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 1.62 2008/09/09 04:06:13 dillon Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 #include "opt_polling.h" 78 #include "opt_ifpoll.h" 79 #include "opt_pctrack.h" 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/callout.h> 84 #include <sys/kernel.h> 85 #include <sys/kinfo.h> 86 #include <sys/proc.h> 87 #include <sys/malloc.h> 88 #include <sys/resourcevar.h> 89 #include <sys/signalvar.h> 90 #include <sys/timex.h> 91 #include <sys/timepps.h> 92 #include <vm/vm.h> 93 #include <sys/lock.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_extern.h> 97 #include <sys/sysctl.h> 98 #include <sys/thread2.h> 99 100 #include <machine/cpu.h> 101 #include <machine/limits.h> 102 #include <machine/smp.h> 103 104 #ifdef GPROF 105 #include <sys/gmon.h> 106 #endif 107 108 #ifdef DEVICE_POLLING 109 extern void init_device_poll_pcpu(int); 110 #endif 111 112 #ifdef IFPOLL_ENABLE 113 extern void ifpoll_init_pcpu(int); 114 #endif 115 116 #ifdef DEBUG_PCTRACK 117 static void do_pctrack(struct intrframe *frame, int which); 118 #endif 119 120 static void initclocks (void *dummy); 121 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 122 123 /* 124 * Some of these don't belong here, but it's easiest to concentrate them. 125 * Note that cpu_time counts in microseconds, but most userland programs 126 * just compare relative times against the total by delta. 127 */ 128 struct kinfo_cputime cputime_percpu[MAXCPU]; 129 #ifdef DEBUG_PCTRACK 130 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE }; 131 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE]; 132 #endif 133 134 #ifdef SMP 135 static int 136 sysctl_cputime(SYSCTL_HANDLER_ARGS) 137 { 138 int cpu, error = 0; 139 size_t size = sizeof(struct kinfo_cputime); 140 141 for (cpu = 0; cpu < ncpus; ++cpu) { 142 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 143 break; 144 } 145 146 return (error); 147 } 148 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 149 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 150 #else 151 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 152 "CPU time statistics"); 153 #endif 154 155 /* 156 * boottime is used to calculate the 'real' uptime. Do not confuse this with 157 * microuptime(). microtime() is not drift compensated. The real uptime 158 * with compensation is nanotime() - bootime. boottime is recalculated 159 * whenever the real time is set based on the compensated elapsed time 160 * in seconds (gd->gd_time_seconds). 161 * 162 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 163 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 164 * the real time. 165 */ 166 struct timespec boottime; /* boot time (realtime) for reference only */ 167 time_t time_second; /* read-only 'passive' uptime in seconds */ 168 169 /* 170 * basetime is used to calculate the compensated real time of day. The 171 * basetime can be modified on a per-tick basis by the adjtime(), 172 * ntp_adjtime(), and sysctl-based time correction APIs. 173 * 174 * Note that frequency corrections can also be made by adjusting 175 * gd_cpuclock_base. 176 * 177 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 178 * used on both SMP and UP systems to avoid MP races between cpu's and 179 * interrupt races on UP systems. 180 */ 181 #define BASETIME_ARYSIZE 16 182 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 183 static struct timespec basetime[BASETIME_ARYSIZE]; 184 static volatile int basetime_index; 185 186 static int 187 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 188 { 189 struct timespec *bt; 190 int error; 191 int index; 192 193 /* 194 * Because basetime data and index may be updated by another cpu, 195 * a load fence is required to ensure that the data we read has 196 * not been speculatively read relative to a possibly updated index. 