1 /* $NetBSD: kern_clock.c,v 1.79 2002/03/17 11:10:43 simonb 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 1.79 2002/03/17 11:10:43 simonb Exp $"); 82 83 #include "opt_callout.h" 84 #include "opt_ntp.h" 85 86 #include <sys/param.h> 87 #include <sys/systm.h> 88 #include <sys/dkstat.h> 89 #include <sys/callout.h> 90 #include <sys/kernel.h> 91 #include <sys/proc.h> 92 #include <sys/resourcevar.h> 93 #include <sys/signalvar.h> 94 #include <sys/sysctl.h> 95 #include <sys/timex.h> 96 #include <sys/sched.h> 97 #ifdef CALLWHEEL_STATS 98 #include <sys/device.h> 99 #endif 100 101 #include <machine/cpu.h> 102 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 103 #include <machine/intr.h> 104 #endif 105 106 #ifdef GPROF 107 #include <sys/gmon.h> 108 #endif 109 110 /* 111 * Clock handling routines. 112 * 113 * This code is written to operate with two timers that run independently of 114 * each other. The main clock, running hz times per second, is used to keep 115 * track of real time. The second timer handles kernel and user profiling, 116 * and does resource use estimation. If the second timer is programmable, 117 * it is randomized to avoid aliasing between the two clocks. For example, 118 * the randomization prevents an adversary from always giving up the cpu 119 * just before its quantum expires. Otherwise, it would never accumulate 120 * cpu ticks. The mean frequency of the second timer is stathz. 121 * 122 * If no second timer exists, stathz will be zero; in this case we drive 123 * profiling and statistics off the main clock. This WILL NOT be accurate; 124 * do not do it unless absolutely necessary. 125 * 126 * The statistics clock may (or may not) be run at a higher rate while 127 * profiling. This profile clock runs at profhz. We require that profhz 128 * be an integral multiple of stathz. 129 * 130 * If the statistics clock is running fast, it must be divided by the ratio 131 * profhz/stathz for statistics. (For profiling, every tick counts.) 132 */ 133 134 #ifdef NTP /* NTP phase-locked loop in kernel */ 135 /* 136 * Phase/frequency-lock loop (PLL/FLL) definitions 137 * 138 * The following variables are read and set by the ntp_adjtime() system 139 * call. 140 * 141 * time_state shows the state of the system clock, with values defined 142 * in the timex.h header file. 143 * 144 * time_status shows the status of the system clock, with bits defined 145 * in the timex.h header file. 146 * 147 * time_offset is used by the PLL/FLL to adjust the system time in small 148 * increments. 149 * 150 * time_constant determines the bandwidth or "stiffness" of the PLL. 151 * 152 * time_tolerance determines maximum frequency error or tolerance of the 153 * CPU clock oscillator and is a property of the architecture; however, 154 * in principle it could change as result of the presence of external 155 * discipline signals, for instance. 156 * 157 * time_precision is usually equal to the kernel tick variable; however, 158 * in cases where a precision clock counter or external clock is 159 * available, the resolution can be much less than this and depend on 160 * whether the external clock is working or not. 161 * 162 * time_maxerror is initialized by a ntp_adjtime() call and increased by 163 * the kernel once each second to reflect the maximum error bound 164 * growth. 165 * 166 * time_esterror is set and read by the ntp_adjtime() call, but 167 * otherwise not used by the kernel. 168 */ 169 int time_state = TIME_OK; /* clock state */ 170 int time_status = STA_UNSYNC; /* clock status bits */ 171 long time_offset = 0; /* time offset (us) */ 172 long time_constant = 0; /* pll time constant */ 173 long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */ 174 long time_precision = 1; /* clock precision (us) */ 175 long time_maxerror = MAXPHASE; /* maximum error (us) */ 176 long time_esterror = MAXPHASE; /* estimated error (us) */ 177 178 /* 179 * The following variables establish the state of the PLL/FLL and the 180 * residual time and frequency offset of the local clock. The scale 181 * factors are defined in the timex.h header file. 182 * 183 * time_phase and time_freq are the phase increment and the frequency 184 * increment, respectively, of the kernel time variable. 185 * 186 * time_freq is set via ntp_adjtime() from a value stored in a file when 187 * the synchronization daemon is first started. Its value is retrieved 188 * via ntp_adjtime() and written to the file about once per hour by the 189 * daemon. 190 * 191 * time_adj is the adjustment added to the value of tick at each timer 192 * interrupt and is recomputed from time_phase and time_freq at each 193 * seconds rollover. 194 * 195 * time_reftime is the second's portion of the system time at the last 196 * call to ntp_adjtime(). It is used to adjust the time_freq variable 197 * and to increase the time_maxerror as the time since last update 198 * increases. 199 */ 200 long time_phase = 0; /* phase offset (scaled us) */ 201 long time_freq = 0; /* frequency offset (scaled ppm) */ 202 long time_adj = 0; /* tick adjust (scaled 1 / hz) */ 203 long time_reftime = 0; /* time at last adjustment (s) */ 204 205 #ifdef PPS_SYNC 206 /* 207 * The following variables are used only if the kernel PPS discipline 208 * code is configured (PPS_SYNC). The scale factors are defined in the 209 * timex.h header file. 210 * 211 * pps_time contains the time at each calibration interval, as read by 212 * microtime(). pps_count counts the seconds of the calibration 213 * interval, the duration of which is nominally pps_shift in powers of 214 * two. 215 * 216 * pps_offset is the time offset produced by the time median filter 217 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by 218 * this filter. 219 * 220 * pps_freq is the frequency offset produced by the frequency median 221 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured 222 * by this filter. 223 * 224 * pps_usec is latched from a high resolution counter or external clock 225 * at pps_time. Here we want the hardware counter contents only, not the 226 * contents plus the time_tv.usec as usual. 227 * 228 * pps_valid counts the number of seconds since the last PPS update. It 229 * is used as a watchdog timer to disable the PPS discipline should the 230 * PPS signal be lost. 231 * 232 * pps_glitch counts the number of seconds since the beginning of an 233 * offset burst more than tick/2 from current nominal offset. It is used 234 * mainly to suppress error bursts due to priority conflicts between the 235 * PPS interrupt and timer interrupt. 