Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.63
      1 /*	$NetBSD: kern_clock.c,v 1.63 2000/08/01 04:57:29 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 <uvm/uvm_extern.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(void)
    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 1:
    402 		shifthz = SHIFT_SCALE - 0;
    403 		break;
    404 	case 2:
    405 		shifthz = SHIFT_SCALE - 1;
    406 		break;
    407 	case 4:
    408 		shifthz = SHIFT_SCALE - 2;
    409 		break;
    410 	case 8:
    411 		shifthz = SHIFT_SCALE - 3;
    412 		break;
    413 	case 16:
    414 		shifthz = SHIFT_SCALE - 4;
    415 		break;
    416 	case 32:
    417 		shifthz = SHIFT_SCALE - 5;
    418 		break;
    419 	case 60:
    420 	case 64:
    421 		shifthz = SHIFT_SCALE - 6;
    422 		break;
    423 	case 96:
    424 	case 100:
    425 	case 128:
    426 		shifthz = SHIFT_SCALE - 7;
    427 		break;
    428 	case 256:
    429 		shifthz = SHIFT_SCALE - 8;
    430 		break;
    431 	case 512:
    432 		shifthz = SHIFT_SCALE - 9;
    433 		break;
    434 	case 1000:
    435 	case 1024:
    436 		shifthz = SHIFT_SCALE - 10;
    437 		break;
    438 	case 1200:
    439 	case 2048:
    440 		shifthz = SHIFT_SCALE - 11;
    441 		break;
    442 	case 4096:
    443 		shifthz = SHIFT_SCALE - 12;
    444 		break;
    445 	case 8192:
    446 		shifthz = SHIFT_SCALE - 13;
    447 		break;
    448 	case 16384:
    449 		shifthz = SHIFT_SCALE - 14;
    450 		break;
    451 	case 32768:
    452 		shifthz = SHIFT_SCALE - 15;
    453 		break;
    454 	case 65536:
    455 		shifthz = SHIFT_SCALE - 16;
    456 		break;
    457 	default:
    458 		panic("weird hz");
    459 	}
    460 	if (fixtick == 0) {
    461 		/*
    462 		 * Give MD code a chance to set this to a better
    463 		 * value; but, if it doesn't, we should.
    464 		 */
    465 		fixtick = (1000000 - (hz*tick));
    466 	}
    467 #endif
    468 }
    469 
    470 /*
    471  * The real-time timer, interrupting hz times per second.
    472  */
    473 void
    474 hardclock(struct clockframe *frame)
    475 {
    476 	struct proc *p;
    477 	int delta;
    478 	extern int tickdelta;
    479 	extern long timedelta;
    480 #ifdef NTP
    481 	int time_update;
    482 	int ltemp;
    483 #endif
    484 
    485 	p = curproc;
    486 	if (p) {
    487 		struct pstats *pstats;
    488 
    489 		/*
    490 		 * Run current process's virtual and profile time, as needed.
    491 		 */
    492 		pstats = p->p_stats;
    493 		if (CLKF_USERMODE(frame) &&
    494 		    timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
    495 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
    496 			psignal(p, SIGVTALRM);
    497 		if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
    498 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
    499 			psignal(p, SIGPROF);
    500 	}
    501 
    502 	/*
    503 	 * If no separate statistics clock is available, run it from here.
    504 	 */
    505 	if (stathz == 0)
    506 		statclock(frame);
    507 
    508 #if defined(MULTIPROCESSOR)
    509 	/*
    510 	 * If we are not the primary CPU, we're not allowed to do
    511 	 * any more work.
    512 	 */
    513 	if (CPU_IS_PRIMARY(curcpu()) == 0)
    514 		return;
    515 #endif
    516 
    517 	/*
    518 	 * Increment the time-of-day.  The increment is normally just
    519 	 * ``tick''.  If the machine is one which has a clock frequency
    520 	 * such that ``hz'' would not divide the second evenly into
    521 	 * milliseconds, a periodic adjustment must be applied.  Finally,
    522 	 * if we are still adjusting the time (see adjtime()),
    523 	 * ``tickdelta'' may also be added in.
    524 	 */
    525 	hardclock_ticks++;
    526 	delta = tick;
    527 
    528 #ifndef NTP
    529 	if (tickfix) {
    530 		tickfixcnt += tickfix;
    531 		if (tickfixcnt >= tickfixinterval) {
    532 			delta++;
    533 			tickfixcnt -= tickfixinterval;
    534 		}
    535 	}
    536 #endif /* !NTP */
    537 	/* Imprecise 4bsd adjtime() handling */
    538 	if (timedelta != 0) {
    539 		delta += tickdelta;
    540 		timedelta -= tickdelta;
    541 	}
    542 
    543 #ifdef notyet
    544 	microset();
    545 #endif
    546 
    547 #ifndef NTP
    548 	BUMPTIME(&time, delta);		/* XXX Now done using NTP code below */
    549 #endif
    550 	BUMPTIME(&mono_time, delta);
    551 
    552 #ifdef NTP
    553 	time_update = delta;
    554 
    555 	/*
    556 	 * Compute the phase adjustment. If the low-order bits
    557 	 * (time_phase) of the update overflow, bump the high-order bits
    558 	 * (time_update).
    559 	 */
    560 	time_phase += time_adj;
    561 	if (time_phase <= -FINEUSEC) {
    562 		ltemp = -time_phase >> SHIFT_SCALE;
    563 		time_phase += ltemp << SHIFT_SCALE;
    564 		time_update -= ltemp;
    565 	} else if (time_phase >= FINEUSEC) {
    566 		ltemp = time_phase >> SHIFT_SCALE;
    567 		time_phase -= ltemp << SHIFT_SCALE;
    568 		time_update += ltemp;
    569 	}
    570 
    571 #ifdef HIGHBALL
    572 	/*
    573 	 * If the HIGHBALL board is installed, we need to adjust the
    574 	 * external clock offset in order to close the hardware feedback
    575 	 * loop. This will adjust the external clock phase and frequency
    576 	 * in small amounts. The additional phase noise and frequency
    577 	 * wander this causes should be minimal. We also need to
    578 	 * discipline the kernel time variable, since the PLL is used to
    579 	 * discipline the external clock. If the Highball board is not
    580 	 * present, we discipline kernel time with the PLL as usual. We
    581 	 * assume that the external clock phase adjustment (time_update)
    582 	 * and kernel phase adjustment (clock_cpu) are less than the
    583 	 * value of tick.
