Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.53
      1 /*	$NetBSD: kern_clock.c,v 1.53 2000/03/23 20:51:09 thorpej Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2000 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
      9  * NASA Ames Research Center.
     10  *
     11  * Redistribution and use in source and binary forms, with or without
     12  * modification, are permitted provided that the following conditions
     13  * are met:
     14  * 1. Redistributions of source code must retain the above copyright
     15  *    notice, this list of conditions and the following disclaimer.
     16  * 2. Redistributions in binary form must reproduce the above copyright
     17  *    notice, this list of conditions and the following disclaimer in the
     18  *    documentation and/or other materials provided with the distribution.
     19  * 3. All advertising materials mentioning features or use of this software
     20  *    must display the following acknowledgement:
     21  *	This product includes software developed by the NetBSD
     22  *	Foundation, Inc. and its contributors.
     23  * 4. Neither the name of The NetBSD Foundation nor the names of its
     24  *    contributors may be used to endorse or promote products derived
     25  *    from this software without specific prior written permission.
     26  *
     27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     37  * POSSIBILITY OF SUCH DAMAGE.
     38  */
     39 
     40 /*-
     41  * Copyright (c) 1982, 1986, 1991, 1993
     42  *	The Regents of the University of California.  All rights reserved.
     43  * (c) UNIX System Laboratories, Inc.
     44  * All or some portions of this file are derived from material licensed
     45  * to the University of California by American Telephone and Telegraph
     46  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     47  * the permission of UNIX System Laboratories, Inc.
     48  *
     49  * Redistribution and use in source and binary forms, with or without
     50  * modification, are permitted provided that the following conditions
     51  * are met:
     52  * 1. Redistributions of source code must retain the above copyright
     53  *    notice, this list of conditions and the following disclaimer.
     54  * 2. Redistributions in binary form must reproduce the above copyright
     55  *    notice, this list of conditions and the following disclaimer in the
     56  *    documentation and/or other materials provided with the distribution.
     57  * 3. All advertising materials mentioning features or use of this software
     58  *    must display the following acknowledgement:
     59  *	This product includes software developed by the University of
     60  *	California, Berkeley and its contributors.
     61  * 4. Neither the name of the University nor the names of its contributors
     62  *    may be used to endorse or promote products derived from this software
     63  *    without specific prior written permission.
     64  *
     65  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     66  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     67  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     68  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     69  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     70  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     71  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     72  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     73  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     74  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     75  * SUCH DAMAGE.
     76  *
     77  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
     78  */
     79 
     80 #include "opt_ntp.h"
     81 
     82 #include <sys/param.h>
     83 #include <sys/systm.h>
     84 #include <sys/dkstat.h>
     85 #include <sys/callout.h>
     86 #include <sys/kernel.h>
     87 #include <sys/proc.h>
     88 #include <sys/resourcevar.h>
     89 #include <sys/signalvar.h>
     90 #include <vm/vm.h>
     91 #include <sys/sysctl.h>
     92 #include <sys/timex.h>
     93 #include <sys/sched.h>
     94 
     95 #include <machine/cpu.h>
     96 
     97 #ifdef GPROF
     98 #include <sys/gmon.h>
     99 #endif
    100 
    101 /*
    102  * Clock handling routines.
    103  *
    104  * This code is written to operate with two timers that run independently of
    105  * each other.  The main clock, running hz times per second, is used to keep
    106  * track of real time.  The second timer handles kernel and user profiling,
    107  * and does resource use estimation.  If the second timer is programmable,
    108  * it is randomized to avoid aliasing between the two clocks.  For example,
    109  * the randomization prevents an adversary from always giving up the cpu
    110  * just before its quantum expires.  Otherwise, it would never accumulate
    111  * cpu ticks.  The mean frequency of the second timer is stathz.
    112  *
    113  * If no second timer exists, stathz will be zero; in this case we drive
    114  * profiling and statistics off the main clock.  This WILL NOT be accurate;
    115  * do not do it unless absolutely necessary.
    116  *
    117  * The statistics clock may (or may not) be run at a higher rate while
    118  * profiling.  This profile clock runs at profhz.  We require that profhz
    119  * be an integral multiple of stathz.
    120  *
    121  * If the statistics clock is running fast, it must be divided by the ratio
    122  * profhz/stathz for statistics.  (For profiling, every tick counts.)
    123  */
    124 
    125 #ifdef NTP	/* NTP phase-locked loop in kernel */
    126 /*
    127  * Phase/frequency-lock loop (PLL/FLL) definitions
    128  *
    129  * The following variables are read and set by the ntp_adjtime() system
    130  * call.
    131  *
    132  * time_state shows the state of the system clock, with values defined
    133  * in the timex.h header file.
    134  *
    135  * time_status shows the status of the system clock, with bits defined
    136  * in the timex.h header file.
    137  *
    138  * time_offset is used by the PLL/FLL to adjust the system time in small
    139  * increments.
    140  *
    141  * time_constant determines the bandwidth or "stiffness" of the PLL.
    142  *
    143  * time_tolerance determines maximum frequency error or tolerance of the
    144  * CPU clock oscillator and is a property of the architecture; however,
    145  * in principle it could change as result of the presence of external
    146  * discipline signals, for instance.
    147  *
    148  * time_precision is usually equal to the kernel tick variable; however,
    149  * in cases where a precision clock counter or external clock is
    150  * available, the resolution can be much less than this and depend on
    151  * whether the external clock is working or not.
    152  *
    153  * time_maxerror is initialized by a ntp_adjtime() call and increased by
    154  * the kernel once each second to reflect the maximum error bound
    155  * growth.
    156  *
    157  * time_esterror is set and read by the ntp_adjtime() call, but
    158  * otherwise not used by the kernel.
    159  */
    160 int time_state = TIME_OK;	/* clock state */
    161 int time_status = STA_UNSYNC;	/* clock status bits */
    162 long time_offset = 0;		/* time offset (us) */
    163 long time_constant = 0;		/* pll time constant */
    164 long time_tolerance = MAXFREQ;	/* frequency tolerance (scaled ppm) */
    165 long time_precision = 1;	/* clock precision (us) */
    166 long time_maxerror = MAXPHASE;	/* maximum error (us) */
    167 long time_esterror = MAXPHASE;	/* estimated error (us) */
    168 
    169 /*
    170  * The following variables establish the state of the PLL/FLL and the
    171  * residual time and frequency offset of the local clock. The scale
    172  * factors are defined in the timex.h header file.
    173  *
    174  * time_phase and time_freq are the phase increment and the frequency
    175  * increment, respectively, of the kernel time variable.
    176  *
    177  * time_freq is set via ntp_adjtime() from a value stored in a file when
    178  * the synchronization daemon is first started. Its value is retrieved
    179  * via ntp_adjtime() and written to the file about once per hour by the
    180  * daemon.
    181  *
    182  * time_adj is the adjustment added to the value of tick at each timer
    183  * interrupt and is recomputed from time_phase and time_freq at each
    184  * seconds rollover.
    185  *
    186  * time_reftime is the second's portion of the system time at the last
    187  * call to ntp_adjtime(). It is used to adjust the time_freq variable
    188  * and to increase the time_maxerror as the time since last update
    189  * increases.
