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. Neither the name of the University nor the names of its contributors 52 * may be used to endorse or promote products derived from this software 53 * without specific prior written permission. 54 * 55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 65 * SUCH DAMAGE. 66 * 67 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 68 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ 69 */ 70 71 #include "opt_ntp.h" 72 #include "opt_pctrack.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/callout.h> 77 #include <sys/kernel.h> 78 #include <sys/kinfo.h> 79 #include <sys/proc.h> 80 #include <sys/malloc.h> 81 #include <sys/resource.h> 82 #include <sys/resourcevar.h> 83 #include <sys/signalvar.h> 84 #include <sys/priv.h> 85 #include <sys/timex.h> 86 #include <sys/timepps.h> 87 #include <sys/upmap.h> 88 #include <sys/lock.h> 89 #include <sys/sysctl.h> 90 #include <sys/kcollect.h> 91 92 #include <vm/vm.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_extern.h> 96 97 #include <sys/thread2.h> 98 #include <sys/spinlock2.h> 99 100 #include <machine/cpu.h> 101 #include <machine/limits.h> 102 #include <machine/smp.h> 103 #include <machine/cpufunc.h> 104 #include <machine/specialreg.h> 105 #include <machine/clock.h> 106 107 #ifdef DEBUG_PCTRACK 108 static void do_pctrack(struct intrframe *frame, int which); 109 #endif 110 111 static void initclocks (void *dummy); 112 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL); 113 114 /* 115 * Some of these don't belong here, but it's easiest to concentrate them. 116 * Note that cpu_time counts in microseconds, but most userland programs 117 * just compare relative times against the total by delta. 118 */ 119 struct kinfo_cputime cputime_percpu[MAXCPU]; 120 #ifdef DEBUG_PCTRACK 121 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE }; 122 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE]; 123 #endif 124 125 static int sniff_enable = 1; 126 static int sniff_target = -1; 127 SYSCTL_INT(_kern, OID_AUTO, sniff_enable, CTLFLAG_RW, &sniff_enable, 0 , ""); 128 SYSCTL_INT(_kern, OID_AUTO, sniff_target, CTLFLAG_RW, &sniff_target, 0 , ""); 129 130 static int 131 sysctl_cputime(SYSCTL_HANDLER_ARGS) 132 { 133 int cpu, error = 0; 134 int root_error; 135 size_t size = sizeof(struct kinfo_cputime); 136 struct kinfo_cputime tmp; 137 138 /* 139 * NOTE: For security reasons, only root can sniff %rip 140 */ 141 root_error = priv_check_cred(curthread->td_ucred, PRIV_ROOT, 0); 142 143 for (cpu = 0; cpu < ncpus; ++cpu) { 144 tmp = cputime_percpu[cpu]; 145 if (root_error == 0) { 146 tmp.cp_sample_pc = 147 (int64_t)globaldata_find(cpu)->gd_sample_pc; 148 tmp.cp_sample_sp = 149 (int64_t)globaldata_find(cpu)->gd_sample_sp; 150 } 151 if ((error = SYSCTL_OUT(req, &tmp, size)) != 0) 152 break; 153 } 154 155 if (root_error == 0) { 156 if (sniff_enable) { 157 int n = sniff_target; 158 if (n < 0) 159 smp_sniff(); 160 else if (n < ncpus) 161 cpu_sniff(n); 162 } 163 } 164 165 return (error); 166 } 167 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 168 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics"); 169 170 static int 171 sysctl_cp_time(SYSCTL_HANDLER_ARGS) 172 { 173 long cpu_states[CPUSTATES] = {0}; 174 int cpu, error = 0; 175 size_t size = sizeof(cpu_states); 176 177 for (cpu = 0; cpu < ncpus; ++cpu) { 178 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user; 179 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice; 180 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys; 181 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr; 182 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle; 183 } 184 185 error = SYSCTL_OUT(req, cpu_states, size); 186 187 return (error); 188 } 189 190 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0, 191 sysctl_cp_time, "LU", "CPU time statistics"); 192 193 static int 194 sysctl_cp_times(SYSCTL_HANDLER_ARGS) 195 { 196 long cpu_states[CPUSTATES] = {0}; 197 int cpu, error; 198 size_t size = sizeof(cpu_states); 199 200 for (error = 0, cpu = 0; error == 0 && cpu < ncpus; ++cpu) { 201 cpu_states[CP_USER] = cputime_percpu[cpu].cp_user; 202 cpu_states[CP_NICE] = cputime_percpu[cpu].cp_nice; 203 cpu_states[CP_SYS] = cputime_percpu[cpu].cp_sys; 204 cpu_states[CP_INTR] = cputime_percpu[cpu].cp_intr; 205 cpu_states[CP_IDLE] = cputime_percpu[cpu].cp_idle; 206 error = SYSCTL_OUT(req, cpu_states, size); 207 } 208 209 return (error); 210 } 211 212 SYSCTL_PROC(_kern, OID_AUTO, cp_times, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0, 213 sysctl_cp_times, "LU", "per-CPU time statistics"); 214 215 /* 216 * boottime is used to calculate the 'real' uptime. Do not confuse this with 217 * microuptime(). microtime() is not drift compensated. The real uptime 218 * with compensation is nanotime() - bootime. boottime is recalculated 219 * whenever the real time is set based on the compensated elapsed time 220 * in seconds (gd->gd_time_seconds). 221 * 222 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic. 223 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to 224 * the real time. 225 * 226 * WARNING! time_second can backstep on time corrections. Also, unlike 227 * time_second, time_uptime is not a "real" time_t (seconds 228 * since the Epoch) but seconds since booting. 229 */ 230 struct timespec boottime; /* boot time (realtime) for reference only */ 231 time_t time_second; /* read-only 'passive' realtime in seconds */ 232 time_t time_uptime; /* read-only 'passive' uptime in seconds */ 233 234 /* 235 * basetime is used to calculate the compensated real time of day. The 236 * basetime can be modified on a per-tick basis by the adjtime(), 237 * ntp_adjtime(), and sysctl-based time correction APIs. 238 * 239 * Note that frequency corrections can also be made by adjusting 240 * gd_cpuclock_base. 241 * 242 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is 243 * used on both SMP and UP systems to avoid MP races between cpu's and 244 * interrupt races on UP systems. 