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> 100 101 #include <machine/cpu.h> 102 #include <machine/limits.h> 103 #include <machine/smp.h> 104 #include <machine/cpufunc.h> 105 #include <machine/specialreg.h> 106 #include <machine/clock.h> 107 108 #ifdef GPROF 109 #include <sys/gmon.h> 110 #endif 111 112 #ifdef DEVICE_POLLING 113 extern void init_device_poll_pcpu(int); 114 #endif 115 116 #ifdef IFPOLL_ENABLE 117 extern void ifpoll_init_pcpu(int); 118 #endif 119 120 #ifdef DEBUG_PCTRACK 121 static void do_pctrack(struct intrframe *frame, int which); 122 #endif 123 124 static void initclocks (void *dummy); 125 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 126 127 /* 128 * Some of these don't belong here, but it's easiest to concentrate them. 129 * Note that cpu_time counts in microseconds, but most userland programs 130 * just compare relative times against the total by delta. 131 */ 132 struct kinfo_cputime cputime_percpu[MAXCPU]; 133 #ifdef DEBUG_PCTRACK 134 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE }; 135 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE]; 136 #endif 137 138 #ifdef SMP 139 static int 140 sysctl_cputime(SYSCTL_HANDLER_ARGS) 141 { 142 int cpu, error = 0; 143 size_t size = sizeof(struct kinfo_cputime); 144 145 for (cpu = 0; cpu < ncpus; ++cpu) { 146 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size))) 147 break; 148 } 149 150 return (error); 151 } 152 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 153 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 154 #else 155 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime, 156 "CPU time statistics"); 157 #endif 158 159 static int 160 sysctl_cp_time(SYSCTL_HANDLER_ARGS) 161 { 162 long cpu_states[5] = {0}; 163 int cpu, error = 0; 164 size_t size = sizeof(cpu_states); 165 166 for (cpu = 0; cpu < ncpus; ++cpu) { 167 cpu_states[0] += cputime_percpu[cpu].cp_user; 168 cpu_states[1] += cputime_percpu[cpu].cp_nice; 169 cpu_states[2] += cputime_percpu[cpu].cp_sys; 170 cpu_states[3] += cputime_percpu[cpu].cp_intr; 171 cpu_states[4] += cputime_percpu[cpu].cp_idle; 172 } 173 174 error = SYSCTL_OUT(req, cpu_states, size); 175 176 return (error); 177 } 178 179 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0, 180 sysctl_cp_time, "LU", "CPU time statistics"); 181 182 /* 183 * boottime is used to calculate the 'real' uptime. Do not confuse this with 184 * microuptime(). microtime() is not drift compensated. The real uptime 185 * with compensation is nanotime() - bootime. boottime is recalculated 186 * whenever the real time is set based on the compensated elapsed time 187 * in seconds (gd->gd_time_seconds). 188 * 189 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 190 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 191 * the real time. 192 */ 193 struct timespec boottime; /* boot time (realtime) for reference only */ 194 time_t time_second; /* read-only 'passive' uptime in seconds */ 195 196 /* 197 * basetime is used to calculate the compensated real time of day. The 198 * basetime can be modified on a per-tick basis by the adjtime(), 199 * ntp_adjtime(), and sysctl-based time correction APIs. 200 * 201 * Note that frequency corrections can also be made by adjusting 202 * gd_cpuclock_base. 203 * 204 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 205 * used on both SMP and UP systems to avoid MP races between cpu's and 206 * interrupt races on UP systems. 207 */ 208 #define BASETIME_ARYSIZE 16 209 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 210 static struct timespec basetime[BASETIME_ARYSIZE]; 211 static volatile int basetime_index; 212 213 static int 214 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 215 { 216 struct timespec *bt; 217 int error; 218 int index; 219 220 /* 221 * Because basetime data and index may be updated by another cpu, 222 * a load fence is required to ensure that the data we read has 223 * not been speculatively read relative to a possibly updated index. 224 */ 225 index = basetime_index; 226 cpu_lfence(); 227 bt = &basetime[index]; 228 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 229 return (error); 230 } 231 232 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 233 &boottime, timespec, "System boottime"); 234 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 235 sysctl_get_basetime, "S,timespec", "System basetime"); 236 237 static void hardclock(systimer_t info, int, struct intrframe *frame); 238 static void statclock(systimer_t info, int, struct intrframe *frame); 239 static void schedclock(systimer_t info, int, struct intrframe *frame); 240 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 241 242 int ticks; /* system master ticks at hz */ 