1 /* 2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> 35 * Copyright (c) 1982, 1986, 1991, 1993 36 * The Regents of the University of California. All rights reserved. 37 * (c) UNIX System Laboratories, Inc. 38 * All or some portions of this file are derived from material licensed 39 * to the University of California by American Telephone and Telegraph 40 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 41 * the permission of UNIX System Laboratories, Inc. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 3. All advertising materials mentioning features or use of this software 52 * must display the following acknowledgement: 53 * This product includes software developed by the University of 54 * California, Berkeley and its contributors. 55 * 4. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 72 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ 73 * $DragonFly: src/sys/kern/kern_clock.c,v 1.62 2008/09/09 04:06:13 dillon Exp $ 74 */ 75 76 #include "opt_ntp.h" 77 #include "opt_polling.h" 78 #include "opt_ifpoll.h" 79 #include "opt_pctrack.h" 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/callout.h> 84 #include <sys/kernel.h> 85 #include <sys/kinfo.h> 86 #include <sys/proc.h> 87 #include <sys/malloc.h> 88 #include <sys/resourcevar.h> 89 #include <sys/signalvar.h> 90 #include <sys/timex.h> 91 #include <sys/timepps.h> 92 #include <vm/vm.h> 93 #include <sys/lock.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_extern.h> 97 #include <sys/sysctl.h> 98 99 #include <sys/thread2.h> 100 #include <sys/mplock2.h> 101 102 #include <machine/cpu.h> 103 #include <machine/limits.h> 104 #include <machine/smp.h> 105 #include <machine/cpufunc.h> 106 #include <machine/specialreg.h> 107 #include <machine/clock.h> 108 109 #ifdef GPROF 110 #include <sys/gmon.h> 111 #endif 112 113 #ifdef DEVICE_POLLING 114 extern void init_device_poll_pcpu(int); 115 #endif 116 117 #ifdef IFPOLL_ENABLE 118 extern void ifpoll_init_pcpu(int); 119 #endif 120 121 #ifdef DEBUG_PCTRACK 122 static void do_pctrack(struct intrframe *frame, int which); 123 #endif 124 125 static void initclocks (void *dummy); 126 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 127 128 /* 129 * Some of these don't belong here, but it's easiest to concentrate them. 130 * Note that cpu_time counts in microseconds, but most userland programs 131 * just compare relative times against the total by delta. 132 */ 133 struct kinfo_cputime cputime_percpu[MAXCPU]; 134 #ifdef DEBUG_PCTRACK 135 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE }; 136 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE]; 137 #endif 138 139 #ifdef SMP 140 static int 141 sysctl_cputime(SYSCTL_HANDLER_ARGS) 142 { 143 int cpu, error = 0; 144 size_t size = sizeof(struct kinfo_cputime); 145 146 for (cpu = 0; cpu < ncpus; ++cpu) { 147 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 148 break; 149 } 150 151 return (error); 152 } 153 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 154 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 155 #else 156 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 157 "CPU time statistics"); 158 #endif 159 160 /* 161 * boottime is used to calculate the 'real' uptime. Do not confuse this with 162 * microuptime(). microtime() is not drift compensated. The real uptime 163 * with compensation is nanotime() - bootime. boottime is recalculated 164 * whenever the real time is set based on the compensated elapsed time 165 * in seconds (gd->gd_time_seconds). 166 * 167 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 168 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 169 * the real time. 170 */ 171 struct timespec boottime; /* boot time (realtime) for reference only */ 172 time_t time_second; /* read-only 'passive' uptime in seconds */ 173 174 /* 175 * basetime is used to calculate the compensated real time of day. The 176 * basetime can be modified on a per-tick basis by the adjtime(), 177 * ntp_adjtime(), and sysctl-based time correction APIs. 178 * 179 * Note that frequency corrections can also be made by adjusting 180 * gd_cpuclock_base. 181 * 182 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 183 * used on both SMP and UP systems to avoid MP races between cpu's and 184 * interrupt races on UP systems. 185 */ 186 #define BASETIME_ARYSIZE 16 187 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 188 static struct timespec basetime[BASETIME_ARYSIZE]; 189 static volatile int basetime_index; 190 191 static int 192 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 193 { 194 struct timespec *bt; 195 int error; 196 int index; 197 198 /* 199 * Because basetime data and index may be updated by another cpu, 200 * a load fence is required to ensure that the data we read has 201 * not been speculatively read relative to a possibly updated index. 