    584 	 */
    585 	clock_offset.tv_usec += time_update;
    586 	if (clock_offset.tv_usec >= 1000000) {
    587 		clock_offset.tv_sec++;
    588 		clock_offset.tv_usec -= 1000000;
    589 	}
    590 	if (clock_offset.tv_usec < 0) {
    591 		clock_offset.tv_sec--;
    592 		clock_offset.tv_usec += 1000000;
    593 	}
    594 	time.tv_usec += clock_cpu;
    595 	clock_cpu = 0;
    596 #else
    597 	time.tv_usec += time_update;
    598 #endif /* HIGHBALL */
    599 
    600 	/*
    601 	 * On rollover of the second the phase adjustment to be used for
    602 	 * the next second is calculated. Also, the maximum error is
    603 	 * increased by the tolerance. If the PPS frequency discipline
    604 	 * code is present, the phase is increased to compensate for the
    605 	 * CPU clock oscillator frequency error.
    606 	 *
    607  	 * On a 32-bit machine and given parameters in the timex.h
    608 	 * header file, the maximum phase adjustment is +-512 ms and
    609 	 * maximum frequency offset is a tad less than) +-512 ppm. On a
    610 	 * 64-bit machine, you shouldn't need to ask.
    611 	 */
    612 	if (time.tv_usec >= 1000000) {
    613 		time.tv_usec -= 1000000;
    614 		time.tv_sec++;
    615 		time_maxerror += time_tolerance >> SHIFT_USEC;
    616 
    617 		/*
    618 		 * Leap second processing. If in leap-insert state at
    619 		 * the end of the day, the system clock is set back one
    620 		 * second; if in leap-delete state, the system clock is
    621 		 * set ahead one second. The microtime() routine or
    622 		 * external clock driver will insure that reported time
    623 		 * is always monotonic. The ugly divides should be
    624 		 * replaced.
    625 		 */
    626 		switch (time_state) {
    627 		case TIME_OK:
    628 			if (time_status & STA_INS)
    629 				time_state = TIME_INS;
    630 			else if (time_status & STA_DEL)
    631 				time_state = TIME_DEL;
    632 			break;
    633 
    634 		case TIME_INS:
    635 			if (time.tv_sec % 86400 == 0) {
    636 				time.tv_sec--;
    637 				time_state = TIME_OOP;
    638 			}
    639 			break;
    640 
    641 		case TIME_DEL:
    642 			if ((time.tv_sec + 1) % 86400 == 0) {
    643 				time.tv_sec++;
    644 				time_state = TIME_WAIT;
    645 			}
    646 			break;
    647 
    648 		case TIME_OOP:
    649 			time_state = TIME_WAIT;
    650 			break;
    651 
    652 		case TIME_WAIT:
    653 			if (!(time_status & (STA_INS | STA_DEL)))
    654 				time_state = TIME_OK;
    655 			break;
    656 		}
    657 
    658 		/*
    659 		 * Compute the phase adjustment for the next second. In
    660 		 * PLL mode, the offset is reduced by a fixed factor
    661 		 * times the time constant. In FLL mode the offset is
    662 		 * used directly. In either mode, the maximum phase
    663 		 * adjustment for each second is clamped so as to spread
    664 		 * the adjustment over not more than the number of
    665 		 * seconds between updates.
    666 		 */
    667 		if (time_offset < 0) {
    668 			ltemp = -time_offset;
    669 			if (!(time_status & STA_FLL))
    670 				ltemp >>= SHIFT_KG + time_constant;
    671 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    672 				ltemp = (MAXPHASE / MINSEC) <<
    673 				    SHIFT_UPDATE;
    674 			time_offset += ltemp;
    675 			time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
    676 		} else if (time_offset > 0) {
    677 			ltemp = time_offset;
    678 			if (!(time_status & STA_FLL))
    679 				ltemp >>= SHIFT_KG + time_constant;
    680 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    681 				ltemp = (MAXPHASE / MINSEC) <<
    682 				    SHIFT_UPDATE;
    683 			time_offset -= ltemp;
    684 			time_adj = ltemp << (shifthz - SHIFT_UPDATE);
    685 		} else
    686 			time_adj = 0;
    687 
    688 		/*
    689 		 * Compute the frequency estimate and additional phase
    690 		 * adjustment due to frequency error for the next
    691 		 * second. When the PPS signal is engaged, gnaw on the
    692 		 * watchdog counter and update the frequency computed by
    693 		 * the pll and the PPS signal.
    694 		 */
    695 #ifdef PPS_SYNC
    696 		pps_valid++;
    697 		if (pps_valid == PPS_VALID) {
    698 			pps_jitter = MAXTIME;
    699 			pps_stabil = MAXFREQ;
    700 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
    701 			    STA_PPSWANDER | STA_PPSERROR);
    702 		}
    703 		ltemp = time_freq + pps_freq;
    704 #else
    705 		ltemp = time_freq;
    706 #endif /* PPS_SYNC */
    707 
    708 		if (ltemp < 0)
    709 			time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
    710 		else
    711 			time_adj += ltemp >> (SHIFT_USEC - shifthz);
    712 		time_adj += (long)fixtick << shifthz;
    713 
    714 		/*
    715 		 * When the CPU clock oscillator frequency is not a
    716 		 * power of 2 in Hz, shifthz is only an approximate
    717 		 * scale factor.