236 * 237 * pps_intcnt counts the calibration intervals for use in the interval- 238 * adaptation algorithm. It's just too complicated for words. 239 */ 240 struct timeval pps_time; /* kernel time at last interval */ 241 long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */ 242 long pps_offset = 0; /* pps time offset (us) */ 243 long pps_jitter = MAXTIME; /* time dispersion (jitter) (us) */ 244 long pps_ff[] = {0, 0, 0}; /* pps frequency offset median filter */ 245 long pps_freq = 0; /* frequency offset (scaled ppm) */ 246 long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */ 247 long pps_usec = 0; /* microsec counter at last interval */ 248 long pps_valid = PPS_VALID; /* pps signal watchdog counter */ 249 int pps_glitch = 0; /* pps signal glitch counter */ 250 int pps_count = 0; /* calibration interval counter (s) */ 251 int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */ 252 int pps_intcnt = 0; /* intervals at current duration */ 253 254 /* 255 * PPS signal quality monitors 256 * 257 * pps_jitcnt counts the seconds that have been discarded because the 258 * jitter measured by the time median filter exceeds the limit MAXTIME 259 * (100 us). 260 * 261 * pps_calcnt counts the frequency calibration intervals, which are 262 * variable from 4 s to 256 s. 263 * 264 * pps_errcnt counts the calibration intervals which have been discarded 265 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the 266 * calibration interval jitter exceeds two ticks. 267 * 268 * pps_stbcnt counts the calibration intervals that have been discarded 269 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us). 270 */ 271 long pps_jitcnt = 0; /* jitter limit exceeded */ 272 long pps_calcnt = 0; /* calibration intervals */ 273 long pps_errcnt = 0; /* calibration errors */ 274 long pps_stbcnt = 0; /* stability limit exceeded */ 275 #endif /* PPS_SYNC */ 276 277 #ifdef EXT_CLOCK 278 /* 279 * External clock definitions 280 * 281 * The following definitions and declarations are used only if an 282 * external clock is configured on the system. 283 */ 284 #define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */ 285 286 /* 287 * The clock_count variable is set to CLOCK_INTERVAL at each PPS 288 * interrupt and decremented once each second. 289 */ 290 int clock_count = 0; /* CPU clock counter */ 291 292 #ifdef HIGHBALL 293 /* 294 * The clock_offset and clock_cpu variables are used by the HIGHBALL 295 * interface. The clock_offset variable defines the offset between 296 * system time and the HIGBALL counters. The clock_cpu variable contains 297 * the offset between the system clock and the HIGHBALL clock for use in 298 * disciplining the kernel time variable. 299 */ 300 extern struct timeval clock_offset; /* Highball clock offset */ 301 long clock_cpu = 0; /* CPU clock adjust */ 302 #endif /* HIGHBALL */ 303 #endif /* EXT_CLOCK */ 304 #endif /* NTP */ 305 306 307 /* 308 * Bump a timeval by a small number of usec's. 309 */ 310 #define BUMPTIME(t, usec) { \ 311 volatile struct timeval *tp = (t); \ 312 long us; \ 313 \ 314 tp->tv_usec = us = tp->tv_usec + (usec); \ 315 if (us >= 1000000) { \ 316 tp->tv_usec = us - 1000000; \ 317 tp->tv_sec++; \ 318 } \ 319 } 320 321 int stathz; 322 int profhz; 323 int schedhz; 324 int profprocs; 325 int softclock_running; /* 1 => softclock() is running */ 326 static int psdiv; /* prof => stat divider */ 327 int psratio; /* ratio: prof / stat */ 328 int tickfix, tickfixinterval; /* used if tick not really integral */ 329 #ifndef NTP 330 static int tickfixcnt; /* accumulated fractional error */ 331 #else 332 int fixtick; /* used by NTP for same */ 333 int shifthz; 334 #endif 335 336 /* 337 * We might want ldd to load the both words from time at once. 338 * To succeed we need to be quadword aligned. 339 * The sparc already does that, and that it has worked so far is a fluke. 340 */ 341 volatile struct timeval time __attribute__((__aligned__(__alignof__(quad_t)))); 342 volatile struct timeval mono_time; 343 344 /* 345 * The callout mechanism is based on the work of Adam M. Costello and 346 * George Varghese, published in a technical report entitled "Redesigning 347 * the BSD Callout and Timer Facilities", and Justin Gibbs's subsequent 348 * integration into FreeBSD, modified for NetBSD by Jason R. Thorpe. 349 * 350 * The original work on the data structures used in this implementation 351 * was published by G. Varghese and A. Lauck in the paper "Hashed and 352 * Hierarchical Timing Wheels: Data Structures for the Efficient 353 * Implementation of a Timer Facility" in the Proceedings of the 11th 354 * ACM Annual Symposium on Operating System Principles, Austin, Texas, 355 * November 1987. 356 */ 357 struct callout_queue *callwheel; 358 int callwheelsize, callwheelbits, callwheelmask; 359 360 static struct callout *nextsoftcheck; /* next callout to be checked */ 361 362 #ifdef CALLWHEEL_STATS 363 int *callwheel_sizes; /* per-bucket length count */ 364 struct evcnt callwheel_collisions; /* number of hash collisions */ 365 struct evcnt callwheel_maxlength; /* length of the longest hash chain */ 366 struct evcnt callwheel_count; /* # callouts currently */ 367 struct evcnt callwheel_established; /* # callouts established */ 368 struct evcnt callwheel_fired; /* # callouts that fired */ 369 struct evcnt callwheel_disestablished; /* # callouts disestablished */ 370 struct evcnt callwheel_changed; /* # callouts changed */ 371 struct evcnt callwheel_softclocks; /* # times softclock() called */ 372 struct evcnt callwheel_softchecks; /* # checks per softclock() */ 373 struct evcnt callwheel_softempty; /* # empty buckets seen */ 374 struct evcnt callwheel_hintworked; /* # times hint saved scan */ 375 #endif /* CALLWHEEL_STATS */ 376 377 /* 378 * This value indicates the number of consecutive callouts that 379 * will be checked before we allow interrupts to have a chance 380 * again. 381 */ 382 #ifndef MAX_SOFTCLOCK_STEPS 383 #define MAX_SOFTCLOCK_STEPS 100 384 #endif 385 386 struct simplelock callwheel_slock; 387 388 #define CALLWHEEL_LOCK(s) \ 389 do { \ 390 s = splclock(); \ 391 simple_lock(&callwheel_slock); \ 392 } while (0) 393 394 #define CALLWHEEL_UNLOCK(s) \ 395 do { \ 396 simple_unlock(&callwheel_slock); \ 397 splx(s); \ 398 } while (0) 399 400 static void callout_stop_locked(struct callout *); 401 402 /* 403 * These are both protected by callwheel_lock. 404 * XXX SHOULD BE STATIC!! 