kern_timeout.c revision 29179
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 1.39 1997/09/02 20:05:37 bde 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 "opt_cpu.h" /* XXX */ 60 61#include <sys/param.h> 62#include <sys/systm.h> 63#include <sys/dkstat.h> 64#include <sys/callout.h> 65#include <sys/kernel.h> 66#include <sys/proc.h> 67#include <sys/resourcevar.h> 68#include <sys/signalvar.h> 69#include <sys/timex.h> 70#include <vm/vm.h> 71#include <sys/lock.h> 72#include <vm/pmap.h> 73#include <vm/vm_map.h> 74#include <sys/sysctl.h> 75 76#include <machine/cpu.h> 77#define CLOCK_HAIR /* XXX */ 78#include <machine/clock.h> 79#include <machine/limits.h> 80 81#ifdef GPROF 82#include <sys/gmon.h> 83#endif 84 85static void initclocks __P((void *dummy)); 86SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 87 88/* Exported to machdep.c. */ 89struct callout *callfree, *callout; 90 91static struct callout calltodo; 92 93/* Some of these don't belong here, but it's easiest to concentrate them. */ 94static long cp_time[CPUSTATES]; 95long dk_seek[DK_NDRIVE]; 96static long dk_time[DK_NDRIVE]; /* time busy (in statclock ticks) */ 97long dk_wds[DK_NDRIVE]; 98long dk_wpms[DK_NDRIVE]; 99long dk_xfer[DK_NDRIVE]; 100 101int dk_busy; 102int dk_ndrive = 0; 103char dk_names[DK_NDRIVE][DK_NAMELEN]; 104 105long tk_cancc; 106long tk_nin; 107long tk_nout; 108long tk_rawcc; 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/* 135 * TODO: 136 * allocate more timeout table slots when table overflows. 137 */ 138 139/* 140 * Bump a timeval by a small number of usec's. 141 */ 142#define BUMPTIME(t, usec) { \ 143 register volatile struct timeval *tp = (t); \ 144 register long us; \ 145 \ 146 tp->tv_usec = us = tp->tv_usec + (usec); \ 147 if (us >= 1000000) { \ 148 tp->tv_usec = us - 1000000; \ 149 tp->tv_sec++; \ 150 } \ 151} 152 153int stathz; 154int profhz; 155static int profprocs; 156int ticks; 157static int psdiv, pscnt; /* prof => stat divider */ 158int psratio; /* ratio: prof / stat */ 159 160volatile struct timeval time; 161volatile struct timeval mono_time; 162 163/* 164 * Phase/frequency-lock loop (PLL/FLL) definitions 165 * 166 * The following variables are read and set by the ntp_adjtime() system 167 * call. 168 * 169 * time_state shows the state of the system clock, with values defined 170 * in the timex.h header file. 171 * 172 * time_status shows the status of the system clock, with bits defined 173 * in the timex.h header file. 174 * 175 * time_offset is used by the PLL/FLL to adjust the system time in small 176 * increments. 177 * 178 * time_constant determines the bandwidth or "stiffness" of the PLL. 179 * 180 * time_tolerance determines maximum frequency error or tolerance of the 181 * CPU clock oscillator and is a property of the architecture; however, 182 * in principle it could change as result of the presence of external 183 * discipline signals, for instance. 184 * 185 * time_precision is usually equal to the kernel tick variable; however, 186 * in cases where a precision clock counter or external clock is 187 * available, the resolution can be much less than this and depend on 188 * whether the external clock is working or not. 189 * 190 * time_maxerror is initialized by a ntp_adjtime() call and increased by 191 * the kernel once each second to reflect the maximum error 192 * bound growth. 193 * 194 * time_esterror is set and read by the ntp_adjtime() call, but 195 * otherwise not used by the kernel. 196 */ 197int time_status = STA_UNSYNC; /* clock status bits */ 198int time_state = TIME_OK; /* clock state */ 199long time_offset = 0; /* time offset (us) */ 200long time_constant = 0; /* pll time constant */ 201long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */ 202long time_precision = 1; /* clock precision (us) */ 203long time_maxerror = MAXPHASE; /* maximum error (us) */ 204long time_esterror = MAXPHASE; /* estimated error (us) */ 205 206/* 207 * The following variables establish the state of the PLL/FLL and the 208 * residual time and frequency offset of the local clock. The scale 209 * factors are defined in the timex.h header file. 210 * 211 * time_phase and time_freq are the phase increment and the frequency 212 * increment, respectively, of the kernel time variable at each tick of 213 * the clock. 214 * 215 * time_freq is set via ntp_adjtime() from a value stored in a file when 216 * the synchronization daemon is first started. Its value is retrieved 217 * via ntp_adjtime() and written to the file about once per hour by the 218 * daemon. 219 * 220 * time_adj is the adjustment added to the value of tick at each timer 221 * interrupt and is recomputed from time_phase and time_freq at each 222 * seconds rollover. 