    718 		 *
    719 		 * To determine the adjustment, you can do the following:
    720 		 *   bc -q
    721 		 *   scale=24
    722 		 *   obase=2
    723 		 *   idealhz/realhz
    724 		 * where `idealhz' is the next higher power of 2, and `realhz'
    725 		 * is the actual value.  You may need to factor this result
    726 		 * into a sequence of 2 multipliers to get better precision.
    727 		 *
    728 		 * Likewise, the error can be calculated with (e.g. for 100Hz):
    729 		 *   bc -q
    730 		 *   scale=24
    731 		 *   ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz
    732 		 * (and then multiply by 1000000 to get ppm).
    733 		 */
    734 		switch (hz) {
    735 		case 60:
    736 			/* A factor of 1.000100010001 gives about 15ppm
    737 			   error. */
    738 			if (time_adj < 0) {
    739 				time_adj -= (-time_adj >> 4);
    740 				time_adj -= (-time_adj >> 8);
    741 			} else {
    742 				time_adj += (time_adj >> 4);
    743 				time_adj += (time_adj >> 8);
    744 			}
    745 			break;
    746 
    747 		case 96:
    748 			/* A factor of 1.0101010101 gives about 244ppm error. */
    749 			if (time_adj < 0) {
    750 				time_adj -= (-time_adj >> 2);
    751 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    752 			} else {
    753 				time_adj += (time_adj >> 2);
    754 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    755 			}
    756 			break;
    757 
    758 		case 100:
    759 			/* A factor of 1.010001111010111 gives about 1ppm
    760 			   error. */
    761 			if (time_adj < 0) {
    762 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    763 				time_adj += (-time_adj >> 10);
    764 			} else {
    765 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    766 				time_adj -= (time_adj >> 10);
    767 			}
    768 			break;
    769 
    770 		case 1000:
    771 			/* A factor of 1.000001100010100001 gives about 50ppm
    772 			   error. */
    773 			if (time_adj < 0) {
    774 				time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
    775 				time_adj -= (-time_adj >> 7);
    776 			} else {
    777 				time_adj += (time_adj >> 6) + (time_adj >> 11);
    778 				time_adj += (time_adj >> 7);
    779 			}
    780 			break;
    781 
    782 		case 1200:
    783 			/* A factor of 1.1011010011100001 gives about 64ppm
    784 			   error. */
    785 			if (time_adj < 0) {
    786 				time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
    787 				time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
    788 			} else {
    789 				time_adj += (time_adj >> 1) + (time_adj >> 6);
    790 				time_adj += (time_adj >> 3) + (time_adj >> 10);
    791 			}
    792 			break;
    793 		}
    794 
    795 #ifdef EXT_CLOCK
    796 		/*
    797 		 * If an external clock is present, it is necessary to
    798 		 * discipline the kernel time variable anyway, since not
    799 		 * all system components use the microtime() interface.
    800 		 * Here, the time offset between the external clock and
    801 		 * kernel time variable is computed every so often.
    802 		 */
    803 		clock_count++;
    804 		if (clock_count > CLOCK_INTERVAL) {
    805 			clock_count = 0;
    806 			microtime(&clock_ext);
    807 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    808 			delta.tv_usec = clock_ext.tv_usec -
    809 			    time.tv_usec;
    810 			if (delta.tv_usec < 0)
    811 				delta.tv_sec--;
    812 			if (delta.tv_usec >= 500000) {
    813 				delta.tv_usec -= 1000000;
    814 				delta.tv_sec++;
    815 			}
    816 			if (delta.tv_usec < -500000) {
    817 				delta.tv_usec += 1000000;
    818 				delta.tv_sec--;
    819 			}
    820 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    821 			    delta.tv_usec > MAXPHASE) ||
    822 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    823 			    delta.tv_usec < -MAXPHASE)) {
    824 				time = clock_ext;
    825 				delta.tv_sec = 0;
    826 				delta.tv_usec = 0;
    827 			}
    828 #ifdef HIGHBALL
    829 			clock_cpu = delta.tv_usec;
    830 #else /* HIGHBALL */
    831 			hardupdate(delta.tv_usec);
    832 #endif /* HIGHBALL */
    833 		}
    834 #endif /* EXT_CLOCK */
    835 	}
    836 
    837 #endif /* NTP */
    838 
    839 	/*
    840 	 * Process callouts at a very low cpu priority, so we don't keep the
    841 	 * relatively high clock interrupt priority any longer than necessary.
    842 	 */
    843 	if (TAILQ_FIRST(&callwheel[hardclock_ticks & callwheelmask]) != NULL) {
    844 		if (CLKF_BASEPRI(frame)) {
    845 			/*
    846 			 * Save the overhead of a software interrupt;
    847 			 * it will happen as soon as we return, so do
    848 			 * it now.
    849 			 *
    850 			 * NOTE: If we're at ``base priority'', softclock()
    851 			 * was not already running.
    852 			 */
    853 			(void)spllowersoftclock();
    854 			softclock();
    855 		} else
    856 			setsoftclock();
    857 	} else if (softclock_running == 0 &&
    858 		   (softclock_ticks + 1) == hardclock_ticks)
    859 		softclock_ticks++;
    860 }
    861 
    862 /*
    863  * Software (low priority) clock interrupt.
    864  * Run periodic events from timeout queue.
