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 1.56 2007/01/01 22:51:17 corecode 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_SUB_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 cpu_initclocks(); 233 #ifdef DEVICE_POLLING 234 init_device_poll(); 235 #endif 236 /*psratio = profhz / stathz;*/ 237 initclocks_pcpu(); 238 clocks_running = 1; 239 } 240 241 /* 242 * Called on a per-cpu basis 243 */ 244 void 245 initclocks_pcpu(void) 246 { 247 struct globaldata *gd = mycpu; 248 249 crit_enter(); 250 if (gd->gd_cpuid == 0) { 251 gd->gd_time_seconds = 1; 252 gd->gd_cpuclock_base = sys_cputimer->count(); 253 } else { 254 /* XXX */ 255 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 256 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 257 } 258 259 /* 260 * Use a non-queued periodic systimer to prevent multiple ticks from 261 * building up if the sysclock jumps forward (8254 gets reset). The 262 * sysclock will never jump backwards. Our time sync is based on 263 * the actual sysclock, not the ticks count. 264 */ 265 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 266 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 267 /* XXX correct the frequency for scheduler / estcpu tests */ 268 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 269 NULL, ESTCPUFREQ); 270 crit_exit(); 271 } 272 273 /* 274 * This sets the current real time of day. Timespecs are in seconds and 275 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 276 * instead we adjust basetime so basetime + gd_* results in the current 277 * time of day. This way the gd_* fields are guarenteed to represent 278 * a monotonically increasing 'uptime' value. 279 * 280 * When set_timeofday() is called from userland, the system call forces it 281 * onto cpu #0 since only cpu #0 can update basetime_index. 282 */ 283 void 284 set_timeofday(struct timespec *ts) 285 { 286 struct timespec *nbt; 287 int ni; 288 289 /* 290 * XXX SMP / non-atomic basetime updates 291 */ 292 crit_enter(); 293 ni = (basetime_index + 1) & BASETIME_ARYMASK; 294 nbt = &basetime[ni]; 295 nanouptime(nbt); 296 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 297 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 298 if (nbt->tv_nsec < 0) { 299 nbt->tv_nsec += 1000000000; 300 --nbt->tv_sec; 301 } 302 303 /* 304 * Note that basetime diverges from boottime as the clock drift is 305 * compensated for, so we cannot do away with boottime. When setting 306 * the absolute time of day the drift is 0 (for an instant) and we 307 * can simply assign boottime to basetime. 308 * 309 * Note that nanouptime() is based on gd_time_seconds which is drift 310 * compensated up to a point (it is guarenteed to remain monotonically 311 * increasing). gd_time_seconds is thus our best uptime guess and 312 * suitable for use in the boottime calculation. It is already taken 313 * into account in the basetime calculation above. 314 */ 315 boottime.tv_sec = nbt->tv_sec; 316 ntp_delta = 0; 317 318 /* 319 * We now have a new basetime, make sure all other cpus have it, 320 * then update the index. 321 */ 322 cpu_sfence(); 323 basetime_index = ni; 324 325 crit_exit(); 326 } 327 328 /* 329 * Each cpu has its own hardclock, but we only increments ticks and softticks 330 * on cpu #0. 331 * 332 * NOTE! systimer! the MP lock might not be held here. We can only safely 333 * manipulate objects owned by the current cpu. 334 */ 335 static void 336 hardclock(systimer_t info, struct intrframe *frame) 337 { 338 sysclock_t cputicks; 339 struct proc *p; 340 struct globaldata *gd = mycpu; 341 342 /* 343 * Realtime updates are per-cpu. Note that timer corrections as 344 * returned by microtime() and friends make an additional adjustment 345 * using a system-wise 'basetime', but the running time is always 346 * taken from the per-cpu globaldata area. Since the same clock 347 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 348 * stay in synch. 