kern_timeout.c revision 3308
1/*- 2 * Copyright (c) 1982, 1986, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 39 * $Id: kern_clock.c,v 1.8 1994/09/29 00:52:06 wollman Exp $ 40 */ 41 42/* Portions of this software are covered by the following: */ 43/****************************************************************************** 44 * * 45 * Copyright (c) David L. Mills 1993, 1994 * 46 * * 47 * Permission to use, copy, modify, and distribute this software and its * 48 * documentation for any purpose and without fee is hereby granted, provided * 49 * that the above copyright notice appears in all copies and that both the * 50 * copyright notice and this permission notice appear in supporting * 51 * documentation, and that the name University of Delaware not be used in * 52 * advertising or publicity pertaining to distribution of the software * 53 * without specific, written prior permission. The University of Delaware * 54 * makes no representations about the suitability this software for any * 55 * purpose. It is provided "as is" without express or implied warranty. * 56 * * 57 *****************************************************************************/ 58 59#include <sys/param.h> 60#include <sys/systm.h> 61#include <sys/dkstat.h> 62#include <sys/callout.h> 63#include <sys/kernel.h> 64#include <sys/proc.h> 65#include <sys/resourcevar.h> 66#include <sys/signalvar.h> 67#include <sys/timex.h> 68#include <vm/vm.h> 69#include <sys/sysctl.h> 70 71#include <machine/cpu.h> 72#include <machine/clock.h> 73 74#ifdef GPROF 75#include <sys/gmon.h> 76#endif 77 78/* Does anybody else really care about these? */ 79struct callout *callfree, *callout, calltodo; 80int ncallout; 81 82/* Some of these don't belong here, but it's easiest to concentrate them. */ 83long cp_time[CPUSTATES]; 84long dk_seek[DK_NDRIVE]; 85long dk_time[DK_NDRIVE]; 86long dk_wds[DK_NDRIVE]; 87long dk_wpms[DK_NDRIVE]; 88long dk_xfer[DK_NDRIVE]; 89 90int dk_busy; 91int dk_ndrive = DK_NDRIVE; 92 93long tk_cancc; 94long tk_nin; 95long tk_nout; 96long tk_rawcc; 97 98/* 99 * Clock handling routines. 100 * 101 * This code is written to operate with two timers that run independently of 102 * each other. The main clock, running hz times per second, is used to keep 103 * track of real time. The second timer handles kernel and user profiling, 104 * and does resource use estimation. If the second timer is programmable, 105 * it is randomized to avoid aliasing between the two clocks. For example, 106 * the randomization prevents an adversary from always giving up the cpu 107 * just before its quantum expires. Otherwise, it would never accumulate 108 * cpu ticks. The mean frequency of the second timer is stathz. 109 * 110 * If no second timer exists, stathz will be zero; in this case we drive 111 * profiling and statistics off the main clock. This WILL NOT be accurate; 112 * do not do it unless absolutely necessary. 113 * 114 * The statistics clock may (or may not) be run at a higher rate while 115 * profiling. This profile clock runs at profhz. We require that profhz 116 * be an integral multiple of stathz. 117 * 118 * If the statistics clock is running fast, it must be divided by the ratio 119 * profhz/stathz for statistics. (For profiling, every tick counts.) 120 */ 121 122/* 123 * TODO: 124 * allocate more timeout table slots when table overflows. 125 */ 126 127/* 128 * Bump a timeval by a small number of usec's. 129 */ 130#define BUMPTIME(t, usec) { \ 131 register volatile struct timeval *tp = (t); \ 132 register long us; \ 133 \ 134 tp->tv_usec = us = tp->tv_usec + (usec); \ 135 if (us >= 1000000) { \ 136 tp->tv_usec = us - 1000000; \ 137 tp->tv_sec++; \ 138 } \ 139} 140 141int stathz; 142int profhz; 143int profprocs; 144int ticks; 145static int psdiv, pscnt; /* prof => stat divider */ 146int psratio; /* ratio: prof / stat */ 147 148volatile struct timeval time; 149volatile struct timeval mono_time; 150 151/* 152 * Phase-lock loop (PLL) definitions 153 * 154 * The following variables are read and set by the ntp_adjtime() system 155 * call. 156 * 157 * time_state shows the state of the system clock, with values defined 158 * in the timex.h header file. 159 * 160 * time_status shows the status of the system clock, with bits defined 161 * in the timex.h header file. 162 * 163 * time_offset is used by the PLL to adjust the system time in small 164 * increments. 