Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.58
      1 /*	$NetBSD: kern_clock.c,v 1.58 2000/05/29 23:48:33 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 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(frame)
    475 	struct clockframe *frame;
    476 {
    477 	struct proc *p;
    478 	int delta;
    479 	extern int tickdelta;
    480 	extern long timedelta;
    481 #ifdef NTP
    482 	int time_update;
    483 	int ltemp;
    484 #endif
    485 
    486 	p = curproc;
    487 	if (p) {
    488 		struct pstats *pstats;
    489 
    490 		/*
    491 		 * Run current process's virtual and profile time, as needed.
    492 		 */
    493 		pstats = p->p_stats;
    494 		if (CLKF_USERMODE(frame) &&
    495 		    timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
    496 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
    497 			psignal(p, SIGVTALRM);
    498 		if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
    499 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
    500 			psignal(p, SIGPROF);
    501 	}
    502 
    503 	/*
    504 	 * If no separate statistics clock is available, run it from here.
    505 	 */
    506 	if (stathz == 0)
    507 		statclock(frame);
    508 
    509 	/*
    510 	 * Increment the time-of-day.  The increment is normally just
    511 	 * ``tick''.  If the machine is one which has a clock frequency
    512 	 * such that ``hz'' would not divide the second evenly into
    513 	 * milliseconds, a periodic adjustment must be applied.  Finally,
    514 	 * if we are still adjusting the time (see adjtime()),
    515 	 * ``tickdelta'' may also be added in.
    516 	 */
    517 	hardclock_ticks++;
    518 	delta = tick;
    519 
    520 #ifndef NTP
    521 	if (tickfix) {
    522 		tickfixcnt += tickfix;
    523 		if (tickfixcnt >= tickfixinterval) {
    524 			delta++;
    525 			tickfixcnt -= tickfixinterval;
    526 		}
    527 	}
    528 #endif /* !NTP */
    529 	/* Imprecise 4bsd adjtime() handling */
    530 	if (timedelta != 0) {
    531 		delta += tickdelta;
    532 		timedelta -= tickdelta;
    533 	}
    534 
    535 #ifdef notyet
    536 	microset();
    537 #endif
    538 
    539 #ifndef NTP
    540 	BUMPTIME(&time, delta);		/* XXX Now done using NTP code below */
    541 #endif
    542 	BUMPTIME(&mono_time, delta);
    543 
    544 #ifdef NTP
    545 	time_update = delta;
    546 
    547 	/*
    548 	 * Compute the phase adjustment. If the low-order bits
    549 	 * (time_phase) of the update overflow, bump the high-order bits
    550 	 * (time_update).
    551 	 */
    552 	time_phase += time_adj;
    553 	if (time_phase <= -FINEUSEC) {
    554 		ltemp = -time_phase >> SHIFT_SCALE;
    555 		time_phase += ltemp << SHIFT_SCALE;
    556 		time_update -= ltemp;
    557 	} else if (time_phase >= FINEUSEC) {
    558 		ltemp = time_phase >> SHIFT_SCALE;
    559 		time_phase -= ltemp << SHIFT_SCALE;
    560 		time_update += ltemp;
    561 	}
    562 
    563 #ifdef HIGHBALL
    564 	/*
    565 	 * If the HIGHBALL board is installed, we need to adjust the
    566 	 * external clock offset in order to close the hardware feedback
    567 	 * loop. This will adjust the external clock phase and frequency
    568 	 * in small amounts. The additional phase noise and frequency
    569 	 * wander this causes should be minimal. We also need to
    570 	 * discipline the kernel time variable, since the PLL is used to
    571 	 * discipline the external clock. If the Highball board is not
    572 	 * present, we discipline kernel time with the PLL as usual. We
    573 	 * assume that the external clock phase adjustment (time_update)
    574 	 * and kernel phase adjustment (clock_cpu) are less than the
    575 	 * value of tick.
    576 	 */
    577 	clock_offset.tv_usec += time_update;
    578 	if (clock_offset.tv_usec >= 1000000) {
    579 		clock_offset.tv_sec++;
    580 		clock_offset.tv_usec -= 1000000;
    581 	}
    582 	if (clock_offset.tv_usec < 0) {
    583 		clock_offset.tv_sec--;
    584 		clock_offset.tv_usec += 1000000;
    585 	}
    586 	time.tv_usec += clock_cpu;
    587 	clock_cpu = 0;
    588 #else
    589 	time.tv_usec += time_update;
    590 #endif /* HIGHBALL */
    591 
    592 	/*
    593 	 * On rollover of the second the phase adjustment to be used for
    594 	 * the next second is calculated. Also, the maximum error is
    595 	 * increased by the tolerance. If the PPS frequency discipline
    596 	 * code is present, the phase is increased to compensate for the
    597 	 * CPU clock oscillator frequency error.
    598 	 *
    599  	 * On a 32-bit machine and given parameters in the timex.h
    600 	 * header file, the maximum phase adjustment is +-512 ms and
    601 	 * maximum frequency offset is a tad less than) +-512 ppm. On a
    602 	 * 64-bit machine, you shouldn't need to ask.