349 * 350 * Note that we never allow info->time (aka gd->gd_hardclock.time) 351 * to reverse index gd_cpuclock_base, but that it is possible for 352 * it to temporarily get behind in the seconds if something in the 353 * system locks interrupts for a long period of time. Since periodic 354 * timers count events, though everything should resynch again 355 * immediately. 356 */ 357 cputicks = info->time - gd->gd_cpuclock_base; 358 if (cputicks >= sys_cputimer->freq) { 359 ++gd->gd_time_seconds; 360 gd->gd_cpuclock_base += sys_cputimer->freq; 361 } 362 363 /* 364 * The system-wide ticks counter and NTP related timedelta/tickdelta 365 * adjustments only occur on cpu #0. NTP adjustments are accomplished 366 * by updating basetime. 367 */ 368 if (gd->gd_cpuid == 0) { 369 struct timespec *nbt; 370 struct timespec nts; 371 int leap; 372 int ni; 373 374 ++ticks; 375 376 #if 0 377 if (tco->tc_poll_pps) 378 tco->tc_poll_pps(tco); 379 #endif 380 381 /* 382 * Calculate the new basetime index. We are in a critical section 383 * on cpu #0 and can safely play with basetime_index. Start 384 * with the current basetime and then make adjustments. 385 */ 386 ni = (basetime_index + 1) & BASETIME_ARYMASK; 387 nbt = &basetime[ni]; 388 *nbt = basetime[basetime_index]; 389 390 /* 391 * Apply adjtime corrections. (adjtime() API) 392 * 393 * adjtime() only runs on cpu #0 so our critical section is 394 * sufficient to access these variables. 395 */ 396 if (ntp_delta != 0) { 397 nbt->tv_nsec += ntp_tick_delta; 398 ntp_delta -= ntp_tick_delta; 399 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 400 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 401 ntp_tick_delta = ntp_delta; 402 } 403 } 404 405 /* 406 * Apply permanent frequency corrections. (sysctl API) 407 */ 408 if (ntp_tick_permanent != 0) { 409 ntp_tick_acc += ntp_tick_permanent; 410 if (ntp_tick_acc >= (1LL << 32)) { 411 nbt->tv_nsec += ntp_tick_acc >> 32; 412 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 413 } else if (ntp_tick_acc <= -(1LL << 32)) { 414 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 415 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 416 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 417 } 418 } 419 420 if (nbt->tv_nsec >= 1000000000) { 421 nbt->tv_sec++; 422 nbt->tv_nsec -= 1000000000; 423 } else if (nbt->tv_nsec < 0) { 424 nbt->tv_sec--; 425 nbt->tv_nsec += 1000000000; 426 } 427 428 /* 429 * Another per-tick compensation. (for ntp_adjtime() API) 430 */ 431 if (nsec_adj != 0) { 432 nsec_acc += nsec_adj; 433 if (nsec_acc >= 0x100000000LL) { 434 nbt->tv_nsec += nsec_acc >> 32; 435 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 436 } else if (nsec_acc <= -0x100000000LL) { 437 nbt->tv_nsec -= -nsec_acc >> 32; 438 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 439 } 440 if (nbt->tv_nsec >= 1000000000) { 441 nbt->tv_nsec -= 1000000000; 442 ++nbt->tv_sec; 443 } else if (nbt->tv_nsec < 0) { 444 nbt->tv_nsec += 1000000000; 445 --nbt->tv_sec; 446 } 447 } 448 449 /************************************************************ 450 * LEAP SECOND CORRECTION * 451 ************************************************************ 452 * 453 * Taking into account all the corrections made above, figure 454 * out the new real time. If the seconds field has changed 455 * then apply any pending leap-second corrections. 