405 */ 406 u_int64_t hardclock_ticks, softclock_ticks; 407 408 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 409 void softclock(void *); 410 void *softclock_si; 411 #endif 412 413 /* 414 * Initialize clock frequencies and start both clocks running. 415 */ 416 void 417 initclocks(void) 418 { 419 int i; 420 421 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 422 softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL); 423 if (softclock_si == NULL) 424 panic("initclocks: unable to register softclock intr"); 425 #endif 426 427 /* 428 * Set divisors to 1 (normal case) and let the machine-specific 429 * code do its bit. 430 */ 431 psdiv = 1; 432 cpu_initclocks(); 433 434 /* 435 * Compute profhz/stathz/rrticks, and fix profhz if needed. 436 */ 437 i = stathz ? stathz : hz; 438 if (profhz == 0) 439 profhz = i; 440 psratio = profhz / i; 441 rrticks = hz / 10; 442 443 #ifdef NTP 444 switch (hz) { 445 case 1: 446 shifthz = SHIFT_SCALE - 0; 447 break; 448 case 2: 449 shifthz = SHIFT_SCALE - 1; 450 break; 451 case 4: 452 shifthz = SHIFT_SCALE - 2; 453 break; 454 case 8: 455 shifthz = SHIFT_SCALE - 3; 456 break; 457 case 16: 458 shifthz = SHIFT_SCALE - 4; 459 break; 460 case 32: 461 shifthz = SHIFT_SCALE - 5; 462 break; 463 case 60: 464 case 64: 465 shifthz = SHIFT_SCALE - 6; 466 break; 467 case 96: 468 case 100: 469 case 128: 470 shifthz = SHIFT_SCALE - 7; 471 break; 472 case 256: 473 shifthz = SHIFT_SCALE - 8; 474 break; 475 case 512: 476 shifthz = SHIFT_SCALE - 9; 477 break; 478 case 1000: 479 case 1024: 480 shifthz = SHIFT_SCALE - 10; 481 break; 482 case 1200: 483 case 2048: 484 shifthz = SHIFT_SCALE - 11; 485 break; 486 case 4096: 487 shifthz = SHIFT_SCALE - 12; 488 break; 489 case 8192: 490 shifthz = SHIFT_SCALE - 13; 491 break; 492 case 16384: 493 shifthz = SHIFT_SCALE - 14; 494 break; 495 case 32768: 496 shifthz = SHIFT_SCALE - 15; 497 break; 498 case 65536: 499 shifthz = SHIFT_SCALE - 16; 500 break; 501 default: 502 panic("weird hz"); 503 } 504 if (fixtick == 0) { 505 /* 506 * Give MD code a chance to set this to a better 507 * value; but, if it doesn't, we should. 508 */ 509 fixtick = (1000000 - (hz*tick)); 510 } 511 #endif 512 } 513 514 /* 515 * The real-time timer, interrupting hz times per second. 516 */ 517 void 518 hardclock(struct clockframe *frame) 519 { 520 struct proc *p; 521 int delta; 522 extern int tickdelta; 523 extern long timedelta; 524 struct cpu_info *ci = curcpu(); 525 #ifdef NTP 526 int time_update; 527 int ltemp; 528 #endif 529 530 p = curproc; 531 if (p) { 532 struct pstats *pstats; 533 534 /* 535 * Run current process's virtual and profile time, as needed. 536 */ 537 pstats = p->p_stats; 538 if (CLKF_USERMODE(frame) && 539 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 540 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 541 psignal(p, SIGVTALRM); 542 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 543 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 544 psignal(p, SIGPROF); 545 } 546 547 /* 548 * If no separate statistics clock is available, run it from here. 549 */ 550 if (stathz == 0) 551 statclock(frame); 552 if ((--ci->ci_schedstate.spc_rrticks) <= 0) 553 roundrobin(ci); 554 555 #if defined(MULTIPROCESSOR) 556 /* 557 * If we are not the primary CPU, we're not allowed to do 558 * any more work. 559 */ 560 if (CPU_IS_PRIMARY(ci) == 0) 561 return; 562 #endif 563 564 /* 565 * Increment the time-of-day. The increment is normally just 566 * ``tick''. If the machine is one which has a clock frequency 567 * such that ``hz'' would not divide the second evenly into 568 * milliseconds, a periodic adjustment must be applied. Finally, 569 * if we are still adjusting the time (see adjtime()), 570 * ``tickdelta'' may also be added in. 571 */ 572 delta = tick; 573 574 #ifndef NTP 575 if (tickfix) { 576 tickfixcnt += tickfix; 577 if (tickfixcnt >= tickfixinterval) { 578 delta++; 579 tickfixcnt -= tickfixinterval; 580 } 581 } 582 #endif /* !NTP */ 583 /* Imprecise 4bsd adjtime() handling */ 584 if (timedelta != 0) { 585 delta += tickdelta; 586 timedelta -= tickdelta; 587 } 588 589 #ifdef notyet 590 microset(); 591 #endif 592 593 #ifndef NTP 594 BUMPTIME(&time, delta); /* XXX Now done using NTP code below */ 595 #endif 596 BUMPTIME(&mono_time, delta); 597 598 #ifdef NTP 599 time_update = delta; 600 601 /* 602 * Compute the phase adjustment. If the low-order bits 603 * (time_phase) of the update overflow, bump the high-order bits 604 * (time_update). 605 */ 606 time_phase += time_adj; 607 if (time_phase <= -FINEUSEC) { 608 ltemp = -time_phase >> SHIFT_SCALE; 609 time_phase += ltemp << SHIFT_SCALE; 610 time_update -= ltemp; 611 } else if (time_phase >= FINEUSEC) { 612 ltemp = time_phase >> SHIFT_SCALE; 613 time_phase -= ltemp << SHIFT_SCALE; 614 time_update += ltemp; 615 } 616 617 #ifdef HIGHBALL 618 /* 619 * If the HIGHBALL board is installed, we need to adjust the 620 * external clock offset in order to close the hardware feedback 621 * loop. This will adjust the external clock phase and frequency 622 * in small amounts. The additional phase noise and frequency 623 * wander this causes should be minimal. We also need to 624 * discipline the kernel time variable, since the PLL is used to 625 * discipline the external clock. If the Highball board is not 626 * present, we discipline kernel time with the PLL as usual. We 627 * assume that the external clock phase adjustment (time_update) 628 * and kernel phase adjustment (clock_cpu) are less than the 629 * value of tick. 630 */ 631 clock_offset.tv_usec += time_update; 632 if (clock_offset.tv_usec >= 1000000) { 633 clock_offset.tv_sec++; 634 clock_offset.tv_usec -= 1000000; 635 } 636 if (clock_offset.tv_usec < 0) { 637 clock_offset.tv_sec--; 638 clock_offset.tv_usec += 1000000; 639 } 640 time.tv_usec += clock_cpu; 641 clock_cpu = 0; 642 #else 643 time.tv_usec += time_update; 644 #endif /* HIGHBALL */ 645 646 /* 647 * On rollover of the second the phase adjustment to be used for 648 * the next second is calculated. Also, the maximum error is 649 * increased by the tolerance. If the PPS frequency discipline 650 * code is present, the phase is increased to compensate for the 651 * CPU clock oscillator frequency error. 