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 1.59 2007/06/30 21:52:19 swildner Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 #include "opt_polling.h" 78 #include "opt_pctrack.h" 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/callout.h> 83 #include <sys/kernel.h> 84 #include <sys/kinfo.h> 85 #include <sys/proc.h> 86 #include <sys/malloc.h> 87 #include <sys/resourcevar.h> 88 #include <sys/signalvar.h> 89 #include <sys/timex.h> 90 #include <sys/timepps.h> 91 #include <vm/vm.h> 92 #include <sys/lock.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_extern.h> 96 #include <sys/sysctl.h> 97 #include <sys/thread2.h> 98 99 #include <machine/cpu.h> 100 #include <machine/limits.h> 101 #include <machine/smp.h> 102 103 #ifdef GPROF 104 #include <sys/gmon.h> 105 #endif 106 107 #ifdef DEVICE_POLLING 108 extern void init_device_poll(void); 109 #endif 110 111 #ifdef DEBUG_PCTRACK 112 static void do_pctrack(struct intrframe *frame, int which); 113 #endif 114 115 static void initclocks (void *dummy); 116 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 117 118 /* 119 * Some of these don't belong here, but it's easiest to concentrate them. 120 * Note that cpu_time counts in microseconds, but most userland programs 121 * just compare relative times against the total by delta. 122 */ 123 struct kinfo_cputime cputime_percpu[MAXCPU]; 124 #ifdef DEBUG_PCTRACK 125 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE }; 126 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE]; 127 #endif 128 129 #ifdef SMP 130 static int 131 sysctl_cputime(SYSCTL_HANDLER_ARGS) 132 { 133 int cpu, error = 0; 134 size_t size = sizeof(struct kinfo_cputime); 135 136 for (cpu = 0; cpu < ncpus; ++cpu) { 137 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 138 break; 139 } 140 141 return (error); 142 } 143 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 144 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 145 #else 146 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 147 "CPU time statistics"); 148 #endif 149 150 /* 151 * boottime is used to calculate the 'real' uptime. Do not confuse this with 152 * microuptime(). microtime() is not drift compensated. The real uptime 153 * with compensation is nanotime() - bootime. boottime is recalculated 154 * whenever the real time is set based on the compensated elapsed time 155 * in seconds (gd->gd_time_seconds). 156 * 157 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 158 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 159 * the real time. 160 */ 161 struct timespec boottime; /* boot time (realtime) for reference only */ 162 time_t time_second; /* read-only 'passive' uptime in seconds */ 163 164 /* 165 * basetime is used to calculate the compensated real time of day. The 166 * basetime can be modified on a per-tick basis by the adjtime(), 167 * ntp_adjtime(), and sysctl-based time correction APIs. 168 * 169 * Note that frequency corrections can also be made by adjusting 170 * gd_cpuclock_base. 171 * 172 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 173 * used on both SMP and UP systems to avoid MP races between cpu's and 174 * interrupt races on UP systems. 175 */ 176 #define BASETIME_ARYSIZE 16 177 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 178 static struct timespec basetime[BASETIME_ARYSIZE]; 179 static volatile int basetime_index; 180 181 static int 182 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 183 { 184 struct timespec *bt; 185 int error; 186 int index; 187 188 /* 189 * Because basetime data and index may be updated by another cpu, 190 * a load fence is required to ensure that the data we read has 191 * not been speculatively read relative to a possibly updated index. 