    190  */
    191 long time_phase = 0;		/* phase offset (scaled us) */
    192 long time_freq = 0;		/* frequency offset (scaled ppm) */
    193 long time_adj = 0;		/* tick adjust (scaled 1 / hz) */
    194 long time_reftime = 0;		/* time at last adjustment (s) */
    195 
    196 #ifdef PPS_SYNC
    197 /*
    198  * The following variables are used only if the kernel PPS discipline
    199  * code is configured (PPS_SYNC). The scale factors are defined in the
    200  * timex.h header file.
    201  *
    202  * pps_time contains the time at each calibration interval, as read by
    203  * microtime(). pps_count counts the seconds of the calibration
    204  * interval, the duration of which is nominally pps_shift in powers of
    205  * two.
    206  *
    207  * pps_offset is the time offset produced by the time median filter
    208  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
    209  * this filter.
    210  *
    211  * pps_freq is the frequency offset produced by the frequency median
    212  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
    213  * by this filter.
    214  *
    215  * pps_usec is latched from a high resolution counter or external clock
    216  * at pps_time. Here we want the hardware counter contents only, not the
    217  * contents plus the time_tv.usec as usual.
    218  *
    219  * pps_valid counts the number of seconds since the last PPS update. It
    220  * is used as a watchdog timer to disable the PPS discipline should the
    221  * PPS signal be lost.
    222  *
    223  * pps_glitch counts the number of seconds since the beginning of an
    224  * offset burst more than tick/2 from current nominal offset. It is used
    225  * mainly to suppress error bursts due to priority conflicts between the
    226  * PPS interrupt and timer interrupt.
    227  *
    228  * pps_intcnt counts the calibration intervals for use in the interval-
    229  * adaptation algorithm. It's just too complicated for words.
    230  */
    231 struct timeval pps_time;	/* kernel time at last interval */
    232 long pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
    233 long pps_offset = 0;		/* pps time offset (us) */
    234 long pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
    235 long pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
    236 long pps_freq = 0;		/* frequency offset (scaled ppm) */
    237 long pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
    238 long pps_usec = 0;		/* microsec counter at last interval */
    239 long pps_valid = PPS_VALID;	/* pps signal watchdog counter */
    240 int pps_glitch = 0;		/* pps signal glitch counter */
    241 int pps_count = 0;		/* calibration interval counter (s) */
    242 int pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
    243 int pps_intcnt = 0;		/* intervals at current duration */
    244 
    245 /*
    246  * PPS signal quality monitors
    247  *
    248  * pps_jitcnt counts the seconds that have been discarded because the
    249  * jitter measured by the time median filter exceeds the limit MAXTIME
    250  * (100 us).
    251  *
    252  * pps_calcnt counts the frequency calibration intervals, which are
    253  * variable from 4 s to 256 s.
    254  *
    255  * pps_errcnt counts the calibration intervals which have been discarded
    256  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
    257  * calibration interval jitter exceeds two ticks.
    258  *
    259  * pps_stbcnt counts the calibration intervals that have been discarded
    260  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
    261  */
    262 long pps_jitcnt = 0;		/* jitter limit exceeded */
    263 long pps_calcnt = 0;		/* calibration intervals */
    264 long pps_errcnt = 0;		/* calibration errors */
    265 long pps_stbcnt = 0;		/* stability limit exceeded */
    266 #endif /* PPS_SYNC */
    267 
    268 #ifdef EXT_CLOCK
    269 /*
    270  * External clock definitions
    271  *
    272  * The following definitions and declarations are used only if an
    273  * external clock is configured on the system.
    274  */
    275 #define CLOCK_INTERVAL 30	/* CPU clock update interval (s) */
    276 
    277 /*
    278  * The clock_count variable is set to CLOCK_INTERVAL at each PPS
    279  * interrupt and decremented once each second.
    280  */
    281 int clock_count = 0;		/* CPU clock counter */
    282 
    283 #ifdef HIGHBALL
    284 /*
    285  * The clock_offset and clock_cpu variables are used by the HIGHBALL
    286  * interface. The clock_offset variable defines the offset between
    287  * system time and the HIGBALL counters. The clock_cpu variable contains
    288  * the offset between the system clock and the HIGHBALL clock for use in
    289  * disciplining the kernel time variable.
    290  */
    291 extern struct timeval clock_offset; /* Highball clock offset */
    292 long clock_cpu = 0;		/* CPU clock adjust */
    293 #endif /* HIGHBALL */
    294 #endif /* EXT_CLOCK */
    295 #endif /* NTP */
    296 
    297 
    298 /*
    299  * Bump a timeval by a small number of usec's.
    300  */
    301 #define BUMPTIME(t, usec) { \
    302 	register volatile struct timeval *tp = (t); \
    303 	register long us; \
    304  \
    305 	tp->tv_usec = us = tp->tv_usec + (usec); \
    306 	if (us >= 1000000) { \
    307 		tp->tv_usec = us - 1000000; \
    308 		tp->tv_sec++; \
    309 	} \
    310 }
    311 
    312 int	stathz;
    313 int	profhz;
    314 int	profprocs;
    315 u_int64_t hardclock_ticks, softclock_ticks;
    316 int	softclock_running;		/* 1 => softclock() is running */
    317 static int psdiv, pscnt;		/* prof => stat divider */
    318 int	psratio;			/* ratio: prof / stat */
    319 int	tickfix, tickfixinterval;	/* used if tick not really integral */
    320 #ifndef NTP
    321 static int tickfixcnt;			/* accumulated fractional error */
    322 #else
    323 int	fixtick;			/* used by NTP for same */
    324 int	shifthz;
    325 #endif
    326 
    327 /*
    328  * We might want ldd to load the both words from time at once.
    329  * To succeed we need to be quadword aligned.
    330  * The sparc already does that, and that it has worked so far is a fluke.
    331  */
    332 volatile struct	timeval time  __attribute__((__aligned__(__alignof__(quad_t))));
    333 volatile struct	timeval mono_time;
    334 
    335 /*
    336  * The callout mechanism is based on the work of Adam M. Costello and
    337  * George Varghese, published in a technical report entitled "Redesigning
    338  * the BSD Callout and Timer Facilities", and Justin Gibbs's subsequent
    339  * integration into FreeBSD, modified for NetBSD by Jason R. Thorpe.
    340  *
    341  * The original work on the data structures used in this implementation
    342  * was published by G. Varghese and A. Lauck in the paper "Hashed and
    343  * Hierarchical Timing Wheels: Data Structures for the Efficient
    344  * Implementation of a Timer Facility" in the Proceedings of the 11th
    345  * ACM Annual Symposium on Operating System Principles, Austin, Texas,
    346  * November 1987.
    347  */
    348 struct callout_queue *callwheel;
    349 int	callwheelsize, callwheelbits, callwheelmask;
    350 
    351 static struct callout *nextsoftcheck;	/* next callout to be checked */
    352 
    353 #ifdef CALLWHEEL_STATS
    354 int	callwheel_collisions;		/* number of hash collisions */
    355 int	callwheel_maxlength;		/* length of the longest hash chain */
    356 int	*callwheel_sizes;		/* per-bucket length count */
    357 u_int64_t callwheel_count;		/* # callouts currently */
    358 u_int64_t callwheel_established;	/* # callouts established */
    359 u_int64_t callwheel_fired;		/* # callouts that fired */
    360 u_int64_t callwheel_disestablished;	/* # callouts disestablished */
    361 u_int64_t callwheel_changed;		/* # callouts changed */
    362 u_int64_t callwheel_softclocks;		/* # times softclock() called */
    363 u_int64_t callwheel_softchecks;		/* # checks per softclock() */
    364 u_int64_t callwheel_softempty;		/* # empty buckets seen */
    365 #endif /* CALLWHEEL_STATS */
    366 
    367 /*
    368  * This value indicates the number of consecutive callouts that
    369  * will be checked before we allow interrupts to have a chance
    370  * again.