456 */ 457 getnanotime_nbt(nbt, &nts); 458 459 if (time_second != nts.tv_sec) { 460 /* 461 * Apply leap second (sysctl API). Adjust nts for changes 462 * so we do not have to call getnanotime_nbt again. 463 */ 464 if (ntp_leap_second) { 465 if (ntp_leap_second == nts.tv_sec) { 466 if (ntp_leap_insert) { 467 nbt->tv_sec++; 468 nts.tv_sec++; 469 } else { 470 nbt->tv_sec--; 471 nts.tv_sec--; 472 } 473 ntp_leap_second--; 474 } 475 } 476 477 /* 478 * Apply leap second (ntp_adjtime() API), calculate a new 479 * nsec_adj field. ntp_update_second() returns nsec_adj 480 * as a per-second value but we need it as a per-tick value. 481 */ 482 leap = ntp_update_second(time_second, &nsec_adj); 483 nsec_adj /= hz; 484 nbt->tv_sec += leap; 485 nts.tv_sec += leap; 486 487 /* 488 * Update the time_second 'approximate time' global. 489 */ 490 time_second = nts.tv_sec; 491 } 492 493 /* 494 * Finally, our new basetime is ready to go live! 495 */ 496 cpu_sfence(); 497 basetime_index = ni; 498 499 /* 500 * Figure out how badly the system is starved for memory 501 */ 502 vm_fault_ratecheck(); 503 } 504 505 /* 506 * softticks are handled for all cpus 507 */ 508 hardclock_softtick(gd); 509 510 /* 511 * ITimer handling is per-tick, per-cpu. I don't think ksignal() 512 * is mpsafe on curproc, so XXX get the mplock. 513 */ 514 if ((p = curproc) != NULL && try_mplock()) { 515 if (frame && CLKF_USERMODE(frame) && 516 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 517 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 518 ksignal(p, SIGVTALRM); 519 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 520 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 521 ksignal(p, SIGPROF); 522 rel_mplock(); 523 } 524 setdelayed(); 525 } 526 527 /* 528 * The statistics clock typically runs at a 125Hz rate, and is intended 529 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 530 * 531 * NOTE! systimer! the MP lock might not be held here. We can only safely 532 * manipulate objects owned by the current cpu. 533 * 534 * The stats clock is responsible for grabbing a profiling sample. 535 * Most of the statistics are only used by user-level statistics programs. 536 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 537 * p->p_estcpu. 538 * 539 * Like the other clocks, the stat clock is called from what is effectively 540 * a fast interrupt, so the context should be the thread/process that got 541 * interrupted. 542 */ 543 static void 544 statclock(systimer_t info, struct intrframe *frame) 545 { 546 #ifdef GPROF 547 struct gmonparam *g; 548 int i; 549 #endif 550 thread_t td; 551 struct proc *p; 552 int bump; 553 struct timeval tv; 554 struct timeval *stv; 555 556 /* 557 * How big was our timeslice relative to the last time? 558 */ 559 microuptime(&tv); /* mpsafe */ 560 stv = &mycpu->gd_stattv; 561 if (stv->tv_sec == 0) { 562 bump = 1; 563 } else { 564 bump = tv.tv_usec - stv->tv_usec + 565 (tv.tv_sec - stv->tv_sec) * 1000000; 566 if (bump < 0) 567 bump = 0; 568 if (bump > 1000000) 569 bump = 1000000; 570 } 571 *stv = tv; 572 573 td = curthread; 574 p = td->td_proc; 575 576 if (frame && CLKF_USERMODE(frame)) { 577 /* 578 * Came from userland, handle user time and deal with 579 * possible process. 580 */ 581 if (p && (p->p_flag & P_PROFIL)) 582 addupc_intr(p, CLKF_PC(frame), 1); 583 td->td_uticks += bump; 584 585 /* 586 * Charge the time as appropriate 587 */ 588 if (p && p->p_nice > NZERO) 589 cpu_time.cp_nice += bump; 590 else 591 cpu_time.cp_user += bump; 592 } else { 593 #ifdef GPROF 594 /* 595 * Kernel statistics are just like addupc_intr, only easier. 