652 * 653 * On a 32-bit machine and given parameters in the timex.h 654 * header file, the maximum phase adjustment is +-512 ms and 655 * maximum frequency offset is a tad less than) +-512 ppm. On a 656 * 64-bit machine, you shouldn't need to ask. 657 */ 658 if (time.tv_usec >= 1000000) { 659 time.tv_usec -= 1000000; 660 time.tv_sec++; 661 time_maxerror += time_tolerance >> SHIFT_USEC; 662 663 /* 664 * Leap second processing. If in leap-insert state at 665 * the end of the day, the system clock is set back one 666 * second; if in leap-delete state, the system clock is 667 * set ahead one second. The microtime() routine or 668 * external clock driver will insure that reported time 669 * is always monotonic. The ugly divides should be 670 * replaced. 671 */ 672 switch (time_state) { 673 case TIME_OK: 674 if (time_status & STA_INS) 675 time_state = TIME_INS; 676 else if (time_status & STA_DEL) 677 time_state = TIME_DEL; 678 break; 679 680 case TIME_INS: 681 if (time.tv_sec % 86400 == 0) { 682 time.tv_sec--; 683 time_state = TIME_OOP; 684 } 685 break; 686 687 case TIME_DEL: 688 if ((time.tv_sec + 1) % 86400 == 0) { 689 time.tv_sec++; 690 time_state = TIME_WAIT; 691 } 692 break; 693 694 case TIME_OOP: 695 time_state = TIME_WAIT; 696 break; 697 698 case TIME_WAIT: 699 if (!(time_status & (STA_INS | STA_DEL))) 700 time_state = TIME_OK; 701 break; 702 } 703 704 /* 705 * Compute the phase adjustment for the next second. In 706 * PLL mode, the offset is reduced by a fixed factor 707 * times the time constant. In FLL mode the offset is 708 * used directly. In either mode, the maximum phase 709 * adjustment for each second is clamped so as to spread 710 * the adjustment over not more than the number of 711 * seconds between updates. 712 */ 713 if (time_offset < 0) { 714 ltemp = -time_offset; 715 if (!(time_status & STA_FLL)) 716 ltemp >>= SHIFT_KG + time_constant; 717 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 718 ltemp = (MAXPHASE / MINSEC) << 719 SHIFT_UPDATE; 720 time_offset += ltemp; 721 time_adj = -ltemp << (shifthz - SHIFT_UPDATE); 722 } else if (time_offset > 0) { 723 ltemp = time_offset; 724 if (!(time_status & STA_FLL)) 725 ltemp >>= SHIFT_KG + time_constant; 726 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 727 ltemp = (MAXPHASE / MINSEC) << 728 SHIFT_UPDATE; 729 time_offset -= ltemp; 730 time_adj = ltemp << (shifthz - SHIFT_UPDATE); 731 } else 732 time_adj = 0; 733 734 /* 735 * Compute the frequency estimate and additional phase 736 * adjustment due to frequency error for the next 737 * second. When the PPS signal is engaged, gnaw on the 738 * watchdog counter and update the frequency computed by 739 * the pll and the PPS signal. 740 */ 741 #ifdef PPS_SYNC 742 pps_valid++; 743 if (pps_valid == PPS_VALID) { 744 pps_jitter = MAXTIME; 745 pps_stabil = MAXFREQ; 746 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 747 STA_PPSWANDER | STA_PPSERROR); 748 } 749 ltemp = time_freq + pps_freq; 750 #else 751 ltemp = time_freq; 752 #endif /* PPS_SYNC */ 753 754 if (ltemp < 0) 755 time_adj -= -ltemp >> (SHIFT_USEC - shifthz); 756 else 757 time_adj += ltemp >> (SHIFT_USEC - shifthz); 758 time_adj += (long)fixtick << shifthz; 759 760 /* 761 * When the CPU clock oscillator frequency is not a 762 * power of 2 in Hz, shifthz is only an approximate 763 * scale factor. 764 * 765 * To determine the adjustment, you can do the following: 766 * bc -q 767 * scale=24 768 * obase=2 769 * idealhz/realhz 770 * where `idealhz' is the next higher power of 2, and `realhz' 771 * is the actual value. You may need to factor this result 772 * into a sequence of 2 multipliers to get better precision. 773 * 774 * Likewise, the error can be calculated with (e.g. for 100Hz): 775 * bc -q 776 * scale=24 777 * ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz 778 * (and then multiply by 1000000 to get ppm). 779 */ 780 switch (hz) { 781 case 60: 782 /* A factor of 1.000100010001 gives about 15ppm 783 error. */ 784 if (time_adj < 0) { 785 time_adj -= (-time_adj >> 4); 786 time_adj -= (-time_adj >> 8); 787 } else { 788 time_adj += (time_adj >> 4); 789 time_adj += (time_adj >> 8); 790 } 791 break; 792 793 case 96: 794 /* A factor of 1.0101010101 gives about 244ppm error. */ 795 if (time_adj < 0) { 796 time_adj -= (-time_adj >> 2); 797 time_adj -= (-time_adj >> 4) + (-time_adj >> 8); 798 } else { 799 time_adj += (time_adj >> 2); 800 time_adj += (time_adj >> 4) + (time_adj >> 8); 801 } 802 break; 803 804 case 100: 805 /* A factor of 1.010001111010111 gives about 1ppm 806 error. */ 807 if (time_adj < 0) { 808 time_adj -= (-time_adj >> 2) + (-time_adj >> 5); 809 time_adj += (-time_adj >> 10); 810 } else { 811 time_adj += (time_adj >> 2) + (time_adj >> 5); 812 time_adj -= (time_adj >> 10); 813 } 814 break; 815 816 case 1000: 817 /* A factor of 1.000001100010100001 gives about 50ppm 818 error. */ 819 if (time_adj < 0) { 820 time_adj -= (-time_adj >> 6) + (-time_adj >> 11); 821 time_adj -= (-time_adj >> 7); 822 } else { 823 time_adj += (time_adj >> 6) + (time_adj >> 11); 824 time_adj += (time_adj >> 7); 825 } 826 break; 827 828 case 1200: 829 /* A factor of 1.1011010011100001 gives about 64ppm 830 error. */ 831 if (time_adj < 0) { 832 time_adj -= (-time_adj >> 1) + (-time_adj >> 6); 833 time_adj -= (-time_adj >> 3) + (-time_adj >> 10); 834 } else { 835 time_adj += (time_adj >> 1) + (time_adj >> 6); 836 time_adj += (time_adj >> 3) + (time_adj >> 10); 837 } 838 break; 839 } 840 841 #ifdef EXT_CLOCK 842 /* 843 * If an external clock is present, it is necessary to 844 * discipline the kernel time variable anyway, since not 845 * all system components use the microtime() interface. 846 * Here, the time offset between the external clock and 847 * kernel time variable is computed every so often. 848 */ 849 clock_count++; 850 if (clock_count > CLOCK_INTERVAL) { 851 clock_count = 0; 852 microtime(&clock_ext); 853 delta.tv_sec = clock_ext.tv_sec - time.tv_sec; 854 delta.tv_usec = clock_ext.tv_usec - 855 time.tv_usec; 856 if (delta.tv_usec < 0) 857 delta.tv_sec--; 858 if (delta.tv_usec >= 500000) { 859 delta.tv_usec -= 1000000; 860 delta.tv_sec++; 861 } 862 if (delta.tv_usec < -500000) { 863 delta.tv_usec += 1000000; 864 delta.tv_sec--; 865 } 866 if (delta.tv_sec > 0 || (delta.tv_sec == 0 && 867 delta.tv_usec > MAXPHASE) || 868 delta.tv_sec < -1 || (delta.tv_sec == -1 && 869 delta.tv_usec < -MAXPHASE)) { 870 time = clock_ext; 871 delta.tv_sec = 0; 872 delta.tv_usec = 0; 873 } 874 #ifdef HIGHBALL 875 clock_cpu = delta.tv_usec; 876 #else /* HIGHBALL */ 877 hardupdate(delta.tv_usec); 878 #endif /* HIGHBALL */ 879 } 880 #endif /* EXT_CLOCK */ 881 } 882 883 #endif /* NTP */ 884 885 /* 886 * Process callouts at a very low cpu priority, so we don't keep the 887 * relatively high clock interrupt priority any longer than necessary. 