    371  */
    372 #ifndef MAX_SOFTCLOCK_STEPS
    373 #define	MAX_SOFTCLOCK_STEPS	100
    374 #endif
    375 
    376 /*
    377  * Initialize clock frequencies and start both clocks running.
    378  */
    379 void
    380 initclocks()
    381 {
    382 	register int i;
    383 
    384 	/*
    385 	 * Set divisors to 1 (normal case) and let the machine-specific
    386 	 * code do its bit.
    387 	 */
    388 	psdiv = pscnt = 1;
    389 	cpu_initclocks();
    390 
    391 	/*
    392 	 * Compute profhz/stathz, and fix profhz if needed.
    393 	 */
    394 	i = stathz ? stathz : hz;
    395 	if (profhz == 0)
    396 		profhz = i;
    397 	psratio = profhz / i;
    398 
    399 #ifdef NTP
    400 	switch (hz) {
    401 	case 60:
    402 	case 64:
    403 		shifthz = SHIFT_SCALE - 6;
    404 		break;
    405 	case 96:
    406 	case 100:
    407 	case 128:
    408 		shifthz = SHIFT_SCALE - 7;
    409 		break;
    410 	case 256:
    411 		shifthz = SHIFT_SCALE - 8;
    412 		break;
    413 	case 512:
    414 		shifthz = SHIFT_SCALE - 9;
    415 		break;
    416 	case 1000:
    417 	case 1024:
    418 		shifthz = SHIFT_SCALE - 10;
    419 		break;
    420 	default:
    421 		panic("weird hz");
    422 	}
    423 	if (fixtick == 0) {
    424 		/*
    425 		 * Give MD code a chance to set this to a better
    426 		 * value; but, if it doesn't, we should.
    427 		 */
    428 		fixtick = (1000000 - (hz*tick));
    429 	}
    430 #endif
    431 }
    432 
    433 /*
    434  * The real-time timer, interrupting hz times per second.
    435  */
    436 void
    437 hardclock(frame)
    438 	register struct clockframe *frame;
    439 {
    440 	register struct proc *p;
    441 	register int delta;
    442 	extern int tickdelta;
    443 	extern long timedelta;
    444 #ifdef NTP
    445 	register int time_update;
    446 	register int ltemp;
    447 #endif
    448 
    449 	p = curproc;
    450 	if (p) {
    451 		register struct pstats *pstats;
    452 
    453 		/*
    454 		 * Run current process's virtual and profile time, as needed.
    455 		 */
    456 		pstats = p->p_stats;
    457 		if (CLKF_USERMODE(frame) &&
    458 		    timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
    459 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
    460 			psignal(p, SIGVTALRM);
    461 		if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
    462 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
    463 			psignal(p, SIGPROF);
    464 	}
    465 
    466 	/*
    467 	 * If no separate statistics clock is available, run it from here.
    468 	 */
    469 	if (stathz == 0)
    470 		statclock(frame);
    471 
    472 	/*
    473 	 * Increment the time-of-day.  The increment is normally just
    474 	 * ``tick''.  If the machine is one which has a clock frequency
    475 	 * such that ``hz'' would not divide the second evenly into
    476 	 * milliseconds, a periodic adjustment must be applied.  Finally,
    477 	 * if we are still adjusting the time (see adjtime()),
    478 	 * ``tickdelta'' may also be added in.
    479 	 */
    480 	hardclock_ticks++;
    481 	delta = tick;
    482 
    483 #ifndef NTP
    484 	if (tickfix) {
    485 		tickfixcnt += tickfix;
    486 		if (tickfixcnt >= tickfixinterval) {
    487 			delta++;
    488 			tickfixcnt -= tickfixinterval;
    489 		}
    490 	}
    491 #endif /* !NTP */
    492 	/* Imprecise 4bsd adjtime() handling */
    493 	if (timedelta != 0) {
    494 		delta += tickdelta;
    495 		timedelta -= tickdelta;
    496 	}
    497 
    498 #ifdef notyet
    499 	microset();
    500 #endif
    501 
    502 #ifndef NTP
    503 	BUMPTIME(&time, delta);		/* XXX Now done using NTP code below */
    504 #endif
    505 	BUMPTIME(&mono_time, delta);
    506 
    507 #ifdef NTP
    508 	time_update = delta;
    509 
    510 	/*
    511 	 * Compute the phase adjustment. If the low-order bits
    512 	 * (time_phase) of the update overflow, bump the high-order bits
    513 	 * (time_update).
    514 	 */
    515 	time_phase += time_adj;
    516 	if (time_phase <= -FINEUSEC) {
    517 		ltemp = -time_phase >> SHIFT_SCALE;
    518 		time_phase += ltemp << SHIFT_SCALE;
    519 		time_update -= ltemp;
    520 	} else if (time_phase >= FINEUSEC) {
    521 		ltemp = time_phase >> SHIFT_SCALE;
    522 		time_phase -= ltemp << SHIFT_SCALE;
    523 		time_update += ltemp;
    524 	}
    525 
    526 #ifdef HIGHBALL
    527 	/*
    528 	 * If the HIGHBALL board is installed, we need to adjust the
    529 	 * external clock offset in order to close the hardware feedback
    530 	 * loop. This will adjust the external clock phase and frequency
    531 	 * in small amounts. The additional phase noise and frequency
    532 	 * wander this causes should be minimal. We also need to
    533 	 * discipline the kernel time variable, since the PLL is used to
    534 	 * discipline the external clock. If the Highball board is not
    535 	 * present, we discipline kernel time with the PLL as usual. We
    536 	 * assume that the external clock phase adjustment (time_update)
    537 	 * and kernel phase adjustment (clock_cpu) are less than the
    538 	 * value of tick.
    539 	 */
    540 	clock_offset.tv_usec += time_update;
    541 	if (clock_offset.tv_usec >= 1000000) {
    542 		clock_offset.tv_sec++;
    543 		clock_offset.tv_usec -= 1000000;
    544 	}
    545 	if (clock_offset.tv_usec < 0) {
    546 		clock_offset.tv_sec--;
    547 		clock_offset.tv_usec += 1000000;
    548 	}
    549 	time.tv_usec += clock_cpu;
    550 	clock_cpu = 0;
    551 #else
    552 	time.tv_usec += time_update;
    553 #endif /* HIGHBALL */
    554 
    555 	/*
    556 	 * On rollover of the second the phase adjustment to be used for
    557 	 * the next second is calculated. Also, the maximum error is
    558 	 * increased by the tolerance. If the PPS frequency discipline
    559 	 * code is present, the phase is increased to compensate for the
    560 	 * CPU clock oscillator frequency error.
    561 	 *
    562  	 * On a 32-bit machine and given parameters in the timex.h
    563 	 * header file, the maximum phase adjustment is +-512 ms and
    564 	 * maximum frequency offset is a tad less than) +-512 ppm. On a
    565 	 * 64-bit machine, you shouldn't need to ask.