192 */ 193 index = basetime_index; 194 cpu_lfence(); 195 bt = &basetime[index]; 196 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 197 return (error); 198 } 199 200 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 201 &boottime, timespec, "System boottime"); 202 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 203 sysctl_get_basetime, "S,timespec", "System basetime"); 204 205 static void hardclock(systimer_t info, struct intrframe *frame); 206 static void statclock(systimer_t info, struct intrframe *frame); 207 static void schedclock(systimer_t info, struct intrframe *frame); 208 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 209 210 int ticks; /* system master ticks at hz */ 211 int clocks_running; /* tsleep/timeout clocks operational */ 212 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 213 int64_t nsec_acc; /* accumulator */ 214 215 /* NTPD time correction fields */ 216 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 217 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 218 int64_t ntp_delta; /* one-time correction in nsec */ 219 int64_t ntp_big_delta = 1000000000; 220 int32_t ntp_tick_delta; /* current adjustment rate */ 221 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 222 time_t ntp_leap_second; /* time of next leap second */ 223 int ntp_leap_insert; /* whether to insert or remove a second */ 224 225 /* 226 * Finish initializing clock frequencies and start all clocks running. 227 */ 228 /* ARGSUSED*/ 229 static void 230 initclocks(void *dummy) 231 { 232 #ifdef DEVICE_POLLING 233 init_device_poll(); 234 #endif 235 /*psratio = profhz / stathz;*/ 236 initclocks_pcpu(); 237 clocks_running = 1; 238 } 239 240 /* 241 * Called on a per-cpu basis 242 */ 243 void 244 initclocks_pcpu(void) 245 { 246 struct globaldata *gd = mycpu; 247 248 crit_enter(); 249 if (gd->gd_cpuid == 0) { 250 gd->gd_time_seconds = 1; 251 gd->gd_cpuclock_base = sys_cputimer->count(); 252 } else { 253 /* XXX */ 254 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 255 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 256 } 257 258 /* 259 * Use a non-queued periodic systimer to prevent multiple ticks from 260 * building up if the sysclock jumps forward (8254 gets reset). The 261 * sysclock will never jump backwards. Our time sync is based on 262 * the actual sysclock, not the ticks count. 263 */ 264 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 265 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 266 /* XXX correct the frequency for scheduler / estcpu tests */ 267 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 268 NULL, ESTCPUFREQ); 269 crit_exit(); 270 } 271 272 /* 273 * This sets the current real time of day. Timespecs are in seconds and 274 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 275 * instead we adjust basetime so basetime + gd_* results in the current 276 * time of day. This way the gd_* fields are guarenteed to represent 277 * a monotonically increasing 'uptime' value. 278 * 279 * When set_timeofday() is called from userland, the system call forces it 280 * onto cpu #0 since only cpu #0 can update basetime_index. 281 */ 282 void 283 set_timeofday(struct timespec *ts) 284 { 285 struct timespec *nbt; 286 int ni; 287 288 /* 289 * XXX SMP / non-atomic basetime updates 290 */ 291 crit_enter(); 292 ni = (basetime_index + 1) & BASETIME_ARYMASK; 293 nbt = &basetime[ni]; 294 nanouptime(nbt); 295 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 296 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 297 if (nbt->tv_nsec < 0) { 298 nbt->tv_nsec += 1000000000; 299 --nbt->tv_sec; 300 } 301 302 /* 303 * Note that basetime diverges from boottime as the clock drift is 304 * compensated for, so we cannot do away with boottime. When setting 305 * the absolute time of day the drift is 0 (for an instant) and we 306 * can simply assign boottime to basetime. 307 * 308 * Note that nanouptime() is based on gd_time_seconds which is drift 309 * compensated up to a point (it is guarenteed to remain monotonically 310 * increasing). gd_time_seconds is thus our best uptime guess and 311 * suitable for use in the boottime calculation. It is already taken 312 * into account in the basetime calculation above. 313 */ 314 boottime.tv_sec = nbt->tv_sec; 315 ntp_delta = 0; 316 317 /* 318 * We now have a new basetime, make sure all other cpus have it, 319 * then update the index. 320 */ 321 cpu_sfence(); 322 basetime_index = ni; 323 324 crit_exit(); 325 } 326 327 /* 328 * Each cpu has its own hardclock, but we only increments ticks and softticks 329 * on cpu #0. 