kern_tc.c revision 9759
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 1.13 1995/05/30 08:05:20 rgrimes 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; 80 81/* Some of these don't belong here, but it's easiest to concentrate them. */ 82long cp_time[CPUSTATES]; 83long dk_seek[DK_NDRIVE]; 84long dk_time[DK_NDRIVE]; 85long dk_wds[DK_NDRIVE]; 86long dk_wpms[DK_NDRIVE]; 87long dk_xfer[DK_NDRIVE]; 88 89int dk_busy; 90int dk_ndrive = 0; 91char dk_names[DK_NDRIVE][DK_NAMELEN]; 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; 417 418 /* 419 * Update real-time timeout queue. 420 * At front of queue are some number of events which are ``due''. 421 * The time to these is <= 0 and if negative represents the 422 * number of ticks which have passed since it was supposed to happen. 423 * The rest of the q elements (times > 0) are events yet to happen, 424 * where the time for each is given as a delta from the previous. 425 * Decrementing just the first of these serves to decrement the time 426 * to all events. 427 */ 428 needsoft = 0; 429 for (p1 = calltodo.c_next; p1 != NULL; p1 = p1->c_next) { 430 if (--p1->c_time > 0) 431 break; 432 needsoft = 1; 433 if (p1->c_time == 0) 434 break; 435 } 436 437 p = curproc; 438 if (p) { 439 register struct pstats *pstats; 440 441 /* 442 * Run current process's virtual and profile time, as needed. 443 */ 444 pstats = p->p_stats; 445 if (CLKF_USERMODE(frame) && 446 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 447 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 448 psignal(p, SIGVTALRM); 449 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 450 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 451 psignal(p, SIGPROF); 452 } 453 454 /* 455 * If no separate statistics clock is available, run it from here. 456 */ 457 if (stathz == 0) 458 statclock(frame); 459 460 /* 461 * Increment the time-of-day. 462 */ 463 ticks++; 464 { 465 int time_update; 466 struct timeval newtime = time; 467 long ltemp; 468 469 if (timedelta == 0) { 470 time_update = tick; 471 } else { 472 time_update = tick + tickdelta; 473 timedelta -= tickdelta; 474 } 475 BUMPTIME(&mono_time, time_update); 476 477 /* 478 * Compute the phase adjustment. If the low-order bits 479 * (time_phase) of the update overflow, bump the high-order bits 480 * (time_update). 481 */ 482 time_phase += time_adj; 483 if (time_phase <= -FINEUSEC) { 484 ltemp = -time_phase >> SHIFT_SCALE; 485 time_phase += ltemp << SHIFT_SCALE; 486 time_update -= ltemp; 487 } 488 else if (time_phase >= FINEUSEC) { 489 ltemp = time_phase >> SHIFT_SCALE; 490 time_phase -= ltemp << SHIFT_SCALE; 491 time_update += ltemp; 492 } 493 494 newtime.tv_usec += time_update; 495 /* 496 * On rollover of the second the phase adjustment to be used for 497 * the next second is calculated. Also, the maximum error is 498 * increased by the tolerance. If the PPS frequency discipline 499 * code is present, the phase is increased to compensate for the 500 * CPU clock oscillator frequency error. 501 * 502 * With SHIFT_SCALE = 23, the maximum frequency adjustment is 503 * +-256 us per tick, or 25.6 ms/s at a clock frequency of 100 504 * Hz. The time contribution is shifted right a minimum of two 505 * bits, while the frequency contribution is a right shift. 