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 1.51 2006/03/24 18:30:33 dillon Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 #include "opt_polling.h" 78 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/callout.h> 82 #include <sys/kernel.h> 83 #include <sys/kinfo.h> 84 #include <sys/proc.h> 85 #include <sys/malloc.h> 86 #include <sys/resourcevar.h> 87 #include <sys/signalvar.h> 88 #include <sys/timex.h> 89 #include <sys/timepps.h> 90 #include <vm/vm.h> 91 #include <sys/lock.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_extern.h> 95 #include <sys/sysctl.h> 96 #include <sys/thread2.h> 97 98 #include <machine/cpu.h> 99 #include <machine/limits.h> 100 #include <machine/smp.h> 101 102 #ifdef GPROF 103 #include <sys/gmon.h> 104 #endif 105 106 #ifdef DEVICE_POLLING 107 extern void init_device_poll(void); 108 #endif 109 110 static void initclocks (void *dummy); 111 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 112 113 /* 114 * Some of these don't belong here, but it's easiest to concentrate them. 115 * Note that cpu_time counts in microseconds, but most userland programs 116 * just compare relative times against the total by delta. 117 */ 118 struct kinfo_cputime cputime_percpu[MAXCPU]; 119 #ifdef SMP 120 static int 121 sysctl_cputime(SYSCTL_HANDLER_ARGS) 122 { 123 int cpu, error = 0; 124 size_t size = sizeof(struct kinfo_cputime); 125 126 for (cpu = 0; cpu < ncpus; ++cpu) { 127 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 128 break; 129 } 130 131 return (error); 132 } 133 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 134 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 135 #else 136 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 137 "CPU time statistics"); 138 #endif 139 140 /* 141 * boottime is used to calculate the 'real' uptime. Do not confuse this with 142 * microuptime(). microtime() is not drift compensated. The real uptime 143 * with compensation is nanotime() - bootime. boottime is recalculated 144 * whenever the real time is set based on the compensated elapsed time 145 * in seconds (gd->gd_time_seconds). 146 * 147 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 148 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 149 * the real time. 150 */ 151 struct timespec boottime; /* boot time (realtime) for reference only */ 152 time_t time_second; /* read-only 'passive' uptime in seconds */ 153 154 /* 155 * basetime is used to calculate the compensated real time of day. The 156 * basetime can be modified on a per-tick basis by the adjtime(), 157 * ntp_adjtime(), and sysctl-based time correction APIs. 158 * 159 * Note that frequency corrections can also be made by adjusting 160 * gd_cpuclock_base. 161 * 162 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 163 * used on both SMP and UP systems to avoid MP races between cpu's and 164 * interrupt races on UP systems. 165 */ 166 #define BASETIME_ARYSIZE 16 167 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 168 static struct timespec basetime[BASETIME_ARYSIZE]; 169 static volatile int basetime_index; 170 171 static int 172 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 173 { 174 struct timespec *bt; 175 int error; 176 int index; 177 178 /* 179 * Because basetime data and index may be updated by another cpu, 180 * a load fence is required to ensure that the data we read has 181 * not been speculatively read relative to a possibly updated index. 182 */ 183 index = basetime_index; 184 cpu_lfence(); 185 bt = &basetime[index]; 186 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 187 return (error); 188 } 189 190 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 191 &boottime, timespec, "System boottime"); 192 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 193 sysctl_get_basetime, "S,timespec", "System basetime"); 194 195 static void hardclock(systimer_t info, struct intrframe *frame); 196 static void statclock(systimer_t info, struct intrframe *frame); 197 static void schedclock(systimer_t info, struct intrframe *frame); 198 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 199 200 int