Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.56
      1 /*	$NetBSD: kern_clock.c,v 1.56 2000/05/29 14:58:59 mycroft 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 	volatile struct timeval *tp = (t); \
    303 	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 	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 	struct clockframe *frame;
    439 {
    440 	struct proc *p;
    441 	int delta;
    442 	extern int tickdelta;
    443 	extern long timedelta;
    444 #ifdef NTP
    445 	int time_update;
    446 	int ltemp;
    447 #endif
    448 
    449 	p = curproc;
    450 	if (p) {
    451 		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 244ppm 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 1.010001111010111 gives about 1ppm
    702 			   error. */
    703 			if (time_adj < 0) {
    704 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    705 				time_adj += (-time_adj >> 10);
    706 			} else {
    707 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    708 				time_adj -= (time_adj >> 10);
    709 			}
    710 			break;
    711 
    712 		case 60:
    713 			/* A factor of 1.00010001 gives about 244ppm error. */
    714 			if (time_adj < 0)
    715 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    716 			else
    717 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    718 			break;
    719 
    720 		case 1000:
    721 			/* A factor of 1.000001100010100001 gives about 50ppm
    722 			   error. */
    723 			if (time_adj < 0) {
    724 				time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
    725 				time_adj -= (-time_adj >> 7);
    726 			} else {
    727 				time_adj += (time_adj >> 6) + (time_adj >> 11);
    728 				time_adj += (time_adj >> 7);
    729 			}
    730 			break;
    731 
    732 		case 1200:
    733 			/* A factor of 1.1011010011100001 gives about 64ppm
    734 			   error. */
    735 			if (time_adj < 0) {
    736 				time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
    737 				time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
    738 			} else {
    739 				time_adj += (time_adj >> 1) + (time_adj >> 6);
    740 				time_adj += (time_adj >> 3) + (time_adj >> 10);
    741 			}
    742 			break;
    743 		}
    744 
    745 #ifdef EXT_CLOCK
    746 		/*
    747 		 * If an external clock is present, it is necessary to
    748 		 * discipline the kernel time variable anyway, since not
    749 		 * all system components use the microtime() interface.
    750 		 * Here, the time offset between the external clock and
    751 		 * kernel time variable is computed every so often.
    752 		 */
    753 		clock_count++;
    754 		if (clock_count > CLOCK_INTERVAL) {
    755 			clock_count = 0;
    756 			microtime(&clock_ext);
    757 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    758 			delta.tv_usec = clock_ext.tv_usec -
    759 			    time.tv_usec;
    760 			if (delta.tv_usec < 0)
    761 				delta.tv_sec--;
    762 			if (delta.tv_usec >= 500000) {
    763 				delta.tv_usec -= 1000000;
    764 				delta.tv_sec++;
    765 			}
    766 			if (delta.tv_usec < -500000) {
    767 				delta.tv_usec += 1000000;
    768 				delta.tv_sec--;
    769 			}
    770 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    771 			    delta.tv_usec > MAXPHASE) ||
    772 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    773 			    delta.tv_usec < -MAXPHASE)) {
    774 				time = clock_ext;
    775 				delta.tv_sec = 0;
    776 				delta.tv_usec = 0;
    777 			}
    778 #ifdef HIGHBALL
    779 			clock_cpu = delta.tv_usec;
    780 #else /* HIGHBALL */
    781 			hardupdate(delta.tv_usec);
    782 #endif /* HIGHBALL */
    783 		}
    784 #endif /* EXT_CLOCK */
    785 	}
    786 
    787 #endif /* NTP */
    788 
    789 	/*
    790 	 * Process callouts at a very low cpu priority, so we don't keep the
    791 	 * relatively high clock interrupt priority any longer than necessary.
    792 	 */
    793 	if (TAILQ_FIRST(&callwheel[hardclock_ticks & callwheelmask]) != NULL) {
    794 		if (CLKF_BASEPRI(frame)) {
    795 			/*
    796 			 * Save the overhead of a software interrupt;
    797 			 * it will happen as soon as we return, so do
    798 			 * it now.
    799 			 *
    800 			 * NOTE: If we're at ``base priority'', softclock()
    801 			 * was not already running.
