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 1.50 2005/10/24 08:06:16 sephe Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 #include "opt_polling.h" 78 79 #include <sys/param.h> 80 #include <sys/systm.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 #endif 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 #if 0 367 if (tco->tc_poll_pps) 368 tco->tc_poll_pps(tco); 369 #endif 370 371 /* 372 * Calculate the new basetime index. We are in a critical section 373 * on cpu #0 and can safely play with basetime_index. Start 374 * with the current basetime and then make adjustments. 375 */ 376 ni = (basetime_index + 1) & BASETIME_ARYMASK; 377 nbt = &basetime[ni]; 378 *nbt = basetime[basetime_index]; 379 380 /* 381 * Apply adjtime corrections. (adjtime() API) 382 * 383 * adjtime() only runs on cpu #0 so our critical section is 384 * sufficient to access these variables. 385 */ 386 if (ntp_delta != 0) { 387 nbt->tv_nsec += ntp_tick_delta; 388 ntp_delta -= ntp_tick_delta; 389 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 390 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 391 ntp_tick_delta = ntp_delta; 392 } 393 } 394 395 /* 396 * Apply permanent frequency corrections. (sysctl API) 397 */ 398 if (ntp_tick_permanent != 0) { 399 ntp_tick_acc += ntp_tick_permanent; 400 if (ntp_tick_acc >= (1LL << 32)) { 401 nbt->tv_nsec += ntp_tick_acc >> 32; 402 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 403 } else if (ntp_tick_acc <= -(1LL << 32)) { 404 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 405 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 406 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 407 } 408 } 409 410 if (nbt->tv_nsec >= 1000000000) { 411 nbt->tv_sec++; 412 nbt->tv_nsec -= 1000000000; 413 } else if (nbt->tv_nsec < 0) { 414 nbt->tv_sec--; 415 nbt->tv_nsec += 1000000000; 416 } 417 418 /* 419 * Another per-tick compensation. (for ntp_adjtime() API) 420 */ 421 if (nsec_adj != 0) { 422 nsec_acc += nsec_adj; 423 if (nsec_acc >= 0x100000000LL) { 424 nbt->tv_nsec += nsec_acc >> 32; 425 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 426 } else if (nsec_acc <= -0x100000000LL) { 427 nbt->tv_nsec -= -nsec_acc >> 32; 428 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 429 } 430 if (nbt->tv_nsec >= 1000000000) { 431 nbt->tv_nsec -= 1000000000; 432 ++nbt->tv_sec; 433 } else if (nbt->tv_nsec < 0) { 434 nbt->tv_nsec += 1000000000; 435 --nbt->tv_sec; 436 } 437 } 438 439 /************************************************************ 440 * LEAP SECOND CORRECTION * 441 ************************************************************ 442 * 443 * Taking into account all the corrections made above, figure 444 * out the new real time. If the seconds field has changed 445 * then apply any pending leap-second corrections. 446 */ 447 getnanotime_nbt(nbt, &nts); 448 449 if (time_second != nts.tv_sec) { 450 /* 451 * Apply leap second (sysctl API). Adjust nts for changes 452 * so we do not have to call getnanotime_nbt again. 453 */ 454 if (ntp_leap_second) { 455 if (ntp_leap_second == nts.tv_sec) { 456 if (ntp_leap_insert) { 457 nbt->tv_sec++; 458 nts.tv_sec++; 459 } else { 460 nbt->tv_sec--; 461 nts.tv_sec--; 462 } 463 ntp_leap_second--; 464 } 465 } 466 467 /* 468 * Apply leap second (ntp_adjtime() API), calculate a new 469 * nsec_adj field. ntp_update_second() returns nsec_adj 470 * as a per-second value but we need it as a per-tick value. 471 */ 472 leap = ntp_update_second(time_second, &nsec_adj); 473 nsec_adj /= hz; 474 nbt->tv_sec += leap; 475 nts.tv_sec += leap; 476 477 /* 478 * Update the time_second 'approximate time' global. 479 */ 480 time_second = nts.tv_sec; 481 } 482 483 /* 484 * Finally, our new basetime is ready to go live! 485 */ 486 cpu_sfence(); 487 basetime_index = ni; 488 } 489 490 /* 491 * softticks are handled for all cpus 492 */ 493 hardclock_softtick(gd); 494 495 /* 496 * ITimer handling is per-tick, per-cpu. I don't think psignal() 497 * is mpsafe on curproc, so XXX get the mplock. 