202 */ 203 index = basetime_index; 204 cpu_lfence(); 205 bt = &basetime[index]; 206 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 207 return (error); 208 } 209 210 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 211 &boottime, timespec, "System boottime"); 212 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 213 sysctl_get_basetime, "S,timespec", "System basetime"); 214 215 static void hardclock(systimer_t info, struct intrframe *frame); 216 static void statclock(systimer_t info, struct intrframe *frame); 217 static void schedclock(systimer_t info, struct intrframe *frame); 218 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 219 220 int ticks; /* system master ticks at hz */ 221 int clocks_running; /* tsleep/timeout clocks operational */ 222 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 223 int64_t nsec_acc; /* accumulator */ 224 225 /* NTPD time correction fields */ 226 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 227 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 228 int64_t ntp_delta; /* one-time correction in nsec */ 229 int64_t ntp_big_delta = 1000000000; 230 int32_t ntp_tick_delta; /* current adjustment rate */ 231 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 232 time_t ntp_leap_second; /* time of next leap second */ 233 int ntp_leap_insert; /* whether to insert or remove a second */ 234 235 /* 236 * Finish initializing clock frequencies and start all clocks running. 237 */ 238 /* ARGSUSED*/ 239 static void 240 initclocks(void *dummy) 241 { 242 /*psratio = profhz / stathz;*/ 243 initclocks_pcpu(); 244 clocks_running = 1; 245 } 246 247 /* 248 * Called on a per-cpu basis 249 */ 250 void 251 initclocks_pcpu(void) 252 { 253 struct globaldata *gd = mycpu; 254 255 crit_enter(); 256 if (gd->gd_cpuid == 0) { 257 gd->gd_time_seconds = 1; 258 gd->gd_cpuclock_base = sys_cputimer->count(); 259 } else { 260 /* XXX */ 261 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 262 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 263 } 264 265 systimer_intr_enable(); 266 267 #ifdef DEVICE_POLLING 268 init_device_poll_pcpu(gd->gd_cpuid); 269 #endif 270 271 #ifdef IFPOLL_ENABLE 272 ifpoll_init_pcpu(gd->gd_cpuid); 273 #endif 274 275 /* 276 * Use a non-queued periodic systimer to prevent multiple ticks from 277 * building up if the sysclock jumps forward (8254 gets reset). The 278 * sysclock will never jump backwards. Our time sync is based on 279 * the actual sysclock, not the ticks count. 280 */ 281 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 282 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 283 /* XXX correct the frequency for scheduler / estcpu tests */ 284 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 285 NULL, ESTCPUFREQ); 286 crit_exit(); 287 } 288 289 /* 290 * This sets the current real time of day. Timespecs are in seconds and 291 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 292 * instead we adjust basetime so basetime + gd_* results in the current 293 * time of day. This way the gd_* fields are guarenteed to represent 294 * a monotonically increasing 'uptime' value. 295 * 296 * When set_timeofday() is called from userland, the system call forces it 297 * onto cpu #0 since only cpu #0 can update basetime_index. 298 */ 299 void 300 set_timeofday(struct timespec *ts) 301 { 302 struct timespec *nbt; 303 int ni; 304 305 /* 306 * XXX SMP / non-atomic basetime updates 307 */ 308 crit_enter(); 309 ni = (basetime_index + 1) & BASETIME_ARYMASK; 310 nbt = &basetime[ni]; 311 nanouptime(nbt); 312 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 313 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 314 if (nbt->tv_nsec < 0) { 315 nbt->tv_nsec += 1000000000; 316 --nbt->tv_sec; 317 } 318 319 /* 320 * Note that basetime diverges from boottime as the clock drift is 321 * compensated for, so we cannot do away with boottime. When setting 322 * the absolute time of day the drift is 0 (for an instant) and we 323 * can simply assign boottime to basetime. 324 * 325 * Note that nanouptime() is based on gd_time_seconds which is drift 326 * compensated up to a point (it is guarenteed to remain monotonically 327 * increasing). gd_time_seconds is thus our best uptime guess and 328 * suitable for use in the boottime calculation. It is already taken 329 * into account in the basetime calculation above. 330 */ 331 boottime.tv_sec = nbt->tv_sec; 332 ntp_delta = 0; 333 334 /* 335 * We now have a new basetime, make sure all other cpus have it, 336 * then update the index. 337 */ 338 cpu_sfence(); 339 basetime_index = ni; 340 341 crit_exit(); 342 } 343 344 /* 345 * Each cpu has its own hardclock, but we only increments ticks and softticks 346 * on cpu #0. 