 void 249 set_timeofday(struct timespec *ts) 250 { 251 struct timespec *nbt; 252 int ni; 253 254 /* 255 * XXX SMP / non-atomic basetime updates 256 */ 257 crit_enter(); 258 ni = (basetime_index + 1) & BASETIME_ARYMASK; 259 nbt = &basetime[ni]; 260 nanouptime(nbt); 261 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 262 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 263 if (nbt->tv_nsec < 0) { 264 nbt->tv_nsec += 1000000000; 265 --nbt->tv_sec; 266 } 267 268 /* 269 * Note that basetime diverges from boottime as the clock drift is 270 * compensated for, so we cannot do away with boottime. When setting 271 * the absolute time of day the drift is 0 (for an instant) and we 272 * can simply assign boottime to basetime. 273 * 274 * Note that nanouptime() is based on gd_time_seconds which is drift 275 * compensated up to a point (it is guarenteed to remain monotonically 276 * increasing). gd_time_seconds is thus our best uptime guess and 277 * suitable for use in the boottime calculation. It is already taken 278 * into account in the basetime calculation above. 279 */ 280 boottime.tv_sec = nbt->tv_sec; 281 ntp_delta = 0; 282 283 /* 284 * We now have a new basetime, update the index. 285 */ 286 cpu_mb1(); 287 basetime_index = ni; 288 289 crit_exit(); 290 } 291 292 /* 293 * Each cpu has its own hardclock, but we only increments ticks and softticks 294 * on cpu #0. 295 * 296 * NOTE! systimer! the MP lock might not be held here. We can only safely 297 * manipulate objects owned by the current cpu. 298 */ 299 static void 300 hardclock(systimer_t info, struct intrframe *frame) 301 { 302 sysclock_t cputicks; 303 struct proc *p; 304 struct pstats *pstats; 305 struct globaldata *gd = mycpu; 306 307 /* 308 * Realtime updates are per-cpu. Note that timer corrections as 309 * returned by microtime() and friends make an additional adjustment 310 * using a system-wise 'basetime', but the running time is always 311 * taken from the per-cpu globaldata area. Since the same clock 312 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 313 * stay in synch. 314 * 315 * Note that we never allow info->time (aka gd->gd_hardclock.time) 316 * to reverse index gd_cpuclock_base, but that it is possible for 317 * it to temporarily get behind in the seconds if something in the 318 * system locks interrupts for a long period of time. Since periodic 319 * timers count events, though everything should resynch again 320 * immediately. 321 */ 322 cputicks = info->time - gd->gd_cpuclock_base; 323 if (cputicks >= cputimer_freq) { 324 ++gd->gd_time_seconds; 325 gd->gd_cpuclock_base += cputimer_freq; 326 } 327 328 /* 329 * The system-wide ticks counter and NTP related timedelta/tickdelta 330 * adjustments only occur on cpu #0. NTP adjustments are accomplished 331 * by updating basetime. 332 */ 333 if (gd->gd_cpuid == 0) { 334 struct timespec *nbt; 335 struct timespec nts; 336 int leap; 337 int ni; 338 339 ++ticks; 340 341 #ifdef DEVICE_POLLING 342 hardclock_device_poll(); /* mpsafe, short and quick */ 343 #endif /* DEVICE_POLLING */ 344 345 #if 0 346 if (tco->tc_poll_pps) 347 tco->tc_poll_pps(tco); 348 #endif 349 350 /* 351 * Calculate the new basetime index. We are in a critical section 352 * on cpu #0 and can safely play with basetime_index. Start 353 * with the current basetime and then make adjustments. 354 */ 355 ni = (basetime_index + 1) & BASETIME_ARYMASK; 356 nbt = &basetime[ni]; 357 *nbt = basetime[basetime_index]; 358 359 /* 360 * Apply adjtime corrections. (adjtime() API) 361 * 362 * adjtime() only runs on cpu #0 so our critical section is 363 * sufficient to access these variables. 