330 * 331 * NOTE! systimer! the MP lock might not be held here. We can only safely 332 * manipulate objects owned by the current cpu. 333 */ 334 static void 335 hardclock(systimer_t info, struct intrframe *frame) 336 { 337 sysclock_t cputicks; 338 struct proc *p; 339 struct globaldata *gd = mycpu; 340 341 /* 342 * Realtime updates are per-cpu. Note that timer corrections as 343 * returned by microtime() and friends make an additional adjustment 344 * using a system-wise 'basetime', but the running time is always 345 * taken from the per-cpu globaldata area. Since the same clock 346 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 347 * stay in synch. 348 * 349 * Note that we never allow info->time (aka gd->gd_hardclock.time) 350 * to reverse index gd_cpuclock_base, but that it is possible for 351 * it to temporarily get behind in the seconds if something in the 352 * system locks interrupts for a long period of time. Since periodic 353 * timers count events, though everything should resynch again 354 * immediately. 355 */ 356 cputicks = info->time - gd->gd_cpuclock_base; 357 if (cputicks >= sys_cputimer->freq) { 358 ++gd->gd_time_seconds; 359 gd->gd_cpuclock_base += sys_cputimer->freq; 360 } 361 362 /* 363 * The system-wide ticks counter and NTP related timedelta/tickdelta 364 * adjustments only occur on cpu #0. NTP adjustments are accomplished 365 * by updating basetime. 366 */ 367 if (gd->gd_cpuid == 0) { 368 struct timespec *nbt; 369 struct timespec nts; 370 int leap; 371 int ni; 372 373 ++ticks; 374 375 #if 0 376 if (tco->tc_poll_pps) 377 tco->tc_poll_pps(tco); 378 #endif 379 380 /* 381 * Calculate the new basetime index. We are in a critical section 382 * on cpu #0 and can safely play with basetime_index. Start 383 * with the current basetime and then make adjustments. 384 */ 385 ni = (basetime_index + 1) & BASETIME_ARYMASK; 386 nbt = &basetime[ni]; 387 *nbt = basetime[basetime_index]; 388 389 /* 390 * Apply adjtime corrections. (adjtime() API) 391 * 392 * adjtime() only runs on cpu #0 so our critical section is 393 * sufficient to access these variables. 394 */ 395 if (ntp_delta != 0) { 396 nbt->tv_nsec += ntp_tick_delta; 397 ntp_delta -= ntp_tick_delta; 398 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 399 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 400 ntp_tick_delta = ntp_delta; 401 } 402 } 403 404 /* 405 * Apply permanent frequency corrections. (sysctl API) 406 */ 407 if (ntp_tick_permanent != 0) { 408 ntp_tick_acc += ntp_tick_permanent; 409 if (ntp_tick_acc >= (1LL << 32)) { 410 nbt->tv_nsec += ntp_tick_acc >> 32; 411 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 412 } else if (ntp_tick_acc <= -(1LL << 32)) { 413 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 414 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 415 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 416 } 417 } 418 419 if (nbt->tv_nsec >= 1000000000) { 420 nbt->tv_sec++; 421 nbt->tv_nsec -= 1000000000; 422 } else if (nbt->tv_nsec < 0) { 423 nbt->tv_sec--; 424 nbt->tv_nsec += 1000000000; 425 } 426 427 /* 428 * Another per-tick compensation. (for ntp_adjtime() API) 429 */ 430 if (nsec_adj != 0) { 431 nsec_acc += nsec_adj; 432 if (nsec_acc >= 0x100000000LL) { 433 nbt->tv_nsec += nsec_acc >> 32; 434 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 435 } else if (nsec_acc <= -0x100000000LL) { 436 nbt->tv_nsec -= -nsec_acc >> 32; 437 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 438 } 439 if (nbt->tv_nsec >= 1000000000) { 440 nbt->tv_nsec -= 1000000000; 441 ++nbt->tv_sec; 442 } else if (nbt->tv_nsec < 0) { 443 nbt->tv_nsec += 1000000000; 444 --nbt->tv_sec; 445 } 446 } 447 448 /************************************************************ 449 * LEAP SECOND CORRECTION * 450 ************************************************************ 451 * 452 * Taking into account all the corrections made above, figure 453 * out the new real time. If the seconds field has changed 454 * then apply any pending leap-second corrections. 455 */ 456 getnanotime_nbt(nbt, &nts); 457 458 if (time_second != nts.tv_sec) { 459 /* 460 * Apply leap second (sysctl API). Adjust nts for changes 461 * so we do not have to call getnanotime_nbt again. 462 */ 463 if (ntp_leap_second) { 464 if (ntp_leap_second == nts.tv_sec) { 465 if (ntp_leap_insert) { 466 nbt->tv_sec++; 467 nts.tv_sec++; 468 } else { 469 nbt->tv_sec--; 470 nts.tv_sec--; 471 } 472 ntp_leap_second--; 473 } 474 } 475 476 /* 477 * Apply leap second (ntp_adjtime() API), calculate a new 478 * nsec_adj field. ntp_update_second() returns nsec_adj 479 * as a per-second value but we need it as a per-tick value. 