888 */ 889 simple_lock(&callwheel_slock); /* already at splclock() */ 890 hardclock_ticks++; 891 if (! TAILQ_EMPTY(&callwheel[hardclock_ticks & callwheelmask].cq_q)) { 892 simple_unlock(&callwheel_slock); 893 if (CLKF_BASEPRI(frame)) { 894 /* 895 * Save the overhead of a software interrupt; 896 * it will happen as soon as we return, so do 897 * it now. 898 * 899 * NOTE: If we're at ``base priority'', softclock() 900 * was not already running. 901 */ 902 spllowersoftclock(); 903 KERNEL_LOCK(LK_CANRECURSE|LK_EXCLUSIVE); 904 softclock(NULL); 905 KERNEL_UNLOCK(); 906 } else { 907 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 908 softintr_schedule(softclock_si); 909 #else 910 setsoftclock(); 911 #endif 912 } 913 return; 914 } else if (softclock_running == 0 && 915 (softclock_ticks + 1) == hardclock_ticks) { 916 softclock_ticks++; 917 } 918 simple_unlock(&callwheel_slock); 919 } 920 921 /* 922 * Software (low priority) clock interrupt. 923 * Run periodic events from timeout queue. 924 */ 925 /*ARGSUSED*/ 926 void 927 softclock(void *v) 928 { 929 struct callout_queue *bucket; 930 struct callout *c; 931 void (*func)(void *); 932 void *arg; 933 int s, idx; 934 int steps = 0; 935 936 CALLWHEEL_LOCK(s); 937 938 softclock_running = 1; 939 940 #ifdef CALLWHEEL_STATS 941 callwheel_softclocks.ev_count++; 942 #endif 943 944 while (softclock_ticks != hardclock_ticks) { 945 softclock_ticks++; 946 idx = (int)(softclock_ticks & callwheelmask); 947 bucket = &callwheel[idx]; 948 c = TAILQ_FIRST(&bucket->cq_q); 949 if (c == NULL) { 950 #ifdef CALLWHEEL_STATS 951 callwheel_softempty.ev_count++; 952 #endif 953 continue; 954 } 955 if (softclock_ticks < bucket->cq_hint) { 956 #ifdef CALLWHEEL_STATS 957 callwheel_hintworked.ev_count++; 958 #endif 959 continue; 960 } 961 bucket->cq_hint = UQUAD_MAX; 962 while (c != NULL) { 963 #ifdef CALLWHEEL_STATS 964 callwheel_softchecks.ev_count++; 965 #endif 966 if (c->c_time != softclock_ticks) { 967 if (c->c_time < bucket->cq_hint) 968 bucket->cq_hint = c->c_time; 969 c = TAILQ_NEXT(c, c_link); 970 if (++steps >= MAX_SOFTCLOCK_STEPS) { 971 nextsoftcheck = c; 972 /* Give interrupts a chance. */ 973 CALLWHEEL_UNLOCK(s); 974 CALLWHEEL_LOCK(s); 975 c = nextsoftcheck; 976 steps = 0; 977 } 978 } else { 979 nextsoftcheck = TAILQ_NEXT(c, c_link); 980 TAILQ_REMOVE(&bucket->cq_q, c, c_link); 981 #ifdef CALLWHEEL_STATS 982 callwheel_sizes[idx]--; 983 callwheel_fired.ev_count++; 984 callwheel_count.ev_count--; 985 #endif 986 func = c->c_func; 987 arg = c->c_arg; 988 c->c_func = NULL; 989 c->c_flags &= ~CALLOUT_PENDING; 990 CALLWHEEL_UNLOCK(s); 991 (*func)(arg); 992 CALLWHEEL_LOCK(s); 993 steps = 0; 994 c = nextsoftcheck; 995 } 996 } 997 if (TAILQ_EMPTY(&bucket->cq_q)) 998 bucket->cq_hint = UQUAD_MAX; 999 } 1000 nextsoftcheck = NULL; 1001 softclock_running = 0; 1002 CALLWHEEL_UNLOCK(s); 1003 } 1004 1005 /* 1006 * callout_setsize: 1007 * 1008 * Determine how many callwheels are necessary and 1009 * set hash mask. Called from allocsys(). 1010 */ 1011 void 1012 callout_setsize(void) 1013 { 1014 1015 for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1) 1016 /* loop */ ; 1017 callwheelmask = callwheelsize - 1; 1018 } 1019 1020 /* 1021 * callout_startup: 1022 * 1023 * Initialize the callwheel buckets. 1024 */ 1025 void 1026 callout_startup(void) 1027 { 1028 int i; 1029 1030 for (i = 0; i < callwheelsize; i++) { 1031 callwheel[i].cq_hint = UQUAD_MAX; 1032 TAILQ_INIT(&callwheel[i].cq_q); 1033 } 1034 1035 simple_lock_init(&callwheel_slock); 1036 1037 #ifdef CALLWHEEL_STATS 1038 evcnt_attach_dynamic(&callwheel_collisions, EVCNT_TYPE_MISC, 1039 NULL, "callwheel", "collisions"); 1040 evcnt_attach_dynamic(&callwheel_maxlength, EVCNT_TYPE_MISC, 1041 NULL, "callwheel", "maxlength"); 1042 evcnt_attach_dynamic(&callwheel_count, EVCNT_TYPE_MISC, 1043 NULL, "callwheel", "count"); 1044 evcnt_attach_dynamic(&callwheel_established, EVCNT_TYPE_MISC, 1045 NULL, "callwheel", "established"); 1046 evcnt_attach_dynamic(&callwheel_fired, EVCNT_TYPE_MISC, 1047 NULL, "callwheel", "fired"); 1048 evcnt_attach_dynamic(&callwheel_disestablished, EVCNT_TYPE_MISC, 1049 NULL, "callwheel", "disestablished"); 1050 evcnt_attach_dynamic(&callwheel_changed, EVCNT_TYPE_MISC, 1051 NULL, "callwheel", "changed"); 1052 evcnt_attach_dynamic(&callwheel_softclocks, EVCNT_TYPE_MISC, 1053 NULL, "callwheel", "softclocks"); 1054 evcnt_attach_dynamic(&callwheel_softempty, EVCNT_TYPE_MISC, 1055 NULL, "callwheel", "softempty"); 1056 evcnt_attach_dynamic(&callwheel_hintworked, EVCNT_TYPE_MISC, 1057 NULL, "callwheel", "hintworked"); 1058 #endif /* CALLWHEEL_STATS */ 1059 } 1060 1061 /* 1062 * callout_init: 1063 * 1064 * Initialize a callout structure so that it can be used 1065 * by callout_reset() and callout_stop(). 1066 */ 1067 void 1068 callout_init(struct callout *c) 1069 { 1070 1071 memset(c, 0, sizeof(*c)); 1072 } 1073 1074 /* 1075 * callout_reset: 1076 * 1077 * Establish or change a timeout. 1078 */ 1079 void 1080 callout_reset(struct callout *c, int ticks, void (*func)(void *), void *arg) 1081 { 1082 struct callout_queue *bucket; 1083 int s; 1084 1085 if (ticks <= 0) 1086 ticks = 1; 1087 1088 CALLWHEEL_LOCK(s); 1089 1090 /* 1091 * If this callout's timer is already running, cancel it 1092 * before we modify it. 1093 */ 1094 if (c->c_flags & CALLOUT_PENDING) { 1095 callout_stop_locked(c); /* Already locked */ 1096 #ifdef CALLWHEEL_STATS 1097 callwheel_changed.ev_count++; 1098 #endif 1099 } 1100 1101 c->c_arg = arg; 1102 c->c_func = func; 1103 c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING; 1104 c->c_time = hardclock_ticks + ticks; 1105 1106 bucket = &callwheel[c->c_time & callwheelmask]; 1107 1108 #ifdef CALLWHEEL_STATS 1109 if (! TAILQ_EMPTY(&bucket->cq_q)) 1110 callwheel_collisions.ev_count++; 1111 #endif 1112 1113 TAILQ_INSERT_TAIL(&bucket->cq_q, c, c_link); 1114 if (c->c_time < bucket->cq_hint) 1115 bucket->cq_hint = c->c_time; 1116 1117 #ifdef CALLWHEEL_STATS 1118 callwheel_count.ev_count++; 1119 callwheel_established.ev_count++; 1120 if (++callwheel_sizes[c->c_time & callwheelmask] > 1121 callwheel_maxlength.ev_count) 1122 callwheel_maxlength.ev_count = 1123 callwheel_sizes[c->c_time & callwheelmask]; 1124 #endif 1125 1126 CALLWHEEL_UNLOCK(s); 1127 } 1128 1129 /* 1130 * callout_stop_locked: 1131 * 1132 * Disestablish a timeout. Callwheel is locked. 1133 */ 1134 static void 1135 callout_stop_locked(struct callout *c) 1136 { 1137 struct callout_queue *bucket; 1138 1139 /* 1140 * Don't attempt to delete a callout that's not on the queue. 