165 * 166 * time_constant determines the bandwidth or "stiffness" of the PLL. 167 * 168 * time_tolerance determines maximum frequency error or tolerance of the 169 * CPU clock oscillator and is a property of the architecture; however, 170 * in principle it could change as result of the presence of external 171 * discipline signals, for instance. 172 * 173 * time_precision is usually equal to the kernel tick variable; however, 174 * in cases where a precision clock counter or external clock is 175 * available, the resolution can be much less than this and depend on 176 * whether the external clock is working or not. 177 * 178 * time_maxerror is initialized by a ntp_adjtime() call and increased by 179 * the kernel once each second to reflect the maximum error 180 * bound growth. 181 * 182 * time_esterror is set and read by the ntp_adjtime() call, but 183 * otherwise not used by the kernel. 184 */ 185int time_status = STA_UNSYNC; /* clock status bits */ 186int time_state = TIME_OK; /* clock state */ 187long time_offset = 0; /* time offset (us) */ 188long time_constant = 0; /* pll time constant */ 189long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */ 190long time_precision = 1; /* clock precision (us) */ 191long time_maxerror = MAXPHASE; /* maximum error (us) */ 192long time_esterror = MAXPHASE; /* estimated error (us) */ 193 194/* 195 * The following variables establish the state of the PLL and the 196 * residual time and frequency offset of the local clock. The scale 197 * factors are defined in the timex.h header file. 198 * 199 * time_phase and time_freq are the phase increment and the frequency 200 * increment, respectively, of the kernel time variable at each tick of 201 * the clock. 202 * 203 * time_freq is set via ntp_adjtime() from a value stored in a file when 204 * the synchronization daemon is first started. Its value is retrieved 205 * via ntp_adjtime() and written to the file about once per hour by the 206 * daemon. 207 * 208 * time_adj is the adjustment added to the value of tick at each timer 209 * interrupt and is recomputed at each timer interrupt. 210 * 211 * time_reftime is the second's portion of the system time on the last 212 * call to ntp_adjtime(). It is used to adjust the time_freq variable 213 * and to increase the time_maxerror as the time since last update 214 * increases. 215 */ 216long time_phase = 0; /* phase offset (scaled us) */ 217long time_freq = 0; /* frequency offset (scaled ppm) */ 218long time_adj = 0; /* tick adjust (scaled 1 / hz) */ 219long time_reftime = 0; /* time at last adjustment (s) */ 220 221#ifdef PPS_SYNC 222/* 223 * The following variables are used only if the if the kernel PPS 224 * discipline code is configured (PPS_SYNC). The scale factors are 225 * defined in the timex.h header file. 226 * 227 * pps_time contains the time at each calibration interval, as read by 228 * microtime(). 229 * 230 * pps_offset is the time offset produced by the time median filter 231 * pps_tf[], while pps_jitter is the dispersion measured by this 232 * filter. 233 * 234 * pps_freq is the frequency offset produced by the frequency median 235 * filter pps_ff[], while pps_stabil is the dispersion measured by 236 * this filter. 237 * 238 * pps_usec is latched from a high resolution counter or external clock 239 * at pps_time. Here we want the hardware counter contents only, not the 240 * contents plus the time_tv.usec as usual. 241 * 242 * pps_valid counts the number of seconds since the last PPS update. It 243 * is used as a watchdog timer to disable the PPS discipline should the 244 * PPS signal be lost. 245 * 246 * pps_glitch counts the number of seconds since the beginning of an 247 * offset burst more than tick/2 from current nominal offset. It is used 248 * mainly to suppress error bursts due to priority conflicts between the 249 * PPS interrupt and timer interrupt. 250 * 251 * pps_count counts the seconds of the calibration interval, the 252 * duration of which is pps_shift in powers of two. 253 * 254 * pps_intcnt counts the calibration intervals for use in the interval- 255 * adaptation algorithm. It's just too complicated for words. 