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 systimer_intr_enable(); 261 262 #ifdef DEVICE_POLLING 263 init_device_poll_pcpu(gd->gd_cpuid); 264 #endif 265 266 #ifdef IFPOLL_ENABLE 267 ifpoll_init_pcpu(gd->gd_cpuid); 268 #endif 269 270 /* 271 * Use a non-queued periodic systimer to prevent multiple ticks from 272 * building up if the sysclock jumps forward (8254 gets reset). The 273 * sysclock will never jump backwards. Our time sync is based on 274 * the actual sysclock, not the ticks count. 275 */ 276 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 277 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 278 /* XXX correct the frequency for scheduler / estcpu tests */ 279 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 280 NULL, ESTCPUFREQ); 281 crit_exit(); 282 } 283 284 /* 285 * This sets the current real time of day. Timespecs are in seconds and 286 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 287 * instead we adjust basetime so basetime + gd_* results in the current 288 * time of day. This way the gd_* fields are guarenteed to represent 289 * a monotonically increasing 'uptime' value. 290 * 291 * When set_timeofday() is called from userland, the system call forces it 292 * onto cpu #0 since only cpu #0 can update basetime_index. 293 */ 294 void 295 set_timeofday(struct timespec *ts) 296 { 297 struct timespec *nbt; 298 int ni; 299 300 /* 301 * XXX SMP / non-atomic basetime updates 302 */ 303 crit_enter(); 304 ni = (basetime_index + 1) & BASETIME_ARYMASK; 305 nbt = &basetime[ni]; 306 nanouptime(nbt); 307 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 308 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 309 if (nbt->tv_nsec < 0) { 310 nbt->tv_nsec += 1000000000; 311 --nbt->tv_sec; 312 } 313 314 /* 315 * Note that basetime diverges from boottime as the clock drift is 316 * compensated for, so we cannot do away with boottime. When setting 317 * the absolute time of day the drift is 0 (for an instant) and we 318 * can simply assign boottime to basetime. 319 * 320 * Note that nanouptime() is based on gd_time_seconds which is drift 321 * compensated up to a point (it is guarenteed to remain monotonically 322 * increasing). gd_time_seconds is thus our best uptime guess and 323 * suitable for use in the boottime calculation. It is already taken 324 * into account in the basetime calculation above. 325 */ 326 boottime.tv_sec = nbt->tv_sec; 327 ntp_delta = 0; 328 329 /* 330 * We now have a new basetime, make sure all other cpus have it, 331 * then update the index. 332 */ 333 cpu_sfence(); 334 basetime_index = ni; 335 336 crit_exit(); 337 } 338 339 /* 340 * Each cpu has its own hardclock, but we only increments ticks and softticks 341 * on cpu #0. 342 * 343 * NOTE! systimer! the MP lock might not be held here. We can only safely 344 * manipulate objects owned by the current cpu. 345 */ 346 static void 347 hardclock(systimer_t info, struct intrframe *frame) 348 { 349 sysclock_t cputicks; 350 struct proc *p; 351 struct globaldata *gd = mycpu; 352 353 /* 354 * Realtime updates are per-cpu. Note that timer corrections as 355 * returned by microtime() and friends make an additional adjustment 356 * using a system-wise 'basetime', but the running time is always 357 * taken from the per-cpu globaldata area. Since the same clock 358 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 359 * stay in synch. 360 * 361 * Note that we never allow info->time (aka gd->gd_hardclock.time) 362 * to reverse index gd_cpuclock_base, but that it is possible for 363 * it to temporarily get behind in the seconds if something in the 364 * system locks interrupts for a long period of time. Since periodic 365 * timers count events, though everything should resynch again 366 * immediately. 