596 */ 597 g = &_gmonparam; 598 if (g->state == GMON_PROF_ON && frame) { 599 i = CLKF_PC(frame) - g->lowpc; 600 if (i < g->textsize) { 601 i /= HISTFRACTION * sizeof(*g->kcount); 602 g->kcount[i]++; 603 } 604 } 605 #endif 606 /* 607 * Came from kernel mode, so we were: 608 * - handling an interrupt, 609 * - doing syscall or trap work on behalf of the current 610 * user process, or 611 * - spinning in the idle loop. 612 * Whichever it is, charge the time as appropriate. 613 * Note that we charge interrupts to the current process, 614 * regardless of whether they are ``for'' that process, 615 * so that we know how much of its real time was spent 616 * in ``non-process'' (i.e., interrupt) work. 617 * 618 * XXX assume system if frame is NULL. A NULL frame 619 * can occur if ipi processing is done from a crit_exit(). 620 */ 621 if (frame && CLKF_INTR(frame)) 622 td->td_iticks += bump; 623 else 624 td->td_sticks += bump; 625 626 if (frame && CLKF_INTR(frame)) { 627 #ifdef DEBUG_PCTRACK 628 do_pctrack(frame, PCTRACK_INT); 629 #endif 630 cpu_time.cp_intr += bump; 631 } else { 632 if (td == &mycpu->gd_idlethread) { 633 cpu_time.cp_idle += bump; 634 } else { 635 #ifdef DEBUG_PCTRACK 636 if (frame) 637 do_pctrack(frame, PCTRACK_SYS); 638 #endif 639 cpu_time.cp_sys += bump; 640 } 641 } 642 } 643 } 644 645 #ifdef DEBUG_PCTRACK 646 /* 647 * Sample the PC when in the kernel or in an interrupt. User code can 648 * retrieve the information and generate a histogram or other output. 649 */ 650 651 static void 652 do_pctrack(struct intrframe *frame, int which) 653 { 654 struct kinfo_pctrack *pctrack; 655 656 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 657 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 658 (void *)CLKF_PC(frame); 659 ++pctrack->pc_index; 660 } 661 662 static int 663 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 664 { 665 struct kinfo_pcheader head; 666 int error; 667 int cpu; 668 int ntrack; 669 670 head.pc_ntrack = PCTRACK_SIZE; 671 head.pc_arysize = PCTRACK_ARYSIZE; 672 673 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 674 return (error); 675 676 for (cpu = 0; cpu < ncpus; ++cpu) { 677 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 678 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 679 sizeof(struct kinfo_pctrack)); 680 if (error) 681 break; 682 } 683 if (error) 684 break; 685 } 686 return (error); 687 } 688 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 689 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 690 691 #endif 692 693 /* 694 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 695 * the MP lock might not be held. We can safely manipulate parts of curproc 696 * but that's about it. 697 * 698 * Each cpu has its own scheduler clock. 699 */ 700 static void 701 schedclock(systimer_t info, struct intrframe *frame) 702 { 703 struct lwp *lp; 704 struct rusage *ru; 705 struct vmspace *vm; 706 long rss; 707 708 if ((lp = lwkt_preempted_proc()) != NULL) { 709 /* 710 * Account for cpu time used and hit the scheduler. Note 711 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 712 * HERE. 713 */ 714 ++lp->lwp_cpticks; 715 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 716 info->time); 717 } 718 if ((lp = curthread->td_lwp) != NULL) { 719 /* 720 * Update resource usage integrals and maximums. 