245 */ 246 struct hardtime { 247 __uint32_t time_second; 248 sysclock_t cpuclock_base; 249 }; 250 251 #define BASETIME_ARYSIZE 16 252 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1) 253 static struct timespec basetime[BASETIME_ARYSIZE]; 254 static struct hardtime hardtime[BASETIME_ARYSIZE]; 255 static volatile int basetime_index; 256 257 static int 258 sysctl_get_basetime(SYSCTL_HANDLER_ARGS) 259 { 260 struct timespec *bt; 261 int error; 262 int index; 263 264 /* 265 * Because basetime data and index may be updated by another cpu, 266 * a load fence is required to ensure that the data we read has 267 * not been speculatively read relative to a possibly updated index. 268 */ 269 index = basetime_index; 270 cpu_lfence(); 271 bt = &basetime[index]; 272 error = SYSCTL_OUT(req, bt, sizeof(*bt)); 273 return (error); 274 } 275 276 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, 277 &boottime, timespec, "System boottime"); 278 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0, 279 sysctl_get_basetime, "S,timespec", "System basetime"); 280 281 static void hardclock(systimer_t info, int, struct intrframe *frame); 282 static void statclock(systimer_t info, int, struct intrframe *frame); 283 static void schedclock(systimer_t info, int, struct intrframe *frame); 284 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp); 285 286 int ticks; /* system master ticks at hz */ 287 int clocks_running; /* tsleep/timeout clocks operational */ 288 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 289 int64_t nsec_acc; /* accumulator */ 290 int sched_ticks; /* global schedule clock ticks */ 291 292 /* NTPD time correction fields */ 293 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 294 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 295 int64_t ntp_delta; /* one-time correction in nsec */ 296 int64_t ntp_big_delta = 1000000000; 297 int32_t ntp_tick_delta; /* current adjustment rate */ 298 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 299 time_t ntp_leap_second; /* time of next leap second */ 300 int ntp_leap_insert; /* whether to insert or remove a second */ 301 struct spinlock ntp_spin; 302 303 /* 304 * Finish initializing clock frequencies and start all clocks running. 305 */ 306 /* ARGSUSED*/ 307 static void 308 initclocks(void *dummy) 309 { 310 /*psratio = profhz / stathz;*/ 311 spin_init(&ntp_spin, "ntp"); 312 initclocks_pcpu(); 313 clocks_running = 1; 314 if (kpmap) { 315 kpmap->tsc_freq = tsc_frequency; 316 kpmap->tick_freq = hz; 317 } 318 } 319 320 /* 321 * Called on a per-cpu basis from the idle thread bootstrap on each cpu 322 * during SMP initialization. 323 * 324 * This routine is called concurrently during low-level SMP initialization 325 * and may not block in any way. Meaning, among other things, we can't 326 * acquire any tokens. 327 */ 328 void 329 initclocks_pcpu(void) 330 { 331 struct globaldata *gd = mycpu; 332 333 crit_enter(); 334 if (gd->gd_cpuid == 0) { 335 gd->gd_time_seconds = 1; 336 gd->gd_cpuclock_base = sys_cputimer->count(); 337 hardtime[0].time_second = gd->gd_time_seconds; 338 hardtime[0].cpuclock_base = gd->gd_cpuclock_base; 339 } else { 340 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 341 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 342 } 343 344 systimer_intr_enable(); 345 346 crit_exit(); 347 } 348 349 /* 350 * Called on a 10-second interval after the system is operational. 351 * Return the collection data for USERPCT and install the data for 352 * SYSTPCT and IDLEPCT. 353 */ 354 static 355 uint64_t 356 collect_cputime_callback(int n) 357 { 358 static long cpu_base[CPUSTATES]; 359 long cpu_states[CPUSTATES]; 360 long total; 361 long acc; 362 long lsb; 363 364 bzero(cpu_states, sizeof(cpu_states)); 365 for (n = 0; n < ncpus; ++n) { 366 cpu_states[CP_USER] += cputime_percpu[n].cp_user; 367 cpu_states[CP_NICE] += cputime_percpu[n].cp_nice; 368 cpu_states[CP_SYS] += cputime_percpu[n].cp_sys; 369 cpu_states[CP_INTR] += cputime_percpu[n].cp_intr; 370 cpu_states[CP_IDLE] += cputime_percpu[n].cp_idle; 371 } 372 373 acc = 0; 374 for (n = 0; n < CPUSTATES; ++n) { 375 total = cpu_states[n] - cpu_base[n]; 376 cpu_base[n] = cpu_states[n]; 377 cpu_states[n] = total; 378 acc += total; 379 } 380 if (acc == 0) /* prevent degenerate divide by 0 */ 381 acc = 1; 382 lsb = acc / (10000 * 2); 383 kcollect_setvalue(KCOLLECT_SYSTPCT, 384 (cpu_states[CP_SYS] + lsb) * 10000 / acc); 385 kcollect_setvalue(KCOLLECT_IDLEPCT, 386 (cpu_states[CP_IDLE] + lsb) * 10000 / acc); 387 kcollect_setvalue(KCOLLECT_INTRPCT, 388 (cpu_states[CP_INTR] + lsb) * 10000 / acc); 389 return((cpu_states[CP_USER] + cpu_states[CP_NICE] + lsb) * 10000 / acc); 390 } 391 392 /* 393 * This routine is called on just the BSP, just after SMP initialization 394 * completes to * finish initializing any clocks that might contend/block 395 * (e.g. like on a token). We can't do this in initclocks_pcpu() because 396 * that function is called from the idle thread bootstrap for each cpu and 397 * not allowed to block at all. 398 */ 399 static 400 void 401 initclocks_other(void *dummy) 402 { 403 struct globaldata *ogd = mycpu; 404 struct globaldata *gd; 405 int n; 406 407 for (n = 0; n < ncpus; ++n) { 408 lwkt_setcpu_self(globaldata_find(n)); 409 gd = mycpu; 410 411 /* 412 * Use a non-queued periodic systimer to prevent multiple 413 * ticks from building up if the sysclock jumps forward 414 * (8254 gets reset). The sysclock will never jump backwards. 415 * Our time sync is based on the actual sysclock, not the 416 * ticks count. 417 * 418 * Install statclock before hardclock to prevent statclock 419 * from misinterpreting gd_flags for tick assignment when 420 * they overlap. 