197 */ 198 index = basetime_index; 199 cpu_lfence(); 200 bt = &basetime[index]; 201 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 202 return (error); 203 } 204 205 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 206 &boottime, timespec, "System boottime"); 207 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 208 sysctl_get_basetime, "S,timespec", "System basetime"); 209 210 static void hardclock(systimer_t info, struct intrframe *frame); 211 static void statclock(systimer_t info, struct intrframe *frame); 212 static void schedclock(systimer_t info, struct intrframe *frame); 213 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 214 215 int ticks; /* system master ticks at hz */ 216 int clocks_running; /* tsleep/timeout clocks operational */ 217 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 218 int64_t nsec_acc; /* accumulator */ 219 220 /* NTPD time correction fields */ 221 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 222 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 223 int64_t ntp_delta; /* one-time correction in nsec */ 224 int64_t ntp_big_delta = 1000000000; 225 int32_t ntp_tick_delta; /* current adjustment rate */ 226 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 227 time_t ntp_leap_second; /* time of next leap second */ 228 int ntp_leap_insert; /* whether to insert or remove a second */ 229 230 /* 231 * Finish initializing clock frequencies and start all clocks running. 232 */ 233 /* ARGSUSED*/ 234 static void 235 initclocks(void *dummy) 236 { 237 /*psratio = profhz / stathz;*/ 238 initclocks_pcpu(); 239 clocks_running = 1; 240 } 241 242 /* 243 * Called on a per-cpu basis 244 */ 245 void 246 initclocks_pcpu(void) 247 { 248 struct globaldata *gd = mycpu; 249 250 crit_enter(); 251 if (gd->gd_cpuid == 0) { 252 gd->gd_time_seconds = 1; 253 gd->gd_cpuclock_base = sys_cputimer->count(); 254 } else { 255 /* XXX */ 256 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 257 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 258 } 259 260 #ifdef DEVICE_POLLING 261 init_device_poll_pcpu(gd->gd_cpuid); 262 #endif 263 264 #ifdef IFPOLL_ENABLE 265 ifpoll_init_pcpu(gd->gd_cpuid); 266 #endif 267 268 /* 269 * Use a non-queued periodic systimer to prevent multiple ticks from 270 * building up if the sysclock jumps forward (8254 gets reset). The 271 * sysclock will never jump backwards. Our time sync is based on 272 * the actual sysclock, not the ticks count. 273 */ 274 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 275 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 276 /* XXX correct the frequency for scheduler / estcpu tests */ 277 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 278 NULL, ESTCPUFREQ); 279 crit_exit(); 280 } 281 282 /* 283 * This sets the current real time of day. Timespecs are in seconds and 284 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 285 * instead we adjust basetime so basetime + gd_* results in the current 286 * time of day. This way the gd_* fields are guarenteed to represent 287 * a monotonically increasing 'uptime' value. 288 * 289 * When set_timeofday() is called from userland, the system call forces it 290 * onto cpu #0 since only cpu #0 can update basetime_index. 291 */ 292 void 293 set_timeofday(struct timespec *ts) 294 { 295 struct timespec *nbt; 296 int ni; 297 298 /* 299 * XXX SMP / non-atomic basetime updates 300 */ 301 crit_enter(); 302 ni = (basetime_index + 1) & BASETIME_ARYMASK; 303 nbt = &basetime[ni]; 304 nanouptime(nbt); 305 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 306 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 307 if (nbt->tv_nsec < 0) { 308 nbt->tv_nsec += 1000000000; 309 --nbt->tv_sec; 310 } 311 312 /* 313 * Note that basetime diverges from boottime as the clock drift is 314 * compensated for, so we cannot do away with boottime. When setting 315 * the absolute time of day the drift is 0 (for an instant) and we 316 * can simply assign boottime to basetime. 317 * 318 * Note that nanouptime() is based on gd_time_seconds which is drift 319 * compensated up to a point (it is guarenteed to remain monotonically 320 * increasing). gd_time_seconds is thus our best uptime guess and 321 * suitable for use in the boottime calculation. It is already taken 322 * into account in the basetime calculation above. 323 */ 324 boottime.tv_sec = nbt->tv_sec; 325 ntp_delta = 0; 326 327 /* 328 * We now have a new basetime, make sure all other cpus have it, 329 * then update the index. 330 */ 331 cpu_sfence(); 332 basetime_index = ni; 333 334 crit_exit(); 335 } 336 337 /* 338 * Each cpu has its own hardclock, but we only increments ticks and softticks 339 * on cpu #0. 