256 */ 257struct timeval pps_time; /* kernel time at last interval */ 258long pps_offset = 0; /* pps time offset (us) */ 259long pps_jitter = MAXTIME; /* pps time dispersion (jitter) (us) */ 260long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */ 261long pps_freq = 0; /* frequency offset (scaled ppm) */ 262long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */ 263long pps_ff[] = {0, 0, 0}; /* frequency offset median filter */ 264long pps_usec = 0; /* microsec counter at last interval */ 265long pps_valid = PPS_VALID; /* pps signal watchdog counter */ 266int pps_glitch = 0; /* pps signal glitch counter */ 267int pps_count = 0; /* calibration interval counter (s) */ 268int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */ 269int pps_intcnt = 0; /* intervals at current duration */ 270 271/* 272 * PPS signal quality monitors 273 * 274 * pps_jitcnt counts the seconds that have been discarded because the 275 * jitter measured by the time median filter exceeds the limit MAXTIME 276 * (100 us). 277 * 278 * pps_calcnt counts the frequency calibration intervals, which are 279 * variable from 4 s to 256 s. 280 * 281 * pps_errcnt counts the calibration intervals which have been discarded 282 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the 283 * calibration interval jitter exceeds two ticks. 284 * 285 * pps_stbcnt counts the calibration intervals that have been discarded 286 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us). 287 */ 288long pps_jitcnt = 0; /* jitter limit exceeded */ 289long pps_calcnt = 0; /* calibration intervals */ 290long pps_errcnt = 0; /* calibration errors */ 291long pps_stbcnt = 0; /* stability limit exceeded */ 292#endif /* PPS_SYNC */ 293 294/* XXX none of this stuff works under FreeBSD */ 295#ifdef EXT_CLOCK 296/* 297 * External clock definitions 298 * 299 * The following definitions and declarations are used only if an 300 * external clock (HIGHBALL or TPRO) is configured on the system. 301 */ 302#define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */ 303 304/* 305 * The clock_count variable is set to CLOCK_INTERVAL at each PPS 306 * interrupt and decremented once each second. 307 */ 308int clock_count = 0; /* CPU clock counter */ 309 310#ifdef HIGHBALL 311/* 312 * The clock_offset and clock_cpu variables are used by the HIGHBALL 313 * interface. The clock_offset variable defines the offset between 314 * system time and the HIGBALL counters. The clock_cpu variable contains 315 * the offset between the system clock and the HIGHBALL clock for use in 316 * disciplining the kernel time variable. 317 */ 318extern struct timeval clock_offset; /* Highball clock offset */ 319long clock_cpu = 0; /* CPU clock adjust */ 320#endif /* HIGHBALL */ 321#endif /* EXT_CLOCK */ 322 323/* 324 * hardupdate() - local clock update 325 * 326 * This routine is called by ntp_adjtime() to update the local clock 327 * phase and frequency. This is used to implement an adaptive-parameter, 328 * first-order, type-II phase-lock loop. The code computes new time and 329 * frequency offsets each time it is called. The hardclock() routine 330 * amortizes these offsets at each tick interrupt. If the kernel PPS 331 * discipline code is configured (PPS_SYNC), the PPS signal itself 332 * determines the new time offset, instead of the calling argument. 333 * Presumably, calls to ntp_adjtime() occur only when the caller 334 * believes the local clock is valid within some bound (+-128 ms with 335 * NTP). If the caller's time is far different than the PPS time, an 336 * argument will ensue, and it's not clear who will lose. 337 * 338 * For default SHIFT_UPDATE = 12, the offset is limited to +-512 ms, the 339 * maximum interval between updates is 4096 s and the maximum frequency 340 * offset is +-31.25 ms/s. 341 * 342 * Note: splclock() is in effect. 343 */ 344void 345hardupdate(offset) 346 long offset; 347{ 348 long ltemp, mtemp; 349 350 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME)) 351 return; 352 ltemp = offset; 353#ifdef PPS_SYNC 354 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL) 355 ltemp = pps_offset; 356#endif /* PPS_SYNC */ 357 if (ltemp > MAXPHASE) 358 time_offset = MAXPHASE << SHIFT_UPDATE; 359 else if (ltemp < -MAXPHASE) 360 time_offset = -(MAXPHASE << SHIFT_UPDATE); 361 else 362 time_offset = ltemp << SHIFT_UPDATE; 363 mtemp = time.tv_sec - time_reftime; 364 time_reftime = time.tv_sec; 365 if (mtemp > MAXSEC) 366 mtemp = 0; 367 368 /* ugly multiply should be replaced */ 369 if (ltemp < 0) 370 time_freq -= (-ltemp * mtemp) >> (time_constant + 371 time_constant + SHIFT_KF - SHIFT_USEC); 372 else 373 time_freq += (ltemp * mtemp) >> (time_constant + 374 time_constant + SHIFT_KF - SHIFT_USEC); 375 if (time_freq > time_tolerance) 376 time_freq = time_tolerance; 377 else if (time_freq < -time_tolerance) 378 time_freq = -time_tolerance; 379} 380 381 382 383/* 384 * Initialize clock frequencies and start both clocks running. 