721 */ 722 if ((ru = &lp->lwp_proc->p_ru) && 723 (vm = lp->lwp_proc->p_vmspace) != NULL) { 724 ru->ru_ixrss += pgtok(vm->vm_tsize); 725 ru->ru_idrss += pgtok(vm->vm_dsize); 726 ru->ru_isrss += pgtok(vm->vm_ssize); 727 rss = pgtok(vmspace_resident_count(vm)); 728 if (ru->ru_maxrss < rss) 729 ru->ru_maxrss = rss; 730 } 731 } 732 } 733 734 /* 735 * Compute number of ticks for the specified amount of time. The 736 * return value is intended to be used in a clock interrupt timed 737 * operation and guarenteed to meet or exceed the requested time. 738 * If the representation overflows, return INT_MAX. The minimum return 739 * value is 1 ticks and the function will average the calculation up. 740 * If any value greater then 0 microseconds is supplied, a value 741 * of at least 2 will be returned to ensure that a near-term clock 742 * interrupt does not cause the timeout to occur (degenerately) early. 743 * 744 * Note that limit checks must take into account microseconds, which is 745 * done simply by using the smaller signed long maximum instead of 746 * the unsigned long maximum. 747 * 748 * If ints have 32 bits, then the maximum value for any timeout in 749 * 10ms ticks is 248 days. 750 */ 751 int 752 tvtohz_high(struct timeval *tv) 753 { 754 int ticks; 755 long sec, usec; 756 757 sec = tv->tv_sec; 758 usec = tv->tv_usec; 759 if (usec < 0) { 760 sec--; 761 usec += 1000000; 762 } 763 if (sec < 0) { 764 #ifdef DIAGNOSTIC 765 if (usec > 0) { 766 sec++; 767 usec -= 1000000; 768 } 769 kprintf("tvotohz: negative time difference %ld sec %ld usec\n", 770 sec, usec); 771 #endif 772 ticks = 1; 773 } else if (sec <= INT_MAX / hz) { 774 ticks = (int)(sec * hz + 775 ((u_long)usec + (tick - 1)) / tick) + 1; 776 } else { 777 ticks = INT_MAX; 778 } 779 return (ticks); 780 } 781 782 /* 783 * Compute number of ticks for the specified amount of time, erroring on 784 * the side of it being too low to ensure that sleeping the returned number 785 * of ticks will not result in a late return. 786 * 787 * The supplied timeval may not be negative and should be normalized. A 788 * return value of 0 is possible if the timeval converts to less then 789 * 1 tick. 790 * 791 * If ints have 32 bits, then the maximum value for any timeout in 792 * 10ms ticks is 248 days. 793 */ 794 int 795 tvtohz_low(struct timeval *tv) 796 { 797 int ticks; 798 long sec; 799 800 sec = tv->tv_sec; 801 if (sec <= INT_MAX / hz) 802 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 803 else 804 ticks = INT_MAX; 805 return (ticks); 806 } 807 808 809 /* 810 * Start profiling on a process. 811 * 812 * Kernel profiling passes proc0 which never exits and hence 813 * keeps the profile clock running constantly. 814 */ 815 void 816 startprofclock(struct proc *p) 817 { 818 if ((p->p_flag & P_PROFIL) == 0) { 819 p->p_flag |= P_PROFIL; 820 #if 0 /* XXX */ 821 if (++profprocs == 1 && stathz != 0) { 822 crit_enter(); 823 psdiv = psratio; 824 setstatclockrate(profhz); 825 crit_exit(); 826 } 827 #endif 828 } 829 } 830 831 /* 832 * Stop profiling on a process. 833 */ 834 void 835 stopprofclock(struct proc *p) 836 { 837 if (p->p_flag & P_PROFIL) { 838 p->p_flag &= ~P_PROFIL; 839 #if 0 /* XXX */ 840 if (--profprocs == 0 && stathz != 0) { 841 crit_enter(); 842 psdiv = 1; 843 setstatclockrate(stathz); 844 crit_exit(); 845 } 846 #endif 847 } 848 } 849 850 /* 851 * Return information about system clocks. 852 */ 853 static int 854 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 855 { 856 struct kinfo_clockinfo clkinfo; 857 /* 858 * Construct clockinfo structure. 859 */ 860 clkinfo.ci_hz = hz; 861 clkinfo.ci_tick = tick; 862 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 863 clkinfo.ci_profhz = profhz; 864 clkinfo.ci_stathz = stathz ? stathz : hz; 865 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 866 } 867 868 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 869 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 870 871 /* 872 * We have eight functions for looking at the clock, four for 873 * microseconds and four for nanoseconds. For each there is fast 874 * but less precise version "get{nano|micro}[up]time" which will 875 * return a time which is up to 1/HZ previous to the call, whereas 876 * the raw version "{nano|micro}[up]time" will return a timestamp 877 * which is as precise as possible. The "up" variants return the 878 * time relative to system boot, these are well suited for time 879 * interval measurements. 