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 1.43 2005/06/06 15:02:27 dillon Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 78 #include <sys/param.h> 79 #include <sys/systm.h> 80 #include <sys/callout.h> 81 #include <sys/kernel.h> 82 #include <sys/kinfo.h> 83 #include <sys/proc.h> 84 #include <sys/malloc.h> 85 #include <sys/resourcevar.h> 86 #include <sys/signalvar.h> 87 #include <sys/timex.h> 88 #include <sys/timepps.h> 89 #include <vm/vm.h> 90 #include <sys/lock.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <sys/sysctl.h> 94 #include <sys/thread2.h> 95 96 #include <machine/cpu.h> 97 #include <machine/limits.h> 98 #include <machine/smp.h> 99 100 #ifdef GPROF 101 #include <sys/gmon.h> 102 #endif 103 104 #ifdef DEVICE_POLLING 105 extern void init_device_poll(void); 106 extern void hardclock_device_poll(void); 107 #endif /* DEVICE_POLLING */ 108 109 static void initclocks (void *dummy); 110 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 111 112 /* 113 * Some of these don't belong here, but it's easiest to concentrate them. 114 * Note that cpu_time counts in microseconds, but most userland programs 115 * just compare relative times against the total by delta. 116 */ 117 struct kinfo_cputime cputime_percpu[MAXCPU]; 118 #ifdef SMP 119 static int 120 sysctl_cputime(SYSCTL_HANDLER_ARGS) 121 { 122 int cpu, error = 0; 123 size_t size = sizeof(struct kinfo_cputime); 124 125 for (cpu = 0; cpu < ncpus; ++cpu) { 126 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 127 break; 128 } 129 130 return (error); 131 } 132 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 133 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 134 #else 135 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 136 "CPU time statistics"); 137 #endif 138 139 /* 140 * boottime is used to calculate the 'real' uptime. Do not confuse this with 141 * microuptime(). microtime() is not drift compensated. The real uptime 142 * with compensation is nanotime() - bootime. boottime is recalculated 143 * whenever the real time is set based on the compensated elapsed time 144 * in seconds (gd->gd_time_seconds). 145 * 146 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 147 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 148 * the real time. 149 */ 150 struct timespec boottime; /* boot time (realtime) for reference only */ 151 time_t time_second; /* read-only 'passive' uptime in seconds */ 152 153 /* 154 * basetime is used to calculate the compensated real time of day. The 155 * basetime can be modified on a per-tick basis by the adjtime(), 156 * ntp_adjtime(), and sysctl-based time correction APIs. 157 * 158 * Note that frequency corrections can also be made by adjusting 159 * gd_cpuclock_base. 160 * 161 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 162 * used on both SMP and UP systems to avoid MP races between cpu's and 163 * interrupt races on UP systems. 164 */ 165 #define BASETIME_ARYSIZE 16 166 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 167 static struct timespec basetime[BASETIME_ARYSIZE]; 168 static volatile int basetime_index; 169 170 static int 171 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 172 { 173 struct timespec *bt; 174 int error; 175 int index; 176 177 /* 178 * Because basetime data and index may be updated by another cpu, 179 * a load fence is required to ensure that the data we read has 180 * not been speculatively read relative to a possibly updated index. 