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 __read_mostly struct timespec boottime; /* boot time (realtime) for ref only */ 231 __read_mostly time_t time_second; /* read-only 'passive' rt in seconds */ 232 __read_mostly time_t time_uptime; /* read-only 'passive' ut 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 /* 287 * Use __read_mostly for ticks and sched_ticks because these variables are 288 * used all over the kernel and only updated once per tick. 289 */ 290 __read_mostly int ticks; /* system master ticks at hz */ 291 __read_mostly int sched_ticks; /* global schedule clock ticks */ 292 __read_mostly int clocks_running; /* tsleep/timeout clocks operational */ 293 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */ 294 int64_t nsec_acc; /* accumulator */ 295 296 /* NTPD time correction fields */ 297 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */ 298 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */ 299 int64_t ntp_delta; /* one-time correction in nsec */ 300 int64_t ntp_big_delta = 1000000000; 301 int32_t ntp_tick_delta; /* current adjustment rate */ 302 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */ 303 time_t ntp_leap_second; /* time of next leap second */ 304 int ntp_leap_insert; /* whether to insert or remove a second */ 305 struct spinlock ntp_spin; 306 307 /* 308 * Finish initializing clock frequencies and start all clocks running. 309 */ 310 /* ARGSUSED*/ 311 static void 312 initclocks(void *dummy) 313 { 314 /*psratio = profhz / stathz;*/ 315 spin_init(&ntp_spin, "ntp"); 316 initclocks_pcpu(); 317 clocks_running = 1; 318 if (kpmap) { 319 kpmap->tsc_freq = tsc_frequency; 320 kpmap->tick_freq = hz; 321 } 322 } 323 324 /* 325 * Called on a per-cpu basis from the idle thread bootstrap on each cpu 326 * during SMP initialization. 327 * 328 * This routine is called concurrently during low-level SMP initialization 329 * and may not block in any way. Meaning, among other things, we can't 330 * acquire any tokens. 331 */ 332 void 333 initclocks_pcpu(void) 334 { 335 struct globaldata *gd = mycpu; 336 337 crit_enter(); 338 if (gd->gd_cpuid == 0) { 339 gd->gd_time_seconds = 1; 340 gd->gd_cpuclock_base = sys_cputimer->count(); 341 hardtime[0].time_second = gd->gd_time_seconds; 342 hardtime[0].cpuclock_base = gd->gd_cpuclock_base; 343 } else { 344 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds; 345 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base; 346 } 347 348 systimer_intr_enable(); 349 350 crit_exit(); 351 } 352 353 /* 354 * Called on a 10-second interval after the system is operational. 355 * Return the collection data for USERPCT and install the data for 356 * SYSTPCT and IDLEPCT. 357 */ 358 static 359 uint64_t 360 collect_cputime_callback(int n) 361 { 362 static long cpu_base[CPUSTATES]; 363 long cpu_states[CPUSTATES]; 364 long total; 365 long acc; 366 long lsb; 367 368 bzero(cpu_states, sizeof(cpu_states)); 369 for (n = 0; n < ncpus; ++n) { 370 cpu_states[CP_USER] += cputime_percpu[n].cp_user; 371 cpu_states[CP_NICE] += cputime_percpu[n].cp_nice; 372 cpu_states[CP_SYS] += cputime_percpu[n].cp_sys; 373 cpu_states[CP_INTR] += cputime_percpu[n].cp_intr; 374 cpu_states[CP_IDLE] += cputime_percpu[n].cp_idle; 375 } 376 377 acc = 0; 378 for (n = 0; n < CPUSTATES; ++n) { 379 total = cpu_states[n] - cpu_base[n]; 380 cpu_base[n] = cpu_states[n]; 381 cpu_states[n] = total; 382 acc += total; 383 } 384 if (acc == 0) /* prevent degenerate divide by 0 */ 385 acc = 1; 386 lsb = acc / (10000 * 2); 387 kcollect_setvalue(KCOLLECT_SYSTPCT, 388 (cpu_states[CP_SYS] + lsb) * 10000 / acc); 389 kcollect_setvalue(KCOLLECT_IDLEPCT, 390 (cpu_states[CP_IDLE] + lsb) * 10000 / acc); 391 kcollect_setvalue(KCOLLECT_INTRPCT, 392 (cpu_states[CP_INTR] + lsb) * 10000 / acc); 393 return((cpu_states[CP_USER] + cpu_states[CP_NICE] + lsb) * 10000 / acc); 394 } 395 396 /* 397 * This routine is called on just the BSP, just after SMP initialization 398 * completes to * finish initializing any clocks that might contend/block 399 * (e.g. like on a token). We can't do this in initclocks_pcpu() because 400 * that function is called from the idle thread bootstrap for each cpu and 401 * not allowed to block at all. 402 */ 403 static 404 void 405 initclocks_other(void *dummy) 406 { 407 struct globaldata *ogd = mycpu; 408 struct globaldata *gd; 409 int n; 410 411 for (n = 0; n < ncpus; ++n) { 412 lwkt_setcpu_self(globaldata_find(n)); 413 gd = mycpu; 414 415 /* 416 * Use a non-queued periodic systimer to prevent multiple 417 * ticks from building up if the sysclock jumps forward 418 * (8254 gets reset). The sysclock will never jump backwards. 419 * Our time sync is based on the actual sysclock, not the 420 * ticks count. 421 * 422 * Install statclock before hardclock to prevent statclock 423 * from misinterpreting gd_flags for tick assignment when 424 * they overlap. Also offset the statclock by half of 425 * its interval to try to avoid being coincident with 426 * callouts. 