385 */ 386void 387initclocks() 388{ 389 register int i; 390 391 /* 392 * Set divisors to 1 (normal case) and let the machine-specific 393 * code do its bit. 394 */ 395 psdiv = pscnt = 1; 396 cpu_initclocks(); 397 398 /* 399 * Compute profhz/stathz, and fix profhz if needed. 400 */ 401 i = stathz ? stathz : hz; 402 if (profhz == 0) 403 profhz = i; 404 psratio = profhz / i; 405} 406 407/* 408 * The real-time timer, interrupting hz times per second. 409 */ 410void 411hardclock(frame) 412 register struct clockframe *frame; 413{ 414 register struct callout *p1; 415 register struct proc *p; 416 register int needsoft; 417 extern int tickdelta; 418 extern long timedelta; 419 420 /* 421 * Update real-time timeout queue. 422 * At front of queue are some number of events which are ``due''. 423 * The time to these is <= 0 and if negative represents the 424 * number of ticks which have passed since it was supposed to happen. 425 * The rest of the q elements (times > 0) are events yet to happen, 426 * where the time for each is given as a delta from the previous. 427 * Decrementing just the first of these serves to decrement the time 428 * to all events. 429 */ 430 needsoft = 0; 431 for (p1 = calltodo.c_next; p1 != NULL; p1 = p1->c_next) { 432 if (--p1->c_time > 0) 433 break; 434 needsoft = 1; 435 if (p1->c_time == 0) 436 break; 437 } 438 439 p = curproc; 440 if (p) { 441 register struct pstats *pstats; 442 443 /* 444 * Run current process's virtual and profile time, as needed. 445 */ 446 pstats = p->p_stats; 447 if (CLKF_USERMODE(frame) && 448 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 449 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 450 psignal(p, SIGVTALRM); 451 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 452 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 453 psignal(p, SIGPROF); 454 } 455 456 /* 457 * If no separate statistics clock is available, run it from here. 458 */ 459 if (stathz == 0) 460 statclock(frame); 461 462 /* 463 * Increment the time-of-day. 464 */ 465 ticks++; 466 { 467 int time_update; 468 struct timeval newtime = time; 469 long ltemp; 470 471 if (timedelta == 0) { 472 time_update = tick; 473 } else { 474 time_update = tick + tickdelta; 475 timedelta -= tickdelta; 476 } 477 BUMPTIME(&mono_time, time_update); 478 479 /* 480 * Compute the phase adjustment. If the low-order bits 481 * (time_phase) of the update overflow, bump the high-order bits 482 * (time_update). 483 */ 484 time_phase += time_adj; 485 if (time_phase <= -FINEUSEC) { 486 ltemp = -time_phase >> SHIFT_SCALE; 487 time_phase += ltemp << SHIFT_SCALE; 488 time_update -= ltemp; 489 } 490 else if (time_phase >= FINEUSEC) { 491 ltemp = time_phase >> SHIFT_SCALE; 492 time_phase -= ltemp << SHIFT_SCALE; 493 time_update += ltemp; 494 } 495 496 newtime.tv_usec += time_update; 497 /* 498 * On rollover of the second the phase adjustment to be used for 499 * the next second is calculated. Also, the maximum error is 500 * increased by the tolerance. If the PPS frequency discipline 501 * code is present, the phase is increased to compensate for the 502 * CPU clock oscillator frequency error. 503 * 504 * With SHIFT_SCALE = 23, the maximum frequency adjustment is 505 * +-256 us per tick, or 25.6 ms/s at a clock frequency of 100 506 * Hz. The time contribution is shifted right a minimum of two 507 * bits, while the frequency contribution is a right shift. 508 * Thus, overflow is prevented if the frequency contribution is 509 * limited to half the maximum or 15.625 ms/s. 510 */ 511 if (newtime.tv_usec >= 1000000) { 512 newtime.tv_usec -= 1000000; 513 newtime.tv_sec++; 514 time_maxerror += time_tolerance >> SHIFT_USEC; 515 if (time_offset < 0) { 516 ltemp = -time_offset >> 517 (SHIFT_KG + time_constant); 518 time_offset += ltemp; 519 time_adj = -ltemp << 520 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE); 521 } else { 522 ltemp = time_offset >> 523 (SHIFT_KG + time_constant); 524 time_offset -= ltemp; 525 time_adj = ltemp << 526 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE); 527 } 528#ifdef PPS_SYNC 529 /* 530 * Gnaw on the watchdog counter and update the frequency 531 * computed by the pll and the PPS signal. 