1141 */ 1142 if ((c->c_flags & CALLOUT_PENDING) == 0) { 1143 c->c_flags &= ~CALLOUT_ACTIVE; 1144 return; 1145 } 1146 1147 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 1148 1149 if (nextsoftcheck == c) 1150 nextsoftcheck = TAILQ_NEXT(c, c_link); 1151 1152 bucket = &callwheel[c->c_time & callwheelmask]; 1153 TAILQ_REMOVE(&bucket->cq_q, c, c_link); 1154 if (TAILQ_EMPTY(&bucket->cq_q)) 1155 bucket->cq_hint = UQUAD_MAX; 1156 #ifdef CALLWHEEL_STATS 1157 callwheel_count.ev_count--; 1158 callwheel_disestablished.ev_count++; 1159 callwheel_sizes[c->c_time & callwheelmask]--; 1160 #endif 1161 1162 c->c_func = NULL; 1163 } 1164 1165 /* 1166 * callout_stop: 1167 * 1168 * Disestablish a timeout. Callwheel is unlocked. This is 1169 * the standard entry point. 1170 */ 1171 void 1172 callout_stop(struct callout *c) 1173 { 1174 int s; 1175 1176 CALLWHEEL_LOCK(s); 1177 callout_stop_locked(c); 1178 CALLWHEEL_UNLOCK(s); 1179 } 1180 1181 #ifdef CALLWHEEL_STATS 1182 /* 1183 * callout_showstats: 1184 * 1185 * Display callout statistics. Call it from DDB. 1186 */ 1187 void 1188 callout_showstats(void) 1189 { 1190 u_int64_t curticks; 1191 int s; 1192 1193 s = splclock(); 1194 curticks = softclock_ticks; 1195 splx(s); 1196 1197 printf("Callwheel statistics:\n"); 1198 printf("\tCallouts currently queued: %llu\n", 1199 (long long) callwheel_count.ev_count); 1200 printf("\tCallouts established: %llu\n", 1201 (long long) callwheel_established.ev_count); 1202 printf("\tCallouts disestablished: %llu\n", 1203 (long long) callwheel_disestablished.ev_count); 1204 if (callwheel_changed.ev_count != 0) 1205 printf("\t\tOf those, %llu were changes\n", 1206 (long long) callwheel_changed.ev_count); 1207 printf("\tCallouts that fired: %llu\n", 1208 (long long) callwheel_fired.ev_count); 1209 printf("\tNumber of buckets: %d\n", callwheelsize); 1210 printf("\tNumber of hash collisions: %llu\n", 1211 (long long) callwheel_collisions.ev_count); 1212 printf("\tMaximum hash chain length: %llu\n", 1213 (long long) callwheel_maxlength.ev_count); 1214 printf("\tSoftclocks: %llu, Softchecks: %llu\n", 1215 (long long) callwheel_softclocks.ev_count, 1216 (long long) callwheel_softchecks.ev_count); 1217 printf("\t\tEmpty buckets seen: %llu\n", 1218 (long long) callwheel_softempty.ev_count); 1219 printf("\t\tTimes hint saved scan: %llu\n", 1220 (long long) callwheel_hintworked.ev_count); 1221 } 1222 #endif 1223 1224 /* 1225 * Compute number of hz until specified time. Used to compute second 1226 * argument to callout_reset() from an absolute time. 1227 */ 1228 int 1229 hzto(struct timeval *tv) 1230 { 1231 unsigned long ticks; 1232 long sec, usec; 1233 int s; 1234 1235 /* 1236 * If the number of usecs in the whole seconds part of the time 1237 * difference fits in a long, then the total number of usecs will 1238 * fit in an unsigned long. Compute the total and convert it to 1239 * ticks, rounding up and adding 1 to allow for the current tick 1240 * to expire. Rounding also depends on unsigned long arithmetic 1241 * to avoid overflow. 1242 * 1243 * Otherwise, if the number of ticks in the whole seconds part of 1244 * the time difference fits in a long, then convert the parts to 1245 * ticks separately and add, using similar rounding methods and 1246 * overflow avoidance. This method would work in the previous 1247 * case, but it is slightly slower and assume that hz is integral. 1248 * 1249 * Otherwise, round the time difference down to the maximum 1250 * representable value. 1251 * 1252 * If ints are 32-bit, then the maximum value for any timeout in 1253 * 10ms ticks is 248 days. 1254 */ 1255 s = splclock(); 1256 sec = tv->tv_sec - time.tv_sec; 1257 usec = tv->tv_usec - time.tv_usec; 1258 splx(s); 1259 1260 if (usec < 0) { 1261 sec--; 1262 usec += 1000000; 1263 } 1264 1265 if (sec < 0 || (sec == 0 && usec <= 0)) { 1266 /* 1267 * Would expire now or in the past. Return 0 ticks. 1268 * This is different from the legacy hzto() interface, 1269 * and callers need to check for it. 1270 */ 1271 ticks = 0; 1272 } else if (sec <= (LONG_MAX / 1000000)) 1273 ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1)) 1274 / tick) + 1; 1275 else if (sec <= (LONG_MAX / hz)) 1276 ticks = (sec * hz) + 1277 (((unsigned long)usec + (tick - 1)) / tick) + 1; 1278 else 1279 ticks = LONG_MAX; 1280 1281 if (ticks > INT_MAX) 1282 ticks = INT_MAX; 1283 1284 return ((int)ticks); 1285 } 1286 1287 /* 1288 * Start profiling on a process. 1289 * 1290 * Kernel profiling passes proc0 which never exits and hence 1291 * keeps the profile clock running constantly. 1292 */ 1293 void 1294 startprofclock(struct proc *p) 1295 { 1296 1297 if ((p->p_flag & P_PROFIL) == 0) { 1298 p->p_flag |= P_PROFIL; 1299 if (++profprocs == 1 && stathz != 0) 1300 psdiv = psratio; 1301 } 1302 } 1303 1304 /* 1305 * Stop profiling on a process. 1306 */ 1307 void 1308 stopprofclock(struct proc *p) 1309 { 1310 1311 if (p->p_flag & P_PROFIL) { 1312 p->p_flag &= ~P_PROFIL; 1313 if (--profprocs == 0 && stathz != 0) 1314 psdiv = 1; 1315 } 1316 } 1317 1318 /* 1319 * Statistics clock. Grab profile sample, and if divider reaches 0, 1320 * do process and kernel statistics. 1321 */ 1322 void 1323 statclock(struct clockframe *frame) 1324 { 1325 #ifdef GPROF 1326 struct gmonparam *g; 1327 intptr_t i; 1328 #endif 1329 struct cpu_info *ci = curcpu(); 1330 struct schedstate_percpu *spc = &ci->ci_schedstate; 1331 struct proc *p; 1332 1333 /* 1334 * Notice changes in divisor frequency, and adjust clock 1335 * frequency accordingly. 1336 */ 1337 if (spc->spc_psdiv != psdiv) { 1338 spc->spc_psdiv = psdiv; 1339 spc->spc_pscnt = psdiv; 1340 if (psdiv == 1) { 1341 setstatclockrate(stathz); 1342 } else { 1343 setstatclockrate(profhz); 1344 } 1345 } 1346 p = curproc; 1347 if (CLKF_USERMODE(frame)) { 1348 if (p->p_flag & P_PROFIL) 1349 addupc_intr(p, CLKF_PC(frame)); 1350 if (--spc->spc_pscnt > 0) 1351 return; 1352 /* 1353 * Came from user mode; CPU was in user state. 1354 * If this process is being profiled record the tick. 1355 */ 1356 p->p_uticks++; 1357 if (p->p_nice > NZERO) 1358 spc->spc_cp_time[CP_NICE]++; 1359 else 1360 spc->spc_cp_time[CP_USER]++; 1361 } else { 1362 #ifdef GPROF 1363 /* 1364 * Kernel statistics are just like addupc_intr, only easier. 