364 */ 365 if (ntp_delta != 0) { 366 nbt->tv_nsec += ntp_tick_delta; 367 ntp_delta -= ntp_tick_delta; 368 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 369 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 370 ntp_tick_delta = ntp_delta; 371 } 372 } 373 374 /* 375 * Apply permanent frequency corrections. (sysctl API) 376 */ 377 if (ntp_tick_permanent != 0) { 378 ntp_tick_acc += ntp_tick_permanent; 379 if (ntp_tick_acc >= (1LL << 32)) { 380 nbt->tv_nsec += ntp_tick_acc >> 32; 381 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 382 } else if (ntp_tick_acc <= -(1LL << 32)) { 383 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 384 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 385 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 386 } 387 } 388 389 if (nbt->tv_nsec >= 1000000000) { 390 nbt->tv_sec++; 391 nbt->tv_nsec -= 1000000000; 392 } else if (nbt->tv_nsec < 0) { 393 nbt->tv_sec--; 394 nbt->tv_nsec += 1000000000; 395 } 396 397 /* 398 * Another per-tick compensation. (for ntp_adjtime() API) 399 */ 400 if (nsec_adj != 0) { 401 nsec_acc += nsec_adj; 402 if (nsec_acc >= 0x100000000LL) { 403 nbt->tv_nsec += nsec_acc >> 32; 404 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 405 } else if (nsec_acc <= -0x100000000LL) { 406 nbt->tv_nsec -= -nsec_acc >> 32; 407 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 408 } 409 if (nbt->tv_nsec >= 1000000000) { 410 nbt->tv_nsec -= 1000000000; 411 ++nbt->tv_sec; 412 } else if (nbt->tv_nsec < 0) { 413 nbt->tv_nsec += 1000000000; 414 --nbt->tv_sec; 415 } 416 } 417 418 /************************************************************ 419 * LEAP SECOND CORRECTION * 420 ************************************************************ 421 * 422 * Taking into account all the corrections made above, figure 423 * out the new real time. If the seconds field has changed 424 * then apply any pending leap-second corrections. 425 */ 426 getnanotime_nbt(nbt, &nts); 427 428 /* 429 * Apply leap second (sysctl API) 430 */ 431 if (ntp_leap_second) { 432 if (ntp_leap_second == nts.tv_sec) { 433 if (ntp_leap_insert) 434 nbt->tv_sec++; 435 else 436 nbt->tv_sec--; 437 ntp_leap_second--; 438 } 439 } 440 441 /* 442 * Apply leap second (ntp_adjtime() API) 443 */ 444 if (time_second != nts.tv_sec) { 445 leap = ntp_update_second(time_second, &nsec_adj); 446 nbt->tv_sec += leap; 447 time_second = nbt->tv_sec; 448 nsec_adj /= hz; 449 } 450 451 /* 452 * Finally, our new basetime is ready to go live! 453 */ 454 cpu_mb1(); 455 basetime_index = ni; 456 } 457 458 /* 459 * softticks are handled for all cpus 460 */ 461 hardclock_softtick(gd); 462 463 /* 464 * ITimer handling is per-tick, per-cpu. I don't think psignal() 465 * is mpsafe on curproc, so XXX get the mplock. 466 */ 467 if ((p = curproc) != NULL && try_mplock()) { 468 pstats = p->p_stats; 469 if (frame && CLKF_USERMODE(frame) && 470 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 471 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 472 psignal(p, SIGVTALRM); 473 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 474 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 475 psignal(p, SIGPROF); 476 rel_mplock(); 477 } 478 setdelayed(); 479 } 480 481 /* 482 * The statistics clock typically runs at a 125Hz rate, and is intended 483 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 484 * 485 * NOTE! systimer! the MP lock might not be held here. We can only safely 486 * manipulate objects owned by the current cpu. 487 * 488 * The stats clock is responsible for grabbing a profiling sample. 