223 * 224 * time_reftime is the second's portion of the system time on the last 225 * call to ntp_adjtime(). It is used to adjust the time_freq variable 226 * and to increase the time_maxerror as the time since last update 227 * increases. 228 */ 229static long time_phase = 0; /* phase offset (scaled us) */ 230long time_freq = 0; /* frequency offset (scaled ppm) */ 231static long time_adj = 0; /* tick adjust (scaled 1 / hz) */ 232static long time_reftime = 0; /* time at last adjustment (s) */ 233 234#ifdef PPS_SYNC 235/* 236 * The following variables are used only if the kernel PPS discipline 237 * code is configured (PPS_SYNC). The scale factors are defined in the 238 * timex.h header file. 239 * 240 * pps_time contains the time at each calibration interval, as read by 241 * microtime(). pps_count counts the seconds of the calibration 242 * interval, the duration of which is nominally pps_shift in powers of 243 * two. 244 * 245 * pps_offset is the time offset produced by the time median filter 246 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by 247 * this filter. 248 * 249 * pps_freq is the frequency offset produced by the frequency median 250 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured 251 * by this filter. 252 * 253 * pps_usec is latched from a high resolution counter or external clock 254 * at pps_time. Here we want the hardware counter contents only, not the 255 * contents plus the time_tv.usec as usual. 256 * 257 * pps_valid counts the number of seconds since the last PPS update. It 258 * is used as a watchdog timer to disable the PPS discipline should the 259 * PPS signal be lost. 260 * 261 * pps_glitch counts the number of seconds since the beginning of an 262 * offset burst more than tick/2 from current nominal offset. It is used 263 * mainly to suppress error bursts due to priority conflicts between the 264 * PPS interrupt and timer interrupt. 265 * 266 * pps_intcnt counts the calibration intervals for use in the interval- 267 * adaptation algorithm. It's just too complicated for words. 268 */ 269struct timeval pps_time; /* kernel time at last interval */ 270long pps_offset = 0; /* pps time offset (us) */ 271long pps_jitter = MAXTIME; /* pps time dispersion (jitter) (us) */ 272long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */ 273long pps_freq = 0; /* frequency offset (scaled ppm) */ 274long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */ 275long pps_ff[] = {0, 0, 0}; /* frequency offset median filter */ 276long pps_usec = 0; /* microsec counter at last interval */ 277long pps_valid = PPS_VALID; /* pps signal watchdog counter */ 278int pps_glitch = 0; /* pps signal glitch counter */ 279int pps_count = 0; /* calibration interval counter (s) */ 280int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */ 281int pps_intcnt = 0; /* intervals at current duration */ 282 283/* 284 * PPS signal quality monitors 285 * 286 * pps_jitcnt counts the seconds that have been discarded because the 287 * jitter measured by the time median filter exceeds the limit MAXTIME 288 * (100 us). 289 * 290 * pps_calcnt counts the frequency calibration intervals, which are 291 * variable from 4 s to 256 s. 292 * 293 * pps_errcnt counts the calibration intervals which have been discarded 294 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the 295 * calibration interval jitter exceeds two ticks. 296 * 297 * pps_stbcnt counts the calibration intervals that have been discarded 298 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us). 299 */ 300long pps_jitcnt = 0; /* jitter limit exceeded */ 301long pps_calcnt = 0; /* calibration intervals */ 302long pps_errcnt = 0; /* calibration errors */ 303long pps_stbcnt = 0; /* stability limit exceeded */ 304#endif /* PPS_SYNC */ 305 306/* XXX none of this stuff works under FreeBSD */ 307#ifdef EXT_CLOCK 308/* 309 * External clock definitions 310 * 311 * The following definitions and declarations are used only if an 312 * external clock (HIGHBALL or TPRO) is configured on the system. 313 */ 314#define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */ 315 316/* 317 * The clock_count variable is set to CLOCK_INTERVAL at each PPS 318 * interrupt and decremented once each second. 319 */ 320int clock_count = 0; /* CPU clock counter */ 321 322#ifdef HIGHBALL 323/* 324 * The clock_offset and clock_cpu variables are used by the HIGHBALL 325 * interface. The clock_offset variable defines the offset between 326 * system time and the HIGBALL counters. The clock_cpu variable contains 327 * the offset between the system clock and the HIGHBALL clock for use in 328 * disciplining the kernel time variable. 