243 int clocks_running; /* tsleep/timeout clocks operational */ 244 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 245 int64_t nsec_acc; /* accumulator */ 246 247 /* NTPD time correction fields */ 248 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 249 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 250 int64_t ntp_delta; /* one-time correction in nsec */ 251 int64_t ntp_big_delta = 1000000000; 252 int32_t ntp_tick_delta; /* current adjustment rate */ 253 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 254 time_t ntp_leap_second; /* time of next leap second */ 255 int ntp_leap_insert; /* whether to insert or remove a second */ 256 257 /* 258 * Finish initializing clock frequencies and start all clocks running. 259 */ 260 /* ARGSUSED*/ 261 static void 262 initclocks(void *dummy) 263 { 264 /*psratio = profhz / stathz;*/ 265 initclocks_pcpu(); 266 clocks_running = 1; 267 } 268 269 /* 270 * Called on a per-cpu basis 271 */ 272 void 273 initclocks_pcpu(void) 274 { 275 struct globaldata *gd = mycpu; 276 277 crit_enter(); 278 if (gd->gd_cpuid == 0) { 279 gd->gd_time_seconds = 1; 280 gd->gd_cpuclock_base = sys_cputimer->count(); 281 } else { 282 /* XXX */ 283 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 284 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 285 } 286 287 systimer_intr_enable(); 288 289 #ifdef DEVICE_POLLING 290 init_device_poll_pcpu(gd->gd_cpuid); 291 #endif 292 293 #ifdef IFPOLL_ENABLE 294 ifpoll_init_pcpu(gd->gd_cpuid); 295 #endif 296 297 /* 298 * Use a non-queued periodic systimer to prevent multiple ticks from 299 * building up if the sysclock jumps forward (8254 gets reset). The 300 * sysclock will never jump backwards. Our time sync is based on 301 * the actual sysclock, not the ticks count. 302 */ 303 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz); 304 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz); 305 /* XXX correct the frequency for scheduler / estcpu tests */ 306 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock, 307 NULL, ESTCPUFREQ); 308 crit_exit(); 309 } 310 311 /* 312 * This sets the current real time of day. Timespecs are in seconds and 313 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 314 * instead we adjust basetime so basetime + gd_* results in the current 315 * time of day. This way the gd_* fields are guarenteed to represent 316 * a monotonically increasing 'uptime' value. 317 * 318 * When set_timeofday() is called from userland, the system call forces it 319 * onto cpu #0 since only cpu #0 can update basetime_index. 320 */ 321 void 322 set_timeofday(struct timespec *ts) 323 { 324 struct timespec *nbt; 325 int ni; 326 327 /* 328 * XXX SMP / non-atomic basetime updates 329 */ 330 crit_enter(); 331 ni = (basetime_index + 1) & BASETIME_ARYMASK; 332 nbt = &basetime[ni]; 333 nanouptime(nbt); 334 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 335 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 336 if (nbt->tv_nsec < 0) { 337 nbt->tv_nsec += 1000000000; 338 --nbt->tv_sec; 339 } 340 341 /* 342 * Note that basetime diverges from boottime as the clock drift is 343 * compensated for, so we cannot do away with boottime. When setting 344 * the absolute time of day the drift is 0 (for an instant) and we 345 * can simply assign boottime to basetime. 346 * 347 * Note that nanouptime() is based on gd_time_seconds which is drift 348 * compensated up to a point (it is guarenteed to remain monotonically 349 * increasing). gd_time_seconds is thus our best uptime guess and 350 * suitable for use in the boottime calculation. It is already taken 351 * into account in the basetime calculation above. 352 */ 353 boottime.tv_sec = nbt->tv_sec; 354 ntp_delta = 0; 355 356 /* 357 * We now have a new basetime, make sure all other cpus have it, 358 * then update the index. 359 */ 360 cpu_sfence(); 361 basetime_index = ni; 362 363 crit_exit(); 364 } 365 366 /* 367 * Each cpu has its own hardclock, but we only increments ticks and softticks 368 * on cpu #0. 369 * 370 * NOTE! systimer! the MP lock might not be held here. We can only safely 371 * manipulate objects owned by the current cpu. 372 */ 373 static void 374 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame) 375 { 376 sysclock_t cputicks; 377 struct proc *p; 378 struct globaldata *gd = mycpu; 379 380 /* 381 * Realtime updates are per-cpu. Note that timer corrections as 382 * returned by microtime() and friends make an additional adjustment 383 * using a system-wise 'basetime', but the running time is always 384 * taken from the per-cpu globaldata area. Since the same clock 385 * is distributing (XXX SMP) to all cpus, the per-cpu timebases 386 * stay in synch. 