427 */ 428 systimer_init_periodic_flags(&gd->gd_statclock, statclock, 429 NULL, stathz, 430 SYSTF_MSSYNC | SYSTF_FIRST | 431 SYSTF_OFFSET50); 432 systimer_init_periodic_flags(&gd->gd_hardclock, hardclock, 433 NULL, hz, SYSTF_MSSYNC); 434 } 435 lwkt_setcpu_self(ogd); 436 437 /* 438 * Regular data collection 439 */ 440 kcollect_register(KCOLLECT_USERPCT, "user", collect_cputime_callback, 441 KCOLLECT_SCALE(KCOLLECT_USERPCT_FORMAT, 0)); 442 kcollect_register(KCOLLECT_SYSTPCT, "syst", NULL, 443 KCOLLECT_SCALE(KCOLLECT_SYSTPCT_FORMAT, 0)); 444 kcollect_register(KCOLLECT_IDLEPCT, "idle", NULL, 445 KCOLLECT_SCALE(KCOLLECT_IDLEPCT_FORMAT, 0)); 446 } 447 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL); 448 449 /* 450 * This method is called on just the BSP, after all the usched implementations 451 * are initialized. This avoids races between usched initialization functions 452 * and usched_schedulerclock(). 453 */ 454 static 455 void 456 initclocks_usched(void *dummy) 457 { 458 struct globaldata *ogd = mycpu; 459 struct globaldata *gd; 460 int n; 461 462 for (n = 0; n < ncpus; ++n) { 463 lwkt_setcpu_self(globaldata_find(n)); 464 gd = mycpu; 465 466 /* XXX correct the frequency for scheduler / estcpu tests */ 467 systimer_init_periodic_flags(&gd->gd_schedclock, schedclock, 468 NULL, ESTCPUFREQ, SYSTF_MSSYNC); 469 } 470 lwkt_setcpu_self(ogd); 471 } 472 SYSINIT(clocks3, SI_BOOT2_USCHED, SI_ORDER_ANY, initclocks_usched, NULL); 473 474 /* 475 * This sets the current real time of day. Timespecs are in seconds and 476 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base, 477 * instead we adjust basetime so basetime + gd_* results in the current 478 * time of day. This way the gd_* fields are guaranteed to represent 479 * a monotonically increasing 'uptime' value. 480 * 481 * When set_timeofday() is called from userland, the system call forces it 482 * onto cpu #0 since only cpu #0 can update basetime_index. 483 */ 484 void 485 set_timeofday(struct timespec *ts) 486 { 487 struct timespec *nbt; 488 int ni; 489 490 /* 491 * XXX SMP / non-atomic basetime updates 492 */ 493 crit_enter(); 494 ni = (basetime_index + 1) & BASETIME_ARYMASK; 495 cpu_lfence(); 496 nbt = &basetime[ni]; 497 nanouptime(nbt); 498 nbt->tv_sec = ts->tv_sec - nbt->tv_sec; 499 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec; 500 if (nbt->tv_nsec < 0) { 501 nbt->tv_nsec += 1000000000; 502 --nbt->tv_sec; 503 } 504 505 /* 506 * Note that basetime diverges from boottime as the clock drift is 507 * compensated for, so we cannot do away with boottime. When setting 508 * the absolute time of day the drift is 0 (for an instant) and we 509 * can simply assign boottime to basetime. 510 * 511 * Note that nanouptime() is based on gd_time_seconds which is drift 512 * compensated up to a point (it is guaranteed to remain monotonically 513 * increasing). gd_time_seconds is thus our best uptime guess and 514 * suitable for use in the boottime calculation. It is already taken 515 * into account in the basetime calculation above. 516 */ 517 spin_lock(&ntp_spin); 518 boottime.tv_sec = nbt->tv_sec; 519 ntp_delta = 0; 520 521 /* 522 * We now have a new basetime, make sure all other cpus have it, 523 * then update the index. 524 */ 525 cpu_sfence(); 526 basetime_index = ni; 527 spin_unlock(&ntp_spin); 528 529 crit_exit(); 530 } 531 532 /* 533 * Each cpu has its own hardclock, but we only increment ticks and softticks 534 * on cpu #0. 535 * 536 * NOTE! systimer! the MP lock might not be held here. We can only safely 537 * manipulate objects owned by the current cpu. 538 */ 539 static void 540 hardclock(systimer_t info, int in_ipi, struct intrframe *frame) 541 { 542 sysclock_t cputicks; 543 struct proc *p; 544 struct globaldata *gd = mycpu; 545 546 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) { 547 /* Defer to doreti on passive IPIQ processing */ 548 need_ipiq(); 549 } 550 551 /* 552 * We update the compensation base to calculate fine-grained time 553 * from the sys_cputimer on a per-cpu basis in order to avoid 554 * having to mess around with locks. sys_cputimer is assumed to 555 * be consistent across all cpus. CPU N copies the base state from 556 * CPU 0 using the same FIFO trick that we use for basetime (so we 557 * don't catch a CPU 0 update in the middle). 558 * 559 * Note that we never allow info->time (aka gd->gd_hardclock.time) 560 * to reverse index gd_cpuclock_base, but that it is possible for 561 * it to temporarily get behind in the seconds if something in the 562 * system locks interrupts for a long period of time. Since periodic 563 * timers count events, though everything should resynch again 564 * immediately. 565 */ 566 if (gd->gd_cpuid == 0) { 567 int ni; 568 569 cputicks = info->time - gd->gd_cpuclock_base; 570 if (cputicks >= sys_cputimer->freq) { 571 cputicks /= sys_cputimer->freq; 572 if (cputicks != 0 && cputicks != 1) 573 kprintf("Warning: hardclock missed > 1 sec\n"); 574 gd->gd_time_seconds += cputicks; 575 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks; 576 /* uncorrected monotonic 1-sec gran */ 577 time_uptime += cputicks; 578 } 579 ni = (basetime_index + 1) & BASETIME_ARYMASK; 580 hardtime[ni].time_second = gd->gd_time_seconds; 581 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base; 582 } else { 583 int ni; 584 585 ni = basetime_index; 586 cpu_lfence(); 587 gd->gd_time_seconds = hardtime[ni].time_second; 588 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base; 589 } 590 591 /* 592 * The system-wide ticks counter and NTP related timedelta/tickdelta 593 * adjustments only occur on cpu #0. NTP adjustments are accomplished 594 * by updating basetime. 