    802 			 */
    803 			(void)spllowersoftclock();
    804 			softclock();
    805 		} else
    806 			setsoftclock();
    807 	} else if (softclock_running == 0 &&
    808 		   (softclock_ticks + 1) == hardclock_ticks)
    809 		softclock_ticks++;
    810 }
    811 
    812 /*
    813  * Software (low priority) clock interrupt.
    814  * Run periodic events from timeout queue.
    815  */
    816 /*ARGSUSED*/
    817 void
    818 softclock()
    819 {
    820 	struct callout_queue *bucket;
    821 	struct callout *c;
    822 	void (*func) __P((void *));
    823 	void *arg;
    824 	int s, idx;
    825 	int steps = 0;
    826 
    827 	s = splhigh();
    828 	softclock_running = 1;
    829 
    830 #ifdef CALLWHEEL_STATS
    831 	callwheel_softclocks++;
    832 #endif
    833 
    834 	while (softclock_ticks != hardclock_ticks) {
    835 		softclock_ticks++;
    836 		idx = (int)(softclock_ticks & callwheelmask);
    837 		bucket = &callwheel[idx];
    838 		c = TAILQ_FIRST(bucket);
    839 #ifdef CALLWHEEL_STATS
    840 		if (c == NULL)
    841 			callwheel_softempty++;
    842 #endif
    843 		while (c != NULL) {
    844 #ifdef CALLWHEEL_STATS
    845 			callwheel_softchecks++;
    846 #endif
    847 			if (c->c_time != softclock_ticks) {
    848 				c = TAILQ_NEXT(c, c_link);
    849 				if (++steps >= MAX_SOFTCLOCK_STEPS) {
    850 					nextsoftcheck = c;
    851 					/* Give interrupts a chance. */
    852 					splx(s);
    853 					(void) splhigh();
    854 					c = nextsoftcheck;
    855 					steps = 0;
    856 				}
    857 			} else {
    858 				nextsoftcheck = TAILQ_NEXT(c, c_link);
    859 				TAILQ_REMOVE(bucket, c, c_link);
    860 #ifdef CALLWHEEL_STATS
    861 				callwheel_sizes[idx]--;
    862 				callwheel_fired++;
    863 				callwheel_count--;
    864 #endif
    865 				func = c->c_func;
    866 				arg = c->c_arg;
    867 				c->c_func = NULL;
    868 				c->c_flags &= ~CALLOUT_PENDING;
    869 				splx(s);
    870 				(*func)(arg);
    871 				(void) splhigh();
    872 				steps = 0;
    873 				c = nextsoftcheck;
    874 			}
    875 		}
    876 	}
    877 	nextsoftcheck = NULL;
    878 	softclock_running = 0;
    879 	splx(s);
    880 }
    881 
    882 /*
    883  * callout_setsize:
    884  *
    885  *	Determine how many callwheels are necessary and
    886  *	set hash mask.  Called from allocsys().
    887  */
    888 void
    889 callout_setsize()
    890 {
    891 
    892 	for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1)
    893 		/* loop */ ;
    894 	callwheelmask = callwheelsize - 1;
    895 }
    896 
    897 /*
    898  * callout_startup:
    899  *
    900  *	Initialize the callwheel buckets.
    901  */
    902 void
    903 callout_startup()
    904 {
    905 	int i;
    906 
    907 	for (i = 0; i < callwheelsize; i++)
    908 		TAILQ_INIT(&callwheel[i]);
    909 }
    910 
    911 /*
    912  * callout_init:
    913  *
    914  *	Initialize a callout structure so that it can be used
    915  *	by callout_reset() and callout_stop().
    916  */
    917 void
    918 callout_init(c)
    919 	struct callout *c;
    920 {
    921 
    922 	memset(c, 0, sizeof(*c));
    923 }
    924 
    925 /*
    926  * callout_reset:
    927  *
    928  *	Establish or change a timeout.
    929  */
    930 void
    931 callout_reset(c, ticks, func, arg)
    932 	struct callout *c;
    933 	int ticks;
    934 	void (*func) __P((void *));
    935 	void *arg;
    936 {
    937 	struct callout_queue *bucket;
    938 	int s;
    939 
    940 	if (ticks <= 0)
    941 		ticks = 1;
    942 
    943 	/* Lock out the clock. */
    944 	s = splhigh();
    945 
    946 	/*
    947 	 * If this callout's timer is already running, cancel it
    948 	 * before we modify it.