506 * Thus, overflow is prevented if the frequency contribution is 507 * limited to half the maximum or 15.625 ms/s. 508 */ 509 if (newtime.tv_usec >= 1000000) { 510 newtime.tv_usec -= 1000000; 511 newtime.tv_sec++; 512 time_maxerror += time_tolerance >> SHIFT_USEC; 513 if (time_offset < 0) { 514 ltemp = -time_offset >> 515 (SHIFT_KG + time_constant); 516 time_offset += ltemp; 517 time_adj = -ltemp << 518 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE); 519 } else { 520 ltemp = time_offset >> 521 (SHIFT_KG + time_constant); 522 time_offset -= ltemp; 523 time_adj = ltemp << 524 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE); 525 } 526#ifdef PPS_SYNC 527 /* 528 * Gnaw on the watchdog counter and update the frequency 529 * computed by the pll and the PPS signal. 530 */ 531 pps_valid++; 532 if (pps_valid == PPS_VALID) { 533 pps_jitter = MAXTIME; 534 pps_stabil = MAXFREQ; 535 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 536 STA_PPSWANDER | STA_PPSERROR); 537 } 538 ltemp = time_freq + pps_freq; 539#else 540 ltemp = time_freq; 541#endif /* PPS_SYNC */ 542 if (ltemp < 0) 543 time_adj -= -ltemp >> 544 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 545 else 546 time_adj += ltemp >> 547 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 548 549 /* 550 * When the CPU clock oscillator frequency is not a 551 * power of two in Hz, the SHIFT_HZ is only an 552 * approximate scale factor. In the SunOS kernel, this 553 * results in a PLL gain factor of 1/1.28 = 0.78 what it 554 * should be. In the following code the overall gain is 555 * increased by a factor of 1.25, which results in a 556 * residual error less than 3 percent. 557 */ 558 /* Same thing applies for FreeBSD --GAW */ 559 if (hz == 100) { 560 if (time_adj < 0) 561 time_adj -= -time_adj >> 2; 562 else 563 time_adj += time_adj >> 2; 564 } 565 566 /* XXX - this is really bogus, but can't be fixed until 567 xntpd's idea of the system clock is fixed to know how 568 the user wants leap seconds handled; in the mean time, 569 we assume that users of NTP are running without proper 570 leap second support (this is now the default anyway) */ 571 /* 572 * Leap second processing. If in leap-insert state at 573 * the end of the day, the system clock is set back one 574 * second; if in leap-delete state, the system clock is 575 * set ahead one second. The microtime() routine or 576 * external clock driver will insure that reported time 577 * is always monotonic. The ugly divides should be 578 * replaced. 579 */ 580 switch (time_state) { 581 582 case TIME_OK: 583 if (time_status & STA_INS) 584 time_state = TIME_INS; 585 else if (time_status & STA_DEL) 586 time_state = TIME_DEL; 587 break; 588 589 case TIME_INS: 590 if (newtime.tv_sec % 86400 == 0) { 591 newtime.tv_sec--; 592 time_state = TIME_OOP; 593 } 594 break; 595 596 case TIME_DEL: 597 if ((newtime.tv_sec + 1) % 86400 == 0) { 598 newtime.tv_sec++; 599 time_state = TIME_WAIT; 600 } 601 break; 602 603 case TIME_OOP: 604 time_state = TIME_WAIT; 605 break; 606 607 case TIME_WAIT: 608 if (!(time_status & (STA_INS | STA_DEL))) 609 time_state = TIME_OK; 610 } 611 } 612 CPU_CLOCKUPDATE(&time, &newtime); 613 } 614 615 /* 616 * Process callouts at a very low cpu priority, so we don't keep the 617 * relatively high clock interrupt priority any longer than necessary. 618 */ 619 if (needsoft) { 620 if (CLKF_BASEPRI(frame)) { 621 /* 622 * Save the overhead of a software interrupt; 623 * it will happen as soon as we return, so do it now. 624 */ 625 (void)splsoftclock(); 626 softclock(); 627 } else 628 setsoftclock(); 629 } 630} 631 632/* 633 * Software (low priority) clock interrupt. 634 * Run periodic events from timeout queue. 