480 */ 481 leap = ntp_update_second(time_second, &nsec_adj); 482 nsec_adj /= hz; 483 nbt->tv_sec += leap; 484 nts.tv_sec += leap; 485 486 /* 487 * Update the time_second 'approximate time' global. 488 */ 489 time_second = nts.tv_sec; 490 } 491 492 /* 493 * Finally, our new basetime is ready to go live! 494 */ 495 cpu_sfence(); 496 basetime_index = ni; 497 498 /* 499 * Figure out how badly the system is starved for memory 500 */ 501 vm_fault_ratecheck(); 502 } 503 504 /* 505 * softticks are handled for all cpus 506 */ 507 hardclock_softtick(gd); 508 509 /* 510 * ITimer handling is per-tick, per-cpu. I don't think ksignal() 511 * is mpsafe on curproc, so XXX get the mplock. 512 */ 513 if ((p = curproc) != NULL && try_mplock()) { 514 if (frame && CLKF_USERMODE(frame) && 515 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 516 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 517 ksignal(p, SIGVTALRM); 518 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 519 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 520 ksignal(p, SIGPROF); 521 rel_mplock(); 522 } 523 setdelayed(); 524 } 525 526 /* 527 * The statistics clock typically runs at a 125Hz rate, and is intended 528 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 529 * 530 * NOTE! systimer! the MP lock might not be held here. We can only safely 531 * manipulate objects owned by the current cpu. 532 * 533 * The stats clock is responsible for grabbing a profiling sample. 534 * Most of the statistics are only used by user-level statistics programs. 535 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 536 * p->p_estcpu. 537 * 538 * Like the other clocks, the stat clock is called from what is effectively 539 * a fast interrupt, so the context should be the thread/process that got 540 * interrupted. 541 */ 542 static void 543 statclock(systimer_t info, struct intrframe *frame) 544 { 545 #ifdef GPROF 546 struct gmonparam *g; 547 int i; 548 #endif 549 thread_t td; 550 struct proc *p; 551 int bump; 552 struct timeval tv; 553 struct timeval *stv; 554 555 /* 556 * How big was our timeslice relative to the last time? 557 */ 558 microuptime(&tv); /* mpsafe */ 559 stv = &mycpu->gd_stattv; 560 if (stv->tv_sec == 0) { 561 bump = 1; 562 } else { 563 bump = tv.tv_usec - stv->tv_usec + 564 (tv.tv_sec - stv->tv_sec) * 1000000; 565 if (bump < 0) 566 bump = 0; 567 if (bump > 1000000) 568 bump = 1000000; 569 } 570 *stv = tv; 571 572 td = curthread; 573 p = td->td_proc; 574 575 if (frame && CLKF_USERMODE(frame)) { 576 /* 577 * Came from userland, handle user time and deal with 578 * possible process. 579 */ 580 if (p && (p->p_flag & P_PROFIL)) 581 addupc_intr(p, CLKF_PC(frame), 1); 582 td->td_uticks += bump; 583 584 /* 585 * Charge the time as appropriate 586 */ 587 if (p && p->p_nice > NZERO) 588 cpu_time.cp_nice += bump; 589 else 590 cpu_time.cp_user += bump; 591 } else { 592 #ifdef GPROF 593 /* 594 * Kernel statistics are just like addupc_intr, only easier. 595 */ 596 g = &_gmonparam; 597 if (g->state == GMON_PROF_ON && frame) { 598 i = CLKF_PC(frame) - g->lowpc; 599 if (i < g->textsize) { 600 i /= HISTFRACTION * sizeof(*g->kcount); 601 g->kcount[i]++; 602 } 603 } 604 #endif 605 /* 606 * Came from kernel mode, so we were: 607 * - handling an interrupt, 608 * - doing syscall or trap work on behalf of the current 609 * user process, or 610 * - spinning in the idle loop. 611 * Whichever it is, charge the time as appropriate. 612 * Note that we charge interrupts to the current process, 613 * regardless of whether they are ``for'' that process, 614 * so that we know how much of its real time was spent 615 * in ``non-process'' (i.e., interrupt) work. 616 * 617 * XXX assume system if frame is NULL. A NULL frame 618 * can occur if ipi processing is done from a crit_exit(). 619 */ 620 if (frame && CLKF_INTR(frame)) 621 td->td_iticks += bump; 622 else 623 td->td_sticks += bump; 624 625 if (frame && CLKF_INTR(frame)) { 626 #ifdef DEBUG_PCTRACK 627 do_pctrack(frame, PCTRACK_INT); 628 #endif 629 cpu_time.cp_intr += bump; 630 } else { 631 if (td == &mycpu->gd_idlethread) { 632 cpu_time.cp_idle += bump; 633 } else { 634 #ifdef DEBUG_PCTRACK 635 if (frame) 636 do_pctrack(frame, PCTRACK_SYS); 637 #endif 638 cpu_time.cp_sys += bump; 639 } 640 } 641 } 642 } 643 644 #ifdef DEBUG_PCTRACK 645 /* 646 * Sample the PC when in the kernel or in an interrupt. User code can 647 * retrieve the information and generate a histogram or other output. 