1 /* $NetBSD: kern_clock.c,v 1.81 2002/11/02 07:25:19 perry Exp $ */ 2 3 /*- 4 * Copyright (c) 2000 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility, 9 * NASA Ames Research Center. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. All advertising materials mentioning features or use of this software 20 * must display the following acknowledgement: 21 * This product includes software developed by the NetBSD 22 * Foundation, Inc. and its contributors. 23 * 4. Neither the name of The NetBSD Foundation nor the names of its 24 * contributors may be used to endorse or promote products derived 25 * from this software without specific prior written permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 28 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 29 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 30 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 31 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 32 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 33 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 34 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 35 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 36 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 37 * POSSIBILITY OF SUCH DAMAGE. 38 */ 39 40 /*- 41 * Copyright (c) 1982, 1986, 1991, 1993 42 * The Regents of the University of California. All rights reserved. 43 * (c) UNIX System Laboratories, Inc. 44 * All or some portions of this file are derived from material licensed 45 * to the University of California by American Telephone and Telegraph 46 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 47 * the permission of UNIX System Laboratories, Inc. 48 * 49 * Redistribution and use in source and binary forms, with or without 50 * modification, are permitted provided that the following conditions 51 * are met: 52 * 1. Redistributions of source code must retain the above copyright 53 * notice, this list of conditions and the following disclaimer. 54 * 2. Redistributions in binary form must reproduce the above copyright 55 * notice, this list of conditions and the following disclaimer in the 56 * documentation and/or other materials provided with the distribution. 57 * 3. All advertising materials mentioning features or use of this software 58 * must display the following acknowledgement: 59 * This product includes software developed by the University of 60 * California, Berkeley and its contributors. 61 * 4. Neither the name of the University nor the names of its contributors 62 * may be used to endorse or promote products derived from this software 63 * without specific prior written permission. 64 * 65 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 66 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 67 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 68 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 69 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 70 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 71 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 72 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 73 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 74 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 75 * SUCH DAMAGE. 76 * 77 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 78 */ 79 80 #include <sys/cdefs.h> 81 __KERNEL_RCSID(0, "$NetBSD: kern_clock.c,v 1.81 2002/11/02 07:25:19 perry Exp $"); 82 83 #include "opt_callout.h" 84 #include "opt_ntp.h" 85 #include "opt_perfctrs.h" 86 87 #include <sys/param.h> 88 #include <sys/systm.h> 89 #include <sys/dkstat.h> 90 #include <sys/callout.h> 91 #include <sys/kernel.h> 92 #include <sys/proc.h> 93 #include <sys/resourcevar.h> 94 #include <sys/signalvar.h> 95 #include <sys/sysctl.h> 96 #include <sys/timex.h> 97 #include <sys/sched.h> 98 #ifdef CALLWHEEL_STATS 99 #include <sys/device.h> 100 #endif 101 102 #include <machine/cpu.h> 103 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 104 #include <machine/intr.h> 105 #endif 106 107 #ifdef GPROF 108 #include <sys/gmon.h> 109 #endif 110 111 /* 112 * Clock handling routines. 113 * 114 * This code is written to operate with two timers that run independently of 115 * each other. The main clock, running hz times per second, is used to keep 116 * track of real time. The second timer handles kernel and user profiling, 117 * and does resource use estimation. If the second timer is programmable, 118 * it is randomized to avoid aliasing between the two clocks. For example, 119 * the randomization prevents an adversary from always giving up the cpu 120 * just before its quantum expires. Otherwise, it would never accumulate 121 * cpu ticks. The mean frequency of the second timer is stathz. 122 * 123 * If no second timer exists, stathz will be zero; in this case we drive 124 * profiling and statistics off the main clock. This WILL NOT be accurate; 125 * do not do it unless absolutely necessary. 126 * 127 * The statistics clock may (or may not) be run at a higher rate while 128 * profiling. This profile clock runs at profhz. We require that profhz 129 * be an integral multiple of stathz. 130 * 131 * If the statistics clock is running fast, it must be divided by the ratio 132 * profhz/stathz for statistics. (For profiling, every tick counts.) 133 */ 134 135 #ifdef NTP /* NTP phase-locked loop in kernel */ 136 /* 137 * Phase/frequency-lock loop (PLL/FLL) definitions 138 * 139 * The following variables are read and set by the ntp_adjtime() system 140 * call. 141 * 142 * time_state shows the state of the system clock, with values defined 143 * in the timex.h header file. 144 * 145 * time_status shows the status of the system clock, with bits defined 146 * in the timex.h header file. 147 * 148 * time_offset is used by the PLL/FLL to adjust the system time in small 149 * increments. 150 * 151 * time_constant determines the bandwidth or "stiffness" of the PLL. 152 * 153 * time_tolerance determines maximum frequency error or tolerance of the 154 * CPU clock oscillator and is a property of the architecture; however, 155 * in principle it could change as result of the presence of external 156 * discipline signals, for instance. 157 * 158 * time_precision is usually equal to the kernel tick variable; however, 159 * in cases where a precision clock counter or external clock is 160 * available, the resolution can be much less than this and depend on 161 * whether the external clock is working or not. 162 * 163 * time_maxerror is initialized by a ntp_adjtime() call and increased by 164 * the kernel once each second to reflect the maximum error bound 165 * growth. 166 * 167 * time_esterror is set and read by the ntp_adjtime() call, but 168 * otherwise not used by the kernel. 