181 */ 182 index = basetime_index; 183 cpu_lfence(); 184 bt = &basetime[index]; 185 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 186 return (error); 187 } 188 189 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 190 &boottime, timespec, "System boottime"); 191 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 192 sysctl_get_basetime, "S,timespec", "System basetime"); 193 194 static void hardclock(systimer_t info, struct intrframe *frame); 195 static void statclock(systimer_t info, struct intrframe *frame); 196 static void schedclock(systimer_t info, struct intrframe *frame); 197 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 198 199 int ticks; /* system master ticks at hz */ 200 int clocks_running; /* tsleep/timeout clocks operational */ 201 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 202 int64_t nsec_acc; /* accumulator */ 203 204 /* NTPD time correction fields */ 205 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 206 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 207 int64_t ntp_delta; /* one-time correction in nsec */ 208 int64_t ntp_big_delta = 1000000000; 209 int32_t ntp_tick_delta; /* current adjustment rate */ 210 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 211 time_t ntp_leap_second; /* time of next leap second */ 212 int ntp_leap_insert; /* whether to insert or remove a second */ 213 214 /* 215 * Finish initializing clock frequencies and start all clocks running. 216 */ 217 /* ARGSUSED*/ 218 static void 219 initclocks(void *dummy) 220 { 221 cpu_initclocks(); 222 #ifdef DEVICE_POLLING 223 init_device_poll(); 224 #endif 225 /*psratio = profhz / stathz;*/ 226 initclocks_pcpu(); 227 clocks_running = 1; 228 } 229 230 /* 231 * Called on a per-cpu basis 232 */ 233 void 234 initclocks_pcpu(void) 235 { 236 struct globaldata *gd = mycpu; 237 238 crit_enter(); 239 if (gd->gd_cpuid == 0) { 240 gd->gd_time_seconds = 1; 241 gd->gd_cpuclock_base = sys_cputimer->count(); 242 } else { 243 /* XXX */ 244 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 245 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 246 } 247 248 /* 249 * Use a non-queued periodic systimer to prevent multiple ticks from 250 * building up if the sysclock jumps forward (8254 gets reset). The 251 * sysclock will never jump backwards. Our time sync is based on 252 * the actual sysclock, not the ticks count. 253 */ 254 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 255 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 256 /* XXX correct the frequency for scheduler / estcpu tests */ 257 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 258 NULL, ESTCPUFREQ); 259 crit_exit(); 260 } 261 262 /* 263 * This sets the current real time of day. Timespecs are in seconds and 264 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 265 * instead we adjust basetime so basetime + gd_* results in the current 266 * time of day. This way the gd_* fields are guarenteed to represent 267 * a monotonically increasing 'uptime' value. 268 * 269 * When set_timeofday() is called from userland, the system call forces it 270 * onto cpu #0 since only cpu #0 can update basetime_index. 271 */ 272 void 273 set_timeofday(struct timespec *ts) 274 { 275 struct timespec *nbt; 276 int ni; 277 278 /* 279 * XXX SMP / non-atomic basetime updates 280 */ 281 crit_enter(); 282 ni = (basetime_index + 1) & BASETIME_ARYMASK; 283 nbt = &basetime[ni]; 284 nanouptime(nbt); 285 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 286 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 287 if (nbt->tv_nsec < 0) { 288 nbt->tv_nsec += 1000000000; 289 --nbt->tv_sec; 290 } 291 292 /* 293 * Note that basetime diverges from boottime as the clock drift is 294 * compensated for, so we cannot do away with boottime. When setting 295 * the absolute time of day the drift is 0 (for an instant) and we 296 * can simply assign boottime to basetime. 297 * 298 * Note that nanouptime() is based on gd_time_seconds which is drift 299 * compensated up to a point (it is guarenteed to remain monotonically 300 * increasing). gd_time_seconds is thus our best uptime guess and 301 * suitable for use in the boottime calculation. It is already taken 302 * into account in the basetime calculation above. 303 */ 304 boottime.tv_sec = nbt->tv_sec; 305 ntp_delta = 0; 306 307 /* 308 * We now have a new basetime, make sure all other cpus have it, 309 * then update the index. 310 */ 311 cpu_sfence(); 312 basetime_index = ni; 313 314 crit_exit(); 315 } 316 317 /* 318 * Each cpu has its own hardclock, but we only increments ticks and softticks 319 * on cpu #0. 