347 * 348 * NOTE! systimer! the MP lock might not be held here. We can only safely 349 * manipulate objects owned by the current cpu. 350 */ 351 static void 352 hardclock(systimer_t info, struct intrframe *frame) 353 { 354 sysclock_t cputicks; 355 struct proc *p; 356 struct globaldata *gd = mycpu; 357 358 /* 359 * Realtime updates are per-cpu. Note that timer corrections as 360 * returned by microtime() and friends make an additional adjustment 361 * using a system-wise 'basetime', but the running time is always 362 * taken from the per-cpu globaldata area. Since the same clock 363 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 364 * stay in synch. 365 * 366 * Note that we never allow info->time (aka gd->gd_hardclock.time) 367 * to reverse index gd_cpuclock_base, but that it is possible for 368 * it to temporarily get behind in the seconds if something in the 369 * system locks interrupts for a long period of time. Since periodic 370 * timers count events, though everything should resynch again 371 * immediately. 372 */ 373 cputicks = info->time - gd->gd_cpuclock_base; 374 if (cputicks >= sys_cputimer->freq) { 375 ++gd->gd_time_seconds; 376 gd->gd_cpuclock_base += sys_cputimer->freq; 377 } 378 379 /* 380 * The system-wide ticks counter and NTP related timedelta/tickdelta 381 * adjustments only occur on cpu #0. NTP adjustments are accomplished 382 * by updating basetime. 383 */ 384 if (gd->gd_cpuid == 0) { 385 struct timespec *nbt; 386 struct timespec nts; 387 int leap; 388 int ni; 389 390 ++ticks; 391 392 #if 0 393 if (tco->tc_poll_pps) 394 tco->tc_poll_pps(tco); 395 #endif 396 397 /* 398 * Calculate the new basetime index. We are in a critical section 399 * on cpu #0 and can safely play with basetime_index. Start 400 * with the current basetime and then make adjustments. 401 */ 402 ni = (basetime_index + 1) & BASETIME_ARYMASK; 403 nbt = &basetime[ni]; 404 *nbt = basetime[basetime_index]; 405 406 /* 407 * Apply adjtime corrections. (adjtime() API) 408 * 409 * adjtime() only runs on cpu #0 so our critical section is 410 * sufficient to access these variables. 411 */ 412 if (ntp_delta != 0) { 413 nbt->tv_nsec += ntp_tick_delta; 414 ntp_delta -= ntp_tick_delta; 415 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 416 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 417 ntp_tick_delta = ntp_delta; 418 } 419 } 420 421 /* 422 * Apply permanent frequency corrections. (sysctl API) 423 */ 424 if (ntp_tick_permanent != 0) { 425 ntp_tick_acc += ntp_tick_permanent; 426 if (ntp_tick_acc >= (1LL << 32)) { 427 nbt->tv_nsec += ntp_tick_acc >> 32; 428 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 429 } else if (ntp_tick_acc <= -(1LL << 32)) { 430 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 431 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 432 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 433 } 434 } 435 436 if (nbt->tv_nsec >= 1000000000) { 437 nbt->tv_sec++; 438 nbt->tv_nsec -= 1000000000; 439 } else if (nbt->tv_nsec < 0) { 440 nbt->tv_sec--; 441 nbt->tv_nsec += 1000000000; 442 } 443 444 /* 445 * Another per-tick compensation. (for ntp_adjtime() API) 446 */ 447 if (nsec_adj != 0) { 448 nsec_acc += nsec_adj; 449 if (nsec_acc >= 0x100000000LL) { 450 nbt->tv_nsec += nsec_acc >> 32; 451 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 452 } else if (nsec_acc <= -0x100000000LL) { 453 nbt->tv_nsec -= -nsec_acc >> 32; 454 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 455 } 456 if (nbt->tv_nsec >= 1000000000) { 457 nbt->tv_nsec -= 1000000000; 458 ++nbt->tv_sec; 459 } else if (nbt->tv_nsec < 0) { 460 nbt->tv_nsec += 1000000000; 461 --nbt->tv_sec; 462 } 463 } 464 465 /************************************************************ 466 * LEAP SECOND CORRECTION * 467 ************************************************************ 468 * 469 * Taking into account all the corrections made above, figure 470 * out the new real time. If the seconds field has changed 471 * then apply any pending leap-second corrections. 472 */ 473 getnanotime_nbt(nbt, &nts); 474 475 if (time_second != nts.tv_sec) { 476 /* 477 * Apply leap second (sysctl API). Adjust nts for changes 478 * so we do not have to call getnanotime_nbt again. 479 */ 480 if (ntp_leap_second) { 481 if (ntp_leap_second == nts.tv_sec) { 482 if (ntp_leap_insert) { 483 nbt->tv_sec++; 484 nts.tv_sec++; 485 } else { 486 nbt->tv_sec--; 487 nts.tv_sec--; 488 } 489 ntp_leap_second--; 490 } 491 } 492 493 /* 494 * Apply leap second (ntp_adjtime() API), calculate a new 495 * nsec_adj field. ntp_update_second() returns nsec_adj 496 * as a per-second value but we need it as a per-tick value. 