635 */ 636/*ARGSUSED*/ 637void 638softclock() 639{ 640 register struct callout *c; 641 register void *arg; 642 register void (*func) __P((void *)); 643 register int s; 644 645 s = splhigh(); 646 while ((c = calltodo.c_next) != NULL && c->c_time <= 0) { 647 func = c->c_func; 648 arg = c->c_arg; 649 calltodo.c_next = c->c_next; 650 c->c_next = callfree; 651 callfree = c; 652 splx(s); 653 (*func)(arg); 654 (void) splhigh(); 655 } 656 splx(s); 657} 658 659/* 660 * timeout -- 661 * Execute a function after a specified length of time. 662 * 663 * untimeout -- 664 * Cancel previous timeout function call. 665 * 666 * See AT&T BCI Driver Reference Manual for specification. This 667 * implementation differs from that one in that no identification 668 * value is returned from timeout, rather, the original arguments 669 * to timeout are used to identify entries for untimeout. 670 */ 671void 672timeout(ftn, arg, ticks) 673 timeout_t ftn; 674 void *arg; 675 register int ticks; 676{ 677 register struct callout *new, *p, *t; 678 register int s; 679 680 if (ticks <= 0) 681 ticks = 1; 682 683 /* Lock out the clock. */ 684 s = splhigh(); 685 686 /* Fill in the next free callout structure. */ 687 if (callfree == NULL) 688 panic("timeout table full"); 689 new = callfree; 690 callfree = new->c_next; 691 new->c_arg = arg; 692 new->c_func = ftn; 693 694 /* 695 * The time for each event is stored as a difference from the time 696 * of the previous event on the queue. Walk the queue, correcting 697 * the ticks argument for queue entries passed. Correct the ticks 698 * value for the queue entry immediately after the insertion point 699 * as well. Watch out for negative c_time values; these represent 700 * overdue events. 701 */ 702 for (p = &calltodo; 703 (t = p->c_next) != NULL && ticks > t->c_time; p = t) 704 if (t->c_time > 0) 705 ticks -= t->c_time; 706 new->c_time = ticks; 707 if (t != NULL) 708 t->c_time -= ticks; 709 710 /* Insert the new entry into the queue. */ 711 p->c_next = new; 712 new->c_next = t; 713 splx(s); 714} 715 716void 717untimeout(ftn, arg) 718 timeout_t ftn; 719 void *arg; 720{ 721 register struct callout *p, *t; 722 register int s; 723 724 s = splhigh(); 725 for (p = &calltodo; (t = p->c_next) != NULL; p = t) 726 if (t->c_func == ftn && t->c_arg == arg) { 727 /* Increment next entry's tick count. */ 728 if (t->c_next && t->c_time > 0) 729 t->c_next->c_time += t->c_time; 730 731 /* Move entry from callout queue to callfree queue. */ 732 p->c_next = t->c_next; 733 t->c_next = callfree; 734 callfree = t; 735 break; 736 } 737 splx(s); 738} 739 740/* 741 * Compute number of hz until specified time. Used to 742 * compute third argument to timeout() from an absolute time. 743 */ 744int 745hzto(tv) 746 struct timeval *tv; 747{ 748 register unsigned long ticks; 749 register long sec, usec; 750 int s; 751 752 /* 753 * If the number of usecs in the whole seconds part of the time 754 * difference fits in a long, then the total number of usecs will 755 * fit in an unsigned long. Compute the total and convert it to 756 * ticks, rounding up and adding 1 to allow for the current tick 757 * to expire. Rounding also depends on unsigned long arithmetic 758 * to avoid overflow. 759 * 760 * Otherwise, if the number of ticks in the whole seconds part of 761 * the time difference fits in a long, then convert the parts to 762 * ticks separately and add, using similar rounding methods and 763 * overflow avoidance. This method would work in the previous 764 * case but it is slightly slower and assumes that hz is integral. 765 * 766 * Otherwise, round the time difference down to the maximum 767 * representable value. 