320 * 321 * NOTE! systimer! the MP lock might not be held here. We can only safely 322 * manipulate objects owned by the current cpu. 323 */ 324 static void 325 hardclock(systimer_t info, struct intrframe *frame) 326 { 327 sysclock_t cputicks; 328 struct proc *p; 329 struct pstats *pstats; 330 struct globaldata *gd = mycpu; 331 332 /* 333 * Realtime updates are per-cpu. Note that timer corrections as 334 * returned by microtime() and friends make an additional adjustment 335 * using a system-wise 'basetime', but the running time is always 336 * taken from the per-cpu globaldata area. Since the same clock 337 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 338 * stay in synch. 339 * 340 * Note that we never allow info->time (aka gd->gd_hardclock.time) 341 * to reverse index gd_cpuclock_base, but that it is possible for 342 * it to temporarily get behind in the seconds if something in the 343 * system locks interrupts for a long period of time. Since periodic 344 * timers count events, though everything should resynch again 345 * immediately. 346 */ 347 cputicks = info->time - gd->gd_cpuclock_base; 348 if (cputicks >= sys_cputimer->freq) { 349 ++gd->gd_time_seconds; 350 gd->gd_cpuclock_base += sys_cputimer->freq; 351 } 352 353 /* 354 * The system-wide ticks counter and NTP related timedelta/tickdelta 355 * adjustments only occur on cpu #0. NTP adjustments are accomplished 356 * by updating basetime. 357 */ 358 if (gd->gd_cpuid == 0) { 359 struct timespec *nbt; 360 struct timespec nts; 361 int leap; 362 int ni; 363 364 ++ticks; 365 366 #ifdef DEVICE_POLLING 367 hardclock_device_poll(); /* mpsafe, short and quick */ 368 #endif /* DEVICE_POLLING */ 369 370 #if 0 371 if (tco->tc_poll_pps) 372 tco->tc_poll_pps(tco); 373 #endif 374 375 /* 376 * Calculate the new basetime index. We are in a critical section 377 * on cpu #0 and can safely play with basetime_index. Start 378 * with the current basetime and then make adjustments. 379 */ 380 ni = (basetime_index + 1) & BASETIME_ARYMASK; 381 nbt = &basetime[ni]; 382 *nbt = basetime[basetime_index]; 383 384 /* 385 * Apply adjtime corrections. (adjtime() API) 386 * 387 * adjtime() only runs on cpu #0 so our critical section is 388 * sufficient to access these variables. 389 */ 390 if (ntp_delta != 0) { 391 nbt->tv_nsec += ntp_tick_delta; 392 ntp_delta -= ntp_tick_delta; 393 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 394 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 395 ntp_tick_delta = ntp_delta; 396 } 397 } 398 399 /* 400 * Apply permanent frequency corrections. (sysctl API) 401 */ 402 if (ntp_tick_permanent != 0) { 403 ntp_tick_acc += ntp_tick_permanent; 404 if (ntp_tick_acc >= (1LL << 32)) { 405 nbt->tv_nsec += ntp_tick_acc >> 32; 406 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 407 } else if (ntp_tick_acc <= -(1LL << 32)) { 408 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 409 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 410 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 411 } 412 } 413 414 if (nbt->tv_nsec >= 1000000000) { 415 nbt->tv_sec++; 416 nbt->tv_nsec -= 1000000000; 417 } else if (nbt->tv_nsec < 0) { 418 nbt->tv_sec--; 419 nbt->tv_nsec += 1000000000; 420 } 421 422 /* 423 * Another per-tick compensation. (for ntp_adjtime() API) 424 */ 425 if (nsec_adj != 0) { 426 nsec_acc += nsec_adj; 427 if (nsec_acc >= 0x100000000LL) { 428 nbt->tv_nsec += nsec_acc >> 32; 429 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 430 } else if (nsec_acc <= -0x100000000LL) { 431 nbt->tv_nsec -= -nsec_acc >> 32; 432 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 433 } 434 if (nbt->tv_nsec >= 1000000000) { 435 nbt->tv_nsec -= 1000000000; 436 ++nbt->tv_sec; 437 } else if (nbt->tv_nsec < 0) { 438 nbt->tv_nsec += 1000000000; 439 --nbt->tv_sec; 440 } 441 } 442 443 /************************************************************ 444 * LEAP SECOND CORRECTION * 445 ************************************************************ 446 * 447 * Taking into account all the corrections made above, figure 448 * out the new real time. If the seconds field has changed 449 * then apply any pending leap-second corrections. 