595 */ 596 if (gd->gd_cpuid == 0) { 597 struct timespec *nbt; 598 struct timespec nts; 599 int leap; 600 int ni; 601 602 ++ticks; 603 604 #if 0 605 if (tco->tc_poll_pps) 606 tco->tc_poll_pps(tco); 607 #endif 608 609 /* 610 * Calculate the new basetime index. We are in a critical section 611 * on cpu #0 and can safely play with basetime_index. Start 612 * with the current basetime and then make adjustments. 613 */ 614 ni = (basetime_index + 1) & BASETIME_ARYMASK; 615 nbt = &basetime[ni]; 616 *nbt = basetime[basetime_index]; 617 618 /* 619 * ntp adjustments only occur on cpu 0 and are protected by 620 * ntp_spin. This spinlock virtually never conflicts. 621 */ 622 spin_lock(&ntp_spin); 623 624 /* 625 * Apply adjtime corrections. (adjtime() API) 626 * 627 * adjtime() only runs on cpu #0 so our critical section is 628 * sufficient to access these variables. 629 */ 630 if (ntp_delta != 0) { 631 nbt->tv_nsec += ntp_tick_delta; 632 ntp_delta -= ntp_tick_delta; 633 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) || 634 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) { 635 ntp_tick_delta = ntp_delta; 636 } 637 } 638 639 /* 640 * Apply permanent frequency corrections. (sysctl API) 641 */ 642 if (ntp_tick_permanent != 0) { 643 ntp_tick_acc += ntp_tick_permanent; 644 if (ntp_tick_acc >= (1LL << 32)) { 645 nbt->tv_nsec += ntp_tick_acc >> 32; 646 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32; 647 } else if (ntp_tick_acc <= -(1LL << 32)) { 648 /* Negate ntp_tick_acc to avoid shifting the sign bit. */ 649 nbt->tv_nsec -= (-ntp_tick_acc) >> 32; 650 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32; 651 } 652 } 653 654 if (nbt->tv_nsec >= 1000000000) { 655 nbt->tv_sec++; 656 nbt->tv_nsec -= 1000000000; 657 } else if (nbt->tv_nsec < 0) { 658 nbt->tv_sec--; 659 nbt->tv_nsec += 1000000000; 660 } 661 662 /* 663 * Another per-tick compensation. (for ntp_adjtime() API) 664 */ 665 if (nsec_adj != 0) { 666 nsec_acc += nsec_adj; 667 if (nsec_acc >= 0x100000000LL) { 668 nbt->tv_nsec += nsec_acc >> 32; 669 nsec_acc = (nsec_acc & 0xFFFFFFFFLL); 670 } else if (nsec_acc <= -0x100000000LL) { 671 nbt->tv_nsec -= -nsec_acc >> 32; 672 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL); 673 } 674 if (nbt->tv_nsec >= 1000000000) { 675 nbt->tv_nsec -= 1000000000; 676 ++nbt->tv_sec; 677 } else if (nbt->tv_nsec < 0) { 678 nbt->tv_nsec += 1000000000; 679 --nbt->tv_sec; 680 } 681 } 682 spin_unlock(&ntp_spin); 683 684 /************************************************************ 685 * LEAP SECOND CORRECTION * 686 ************************************************************ 687 * 688 * Taking into account all the corrections made above, figure 689 * out the new real time. If the seconds field has changed 690 * then apply any pending leap-second corrections. 691 */ 692 getnanotime_nbt(nbt, &nts); 693 694 if (time_second != nts.tv_sec) { 695 /* 696 * Apply leap second (sysctl API). Adjust nts for changes 697 * so we do not have to call getnanotime_nbt again. 698 */ 699 if (ntp_leap_second) { 700 if (ntp_leap_second == nts.tv_sec) { 701 if (ntp_leap_insert) { 702 nbt->tv_sec++; 703 nts.tv_sec++; 704 } else { 705 nbt->tv_sec--; 706 nts.tv_sec--; 707 } 708 ntp_leap_second--; 709 } 710 } 711 712 /* 713 * Apply leap second (ntp_adjtime() API), calculate a new 714 * nsec_adj field. ntp_update_second() returns nsec_adj 715 * as a per-second value but we need it as a per-tick value. 716 */ 717 leap = ntp_update_second(time_second, &nsec_adj); 718 nsec_adj /= hz; 719 nbt->tv_sec += leap; 720 nts.tv_sec += leap; 721 722 /* 723 * Update the time_second 'approximate time' global. 724 */ 725 time_second = nts.tv_sec; 726 727 /* 728 * Clear the IPC hint for the currently running thread once 729 * per second, allowing us to disconnect the hint from a 730 * thread which may no longer care. 731 */ 732 curthread->td_wakefromcpu = -1; 733 734 } 735 736 /* 737 * Finally, our new basetime is ready to go live! 738 */ 739 cpu_sfence(); 740 basetime_index = ni; 741 742 /* 743 * Update kpmap on each tick. TS updates are integrated with 744 * fences and upticks allowing userland to read the data 745 * deterministically. 746 */ 747 if (kpmap) { 748 int w; 749 750 w = (kpmap->upticks + 1) & 1; 751 getnanouptime(&kpmap->ts_uptime[w]); 752 getnanotime(&kpmap->ts_realtime[w]); 753 cpu_sfence(); 754 ++kpmap->upticks; 755 cpu_sfence(); 756 } 757 } 758 759 /* 760 * lwkt thread scheduler fair queueing 761 */ 762 lwkt_schedulerclock(curthread); 763 764 /* 765 * softticks are handled for all cpus 766 */ 767 hardclock_softtick(gd); 768 769 /* 770 * Rollup accumulated vmstats, copy-back for critical path checks. 771 */ 772 vmstats_rollup_cpu(gd); 773 vfscache_rollup_cpu(gd); 774 mycpu->gd_vmstats = vmstats; 775 776 /* 777 * ITimer handling is per-tick, per-cpu. 778 * 779 * We must acquire the per-process token in order for ksignal() 780 * to be non-blocking. For the moment this requires an AST fault, 781 * the ksignal() cannot be safely issued from this hard interrupt. 