169 */ 170 int time_state = TIME_OK; /* clock state */ 171 int time_status = STA_UNSYNC; /* clock status bits */ 172 long time_offset = 0; /* time offset (us) */ 173 long time_constant = 0; /* pll time constant */ 174 long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */ 175 long time_precision = 1; /* clock precision (us) */ 176 long time_maxerror = MAXPHASE; /* maximum error (us) */ 177 long time_esterror = MAXPHASE; /* estimated error (us) */ 178 179 /* 180 * The following variables establish the state of the PLL/FLL and the 181 * residual time and frequency offset of the local clock. The scale 182 * factors are defined in the timex.h header file. 183 * 184 * time_phase and time_freq are the phase increment and the frequency 185 * increment, respectively, of the kernel time variable. 186 * 187 * time_freq is set via ntp_adjtime() from a value stored in a file when 188 * the synchronization daemon is first started. Its value is retrieved 189 * via ntp_adjtime() and written to the file about once per hour by the 190 * daemon. 191 * 192 * time_adj is the adjustment added to the value of tick at each timer 193 * interrupt and is recomputed from time_phase and time_freq at each 194 * seconds rollover. 195 * 196 * time_reftime is the second's portion of the system time at the last 197 * call to ntp_adjtime(). It is used to adjust the time_freq variable 198 * and to increase the time_maxerror as the time since last update 199 * increases. 200 */ 201 long time_phase = 0; /* phase offset (scaled us) */ 202 long time_freq = 0; /* frequency offset (scaled ppm) */ 203 long time_adj = 0; /* tick adjust (scaled 1 / hz) */ 204 long time_reftime = 0; /* time at last adjustment (s) */ 205 206 #ifdef PPS_SYNC 207 /* 208 * The following variables are used only if the kernel PPS discipline 209 * code is configured (PPS_SYNC). The scale factors are defined in the 210 * timex.h header file. 211 * 212 * pps_time contains the time at each calibration interval, as read by 213 * microtime(). pps_count counts the seconds of the calibration 214 * interval, the duration of which is nominally pps_shift in powers of 215 * two. 216 * 217 * pps_offset is the time offset produced by the time median filter 218 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by 219 * this filter. 220 * 221 * pps_freq is the frequency offset produced by the frequency median 222 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured 223 * by this filter. 224 * 225 * pps_usec is latched from a high resolution counter or external clock 226 * at pps_time. Here we want the hardware counter contents only, not the 227 * contents plus the time_tv.usec as usual. 228 * 229 * pps_valid counts the number of seconds since the last PPS update. It 230 * is used as a watchdog timer to disable the PPS discipline should the 231 * PPS signal be lost. 232 * 233 * pps_glitch counts the number of seconds since the beginning of an 234 * offset burst more than tick/2 from current nominal offset. It is used 235 * mainly to suppress error bursts due to priority conflicts between the 236 * PPS interrupt and timer interrupt. 237 * 238 * pps_intcnt counts the calibration intervals for use in the interval- 239 * adaptation algorithm. It's just too complicated for words. 240 */ 241 struct timeval pps_time; /* kernel time at last interval */ 242 long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */ 243 long pps_offset = 0; /* pps time offset (us) */ 244 long pps_jitter = MAXTIME; /* time dispersion (jitter) (us) */ 245 long pps_ff[] = {0, 0, 0}; /* pps frequency offset median filter */ 246 long pps_freq = 0; /* frequency offset (scaled ppm) */ 247 long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */ 248 long pps_usec = 0; /* microsec counter at last interval */ 249 long pps_valid = PPS_VALID; /* pps signal watchdog counter */ 250 int pps_glitch = 0; /* pps signal glitch counter */ 251 int pps_count = 0; /* calibration interval counter (s) */ 252 int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */ 253 int pps_intcnt = 0; /* intervals at current duration */ 254 255 /* 256 * PPS signal quality monitors 257 * 258 * pps_jitcnt counts the seconds that have been discarded because the 259 * jitter measured by the time median filter exceeds the limit MAXTIME 260 * (100 us). 261 * 262 * pps_calcnt counts the frequency calibration intervals, which are 263 * variable from 4 s to 256 s. 264 * 265 * pps_errcnt counts the calibration intervals which have been discarded 266 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the 267 * calibration interval jitter exceeds two ticks. 268 * 269 * pps_stbcnt counts the calibration intervals that have been discarded 270 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us). 271 */ 272 long pps_jitcnt = 0; /* jitter limit exceeded */ 273 long pps_calcnt = 0; /* calibration intervals */ 274 long pps_errcnt = 0; /* calibration errors */ 275 long pps_stbcnt = 0; /* stability limit exceeded */ 276 #endif /* PPS_SYNC */ 277 278 #ifdef EXT_CLOCK 279 /* 280 * External clock definitions 281 * 282 * The following definitions and declarations are used only if an 283 * external clock is configured on the system. 284 */ 285 #define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */ 286 287 /* 288 * The clock_count variable is set to CLOCK_INTERVAL at each PPS 289 * interrupt and decremented once each second. 290 */ 291 int clock_count = 0; /* CPU clock counter */ 292 293 #ifdef HIGHBALL 294 /* 295 * The clock_offset and clock_cpu variables are used by the HIGHBALL 296 * interface. The clock_offset variable defines the offset between 297 * system time and the HIGBALL counters. The clock_cpu variable contains 298 * the offset between the system clock and the HIGHBALL clock for use in 299 * disciplining the kernel time variable. 300 */ 301 extern struct timeval clock_offset; /* Highball clock offset */ 302 long clock_cpu = 0; /* CPU clock adjust */ 303 #endif /* HIGHBALL */ 304 #endif /* EXT_CLOCK */ 305 #endif /* NTP */ 306 307 308 /* 309 * Bump a timeval by a small number of usec's. 310 */ 311 #define BUMPTIME(t, usec) { \ 312 volatile struct timeval *tp = (t); \ 313 long us; \ 314 \ 315 tp->tv_usec = us = tp->tv_usec + (usec); \ 316 if (us >= 1000000) { \ 317 tp->tv_usec = us - 1000000; \ 318 tp->tv_sec++; \ 319 } \ 320 } 321 322 int stathz; 323 int profhz; 324 int profsrc; 325 int schedhz; 326 int profprocs; 327 int softclock_running; /* 1 => softclock() is running */ 328 static int psdiv; /* prof => stat divider */ 329 int psratio; /* ratio: prof / stat */ 330 int tickfix, tickfixinterval; /* used if tick not really integral */ 331 #ifndef NTP 332 static int tickfixcnt; /* accumulated fractional error */ 333 #else 334 int fixtick; /* used by NTP for same */ 335 int shifthz; 336 #endif 337 338 /* 339 * We might want ldd to load the both words from time at once. 340 * To succeed we need to be quadword aligned. 341 * The sparc already does that, and that it has worked so far is a fluke. 342 */ 343 volatile struct timeval time __attribute__((__aligned__(__alignof__(quad_t)))); 344 volatile struct timeval mono_time; 345 346 /* 347 * The callout mechanism is based on the work of Adam M. Costello and 348 * George Varghese, published in a technical report entitled "Redesigning 349 * the BSD Callout and Timer Facilities", and Justin Gibbs's subsequent 350 * integration into FreeBSD, modified for NetBSD by Jason R. Thorpe. 