450 */ 451 getnanotime_nbt(nbt, &nts); 452 453 if (time_second != nts.tv_sec) { 454 /* 455 * Apply leap second (sysctl API). Adjust nts for changes 456 * so we do not have to call getnanotime_nbt again. 457 */ 458 if (ntp_leap_second) { 459 if (ntp_leap_second == nts.tv_sec) { 460 if (ntp_leap_insert) { 461 nbt->tv_sec++; 462 nts.tv_sec++; 463 } else { 464 nbt->tv_sec--; 465 nts.tv_sec--; 466 } 467 ntp_leap_second--; 468 } 469 } 470 471 /* 472 * Apply leap second (ntp_adjtime() API), calculate a new 473 * nsec_adj field. ntp_update_second() returns nsec_adj 474 * as a per-second value but we need it as a per-tick value. 475 */ 476 leap = ntp_update_second(time_second, &nsec_adj); 477 nsec_adj /= hz; 478 nbt->tv_sec += leap; 479 nts.tv_sec += leap; 480 481 /* 482 * Update the time_second 'approximate time' global. 483 */ 484 time_second = nts.tv_sec; 485 } 486 487 /* 488 * Finally, our new basetime is ready to go live! 489 */ 490 cpu_sfence(); 491 basetime_index = ni; 492 } 493 494 /* 495 * softticks are handled for all cpus 496 */ 497 hardclock_softtick(gd); 498 499 /* 500 * ITimer handling is per-tick, per-cpu. I don't think psignal() 501 * is mpsafe on curproc, so XXX get the mplock. 502 */ 503 if ((p = curproc) != NULL && try_mplock()) { 504 pstats = p->p_stats; 505 if (frame && CLKF_USERMODE(frame) && 506 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 507 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 508 psignal(p, SIGVTALRM); 509 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 510 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 511 psignal(p, SIGPROF); 512 rel_mplock(); 513 } 514 setdelayed(); 515 } 516 517 /* 518 * The statistics clock typically runs at a 125Hz rate, and is intended 519 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 520 * 521 * NOTE! systimer! the MP lock might not be held here. We can only safely 522 * manipulate objects owned by the current cpu. 523 * 524 * The stats clock is responsible for grabbing a profiling sample. 525 * Most of the statistics are only used by user-level statistics programs. 526 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 527 * p->p_estcpu. 528 * 529 * Like the other clocks, the stat clock is called from what is effectively 530 * a fast interrupt, so the context should be the thread/process that got 531 * interrupted. 532 */ 533 static void 534 statclock(systimer_t info, struct intrframe *frame) 535 { 536 #ifdef GPROF 537 struct gmonparam *g; 538 int i; 539 #endif 540 thread_t td; 541 struct proc *p; 542 int bump; 543 struct timeval tv; 544 struct timeval *stv; 545 546 /* 547 * How big was our timeslice relative to the last time? 548 */ 549 microuptime(&tv); /* mpsafe */ 550 stv = &mycpu->gd_stattv; 551 if (stv->tv_sec == 0) { 552 bump = 1; 553 } else { 554 bump = tv.tv_usec - stv->tv_usec + 555 (tv.tv_sec - stv->tv_sec) * 1000000; 556 if (bump < 0) 557 bump = 0; 558 if (bump > 1000000) 559 bump = 1000000; 560 } 561 *stv = tv; 562 563 td = curthread; 564 p = td->td_proc; 565 566 if (frame && CLKF_USERMODE(frame)) { 567 /* 568 * Came from userland, handle user time and deal with 569 * possible process. 570 */ 571 if (p && (p->p_flag & P_PROFIL)) 572 addupc_intr(p, CLKF_PC(frame), 1); 573 td->td_uticks += bump; 574 575 /* 576 * Charge the time as appropriate 577 */ 578 if (p && p->p_nice > NZERO) 579 cpu_time.cp_nice += bump; 580 else 581 cpu_time.cp_user += bump; 582 } else { 583 #ifdef GPROF 584 /* 585 * Kernel statistics are just like addupc_intr, only easier. 586 */ 587 g = &_gmonparam; 588 if (g->state == GMON_PROF_ON && frame) { 589 i = CLKF_PC(frame) - g->lowpc; 590 if (i < g->textsize) { 591 i /= HISTFRACTION * sizeof(*g->kcount); 592 g->kcount[i]++; 593 } 594 } 595 #endif 596 /* 597 * Came from kernel mode, so we were: 598 * - handling an interrupt, 599 * - doing syscall or trap work on behalf of the current 600 * user process, or 601 * - spinning in the idle loop. 602 * Whichever it is, charge the time as appropriate. 