489 * Most of the statistics are only used by user-level statistics programs. 490 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 491 * p->p_estcpu. 492 * 493 * Like the other clocks, the stat clock is called from what is effectively 494 * a fast interrupt, so the context should be the thread/process that got 495 * interrupted. 496 */ 497 static void 498 statclock(systimer_t info, struct intrframe *frame) 499 { 500 #ifdef GPROF 501 struct gmonparam *g; 502 int i; 503 #endif 504 thread_t td; 505 struct proc *p; 506 int bump; 507 struct timeval tv; 508 struct timeval *stv; 509 510 /* 511 * How big was our timeslice relative to the last time? 512 */ 513 microuptime(&tv); /* mpsafe */ 514 stv = &mycpu->gd_stattv; 515 if (stv->tv_sec == 0) { 516 bump = 1; 517 } else { 518 bump = tv.tv_usec - stv->tv_usec + 519 (tv.tv_sec - stv->tv_sec) * 1000000; 520 if (bump < 0) 521 bump = 0; 522 if (bump > 1000000) 523 bump = 1000000; 524 } 525 *stv = tv; 526 527 td = curthread; 528 p = td->td_proc; 529 530 if (frame && CLKF_USERMODE(frame)) { 531 /* 532 * Came from userland, handle user time and deal with 533 * possible process. 534 */ 535 if (p && (p->p_flag & P_PROFIL)) 536 addupc_intr(p, CLKF_PC(frame), 1); 537 td->td_uticks += bump; 538 539 /* 540 * Charge the time as appropriate 541 */ 542 if (p && p->p_nice > NZERO) 543 cp_time.cp_nice += bump; 544 else 545 cp_time.cp_user += bump; 546 } else { 547 #ifdef GPROF 548 /* 549 * Kernel statistics are just like addupc_intr, only easier. 550 */ 551 g = &_gmonparam; 552 if (g->state == GMON_PROF_ON && frame) { 553 i = CLKF_PC(frame) - g->lowpc; 554 if (i < g->textsize) { 555 i /= HISTFRACTION * sizeof(*g->kcount); 556 g->kcount[i]++; 557 } 558 } 559 #endif 560 /* 561 * Came from kernel mode, so we were: 562 * - handling an interrupt, 563 * - doing syscall or trap work on behalf of the current 564 * user process, or 565 * - spinning in the idle loop. 566 * Whichever it is, charge the time as appropriate. 567 * Note that we charge interrupts to the current process, 568 * regardless of whether they are ``for'' that process, 569 * so that we know how much of its real time was spent 570 * in ``non-process'' (i.e., interrupt) work. 571 * 572 * XXX assume system if frame is NULL. A NULL frame 573 * can occur if ipi processing is done from an splx(). 574 */ 575 if (frame && CLKF_INTR(frame)) 576 td->td_iticks += bump; 577 else 578 td->td_sticks += bump; 579 580 if (frame && CLKF_INTR(frame)) { 581 cp_time.cp_intr += bump; 582 } else { 583 if (td == &mycpu->gd_idlethread) 584 cp_time.cp_idle += bump; 585 else 586 cp_time.cp_sys += bump; 587 } 588 } 589 } 590 591 /* 592 * The scheduler clock typically runs at a 20Hz rate. NOTE! systimer, 593 * the MP lock might not be held. We can safely manipulate parts of curproc 594 * but that's about it. 595 */ 596 static void 597 schedclock(systimer_t info, struct intrframe *frame) 598 { 599 struct proc *p; 600 struct pstats *pstats; 601 struct rusage *ru; 602 struct vmspace *vm; 603 long rss; 604 605 schedulerclock(NULL); /* mpsafe */ 606 if ((p = curproc) != NULL) { 607 /* Update resource usage integrals and maximums. */ 608 if ((pstats = p->p_stats) != NULL && 609 (ru = &pstats->p_ru) != NULL && 610 (vm = p->p_vmspace) != NULL) { 611 ru->ru_ixrss += pgtok(vm->vm_tsize); 612 ru->ru_idrss += pgtok(vm->vm_dsize); 613 ru->ru_isrss += pgtok(vm->vm_ssize); 614 rss = pgtok(vmspace_resident_count(vm)); 615 if (ru->ru_maxrss < rss) 616 ru->ru_maxrss = rss; 617 } 618 } 619 } 620 621 /* 622 * Compute number of ticks for the specified amount of time. The 623 * return value is intended to be used in a clock interrupt timed 624 * operation and guarenteed to meet or exceed the requested time. 625 * If the representation overflows, return INT_MAX. The minimum return 626 * value is 1 ticks and the function will average the calculation up. 627 * If any value greater then 0 microseconds is supplied, a value 628 * of at least 2 will be returned to ensure that a near-term clock 629 * interrupt does not cause the timeout to occur (degenerately) early. 