367 */ 368 cputicks = info->time - gd->gd_cpuclock_base; 369 if (cputicks >= sys_cputimer->freq) { 370 ++gd->gd_time_seconds; 371 gd->gd_cpuclock_base += sys_cputimer->freq; 372 } 373 374 /* 375 * The system-wide ticks counter and NTP related timedelta/tickdelta 376 * adjustments only occur on cpu #0. NTP adjustments are accomplished 377 * by updating basetime. 378 */ 379 if (gd->gd_cpuid == 0) { 380 struct timespec *nbt; 381 struct timespec nts; 382 int leap; 383 int ni; 384 385 ++ticks; 386 387 #if 0 388 if (tco->tc_poll_pps) 389 tco->tc_poll_pps(tco); 390 #endif 391 392 /* 393 * Calculate the new basetime index. We are in a critical section 394 * on cpu #0 and can safely play with basetime_index. Start 395 * with the current basetime and then make adjustments. 396 */ 397 ni = (basetime_index + 1) & BASETIME_ARYMASK; 398 nbt = &basetime[ni]; 399 *nbt = basetime[basetime_index]; 400 401 /* 402 * Apply adjtime corrections. (adjtime() API) 403 * 404 * adjtime() only runs on cpu #0 so our critical section is 405 * sufficient to access these variables. 406 */ 407 if (ntp_delta != 0) { 408 nbt->tv_nsec += ntp_tick_delta; 409 ntp_delta -= ntp_tick_delta; 410 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 411 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 412 ntp_tick_delta = ntp_delta; 413 } 414 } 415 416 /* 417 * Apply permanent frequency corrections. (sysctl API) 418 */ 419 if (ntp_tick_permanent != 0) { 420 ntp_tick_acc += ntp_tick_permanent; 421 if (ntp_tick_acc >= (1LL << 32)) { 422 nbt->tv_nsec += ntp_tick_acc >> 32; 423 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 424 } else if (ntp_tick_acc <= -(1LL << 32)) { 425 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 426 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 427 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 428 } 429 } 430 431 if (nbt->tv_nsec >= 1000000000) { 432 nbt->tv_sec++; 433 nbt->tv_nsec -= 1000000000; 434 } else if (nbt->tv_nsec < 0) { 435 nbt->tv_sec--; 436 nbt->tv_nsec += 1000000000; 437 } 438 439 /* 440 * Another per-tick compensation. (for ntp_adjtime() API) 441 */ 442 if (nsec_adj != 0) { 443 nsec_acc += nsec_adj; 444 if (nsec_acc >= 0x100000000LL) { 445 nbt->tv_nsec += nsec_acc >> 32; 446 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 447 } else if (nsec_acc <= -0x100000000LL) { 448 nbt->tv_nsec -= -nsec_acc >> 32; 449 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 450 } 451 if (nbt->tv_nsec >= 1000000000) { 452 nbt->tv_nsec -= 1000000000; 453 ++nbt->tv_sec; 454 } else if (nbt->tv_nsec < 0) { 455 nbt->tv_nsec += 1000000000; 456 --nbt->tv_sec; 457 } 458 } 459 460 /************************************************************ 461 * LEAP SECOND CORRECTION * 462 ************************************************************ 463 * 464 * Taking into account all the corrections made above, figure 465 * out the new real time. If the seconds field has changed 466 * then apply any pending leap-second corrections. 467 */ 468 getnanotime_nbt(nbt, &nts); 469 470 if (time_second != nts.tv_sec) { 471 /* 472 * Apply leap second (sysctl API). Adjust nts for changes 473 * so we do not have to call getnanotime_nbt again. 474 */ 475 if (ntp_leap_second) { 476 if (ntp_leap_second == nts.tv_sec) { 477 if (ntp_leap_insert) { 478 nbt->tv_sec++; 479 nts.tv_sec++; 480 } else { 481 nbt->tv_sec--; 482 nts.tv_sec--; 483 } 484 ntp_leap_second--; 485 } 486 } 487 488 /* 489 * Apply leap second (ntp_adjtime() API), calculate a new 490 * nsec_adj field. ntp_update_second() returns nsec_adj 491 * as a per-second value but we need it as a per-tick value. 