497 */ 498 leap = ntp_update_second(time_second, &nsec_adj); 499 nsec_adj /= hz; 500 nbt->tv_sec += leap; 501 nts.tv_sec += leap; 502 503 /* 504 * Update the time_second 'approximate time' global. 505 */ 506 time_second = nts.tv_sec; 507 } 508 509 /* 510 * Finally, our new basetime is ready to go live! 511 */ 512 cpu_sfence(); 513 basetime_index = ni; 514 515 /* 516 * Figure out how badly the system is starved for memory 517 */ 518 vm_fault_ratecheck(); 519 } 520 521 /* 522 * lwkt thread scheduler fair queueing 523 */ 524 lwkt_fairq_schedulerclock(curthread); 525 526 /* 527 * softticks are handled for all cpus 528 */ 529 hardclock_softtick(gd); 530 531 /* 532 * The LWKT scheduler will generally allow the current process to 533 * return to user mode even if there are other runnable LWKT threads 534 * running in kernel mode on behalf of a user process. This will 535 * ensure that those other threads have an opportunity to run in 536 * fairly short order (but not instantly). 537 */ 538 need_lwkt_resched(); 539 540 /* 541 * ITimer handling is per-tick, per-cpu. 542 * 543 * We must acquire the per-process token in order for ksignal() 544 * to be non-blocking. 545 */ 546 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) { 547 crit_enter_hard(); 548 if (frame && CLKF_USERMODE(frame) && 549 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 550 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) 551 ksignal(p, SIGVTALRM); 552 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 553 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) 554 ksignal(p, SIGPROF); 555 crit_exit_hard(); 556 lwkt_reltoken(&p->p_token); 557 } 558 setdelayed(); 559 } 560 561 /* 562 * The statistics clock typically runs at a 125Hz rate, and is intended 563 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 564 * 565 * NOTE! systimer! the MP lock might not be held here. We can only safely 566 * manipulate objects owned by the current cpu. 567 * 568 * The stats clock is responsible for grabbing a profiling sample. 569 * Most of the statistics are only used by user-level statistics programs. 570 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 571 * p->p_estcpu. 572 * 573 * Like the other clocks, the stat clock is called from what is effectively 574 * a fast interrupt, so the context should be the thread/process that got 575 * interrupted. 576 */ 577 static void 578 statclock(systimer_t info, struct intrframe *frame) 579 { 580 #ifdef GPROF 581 struct gmonparam *g; 582 int i; 583 #endif 584 thread_t td; 585 struct proc *p; 586 int bump; 587 struct timeval tv; 588 struct timeval *stv; 589 590 /* 591 * How big was our timeslice relative to the last time? 592 */ 593 microuptime(&tv); /* mpsafe */ 594 stv = &mycpu->gd_stattv; 595 if (stv->tv_sec == 0) { 596 bump = 1; 597 } else { 598 bump = tv.tv_usec - stv->tv_usec + 599 (tv.tv_sec - stv->tv_sec) * 1000000; 600 if (bump < 0) 601 bump = 0; 602 if (bump > 1000000) 603 bump = 1000000; 604 } 605 *stv = tv; 606 607 td = curthread; 608 p = td->td_proc; 609 610 if (frame && CLKF_USERMODE(frame)) { 611 /* 612 * Came from userland, handle user time and deal with 613 * possible process. 614 */ 615 if (p && (p->p_flag & P_PROFIL)) 616 addupc_intr(p, CLKF_PC(frame), 1); 617 td->td_uticks += bump; 618 619 /* 620 * Charge the time as appropriate 621 */ 622 if (p && p->p_nice > NZERO) 623 cpu_time.cp_nice += bump; 624 else 625 cpu_time.cp_user += bump; 626 } else { 627 #ifdef GPROF 628 /* 629 * Kernel statistics are just like addupc_intr, only easier. 630 */ 631 g = &_gmonparam; 632 if (g->state == GMON_PROF_ON && frame) { 633 i = CLKF_PC(frame) - g->lowpc; 634 if (i < g->textsize) { 635 i /= HISTFRACTION * sizeof(*g->kcount); 636 g->kcount[i]++; 637 } 638 } 639 #endif 640 /* 641 * Came from kernel mode, so we were: 642 * - handling an interrupt, 643 * - doing syscall or trap work on behalf of the current 644 * user process, or 645 * - spinning in the idle loop. 646 * Whichever it is, charge the time as appropriate. 647 * Note that we charge interrupts to the current process, 648 * regardless of whether they are ``for'' that process, 649 * so that we know how much of its real time was spent 650 * in ``non-process'' (i.e., interrupt) work. 651 * 652 * XXX assume system if frame is NULL. A NULL frame 653 * can occur if ipi processing is done from a crit_exit(). 654 */ 655 if (frame && CLKF_INTR(frame)) 656 td->td_iticks += bump; 657 else 658 td->td_sticks += bump; 659 660 if (frame && CLKF_INTR(frame)) { 661 #ifdef DEBUG_PCTRACK 662 do_pctrack(frame, PCTRACK_INT); 663 #endif 664 cpu_time.cp_intr += bump; 665 } else { 666 if (td == &mycpu->gd_idlethread) { 667 cpu_time.cp_idle += bump; 668 } else { 669 #ifdef DEBUG_PCTRACK 670 if (frame) 671 do_pctrack(frame, PCTRACK_SYS); 672 #endif 673 cpu_time.cp_sys += bump; 674 } 675 } 676 } 677 } 678 679 #ifdef DEBUG_PCTRACK 680 /* 681 * Sample the PC when in the kernel or in an interrupt. User code can 682 * retrieve the information and generate a histogram or other output. 