421 */ 422 systimer_init_periodic_flags(&gd->gd_statclock, statclock, 423 NULL, stathz, 424 SYSTF_MSSYNC | SYSTF_FIRST); 425 systimer_init_periodic_flags(&gd->gd_hardclock, hardclock, 426 NULL, hz, SYSTF_MSSYNC); 427 } 428 lwkt_setcpu_self(ogd); 429 430 /* 431 * Regular data collection 432 */ 433 kcollect_register(KCOLLECT_USERPCT, "user", collect_cputime_callback, 434 KCOLLECT_SCALE(KCOLLECT_USERPCT_FORMAT, 0)); 435 kcollect_register(KCOLLECT_SYSTPCT, "syst", NULL, 436 KCOLLECT_SCALE(KCOLLECT_SYSTPCT_FORMAT, 0)); 437 kcollect_register(KCOLLECT_IDLEPCT, "idle", NULL, 438 KCOLLECT_SCALE(KCOLLECT_IDLEPCT_FORMAT, 0)); 439 } 440 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL); 441 442 /* 443 * This method is called on just the BSP, after all the usched implementations 444 * are initialized. This avoids races between usched initialization functions 445 * and usched_schedulerclock(). 446 */ 447 static 448 void 449 initclocks_usched(void *dummy) 450 { 451 struct globaldata *ogd = mycpu; 452 struct globaldata *gd; 453 int n; 454 455 for (n = 0; n < ncpus; ++n) { 456 lwkt_setcpu_self(globaldata_find(n)); 457 gd = mycpu; 458 459 /* XXX correct the frequency for scheduler / estcpu tests */ 460 systimer_init_periodic_flags(&gd->gd_schedclock, schedclock, 461 NULL, ESTCPUFREQ, SYSTF_MSSYNC); 462 } 463 lwkt_setcpu_self(ogd); 464 } 465 SYSINIT(clocks3, SI_BOOT2_USCHED, SI_ORDER_ANY, initclocks_usched, NULL); 466 467 /* 468 * This sets the current real time of day. Timespecs are in seconds and 469 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 470 * instead we adjust basetime so basetime + gd_* results in the current 471 * time of day. This way the gd_* fields are guaranteed to represent 472 * a monotonically increasing 'uptime' value. 473 * 474 * When set_timeofday() is called from userland, the system call forces it 475 * onto cpu #0 since only cpu #0 can update basetime_index. 476 */ 477 void 478 set_timeofday(struct timespec *ts) 479 { 480 struct timespec *nbt; 481 int ni; 482 483 /* 484 * XXX SMP / non-atomic basetime updates 485 */ 486 crit_enter(); 487 ni = (basetime_index + 1) & BASETIME_ARYMASK; 488 cpu_lfence(); 489 nbt = &basetime[ni]; 490 nanouptime(nbt); 491 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 492 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 493 if (nbt->tv_nsec < 0) { 494 nbt->tv_nsec += 1000000000; 495 --nbt->tv_sec; 496 } 497 498 /* 499 * Note that basetime diverges from boottime as the clock drift is 500 * compensated for, so we cannot do away with boottime. When setting 501 * the absolute time of day the drift is 0 (for an instant) and we 502 * can simply assign boottime to basetime. 503 * 504 * Note that nanouptime() is based on gd_time_seconds which is drift 505 * compensated up to a point (it is guaranteed to remain monotonically 506 * increasing). gd_time_seconds is thus our best uptime guess and 507 * suitable for use in the boottime calculation. It is already taken 508 * into account in the basetime calculation above. 509 */ 510 spin_lock(&ntp_spin); 511 boottime.tv_sec = nbt->tv_sec; 512 ntp_delta = 0; 513 514 /* 515 * We now have a new basetime, make sure all other cpus have it, 516 * then update the index. 517 */ 518 cpu_sfence(); 519 basetime_index = ni; 520 spin_unlock(&ntp_spin); 521 522 crit_exit(); 523 } 524 525 /* 526 * Each cpu has its own hardclock, but we only increments ticks and softticks 527 * on cpu #0. 528 * 529 * NOTE! systimer! the MP lock might not be held here. We can only safely 530 * manipulate objects owned by the current cpu. 531 */ 532 static void 533 hardclock(systimer_t info, int in_ipi, struct intrframe *frame) 534 { 535 sysclock_t cputicks; 536 struct proc *p; 537 struct globaldata *gd = mycpu; 538 539 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) { 540 /* Defer to doreti on passive IPIQ processing */ 541 need_ipiq(); 542 } 543 544 /* 545 * We update the compensation base to calculate fine-grained time 546 * from the sys_cputimer on a per-cpu basis in order to avoid 547 * having to mess around with locks. sys_cputimer is assumed to 548 * be consistent across all cpus. CPU N copies the base state from 549 * CPU 0 using the same FIFO trick that we use for basetime (so we 550 * don't catch a CPU 0 update in the middle). 551 * 552 * Note that we never allow info->time (aka gd->gd_hardclock.time) 553 * to reverse index gd_cpuclock_base, but that it is possible for 554 * it to temporarily get behind in the seconds if something in the 555 * system locks interrupts for a long period of time. Since periodic 556 * timers count events, though everything should resynch again 557 * immediately. 558 */ 559 if (gd->gd_cpuid == 0) { 560 int ni; 561 562 cputicks = info->time - gd->gd_cpuclock_base; 563 if (cputicks >= sys_cputimer->freq) { 564 cputicks /= sys_cputimer->freq; 565 if (cputicks != 0 && cputicks != 1) 566 kprintf("Warning: hardclock missed > 1 sec\n"); 567 gd->gd_time_seconds += cputicks; 568 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks; 569 /* uncorrected monotonic 1-sec gran */ 570 time_uptime += cputicks; 571 } 572 ni = (basetime_index + 1) & BASETIME_ARYMASK; 573 hardtime[ni].time_second = gd->gd_time_seconds; 574 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base; 575 } else { 576 int ni; 577 578 ni = basetime_index; 579 cpu_lfence(); 580 gd->gd_time_seconds = hardtime[ni].time_second; 581 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base; 582 } 583 584 /* 585 * The system-wide ticks counter and NTP related timedelta/tickdelta 586 * adjustments only occur on cpu #0. NTP adjustments are accomplished 587 * by updating basetime. 588 */ 589 if (gd->gd_cpuid == 0) { 590 struct timespec *nbt; 591 struct timespec nts; 592 int leap; 593 int ni; 594 595 ++ticks; 596 597 #if 0 598 if (tco->tc_poll_pps) 599 tco->tc_poll_pps(tco); 600 #endif 601 602 /* 603 * Calculate the new basetime index. We are in a critical section 604 * on cpu #0 and can safely play with basetime_index. Start 605 * with the current basetime and then make adjustments. 606 */ 607 ni = (basetime_index + 1) & BASETIME_ARYMASK; 608 nbt = &basetime[ni]; 609 *nbt = basetime[basetime_index]; 610 611 /* 612 * ntp adjustments only occur on cpu 0 and are protected by 613 * ntp_spin. This spinlock virtually never conflicts. 