351 * 352 * The original work on the data structures used in this implementation 353 * was published by G. Varghese and A. Lauck in the paper "Hashed and 354 * Hierarchical Timing Wheels: Data Structures for the Efficient 355 * Implementation of a Timer Facility" in the Proceedings of the 11th 356 * ACM Annual Symposium on Operating System Principles, Austin, Texas, 357 * November 1987. 358 */ 359 struct callout_queue *callwheel; 360 int callwheelsize, callwheelbits, callwheelmask; 361 362 static struct callout *nextsoftcheck; /* next callout to be checked */ 363 364 #ifdef CALLWHEEL_STATS 365 int *callwheel_sizes; /* per-bucket length count */ 366 struct evcnt callwheel_collisions; /* number of hash collisions */ 367 struct evcnt callwheel_maxlength; /* length of the longest hash chain */ 368 struct evcnt callwheel_count; /* # callouts currently */ 369 struct evcnt callwheel_established; /* # callouts established */ 370 struct evcnt callwheel_fired; /* # callouts that fired */ 371 struct evcnt callwheel_disestablished; /* # callouts disestablished */ 372 struct evcnt callwheel_changed; /* # callouts changed */ 373 struct evcnt callwheel_softclocks; /* # times softclock() called */ 374 struct evcnt callwheel_softchecks; /* # checks per softclock() */ 375 struct evcnt callwheel_softempty; /* # empty buckets seen */ 376 struct evcnt callwheel_hintworked; /* # times hint saved scan */ 377 #endif /* CALLWHEEL_STATS */ 378 379 /* 380 * This value indicates the number of consecutive callouts that 381 * will be checked before we allow interrupts to have a chance 382 * again. 383 */ 384 #ifndef MAX_SOFTCLOCK_STEPS 385 #define MAX_SOFTCLOCK_STEPS 100 386 #endif 387 388 struct simplelock callwheel_slock; 389 390 #define CALLWHEEL_LOCK(s) \ 391 do { \ 392 s = splclock(); \ 393 simple_lock(&callwheel_slock); \ 394 } while (/*CONSTCOND*/ 0) 395 396 #define CALLWHEEL_UNLOCK(s) \ 397 do { \ 398 simple_unlock(&callwheel_slock); \ 399 splx(s); \ 400 } while (/*CONSTCOND*/ 0) 401 402 static void callout_stop_locked(struct callout *); 403 404 /* 405 * These are both protected by callwheel_lock. 406 * XXX SHOULD BE STATIC!! 407 */ 408 u_int64_t hardclock_ticks, softclock_ticks; 409 410 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 411 void softclock(void *); 412 void *softclock_si; 413 #endif 414 415 /* 416 * Initialize clock frequencies and start both clocks running. 417 */ 418 void 419 initclocks(void) 420 { 421 int i; 422 423 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 424 softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL); 425 if (softclock_si == NULL) 426 panic("initclocks: unable to register softclock intr"); 427 #endif 428 429 /* 430 * Set divisors to 1 (normal case) and let the machine-specific 431 * code do its bit. 432 */ 433 psdiv = 1; 434 cpu_initclocks(); 435 436 /* 437 * Compute profhz/stathz/rrticks, and fix profhz if needed. 438 */ 439 i = stathz ? stathz : hz; 440 if (profhz == 0) 441 profhz = i; 442 psratio = profhz / i; 443 rrticks = hz / 10; 444 445 #ifdef NTP 446 switch (hz) { 447 case 1: 448 shifthz = SHIFT_SCALE - 0; 449 break; 450 case 2: 451 shifthz = SHIFT_SCALE - 1; 452 break; 453 case 4: 454 shifthz = SHIFT_SCALE - 2; 455 break; 456 case 8: 457 shifthz = SHIFT_SCALE - 3; 458 break; 459 case 16: 460 shifthz = SHIFT_SCALE - 4; 461 break; 462 case 32: 463 shifthz = SHIFT_SCALE - 5; 464 break; 465 case 60: 466 case 64: 467 shifthz = SHIFT_SCALE - 6; 468 break; 469 case 96: 470 case 100: 471 case 128: 472 shifthz = SHIFT_SCALE - 7; 473 break; 474 case 256: 475 shifthz = SHIFT_SCALE - 8; 476 break; 477 case 512: 478 shifthz = SHIFT_SCALE - 9; 479 break; 480 case 1000: 481 case 1024: 482 shifthz = SHIFT_SCALE - 10; 483 break; 484 case 1200: 485 case 2048: 486 shifthz = SHIFT_SCALE - 11; 487 break; 488 case 4096: 489 shifthz = SHIFT_SCALE - 12; 490 break; 491 case 8192: 492 shifthz = SHIFT_SCALE - 13; 493 break; 494 case 16384: 495 shifthz = SHIFT_SCALE - 14; 496 break; 497 case 32768: 498 shifthz = SHIFT_SCALE - 15; 499 break; 500 case 65536: 501 shifthz = SHIFT_SCALE - 16; 502 break; 503 default: 504 panic("weird hz"); 505 } 506 if (fixtick == 0) { 507 /* 508 * Give MD code a chance to set this to a better 509 * value; but, if it doesn't, we should. 510 */ 511 fixtick = (1000000 - (hz*tick)); 512 } 513 #endif 514 } 515 516 /* 517 * The real-time timer, interrupting hz times per second. 518 */ 519 void 520 hardclock(struct clockframe *frame) 521 { 522 struct proc *p; 523 int delta; 524 extern int tickdelta; 525 extern long timedelta; 526 struct cpu_info *ci = curcpu(); 527 #ifdef NTP 528 int time_update; 529 int ltemp; 530 #endif 531 532 p = curproc; 533 if (p) { 534 struct pstats *pstats; 535 536 /* 537 * Run current process's virtual and profile time, as needed. 538 */ 539 pstats = p->p_stats; 540 if (CLKF_USERMODE(frame) && 541 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 542 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 543 psignal(p, SIGVTALRM); 544 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 545 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 546 psignal(p, SIGPROF); 547 } 548 549 /* 550 * If no separate statistics clock is available, run it from here. 551 */ 552 if (stathz == 0) 553 statclock(frame); 554 if ((--ci->ci_schedstate.spc_rrticks) <= 0) 555 roundrobin(ci); 556 557 #if defined(MULTIPROCESSOR) 558 /* 559 * If we are not the primary CPU, we're not allowed to do 560 * any more work. 561 */ 562 if (CPU_IS_PRIMARY(ci) == 0) 563 return; 564 #endif 565 566 /* 567 * Increment the time-of-day. The increment is normally just 568 * ``tick''. If the machine is one which has a clock frequency 569 * such that ``hz'' would not divide the second evenly into 570 * milliseconds, a periodic adjustment must be applied. Finally, 571 * if we are still adjusting the time (see adjtime()), 572 * ``tickdelta'' may also be added in. 573 */ 574 delta = tick; 575 576 #ifndef NTP 577 if (tickfix) { 578 tickfixcnt += tickfix; 579 if (tickfixcnt >= tickfixinterval) { 580 delta++; 581 tickfixcnt -= tickfixinterval; 582 } 583 } 584 #endif /* !NTP */ 585 /* Imprecise 4bsd adjtime() handling */ 586 if (timedelta != 0) { 587 delta += tickdelta; 588 timedelta -= tickdelta; 589 } 590 591 #ifdef notyet 592 microset(); 593 #endif 594 595 #ifndef NTP 596 BUMPTIME(&time, delta); /* XXX Now done using NTP code below */ 597 #endif 598 BUMPTIME(&mono_time, delta); 599 600 #ifdef NTP 601 time_update = delta; 602 603 /* 604 * Compute the phase adjustment. If the low-order bits 605 * (time_phase) of the update overflow, bump the high-order bits 606 * (time_update). 