    865  */
    866 /*ARGSUSED*/
    867 void
    868 softclock(void)
    869 {
    870 	struct callout_queue *bucket;
    871 	struct callout *c;
    872 	void (*func)(void *);
    873 	void *arg;
    874 	int s, idx;
    875 	int steps = 0;
    876 
    877 	s = splhigh();
    878 	softclock_running = 1;
    879 
    880 #ifdef CALLWHEEL_STATS
    881 	callwheel_softclocks++;
    882 #endif
    883 
    884 	while (softclock_ticks != hardclock_ticks) {
    885 		softclock_ticks++;
    886 		idx = (int)(softclock_ticks & callwheelmask);
    887 		bucket = &callwheel[idx];
    888 		c = TAILQ_FIRST(bucket);
    889 #ifdef CALLWHEEL_STATS
    890 		if (c == NULL)
    891 			callwheel_softempty++;
    892 #endif
    893 		while (c != NULL) {
    894 #ifdef CALLWHEEL_STATS
    895 			callwheel_softchecks++;
    896 #endif
    897 			if (c->c_time != softclock_ticks) {
    898 				c = TAILQ_NEXT(c, c_link);
    899 				if (++steps >= MAX_SOFTCLOCK_STEPS) {
    900 					nextsoftcheck = c;
    901 					/* Give interrupts a chance. */
    902 					splx(s);
    903 					(void) splhigh();
    904 					c = nextsoftcheck;
    905 					steps = 0;
    906 				}
    907 			} else {
    908 				nextsoftcheck = TAILQ_NEXT(c, c_link);
    909 				TAILQ_REMOVE(bucket, c, c_link);
    910 #ifdef CALLWHEEL_STATS
    911 				callwheel_sizes[idx]--;
    912 				callwheel_fired++;
    913 				callwheel_count--;
    914 #endif
    915 				func = c->c_func;
    916 				arg = c->c_arg;
    917 				c->c_func = NULL;
    918 				c->c_flags &= ~CALLOUT_PENDING;
    919 				splx(s);
    920 				(*func)(arg);
    921 				(void) splhigh();
    922 				steps = 0;
    923 				c = nextsoftcheck;
    924 			}
    925 		}
    926 	}
    927 	nextsoftcheck = NULL;
    928 	softclock_running = 0;
    929 	splx(s);
    930 }
    931 
    932 /*
    933  * callout_setsize:
    934  *
    935  *	Determine how many callwheels are necessary and
    936  *	set hash mask.  Called from allocsys().
    937  */
    938 void
    939 callout_setsize(void)
    940 {
    941 
    942 	for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1)
    943 		/* loop */ ;
    944 	callwheelmask = callwheelsize - 1;
    945 }
    946 
    947 /*
    948  * callout_startup:
    949  *
    950  *	Initialize the callwheel buckets.
    951  */
    952 void
    953 callout_startup(void)
    954 {
    955 	int i;
    956 
    957 	for (i = 0; i < callwheelsize; i++)
    958 		TAILQ_INIT(&callwheel[i]);
    959 }
    960 
    961 /*
    962  * callout_init:
    963  *
    964  *	Initialize a callout structure so that it can be used
    965  *	by callout_reset() and callout_stop().
    966  */
    967 void
    968 callout_init(struct callout *c)
    969 {
    970 
    971 	memset(c, 0, sizeof(*c));
    972 }
    973 
    974 /*
    975  * callout_reset:
    976  *
    977  *	Establish or change a timeout.
    978  */
    979 void
    980 callout_reset(struct callout *c, int ticks, void (*func)(void *), void *arg)
    981 {
    982 	struct callout_queue *bucket;
    983 	int s;
    984 
    985 	if (ticks <= 0)
    986 		ticks = 1;
    987 
    988 	/* Lock out the clock. */
    989 	s = splhigh();
    990 
    991 	/*
    992 	 * If this callout's timer is already running, cancel it
    993 	 * before we modify it.
    994 	 */
    995 	if (c->c_flags & CALLOUT_PENDING) {
    996 		callout_stop(c);
    997 #ifdef CALLWHEEL_STATS
    998 		callwheel_changed++;
    999 #endif
   1000 	}
   1001 
   1002 	c->c_arg = arg;
   1003 	c->c_func = func;
   1004 	c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING;
   1005 	c->c_time = hardclock_ticks + ticks;
   1006 
   1007 	bucket = &callwheel[c->c_time & callwheelmask];
   1008 
   1009 #ifdef CALLWHEEL_STATS
   1010 	if (TAILQ_FIRST(bucket) != NULL)
   1011 		callwheel_collisions++;
   1012 #endif
   1013 
   1014 	TAILQ_INSERT_TAIL(bucket, c, c_link);
   1015 
   1016 #ifdef CALLWHEEL_STATS
   1017 	callwheel_count++;
   1018 	callwheel_established++;
   1019 	if (++callwheel_sizes[c->c_time & callwheelmask] > callwheel_maxlength)
   1020 		callwheel_maxlength =
   1021 		    callwheel_sizes[c->c_time & callwheelmask];
   1022 #endif
   1023 
   1024 	splx(s);
   1025 }
   1026 
   1027 /*
   1028  * callout_stop:
   1029  *
   1030  *	Disestablish a timeout.
   1031  */
   1032 void
   1033 callout_stop(struct callout *c)
   1034 {
   1035 	int s;
   1036 
   1037 	/* Lock out the clock. */
   1038 	s = splhigh();
   1039 
   1040 	/*
   1041 	 * Don't attempt to delete a callout that's not on the queue.
   1042 	 */
   1043 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
   1044 		c->c_flags &= ~CALLOUT_ACTIVE;
   1045 		splx(s);
   1046 		return;
   1047 	}
   1048 
   1049 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
   1050 
   1051 	if (nextsoftcheck == c)
   1052 		nextsoftcheck = TAILQ_NEXT(c, c_link);
   1053 
   1054 	TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_link);
   1055 #ifdef CALLWHEEL_STATS
   1056 	callwheel_count--;
   1057 	callwheel_disestablished++;
   1058 	callwheel_sizes[c->c_time & callwheelmask]--;
   1059 #endif
   1060 
   1061 	c->c_func = NULL;
   1062 
   1063 	splx(s);
   1064 }
   1065 
   1066 #ifdef CALLWHEEL_STATS
   1067 /*
   1068  * callout_showstats:
   1069  *
   1070  *	Display callout statistics.  Call it from DDB.