    949 	 */
    950 	if (c->c_flags & CALLOUT_PENDING) {
    951 		callout_stop(c);
    952 #ifdef CALLWHEEL_STATS
    953 		callwheel_changed++;
    954 #endif
    955 	}
    956 
    957 	c->c_arg = arg;
    958 	c->c_func = func;
    959 	c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING;
    960 	c->c_time = hardclock_ticks + ticks;
    961 
    962 	bucket = &callwheel[c->c_time & callwheelmask];
    963 
    964 #ifdef CALLWHEEL_STATS
    965 	if (TAILQ_FIRST(bucket) != NULL)
    966 		callwheel_collisions++;
    967 #endif
    968 
    969 	TAILQ_INSERT_TAIL(bucket, c, c_link);
    970 
    971 #ifdef CALLWHEEL_STATS
    972 	callwheel_count++;
    973 	callwheel_established++;
    974 	if (++callwheel_sizes[c->c_time & callwheelmask] > callwheel_maxlength)
    975 		callwheel_maxlength =
    976 		    callwheel_sizes[c->c_time & callwheelmask];
    977 #endif
    978 
    979 	splx(s);
    980 }
    981 
    982 /*
    983  * callout_stop:
    984  *
    985  *	Disestablish a timeout.
    986  */
    987 void
    988 callout_stop(c)
    989 	struct callout *c;
    990 {
    991 	int s;
    992 
    993 	/* Lock out the clock. */
    994 	s = splhigh();
    995 
    996 	/*
    997 	 * Don't attempt to delete a callout that's not on the queue.
    998 	 */
    999 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
   1000 		c->c_flags &= ~CALLOUT_ACTIVE;
   1001 		splx(s);
   1002 		return;
   1003 	}
   1004 
   1005 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
   1006 
   1007 	if (nextsoftcheck == c)
   1008 		nextsoftcheck = TAILQ_NEXT(c, c_link);
   1009 
   1010 	TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_link);
   1011 #ifdef CALLWHEEL_STATS
   1012 	callwheel_count--;
   1013 	callwheel_disestablished++;
   1014 	callwheel_sizes[c->c_time & callwheelmask]--;
   1015 #endif
   1016 
   1017 	c->c_func = NULL;
   1018 
   1019 	splx(s);
   1020 }
   1021 
   1022 #ifdef CALLWHEEL_STATS
   1023 /*
   1024  * callout_showstats:
   1025  *
   1026  *	Display callout statistics.  Call it from DDB.
   1027  */
   1028 void
   1029 callout_showstats()
   1030 {
   1031 	u_int64_t curticks;
   1032 	int s;
   1033 
   1034 	s = splclock();
   1035 	curticks = softclock_ticks;
   1036 	splx(s);
   1037 
   1038 	printf("Callwheel statistics:\n");
   1039 	printf("\tCallouts currently queued: %llu\n", callwheel_count);
   1040 	printf("\tCallouts established: %llu\n", callwheel_established);
   1041 	printf("\tCallouts disestablished: %llu\n", callwheel_disestablished);
   1042 	if (callwheel_changed != 0)
   1043 		printf("\t\tOf those, %llu were changes\n", callwheel_changed);
   1044 	printf("\tCallouts that fired: %llu\n", callwheel_fired);
   1045 	printf("\tNumber of buckets: %d\n", callwheelsize);
   1046 	printf("\tNumber of hash collisions: %d\n", callwheel_collisions);
   1047 	printf("\tMaximum hash chain length: %d\n", callwheel_maxlength);
   1048 	printf("\tSoftclocks: %llu, Softchecks: %llu\n",
   1049 	    callwheel_softclocks, callwheel_softchecks);
   1050 	printf("\t\tEmpty buckets seen: %llu\n", callwheel_softempty);
   1051 }
   1052 #endif
   1053 
   1054 /*
   1055  * Compute number of hz until specified time.  Used to compute second
   1056  * argument to callout_reset() from an absolute time.
