kern_clock.c revision 32391
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 1.49 1998/01/10 13:16:19 phk 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/lock.h> 70#include <vm/pmap.h> 71#include <vm/vm_map.h> 72#include <sys/sysctl.h> 73 74#include <machine/cpu.h> 75#define CLOCK_HAIR /* XXX */ 76#include <machine/clock.h> 77#include <machine/limits.h> 78 79#ifdef GPROF 80#include <sys/gmon.h> 81#endif 82 83#if defined(SMP) && defined(BETTER_CLOCK) 84#include <machine/smp.h> 85#endif 86 87static void initclocks __P((void *dummy)); 88SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 89 90/* Some of these don't belong here, but it's easiest to concentrate them. */ 91#if defined(SMP) && defined(BETTER_CLOCK) 92long cp_time[CPUSTATES]; 93#else 94static long cp_time[CPUSTATES]; 95#endif 96long dk_seek[DK_NDRIVE]; 97static long dk_time[DK_NDRIVE]; /* time busy (in statclock ticks) */ 98long dk_wds[DK_NDRIVE]; 99long dk_wpms[DK_NDRIVE]; 100long dk_xfer[DK_NDRIVE]; 101 102int dk_busy; 103int dk_ndrive = 0; 104char dk_names[DK_NDRIVE][DK_NAMELEN]; 105 106long tk_cancc; 107long tk_nin; 108long tk_nout; 109long tk_rawcc; 110 111/* 112 * Clock handling routines. 113 * 114 * This code is written to operate with two timers that run independently of 115 * each other. The main clock, running hz times per second, is used to keep 116 * track of real time. The second timer handles kernel and user profiling, 117 * and does resource use estimation. If the second timer is programmable, 118 * it is randomized to avoid aliasing between the two clocks. For example, 119 * the randomization prevents an adversary from always giving up the cpu 120 * just before its quantum expires. Otherwise, it would never accumulate 121 * cpu ticks. The mean frequency of the second timer is stathz. 122 * 123 * If no second timer exists, stathz will be zero; in this case we drive 124 * profiling and statistics off the main clock. This WILL NOT be accurate; 125 * do not do it unless absolutely necessary. 126 * 127 * The statistics clock may (or may not) be run at a higher rate while 128 * profiling. This profile clock runs at profhz. We require that profhz 129 * be an integral multiple of stathz. 130 * 131 * If the statistics clock is running fast, it must be divided by the ratio 132 * profhz/stathz for statistics. (For profiling, every tick counts.) 133 */ 134 135/* 136 * TODO: 137 * allocate more timeout table slots when table overflows. 138 */ 139 140/* 141 * Bump a timeval by a small number of usec's. 142 */ 143#define BUMPTIME(t, usec) { \ 144 register volatile struct timeval *tp = (t); \ 145 register long us; \ 146 \ 147 tp->tv_usec = us = tp->tv_usec + (usec); \ 148 if (us >= 1000000) { \ 149 tp->tv_usec = us - 1000000; \ 150 tp->tv_sec++; \ 151 } \ 152} 153 154int stathz; 155int profhz; 156static int profprocs; 157int ticks; 158static int psdiv, pscnt; /* prof => stat divider */ 159int psratio; /* ratio: prof / stat */ 160 161volatile struct timeval time; 162volatile struct timeval mono_time; 163 164/* 165 * Phase/frequency-lock loop (PLL/FLL) definitions 166 * 167 * The following variables are read and set by the ntp_adjtime() system 168 * call. 169 * 170 * time_state shows the state of the system clock, with values defined 171 * in the timex.h header file. 172 * 173 * time_status shows the status of the system clock, with bits defined 174 * in the timex.h header file. 175 * 176 * time_offset is used by the PLL/FLL to adjust the system time in small 177 * increments. 178 * 179 * time_constant determines the bandwidth or "stiffness" of the PLL. 180 * 181 * time_tolerance determines maximum frequency error or tolerance of the 182 * CPU clock oscillator and is a property of the architecture; however, 183 * in principle it could change as result of the presence of external 184 * discipline signals, for instance. 185 * 186 * time_precision is usually equal to the kernel tick variable; however, 187 * in cases where a precision clock counter or external clock is 188 * available, the resolution can be much less than this and depend on 189 * whether the external clock is working or not. 190 * 191 * time_maxerror is initialized by a ntp_adjtime() call and increased by 192 * the kernel once each second to reflect the maximum error 193 * bound growth. 