1365 */ 1366 g = &_gmonparam; 1367 if (g->state == GMON_PROF_ON) { 1368 i = CLKF_PC(frame) - g->lowpc; 1369 if (i < g->textsize) { 1370 i /= HISTFRACTION * sizeof(*g->kcount); 1371 g->kcount[i]++; 1372 } 1373 } 1374 #endif 1375 #ifdef PROC_PC 1376 if (p && p->p_flag & P_PROFIL) 1377 addupc_intr(p, PROC_PC(p)); 1378 #endif 1379 if (--spc->spc_pscnt > 0) 1380 return; 1381 /* 1382 * Came from kernel mode, so we were: 1383 * - handling an interrupt, 1384 * - doing syscall or trap work on behalf of the current 1385 * user process, or 1386 * - spinning in the idle loop. 1387 * Whichever it is, charge the time as appropriate. 1388 * Note that we charge interrupts to the current process, 1389 * regardless of whether they are ``for'' that process, 1390 * so that we know how much of its real time was spent 1391 * in ``non-process'' (i.e., interrupt) work. 1392 */ 1393 if (CLKF_INTR(frame)) { 1394 if (p != NULL) 1395 p->p_iticks++; 1396 spc->spc_cp_time[CP_INTR]++; 1397 } else if (p != NULL) { 1398 p->p_sticks++; 1399 spc->spc_cp_time[CP_SYS]++; 1400 } else 1401 spc->spc_cp_time[CP_IDLE]++; 1402 } 1403 spc->spc_pscnt = psdiv; 1404 1405 if (p != NULL) { 1406 ++p->p_cpticks; 1407 /* 1408 * If no separate schedclock is provided, call it here 1409 * at ~~12-25 Hz, ~~16 Hz is best 1410 */ 1411 if (schedhz == 0) 1412 if ((++ci->ci_schedstate.spc_schedticks & 3) == 0) 1413 schedclock(p); 1414 } 1415 } 1416 1417 1418 #ifdef NTP /* NTP phase-locked loop in kernel */ 1419 1420 /* 1421 * hardupdate() - local clock update 1422 * 1423 * This routine is called by ntp_adjtime() to update the local clock 1424 * phase and frequency. The implementation is of an adaptive-parameter, 1425 * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new 1426 * time and frequency offset estimates for each call. If the kernel PPS 1427 * discipline code is configured (PPS_SYNC), the PPS signal itself 1428 * determines the new time offset, instead of the calling argument. 1429 * Presumably, calls to ntp_adjtime() occur only when the caller 1430 * believes the local clock is valid within some bound (+-128 ms with 1431 * NTP). If the caller's time is far different than the PPS time, an 1432 * argument will ensue, and it's not clear who will lose. 1433 * 1434 * For uncompensated quartz crystal oscillatores and nominal update 1435 * intervals less than 1024 s, operation should be in phase-lock mode 1436 * (STA_FLL = 0), where the loop is disciplined to phase. For update 1437 * intervals greater than thiss, operation should be in frequency-lock 1438 * mode (STA_FLL = 1), where the loop is disciplined to frequency. 1439 * 1440 * Note: splclock() is in effect. 1441 */ 1442 void 1443 hardupdate(long offset) 1444 { 1445 long ltemp, mtemp; 1446 1447 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME)) 1448 return; 1449 ltemp = offset; 1450 #ifdef PPS_SYNC 1451 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL) 1452 ltemp = pps_offset; 1453 #endif /* PPS_SYNC */ 1454 1455 /* 1456 * Scale the phase adjustment and clamp to the operating range. 1457 */ 1458 if (ltemp > MAXPHASE) 1459 time_offset = MAXPHASE << SHIFT_UPDATE; 1460 else if (ltemp < -MAXPHASE) 1461 time_offset = -(MAXPHASE << SHIFT_UPDATE); 1462 else 1463 time_offset = ltemp << SHIFT_UPDATE; 1464 1465 /* 1466 * Select whether the frequency is to be controlled and in which 1467 * mode (PLL or FLL). Clamp to the operating range. Ugly 1468 * multiply/divide should be replaced someday. 1469 */ 1470 if (time_status & STA_FREQHOLD || time_reftime == 0) 1471 time_reftime = time.tv_sec; 1472 mtemp = time.tv_sec - time_reftime; 1473 time_reftime = time.tv_sec; 1474 if (time_status & STA_FLL) { 1475 if (mtemp >= MINSEC) { 1476 ltemp = ((time_offset / mtemp) << (SHIFT_USEC - 1477 SHIFT_UPDATE)); 1478 if (ltemp < 0) 1479 time_freq -= -ltemp >> SHIFT_KH; 1480 else 1481 time_freq += ltemp >> SHIFT_KH; 1482 } 1483 } else { 1484 if (mtemp < MAXSEC) { 1485 ltemp *= mtemp; 1486 if (ltemp < 0) 1487 time_freq -= -ltemp >> (time_constant + 1488 time_constant + SHIFT_KF - 1489 SHIFT_USEC); 1490 else 1491 time_freq += ltemp >> (time_constant + 1492 time_constant + SHIFT_KF - 1493 SHIFT_USEC); 1494 } 1495 } 1496 if (time_freq > time_tolerance) 1497 time_freq = time_tolerance; 1498 else if (time_freq < -time_tolerance) 1499 time_freq = -time_tolerance; 1500 } 1501 1502 #ifdef PPS_SYNC 1503 /* 1504 * hardpps() - discipline CPU clock oscillator to external PPS signal 1505 * 1506 * This routine is called at each PPS interrupt in order to discipline 1507 * the CPU clock oscillator to the PPS signal. It measures the PPS phase 1508 * and leaves it in a handy spot for the hardclock() routine. It 1509 * integrates successive PPS phase differences and calculates the 1510 * frequency offset. This is used in hardclock() to discipline the CPU 1511 * clock oscillator so that intrinsic frequency error is cancelled out. 1512 * The code requires the caller to capture the time and hardware counter 1513 * value at the on-time PPS signal transition. 1514 * 1515 * Note that, on some Unix systems, this routine runs at an interrupt 1516 * priority level higher than the timer interrupt routine hardclock(). 1517 * Therefore, the variables used are distinct from the hardclock() 1518 * variables, except for certain exceptions: The PPS frequency pps_freq 1519 * and phase pps_offset variables are determined by this routine and 1520 * updated atomically. The time_tolerance variable can be considered a 1521 * constant, since it is infrequently changed, and then only when the 1522 * PPS signal is disabled. The watchdog counter pps_valid is updated 1523 * once per second by hardclock() and is atomically cleared in this 1524 * routine. 1525 */ 1526 void 1527 hardpps(struct timeval *tvp, /* time at PPS */ 1528 long usec /* hardware counter at PPS */) 1529 { 1530 long u_usec, v_usec, bigtick; 1531 long cal_sec, cal_usec; 1532 1533 /* 1534 * An occasional glitch can be produced when the PPS interrupt 1535 * occurs in the hardclock() routine before the time variable is 1536 * updated. Here the offset is discarded when the difference 1537 * between it and the last one is greater than tick/2, but not 1538 * if the interval since the first discard exceeds 30 s. 1539 */ 1540 time_status |= STA_PPSSIGNAL; 1541 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR); 1542 pps_valid = 0; 1543 u_usec = -tvp->tv_usec; 1544 if (u_usec < -500000) 1545 u_usec += 1000000; 1546 v_usec = pps_offset - u_usec; 1547 if (v_usec < 0) 1548 v_usec = -v_usec; 1549 if (v_usec > (tick >> 1)) { 1550 if (pps_glitch > MAXGLITCH) { 1551 pps_glitch = 0; 1552 pps_tf[2] = u_usec; 1553 pps_tf[1] = u_usec; 1554 } else { 1555 pps_glitch++; 1556 u_usec = pps_offset; 1557 } 1558 } else 1559 pps_glitch = 0; 1560 1561 /* 1562 * A three-stage median filter is used to help deglitch the pps 1563 * time. The median sample becomes the time offset estimate; the 1564 * difference between the other two samples becomes the time 1565 * dispersion (jitter) estimate. 