340 * 341 * NOTE! systimer! the MP lock might not be held here. We can only safely 342 * manipulate objects owned by the current cpu. 343 */ 344 static void 345 hardclock(systimer_t info, struct intrframe *frame) 346 { 347 sysclock_t cputicks; 348 struct proc *p; 349 struct globaldata *gd = mycpu; 350 351 /* 352 * Realtime updates are per-cpu. Note that timer corrections as 353 * returned by microtime() and friends make an additional adjustment 354 * using a system-wise 'basetime', but the running time is always 355 * taken from the per-cpu globaldata area. Since the same clock 356 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 357 * stay in synch. 358 * 359 * Note that we never allow info->time (aka gd->gd_hardclock.time) 360 * to reverse index gd_cpuclock_base, but that it is possible for 361 * it to temporarily get behind in the seconds if something in the 362 * system locks interrupts for a long period of time. Since periodic 363 * timers count events, though everything should resynch again 364 * immediately. 365 */ 366 cputicks = info->time - gd->gd_cpuclock_base; 367 if (cputicks >= sys_cputimer->freq) { 368 ++gd->gd_time_seconds; 369 gd->gd_cpuclock_base += sys_cputimer->freq; 370 } 371 372 /* 373 * The system-wide ticks counter and NTP related timedelta/tickdelta 374 * adjustments only occur on cpu #0. NTP adjustments are accomplished 375 * by updating basetime. 376 */ 377 if (gd->gd_cpuid == 0) { 378 struct timespec *nbt; 379 struct timespec nts; 380 int leap; 381 int ni; 382 383 ++ticks; 384 385 #if 0 386 if (tco->tc_poll_pps) 387 tco->tc_poll_pps(tco); 388 #endif 389 390 /* 391 * Calculate the new basetime index. We are in a critical section 392 * on cpu #0 and can safely play with basetime_index. Start 393 * with the current basetime and then make adjustments. 394 */ 395 ni = (basetime_index + 1) & BASETIME_ARYMASK; 396 nbt = &basetime[ni]; 397 *nbt = basetime[basetime_index]; 398 399 /* 400 * Apply adjtime corrections. (adjtime() API) 401 * 402 * adjtime() only runs on cpu #0 so our critical section is 403 * sufficient to access these variables. 404 */ 405 if (ntp_delta != 0) { 406 nbt->tv_nsec += ntp_tick_delta; 407 ntp_delta -= ntp_tick_delta; 408 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 409 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 410 ntp_tick_delta = ntp_delta; 411 } 412 } 413 414 /* 415 * Apply permanent frequency corrections. (sysctl API) 416 */ 417 if (ntp_tick_permanent != 0) { 418 ntp_tick_acc += ntp_tick_permanent; 419 if (ntp_tick_acc >= (1LL << 32)) { 420 nbt->tv_nsec += ntp_tick_acc >> 32; 421 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 422 } else if (ntp_tick_acc <= -(1LL << 32)) { 423 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 424 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 425 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 426 } 427 } 428 429 if (nbt->tv_nsec >= 1000000000) { 430 nbt->tv_sec++; 431 nbt->tv_nsec -= 1000000000; 432 } else if (nbt->tv_nsec < 0) { 433 nbt->tv_sec--; 434 nbt->tv_nsec += 1000000000; 435 } 436 437 /* 438 * Another per-tick compensation. (for ntp_adjtime() API) 439 */ 440 if (nsec_adj != 0) { 441 nsec_acc += nsec_adj; 442 if (nsec_acc >= 0x100000000LL) { 443 nbt->tv_nsec += nsec_acc >> 32; 444 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 445 } else if (nsec_acc <= -0x100000000LL) { 446 nbt->tv_nsec -= -nsec_acc >> 32; 447 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 448 } 449 if (nbt->tv_nsec >= 1000000000) { 450 nbt->tv_nsec -= 1000000000; 451 ++nbt->tv_sec; 452 } else if (nbt->tv_nsec < 0) { 453 nbt->tv_nsec += 1000000000; 454 --nbt->tv_sec; 455 } 456 } 457 458 /************************************************************ 459 * LEAP SECOND CORRECTION * 460 ************************************************************ 461 * 462 * Taking into account all the corrections made above, figure 463 * out the new real time. If the seconds field has changed 464 * then apply any pending leap-second corrections. 465 */ 466 getnanotime_nbt(nbt, &nts); 467 468 if (time_second != nts.tv_sec) { 469 /* 470 * Apply leap second (sysctl API). Adjust nts for changes 471 * so we do not have to call getnanotime_nbt again. 472 */ 473 if (ntp_leap_second) { 474 if (ntp_leap_second == nts.tv_sec) { 475 if (ntp_leap_insert) { 476 nbt->tv_sec++; 477 nts.tv_sec++; 478 } else { 479 nbt->tv_sec--; 480 nts.tv_sec--; 481 } 482 ntp_leap_second--; 483 } 484 } 485 486 /* 487 * Apply leap second (ntp_adjtime() API), calculate a new 488 * nsec_adj field. ntp_update_second() returns nsec_adj 489 * as a per-second value but we need it as a per-tick value. 