782 * 783 * XXX Even the trytoken here isn't right, and itimer operation in 784 * a multi threaded environment is going to be weird at the 785 * very least. 786 */ 787 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) { 788 crit_enter_hard(); 789 if (p->p_upmap) 790 ++p->p_upmap->runticks; 791 792 if (frame && CLKF_USERMODE(frame) && 793 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) && 794 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) { 795 p->p_flags |= P_SIGVTALRM; 796 need_user_resched(); 797 } 798 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) && 799 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) { 800 p->p_flags |= P_SIGPROF; 801 need_user_resched(); 802 } 803 crit_exit_hard(); 804 lwkt_reltoken(&p->p_token); 805 } 806 setdelayed(); 807 } 808 809 /* 810 * The statistics clock typically runs at a 125Hz rate, and is intended 811 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu. 812 * 813 * NOTE! systimer! the MP lock might not be held here. We can only safely 814 * manipulate objects owned by the current cpu. 815 * 816 * The stats clock is responsible for grabbing a profiling sample. 817 * Most of the statistics are only used by user-level statistics programs. 818 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and 819 * p->p_estcpu. 820 * 821 * Like the other clocks, the stat clock is called from what is effectively 822 * a fast interrupt, so the context should be the thread/process that got 823 * interrupted. 824 */ 825 static void 826 statclock(systimer_t info, int in_ipi, struct intrframe *frame) 827 { 828 globaldata_t gd = mycpu; 829 thread_t td; 830 struct proc *p; 831 int bump; 832 sysclock_t cv; 833 sysclock_t scv; 834 835 /* 836 * How big was our timeslice relative to the last time? Calculate 837 * in microseconds. 838 * 839 * NOTE: Use of microuptime() is typically MPSAFE, but usually not 840 * during early boot. Just use the systimer count to be nice 841 * to e.g. qemu. The systimer has a better chance of being 842 * MPSAFE at early boot. 843 */ 844 cv = sys_cputimer->count(); 845 scv = gd->statint.gd_statcv; 846 if (scv == 0) { 847 bump = 1; 848 } else { 849 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32; 850 if (bump < 0) 851 bump = 0; 852 if (bump > 1000000) 853 bump = 1000000; 854 } 855 gd->statint.gd_statcv = cv; 856 857 #if 0 858 stv = &gd->gd_stattv; 859 if (stv->tv_sec == 0) { 860 bump = 1; 861 } else { 862 bump = tv.tv_usec - stv->tv_usec + 863 (tv.tv_sec - stv->tv_sec) * 1000000; 864 if (bump < 0) 865 bump = 0; 866 if (bump > 1000000) 867 bump = 1000000; 868 } 869 *stv = tv; 870 #endif 871 872 td = curthread; 873 p = td->td_proc; 874 875 if (frame && CLKF_USERMODE(frame)) { 876 /* 877 * Came from userland, handle user time and deal with 878 * possible process. 879 */ 880 if (p && (p->p_flags & P_PROFIL)) 881 addupc_intr(p, CLKF_PC(frame), 1); 882 td->td_uticks += bump; 883 884 /* 885 * Charge the time as appropriate 886 */ 887 if (p && p->p_nice > NZERO) 888 cpu_time.cp_nice += bump; 889 else 890 cpu_time.cp_user += bump; 891 } else { 892 int intr_nest = gd->gd_intr_nesting_level; 893 894 if (in_ipi) { 895 /* 896 * IPI processing code will bump gd_intr_nesting_level 897 * up by one, which breaks following CLKF_INTR testing, 898 * so we subtract it by one here. 899 */ 900 --intr_nest; 901 } 902 903 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td)) 904 905 /* 906 * Came from kernel mode, so we were: 907 * - handling an interrupt, 908 * - doing syscall or trap work on behalf of the current 909 * user process, or 910 * - spinning in the idle loop. 911 * Whichever it is, charge the time as appropriate. 912 * Note that we charge interrupts to the current process, 913 * regardless of whether they are ``for'' that process, 914 * so that we know how much of its real time was spent 915 * in ``non-process'' (i.e., interrupt) work. 916 * 917 * XXX assume system if frame is NULL. A NULL frame 918 * can occur if ipi processing is done from a crit_exit(). 919 */ 920 if (IS_INTR_RUNNING || 921 (gd->gd_reqflags & RQF_INTPEND)) { 922 /* 923 * If we interrupted an interrupt thread, well, 924 * count it as interrupt time. 925 */ 926 td->td_iticks += bump; 927 #ifdef DEBUG_PCTRACK 928 if (frame) 929 do_pctrack(frame, PCTRACK_INT); 930 #endif 931 cpu_time.cp_intr += bump; 932 } else if (gd->gd_flags & GDF_VIRTUSER) { 933 /* 934 * The vkernel doesn't do a good job providing trap 935 * frames that we can test. If the GDF_VIRTUSER 936 * flag is set we probably interrupted user mode. 937 * 938 * We also use this flag on the host when entering 939 * VMM mode. 940 */ 941 td->td_uticks += bump; 942 943 /* 944 * Charge the time as appropriate 945 */ 946 if (p && p->p_nice > NZERO) 947 cpu_time.cp_nice += bump; 948 else 949 cpu_time.cp_user += bump; 950 } else { 951 td->td_sticks += bump; 952 if (td == &gd->gd_idlethread) { 953 /* 954 * We want to count token contention as 955 * system time. When token contention occurs 956 * the cpu may only be outside its critical 957 * section while switching through the idle 958 * thread. In this situation, various flags 959 * will be set in gd_reqflags. 