329 */ 330extern struct timeval clock_offset; /* Highball clock offset */ 331long clock_cpu = 0; /* CPU clock adjust */ 332#endif /* HIGHBALL */ 333#endif /* EXT_CLOCK */ 334 335/* 336 * hardupdate() - local clock update 337 * 338 * This routine is called by ntp_adjtime() to update the local clock 339 * phase and frequency. The implementation is of an adaptive-parameter, 340 * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new 341 * time and frequency offset estimates for each call. If the kernel PPS 342 * discipline code is configured (PPS_SYNC), the PPS signal itself 343 * determines the new time offset, instead of the calling argument. 344 * Presumably, calls to ntp_adjtime() occur only when the caller 345 * believes the local clock is valid within some bound (+-128 ms with 346 * NTP). If the caller's time is far different than the PPS time, an 347 * argument will ensue, and it's not clear who will lose. 348 * 349 * For uncompensated quartz crystal oscillatores and nominal update 350 * intervals less than 1024 s, operation should be in phase-lock mode 351 * (STA_FLL = 0), where the loop is disciplined to phase. For update 352 * intervals greater than thiss, operation should be in frequency-lock 353 * mode (STA_FLL = 1), where the loop is disciplined to frequency. 354 * 355 * Note: splclock() is in effect. 356 */ 357void 358hardupdate(offset) 359 long offset; 360{ 361 long ltemp, mtemp; 362 363 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME)) 364 return; 365 ltemp = offset; 366#ifdef PPS_SYNC 367 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL) 368 ltemp = pps_offset; 369#endif /* PPS_SYNC */ 370 371 /* 372 * Scale the phase adjustment and clamp to the operating range. 373 */ 374 if (ltemp > MAXPHASE) 375 time_offset = MAXPHASE << SHIFT_UPDATE; 376 else if (ltemp < -MAXPHASE) 377 time_offset = -(MAXPHASE << SHIFT_UPDATE); 378 else 379 time_offset = ltemp << SHIFT_UPDATE; 380 381 /* 382 * Select whether the frequency is to be controlled and in which 383 * mode (PLL or FLL). Clamp to the operating range. Ugly 384 * multiply/divide should be replaced someday. 385 */ 386 if (time_status & STA_FREQHOLD || time_reftime == 0) 387 time_reftime = time.tv_sec; 388 mtemp = time.tv_sec - time_reftime; 389 time_reftime = time.tv_sec; 390 if (time_status & STA_FLL) { 391 if (mtemp >= MINSEC) { 392 ltemp = ((time_offset / mtemp) << (SHIFT_USEC - 393 SHIFT_UPDATE)); 394 if (ltemp < 0) 395 time_freq -= -ltemp >> SHIFT_KH; 396 else 397 time_freq += ltemp >> SHIFT_KH; 398 } 399 } else { 400 if (mtemp < MAXSEC) { 401 ltemp *= mtemp; 402 if (ltemp < 0) 403 time_freq -= -ltemp >> (time_constant + 404 time_constant + SHIFT_KF - 405 SHIFT_USEC); 406 else 407 time_freq += ltemp >> (time_constant + 408 time_constant + SHIFT_KF - 409 SHIFT_USEC); 410 } 411 } 412 if (time_freq > time_tolerance) 413 time_freq = time_tolerance; 414 else if (time_freq < -time_tolerance) 415 time_freq = -time_tolerance; 416} 417 418 419 420/* 421 * Initialize clock frequencies and start both clocks running. 422 */ 423/* ARGSUSED*/ 424static void 425initclocks(dummy) 426 void *dummy; 427{ 428 register int i; 429 430 /* 431 * Set divisors to 1 (normal case) and let the machine-specific 432 * code do its bit. 433 */ 434 psdiv = pscnt = 1; 435 cpu_initclocks(); 436 437 /* 438 * Compute profhz/stathz, and fix profhz if needed. 439 */ 440 i = stathz ? stathz : hz; 441 if (profhz == 0) 442 profhz = i; 443 psratio = profhz / i; 444} 445 446/* 447 * The real-time timer, interrupting hz times per second. 448 */ 449void 450hardclock(frame) 451 register struct clockframe *frame; 452{ 453 register struct callout *p1; 454 register struct proc *p; 455 register int needsoft; 456 457 /* 458 * Update real-time timeout queue. 459 * At front of queue are some number of events which are ``due''. 460 * The time to these is <= 0 and if negative represents the 461 * number of ticks which have passed since it was supposed to happen. 462 * The rest of the q elements (times > 0) are events yet to happen, 463 * where the time for each is given as a delta from the previous. 464 * Decrementing just the first of these serves to decrement the time 465 * to all events. 466 */ 467 needsoft = 0; 468 for (p1 = calltodo.c_next; p1 != NULL; p1 = p1->c_next) { 469 if (--p1->c_time > 0) 470 break; 471 needsoft = 1; 472 if (p1->c_time == 0) 473 break; 474 } 475 476 p = curproc; 477 if (p) { 478 register struct pstats *pstats; 479 480 /* 481 * Run current process's virtual and profile time, as needed. 482 */ 483 pstats = p->p_stats; 484 if (CLKF_USERMODE(frame) && 485 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 486 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 487 psignal(p, SIGVTALRM); 488 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 489 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 490 psignal(p, SIGPROF); 491 } 492 493 /* 494 * If no separate statistics clock is available, run it from here. 495 */ 496 if (stathz == 0) 497 statclock(frame); 498 499 /* 500 * Increment the time-of-day. 501 */ 502 ticks++; 503 { 504 int time_update; 505 struct timeval newtime = time; 506 long ltemp; 507 508 if (timedelta == 0) { 509 time_update = CPU_THISTICKLEN(tick); 510 } else { 511 time_update = CPU_THISTICKLEN(tick) + tickdelta; 512 timedelta -= tickdelta; 513 } 514 BUMPTIME(&mono_time, time_update); 515 516 /* 517 * Compute the phase adjustment. If the low-order bits 518 * (time_phase) of the update overflow, bump the high-order bits 519 * (time_update). 