    566 	 */
    567 	if (time.tv_usec >= 1000000) {
    568 		time.tv_usec -= 1000000;
    569 		time.tv_sec++;
    570 		time_maxerror += time_tolerance >> SHIFT_USEC;
    571 
    572 		/*
    573 		 * Leap second processing. If in leap-insert state at
    574 		 * the end of the day, the system clock is set back one
    575 		 * second; if in leap-delete state, the system clock is
    576 		 * set ahead one second. The microtime() routine or
    577 		 * external clock driver will insure that reported time
    578 		 * is always monotonic. The ugly divides should be
    579 		 * replaced.
    580 		 */
    581 		switch (time_state) {
    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 (time.tv_sec % 86400 == 0) {
    591 				time.tv_sec--;
    592 				time_state = TIME_OOP;
    593 			}
    594 			break;
    595 
    596 		case TIME_DEL:
    597 			if ((time.tv_sec + 1) % 86400 == 0) {
    598 				time.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 			break;
    611 		}
    612 
    613 		/*
    614 		 * Compute the phase adjustment for the next second. In
    615 		 * PLL mode, the offset is reduced by a fixed factor
    616 		 * times the time constant. In FLL mode the offset is
    617 		 * used directly. In either mode, the maximum phase
    618 		 * adjustment for each second is clamped so as to spread
    619 		 * the adjustment over not more than the number of
    620 		 * seconds between updates.
    621 		 */
    622 		if (time_offset < 0) {
    623 			ltemp = -time_offset;
    624 			if (!(time_status & STA_FLL))
    625 				ltemp >>= SHIFT_KG + time_constant;
    626 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    627 				ltemp = (MAXPHASE / MINSEC) <<
    628 				    SHIFT_UPDATE;
    629 			time_offset += ltemp;
    630 			time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
    631 		} else if (time_offset > 0) {
    632 			ltemp = time_offset;
    633 			if (!(time_status & STA_FLL))
    634 				ltemp >>= SHIFT_KG + time_constant;
    635 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    636 				ltemp = (MAXPHASE / MINSEC) <<
    637 				    SHIFT_UPDATE;
    638 			time_offset -= ltemp;
    639 			time_adj = ltemp << (shifthz - SHIFT_UPDATE);
    640 		} else
    641 			time_adj = 0;
    642 
    643 		/*
    644 		 * Compute the frequency estimate and additional phase
    645 		 * adjustment due to frequency error for the next
    646 		 * second. When the PPS signal is engaged, gnaw on the
    647 		 * watchdog counter and update the frequency computed by
    648 		 * the pll and the PPS signal.
    649 		 */
    650 #ifdef PPS_SYNC
    651 		pps_valid++;
    652 		if (pps_valid == PPS_VALID) {
    653 			pps_jitter = MAXTIME;
    654 			pps_stabil = MAXFREQ;
    655 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
    656 			    STA_PPSWANDER | STA_PPSERROR);
    657 		}
    658 		ltemp = time_freq + pps_freq;
    659 #else
    660 		ltemp = time_freq;
    661 #endif /* PPS_SYNC */
    662 
    663 		if (ltemp < 0)
    664 			time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
    665 		else
    666 			time_adj += ltemp >> (SHIFT_USEC - shifthz);
    667 		time_adj += (long)fixtick << shifthz;
    668 
    669 		/*
    670 		 * When the CPU clock oscillator frequency is not a
    671 		 * power of 2 in Hz, shifthz is only an approximate
    672 		 * scale factor.
    673 		 *
    674 		 * To determine the adjustment, you can do the following:
    675 		 *   bc -q
    676 		 *   scale=24
    677 		 *   obase=2
    678 		 *   idealhz/realhz
    679 		 * where `idealhz' is the next higher power of 2, and `realhz'
    680 		 * is the actual value.
    681 		 *
    682 		 * Likewise, the error can be calculated with (e.g. for 100Hz):
    683 		 *   bc -q
    684 		 *   scale=24
    685 		 *   ((1+2^-2+2^-5)*realhz-idealhz)/idealhz
    686 		 * (and then multiply by 100 to get %).
    687 		 */
    688 		switch (hz) {
    689 		case 96:
    690 			/* A factor of 1.0101010101 gives about .025% error. */
    691 			if (time_adj < 0) {
    692 				time_adj -= (-time_adj >> 2);
    693 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    694 			} else {
    695 				time_adj += (time_adj >> 2);
    696 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    697 			}
    698 			break;
    699 
    700 		case 100:
    701 			/* A factor of 1.01001 gives about .1% error. */
    702 			if (time_adj < 0)
    703 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    704 			else
    705 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    706 			break;
    707 
    708 		case 60:
    709 			/* A factor of 1.00010001 gives about .025% error. */
    710 			if (time_adj < 0)
    711 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    712 			else
    713 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    714 			break;
    715 
    716 		case 1000:
    717 			 /* A factor of 1.0000011 gives about .055% error. */
    718 			if (time_adj < 0)
    719 				time_adj -= (-time_adj >> 6) + (-time_adj >> 7);
    720 			else
    721 				time_adj += (time_adj >> 6) + (time_adj >> 7);
    722 			break;
    723 		}
    724 
    725 #ifdef EXT_CLOCK
    726 		/*
    727 		 * If an external clock is present, it is necessary to
    728 		 * discipline the kernel time variable anyway, since not
    729 		 * all system components use the microtime() interface.
    730 		 * Here, the time offset between the external clock and
    731 		 * kernel time variable is computed every so often.
    732 		 */
    733 		clock_count++;
    734 		if (clock_count > CLOCK_INTERVAL) {
    735 			clock_count = 0;
    736 			microtime(&clock_ext);
    737 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    738 			delta.tv_usec = clock_ext.tv_usec -
    739 			    time.tv_usec;
    740 			if (delta.tv_usec < 0)
    741 				delta.tv_sec--;
    742 			if (delta.tv_usec >= 500000) {
    743 				delta.tv_usec -= 1000000;
    744 				delta.tv_sec++;
    745 			}
    746 			if (delta.tv_usec < -500000) {
    747 				delta.tv_usec += 1000000;
    748 				delta.tv_sec--;
    749 			}
    750 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    751 			    delta.tv_usec > MAXPHASE) ||
    752 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    753 			    delta.tv_usec < -MAXPHASE)) {
    754 				time = clock_ext;
    755 				delta.tv_sec = 0;
    756 				delta.tv_usec = 0;
    757 			}
    758 #ifdef HIGHBALL
    759 			clock_cpu = delta.tv_usec;
    760 #else /* HIGHBALL */
    761 			hardupdate(delta.tv_usec);
    762 #endif /* HIGHBALL */
    763 		}
    764 #endif /* EXT_CLOCK */
    765 	}
    766 
    767 #endif /* NTP */
    768 
    769 	/*
    770 	 * Process callouts at a very low cpu priority, so we don't keep the
    771 	 * relatively high clock interrupt priority any longer than necessary.
    772 	 */
    773 	if (TAILQ_FIRST(&callwheel[hardclock_ticks & callwheelmask]) != NULL) {
    774 		if (CLKF_BASEPRI(frame)) {
    775 			/*
    776 			 * Save the overhead of a software interrupt;
    777 			 * it will happen as soon as we return, so do
    778 			 * it now.
    779 			 *
    780 			 * NOTE: If we're at ``base priority'', softclock()
    781 			 * was not already running.