498 */ 499 if ((p = curproc) != NULL && try_mplock()) { 500 pstats = p->p_stats; 501 if (frame && CLKF_USERMODE(frame) && 502 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 503 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 504 psignal(p, SIGVTALRM); 505 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 506 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 507 psignal(p, SIGPROF); 508 rel_mplock(); 509 } 510 setdelayed(); 511 } 512 513 /* 514 * The statistics clock typically runs at a 125Hz rate, and is intended 515 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 516 * 517 * NOTE! systimer! the MP lock might not be held here. We can only safely 518 * manipulate objects owned by the current cpu. 519 * 520 * The stats clock is responsible for grabbing a profiling sample. 521 * Most of the statistics are only used by user-level statistics programs. 522 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 523 * p->p_estcpu. 524 * 525 * Like the other clocks, the stat clock is called from what is effectively 526 * a fast interrupt, so the context should be the thread/process that got 527 * interrupted. 528 */ 529 static void 530 statclock(systimer_t info, struct intrframe *frame) 531 { 532 #ifdef GPROF 533 struct gmonparam *g; 534 int i; 535 #endif 536 thread_t td; 537 struct proc *p; 538 int bump; 539 struct timeval tv; 540 struct timeval *stv; 541 542 /* 543 * How big was our timeslice relative to the last time? 544 */ 545 microuptime(&tv); /* mpsafe */ 546 stv = &mycpu->gd_stattv; 547 if (stv->tv_sec == 0) { 548 bump = 1; 549 } else { 550 bump = tv.tv_usec - stv->tv_usec + 551 (tv.tv_sec - stv->tv_sec) * 1000000; 552 if (bump < 0) 553 bump = 0; 554 if (bump > 1000000) 555 bump = 1000000; 556 } 557 *stv = tv; 558 559 td = curthread; 560 p = td->td_proc; 561 562 if (frame && CLKF_USERMODE(frame)) { 563 /* 564 * Came from userland, handle user time and deal with 565 * possible process. 566 */ 567 if (p && (p->p_flag & P_PROFIL)) 568 addupc_intr(p, CLKF_PC(frame), 1); 569 td->td_uticks += bump; 570 571 /* 572 * Charge the time as appropriate 573 */ 574 if (p && p->p_nice > NZERO) 575 cpu_time.cp_nice += bump; 576 else 577 cpu_time.cp_user += bump; 578 } else { 579 #ifdef GPROF 580 /* 581 * Kernel statistics are just like addupc_intr, only easier. 582 */ 583 g = &_gmonparam; 584 if (g->state == GMON_PROF_ON && frame) { 585 i = CLKF_PC(frame) - g->lowpc; 586 if (i < g->textsize) { 587 i /= HISTFRACTION * sizeof(*g->kcount); 588 g->kcount[i]++; 589 } 590 } 591 #endif 592 /* 593 * Came from kernel mode, so we were: 594 * - handling an interrupt, 595 * - doing syscall or trap work on behalf of the current 596 * user process, or 597 * - spinning in the idle loop. 598 * Whichever it is, charge the time as appropriate. 599 * Note that we charge interrupts to the current process, 600 * regardless of whether they are ``for'' that process, 601 * so that we know how much of its real time was spent 602 * in ``non-process'' (i.e., interrupt) work. 603 * 604 * XXX assume system if frame is NULL. A NULL frame 605 * can occur if ipi processing is done from a crit_exit(). 606 */ 607 if (frame && CLKF_INTR(frame)) 608 td->td_iticks += bump; 609 else 610 td->td_sticks += bump; 611 612 if (frame && CLKF_INTR(frame)) { 613 cpu_time.cp_intr += bump; 614 } else { 615 if (td == &mycpu->gd_idlethread) 616 cpu_time.cp_idle += bump; 617 else 618 cpu_time.cp_sys += bump; 619 } 620 } 621 } 622 623 /* 624 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 625 * the MP lock might not be held. We can safely manipulate parts of curproc 626 * but that's about it. 627 * 628 * Each cpu has its own scheduler clock. 