   1057  */
   1058 int
   1059 hzto(tv)
   1060 	struct timeval *tv;
   1061 {
   1062 	long ticks, sec;
   1063 	int s;
   1064 
   1065 	/*
   1066 	 * If number of microseconds will fit in 32 bit arithmetic,
   1067 	 * then compute number of microseconds to time and scale to
   1068 	 * ticks.  Otherwise just compute number of hz in time, rounding
   1069 	 * times greater than representible to maximum value.  (We must
   1070 	 * compute in microseconds, because hz can be greater than 1000,
   1071 	 * and thus tick can be less than one millisecond).
   1072 	 *
   1073 	 * Delta times less than 14 hours can be computed ``exactly''.
   1074 	 * (Note that if hz would yeild a non-integral number of us per
   1075 	 * tick, i.e. tickfix is nonzero, timouts can be a tick longer
   1076 	 * than they should be.)  Maximum value for any timeout in 10ms
   1077 	 * ticks is 250 days.
   1078 	 */
   1079 	s = splclock();
   1080 	sec = tv->tv_sec - time.tv_sec;
   1081 	if (sec <= 0x7fffffff / 1000000 - 1)
   1082 		ticks = ((tv->tv_sec - time.tv_sec) * 1000000 +
   1083 			(tv->tv_usec - time.tv_usec)) / tick;
   1084 	else if (sec <= 0x7fffffff / hz)
   1085 		ticks = sec * hz;
   1086 	else
   1087 		ticks = 0x7fffffff;
   1088 	splx(s);
   1089 	return (ticks);
   1090 }
   1091 
   1092 /*
   1093  * Start profiling on a process.
   1094  *
   1095  * Kernel profiling passes proc0 which never exits and hence
   1096  * keeps the profile clock running constantly.
   1097  */
   1098 void
   1099 startprofclock(p)
   1100 	struct proc *p;
   1101 {
   1102 	int s;
   1103 
   1104 	if ((p->p_flag & P_PROFIL) == 0) {
   1105 		p->p_flag |= P_PROFIL;
   1106 		if (++profprocs == 1 && stathz != 0) {
   1107 			s = splstatclock();
   1108 			psdiv = pscnt = psratio;
   1109 			setstatclockrate(profhz);
   1110 			splx(s);
   1111 		}
   1112 	}
   1113 }
   1114 
   1115 /*
   1116  * Stop profiling on a process.
   1117  */
   1118 void
   1119 stopprofclock(p)
   1120 	struct proc *p;
   1121 {
   1122 	int s;
   1123 
   1124 	if (p->p_flag & P_PROFIL) {
   1125 		p->p_flag &= ~P_PROFIL;
   1126 		if (--profprocs == 0 && stathz != 0) {
   1127 			s = splstatclock();
   1128 			psdiv = pscnt = 1;
   1129 			setstatclockrate(stathz);
   1130 			splx(s);
   1131 		}
   1132 	}
   1133 }
   1134 
   1135 /*
   1136  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1137  * do process and kernel statistics.
   1138  */
   1139 void
   1140 statclock(frame)
   1141 	struct clockframe *frame;
   1142 {
   1143 #ifdef GPROF
   1144 	struct gmonparam *g;
   1145 	int i;
   1146 #endif
   1147 	static int schedclk;
   1148 	struct proc *p;
   1149 
   1150 	if (CLKF_USERMODE(frame)) {
   1151 		p = curproc;
   1152 		if (p->p_flag & P_PROFIL)
   1153 			addupc_intr(p, CLKF_PC(frame), 1);
   1154 		if (--pscnt > 0)
   1155 			return;
   1156 		/*
   1157 		 * Came from user mode; CPU was in user state.
   1158 		 * If this process is being profiled record the tick.
   1159 		 */
   1160 		p->p_uticks++;
   1161 		if (p->p_nice > NZERO)
   1162 			cp_time[CP_NICE]++;
   1163 		else
   1164 			cp_time[CP_USER]++;
   1165 	} else {
   1166 #ifdef GPROF
   1167 		/*
   1168 		 * Kernel statistics are just like addupc_intr, only easier.