768 * 769 * If ints have 32 bits, then the maximum value for any timeout in 770 * 10ms ticks is 248 days. 771 */ 772 s = splclock(); 773 sec = tv->tv_sec - time.tv_sec; 774 usec = tv->tv_usec - time.tv_usec; 775 splx(s); 776 if (usec < 0) { 777 sec--; 778 usec += 1000000; 779 } 780 if (sec < 0) { 781#ifdef DIAGNOSTIC 782 printf("hzto: negative time difference %ld sec %ld usec\n", 783 sec, usec); 784#endif 785 ticks = 1; 786 } else if (sec <= LONG_MAX / 1000000) 787 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 788 / tick + 1; 789 else if (sec <= LONG_MAX / hz) 790 ticks = sec * hz 791 + ((unsigned long)usec + (tick - 1)) / tick + 1; 792 else 793 ticks = LONG_MAX; 794 if (ticks > INT_MAX) 795 ticks = INT_MAX; 796 return (ticks); 797} 798 799/* 800 * Start profiling on a process. 801 * 802 * Kernel profiling passes proc0 which never exits and hence 803 * keeps the profile clock running constantly. 804 */ 805void 806startprofclock(p) 807 register struct proc *p; 808{ 809 int s; 810 811 if ((p->p_flag & P_PROFIL) == 0) { 812 p->p_flag |= P_PROFIL; 813 if (++profprocs == 1 && stathz != 0) { 814 s = splstatclock(); 815 psdiv = pscnt = psratio; 816 setstatclockrate(profhz); 817 splx(s); 818 } 819 } 820} 821 822/* 823 * Stop profiling on a process. 824 */ 825void 826stopprofclock(p) 827 register struct proc *p; 828{ 829 int s; 830 831 if (p->p_flag & P_PROFIL) { 832 p->p_flag &= ~P_PROFIL; 833 if (--profprocs == 0 && stathz != 0) { 834 s = splstatclock(); 835 psdiv = pscnt = 1; 836 setstatclockrate(stathz); 837 splx(s); 838 } 839 } 840} 841 842/* 843 * Statistics clock. Grab profile sample, and if divider reaches 0, 844 * do process and kernel statistics. 845 */ 846void 847statclock(frame) 848 register struct clockframe *frame; 849{ 850#ifdef GPROF 851 register struct gmonparam *g; 852#endif 853 register struct proc *p = curproc; 854 register int i; 855 856 if (p) { 857 struct pstats *pstats; 858 struct rusage *ru; 859 struct vmspace *vm; 860 861 /* bump the resource usage of integral space use */ 862 if ((pstats = p->p_stats) && (ru = &pstats->p_ru) && (vm = p->p_vmspace)) { 863 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024; 864 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024; 865 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024; 866 if ((vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024) > 867 ru->ru_maxrss) { 868 ru->ru_maxrss = 869 vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024; 870 } 871 } 872 } 873 874 if (CLKF_USERMODE(frame)) { 875 if (p->p_flag & P_PROFIL) 876 addupc_intr(p, CLKF_PC(frame), 1); 877 if (--pscnt > 0) 878 return; 879 /* 880 * Came from user mode; CPU was in user state. 881 * If this process is being profiled record the tick. 882 */ 883 p->p_uticks++; 884 if (p->p_nice > NZERO) 885 cp_time[CP_NICE]++; 886 else 887 cp_time[CP_USER]++; 888 } else { 889#ifdef GPROF 890 /* 891 * Kernel statistics are just like addupc_intr, only easier. 892 */ 893 g = &_gmonparam; 894 if (g->state == GMON_PROF_ON) { 895 i = CLKF_PC(frame) - g->lowpc; 896 if (i < g->textsize) { 897 i /= HISTFRACTION * sizeof(*g->kcount); 898 g->kcount[i]++; 899 } 900 } 901#endif 902 if (--pscnt > 0) 903 return; 904 /* 905 * Came from kernel mode, so we were: 906 * - handling an interrupt, 907 * - doing syscall or trap work on behalf of the current 908 * user process, or 909 * - spinning in the idle loop. 910 * Whichever it is, charge the time as appropriate. 