683 */ 684 685 static void 686 do_pctrack(struct intrframe *frame, int which) 687 { 688 struct kinfo_pctrack *pctrack; 689 690 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 691 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 692 (void *)CLKF_PC(frame); 693 ++pctrack->pc_index; 694 } 695 696 static int 697 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 698 { 699 struct kinfo_pcheader head; 700 int error; 701 int cpu; 702 int ntrack; 703 704 head.pc_ntrack = PCTRACK_SIZE; 705 head.pc_arysize = PCTRACK_ARYSIZE; 706 707 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 708 return (error); 709 710 for (cpu = 0; cpu < ncpus; ++cpu) { 711 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 712 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 713 sizeof(struct kinfo_pctrack)); 714 if (error) 715 break; 716 } 717 if (error) 718 break; 719 } 720 return (error); 721 } 722 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 723 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 724 725 #endif 726 727 /* 728 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 729 * the MP lock might not be held. We can safely manipulate parts of curproc 730 * but that's about it. 731 * 732 * Each cpu has its own scheduler clock. 733 */ 734 static void 735 schedclock(systimer_t info, struct intrframe *frame) 736 { 737 struct lwp *lp; 738 struct rusage *ru; 739 struct vmspace *vm; 740 long rss; 741 742 if ((lp = lwkt_preempted_proc()) != NULL) { 743 /* 744 * Account for cpu time used and hit the scheduler. Note 745 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 746 * HERE. 747 */ 748 ++lp->lwp_cpticks; 749 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 750 info->time); 751 } 752 if ((lp = curthread->td_lwp) != NULL) { 753 /* 754 * Update resource usage integrals and maximums. 755 */ 756 if ((ru = &lp->lwp_proc->p_ru) && 757 (vm = lp->lwp_proc->p_vmspace) != NULL) { 758 ru->ru_ixrss += pgtok(vm->vm_tsize); 759 ru->ru_idrss += pgtok(vm->vm_dsize); 760 ru->ru_isrss += pgtok(vm->vm_ssize); 761 rss = pgtok(vmspace_resident_count(vm)); 762 if (ru->ru_maxrss < rss) 763 ru->ru_maxrss = rss; 764 } 765 } 766 } 767 768 /* 769 * Compute number of ticks for the specified amount of time. The 770 * return value is intended to be used in a clock interrupt timed 771 * operation and guarenteed to meet or exceed the requested time. 772 * If the representation overflows, return INT_MAX. The minimum return 773 * value is 1 ticks and the function will average the calculation up. 774 * If any value greater then 0 microseconds is supplied, a value 775 * of at least 2 will be returned to ensure that a near-term clock 776 * interrupt does not cause the timeout to occur (degenerately) early. 777 * 778 * Note that limit checks must take into account microseconds, which is 779 * done simply by using the smaller signed long maximum instead of 780 * the unsigned long maximum. 781 * 782 * If ints have 32 bits, then the maximum value for any timeout in 783 * 10ms ticks is 248 days. 784 */ 785 int 786 tvtohz_high(struct timeval *tv) 787 { 788 int ticks; 789 long sec, usec; 790 791 sec = tv->tv_sec; 792 usec = tv->tv_usec; 793 if (usec < 0) { 794 sec--; 795 usec += 1000000; 796 } 797 if (sec < 0) { 798 #ifdef DIAGNOSTIC 799 if (usec > 0) { 800 sec++; 801 usec -= 1000000; 802 } 803 kprintf("tvtohz_high: negative time difference " 804 "%ld sec %ld usec\n", 805 sec, usec); 806 #endif 807 ticks = 1; 808 } else if (sec <= INT_MAX / hz) { 809 ticks = (int)(sec * hz + 810 ((u_long)usec + (ustick - 1)) / ustick) + 1; 811 } else { 812 ticks = INT_MAX; 813 } 814 return (ticks); 815 } 816 817 int 818 tstohz_high(struct timespec *ts) 819 { 820 int ticks; 821 long sec, nsec; 822 823 sec = ts->tv_sec; 824 nsec = ts->tv_nsec; 825 if (nsec < 0) { 826 sec--; 827 nsec += 1000000000; 828 } 829 if (sec < 0) { 830 #ifdef DIAGNOSTIC 831 if (nsec > 0) { 832 sec++; 833 nsec -= 1000000000; 834 } 835 kprintf("tstohz_high: negative time difference " 836 "%ld sec %ld nsec\n", 837 sec, nsec); 838 #endif 839 ticks = 1; 840 } else if (sec <= INT_MAX / hz) { 841 ticks = (int)(sec * hz + 842 ((u_long)nsec + (nstick - 1)) / nstick) + 1; 843 } else { 844 ticks = INT_MAX; 845 } 846 return (ticks); 847 } 848 849 850 /* 851 * Compute number of ticks for the specified amount of time, erroring on 852 * the side of it being too low to ensure that sleeping the returned number 853 * of ticks will not result in a late return. 