648 */ 649 650 static void 651 do_pctrack(struct intrframe *frame, int which) 652 { 653 struct kinfo_pctrack *pctrack; 654 655 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 656 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 657 (void *)CLKF_PC(frame); 658 ++pctrack->pc_index; 659 } 660 661 static int 662 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 663 { 664 struct kinfo_pcheader head; 665 int error; 666 int cpu; 667 int ntrack; 668 669 head.pc_ntrack = PCTRACK_SIZE; 670 head.pc_arysize = PCTRACK_ARYSIZE; 671 672 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 673 return (error); 674 675 for (cpu = 0; cpu < ncpus; ++cpu) { 676 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 677 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 678 sizeof(struct kinfo_pctrack)); 679 if (error) 680 break; 681 } 682 if (error) 683 break; 684 } 685 return (error); 686 } 687 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 688 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 689 690 #endif 691 692 /* 693 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 694 * the MP lock might not be held. We can safely manipulate parts of curproc 695 * but that's about it. 696 * 697 * Each cpu has its own scheduler clock. 698 */ 699 static void 700 schedclock(systimer_t info, struct intrframe *frame) 701 { 702 struct lwp *lp; 703 struct rusage *ru; 704 struct vmspace *vm; 705 long rss; 706 707 if ((lp = lwkt_preempted_proc()) != NULL) { 708 /* 709 * Account for cpu time used and hit the scheduler. Note 710 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 711 * HERE. 712 */ 713 ++lp->lwp_cpticks; 714 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 715 info->time); 716 } 717 if ((lp = curthread->td_lwp) != NULL) { 718 /* 719 * Update resource usage integrals and maximums. 720 */ 721 if ((ru = &lp->lwp_proc->p_ru) && 722 (vm = lp->lwp_proc->p_vmspace) != NULL) { 723 ru->ru_ixrss += pgtok(vm->vm_tsize); 724 ru->ru_idrss += pgtok(vm->vm_dsize); 725 ru->ru_isrss += pgtok(vm->vm_ssize); 726 rss = pgtok(vmspace_resident_count(vm)); 727 if (ru->ru_maxrss < rss) 728 ru->ru_maxrss = rss; 729 } 730 } 731 } 732 733 /* 734 * Compute number of ticks for the specified amount of time. The 735 * return value is intended to be used in a clock interrupt timed 736 * operation and guarenteed to meet or exceed the requested time. 737 * If the representation overflows, return INT_MAX. The minimum return 738 * value is 1 ticks and the function will average the calculation up. 739 * If any value greater then 0 microseconds is supplied, a value 740 * of at least 2 will be returned to ensure that a near-term clock 741 * interrupt does not cause the timeout to occur (degenerately) early. 742 * 743 * Note that limit checks must take into account microseconds, which is 744 * done simply by using the smaller signed long maximum instead of 745 * the unsigned long maximum. 746 * 747 * If ints have 32 bits, then the maximum value for any timeout in 748 * 10ms ticks is 248 days. 749 */ 750 int 751 tvtohz_high(struct timeval *tv) 752 { 753 int ticks; 754 long sec, usec; 755 756 sec = tv->tv_sec; 757 usec = tv->tv_usec; 758 if (usec < 0) { 759 sec--; 760 usec += 1000000; 761 } 762 if (sec < 0) { 763 #ifdef DIAGNOSTIC 764 if (usec > 0) { 765 sec++; 766 usec -= 1000000; 767 } 768 kprintf("tvtohz_high: negative time difference %ld sec %ld usec\n", 769 sec, usec); 770 #endif 771 ticks = 1; 772 } else if (sec <= INT_MAX / hz) { 773 ticks = (int)(sec * hz + 774 ((u_long)usec + (tick - 1)) / tick) + 1; 775 } else { 776 ticks = INT_MAX; 777 } 778 return (ticks); 779 } 780 781 /* 782 * Compute number of ticks for the specified amount of time, erroring on 783 * the side of it being too low to ensure that sleeping the returned number 784 * of ticks will not result in a late return. 