614 */ 615 spin_lock(&ntp_spin); 616 617 /* 618 * Apply adjtime corrections. (adjtime() API) 619 * 620 * adjtime() only runs on cpu #0 so our critical section is 621 * sufficient to access these variables. 622 */ 623 if (ntp_delta != 0) { 624 nbt->tv_nsec += ntp_tick_delta; 625 ntp_delta -= ntp_tick_delta; 626 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 627 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 628 ntp_tick_delta = ntp_delta; 629 } 630 } 631 632 /* 633 * Apply permanent frequency corrections. (sysctl API) 634 */ 635 if (ntp_tick_permanent != 0) { 636 ntp_tick_acc += ntp_tick_permanent; 637 if (ntp_tick_acc >= (1LL << 32)) { 638 nbt->tv_nsec += ntp_tick_acc >> 32; 639 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 640 } else if (ntp_tick_acc <= -(1LL << 32)) { 641 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 642 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 643 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 644 } 645 } 646 647 if (nbt->tv_nsec >= 1000000000) { 648 nbt->tv_sec++; 649 nbt->tv_nsec -= 1000000000; 650 } else if (nbt->tv_nsec < 0) { 651 nbt->tv_sec--; 652 nbt->tv_nsec += 1000000000; 653 } 654 655 /* 656 * Another per-tick compensation. (for ntp_adjtime() API) 657 */ 658 if (nsec_adj != 0) { 659 nsec_acc += nsec_adj; 660 if (nsec_acc >= 0x100000000LL) { 661 nbt->tv_nsec += nsec_acc >> 32; 662 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 663 } else if (nsec_acc <= -0x100000000LL) { 664 nbt->tv_nsec -= -nsec_acc >> 32; 665 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 666 } 667 if (nbt->tv_nsec >= 1000000000) { 668 nbt->tv_nsec -= 1000000000; 669 ++nbt->tv_sec; 670 } else if (nbt->tv_nsec < 0) { 671 nbt->tv_nsec += 1000000000; 672 --nbt->tv_sec; 673 } 674 } 675 spin_unlock(&ntp_spin); 676 677 /************************************************************ 678 * LEAP SECOND CORRECTION * 679 ************************************************************ 680 * 681 * Taking into account all the corrections made above, figure 682 * out the new real time. If the seconds field has changed 683 * then apply any pending leap-second corrections. 684 */ 685 getnanotime_nbt(nbt, &nts); 686 687 if (time_second != nts.tv_sec) { 688 /* 689 * Apply leap second (sysctl API). Adjust nts for changes 690 * so we do not have to call getnanotime_nbt again. 691 */ 692 if (ntp_leap_second) { 693 if (ntp_leap_second == nts.tv_sec) { 694 if (ntp_leap_insert) { 695 nbt->tv_sec++; 696 nts.tv_sec++; 697 } else { 698 nbt->tv_sec--; 699 nts.tv_sec--; 700 } 701 ntp_leap_second--; 702 } 703 } 704 705 /* 706 * Apply leap second (ntp_adjtime() API), calculate a new 707 * nsec_adj field. ntp_update_second() returns nsec_adj 708 * as a per-second value but we need it as a per-tick value. 709 */ 710 leap = ntp_update_second(time_second, &nsec_adj); 711 nsec_adj /= hz; 712 nbt->tv_sec += leap; 713 nts.tv_sec += leap; 714 715 /* 716 * Update the time_second 'approximate time' global. 717 */ 718 time_second = nts.tv_sec; 719 } 720 721 /* 722 * Finally, our new basetime is ready to go live! 723 */ 724 cpu_sfence(); 725 basetime_index = ni; 726 727 /* 728 * Update kpmap on each tick. TS updates are integrated with 729 * fences and upticks allowing userland to read the data 730 * deterministically. 731 */ 732 if (kpmap) { 733 int w; 734 735 w = (kpmap->upticks + 1) & 1; 736 getnanouptime(&kpmap->ts_uptime[w]); 737 getnanotime(&kpmap->ts_realtime[w]); 738 cpu_sfence(); 739 ++kpmap->upticks; 740 cpu_sfence(); 741 } 742 } 743 744 /* 745 * lwkt thread scheduler fair queueing 746 */ 747 lwkt_schedulerclock(curthread); 748 749 /* 750 * softticks are handled for all cpus 751 */ 752 hardclock_softtick(gd); 753 754 /* 755 * Rollup accumulated vmstats, copy-back for critical path checks. 756 */ 757 vmstats_rollup_cpu(gd); 758 vfscache_rollup_cpu(gd); 759 mycpu->gd_vmstats = vmstats; 760 761 /* 762 * ITimer handling is per-tick, per-cpu. 763 * 764 * We must acquire the per-process token in order for ksignal() 765 * to be non-blocking. For the moment this requires an AST fault, 766 * the ksignal() cannot be safely issued from this hard interrupt. 767 * 768 * XXX Even the trytoken here isn't right, and itimer operation in 769 * a multi threaded environment is going to be weird at the 770 * very least. 771 */ 772 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) { 773 crit_enter_hard(); 774 if (p->p_upmap) 775 ++p->p_upmap->runticks; 776 777 if (frame && CLKF_USERMODE(frame) && 778 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 779 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) { 780 p->p_flags |= P_SIGVTALRM; 781 need_user_resched(); 782 } 783 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 784 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) { 785 p->p_flags |= P_SIGPROF; 786 need_user_resched(); 787 } 788 crit_exit_hard(); 789 lwkt_reltoken(&p->p_token); 790 } 791 setdelayed(); 792 } 793 794 /* 795 * The statistics clock typically runs at a 125Hz rate, and is intended 796 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 797 * 798 * NOTE! systimer! the MP lock might not be held here. We can only safely 799 * manipulate objects owned by the current cpu. 800 * 801 * The stats clock is responsible for grabbing a profiling sample. 802 * Most of the statistics are only used by user-level statistics programs. 803 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 804 * p->p_estcpu. 805 * 806 * Like the other clocks, the stat clock is called from what is effectively 807 * a fast interrupt, so the context should be the thread/process that got 808 * interrupted. 809 */ 810 static void 811 statclock(systimer_t info, int in_ipi, struct intrframe *frame) 812 { 813 globaldata_t gd = mycpu; 814 thread_t td; 815 struct proc *p; 816 int bump; 817 sysclock_t cv; 818 sysclock_t scv; 819 820 /* 821 * How big was our timeslice relative to the last time? Calculate 822 * in microseconds. 