    782 			 */
    783 			(void)spllowersoftclock();
    784 			softclock();
    785 		} else
    786 			setsoftclock();
    787 	} else if (softclock_running == 0 &&
    788 		   (softclock_ticks + 1) == hardclock_ticks)
    789 		softclock_ticks++;
    790 }
    791 
    792 /*
    793  * Software (low priority) clock interrupt.
    794  * Run periodic events from timeout queue.
    795  */
    796 /*ARGSUSED*/
    797 void
    798 softclock()
    799 {
    800 	struct callout_queue *bucket;
    801 	struct callout *c;
    802 	void (*func) __P((void *));
    803 	void *arg;
    804 	int s, idx;
    805 	int steps = 0;
    806 
    807 	s = splhigh();
    808 	softclock_running = 1;
    809 
    810 #ifdef CALLWHEEL_STATS
    811 	callwheel_softclocks++;
    812 #endif
    813 
    814 	while (softclock_ticks != hardclock_ticks) {
    815 		softclock_ticks++;
    816 		idx = (int)(softclock_ticks & callwheelmask);
    817 		bucket = &callwheel[idx];
    818 		c = TAILQ_FIRST(bucket);
    819 #ifdef CALLWHEEL_STATS
    820 		if (c == NULL)
    821 			callwheel_softempty++;
    822 #endif
    823 		while (c != NULL) {
    824 #ifdef CALLWHEEL_STATS
    825 			callwheel_softchecks++;
    826 #endif
    827 			if (c->c_time != softclock_ticks) {
    828 				c = TAILQ_NEXT(c, c_link);
    829 				if (++steps >= MAX_SOFTCLOCK_STEPS) {
    830 					nextsoftcheck = c;
    831 					/* Give interrupts a chance. */
    832 					splx(s);
    833 					(void) splhigh();
    834 					c = nextsoftcheck;
    835 					steps = 0;
    836 				}
    837 			} else {
    838 				nextsoftcheck = TAILQ_NEXT(c, c_link);
    839 				TAILQ_REMOVE(bucket, c, c_link);
    840 #ifdef CALLWHEEL_STATS
    841 				callwheel_sizes[idx]--;
    842 				callwheel_fired++;
    843 				callwheel_count--;
    844 #endif
    845 				func = c->c_func;
    846 				arg = c->c_arg;
    847 				c->c_func = NULL;
    848 				c->c_flags &= ~CALLOUT_PENDING;
    849 				splx(s);
    850 				(*func)(arg);
    851 				(void) splhigh();
    852 				steps = 0;
    853 				c = nextsoftcheck;
    854 			}
    855 		}
    856 	}
    857 	nextsoftcheck = NULL;
    858 	softclock_running = 0;
    859 	splx(s);
    860 }
    861 
    862 /*
    863  * callout_startup:
    864  *
    865  *	Initialize the callout subsystem.  Called from main().
    866  *
    867  *	This must be called before cpu_startup(), which calls
    868  *	allocsys(), which is what actually allocates the callwheel.
    869  */
    870 void
    871 callout_startup()
    872 {
    873 
    874 	for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1)
    875 		/* loop */ ;
    876 	callwheelmask = callwheelsize - 1;
    877 }
    878 
    879 /*
    880  * callout_startup1:
    881  *
    882  *	Initialize the callwheel buckets.
    883  */
    884 void
    885 callout_startup1()
    886 {
    887 	int i;
    888 
    889 	for (i = 0; i < callwheelsize; i++)
    890 		TAILQ_INIT(&callwheel[i]);
    891 }
    892 
    893 /*
    894  * callout_init:
    895  *
    896  *	Initialize a callout structure so that it can be used
    897  *	by callout_reset() and callout_stop().
    898  */
    899 void
    900 callout_init(c)
    901 	struct callout *c;
    902 {
    903 
    904 	memset(c, 0, sizeof(*c));
    905 }
    906 
    907 /*
    908  * callout_reset:
    909  *
    910  *	Establish or change a timeout.
    911  */
    912 void
    913 callout_reset(c, ticks, func, arg)
    914 	struct callout *c;
    915 	int ticks;
    916 	void (*func) __P((void *));
    917 	void *arg;
    918 {
    919 	struct callout_queue *bucket;
    920 	int s;
    921 
    922 	if (ticks <= 0)
    923 		ticks = 1;
    924 
    925 	/* Lock out the clock. */
    926 	s = splhigh();
    927 
    928 	/*
    929 	 * If this callout's timer is already running, cancel it
    930 	 * before we modify it.
    931 	 */
    932 	if (c->c_flags & CALLOUT_PENDING) {
    933 		callout_stop(c);
    934 #ifdef CALLWHEEL_STATS
    935 		callwheel_changed++;
    936 #endif
    937 	}
    938 
    939 	c->c_arg = arg;
    940 	c->c_func = func;
    941 	c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING;
    942 	c->c_time = hardclock_ticks + ticks;
    943 
    944 	bucket = &callwheel[c->c_time & callwheelmask];
    945 
    946 #ifdef CALLWHEEL_STATS
    947 	if (TAILQ_FIRST(bucket) != NULL)
    948 		callwheel_collisions++;
    949 #endif
    950 
    951 	TAILQ_INSERT_TAIL(bucket, c, c_link);
    952 
    953 #ifdef CALLWHEEL_STATS
    954 	callwheel_count++;
    955 	callwheel_established++;
    956 	if (++callwheel_sizes[c->c_time & callwheelmask] > callwheel_maxlength)
    957 		callwheel_maxlength =
    958 		    callwheel_sizes[c->c_time & callwheelmask];
    959 #endif
    960 
    961 	splx(s);
    962 }
    963 
    964 /*
    965  * callout_stop:
    966  *
    967  *	Disestablish a timeout.
    968  */
    969 void
    970 callout_stop(c)
    971 	struct callout *c;
    972 {
    973 	int s;
    974 
    975 	/* Lock out the clock. */
    976 	s = splhigh();
    977 
    978 	/*
    979 	 * Don't attempt to delete a callout that's not on the queue.
    980 	 */
    981 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
    982 		c->c_flags &= ~CALLOUT_ACTIVE;
    983 		splx(s);
    984 		return;
    985 	}
    986 
    987 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
    988 
    989 	if (nextsoftcheck == c)
    990 		nextsoftcheck = TAILQ_NEXT(c, c_link);
    991 
    992 	TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_link);
    993 #ifdef CALLWHEEL_STATS
    994 	callwheel_count--;
    995 	callwheel_disestablished++;
    996 	callwheel_sizes[c->c_time & callwheelmask]--;
    997 #endif
    998 
    999 	c->c_func = NULL;
   1000 
   1001 	splx(s);
   1002 }
   1003 
   1004 #ifdef CALLWHEEL_STATS
   1005 /*
   1006  * callout_showstats:
   1007  *
   1008  *	Display callout statistics.  Call it from DDB.