387 * 388 * Note that we never allow info->time (aka gd->gd_hardclock.time) 389 * to reverse index gd_cpuclock_base, but that it is possible for 390 * it to temporarily get behind in the seconds if something in the 391 * system locks interrupts for a long period of time. Since periodic 392 * timers count events, though everything should resynch again 393 * immediately. 394 */ 395 cputicks = info->time - gd->gd_cpuclock_base; 396 if (cputicks >= sys_cputimer->freq) { 397 ++gd->gd_time_seconds; 398 gd->gd_cpuclock_base += sys_cputimer->freq; 399 } 400 401 /* 402 * The system-wide ticks counter and NTP related timedelta/tickdelta 403 * adjustments only occur on cpu #0. NTP adjustments are accomplished 404 * by updating basetime. 405 */ 406 if (gd->gd_cpuid == 0) { 407 struct timespec *nbt; 408 struct timespec nts; 409 int leap; 410 int ni; 411 412 ++ticks; 413 414 #if 0 415 if (tco->tc_poll_pps) 416 tco->tc_poll_pps(tco); 417 #endif 418 419 /* 420 * Calculate the new basetime index. We are in a critical section 421 * on cpu #0 and can safely play with basetime_index. Start 422 * with the current basetime and then make adjustments. 423 */ 424 ni = (basetime_index + 1) & BASETIME_ARYMASK; 425 nbt = &basetime[ni]; 426 *nbt = basetime[basetime_index]; 427 428 /* 429 * Apply adjtime corrections. (adjtime() API) 430 * 431 * adjtime() only runs on cpu #0 so our critical section is 432 * sufficient to access these variables. 433 */ 434 if (ntp_delta != 0) { 435 nbt->tv_nsec += ntp_tick_delta; 436 ntp_delta -= ntp_tick_delta; 437 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 438 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 439 ntp_tick_delta = ntp_delta; 440 } 441 } 442 443 /* 444 * Apply permanent frequency corrections. (sysctl API) 445 */ 446 if (ntp_tick_permanent != 0) { 447 ntp_tick_acc += ntp_tick_permanent; 448 if (ntp_tick_acc >= (1LL << 32)) { 449 nbt->tv_nsec += ntp_tick_acc >> 32; 450 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 451 } else if (ntp_tick_acc <= -(1LL << 32)) { 452 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 453 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 454 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 455 } 456 } 457 458 if (nbt->tv_nsec >= 1000000000) { 459 nbt->tv_sec++; 460 nbt->tv_nsec -= 1000000000; 461 } else if (nbt->tv_nsec < 0) { 462 nbt->tv_sec--; 463 nbt->tv_nsec += 1000000000; 464 } 465 466 /* 467 * Another per-tick compensation. (for ntp_adjtime() API) 468 */ 469 if (nsec_adj != 0) { 470 nsec_acc += nsec_adj; 471 if (nsec_acc >= 0x100000000LL) { 472 nbt->tv_nsec += nsec_acc >> 32; 473 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 474 } else if (nsec_acc <= -0x100000000LL) { 475 nbt->tv_nsec -= -nsec_acc >> 32; 476 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 477 } 478 if (nbt->tv_nsec >= 1000000000) { 479 nbt->tv_nsec -= 1000000000; 480 ++nbt->tv_sec; 481 } else if (nbt->tv_nsec < 0) { 482 nbt->tv_nsec += 1000000000; 483 --nbt->tv_sec; 484 } 485 } 486 487 /************************************************************ 488 * LEAP SECOND CORRECTION * 489 ************************************************************ 490 * 491 * Taking into account all the corrections made above, figure 492 * out the new real time. If the seconds field has changed 493 * then apply any pending leap-second corrections. 494 */ 495 getnanotime_nbt(nbt, &nts); 496 497 if (time_second != nts.tv_sec) { 498 /* 499 * Apply leap second (sysctl API). Adjust nts for changes 500 * so we do not have to call getnanotime_nbt again. 501 */ 502 if (ntp_leap_second) { 503 if (ntp_leap_second == nts.tv_sec) { 504 if (ntp_leap_insert) { 505 nbt->tv_sec++; 506 nts.tv_sec++; 507 } else { 508 nbt->tv_sec--; 509 nts.tv_sec--; 510 } 511 ntp_leap_second--; 512 } 513 } 514 515 /* 516 * Apply leap second (ntp_adjtime() API), calculate a new 517 * nsec_adj field. ntp_update_second() returns nsec_adj 518 * as a per-second value but we need it as a per-tick value. 519 */ 520 leap = ntp_update_second(time_second, &nsec_adj); 521 nsec_adj /= hz; 522 nbt->tv_sec += leap; 523 nts.tv_sec += leap; 524 525 /* 526 * Update the time_second 'approximate time' global. 527 */ 528 time_second = nts.tv_sec; 529 } 530 531 /* 532 * Finally, our new basetime is ready to go live! 533 */ 534 cpu_sfence(); 535 basetime_index = ni; 536 537 /* 538 * Figure out how badly the system is starved for memory 539 */ 540 vm_fault_ratecheck(); 541 } 542 543 /* 544 * lwkt thread scheduler fair queueing 545 */ 546 lwkt_schedulerclock(curthread); 547 548 /* 549 * softticks are handled for all cpus 550 */ 551 hardclock_softtick(gd); 552 553 /* 554 * ITimer handling is per-tick, per-cpu. 555 * 556 * We must acquire the per-process token in order for ksignal() 557 * to be non-blocking. 