607 */ 608 time_phase += time_adj; 609 if (time_phase <= -FINEUSEC) { 610 ltemp = -time_phase >> SHIFT_SCALE; 611 time_phase += ltemp << SHIFT_SCALE; 612 time_update -= ltemp; 613 } else if (time_phase >= FINEUSEC) { 614 ltemp = time_phase >> SHIFT_SCALE; 615 time_phase -= ltemp << SHIFT_SCALE; 616 time_update += ltemp; 617 } 618 619 #ifdef HIGHBALL 620 /* 621 * If the HIGHBALL board is installed, we need to adjust the 622 * external clock offset in order to close the hardware feedback 623 * loop. This will adjust the external clock phase and frequency 624 * in small amounts. The additional phase noise and frequency 625 * wander this causes should be minimal. We also need to 626 * discipline the kernel time variable, since the PLL is used to 627 * discipline the external clock. If the Highball board is not 628 * present, we discipline kernel time with the PLL as usual. We 629 * assume that the external clock phase adjustment (time_update) 630 * and kernel phase adjustment (clock_cpu) are less than the 631 * value of tick. 632 */ 633 clock_offset.tv_usec += time_update; 634 if (clock_offset.tv_usec >= 1000000) { 635 clock_offset.tv_sec++; 636 clock_offset.tv_usec -= 1000000; 637 } 638 if (clock_offset.tv_usec < 0) { 639 clock_offset.tv_sec--; 640 clock_offset.tv_usec += 1000000; 641 } 642 time.tv_usec += clock_cpu; 643 clock_cpu = 0; 644 #else 645 time.tv_usec += time_update; 646 #endif /* HIGHBALL */ 647 648 /* 649 * On rollover of the second the phase adjustment to be used for 650 * the next second is calculated. Also, the maximum error is 651 * increased by the tolerance. If the PPS frequency discipline 652 * code is present, the phase is increased to compensate for the 653 * CPU clock oscillator frequency error. 654 * 655 * On a 32-bit machine and given parameters in the timex.h 656 * header file, the maximum phase adjustment is +-512 ms and 657 * maximum frequency offset is a tad less than) +-512 ppm. On a 658 * 64-bit machine, you shouldn't need to ask. 659 */ 660 if (time.tv_usec >= 1000000) { 661 time.tv_usec -= 1000000; 662 time.tv_sec++; 663 time_maxerror += time_tolerance >> SHIFT_USEC; 664 665 /* 666 * Leap second processing. If in leap-insert state at 667 * the end of the day, the system clock is set back one 668 * second; if in leap-delete state, the system clock is 669 * set ahead one second. The microtime() routine or 670 * external clock driver will insure that reported time 671 * is always monotonic. The ugly divides should be 672 * replaced. 673 */ 674 switch (time_state) { 675 case TIME_OK: 676 if (time_status & STA_INS) 677 time_state = TIME_INS; 678 else if (time_status & STA_DEL) 679 time_state = TIME_DEL; 680 break; 681 682 case TIME_INS: 683 if (time.tv_sec % 86400 == 0) { 684 time.tv_sec--; 685 time_state = TIME_OOP; 686 } 687 break; 688 689 case TIME_DEL: 690 if ((time.tv_sec + 1) % 86400 == 0) { 691 time.tv_sec++; 692 time_state = TIME_WAIT; 693 } 694 break; 695 696 case TIME_OOP: 697 time_state = TIME_WAIT; 698 break; 699 700 case TIME_WAIT: 701 if (!(time_status & (STA_INS | STA_DEL))) 702 time_state = TIME_OK; 703 break; 704 } 705 706 /* 707 * Compute the phase adjustment for the next second. In 708 * PLL mode, the offset is reduced by a fixed factor 709 * times the time constant. In FLL mode the offset is 710 * used directly. In either mode, the maximum phase 711 * adjustment for each second is clamped so as to spread 712 * the adjustment over not more than the number of 713 * seconds between updates. 714 */ 715 if (time_offset < 0) { 716 ltemp = -time_offset; 717 if (!(time_status & STA_FLL)) 718 ltemp >>= SHIFT_KG + time_constant; 719 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 720 ltemp = (MAXPHASE / MINSEC) << 721 SHIFT_UPDATE; 722 time_offset += ltemp; 723 time_adj = -ltemp << (shifthz - SHIFT_UPDATE); 724 } else if (time_offset > 0) { 725 ltemp = time_offset; 726 if (!(time_status & STA_FLL)) 727 ltemp >>= SHIFT_KG + time_constant; 728 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 729 ltemp = (MAXPHASE / MINSEC) << 730 SHIFT_UPDATE; 731 time_offset -= ltemp; 732 time_adj = ltemp << (shifthz - SHIFT_UPDATE); 733 } else 734 time_adj = 0; 735 736 /* 737 * Compute the frequency estimate and additional phase 738 * adjustment due to frequency error for the next 739 * second. When the PPS signal is engaged, gnaw on the 740 * watchdog counter and update the frequency computed by 741 * the pll and the PPS signal. 742 */ 743 #ifdef PPS_SYNC 744 pps_valid++; 745 if (pps_valid == PPS_VALID) { 746 pps_jitter = MAXTIME; 747 pps_stabil = MAXFREQ; 748 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 749 STA_PPSWANDER | STA_PPSERROR); 750 } 751 ltemp = time_freq + pps_freq; 752 #else 753 ltemp = time_freq; 754 #endif /* PPS_SYNC */ 755 756 if (ltemp < 0) 757 time_adj -= -ltemp >> (SHIFT_USEC - shifthz); 758 else 759 time_adj += ltemp >> (SHIFT_USEC - shifthz); 760 time_adj += (long)fixtick << shifthz; 761 762 /* 763 * When the CPU clock oscillator frequency is not a 764 * power of 2 in Hz, shifthz is only an approximate 765 * scale factor. 766 * 767 * To determine the adjustment, you can do the following: 768 * bc -q 769 * scale=24 770 * obase=2 771 * idealhz/realhz 772 * where `idealhz' is the next higher power of 2, and `realhz' 773 * is the actual value. You may need to factor this result 774 * into a sequence of 2 multipliers to get better precision. 775 * 776 * Likewise, the error can be calculated with (e.g. for 100Hz): 777 * bc -q 778 * scale=24 779 * ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz 780 * (and then multiply by 1000000 to get ppm). 781 */ 782 switch (hz) { 783 case 60: 784 /* A factor of 1.000100010001 gives about 15ppm 785 error. */ 786 if (time_adj < 0) { 787 time_adj -= (-time_adj >> 4); 788 time_adj -= (-time_adj >> 8); 789 } else { 790 time_adj += (time_adj >> 4); 791 time_adj += (time_adj >> 8); 792 } 793 break; 794 795 case 96: 796 /* A factor of 1.0101010101 gives about 244ppm error. */ 797 if (time_adj < 0) { 798 time_adj -= (-time_adj >> 2); 799 time_adj -= (-time_adj >> 4) + (-time_adj >> 8); 800 } else { 801 time_adj += (time_adj >> 2); 802 time_adj += (time_adj >> 4) + (time_adj >> 8); 803 } 804 break; 805 806 case 100: 807 /* A factor of 1.010001111010111 gives about 1ppm 808 error. */ 809 if (time_adj < 0) { 810 time_adj -= (-time_adj >> 2) + (-time_adj >> 5); 811 time_adj += (-time_adj >> 10); 812 } else { 813 time_adj += (time_adj >> 2) + (time_adj >> 5); 814 time_adj -= (time_adj >> 10); 815 } 816 break; 817 818 case 1000: 819 /* A factor of 1.000001100010100001 gives about 50ppm 820 error. */ 821 if (time_adj < 0) { 822 time_adj -= (-time_adj >> 6) + (-time_adj >> 11); 823 time_adj -= (-time_adj >> 7); 824 } else { 825 time_adj += (time_adj >> 6) + (time_adj >> 11); 826 time_adj += (time_adj >> 7); 827 } 828 break; 829 830 case 1200: 831 /* A factor of 1.1011010011100001 gives about 64ppm 832 error. */ 833 if (time_adj < 0) { 834 time_adj -= (-time_adj >> 1) + (-time_adj >> 6); 835 time_adj -= (-time_adj >> 3) + (-time_adj >> 10); 836 } else { 837 time_adj += (time_adj >> 1) + (time_adj >> 6); 838 time_adj += (time_adj >> 3) + (time_adj >> 10); 839 } 840 break; 841 } 842 843 #ifdef EXT_CLOCK 844 /* 845 * If an external clock is present, it is necessary to 846 * discipline the kernel time variable anyway, since not 847 * all system components use the microtime() interface. 