194 * 195 * time_esterror is set and read by the ntp_adjtime() call, but 196 * otherwise not used by the kernel. 197 */ 198int time_status = STA_UNSYNC; /* clock status bits */ 199int time_state = TIME_OK; /* clock state */ 200long time_offset = 0; /* time offset (us) */ 201long time_constant = 0; /* pll time constant */ 202long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */ 203long time_precision = 1; /* clock precision (us) */ 204long time_maxerror = MAXPHASE; /* maximum error (us) */ 205long time_esterror = MAXPHASE; /* estimated error (us) */ 206 207/* 208 * The following variables establish the state of the PLL/FLL and the 209 * residual time and frequency offset of the local clock. The scale 210 * factors are defined in the timex.h header file. 211 * 212 * time_phase and time_freq are the phase increment and the frequency 213 * increment, respectively, of the kernel time variable at each tick of 214 * the clock. 215 * 216 * time_freq is set via ntp_adjtime() from a value stored in a file when 217 * the synchronization daemon is first started. Its value is retrieved 218 * via ntp_adjtime() and written to the file about once per hour by the 219 * daemon. 220 * 221 * time_adj is the adjustment added to the value of tick at each timer 222 * interrupt and is recomputed from time_phase and time_freq at each 223 * seconds rollover. 224 * 225 * time_reftime is the second's portion of the system time on the last 226 * call to ntp_adjtime(). It is used to adjust the time_freq variable 227 * and to increase the time_maxerror as the time since last update 228 * increases. 229 */ 230static long time_phase = 0; /* phase offset (scaled us) */ 231long time_freq = 0; /* frequency offset (scaled ppm) */ 232static long time_adj = 0; /* tick adjust (scaled 1 / hz) */ 233static long time_reftime = 0; /* time at last adjustment (s) */ 234 235#ifdef PPS_SYNC 236/* 237 * The following variables are used only if the kernel PPS discipline 238 * code is configured (PPS_SYNC). The scale factors are defined in the 239 * timex.h header file. 240 * 241 * pps_time contains the time at each calibration interval, as read by 242 * microtime(). pps_count counts the seconds of the calibration 243 * interval, the duration of which is nominally pps_shift in powers of 244 * two. 245 * 246 * pps_offset is the time offset produced by the time median filter 247 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by 248 * this filter. 249 * 250 * pps_freq is the frequency offset produced by the frequency median 251 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured 252 * by this filter. 253 * 254 * pps_usec is latched from a high resolution counter or external clock 255 * at pps_time. Here we want the hardware counter contents only, not the 256 * contents plus the time_tv.usec as usual. 257 * 258 * pps_valid counts the number of seconds since the last PPS update. It 259 * is used as a watchdog timer to disable the PPS discipline should the 260 * PPS signal be lost. 261 * 262 * pps_glitch counts the number of seconds since the beginning of an 263 * offset burst more than tick/2 from current nominal offset. It is used 264 * mainly to suppress error bursts due to priority conflicts between the 265 * PPS interrupt and timer interrupt. 266 * 267 * pps_intcnt counts the calibration intervals for use in the interval- 268 * adaptation algorithm. It's just too complicated for words. 269 */ 270struct timeval pps_time; /* kernel time at last interval */ 271long pps_offset = 0; /* pps time offset (us) */ 272long pps_jitter = MAXTIME; /* pps time dispersion (jitter) (us) */ 273long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */ 274long pps_freq = 0; /* frequency offset (scaled ppm) */ 275long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */ 276long pps_ff[] = {0, 0, 0}; /* frequency offset median filter */ 277long pps_usec = 0; /* microsec counter at last interval */ 278long pps_valid = PPS_VALID; /* pps signal watchdog counter */ 279int pps_glitch = 0; /* pps signal glitch counter */ 280int pps_count = 0; /* calibration interval counter (s) */ 281int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */ 282int pps_intcnt = 0; /* intervals at current duration */ 283 284/* 285 * PPS signal quality monitors 286 * 287 * pps_jitcnt counts the seconds that have been discarded because the 288 * jitter measured by the time median filter exceeds the limit MAXTIME 289 * (100 us). 