520 */ 521 time_phase += time_adj; 522 if (time_phase <= -FINEUSEC) { 523 ltemp = -time_phase >> SHIFT_SCALE; 524 time_phase += ltemp << SHIFT_SCALE; 525 time_update -= ltemp; 526 } 527 else if (time_phase >= FINEUSEC) { 528 ltemp = time_phase >> SHIFT_SCALE; 529 time_phase -= ltemp << SHIFT_SCALE; 530 time_update += ltemp; 531 } 532 533 newtime.tv_usec += time_update; 534 /* 535 * On rollover of the second the phase adjustment to be used for 536 * the next second is calculated. Also, the maximum error is 537 * increased by the tolerance. If the PPS frequency discipline 538 * code is present, the phase is increased to compensate for the 539 * CPU clock oscillator frequency error. 540 * 541 * On a 32-bit machine and given parameters in the timex.h 542 * header file, the maximum phase adjustment is +-512 ms and 543 * maximum frequency offset is a tad less than) +-512 ppm. On a 544 * 64-bit machine, you shouldn't need to ask. 545 */ 546 if (newtime.tv_usec >= 1000000) { 547 newtime.tv_usec -= 1000000; 548 newtime.tv_sec++; 549 time_maxerror += time_tolerance >> SHIFT_USEC; 550 551 /* 552 * Compute the phase adjustment for the next second. In 553 * PLL mode, the offset is reduced by a fixed factor 554 * times the time constant. In FLL mode the offset is 555 * used directly. In either mode, the maximum phase 556 * adjustment for each second is clamped so as to spread 557 * the adjustment over not more than the number of 558 * seconds between updates. 559 */ 560 if (time_offset < 0) { 561 ltemp = -time_offset; 562 if (!(time_status & STA_FLL)) 563 ltemp >>= SHIFT_KG + time_constant; 564 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 565 ltemp = (MAXPHASE / MINSEC) << 566 SHIFT_UPDATE; 567 time_offset += ltemp; 568 time_adj = -ltemp << (SHIFT_SCALE - SHIFT_HZ - 569 SHIFT_UPDATE); 570 } else { 571 ltemp = time_offset; 572 if (!(time_status & STA_FLL)) 573 ltemp >>= SHIFT_KG + time_constant; 574 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 575 ltemp = (MAXPHASE / MINSEC) << 576 SHIFT_UPDATE; 577 time_offset -= ltemp; 578 time_adj = ltemp << (SHIFT_SCALE - SHIFT_HZ - 579 SHIFT_UPDATE); 580 } 581 582 /* 583 * Compute the frequency estimate and additional phase 584 * adjustment due to frequency error for the next 585 * second. When the PPS signal is engaged, gnaw on the 586 * watchdog counter and update the frequency computed by 587 * the pll and the PPS signal. 588 */ 589#ifdef PPS_SYNC 590 pps_valid++; 591 if (pps_valid == PPS_VALID) { 592 pps_jitter = MAXTIME; 593 pps_stabil = MAXFREQ; 594 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 595 STA_PPSWANDER | STA_PPSERROR); 596 } 597 ltemp = time_freq + pps_freq; 598#else 599 ltemp = time_freq; 600#endif /* PPS_SYNC */ 601 if (ltemp < 0) 602 time_adj -= -ltemp >> 603 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 604 else 605 time_adj += ltemp >> 606 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 607 608#if SHIFT_HZ == 7 609 /* 610 * When the CPU clock oscillator frequency is not a 611 * power of two in Hz, the SHIFT_HZ is only an 612 * approximate scale factor. In the SunOS kernel, this 613 * results in a PLL gain factor of 1/1.28 = 0.78 what it 614 * should be. In the following code the overall gain is 615 * increased by a factor of 1.25, which results in a 616 * residual error less than 3 percent. 617 */ 618 /* Same thing applies for FreeBSD --GAW */ 619 if (hz == 100) { 620 if (time_adj < 0) 621 time_adj -= -time_adj >> 2; 622 else 623 time_adj += time_adj >> 2; 624 } 625#endif /* SHIFT_HZ */ 626 627 /* XXX - this is really bogus, but can't be fixed until 628 xntpd's idea of the system clock is fixed to know how 629 the user wants leap seconds handled; in the mean time, 630 we assume that users of NTP are running without proper 631 leap second support (this is now the default anyway) */ 632 /* 633 * Leap second processing. If in leap-insert state at 634 * the end of the day, the system clock is set back one 635 * second; if in leap-delete state, the system clock is 636 * set ahead one second. The microtime() routine or 637 * external clock driver will insure that reported time 638 * is always monotonic. The ugly divides should be 639 * replaced. 