603 * Note that we charge interrupts to the current process, 604 * regardless of whether they are ``for'' that process, 605 * so that we know how much of its real time was spent 606 * in ``non-process'' (i.e., interrupt) work. 607 * 608 * XXX assume system if frame is NULL. A NULL frame 609 * can occur if ipi processing is done from a crit_exit(). 610 */ 611 if (frame && CLKF_INTR(frame)) 612 td->td_iticks += bump; 613 else 614 td->td_sticks += bump; 615 616 if (frame && CLKF_INTR(frame)) { 617 cpu_time.cp_intr += bump; 618 } else { 619 if (td == &mycpu->gd_idlethread) 620 cpu_time.cp_idle += bump; 621 else 622 cpu_time.cp_sys += bump; 623 } 624 } 625 } 626 627 /* 628 * The scheduler clock typically runs at a 20Hz rate. NOTE! systimer, 629 * the MP lock might not be held. We can safely manipulate parts of curproc 630 * but that's about it. 631 */ 632 static void 633 schedclock(systimer_t info, struct intrframe *frame) 634 { 635 struct proc *p; 636 struct pstats *pstats; 637 struct rusage *ru; 638 struct vmspace *vm; 639 long rss; 640 641 schedulerclock(NULL); /* mpsafe */ 642 if ((p = curproc) != NULL) { 643 /* Update resource usage integrals and maximums. */ 644 if ((pstats = p->p_stats) != NULL && 645 (ru = &pstats->p_ru) != NULL && 646 (vm = p->p_vmspace) != NULL) { 647 ru->ru_ixrss += pgtok(vm->vm_tsize); 648 ru->ru_idrss += pgtok(vm->vm_dsize); 649 ru->ru_isrss += pgtok(vm->vm_ssize); 650 rss = pgtok(vmspace_resident_count(vm)); 651 if (ru->ru_maxrss < rss) 652 ru->ru_maxrss = rss; 653 } 654 } 655 } 656 657 /* 658 * Compute number of ticks for the specified amount of time. The 659 * return value is intended to be used in a clock interrupt timed 660 * operation and guarenteed to meet or exceed the requested time. 661 * If the representation overflows, return INT_MAX. The minimum return 662 * value is 1 ticks and the function will average the calculation up. 663 * If any value greater then 0 microseconds is supplied, a value 664 * of at least 2 will be returned to ensure that a near-term clock 665 * interrupt does not cause the timeout to occur (degenerately) early. 666 * 667 * Note that limit checks must take into account microseconds, which is 668 * done simply by using the smaller signed long maximum instead of 669 * the unsigned long maximum. 670 * 671 * If ints have 32 bits, then the maximum value for any timeout in 672 * 10ms ticks is 248 days. 673 */ 674 int 675 tvtohz_high(struct timeval *tv) 676 { 677 int ticks; 678 long sec, usec; 679 680 sec = tv->tv_sec; 681 usec = tv->tv_usec; 682 if (usec < 0) { 683 sec--; 684 usec += 1000000; 685 } 686 if (sec < 0) { 687 #ifdef DIAGNOSTIC 688 if (usec > 0) { 689 sec++; 690 usec -= 1000000; 691 } 692 printf("tvotohz: negative time difference %ld sec %ld usec\n", 693 sec, usec); 694 #endif 695 ticks = 1; 696 } else if (sec <= INT_MAX / hz) { 697 ticks = (int)(sec * hz + 698 ((u_long)usec + (tick - 1)) / tick) + 1; 699 } else { 700 ticks = INT_MAX; 701 } 702 return (ticks); 703 } 704 705 /* 706 * Compute number of ticks for the specified amount of time, erroring on 707 * the side of it being too low to ensure that sleeping the returned number 708 * of ticks will not result in a late return. 709 * 710 * The supplied timeval may not be negative and should be normalized. A 711 * return value of 0 is possible if the timeval converts to less then 712 * 1 tick. 713 * 714 * If ints have 32 bits, then the maximum value for any timeout in 715 * 10ms ticks is 248 days. 716 */ 717 int 718 tvtohz_low(struct timeval *tv) 719 { 720 int ticks; 721 long sec; 722 723 sec = tv->tv_sec; 724 if (sec <= INT_MAX / hz) 725 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 726 else 727 ticks = INT_MAX; 728 return (ticks); 729 } 730 731 732 /* 733 * Start profiling on a process. 734 * 735 * Kernel profiling passes proc0 which never exits and hence 736 * keeps the profile clock running constantly. 