   1071  */
   1072 void
   1073 callout_showstats(void)
   1074 {
   1075 	u_int64_t curticks;
   1076 	int s;
   1077 
   1078 	s = splclock();
   1079 	curticks = softclock_ticks;
   1080 	splx(s);
   1081 
   1082 	printf("Callwheel statistics:\n");
   1083 	printf("\tCallouts currently queued: %llu\n", callwheel_count);
   1084 	printf("\tCallouts established: %llu\n", callwheel_established);
   1085 	printf("\tCallouts disestablished: %llu\n", callwheel_disestablished);
   1086 	if (callwheel_changed != 0)
   1087 		printf("\t\tOf those, %llu were changes\n", callwheel_changed);
   1088 	printf("\tCallouts that fired: %llu\n", callwheel_fired);
   1089 	printf("\tNumber of buckets: %d\n", callwheelsize);
   1090 	printf("\tNumber of hash collisions: %d\n", callwheel_collisions);
   1091 	printf("\tMaximum hash chain length: %d\n", callwheel_maxlength);
   1092 	printf("\tSoftclocks: %llu, Softchecks: %llu\n",
   1093 	    callwheel_softclocks, callwheel_softchecks);
   1094 	printf("\t\tEmpty buckets seen: %llu\n", callwheel_softempty);
   1095 }
   1096 #endif
   1097 
   1098 /*
   1099  * Compute number of hz until specified time.  Used to compute second
   1100  * argument to callout_reset() from an absolute time.
   1101  */
   1102 int
   1103 hzto(struct timeval *tv)
   1104 {
   1105 	unsigned long ticks;
   1106 	long sec, usec;
   1107 	int s;
   1108 
   1109 	/*
   1110 	 * If the number of usecs in the whole seconds part of the time
   1111 	 * difference fits in a long, then the total number of usecs will
   1112 	 * fit in an unsigned long.  Compute the total and convert it to
   1113 	 * ticks, rounding up and adding 1 to allow for the current tick
   1114 	 * to expire.  Rounding also depends on unsigned long arithmetic
   1115 	 * to avoid overflow.
   1116 	 *
   1117 	 * Otherwise, if the number of ticks in the whole seconds part of
   1118 	 * the time difference fits in a long, then convert the parts to
   1119 	 * ticks separately and add, using similar rounding methods and
   1120 	 * overflow avoidance.  This method would work in the previous
   1121 	 * case, but it is slightly slower and assume that hz is integral.
   1122 	 *
   1123 	 * Otherwise, round the time difference down to the maximum
   1124 	 * representable value.
   1125 	 *
   1126 	 * If ints are 32-bit, then the maximum value for any timeout in
   1127 	 * 10ms ticks is 248 days.
   1128 	 */
   1129 	s = splclock();
   1130 	sec = tv->tv_sec - time.tv_sec;
   1131 	usec = tv->tv_usec - time.tv_usec;
   1132 	splx(s);
   1133 
   1134 	if (usec < 0) {
   1135 		sec--;
   1136 		usec += 1000000;
   1137 	}
   1138 
   1139 	if (sec < 0 || (sec == 0 && usec <= 0)) {
   1140 		/*
   1141 		 * Would expire now or in the past.  Return 0 ticks.
   1142 		 * This is different from the legacy hzto() interface,
   1143 		 * and callers need to check for it.
   1144 		 */
   1145 		ticks = 0;
   1146 	} else if (sec <= (LONG_MAX / 1000000))
   1147 		ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1))
   1148 		    / tick) + 1;
   1149 	else if (sec <= (LONG_MAX / hz))
   1150 		ticks = (sec * hz) +
   1151 		    (((unsigned long)usec + (tick - 1)) / tick) + 1;
   1152 	else
   1153 		ticks = LONG_MAX;
   1154 
   1155 	if (ticks > INT_MAX)
   1156 		ticks = INT_MAX;
   1157 
   1158 	return ((int)ticks);
   1159 }
   1160 
   1161 /*
   1162  * Start profiling on a process.
   1163  *
   1164  * Kernel profiling passes proc0 which never exits and hence
   1165  * keeps the profile clock running constantly.
   1166  */
   1167 void
   1168 startprofclock(struct proc *p)
   1169 {
   1170 	int s;
   1171 
   1172 	if ((p->p_flag & P_PROFIL) == 0) {
   1173 		p->p_flag |= P_PROFIL;
   1174 		if (++profprocs == 1 && stathz != 0) {
   1175 			s = splstatclock();
   1176 			psdiv = pscnt = psratio;
   1177 			setstatclockrate(profhz);
   1178 			splx(s);
   1179 		}
   1180 	}
   1181 }
   1182 
   1183 /*
   1184  * Stop profiling on a process.
   1185  */
   1186 void
   1187 stopprofclock(struct proc *p)
   1188 {
   1189 	int s;
   1190 
   1191 	if (p->p_flag & P_PROFIL) {
   1192 		p->p_flag &= ~P_PROFIL;
   1193 		if (--profprocs == 0 && stathz != 0) {
   1194 			s = splstatclock();
   1195 			psdiv = pscnt = 1;
   1196 			setstatclockrate(stathz);
   1197 			splx(s);
   1198 		}
   1199 	}
   1200 }
   1201 
   1202 /*
   1203  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1204  * do process and kernel statistics.
   1205  */
   1206 void
   1207 statclock(struct clockframe *frame)
   1208 {
   1209 #ifdef GPROF
   1210 	struct gmonparam *g;
   1211 	int i;
   1212 #endif
   1213 	struct cpu_info *ci = curcpu();
   1214 	struct schedstate_percpu *spc = &ci->ci_schedstate;
   1215 	struct proc *p;
   1216 
   1217 	if (CLKF_USERMODE(frame)) {
   1218 		p = curproc;
   1219 		if (p->p_flag & P_PROFIL)
   1220 			addupc_intr(p, CLKF_PC(frame), 1);
   1221 		if (--pscnt > 0)
   1222 			return;
   1223 		/*
   1224 		 * Came from user mode; CPU was in user state.