490 */ 491 leap = ntp_update_second(time_second, &nsec_adj); 492 nsec_adj /= hz; 493 nbt->tv_sec += leap; 494 nts.tv_sec += leap; 495 496 /* 497 * Update the time_second 'approximate time' global. 498 */ 499 time_second = nts.tv_sec; 500 } 501 502 /* 503 * Finally, our new basetime is ready to go live! 504 */ 505 cpu_sfence(); 506 basetime_index = ni; 507 508 /* 509 * Figure out how badly the system is starved for memory 510 */ 511 vm_fault_ratecheck(); 512 } 513 514 /* 515 * softticks are handled for all cpus 516 */ 517 hardclock_softtick(gd); 518 519 /* 520 * The LWKT scheduler will generally allow the current process to 521 * return to user mode even if there are other runnable LWKT threads 522 * running in kernel mode on behalf of a user process. This will 523 * ensure that those other threads have an opportunity to run in 524 * fairly short order (but not instantly). 525 */ 526 need_lwkt_resched(); 527 528 /* 529 * ITimer handling is per-tick, per-cpu. I don't think ksignal() 530 * is mpsafe on curproc, so XXX get the mplock. 531 */ 532 if ((p = curproc) != NULL && try_mplock()) { 533 if (frame && CLKF_USERMODE(frame) && 534 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 535 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 536 ksignal(p, SIGVTALRM); 537 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 538 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 539 ksignal(p, SIGPROF); 540 rel_mplock(); 541 } 542 setdelayed(); 543 } 544 545 /* 546 * The statistics clock typically runs at a 125Hz rate, and is intended 547 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 548 * 549 * NOTE! systimer! the MP lock might not be held here. We can only safely 550 * manipulate objects owned by the current cpu. 551 * 552 * The stats clock is responsible for grabbing a profiling sample. 553 * Most of the statistics are only used by user-level statistics programs. 554 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 555 * p->p_estcpu. 556 * 557 * Like the other clocks, the stat clock is called from what is effectively 558 * a fast interrupt, so the context should be the thread/process that got 559 * interrupted. 560 */ 561 static void 562 statclock(systimer_t info, struct intrframe *frame) 563 { 564 #ifdef GPROF 565 struct gmonparam *g; 566 int i; 567 #endif 568 thread_t td; 569 struct proc *p; 570 int bump; 571 struct timeval tv; 572 struct timeval *stv; 573 574 /* 575 * How big was our timeslice relative to the last time? 576 */ 577 microuptime(&tv); /* mpsafe */ 578 stv = &mycpu->gd_stattv; 579 if (stv->tv_sec == 0) { 580 bump = 1; 581 } else { 582 bump = tv.tv_usec - stv->tv_usec + 583 (tv.tv_sec - stv->tv_sec) * 1000000; 584 if (bump < 0) 585 bump = 0; 586 if (bump > 1000000) 587 bump = 1000000; 588 } 589 *stv = tv; 590 591 td = curthread; 592 p = td->td_proc; 593 594 if (frame && CLKF_USERMODE(frame)) { 595 /* 596 * Came from userland, handle user time and deal with 597 * possible process. 598 */ 599 if (p && (p->p_flag & P_PROFIL)) 600 addupc_intr(p, CLKF_PC(frame), 1); 601 td->td_uticks += bump; 602 603 /* 604 * Charge the time as appropriate 605 */ 606 if (p && p->p_nice > NZERO) 607 cpu_time.cp_nice += bump; 608 else 609 cpu_time.cp_user += bump; 610 } else { 611 #ifdef GPROF 612 /* 613 * Kernel statistics are just like addupc_intr, only easier. 614 */ 615 g = &_gmonparam; 616 if (g->state == GMON_PROF_ON && frame) { 617 i = CLKF_PC(frame) - g->lowpc; 618 if (i < g->textsize) { 619 i /= HISTFRACTION * sizeof(*g->kcount); 620 g->kcount[i]++; 621 } 622 } 623 #endif 624 /* 625 * Came from kernel mode, so we were: 626 * - handling an interrupt, 627 * - doing syscall or trap work on behalf of the current 628 * user process, or 629 * - spinning in the idle loop. 630 * Whichever it is, charge the time as appropriate. 631 * Note that we charge interrupts to the current process, 632 * regardless of whether they are ``for'' that process, 633 * so that we know how much of its real time was spent 634 * in ``non-process'' (i.e., interrupt) work. 635 * 636 * XXX assume system if frame is NULL. A NULL frame 637 * can occur if ipi processing is done from a crit_exit(). 