532 */ 533 pps_valid++; 534 if (pps_valid == PPS_VALID) { 535 pps_jitter = MAXTIME; 536 pps_stabil = MAXFREQ; 537 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 538 STA_PPSWANDER | STA_PPSERROR); 539 } 540 ltemp = time_freq + pps_freq; 541#else 542 ltemp = time_freq; 543#endif /* PPS_SYNC */ 544 if (ltemp < 0) 545 time_adj -= -ltemp >> 546 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 547 else 548 time_adj += ltemp >> 549 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 550 551 /* 552 * When the CPU clock oscillator frequency is not a 553 * power of two in Hz, the SHIFT_HZ is only an 554 * approximate scale factor. In the SunOS kernel, this 555 * results in a PLL gain factor of 1/1.28 = 0.78 what it 556 * should be. In the following code the overall gain is 557 * increased by a factor of 1.25, which results in a 558 * residual error less than 3 percent. 559 */ 560 /* Same thing applies for FreeBSD --GAW */ 561 if (hz == 100) { 562 if (time_adj < 0) 563 time_adj -= -time_adj >> 2; 564 else 565 time_adj += time_adj >> 2; 566 } 567 568 /* XXX - this is really bogus, but can't be fixed until 569 xntpd's idea of the system clock is fixed to know how 570 the user wants leap seconds handled; in the mean time, 571 we assume that users of NTP are running without proper 572 leap second support (this is now the default anyway) */ 573 /* 574 * Leap second processing. If in leap-insert state at 575 * the end of the day, the system clock is set back one 576 * second; if in leap-delete state, the system clock is 577 * set ahead one second. The microtime() routine or 578 * external clock driver will insure that reported time 579 * is always monotonic. The ugly divides should be 580 * replaced. 581 */ 582 switch (time_state) { 583 584 case TIME_OK: 585 if (time_status & STA_INS) 586 time_state = TIME_INS; 587 else if (time_status & STA_DEL) 588 time_state = TIME_DEL; 589 break; 590 591 case TIME_INS: 592 if (newtime.tv_sec % 86400 == 0) { 593 newtime.tv_sec--; 594 time_state = TIME_OOP; 595 } 596 break; 597 598 case TIME_DEL: 599 if ((newtime.tv_sec + 1) % 86400 == 0) { 600 newtime.tv_sec++; 601 time_state = TIME_WAIT; 602 } 603 break; 604 605 case TIME_OOP: 606 time_state = TIME_WAIT; 607 break; 608 609 case TIME_WAIT: 610 if (!(time_status & (STA_INS | STA_DEL))) 611 time_state = TIME_OK; 612 } 613 } 614 CPU_CLOCKUPDATE(&time, &newtime); 615 } 616 617 /* 618 * Process callouts at a very low cpu priority, so we don't keep the 619 * relatively high clock interrupt priority any longer than necessary. 620 */ 621 if (needsoft) { 622 if (CLKF_BASEPRI(frame)) { 623 /* 624 * Save the overhead of a software interrupt; 625 * it will happen as soon as we return, so do it now. 626 */ 627 (void)splsoftclock(); 628 softclock(); 629 } else 630 setsoftclock(); 631 } 632} 633 634/* 635 * Software (low priority) clock interrupt. 636 * Run periodic events from timeout queue. 637 */ 638/*ARGSUSED*/ 639void 640softclock() 641{ 642 register struct callout *c; 643 register void *arg; 644 register void (*func) __P((void *)); 645 register int s; 646 647 s = splhigh(); 648 while ((c = calltodo.c_next) != NULL && c->c_time <= 0) { 649 func = c->c_func; 650 arg = c->c_arg; 651 calltodo.c_next = c->c_next; 652 c->c_next = callfree; 653 callfree = c; 654 splx(s); 655 (*func)(arg); 656 (void) splhigh(); 657 } 658 splx(s); 659} 660 661/* 662 * timeout -- 663 * Execute a function after a specified length of time. 664 * 665 * untimeout -- 666 * Cancel previous timeout function call. 667 * 668 * See AT&T BCI Driver Reference Manual for specification. This 669 * implementation differs from that one in that no identification 670 * value is returned from timeout, rather, the original arguments 671 * to timeout are used to identify entries for untimeout. 