   1225 		 * If this process is being profiled record the tick.
   1226 		 */
   1227 		p->p_uticks++;
   1228 		if (p->p_nice > NZERO)
   1229 			spc->spc_cp_time[CP_NICE]++;
   1230 		else
   1231 			spc->spc_cp_time[CP_USER]++;
   1232 	} else {
   1233 #ifdef GPROF
   1234 		/*
   1235 		 * Kernel statistics are just like addupc_intr, only easier.
   1236 		 */
   1237 		g = &_gmonparam;
   1238 		if (g->state == GMON_PROF_ON) {
   1239 			i = CLKF_PC(frame) - g->lowpc;
   1240 			if (i < g->textsize) {
   1241 				i /= HISTFRACTION * sizeof(*g->kcount);
   1242 				g->kcount[i]++;
   1243 			}
   1244 		}
   1245 #endif
   1246 		if (--pscnt > 0)
   1247 			return;
   1248 		/*
   1249 		 * Came from kernel mode, so we were:
   1250 		 * - handling an interrupt,
   1251 		 * - doing syscall or trap work on behalf of the current
   1252 		 *   user process, or
   1253 		 * - spinning in the idle loop.
   1254 		 * Whichever it is, charge the time as appropriate.
   1255 		 * Note that we charge interrupts to the current process,
   1256 		 * regardless of whether they are ``for'' that process,
   1257 		 * so that we know how much of its real time was spent
   1258 		 * in ``non-process'' (i.e., interrupt) work.
   1259 		 */
   1260 		p = curproc;
   1261 		if (CLKF_INTR(frame)) {
   1262 			if (p != NULL)
   1263 				p->p_iticks++;
   1264 			spc->spc_cp_time[CP_INTR]++;
   1265 		} else if (p != NULL) {
   1266 			p->p_sticks++;
   1267 			spc->spc_cp_time[CP_SYS]++;
   1268 		} else
   1269 			spc->spc_cp_time[CP_IDLE]++;
   1270 	}
   1271 	pscnt = psdiv;
   1272 
   1273 	if (p != NULL) {
   1274 		++p->p_cpticks;
   1275 		/*
   1276 		 * If no separate schedclock is provided, call it here
   1277 		 * at ~~12-25 Hz, ~~16 Hz is best
   1278 		 */
   1279 		if (schedhz == 0)
   1280 			if ((++ci->ci_schedstate.spc_schedticks & 3) == 0)
   1281 				schedclock(p);
   1282 	}
   1283 }
   1284 
   1285 
   1286 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1287 
   1288 /*
   1289  * hardupdate() - local clock update
   1290  *
   1291  * This routine is called by ntp_adjtime() to update the local clock
   1292  * phase and frequency. The implementation is of an adaptive-parameter,
   1293  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1294  * time and frequency offset estimates for each call. If the kernel PPS
   1295  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1296  * determines the new time offset, instead of the calling argument.
   1297  * Presumably, calls to ntp_adjtime() occur only when the caller
   1298  * believes the local clock is valid within some bound (+-128 ms with
   1299  * NTP). If the caller's time is far different than the PPS time, an
   1300  * argument will ensue, and it's not clear who will lose.
   1301  *
   1302  * For uncompensated quartz crystal oscillatores and nominal update
   1303  * intervals less than 1024 s, operation should be in phase-lock mode
   1304  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1305  * intervals greater than thiss, operation should be in frequency-lock
   1306  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1307  *
   1308  * Note: splclock() is in effect.
   1309  */
   1310 void
   1311 hardupdate(long offset)
   1312 {
   1313 	long ltemp, mtemp;
   1314 
   1315 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1316 		return;
   1317 	ltemp = offset;
   1318 #ifdef PPS_SYNC
   1319 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1320 		ltemp = pps_offset;
   1321 #endif /* PPS_SYNC */
   1322 
   1323 	/*
   1324 	 * Scale the phase adjustment and clamp to the operating range.
   1325 	 */
   1326 	if (ltemp > MAXPHASE)
   1327 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1328 	else if (ltemp < -MAXPHASE)
   1329 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1330 	else
   1331 		time_offset = ltemp << SHIFT_UPDATE;
   1332 
   1333 	/*
   1334 	 * Select whether the frequency is to be controlled and in which
   1335 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1336 	 * multiply/divide should be replaced someday.
   1337 	 */
   1338 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1339 		time_reftime = time.tv_sec;
   1340 	mtemp = time.tv_sec - time_reftime;
   1341 	time_reftime = time.tv_sec;
   1342 	if (time_status & STA_FLL) {
   1343 		if (mtemp >= MINSEC) {
   1344 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1345 			    SHIFT_UPDATE));
   1346 			if (ltemp < 0)
   1347 				time_freq -= -ltemp >> SHIFT_KH;
   1348 			else
   1349 				time_freq += ltemp >> SHIFT_KH;
   1350 		}
   1351 	} else {
   1352 		if (mtemp < MAXSEC) {
   1353 			ltemp *= mtemp;
   1354 			if (ltemp < 0)
   1355 				time_freq -= -ltemp >> (time_constant +
   1356 				    time_constant + SHIFT_KF -
   1357 				    SHIFT_USEC);
   1358 			else
   1359 				time_freq += ltemp >> (time_constant +
   1360 				    time_constant + SHIFT_KF -
   1361 				    SHIFT_USEC);
   1362 		}
   1363 	}
   1364 	if (time_freq > time_tolerance)
   1365 		time_freq = time_tolerance;
   1366 	else if (time_freq < -time_tolerance)
   1367 		time_freq = -time_tolerance;
   1368 }
   1369 
   1370 #ifdef PPS_SYNC
   1371 /*
   1372  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1373  *
   1374  * This routine is called at each PPS interrupt in order to discipline
   1375  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1376  * and leaves it in a handy spot for the hardclock() routine. It
   1377  * integrates successive PPS phase differences and calculates the
   1378  * frequency offset. This is used in hardclock() to discipline the CPU
   1379  * clock oscillator so that intrinsic frequency error is cancelled out.