558 */ 559 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) { 560 crit_enter_hard(); 561 if (frame && CLKF_USERMODE(frame) && 562 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 563 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) 564 ksignal(p, SIGVTALRM); 565 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 566 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) 567 ksignal(p, SIGPROF); 568 crit_exit_hard(); 569 lwkt_reltoken(&p->p_token); 570 } 571 setdelayed(); 572 } 573 574 /* 575 * The statistics clock typically runs at a 125Hz rate, and is intended 576 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 577 * 578 * NOTE! systimer! the MP lock might not be held here. We can only safely 579 * manipulate objects owned by the current cpu. 580 * 581 * The stats clock is responsible for grabbing a profiling sample. 582 * Most of the statistics are only used by user-level statistics programs. 583 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 584 * p->p_estcpu. 585 * 586 * Like the other clocks, the stat clock is called from what is effectively 587 * a fast interrupt, so the context should be the thread/process that got 588 * interrupted. 589 */ 590 static void 591 statclock(systimer_t info, int in_ipi, struct intrframe *frame) 592 { 593 #ifdef GPROF 594 struct gmonparam *g; 595 int i; 596 #endif 597 thread_t td; 598 struct proc *p; 599 int bump; 600 struct timeval tv; 601 struct timeval *stv; 602 603 /* 604 * How big was our timeslice relative to the last time? 605 */ 606 microuptime(&tv); /* mpsafe */ 607 stv = &mycpu->gd_stattv; 608 if (stv->tv_sec == 0) { 609 bump = 1; 610 } else { 611 bump = tv.tv_usec - stv->tv_usec + 612 (tv.tv_sec - stv->tv_sec) * 1000000; 613 if (bump < 0) 614 bump = 0; 615 if (bump > 1000000) 616 bump = 1000000; 617 } 618 *stv = tv; 619 620 td = curthread; 621 p = td->td_proc; 622 623 if (frame && CLKF_USERMODE(frame)) { 624 /* 625 * Came from userland, handle user time and deal with 626 * possible process. 627 */ 628 if (p && (p->p_flag & P_PROFIL)) 629 addupc_intr(p, CLKF_PC(frame), 1); 630 td->td_uticks += bump; 631 632 /* 633 * Charge the time as appropriate 634 */ 635 if (p && p->p_nice > NZERO) 636 cpu_time.cp_nice += bump; 637 else 638 cpu_time.cp_user += bump; 639 } else { 640 int intr_nest = mycpu->gd_intr_nesting_level; 641 642 if (in_ipi) { 643 /* 644 * IPI processing code will bump gd_intr_nesting_level 645 * up by one, which breaks following CLKF_INTR testing, 646 * so we substract it by one here. 647 */ 648 --intr_nest; 649 } 650 #ifdef GPROF 651 /* 652 * Kernel statistics are just like addupc_intr, only easier. 653 */ 654 g = &_gmonparam; 655 if (g->state == GMON_PROF_ON && frame) { 656 i = CLKF_PC(frame) - g->lowpc; 657 if (i < g->textsize) { 658 i /= HISTFRACTION * sizeof(*g->kcount); 659 g->kcount[i]++; 660 } 661 } 662 #endif 663 /* 664 * Came from kernel mode, so we were: 665 * - handling an interrupt, 666 * - doing syscall or trap work on behalf of the current 667 * user process, or 668 * - spinning in the idle loop. 669 * Whichever it is, charge the time as appropriate. 670 * Note that we charge interrupts to the current process, 671 * regardless of whether they are ``for'' that process, 672 * so that we know how much of its real time was spent 673 * in ``non-process'' (i.e., interrupt) work. 674 * 675 * XXX assume system if frame is NULL. A NULL frame 676 * can occur if ipi processing is done from a crit_exit(). 677 */ 678 if (frame && CLKF_INTR(intr_nest)) 679 td->td_iticks += bump; 680 else 681 td->td_sticks += bump; 682 683 if (frame && CLKF_INTR(intr_nest)) { 684 #ifdef DEBUG_PCTRACK 685 do_pctrack(frame, PCTRACK_INT); 686 #endif 687 cpu_time.cp_intr += bump; 688 } else { 689 if (td == &mycpu->gd_idlethread) { 690 cpu_time.cp_idle += bump; 691 } else { 692 #ifdef DEBUG_PCTRACK 693 if (frame) 694 do_pctrack(frame, PCTRACK_SYS); 695 #endif 696 cpu_time.cp_sys += bump; 697 } 698 } 699 } 700 } 701 702 #ifdef DEBUG_PCTRACK 703 /* 704 * Sample the PC when in the kernel or in an interrupt. User code can 705 * retrieve the information and generate a histogram or other output. 