672 */ 673void 674timeout(ftn, arg, ticks) 675 timeout_t ftn; 676 void *arg; 677 register int ticks; 678{ 679 register struct callout *new, *p, *t; 680 register int s; 681 682 if (ticks <= 0) 683 ticks = 1; 684 685 /* Lock out the clock. */ 686 s = splhigh(); 687 688 /* Fill in the next free callout structure. */ 689 if (callfree == NULL) 690 panic("timeout table full"); 691 new = callfree; 692 callfree = new->c_next; 693 new->c_arg = arg; 694 new->c_func = ftn; 695 696 /* 697 * The time for each event is stored as a difference from the time 698 * of the previous event on the queue. Walk the queue, correcting 699 * the ticks argument for queue entries passed. Correct the ticks 700 * value for the queue entry immediately after the insertion point 701 * as well. Watch out for negative c_time values; these represent 702 * overdue events. 703 */ 704 for (p = &calltodo; 705 (t = p->c_next) != NULL && ticks > t->c_time; p = t) 706 if (t->c_time > 0) 707 ticks -= t->c_time; 708 new->c_time = ticks; 709 if (t != NULL) 710 t->c_time -= ticks; 711 712 /* Insert the new entry into the queue. */ 713 p->c_next = new; 714 new->c_next = t; 715 splx(s); 716} 717 718void 719untimeout(ftn, arg) 720 timeout_t ftn; 721 void *arg; 722{ 723 register struct callout *p, *t; 724 register int s; 725 726 s = splhigh(); 727 for (p = &calltodo; (t = p->c_next) != NULL; p = t) 728 if (t->c_func == ftn && t->c_arg == arg) { 729 /* Increment next entry's tick count. */ 730 if (t->c_next && t->c_time > 0) 731 t->c_next->c_time += t->c_time; 732 733 /* Move entry from callout queue to callfree queue. */ 734 p->c_next = t->c_next; 735 t->c_next = callfree; 736 callfree = t; 737 break; 738 } 739 splx(s); 740} 741 742/* 743 * Compute number of hz until specified time. Used to 744 * compute third argument to timeout() from an absolute time. 745 */ 746int 747hzto(tv) 748 struct timeval *tv; 749{ 750 register long ticks, sec; 751 int s; 752 753 /* 754 * If number of milliseconds will fit in 32 bit arithmetic, 755 * then compute number of milliseconds to time and scale to 756 * ticks. Otherwise just compute number of hz in time, rounding 757 * times greater than representible to maximum value. 758 * 759 * Delta times less than 25 days can be computed ``exactly''. 760 * Maximum value for any timeout in 10ms ticks is 250 days. 761 */ 762 s = splhigh(); 763 sec = tv->tv_sec - time.tv_sec; 764 if (sec <= 0x7fffffff / 1000 - 1000) 765 ticks = ((tv->tv_sec - time.tv_sec) * 1000 + 766 (tv->tv_usec - time.tv_usec) / 1000) / (tick / 1000); 767 else if (sec <= 0x7fffffff / hz) 768 ticks = sec * hz; 769 else 770 ticks = 0x7fffffff; 771 splx(s); 772 return (ticks); 773} 774 775/* 776 * Start profiling on a process. 777 * 778 * Kernel profiling passes proc0 which never exits and hence 779 * keeps the profile clock running constantly. 780 */ 781void 782startprofclock(p) 783 register struct proc *p; 784{ 785 int s; 786 787 if ((p->p_flag & P_PROFIL) == 0) { 788 p->p_flag |= P_PROFIL; 789 if (++profprocs == 1 && stathz != 0) { 790 s = splstatclock(); 791 psdiv = pscnt = psratio; 792 setstatclockrate(profhz); 793 splx(s); 794 } 795 } 796} 797 798/* 799 * Stop profiling on a process. 800 */ 801void 802stopprofclock(p) 803 register struct proc *p; 804{ 805 int s; 806 807 if (p->p_flag & P_PROFIL) { 808 p->p_flag &= ~P_PROFIL; 809 if (--profprocs == 0 && stathz != 0) { 810 s = splstatclock(); 811 psdiv = pscnt = 1; 812 setstatclockrate(stathz); 813 splx(s); 814 } 815 } 816} 817 818/* 819 * Statistics clock. Grab profile sample, and if divider reaches 0, 820 * do process and kernel statistics. 821 */ 822void 823statclock(frame) 824 register struct clockframe *frame; 825{ 826#ifdef GPROF 827 register struct gmonparam *g; 828#endif 829 register struct proc *p = curproc; 830 register int i; 831 832 if (p) { 833 struct pstats *pstats; 834 struct rusage *ru; 835 struct vmspace *vm; 836 837 /* bump the resource usage of integral space use */ 838 if ((pstats = p->p_stats) && (ru = &pstats->p_ru) && (vm = p->p_vmspace)) { 839 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024; 840 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024; 841 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024; 842 if ((vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024) > 843 ru->ru_maxrss) { 844 ru->ru_maxrss = 845 vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024; 846 } 847 } 848 } 849 850 if (CLKF_USERMODE(frame)) { 851 if (p->p_flag & P_PROFIL) 852 addupc_intr(p, CLKF_PC(frame), 1); 853 if (--pscnt > 0) 854 return; 855 /* 856 * Came from user mode; CPU was in user state. 