638 */ 639 if (frame && CLKF_INTR(frame)) 640 td->td_iticks += bump; 641 else 642 td->td_sticks += bump; 643 644 if (frame && CLKF_INTR(frame)) { 645 #ifdef DEBUG_PCTRACK 646 do_pctrack(frame, PCTRACK_INT); 647 #endif 648 cpu_time.cp_intr += bump; 649 } else { 650 if (td == &mycpu->gd_idlethread) { 651 cpu_time.cp_idle += bump; 652 } else { 653 #ifdef DEBUG_PCTRACK 654 if (frame) 655 do_pctrack(frame, PCTRACK_SYS); 656 #endif 657 cpu_time.cp_sys += bump; 658 } 659 } 660 } 661 } 662 663 #ifdef DEBUG_PCTRACK 664 /* 665 * Sample the PC when in the kernel or in an interrupt. User code can 666 * retrieve the information and generate a histogram or other output. 667 */ 668 669 static void 670 do_pctrack(struct intrframe *frame, int which) 671 { 672 struct kinfo_pctrack *pctrack; 673 674 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 675 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 676 (void *)CLKF_PC(frame); 677 ++pctrack->pc_index; 678 } 679 680 static int 681 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 682 { 683 struct kinfo_pcheader head; 684 int error; 685 int cpu; 686 int ntrack; 687 688 head.pc_ntrack = PCTRACK_SIZE; 689 head.pc_arysize = PCTRACK_ARYSIZE; 690 691 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 692 return (error); 693 694 for (cpu = 0; cpu < ncpus; ++cpu) { 695 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 696 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 697 sizeof(struct kinfo_pctrack)); 698 if (error) 699 break; 700 } 701 if (error) 702 break; 703 } 704 return (error); 705 } 706 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 707 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 708 709 #endif 710 711 /* 712 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 713 * the MP lock might not be held. We can safely manipulate parts of curproc 714 * but that's about it. 715 * 716 * Each cpu has its own scheduler clock. 717 */ 718 static void 719 schedclock(systimer_t info, struct intrframe *frame) 720 { 721 struct lwp *lp; 722 struct rusage *ru; 723 struct vmspace *vm; 724 long rss; 725 726 if ((lp = lwkt_preempted_proc()) != NULL) { 727 /* 728 * Account for cpu time used and hit the scheduler. Note 729 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 730 * HERE. 731 */ 732 ++lp->lwp_cpticks; 733 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 734 info->time); 735 } 736 if ((lp = curthread->td_lwp) != NULL) { 737 /* 738 * Update resource usage integrals and maximums. 739 */ 740 if ((ru = &lp->lwp_proc->p_ru) && 741 (vm = lp->lwp_proc->p_vmspace) != NULL) { 742 ru->ru_ixrss += pgtok(vm->vm_tsize); 743 ru->ru_idrss += pgtok(vm->vm_dsize); 744 ru->ru_isrss += pgtok(vm->vm_ssize); 745 rss = pgtok(vmspace_resident_count(vm)); 746 if (ru->ru_maxrss < rss) 747 ru->ru_maxrss = rss; 748 } 749 } 750 } 751 752 /* 753 * Compute number of ticks for the specified amount of time. The 754 * return value is intended to be used in a clock interrupt timed 755 * operation and guarenteed to meet or exceed the requested time. 756 * If the representation overflows, return INT_MAX. The minimum return 757 * value is 1 ticks and the function will average the calculation up. 758 * If any value greater then 0 microseconds is supplied, a value 759 * of at least 2 will be returned to ensure that a near-term clock 760 * interrupt does not cause the timeout to occur (degenerately) early. 761 * 762 * Note that limit checks must take into account microseconds, which is 763 * done simply by using the smaller signed long maximum instead of 764 * the unsigned long maximum. 765 * 766 * If ints have 32 bits, then the maximum value for any timeout in 767 * 10ms ticks is 248 days. 768 */ 769 int 770 tvtohz_high(struct timeval *tv) 771 { 772 int ticks; 773 long sec, usec; 774 775 sec = tv->tv_sec; 776 usec = tv->tv_usec; 777 if (usec < 0) { 778 sec--; 779 usec += 1000000; 780 } 781 if (sec < 0) { 782 #ifdef DIAGNOSTIC 783 if (usec > 0) { 784 sec++; 785 usec -= 1000000; 786 } 787 kprintf("tvtohz_high: negative time difference %ld sec %ld usec\n", 788 sec, usec); 789 #endif 790 ticks = 1; 791 } else if (sec <= INT_MAX / hz) { 792 ticks = (int)(sec * hz + 793 ((u_long)usec + (tick - 1)) / tick) + 1; 794 } else { 795 ticks = INT_MAX; 796 } 797 return (ticks); 798 } 799 800 /* 801 * Compute number of ticks for the specified amount of time, erroring on 802 * the side of it being too low to ensure that sleeping the returned number 803 * of ticks will not result in a late return. 