823 * 824 * NOTE: Use of microuptime() is typically MPSAFE, but usually not 825 * during early boot. Just use the systimer count to be nice 826 * to e.g. qemu. The systimer has a better chance of being 827 * MPSAFE at early boot. 828 */ 829 cv = sys_cputimer->count(); 830 scv = gd->statint.gd_statcv; 831 if (scv == 0) { 832 bump = 1; 833 } else { 834 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32; 835 if (bump < 0) 836 bump = 0; 837 if (bump > 1000000) 838 bump = 1000000; 839 } 840 gd->statint.gd_statcv = cv; 841 842 #if 0 843 stv = &gd->gd_stattv; 844 if (stv->tv_sec == 0) { 845 bump = 1; 846 } else { 847 bump = tv.tv_usec - stv->tv_usec + 848 (tv.tv_sec - stv->tv_sec) * 1000000; 849 if (bump < 0) 850 bump = 0; 851 if (bump > 1000000) 852 bump = 1000000; 853 } 854 *stv = tv; 855 #endif 856 857 td = curthread; 858 p = td->td_proc; 859 860 if (frame && CLKF_USERMODE(frame)) { 861 /* 862 * Came from userland, handle user time and deal with 863 * possible process. 864 */ 865 if (p && (p->p_flags & P_PROFIL)) 866 addupc_intr(p, CLKF_PC(frame), 1); 867 td->td_uticks += bump; 868 869 /* 870 * Charge the time as appropriate 871 */ 872 if (p && p->p_nice > NZERO) 873 cpu_time.cp_nice += bump; 874 else 875 cpu_time.cp_user += bump; 876 } else { 877 int intr_nest = gd->gd_intr_nesting_level; 878 879 if (in_ipi) { 880 /* 881 * IPI processing code will bump gd_intr_nesting_level 882 * up by one, which breaks following CLKF_INTR testing, 883 * so we subtract it by one here. 884 */ 885 --intr_nest; 886 } 887 888 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td)) 889 890 /* 891 * Came from kernel mode, so we were: 892 * - handling an interrupt, 893 * - doing syscall or trap work on behalf of the current 894 * user process, or 895 * - spinning in the idle loop. 896 * Whichever it is, charge the time as appropriate. 897 * Note that we charge interrupts to the current process, 898 * regardless of whether they are ``for'' that process, 899 * so that we know how much of its real time was spent 900 * in ``non-process'' (i.e., interrupt) work. 901 * 902 * XXX assume system if frame is NULL. A NULL frame 903 * can occur if ipi processing is done from a crit_exit(). 904 */ 905 if (IS_INTR_RUNNING || 906 (gd->gd_reqflags & RQF_INTPEND)) { 907 /* 908 * If we interrupted an interrupt thread, well, 909 * count it as interrupt time. 910 */ 911 td->td_iticks += bump; 912 #ifdef DEBUG_PCTRACK 913 if (frame) 914 do_pctrack(frame, PCTRACK_INT); 915 #endif 916 cpu_time.cp_intr += bump; 917 } else if (gd->gd_flags & GDF_VIRTUSER) { 918 /* 919 * The vkernel doesn't do a good job providing trap 920 * frames that we can test. If the GDF_VIRTUSER 921 * flag is set we probably interrupted user mode. 922 * 923 * We also use this flag on the host when entering 924 * VMM mode. 925 */ 926 td->td_uticks += bump; 927 928 /* 929 * Charge the time as appropriate 930 */ 931 if (p && p->p_nice > NZERO) 932 cpu_time.cp_nice += bump; 933 else 934 cpu_time.cp_user += bump; 935 } else { 936 td->td_sticks += bump; 937 if (td == &gd->gd_idlethread) { 938 /* 939 * We want to count token contention as 940 * system time. When token contention occurs 941 * the cpu may only be outside its critical 942 * section while switching through the idle 943 * thread. In this situation, various flags 944 * will be set in gd_reqflags. 945 */ 946 if (gd->gd_reqflags & RQF_IDLECHECK_WK_MASK) 947 cpu_time.cp_sys += bump; 948 else 949 cpu_time.cp_idle += bump; 950 } else { 951 /* 952 * System thread was running. 953 */ 954 #ifdef DEBUG_PCTRACK 955 if (frame) 956 do_pctrack(frame, PCTRACK_SYS); 957 #endif 958 cpu_time.cp_sys += bump; 959 } 960 } 961 962 #undef IS_INTR_RUNNING 963 } 964 } 965 966 #ifdef DEBUG_PCTRACK 967 /* 968 * Sample the PC when in the kernel or in an interrupt. User code can 969 * retrieve the information and generate a histogram or other output. 970 */ 971 972 static void 973 do_pctrack(struct intrframe *frame, int which) 974 { 975 struct kinfo_pctrack *pctrack; 976 977 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 978 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 979 (void *)CLKF_PC(frame); 980 ++pctrack->pc_index; 981 } 982 983 static int 984 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 985 { 986 struct kinfo_pcheader head; 987 int error; 988 int cpu; 989 int ntrack; 990 991 head.pc_ntrack = PCTRACK_SIZE; 992 head.pc_arysize = PCTRACK_ARYSIZE; 993 994 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 995 return (error); 996 997 for (cpu = 0; cpu < ncpus; ++cpu) { 998 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 999 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 1000 sizeof(struct kinfo_pctrack)); 1001 if (error) 1002 break; 1003 } 1004 if (error) 1005 break; 1006 } 1007 return (error); 1008 } 1009 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 1010 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 1011 1012 #endif 1013 1014 /* 1015 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 1016 * the MP lock might not be held. We can safely manipulate parts of curproc 1017 * but that's about it. 1018 * 1019 * Each cpu has its own scheduler clock. 1020 */ 1021 static void 1022 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame) 1023 { 1024 struct lwp *lp; 1025 struct rusage *ru; 1026 struct vmspace *vm; 1027 long rss; 1028 1029 if ((lp = lwkt_preempted_proc()) != NULL) { 1030 /* 1031 * Account for cpu time used and hit the scheduler. Note 1032 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 1033 * HERE. 1034 */ 1035 ++lp->lwp_cpticks; 1036 usched_schedulerclock(lp, info->periodic, info->time); 1037 } else { 1038 usched_schedulerclock(NULL, info->periodic, info->time); 1039 } 1040 if ((lp = curthread->td_lwp) != NULL) { 1041 /* 1042 * Update resource usage integrals and maximums. 