   1009  */
   1010 void
   1011 callout_showstats()
   1012 {
   1013 	u_int64_t curticks;
   1014 	int s;
   1015 
   1016 	s = splclock();
   1017 	curticks = softclock_ticks;
   1018 	splx(s);
   1019 
   1020 	printf("Callwheel statistics:\n");
   1021 	printf("\tCallouts currently queued: %llu\n", callwheel_count);
   1022 	printf("\tCallouts established: %llu\n", callwheel_established);
   1023 	printf("\tCallouts disestablished: %llu\n", callwheel_disestablished);
   1024 	if (callwheel_changed != 0)
   1025 		printf("\t\tOf those, %llu were changes\n", callwheel_changed);
   1026 	printf("\tCallouts that fired: %llu\n", callwheel_fired);
   1027 	printf("\tNumber of buckets: %d\n", callwheelsize);
   1028 	printf("\tNumber of hash collisions: %d\n", callwheel_collisions);
   1029 	printf("\tMaximum hash chain length: %d\n", callwheel_maxlength);
   1030 	printf("\tSoftclocks: %llu, Softchecks: %llu\n",
   1031 	    callwheel_softclocks, callwheel_softchecks);
   1032 	printf("\t\tEmpty buckets seen: %llu\n", callwheel_softempty);
   1033 }
   1034 #endif
   1035 
   1036 /*
   1037  * Compute number of hz until specified time.  Used to compute second
   1038  * argument to callout_reset() from an absolute time.
   1039  */
   1040 int
   1041 hzto(tv)
   1042 	struct timeval *tv;
   1043 {
   1044 	register long ticks, sec;
   1045 	int s;
   1046 
   1047 	/*
   1048 	 * If number of microseconds will fit in 32 bit arithmetic,
   1049 	 * then compute number of microseconds to time and scale to
   1050 	 * ticks.  Otherwise just compute number of hz in time, rounding
   1051 	 * times greater than representible to maximum value.  (We must
   1052 	 * compute in microseconds, because hz can be greater than 1000,
   1053 	 * and thus tick can be less than one millisecond).
   1054 	 *
   1055 	 * Delta times less than 14 hours can be computed ``exactly''.
   1056 	 * (Note that if hz would yeild a non-integral number of us per
   1057 	 * tick, i.e. tickfix is nonzero, timouts can be a tick longer
   1058 	 * than they should be.)  Maximum value for any timeout in 10ms
   1059 	 * ticks is 250 days.
   1060 	 */
   1061 	s = splclock();
   1062 	sec = tv->tv_sec - time.tv_sec;
   1063 	if (sec <= 0x7fffffff / 1000000 - 1)
   1064 		ticks = ((tv->tv_sec - time.tv_sec) * 1000000 +
   1065 			(tv->tv_usec - time.tv_usec)) / tick;
   1066 	else if (sec <= 0x7fffffff / hz)
   1067 		ticks = sec * hz;
   1068 	else
   1069 		ticks = 0x7fffffff;
   1070 	splx(s);
   1071 	return (ticks);
   1072 }
   1073 
   1074 /*
   1075  * Start profiling on a process.
   1076  *
   1077  * Kernel profiling passes proc0 which never exits and hence
   1078  * keeps the profile clock running constantly.
   1079  */
   1080 void
   1081 startprofclock(p)
   1082 	register struct proc *p;
   1083 {
   1084 	int s;
   1085 
   1086 	if ((p->p_flag & P_PROFIL) == 0) {
   1087 		p->p_flag |= P_PROFIL;
   1088 		if (++profprocs == 1 && stathz != 0) {
   1089 			s = splstatclock();
   1090 			psdiv = pscnt = psratio;
   1091 			setstatclockrate(profhz);
   1092 			splx(s);
   1093 		}
   1094 	}
   1095 }
   1096 
   1097 /*
   1098  * Stop profiling on a process.
   1099  */
   1100 void
   1101 stopprofclock(p)
   1102 	register struct proc *p;
   1103 {
   1104 	int s;
   1105 
   1106 	if (p->p_flag & P_PROFIL) {
   1107 		p->p_flag &= ~P_PROFIL;
   1108 		if (--profprocs == 0 && stathz != 0) {
   1109 			s = splstatclock();
   1110 			psdiv = pscnt = 1;
   1111 			setstatclockrate(stathz);
   1112 			splx(s);
   1113 		}
   1114 	}
   1115 }
   1116 
   1117 /*
   1118  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1119  * do process and kernel statistics.
   1120  */
   1121 void
   1122 statclock(frame)
   1123 	register struct clockframe *frame;
   1124 {
   1125 #ifdef GPROF
   1126 	register struct gmonparam *g;
   1127 	register int i;
   1128 #endif
   1129 	static int schedclk;
   1130 	register struct proc *p;
   1131 
   1132 	if (CLKF_USERMODE(frame)) {
   1133 		p = curproc;
   1134 		if (p->p_flag & P_PROFIL)
   1135 			addupc_intr(p, CLKF_PC(frame), 1);
   1136 		if (--pscnt > 0)
   1137 			return;
   1138 		/*
   1139 		 * Came from user mode; CPU was in user state.
   1140 		 * If this process is being profiled record the tick.
   1141 		 */
   1142 		p->p_uticks++;
   1143 		if (p->p_nice > NZERO)
   1144 			cp_time[CP_NICE]++;
   1145 		else
   1146 			cp_time[CP_USER]++;
   1147 	} else {
   1148 #ifdef GPROF
   1149 		/*
   1150 		 * Kernel statistics are just like addupc_intr, only easier.
   1151 		 */
   1152 		g = &_gmonparam;
   1153 		if (g->state == GMON_PROF_ON) {
   1154 			i = CLKF_PC(frame) - g->lowpc;
   1155 			if (i < g->textsize) {
   1156 				i /= HISTFRACTION * sizeof(*g->kcount);
   1157 				g->kcount[i]++;
   1158 			}
   1159 		}
   1160 #endif
   1161 		if (--pscnt > 0)
   1162 			return;
   1163 		/*
   1164 		 * Came from kernel mode, so we were:
   1165 		 * - handling an interrupt,
   1166 		 * - doing syscall or trap work on behalf of the current
   1167 		 *   user process, or
   1168 		 * - spinning in the idle loop.
   1169 		 * Whichever it is, charge the time as appropriate.
   1170 		 * Note that we charge interrupts to the current process,
   1171 		 * regardless of whether they are ``for'' that process,
   1172 		 * so that we know how much of its real time was spent
   1173 		 * in ``non-process'' (i.e., interrupt) work.
   1174 		 */
   1175 		p = curproc;
   1176 		if (CLKF_INTR(frame)) {
   1177 			if (p != NULL)
   1178 				p->p_iticks++;
   1179 			cp_time[CP_INTR]++;
   1180 		} else if (p != NULL) {
   1181 			p->p_sticks++;
   1182 			cp_time[CP_SYS]++;
   1183 		} else
   1184 			cp_time[CP_IDLE]++;
   1185 	}
   1186 	pscnt = psdiv;
   1187 
   1188 	if (p != NULL) {
   1189 		++p->p_cpticks;
   1190 		/*
   1191 		 * If no schedclock is provided, call it here at ~~12-25 Hz,
   1192 		 * ~~16 Hz is best
   1193 		 */
   1194 		if(schedhz == 0)
   1195 			if ((++schedclk & 3) == 0)
   1196 				schedclock(p);
   1197 	}
   1198 }
   1199 
   1200 
   1201 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1202 
   1203 /*
   1204  * hardupdate() - local clock update
   1205  *
   1206  * This routine is called by ntp_adjtime() to update the local clock
   1207  * phase and frequency. The implementation is of an adaptive-parameter,
   1208  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1209  * time and frequency offset estimates for each call. If the kernel PPS
   1210  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1211  * determines the new time offset, instead of the calling argument.