854 * 855 * The supplied timeval may not be negative and should be normalized. A 856 * return value of 0 is possible if the timeval converts to less then 857 * 1 tick. 858 * 859 * If ints have 32 bits, then the maximum value for any timeout in 860 * 10ms ticks is 248 days. 861 */ 862 int 863 tvtohz_low(struct timeval *tv) 864 { 865 int ticks; 866 long sec; 867 868 sec = tv->tv_sec; 869 if (sec <= INT_MAX / hz) 870 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick); 871 else 872 ticks = INT_MAX; 873 return (ticks); 874 } 875 876 int 877 tstohz_low(struct timespec *ts) 878 { 879 int ticks; 880 long sec; 881 882 sec = ts->tv_sec; 883 if (sec <= INT_MAX / hz) 884 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick); 885 else 886 ticks = INT_MAX; 887 return (ticks); 888 } 889 890 /* 891 * Start profiling on a process. 892 * 893 * Kernel profiling passes proc0 which never exits and hence 894 * keeps the profile clock running constantly. 895 */ 896 void 897 startprofclock(struct proc *p) 898 { 899 if ((p->p_flag & P_PROFIL) == 0) { 900 p->p_flag |= P_PROFIL; 901 #if 0 /* XXX */ 902 if (++profprocs == 1 && stathz != 0) { 903 crit_enter(); 904 psdiv = psratio; 905 setstatclockrate(profhz); 906 crit_exit(); 907 } 908 #endif 909 } 910 } 911 912 /* 913 * Stop profiling on a process. 914 */ 915 void 916 stopprofclock(struct proc *p) 917 { 918 if (p->p_flag & P_PROFIL) { 919 p->p_flag &= ~P_PROFIL; 920 #if 0 /* XXX */ 921 if (--profprocs == 0 && stathz != 0) { 922 crit_enter(); 923 psdiv = 1; 924 setstatclockrate(stathz); 925 crit_exit(); 926 } 927 #endif 928 } 929 } 930 931 /* 932 * Return information about system clocks. 933 */ 934 static int 935 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 936 { 937 struct kinfo_clockinfo clkinfo; 938 /* 939 * Construct clockinfo structure. 940 */ 941 clkinfo.ci_hz = hz; 942 clkinfo.ci_tick = ustick; 943 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 944 clkinfo.ci_profhz = profhz; 945 clkinfo.ci_stathz = stathz ? stathz : hz; 946 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 947 } 948 949 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 950 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 951 952 /* 953 * We have eight functions for looking at the clock, four for 954 * microseconds and four for nanoseconds. For each there is fast 955 * but less precise version "get{nano|micro}[up]time" which will 956 * return a time which is up to 1/HZ previous to the call, whereas 957 * the raw version "{nano|micro}[up]time" will return a timestamp 958 * which is as precise as possible. The "up" variants return the 959 * time relative to system boot, these are well suited for time 960 * interval measurements. 961 * 962 * Each cpu independantly maintains the current time of day, so all 963 * we need to do to protect ourselves from changes is to do a loop 964 * check on the seconds field changing out from under us. 965 * 966 * The system timer maintains a 32 bit count and due to various issues 967 * it is possible for the calculated delta to occassionally exceed 968 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 969 * multiplication can easily overflow, so we deal with the case. For 970 * uniformity we deal with the case in the usec case too. 971 * 972 * All the [get][micro,nano][time,uptime]() routines are MPSAFE. 973 */ 974 void 975 getmicrouptime(struct timeval *tvp) 976 { 977 struct globaldata *gd = mycpu; 978 sysclock_t delta; 979 980 do { 981 tvp->tv_sec = gd->gd_time_seconds; 982 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 983 } while (tvp->tv_sec != gd->gd_time_seconds); 984 985 if (delta >= sys_cputimer->freq) { 986 tvp->tv_sec += delta / sys_cputimer->freq; 987 delta %= sys_cputimer->freq; 988 } 989 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 990 if (tvp->tv_usec >= 1000000) { 991 tvp->tv_usec -= 1000000; 992 ++tvp->tv_sec; 993 } 994 } 995 996 void 997 getnanouptime(struct timespec *tsp) 998 { 999 struct globaldata *gd = mycpu; 1000 sysclock_t delta; 1001 1002 do { 1003 tsp->tv_sec = gd->gd_time_seconds; 1004 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1005 } while (tsp->tv_sec != gd->gd_time_seconds); 1006 1007 if (delta >= sys_cputimer->freq) { 1008 tsp->tv_sec += delta / sys_cputimer->freq; 1009 delta %= sys_cputimer->freq; 1010 } 1011 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1012 } 1013 1014 void 1015 microuptime(struct timeval *tvp) 1016 { 1017 struct globaldata *gd = mycpu; 1018 sysclock_t delta; 1019 1020 do { 1021 tvp->tv_sec = gd->gd_time_seconds; 1022 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1023 } while (tvp->tv_sec != gd->gd_time_seconds); 1024 1025 if (delta >= sys_cputimer->freq) { 1026 tvp->tv_sec += delta / sys_cputimer->freq; 1027 delta %= sys_cputimer->freq; 1028 } 1029 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1030 } 1031 1032 void 1033 