1043 */ 1044 if ((ru = &lp->lwp_proc->p_ru) && 1045 (vm = lp->lwp_proc->p_vmspace) != NULL) { 1046 ru->ru_ixrss += pgtok(btoc(vm->vm_tsize)); 1047 ru->ru_idrss += pgtok(btoc(vm->vm_dsize)); 1048 ru->ru_isrss += pgtok(btoc(vm->vm_ssize)); 1049 if (lwkt_trytoken(&vm->vm_map.token)) { 1050 rss = pgtok(vmspace_resident_count(vm)); 1051 if (ru->ru_maxrss < rss) 1052 ru->ru_maxrss = rss; 1053 lwkt_reltoken(&vm->vm_map.token); 1054 } 1055 } 1056 } 1057 /* Increment the global sched_ticks */ 1058 if (mycpu->gd_cpuid == 0) 1059 ++sched_ticks; 1060 } 1061 1062 /* 1063 * Compute number of ticks for the specified amount of time. The 1064 * return value is intended to be used in a clock interrupt timed 1065 * operation and guaranteed to meet or exceed the requested time. 1066 * If the representation overflows, return INT_MAX. The minimum return 1067 * value is 1 ticks and the function will average the calculation up. 1068 * If any value greater then 0 microseconds is supplied, a value 1069 * of at least 2 will be returned to ensure that a near-term clock 1070 * interrupt does not cause the timeout to occur (degenerately) early. 1071 * 1072 * Note that limit checks must take into account microseconds, which is 1073 * done simply by using the smaller signed long maximum instead of 1074 * the unsigned long maximum. 1075 * 1076 * If ints have 32 bits, then the maximum value for any timeout in 1077 * 10ms ticks is 248 days. 1078 */ 1079 int 1080 tvtohz_high(struct timeval *tv) 1081 { 1082 int ticks; 1083 long sec, usec; 1084 1085 sec = tv->tv_sec; 1086 usec = tv->tv_usec; 1087 if (usec < 0) { 1088 sec--; 1089 usec += 1000000; 1090 } 1091 if (sec < 0) { 1092 #ifdef DIAGNOSTIC 1093 if (usec > 0) { 1094 sec++; 1095 usec -= 1000000; 1096 } 1097 kprintf("tvtohz_high: negative time difference " 1098 "%ld sec %ld usec\n", 1099 sec, usec); 1100 #endif 1101 ticks = 1; 1102 } else if (sec <= INT_MAX / hz) { 1103 ticks = (int)(sec * hz + 1104 ((u_long)usec + (ustick - 1)) / ustick) + 1; 1105 } else { 1106 ticks = INT_MAX; 1107 } 1108 return (ticks); 1109 } 1110 1111 int 1112 tstohz_high(struct timespec *ts) 1113 { 1114 int ticks; 1115 long sec, nsec; 1116 1117 sec = ts->tv_sec; 1118 nsec = ts->tv_nsec; 1119 if (nsec < 0) { 1120 sec--; 1121 nsec += 1000000000; 1122 } 1123 if (sec < 0) { 1124 #ifdef DIAGNOSTIC 1125 if (nsec > 0) { 1126 sec++; 1127 nsec -= 1000000000; 1128 } 1129 kprintf("tstohz_high: negative time difference " 1130 "%ld sec %ld nsec\n", 1131 sec, nsec); 1132 #endif 1133 ticks = 1; 1134 } else if (sec <= INT_MAX / hz) { 1135 ticks = (int)(sec * hz + 1136 ((u_long)nsec + (nstick - 1)) / nstick) + 1; 1137 } else { 1138 ticks = INT_MAX; 1139 } 1140 return (ticks); 1141 } 1142 1143 1144 /* 1145 * Compute number of ticks for the specified amount of time, erroring on 1146 * the side of it being too low to ensure that sleeping the returned number 1147 * of ticks will not result in a late return. 1148 * 1149 * The supplied timeval may not be negative and should be normalized. A 1150 * return value of 0 is possible if the timeval converts to less then 1151 * 1 tick. 1152 * 1153 * If ints have 32 bits, then the maximum value for any timeout in 1154 * 10ms ticks is 248 days. 1155 */ 1156 int 1157 tvtohz_low(struct timeval *tv) 1158 { 1159 int ticks; 1160 long sec; 1161 1162 sec = tv->tv_sec; 1163 if (sec <= INT_MAX / hz) 1164 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick); 1165 else 1166 ticks = INT_MAX; 1167 return (ticks); 1168 } 1169 1170 int 1171 tstohz_low(struct timespec *ts) 1172 { 1173 int ticks; 1174 long sec; 1175 1176 sec = ts->tv_sec; 1177 if (sec <= INT_MAX / hz) 1178 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick); 1179 else 1180 ticks = INT_MAX; 1181 return (ticks); 1182 } 1183 1184 /* 1185 * Start profiling on a process. 1186 * 1187 * Caller must hold p->p_token(); 1188 * 1189 * Kernel profiling passes proc0 which never exits and hence 1190 * keeps the profile clock running constantly. 1191 */ 1192 void 1193 startprofclock(struct proc *p) 1194 { 1195 if ((p->p_flags & P_PROFIL) == 0) { 1196 p->p_flags |= P_PROFIL; 1197 #if 0 /* XXX */ 1198 if (++profprocs == 1 && stathz != 0) { 1199 crit_enter(); 1200 psdiv = psratio; 1201 setstatclockrate(profhz); 1202 crit_exit(); 1203 } 1204 #endif 1205 } 1206 } 1207 1208 /* 1209 * Stop profiling on a process. 1210 * 1211 * caller must hold p->p_token 1212 */ 1213 void 1214 stopprofclock(struct proc *p) 1215 { 1216 if (p->p_flags & P_PROFIL) { 1217 p->p_flags &= ~P_PROFIL; 1218 #if 0 /* XXX */ 1219 if (--profprocs == 0 && stathz != 0) { 1220 crit_enter(); 1221 psdiv = 1; 1222 setstatclockrate(stathz); 1223 crit_exit(); 1224 } 1225 #endif 1226 } 1227 } 1228 1229 /* 1230 * Return information about system clocks. 1231 */ 1232 static int 1233 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 1234 { 1235 struct kinfo_clockinfo clkinfo; 1236 /* 1237 * Construct clockinfo structure. 1238 */ 1239 clkinfo.ci_hz = hz; 1240 clkinfo.ci_tick = ustick; 1241 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 1242 clkinfo.ci_profhz = profhz; 1243 clkinfo.ci_stathz = stathz ? stathz : hz; 1244 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 1245 } 1246 1247 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 1248 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 1249 1250 /* 1251 * We have eight functions for looking at the clock, four for 1252 * microseconds and four for nanoseconds. For each there is fast 1253 * but less precise version "get{nano|micro}[up]time" which will 1254 * return a time which is up to 1/HZ previous to the call, whereas 1255 * the raw version "{nano|micro}[up]time" will return a timestamp 1256 * which is as precise as possible. The "up" variants return the 1257 * time relative to system boot, these are well suited for time 1258 * interval measurements. 1259 * 1260 * Each cpu independently maintains the current time of day, so all 1261 * we need to do to protect ourselves from changes is to do a loop 1262 * check on the seconds field changing out from under us. 1263 * 1264 * The system timer maintains a 32 bit count and due to various issues 1265 * it is possible for the calculated delta to occasionally exceed 1266 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 1267 * multiplication can easily overflow, so we deal with the case. For 1268 * uniformity we deal with the case in the usec case too. 1269 * 1270 * All the [get][micro,nano][time,uptime]() routines are MPSAFE. 