   1212  * Presumably, calls to ntp_adjtime() occur only when the caller
   1213  * believes the local clock is valid within some bound (+-128 ms with
   1214  * NTP). If the caller's time is far different than the PPS time, an
   1215  * argument will ensue, and it's not clear who will lose.
   1216  *
   1217  * For uncompensated quartz crystal oscillatores and nominal update
   1218  * intervals less than 1024 s, operation should be in phase-lock mode
   1219  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1220  * intervals greater than thiss, operation should be in frequency-lock
   1221  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1222  *
   1223  * Note: splclock() is in effect.
   1224  */
   1225 void
   1226 hardupdate(offset)
   1227 	long offset;
   1228 {
   1229 	long ltemp, mtemp;
   1230 
   1231 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1232 		return;
   1233 	ltemp = offset;
   1234 #ifdef PPS_SYNC
   1235 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1236 		ltemp = pps_offset;
   1237 #endif /* PPS_SYNC */
   1238 
   1239 	/*
   1240 	 * Scale the phase adjustment and clamp to the operating range.
   1241 	 */
   1242 	if (ltemp > MAXPHASE)
   1243 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1244 	else if (ltemp < -MAXPHASE)
   1245 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1246 	else
   1247 		time_offset = ltemp << SHIFT_UPDATE;
   1248 
   1249 	/*
   1250 	 * Select whether the frequency is to be controlled and in which
   1251 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1252 	 * multiply/divide should be replaced someday.
   1253 	 */
   1254 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1255 		time_reftime = time.tv_sec;
   1256 	mtemp = time.tv_sec - time_reftime;
   1257 	time_reftime = time.tv_sec;
   1258 	if (time_status & STA_FLL) {
   1259 		if (mtemp >= MINSEC) {
   1260 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1261 			    SHIFT_UPDATE));
   1262 			if (ltemp < 0)
   1263 				time_freq -= -ltemp >> SHIFT_KH;
   1264 			else
   1265 				time_freq += ltemp >> SHIFT_KH;
   1266 		}
   1267 	} else {
   1268 		if (mtemp < MAXSEC) {
   1269 			ltemp *= mtemp;
   1270 			if (ltemp < 0)
   1271 				time_freq -= -ltemp >> (time_constant +
   1272 				    time_constant + SHIFT_KF -
   1273 				    SHIFT_USEC);
   1274 			else
   1275 				time_freq += ltemp >> (time_constant +
   1276 				    time_constant + SHIFT_KF -
   1277 				    SHIFT_USEC);
   1278 		}
   1279 	}
   1280 	if (time_freq > time_tolerance)
   1281 		time_freq = time_tolerance;
   1282 	else if (time_freq < -time_tolerance)
   1283 		time_freq = -time_tolerance;
   1284 }
   1285 
   1286 #ifdef PPS_SYNC
   1287 /*
   1288  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1289  *
   1290  * This routine is called at each PPS interrupt in order to discipline
   1291  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1292  * and leaves it in a handy spot for the hardclock() routine. It
   1293  * integrates successive PPS phase differences and calculates the
   1294  * frequency offset. This is used in hardclock() to discipline the CPU
   1295  * clock oscillator so that intrinsic frequency error is cancelled out.
   1296  * The code requires the caller to capture the time and hardware counter
   1297  * value at the on-time PPS signal transition.
   1298  *
   1299  * Note that, on some Unix systems, this routine runs at an interrupt
   1300  * priority level higher than the timer interrupt routine hardclock().
   1301  * Therefore, the variables used are distinct from the hardclock()
   1302  * variables, except for certain exceptions: The PPS frequency pps_freq
   1303  * and phase pps_offset variables are determined by this routine and
   1304  * updated atomically. The time_tolerance variable can be considered a
   1305  * constant, since it is infrequently changed, and then only when the
   1306  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1307  * once per second by hardclock() and is atomically cleared in this
   1308  * routine.
   1309  */
   1310 void
   1311 hardpps(tvp, usec)
   1312 	struct timeval *tvp;		/* time at PPS */
   1313 	long usec;			/* hardware counter at PPS */
   1314 {
   1315 	long u_usec, v_usec, bigtick;
   1316 	long cal_sec, cal_usec;
   1317 
   1318 	/*
   1319 	 * An occasional glitch can be produced when the PPS interrupt
   1320 	 * occurs in the hardclock() routine before the time variable is
   1321 	 * updated. Here the offset is discarded when the difference
   1322 	 * between it and the last one is greater than tick/2, but not
   1323 	 * if the interval since the first discard exceeds 30 s.
   1324 	 */
   1325 	time_status |= STA_PPSSIGNAL;
   1326 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1327 	pps_valid = 0;
   1328 	u_usec = -tvp->tv_usec;
   1329 	if (u_usec < -500000)
   1330 		u_usec += 1000000;
   1331 	v_usec = pps_offset - u_usec;
   1332 	if (v_usec < 0)
   1333 		v_usec = -v_usec;
   1334 	if (v_usec > (tick >> 1)) {
   1335 		if (pps_glitch > MAXGLITCH) {
   1336 			pps_glitch = 0;
   1337 			pps_tf[2] = u_usec;
   1338 			pps_tf[1] = u_usec;
   1339 		} else {
   1340 			pps_glitch++;
   1341 			u_usec = pps_offset;
   1342 		}
   1343 	} else
   1344 		pps_glitch = 0;
   1345 
   1346 	/*
   1347 	 * A three-stage median filter is used to help deglitch the pps
   1348 	 * time. The median sample becomes the time offset estimate; the
   1349 	 * difference between the other two samples becomes the time
   1350 	 * dispersion (jitter) estimate.
   1351 	 */
   1352 	pps_tf[2] = pps_tf[1];
   1353 	pps_tf[1] = pps_tf[0];
   1354 	pps_tf[0] = u_usec;
   1355 	if (pps_tf[0] > pps_tf[1]) {
   1356 		if (pps_tf[1] > pps_tf[2]) {
   1357 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1358 			v_usec = pps_tf[0] - pps_tf[2];
   1359 		} else if (pps_tf[2] > pps_tf[0]) {
   1360 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1361 			v_usec = pps_tf[2] - pps_tf[1];
   1362 		} else {
   1363 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1364 			v_usec = pps_tf[0] - pps_tf[1];
   1365 		}
   1366 	} else {
   1367 		if (pps_tf[1] < pps_tf[2]) {
   1368 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1369 			v_usec = pps_tf[2] - pps_tf[0];
   1370 		} else  if (pps_tf[2] < pps_tf[0]) {
   1371 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1372 			v_usec = pps_tf[1] - pps_tf[2];
   1373 		} else {
   1374 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1375 			v_usec = pps_tf[1] - pps_tf[0];
   1376 		}
   1377 	}
   1378 	if (v_usec > MAXTIME)
   1379 		pps_jitcnt++;
   1380 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1381 	if (v_usec < 0)
   1382 		pps_jitter -= -v_usec >> PPS_AVG;
   1383 	else
   1384 		pps_jitter += v_usec >> PPS_AVG;
   1385 	if (pps_jitter > (MAXTIME >> 1))
   1386 		time_status |= STA_PPSJITTER;
   1387 
   1388 	/*
   1389 	 * During the calibration interval adjust the starting time when
   1390 	 * the tick overflows. At the end of the interval compute the
   1391 	 * duration of the interval and the difference of the hardware
   1392 	 * counters at the beginning and end of the interval. This code
   1393 	 * is deliciously complicated by the fact valid differences may
   1394 	 * exceed the value of tick when using long calibration
   1395 	 * intervals and small ticks. Note that the counter can be
   1396 	 * greater than tick if caught at just the wrong instant, but
   1397 	 * the values returned and used here are correct.