785 * 786 * The supplied timeval may not be negative and should be normalized. A 787 * return value of 0 is possible if the timeval converts to less then 788 * 1 tick. 789 * 790 * If ints have 32 bits, then the maximum value for any timeout in 791 * 10ms ticks is 248 days. 792 */ 793 int 794 tvtohz_low(struct timeval *tv) 795 { 796 int ticks; 797 long sec; 798 799 sec = tv->tv_sec; 800 if (sec <= INT_MAX / hz) 801 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 802 else 803 ticks = INT_MAX; 804 return (ticks); 805 } 806 807 808 /* 809 * Start profiling on a process. 810 * 811 * Kernel profiling passes proc0 which never exits and hence 812 * keeps the profile clock running constantly. 813 */ 814 void 815 startprofclock(struct proc *p) 816 { 817 if ((p->p_flag & P_PROFIL) == 0) { 818 p->p_flag |= P_PROFIL; 819 #if 0 /* XXX */ 820 if (++profprocs == 1 && stathz != 0) { 821 crit_enter(); 822 psdiv = psratio; 823 setstatclockrate(profhz); 824 crit_exit(); 825 } 826 #endif 827 } 828 } 829 830 /* 831 * Stop profiling on a process. 832 */ 833 void 834 stopprofclock(struct proc *p) 835 { 836 if (p->p_flag & P_PROFIL) { 837 p->p_flag &= ~P_PROFIL; 838 #if 0 /* XXX */ 839 if (--profprocs == 0 && stathz != 0) { 840 crit_enter(); 841 psdiv = 1; 842 setstatclockrate(stathz); 843 crit_exit(); 844 } 845 #endif 846 } 847 } 848 849 /* 850 * Return information about system clocks. 851 */ 852 static int 853 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 854 { 855 struct kinfo_clockinfo clkinfo; 856 /* 857 * Construct clockinfo structure. 858 */ 859 clkinfo.ci_hz = hz; 860 clkinfo.ci_tick = tick; 861 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 862 clkinfo.ci_profhz = profhz; 863 clkinfo.ci_stathz = stathz ? stathz : hz; 864 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 865 } 866 867 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 868 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 869 870 /* 871 * We have eight functions for looking at the clock, four for 872 * microseconds and four for nanoseconds. For each there is fast 873 * but less precise version "get{nano|micro}[up]time" which will 874 * return a time which is up to 1/HZ previous to the call, whereas 875 * the raw version "{nano|micro}[up]time" will return a timestamp 876 * which is as precise as possible. The "up" variants return the 877 * time relative to system boot, these are well suited for time 878 * interval measurements. 879 * 880 * Each cpu independantly maintains the current time of day, so all 881 * we need to do to protect ourselves from changes is to do a loop 882 * check on the seconds field changing out from under us. 883 * 884 * The system timer maintains a 32 bit count and due to various issues 885 * it is possible for the calculated delta to occassionally exceed 886 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 887 * multiplication can easily overflow, so we deal with the case. For 888 * uniformity we deal with the case in the usec case too. 889 */ 890 void 891 getmicrouptime(struct timeval *tvp) 892 { 893 struct globaldata *gd = mycpu; 894 sysclock_t delta; 895 896 do { 897 tvp->tv_sec = gd->gd_time_seconds; 898 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 899 } while (tvp->tv_sec != gd->gd_time_seconds); 900 901 if (delta >= sys_cputimer->freq) { 902 tvp->tv_sec += delta / sys_cputimer->freq; 903 delta %= sys_cputimer->freq; 904 } 905 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 906 if (tvp->tv_usec >= 1000000) { 907 tvp->tv_usec -= 1000000; 908 ++tvp->tv_sec; 909 } 910 } 911 912 void 913 getnanouptime(struct timespec *tsp) 914 { 915 struct globaldata *gd = mycpu; 916 sysclock_t delta; 917 918 do { 919 tsp->tv_sec = gd->gd_time_seconds; 920 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 921 } while (tsp->tv_sec != gd->gd_time_seconds); 922 923 if (delta >= sys_cputimer->freq) { 924 tsp->tv_sec += delta / sys_cputimer->freq; 925 delta %= sys_cputimer->freq; 926 } 927 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 928 } 929 930 void 931 microuptime(struct timeval *tvp) 932 { 933 struct globaldata *gd = mycpu; 934 sysclock_t delta; 935 936 do { 937 tvp->tv_sec = gd->gd_time_seconds; 938 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 939 } while (tvp->tv_sec != gd->gd_time_seconds); 940 941 if (delta >= sys_cputimer->freq) { 942 tvp->tv_sec += delta / sys_cputimer->freq; 943 delta %= sys_cputimer->freq; 944 } 945 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 946 } 947 948 