290 * 291 * pps_calcnt counts the frequency calibration intervals, which are 292 * variable from 4 s to 256 s. 293 * 294 * pps_errcnt counts the calibration intervals which have been discarded 295 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the 296 * calibration interval jitter exceeds two ticks. 297 * 298 * pps_stbcnt counts the calibration intervals that have been discarded 299 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us). 300 */ 301long pps_jitcnt = 0; /* jitter limit exceeded */ 302long pps_calcnt = 0; /* calibration intervals */ 303long pps_errcnt = 0; /* calibration errors */ 304long pps_stbcnt = 0; /* stability limit exceeded */ 305#endif /* PPS_SYNC */ 306 307/* XXX none of this stuff works under FreeBSD */ 308 309/* 310 * hardupdate() - local clock update 311 * 312 * This routine is called by ntp_adjtime() to update the local clock 313 * phase and frequency. The implementation is of an adaptive-parameter, 314 * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new 315 * time and frequency offset estimates for each call. If the kernel PPS 316 * discipline code is configured (PPS_SYNC), the PPS signal itself 317 * determines the new time offset, instead of the calling argument. 318 * Presumably, calls to ntp_adjtime() occur only when the caller 319 * believes the local clock is valid within some bound (+-128 ms with 320 * NTP). If the caller's time is far different than the PPS time, an 321 * argument will ensue, and it's not clear who will lose. 322 * 323 * For uncompensated quartz crystal oscillatores and nominal update 324 * intervals less than 1024 s, operation should be in phase-lock mode 325 * (STA_FLL = 0), where the loop is disciplined to phase. For update 326 * intervals greater than thiss, operation should be in frequency-lock 327 * mode (STA_FLL = 1), where the loop is disciplined to frequency. 328 * 329 * Note: splclock() is in effect. 330 */ 331void 332hardupdate(offset) 333 long offset; 334{ 335 long ltemp, mtemp; 336 337 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME)) 338 return; 339 ltemp = offset; 340#ifdef PPS_SYNC 341 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL) 342 ltemp = pps_offset; 343#endif /* PPS_SYNC */ 344 345 /* 346 * Scale the phase adjustment and clamp to the operating range. 347 */ 348 if (ltemp > MAXPHASE) 349 time_offset = MAXPHASE << SHIFT_UPDATE; 350 else if (ltemp < -MAXPHASE) 351 time_offset = -(MAXPHASE << SHIFT_UPDATE); 352 else 353 time_offset = ltemp << SHIFT_UPDATE; 354 355 /* 356 * Select whether the frequency is to be controlled and in which 357 * mode (PLL or FLL). Clamp to the operating range. Ugly 358 * multiply/divide should be replaced someday. 359 */ 360 if (time_status & STA_FREQHOLD || time_reftime == 0) 361 time_reftime = time.tv_sec; 362 mtemp = time.tv_sec - time_reftime; 363 time_reftime = time.tv_sec; 364 if (time_status & STA_FLL) { 365 if (mtemp >= MINSEC) { 366 ltemp = ((time_offset / mtemp) << (SHIFT_USEC - 367 SHIFT_UPDATE)); 368 if (ltemp < 0) 369 time_freq -= -ltemp >> SHIFT_KH; 370 else 371 time_freq += ltemp >> SHIFT_KH; 372 } 373 } else { 374 if (mtemp < MAXSEC) { 375 ltemp *= mtemp; 376 if (ltemp < 0) 377 time_freq -= -ltemp >> (time_constant + 378 time_constant + SHIFT_KF - 379 SHIFT_USEC); 380 else 381 time_freq += ltemp >> (time_constant + 382 time_constant + SHIFT_KF - 383 SHIFT_USEC); 384 } 385 } 386 if (time_freq > time_tolerance) 387 time_freq = time_tolerance; 388 else if (time_freq < -time_tolerance) 389 time_freq = -time_tolerance; 390} 391 392 393 394/* 395 * Initialize clock frequencies and start both clocks running. 396 */ 397/* ARGSUSED*/ 398static void 399initclocks(dummy) 400 void *dummy; 401{ 402 register int i; 403 404 /* 405 * Set divisors to 1 (normal case) and let the machine-specific 406 * code do its bit. 407 */ 408 psdiv = pscnt = 1; 409 cpu_initclocks(); 410 411 /* 412 * Compute profhz/stathz, and fix profhz if needed. 413 */ 414 i = stathz ? stathz : hz; 415 if (profhz == 0) 416 profhz = i; 417 psratio = profhz / i; 418} 419 420/* 421 * The real-time timer, interrupting hz times per second. 