911 * Note that we charge interrupts to the current process, 912 * regardless of whether they are ``for'' that process, 913 * so that we know how much of its real time was spent 914 * in ``non-process'' (i.e., interrupt) work. 915 */ 916 if (CLKF_INTR(frame)) { 917 if (p != NULL) 918 p->p_iticks++; 919 cp_time[CP_INTR]++; 920 } else if (p != NULL) { 921 p->p_sticks++; 922 cp_time[CP_SYS]++; 923 } else 924 cp_time[CP_IDLE]++; 925 } 926 pscnt = psdiv; 927 928 /* 929 * We maintain statistics shown by user-level statistics 930 * programs: the amount of time in each cpu state, and 931 * the amount of time each of DK_NDRIVE ``drives'' is busy. 932 * 933 * XXX should either run linked list of drives, or (better) 934 * grab timestamps in the start & done code. 935 */ 936 for (i = 0; i < DK_NDRIVE; i++) 937 if (dk_busy & (1 << i)) 938 dk_time[i]++; 939 940 /* 941 * We adjust the priority of the current process. The priority of 942 * a process gets worse as it accumulates CPU time. The cpu usage 943 * estimator (p_estcpu) is increased here. The formula for computing 944 * priorities (in kern_synch.c) will compute a different value each 945 * time p_estcpu increases by 4. The cpu usage estimator ramps up 946 * quite quickly when the process is running (linearly), and decays 947 * away exponentially, at a rate which is proportionally slower when 948 * the system is busy. The basic principal is that the system will 949 * 90% forget that the process used a lot of CPU time in 5 * loadav 950 * seconds. This causes the system to favor processes which haven't 951 * run much recently, and to round-robin among other processes. 952 */ 953 if (p != NULL) { 954 p->p_cpticks++; 955 if (++p->p_estcpu == 0) 956 p->p_estcpu--; 957 if ((p->p_estcpu & 3) == 0) { 958 resetpriority(p); 959 if (p->p_priority >= PUSER) 960 p->p_priority = p->p_usrpri; 961 } 962 } 963} 964 965/* 966 * Return information about system clocks. 967 */ 968int 969sysctl_clockrate(where, sizep) 970 register char *where; 971 size_t *sizep; 972{ 973 struct clockinfo clkinfo; 974 975 /* 976 * Construct clockinfo structure. 977 */ 978 clkinfo.hz = hz; 979 clkinfo.tick = tick; 980 clkinfo.profhz = profhz; 981 clkinfo.stathz = stathz ? stathz : hz; 982 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo))); 983} 984 985/*#ifdef PPS_SYNC*/ 986#if 0 987/* This code is completely bogus; if anybody ever wants to use it, get 988 * the current version from Dave Mills. */ 989 990/* 991 * hardpps() - discipline CPU clock oscillator to external pps signal 992 * 993 * This routine is called at each PPS interrupt in order to discipline 994 * the CPU clock oscillator to the PPS signal. It integrates successive 995 * phase differences between the two oscillators and calculates the 996 * frequency offset. This is used in hardclock() to discipline the CPU 997 * clock oscillator so that intrinsic frequency error is cancelled out. 998 * The code requires the caller to capture the time and hardware 999 * counter value at the designated PPS signal transition. 1000 */ 1001void 1002hardpps(tvp, usec) 1003 struct timeval *tvp; /* time at PPS */ 1004 long usec; /* hardware counter at PPS */ 1005{ 1006 long u_usec, v_usec, bigtick; 1007 long cal_sec, cal_usec; 1008 1009 /* 1010 * During the calibration interval adjust the starting time when 1011 * the tick overflows. At the end of the interval compute the 1012 * duration of the interval and the difference of the hardware 1013 * counters at the beginning and end of the interval. This code 1014 * is deliciously complicated by the fact valid differences may 1015 * exceed the value of tick when using long calibration 1016 * intervals and small ticks. Note that the counter can be 1017 * greater than tick if caught at just the wrong instant, but 1018 * the values returned and used here are correct. 