737 */ 738 void 739 startprofclock(struct proc *p) 740 { 741 if ((p->p_flag & P_PROFIL) == 0) { 742 p->p_flag |= P_PROFIL; 743 #if 0 /* XXX */ 744 if (++profprocs == 1 && stathz != 0) { 745 crit_enter(); 746 psdiv = psratio; 747 setstatclockrate(profhz); 748 crit_exit(); 749 } 750 #endif 751 } 752 } 753 754 /* 755 * Stop profiling on a process. 756 */ 757 void 758 stopprofclock(struct proc *p) 759 { 760 if (p->p_flag & P_PROFIL) { 761 p->p_flag &= ~P_PROFIL; 762 #if 0 /* XXX */ 763 if (--profprocs == 0 && stathz != 0) { 764 crit_enter(); 765 psdiv = 1; 766 setstatclockrate(stathz); 767 crit_exit(); 768 } 769 #endif 770 } 771 } 772 773 /* 774 * Return information about system clocks. 775 */ 776 static int 777 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 778 { 779 struct kinfo_clockinfo clkinfo; 780 /* 781 * Construct clockinfo structure. 782 */ 783 clkinfo.ci_hz = hz; 784 clkinfo.ci_tick = tick; 785 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 786 clkinfo.ci_profhz = profhz; 787 clkinfo.ci_stathz = stathz ? stathz : hz; 788 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 789 } 790 791 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 792 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 793 794 /* 795 * We have eight functions for looking at the clock, four for 796 * microseconds and four for nanoseconds. For each there is fast 797 * but less precise version "get{nano|micro}[up]time" which will 798 * return a time which is up to 1/HZ previous to the call, whereas 799 * the raw version "{nano|micro}[up]time" will return a timestamp 800 * which is as precise as possible. The "up" variants return the 801 * time relative to system boot, these are well suited for time 802 * interval measurements. 803 * 804 * Each cpu independantly maintains the current time of day, so all 805 * we need to do to protect ourselves from changes is to do a loop 806 * check on the seconds field changing out from under us. 807 * 808 * The system timer maintains a 32 bit count and due to various issues 809 * it is possible for the calculated delta to occassionally exceed 810 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 811 * multiplication can easily overflow, so we deal with the case. For 812 * uniformity we deal with the case in the usec case too. 813 */ 814 void 815 getmicrouptime(struct timeval *tvp) 816 { 817 struct globaldata *gd = mycpu; 818 sysclock_t delta; 819 820 do { 821 tvp->tv_sec = gd->gd_time_seconds; 822 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 823 } while (tvp->tv_sec != gd->gd_time_seconds); 824 825 if (delta >= sys_cputimer->freq) { 826 tvp->tv_sec += delta / sys_cputimer->freq; 827 delta %= sys_cputimer->freq; 828 } 829 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 830 if (tvp->tv_usec >= 1000000) { 831 tvp->tv_usec -= 1000000; 832 ++tvp->tv_sec; 833 } 834 } 835 836 void 837 getnanouptime(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 = gd->gd_hardclock.time - gd->gd_cpuclock_base; 845 } while (tsp->tv_sec != gd->gd_time_seconds); 846 847 if (delta >= sys_cputimer->freq) { 848 tsp->tv_sec += delta / sys_cputimer->freq; 849 delta %= sys_cputimer->freq; 850 } 851 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 852 } 853 854 void 855 microuptime(struct timeval *tvp) 856 { 857 struct globaldata *gd = mycpu; 858 sysclock_t delta; 859 860 do { 861 tvp->tv_sec = gd->gd_time_seconds; 862 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 863 } while (tvp->tv_sec != gd->gd_time_seconds); 864 865 if (delta >= sys_cputimer->freq) { 866 tvp->tv_sec += delta / sys_cputimer->freq; 867 delta %= sys_cputimer->freq; 868 } 869 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 870 } 871 872 void 873 nanouptime(struct timespec *tsp) 874 { 875 struct globaldata *gd = mycpu; 876 sysclock_t delta; 877 878 do { 879 tsp->tv_sec = gd->gd_time_seconds; 880 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 881 } while (tsp->tv_sec != gd->gd_time_seconds); 882 883 if (delta >= sys_cputimer->freq) { 884 tsp->tv_sec += delta / sys_cputimer->freq; 885 delta %= sys_cputimer->freq; 886 } 887 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 