492 */ 493 leap = ntp_update_second(time_second, &nsec_adj); 494 nsec_adj /= hz; 495 nbt->tv_sec += leap; 496 nts.tv_sec += leap; 497 498 /* 499 * Update the time_second 'approximate time' global. 500 */ 501 time_second = nts.tv_sec; 502 } 503 504 /* 505 * Finally, our new basetime is ready to go live! 506 */ 507 cpu_sfence(); 508 basetime_index = ni; 509 510 /* 511 * Figure out how badly the system is starved for memory 512 */ 513 vm_fault_ratecheck(); 514 } 515 516 /* 517 * softticks are handled for all cpus 518 */ 519 hardclock_softtick(gd); 520 521 /* 522 * The LWKT scheduler will generally allow the current process to 523 * return to user mode even if there are other runnable LWKT threads 524 * running in kernel mode on behalf of a user process. This will 525 * ensure that those other threads have an opportunity to run in 526 * fairly short order (but not instantly). 527 */ 528 need_lwkt_resched(); 529 530 /* 531 * ITimer handling is per-tick, per-cpu. I don't think ksignal() 532 * is mpsafe on curproc, so XXX get the mplock. 533 */ 534 if ((p = curproc) != NULL && try_mplock()) { 535 if (frame && CLKF_USERMODE(frame) && 536 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 537 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 538 ksignal(p, SIGVTALRM); 539 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 540 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 541 ksignal(p, SIGPROF); 542 rel_mplock(); 543 } 544 setdelayed(); 545 } 546 547 /* 548 * The statistics clock typically runs at a 125Hz rate, and is intended 549 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 550 * 551 * NOTE! systimer! the MP lock might not be held here. We can only safely 552 * manipulate objects owned by the current cpu. 553 * 554 * The stats clock is responsible for grabbing a profiling sample. 555 * Most of the statistics are only used by user-level statistics programs. 556 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 557 * p->p_estcpu. 558 * 559 * Like the other clocks, the stat clock is called from what is effectively 560 * a fast interrupt, so the context should be the thread/process that got 561 * interrupted. 562 */ 563 static void 564 statclock(systimer_t info, struct intrframe *frame) 565 { 566 #ifdef GPROF 567 struct gmonparam *g; 568 int i; 569 #endif 570 thread_t td; 571 struct proc *p; 572 int bump; 573 struct timeval tv; 574 struct timeval *stv; 575 576 /* 577 * How big was our timeslice relative to the last time? 578 */ 579 microuptime(&tv); /* mpsafe */ 580 stv = &mycpu->gd_stattv; 581 if (stv->tv_sec == 0) { 582 bump = 1; 583 } else { 584 bump = tv.tv_usec - stv->tv_usec + 585 (tv.tv_sec - stv->tv_sec) * 1000000; 586 if (bump < 0) 587 bump = 0; 588 if (bump > 1000000) 589 bump = 1000000; 590 } 591 *stv = tv; 592 593 td = curthread; 594 p = td->td_proc; 595 596 if (frame && CLKF_USERMODE(frame)) { 597 /* 598 * Came from userland, handle user time and deal with 599 * possible process. 600 */ 601 if (p && (p->p_flag & P_PROFIL)) 602 addupc_intr(p, CLKF_PC(frame), 1); 603 td->td_uticks += bump; 604 605 /* 606 * Charge the time as appropriate 607 */ 608 if (p && p->p_nice > NZERO) 609 cpu_time.cp_nice += bump; 610 else 611 cpu_time.cp_user += bump; 612 } else { 613 #ifdef GPROF 614 /* 615 * Kernel statistics are just like addupc_intr, only easier. 