422 */ 423void 424hardclock(frame) 425 register struct clockframe *frame; 426{ 427 register struct proc *p; 428 429 p = curproc; 430 if (p) { 431 register struct pstats *pstats; 432 433 /* 434 * Run current process's virtual and profile time, as needed. 435 */ 436 pstats = p->p_stats; 437 if (CLKF_USERMODE(frame) && 438 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 439 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 440 psignal(p, SIGVTALRM); 441 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 442 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 443 psignal(p, SIGPROF); 444 } 445 446#if defined(SMP) && defined(BETTER_CLOCK) 447 forward_hardclock(pscnt); 448#endif 449 /* 450 * If no separate statistics clock is available, run it from here. 451 */ 452 if (stathz == 0) 453 statclock(frame); 454 455 /* 456 * Increment the time-of-day. 457 */ 458 ticks++; 459 { 460 int time_update; 461 struct timeval newtime = time; 462 long ltemp; 463 464 if (timedelta == 0) { 465 time_update = CPU_THISTICKLEN(tick); 466 } else { 467 time_update = CPU_THISTICKLEN(tick) + tickdelta; 468 timedelta -= tickdelta; 469 } 470 BUMPTIME(&mono_time, time_update); 471 472 /* 473 * Compute the phase adjustment. If the low-order bits 474 * (time_phase) of the update overflow, bump the high-order bits 475 * (time_update). 476 */ 477 time_phase += time_adj; 478 if (time_phase <= -FINEUSEC) { 479 ltemp = -time_phase >> SHIFT_SCALE; 480 time_phase += ltemp << SHIFT_SCALE; 481 time_update -= ltemp; 482 } 483 else if (time_phase >= FINEUSEC) { 484 ltemp = time_phase >> SHIFT_SCALE; 485 time_phase -= ltemp << SHIFT_SCALE; 486 time_update += ltemp; 487 } 488 489 newtime.tv_usec += time_update; 490 /* 491 * On rollover of the second the phase adjustment to be used for 492 * the next second is calculated. Also, the maximum error is 493 * increased by the tolerance. If the PPS frequency discipline 494 * code is present, the phase is increased to compensate for the 495 * CPU clock oscillator frequency error. 496 * 497 * On a 32-bit machine and given parameters in the timex.h 498 * header file, the maximum phase adjustment is +-512 ms and 499 * maximum frequency offset is a tad less than) +-512 ppm. On a 500 * 64-bit machine, you shouldn't need to ask. 501 */ 502 if (newtime.tv_usec >= 1000000) { 503 newtime.tv_usec -= 1000000; 504 newtime.tv_sec++; 505 time_maxerror += time_tolerance >> SHIFT_USEC; 506 507 /* 508 * Compute the phase adjustment for the next second. In 509 * PLL mode, the offset is reduced by a fixed factor 510 * times the time constant. In FLL mode the offset is 511 * used directly. In either mode, the maximum phase 512 * adjustment for each second is clamped so as to spread 513 * the adjustment over not more than the number of 514 * seconds between updates. 515 */ 516 if (time_offset < 0) { 517 ltemp = -time_offset; 518 if (!(time_status & STA_FLL)) 519 ltemp >>= SHIFT_KG + time_constant; 520 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 521 ltemp = (MAXPHASE / MINSEC) << 522 SHIFT_UPDATE; 523 time_offset += ltemp; 524 time_adj = -ltemp << (SHIFT_SCALE - SHIFT_HZ - 525 SHIFT_UPDATE); 526 } else { 527 ltemp = time_offset; 528 if (!(time_status & STA_FLL)) 529 ltemp >>= SHIFT_KG + time_constant; 530 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE) 531 ltemp = (MAXPHASE / MINSEC) << 532 SHIFT_UPDATE; 533 time_offset -= ltemp; 534 time_adj = ltemp << (SHIFT_SCALE - SHIFT_HZ - 535 SHIFT_UPDATE); 536 } 537 538 /* 539 * Compute the frequency estimate and additional phase 540 * adjustment due to frequency error for the next 541 * second. When the PPS signal is engaged, gnaw on the 542 * watchdog counter and update the frequency computed by 543 * the pll and the PPS signal. 