888 } 889 890 /* 891 * realtime routines 892 */ 893 894 void 895 getmicrotime(struct timeval *tvp) 896 { 897 struct globaldata *gd = mycpu; 898 struct timespec *bt; 899 sysclock_t delta; 900 901 do { 902 tvp->tv_sec = gd->gd_time_seconds; 903 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 904 } while (tvp->tv_sec != gd->gd_time_seconds); 905 906 if (delta >= sys_cputimer->freq) { 907 tvp->tv_sec += delta / sys_cputimer->freq; 908 delta %= sys_cputimer->freq; 909 } 910 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 911 912 bt = &basetime[basetime_index]; 913 tvp->tv_sec += bt->tv_sec; 914 tvp->tv_usec += bt->tv_nsec / 1000; 915 while (tvp->tv_usec >= 1000000) { 916 tvp->tv_usec -= 1000000; 917 ++tvp->tv_sec; 918 } 919 } 920 921 void 922 getnanotime(struct timespec *tsp) 923 { 924 struct globaldata *gd = mycpu; 925 struct timespec *bt; 926 sysclock_t delta; 927 928 do { 929 tsp->tv_sec = gd->gd_time_seconds; 930 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 931 } while (tsp->tv_sec != gd->gd_time_seconds); 932 933 if (delta >= sys_cputimer->freq) { 934 tsp->tv_sec += delta / sys_cputimer->freq; 935 delta %= sys_cputimer->freq; 936 } 937 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 938 939 bt = &basetime[basetime_index]; 940 tsp->tv_sec += bt->tv_sec; 941 tsp->tv_nsec += bt->tv_nsec; 942 while (tsp->tv_nsec >= 1000000000) { 943 tsp->tv_nsec -= 1000000000; 944 ++tsp->tv_sec; 945 } 946 } 947 948 static void 949 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 950 { 951 struct globaldata *gd = mycpu; 952 sysclock_t delta; 953 954 do { 955 tsp->tv_sec = gd->gd_time_seconds; 956 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 957 } while (tsp->tv_sec != gd->gd_time_seconds); 958 959 if (delta >= sys_cputimer->freq) { 960 tsp->tv_sec += delta / sys_cputimer->freq; 961 delta %= sys_cputimer->freq; 962 } 963 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 964 965 tsp->tv_sec += nbt->tv_sec; 966 tsp->tv_nsec += nbt->tv_nsec; 967 while (tsp->tv_nsec >= 1000000000) { 968 tsp->tv_nsec -= 1000000000; 969 ++tsp->tv_sec; 970 } 971 } 972 973 974 void 975 microtime(struct timeval *tvp) 976 { 977 struct globaldata *gd = mycpu; 978 struct timespec *bt; 979 sysclock_t delta; 980 981 do { 982 tvp->tv_sec = gd->gd_time_seconds; 983 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 984 } while (tvp->tv_sec != gd->gd_time_seconds); 985 986 if (delta >= sys_cputimer->freq) { 987 tvp->tv_sec += delta / sys_cputimer->freq; 988 delta %= sys_cputimer->freq; 989 } 990 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 991 992 bt = &basetime[basetime_index]; 993 tvp->tv_sec += bt->tv_sec; 994 tvp->tv_usec += bt->tv_nsec / 1000; 995 while (tvp->tv_usec >= 1000000) { 996 tvp->tv_usec -= 1000000; 997 ++tvp->tv_sec; 998 } 999 } 1000 1001 void 1002 nanotime(struct timespec *tsp) 1003 { 1004 struct globaldata *gd = mycpu; 1005 struct timespec *bt; 1006 sysclock_t delta; 1007 1008 do { 1009 tsp->tv_sec = gd->gd_time_seconds; 1010 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1011 } while (tsp->tv_sec != gd->gd_time_seconds); 1012 1013 if (delta >= sys_cputimer->freq) { 1014 tsp->tv_sec += delta / sys_cputimer->freq; 1015 delta %= sys_cputimer->freq; 1016 } 1017 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1018 1019 bt = &basetime[basetime_index]; 1020 tsp->tv_sec += bt->tv_sec; 1021 tsp->tv_nsec += bt->tv_nsec; 1022 while (tsp->tv_nsec >= 1000000000) { 1023 tsp->tv_nsec -= 1000000000; 1024 ++tsp->tv_sec; 1025 } 1026 } 1027 1028 /* 1029 * note: this is not exactly synchronized with real time. To do that we 1030 * would have to do what microtime does and check for a nanoseconds overflow. 