640 */ 641 switch (time_state) { 642 643 case TIME_OK: 644 if (time_status & STA_INS) 645 time_state = TIME_INS; 646 else if (time_status & STA_DEL) 647 time_state = TIME_DEL; 648 break; 649 650 case TIME_INS: 651 if (newtime.tv_sec % 86400 == 0) { 652 newtime.tv_sec--; 653 time_state = TIME_OOP; 654 } 655 break; 656 657 case TIME_DEL: 658 if ((newtime.tv_sec + 1) % 86400 == 0) { 659 newtime.tv_sec++; 660 time_state = TIME_WAIT; 661 } 662 break; 663 664 case TIME_OOP: 665 time_state = TIME_WAIT; 666 break; 667 668 case TIME_WAIT: 669 if (!(time_status & (STA_INS | STA_DEL))) 670 time_state = TIME_OK; 671 } 672 } 673 CPU_CLOCKUPDATE(&time, &newtime); 674 } 675 676 /* 677 * Process callouts at a very low cpu priority, so we don't keep the 678 * relatively high clock interrupt priority any longer than necessary. 679 */ 680 if (needsoft) { 681 if (CLKF_BASEPRI(frame)) { 682 /* 683 * Save the overhead of a software interrupt; 684 * it will happen as soon as we return, so do it now. 685 */ 686 (void)splsoftclock(); 687 softclock(); 688 } else 689 setsoftclock(); 690 } 691} 692 693/* 694 * Software (low priority) clock interrupt. 695 * Run periodic events from timeout queue. 696 */ 697/*ARGSUSED*/ 698void 699softclock() 700{ 701 register struct callout *c; 702 register void *arg; 703 register void (*func) __P((void *)); 704 register int s; 705 706 s = splhigh(); 707 while ((c = calltodo.c_next) != NULL && c->c_time <= 0) { 708 func = c->c_func; 709 arg = c->c_arg; 710 calltodo.c_next = c->c_next; 711 c->c_next = callfree; 712 callfree = c; 713 splx(s); 714 (*func)(arg); 715 (void) splhigh(); 716 } 717 splx(s); 718} 719 720/* 721 * timeout -- 722 * Execute a function after a specified length of time. 723 * 724 * untimeout -- 725 * Cancel previous timeout function call. 726 * 727 * See AT&T BCI Driver Reference Manual for specification. This 728 * implementation differs from that one in that no identification 729 * value is returned from timeout, rather, the original arguments 730 * to timeout are used to identify entries for untimeout. 731 */ 732void 733timeout(ftn, arg, ticks) 734 timeout_t ftn; 735 void *arg; 736 register int ticks; 737{ 738 register struct callout *new, *p, *t; 739 register int s; 740 741 if (ticks <= 0) 742 ticks = 1; 743 744 /* Lock out the clock. */ 745 s = splhigh(); 746 747 /* Fill in the next free callout structure. */ 748 if (callfree == NULL) 749 panic("timeout table full"); 750 new = callfree; 751 callfree = new->c_next; 752 new->c_arg = arg; 753 new->c_func = ftn; 754 755 /* 756 * The time for each event is stored as a difference from the time 757 * of the previous event on the queue. Walk the queue, correcting 758 * the ticks argument for queue entries passed. Correct the ticks 759 * value for the queue entry immediately after the insertion point 760 * as well. Watch out for negative c_time values; these represent 761 * overdue events. 762 */ 763 for (p = &calltodo; 764 (t = p->c_next) != NULL && ticks > t->c_time; p = t) 765 if (t->c_time > 0) 766 ticks -= t->c_time; 767 new->c_time = ticks; 768 if (t != NULL) 769 t->c_time -= ticks; 770 771 /* Insert the new entry into the queue. */ 772 p->c_next = new; 773 new->c_next = t; 774 splx(s); 775} 776 777void 778untimeout(ftn, arg) 779 timeout_t ftn; 780 void *arg; 781{ 782 register struct callout *p, *t; 783 register int s; 784 785 s = splhigh(); 786 for (p = &calltodo; (t = p->c_next) != NULL; p = t) 787 if (t->c_func == ftn && t->c_arg == arg) { 788 /* Increment next entry's tick count. */ 789 if (t->c_next && t->c_time > 0) 790 t->c_next->c_time += t->c_time; 791 792 /* Move entry from callout queue to callfree queue. */ 793 p->c_next = t->c_next; 794 t->c_next = callfree; 795 callfree = t; 796 break; 797 } 798 splx(s); 799} 800 801void 802gettime(struct timeval *tvp) 803{ 804 int s; 805 806 s = splclock(); 807 /* XXX should use microtime() iff tv_usec is used. */ 808 *tvp = time; 809 splx(s); 810} 811 812/* 813 * Compute number of hz until specified time. Used to 814 * compute third argument to timeout() from an absolute time. 815 */ 816int 817hzto(tv) 818 struct timeval *tv; 819{ 820 register unsigned long ticks; 821 register long sec, usec; 822 int s; 823 824 /* 825 * If the number of usecs in the whole seconds part of the time 826 * difference fits in a long, then the total number of usecs will 827 * fit in an unsigned long. Compute the total and convert it to 828 * ticks, rounding up and adding 1 to allow for the current tick 829 * to expire. Rounding also depends on unsigned long arithmetic 830 * to avoid overflow. 831 * 832 * Otherwise, if the number of ticks in the whole seconds part of 833 * the time difference fits in a long, then convert the parts to 834 * ticks separately and add, using similar rounding methods and 835 * overflow avoidance. This method would work in the previous 836 * case but it is slightly slower and assumes that hz is integral. 