706 */ 707 708 static void 709 do_pctrack(struct intrframe *frame, int which) 710 { 711 struct kinfo_pctrack *pctrack; 712 713 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 714 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 715 (void *)CLKF_PC(frame); 716 ++pctrack->pc_index; 717 } 718 719 static int 720 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 721 { 722 struct kinfo_pcheader head; 723 int error; 724 int cpu; 725 int ntrack; 726 727 head.pc_ntrack = PCTRACK_SIZE; 728 head.pc_arysize = PCTRACK_ARYSIZE; 729 730 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 731 return (error); 732 733 for (cpu = 0; cpu < ncpus; ++cpu) { 734 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 735 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 736 sizeof(struct kinfo_pctrack)); 737 if (error) 738 break; 739 } 740 if (error) 741 break; 742 } 743 return (error); 744 } 745 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 746 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 747 748 #endif 749 750 /* 751 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 752 * the MP lock might not be held. We can safely manipulate parts of curproc 753 * but that's about it. 754 * 755 * Each cpu has its own scheduler clock. 756 */ 757 static void 758 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame) 759 { 760 struct lwp *lp; 761 struct rusage *ru; 762 struct vmspace *vm; 763 long rss; 764 765 if ((lp = lwkt_preempted_proc()) != NULL) { 766 /* 767 * Account for cpu time used and hit the scheduler. Note 768 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 769 * HERE. 770 */ 771 ++lp->lwp_cpticks; 772 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic, 773 info->time); 774 } 775 if ((lp = curthread->td_lwp) != NULL) { 776 /* 777 * Update resource usage integrals and maximums. 778 */ 779 if ((ru = &lp->lwp_proc->p_ru) && 780 (vm = lp->lwp_proc->p_vmspace) != NULL) { 781 ru->ru_ixrss += pgtok(vm->vm_tsize); 782 ru->ru_idrss += pgtok(vm->vm_dsize); 783 ru->ru_isrss += pgtok(vm->vm_ssize); 784 if (lwkt_trytoken(&vm->vm_map.token)) { 785 rss = pgtok(vmspace_resident_count(vm)); 786 if (ru->ru_maxrss < rss) 787 ru->ru_maxrss = rss; 788 lwkt_reltoken(&vm->vm_map.token); 789 } 790 } 791 } 792 } 793 794 /* 795 * Compute number of ticks for the specified amount of time. The 796 * return value is intended to be used in a clock interrupt timed 797 * operation and guarenteed to meet or exceed the requested time. 798 * If the representation overflows, return INT_MAX. The minimum return 799 * value is 1 ticks and the function will average the calculation up. 800 * If any value greater then 0 microseconds is supplied, a value 801 * of at least 2 will be returned to ensure that a near-term clock 802 * interrupt does not cause the timeout to occur (degenerately) early. 803 * 804 * Note that limit checks must take into account microseconds, which is 805 * done simply by using the smaller signed long maximum instead of 806 * the unsigned long maximum. 807 * 808 * If ints have 32 bits, then the maximum value for any timeout in 809 * 10ms ticks is 248 days. 810 */ 811 int 812 tvtohz_high(struct timeval *tv) 813 { 814 int ticks; 815 long sec, usec; 816 817 sec = tv->tv_sec; 818 usec = tv->tv_usec; 819 if (usec < 0) { 820 sec--; 821 usec += 1000000; 822 } 823 if (sec < 0) { 824 #ifdef DIAGNOSTIC 825 if (usec > 0) { 826 sec++; 827 usec -= 1000000; 828 } 829 kprintf("tvtohz_high: negative time difference " 830 "%ld sec %ld usec\n", 831 sec, usec); 832 #endif 833 ticks = 1; 834 } else if (sec <= INT_MAX / hz) { 835 ticks = (int)(sec * hz + 836 ((u_long)usec + (ustick - 1)) / ustick) + 1; 837 } else { 838 ticks = INT_MAX; 839 } 840 return (ticks); 841 } 842 843 int 844 tstohz_high(struct timespec *ts) 845 { 846 int ticks; 847 long sec, nsec; 848 849 sec = ts->tv_sec; 850 nsec = ts->tv_nsec; 851 if (nsec < 0) { 852 sec--; 853 nsec += 1000000000; 854 } 855 if (sec < 0) { 856 #ifdef DIAGNOSTIC 857 if (nsec > 0) { 858 sec++; 859 nsec -= 1000000000; 860 } 861 kprintf("tstohz_high: negative time difference " 862 "%ld sec %ld nsec\n", 863 sec, nsec); 864 #endif 865 ticks = 1; 866 } else if (sec <= INT_MAX / hz) { 867 ticks = (int)(sec * hz + 868 ((u_long)nsec + (nstick - 1)) / nstick) + 1; 869 } else { 870 ticks = INT_MAX; 871 } 872 return (ticks); 873 } 874 875 876 /* 877 * Compute number of ticks for the specified amount of time, erroring on 878 * the side of it being too low to ensure that sleeping the returned number 879 * of ticks will not result in a late return. 880 * 881 * The supplied timeval may not be negative and should be normalized. A 882 * return value of 0 is possible if the timeval converts to less then 883 * 1 tick. 884 * 885 * If ints have 32 bits, then the maximum value for any timeout in 886 * 10ms ticks is 248 days. 887 */ 888 int 889 tvtohz_low(struct timeval *tv) 890 { 891 int ticks; 892 long sec; 893 894 sec = tv->tv_sec; 895 if (sec <= INT_MAX / hz) 896 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick); 897 else 898 ticks = INT_MAX; 899 return (ticks); 900 } 901 902 int 903 tstohz_low(struct timespec *ts) 904 { 905 int ticks; 906 long sec; 907 908 sec = ts->tv_sec; 909 if (sec <= INT_MAX / hz) 910 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick); 911 else 912 ticks = INT_MAX; 913 return (ticks); 914 } 915 916 /* 917 * Start profiling on a process. 