1271 */ 1272 void 1273 getmicrouptime(struct timeval *tvp) 1274 { 1275 struct globaldata *gd = mycpu; 1276 sysclock_t delta; 1277 1278 do { 1279 tvp->tv_sec = gd->gd_time_seconds; 1280 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1281 } while (tvp->tv_sec != gd->gd_time_seconds); 1282 1283 if (delta >= sys_cputimer->freq) { 1284 tvp->tv_sec += delta / sys_cputimer->freq; 1285 delta %= sys_cputimer->freq; 1286 } 1287 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1288 if (tvp->tv_usec >= 1000000) { 1289 tvp->tv_usec -= 1000000; 1290 ++tvp->tv_sec; 1291 } 1292 } 1293 1294 void 1295 getnanouptime(struct timespec *tsp) 1296 { 1297 struct globaldata *gd = mycpu; 1298 sysclock_t delta; 1299 1300 do { 1301 tsp->tv_sec = gd->gd_time_seconds; 1302 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1303 } while (tsp->tv_sec != gd->gd_time_seconds); 1304 1305 if (delta >= sys_cputimer->freq) { 1306 tsp->tv_sec += delta / sys_cputimer->freq; 1307 delta %= sys_cputimer->freq; 1308 } 1309 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1310 } 1311 1312 void 1313 microuptime(struct timeval *tvp) 1314 { 1315 struct globaldata *gd = mycpu; 1316 sysclock_t delta; 1317 1318 do { 1319 tvp->tv_sec = gd->gd_time_seconds; 1320 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1321 } while (tvp->tv_sec != gd->gd_time_seconds); 1322 1323 if (delta >= sys_cputimer->freq) { 1324 tvp->tv_sec += delta / sys_cputimer->freq; 1325 delta %= sys_cputimer->freq; 1326 } 1327 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1328 } 1329 1330 void 1331 nanouptime(struct timespec *tsp) 1332 { 1333 struct globaldata *gd = mycpu; 1334 sysclock_t delta; 1335 1336 do { 1337 tsp->tv_sec = gd->gd_time_seconds; 1338 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1339 } while (tsp->tv_sec != gd->gd_time_seconds); 1340 1341 if (delta >= sys_cputimer->freq) { 1342 tsp->tv_sec += delta / sys_cputimer->freq; 1343 delta %= sys_cputimer->freq; 1344 } 1345 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1346 } 1347 1348 /* 1349 * realtime routines 1350 */ 1351 void 1352 getmicrotime(struct timeval *tvp) 1353 { 1354 struct globaldata *gd = mycpu; 1355 struct timespec *bt; 1356 sysclock_t delta; 1357 1358 do { 1359 tvp->tv_sec = gd->gd_time_seconds; 1360 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1361 } while (tvp->tv_sec != gd->gd_time_seconds); 1362 1363 if (delta >= sys_cputimer->freq) { 1364 tvp->tv_sec += delta / sys_cputimer->freq; 1365 delta %= sys_cputimer->freq; 1366 } 1367 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1368 1369 bt = &basetime[basetime_index]; 1370 cpu_lfence(); 1371 tvp->tv_sec += bt->tv_sec; 1372 tvp->tv_usec += bt->tv_nsec / 1000; 1373 while (tvp->tv_usec >= 1000000) { 1374 tvp->tv_usec -= 1000000; 1375 ++tvp->tv_sec; 1376 } 1377 } 1378 1379 void 1380 getnanotime(struct timespec *tsp) 1381 { 1382 struct globaldata *gd = mycpu; 1383 struct timespec *bt; 1384 sysclock_t delta; 1385 1386 do { 1387 tsp->tv_sec = gd->gd_time_seconds; 1388 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1389 } while (tsp->tv_sec != gd->gd_time_seconds); 1390 1391 if (delta >= sys_cputimer->freq) { 1392 tsp->tv_sec += delta / sys_cputimer->freq; 1393 delta %= sys_cputimer->freq; 1394 } 1395 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1396 1397 bt = &basetime[basetime_index]; 1398 cpu_lfence(); 1399 tsp->tv_sec += bt->tv_sec; 1400 tsp->tv_nsec += bt->tv_nsec; 1401 while (tsp->tv_nsec >= 1000000000) { 1402 tsp->tv_nsec -= 1000000000; 1403 ++tsp->tv_sec; 1404 } 1405 } 1406 1407 static void 1408 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1409 { 1410 struct globaldata *gd = mycpu; 1411 sysclock_t delta; 1412 1413 do { 1414 tsp->tv_sec = gd->gd_time_seconds; 1415 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1416 } while (tsp->tv_sec != gd->gd_time_seconds); 1417 1418 if (delta >= sys_cputimer->freq) { 1419 tsp->tv_sec += delta / sys_cputimer->freq; 1420 delta %= sys_cputimer->freq; 1421 } 1422 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1423 1424 tsp->tv_sec += nbt->tv_sec; 1425 tsp->tv_nsec += nbt->tv_nsec; 1426 while (tsp->tv_nsec >= 1000000000) { 1427 tsp->tv_nsec -= 1000000000; 1428 ++tsp->tv_sec; 1429 } 1430 } 1431 1432 1433 void 1434 microtime(struct timeval *tvp) 1435 { 1436 struct globaldata *gd = mycpu; 1437 struct timespec *bt; 1438 sysclock_t delta; 1439 1440 do { 1441 tvp->tv_sec = gd->gd_time_seconds; 1442 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1443 } while (tvp->tv_sec != gd->gd_time_seconds); 1444 1445 if (delta >= sys_cputimer->freq) { 1446 tvp->tv_sec += delta / sys_cputimer->freq; 1447 delta %= sys_cputimer->freq; 1448 } 1449 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1450 1451 bt = &basetime[basetime_index]; 1452 cpu_lfence(); 1453 tvp->tv_sec += bt->tv_sec; 1454 tvp->tv_usec += bt->tv_nsec / 1000; 1455 while (tvp->tv_usec >= 1000000) { 1456 tvp->tv_usec -= 1000000; 1457 ++tvp->tv_sec; 1458 } 1459 } 1460 1461 void 1462 nanotime(struct timespec *tsp) 1463 { 1464 struct globaldata *gd = mycpu; 1465 struct timespec *bt; 1466 sysclock_t delta; 1467 1468 do { 1469 tsp->tv_sec = gd->gd_time_seconds; 1470 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1471 } while (tsp->tv_sec != gd->gd_time_seconds); 1472 1473 if (delta >= sys_cputimer->freq) { 1474 tsp->tv_sec += delta / sys_cputimer->freq; 1475 delta %= sys_cputimer->freq; 1476 } 1477 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1478 1479 bt = &basetime[basetime_index]; 1480 cpu_lfence(); 1481 tsp->tv_sec += bt->tv_sec; 1482 tsp->tv_nsec += bt->tv_nsec; 1483 while (tsp->tv_nsec >= 1000000000) { 1484 tsp->tv_nsec -= 1000000000; 1485 ++tsp->tv_sec; 1486 } 1487 } 1488 1489 /* 1490 * Get an approximate time_t. It does not have to be accurate. This 1491 * function is called only from KTR and can be called with the system in 1492 * any state so do not use a critical section or other complex operation 1493 * here. 1494 * 1495 * NOTE: This is not exactly synchronized with real time. To do that we 1496 * would have to do what microtime does and check for a nanoseconds 1497 * overflow. 