880 * 881 * Each cpu independantly maintains the current time of day, so all 882 * we need to do to protect ourselves from changes is to do a loop 883 * check on the seconds field changing out from under us. 884 * 885 * The system timer maintains a 32 bit count and due to various issues 886 * it is possible for the calculated delta to occassionally exceed 887 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 888 * multiplication can easily overflow, so we deal with the case. For 889 * uniformity we deal with the case in the usec case too. 890 */ 891 void 892 getmicrouptime(struct timeval *tvp) 893 { 894 struct globaldata *gd = mycpu; 895 sysclock_t delta; 896 897 do { 898 tvp->tv_sec = gd->gd_time_seconds; 899 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 900 } while (tvp->tv_sec != gd->gd_time_seconds); 901 902 if (delta >= sys_cputimer->freq) { 903 tvp->tv_sec += delta / sys_cputimer->freq; 904 delta %= sys_cputimer->freq; 905 } 906 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 907 if (tvp->tv_usec >= 1000000) { 908 tvp->tv_usec -= 1000000; 909 ++tvp->tv_sec; 910 } 911 } 912 913 void 914 getnanouptime(struct timespec *tsp) 915 { 916 struct globaldata *gd = mycpu; 917 sysclock_t delta; 918 919 do { 920 tsp->tv_sec = gd->gd_time_seconds; 921 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 922 } while (tsp->tv_sec != gd->gd_time_seconds); 923 924 if (delta >= sys_cputimer->freq) { 925 tsp->tv_sec += delta / sys_cputimer->freq; 926 delta %= sys_cputimer->freq; 927 } 928 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 929 } 930 931 void 932 microuptime(struct timeval *tvp) 933 { 934 struct globaldata *gd = mycpu; 935 sysclock_t delta; 936 937 do { 938 tvp->tv_sec = gd->gd_time_seconds; 939 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 940 } while (tvp->tv_sec != gd->gd_time_seconds); 941 942 if (delta >= sys_cputimer->freq) { 943 tvp->tv_sec += delta / sys_cputimer->freq; 944 delta %= sys_cputimer->freq; 945 } 946 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 947 } 948 949 void 950 nanouptime(struct timespec *tsp) 951 { 952 struct globaldata *gd = mycpu; 953 sysclock_t delta; 954 955 do { 956 tsp->tv_sec = gd->gd_time_seconds; 957 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 958 } while (tsp->tv_sec != gd->gd_time_seconds); 959 960 if (delta >= sys_cputimer->freq) { 961 tsp->tv_sec += delta / sys_cputimer->freq; 962 delta %= sys_cputimer->freq; 963 } 964 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 965 } 966 967 /* 968 * realtime routines 969 */ 970 971 void 972 getmicrotime(struct timeval *tvp) 973 { 974 struct globaldata *gd = mycpu; 975 struct timespec *bt; 976 sysclock_t delta; 977 978 do { 979 tvp->tv_sec = gd->gd_time_seconds; 980 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 981 } while (tvp->tv_sec != gd->gd_time_seconds); 982 983 if (delta >= sys_cputimer->freq) { 984 tvp->tv_sec += delta / sys_cputimer->freq; 985 delta %= sys_cputimer->freq; 986 } 987 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 988 989 bt = &basetime[basetime_index]; 990 tvp->tv_sec += bt->tv_sec; 991 tvp->tv_usec += bt->tv_nsec / 1000; 992 while (tvp->tv_usec >= 1000000) { 993 tvp->tv_usec -= 1000000; 994 ++tvp->tv_sec; 995 } 996 } 997 998 void 999 getnanotime(struct timespec *tsp) 1000 { 1001 struct