616 */ 617 g = &_gmonparam; 618 if (g->state == GMON_PROF_ON && frame) { 619 i = CLKF_PC(frame) - g->lowpc; 620 if (i < g->textsize) { 621 i /= HISTFRACTION * sizeof(*g->kcount); 622 g->kcount[i]++; 623 } 624 } 625 #endif 626 /* 627 * Came from kernel mode, so we were: 628 * - handling an interrupt, 629 * - doing syscall or trap work on behalf of the current 630 * user process, or 631 * - spinning in the idle loop. 632 * Whichever it is, charge the time as appropriate. 633 * Note that we charge interrupts to the current process, 634 * regardless of whether they are ``for'' that process, 635 * so that we know how much of its real time was spent 636 * in ``non-process'' (i.e., interrupt) work. 637 * 638 * XXX assume system if frame is NULL. A NULL frame 639 * can occur if ipi processing is done from a crit_exit(). 640 */ 641 if (frame && CLKF_INTR(frame)) 642 td->td_iticks += bump; 643 else 644 td->td_sticks += bump; 645 646 if (frame && CLKF_INTR(frame)) { 647 #ifdef DEBUG_PCTRACK 648 do_pctrack(frame, PCTRACK_INT); 649 #endif 650 cpu_time.cp_intr += bump; 651 } else { 652 if (td == &mycpu->gd_idlethread) { 653 cpu_time.cp_idle += bump; 654 } else { 655 #ifdef DEBUG_PCTRACK 656 if (frame) 657 do_pctrack(frame, PCTRACK_SYS); 658 #endif 659 cpu_time.cp_sys += bump; 660 } 661 } 662 } 663 } 664 665 #ifdef DEBUG_PCTRACK 666 /* 667 * Sample the PC when in the kernel or in an interrupt. User code can 668 * retrieve the information and generate a histogram or other output. 669 */ 670 671 static void 672 do_pctrack(struct intrframe *frame, int which) 673 { 674 struct kinfo_pctrack *pctrack; 675 676 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 677 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 678 (void *)CLKF_PC(frame); 679 ++pctrack->pc_index; 680 } 681 682 static int 683 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 684 { 685 struct kinfo_pcheader head; 686 int error; 687 int cpu; 688 int ntrack; 689 690 head.pc_ntrack = PCTRACK_SIZE; 691 head.pc_arysize = PCTRACK_ARYSIZE; 692 693 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 694 return (error); 695 696 for (cpu = 0; cpu < ncpus; ++cpu) { 697 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 698 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 699 sizeof(struct kinfo_pctrack)); 700 if (error) 701 break; 702 } 703 if (error) 704 break; 705 } 706 return (error); 707 } 708 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 709 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 710 711 #endif 712 713 /* 714 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 715 * the MP lock might not be held. We can safely manipulate parts of curproc 716 * but that's about it. 717 * 718 * Each cpu has its own scheduler clock. 719 */ 720 static void 721 schedclock(systimer_t info, struct intrframe *frame) 722 { 723 struct lwp *lp; 724 struct rusage *ru; 725 struct vmspace *vm; 726 long rss; 727 728 if ((lp = lwkt_preempted_proc()) != NULL) { 729 /* 730 * Account for cpu time used and hit the scheduler. Note 731 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 732 * HERE. 733 */ 734 ++lp->lwp_cpticks; 735 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 736 info->time); 737 } 738 if ((lp = curthread->td_lwp) != NULL) { 739 /* 740 * Update resource usage integrals and maximums. 741 */ 742 if ((ru = &lp->lwp_proc->p_ru) && 743 (vm = lp->lwp_proc->p_vmspace) != NULL) { 744 ru->ru_ixrss += pgtok(vm->vm_tsize); 745 ru->ru_idrss += pgtok(vm->vm_dsize); 746 ru->ru_isrss += pgtok(vm->vm_ssize); 747 rss = pgtok(vmspace_resident_count(vm)); 748 if (ru->ru_maxrss < rss) 749 ru->ru_maxrss = rss; 750 } 751 } 752 } 753 754 /* 755 * Compute number of ticks for the specified amount of time. The 756 * return value is intended to be used in a clock interrupt timed 757 * operation and guarenteed to meet or exceed the requested time. 758 * If the representation overflows, return INT_MAX. The minimum return 759 * value is 1 ticks and the function will average the calculation up. 760 * If any value greater then 0 microseconds is supplied, a value 761 * of at least 2 will be returned to ensure that a near-term clock 762 * interrupt does not cause the timeout to occur (degenerately) early. 