544 */ 545#ifdef PPS_SYNC 546 pps_valid++; 547 if (pps_valid == PPS_VALID) { 548 pps_jitter = MAXTIME; 549 pps_stabil = MAXFREQ; 550 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER | 551 STA_PPSWANDER | STA_PPSERROR); 552 } 553 ltemp = time_freq + pps_freq; 554#else 555 ltemp = time_freq; 556#endif /* PPS_SYNC */ 557 if (ltemp < 0) 558 time_adj -= -ltemp >> 559 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 560 else 561 time_adj += ltemp >> 562 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE); 563 564#if SHIFT_HZ == 7 565 /* 566 * When the CPU clock oscillator frequency is not a 567 * power of two in Hz, the SHIFT_HZ is only an 568 * approximate scale factor. In the SunOS kernel, this 569 * results in a PLL gain factor of 1/1.28 = 0.78 what it 570 * should be. In the following code the overall gain is 571 * increased by a factor of 1.25, which results in a 572 * residual error less than 3 percent. 573 */ 574 /* Same thing applies for FreeBSD --GAW */ 575 if (hz == 100) { 576 if (time_adj < 0) 577 time_adj -= -time_adj >> 2; 578 else 579 time_adj += time_adj >> 2; 580 } 581#endif /* SHIFT_HZ */ 582 583 /* XXX - this is really bogus, but can't be fixed until 584 xntpd's idea of the system clock is fixed to know how 585 the user wants leap seconds handled; in the mean time, 586 we assume that users of NTP are running without proper 587 leap second support (this is now the default anyway) */ 588 /* 589 * Leap second processing. If in leap-insert state at 590 * the end of the day, the system clock is set back one 591 * second; if in leap-delete state, the system clock is 592 * set ahead one second. The microtime() routine or 593 * external clock driver will insure that reported time 594 * is always monotonic. The ugly divides should be 595 * replaced. 596 */ 597 switch (time_state) { 598 599 case TIME_OK: 600 if (time_status & STA_INS) 601 time_state = TIME_INS; 602 else if (time_status & STA_DEL) 603 time_state = TIME_DEL; 604 break; 605 606 case TIME_INS: 607 if (newtime.tv_sec % 86400 == 0) { 608 newtime.tv_sec--; 609 time_state = TIME_OOP; 610 } 611 break; 612 613 case TIME_DEL: 614 if ((newtime.tv_sec + 1) % 86400 == 0) { 615 newtime.tv_sec++; 616 time_state = TIME_WAIT; 617 } 618 break; 619 620 case TIME_OOP: 621 time_state = TIME_WAIT; 622 break; 623 624 case TIME_WAIT: 625 if (!(time_status & (STA_INS | STA_DEL))) 626 time_state = TIME_OK; 627 } 628 } 629 CPU_CLOCKUPDATE(&time, &newtime); 630 } 631 632 setsoftclock(); 633} 634 635void 636gettime(struct timeval *tvp) 637{ 638 int s; 639 640 s = splclock(); 641 /* XXX should use microtime() iff tv_usec is used. */ 642 *tvp = time; 643 splx(s); 644} 645 646/* 647 * Compute number of hz until specified time. Used to 648 * compute third argument to timeout() from an absolute time. 649 */ 650int 651hzto(tv) 652 struct timeval *tv; 653{ 654 register unsigned long ticks; 655 register long sec, usec; 656 int s; 657 658 /* 659 * If the number of usecs in the whole seconds part of the time 660 * difference fits in a long, then the total number of usecs will 661 * fit in an unsigned long. Compute the total and convert it to 662 * ticks, rounding up and adding 1 to allow for the current tick 663 * to expire. Rounding also depends on unsigned long arithmetic 664 * to avoid overflow. 665 * 666 * Otherwise, if the number of ticks in the whole seconds part of 667 * the time difference fits in a long, then convert the parts to 668 * ticks separately and add, using similar rounding methods and 669 * overflow avoidance. This method would work in the previous 670 * case but it is slightly slower and assumes that hz is integral. 671 * 672 * Otherwise, round the time difference down to the maximum 673 * representable value. 674 * 675 * If ints have 32 bits, then the maximum value for any timeout in 676 * 10ms ticks is 248 days. 677 */ 678 s = splclock(); 679 sec = tv->tv_sec - time.tv_sec; 680 usec = tv->tv_usec - time.tv_usec; 681 splx(s); 682 if (usec < 0) { 683 sec--; 684 usec += 1000000; 685 } 686 if (sec < 0) { 687#ifdef DIAGNOSTIC 688 printf("hzto: negative time difference %ld sec %ld usec\n", 689 sec, usec); 690#endif 691 ticks = 1; 692 } else if (sec <= LONG_MAX / 1000000) 693 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 694 / tick + 1; 695 else if (sec <= LONG_MAX / hz) 696 ticks = sec * hz 697 + ((unsigned long)usec + (tick - 1)) / tick + 1; 698 else 699 ticks = LONG_MAX; 700 if (ticks > INT_MAX) 701 ticks = INT_MAX; 702 return (ticks); 703} 704 705/* 706 * Start profiling on a process. 707 * 708 * Kernel profiling passes proc0 which never exits and hence 709 * keeps the profile clock running constantly. 