629 */ 630 static void 631 schedclock(systimer_t info, struct intrframe *frame) 632 { 633 struct lwp *lp; 634 struct pstats *pstats; 635 struct rusage *ru; 636 struct vmspace *vm; 637 long rss; 638 639 if ((lp = lwkt_preempted_proc()) != NULL) { 640 /* 641 * Account for cpu time used and hit the scheduler. Note 642 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 643 * HERE. 644 */ 645 ++lp->lwp_cpticks; 646 /* 647 * XXX I think accessing lwp_proc's p_usched is 648 * reasonably MP safe. This needs to be revisited 649 * when we have pluggable schedulers. 650 */ 651 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, info->time); 652 } 653 if ((lp = curthread->td_lwp) != NULL) { 654 /* 655 * Update resource usage integrals and maximums. 656 */ 657 if ((pstats = lp->lwp_stats) != NULL && 658 (ru = &pstats->p_ru) != NULL && 659 (vm = lp->lwp_proc->p_vmspace) != NULL) { 660 ru->ru_ixrss += pgtok(vm->vm_tsize); 661 ru->ru_idrss += pgtok(vm->vm_dsize); 662 ru->ru_isrss += pgtok(vm->vm_ssize); 663 rss = pgtok(vmspace_resident_count(vm)); 664 if (ru->ru_maxrss < rss) 665 ru->ru_maxrss = rss; 666 } 667 } 668 } 669 670 /* 671 * Compute number of ticks for the specified amount of time. The 672 * return value is intended to be used in a clock interrupt timed 673 * operation and guarenteed to meet or exceed the requested time. 674 * If the representation overflows, return INT_MAX. The minimum return 675 * value is 1 ticks and the function will average the calculation up. 676 * If any value greater then 0 microseconds is supplied, a value 677 * of at least 2 will be returned to ensure that a near-term clock 678 * interrupt does not cause the timeout to occur (degenerately) early. 679 * 680 * Note that limit checks must take into account microseconds, which is 681 * done simply by using the smaller signed long maximum instead of 682 * the unsigned long maximum. 683 * 684 * If ints have 32 bits, then the maximum value for any timeout in 685 * 10ms ticks is 248 days. 686 */ 687 int 688 tvtohz_high(struct timeval *tv) 689 { 690 int ticks; 691 long sec, usec; 692 693 sec = tv->tv_sec; 694 usec = tv->tv_usec; 695 if (usec < 0) { 696 sec--; 697 usec += 1000000; 698 } 699 if (sec < 0) { 700 #ifdef DIAGNOSTIC 701 if (usec > 0) { 702 sec++; 703 usec -= 1000000; 704 } 705 printf("tvotohz: negative time difference %ld sec %ld usec\n", 706 sec, usec); 707 #endif 708 ticks = 1; 709 } else if (sec <= INT_MAX / hz) { 710 ticks = (int)(sec * hz + 711 ((u_long)usec + (tick - 1)) / tick) + 1; 712 } else { 713 ticks = INT_MAX; 714 } 715 return (ticks); 716 } 717 718 /* 719 * Compute number of ticks for the specified amount of time, erroring on 720 * the side of it being too low to ensure that sleeping the returned number 721 * of ticks will not result in a late return. 722 * 723 * The supplied timeval may not be negative and should be normalized. A 724 * return value of 0 is possible if the timeval converts to less then 725 * 1 tick. 726 * 727 * If ints have 32 bits, then the maximum value for any timeout in 728 * 10ms ticks is 248 days. 729 */ 730 int 731 tvtohz_low(struct timeval *tv) 732 { 733 int ticks; 734 long sec; 735 736 sec = tv->tv_sec; 737 if (sec <= INT_MAX / hz) 738 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 739 else 740 ticks = INT_MAX; 741 return (ticks); 742 } 743 744 745 /* 746 * Start profiling on a process. 747 * 748 * Kernel profiling passes proc0 which never exits and hence 749 * keeps the profile clock running constantly. 750 */ 751 void 752 startprofclock(struct proc *p) 753 { 754 if ((p->p_flag & P_PROFIL) == 0) { 755 p->p_flag |= P_PROFIL; 756 #if 0 /* XXX */ 757 if (++profprocs == 1 && stathz != 0) { 758 crit_enter(); 759 psdiv = psratio; 760 setstatclockrate(profhz); 761 crit_exit(); 762 } 763 #endif 764 } 765 } 766 767 /* 768 * Stop profiling on a process. 769 */ 770 void 771 stopprofclock(struct proc *p) 772 { 773 if (p->p_flag & P_PROFIL) { 774 p->p_flag &= ~P_PROFIL; 775 #if 0 /* XXX */ 776 if (--profprocs == 0 && stathz != 0) { 777 crit_enter(); 778 psdiv = 1; 779 setstatclockrate(stathz); 780 crit_exit(); 781 } 782 #endif 783 } 784 } 785 786 /* 787 * Return information about system clocks. 