918 * 919 * Kernel profiling passes proc0 which never exits and hence 920 * keeps the profile clock running constantly. 921 */ 922 void 923 startprofclock(struct proc *p) 924 { 925 if ((p->p_flag & P_PROFIL) == 0) { 926 p->p_flag |= P_PROFIL; 927 #if 0 /* XXX */ 928 if (++profprocs == 1 && stathz != 0) { 929 crit_enter(); 930 psdiv = psratio; 931 setstatclockrate(profhz); 932 crit_exit(); 933 } 934 #endif 935 } 936 } 937 938 /* 939 * Stop profiling on a process. 940 * 941 * caller must hold p->p_token 942 */ 943 void 944 stopprofclock(struct proc *p) 945 { 946 if (p->p_flag & P_PROFIL) { 947 p->p_flag &= ~P_PROFIL; 948 #if 0 /* XXX */ 949 if (--profprocs == 0 && stathz != 0) { 950 crit_enter(); 951 psdiv = 1; 952 setstatclockrate(stathz); 953 crit_exit(); 954 } 955 #endif 956 } 957 } 958 959 /* 960 * Return information about system clocks. 961 */ 962 static int 963 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 964 { 965 struct kinfo_clockinfo clkinfo; 966 /* 967 * Construct clockinfo structure. 968 */ 969 clkinfo.ci_hz = hz; 970 clkinfo.ci_tick = ustick; 971 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 972 clkinfo.ci_profhz = profhz; 973 clkinfo.ci_stathz = stathz ? stathz : hz; 974 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 975 } 976 977 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 978 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 979 980 /* 981 * We have eight functions for looking at the clock, four for 982 * microseconds and four for nanoseconds. For each there is fast 983 * but less precise version "get{nano|micro}[up]time" which will 984 * return a time which is up to 1/HZ previous to the call, whereas 985 * the raw version "{nano|micro}[up]time" will return a timestamp 986 * which is as precise as possible. The "up" variants return the 987 * time relative to system boot, these are well suited for time 988 * interval measurements. 989 * 990 * Each cpu independantly maintains the current time of day, so all 991 * we need to do to protect ourselves from changes is to do a loop 992 * check on the seconds field changing out from under us. 993 * 994 * The system timer maintains a 32 bit count and due to various issues 995 * it is possible for the calculated delta to occassionally exceed 996 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 997 * multiplication can easily overflow, so we deal with the case. For 998 * uniformity we deal with the case in the usec case too. 999 * 1000 * All the [get][micro,nano][time,uptime]() routines are MPSAFE. 1001 */ 1002 void 1003 getmicrouptime(struct timeval *tvp) 1004 { 1005 struct globaldata *gd = mycpu; 1006 sysclock_t delta; 1007 1008 do { 1009 tvp->tv_sec = gd->gd_time_seconds; 1010 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1011 } while (tvp->tv_sec != gd->gd_time_seconds); 1012 1013 if (delta >= sys_cputimer->freq) { 1014 tvp->tv_sec += delta / sys_cputimer->freq; 1015 delta %= sys_cputimer->freq; 1016 } 1017 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1018 if (tvp->tv_usec >= 1000000) { 1019 tvp->tv_usec -= 1000000; 1020 ++tvp->tv_sec; 1021 } 1022 } 1023 1024 void 1025 getnanouptime(struct timespec *tsp) 1026 { 1027 struct globaldata *gd = mycpu; 1028 sysclock_t delta; 1029 1030 do { 1031 tsp->tv_sec = gd->gd_time_seconds; 1032 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1033 } while (tsp->tv_sec != gd->gd_time_seconds); 1034 1035 if (delta >= sys_cputimer->freq) { 1036 tsp->tv_sec += delta / sys_cputimer->freq; 1037 delta %= sys_cputimer->freq; 1038 } 1039 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1040 } 1041 1042 void 1043 microuptime(struct timeval *tvp) 1044 { 1045 struct globaldata *gd = mycpu; 1046 sysclock_t delta; 1047 1048 do { 1049 tvp->tv_sec = gd->gd_time_seconds; 1050 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1051 } while (tvp->tv_sec != gd->gd_time_seconds); 1052 1053 if (delta >= sys_cputimer->freq) { 1054 tvp->tv_sec += delta / sys_cputimer->freq; 1055 delta %= sys_cputimer->freq; 1056 } 1057 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1058 } 1059 1060 void 1061 nanouptime(struct timespec *tsp) 1062 { 1063 struct globaldata *gd = mycpu; 1064 sysclock_t delta; 1065 1066 do { 1067 tsp->tv_sec = gd->gd_time_seconds; 1068 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1069 } while (tsp->tv_sec != gd->gd_time_seconds); 1070 1071 if (delta >= sys_cputimer->freq) { 1072 tsp->tv_sec += delta / sys_cputimer->freq; 1073 delta %= sys_cputimer->freq; 1074 } 1075 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1076 } 1077 1078 /* 1079 * realtime routines 1080 */ 1081 void 1082 getmicrotime(struct timeval *tvp) 1083 { 1084 struct globaldata *gd = mycpu; 1085 struct timespec *bt; 1086 sysclock_t delta; 1087 1088 do { 1089 tvp->tv_sec = gd->gd_time_seconds; 1090 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1091 } while (tvp->tv_sec != gd->gd_time_seconds); 1092 1093 if (delta >= sys_cputimer->freq) { 1094 tvp->tv_sec += delta / sys_cputimer->freq; 1095 delta %= sys_cputimer->freq; 1096 } 1097 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1098 1099 bt = &basetime[basetime_index]; 1100 tvp->tv_sec += bt->tv_sec; 1101 tvp->tv_usec += bt->tv_nsec / 1000; 1102 while (tvp->tv_usec >= 1000000) { 1103 tvp->tv_usec -= 1000000; 1104 ++tvp->tv_sec; 1105 } 1106 } 1107 1108 void 1109 getnanotime(struct timespec *tsp) 1110 { 1111 struct globaldata *gd = mycpu; 1112 struct timespec *bt; 1113 sysclock_t delta; 1114 1115 do { 1116 tsp->tv_sec = gd->gd_time_seconds; 1117 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1118 } while (tsp->tv_sec != gd->gd_time_seconds); 1119 1120 if (delta >= sys_cputimer->freq) { 1121 tsp->tv_sec += delta / sys_cputimer->freq; 1122 delta %= sys_cputimer->freq; 1123 } 1124 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1125 1126 bt = &basetime[basetime_index]; 1127 tsp->tv_sec += bt->tv_sec; 1128 tsp->tv_nsec += bt->tv_nsec; 1129 while (tsp->tv_nsec >= 1000000000) { 1130 tsp->tv_nsec -= 1000000000; 1131 ++tsp->tv_sec; 1132 } 1133 } 1134 1135 static void 1136 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1137 { 1138 struct globaldata *gd = mycpu; 1139 sysclock_t delta; 1140 1141 do { 1142 tsp->tv_sec = gd->gd_time_seconds; 1143 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1144 } while (tsp->tv_sec != gd->gd_time_seconds); 1145 1146 if (delta >= sys_cputimer->freq) { 1147 tsp->tv_sec += delta / sys_cputimer->freq; 1148 delta %= sys_cputimer->freq; 1149 } 1150 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1151 1152 tsp->tv_sec += nbt->tv_sec; 1153 tsp->tv_nsec += nbt->tv_nsec; 1154 while (tsp->tv_nsec >= 1000000000) { 1155 tsp->tv_nsec -= 1000000000; 1156 ++tsp->tv_sec; 1157 } 1158 } 1159 1160 1161 void 1162 microtime(struct timeval *tvp) 1163 { 1164 struct globaldata *gd = mycpu; 1165 struct timespec *bt; 1166 sysclock_t delta; 1167 1168 do { 1169 tvp->tv_sec = gd->gd_time_seconds; 1170 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1171 } while (tvp->tv_sec != gd->gd_time_seconds); 1172 1173 if (delta >= sys_cputimer->freq) { 1174 tvp->tv_sec += delta / sys_cputimer->freq; 1175 delta %= sys_cputimer->freq; 1176 } 1177 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1178 1179 bt = &basetime[basetime_index]; 1180 tvp->tv_sec += bt->tv_sec; 1181 tvp->tv_usec += bt->tv_nsec / 1000; 1182 while (tvp->tv_usec >= 1000000) { 1183 tvp->tv_usec -= 1000000; 1184 ++tvp->tv_sec; 1185 } 1186 } 1187 1188 void 1189 nanotime(struct timespec *tsp) 1190 { 1191 struct globaldata *gd = mycpu; 1192 struct timespec *bt; 1193 sysclock_t delta; 1194 1195 do { 1196 tsp->tv_sec = gd->gd_time_seconds; 1197 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1198 } while (tsp->tv_sec != gd->gd_time_seconds); 1199 1200 if (delta >= sys_cputimer->freq) { 1201 tsp->tv_sec += delta / sys_cputimer->freq; 1202 delta %= sys_cputimer->freq; 1203 } 1204 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1205 1206 bt = &basetime[basetime_index]; 1207 tsp->tv_sec += bt->tv_sec; 1208 tsp->tv_nsec += bt->tv_nsec; 1209 while (tsp->tv_nsec >= 1000000000) { 1210 tsp->tv_nsec -= 1000000000; 1211 ++tsp->tv_sec; 1212 } 1213 } 1214 1215 /* 1216 * note: this is not exactly synchronized with real time. To do that we 1217 * would have to do what microtime does and check for a nanoseconds overflow. 1218 */ 1219 time_t 1220 get_approximate_time_t(void) 1221 { 1222 struct globaldata *gd = mycpu; 1223 struct timespec *bt; 1224 1225 bt = &basetime[basetime_index]; 1226 return(gd->gd_time_seconds + bt->tv_sec); 1227 } 1228 1229 int 1230 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1231 { 1232 pps_params_t *app; 1233 struct pps_fetch_args *fapi; 1234 #ifdef PPS_SYNC 1235 struct pps_kcbind_args *kapi; 1236 #endif 1237 1238 switch (cmd) { 1239 case PPS_IOC_CREATE: 1240 return (0); 1241 case PPS_IOC_DESTROY: 1242 return (0); 1243 case PPS_IOC_SETPARAMS: 1244 app = (pps_params_t *)data; 1245 if (app->mode & ~pps->ppscap) 1246 return (EINVAL); 1247 pps->ppsparam = *app; 1248 return (0); 1249 case PPS_IOC_GETPARAMS: 1250 app = (pps_params_t *)data; 1251 *app = pps->ppsparam; 1252 app->api_version = PPS_API_VERS_1; 1253 return (0); 1254 case PPS_IOC_GETCAP: 1255 *(int*)data = pps->ppscap; 1256 return (0); 1257 case PPS_IOC_FETCH: 1258 