710 */ 711void 712startprofclock(p) 713 register struct proc *p; 714{ 715 int s; 716 717 if ((p->p_flag & P_PROFIL) == 0) { 718 p->p_flag |= P_PROFIL; 719 if (++profprocs == 1 && stathz != 0) { 720 s = splstatclock(); 721 psdiv = pscnt = psratio; 722 setstatclockrate(profhz); 723 splx(s); 724 } 725 } 726} 727 728/* 729 * Stop profiling on a process. 730 */ 731void 732stopprofclock(p) 733 register struct proc *p; 734{ 735 int s; 736 737 if (p->p_flag & P_PROFIL) { 738 p->p_flag &= ~P_PROFIL; 739 if (--profprocs == 0 && stathz != 0) { 740 s = splstatclock(); 741 psdiv = pscnt = 1; 742 setstatclockrate(stathz); 743 splx(s); 744 } 745 } 746} 747 748/* 749 * Statistics clock. Grab profile sample, and if divider reaches 0, 750 * do process and kernel statistics. 751 */ 752void 753statclock(frame) 754 register struct clockframe *frame; 755{ 756#ifdef GPROF 757 register struct gmonparam *g; 758#endif 759 register struct proc *p; 760 register int i; 761 struct pstats *pstats; 762 long rss; 763 struct rusage *ru; 764 struct vmspace *vm; 765 766 if (CLKF_USERMODE(frame)) { 767 p = curproc; 768 if (p->p_flag & P_PROFIL) 769 addupc_intr(p, CLKF_PC(frame), 1); 770#if defined(SMP) && defined(BETTER_CLOCK) 771 if (stathz != 0) 772 forward_statclock(pscnt); 773#endif 774 if (--pscnt > 0) 775 return; 776 /* 777 * Came from user mode; CPU was in user state. 778 * If this process is being profiled record the tick. 779 */ 780 p->p_uticks++; 781 if (p->p_nice > NZERO) 782 cp_time[CP_NICE]++; 783 else 784 cp_time[CP_USER]++; 785 } else { 786#ifdef GPROF 787 /* 788 * Kernel statistics are just like addupc_intr, only easier. 789 */ 790 g = &_gmonparam; 791 if (g->state == GMON_PROF_ON) { 792 i = CLKF_PC(frame) - g->lowpc; 793 if (i < g->textsize) { 794 i /= HISTFRACTION * sizeof(*g->kcount); 795 g->kcount[i]++; 796 } 797 } 798#endif 799#if defined(SMP) && defined(BETTER_CLOCK) 800 if (stathz != 0) 801 forward_statclock(pscnt); 802#endif 803 if (--pscnt > 0) 804 return; 805 /* 806 * Came from kernel mode, so we were: 807 * - handling an interrupt, 808 * - doing syscall or trap work on behalf of the current 809 * user process, or 810 * - spinning in the idle loop. 811 * Whichever it is, charge the time as appropriate. 812 * Note that we charge interrupts to the current process, 813 * regardless of whether they are ``for'' that process, 814 * so that we know how much of its real time was spent 815 * in ``non-process'' (i.e., interrupt) work. 816 */ 817 p = curproc; 818 if (CLKF_INTR(frame)) { 819 if (p != NULL) 820 p->p_iticks++; 821 cp_time[CP_INTR]++; 822 } else if (p != NULL) { 823 p->p_sticks++; 824 cp_time[CP_SYS]++; 825 } else 826 cp_time[CP_IDLE]++; 827 } 828 pscnt = psdiv; 829 830 /* 831 * We maintain statistics shown by user-level statistics 832 * programs: the amount of time in each cpu state, and 833 * the amount of time each of DK_NDRIVE ``drives'' is busy. 834 * 835 * XXX should either run linked list of drives, or (better) 836 * grab timestamps in the start & done code. 837 */ 838 for (i = 0; i < DK_NDRIVE; i++) 839 if (dk_busy & (1 << i)) 840 dk_time[i]++; 841 842 /* 843 * We adjust the priority of the current process. The priority of 844 * a process gets worse as it accumulates CPU time. The cpu usage 845 * estimator (p_estcpu) is increased here. The formula for computing 846 * priorities (in kern_synch.c) will compute a different value each 847 * time p_estcpu increases by 4. The cpu usage estimator ramps up 848 * quite quickly when the process is running (linearly), and decays 849 * away exponentially, at a rate which is proportionally slower when 850 * the system is busy. The basic principal is that the system will 851 * 90% forget that the process used a lot of CPU time in 5 * loadav 852 * seconds. This causes the system to favor processes which haven't 853 * run much recently, and to round-robin among other processes. 854 */ 855 if (p != NULL) { 856 p->p_cpticks++; 857 if (++p->p_estcpu == 0) 858 p->p_estcpu--; 859 if ((p->p_estcpu & 3) == 0) { 860 resetpriority(p); 861 if (p->p_priority >= PUSER) 862 p->p_priority = p->p_usrpri; 863 } 864 865 /* Update resource usage integrals and maximums. */ 866 if ((pstats = p->p_stats) != NULL && 867 (ru = &pstats->p_ru) != NULL && 868 (vm = p->p_vmspace) != NULL) { 869 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024; 870 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024; 871 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024; 872 rss = vm->vm_pmap.pm_stats.resident_count * 873 PAGE_SIZE / 1024; 874 if (ru->ru_maxrss < rss) 875 ru->ru_maxrss = rss; 876 } 877 } 878} 879 880/* 881 * Return information about system clocks. 