804 * 805 * The supplied timeval may not be negative and should be normalized. A 806 * return value of 0 is possible if the timeval converts to less then 807 * 1 tick. 808 * 809 * If ints have 32 bits, then the maximum value for any timeout in 810 * 10ms ticks is 248 days. 811 */ 812 int 813 tvtohz_low(struct timeval *tv) 814 { 815 int ticks; 816 long sec; 817 818 sec = tv->tv_sec; 819 if (sec <= INT_MAX / hz) 820 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 821 else 822 ticks = INT_MAX; 823 return (ticks); 824 } 825 826 827 /* 828 * Start profiling on a process. 829 * 830 * Kernel profiling passes proc0 which never exits and hence 831 * keeps the profile clock running constantly. 832 */ 833 void 834 startprofclock(struct proc *p) 835 { 836 if ((p->p_flag & P_PROFIL) == 0) { 837 p->p_flag |= P_PROFIL; 838 #if 0 /* XXX */ 839 if (++profprocs == 1 && stathz != 0) { 840 crit_enter(); 841 psdiv = psratio; 842 setstatclockrate(profhz); 843 crit_exit(); 844 } 845 #endif 846 } 847 } 848 849 /* 850 * Stop profiling on a process. 851 */ 852 void 853 stopprofclock(struct proc *p) 854 { 855 if (p->p_flag & P_PROFIL) { 856 p->p_flag &= ~P_PROFIL; 857 #if 0 /* XXX */ 858 if (--profprocs == 0 && stathz != 0) { 859 crit_enter(); 860 psdiv = 1; 861 setstatclockrate(stathz); 862 crit_exit(); 863 } 864 #endif 865 } 866 } 867 868 /* 869 * Return information about system clocks. 870 */ 871 static int 872 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 873 { 874 struct kinfo_clockinfo clkinfo; 875 /* 876 * Construct clockinfo structure. 877 */ 878 clkinfo.ci_hz = hz; 879 clkinfo.ci_tick = tick; 880 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 881 clkinfo.ci_profhz = profhz; 882 clkinfo.ci_stathz = stathz ? stathz : hz; 883 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 884 } 885 886 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 887 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 888 889 /* 890 * We have eight functions for looking at the clock, four for 891 * microseconds and four for nanoseconds. For each there is fast 892 * but less precise version "get{nano|micro}[up]time" which will 893 * return a time which is up to 1/HZ previous to the call, whereas 894 * the raw version "{nano|micro}[up]time" will return a timestamp 895 * which is as precise as possible. The "up" variants return the 896 * time relative to system boot, these are well suited for time 897 * interval measurements. 898 * 899 * Each cpu independantly maintains the current time of day, so all 900 * we need to do to protect ourselves from changes is to do a loop 901 * check on the seconds field changing out from under us. 902 * 903 * The system timer maintains a 32 bit count and due to various issues 904 * it is possible for the calculated delta to occassionally exceed 905 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 906 * multiplication can easily overflow, so we deal with the case. For 907 * uniformity we deal with the case in the usec case too. 908 */ 909 void 910 getmicrouptime(struct timeval *tvp) 911 { 912 struct globaldata *gd = mycpu; 913 sysclock_t delta; 914 915 do { 916 tvp->tv_sec = gd->gd_time_seconds; 917 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 918 } while (tvp->tv_sec != gd->gd_time_seconds); 919 920 if (delta >= sys_cputimer->freq) { 921 tvp->tv_sec += delta / sys_cputimer->freq; 922 delta %= sys_cputimer->freq; 923 } 924 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 925 if (tvp->tv_usec >= 1000000) { 926 tvp->tv_usec -= 1000000; 927 ++tvp->tv_sec; 928 } 929 } 930 931 void 932 getnanouptime(struct timespec *tsp) 933 { 934 struct globaldata *gd = mycpu; 935 sysclock_t delta; 936 937 do { 938 tsp->tv_sec = gd->gd_time_seconds; 939 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 940 } while (tsp->tv_sec != gd->gd_time_seconds); 941 942 if (delta >= sys_cputimer->freq) { 943 tsp->tv_sec += delta / sys_cputimer->freq; 944 delta %= sys_cputimer->freq; 945 } 946 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 947 } 948 949 void 950 microuptime(struct timeval *tvp) 951 { 952 struct globaldata *gd = mycpu; 953 sysclock_t delta; 954 955 do { 956 tvp->tv_sec = gd->gd_time_seconds; 957 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 958 } while (tvp->tv_sec != gd->gd_time_seconds); 959 960 if (delta >= sys_cputimer->freq) { 961 tvp->tv_sec += delta / sys_cputimer->freq; 962 delta %= sys_cputimer->freq; 963 } 964 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 965 } 966 967 void 968 nanouptime(struct timespec *tsp) 969 { 970 struct globaldata *gd = mycpu; 971 sysclock_t delta; 972 973 do { 974 tsp->tv_sec = gd->gd_time_seconds; 975 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 976 } while (tsp->tv_sec != gd->gd_time_seconds); 977 978 if (delta >= sys_cputimer->freq) { 979 tsp->tv_sec += delta / sys_cputimer->freq; 980 delta %= sys_cputimer->freq; 981 } 982 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 983 } 984 985 /* 986 * realtime routines 987 */ 988 989 void 990 getmicrotime(struct timeval *tvp) 991 { 992 struct globaldata *gd = mycpu; 993 struct timespec *bt; 994 sysclock_t delta; 995 996 do { 997 tvp->tv_sec = gd->gd_time_seconds; 998 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 999 } while (tvp->tv_sec != gd->gd_time_seconds); 1000 1001 if (delta >= sys_cputimer->freq) { 1002 tvp->tv_sec += delta / sys_cputimer->freq; 1003 delta %= sys_cputimer->freq; 1004 } 1005 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1006 1007 bt = &basetime[basetime_index]; 1008 tvp->tv_sec += bt->tv_sec; 1009 tvp->tv_usec += bt->tv_nsec / 1000; 1010 while (tvp->tv_usec >= 1000000) { 1011 tvp->tv_usec -= 1000000; 1012 ++tvp->tv_sec; 1013 } 1014 } 1015 1016 void 1017 getnanotime(struct timespec *tsp) 1018 { 1019 struct globaldata *gd = mycpu; 1020 struct timespec *bt; 1021 sysclock_t delta; 1022 1023 do { 1024 tsp->tv_sec = gd->gd_time_seconds; 1025 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1026 } while (tsp->tv_sec != gd->gd_time_seconds); 1027 1028 if (delta >= sys_cputimer->freq) { 1029 tsp->tv_sec += delta / sys_cputimer->freq; 1030 delta %= sys_cputimer->freq; 1031 } 1032 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1033 1034 bt = &basetime[basetime_index]; 1035 tsp->tv_sec += bt->tv_sec; 1036 tsp->tv_nsec += bt->tv_nsec; 1037 while (tsp->tv_nsec >= 1000000000) { 1038 tsp->tv_nsec -= 1000000000; 1039 ++tsp->tv_sec; 1040 } 1041 } 1042 1043 static void 1044 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1045 { 1046 struct globaldata *gd = mycpu; 1047 sysclock_t delta; 1048 1049 do { 1050 tsp->tv_sec = gd->gd_time_seconds; 1051 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1052 } while (tsp->tv_sec != gd->gd_time_seconds); 1053 1054 if (delta >= sys_cputimer->freq) { 1055 tsp->tv_sec += delta / sys_cputimer->freq; 1056 delta %= sys_cputimer->freq; 1057 } 1058 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1059 1060 tsp->tv_sec += nbt->tv_sec; 1061 tsp->tv_nsec += nbt->tv_nsec; 1062 while (tsp->tv_nsec >= 1000000000) { 1063 tsp->tv_nsec -= 1000000000; 1064 ++tsp->tv_sec; 1065 } 1066 } 1067 1068 1069 void 1070 microtime(struct timeval *tvp) 1071 { 1072 struct globaldata *gd = mycpu; 1073 struct timespec *bt; 1074 sysclock_t delta; 1075 1076 do { 1077 tvp->tv_sec = gd->gd_time_seconds; 1078 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1079 } while (tvp->tv_sec != gd->gd_time_seconds); 1080 1081 if (delta >= sys_cputimer->freq) { 1082 tvp->tv_sec += delta / sys_cputimer->freq; 1083 delta %= sys_cputimer->freq; 1084 } 1085 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1086 1087 bt = &basetime[basetime_index]; 1088 tvp->tv_sec += bt->tv_sec; 1089 tvp->tv_usec += bt->tv_nsec / 1000; 1090 while (tvp->tv_usec >= 1000000) { 1091 tvp->tv_usec -= 1000000; 1092 ++tvp->tv_sec; 1093 } 1094 } 1095 1096 void 1097 nanotime(struct timespec *tsp) 1098 { 1099 struct globaldata *gd = mycpu; 1100 struct timespec *bt; 1101 sysclock_t delta; 1102 1103 do { 1104 tsp->tv_sec = gd->gd_time_seconds; 1105 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1106 } while (tsp->tv_sec != gd->gd_time_seconds); 1107 1108 if (delta >= sys_cputimer->freq) { 1109 tsp->tv_sec += delta / sys_cputimer->freq; 1110 delta %= sys_cputimer->freq; 1111 } 1112 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1113 1114 bt = &basetime[basetime_index]; 1115 tsp->tv_sec += bt->tv_sec; 1116 tsp->tv_nsec += bt->tv_nsec; 1117 while (tsp->tv_nsec >= 1000000000) { 1118 tsp->tv_nsec -= 1000000000; 1119 ++tsp->tv_sec; 1120 } 1121 } 1122 1123 /* 1124 * note: this is not exactly synchronized with real time. To do that we 1125 * would have to do what microtime does and check for a nanoseconds overflow. 