763 * 764 * Note that limit checks must take into account microseconds, which is 765 * done simply by using the smaller signed long maximum instead of 766 * the unsigned long maximum. 767 * 768 * If ints have 32 bits, then the maximum value for any timeout in 769 * 10ms ticks is 248 days. 770 */ 771 int 772 tvtohz_high(struct timeval *tv) 773 { 774 int ticks; 775 long sec, usec; 776 777 sec = tv->tv_sec; 778 usec = tv->tv_usec; 779 if (usec < 0) { 780 sec--; 781 usec += 1000000; 782 } 783 if (sec < 0) { 784 #ifdef DIAGNOSTIC 785 if (usec > 0) { 786 sec++; 787 usec -= 1000000; 788 } 789 kprintf("tvtohz_high: negative time difference %ld sec %ld usec\n", 790 sec, usec); 791 #endif 792 ticks = 1; 793 } else if (sec <= INT_MAX / hz) { 794 ticks = (int)(sec * hz + 795 ((u_long)usec + (tick - 1)) / tick) + 1; 796 } else { 797 ticks = INT_MAX; 798 } 799 return (ticks); 800 } 801 802 /* 803 * Compute number of ticks for the specified amount of time, erroring on 804 * the side of it being too low to ensure that sleeping the returned number 805 * of ticks will not result in a late return. 806 * 807 * The supplied timeval may not be negative and should be normalized. A 808 * return value of 0 is possible if the timeval converts to less then 809 * 1 tick. 810 * 811 * If ints have 32 bits, then the maximum value for any timeout in 812 * 10ms ticks is 248 days. 813 */ 814 int 815 tvtohz_low(struct timeval *tv) 816 { 817 int ticks; 818 long sec; 819 820 sec = tv->tv_sec; 821 if (sec <= INT_MAX / hz) 822 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 823 else 824 ticks = INT_MAX; 825 return (ticks); 826 } 827 828 829 /* 830 * Start profiling on a process. 831 * 832 * Kernel profiling passes proc0 which never exits and hence 833 * keeps the profile clock running constantly. 834 */ 835 void 836 startprofclock(struct proc *p) 837 { 838 if ((p->p_flag & P_PROFIL) == 0) { 839 p->p_flag |= P_PROFIL; 840 #if 0 /* XXX */ 841 if (++profprocs == 1 && stathz != 0) { 842 crit_enter(); 843 psdiv = psratio; 844 setstatclockrate(profhz); 845 crit_exit(); 846 } 847 #endif 848 } 849 } 850 851 /* 852 * Stop profiling on a process. 853 */ 854 void 855 stopprofclock(struct proc *p) 856 { 857 if (p->p_flag & P_PROFIL) { 858 p->p_flag &= ~P_PROFIL; 859 #if 0 /* XXX */ 860 if (--profprocs == 0 && stathz != 0) { 861 crit_enter(); 862 psdiv = 1; 863 setstatclockrate(stathz); 864 crit_exit(); 865 } 866 #endif 867 } 868 } 869 870 /* 871 * Return information about system clocks. 872 */ 873 static int 874 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 875 { 876 struct kinfo_clockinfo clkinfo; 877 /* 878 * Construct clockinfo structure. 879 */ 880 clkinfo.ci_hz = hz; 881 clkinfo.ci_tick = tick; 882 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 883 clkinfo.ci_profhz = profhz; 884 clkinfo.ci_stathz = stathz ? stathz : hz; 885 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 886 } 887 888 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 889 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 890 891 /* 892 * We have eight functions for looking at the clock, four for 893 * microseconds and four for nanoseconds. For each there is fast 894 * but less precise version "get{nano|micro}[up]time" which will 895 * return a time which is up to 1/HZ previous to the call, whereas 896 * the raw version "{nano|micro}[up]time" will return a timestamp 897 * which is as precise as possible. The "up" variants return the 898 * time relative to system boot, these are well suited for time 899 * interval measurements. 