788 */ 789 static int 790 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 791 { 792 struct kinfo_clockinfo clkinfo; 793 /* 794 * Construct clockinfo structure. 795 */ 796 clkinfo.ci_hz = hz; 797 clkinfo.ci_tick = tick; 798 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 799 clkinfo.ci_profhz = profhz; 800 clkinfo.ci_stathz = stathz ? stathz : hz; 801 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 802 } 803 804 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 805 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 806 807 /* 808 * We have eight functions for looking at the clock, four for 809 * microseconds and four for nanoseconds. For each there is fast 810 * but less precise version "get{nano|micro}[up]time" which will 811 * return a time which is up to 1/HZ previous to the call, whereas 812 * the raw version "{nano|micro}[up]time" will return a timestamp 813 * which is as precise as possible. The "up" variants return the 814 * time relative to system boot, these are well suited for time 815 * interval measurements. 816 * 817 * Each cpu independantly maintains the current time of day, so all 818 * we need to do to protect ourselves from changes is to do a loop 819 * check on the seconds field changing out from under us. 820 * 821 * The system timer maintains a 32 bit count and due to various issues 822 * it is possible for the calculated delta to occassionally exceed 823 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 824 * multiplication can easily overflow, so we deal with the case. For 825 * uniformity we deal with the case in the usec case too. 826 */ 827 void 828 getmicrouptime(struct timeval *tvp) 829 { 830 struct globaldata *gd = mycpu; 831 sysclock_t delta; 832 833 do { 834 tvp->tv_sec = gd->gd_time_seconds; 835 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 836 } while (tvp->tv_sec != gd->gd_time_seconds); 837 838 if (delta >= sys_cputimer->freq) { 839 tvp->tv_sec += delta / sys_cputimer->freq; 840 delta %= sys_cputimer->freq; 841 } 842 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 843 if (tvp->tv_usec >= 1000000) { 844 tvp->tv_usec -= 1000000; 845 ++tvp->tv_sec; 846 } 847 } 848 849 void 850 getnanouptime(struct timespec *tsp) 851 { 852 struct globaldata *gd = mycpu; 853 sysclock_t delta; 854 855 do { 856 tsp->tv_sec = gd->gd_time_seconds; 857 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 858 } while (tsp->tv_sec != gd->gd_time_seconds); 859 860 if (delta >= sys_cputimer->freq) { 861 tsp->tv_sec += delta / sys_cputimer->freq; 862 delta %= sys_cputimer->freq; 863 } 864 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 865 } 866 867 void 868 microuptime(struct timeval *tvp) 869 { 870 struct globaldata *gd = mycpu; 871 sysclock_t delta; 872 873 do { 874 tvp->tv_sec = gd->gd_time_seconds; 875 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 876 } while (tvp->tv_sec != gd->gd_time_seconds); 877 878 if (delta >= sys_cputimer->freq) { 879 tvp->tv_sec += delta / sys_cputimer->freq; 880 delta %= sys_cputimer->freq; 881 } 882 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 883 } 884 885 void 886 nanouptime(struct timespec *tsp) 887 { 888 struct globaldata *gd = mycpu; 889 sysclock_t delta; 890 891 do { 892 tsp->tv_sec = gd->gd_time_seconds; 893 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 894 } while (tsp->tv_sec != gd->gd_time_seconds); 895 896 if (delta >= sys_cputimer->freq) { 897 tsp->tv_sec += delta / sys_cputimer->freq; 898 delta %= sys_cputimer->freq; 899 } 900 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 901 } 902 903 /* 904 * realtime routines 905 */ 906 907 void 908 getmicrotime(struct timeval *tvp) 909 { 910 struct globaldata *gd = mycpu; 911 struct timespec *bt; 912 sysclock_t delta; 913 914 do { 915 tvp->tv_sec = gd->gd_time_seconds; 