   1169 		 */
   1170 		g = &_gmonparam;
   1171 		if (g->state == GMON_PROF_ON) {
   1172 			i = CLKF_PC(frame) - g->lowpc;
   1173 			if (i < g->textsize) {
   1174 				i /= HISTFRACTION * sizeof(*g->kcount);
   1175 				g->kcount[i]++;
   1176 			}
   1177 		}
   1178 #endif
   1179 		if (--pscnt > 0)
   1180 			return;
   1181 		/*
   1182 		 * Came from kernel mode, so we were:
   1183 		 * - handling an interrupt,
   1184 		 * - doing syscall or trap work on behalf of the current
   1185 		 *   user process, or
   1186 		 * - spinning in the idle loop.
   1187 		 * Whichever it is, charge the time as appropriate.
   1188 		 * Note that we charge interrupts to the current process,
   1189 		 * regardless of whether they are ``for'' that process,
   1190 		 * so that we know how much of its real time was spent
   1191 		 * in ``non-process'' (i.e., interrupt) work.
   1192 		 */
   1193 		p = curproc;
   1194 		if (CLKF_INTR(frame)) {
   1195 			if (p != NULL)
   1196 				p->p_iticks++;
   1197 			cp_time[CP_INTR]++;
   1198 		} else if (p != NULL) {
   1199 			p->p_sticks++;
   1200 			cp_time[CP_SYS]++;
   1201 		} else
   1202 			cp_time[CP_IDLE]++;
   1203 	}
   1204 	pscnt = psdiv;
   1205 
   1206 	if (p != NULL) {
   1207 		++p->p_cpticks;
   1208 		/*
   1209 		 * If no schedclock is provided, call it here at ~~12-25 Hz,
   1210 		 * ~~16 Hz is best
   1211 		 */
   1212 		if(schedhz == 0)
   1213 			if ((++schedclk & 3) == 0)
   1214 				schedclock(p);
   1215 	}
   1216 }
   1217 
   1218 
   1219 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1220 
   1221 /*
   1222  * hardupdate() - local clock update
   1223  *
   1224  * This routine is called by ntp_adjtime() to update the local clock
   1225  * phase and frequency. The implementation is of an adaptive-parameter,
   1226  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1227  * time and frequency offset estimates for each call. If the kernel PPS
   1228  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1229  * determines the new time offset, instead of the calling argument.
   1230  * Presumably, calls to ntp_adjtime() occur only when the caller
   1231  * believes the local clock is valid within some bound (+-128 ms with
   1232  * NTP). If the caller's time is far different than the PPS time, an
   1233  * argument will ensue, and it's not clear who will lose.
   1234  *
   1235  * For uncompensated quartz crystal oscillatores and nominal update
   1236  * intervals less than 1024 s, operation should be in phase-lock mode
   1237  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1238  * intervals greater than thiss, operation should be in frequency-lock
   1239  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1240  *
   1241  * Note: splclock() is in effect.
   1242  */
   1243 void
   1244 hardupdate(offset)
   1245 	long offset;
   1246 {
   1247 	long ltemp, mtemp;
   1248 
   1249 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1250 		return;
   1251 	ltemp = offset;
   1252 #ifdef PPS_SYNC
   1253 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1254 		ltemp = pps_offset;
   1255 #endif /* PPS_SYNC */
   1256 
   1257 	/*
   1258 	 * Scale the phase adjustment and clamp to the operating range.
   1259 	 */
   1260 	if (ltemp > MAXPHASE)
   1261 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1262 	else if (ltemp < -MAXPHASE)
   1263 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1264 	else
   1265 		time_offset = ltemp << SHIFT_UPDATE;
   1266 
   1267 	/*
   1268 	 * Select whether the frequency is to be controlled and in which
   1269 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1270 	 * multiply/divide should be replaced someday.