857 * If this process is being profiled record the tick. 858 */ 859 p->p_uticks++; 860 if (p->p_nice > NZERO) 861 cp_time[CP_NICE]++; 862 else 863 cp_time[CP_USER]++; 864 } else { 865#ifdef GPROF 866 /* 867 * Kernel statistics are just like addupc_intr, only easier. 868 */ 869 g = &_gmonparam; 870 if (g->state == GMON_PROF_ON) { 871 i = CLKF_PC(frame) - g->lowpc; 872 if (i < g->textsize) { 873 i /= HISTFRACTION * sizeof(*g->kcount); 874 g->kcount[i]++; 875 } 876 } 877#endif 878 if (--pscnt > 0) 879 return; 880 /* 881 * Came from kernel mode, so we were: 882 * - handling an interrupt, 883 * - doing syscall or trap work on behalf of the current 884 * user process, or 885 * - spinning in the idle loop. 886 * Whichever it is, charge the time as appropriate. 887 * Note that we charge interrupts to the current process, 888 * regardless of whether they are ``for'' that process, 889 * so that we know how much of its real time was spent 890 * in ``non-process'' (i.e., interrupt) work. 891 */ 892 if (CLKF_INTR(frame)) { 893 if (p != NULL) 894 p->p_iticks++; 895 cp_time[CP_INTR]++; 896 } else if (p != NULL) { 897 p->p_sticks++; 898 cp_time[CP_SYS]++; 899 } else 900 cp_time[CP_IDLE]++; 901 } 902 pscnt = psdiv; 903 904 /* 905 * We maintain statistics shown by user-level statistics 906 * programs: the amount of time in each cpu state, and 907 * the amount of time each of DK_NDRIVE ``drives'' is busy. 908 * 909 * XXX should either run linked list of drives, or (better) 910 * grab timestamps in the start & done code. 911 */ 912 for (i = 0; i < DK_NDRIVE; i++) 913 if (dk_busy & (1 << i)) 914 dk_time[i]++; 915 916 /* 917 * We adjust the priority of the current process. The priority of 918 * a process gets worse as it accumulates CPU time. The cpu usage 919 * estimator (p_estcpu) is increased here. The formula for computing 920 * priorities (in kern_synch.c) will compute a different value each 921 * time p_estcpu increases by 4. The cpu usage estimator ramps up 922 * quite quickly when the process is running (linearly), and decays 923 * away exponentially, at a rate which is proportionally slower when 924 * the system is busy. The basic principal is that the system will 925 * 90% forget that the process used a lot of CPU time in 5 * loadav 926 * seconds. This causes the system to favor processes which haven't 927 * run much recently, and to round-robin among other processes. 928 */ 929 if (p != NULL) { 930 p->p_cpticks++; 931 if (++p->p_estcpu == 0) 932 p->p_estcpu--; 933 if ((p->p_estcpu & 3) == 0) { 934 resetpriority(p); 935 if (p->p_priority >= PUSER) 936 p->p_priority = p->p_usrpri; 937 } 938 } 939} 940 941/* 942 * Return information about system clocks. 943 */ 944int 945sysctl_clockrate(where, sizep) 946 register char *where; 947 size_t *sizep; 948{ 949 struct clockinfo clkinfo; 950 951 /* 952 * Construct clockinfo structure. 953 */ 954 clkinfo.hz = hz; 955 clkinfo.tick = tick; 956 clkinfo.profhz = profhz; 957 clkinfo.stathz = stathz ? stathz : hz; 958 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo))); 959} 960 961/*#ifdef PPS_SYNC*/ 962#if 0 963/* This code is completely bogus; if anybody ever wants to use it, get 964 * the current version from Dave Mills. */ 965 966/* 967 * hardpps() - discipline CPU clock oscillator to external pps signal 968 * 969 * This routine is called at each PPS interrupt in order to discipline 970 * the CPU clock oscillator to the PPS signal. It integrates successive 971 * phase differences between the two oscillators and calculates the 972 * frequency offset. This is used in hardclock() to discipline the CPU 973 * clock oscillator so that intrinsic frequency error is cancelled out. 974 * The code requires the caller to capture the time and hardware 975 * counter value at the designated PPS signal transition. 976 */ 977void 978hardpps(tvp, usec) 979 struct timeval *tvp; /* time at PPS */ 980 long usec; /* hardware counter at PPS */ 981{ 982 long u_usec, v_usec, bigtick; 983 long cal_sec, cal_usec; 984 985 /* 986 * During the calibration interval adjust the starting time when 987 * the tick overflows. At the end of the interval compute the 988 * duration of the interval and the difference of the hardware 989 * counters at the beginning and end of the interval. This code 990 * is deliciously complicated by the fact valid differences may 991 * exceed the value of tick when using long calibration 992 * intervals and small ticks. Note that the counter can be 993 * greater than tick if caught at just the wrong instant, but 994 * the values returned and used here are correct. 