916 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 917 } while (tvp->tv_sec != gd->gd_time_seconds); 918 919 if (delta >= sys_cputimer->freq) { 920 tvp->tv_sec += delta / sys_cputimer->freq; 921 delta %= sys_cputimer->freq; 922 } 923 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 924 925 bt = &basetime[basetime_index]; 926 tvp->tv_sec += bt->tv_sec; 927 tvp->tv_usec += bt->tv_nsec / 1000; 928 while (tvp->tv_usec >= 1000000) { 929 tvp->tv_usec -= 1000000; 930 ++tvp->tv_sec; 931 } 932 } 933 934 void 935 getnanotime(struct timespec *tsp) 936 { 937 struct globaldata *gd = mycpu; 938 struct timespec *bt; 939 sysclock_t delta; 940 941 do { 942 tsp->tv_sec = gd->gd_time_seconds; 943 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 944 } while (tsp->tv_sec != gd->gd_time_seconds); 945 946 if (delta >= sys_cputimer->freq) { 947 tsp->tv_sec += delta / sys_cputimer->freq; 948 delta %= sys_cputimer->freq; 949 } 950 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 951 952 bt = &basetime[basetime_index]; 953 tsp->tv_sec += bt->tv_sec; 954 tsp->tv_nsec += bt->tv_nsec; 955 while (tsp->tv_nsec >= 1000000000) { 956 tsp->tv_nsec -= 1000000000; 957 ++tsp->tv_sec; 958 } 959 } 960 961 static void 962 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 963 { 964 struct globaldata *gd = mycpu; 965 sysclock_t delta; 966 967 do { 968 tsp->tv_sec = gd->gd_time_seconds; 969 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 970 } while (tsp->tv_sec != gd->gd_time_seconds); 971 972 if (delta >= sys_cputimer->freq) { 973 tsp->tv_sec += delta / sys_cputimer->freq; 974 delta %= sys_cputimer->freq; 975 } 976 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 977 978 tsp->tv_sec += nbt->tv_sec; 979 tsp->tv_nsec += nbt->tv_nsec; 980 while (tsp->tv_nsec >= 1000000000) { 981 tsp->tv_nsec -= 1000000000; 982 ++tsp->tv_sec; 983 } 984 } 985 986 987 void 988 microtime(struct timeval *tvp) 989 { 990 struct globaldata *gd = mycpu; 991 struct timespec *bt; 992 sysclock_t delta; 993 994 do { 995 tvp->tv_sec = gd->gd_time_seconds; 996 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 997 } while (tvp->tv_sec != gd->gd_time_seconds); 998 999 if (delta >= sys_cputimer->freq) { 1000 tvp->tv_sec += delta / sys_cputimer->freq; 1001 delta %= sys_cputimer->freq; 1002 } 1003 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1004 1005 bt = &basetime[basetime_index]; 1006 tvp->tv_sec += bt->tv_sec; 1007 tvp->tv_usec += bt->tv_nsec / 1000; 1008 while (tvp->tv_usec >= 1000000) { 1009 tvp->tv_usec -= 1000000; 1010 ++tvp->tv_sec; 1011 } 1012 } 1013 1014 void 1015 nanotime(struct timespec *tsp) 1016 { 1017 struct globaldata *gd = mycpu; 1018 struct timespec *bt; 1019 sysclock_t delta; 1020 1021 do { 1022 tsp->tv_sec = gd->gd_time_seconds; 1023 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1024 } while (tsp->tv_sec != gd->gd_time_seconds); 1025 1026 if (delta >= sys_cputimer->freq) { 1027 tsp->tv_sec += delta / sys_cputimer->freq; 1028 delta %= sys_cputimer->freq; 1029 } 1030 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1031 1032 bt = &basetime[basetime_index]; 1033 tsp->tv_sec += bt->tv_sec; 1034 tsp->tv_nsec += bt->tv_nsec; 1035 while (tsp->tv_nsec >= 1000000000) { 1036 tsp->tv_nsec -= 1000000000; 1037 ++tsp->tv_sec; 1038 } 1039 } 1040 1041 /* 1042 * note: this is not exactly synchronized with real time. To do that we 1043 * would have to do what microtime does and check for a nanoseconds overflow. 