848 * Here, the time offset between the external clock and 849 * kernel time variable is computed every so often. 850 */ 851 clock_count++; 852 if (clock_count > CLOCK_INTERVAL) { 853 clock_count = 0; 854 microtime(&clock_ext); 855 delta.tv_sec = clock_ext.tv_sec - time.tv_sec; 856 delta.tv_usec = clock_ext.tv_usec - 857 time.tv_usec; 858 if (delta.tv_usec < 0) 859 delta.tv_sec--; 860 if (delta.tv_usec >= 500000) { 861 delta.tv_usec -= 1000000; 862 delta.tv_sec++; 863 } 864 if (delta.tv_usec < -500000) { 865 delta.tv_usec += 1000000; 866 delta.tv_sec--; 867 } 868 if (delta.tv_sec > 0 || (delta.tv_sec == 0 && 869 delta.tv_usec > MAXPHASE) || 870 delta.tv_sec < -1 || (delta.tv_sec == -1 && 871 delta.tv_usec < -MAXPHASE)) { 872 time = clock_ext; 873 delta.tv_sec = 0; 874 delta.tv_usec = 0; 875 } 876 #ifdef HIGHBALL 877 clock_cpu = delta.tv_usec; 878 #else /* HIGHBALL */ 879 hardupdate(delta.tv_usec); 880 #endif /* HIGHBALL */ 881 } 882 #endif /* EXT_CLOCK */ 883 } 884 885 #endif /* NTP */ 886 887 /* 888 * Process callouts at a very low cpu priority, so we don't keep the 889 * relatively high clock interrupt priority any longer than necessary. 890 */ 891 simple_lock(&callwheel_slock); /* already at splclock() */ 892 hardclock_ticks++; 893 if (! TAILQ_EMPTY(&callwheel[hardclock_ticks & callwheelmask].cq_q)) { 894 simple_unlock(&callwheel_slock); 895 if (CLKF_BASEPRI(frame)) { 896 /* 897 * Save the overhead of a software interrupt; 898 * it will happen as soon as we return, so do 899 * it now. 900 * 901 * NOTE: If we're at ``base priority'', softclock() 902 * was not already running. 903 */ 904 spllowersoftclock(); 905 KERNEL_LOCK(LK_CANRECURSE|LK_EXCLUSIVE); 906 softclock(NULL); 907 KERNEL_UNLOCK(); 908 } else { 909 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 910 softintr_schedule(softclock_si); 911 #else 912 setsoftclock(); 913 #endif 914 } 915 return; 916 } else if (softclock_running == 0 && 917 (softclock_ticks + 1) == hardclock_ticks) { 918 softclock_ticks++; 919 } 920 simple_unlock(&callwheel_slock); 921 } 922 923 /* 924 * Software (low priority) clock interrupt. 925 * Run periodic events from timeout queue. 926 */ 927 /*ARGSUSED*/ 928 void 929 softclock(void *v) 930 { 931 struct callout_queue *bucket; 932 struct callout *c; 933 void (*func)(void *); 934 void *arg; 935 int s, idx; 936 int steps = 0; 937 938 CALLWHEEL_LOCK(s); 939 940 softclock_running = 1; 941 942 #ifdef CALLWHEEL_STATS 943 callwheel_softclocks.ev_count++; 944 #endif 945 946 while (softclock_ticks != hardclock_ticks) { 947 softclock_ticks++; 948 idx = (int)(softclock_ticks & callwheelmask); 949 bucket = &callwheel[idx]; 950 c = TAILQ_FIRST(&bucket->cq_q); 951 if (c == NULL) { 952 #ifdef CALLWHEEL_STATS 953 callwheel_softempty.ev_count++; 954 #endif 955 continue; 956 } 957 if (softclock_ticks < bucket->cq_hint) { 958 #ifdef CALLWHEEL_STATS 959 callwheel_hintworked.ev_count++; 960 #endif 961 continue; 962 } 963 bucket->cq_hint = UQUAD_MAX; 964 while (c != NULL) { 965 #ifdef CALLWHEEL_STATS 966 callwheel_softchecks.ev_count++; 967 #endif 968 if (c->c_time != softclock_ticks) { 969 if (c->c_time < bucket->cq_hint) 970 bucket->cq_hint = c->c_time; 971 c = TAILQ_NEXT(c, c_link); 972 if (++steps >= MAX_SOFTCLOCK_STEPS) { 973 nextsoftcheck = c; 974 /* Give interrupts a chance. */ 975 CALLWHEEL_UNLOCK(s); 976 CALLWHEEL_LOCK(s); 977 c = nextsoftcheck; 978 steps = 0; 979 } 980 } else { 981 nextsoftcheck = TAILQ_NEXT(c, c_link); 982 TAILQ_REMOVE(&bucket->cq_q, c, c_link); 983 #ifdef CALLWHEEL_STATS 984 callwheel_sizes[idx]--; 985 callwheel_fired.ev_count++; 986 callwheel_count.ev_count--; 987 #endif 988 func = c->c_func; 989 arg = c->c_arg; 990 c->c_func = NULL; 991 c->c_flags &= ~CALLOUT_PENDING; 992 CALLWHEEL_UNLOCK(s); 993 (*func)(arg); 994 CALLWHEEL_LOCK(s); 995 steps = 0; 996 c = nextsoftcheck; 997 } 998 } 999 if (TAILQ_EMPTY(&bucket->cq_q)) 1000 bucket->cq_hint = UQUAD_MAX; 1001 } 1002 nextsoftcheck = NULL; 1003 softclock_running = 0; 1004 CALLWHEEL_UNLOCK(s); 1005 } 1006 1007 /* 1008 * callout_setsize: 1009 * 1010 * Determine how many callwheels are necessary and 1011 * set hash mask. Called from allocsys(). 1012 */ 1013 void 1014 callout_setsize(void) 1015 { 1016 1017 for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1) 1018 /* loop */ ; 1019 callwheelmask = callwheelsize - 1; 1020 } 1021 1022 /* 1023 * callout_startup: 1024 * 1025 * Initialize the callwheel buckets. 1026 */ 1027 void 1028 callout_startup(void) 1029 { 1030 int i; 1031 1032 for (i = 0; i < callwheelsize; i++) { 1033 callwheel[i].cq_hint = UQUAD_MAX; 1034 TAILQ_INIT(&callwheel[i].cq_q); 1035 } 1036 1037 simple_lock_init(&callwheel_slock); 1038 1039 #ifdef CALLWHEEL_STATS 1040 evcnt_attach_dynamic(&callwheel_collisions, EVCNT_TYPE_MISC, 1041 NULL, "callwheel", "collisions"); 1042 evcnt_attach_dynamic(&callwheel_maxlength, EVCNT_TYPE_MISC, 1043 NULL, "callwheel", "maxlength"); 1044 evcnt_attach_dynamic(&callwheel_count, EVCNT_TYPE_MISC, 1045 NULL, "callwheel", "count"); 1046 evcnt_attach_dynamic(&callwheel_established, EVCNT_TYPE_MISC, 1047 NULL, "callwheel", "established"); 1048 evcnt_attach_dynamic(&callwheel_fired, EVCNT_TYPE_MISC, 1049 NULL, "callwheel", "fired"); 1050 evcnt_attach_dynamic(&callwheel_disestablished, EVCNT_TYPE_MISC, 1051 NULL, "callwheel", "disestablished"); 1052 evcnt_attach_dynamic(&callwheel_changed, EVCNT_TYPE_MISC, 1053 NULL, "callwheel", "changed"); 1054 evcnt_attach_dynamic(&callwheel_softclocks, EVCNT_TYPE_MISC, 1055 NULL, "callwheel", "softclocks"); 1056 evcnt_attach_dynamic(&callwheel_softempty, EVCNT_TYPE_MISC, 1057 NULL, "callwheel", "softempty"); 1058 evcnt_attach_dynamic(&callwheel_hintworked, EVCNT_TYPE_MISC, 1059 NULL, "callwheel", "hintworked"); 1060 #endif /* CALLWHEEL_STATS */ 1061 } 1062 1063 /* 1064 * callout_init: 1065 * 1066 * Initialize a callout structure so that it can be used 1067 * by callout_reset() and callout_stop(). 1068 */ 1069 void 1070 callout_init(struct callout *c) 1071 { 1072 1073 memset(c, 0, sizeof(*c)); 1074 } 1075 1076 /* 1077 * callout_reset: 1078 * 1079 * Establish or change a timeout. 1080 */ 1081 void 1082 callout_reset(struct callout *c, int ticks, void (*func)(void *), void *arg) 1083 { 1084 struct callout_queue *bucket; 1085 int s; 1086 1087 if (ticks <= 0) 1088 ticks = 1; 1089 1090 CALLWHEEL_LOCK(s); 1091 1092 /* 1093 * If this callout's timer is already running, cancel it 1094 * before we modify it. 1095 */ 1096 if (c->c_flags & CALLOUT_PENDING) { 1097 callout_stop_locked(c); /* Already locked */ 1098 #ifdef CALLWHEEL_STATS 1099 callwheel_changed.ev_count++; 1100 #endif 1101 } 1102 1103 c->c_arg = arg; 1104 c->c_func = func; 1105 c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING; 1106 c->c_time = hardclock_ticks + ticks; 1107 1108 bucket = &callwheel[c->c_time & callwheelmask]; 1109 1110 #ifdef CALLWHEEL_STATS 1111 if (! TAILQ_EMPTY(&bucket->cq_q)) 1112 callwheel_collisions.ev_count++; 1113 #endif 1114 1115 TAILQ_INSERT_TAIL(&bucket->cq_q, c, c_link); 1116 if (c->c_time < bucket->cq_hint) 1117 bucket->cq_hint = c->c_time; 1118 1119 #ifdef CALLWHEEL_STATS 1120 callwheel_count.ev_count++; 1121 callwheel_established.ev_count++; 1122 if (++callwheel_sizes[c->c_time & callwheelmask] > 1123 callwheel_maxlength.ev_count) 1124 callwheel_maxlength.ev_count = 1125 callwheel_sizes[c->c_time & callwheelmask]; 1126 #endif 1127 1128 CALLWHEEL_UNLOCK(s); 1129 } 1130 1131 /* 1132 * callout_stop_locked: 1133 * 1134 * Disestablish a timeout. Callwheel is locked. 