   1380  * The code requires the caller to capture the time and hardware counter
   1381  * value at the on-time PPS signal transition.
   1382  *
   1383  * Note that, on some Unix systems, this routine runs at an interrupt
   1384  * priority level higher than the timer interrupt routine hardclock().
   1385  * Therefore, the variables used are distinct from the hardclock()
   1386  * variables, except for certain exceptions: The PPS frequency pps_freq
   1387  * and phase pps_offset variables are determined by this routine and
   1388  * updated atomically. The time_tolerance variable can be considered a
   1389  * constant, since it is infrequently changed, and then only when the
   1390  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1391  * once per second by hardclock() and is atomically cleared in this
   1392  * routine.
   1393  */
   1394 void
   1395 hardpps(struct timeval *tvp,		/* time at PPS */
   1396 	long usec			/* hardware counter at PPS */)
   1397 {
   1398 	long u_usec, v_usec, bigtick;
   1399 	long cal_sec, cal_usec;
   1400 
   1401 	/*
   1402 	 * An occasional glitch can be produced when the PPS interrupt
   1403 	 * occurs in the hardclock() routine before the time variable is
   1404 	 * updated. Here the offset is discarded when the difference
   1405 	 * between it and the last one is greater than tick/2, but not
   1406 	 * if the interval since the first discard exceeds 30 s.
   1407 	 */
   1408 	time_status |= STA_PPSSIGNAL;
   1409 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1410 	pps_valid = 0;
   1411 	u_usec = -tvp->tv_usec;
   1412 	if (u_usec < -500000)
   1413 		u_usec += 1000000;
   1414 	v_usec = pps_offset - u_usec;
   1415 	if (v_usec < 0)
   1416 		v_usec = -v_usec;
   1417 	if (v_usec > (tick >> 1)) {
   1418 		if (pps_glitch > MAXGLITCH) {
   1419 			pps_glitch = 0;
   1420 			pps_tf[2] = u_usec;
   1421 			pps_tf[1] = u_usec;
   1422 		} else {
   1423 			pps_glitch++;
   1424 			u_usec = pps_offset;
   1425 		}
   1426 	} else
   1427 		pps_glitch = 0;
   1428 
   1429 	/*
   1430 	 * A three-stage median filter is used to help deglitch the pps
   1431 	 * time. The median sample becomes the time offset estimate; the
   1432 	 * difference between the other two samples becomes the time
   1433 	 * dispersion (jitter) estimate.
   1434 	 */
   1435 	pps_tf[2] = pps_tf[1];
   1436 	pps_tf[1] = pps_tf[0];
   1437 	pps_tf[0] = u_usec;
   1438 	if (pps_tf[0] > pps_tf[1]) {
   1439 		if (pps_tf[1] > pps_tf[2]) {
   1440 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1441 			v_usec = pps_tf[0] - pps_tf[2];
   1442 		} else if (pps_tf[2] > pps_tf[0]) {
   1443 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1444 			v_usec = pps_tf[2] - pps_tf[1];
   1445 		} else {
   1446 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1447 			v_usec = pps_tf[0] - pps_tf[1];
   1448 		}
   1449 	} else {
   1450 		if (pps_tf[1] < pps_tf[2]) {
   1451 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1452 			v_usec = pps_tf[2] - pps_tf[0];
   1453 		} else  if (pps_tf[2] < pps_tf[0]) {
   1454 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1455 			v_usec = pps_tf[1] - pps_tf[2];
   1456 		} else {
   1457 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1458 			v_usec = pps_tf[1] - pps_tf[0];
   1459 		}
   1460 	}
   1461 	if (v_usec > MAXTIME)
   1462 		pps_jitcnt++;
   1463 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1464 	if (v_usec < 0)
   1465 		pps_jitter -= -v_usec >> PPS_AVG;
   1466 	else
   1467 		pps_jitter += v_usec >> PPS_AVG;
   1468 	if (pps_jitter > (MAXTIME >> 1))
   1469 		time_status |= STA_PPSJITTER;
   1470 
   1471 	/*
   1472 	 * During the calibration interval adjust the starting time when
   1473 	 * the tick overflows. At the end of the interval compute the
   1474 	 * duration of the interval and the difference of the hardware
   1475 	 * counters at the beginning and end of the interval. This code
   1476 	 * is deliciously complicated by the fact valid differences may
   1477 	 * exceed the value of tick when using long calibration
   1478 	 * intervals and small ticks. Note that the counter can be
   1479 	 * greater than tick if caught at just the wrong instant, but
   1480 	 * the values returned and used here are correct.
   1481 	 */
   1482 	bigtick = (long)tick << SHIFT_USEC;
   1483 	pps_usec -= pps_freq;
   1484 	if (pps_usec >= bigtick)
   1485 		pps_usec -= bigtick;
   1486 	if (pps_usec < 0)
   1487 		pps_usec += bigtick;
   1488 	pps_time.tv_sec++;
   1489 	pps_count++;
   1490 	if (pps_count < (1 << pps_shift))
   1491 		return;
   1492 	pps_count = 0;
   1493 	pps_calcnt++;
   1494 	u_usec = usec << SHIFT_USEC;
   1495 	v_usec = pps_usec - u_usec;
   1496 	if (v_usec >= bigtick >> 1)
   1497 		v_usec -= bigtick;
   1498 	if (v_usec < -(bigtick >> 1))
   1499 		v_usec += bigtick;
   1500 	if (v_usec < 0)
   1501 		v_usec = -(-v_usec >> pps_shift);
   1502 	else
   1503 		v_usec = v_usec >> pps_shift;
   1504 	pps_usec = u_usec;
   1505 	cal_sec = tvp->tv_sec;
   1506 	cal_usec = tvp->tv_usec;
   1507 	cal_sec -= pps_time.tv_sec;
   1508 	cal_usec -= pps_time.tv_usec;
   1509 	if (cal_usec < 0) {
   1510 		cal_usec += 1000000;
   1511 		cal_sec--;
   1512 	}
   1513 	pps_time = *tvp;
   1514 
   1515 	/*
   1516 	 * Check for lost interrupts, noise, excessive jitter and
   1517 	 * excessive frequency error. The number of timer ticks during
   1518 	 * the interval may vary +-1 tick. Add to this a margin of one
   1519 	 * tick for the PPS signal jitter and maximum frequency
   1520 	 * deviation. If the limits are exceeded, the calibration
   1521 	 * interval is reset to the minimum and we start over.