1044 */ 1045 time_t 1046 get_approximate_time_t(void) 1047 { 1048 struct globaldata *gd = mycpu; 1049 struct timespec *bt; 1050 1051 bt = &basetime[basetime_index]; 1052 return(gd->gd_time_seconds + bt->tv_sec); 1053 } 1054 1055 int 1056 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1057 { 1058 pps_params_t *app; 1059 struct pps_fetch_args *fapi; 1060 #ifdef PPS_SYNC 1061 struct pps_kcbind_args *kapi; 1062 #endif 1063 1064 switch (cmd) { 1065 case PPS_IOC_CREATE: 1066 return (0); 1067 case PPS_IOC_DESTROY: 1068 return (0); 1069 case PPS_IOC_SETPARAMS: 1070 app = (pps_params_t *)data; 1071 if (app->mode & ~pps->ppscap) 1072 return (EINVAL); 1073 pps->ppsparam = *app; 1074 return (0); 1075 case PPS_IOC_GETPARAMS: 1076 app = (pps_params_t *)data; 1077 *app = pps->ppsparam; 1078 app->api_version = PPS_API_VERS_1; 1079 return (0); 1080 case PPS_IOC_GETCAP: 1081 *(int*)data = pps->ppscap; 1082 return (0); 1083 case PPS_IOC_FETCH: 1084 fapi = (struct pps_fetch_args *)data; 1085 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1086 return (EINVAL); 1087 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1088 return (EOPNOTSUPP); 1089 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1090 fapi->pps_info_buf = pps->ppsinfo; 1091 return (0); 1092 case PPS_IOC_KCBIND: 1093 #ifdef PPS_SYNC 1094 kapi = (struct pps_kcbind_args *)data; 1095 /* XXX Only root should be able to do this */ 1096 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1097 return (EINVAL); 1098 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1099 return (EINVAL); 1100 if (kapi->edge & ~pps->ppscap) 1101 return (EINVAL); 1102 pps->kcmode = kapi->edge; 1103 return (0); 1104 #else 1105 return (EOPNOTSUPP); 1106 #endif 1107 default: 1108 return (ENOTTY); 1109 } 1110 } 1111 1112 void 1113 pps_init(struct pps_state *pps) 1114 { 1115 pps->ppscap |= PPS_TSFMT_TSPEC; 1116 if (pps->ppscap & PPS_CAPTUREASSERT) 1117 pps->ppscap |= PPS_OFFSETASSERT; 1118 if (pps->ppscap & PPS_CAPTURECLEAR) 1119 pps->ppscap |= PPS_OFFSETCLEAR; 1120 } 1121 1122 void 1123 pps_event(struct pps_state *pps, sysclock_t count, int event) 1124 { 1125 struct globaldata *gd; 1126 struct timespec *tsp; 1127 struct timespec *osp; 1128 struct timespec *bt; 1129 struct timespec ts; 1130 sysclock_t *pcount; 1131 #ifdef PPS_SYNC 1132 sysclock_t tcount; 1133 #endif 1134 sysclock_t delta; 1135 pps_seq_t *pseq; 1136 int foff; 1137 int fhard; 1138 1139 gd = mycpu; 1140 1141 /* Things would be easier with arrays... */ 1142 if (event == PPS_CAPTUREASSERT) { 1143 tsp = &pps->ppsinfo.assert_timestamp; 1144 osp = &pps->ppsparam.assert_offset; 1145 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1146 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1147 pcount = &pps->ppscount[0]; 1148 pseq = &pps->ppsinfo.assert_sequence; 1149 } else { 1150 tsp = &pps->ppsinfo.clear_timestamp; 1151 osp = &pps->ppsparam.clear_offset; 1152 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1153 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1154 pcount = &pps->ppscount[1]; 1155 pseq = &pps->ppsinfo.clear_sequence; 1156 } 1157 1158 /* Nothing really happened */ 1159 if (*pcount == count) 1160 return; 1161 1162 *pcount = count; 1163 1164 do { 1165 ts.tv_sec = gd->gd_time_seconds; 1166 delta = count - gd->gd_cpuclock_base; 1167 } while (ts.tv_sec != gd->gd_time_seconds); 1168 1169 if (delta >= sys_cputimer->freq) { 1170 ts.tv_sec += delta / sys_cputimer->freq; 1171 delta %= sys_cputimer->freq; 1172 } 1173 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1174 bt = &basetime[basetime_index]; 1175 ts.tv_sec += bt->tv_sec; 1176 ts.tv_nsec += bt->tv_nsec; 1177 while (ts.tv_nsec >= 1000000000) { 1178 ts.tv_nsec -= 1000000000; 1179 ++ts.tv_sec; 1180 } 1181 1182 (*pseq)++; 1183 *tsp = ts; 1184 1185 if (foff) { 1186 timespecadd(tsp, osp); 1187 if (tsp->tv_nsec < 0) { 1188 tsp->tv_nsec += 1000000000; 1189 tsp->tv_sec -= 1; 1190 } 1191 } 1192 #ifdef PPS_SYNC 1193 if (fhard) { 1194 /* magic, at its best... */ 1195 tcount = count - pps->ppscount[2]; 1196 pps->ppscount[2] = count; 1197 if (tcount >= sys_cputimer->freq) { 1198 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1199 sys_cputimer->freq64_nsec * 1200 (tcount % sys_cputimer->freq)) >> 32; 1201 } else { 1202 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1203 } 1204 hardpps(tsp, delta); 1205 } 1206 #endif 1207 } 1208 1209