void 949 nanouptime(struct timespec *tsp) 950 { 951 struct globaldata *gd = mycpu; 952 sysclock_t delta; 953 954 do { 955 tsp->tv_sec = gd->gd_time_seconds; 956 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 957 } while (tsp->tv_sec != gd->gd_time_seconds); 958 959 if (delta >= sys_cputimer->freq) { 960 tsp->tv_sec += delta / sys_cputimer->freq; 961 delta %= sys_cputimer->freq; 962 } 963 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 964 } 965 966 /* 967 * realtime routines 968 */ 969 970 void 971 getmicrotime(struct timeval *tvp) 972 { 973 struct globaldata *gd = mycpu; 974 struct timespec *bt; 975 sysclock_t delta; 976 977 do { 978 tvp->tv_sec = gd->gd_time_seconds; 979 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 980 } while (tvp->tv_sec != gd->gd_time_seconds); 981 982 if (delta >= sys_cputimer->freq) { 983 tvp->tv_sec += delta / sys_cputimer->freq; 984 delta %= sys_cputimer->freq; 985 } 986 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 987 988 bt = &basetime[basetime_index]; 989 tvp->tv_sec += bt->tv_sec; 990 tvp->tv_usec += bt->tv_nsec / 1000; 991 while (tvp->tv_usec >= 1000000) { 992 tvp->tv_usec -= 1000000; 993 ++tvp->tv_sec; 994 } 995 } 996 997 void 998 getnanotime(struct timespec *tsp) 999 { 1000 struct globaldata *gd = mycpu; 1001 struct timespec *bt; 1002 sysclock_t delta; 1003 1004 do { 1005 tsp->tv_sec = gd->gd_time_seconds; 1006 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1007 } while (tsp->tv_sec != gd->gd_time_seconds); 1008 1009 if (delta >= sys_cputimer->freq) { 1010 tsp->tv_sec += delta / sys_cputimer->freq; 1011 delta %= sys_cputimer->freq; 1012 } 1013 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1014 1015 bt = &basetime[basetime_index]; 1016 tsp->tv_sec += bt->tv_sec; 1017 tsp->tv_nsec += bt->tv_nsec; 1018 while (tsp->tv_nsec >= 1000000000) { 1019 tsp->tv_nsec -= 1000000000; 1020 ++tsp->tv_sec; 1021 } 1022 } 1023 1024 static void 1025 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1026 { 1027 struct globaldata *gd = mycpu; 1028 sysclock_t delta; 1029 1030 do { 1031 tsp->tv_sec = gd->gd_time_seconds; 1032 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1033 } while (tsp->tv_sec != gd->gd_time_seconds); 1034 1035 if (delta >= sys_cputimer->freq) { 1036 tsp->tv_sec += delta / sys_cputimer->freq; 1037 delta %= sys_cputimer->freq; 1038 } 1039 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1040 1041 tsp->tv_sec += nbt->tv_sec; 1042 tsp->tv_nsec += nbt->tv_nsec; 1043 while (tsp->tv_nsec >= 1000000000) { 1044 tsp->tv_nsec -= 1000000000; 1045 ++tsp->tv_sec; 1046 } 1047 } 1048 1049 1050 void 1051 microtime(struct timeval *tvp) 1052 { 1053 struct globaldata *gd = mycpu; 1054 struct timespec *bt; 1055 sysclock_t delta; 1056 1057 do { 1058 tvp->tv_sec = gd->gd_time_seconds; 1059 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1060 } while (tvp->tv_sec != gd->gd_time_seconds); 1061 1062 if (delta >= sys_cputimer->freq) { 1063 tvp->tv_sec += delta / sys_cputimer->freq; 1064 delta %= sys_cputimer->freq; 1065 } 1066 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1067 1068 bt = &basetime[basetime_index]; 1069 tvp->tv_sec += bt->tv_sec; 1070 tvp->tv_usec += bt->tv_nsec / 1000; 1071 while (tvp->tv_usec >= 1000000) { 1072 tvp->tv_usec -= 1000000; 1073 ++tvp->tv_sec; 1074 } 1075 } 1076 1077 void 1078 nanotime(struct timespec *tsp) 1079 { 1080 struct globaldata *gd = mycpu; 1081 struct timespec *bt; 1082 sysclock_t delta; 1083 1084 do { 1085 tsp->tv_sec = gd->gd_time_seconds; 1086 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1087 } while (tsp->tv_sec != gd->gd_time_seconds); 1088 1089 if (delta >= sys_cputimer->freq) { 1090 tsp->tv_sec += delta / sys_cputimer->freq; 1091 delta %= sys_cputimer->freq; 1092 } 1093 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1094 1095 bt = &basetime[basetime_index]; 1096 tsp->tv_sec += bt->tv_sec; 1097 tsp->tv_nsec += bt->tv_nsec; 1098 while (tsp->tv_nsec >= 1000000000) { 1099 tsp->tv_nsec -= 1000000000; 1100 ++tsp->tv_sec; 1101 } 1102 } 1103 1104 /* 1105 * note: this is not exactly synchronized with real time. To do that we 1106 * would have to do what microtime does and check for a nanoseconds overflow. 