   1522 	 */
   1523 	u_usec = (long)tick << 1;
   1524 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1525 	    || (cal_sec == 0 && cal_usec < u_usec))
   1526 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1527 		pps_errcnt++;
   1528 		pps_shift = PPS_SHIFT;
   1529 		pps_intcnt = 0;
   1530 		time_status |= STA_PPSERROR;
   1531 		return;
   1532 	}
   1533 
   1534 	/*
   1535 	 * A three-stage median filter is used to help deglitch the pps
   1536 	 * frequency. The median sample becomes the frequency offset
   1537 	 * estimate; the difference between the other two samples
   1538 	 * becomes the frequency dispersion (stability) estimate.
   1539 	 */
   1540 	pps_ff[2] = pps_ff[1];
   1541 	pps_ff[1] = pps_ff[0];
   1542 	pps_ff[0] = v_usec;
   1543 	if (pps_ff[0] > pps_ff[1]) {
   1544 		if (pps_ff[1] > pps_ff[2]) {
   1545 			u_usec = pps_ff[1];		/* 0 1 2 */
   1546 			v_usec = pps_ff[0] - pps_ff[2];
   1547 		} else if (pps_ff[2] > pps_ff[0]) {
   1548 			u_usec = pps_ff[0];		/* 2 0 1 */
   1549 			v_usec = pps_ff[2] - pps_ff[1];
   1550 		} else {
   1551 			u_usec = pps_ff[2];		/* 0 2 1 */
   1552 			v_usec = pps_ff[0] - pps_ff[1];
   1553 		}
   1554 	} else {
   1555 		if (pps_ff[1] < pps_ff[2]) {
   1556 			u_usec = pps_ff[1];		/* 2 1 0 */
   1557 			v_usec = pps_ff[2] - pps_ff[0];
   1558 		} else  if (pps_ff[2] < pps_ff[0]) {
   1559 			u_usec = pps_ff[0];		/* 1 0 2 */
   1560 			v_usec = pps_ff[1] - pps_ff[2];
   1561 		} else {
   1562 			u_usec = pps_ff[2];		/* 1 2 0 */
   1563 			v_usec = pps_ff[1] - pps_ff[0];
   1564 		}
   1565 	}
   1566 
   1567 	/*
   1568 	 * Here the frequency dispersion (stability) is updated. If it
   1569 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1570 	 * offset is updated as well, but clamped to the tolerance. It
   1571 	 * will be processed later by the hardclock() routine.
   1572 	 */
   1573 	v_usec = (v_usec >> 1) - pps_stabil;
   1574 	if (v_usec < 0)
   1575 		pps_stabil -= -v_usec >> PPS_AVG;
   1576 	else
   1577 		pps_stabil += v_usec >> PPS_AVG;
   1578 	if (pps_stabil > MAXFREQ >> 2) {
   1579 		pps_stbcnt++;
   1580 		time_status |= STA_PPSWANDER;
   1581 		return;
   1582 	}
   1583 	if (time_status & STA_PPSFREQ) {
   1584 		if (u_usec < 0) {
   1585 			pps_freq -= -u_usec >> PPS_AVG;
   1586 			if (pps_freq < -time_tolerance)
   1587 				pps_freq = -time_tolerance;
   1588 			u_usec = -u_usec;
   1589 		} else {
   1590 			pps_freq += u_usec >> PPS_AVG;
   1591 			if (pps_freq > time_tolerance)
   1592 				pps_freq = time_tolerance;
   1593 		}
   1594 	}
   1595 
   1596 	/*
   1597 	 * Here the calibration interval is adjusted. If the maximum
   1598 	 * time difference is greater than tick / 4, reduce the interval
   1599 	 * by half. If this is not the case for four consecutive
   1600 	 * intervals, double the interval.
   1601 	 */
   1602 	if (u_usec << pps_shift > bigtick >> 2) {
   1603 		pps_intcnt = 0;
   1604 		if (pps_shift > PPS_SHIFT)
   1605 			pps_shift--;
   1606 	} else if (pps_intcnt >= 4) {
   1607 		pps_intcnt = 0;
   1608 		if (pps_shift < PPS_SHIFTMAX)
   1609 			pps_shift++;
   1610 	} else
   1611 		pps_intcnt++;
   1612 }
   1613 #endif /* PPS_SYNC */
   1614 #endif /* NTP  */
   1615 
   1616 /*
   1617  * Return information about system clocks.
   1618  */
   1619 int
   1620 sysctl_clockrate(void *where, size_t *sizep)
   1621 {
   1622 	struct clockinfo clkinfo;
   1623 
   1624 	/*
   1625 	 * Construct clockinfo structure.
   1626 	 */
   1627 	clkinfo.tick = tick;
   1628 	clkinfo.tickadj = tickadj;
   1629 	clkinfo.hz = hz;
   1630 	clkinfo.profhz = profhz;
   1631 	clkinfo.stathz = stathz ? stathz : hz;
   1632 	return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
   1633 }
   1634