882 */ 883static int 884sysctl_kern_clockrate SYSCTL_HANDLER_ARGS 885{ 886 struct clockinfo clkinfo; 887 /* 888 * Construct clockinfo structure. 889 */ 890 clkinfo.hz = hz; 891 clkinfo.tick = tick; 892 clkinfo.tickadj = tickadj; 893 clkinfo.profhz = profhz; 894 clkinfo.stathz = stathz ? stathz : hz; 895 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 896} 897 898SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 899 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 900 901#ifdef PPS_SYNC 902 903/* We need this ugly monster twice, so lets macroize it... */ 904 905#define MEDIAN3(a, m, s) \ 906 do { \ 907 if (a[0] > a[1]) { \ 908 if (a[1] > a[2]) { \ 909 /* 0 1 2 */ \ 910 m = a[1]; \ 911 s = a[0] - a[2]; \ 912 } else if (a[2] > a[0]) { \ 913 /* 2 0 1 */ \ 914 m = a[0]; \ 915 s = a[2] - a[1]; \ 916 } else { \ 917 /* 0 2 1 */ \ 918 m = a[2]; \ 919 s = a[0] - a[1]; \ 920 } \ 921 } else { \ 922 if (a[1] < a[2]) { \ 923 /* 2 1 0 */ \ 924 m = a[1]; \ 925 s = a[2] - a[0]; \ 926 } else if (a[2] < a[0]) { \ 927 /* 1 0 2 */ \ 928 m = a[0]; \ 929 s = a[1] - a[2]; \ 930 } else { \ 931 /* 1 2 0 */ \ 932 m = a[2]; \ 933 s = a[1] - a[0]; \ 934 } \ 935 } \ 936 } while (0) 937 938/* 939 * hardpps() - discipline CPU clock oscillator to external PPS signal 940 * 941 * This routine is called at each PPS interrupt in order to discipline 942 * the CPU clock oscillator to the PPS signal. It measures the PPS phase 943 * and leaves it in a handy spot for the hardclock() routine. It 944 * integrates successive PPS phase differences and calculates the 945 * frequency offset. This is used in hardclock() to discipline the CPU 946 * clock oscillator so that intrinsic frequency error is cancelled out. 947 * The code requires the caller to capture the time and hardware counter 948 * value at the on-time PPS signal transition. 949 * 950 * Note that, on some Unix systems, this routine runs at an interrupt 951 * priority level higher than the timer interrupt routine hardclock(). 952 * Therefore, the variables used are distinct from the hardclock() 953 * variables, except for certain exceptions: The PPS frequency pps_freq 954 * and phase pps_offset variables are determined by this routine and 955 * updated atomically. The time_tolerance variable can be considered a 956 * constant, since it is infrequently changed, and then only when the 957 * PPS signal is disabled. The watchdog counter pps_valid is updated 958 * once per second by hardclock() and is atomically cleared in this 959 * routine. 960 */ 961void 962hardpps(tvp, p_usec) 963 struct timeval *tvp; /* time at PPS */ 964 long p_usec; /* hardware counter at PPS */ 965{ 966 long u_usec, v_usec, bigtick; 967 long cal_sec, cal_usec; 968 969 /* 970 * An occasional glitch can be produced when the PPS interrupt 971 * occurs in the hardclock() routine before the time variable is 972 * updated. Here the offset is discarded when the difference 973 * between it and the last one is greater than tick/2, but not 974 * if the interval since the first discard exceeds 30 s. 975 */ 976 time_status |= STA_PPSSIGNAL; 977 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR); 978 pps_valid = 0; 979 u_usec = -tvp->tv_usec; 980 if (u_usec < -500000) 981 u_usec += 1000000; 982 v_usec = pps_offset - u_usec; 983 if (v_usec < 0) 984 v_usec = -v_usec; 985 if (v_usec > (tick >> 1)) { 986 if (pps_glitch > MAXGLITCH) { 987 pps_glitch = 0; 988 pps_tf[2] = u_usec; 989 pps_tf[1] = u_usec; 990 } else { 991 pps_glitch++; 992 u_usec = pps_offset; 993 } 994 } else 995 pps_glitch = 0; 996 997 /* 998 * A three-stage median filter is used to help deglitch the pps 999 * time. The median sample becomes the time offset estimate; the 1000 * difference between the other two samples becomes the time 1001 * dispersion (jitter) estimate. 