ticks; /* system master ticks at hz */ 201 int clocks_running; /* tsleep/timeout clocks operational */ 202 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 203 int64_t nsec_acc; /* accumulator */ 204 205 /* NTPD time correction fields */ 206 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 207 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 208 int64_t ntp_delta; /* one-time correction in nsec */ 209 int64_t ntp_big_delta = 1000000000; 210 int32_t ntp_tick_delta; /* current adjustment rate */ 211 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 212 time_t ntp_leap_second; /* time of next leap second */ 213 int ntp_leap_insert; /* whether to insert or remove a second */ 214 215 /* 216 * Finish initializing clock frequencies and start all clocks running. 217 */ 218 /* ARGSUSED*/ 219 static void 220 initclocks(void *dummy) 221 { 222 cpu_initclocks(); 223 #ifdef DEVICE_POLLING 224 init_device_poll(); 225 #endif 226 /*psratio = profhz / stathz;*/ 227 initclocks_pcpu(); 228 clocks_running = 1; 229 } 230 231 /* 232 * Called on a per-cpu basis 233 */ 234 void 235 initclocks_pcpu(void) 236 { 237 struct globaldata *gd = mycpu; 238 239 crit_enter(); 240 if (gd->gd_cpuid == 0) { 241 gd->gd_time_seconds = 1; 242 gd->gd_cpuclock_base = sys_cputimer->count(); 243 } else { 244 /* XXX */ 245 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 246 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 247 } 248 249 /* 250 * Use a non-queued periodic systimer to prevent multiple ticks from 251 * building up if the sysclock jumps forward (8254 gets reset). The 252 * sysclock will never jump backwards. Our time sync is based on 253 * the actual sysclock, not the ticks count. 254 */ 255 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 256 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 257 /* XXX correct the frequency for scheduler / estcpu tests */ 258 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 259 NULL, ESTCPUFREQ); 260 crit_exit(); 261 } 262 263 /* 264 * This sets the current real time of day. Timespecs are in seconds and 265 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 266 * instead we adjust basetime so basetime + gd_* results in the current 267 * time of day. This way the gd_* fields are guarenteed to represent 268 * a monotonically increasing 'uptime' value. 269 * 270 * When set_timeofday() is called from userland, the system call forces it 271 * onto cpu #0 since only cpu #0 can update basetime_index. 272 */ 273 void 274 set_timeofday(struct timespec *ts) 275 { 276 struct timespec *nbt; 277 int ni; 278 279 /* 280 * XXX SMP / non-atomic basetime updates 281 */ 282 crit_enter(); 283 ni = (basetime_index + 1) & BASETIME_ARYMASK; 284 nbt = &basetime[ni]; 285 nanouptime(nbt); 286 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 287 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 288 if (nbt->tv_nsec < 0) { 289 nbt->tv_nsec += 1000000000; 290 --nbt->tv_sec; 291 } 292 293 /* 294 * Note that basetime diverges from boottime as the clock drift is 295 * compensated for, so we cannot do away with boottime. When setting 296 * the absolute time of day the drift is 0 (for an instant) and we 297 * can simply assign boottime to basetime. 298 * 299 * Note that nanouptime() is based on gd_time_seconds which is drift 300 * compensated up to a point (it is guarenteed to remain monotonically 301 * increasing). gd_time_seconds is thus our best uptime guess and 302 * suitable for use in the boottime calculation. It is already taken 303 * into account in the basetime calculation above. 304 */ 305 boottime.tv_sec = nbt->tv_sec; 306 ntp_delta = 0; 307 308 /* 309 * We now have a new basetime, make sure all other cpus have it, 310 * then update the index. 311 */ 312 cpu_sfence(); 313 basetime_index = ni; 314 315 crit_exit(); 316 } 317 318 /* 319 * Each cpu has its own hardclock, but we only increments ticks and softticks 320 * on cpu #0. 