960 */ 961 if (gd->gd_reqflags & RQF_IDLECHECK_WK_MASK) 962 cpu_time.cp_sys += bump; 963 else 964 cpu_time.cp_idle += bump; 965 } else { 966 /* 967 * System thread was running. 968 */ 969 #ifdef DEBUG_PCTRACK 970 if (frame) 971 do_pctrack(frame, PCTRACK_SYS); 972 #endif 973 cpu_time.cp_sys += bump; 974 } 975 } 976 977 #undef IS_INTR_RUNNING 978 } 979 } 980 981 #ifdef DEBUG_PCTRACK 982 /* 983 * Sample the PC when in the kernel or in an interrupt. User code can 984 * retrieve the information and generate a histogram or other output. 985 */ 986 987 static void 988 do_pctrack(struct intrframe *frame, int which) 989 { 990 struct kinfo_pctrack *pctrack; 991 992 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which]; 993 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] = 994 (void *)CLKF_PC(frame); 995 ++pctrack->pc_index; 996 } 997 998 static int 999 sysctl_pctrack(SYSCTL_HANDLER_ARGS) 1000 { 1001 struct kinfo_pcheader head; 1002 int error; 1003 int cpu; 1004 int ntrack; 1005 1006 head.pc_ntrack = PCTRACK_SIZE; 1007 head.pc_arysize = PCTRACK_ARYSIZE; 1008 1009 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0) 1010 return (error); 1011 1012 for (cpu = 0; cpu < ncpus; ++cpu) { 1013 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) { 1014 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack], 1015 sizeof(struct kinfo_pctrack)); 1016 if (error) 1017 break; 1018 } 1019 if (error) 1020 break; 1021 } 1022 return (error); 1023 } 1024 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0, 1025 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking"); 1026 1027 #endif 1028 1029 /* 1030 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer, 1031 * the MP lock might not be held. We can safely manipulate parts of curproc 1032 * but that's about it. 1033 * 1034 * Each cpu has its own scheduler clock. 1035 */ 1036 static void 1037 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame) 1038 { 1039 struct lwp *lp; 1040 struct rusage *ru; 1041 struct vmspace *vm; 1042 long rss; 1043 1044 if ((lp = lwkt_preempted_proc()) != NULL) { 1045 /* 1046 * Account for cpu time used and hit the scheduler. Note 1047 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD 1048 * HERE. 1049 */ 1050 ++lp->lwp_cpticks; 1051 usched_schedulerclock(lp, info->periodic, info->time); 1052 } else { 1053 usched_schedulerclock(NULL, info->periodic, info->time); 1054 } 1055 if ((lp = curthread->td_lwp) != NULL) { 1056 /* 1057 * Update resource usage integrals and maximums. 1058 */ 1059 if ((ru = &lp->lwp_proc->p_ru) && 1060 (vm = lp->lwp_proc->p_vmspace) != NULL) { 1061 ru->ru_ixrss += pgtok(btoc(vm->vm_tsize)); 1062 ru->ru_idrss += pgtok(btoc(vm->vm_dsize)); 1063 ru->ru_isrss += pgtok(btoc(vm->vm_ssize)); 1064 if (lwkt_trytoken(&vm->vm_map.token)) { 1065 rss = pgtok(vmspace_resident_count(vm)); 1066 if (ru->ru_maxrss < rss) 1067 ru->ru_maxrss = rss; 1068 lwkt_reltoken(&vm->vm_map.token); 1069 } 1070 } 1071 } 1072 /* Increment the global sched_ticks */ 1073 if (mycpu->gd_cpuid == 0) 1074 ++sched_ticks; 1075 } 1076 1077 /* 1078 * Compute number of ticks for the specified amount of time. The 1079 * return value is intended to be used in a clock interrupt timed 1080 * operation and guaranteed to meet or exceed the requested time. 1081 * If the representation overflows, return INT_MAX. The minimum return 1082 * value is 1 ticks and the function will average the calculation up. 1083 * If any value greater then 0 microseconds is supplied, a value 1084 * of at least 2 will be returned to ensure that a near-term clock 1085 * interrupt does not cause the timeout to occur (degenerately) early. 1086 * 1087 * Note that limit checks must take into account microseconds, which is 1088 * done simply by using the smaller signed long maximum instead of 1089 * the unsigned long maximum. 1090 * 1091 * If ints have 32 bits, then the maximum value for any timeout in 1092 * 10ms ticks is 248 days. 1093 */ 1094 int 1095 tvtohz_high(struct timeval *tv) 1096 { 1097 int ticks; 1098 long sec, usec; 1099 1100 sec = tv->tv_sec; 1101 usec = tv->tv_usec; 1102 if (usec < 0) { 1103 sec--; 1104 usec += 1000000; 1105 } 1106 if (sec < 0) { 1107 #ifdef DIAGNOSTIC 1108 if (usec > 0) { 1109 sec++; 1110 usec -= 1000000; 1111 } 1112 kprintf("tvtohz_high: negative time difference " 1113 "%ld sec %ld usec\n", 1114 sec, usec); 1115 #endif 1116 ticks = 1; 1117 } else if (sec <= INT_MAX / hz) { 1118 ticks = (int)(sec * hz + howmany((u_long)usec, ustick)) + 1; 1119 } else { 1120 ticks = INT_MAX; 1121 } 1122 return (ticks); 1123 } 1124 1125 int 1126 tstohz_high(struct timespec *ts) 1127 { 1128 int ticks; 1129 long sec, nsec; 1130 1131 sec = ts->tv_sec; 1132 nsec = ts->tv_nsec; 1133 if (nsec < 0) { 1134 sec--; 1135 nsec += 1000000000; 1136 } 1137 if (sec < 0) { 1138 #ifdef DIAGNOSTIC 1139 if (nsec > 0) { 1140 sec++; 1141 nsec -= 1000000000; 1142 } 1143 kprintf("tstohz_high: negative time difference " 1144 "%ld sec %ld nsec\n", 1145 sec, nsec); 1146 #endif 1147 ticks = 1; 1148 } else if (sec <= INT_MAX / hz) { 1149 ticks = (int)(sec * hz + howmany((u_long)nsec, nstick)) + 1; 1150 } else { 1151 ticks = INT_MAX; 1152 } 1153 return (ticks); 1154 } 1155 1156 1157 /* 1158 * Compute number of ticks for the specified amount of time, erroring on 1159 * the side of it being too low to ensure that sleeping the returned number 1160 * of ticks will not result in a late return. 