   1271 	 */
   1272 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1273 		time_reftime = time.tv_sec;
   1274 	mtemp = time.tv_sec - time_reftime;
   1275 	time_reftime = time.tv_sec;
   1276 	if (time_status & STA_FLL) {
   1277 		if (mtemp >= MINSEC) {
   1278 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1279 			    SHIFT_UPDATE));
   1280 			if (ltemp < 0)
   1281 				time_freq -= -ltemp >> SHIFT_KH;
   1282 			else
   1283 				time_freq += ltemp >> SHIFT_KH;
   1284 		}
   1285 	} else {
   1286 		if (mtemp < MAXSEC) {
   1287 			ltemp *= mtemp;
   1288 			if (ltemp < 0)
   1289 				time_freq -= -ltemp >> (time_constant +
   1290 				    time_constant + SHIFT_KF -
   1291 				    SHIFT_USEC);
   1292 			else
   1293 				time_freq += ltemp >> (time_constant +
   1294 				    time_constant + SHIFT_KF -
   1295 				    SHIFT_USEC);
   1296 		}
   1297 	}
   1298 	if (time_freq > time_tolerance)
   1299 		time_freq = time_tolerance;
   1300 	else if (time_freq < -time_tolerance)
   1301 		time_freq = -time_tolerance;
   1302 }
   1303 
   1304 #ifdef PPS_SYNC
   1305 /*
   1306  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1307  *
   1308  * This routine is called at each PPS interrupt in order to discipline
   1309  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1310  * and leaves it in a handy spot for the hardclock() routine. It
   1311  * integrates successive PPS phase differences and calculates the
   1312  * frequency offset. This is used in hardclock() to discipline the CPU
   1313  * clock oscillator so that intrinsic frequency error is cancelled out.
   1314  * The code requires the caller to capture the time and hardware counter
   1315  * value at the on-time PPS signal transition.
   1316  *
   1317  * Note that, on some Unix systems, this routine runs at an interrupt
   1318  * priority level higher than the timer interrupt routine hardclock().
   1319  * Therefore, the variables used are distinct from the hardclock()
   1320  * variables, except for certain exceptions: The PPS frequency pps_freq
   1321  * and phase pps_offset variables are determined by this routine and
   1322  * updated atomically. The time_tolerance variable can be considered a
   1323  * constant, since it is infrequently changed, and then only when the
   1324  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1325  * once per second by hardclock() and is atomically cleared in this
   1326  * routine.
   1327  */
   1328 void
   1329 hardpps(tvp, usec)
   1330 	struct timeval *tvp;		/* time at PPS */
   1331 	long usec;			/* hardware counter at PPS */
   1332 {
   1333 	long u_usec, v_usec, bigtick;
   1334 	long cal_sec, cal_usec;
   1335 
   1336 	/*
   1337 	 * An occasional glitch can be produced when the PPS interrupt
   1338 	 * occurs in the hardclock() routine before the time variable is
   1339 	 * updated. Here the offset is discarded when the difference
   1340 	 * between it and the last one is greater than tick/2, but not
   1341 	 * if the interval since the first discard exceeds 30 s.
   1342 	 */
   1343 	time_status |= STA_PPSSIGNAL;
   1344 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1345 	pps_valid = 0;
   1346 	u_usec = -tvp->tv_usec;
   1347 	if (u_usec < -500000)
   1348 		u_usec += 1000000;
   1349 	v_usec = pps_offset - u_usec;
   1350 	if (v_usec < 0)
   1351 		v_usec = -v_usec;
   1352 	if (v_usec > (tick >> 1)) {
   1353 		if (pps_glitch > MAXGLITCH) {
   1354 			pps_glitch = 0;
   1355 			pps_tf[2] = u_usec;
   1356 			pps_tf[1] = u_usec;
   1357 		} else {
   1358 			pps_glitch++;
   1359 			u_usec = pps_offset;
   1360 		}
   1361 	} else
   1362 		pps_glitch = 0;
   1363 
   1364 	/*
   1365 	 * A three-stage median filter is used to help deglitch the pps
   1366 	 * time. The median sample becomes the time offset estimate; the
   1367 	 * difference between the other two samples becomes the time
   1368 	 * dispersion (jitter) estimate.