1019 */ 1020 bigtick = (long)tick << SHIFT_USEC; 1021 pps_usec -= ntp_pll.ybar; 1022 if (pps_usec >= bigtick) 1023 pps_usec -= bigtick; 1024 if (pps_usec < 0) 1025 pps_usec += bigtick; 1026 pps_time.tv_sec++; 1027 pps_count++; 1028 if (pps_count < (1 << pps_shift)) 1029 return; 1030 pps_count = 0; 1031 ntp_pll.calcnt++; 1032 u_usec = usec << SHIFT_USEC; 1033 v_usec = pps_usec - u_usec; 1034 if (v_usec >= bigtick >> 1) 1035 v_usec -= bigtick; 1036 if (v_usec < -(bigtick >> 1)) 1037 v_usec += bigtick; 1038 if (v_usec < 0) 1039 v_usec = -(-v_usec >> ntp_pll.shift); 1040 else 1041 v_usec = v_usec >> ntp_pll.shift; 1042 pps_usec = u_usec; 1043 cal_sec = tvp->tv_sec; 1044 cal_usec = tvp->tv_usec; 1045 cal_sec -= pps_time.tv_sec; 1046 cal_usec -= pps_time.tv_usec; 1047 if (cal_usec < 0) { 1048 cal_usec += 1000000; 1049 cal_sec--; 1050 } 1051 pps_time = *tvp; 1052 1053 /* 1054 * Check for lost interrupts, noise, excessive jitter and 1055 * excessive frequency error. The number of timer ticks during 1056 * the interval may vary +-1 tick. Add to this a margin of one 1057 * tick for the PPS signal jitter and maximum frequency 1058 * deviation. If the limits are exceeded, the calibration 1059 * interval is reset to the minimum and we start over. 1060 */ 1061 u_usec = (long)tick << 1; 1062 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1063 || (cal_sec == 0 && cal_usec < u_usec)) 1064 || v_usec > ntp_pll.tolerance || v_usec < -ntp_pll.tolerance) { 1065 ntp_pll.jitcnt++; 1066 ntp_pll.shift = NTP_PLL.SHIFT; 1067 pps_dispinc = PPS_DISPINC; 1068 ntp_pll.intcnt = 0; 1069 return; 1070 } 1071 1072 /* 1073 * A three-stage median filter is used to help deglitch the pps 1074 * signal. The median sample becomes the offset estimate; the 1075 * difference between the other two samples becomes the 1076 * dispersion estimate. 1077 */ 1078 pps_mf[2] = pps_mf[1]; 1079 pps_mf[1] = pps_mf[0]; 1080 pps_mf[0] = v_usec; 1081 if (pps_mf[0] > pps_mf[1]) { 1082 if (pps_mf[1] > pps_mf[2]) { 1083 u_usec = pps_mf[1]; /* 0 1 2 */ 1084 v_usec = pps_mf[0] - pps_mf[2]; 1085 } else if (pps_mf[2] > pps_mf[0]) { 1086 u_usec = pps_mf[0]; /* 2 0 1 */ 1087 v_usec = pps_mf[2] - pps_mf[1]; 1088 } else { 1089 u_usec = pps_mf[2]; /* 0 2 1 */ 1090 v_usec = pps_mf[0] - pps_mf[1]; 1091 } 1092 } else { 1093 if (pps_mf[1] < pps_mf[2]) { 1094 u_usec = pps_mf[1]; /* 2 1 0 */ 1095 v_usec = pps_mf[2] - pps_mf[0]; 1096 } else if (pps_mf[2] < pps_mf[0]) { 1097 u_usec = pps_mf[0]; /* 1 0 2 */ 1098 v_usec = pps_mf[1] - pps_mf[2]; 1099 } else { 1100 u_usec = pps_mf[2]; /* 1 2 0 */ 1101 v_usec = pps_mf[1] - pps_mf[0]; 1102 } 1103 } 1104 1105 /* 1106 * Here the dispersion average is updated. If it is less than 1107 * the threshold pps_dispmax, the frequency average is updated 1108 * as well, but clamped to the tolerance. 1109 */ 1110 v_usec = (v_usec >> 1) - ntp_pll.disp; 1111 if (v_usec < 0) 1112 ntp_pll.disp -= -v_usec >> PPS_AVG; 1113 else 1114 ntp_pll.disp += v_usec >> PPS_AVG; 1115 if (ntp_pll.disp > pps_dispmax) { 1116 ntp_pll.discnt++; 1117 return; 1118 } 1119 if (u_usec < 0) { 1120 ntp_pll.ybar -= -u_usec >> PPS_AVG; 1121 if (ntp_pll.ybar < -ntp_pll.tolerance) 1122 ntp_pll.ybar = -ntp_pll.tolerance; 1123 u_usec = -u_usec; 1124 } else { 1125 ntp_pll.ybar += u_usec >> PPS_AVG; 1126 if (ntp_pll.ybar > ntp_pll.tolerance) 1127 ntp_pll.ybar = ntp_pll.tolerance; 1128 } 1129 1130 /* 1131 * Here the calibration interval is adjusted. If the maximum 1132 * time difference is greater than tick/4, reduce the interval 1133 * by half. If this is not the case for four consecutive 1134 * intervals, double the interval. 1135 */ 1136 if (u_usec << ntp_pll.shift > bigtick >> 2) { 1137 ntp_pll.intcnt = 0; 1138 if (ntp_pll.shift > NTP_PLL.SHIFT) { 1139 ntp_pll.shift--; 1140 pps_dispinc <<= 1; 1141 } 1142 } else if (ntp_pll.intcnt >= 4) { 1143 ntp_pll.intcnt = 0; 1144 if (ntp_pll.shift < NTP_PLL.SHIFTMAX) { 1145 ntp_pll.shift++; 1146 pps_dispinc >>= 1; 1147 } 1148 } else 1149 ntp_pll.intcnt++; 1150} 1151#endif /* PPS_SYNC */ 1152