1031 */ 1032 time_t 1033 get_approximate_time_t(void) 1034 { 1035 struct globaldata *gd = mycpu; 1036 struct timespec *bt; 1037 1038 bt = &basetime[basetime_index]; 1039 return(gd->gd_time_seconds + bt->tv_sec); 1040 } 1041 1042 int 1043 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1044 { 1045 pps_params_t *app; 1046 struct pps_fetch_args *fapi; 1047 #ifdef PPS_SYNC 1048 struct pps_kcbind_args *kapi; 1049 #endif 1050 1051 switch (cmd) { 1052 case PPS_IOC_CREATE: 1053 return (0); 1054 case PPS_IOC_DESTROY: 1055 return (0); 1056 case PPS_IOC_SETPARAMS: 1057 app = (pps_params_t *)data; 1058 if (app->mode & ~pps->ppscap) 1059 return (EINVAL); 1060 pps->ppsparam = *app; 1061 return (0); 1062 case PPS_IOC_GETPARAMS: 1063 app = (pps_params_t *)data; 1064 *app = pps->ppsparam; 1065 app->api_version = PPS_API_VERS_1; 1066 return (0); 1067 case PPS_IOC_GETCAP: 1068 *(int*)data = pps->ppscap; 1069 return (0); 1070 case PPS_IOC_FETCH: 1071 fapi = (struct pps_fetch_args *)data; 1072 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1073 return (EINVAL); 1074 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1075 return (EOPNOTSUPP); 1076 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1077 fapi->pps_info_buf = pps->ppsinfo; 1078 return (0); 1079 case PPS_IOC_KCBIND: 1080 #ifdef PPS_SYNC 1081 kapi = (struct pps_kcbind_args *)data; 1082 /* XXX Only root should be able to do this */ 1083 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1084 return (EINVAL); 1085 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1086 return (EINVAL); 1087 if (kapi->edge & ~pps->ppscap) 1088 return (EINVAL); 1089 pps->kcmode = kapi->edge; 1090 return (0); 1091 #else 1092 return (EOPNOTSUPP); 1093 #endif 1094 default: 1095 return (ENOTTY); 1096 } 1097 } 1098 1099 void 1100 pps_init(struct pps_state *pps) 1101 { 1102 pps->ppscap |= PPS_TSFMT_TSPEC; 1103 if (pps->ppscap & PPS_CAPTUREASSERT) 1104 pps->ppscap |= PPS_OFFSETASSERT; 1105 if (pps->ppscap & PPS_CAPTURECLEAR) 1106 pps->ppscap |= PPS_OFFSETCLEAR; 1107 } 1108 1109 void 1110 pps_event(struct pps_state *pps, sysclock_t count, int event) 1111 { 1112 struct globaldata *gd; 1113 struct timespec *tsp; 1114 struct timespec *osp; 1115 struct timespec *bt; 1116 struct timespec ts; 1117 sysclock_t *pcount; 1118 #ifdef PPS_SYNC 1119 sysclock_t tcount; 1120 #endif 1121 sysclock_t delta; 1122 pps_seq_t *pseq; 1123 int foff; 1124 int fhard; 1125 1126 gd = mycpu; 1127 1128 /* Things would be easier with arrays... */ 1129 if (event == PPS_CAPTUREASSERT) { 1130 tsp = &pps->ppsinfo.assert_timestamp; 1131 osp = &pps->ppsparam.assert_offset; 1132 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1133 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1134 pcount = &pps->ppscount[0]; 1135 pseq = &pps->ppsinfo.assert_sequence; 1136 } else { 1137 tsp = &pps->ppsinfo.clear_timestamp; 1138 osp = &pps->ppsparam.clear_offset; 1139 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1140 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1141 pcount = &pps->ppscount[1]; 1142 pseq = &pps->ppsinfo.clear_sequence; 1143 } 1144 1145 /* Nothing really happened */ 1146 if (*pcount == count) 1147 return; 1148 1149 *pcount = count; 1150 1151 do { 1152 ts.tv_sec = gd->gd_time_seconds; 1153 delta = count - gd->gd_cpuclock_base; 1154 } while (ts.tv_sec != gd->gd_time_seconds); 1155 1156 if (delta >= sys_cputimer->freq) { 1157 ts.tv_sec += delta / sys_cputimer->freq; 1158 delta %= sys_cputimer->freq; 1159 } 1160 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1161 bt = &basetime[basetime_index]; 1162 ts.tv_sec += bt->tv_sec; 1163 ts.tv_nsec += bt->tv_nsec; 1164 while (ts.tv_nsec >= 1000000000) { 1165 ts.tv_nsec -= 1000000000; 1166 ++ts.tv_sec; 1167 } 1168 1169 (*pseq)++; 1170 *tsp = ts; 1171 1172 if (foff) { 1173 timespecadd(tsp, osp); 1174 if (tsp->tv_nsec < 0) { 1175 tsp->tv_nsec += 1000000000; 1176 tsp->tv_sec -= 1; 1177 } 1178 } 1179 #ifdef PPS_SYNC 1180 if (fhard) { 1181 /* magic, at its best... */ 1182 tcount = count - pps->ppscount[2]; 1183 pps->ppscount[2] = count; 1184 if (tcount >= sys_cputimer->freq) { 1185 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1186 sys_cputimer->freq64_nsec * 1187 (tcount % sys_cputimer->freq)) >> 32; 1188 } else { 1189 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1190 } 1191 hardpps(tsp, delta); 1192 } 1193 #endif 1194 } 1195 1196