321 * 322 * NOTE! systimer! the MP lock might not be held here. We can only safely 323 * manipulate objects owned by the current cpu. 324 */ 325 static void 326 hardclock(systimer_t info, struct intrframe *frame) 327 { 328 sysclock_t cputicks; 329 struct proc *p; 330 struct pstats *pstats; 331 struct globaldata *gd = mycpu; 332 333 /* 334 * Realtime updates are per-cpu. Note that timer corrections as 335 * returned by microtime() and friends make an additional adjustment 336 * using a system-wise 'basetime', but the running time is always 337 * taken from the per-cpu globaldata area. Since the same clock 338 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 339 * stay in synch. 340 * 341 * Note that we never allow info->time (aka gd->gd_hardclock.time) 342 * to reverse index gd_cpuclock_base, but that it is possible for 343 * it to temporarily get behind in the seconds if something in the 344 * system locks interrupts for a long period of time. Since periodic 345 * timers count events, though everything should resynch again 346 * immediately. 347 */ 348 cputicks = info->time - gd->gd_cpuclock_base; 349 if (cputicks >= sys_cputimer->freq) { 350 ++gd->gd_time_seconds; 351 gd->gd_cpuclock_base += sys_cputimer->freq; 352 } 353 354 /* 355 * The system-wide ticks counter and NTP related timedelta/tickdelta 356 * adjustments only occur on cpu #0. NTP adjustments are accomplished 357 * by updating basetime. 358 */ 359 if (gd->gd_cpuid == 0) { 360 struct timespec *nbt; 361 struct timespec nts; 362 int leap; 363 int ni; 364 365 ++ticks; 366 367 #if 0 368 if (tco->tc_poll_pps) 369 tco->tc_poll_pps(tco); 370 #endif 371 372 /* 373 * Calculate the new basetime index. We are in a critical section 374 * on cpu #0 and can safely play with basetime_index. Start 375 * with the current basetime and then make adjustments. 376 */ 377 ni = (basetime_index + 1) & BASETIME_ARYMASK; 378 nbt = &basetime[ni]; 379 *nbt = basetime[basetime_index]; 380 381 /* 382 * Apply adjtime corrections. (adjtime() API) 383 * 384 * adjtime() only runs on cpu #0 so our critical section is 385 * sufficient to access these variables. 386 */ 387 if (ntp_delta != 0) { 388 nbt->tv_nsec += ntp_tick_delta; 389 ntp_delta -= ntp_tick_delta; 390 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 391 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 392 ntp_tick_delta = ntp_delta; 393 } 394 } 395 396 /* 397 * Apply permanent frequency corrections. (sysctl API) 398 */ 399 if (ntp_tick_permanent != 0) { 400 ntp_tick_acc += ntp_tick_permanent; 401 if (ntp_tick_acc >= (1LL << 32)) { 402 nbt->tv_nsec += ntp_tick_acc >> 32; 403 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 404 } else if (ntp_tick_acc <= -(1LL << 32)) { 405 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 406 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 407 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 408 } 409 } 410 411 if (nbt->tv_nsec >= 1000000000) { 412 nbt->tv_sec++; 413 nbt->tv_nsec -= 1000000000; 414 } else if (nbt->tv_nsec < 0) { 415 nbt->tv_sec--; 416 nbt->tv_nsec += 1000000000; 417 } 418 419 /* 420 * Another per-tick compensation. (for ntp_adjtime() API) 421 */ 422 if (nsec_adj != 0) { 423 nsec_acc += nsec_adj; 424 if (nsec_acc >= 0x100000000LL) { 425 nbt->tv_nsec += nsec_acc >> 32; 426 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 427 } else if (nsec_acc <= -0x100000000LL) { 428 nbt->tv_nsec -= -nsec_acc >> 32; 429 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 430 } 431 if (nbt->tv_nsec >= 1000000000) { 432 nbt->tv_nsec -= 1000000000; 433 ++nbt->tv_sec; 434 } else if (nbt->tv_nsec < 0) { 435 nbt->tv_nsec += 1000000000; 436 --nbt->tv_sec; 437 } 438 } 439 440 /************************************************************ 441 * LEAP SECOND CORRECTION * 442 ************************************************************ 443 * 444 * Taking into account all the corrections made above, figure 445 * out the new real time. If the seconds field has changed 446 * then apply any pending leap-second corrections. 447 */ 448 getnanotime_nbt(nbt, &nts); 449 450 if (time_second != nts.tv_sec) { 451 /* 452 * Apply leap second (sysctl API). Adjust nts for changes 453 * so we do not have to call getnanotime_nbt again. 