   1369 	 */
   1370 	pps_tf[2] = pps_tf[1];
   1371 	pps_tf[1] = pps_tf[0];
   1372 	pps_tf[0] = u_usec;
   1373 	if (pps_tf[0] > pps_tf[1]) {
   1374 		if (pps_tf[1] > pps_tf[2]) {
   1375 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1376 			v_usec = pps_tf[0] - pps_tf[2];
   1377 		} else if (pps_tf[2] > pps_tf[0]) {
   1378 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1379 			v_usec = pps_tf[2] - pps_tf[1];
   1380 		} else {
   1381 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1382 			v_usec = pps_tf[0] - pps_tf[1];
   1383 		}
   1384 	} else {
   1385 		if (pps_tf[1] < pps_tf[2]) {
   1386 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1387 			v_usec = pps_tf[2] - pps_tf[0];
   1388 		} else  if (pps_tf[2] < pps_tf[0]) {
   1389 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1390 			v_usec = pps_tf[1] - pps_tf[2];
   1391 		} else {
   1392 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1393 			v_usec = pps_tf[1] - pps_tf[0];
   1394 		}
   1395 	}
   1396 	if (v_usec > MAXTIME)
   1397 		pps_jitcnt++;
   1398 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1399 	if (v_usec < 0)
   1400 		pps_jitter -= -v_usec >> PPS_AVG;
   1401 	else
   1402 		pps_jitter += v_usec >> PPS_AVG;
   1403 	if (pps_jitter > (MAXTIME >> 1))
   1404 		time_status |= STA_PPSJITTER;
   1405 
   1406 	/*
   1407 	 * During the calibration interval adjust the starting time when
   1408 	 * the tick overflows. At the end of the interval compute the
   1409 	 * duration of the interval and the difference of the hardware
   1410 	 * counters at the beginning and end of the interval. This code
   1411 	 * is deliciously complicated by the fact valid differences may
   1412 	 * exceed the value of tick when using long calibration
   1413 	 * intervals and small ticks. Note that the counter can be
   1414 	 * greater than tick if caught at just the wrong instant, but
   1415 	 * the values returned and used here are correct.
   1416 	 */
   1417 	bigtick = (long)tick << SHIFT_USEC;
   1418 	pps_usec -= pps_freq;
   1419 	if (pps_usec >= bigtick)
   1420 		pps_usec -= bigtick;
   1421 	if (pps_usec < 0)
   1422 		pps_usec += bigtick;
   1423 	pps_time.tv_sec++;
   1424 	pps_count++;
   1425 	if (pps_count < (1 << pps_shift))
   1426 		return;
   1427 	pps_count = 0;
   1428 	pps_calcnt++;
   1429 	u_usec = usec << SHIFT_USEC;
   1430 	v_usec = pps_usec - u_usec;
   1431 	if (v_usec >= bigtick >> 1)
   1432 		v_usec -= bigtick;
   1433 	if (v_usec < -(bigtick >> 1))
   1434 		v_usec += bigtick;
   1435 	if (v_usec < 0)
   1436 		v_usec = -(-v_usec >> pps_shift);
   1437 	else
   1438 		v_usec = v_usec >> pps_shift;
   1439 	pps_usec = u_usec;
   1440 	cal_sec = tvp->tv_sec;
   1441 	cal_usec = tvp->tv_usec;
   1442 	cal_sec -= pps_time.tv_sec;
   1443 	cal_usec -= pps_time.tv_usec;
   1444 	if (cal_usec < 0) {
   1445 		cal_usec += 1000000;
   1446 		cal_sec--;
   1447 	}
   1448 	pps_time = *tvp;
   1449 
   1450 	/*
   1451 	 * Check for lost interrupts, noise, excessive jitter and
   1452 	 * excessive frequency error. The number of timer ticks during
   1453 	 * the interval may vary +-1 tick. Add to this a margin of one
   1454 	 * tick for the PPS signal jitter and maximum frequency
   1455 	 * deviation. If the limits are exceeded, the calibration
   1456 	 * interval is reset to the minimum and we start over.
   1457 	 */
   1458 	u_usec = (long)tick << 1;
   1459 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1460 	    || (cal_sec == 0 && cal_usec < u_usec))
   1461 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1462 		pps_errcnt++;
   1463 		pps_shift = PPS_SHIFT;
   1464 		pps_intcnt = 0;
   1465 		time_status |= STA_PPSERROR;
   1466 		return;
   1467 	}
   1468 
   1469 	/*
   1470 	 * A three-stage median filter is used to help deglitch the pps
   1471 	 * frequency. The median sample becomes the frequency offset
   1472 	 * estimate; the difference between the other two samples
   1473 	 * becomes the frequency dispersion (stability) estimate.