1161 * 1162 * The supplied timeval may not be negative and should be normalized. A 1163 * return value of 0 is possible if the timeval converts to less then 1164 * 1 tick. 1165 * 1166 * If ints have 32 bits, then the maximum value for any timeout in 1167 * 10ms ticks is 248 days. 1168 */ 1169 int 1170 tvtohz_low(struct timeval *tv) 1171 { 1172 int ticks; 1173 long sec; 1174 1175 sec = tv->tv_sec; 1176 if (sec <= INT_MAX / hz) 1177 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick); 1178 else 1179 ticks = INT_MAX; 1180 return (ticks); 1181 } 1182 1183 int 1184 tstohz_low(struct timespec *ts) 1185 { 1186 int ticks; 1187 long sec; 1188 1189 sec = ts->tv_sec; 1190 if (sec <= INT_MAX / hz) 1191 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick); 1192 else 1193 ticks = INT_MAX; 1194 return (ticks); 1195 } 1196 1197 /* 1198 * Start profiling on a process. 1199 * 1200 * Caller must hold p->p_token(); 1201 * 1202 * Kernel profiling passes proc0 which never exits and hence 1203 * keeps the profile clock running constantly. 1204 */ 1205 void 1206 startprofclock(struct proc *p) 1207 { 1208 if ((p->p_flags & P_PROFIL) == 0) { 1209 p->p_flags |= P_PROFIL; 1210 #if 0 /* XXX */ 1211 if (++profprocs == 1 && stathz != 0) { 1212 crit_enter(); 1213 psdiv = psratio; 1214 setstatclockrate(profhz); 1215 crit_exit(); 1216 } 1217 #endif 1218 } 1219 } 1220 1221 /* 1222 * Stop profiling on a process. 1223 * 1224 * caller must hold p->p_token 1225 */ 1226 void 1227 stopprofclock(struct proc *p) 1228 { 1229 if (p->p_flags & P_PROFIL) { 1230 p->p_flags &= ~P_PROFIL; 1231 #if 0 /* XXX */ 1232 if (--profprocs == 0 && stathz != 0) { 1233 crit_enter(); 1234 psdiv = 1; 1235 setstatclockrate(stathz); 1236 crit_exit(); 1237 } 1238 #endif 1239 } 1240 } 1241 1242 /* 1243 * Return information about system clocks. 1244 */ 1245 static int 1246 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 1247 { 1248 struct kinfo_clockinfo clkinfo; 1249 /* 1250 * Construct clockinfo structure. 1251 */ 1252 clkinfo.ci_hz = hz; 1253 clkinfo.ci_tick = ustick; 1254 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000; 1255 clkinfo.ci_profhz = profhz; 1256 clkinfo.ci_stathz = stathz ? stathz : hz; 1257 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 1258 } 1259 1260 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 1261 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 1262 1263 /* 1264 * We have eight functions for looking at the clock, four for 1265 * microseconds and four for nanoseconds. For each there is fast 1266 * but less precise version "get{nano|micro}[up]time" which will 1267 * return a time which is up to 1/HZ previous to the call, whereas 1268 * the raw version "{nano|micro}[up]time" will return a timestamp 1269 * which is as precise as possible. The "up" variants return the 1270 * time relative to system boot, these are well suited for time 1271 * interval measurements. 1272 * 1273 * Each cpu independently maintains the current time of day, so all 1274 * we need to do to protect ourselves from changes is to do a loop 1275 * check on the seconds field changing out from under us. 1276 * 1277 * The system timer maintains a 32 bit count and due to various issues 1278 * it is possible for the calculated delta to occasionally exceed 1279 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec 1280 * multiplication can easily overflow, so we deal with the case. For 1281 * uniformity we deal with the case in the usec case too. 1282 * 1283 * All the [get][micro,nano][time,uptime]() routines are MPSAFE. 1284 */ 1285 void 1286 getmicrouptime(struct timeval *tvp) 1287 { 1288 struct globaldata *gd = mycpu; 1289 sysclock_t delta; 1290 1291 do { 1292 tvp->tv_sec = gd->gd_time_seconds; 1293 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1294 } while (tvp->tv_sec != gd->gd_time_seconds); 1295 1296 if (delta >= sys_cputimer->freq) { 1297 tvp->tv_sec += delta / sys_cputimer->freq; 1298 delta %= sys_cputimer->freq; 1299 } 1300 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1301 if (tvp->tv_usec >= 1000000) { 1302 tvp->tv_usec -= 1000000; 1303 ++tvp->tv_sec; 1304 } 1305 } 1306 1307 void 1308 getnanouptime(struct timespec *tsp) 1309 { 1310 struct globaldata *gd = mycpu; 1311 sysclock_t delta; 1312 1313 do { 1314 tsp->tv_sec = gd->gd_time_seconds; 1315 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1316 } while (tsp->tv_sec != gd->gd_time_seconds); 1317 1318 if (delta >= sys_cputimer->freq) { 1319 tsp->tv_sec += delta / sys_cputimer->freq; 1320 delta %= sys_cputimer->freq; 1321 } 1322 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1323 } 1324 1325 void 1326 microuptime(struct timeval *tvp) 1327 { 1328 struct globaldata *gd = mycpu; 1329 sysclock_t delta; 1330 1331 do { 1332 tvp->tv_sec = gd->gd_time_seconds; 1333 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1334 } while (tvp->tv_sec != gd->gd_time_seconds); 1335 1336 if (delta >= sys_cputimer->freq) { 1337 tvp->tv_sec += delta / sys_cputimer->freq; 1338 delta %= sys_cputimer->freq; 1339 } 1340 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1341 } 1342 1343 void 1344 nanouptime(struct