837 * 838 * Otherwise, round the time difference down to the maximum 839 * representable value. 840 * 841 * If ints have 32 bits, then the maximum value for any timeout in 842 * 10ms ticks is 248 days. 843 */ 844 s = splclock(); 845 sec = tv->tv_sec - time.tv_sec; 846 usec = tv->tv_usec - time.tv_usec; 847 splx(s); 848 if (usec < 0) { 849 sec--; 850 usec += 1000000; 851 } 852 if (sec < 0) { 853#ifdef DIAGNOSTIC 854 printf("hzto: negative time difference %ld sec %ld usec\n", 855 sec, usec); 856#endif 857 ticks = 1; 858 } else if (sec <= LONG_MAX / 1000000) 859 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 860 / tick + 1; 861 else if (sec <= LONG_MAX / hz) 862 ticks = sec * hz 863 + ((unsigned long)usec + (tick - 1)) / tick + 1; 864 else 865 ticks = LONG_MAX; 866 if (ticks > INT_MAX) 867 ticks = INT_MAX; 868 return (ticks); 869} 870 871/* 872 * Start profiling on a process. 873 * 874 * Kernel profiling passes proc0 which never exits and hence 875 * keeps the profile clock running constantly. 876 */ 877void 878startprofclock(p) 879 register struct proc *p; 880{ 881 int s; 882 883 if ((p->p_flag & P_PROFIL) == 0) { 884 p->p_flag |= P_PROFIL; 885 if (++profprocs == 1 && stathz != 0) { 886 s = splstatclock(); 887 psdiv = pscnt = psratio; 888 setstatclockrate(profhz); 889 splx(s); 890 } 891 } 892} 893 894/* 895 * Stop profiling on a process. 896 */ 897void 898stopprofclock(p) 899 register struct proc *p; 900{ 901 int s; 902 903 if (p->p_flag & P_PROFIL) { 904 p->p_flag &= ~P_PROFIL; 905 if (--profprocs == 0 && stathz != 0) { 906 s = splstatclock(); 907 psdiv = pscnt = 1; 908 setstatclockrate(stathz); 909 splx(s); 910 } 911 } 912} 913 914/* 915 * Statistics clock. Grab profile sample, and if divider reaches 0, 916 * do process and kernel statistics. 917 */ 918void 919statclock(frame) 920 register struct clockframe *frame; 921{ 922#ifdef GPROF 923 register struct gmonparam *g; 924#endif 925 register struct proc *p; 926 register int i; 927 struct pstats *pstats; 928 long rss; 929 struct rusage *ru; 930 struct vmspace *vm; 931 932 if (CLKF_USERMODE(frame)) { 933 p = curproc; 934 if (p->p_flag & P_PROFIL) 935 addupc_intr(p, CLKF_PC(frame), 1); 936 if (--pscnt > 0) 937 return; 938 /* 939 * Came from user mode; CPU was in user state. 940 * If this process is being profiled record the tick. 941 */ 942 p->p_uticks++; 943 if (p->p_nice > NZERO) 944 cp_time[CP_NICE]++; 945 else 946 cp_time[CP_USER]++; 947 } else { 948#ifdef GPROF 949 /* 950 * Kernel statistics are just like addupc_intr, only easier. 951 */ 952 g = &_gmonparam; 953 if (g->state == GMON_PROF_ON) { 954 i = CLKF_PC(frame) - g->lowpc; 955 if (i < g->textsize) { 956 i /= HISTFRACTION * sizeof(*g->kcount); 957 g->kcount[i]++; 958 } 959 } 960#endif 961 if (--pscnt > 0) 962 return; 963 /* 964 * Came from kernel mode, so we were: 965 * - handling an interrupt, 966 * - doing syscall or trap work on behalf of the current 967 * user process, or 968 * - spinning in the idle loop. 969 * Whichever it is, charge the time as appropriate. 970 * Note that we charge interrupts to the current process, 971 * regardless of whether they are ``for'' that process, 972 * so that we know how much of its real time was spent 973 * in ``non-process'' (i.e., interrupt) work. 974 */ 975 p = curproc; 976 if (CLKF_INTR(frame)) { 977 if (p != NULL) 978 p->p_iticks++; 979 cp_time[CP_INTR]++; 980 } else if (p != NULL && !(p->p_flag & P_IDLEPROC)) { 981 p->p_sticks++; 982 cp_time[CP_SYS]++; 983 } else 984 cp_time[CP_IDLE]++; 985 } 986 pscnt = psdiv; 987 988 /* 989 * We maintain statistics shown by user-level statistics 990 * programs: the amount of time in each cpu state, and 991 * the amount of time each of DK_NDRIVE ``drives'' is busy. 992 * 993 * XXX should either run linked list of drives, or (better) 994 * grab timestamps in the start & done code. 995 */ 996 for (i = 0; i < DK_NDRIVE; i++) 997 if (dk_busy & (1 << i)) 998 dk_time[i]++; 999 1000 /* 1001 * We adjust the priority of the current process. The priority of 1002 * a process gets worse as it accumulates CPU time. The cpu usage 1003 * estimator (p_estcpu) is increased here. The formula for computing 1004 * priorities (in kern_synch.c) will compute a different value each 1005 * time p_estcpu increases by 4. The cpu usage estimator ramps up 1006 * quite quickly when the process is running (linearly), and decays 1007 * away exponentially, at a rate which is proportionally slower when 1008 * the system is busy. The basic principal is that the system will 1009 * 90% forget that the process used a lot of CPU time in 5 * loadav 1010 * seconds. This causes the system to favor processes which haven't 1011 * run much recently, and to round-robin among other processes. 