1135 */ 1136 static void 1137 callout_stop_locked(struct callout *c) 1138 { 1139 struct callout_queue *bucket; 1140 1141 /* 1142 * Don't attempt to delete a callout that's not on the queue. 1143 */ 1144 if ((c->c_flags & CALLOUT_PENDING) == 0) { 1145 c->c_flags &= ~CALLOUT_ACTIVE; 1146 return; 1147 } 1148 1149 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 1150 1151 if (nextsoftcheck == c) 1152 nextsoftcheck = TAILQ_NEXT(c, c_link); 1153 1154 bucket = &callwheel[c->c_time & callwheelmask]; 1155 TAILQ_REMOVE(&bucket->cq_q, c, c_link); 1156 if (TAILQ_EMPTY(&bucket->cq_q)) 1157 bucket->cq_hint = UQUAD_MAX; 1158 #ifdef CALLWHEEL_STATS 1159 callwheel_count.ev_count--; 1160 callwheel_disestablished.ev_count++; 1161 callwheel_sizes[c->c_time & callwheelmask]--; 1162 #endif 1163 1164 c->c_func = NULL; 1165 } 1166 1167 /* 1168 * callout_stop: 1169 * 1170 * Disestablish a timeout. Callwheel is unlocked. This is 1171 * the standard entry point. 1172 */ 1173 void 1174 callout_stop(struct callout *c) 1175 { 1176 int s; 1177 1178 CALLWHEEL_LOCK(s); 1179 callout_stop_locked(c); 1180 CALLWHEEL_UNLOCK(s); 1181 } 1182 1183 #ifdef CALLWHEEL_STATS 1184 /* 1185 * callout_showstats: 1186 * 1187 * Display callout statistics. Call it from DDB. 1188 */ 1189 void 1190 callout_showstats(void) 1191 { 1192 u_int64_t curticks; 1193 int s; 1194 1195 s = splclock(); 1196 curticks = softclock_ticks; 1197 splx(s); 1198 1199 printf("Callwheel statistics:\n"); 1200 printf("\tCallouts currently queued: %llu\n", 1201 (long long) callwheel_count.ev_count); 1202 printf("\tCallouts established: %llu\n", 1203 (long long) callwheel_established.ev_count); 1204 printf("\tCallouts disestablished: %llu\n", 1205 (long long) callwheel_disestablished.ev_count); 1206 if (callwheel_changed.ev_count != 0) 1207 printf("\t\tOf those, %llu were changes\n", 1208 (long long) callwheel_changed.ev_count); 1209 printf("\tCallouts that fired: %llu\n", 1210 (long long) callwheel_fired.ev_count); 1211 printf("\tNumber of buckets: %d\n", callwheelsize); 1212 printf("\tNumber of hash collisions: %llu\n", 1213 (long long) callwheel_collisions.ev_count); 1214 printf("\tMaximum hash chain length: %llu\n", 1215 (long long) callwheel_maxlength.ev_count); 1216 printf("\tSoftclocks: %llu, Softchecks: %llu\n", 1217 (long long) callwheel_softclocks.ev_count, 1218 (long long) callwheel_softchecks.ev_count); 1219 printf("\t\tEmpty buckets seen: %llu\n", 1220 (long long) callwheel_softempty.ev_count); 1221 printf("\t\tTimes hint saved scan: %llu\n", 1222 (long long) callwheel_hintworked.ev_count); 1223 } 1224 #endif 1225 1226 /* 1227 * Compute number of hz until specified time. Used to compute second 1228 * argument to callout_reset() from an absolute time. 1229 */ 1230 int 1231 hzto(struct timeval *tv) 1232 { 1233 unsigned long ticks; 1234 long sec, usec; 1235 int s; 1236 1237 /* 1238 * If the number of usecs in the whole seconds part of the time 1239 * difference fits in a long, then the total number of usecs will 1240 * fit in an unsigned long. Compute the total and convert it to 1241 * ticks, rounding up and adding 1 to allow for the current tick 1242 * to expire. Rounding also depends on unsigned long arithmetic 1243 * to avoid overflow. 1244 * 1245 * Otherwise, if the number of ticks in the whole seconds part of 1246 * the time difference fits in a long, then convert the parts to 1247 * ticks separately and add, using similar rounding methods and 1248 * overflow avoidance. This method would work in the previous 1249 * case, but it is slightly slower and assume that hz is integral. 1250 * 1251 * Otherwise, round the time difference down to the maximum 1252 * representable value. 1253 * 1254 * If ints are 32-bit, then the maximum value for any timeout in 1255 * 10ms ticks is 248 days. 1256 */ 1257 s = splclock(); 1258 sec = tv->tv_sec - time.tv_sec; 1259 usec = tv->tv_usec - time.tv_usec; 1260 splx(s); 1261 1262 if (usec < 0) { 1263 sec--; 1264 usec += 1000000; 1265 } 1266 1267 if (sec < 0 || (sec == 0 && usec <= 0)) { 1268 /* 1269 * Would expire now or in the past. Return 0 ticks. 1270 * This is different from the legacy hzto() interface, 1271 * and callers need to check for it. 1272 */ 1273 ticks = 0; 1274 } else if (sec <= (LONG_MAX / 1000000)) 1275 ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1)) 1276 / tick) + 1; 1277 else if (sec <= (LONG_MAX / hz)) 1278 ticks = (sec * hz) + 1279 (((unsigned long)usec + (tick - 1)) / tick) + 1; 1280 else 1281 ticks = LONG_MAX; 1282 1283 if (ticks > INT_MAX) 1284 ticks = INT_MAX; 1285 1286 return ((int)ticks); 1287 } 1288 1289 /* 1290 * Start profiling on a process. 1291 * 1292 * Kernel profiling passes proc0 which never exits and hence 1293 * keeps the profile clock running constantly. 1294 */ 1295 void 1296 startprofclock(struct proc *p) 1297 { 1298 1299 if ((p->p_flag & P_PROFIL) == 0) { 1300 p->p_flag |= P_PROFIL; 1301 /* 1302 * This is only necessary if using the clock as the 1303 * profiling source. 1304 */ 1305 if (++profprocs == 1 && stathz != 0) 1306 psdiv = psratio; 1307 } 1308 } 1309 1310 /* 1311 * Stop profiling on a process. 1312 */ 1313 void 1314 stopprofclock(struct proc *p) 1315 { 1316 1317 if (p->p_flag & P_PROFIL) { 1318 p->p_flag &= ~P_PROFIL; 1319 /* 1320 * This is only necessary if using the clock as the 1321 * profiling source. 1322 */ 1323 if (--profprocs == 0 && stathz != 0) 1324 psdiv = 1; 1325 } 1326 } 1327 1328 #if defined(PERFCTRS) 1329 /* 1330 * Independent profiling "tick" in case we're using a separate 1331 * clock or profiling event source. Currently, that's just 1332 * performance counters--hence the wrapper. 1333 */ 1334 void 1335 proftick(struct clockframe *frame) 1336 { 1337 #ifdef GPROF 1338 struct gmonparam *g; 1339 intptr_t i; 1340 #endif 1341 struct proc *p; 1342 1343 p = curproc; 1344 if (CLKF_USERMODE(frame)) { 1345 if (p->p_flag & P_PROFIL) 1346 addupc_intr(p, CLKF_PC(frame)); 1347 } else { 1348 #ifdef GPROF 1349 g = &_gmonparam; 1350 if (g->state == GMON_PROF_ON) { 1351 i = CLKF_PC(frame) - g->lowpc; 1352 if (i < g->textsize) { 1353 i /= HISTFRACTION * sizeof(*g->kcount); 1354 g->kcount[i]++; 1355 } 1356 } 1357 #endif 1358 #ifdef PROC_PC 1359 if (p && p->p_flag & P_PROFIL) 1360 addupc_intr(p, PROC_PC(p)); 1361 #endif 1362 } 1363 } 1364 #endif 1365 1366 /* 1367 * Statistics clock. Grab profile sample, and if divider reaches 0, 1368 * do process and kernel statistics. 1369 */ 1370 void 1371 statclock(struct clockframe *frame) 1372 { 1373 #ifdef GPROF 1374 struct gmonparam *g; 1375 intptr_t i; 1376 #endif 1377 struct cpu_info *ci = curcpu(); 1378 struct schedstate_percpu *spc = &ci->ci_schedstate; 1379 struct proc *p; 1380 1381 /* 1382 * Notice changes in divisor frequency, and adjust clock 1383 * frequency accordingly. 