454 */ 455 if (ntp_leap_second) { 456 if (ntp_leap_second == nts.tv_sec) { 457 if (ntp_leap_insert) { 458 nbt->tv_sec++; 459 nts.tv_sec++; 460 } else { 461 nbt->tv_sec--; 462 nts.tv_sec--; 463 } 464 ntp_leap_second--; 465 } 466 } 467 468 /* 469 * Apply leap second (ntp_adjtime() API), calculate a new 470 * nsec_adj field. ntp_update_second() returns nsec_adj 471 * as a per-second value but we need it as a per-tick value. 472 */ 473 leap = ntp_update_second(time_second, &nsec_adj); 474 nsec_adj /= hz; 475 nbt->tv_sec += leap; 476 nts.tv_sec += leap; 477 478 /* 479 * Update the time_second 'approximate time' global. 480 */ 481 time_second = nts.tv_sec; 482 } 483 484 /* 485 * Finally, our new basetime is ready to go live! 486 */ 487 cpu_sfence(); 488 basetime_index = ni; 489 490 /* 491 * Figure out how badly the system is starved for memory 492 */ 493 vm_fault_ratecheck(); 494 } 495 496 /* 497 * softticks are handled for all cpus 498 */ 499 hardclock_softtick(gd); 500 501 /* 502 * ITimer handling is per-tick, per-cpu. I don't think psignal() 503 * is mpsafe on curproc, so XXX get the mplock. 504 */ 505 if ((p = curproc) != NULL && try_mplock()) { 506 pstats = p->p_stats; 507 if (frame && CLKF_USERMODE(frame) && 508 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 509 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0) 510 psignal(p, SIGVTALRM); 511 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 512 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0) 513 psignal(p, SIGPROF); 514 rel_mplock(); 515 } 516 setdelayed(); 517 } 518 519 /* 520 * The statistics clock typically runs at a 125Hz rate, and is intended 521 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 522 * 523 * NOTE! systimer! the MP lock might not be held here. We can only safely 524 * manipulate objects owned by the current cpu. 525 * 526 * The stats clock is responsible for grabbing a profiling sample. 527 * Most of the statistics are only used by user-level statistics programs. 528 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 529 * p->p_estcpu. 530 * 531 * Like the other clocks, the stat clock is called from what is effectively 532 * a fast interrupt, so the context should be the thread/process that got 533 * interrupted. 534 */ 535 static void 536 statclock(systimer_t info, struct intrframe *frame) 537 { 538 #ifdef GPROF 539 struct gmonparam *g; 540 int i; 541 #endif 542 thread_t td; 543 struct proc *p; 544 int bump; 545 struct timeval tv; 546 struct timeval *stv; 547 548 /* 549 * How big was our timeslice relative to the last time? 550 */ 551 microuptime(&tv); /* mpsafe */ 552 stv = &mycpu->gd_stattv; 553 if (stv->tv_sec == 0) { 554 bump = 1; 555 } else { 556 bump = tv.tv_usec - stv->tv_usec + 557 (tv.tv_sec - stv->tv_sec) * 1000000; 558 if (bump < 0) 559 bump = 0; 560 if (bump > 1000000) 561 bump = 1000000; 562 } 563 *stv = tv; 564 565 td = curthread; 566 p = td->td_proc; 567 568 if (frame && CLKF_USERMODE(frame)) { 569 /* 570 * Came from userland, handle user time and deal with 571 * possible process. 572 */ 573 if (p && (p->p_flag & P_PROFIL)) 574 addupc_intr(p, CLKF_PC(frame), 1); 575 td->td_uticks += bump; 576 577 /* 578 * Charge the time as appropriate 579 */ 580 if (p && p->p_nice > NZERO) 581 cpu_time.cp_nice += bump; 582 else 583 cpu_time.cp_user += bump; 584 } else { 585 #ifdef GPROF 586 /* 587 * Kernel statistics are just like addupc_intr, only easier. 588 */ 589 g = &_gmonparam; 590 if (g->state == GMON_PROF_ON && frame) { 591 i = CLKF_PC(frame) - g->lowpc; 592 if (i < g->textsize) { 593 i /= HISTFRACTION * sizeof(*g->kcount); 594 g->kcount[i]++; 595 } 596 } 597 #endif 598 /* 599 * Came from kernel mode, so we were: 600 * - handling an interrupt, 601 * - doing syscall or trap work on behalf of the current 602 * user process, or 603 * - spinning in the idle loop. 604 * Whichever it is, charge the time as appropriate. 605 * Note that we charge interrupts to the current process, 606 * regardless of whether they are ``for'' that process, 607 * so that we know how much of its real time was spent 608 * in ``non-process'' (i.e., interrupt) work. 609 * 610 * XXX assume system if frame is NULL. A NULL frame 611 * can occur if ipi processing is done from a crit_exit(). 