timespec *tsp) 1345 { 1346 struct globaldata *gd = mycpu; 1347 sysclock_t delta; 1348 1349 do { 1350 tsp->tv_sec = gd->gd_time_seconds; 1351 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1352 } while (tsp->tv_sec != gd->gd_time_seconds); 1353 1354 if (delta >= sys_cputimer->freq) { 1355 tsp->tv_sec += delta / sys_cputimer->freq; 1356 delta %= sys_cputimer->freq; 1357 } 1358 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1359 } 1360 1361 /* 1362 * realtime routines 1363 */ 1364 void 1365 getmicrotime(struct timeval *tvp) 1366 { 1367 struct globaldata *gd = mycpu; 1368 struct timespec *bt; 1369 sysclock_t delta; 1370 1371 do { 1372 tvp->tv_sec = gd->gd_time_seconds; 1373 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1374 } while (tvp->tv_sec != gd->gd_time_seconds); 1375 1376 if (delta >= sys_cputimer->freq) { 1377 tvp->tv_sec += delta / sys_cputimer->freq; 1378 delta %= sys_cputimer->freq; 1379 } 1380 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1381 1382 bt = &basetime[basetime_index]; 1383 cpu_lfence(); 1384 tvp->tv_sec += bt->tv_sec; 1385 tvp->tv_usec += bt->tv_nsec / 1000; 1386 while (tvp->tv_usec >= 1000000) { 1387 tvp->tv_usec -= 1000000; 1388 ++tvp->tv_sec; 1389 } 1390 } 1391 1392 void 1393 getnanotime(struct timespec *tsp) 1394 { 1395 struct globaldata *gd = mycpu; 1396 struct timespec *bt; 1397 sysclock_t delta; 1398 1399 do { 1400 tsp->tv_sec = gd->gd_time_seconds; 1401 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1402 } while (tsp->tv_sec != gd->gd_time_seconds); 1403 1404 if (delta >= sys_cputimer->freq) { 1405 tsp->tv_sec += delta / sys_cputimer->freq; 1406 delta %= sys_cputimer->freq; 1407 } 1408 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1409 1410 bt = &basetime[basetime_index]; 1411 cpu_lfence(); 1412 tsp->tv_sec += bt->tv_sec; 1413 tsp->tv_nsec += bt->tv_nsec; 1414 while (tsp->tv_nsec >= 1000000000) { 1415 tsp->tv_nsec -= 1000000000; 1416 ++tsp->tv_sec; 1417 } 1418 } 1419 1420 static void 1421 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp) 1422 { 1423 struct globaldata *gd = mycpu; 1424 sysclock_t delta; 1425 1426 do { 1427 tsp->tv_sec = gd->gd_time_seconds; 1428 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base; 1429 } while (tsp->tv_sec != gd->gd_time_seconds); 1430 1431 if (delta >= sys_cputimer->freq) { 1432 tsp->tv_sec += delta / sys_cputimer->freq; 1433 delta %= sys_cputimer->freq; 1434 } 1435 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1436 1437 tsp->tv_sec += nbt->tv_sec; 1438 tsp->tv_nsec += nbt->tv_nsec; 1439 while (tsp->tv_nsec >= 1000000000) { 1440 tsp->tv_nsec -= 1000000000; 1441 ++tsp->tv_sec; 1442 } 1443 } 1444 1445 1446 void 1447 microtime(struct timeval *tvp) 1448 { 1449 struct globaldata *gd = mycpu; 1450 struct timespec *bt; 1451 sysclock_t delta; 1452 1453 do { 1454 tvp->tv_sec = gd->gd_time_seconds; 1455 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1456 } while (tvp->tv_sec != gd->gd_time_seconds); 1457 1458 if (delta >= sys_cputimer->freq) { 1459 tvp->tv_sec += delta / sys_cputimer->freq; 1460 delta %= sys_cputimer->freq; 1461 } 1462 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32; 1463 1464 bt = &basetime[basetime_index]; 1465 cpu_lfence(); 1466 tvp->tv_sec += bt->tv_sec; 1467 tvp->tv_usec += bt->tv_nsec / 1000; 1468 while (tvp->tv_usec >= 1000000) { 1469 tvp->tv_usec -= 1000000; 1470 ++tvp->tv_sec; 1471 } 1472 } 1473 1474 void 1475 nanotime(struct timespec *tsp) 1476 { 1477 struct globaldata *gd = mycpu; 1478 struct timespec *bt; 1479 sysclock_t delta; 1480 1481 do { 1482 tsp->tv_sec = gd->gd_time_seconds; 1483 delta = sys_cputimer->count() - gd->gd_cpuclock_base; 1484 } while (tsp->tv_sec != gd->gd_time_seconds); 1485 1486 if (delta >= sys_cputimer->freq) { 1487 tsp->tv_sec += delta / sys_cputimer->freq; 1488 delta %= sys_cputimer->freq; 1489 } 1490 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1491 1492 bt = &basetime[basetime_index]; 1493 cpu_lfence(); 1494 tsp->tv_sec += bt->tv_sec; 1495 tsp->tv_nsec += bt->tv_nsec; 1496 while (tsp->tv_nsec >= 1000000000) { 1497 tsp->tv_nsec -= 1000000000; 1498 ++tsp->tv_sec; 1499 } 1500 } 1501 1502 /* 1503 * Get an approximate time_t. It does not have to be accurate. This 1504 * function is called only from KTR and can be called with the system in 1505 * any state so do not use a critical section or other complex operation 1506 * here. 1507 * 1508 * NOTE: This is not exactly synchronized with real time. To do that we 1509 * would have to do what microtime does and check for a nanoseconds 1510 * overflow. 