900 * 901 * Each cpu independantly maintains the current time of day, so all 902 * we need to do to protect ourselves from changes is to do a loop 903 * check on the seconds field changing out from under us. 904 * 905 * The system timer maintains a 32 bit count and due to various issues 906 * it is possible for the calculated delta to occassionally exceed 907 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 908 * multiplication can easily overflow, so we deal with the case. For 909 * uniformity we deal with the case in the usec case too. 910 */ 911 void 912 getmicrouptime(struct timeval *tvp) 913 { 914 struct globaldata *gd = mycpu; 915 sysclock_t delta; 916 917 do { 918 tvp->tv_sec = gd->gd_time_seconds; 919 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 920 } while (tvp->tv_sec != gd->gd_time_seconds); 921 922 if (delta >= sys_cputimer->freq) { 923 tvp->tv_sec += delta / sys_cputimer->freq; 924 delta %= sys_cputimer->freq; 925 } 926 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 927 if (tvp->tv_usec >= 1000000) { 928 tvp->tv_usec -= 1000000; 929 ++tvp->tv_sec; 930 } 931 } 932 933 void 934 getnanouptime(struct timespec *tsp) 935 { 936 struct globaldata *gd = mycpu; 937 sysclock_t delta; 938 939 do { 940 tsp->tv_sec = gd->gd_time_seconds; 941 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 942 } while (tsp->tv_sec != gd->gd_time_seconds); 943 944 if (delta >= sys_cputimer->freq) { 945 tsp->tv_sec += delta / sys_cputimer->freq; 946 delta %= sys_cputimer->freq; 947 } 948 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 949 } 950 951 void 952 microuptime(struct timeval *tvp) 953 { 954 struct globaldata *gd = mycpu; 955 sysclock_t delta; 956 957 do { 958 tvp->tv_sec = gd->gd_time_seconds; 959 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 960 } while (tvp->tv_sec != gd->gd_time_seconds); 961 962 if (delta >= sys_cputimer->freq) { 963 tvp->tv_sec += delta / sys_cputimer->freq; 964 delta %= sys_cputimer->freq; 965 } 966 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 967 } 968 969 void 970 nanouptime(struct timespec *tsp) 971 { 972 struct globaldata *gd = mycpu; 973 sysclock_t delta; 974 975 do { 976 tsp->tv_sec = gd->gd_time_seconds; 977 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 978 } while (tsp->tv_sec != gd->gd_time_seconds); 979 980 if (delta >= sys_cputimer->freq) { 981 tsp->tv_sec += delta / sys_cputimer->freq; 982 delta %= sys_cputimer->freq; 983 } 984 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 985 } 986 987 /* 988 * realtime routines 989 */ 990 991 void 992 getmicrotime(struct timeval *tvp) 993 { 994 struct globaldata *gd = mycpu; 995 struct timespec *bt; 996 sysclock_t delta; 997 998 do { 999 tvp->tv_sec = gd->gd_time_seconds; 1000 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1001 } while (tvp->tv_sec != gd->gd_time_seconds); 1002 1003 if (delta >= sys_cputimer->freq) { 1004 tvp->tv_sec += delta / sys_cputimer->freq; 1005 delta %= sys_cputimer->freq; 1006 } 1007 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1008 1009 bt = &basetime[basetime_index]; 1010 tvp->tv_sec += bt->tv_sec; 1011 tvp->tv_usec += bt->tv_nsec / 1000; 1012 while (tvp->tv_usec >= 1000000) { 1013 tvp->tv_usec -= 1000000; 1014 ++tvp->tv_sec; 1015 } 1016 } 1017 1018 void 1019 getnanotime(struct timespec *tsp) 1020 { 1021 struct globaldata *gd = mycpu; 1022 struct