nanouptime(struct timespec *tsp) 1034 { 1035 struct globaldata *gd = mycpu; 1036 sysclock_t delta; 1037 1038 do { 1039 tsp->tv_sec = gd->gd_time_seconds; 1040 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1041 } while (tsp->tv_sec != gd->gd_time_seconds); 1042 1043 if (delta >= sys_cputimer->freq) { 1044 tsp->tv_sec += delta / sys_cputimer->freq; 1045 delta %= sys_cputimer->freq; 1046 } 1047 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1048 } 1049 1050 /* 1051 * realtime routines 1052 */ 1053 void 1054 getmicrotime(struct timeval *tvp) 1055 { 1056 struct globaldata *gd = mycpu; 1057 struct timespec *bt; 1058 sysclock_t delta; 1059 1060 do { 1061 tvp->tv_sec = gd->gd_time_seconds; 1062 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1063 } while (tvp->tv_sec != gd->gd_time_seconds); 1064 1065 if (delta >= sys_cputimer->freq) { 1066 tvp->tv_sec += delta / sys_cputimer->freq; 1067 delta %= sys_cputimer->freq; 1068 } 1069 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1070 1071 bt = &basetime[basetime_index]; 1072 tvp->tv_sec += bt->tv_sec; 1073 tvp->tv_usec += bt->tv_nsec / 1000; 1074 while (tvp->tv_usec >= 1000000) { 1075 tvp->tv_usec -= 1000000; 1076 ++tvp->tv_sec; 1077 } 1078 } 1079 1080 void 1081 getnanotime(struct timespec *tsp) 1082 { 1083 struct globaldata *gd = mycpu; 1084 struct timespec *bt; 1085 sysclock_t delta; 1086 1087 do { 1088 tsp->tv_sec = gd->gd_time_seconds; 1089 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1090 } while (tsp->tv_sec != gd->gd_time_seconds); 1091 1092 if (delta >= sys_cputimer->freq) { 1093 tsp->tv_sec += delta / sys_cputimer->freq; 1094 delta %= sys_cputimer->freq; 1095 } 1096 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1097 1098 bt = &basetime[basetime_index]; 1099 tsp->tv_sec += bt->tv_sec; 1100 tsp->tv_nsec += bt->tv_nsec; 1101 while (tsp->tv_nsec >= 1000000000) { 1102 tsp->tv_nsec -= 1000000000; 1103 ++tsp->tv_sec; 1104 } 1105 } 1106 1107 static void 1108 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1109 { 1110 struct globaldata *gd = mycpu; 1111 sysclock_t delta; 1112 1113 do { 1114 tsp->tv_sec = gd->gd_time_seconds; 1115 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1116 } while (tsp->tv_sec != gd->gd_time_seconds); 1117 1118 if (delta >= sys_cputimer->freq) { 1119 tsp->tv_sec += delta / sys_cputimer->freq; 1120 delta %= sys_cputimer->freq; 1121 } 1122 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1123 1124 tsp->tv_sec += nbt->tv_sec; 1125 tsp->tv_nsec += nbt->tv_nsec; 1126 while (tsp->tv_nsec >= 1000000000) { 1127 tsp->tv_nsec -= 1000000000; 1128 ++tsp->tv_sec; 1129 } 1130 } 1131 1132 1133 void 1134 microtime(struct timeval *tvp) 1135 { 1136 struct globaldata *gd = mycpu; 1137 struct timespec *bt; 1138 sysclock_t delta; 1139 1140 do { 1141 tvp->tv_sec = gd->gd_time_seconds; 1142 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1143 } while (tvp->tv_sec != gd->gd_time_seconds); 1144 1145 if (delta >= sys_cputimer->freq) { 1146 tvp->tv_sec += delta / sys_cputimer->freq; 1147 delta %= sys_cputimer->freq; 1148 } 1149 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1150 1151 bt = &basetime[basetime_index]; 1152 tvp->tv_sec += bt->tv_sec; 1153 tvp->tv_usec += bt->tv_nsec / 1000; 1154 while (tvp->tv_usec >= 1000000) { 1155 tvp->tv_usec -= 1000000; 1156 ++tvp->tv_sec; 1157 } 1158 } 1159 1160 void 1161 nanotime(struct timespec *tsp) 1162 { 1163 struct globaldata *gd = mycpu; 1164 struct timespec *bt; 1165 sysclock_t delta; 1166 1167 do { 1168 tsp->tv_sec = gd->gd_time_seconds; 1169 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1170 } while (tsp->tv_sec != gd->gd_time_seconds); 1171 1172 if (delta >= sys_cputimer->freq) { 1173 tsp->tv_sec += delta / sys_cputimer->freq; 1174 delta %= sys_cputimer->freq; 1175 } 1176 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1177 1178 bt = &basetime[basetime_index]; 1179 tsp->tv_sec += bt->tv_sec; 1180 tsp->tv_nsec += bt->tv_nsec; 1181 while (tsp->tv_nsec >= 1000000000) { 1182 tsp->tv_nsec -= 1000000000; 1183 ++tsp->tv_sec; 1184 } 1185 } 1186 1187 /* 1188 * note: this is not exactly synchronized with real time. To do that we 1189 * would have to do what microtime does and check for a nanoseconds overflow. 