1012 */ 1013 if (p != NULL) { 1014 p->p_cpticks++; 1015 if (++p->p_estcpu == 0) 1016 p->p_estcpu--; 1017 if ((p->p_estcpu & 3) == 0) { 1018 resetpriority(p); 1019 if (p->p_priority >= PUSER) 1020 p->p_priority = p->p_usrpri; 1021 } 1022 1023 /* Update resource usage integrals and maximums. */ 1024 if ((pstats = p->p_stats) != NULL && 1025 (ru = &pstats->p_ru) != NULL && 1026 (vm = p->p_vmspace) != NULL) { 1027 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024; 1028 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024; 1029 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024; 1030 rss = vm->vm_pmap.pm_stats.resident_count * 1031 PAGE_SIZE / 1024; 1032 if (ru->ru_maxrss < rss) 1033 ru->ru_maxrss = rss; 1034 } 1035 } 1036} 1037 1038/* 1039 * Return information about system clocks. 1040 */ 1041static int 1042sysctl_kern_clockrate SYSCTL_HANDLER_ARGS 1043{ 1044 struct clockinfo clkinfo; 1045 /* 1046 * Construct clockinfo structure. 1047 */ 1048 clkinfo.hz = hz; 1049 clkinfo.tick = tick; 1050 clkinfo.tickadj = tickadj; 1051 clkinfo.profhz = profhz; 1052 clkinfo.stathz = stathz ? stathz : hz; 1053 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 1054} 1055 1056SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 1057 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 1058 1059#ifdef PPS_SYNC 1060/* 1061 * hardpps() - discipline CPU clock oscillator to external PPS signal 1062 * 1063 * This routine is called at each PPS interrupt in order to discipline 1064 * the CPU clock oscillator to the PPS signal. It measures the PPS phase 1065 * and leaves it in a handy spot for the hardclock() routine. It 1066 * integrates successive PPS phase differences and calculates the 1067 * frequency offset. This is used in hardclock() to discipline the CPU 1068 * clock oscillator so that intrinsic frequency error is cancelled out. 1069 * The code requires the caller to capture the time and hardware counter 1070 * value at the on-time PPS signal transition. 1071 * 1072 * Note that, on some Unix systems, this routine runs at an interrupt 1073 * priority level higher than the timer interrupt routine hardclock(). 1074 * Therefore, the variables used are distinct from the hardclock() 1075 * variables, except for certain exceptions: The PPS frequency pps_freq 1076 * and phase pps_offset variables are determined by this routine and 1077 * updated atomically. The time_tolerance variable can be considered a 1078 * constant, since it is infrequently changed, and then only when the 1079 * PPS signal is disabled. The watchdog counter pps_valid is updated 1080 * once per second by hardclock() and is atomically cleared in this 1081 * routine. 1082 */ 1083void 1084hardpps(tvp, usec) 1085 struct timeval *tvp; /* time at PPS */ 1086 long usec; /* hardware counter at PPS */ 1087{ 1088 long u_usec, v_usec, bigtick; 1089 long cal_sec, cal_usec; 1090 1091 /* 1092 * An occasional glitch can be produced when the PPS interrupt 1093 * occurs in the hardclock() routine before the time variable is 1094 * updated. Here the offset is discarded when the difference 1095 * between it and the last one is greater than tick/2, but not 1096 * if the interval since the first discard exceeds 30 s. 1097 */ 1098 time_status |= STA_PPSSIGNAL; 1099 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR); 1100 pps_valid = 0; 1101 u_usec = -tvp->tv_usec; 1102 if (u_usec < -500000) 1103 u_usec += 1000000; 1104 v_usec = pps_offset - u_usec; 1105 if (v_usec < 0) 1106 v_usec = -v_usec; 1107 if (v_usec > (tick >> 1)) { 1108 if (pps_glitch > MAXGLITCH) { 1109 pps_glitch = 0; 1110 pps_tf[2] = u_usec; 1111 pps_tf[1] = u_usec; 1112 } else { 1113 pps_glitch++; 1114 u_usec = pps_offset; 1115 } 1116 } else 1117 pps_glitch = 0; 1118 1119 /* 1120 * A three-stage median filter is used to help deglitch the pps 1121 * time. The median sample becomes the time offset estimate; the 1122 * difference between the other two samples becomes the time 1123 * dispersion (jitter) estimate. 1124 */ 1125 pps_tf[2] = pps_tf[1]; 1126 pps_tf[1] = pps_tf[0]; 1127 pps_tf[0] = u_usec; 1128 if (pps_tf[0] > pps_tf[1]) { 1129 if (pps_tf[1] > pps_tf[2]) { 1130 pps_offset = pps_tf[1]; /* 0 1 2 */ 1131 v_usec = pps_tf[0] - pps_tf[2]; 1132 } else if (pps_tf[2] > pps_tf[0]) { 1133 pps_offset = pps_tf[0]; /* 2 0 1 */ 1134 v_usec = pps_tf[2] - pps_tf[1]; 1135 } else { 1136 pps_offset = pps_tf[2]; /* 0 2 1 */ 1137 v_usec = pps_tf[0] - pps_tf[1]; 1138 } 1139 } else { 1140 if (pps_tf[1] < pps_tf[2]) { 1141 pps_offset = pps_tf[1]; /* 2 1 0 */ 1142 v_usec = pps_tf[2] - pps_tf[0]; 1143 } else if (pps_tf[2] < pps_tf[0]) { 1144 pps_offset = pps_tf[0]; /* 1 0 2 */ 1145 v_usec = pps_tf[1] - pps_tf[2]; 1146 } else { 1147 pps_offset = pps_tf[2]; /* 1 2 0 */ 1148 v_usec = pps_tf[1] - pps_tf[0]; 1149 } 1150 } 1151 if (v_usec > MAXTIME) 1152 pps_jitcnt++; 1153 v_usec = (v_usec << PPS_AVG) - pps_jitter; 1154 