1498 */ 1499 time_t 1500 get_approximate_time_t(void) 1501 { 1502 struct globaldata *gd = mycpu; 1503 struct timespec *bt; 1504 1505 bt = &basetime[basetime_index]; 1506 return(gd->gd_time_seconds + bt->tv_sec); 1507 } 1508 1509 int 1510 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1511 { 1512 pps_params_t *app; 1513 struct pps_fetch_args *fapi; 1514 #ifdef PPS_SYNC 1515 struct pps_kcbind_args *kapi; 1516 #endif 1517 1518 switch (cmd) { 1519 case PPS_IOC_CREATE: 1520 return (0); 1521 case PPS_IOC_DESTROY: 1522 return (0); 1523 case PPS_IOC_SETPARAMS: 1524 app = (pps_params_t *)data; 1525 if (app->mode & ~pps->ppscap) 1526 return (EINVAL); 1527 pps->ppsparam = *app; 1528 return (0); 1529 case PPS_IOC_GETPARAMS: 1530 app = (pps_params_t *)data; 1531 *app = pps->ppsparam; 1532 app->api_version = PPS_API_VERS_1; 1533 return (0); 1534 case PPS_IOC_GETCAP: 1535 *(int*)data = pps->ppscap; 1536 return (0); 1537 case PPS_IOC_FETCH: 1538 fapi = (struct pps_fetch_args *)data; 1539 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1540 return (EINVAL); 1541 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1542 return (EOPNOTSUPP); 1543 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1544 fapi->pps_info_buf = pps->ppsinfo; 1545 return (0); 1546 case PPS_IOC_KCBIND: 1547 #ifdef PPS_SYNC 1548 kapi = (struct pps_kcbind_args *)data; 1549 /* XXX Only root should be able to do this */ 1550 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1551 return (EINVAL); 1552 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1553 return (EINVAL); 1554 if (kapi->edge & ~pps->ppscap) 1555 return (EINVAL); 1556 pps->kcmode = kapi->edge; 1557 return (0); 1558 #else 1559 return (EOPNOTSUPP); 1560 #endif 1561 default: 1562 return (ENOTTY); 1563 } 1564 } 1565 1566 void 1567 pps_init(struct pps_state *pps) 1568 { 1569 pps->ppscap |= PPS_TSFMT_TSPEC; 1570 if (pps->ppscap & PPS_CAPTUREASSERT) 1571 pps->ppscap |= PPS_OFFSETASSERT; 1572 if (pps->ppscap & PPS_CAPTURECLEAR) 1573 pps->ppscap |= PPS_OFFSETCLEAR; 1574 } 1575 1576 void 1577 pps_event(struct pps_state *pps, sysclock_t count, int event) 1578 { 1579 struct globaldata *gd; 1580 struct timespec *tsp; 1581 struct timespec *osp; 1582 struct timespec *bt; 1583 struct timespec ts; 1584 sysclock_t *pcount; 1585 #ifdef PPS_SYNC 1586 sysclock_t tcount; 1587 #endif 1588 sysclock_t delta; 1589 pps_seq_t *pseq; 1590 int foff; 1591 #ifdef PPS_SYNC 1592 int fhard; 1593 #endif 1594 int ni; 1595 1596 gd = mycpu; 1597 1598 /* Things would be easier with arrays... */ 1599 if (event == PPS_CAPTUREASSERT) { 1600 tsp = &pps->ppsinfo.assert_timestamp; 1601 osp = &pps->ppsparam.assert_offset; 1602 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1603 #ifdef PPS_SYNC 1604 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1605 #endif 1606 pcount = &pps->ppscount[0]; 1607 pseq = &pps->ppsinfo.assert_sequence; 1608 } else { 1609 tsp = &pps->ppsinfo.clear_timestamp; 1610 osp = &pps->ppsparam.clear_offset; 1611 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1612 #ifdef PPS_SYNC 1613 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1614 #endif 1615 pcount = &pps->ppscount[1]; 1616 pseq = &pps->ppsinfo.clear_sequence; 1617 } 1618 1619 /* Nothing really happened */ 1620 if (*pcount == count) 1621 return; 1622 1623 *pcount = count; 1624 1625 do { 1626 ts.tv_sec = gd->gd_time_seconds; 1627 delta = count - gd->gd_cpuclock_base; 1628 } while (ts.tv_sec != gd->gd_time_seconds); 1629 1630 if (delta >= sys_cputimer->freq) { 1631 ts.tv_sec += delta / sys_cputimer->freq; 1632 delta %= sys_cputimer->freq; 1633 } 1634 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1635 ni = basetime_index; 1636 cpu_lfence(); 1637 bt = &basetime[ni]; 1638 ts.tv_sec += bt->tv_sec; 1639 ts.tv_nsec += bt->tv_nsec; 1640 while (ts.tv_nsec >= 1000000000) { 1641 ts.tv_nsec -= 1000000000; 1642 ++ts.tv_sec; 1643 } 1644 1645 (*pseq)++; 1646 *tsp = ts; 1647 1648 if (foff) { 1649 timespecadd(tsp, osp); 1650 if (tsp->tv_nsec < 0) { 1651 tsp->tv_nsec += 1000000000; 1652 tsp->tv_sec -= 1; 1653 } 1654 } 1655 #ifdef PPS_SYNC 1656 if (fhard) { 1657 /* magic, at its best... */ 1658 tcount = count - pps->ppscount[2]; 1659 pps->ppscount[2] = count; 1660 if (tcount >= sys_cputimer->freq) { 1661 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1662 sys_cputimer->freq64_nsec * 1663 (tcount % sys_cputimer->freq)) >> 32; 1664 } else { 1665 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1666 } 1667 hardpps(tsp, delta); 1668 } 1669 #endif 1670 } 1671 1672 /* 1673 * Return the tsc target value for a delay of (ns). 1674 * 1675 * Returns -1 if the TSC is not supported. 1676 */ 1677 tsc_uclock_t 1678 tsc_get_target(int ns) 1679 { 1680 #if defined(_RDTSC_SUPPORTED_) 1681 if (cpu_feature & CPUID_TSC) { 1682 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000); 1683 } 1684 #endif 1685 return(-1); 1686 } 1687 1688 /* 1689 * Compare the tsc against the passed target 1690 * 1691 * Returns +1 if the target has been reached 1692 * Returns 0 if the target has not yet been reached 1693 * Returns -1 if the TSC is not supported. 1694 * 1695 * Typical use: while (tsc_test_target(target) == 0) { ...poll... } 1696 */ 1697 int 1698 tsc_test_target(int64_t target) 1699 { 1700 #if defined(_RDTSC_SUPPORTED_) 1701 if (cpu_feature & CPUID_TSC) { 1702 if ((int64_t)(target - rdtsc()) <= 0) 1703 return(1); 1704 return(0); 1705 } 1706 #endif 1707 return(-1); 1708 } 1709 1710 /* 1711 * Delay the specified number of nanoseconds using the tsc. This function 1712 * returns immediately if the TSC is not supported. At least one cpu_pause() 1713 * will be issued. 1714 */ 1715 void 1716 tsc_delay(int ns) 1717 { 1718 int64_t clk; 1719 1720 clk = tsc_get_target(ns); 1721 cpu_pause(); 1722 cpu_pause(); 1723 while (tsc_test_target(clk) == 0) { 1724 cpu_pause(); 1725 cpu_pause(); 1726 cpu_pause(); 1727 cpu_pause(); 1728 } 1729 } 1730