612 */ 613 if (frame && CLKF_INTR(frame)) 614 td->td_iticks += bump; 615 else 616 td->td_sticks += bump; 617 618 if (frame && CLKF_INTR(frame)) { 619 cpu_time.cp_intr += bump; 620 } else { 621 if (td == &mycpu->gd_idlethread) 622 cpu_time.cp_idle += bump; 623 else 624 cpu_time.cp_sys += bump; 625 } 626 } 627 } 628 629 /* 630 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 631 * the MP lock might not be held. We can safely manipulate parts of curproc 632 * but that's about it. 633 * 634 * Each cpu has its own scheduler clock. 635 */ 636 static void 637 schedclock(systimer_t info, struct intrframe *frame) 638 { 639 struct lwp *lp; 640 struct pstats *pstats; 641 struct rusage *ru; 642 struct vmspace *vm; 643 long rss; 644 645 if ((lp = lwkt_preempted_proc()) != NULL) { 646 /* 647 * Account for cpu time used and hit the scheduler. Note 648 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 649 * HERE. 650 */ 651 ++lp->lwp_cpticks; 652 /* 653 * XXX I think accessing lwp_proc's p_usched is 654 * reasonably MP safe. This needs to be revisited 655 * when we have pluggable schedulers. 656 */ 657 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, info->time); 658 } 659 if ((lp = curthread->td_lwp) != NULL) { 660 /* 661 * Update resource usage integrals and maximums. 662 */ 663 if ((pstats = lp->lwp_stats) != NULL && 664 (ru = &pstats->p_ru) != NULL && 665 (vm = lp->lwp_proc->p_vmspace) != NULL) { 666 ru->ru_ixrss += pgtok(vm->vm_tsize); 667 ru->ru_idrss += pgtok(vm->vm_dsize); 668 ru->ru_isrss += pgtok(vm->vm_ssize); 669 rss = pgtok(vmspace_resident_count(vm)); 670 if (ru->ru_maxrss < rss) 671 ru->ru_maxrss = rss; 672 } 673 } 674 } 675 676 /* 677 * Compute number of ticks for the specified amount of time. The 678 * return value is intended to be used in a clock interrupt timed 679 * operation and guarenteed to meet or exceed the requested time. 680 * If the representation overflows, return INT_MAX. The minimum return 681 * value is 1 ticks and the function will average the calculation up. 682 * If any value greater then 0 microseconds is supplied, a value 683 * of at least 2 will be returned to ensure that a near-term clock 684 * interrupt does not cause the timeout to occur (degenerately) early. 685 * 686 * Note that limit checks must take into account microseconds, which is 687 * done simply by using the smaller signed long maximum instead of 688 * the unsigned long maximum. 689 * 690 * If ints have 32 bits, then the maximum value for any timeout in 691 * 10ms ticks is 248 days. 692 */ 693 int 694 tvtohz_high(struct timeval *tv) 695 { 696 int ticks; 697 long sec, usec; 698 699 sec = tv->tv_sec; 700 usec = tv->tv_usec; 701 if (usec < 0) { 702 sec--; 703 usec += 1000000; 704 } 705 if (sec < 0) { 706 #ifdef DIAGNOSTIC 707 if (usec > 0) { 708 sec++; 709 usec -= 1000000; 710 } 711 printf("tvotohz: negative time difference %ld sec %ld usec\n", 712 sec, usec); 713 #endif 714 ticks = 1; 715 } else if (sec <= INT_MAX / hz) { 716 ticks = (int)(sec * hz + 717 ((u_long)usec + (tick - 1)) / tick) + 1; 718 } else { 719 ticks = INT_MAX; 720 } 721 return (ticks); 722 } 723 724 /* 725 * Compute number of ticks for the specified amount of time, erroring on 726 * the side of it being too low to ensure that sleeping the returned number 727 * of ticks will not result in a late return. 728 * 729 * The supplied timeval may not be negative and should be normalized. A 730 * return value of 0 is possible if the timeval converts to less then 731 * 1 tick. 732 * 733 * If ints have 32 bits, then the maximum value for any timeout in 734 * 10ms ticks is 248 days. 735 */ 736 int 737 tvtohz_low(struct timeval *tv) 738 { 739 int ticks; 740 long sec; 741 742 sec = tv->tv_sec; 743 if (sec <= INT_MAX / hz) 744 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick); 745 else 746 ticks = INT_MAX; 747 return (ticks); 748 } 749 750 751 /* 752 * Start profiling on a process. 753 * 754 * Kernel profiling passes proc0 which never exits and hence 755 * keeps the profile clock running constantly. 