   1398 	 */
   1399 	bigtick = (long)tick << SHIFT_USEC;
   1400 	pps_usec -= pps_freq;
   1401 	if (pps_usec >= bigtick)
   1402 		pps_usec -= bigtick;
   1403 	if (pps_usec < 0)
   1404 		pps_usec += bigtick;
   1405 	pps_time.tv_sec++;
   1406 	pps_count++;
   1407 	if (pps_count < (1 << pps_shift))
   1408 		return;
   1409 	pps_count = 0;
   1410 	pps_calcnt++;
   1411 	u_usec = usec << SHIFT_USEC;
   1412 	v_usec = pps_usec - u_usec;
   1413 	if (v_usec >= bigtick >> 1)
   1414 		v_usec -= bigtick;
   1415 	if (v_usec < -(bigtick >> 1))
   1416 		v_usec += bigtick;
   1417 	if (v_usec < 0)
   1418 		v_usec = -(-v_usec >> pps_shift);
   1419 	else
   1420 		v_usec = v_usec >> pps_shift;
   1421 	pps_usec = u_usec;
   1422 	cal_sec = tvp->tv_sec;
   1423 	cal_usec = tvp->tv_usec;
   1424 	cal_sec -= pps_time.tv_sec;
   1425 	cal_usec -= pps_time.tv_usec;
   1426 	if (cal_usec < 0) {
   1427 		cal_usec += 1000000;
   1428 		cal_sec--;
   1429 	}
   1430 	pps_time = *tvp;
   1431 
   1432 	/*
   1433 	 * Check for lost interrupts, noise, excessive jitter and
   1434 	 * excessive frequency error. The number of timer ticks during
   1435 	 * the interval may vary +-1 tick. Add to this a margin of one
   1436 	 * tick for the PPS signal jitter and maximum frequency
   1437 	 * deviation. If the limits are exceeded, the calibration
   1438 	 * interval is reset to the minimum and we start over.
   1439 	 */
   1440 	u_usec = (long)tick << 1;
   1441 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1442 	    || (cal_sec == 0 && cal_usec < u_usec))
   1443 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1444 		pps_errcnt++;
   1445 		pps_shift = PPS_SHIFT;
   1446 		pps_intcnt = 0;
   1447 		time_status |= STA_PPSERROR;
   1448 		return;
   1449 	}
   1450 
   1451 	/*
   1452 	 * A three-stage median filter is used to help deglitch the pps
   1453 	 * frequency. The median sample becomes the frequency offset
   1454 	 * estimate; the difference between the other two samples
   1455 	 * becomes the frequency dispersion (stability) estimate.
   1456 	 */
   1457 	pps_ff[2] = pps_ff[1];
   1458 	pps_ff[1] = pps_ff[0];
   1459 	pps_ff[0] = v_usec;
   1460 	if (pps_ff[0] > pps_ff[1]) {
   1461 		if (pps_ff[1] > pps_ff[2]) {
   1462 			u_usec = pps_ff[1];		/* 0 1 2 */
   1463 			v_usec = pps_ff[0] - pps_ff[2];
   1464 		} else if (pps_ff[2] > pps_ff[0]) {
   1465 			u_usec = pps_ff[0];		/* 2 0 1 */
   1466 			v_usec = pps_ff[2] - pps_ff[1];
   1467 		} else {
   1468 			u_usec = pps_ff[2];		/* 0 2 1 */
   1469 			v_usec = pps_ff[0] - pps_ff[1];
   1470 		}
   1471 	} else {
   1472 		if (pps_ff[1] < pps_ff[2]) {
   1473 			u_usec = pps_ff[1];		/* 2 1 0 */
   1474 			v_usec = pps_ff[2] - pps_ff[0];
   1475 		} else  if (pps_ff[2] < pps_ff[0]) {
   1476 			u_usec = pps_ff[0];		/* 1 0 2 */
   1477 			v_usec = pps_ff[1] - pps_ff[2];
   1478 		} else {
   1479 			u_usec = pps_ff[2];		/* 1 2 0 */
   1480 			v_usec = pps_ff[1] - pps_ff[0];
   1481 		}
   1482 	}
   1483 
   1484 	/*
   1485 	 * Here the frequency dispersion (stability) is updated. If it
   1486 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1487 	 * offset is updated as well, but clamped to the tolerance. It
   1488 	 * will be processed later by the hardclock() routine.
   1489 	 */
   1490 	v_usec = (v_usec >> 1) - pps_stabil;
   1491 	if (v_usec < 0)
   1492 		pps_stabil -= -v_usec >> PPS_AVG;
   1493 	else
   1494 		pps_stabil += v_usec >> PPS_AVG;
   1495 	if (pps_stabil > MAXFREQ >> 2) {
   1496 		pps_stbcnt++;
   1497 		time_status |= STA_PPSWANDER;
   1498 		return;
   1499 	}
   1500 	if (time_status & STA_PPSFREQ) {
   1501 		if (u_usec < 0) {
   1502 			pps_freq -= -u_usec >> PPS_AVG;
   1503 			if (pps_freq < -time_tolerance)
   1504 				pps_freq = -time_tolerance;
   1505 			u_usec = -u_usec;
   1506 		} else {
   1507 			pps_freq += u_usec >> PPS_AVG;
   1508 			if (pps_freq > time_tolerance)
   1509 				pps_freq = time_tolerance;
   1510 		}
   1511 	}
   1512 
   1513 	/*
   1514 	 * Here the calibration interval is adjusted. If the maximum
   1515 	 * time difference is greater than tick / 4, reduce the interval
   1516 	 * by half. If this is not the case for four consecutive
   1517 	 * intervals, double the interval.
   1518 	 */
   1519 	if (u_usec << pps_shift > bigtick >> 2) {
   1520 		pps_intcnt = 0;
   1521 		if (pps_shift > PPS_SHIFT)
   1522 			pps_shift--;
   1523 	} else if (pps_intcnt >= 4) {
   1524 		pps_intcnt = 0;
   1525 		if (pps_shift < PPS_SHIFTMAX)
   1526 			pps_shift++;
   1527 	} else
   1528 		pps_intcnt++;
   1529 }
   1530 #endif /* PPS_SYNC */
   1531 #endif /* NTP  */
   1532 
   1533 
   1534 /*
   1535  * Return information about system clocks.
   1536  */
   1537 int
   1538 sysctl_clockrate(where, sizep)
   1539 	register char *where;
   1540 	size_t *sizep;
   1541 {
   1542 	struct clockinfo clkinfo;
   1543 
   1544 	/*
   1545 	 * Construct clockinfo structure.
   1546 	 */
   1547 	clkinfo.tick = tick;
   1548 	clkinfo.tickadj = tickadj;
   1549 	clkinfo.hz = hz;
   1550 	clkinfo.profhz = profhz;
   1551 	clkinfo.stathz = stathz ? stathz : hz;
   1552 	return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
   1553 }
   1554