1384 */ 1385 if (spc->spc_psdiv != psdiv) { 1386 spc->spc_psdiv = psdiv; 1387 spc->spc_pscnt = psdiv; 1388 if (psdiv == 1) { 1389 setstatclockrate(stathz); 1390 } else { 1391 setstatclockrate(profhz); 1392 } 1393 } 1394 p = curproc; 1395 if (CLKF_USERMODE(frame)) { 1396 if (p->p_flag & P_PROFIL && profsrc == PROFSRC_CLOCK) 1397 addupc_intr(p, CLKF_PC(frame)); 1398 if (--spc->spc_pscnt > 0) 1399 return; 1400 /* 1401 * Came from user mode; CPU was in user state. 1402 * If this process is being profiled record the tick. 1403 */ 1404 p->p_uticks++; 1405 if (p->p_nice > NZERO) 1406 spc->spc_cp_time[CP_NICE]++; 1407 else 1408 spc->spc_cp_time[CP_USER]++; 1409 } else { 1410 #ifdef GPROF 1411 /* 1412 * Kernel statistics are just like addupc_intr, only easier. 1413 */ 1414 g = &_gmonparam; 1415 if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) { 1416 i = CLKF_PC(frame) - g->lowpc; 1417 if (i < g->textsize) { 1418 i /= HISTFRACTION * sizeof(*g->kcount); 1419 g->kcount[i]++; 1420 } 1421 } 1422 #endif 1423 #ifdef PROC_PC 1424 if (p && profsrc == PROFSRC_CLOCK && p->p_flag & P_PROFIL) 1425 addupc_intr(p, PROC_PC(p)); 1426 #endif 1427 if (--spc->spc_pscnt > 0) 1428 return; 1429 /* 1430 * Came from kernel mode, so we were: 1431 * - handling an interrupt, 1432 * - doing syscall or trap work on behalf of the current 1433 * user process, or 1434 * - spinning in the idle loop. 1435 * Whichever it is, charge the time as appropriate. 1436 * Note that we charge interrupts to the current process, 1437 * regardless of whether they are ``for'' that process, 1438 * so that we know how much of its real time was spent 1439 * in ``non-process'' (i.e., interrupt) work. 1440 */ 1441 if (CLKF_INTR(frame)) { 1442 if (p != NULL) 1443 p->p_iticks++; 1444 spc->spc_cp_time[CP_INTR]++; 1445 } else if (p != NULL) { 1446 p->p_sticks++; 1447 spc->spc_cp_time[CP_SYS]++; 1448 } else 1449 spc->spc_cp_time[CP_IDLE]++; 1450 } 1451 spc->spc_pscnt = psdiv; 1452 1453 if (p != NULL) { 1454 ++p->p_cpticks; 1455 /* 1456 * If no separate schedclock is provided, call it here 1457 * at ~~12-25 Hz, ~~16 Hz is best 1458 */ 1459 if (schedhz == 0) 1460 if ((++ci->ci_schedstate.spc_schedticks & 3) == 0) 1461 schedclock(p); 1462 } 1463 } 1464 1465 1466 #ifdef NTP /* NTP phase-locked loop in kernel */ 1467 1468 /* 1469 * hardupdate() - local clock update 1470 * 1471 * This routine is called by ntp_adjtime() to update the local clock 1472 * phase and frequency. The implementation is of an adaptive-parameter, 1473 * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new 1474 * time and frequency offset estimates for each call. If the kernel PPS 1475 * discipline code is configured (PPS_SYNC), the PPS signal itself 1476 * determines the new time offset, instead of the calling argument. 1477 * Presumably, calls to ntp_adjtime() occur only when the caller 1478 * believes the local clock is valid within some bound (+-128 ms with 1479 * NTP). If the caller's time is far different than the PPS time, an 1480 * argument will ensue, and it's not clear who will lose. 1481 * 1482 * For uncompensated quartz crystal oscillatores and nominal update 1483 * intervals less than 1024 s, operation should be in phase-lock mode 1484 * (STA_FLL = 0), where the loop is disciplined to phase. For update 1485 * intervals greater than thiss, operation should be in frequency-lock 1486 * mode (STA_FLL = 1), where the loop is disciplined to frequency. 1487 * 1488 * Note: splclock() is in effect. 1489 */ 1490 void 1491 hardupdate(long offset) 1492 { 1493 long ltemp, mtemp; 1494 1495 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME)) 1496 return; 1497 ltemp = offset; 1498 #ifdef PPS_SYNC 1499 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL) 1500 ltemp = pps_offset; 1501 #endif /* PPS_SYNC */ 1502 1503 /* 1504 * Scale the phase adjustment and clamp to the operating range. 1505 */ 1506 if (ltemp > MAXPHASE) 1507 time_offset = MAXPHASE << SHIFT_UPDATE; 1508 else if (ltemp < -MAXPHASE) 1509 time_offset = -(MAXPHASE << SHIFT_UPDATE); 1510 else 1511 time_offset = ltemp << SHIFT_UPDATE; 1512 1513 /* 1514 * Select whether the frequency is to be controlled and in which 1515 * mode (PLL or FLL). Clamp to the operating range. Ugly 1516 * multiply/divide should be replaced someday. 1517 */ 1518 if (time_status & STA_FREQHOLD || time_reftime == 0) 1519 time_reftime = time.tv_sec; 1520 mtemp = time.tv_sec - time_reftime; 1521 time_reftime = time.tv_sec; 1522 if (time_status & STA_FLL) { 1523 if (mtemp >= MINSEC) { 1524 ltemp = ((time_offset / mtemp) << (SHIFT_USEC - 1525 SHIFT_UPDATE)); 1526 if (ltemp < 0) 1527 time_freq -= -ltemp >> SHIFT_KH; 1528 else 1529 time_freq += ltemp >> SHIFT_KH; 1530 } 1531 } else { 1532 if (mtemp < MAXSEC) { 1533 ltemp *= mtemp; 1534 if (ltemp < 0) 1535 time_freq -= -ltemp >> (time_constant + 1536 time_constant + SHIFT_KF - 1537 SHIFT_USEC); 1538 else 1539 time_freq += ltemp >> (time_constant + 1540 time_constant + SHIFT_KF - 1541 SHIFT_USEC); 1542 } 1543 } 1544 if (time_freq > time_tolerance) 1545 time_freq = time_tolerance; 1546 else if (time_freq < -time_tolerance) 1547 time_freq = -time_tolerance; 1548 } 1549 1550 #ifdef PPS_SYNC 1551 /* 1552 * hardpps() - discipline CPU clock oscillator to external PPS signal 1553 * 1554 * This routine is called at each PPS interrupt in order to discipline 1555 * the CPU clock oscillator to the PPS signal. It measures the PPS phase 1556 * and leaves it in a handy spot for the hardclock() routine. It 1557 * integrates successive PPS phase differences and calculates the 1558 * frequency offset. This is used in hardclock() to discipline the CPU 1559 * clock oscillator so that intrinsic frequency error is cancelled out. 1560 * The code requires the caller to capture the time and hardware counter 1561 * value at the on-time PPS signal transition. 1562 * 1563 * Note that, on some Unix systems, this routine runs at an interrupt 1564 * priority level higher than the timer interrupt routine hardclock(). 1565 * Therefore, the variables used are distinct from the hardclock() 1566 * variables, except for certain exceptions: The PPS frequency pps_freq 1567 * and phase pps_offset variables are determined by this routine and 1568 * updated atomically. The time_tolerance variable can be considered a 1569 * constant, since it is infrequently changed, and then only when the 1570 * PPS signal is disabled. The watchdog counter pps_valid is updated 1571 * once per second by hardclock() and is atomically cleared in this 1572 * routine. 