756 */ 757 void 758 startprofclock(struct proc *p) 759 { 760 if ((p->p_flag & P_PROFIL) == 0) { 761 p->p_flag |= P_PROFIL; 762 #if 0 /* XXX */ 763 if (++profprocs == 1 && stathz != 0) { 764 crit_enter(); 765 psdiv = psratio; 766 setstatclockrate(profhz); 767 crit_exit(); 768 } 769 #endif 770 } 771 } 772 773 /* 774 * Stop profiling on a process. 775 */ 776 void 777 stopprofclock(struct proc *p) 778 { 779 if (p->p_flag & P_PROFIL) { 780 p->p_flag &= ~P_PROFIL; 781 #if 0 /* XXX */ 782 if (--profprocs == 0 && stathz != 0) { 783 crit_enter(); 784 psdiv = 1; 785 setstatclockrate(stathz); 786 crit_exit(); 787 } 788 #endif 789 } 790 } 791 792 /* 793 * Return information about system clocks. 794 */ 795 static int 796 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 797 { 798 struct kinfo_clockinfo clkinfo; 799 /* 800 * Construct clockinfo structure. 801 */ 802 clkinfo.ci_hz = hz; 803 clkinfo.ci_tick = tick; 804 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 805 clkinfo.ci_profhz = profhz; 806 clkinfo.ci_stathz = stathz ? stathz : hz; 807 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 808 } 809 810 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 811 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 812 813 /* 814 * We have eight functions for looking at the clock, four for 815 * microseconds and four for nanoseconds. For each there is fast 816 * but less precise version "get{nano|micro}[up]time" which will 817 * return a time which is up to 1/HZ previous to the call, whereas 818 * the raw version "{nano|micro}[up]time" will return a timestamp 819 * which is as precise as possible. The "up" variants return the 820 * time relative to system boot, these are well suited for time 821 * interval measurements. 822 * 823 * Each cpu independantly maintains the current time of day, so all 824 * we need to do to protect ourselves from changes is to do a loop 825 * check on the seconds field changing out from under us. 826 * 827 * The system timer maintains a 32 bit count and due to various issues 828 * it is possible for the calculated delta to occassionally exceed 829 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 830 * multiplication can easily overflow, so we deal with the case. For 831 * uniformity we deal with the case in the usec case too. 832 */ 833 void 834 getmicrouptime(struct timeval *tvp) 835 { 836 struct globaldata *gd = mycpu; 837 sysclock_t delta; 838 839 do { 840 tvp->tv_sec = gd->gd_time_seconds; 841 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 842 } while (tvp->tv_sec != gd->gd_time_seconds); 843 844 if (delta >= sys_cputimer->freq) { 845 tvp->tv_sec += delta / sys_cputimer->freq; 846 delta %= sys_cputimer->freq; 847 } 848 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 849 if (tvp->tv_usec >= 1000000) { 850 tvp->tv_usec -= 1000000; 851 ++tvp->tv_sec; 852 } 853 } 854 855 void 856 getnanouptime(struct timespec *tsp) 857 { 858 struct globaldata *gd = mycpu; 859 sysclock_t delta; 860 861 do { 862 tsp->tv_sec = gd->gd_time_seconds; 863 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 864 } while (tsp->tv_sec != gd->gd_time_seconds); 865 866 if (delta >= sys_cputimer->freq) { 867 tsp->tv_sec += delta / sys_cputimer->freq; 868 delta %= sys_cputimer->freq; 869 } 870 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 871 } 872 873 void 874 microuptime(struct timeval *tvp) 875 { 876 struct globaldata *gd = mycpu; 877 sysclock_t delta; 878 879 do { 880 tvp->tv_sec = gd->gd_time_seconds; 881 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 882 } while (tvp->tv_sec != gd->gd_time_seconds); 883 884 if (delta >= sys_cputimer->freq) { 885 tvp->tv_sec += delta / sys_cputimer->freq; 886 delta %= sys_cputimer->freq; 887 } 888 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 889 } 890 891 void 892 nanouptime(struct timespec *tsp) 893 { 894 struct globaldata *gd = mycpu; 895 sysclock_t delta; 896 897 do { 898 tsp->tv_sec = gd->gd_time_seconds; 899 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 900 } while (tsp->tv_sec != gd->gd_time_seconds); 901 902 if (delta >= sys_cputimer->freq) { 