1126 */ 1127 time_t 1128 get_approximate_time_t(void) 1129 { 1130 struct globaldata *gd = mycpu; 1131 struct timespec *bt; 1132 1133 bt = &basetime[basetime_index]; 1134 return(gd->gd_time_seconds + bt->tv_sec); 1135 } 1136 1137 int 1138 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1139 { 1140 pps_params_t *app; 1141 struct pps_fetch_args *fapi; 1142 #ifdef PPS_SYNC 1143 struct pps_kcbind_args *kapi; 1144 #endif 1145 1146 switch (cmd) { 1147 case PPS_IOC_CREATE: 1148 return (0); 1149 case PPS_IOC_DESTROY: 1150 return (0); 1151 case PPS_IOC_SETPARAMS: 1152 app = (pps_params_t *)data; 1153 if (app->mode & ~pps->ppscap) 1154 return (EINVAL); 1155 pps->ppsparam = *app; 1156 return (0); 1157 case PPS_IOC_GETPARAMS: 1158 app = (pps_params_t *)data; 1159 *app = pps->ppsparam; 1160 app->api_version = PPS_API_VERS_1; 1161 return (0); 1162 case PPS_IOC_GETCAP: 1163 *(int*)data = pps->ppscap; 1164 return (0); 1165 case PPS_IOC_FETCH: 1166 fapi = (struct pps_fetch_args *)data; 1167 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1168 return (EINVAL); 1169 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1170 return (EOPNOTSUPP); 1171 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1172 fapi->pps_info_buf = pps->ppsinfo; 1173 return (0); 1174 case PPS_IOC_KCBIND: 1175 #ifdef PPS_SYNC 1176 kapi = (struct pps_kcbind_args *)data; 1177 /* XXX Only root should be able to do this */ 1178 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1179 return (EINVAL); 1180 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1181 return (EINVAL); 1182 if (kapi->edge & ~pps->ppscap) 1183 return (EINVAL); 1184 pps->kcmode = kapi->edge; 1185 return (0); 1186 #else 1187 return (EOPNOTSUPP); 1188 #endif 1189 default: 1190 return (ENOTTY); 1191 } 1192 } 1193 1194 void 1195 pps_init(struct pps_state *pps) 1196 { 1197 pps->ppscap |= PPS_TSFMT_TSPEC; 1198 if (pps->ppscap & PPS_CAPTUREASSERT) 1199 pps->ppscap |= PPS_OFFSETASSERT; 1200 if (pps->ppscap & PPS_CAPTURECLEAR) 1201 pps->ppscap |= PPS_OFFSETCLEAR; 1202 } 1203 1204 void 1205 pps_event(struct pps_state *pps, sysclock_t count, int event) 1206 { 1207 struct globaldata *gd; 1208 struct timespec *tsp; 1209 struct timespec *osp; 1210 struct timespec *bt; 1211 struct timespec ts; 1212 sysclock_t *pcount; 1213 #ifdef PPS_SYNC 1214 sysclock_t tcount; 1215 #endif 1216 sysclock_t delta; 1217 pps_seq_t *pseq; 1218 int foff; 1219 int fhard; 1220 1221 gd = mycpu; 1222 1223 /* Things would be easier with arrays... */ 1224 if (event == PPS_CAPTUREASSERT) { 1225 tsp = &pps->ppsinfo.assert_timestamp; 1226 osp = &pps->ppsparam.assert_offset; 1227 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1228 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1229 pcount = &pps->ppscount[0]; 1230 pseq = &pps->ppsinfo.assert_sequence; 1231 } else { 1232 tsp = &pps->ppsinfo.clear_timestamp; 1233 osp = &pps->ppsparam.clear_offset; 1234 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1235 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1236 pcount = &pps->ppscount[1]; 1237 pseq = &pps->ppsinfo.clear_sequence; 1238 } 1239 1240 /* Nothing really happened */ 1241 if (*pcount == count) 1242 return; 1243 1244 *pcount = count; 1245 1246 do { 1247 ts.tv_sec = gd->gd_time_seconds; 1248 delta = count - gd->gd_cpuclock_base; 1249 } while (ts.tv_sec != gd->gd_time_seconds); 1250 1251 if (delta >= sys_cputimer->freq) { 1252 ts.tv_sec += delta / sys_cputimer->freq; 1253 delta %= sys_cputimer->freq; 1254 } 1255 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1256 bt = &basetime[basetime_index]; 1257 ts.tv_sec += bt->tv_sec; 1258 ts.tv_nsec += bt->tv_nsec; 1259 while (ts.tv_nsec >= 1000000000) { 1260 ts.tv_nsec -= 1000000000; 1261 ++ts.tv_sec; 1262 } 1263 1264 (*pseq)++; 1265 *tsp = ts; 1266 1267 if (foff) { 1268 timespecadd(tsp, osp); 1269 if (tsp->tv_nsec < 0) { 1270 tsp->tv_nsec += 1000000000; 1271 tsp->tv_sec -= 1; 1272 } 1273 } 1274 #ifdef PPS_SYNC 1275 if (fhard) { 1276 /* magic, at its best... */ 1277 tcount = count - pps->ppscount[2]; 1278 pps->ppscount[2] = count; 1279 if (tcount >= sys_cputimer->freq) { 1280 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1281 sys_cputimer->freq64_nsec * 1282 (tcount % sys_cputimer->freq)) >> 32; 1283 } else { 1284 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1285 } 1286 hardpps(tsp, delta); 1287 } 1288 #endif 1289 } 1290 1291