globaldata *gd = mycpu; 1002 struct timespec *bt; 1003 sysclock_t delta; 1004 1005 do { 1006 tsp->tv_sec = gd->gd_time_seconds; 1007 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1008 } while (tsp->tv_sec != gd->gd_time_seconds); 1009 1010 if (delta >= sys_cputimer->freq) { 1011 tsp->tv_sec += delta / sys_cputimer->freq; 1012 delta %= sys_cputimer->freq; 1013 } 1014 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1015 1016 bt = &basetime[basetime_index]; 1017 tsp->tv_sec += bt->tv_sec; 1018 tsp->tv_nsec += bt->tv_nsec; 1019 while (tsp->tv_nsec >= 1000000000) { 1020 tsp->tv_nsec -= 1000000000; 1021 ++tsp->tv_sec; 1022 } 1023 } 1024 1025 static void 1026 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1027 { 1028 struct globaldata *gd = mycpu; 1029 sysclock_t delta; 1030 1031 do { 1032 tsp->tv_sec = gd->gd_time_seconds; 1033 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1034 } while (tsp->tv_sec != gd->gd_time_seconds); 1035 1036 if (delta >= sys_cputimer->freq) { 1037 tsp->tv_sec += delta / sys_cputimer->freq; 1038 delta %= sys_cputimer->freq; 1039 } 1040 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1041 1042 tsp->tv_sec += nbt->tv_sec; 1043 tsp->tv_nsec += nbt->tv_nsec; 1044 while (tsp->tv_nsec >= 1000000000) { 1045 tsp->tv_nsec -= 1000000000; 1046 ++tsp->tv_sec; 1047 } 1048 } 1049 1050 1051 void 1052 microtime(struct timeval *tvp) 1053 { 1054 struct globaldata *gd = mycpu; 1055 struct timespec *bt; 1056 sysclock_t delta; 1057 1058 do { 1059 tvp->tv_sec = gd->gd_time_seconds; 1060 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1061 } while (tvp->tv_sec != gd->gd_time_seconds); 1062 1063 if (delta >= sys_cputimer->freq) { 1064 tvp->tv_sec += delta / sys_cputimer->freq; 1065 delta %= sys_cputimer->freq; 1066 } 1067 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1068 1069 bt = &basetime[basetime_index]; 1070 tvp->tv_sec += bt->tv_sec; 1071 tvp->tv_usec += bt->tv_nsec / 1000; 1072 while (tvp->tv_usec >= 1000000) { 1073 tvp->tv_usec -= 1000000; 1074 ++tvp->tv_sec; 1075 } 1076 } 1077 1078 void 1079 nanotime(struct timespec *tsp) 1080 { 1081 struct globaldata *gd = mycpu; 1082 struct timespec *bt; 1083 sysclock_t delta; 1084 1085 do { 1086 tsp->tv_sec = gd->gd_time_seconds; 1087 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1088 } while (tsp->tv_sec != gd->gd_time_seconds); 1089 1090 if (delta >= sys_cputimer->freq) { 1091 tsp->tv_sec += delta / sys_cputimer->freq; 1092 delta %= sys_cputimer->freq; 1093 } 1094 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1095 1096 bt = &basetime[basetime_index]; 1097 tsp->tv_sec += bt->tv_sec; 1098 tsp->tv_nsec += bt->tv_nsec; 1099 while (tsp->tv_nsec >= 1000000000) { 1100 tsp->tv_nsec -= 1000000000; 1101 ++tsp->tv_sec; 1102 } 1103 } 1104 1105 /* 1106 * note: this is not exactly synchronized with real time. To do that we 1107 * would have to do what microtime does and check for a nanoseconds overflow. 1108 */ 1109 time_t 1110 get_approximate_time_t(void) 1111 { 1112 struct globaldata *gd = mycpu; 1113 struct timespec *bt; 1114 1115 bt = &basetime[basetime_index]; 1116 return(gd->gd_time_seconds + bt->tv_sec); 1117 } 1118 1119 int 1120 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1121 { 1122 pps_params_t *app; 1123 struct pps_fetch_args *fapi; 1124 #ifdef PPS_SYNC 1125 struct pps_kcbind_args *kapi; 1126 #endif 1127 1128 switch (cmd) { 1129 case PPS_IOC_CREATE: 1130 return (0); 1131 case PPS_IOC_DESTROY: 1132 return (0); 1133 case PPS_IOC_SETPARAMS: 1134 app = (pps_params_t *)data; 1135 if (app->mode & ~pps->ppscap) 1136 return (EINVAL); 1137 pps->ppsparam = *app; 1138 return (0); 1139 case PPS_IOC_GETPARAMS: 