   1474 	 */
   1475 	pps_ff[2] = pps_ff[1];
   1476 	pps_ff[1] = pps_ff[0];
   1477 	pps_ff[0] = v_usec;
   1478 	if (pps_ff[0] > pps_ff[1]) {
   1479 		if (pps_ff[1] > pps_ff[2]) {
   1480 			u_usec = pps_ff[1];		/* 0 1 2 */
   1481 			v_usec = pps_ff[0] - pps_ff[2];
   1482 		} else if (pps_ff[2] > pps_ff[0]) {
   1483 			u_usec = pps_ff[0];		/* 2 0 1 */
   1484 			v_usec = pps_ff[2] - pps_ff[1];
   1485 		} else {
   1486 			u_usec = pps_ff[2];		/* 0 2 1 */
   1487 			v_usec = pps_ff[0] - pps_ff[1];
   1488 		}
   1489 	} else {
   1490 		if (pps_ff[1] < pps_ff[2]) {
   1491 			u_usec = pps_ff[1];		/* 2 1 0 */
   1492 			v_usec = pps_ff[2] - pps_ff[0];
   1493 		} else  if (pps_ff[2] < pps_ff[0]) {
   1494 			u_usec = pps_ff[0];		/* 1 0 2 */
   1495 			v_usec = pps_ff[1] - pps_ff[2];
   1496 		} else {
   1497 			u_usec = pps_ff[2];		/* 1 2 0 */
   1498 			v_usec = pps_ff[1] - pps_ff[0];
   1499 		}
   1500 	}
   1501 
   1502 	/*
   1503 	 * Here the frequency dispersion (stability) is updated. If it
   1504 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1505 	 * offset is updated as well, but clamped to the tolerance. It
   1506 	 * will be processed later by the hardclock() routine.
   1507 	 */
   1508 	v_usec = (v_usec >> 1) - pps_stabil;
   1509 	if (v_usec < 0)
   1510 		pps_stabil -= -v_usec >> PPS_AVG;
   1511 	else
   1512 		pps_stabil += v_usec >> PPS_AVG;
   1513 	if (pps_stabil > MAXFREQ >> 2) {
   1514 		pps_stbcnt++;
   1515 		time_status |= STA_PPSWANDER;
   1516 		return;
   1517 	}
   1518 	if (time_status & STA_PPSFREQ) {
   1519 		if (u_usec < 0) {
   1520 			pps_freq -= -u_usec >> PPS_AVG;
   1521 			if (pps_freq < -time_tolerance)
   1522 				pps_freq = -time_tolerance;
   1523 			u_usec = -u_usec;
   1524 		} else {
   1525 			pps_freq += u_usec >> PPS_AVG;
   1526 			if (pps_freq > time_tolerance)
   1527 				pps_freq = time_tolerance;
   1528 		}
   1529 	}
   1530 
   1531 	/*
   1532 	 * Here the calibration interval is adjusted. If the maximum
   1533 	 * time difference is greater than tick / 4, reduce the interval
   1534 	 * by half. If this is not the case for four consecutive
   1535 	 * intervals, double the interval.
   1536 	 */
   1537 	if (u_usec << pps_shift > bigtick >> 2) {
   1538 		pps_intcnt = 0;
   1539 		if (pps_shift > PPS_SHIFT)
   1540 			pps_shift--;
   1541 	} else if (pps_intcnt >= 4) {
   1542 		pps_intcnt = 0;
   1543 		if (pps_shift < PPS_SHIFTMAX)
   1544 			pps_shift++;
   1545 	} else
   1546 		pps_intcnt++;
   1547 }
   1548 #endif /* PPS_SYNC */
   1549 #endif /* NTP  */
   1550 
   1551 
   1552 /*
   1553  * Return information about system clocks.
   1554  */
   1555 int
   1556 sysctl_clockrate(where, sizep)
   1557 	char *where;
   1558 	size_t *sizep;
   1559 {
   1560 	struct clockinfo clkinfo;
   1561 
   1562 	/*
   1563 	 * Construct clockinfo structure.
   1564 	 */
   1565 	clkinfo.tick = tick;
   1566 	clkinfo.tickadj = tickadj;
   1567 	clkinfo.hz = hz;
   1568 	clkinfo.profhz = profhz;
   1569 	clkinfo.stathz = stathz ? stathz : hz;
   1570 	return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
   1571 }
   1572