if (v_usec < 0) 1155 pps_jitter -= -v_usec >> PPS_AVG; 1156 else 1157 pps_jitter += v_usec >> PPS_AVG; 1158 if (pps_jitter > (MAXTIME >> 1)) 1159 time_status |= STA_PPSJITTER; 1160 1161 /* 1162 * During the calibration interval adjust the starting time when 1163 * the tick overflows. At the end of the interval compute the 1164 * duration of the interval and the difference of the hardware 1165 * counters at the beginning and end of the interval. This code 1166 * is deliciously complicated by the fact valid differences may 1167 * exceed the value of tick when using long calibration 1168 * intervals and small ticks. Note that the counter can be 1169 * greater than tick if caught at just the wrong instant, but 1170 * the values returned and used here are correct. 1171 */ 1172 bigtick = (long)tick << SHIFT_USEC; 1173 pps_usec -= pps_freq; 1174 if (pps_usec >= bigtick) 1175 pps_usec -= bigtick; 1176 if (pps_usec < 0) 1177 pps_usec += bigtick; 1178 pps_time.tv_sec++; 1179 pps_count++; 1180 if (pps_count < (1 << pps_shift)) 1181 return; 1182 pps_count = 0; 1183 pps_calcnt++; 1184 u_usec = usec << SHIFT_USEC; 1185 v_usec = pps_usec - u_usec; 1186 if (v_usec >= bigtick >> 1) 1187 v_usec -= bigtick; 1188 if (v_usec < -(bigtick >> 1)) 1189 v_usec += bigtick; 1190 if (v_usec < 0) 1191 v_usec = -(-v_usec >> pps_shift); 1192 else 1193 v_usec = v_usec >> pps_shift; 1194 pps_usec = u_usec; 1195 cal_sec = tvp->tv_sec; 1196 cal_usec = tvp->tv_usec; 1197 cal_sec -= pps_time.tv_sec; 1198 cal_usec -= pps_time.tv_usec; 1199 if (cal_usec < 0) { 1200 cal_usec += 1000000; 1201 cal_sec--; 1202 } 1203 pps_time = *tvp; 1204 1205 /* 1206 * Check for lost interrupts, noise, excessive jitter and 1207 * excessive frequency error. The number of timer ticks during 1208 * the interval may vary +-1 tick. Add to this a margin of one 1209 * tick for the PPS signal jitter and maximum frequency 1210 * deviation. If the limits are exceeded, the calibration 1211 * interval is reset to the minimum and we start over. 1212 */ 1213 u_usec = (long)tick << 1; 1214 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1215 || (cal_sec == 0 && cal_usec < u_usec)) 1216 || v_usec > time_tolerance || v_usec < -time_tolerance) { 1217 pps_errcnt++; 1218 pps_shift = PPS_SHIFT; 1219 pps_intcnt = 0; 1220 time_status |= STA_PPSERROR; 1221 return; 1222 } 1223 1224 /* 1225 * A three-stage median filter is used to help deglitch the pps 1226 * frequency. The median sample becomes the frequency offset 1227 * estimate; the difference between the other two samples 1228 * becomes the frequency dispersion (stability) estimate. 1229 */ 1230 pps_ff[2] = pps_ff[1]; 1231 pps_ff[1] = pps_ff[0]; 1232 pps_ff[0] = v_usec; 1233 if (pps_ff[0] > pps_ff[1]) { 1234 if (pps_ff[1] > pps_ff[2]) { 1235 u_usec = pps_ff[1]; /* 0 1 2 */ 1236 v_usec = pps_ff[0] - pps_ff[2]; 1237 } else if (pps_ff[2] > pps_ff[0]) { 1238 u_usec = pps_ff[0]; /* 2 0 1 */ 1239 v_usec = pps_ff[2] - pps_ff[1]; 1240 } else { 1241 u_usec = pps_ff[2]; /* 0 2 1 */ 1242 v_usec = pps_ff[0] - pps_ff[1]; 1243 } 1244 } else { 1245 if (pps_ff[1] < pps_ff[2]) { 1246 u_usec = pps_ff[1]; /* 2 1 0 */ 1247 v_usec = pps_ff[2] - pps_ff[0]; 1248 } else if (pps_ff[2] < pps_ff[0]) { 1249 u_usec = pps_ff[0]; /* 1 0 2 */ 1250 v_usec = pps_ff[1] - pps_ff[2]; 1251 } else { 1252 u_usec = pps_ff[2]; /* 1 2 0 */ 1253 v_usec = pps_ff[1] - pps_ff[0]; 1254 } 1255 } 1256 1257 /* 1258 * Here the frequency dispersion (stability) is updated. If it 1259 * is less than one-fourth the maximum (MAXFREQ), the frequency 1260 * offset is updated as well, but clamped to the tolerance. It 1261 * will be processed later by the hardclock() routine. 1262 */ 1263 v_usec = (v_usec >> 1) - pps_stabil; 1264 if (v_usec < 0) 1265 pps_stabil -= -v_usec >> PPS_AVG; 1266 else 1267 pps_stabil += v_usec >> PPS_AVG; 1268 if (pps_stabil > MAXFREQ >> 2) { 1269 pps_stbcnt++; 1270 time_status |= STA_PPSWANDER; 1271 return; 1272 } 1273 if (time_status & STA_PPSFREQ) { 1274 if (u_usec < 0) { 1275 pps_freq -= -u_usec >> PPS_AVG; 1276 if (pps_freq < -time_tolerance) 1277 pps_freq = -time_tolerance; 1278 u_usec = -u_usec; 1279 } else { 1280 pps_freq += u_usec >> PPS_AVG; 1281 if (pps_freq > time_tolerance) 1282 pps_freq = time_tolerance; 1283 } 1284 } 1285 1286 /* 1287 * Here the calibration interval is adjusted. If the maximum 1288 * time difference is greater than tick / 4, reduce the interval 1289 * by half. If this is not the case for four consecutive 1290 * intervals, double the interval. 1291 */ 1292 if (u_usec << pps_shift > bigtick >> 2) { 1293 pps_intcnt = 0; 1294 if (pps_shift > PPS_SHIFT) 1295 pps_shift--; 1296 } else if (pps_intcnt >= 4) { 1297 pps_intcnt = 0; 1298 if (pps_shift < PPS_SHIFTMAX) 1299 pps_shift++; 1300 } else 1301 pps_intcnt++; 1302} 1303#endif /* PPS_SYNC */ 1304