630 * 631 * Note that limit checks must take into account microseconds, which is 632 * done simply by using the smaller signed long maximum instead of 633 * the unsigned long maximum. 634 * 635 * If ints have 32 bits, then the maximum value for any timeout in 636 * 10ms ticks is 248 days. 637 */ 638 int 639 tvtohz_high(struct timeval *tv) 640 { 641 int ticks; 642 long sec, usec; 643 644 sec = tv->tv_sec; 645 usec = tv->tv_usec; 646 if (usec < 0) { 647 sec--; 648 usec += 1000000; 649 } 650 if (sec < 0) { 651 #ifdef DIAGNOSTIC 652 if (usec > 0) { 653 sec++; 654 usec -= 1000000; 655 } 656 printf("tvotohz: negative time difference %ld sec %ld usec\n", 657 sec, usec); 658 #endif 659 ticks = 1; 660 } else if (sec <= INT_MAX / hz) { 661 ticks = (int)(sec * hz + 662 ((u_long)usec + (tick - 1)) / tick) + 1; 663 } else { 664 ticks = INT_MAX; 665 } 666 return (ticks); 667 } 668 669 /* 670 * Compute number of ticks for the specified amount of time, erroring on 671 * the side of it being too low to ensure that sleeping the returned number 672 * of ticks will not result in a late return. 673 * 674 * The supplied timeval may not be negative and should be normalized. A 675 * return value of 0 is possible if the timeval converts to less then 676 * 1 tick. 677 * 678 * If ints have 32 bits, then the maximum value for any timeout in 679 * 10ms ticks is 248 days. 680 */ 681 int 682 tvtohz_low(struct timeval *tv) 683 { 684 int ticks; 685 long sec; 686 687 sec = tv->tv_sec; 688 if (sec <= INT_MAX / hz) 689 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 690 else 691 ticks = INT_MAX; 692 return (ticks); 693 } 694 695 696 /* 697 * Start profiling on a process. 698 * 699 * Kernel profiling passes proc0 which never exits and hence 700 * keeps the profile clock running constantly. 701 */ 702 void 703 startprofclock(struct proc *p) 704 { 705 if ((p->p_flag & P_PROFIL) == 0) { 706 p->p_flag |= P_PROFIL; 707 #if 0 /* XXX */ 708 if (++profprocs == 1 && stathz != 0) { 709 s = splstatclock(); 710 psdiv = psratio; 711 setstatclockrate(profhz); 712 splx(s); 713 } 714 #endif 715 } 716 } 717 718 /* 719 * Stop profiling on a process. 720 */ 721 void 722 stopprofclock(struct proc *p) 723 { 724 if (p->p_flag & P_PROFIL) { 725 p->p_flag &= ~P_PROFIL; 726 #if 0 /* XXX */ 727 if (--profprocs == 0 && stathz != 0) { 728 s = splstatclock(); 729 psdiv = 1; 730 setstatclockrate(stathz); 731 splx(s); 732 } 733 #endif 734 } 735 } 736 737 /* 738 * Return information about system clocks. 739 */ 740 static int 741 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 742 { 743 struct kinfo_clockinfo clkinfo; 744 /* 745 * Construct clockinfo structure. 746 */ 747 clkinfo.ci_hz = hz; 748 clkinfo.ci_tick = tick; 749 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 750 clkinfo.ci_profhz = profhz; 751 clkinfo.ci_stathz = stathz ? stathz : hz; 752 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 753 } 754 755 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 756 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 757 758 /* 759 * We have eight functions for looking at the clock, four for 760 * microseconds and four for nanoseconds. For each there is fast 761 * but less precise version "get{nano|micro}[up]time" which will 762 * return a time which is up to 1/HZ previous to the call, whereas 763 * the raw version "{nano|micro}[up]time" will return a timestamp 764 * which is as precise as possible. The "up" variants return the 765 * time relative to system boot, these are well suited for time 766 * interval measurements. 