fapi = (struct pps_fetch_args *)data; 1259 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1260 return (EINVAL); 1261 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1262 return (EOPNOTSUPP); 1263 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1264 fapi->pps_info_buf = pps->ppsinfo; 1265 return (0); 1266 case PPS_IOC_KCBIND: 1267 #ifdef PPS_SYNC 1268 kapi = (struct pps_kcbind_args *)data; 1269 /* XXX Only root should be able to do this */ 1270 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1271 return (EINVAL); 1272 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1273 return (EINVAL); 1274 if (kapi->edge & ~pps->ppscap) 1275 return (EINVAL); 1276 pps->kcmode = kapi->edge; 1277 return (0); 1278 #else 1279 return (EOPNOTSUPP); 1280 #endif 1281 default: 1282 return (ENOTTY); 1283 } 1284 } 1285 1286 void 1287 pps_init(struct pps_state *pps) 1288 { 1289 pps->ppscap |= PPS_TSFMT_TSPEC; 1290 if (pps->ppscap & PPS_CAPTUREASSERT) 1291 pps->ppscap |= PPS_OFFSETASSERT; 1292 if (pps->ppscap & PPS_CAPTURECLEAR) 1293 pps->ppscap |= PPS_OFFSETCLEAR; 1294 } 1295 1296 void 1297 pps_event(struct pps_state *pps, sysclock_t count, int event) 1298 { 1299 struct globaldata *gd; 1300 struct timespec *tsp; 1301 struct timespec *osp; 1302 struct timespec *bt; 1303 struct timespec ts; 1304 sysclock_t *pcount; 1305 #ifdef PPS_SYNC 1306 sysclock_t tcount; 1307 #endif 1308 sysclock_t delta; 1309 pps_seq_t *pseq; 1310 int foff; 1311 int fhard; 1312 1313 gd = mycpu; 1314 1315 /* Things would be easier with arrays... */ 1316 if (event == PPS_CAPTUREASSERT) { 1317 tsp = &pps->ppsinfo.assert_timestamp; 1318 osp = &pps->ppsparam.assert_offset; 1319 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1320 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1321 pcount = &pps->ppscount[0]; 1322 pseq = &pps->ppsinfo.assert_sequence; 1323 } else { 1324 tsp = &pps->ppsinfo.clear_timestamp; 1325 osp = &pps->ppsparam.clear_offset; 1326 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1327 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1328 pcount = &pps->ppscount[1]; 1329 pseq = &pps->ppsinfo.clear_sequence; 1330 } 1331 1332 /* Nothing really happened */ 1333 if (*pcount == count) 1334 return; 1335 1336 *pcount = count; 1337 1338 do { 1339 ts.tv_sec = gd->gd_time_seconds; 1340 delta = count - gd->gd_cpuclock_base; 1341 } while (ts.tv_sec != gd->gd_time_seconds); 1342 1343 if (delta >= sys_cputimer->freq) { 1344 ts.tv_sec += delta / sys_cputimer->freq; 1345 delta %= sys_cputimer->freq; 1346 } 1347 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1348 bt = &basetime[basetime_index]; 1349 ts.tv_sec += bt->tv_sec; 1350 ts.tv_nsec += bt->tv_nsec; 1351 while (ts.tv_nsec >= 1000000000) { 1352 ts.tv_nsec -= 1000000000; 1353 ++ts.tv_sec; 1354 } 1355 1356 (*pseq)++; 1357 *tsp = ts; 1358 1359 if (foff) { 1360 timespecadd(tsp, osp); 1361 if (tsp->tv_nsec < 0) { 1362 tsp->tv_nsec += 1000000000; 1363 tsp->tv_sec -= 1; 1364 } 1365 } 1366 #ifdef PPS_SYNC 1367 if (fhard) { 1368 /* magic, at its best... */ 1369 tcount = count - pps->ppscount[2]; 1370 pps->ppscount[2] = count; 1371 if (tcount >= sys_cputimer->freq) { 1372 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1373 sys_cputimer->freq64_nsec * 1374 (tcount % sys_cputimer->freq)) >> 32; 1375 } else { 1376 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1377 } 1378 hardpps(tsp, delta); 1379 } 1380 #endif 1381 } 1382 1383 /* 1384 * Return the tsc target value for a delay of (ns). 1385 * 1386 * Returns -1 if the TSC is not supported. 1387 */ 1388 int64_t 1389 tsc_get_target(int ns) 1390 { 1391 #if defined(_RDTSC_SUPPORTED_) 1392 if (cpu_feature & CPUID_TSC) { 1393 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000); 1394 } 1395 #endif 1396 return(-1); 1397 } 1398 1399 /* 1400 * Compare the tsc against the passed target 1401 * 1402 * Returns +1 if the target has been reached 1403 * Returns 0 if the target has not yet been reached 1404 * Returns -1 if the TSC is not supported. 1405 * 1406 * Typical use: while (tsc_test_target(target) == 0) { ...poll... } 1407 */ 1408 int 1409 tsc_test_target(int64_t target) 1410 { 1411 #if defined(_RDTSC_SUPPORTED_) 1412 if (cpu_feature & CPUID_TSC) { 1413 if ((int64_t)(target - rdtsc()) <= 0) 1414 return(1); 1415 return(0); 1416 } 1417 #endif 1418 return(-1); 1419 } 1420 1421 /* 1422 * Delay the specified number of nanoseconds using the tsc. This function 1423 * returns immediately if the TSC is not supported. At least one cpu_pause() 1424 * will be issued. 1425 */ 1426 void 1427 tsc_delay(int ns) 1428 { 1429 int64_t clk; 1430 1431 clk = tsc_get_target(ns); 1432 cpu_pause(); 1433 while (tsc_test_target(clk) == 0) 1434 cpu_pause(); 1435 } 1436