1140 app = (pps_params_t *)data; 1141 *app = pps->ppsparam; 1142 app->api_version = PPS_API_VERS_1; 1143 return (0); 1144 case PPS_IOC_GETCAP: 1145 *(int*)data = pps->ppscap; 1146 return (0); 1147 case PPS_IOC_FETCH: 1148 fapi = (struct pps_fetch_args *)data; 1149 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1150 return (EINVAL); 1151 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1152 return (EOPNOTSUPP); 1153 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1154 fapi->pps_info_buf = pps->ppsinfo; 1155 return (0); 1156 case PPS_IOC_KCBIND: 1157 #ifdef PPS_SYNC 1158 kapi = (struct pps_kcbind_args *)data; 1159 /* XXX Only root should be able to do this */ 1160 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1161 return (EINVAL); 1162 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1163 return (EINVAL); 1164 if (kapi->edge & ~pps->ppscap) 1165 return (EINVAL); 1166 pps->kcmode = kapi->edge; 1167 return (0); 1168 #else 1169 return (EOPNOTSUPP); 1170 #endif 1171 default: 1172 return (ENOTTY); 1173 } 1174 } 1175 1176 void 1177 pps_init(struct pps_state *pps) 1178 { 1179 pps->ppscap |= PPS_TSFMT_TSPEC; 1180 if (pps->ppscap & PPS_CAPTUREASSERT) 1181 pps->ppscap |= PPS_OFFSETASSERT; 1182 if (pps->ppscap & PPS_CAPTURECLEAR) 1183 pps->ppscap |= PPS_OFFSETCLEAR; 1184 } 1185 1186 void 1187 pps_event(struct pps_state *pps, sysclock_t count, int event) 1188 { 1189 struct globaldata *gd; 1190 struct timespec *tsp; 1191 struct timespec *osp; 1192 struct timespec *bt; 1193 struct timespec ts; 1194 sysclock_t *pcount; 1195 #ifdef PPS_SYNC 1196 sysclock_t tcount; 1197 #endif 1198 sysclock_t delta; 1199 pps_seq_t *pseq; 1200 int foff; 1201 int fhard; 1202 1203 gd = mycpu; 1204 1205 /* Things would be easier with arrays... */ 1206 if (event == PPS_CAPTUREASSERT) { 1207 tsp = &pps->ppsinfo.assert_timestamp; 1208 osp = &pps->ppsparam.assert_offset; 1209 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1210 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1211 pcount = &pps->ppscount[0]; 1212 pseq = &pps->ppsinfo.assert_sequence; 1213 } else { 1214 tsp = &pps->ppsinfo.clear_timestamp; 1215 osp = &pps->ppsparam.clear_offset; 1216 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1217 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1218 pcount = &pps->ppscount[1]; 1219 pseq = &pps->ppsinfo.clear_sequence; 1220 } 1221 1222 /* Nothing really happened */ 1223 if (*pcount == count) 1224 return; 1225 1226 *pcount = count; 1227 1228 do { 1229 ts.tv_sec = gd->gd_time_seconds; 1230 delta = count - gd->gd_cpuclock_base; 1231 } while (ts.tv_sec != gd->gd_time_seconds); 1232 1233 if (delta >= sys_cputimer->freq) { 1234 ts.tv_sec += delta / sys_cputimer->freq; 1235 delta %= sys_cputimer->freq; 1236 } 1237 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1238 bt = &basetime[basetime_index]; 1239 ts.tv_sec += bt->tv_sec; 1240 ts.tv_nsec += bt->tv_nsec; 1241 while (ts.tv_nsec >= 1000000000) { 1242 ts.tv_nsec -= 1000000000; 1243 ++ts.tv_sec; 1244 } 1245 1246 (*pseq)++; 1247 *tsp = ts; 1248 1249 if (foff) { 1250 timespecadd(tsp, osp); 1251 if (tsp->tv_nsec < 0) { 1252 tsp->tv_nsec += 1000000000; 1253 tsp->tv_sec -= 1; 1254 } 1255 } 1256 #ifdef PPS_SYNC 1257 if (fhard) { 1258 /* magic, at its best... */ 1259 tcount = count - pps->ppscount[2]; 1260 pps->ppscount[2] = count; 1261 if (tcount >= sys_cputimer->freq) { 1262 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1263 sys_cputimer->freq64_nsec * 1264 (tcount % sys_cputimer->freq)) >> 32; 1265 } else { 1266 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1267 } 1268 hardpps(tsp, delta); 1269 } 1270 #endif 1271 } 1272 1273