903 tsp->tv_sec += delta / sys_cputimer->freq; 904 delta %= sys_cputimer->freq; 905 } 906 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 907 } 908 909 /* 910 * realtime routines 911 */ 912 913 void 914 getmicrotime(struct timeval *tvp) 915 { 916 struct globaldata *gd = mycpu; 917 struct timespec *bt; 918 sysclock_t delta; 919 920 do { 921 tvp->tv_sec = gd->gd_time_seconds; 922 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 923 } while (tvp->tv_sec != gd->gd_time_seconds); 924 925 if (delta >= sys_cputimer->freq) { 926 tvp->tv_sec += delta / sys_cputimer->freq; 927 delta %= sys_cputimer->freq; 928 } 929 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 930 931 bt = &basetime[basetime_index]; 932 tvp->tv_sec += bt->tv_sec; 933 tvp->tv_usec += bt->tv_nsec / 1000; 934 while (tvp->tv_usec >= 1000000) { 935 tvp->tv_usec -= 1000000; 936 ++tvp->tv_sec; 937 } 938 } 939 940 void 941 getnanotime(struct timespec *tsp) 942 { 943 struct globaldata *gd = mycpu; 944 struct timespec *bt; 945 sysclock_t delta; 946 947 do { 948 tsp->tv_sec = gd->gd_time_seconds; 949 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 950 } while (tsp->tv_sec != gd->gd_time_seconds); 951 952 if (delta >= sys_cputimer->freq) { 953 tsp->tv_sec += delta / sys_cputimer->freq; 954 delta %= sys_cputimer->freq; 955 } 956 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 957 958 bt = &basetime[basetime_index]; 959 tsp->tv_sec += bt->tv_sec; 960 tsp->tv_nsec += bt->tv_nsec; 961 while (tsp->tv_nsec >= 1000000000) { 962 tsp->tv_nsec -= 1000000000; 963 ++tsp->tv_sec; 964 } 965 } 966 967 static void 968 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 969 { 970 struct globaldata *gd = mycpu; 971 sysclock_t delta; 972 973 do { 974 tsp->tv_sec = gd->gd_time_seconds; 975 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 976 } while (tsp->tv_sec != gd->gd_time_seconds); 977 978 if (delta >= sys_cputimer->freq) { 979 tsp->tv_sec += delta / sys_cputimer->freq; 980 delta %= sys_cputimer->freq; 981 } 982 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 983 984 tsp->tv_sec += nbt->tv_sec; 985 tsp->tv_nsec += nbt->tv_nsec; 986 while (tsp->tv_nsec >= 1000000000) { 987 tsp->tv_nsec -= 1000000000; 988 ++tsp->tv_sec; 989 } 990 } 991 992 993 void 994 microtime(struct timeval *tvp) 995 { 996 struct globaldata *gd = mycpu; 997 struct timespec *bt; 998 sysclock_t delta; 999 1000 do { 1001 tvp->tv_sec = gd->gd_time_seconds; 1002 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1003 } while (tvp->tv_sec != gd->gd_time_seconds); 1004 1005 if (delta >= sys_cputimer->freq) { 1006 tvp->tv_sec += delta / sys_cputimer->freq; 1007 delta %= sys_cputimer->freq; 1008 } 1009 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1010 1011 bt = &basetime[basetime_index]; 1012 tvp->tv_sec += bt->tv_sec; 1013 tvp->tv_usec += bt->tv_nsec / 1000; 1014 while (tvp->tv_usec >= 1000000) { 1015 tvp->tv_usec -= 1000000; 1016 ++tvp->tv_sec; 1017 } 1018 } 1019 1020 void 1021 nanotime(struct timespec *tsp) 1022 { 1023 struct globaldata *gd = mycpu; 1024 struct timespec *bt; 1025 sysclock_t delta; 1026 1027 do { 1028 tsp->tv_sec = gd->gd_time_seconds; 1029 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1030 } while (tsp->tv_sec != gd->gd_time_seconds); 1031 1032 if (delta >= sys_cputimer->freq) { 1033 tsp->tv_sec += delta / sys_cputimer->freq; 1034 delta %= sys_cputimer->freq; 1035 } 1036 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1037 1038 bt = &basetime[basetime_index]; 1039 tsp->tv_sec += bt->tv_sec; 1040 tsp->tv_nsec += bt->tv_nsec; 1041 while (tsp->tv_nsec >= 1000000000) { 1042 tsp->tv_nsec -= 1000000000; 1043 ++tsp->tv_sec; 1044 } 1045 } 1046 1047 /* 1048 * note: this is not exactly synchronized with real time. To do that we 1049 * would have to do what microtime does and check for a nanoseconds overflow. 