timespec *bt; 1023 sysclock_t delta; 1024 1025 do { 1026 tsp->tv_sec = gd->gd_time_seconds; 1027 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1028 } while (tsp->tv_sec != gd->gd_time_seconds); 1029 1030 if (delta >= sys_cputimer->freq) { 1031 tsp->tv_sec += delta / sys_cputimer->freq; 1032 delta %= sys_cputimer->freq; 1033 } 1034 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1035 1036 bt = &basetime[basetime_index]; 1037 tsp->tv_sec += bt->tv_sec; 1038 tsp->tv_nsec += bt->tv_nsec; 1039 while (tsp->tv_nsec >= 1000000000) { 1040 tsp->tv_nsec -= 1000000000; 1041 ++tsp->tv_sec; 1042 } 1043 } 1044 1045 static void 1046 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1047 { 1048 struct globaldata *gd = mycpu; 1049 sysclock_t delta; 1050 1051 do { 1052 tsp->tv_sec = gd->gd_time_seconds; 1053 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1054 } while (tsp->tv_sec != gd->gd_time_seconds); 1055 1056 if (delta >= sys_cputimer->freq) { 1057 tsp->tv_sec += delta / sys_cputimer->freq; 1058 delta %= sys_cputimer->freq; 1059 } 1060 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1061 1062 tsp->tv_sec += nbt->tv_sec; 1063 tsp->tv_nsec += nbt->tv_nsec; 1064 while (tsp->tv_nsec >= 1000000000) { 1065 tsp->tv_nsec -= 1000000000; 1066 ++tsp->tv_sec; 1067 } 1068 } 1069 1070 1071 void 1072 microtime(struct timeval *tvp) 1073 { 1074 struct globaldata *gd = mycpu; 1075 struct timespec *bt; 1076 sysclock_t delta; 1077 1078 do { 1079 tvp->tv_sec = gd->gd_time_seconds; 1080 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1081 } while (tvp->tv_sec != gd->gd_time_seconds); 1082 1083 if (delta >= sys_cputimer->freq) { 1084 tvp->tv_sec += delta / sys_cputimer->freq; 1085 delta %= sys_cputimer->freq; 1086 } 1087 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1088 1089 bt = &basetime[basetime_index]; 1090 tvp->tv_sec += bt->tv_sec; 1091 tvp->tv_usec += bt->tv_nsec / 1000; 1092 while (tvp->tv_usec >= 1000000) { 1093 tvp->tv_usec -= 1000000; 1094 ++tvp->tv_sec; 1095 } 1096 } 1097 1098 void 1099 nanotime(struct timespec *tsp) 1100 { 1101 struct globaldata *gd = mycpu; 1102 struct timespec *bt; 1103 sysclock_t delta; 1104 1105 do { 1106 tsp->tv_sec = gd->gd_time_seconds; 1107 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1108 } while (tsp->tv_sec != gd->gd_time_seconds); 1109 1110 if (delta >= sys_cputimer->freq) { 1111 tsp->tv_sec += delta / sys_cputimer->freq; 1112 delta %= sys_cputimer->freq; 1113 } 1114 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1115 1116 bt = &basetime[basetime_index]; 1117 tsp->tv_sec += bt->tv_sec; 1118 tsp->tv_nsec += bt->tv_nsec; 1119 while (tsp->tv_nsec >= 1000000000) { 1120 tsp->tv_nsec -= 1000000000; 1121 ++tsp->tv_sec; 1122 } 1123 } 1124 1125 /* 1126 * note: this is not exactly synchronized with real time. To do that we 1127 * would have to do what microtime does and check for a nanoseconds overflow. 1128 */ 1129 time_t 1130 get_approximate_time_t(void) 1131 { 1132 struct globaldata *gd = mycpu; 1133 struct timespec *bt; 1134 1135 bt = &basetime[basetime_index]; 1136 return(gd->gd_time_seconds + bt->tv_sec); 1137 } 1138 1139 int 1140 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1141 { 1142 pps_params_t *app; 1143 struct pps_fetch_args *fapi; 1144 #ifdef PPS_SYNC 1145 struct pps_kcbind_args *kapi; 1146 #endif 1147 1148 switch (cmd) { 1149 case PPS_IOC_CREATE: 1150 return (0); 1151 case PPS_IOC_DESTROY: 1152 return (0); 1153 case PPS_IOC_SETPARAMS: 1154 app = (pps_params_t *)data; 1155 if (app->mode & ~pps->ppscap) 1156 return (EINVAL); 1157 pps->ppsparam = *app; 1158 return (0); 1159 case PPS_IOC_GETPARAMS: 1160 app = (pps_params_t *)data; 1161 *app = pps->ppsparam; 1162 app->api_version = PPS_API_VERS_1; 1163 return (0); 1164 case PPS_IOC_GETCAP: 1165 *(int*)data = pps->ppscap; 1166 return (0); 1167 case PPS_IOC_FETCH: 1168 fapi = (struct pps_fetch_args *)data; 1169 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1170 return (EINVAL); 1171 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1172 return (EOPNOTSUPP); 1173 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1174 fapi->pps_info_buf = pps->ppsinfo; 1175 return (0); 1176 case PPS_IOC_KCBIND: 1177 #ifdef PPS_SYNC 1178 kapi = (struct pps_kcbind_args *)data; 1179 /* XXX Only root should be able to do this */ 1180 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1181 return (EINVAL); 1182 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1183 return (EINVAL); 1184 if (kapi->edge & ~pps->ppscap) 1185 return (EINVAL); 1186 pps->kcmode = kapi->edge; 1187 return (0); 1188 #else 1189 return (EOPNOTSUPP); 1190 #endif 1191 default: 1192 return (ENOTTY); 1193 } 1194 } 1195 1196 void 1197 pps_init(struct pps_state *pps) 1198 { 1199 pps->ppscap |= PPS_TSFMT_TSPEC; 1200 if (pps->ppscap & PPS_CAPTUREASSERT) 1201 pps->ppscap |= PPS_OFFSETASSERT; 1202 if (pps->ppscap & PPS_CAPTURECLEAR) 1203 pps->ppscap |= PPS_OFFSETCLEAR; 1204 } 1205 1206 void 1207 pps_event(struct pps_state *pps, sysclock_t count, int event) 1208 { 1209 struct globaldata *gd; 1210 struct timespec *tsp; 1211 struct timespec *osp; 1212 struct timespec *bt; 1213 struct timespec ts; 1214 sysclock_t *pcount; 1215 #ifdef PPS_SYNC 1216 sysclock_t tcount; 1217 #endif 1218 sysclock_t delta; 1219 pps_seq_t *pseq; 1220 int foff; 1221 int fhard; 1222 1223 gd = mycpu; 1224 1225 /* Things would be easier with arrays... */ 1226 if (event == PPS_CAPTUREASSERT) { 1227 tsp = &pps->ppsinfo.assert_timestamp; 1228 osp = &pps->ppsparam.assert_offset; 1229 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1230 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1231 pcount = &pps->ppscount[0]; 1232 pseq = &pps->ppsinfo.assert_sequence; 1233 } else { 1234 tsp = &pps->ppsinfo.clear_timestamp; 1235 osp = &pps->ppsparam.clear_offset; 1236 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1237 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1238 pcount = &pps->ppscount[1]; 1239 pseq = &pps->ppsinfo.clear_sequence; 1240 } 1241 1242 /* Nothing really happened */ 1243 if (*pcount == count) 1244 return; 1245 1246 *pcount = count; 1247 1248 do { 1249 ts.tv_sec = gd->gd_time_seconds; 1250 delta = count - gd->gd_cpuclock_base; 1251 } while (ts.tv_sec != gd->gd_time_seconds); 1252 1253 if (delta >= sys_cputimer->freq) { 1254 ts.tv_sec += delta / sys_cputimer->freq; 1255 delta %= sys_cputimer->freq; 1256 } 1257 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1258 bt = &basetime[basetime_index]; 1259 ts.tv_sec += bt->tv_sec; 1260 ts.tv_nsec += bt->tv_nsec; 1261 while (ts.tv_nsec >= 1000000000) { 1262 ts.tv_nsec -= 1000000000; 1263 ++ts.tv_sec; 1264 } 1265 1266 (*pseq)++; 1267 *tsp = ts; 1268 1269 if (foff) { 1270 timespecadd(tsp, osp); 1271 if (tsp->tv_nsec < 0) { 1272 tsp->tv_nsec += 1000000000; 1273 tsp->tv_sec -= 1; 1274 } 1275 } 1276 #ifdef PPS_SYNC 1277 if (fhard) { 1278 /* magic, at its best... */ 1279 tcount = count - pps->ppscount[2]; 1280 pps->ppscount[2] = count; 1281 if (tcount >= sys_cputimer->freq) { 1282 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1283 sys_cputimer->freq64_nsec * 1284 (tcount % sys_cputimer->freq)) >> 32; 1285 } else { 1286 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1287 } 1288 hardpps(tsp, delta); 1289 } 1290 #endif 1291 } 1292 1293