1107 */ 1108 time_t 1109 get_approximate_time_t(void) 1110 { 1111 struct globaldata *gd = mycpu; 1112 struct timespec *bt; 1113 1114 bt = &basetime[basetime_index]; 1115 return(gd->gd_time_seconds + bt->tv_sec); 1116 } 1117 1118 int 1119 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1120 { 1121 pps_params_t *app; 1122 struct pps_fetch_args *fapi; 1123 #ifdef PPS_SYNC 1124 struct pps_kcbind_args *kapi; 1125 #endif 1126 1127 switch (cmd) { 1128 case PPS_IOC_CREATE: 1129 return (0); 1130 case PPS_IOC_DESTROY: 1131 return (0); 1132 case PPS_IOC_SETPARAMS: 1133 app = (pps_params_t *)data; 1134 if (app->mode & ~pps->ppscap) 1135 return (EINVAL); 1136 pps->ppsparam = *app; 1137 return (0); 1138 case PPS_IOC_GETPARAMS: 1139 app = (pps_params_t *)data; 1140 *app = pps->ppsparam; 1141 app->api_version = PPS_API_VERS_1; 1142 return (0); 1143 case PPS_IOC_GETCAP: 1144 *(int*)data = pps->ppscap; 1145 return (0); 1146 case PPS_IOC_FETCH: 1147 fapi = (struct pps_fetch_args *)data; 1148 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1149 return (EINVAL); 1150 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1151 return (EOPNOTSUPP); 1152 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1153 fapi->pps_info_buf = pps->ppsinfo; 1154 return (0); 1155 case PPS_IOC_KCBIND: 1156 #ifdef PPS_SYNC 1157 kapi = (struct pps_kcbind_args *)data; 1158 /* XXX Only root should be able to do this */ 1159 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1160 return (EINVAL); 1161 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1162 return (EINVAL); 1163 if (kapi->edge & ~pps->ppscap) 1164 return (EINVAL); 1165 pps->kcmode = kapi->edge; 1166 return (0); 1167 #else 1168 return (EOPNOTSUPP); 1169 #endif 1170 default: 1171 return (ENOTTY); 1172 } 1173 } 1174 1175 void 1176 pps_init(struct pps_state *pps) 1177 { 1178 pps->ppscap |= PPS_TSFMT_TSPEC; 1179 if (pps->ppscap & PPS_CAPTUREASSERT) 1180 pps->ppscap |= PPS_OFFSETASSERT; 1181 if (pps->ppscap & PPS_CAPTURECLEAR) 1182 pps->ppscap |= PPS_OFFSETCLEAR; 1183 } 1184 1185 void 1186 pps_event(struct pps_state *pps, sysclock_t count, int event) 1187 { 1188 struct globaldata *gd; 1189 struct timespec *tsp; 1190 struct timespec *osp; 1191 struct timespec *bt; 1192 struct timespec ts; 1193 sysclock_t *pcount; 1194 #ifdef PPS_SYNC 1195 sysclock_t tcount; 1196 #endif 1197 sysclock_t delta; 1198 pps_seq_t *pseq; 1199 int foff; 1200 int fhard; 1201 1202 gd = mycpu; 1203 1204 /* Things would be easier with arrays... */ 1205 if (event == PPS_CAPTUREASSERT) { 1206 tsp = &pps->ppsinfo.assert_timestamp; 1207 osp = &pps->ppsparam.assert_offset; 1208 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1209 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1210 pcount = &pps->ppscount[0]; 1211 pseq = &pps->ppsinfo.assert_sequence; 1212 } else { 1213 tsp = &pps->ppsinfo.clear_timestamp; 1214 osp = &pps->ppsparam.clear_offset; 1215 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1216 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1217 pcount = &pps->ppscount[1]; 1218 pseq = &pps->ppsinfo.clear_sequence; 1219 } 1220 1221 /* Nothing really happened */ 1222 if (*pcount == count) 1223 return; 1224 1225 *pcount = count; 1226 1227 do { 1228 ts.tv_sec = gd->gd_time_seconds; 1229 delta = count - gd->gd_cpuclock_base; 1230 } while (ts.tv_sec != gd->gd_time_seconds); 1231 1232 if (delta >= sys_cputimer->freq) { 1233 ts.tv_sec += delta / sys_cputimer->freq; 1234 delta %= sys_cputimer->freq; 1235 } 1236 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1237 bt = &basetime[basetime_index]; 1238 ts.tv_sec += bt->tv_sec; 1239 ts.tv_nsec += bt->tv_nsec; 1240 while (ts.tv_nsec >= 1000000000) { 1241 ts.tv_nsec -= 1000000000; 1242 ++ts.tv_sec; 1243 } 1244 1245 (*pseq)++; 1246 *tsp = ts; 1247 1248 if (foff) { 1249 timespecadd(tsp, osp); 1250 if (tsp->tv_nsec < 0) { 1251 tsp->tv_nsec += 1000000000; 1252 tsp->tv_sec -= 1; 1253 } 1254 } 1255 #ifdef PPS_SYNC 1256 if (fhard) { 1257 /* magic, at its best... */ 1258 tcount = count - pps->ppscount[2]; 1259 pps->ppscount[2] = count; 1260 if (tcount >= sys_cputimer->freq) { 1261 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1262 sys_cputimer->freq64_nsec * 1263 (tcount % sys_cputimer->freq)) >> 32; 1264 } else { 1265 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1266 } 1267 hardpps(tsp, delta); 1268 } 1269 #endif 1270 } 1271 1272