1050 */ 1051 time_t 1052 get_approximate_time_t(void) 1053 { 1054 struct globaldata *gd = mycpu; 1055 struct timespec *bt; 1056 1057 bt = &basetime[basetime_index]; 1058 return(gd->gd_time_seconds + bt->tv_sec); 1059 } 1060 1061 int 1062 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1063 { 1064 pps_params_t *app; 1065 struct pps_fetch_args *fapi; 1066 #ifdef PPS_SYNC 1067 struct pps_kcbind_args *kapi; 1068 #endif 1069 1070 switch (cmd) { 1071 case PPS_IOC_CREATE: 1072 return (0); 1073 case PPS_IOC_DESTROY: 1074 return (0); 1075 case PPS_IOC_SETPARAMS: 1076 app = (pps_params_t *)data; 1077 if (app->mode & ~pps->ppscap) 1078 return (EINVAL); 1079 pps->ppsparam = *app; 1080 return (0); 1081 case PPS_IOC_GETPARAMS: 1082 app = (pps_params_t *)data; 1083 *app = pps->ppsparam; 1084 app->api_version = PPS_API_VERS_1; 1085 return (0); 1086 case PPS_IOC_GETCAP: 1087 *(int*)data = pps->ppscap; 1088 return (0); 1089 case PPS_IOC_FETCH: 1090 fapi = (struct pps_fetch_args *)data; 1091 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1092 return (EINVAL); 1093 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1094 return (EOPNOTSUPP); 1095 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1096 fapi->pps_info_buf = pps->ppsinfo; 1097 return (0); 1098 case PPS_IOC_KCBIND: 1099 #ifdef PPS_SYNC 1100 kapi = (struct pps_kcbind_args *)data; 1101 /* XXX Only root should be able to do this */ 1102 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1103 return (EINVAL); 1104 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1105 return (EINVAL); 1106 if (kapi->edge & ~pps->ppscap) 1107 return (EINVAL); 1108 pps->kcmode = kapi->edge; 1109 return (0); 1110 #else 1111 return (EOPNOTSUPP); 1112 #endif 1113 default: 1114 return (ENOTTY); 1115 } 1116 } 1117 1118 void 1119 pps_init(struct pps_state *pps) 1120 { 1121 pps->ppscap |= PPS_TSFMT_TSPEC; 1122 if (pps->ppscap & PPS_CAPTUREASSERT) 1123 pps->ppscap |= PPS_OFFSETASSERT; 1124 if (pps->ppscap & PPS_CAPTURECLEAR) 1125 pps->ppscap |= PPS_OFFSETCLEAR; 1126 } 1127 1128 void 1129 pps_event(struct pps_state *pps, sysclock_t count, int event) 1130 { 1131 struct globaldata *gd; 1132 struct timespec *tsp; 1133 struct timespec *osp; 1134 struct timespec *bt; 1135 struct timespec ts; 1136 sysclock_t *pcount; 1137 #ifdef PPS_SYNC 1138 sysclock_t tcount; 1139 #endif 1140 sysclock_t delta; 1141 pps_seq_t *pseq; 1142 int foff; 1143 int fhard; 1144 1145 gd = mycpu; 1146 1147 /* Things would be easier with arrays... */ 1148 if (event == PPS_CAPTUREASSERT) { 1149 tsp = &pps->ppsinfo.assert_timestamp; 1150 osp = &pps->ppsparam.assert_offset; 1151 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1152 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1153 pcount = &pps->ppscount[0]; 1154 pseq = &pps->ppsinfo.assert_sequence; 1155 } else { 1156 tsp = &pps->ppsinfo.clear_timestamp; 1157 osp = &pps->ppsparam.clear_offset; 1158 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1159 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1160 pcount = &pps->ppscount[1]; 1161 pseq = &pps->ppsinfo.clear_sequence; 1162 } 1163 1164 /* Nothing really happened */ 1165 if (*pcount == count) 1166 return; 1167 1168 *pcount = count; 1169 1170 do { 1171 ts.tv_sec = gd->gd_time_seconds; 1172 delta = count - gd->gd_cpuclock_base; 1173 } while (ts.tv_sec != gd->gd_time_seconds); 1174 1175 if (delta >= sys_cputimer->freq) { 1176 ts.tv_sec += delta / sys_cputimer->freq; 1177 delta %= sys_cputimer->freq; 1178 } 1179 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1180 bt = &basetime[basetime_index]; 1181 ts.tv_sec += bt->tv_sec; 1182 ts.tv_nsec += bt->tv_nsec; 1183 while (ts.tv_nsec >= 1000000000) { 1184 ts.tv_nsec -= 1000000000; 1185 ++ts.tv_sec; 1186 } 1187 1188 (*pseq)++; 1189 *tsp = ts; 1190 1191 if (foff) { 1192 timespecadd(tsp, osp); 1193 if (tsp->tv_nsec < 0) { 1194 tsp->tv_nsec += 1000000000; 1195 tsp->tv_sec -= 1; 1196 } 1197 } 1198 #ifdef PPS_SYNC 1199 if (fhard) { 1200 /* magic, at its best... */ 1201 tcount = count - pps->ppscount[2]; 1202 pps->ppscount[2] = count; 1203 if (tcount >= sys_cputimer->freq) { 1204 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1205 sys_cputimer->freq64_nsec * 1206 (tcount % sys_cputimer->freq)) >> 32; 1207 } else { 1208 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1209 } 1210 hardpps(tsp, delta); 1211 } 1212 #endif 1213 } 1214 1215