995 */ 996 bigtick = (long)tick << SHIFT_USEC; 997 pps_usec -= ntp_pll.ybar; 998 if (pps_usec >= bigtick) 999 pps_usec -= bigtick; 1000 if (pps_usec < 0) 1001 pps_usec += bigtick; 1002 pps_time.tv_sec++; 1003 pps_count++; 1004 if (pps_count < (1 << pps_shift)) 1005 return; 1006 pps_count = 0; 1007 ntp_pll.calcnt++; 1008 u_usec = usec << SHIFT_USEC; 1009 v_usec = pps_usec - u_usec; 1010 if (v_usec >= bigtick >> 1) 1011 v_usec -= bigtick; 1012 if (v_usec < -(bigtick >> 1)) 1013 v_usec += bigtick; 1014 if (v_usec < 0) 1015 v_usec = -(-v_usec >> ntp_pll.shift); 1016 else 1017 v_usec = v_usec >> ntp_pll.shift; 1018 pps_usec = u_usec; 1019 cal_sec = tvp->tv_sec; 1020 cal_usec = tvp->tv_usec; 1021 cal_sec -= pps_time.tv_sec; 1022 cal_usec -= pps_time.tv_usec; 1023 if (cal_usec < 0) { 1024 cal_usec += 1000000; 1025 cal_sec--; 1026 } 1027 pps_time = *tvp; 1028 1029 /* 1030 * Check for lost interrupts, noise, excessive jitter and 1031 * excessive frequency error. The number of timer ticks during 1032 * the interval may vary +-1 tick. Add to this a margin of one 1033 * tick for the PPS signal jitter and maximum frequency 1034 * deviation. If the limits are exceeded, the calibration 1035 * interval is reset to the minimum and we start over. 1036 */ 1037 u_usec = (long)tick << 1; 1038 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1039 || (cal_sec == 0 && cal_usec < u_usec)) 1040 || v_usec > ntp_pll.tolerance || v_usec < -ntp_pll.tolerance) { 1041 ntp_pll.jitcnt++; 1042 ntp_pll.shift = NTP_PLL.SHIFT; 1043 pps_dispinc = PPS_DISPINC; 1044 ntp_pll.intcnt = 0; 1045 return; 1046 } 1047 1048 /* 1049 * A three-stage median filter is used to help deglitch the pps 1050 * signal. The median sample becomes the offset estimate; the 1051 * difference between the other two samples becomes the 1052 * dispersion estimate. 1053 */ 1054 pps_mf[2] = pps_mf[1]; 1055 pps_mf[1] = pps_mf[0]; 1056 pps_mf[0] = v_usec; 1057 if (pps_mf[0] > pps_mf[1]) { 1058 if (pps_mf[1] > pps_mf[2]) { 1059 u_usec = pps_mf[1]; /* 0 1 2 */ 1060 v_usec = pps_mf[0] - pps_mf[2]; 1061 } else if (pps_mf[2] > pps_mf[0]) { 1062 u_usec = pps_mf[0]; /* 2 0 1 */ 1063 v_usec = pps_mf[2] - pps_mf[1]; 1064 } else { 1065 u_usec = pps_mf[2]; /* 0 2 1 */ 1066 v_usec = pps_mf[0] - pps_mf[1]; 1067 } 1068 } else { 1069 if (pps_mf[1] < pps_mf[2]) { 1070 u_usec = pps_mf[1]; /* 2 1 0 */ 1071 v_usec = pps_mf[2] - pps_mf[0]; 1072 } else if (pps_mf[2] < pps_mf[0]) { 1073 u_usec = pps_mf[0]; /* 1 0 2 */ 1074 v_usec = pps_mf[1] - pps_mf[2]; 1075 } else { 1076 u_usec = pps_mf[2]; /* 1 2 0 */ 1077 v_usec = pps_mf[1] - pps_mf[0]; 1078 } 1079 } 1080 1081 /* 1082 * Here the dispersion average is updated. If it is less than 1083 * the threshold pps_dispmax, the frequency average is updated 1084 * as well, but clamped to the tolerance. 1085 */ 1086 v_usec = (v_usec >> 1) - ntp_pll.disp; 1087 if (v_usec < 0) 1088 ntp_pll.disp -= -v_usec >> PPS_AVG; 1089 else 1090 ntp_pll.disp += v_usec >> PPS_AVG; 1091 if (ntp_pll.disp > pps_dispmax) { 1092 ntp_pll.discnt++; 1093 return; 1094 } 1095 if (u_usec < 0) { 1096 ntp_pll.ybar -= -u_usec >> PPS_AVG; 1097 if (ntp_pll.ybar < -ntp_pll.tolerance) 1098 ntp_pll.ybar = -ntp_pll.tolerance; 1099 u_usec = -u_usec; 1100 } else { 1101 ntp_pll.ybar += u_usec >> PPS_AVG; 1102 if (ntp_pll.ybar > ntp_pll.tolerance) 1103 ntp_pll.ybar = ntp_pll.tolerance; 1104 } 1105 1106 /* 1107 * Here the calibration interval is adjusted. If the maximum 1108 * time difference is greater than tick/4, reduce the interval 1109 * by half. If this is not the case for four consecutive 1110 * intervals, double the interval. 1111 */ 1112 if (u_usec << ntp_pll.shift > bigtick >> 2) { 1113 ntp_pll.intcnt = 0; 1114 if (ntp_pll.shift > NTP_PLL.SHIFT) { 1115 ntp_pll.shift--; 1116 pps_dispinc <<= 1; 1117 } 1118 } else if (ntp_pll.intcnt >= 4) { 1119 ntp_pll.intcnt = 0; 1120 if (ntp_pll.shift < NTP_PLL.SHIFTMAX) { 1121 ntp_pll.shift++; 1122 pps_dispinc >>= 1; 1123 } 1124 } else 1125 ntp_pll.intcnt++; 1126} 1127#endif /* PPS_SYNC */ 1128