1573 */ 1574 void 1575 hardpps(struct timeval *tvp, /* time at PPS */ 1576 long usec /* hardware counter at PPS */) 1577 { 1578 long u_usec, v_usec, bigtick; 1579 long cal_sec, cal_usec; 1580 1581 /* 1582 * An occasional glitch can be produced when the PPS interrupt 1583 * occurs in the hardclock() routine before the time variable is 1584 * updated. Here the offset is discarded when the difference 1585 * between it and the last one is greater than tick/2, but not 1586 * if the interval since the first discard exceeds 30 s. 1587 */ 1588 time_status |= STA_PPSSIGNAL; 1589 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR); 1590 pps_valid = 0; 1591 u_usec = -tvp->tv_usec; 1592 if (u_usec < -500000) 1593 u_usec += 1000000; 1594 v_usec = pps_offset - u_usec; 1595 if (v_usec < 0) 1596 v_usec = -v_usec; 1597 if (v_usec > (tick >> 1)) { 1598 if (pps_glitch > MAXGLITCH) { 1599 pps_glitch = 0; 1600 pps_tf[2] = u_usec; 1601 pps_tf[1] = u_usec; 1602 } else { 1603 pps_glitch++; 1604 u_usec = pps_offset; 1605 } 1606 } else 1607 pps_glitch = 0; 1608 1609 /* 1610 * A three-stage median filter is used to help deglitch the pps 1611 * time. The median sample becomes the time offset estimate; the 1612 * difference between the other two samples becomes the time 1613 * dispersion (jitter) estimate. 1614 */ 1615 pps_tf[2] = pps_tf[1]; 1616 pps_tf[1] = pps_tf[0]; 1617 pps_tf[0] = u_usec; 1618 if (pps_tf[0] > pps_tf[1]) { 1619 if (pps_tf[1] > pps_tf[2]) { 1620 pps_offset = pps_tf[1]; /* 0 1 2 */ 1621 v_usec = pps_tf[0] - pps_tf[2]; 1622 } else if (pps_tf[2] > pps_tf[0]) { 1623 pps_offset = pps_tf[0]; /* 2 0 1 */ 1624 v_usec = pps_tf[2] - pps_tf[1]; 1625 } else { 1626 pps_offset = pps_tf[2]; /* 0 2 1 */ 1627 v_usec = pps_tf[0] - pps_tf[1]; 1628 } 1629 } else { 1630 if (pps_tf[1] < pps_tf[2]) { 1631 pps_offset = pps_tf[1]; /* 2 1 0 */ 1632 v_usec = pps_tf[2] - pps_tf[0]; 1633 } else if (pps_tf[2] < pps_tf[0]) { 1634 pps_offset = pps_tf[0]; /* 1 0 2 */ 1635 v_usec = pps_tf[1] - pps_tf[2]; 1636 } else { 1637 pps_offset = pps_tf[2]; /* 1 2 0 */ 1638 v_usec = pps_tf[1] - pps_tf[0]; 1639 } 1640 } 1641 if (v_usec > MAXTIME) 1642 pps_jitcnt++; 1643 v_usec = (v_usec << PPS_AVG) - pps_jitter; 1644 if (v_usec < 0) 1645 pps_jitter -= -v_usec >> PPS_AVG; 1646 else 1647 pps_jitter += v_usec >> PPS_AVG; 1648 if (pps_jitter > (MAXTIME >> 1)) 1649 time_status |= STA_PPSJITTER; 1650 1651 /* 1652 * During the calibration interval adjust the starting time when 1653 * the tick overflows. At the end of the interval compute the 1654 * duration of the interval and the difference of the hardware 1655 * counters at the beginning and end of the interval. This code 1656 * is deliciously complicated by the fact valid differences may 1657 * exceed the value of tick when using long calibration 1658 * intervals and small ticks. Note that the counter can be 1659 * greater than tick if caught at just the wrong instant, but 1660 * the values returned and used here are correct. 1661 */ 1662 bigtick = (long)tick << SHIFT_USEC; 1663 pps_usec -= pps_freq; 1664 if (pps_usec >= bigtick) 1665 pps_usec -= bigtick; 1666 if (pps_usec < 0) 1667 pps_usec += bigtick; 1668 pps_time.tv_sec++; 1669 pps_count++; 1670 if (pps_count < (1 << pps_shift)) 1671 return; 1672 pps_count = 0; 1673 pps_calcnt++; 1674 u_usec = usec << SHIFT_USEC; 1675 v_usec = pps_usec - u_usec; 1676 if (v_usec >= bigtick >> 1) 1677 v_usec -= bigtick; 1678 if (v_usec < -(bigtick >> 1)) 1679 v_usec += bigtick; 1680 if (v_usec < 0) 1681 v_usec = -(-v_usec >> pps_shift); 1682 else 1683 v_usec = v_usec >> pps_shift; 1684 pps_usec = u_usec; 1685 cal_sec = tvp->tv_sec; 1686 cal_usec = tvp->tv_usec; 1687 cal_sec -= pps_time.tv_sec; 1688 cal_usec -= pps_time.tv_usec; 1689 if (cal_usec < 0) { 1690 cal_usec += 1000000; 1691 cal_sec--; 1692 } 1693 pps_time = *tvp; 1694 1695 /* 1696 * Check for lost interrupts, noise, excessive jitter and 1697 * excessive frequency error. The number of timer ticks during 1698 * the interval may vary +-1 tick. Add to this a margin of one 1699 * tick for the PPS signal jitter and maximum frequency 1700 * deviation. If the limits are exceeded, the calibration 1701 * interval is reset to the minimum and we start over. 1702 */ 1703 u_usec = (long)tick << 1; 1704 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1705 || (cal_sec == 0 && cal_usec < u_usec)) 1706 || v_usec > time_tolerance || v_usec < -time_tolerance) { 1707 pps_errcnt++; 1708 pps_shift = PPS_SHIFT; 1709 pps_intcnt = 0; 1710 time_status |= STA_PPSERROR; 1711 return; 1712 } 1713 1714 /* 1715 * A three-stage median filter is used to help deglitch the pps 1716 * frequency. The median sample becomes the frequency offset 1717 * estimate; the difference between the other two samples 1718 * becomes the frequency dispersion (stability) estimate. 1719 */ 1720 pps_ff[2] = pps_ff[1]; 1721 pps_ff[1] = pps_ff[0]; 1722 pps_ff[0] = v_usec; 1723 if (pps_ff[0] > pps_ff[1]) { 1724 if (pps_ff[1] > pps_ff[2]) { 1725 u_usec = pps_ff[1]; /* 0 1 2 */ 1726 v_usec = pps_ff[0] - pps_ff[2]; 1727 } else if (pps_ff[2] > pps_ff[0]) { 1728 u_usec = pps_ff[0]; /* 2 0 1 */ 1729 v_usec = pps_ff[2] - pps_ff[1]; 1730 } else { 1731 u_usec = pps_ff[2]; /* 0 2 1 */ 1732 v_usec = pps_ff[0] - pps_ff[1]; 1733 } 1734 } else { 1735 if (pps_ff[1] < pps_ff[2]) { 1736 u_usec = pps_ff[1]; /* 2 1 0 */ 1737 v_usec = pps_ff[2] - pps_ff[0]; 1738 } else if (pps_ff[2] < pps_ff[0]) { 1739 u_usec = pps_ff[0]; /* 1 0 2 */ 1740 v_usec = pps_ff[1] - pps_ff[2]; 1741 } else { 1742 u_usec = pps_ff[2]; /* 1 2 0 */ 1743 v_usec = pps_ff[1] - pps_ff[0]; 1744 } 1745 } 1746 1747 /* 1748 * Here the frequency dispersion (stability) is updated. If it 1749 * is less than one-fourth the maximum (MAXFREQ), the frequency 1750 * offset is updated as well, but clamped to the tolerance. It 1751 * will be processed later by the hardclock() routine. 1752 */ 1753 v_usec = (v_usec >> 1) - pps_stabil; 1754 if (v_usec < 0) 1755 pps_stabil -= -v_usec >> PPS_AVG; 1756 else 1757 pps_stabil += v_usec >> PPS_AVG; 1758 if (pps_stabil > MAXFREQ >> 2) { 1759 pps_stbcnt++; 1760 time_status |= STA_PPSWANDER; 1761 return; 1762 } 1763 if (time_status & STA_PPSFREQ) { 1764 if (u_usec < 0) { 1765 pps_freq -= -u_usec >> PPS_AVG; 1766 if (pps_freq < -time_tolerance) 1767 pps_freq = -time_tolerance; 1768 u_usec = -u_usec; 1769 } else { 1770 pps_freq += u_usec >> PPS_AVG; 1771 if (pps_freq > time_tolerance) 1772 pps_freq = time_tolerance; 1773 } 1774 } 1775 1776 /* 1777 * Here the calibration interval is adjusted. If the maximum 1778 * time difference is greater than tick / 4, reduce the interval 1779 * by half. If this is not the case for four consecutive 1780 * intervals, double the interval. 1781 */ 1782 if (u_usec << pps_shift > bigtick >> 2) { 1783 pps_intcnt = 0; 1784 if (pps_shift > PPS_SHIFT) 1785 pps_shift--; 1786 } else if (pps_intcnt >= 4) { 1787 pps_intcnt = 0; 1788 if (pps_shift < PPS_SHIFTMAX) 1789 pps_shift++; 1790 } else 1791 pps_intcnt++; 1792 } 1793 #endif /* PPS_SYNC */ 1794 #endif /* NTP */ 1795 1796 /* 1797 * Return information about system clocks. 1798 */ 1799 int 1800 sysctl_clockrate(void *where, size_t *sizep) 1801 { 1802 struct clockinfo clkinfo; 1803 1804 /* 1805 * Construct clockinfo structure. 1806 */ 1807 clkinfo.tick = tick; 1808 clkinfo.tickadj = tickadj; 1809 clkinfo.hz = hz; 1810 clkinfo.profhz = profhz; 1811 clkinfo.stathz = stathz ? stathz : hz; 1812 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo))); 1813 } 1814