767 * 768 * Each cpu independantly maintains the current time of day, so all 769 * we need to do to protect ourselves from changes is to do a loop 770 * check on the seconds field changing out from under us. 771 * 772 * The system timer maintains a 32 bit count and due to various issues 773 * it is possible for the calculated delta to occassionally exceed 774 * cputimer_freq. If this occurs the cputimer_freq64_nsec multiplication 775 * can easily overflow, so we deal with the case. For uniformity we deal 776 * with the case in the usec case too. 777 */ 778 void 779 getmicrouptime(struct timeval *tvp) 780 { 781 struct globaldata *gd = mycpu; 782 sysclock_t delta; 783 784 do { 785 tvp->tv_sec = gd->gd_time_seconds; 786 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 787 } while (tvp->tv_sec != gd->gd_time_seconds); 788 789 if (delta >= cputimer_freq) { 790 tvp->tv_sec += delta / cputimer_freq; 791 delta %= cputimer_freq; 792 } 793 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 794 if (tvp->tv_usec >= 1000000) { 795 tvp->tv_usec -= 1000000; 796 ++tvp->tv_sec; 797 } 798 } 799 800 void 801 getnanouptime(struct timespec *tsp) 802 { 803 struct globaldata *gd = mycpu; 804 sysclock_t delta; 805 806 do { 807 tsp->tv_sec = gd->gd_time_seconds; 808 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 809 } while (tsp->tv_sec != gd->gd_time_seconds); 810 811 if (delta >= cputimer_freq) { 812 tsp->tv_sec += delta / cputimer_freq; 813 delta %= cputimer_freq; 814 } 815 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 816 } 817 818 void 819 microuptime(struct timeval *tvp) 820 { 821 struct globaldata *gd = mycpu; 822 sysclock_t delta; 823 824 do { 825 tvp->tv_sec = gd->gd_time_seconds; 826 delta = cputimer_count() - gd->gd_cpuclock_base; 827 } while (tvp->tv_sec != gd->gd_time_seconds); 828 829 if (delta >= cputimer_freq) { 830 tvp->tv_sec += delta / cputimer_freq; 831 delta %= cputimer_freq; 832 } 833 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 834 } 835 836 void 837 nanouptime(struct timespec *tsp) 838 { 839 struct globaldata *gd = mycpu; 840 sysclock_t delta; 841 842 do { 843 tsp->tv_sec = gd->gd_time_seconds; 844 delta = cputimer_count() - gd->gd_cpuclock_base; 845 } while (tsp->tv_sec != gd->gd_time_seconds); 846 847 if (delta >= cputimer_freq) { 848 tsp->tv_sec += delta / cputimer_freq; 849 delta %= cputimer_freq; 850 } 851 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 852 } 853 854 /* 855 * realtime routines 856 */ 857 858 void 859 getmicrotime(struct timeval *tvp) 860 { 861 struct globaldata *gd = mycpu; 862 struct timespec *bt; 863 sysclock_t delta; 864 865 do { 866 tvp->tv_sec = gd->gd_time_seconds; 867 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 868 } while (tvp->tv_sec != gd->gd_time_seconds); 869 870 if (delta >= cputimer_freq) { 871 tvp->tv_sec += delta / cputimer_freq; 872 delta %= cputimer_freq; 873 } 874 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 875 876 bt = &basetime[basetime_index]; 877 tvp->tv_sec += bt->tv_sec; 878 tvp->tv_usec += bt->tv_nsec / 1000; 879 while (tvp->tv_usec >= 1000000) { 880 tvp->tv_usec -= 1000000; 881 ++tvp->tv_sec; 882 } 883 } 884 885 void 886 getnanotime(struct timespec *tsp) 887 { 888 struct globaldata *gd = mycpu; 889 struct timespec *bt; 890 sysclock_t delta; 891 892 do { 893 tsp->tv_sec = gd->gd_time_seconds; 894 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 895 } while (tsp->tv_sec != gd->gd_time_seconds); 896 897 if (delta >= cputimer_freq) { 898 tsp->tv_sec += delta / cputimer_freq; 899 delta %= cputimer_freq; 900 } 901 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 902 903 bt = &basetime[basetime_index]; 904 tsp->tv_sec += bt->tv_sec; 905 tsp->tv_nsec += bt->tv_nsec; 906 while (tsp->tv_nsec >= 1000000000) { 907 tsp->tv_nsec -= 1000000000; 908 ++tsp->tv_sec; 909 } 910 } 911 912 static void 913 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 914 { 915 struct globaldata *gd = mycpu; 916 sysclock_t delta; 917 918 do { 919 tsp->tv_sec = gd->gd_time_seconds; 920 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 921 } while (tsp->tv_sec != gd->gd_time_seconds); 922 923 if (delta >= cputimer_freq) { 924 tsp->tv_sec += delta / cputimer_freq; 925 delta %= cputimer_freq; 926 } 927 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 928 929 tsp->tv_sec += nbt->tv_sec; 930 tsp->tv_nsec += nbt->tv_nsec; 931 while (tsp->tv_nsec >= 1000000000) { 932 tsp->tv_nsec -= 1000000000; 933 ++tsp->tv_sec; 934 } 935 } 936 937 938 void 939 microtime(struct timeval *tvp) 940 { 941 struct globaldata *gd = mycpu; 942 struct timespec *bt; 943 sysclock_t delta; 944 945 do { 946 tvp->tv_sec = gd->gd_time_seconds; 947 delta = cputimer_count() - gd->gd_cpuclock_base; 948 } while (tvp->tv_sec != gd->gd_time_seconds); 949 950 if (delta >= cputimer_freq) { 951 tvp->tv_sec += delta / cputimer_freq; 952 delta %= cputimer_freq; 953 } 954 tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32; 955 956 bt = &basetime[basetime_index]; 957 tvp->tv_sec += bt->tv_sec; 958 tvp->tv_usec += bt->tv_nsec / 1000; 959 while (tvp->tv_usec >= 1000000) { 960 tvp->tv_usec -= 1000000; 961 ++tvp->tv_sec; 962 } 963 } 964 965 void 966 nanotime(struct timespec *tsp) 967 { 968 struct globaldata *gd = mycpu; 969 struct timespec *bt; 970 sysclock_t delta; 971 972 do { 973 tsp->tv_sec = gd->gd_time_seconds; 974 delta = cputimer_count() - gd->gd_cpuclock_base; 975 } while (tsp->tv_sec != gd->gd_time_seconds); 976 977 if (delta >= cputimer_freq) { 978 tsp->tv_sec += delta / cputimer_freq; 979 delta %= cputimer_freq; 980 } 981 tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 982 983 bt = &basetime[basetime_index]; 984 tsp->tv_sec += bt->tv_sec; 985 tsp->tv_nsec += bt->tv_nsec; 986 while (tsp->tv_nsec >= 1000000000) { 987 tsp->tv_nsec -= 1000000000; 988 ++tsp->tv_sec; 989 } 990 } 991 992 /* 993 * note: this is not exactly synchronized with real time. To do that we 994 * would have to do what microtime does and check for a nanoseconds overflow. 995 */ 996 time_t 997 get_approximate_time_t(void) 998 { 999 struct globaldata *gd = mycpu; 1000 struct timespec *bt; 1001 1002 bt = &basetime[basetime_index]; 1003 return(gd->gd_time_seconds + bt->tv_sec); 1004 } 1005 1006 int 1007 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1008 { 1009 pps_params_t *app; 1010 struct pps_fetch_args *fapi; 1011 #ifdef PPS_SYNC 1012 struct pps_kcbind_args *kapi; 1013 #endif 1014 1015 switch (cmd) { 1016 case PPS_IOC_CREATE: 1017 return (0); 1018 case PPS_IOC_DESTROY: 1019 return (0); 1020 case PPS_IOC_SETPARAMS: 1021 app = (pps_params_t *)data; 1022 if (app->mode & ~pps->ppscap) 1023 return (EINVAL); 1024 pps->ppsparam = *app; 1025 return (0); 1026 case PPS_IOC_GETPARAMS: 1027 app = (pps_params_t *)data; 1028 *app = pps->ppsparam; 1029 app->api_version = PPS_API_VERS_1; 1030 return (0); 1031 case PPS_IOC_GETCAP: 1032 *(int*)data = pps->ppscap; 1033 return (0); 1034 case PPS_IOC_FETCH: 1035 fapi = (struct pps_fetch_args *)data; 1036 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1037 return (EINVAL); 1038 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1039 return (EOPNOTSUPP); 1040 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1041 fapi->pps_info_buf = pps->ppsinfo; 1042 return (0); 1043 case PPS_IOC_KCBIND: 1044 #ifdef PPS_SYNC 1045 kapi = (struct pps_kcbind_args *)data; 1046 /* XXX Only root should be able to do this */ 1047 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1048 return (EINVAL); 1049 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1050 return (EINVAL); 1051 if (kapi->edge & ~pps->ppscap) 1052 return (EINVAL); 1053 pps->kcmode = kapi->edge; 1054 return (0); 1055 #else 1056 return (EOPNOTSUPP); 1057 #endif 1058 default: 1059 return (ENOTTY); 1060 } 1061 } 1062 1063 void 1064 pps_init(struct pps_state *pps) 1065 { 1066 pps->ppscap |= PPS_TSFMT_TSPEC; 1067 if (pps->ppscap & PPS_CAPTUREASSERT) 1068 pps->ppscap |= PPS_OFFSETASSERT; 1069 if (pps->ppscap & PPS_CAPTURECLEAR) 1070 pps->ppscap |= PPS_OFFSETCLEAR; 1071 } 1072 1073 void 1074 pps_event(struct pps_state *pps, sysclock_t count, int event) 1075 { 1076 struct globaldata *gd; 1077 struct timespec *tsp; 1078 struct timespec *osp; 1079 struct timespec *bt; 1080 struct timespec ts; 1081 sysclock_t *pcount; 1082 #ifdef PPS_SYNC 1083 sysclock_t tcount; 1084 #endif 1085 sysclock_t delta; 1086 pps_seq_t *pseq; 1087 int foff; 1088 int fhard; 1089 1090 gd = mycpu; 1091 1092 /* Things would be easier with arrays... */ 1093 if (event == PPS_CAPTUREASSERT) { 1094 tsp = &pps->ppsinfo.assert_timestamp; 1095 osp = &pps->ppsparam.assert_offset; 1096 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1097 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1098 pcount = &pps->ppscount[0]; 1099 pseq = &pps->ppsinfo.assert_sequence; 1100 } else { 1101 tsp = &pps->ppsinfo.clear_timestamp; 1102 osp = &pps->ppsparam.clear_offset; 1103 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1104 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1105 pcount = &pps->ppscount[1]; 1106 pseq = &pps->ppsinfo.clear_sequence; 1107 } 1108 1109 /* Nothing really happened */ 1110 if (*pcount == count) 1111 return; 1112 1113 *pcount = count; 1114 1115 do { 1116 ts.tv_sec = gd->gd_time_seconds; 1117 delta = count - gd->gd_cpuclock_base; 1118 } while (ts.tv_sec != gd->gd_time_seconds); 1119 1120 if (delta >= cputimer_freq) { 1121 ts.tv_sec += delta / cputimer_freq; 1122 delta %= cputimer_freq; 1123 } 1124 ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32; 1125 bt = &basetime[basetime_index]; 1126 ts.tv_sec += bt->tv_sec; 1127 ts.tv_nsec += bt->tv_nsec; 1128 while (ts.tv_nsec >= 1000000000) { 1129 ts.tv_nsec -= 1000000000; 1130 ++ts.tv_sec; 1131 } 1132 1133 (*pseq)++; 1134 *tsp = ts; 1135 1136 if (foff) { 1137 timespecadd(tsp, osp); 1138 if (tsp->tv_nsec < 0) { 1139 tsp->tv_nsec += 1000000000; 1140 tsp->tv_sec -= 1; 1141 } 1142 } 1143 #ifdef PPS_SYNC 1144 if (fhard) { 1145 /* magic, at its best... */ 1146 tcount = count - pps->ppscount[2]; 1147 pps->ppscount[2] = count; 1148 if (tcount >= cputimer_freq) { 1149 delta = (1000000000 * (tcount / cputimer_freq) + 1150 cputimer_freq64_nsec * 1151 (tcount % cputimer_freq)) >> 32; 1152 } else { 1153 delta = (cputimer_freq64_nsec * tcount) >> 32; 1154 } 1155 hardpps(tsp, delta); 1156 } 1157 #endif 1158 } 1159 1160