1002 */ 1003 pps_tf[2] = pps_tf[1]; 1004 pps_tf[1] = pps_tf[0]; 1005 pps_tf[0] = u_usec; 1006 1007 MEDIAN3(pps_tf, pps_offset, v_usec); 1008 1009 if (v_usec > MAXTIME) 1010 pps_jitcnt++; 1011 v_usec = (v_usec << PPS_AVG) - pps_jitter; 1012 if (v_usec < 0) 1013 pps_jitter -= -v_usec >> PPS_AVG; 1014 else 1015 pps_jitter += v_usec >> PPS_AVG; 1016 if (pps_jitter > (MAXTIME >> 1)) 1017 time_status |= STA_PPSJITTER; 1018 1019 /* 1020 * During the calibration interval adjust the starting time when 1021 * the tick overflows. At the end of the interval compute the 1022 * duration of the interval and the difference of the hardware 1023 * counters at the beginning and end of the interval. This code 1024 * is deliciously complicated by the fact valid differences may 1025 * exceed the value of tick when using long calibration 1026 * intervals and small ticks. Note that the counter can be 1027 * greater than tick if caught at just the wrong instant, but 1028 * the values returned and used here are correct. 1029 */ 1030 bigtick = (long)tick << SHIFT_USEC; 1031 pps_usec -= pps_freq; 1032 if (pps_usec >= bigtick) 1033 pps_usec -= bigtick; 1034 if (pps_usec < 0) 1035 pps_usec += bigtick; 1036 pps_time.tv_sec++; 1037 pps_count++; 1038 if (pps_count < (1 << pps_shift)) 1039 return; 1040 pps_count = 0; 1041 pps_calcnt++; 1042 u_usec = p_usec << SHIFT_USEC; 1043 v_usec = pps_usec - u_usec; 1044 if (v_usec >= bigtick >> 1) 1045 v_usec -= bigtick; 1046 if (v_usec < -(bigtick >> 1)) 1047 v_usec += bigtick; 1048 if (v_usec < 0) 1049 v_usec = -(-v_usec >> pps_shift); 1050 else 1051 v_usec = v_usec >> pps_shift; 1052 pps_usec = u_usec; 1053 cal_sec = tvp->tv_sec; 1054 cal_usec = tvp->tv_usec; 1055 cal_sec -= pps_time.tv_sec; 1056 cal_usec -= pps_time.tv_usec; 1057 if (cal_usec < 0) { 1058 cal_usec += 1000000; 1059 cal_sec--; 1060 } 1061 pps_time = *tvp; 1062 1063 /* 1064 * Check for lost interrupts, noise, excessive jitter and 1065 * excessive frequency error. The number of timer ticks during 1066 * the interval may vary +-1 tick. Add to this a margin of one 1067 * tick for the PPS signal jitter and maximum frequency 1068 * deviation. If the limits are exceeded, the calibration 1069 * interval is reset to the minimum and we start over. 1070 */ 1071 u_usec = (long)tick << 1; 1072 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec)) 1073 || (cal_sec == 0 && cal_usec < u_usec)) 1074 || v_usec > time_tolerance || v_usec < -time_tolerance) { 1075 pps_errcnt++; 1076 pps_shift = PPS_SHIFT; 1077 pps_intcnt = 0; 1078 time_status |= STA_PPSERROR; 1079 return; 1080 } 1081 1082 /* 1083 * A three-stage median filter is used to help deglitch the pps 1084 * frequency. The median sample becomes the frequency offset 1085 * estimate; the difference between the other two samples 1086 * becomes the frequency dispersion (stability) estimate. 1087 */ 1088 pps_ff[2] = pps_ff[1]; 1089 pps_ff[1] = pps_ff[0]; 1090 pps_ff[0] = v_usec; 1091 1092 MEDIAN3(pps_ff, u_usec, v_usec); 1093 1094 /* 1095 * Here the frequency dispersion (stability) is updated. If it 1096 * is less than one-fourth the maximum (MAXFREQ), the frequency 1097 * offset is updated as well, but clamped to the tolerance. It 1098 * will be processed later by the hardclock() routine. 1099 */ 1100 v_usec = (v_usec >> 1) - pps_stabil; 1101 if (v_usec < 0) 1102 pps_stabil -= -v_usec >> PPS_AVG; 1103 else 1104 pps_stabil += v_usec >> PPS_AVG; 1105 if (pps_stabil > MAXFREQ >> 2) { 1106 pps_stbcnt++; 1107 time_status |= STA_PPSWANDER; 1108 return; 1109 } 1110 if (time_status & STA_PPSFREQ) { 1111 if (u_usec < 0) { 1112 pps_freq -= -u_usec >> PPS_AVG; 1113 if (pps_freq < -time_tolerance) 1114 pps_freq = -time_tolerance; 1115 u_usec = -u_usec; 1116 } else { 1117 pps_freq += u_usec >> PPS_AVG; 1118 if (pps_freq > time_tolerance) 1119 pps_freq = time_tolerance; 1120 } 1121 } 1122 1123 /* 1124 * Here the calibration interval is adjusted. If the maximum 1125 * time difference is greater than tick / 4, reduce the interval 1126 * by half. If this is not the case for four consecutive 1127 * intervals, double the interval. 1128 */ 1129 if (u_usec << pps_shift > bigtick >> 2) { 1130 pps_intcnt = 0; 1131 if (pps_shift > PPS_SHIFT) 1132 pps_shift--; 1133 } else if (pps_intcnt >= 4) { 1134 pps_intcnt = 0; 1135 if (pps_shift < PPS_SHIFTMAX) 1136 pps_shift++; 1137 } else 1138 pps_intcnt++; 1139} 1140#endif /* PPS_SYNC */ 1141 1142