1190 */ 1191 time_t 1192 get_approximate_time_t(void) 1193 { 1194 struct globaldata *gd = mycpu; 1195 struct timespec *bt; 1196 1197 bt = &basetime[basetime_index]; 1198 return(gd->gd_time_seconds + bt->tv_sec); 1199 } 1200 1201 int 1202 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1203 { 1204 pps_params_t *app; 1205 struct pps_fetch_args *fapi; 1206 #ifdef PPS_SYNC 1207 struct pps_kcbind_args *kapi; 1208 #endif 1209 1210 switch (cmd) { 1211 case PPS_IOC_CREATE: 1212 return (0); 1213 case PPS_IOC_DESTROY: 1214 return (0); 1215 case PPS_IOC_SETPARAMS: 1216 app = (pps_params_t *)data; 1217 if (app->mode & ~pps->ppscap) 1218 return (EINVAL); 1219 pps->ppsparam = *app; 1220 return (0); 1221 case PPS_IOC_GETPARAMS: 1222 app = (pps_params_t *)data; 1223 *app = pps->ppsparam; 1224 app->api_version = PPS_API_VERS_1; 1225 return (0); 1226 case PPS_IOC_GETCAP: 1227 *(int*)data = pps->ppscap; 1228 return (0); 1229 case PPS_IOC_FETCH: 1230 fapi = (struct pps_fetch_args *)data; 1231 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1232 return (EINVAL); 1233 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1234 return (EOPNOTSUPP); 1235 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1236 fapi->pps_info_buf = pps->ppsinfo; 1237 return (0); 1238 case PPS_IOC_KCBIND: 1239 #ifdef PPS_SYNC 1240 kapi = (struct pps_kcbind_args *)data; 1241 /* XXX Only root should be able to do this */ 1242 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1243 return (EINVAL); 1244 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1245 return (EINVAL); 1246 if (kapi->edge & ~pps->ppscap) 1247 return (EINVAL); 1248 pps->kcmode = kapi->edge; 1249 return (0); 1250 #else 1251 return (EOPNOTSUPP); 1252 #endif 1253 default: 1254 return (ENOTTY); 1255 } 1256 } 1257 1258 void 1259 pps_init(struct pps_state *pps) 1260 { 1261 pps->ppscap |= PPS_TSFMT_TSPEC; 1262 if (pps->ppscap & PPS_CAPTUREASSERT) 1263 pps->ppscap |= PPS_OFFSETASSERT; 1264 if (pps->ppscap & PPS_CAPTURECLEAR) 1265 pps->ppscap |= PPS_OFFSETCLEAR; 1266 } 1267 1268 void 1269 pps_event(struct pps_state *pps, sysclock_t count, int event) 1270 { 1271 struct globaldata *gd; 1272 struct timespec *tsp; 1273 struct timespec *osp; 1274 struct timespec *bt; 1275 struct timespec ts; 1276 sysclock_t *pcount; 1277 #ifdef PPS_SYNC 1278 sysclock_t tcount; 1279 #endif 1280 sysclock_t delta; 1281 pps_seq_t *pseq; 1282 int foff; 1283 int fhard; 1284 1285 gd = mycpu; 1286 1287 /* Things would be easier with arrays... */ 1288 if (event == PPS_CAPTUREASSERT) { 1289 tsp = &pps->ppsinfo.assert_timestamp; 1290 osp = &pps->ppsparam.assert_offset; 1291 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1292 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1293 pcount = &pps->ppscount[0]; 1294 pseq = &pps->ppsinfo.assert_sequence; 1295 } else { 1296 tsp = &pps->ppsinfo.clear_timestamp; 1297 osp = &pps->ppsparam.clear_offset; 1298 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1299 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1300 pcount = &pps->ppscount[1]; 1301 pseq = &pps->ppsinfo.clear_sequence; 1302 } 1303 1304 /* Nothing really happened */ 1305 if (*pcount == count) 1306 return; 1307 1308 *pcount = count; 1309 1310 do { 1311 ts.tv_sec = gd->gd_time_seconds; 1312 delta = count - gd->gd_cpuclock_base; 1313 } while (ts.tv_sec != gd->gd_time_seconds); 1314 1315 if (delta >= sys_cputimer->freq) { 1316 ts.tv_sec += delta / sys_cputimer->freq; 1317 delta %= sys_cputimer->freq; 1318 } 1319 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1320 bt = &basetime[basetime_index]; 1321 ts.tv_sec += bt->tv_sec; 1322 ts.tv_nsec += bt->tv_nsec; 1323 while (ts.tv_nsec >= 1000000000) { 1324 ts.tv_nsec -= 1000000000; 1325 ++ts.tv_sec; 1326 } 1327 1328 (*pseq)++; 1329 *tsp = ts; 1330 1331 if (foff) { 1332 timespecadd(tsp, osp); 1333 if (tsp->tv_nsec < 0) { 1334 tsp->tv_nsec += 1000000000; 1335 tsp->tv_sec -= 1; 1336 } 1337 } 1338 #ifdef PPS_SYNC 1339 if (fhard) { 1340 /* magic, at its best... */ 1341 tcount = count - pps->ppscount[2]; 1342 pps->ppscount[2] = count; 1343 if (tcount >= sys_cputimer->freq) { 1344 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1345 sys_cputimer->freq64_nsec * 1346 (tcount % sys_cputimer->freq)) >> 32; 1347 } else { 1348 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1349 } 1350 hardpps(tsp, delta); 1351 } 1352 #endif 1353 } 1354 1355 /* 1356 * Return the tsc target value for a delay of (ns). 1357 * 1358 * Returns -1 if the TSC is not supported. 1359 */ 1360 int64_t 1361 tsc_get_target(int ns) 1362 { 1363 #if defined(_RDTSC_SUPPORTED_) 1364 if (cpu_feature & CPUID_TSC) { 1365 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000); 1366 } 1367 #endif 1368 return(-1); 1369 } 1370 1371 /* 1372 * Compare the tsc against the passed target 1373 * 1374 * Returns +1 if the target has been reached 1375 * Returns 0 if the target has not yet been reached 1376 * Returns -1 if the TSC is not supported. 1377 * 1378 * Typical use: while (tsc_test_target(target) == 0) { ...poll... } 1379 */ 1380 int 1381 tsc_test_target(int64_t target) 1382 { 1383 #if defined(_RDTSC_SUPPORTED_) 1384 if (cpu_feature & CPUID_TSC) { 1385 if ((int64_t)(target - rdtsc()) <= 0) 1386 return(1); 1387 return(0); 1388 } 1389 #endif 1390 return(-1); 1391 } 1392