1511 */ 1512 time_t 1513 get_approximate_time_t(void) 1514 { 1515 struct globaldata *gd = mycpu; 1516 struct timespec *bt; 1517 1518 bt = &basetime[basetime_index]; 1519 return(gd->gd_time_seconds + bt->tv_sec); 1520 } 1521 1522 int 1523 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 1524 { 1525 pps_params_t *app; 1526 struct pps_fetch_args *fapi; 1527 #ifdef PPS_SYNC 1528 struct pps_kcbind_args *kapi; 1529 #endif 1530 1531 switch (cmd) { 1532 case PPS_IOC_CREATE: 1533 return (0); 1534 case PPS_IOC_DESTROY: 1535 return (0); 1536 case PPS_IOC_SETPARAMS: 1537 app = (pps_params_t *)data; 1538 if (app->mode & ~pps->ppscap) 1539 return (EINVAL); 1540 pps->ppsparam = *app; 1541 return (0); 1542 case PPS_IOC_GETPARAMS: 1543 app = (pps_params_t *)data; 1544 *app = pps->ppsparam; 1545 app->api_version = PPS_API_VERS_1; 1546 return (0); 1547 case PPS_IOC_GETCAP: 1548 *(int*)data = pps->ppscap; 1549 return (0); 1550 case PPS_IOC_FETCH: 1551 fapi = (struct pps_fetch_args *)data; 1552 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) 1553 return (EINVAL); 1554 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) 1555 return (EOPNOTSUPP); 1556 pps->ppsinfo.current_mode = pps->ppsparam.mode; 1557 fapi->pps_info_buf = pps->ppsinfo; 1558 return (0); 1559 case PPS_IOC_KCBIND: 1560 #ifdef PPS_SYNC 1561 kapi = (struct pps_kcbind_args *)data; 1562 /* XXX Only root should be able to do this */ 1563 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) 1564 return (EINVAL); 1565 if (kapi->kernel_consumer != PPS_KC_HARDPPS) 1566 return (EINVAL); 1567 if (kapi->edge & ~pps->ppscap) 1568 return (EINVAL); 1569 pps->kcmode = kapi->edge; 1570 return (0); 1571 #else 1572 return (EOPNOTSUPP); 1573 #endif 1574 default: 1575 return (ENOTTY); 1576 } 1577 } 1578 1579 void 1580 pps_init(struct pps_state *pps) 1581 { 1582 pps->ppscap |= PPS_TSFMT_TSPEC; 1583 if (pps->ppscap & PPS_CAPTUREASSERT) 1584 pps->ppscap |= PPS_OFFSETASSERT; 1585 if (pps->ppscap & PPS_CAPTURECLEAR) 1586 pps->ppscap |= PPS_OFFSETCLEAR; 1587 } 1588 1589 void 1590 pps_event(struct pps_state *pps, sysclock_t count, int event) 1591 { 1592 struct globaldata *gd; 1593 struct timespec *tsp; 1594 struct timespec *osp; 1595 struct timespec *bt; 1596 struct timespec ts; 1597 sysclock_t *pcount; 1598 #ifdef PPS_SYNC 1599 sysclock_t tcount; 1600 #endif 1601 sysclock_t delta; 1602 pps_seq_t *pseq; 1603 int foff; 1604 #ifdef PPS_SYNC 1605 int fhard; 1606 #endif 1607 int ni; 1608 1609 gd = mycpu; 1610 1611 /* Things would be easier with arrays... */ 1612 if (event == PPS_CAPTUREASSERT) { 1613 tsp = &pps->ppsinfo.assert_timestamp; 1614 osp = &pps->ppsparam.assert_offset; 1615 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 1616 #ifdef PPS_SYNC 1617 fhard = pps->kcmode & PPS_CAPTUREASSERT; 1618 #endif 1619 pcount = &pps->ppscount[0]; 1620 pseq = &pps->ppsinfo.assert_sequence; 1621 } else { 1622 tsp = &pps->ppsinfo.clear_timestamp; 1623 osp = &pps->ppsparam.clear_offset; 1624 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 1625 #ifdef PPS_SYNC 1626 fhard = pps->kcmode & PPS_CAPTURECLEAR; 1627 #endif 1628 pcount = &pps->ppscount[1]; 1629 pseq = &pps->ppsinfo.clear_sequence; 1630 } 1631 1632 /* Nothing really happened */ 1633 if (*pcount == count) 1634 return; 1635 1636 *pcount = count; 1637 1638 do { 1639 ts.tv_sec = gd->gd_time_seconds; 1640 delta = count - gd->gd_cpuclock_base; 1641 } while (ts.tv_sec != gd->gd_time_seconds); 1642 1643 if (delta >= sys_cputimer->freq) { 1644 ts.tv_sec += delta / sys_cputimer->freq; 1645 delta %= sys_cputimer->freq; 1646 } 1647 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32; 1648 ni = basetime_index; 1649 cpu_lfence(); 1650 bt = &basetime[ni]; 1651 ts.tv_sec += bt->tv_sec; 1652 ts.tv_nsec += bt->tv_nsec; 1653 while (ts.tv_nsec >= 1000000000) { 1654 ts.tv_nsec -= 1000000000; 1655 ++ts.tv_sec; 1656 } 1657 1658 (*pseq)++; 1659 *tsp = ts; 1660 1661 if (foff) { 1662 timespecadd(tsp, osp, tsp); 1663 if (tsp->tv_nsec < 0) { 1664 tsp->tv_nsec += 1000000000; 1665 tsp->tv_sec -= 1; 1666 } 1667 } 1668 #ifdef PPS_SYNC 1669 if (fhard) { 1670 /* magic, at its best... */ 1671 tcount = count - pps->ppscount[2]; 1672 pps->ppscount[2] = count; 1673 if (tcount >= sys_cputimer->freq) { 1674 delta = (1000000000 * (tcount / sys_cputimer->freq) + 1675 sys_cputimer->freq64_nsec * 1676 (tcount % sys_cputimer->freq)) >> 32; 1677 } else { 1678 delta = (sys_cputimer->freq64_nsec * tcount) >> 32; 1679 } 1680 hardpps(tsp, delta); 1681 } 1682 #endif 1683 } 1684 1685 /* 1686 * Return the tsc target value for a delay of (ns). 1687 * 1688 * Returns -1 if the TSC is not supported. 1689 */ 1690 tsc_uclock_t 1691 tsc_get_target(int ns) 1692 { 1693 #if defined(_RDTSC_SUPPORTED_) 1694 if (cpu_feature & CPUID_TSC) { 1695 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000); 1696 } 1697 #endif 1698 return(-1); 1699 } 1700 1701 /* 1702 * Compare the tsc against the passed target 1703 * 1704 * Returns +1 if the target has been reached 1705 * Returns 0 if the target has not yet been reached 1706 * Returns -1 if the TSC is not supported. 1707 * 1708 * Typical use: while (tsc_test_target(target) == 0) { ...poll... } 1709 */ 1710 int 1711 tsc_test_target(int64_t target) 1712 { 1713 #if defined(_RDTSC_SUPPORTED_) 1714 if (cpu_feature & CPUID_TSC) { 1715 if ((int64_t)(target - rdtsc()) <= 0) 1716 return(1); 1717 return(0); 1718 } 1719 #endif 1720 return(-1); 1721 } 1722 1723 /* 1724 * Delay the specified number of nanoseconds using the tsc. This function 1725 * returns immediately if the TSC is not supported. At least one cpu_pause() 1726 * will be issued. 1727 */ 1728 void 1729 tsc_delay(int ns) 1730 { 1731 int64_t clk; 1732 1733 clk = tsc_get_target(ns); 1734 cpu_pause(); 1735 cpu_pause(); 1736 while (tsc_test_target(clk) == 0) { 1737 cpu_pause(); 1738 cpu_pause(); 1739 cpu_pause(); 1740 cpu_pause(); 1741 } 1742 } 1743