1566 */ 1567 pps_tf[2] = pps_tf[1]; 1568 pps_tf[1] = pps_tf[0]; 1569 pps_tf[0] = u_usec; 1570 if (pps_tf[0] > pps_tf[1]) { 1571 if (pps_tf[1] > pps_tf[2]) { 1572 pps_offset = pps_tf[1]; /* 0 1 2 */ 1573 v_usec = pps_tf[0] - pps_tf[2]; 1574 } else if (pps_tf[2] > pps_tf[0]) { 1575 pps_offset = pps_tf[0]; /* 2 0 1 */ 1576 v_usec = pps_tf[2] - pps_tf[1]; 1577 } else { 1578 pps_offset = pps_tf[2]; /* 0 2 1 */ 1579 v_usec = pps_tf[0] - pps_tf[1]; 1580 } 1581 } else { 1582 if (pps_tf[1] < pps_tf[2]) { 1583 pps_offset = pps_tf[1]; /* 2 1 0 */ 1584 v_usec = pps_tf[2] - pps_tf[0]; 1585 } else if (pps_tf[2] < pps_tf[0]) { 1586 pps_offset = pps_tf[0]; /* 1 0 2 */ 1587 v_usec = pps_tf[1] - pps_tf[2]; 1588 } else { 1589 pps_offset = pps_tf[2]; /* 1 2 0 */ 1590 v_usec = pps_tf[1] - pps_tf[0]; 1591 } 1592 } 1593 if (v_usec > MAXTIME) 1594 pps_jitcnt++; 1595 v_usec = (v_usec << PPS_AVG) - pps_jitter; 1596 if (v_usec < 0) 1597 pps_jitter -= -v_usec >> PPS_AVG; 1598 else 1599 pps_jitter += v_usec >> PPS_AVG; 1600 if (pps_jitter > (MAXTIME >> 1)) 1601 time_status |= STA_PPSJITTER; 1602 1603 /* 1604 * During the calibration interval adjust the starting time when 1605 * the tick overflows. At the end of the interval compute the 1606 * duration of the interval and the difference of the hardware 1607 * counters at the beginning and end of the interval. This code 1608 * is deliciously complicated by the fact valid differences may 1609 * exceed the value of tick when using long calibration 1610 * intervals and small ticks. Note that the counter can be 1611 * greater than tick if caught at just the wrong instant, but 1612 * the values returned and used here are correct. 1613 */ 1614 bigtick = (long)tick << SHIFT_USEC; 1615 pps_usec -= pps_freq; 1616 if (pps_usec >= bigtick) 1617 pps_usec -= bigtick; 1618 if (pps_usec < 0) 1619 pps_usec += bigtick; 1620 pps_time.tv_sec++; 1621 pps_count++; 1622 if (pps_count < (1 << pps_shift)) 1623 return; 1624 pps_count = 0; 1625 pps_calcnt++; 1626 u_usec = usec << SHIFT_USEC; 1627 v_usec = pps_usec - u_usec; 1628 if (v_usec >= bigtick >> 1) 1629 v_usec -= bigtick; 1630 if (v_usec < -(bigtick >> 1)) 1631 v_usec += bigtick; 1632 if (v_usec < 0) 1633 v_usec = -(-v_usec >> pps_shift); 1634 else 1635 v_usec = v_usec >> pps_shift; 1636 pps_usec = u_usec; 1637 cal_sec = tvp->tv_sec; 1638 cal_usec = tvp->tv_usec; 1639 cal_sec -= pps_time.tv_sec; 1640 cal_usec -= pps_time.tv_usec; 1641 if (cal_usec < 0) { 1642 cal_usec += 1000000; 1643 cal_sec--; 1644 } 1645 pps_time = *tvp; 1646 1647 /* 1648 * Check for lost interrupts, noise, excessive jitter and 1649 * excessive frequency error. The number of timer ticks during 1650 * the interval may vary +-1 tick. Add to this a margin of one 1651 * tick for the PPS signal jitter and maximum frequency 1652 * deviation. If the limits are exceeded, the calibration 1653 * interval is reset to the minimum and we start over. 1654 */ 1655 u_usec = (long)tick << 1; 1656 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1657 || (cal_sec == 0 && cal_usec < u_usec)) 1658 || v_usec > time_tolerance || v_usec < -time_tolerance) { 1659 pps_errcnt++; 1660 pps_shift = PPS_SHIFT; 1661 pps_intcnt = 0; 1662 time_status |= STA_PPSERROR; 1663 return; 1664 } 1665 1666 /* 1667 * A three-stage median filter is used to help deglitch the pps 1668 * frequency. The median sample becomes the frequency offset 1669 * estimate; the difference between the other two samples 1670 * becomes the frequency dispersion (stability) estimate. 1671 */ 1672 pps_ff[2] = pps_ff[1]; 1673 pps_ff[1] = pps_ff[0]; 1674 pps_ff[0] = v_usec; 1675 if (pps_ff[0] > pps_ff[1]) { 1676 if (pps_ff[1] > pps_ff[2]) { 1677 u_usec = pps_ff[1]; /* 0 1 2 */ 1678 v_usec = pps_ff[0] - pps_ff[2]; 1679 } else if (pps_ff[2] > pps_ff[0]) { 1680 u_usec = pps_ff[0]; /* 2 0 1 */ 1681 v_usec = pps_ff[2] - pps_ff[1]; 1682 } else { 1683 u_usec = pps_ff[2]; /* 0 2 1 */ 1684 v_usec = pps_ff[0] - pps_ff[1]; 1685 } 1686 } else { 1687 if (pps_ff[1] < pps_ff[2]) { 1688 u_usec = pps_ff[1]; /* 2 1 0 */ 1689 v_usec = pps_ff[2] - pps_ff[0]; 1690 } else if (pps_ff[2] < pps_ff[0]) { 1691 u_usec = pps_ff[0]; /* 1 0 2 */ 1692 v_usec = pps_ff[1] - pps_ff[2]; 1693 } else { 1694 u_usec = pps_ff[2]; /* 1 2 0 */ 1695 v_usec = pps_ff[1] - pps_ff[0]; 1696 } 1697 } 1698 1699 /* 1700 * Here the frequency dispersion (stability) is updated. If it 1701 * is less than one-fourth the maximum (MAXFREQ), the frequency 1702 * offset is updated as well, but clamped to the tolerance. It 1703 * will be processed later by the hardclock() routine. 1704 */ 1705 v_usec = (v_usec >> 1) - pps_stabil; 1706 if (v_usec < 0) 1707 pps_stabil -= -v_usec >> PPS_AVG; 1708 else 1709 pps_stabil += v_usec >> PPS_AVG; 1710 if (pps_stabil > MAXFREQ >> 2) { 1711 pps_stbcnt++; 1712 time_status |= STA_PPSWANDER; 1713 return; 1714 } 1715 if (time_status & STA_PPSFREQ) { 1716 if (u_usec < 0) { 1717 pps_freq -= -u_usec >> PPS_AVG; 1718 if (pps_freq < -time_tolerance) 1719 pps_freq = -time_tolerance; 1720 u_usec = -u_usec; 1721 } else { 1722 pps_freq += u_usec >> PPS_AVG; 1723 if (pps_freq > time_tolerance) 1724 pps_freq = time_tolerance; 1725 } 1726 } 1727 1728 /* 1729 * Here the calibration interval is adjusted. If the maximum 1730 * time difference is greater than tick / 4, reduce the interval 1731 * by half. If this is not the case for four consecutive 1732 * intervals, double the interval. 1733 */ 1734 if (u_usec << pps_shift > bigtick >> 2) { 1735 pps_intcnt = 0; 1736 if (pps_shift > PPS_SHIFT) 1737 pps_shift--; 1738 } else if (pps_intcnt >= 4) { 1739 pps_intcnt = 0; 1740 if (pps_shift < PPS_SHIFTMAX) 1741 pps_shift++; 1742 } else 1743 pps_intcnt++; 1744 } 1745 #endif /* PPS_SYNC */ 1746 #endif /* NTP */ 1747 1748 /* 1749 * Return information about system clocks. 1750 */ 1751 int 1752 sysctl_clockrate(void *where, size_t *sizep) 1753 { 1754 struct clockinfo clkinfo; 1755 1756 /* 1757 * Construct clockinfo structure. 1758 */ 1759 clkinfo.tick = tick; 1760 clkinfo.tickadj = tickadj; 1761 clkinfo.hz = hz; 1762 clkinfo.profhz = profhz; 1763 clkinfo.stathz = stathz ? stathz : hz; 1764 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo))); 1765 } 1766