    603 	 */
    604 	if (time.tv_usec >= 1000000) {
    605 		time.tv_usec -= 1000000;
    606 		time.tv_sec++;
    607 		time_maxerror += time_tolerance >> SHIFT_USEC;
    608 
    609 		/*
    610 		 * Leap second processing. If in leap-insert state at
    611 		 * the end of the day, the system clock is set back one
    612 		 * second; if in leap-delete state, the system clock is
    613 		 * set ahead one second. The microtime() routine or
    614 		 * external clock driver will insure that reported time
    615 		 * is always monotonic. The ugly divides should be
    616 		 * replaced.
    617 		 */
    618 		switch (time_state) {
    619 		case TIME_OK:
    620 			if (time_status & STA_INS)
    621 				time_state = TIME_INS;
    622 			else if (time_status & STA_DEL)
    623 				time_state = TIME_DEL;
    624 			break;
    625 
    626 		case TIME_INS:
    627 			if (time.tv_sec % 86400 == 0) {
    628 				time.tv_sec--;
    629 				time_state = TIME_OOP;
    630 			}
    631 			break;
    632 
    633 		case TIME_DEL:
    634 			if ((time.tv_sec + 1) % 86400 == 0) {
    635 				time.tv_sec++;
    636 				time_state = TIME_WAIT;
    637 			}
    638 			break;
    639 
    640 		case TIME_OOP:
    641 			time_state = TIME_WAIT;
    642 			break;
    643 
    644 		case TIME_WAIT:
    645 			if (!(time_status & (STA_INS | STA_DEL)))
    646 				time_state = TIME_OK;
    647 			break;
    648 		}
    649 
    650 		/*
    651 		 * Compute the phase adjustment for the next second. In
    652 		 * PLL mode, the offset is reduced by a fixed factor
    653 		 * times the time constant. In FLL mode the offset is
    654 		 * used directly. In either mode, the maximum phase
    655 		 * adjustment for each second is clamped so as to spread
    656 		 * the adjustment over not more than the number of
    657 		 * seconds between updates.
    658 		 */
    659 		if (time_offset < 0) {
    660 			ltemp = -time_offset;
    661 			if (!(time_status & STA_FLL))
    662 				ltemp >>= SHIFT_KG + time_constant;
    663 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    664 				ltemp = (MAXPHASE / MINSEC) <<
    665 				    SHIFT_UPDATE;
    666 			time_offset += ltemp;
    667 			time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
    668 		} else if (time_offset > 0) {
    669 			ltemp = time_offset;
    670 			if (!(time_status & STA_FLL))
    671 				ltemp >>= SHIFT_KG + time_constant;
    672 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    673 				ltemp = (MAXPHASE / MINSEC) <<
    674 				    SHIFT_UPDATE;
    675 			time_offset -= ltemp;
    676 			time_adj = ltemp << (shifthz - SHIFT_UPDATE);
    677 		} else
    678 			time_adj = 0;
    679 
    680 		/*
    681 		 * Compute the frequency estimate and additional phase
    682 		 * adjustment due to frequency error for the next
    683 		 * second. When the PPS signal is engaged, gnaw on the
    684 		 * watchdog counter and update the frequency computed by
    685 		 * the pll and the PPS signal.
    686 		 */
    687 #ifdef PPS_SYNC
    688 		pps_valid++;
    689 		if (pps_valid == PPS_VALID) {
    690 			pps_jitter = MAXTIME;
    691 			pps_stabil = MAXFREQ;
    692 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
    693 			    STA_PPSWANDER | STA_PPSERROR);
    694 		}
    695 		ltemp = time_freq + pps_freq;
    696 #else
    697 		ltemp = time_freq;
    698 #endif /* PPS_SYNC */
    699 
    700 		if (ltemp < 0)
    701 			time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
    702 		else
    703 			time_adj += ltemp >> (SHIFT_USEC - shifthz);
    704 		time_adj += (long)fixtick << shifthz;
    705 
    706 		/*
    707 		 * When the CPU clock oscillator frequency is not a
    708 		 * power of 2 in Hz, shifthz is only an approximate
    709 		 * scale factor.
    710 		 *
    711 		 * To determine the adjustment, you can do the following:
    712 		 *   bc -q
    713 		 *   scale=24
    714 		 *   obase=2
    715 		 *   idealhz/realhz
    716 		 * where `idealhz' is the next higher power of 2, and `realhz'
    717 		 * is the actual value.  You may need to factor this result
    718 		 * into a sequence of 2 multipliers to get better precision.
    719 		 *
    720 		 * Likewise, the error can be calculated with (e.g. for 100Hz):
    721 		 *   bc -q
    722 		 *   scale=24
    723 		 *   ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz
    724 		 * (and then multiply by 1000000 to get ppm).
    725 		 */
    726 		switch (hz) {
    727 		case 60:
    728 			/* A factor of 1.000100010001 gives about 15ppm
    729 			   error. */
    730 			if (time_adj < 0) {
    731 				time_adj -= (-time_adj >> 4);
    732 				time_adj -= (-time_adj >> 8);
    733 			} else {
    734 				time_adj += (time_adj >> 4);
    735 				time_adj += (time_adj >> 8);
    736 			}
    737 			break;
    738 
    739 		case 96:
    740 			/* A factor of 1.0101010101 gives about 244ppm error. */
    741 			if (time_adj < 0) {
    742 				time_adj -= (-time_adj >> 2);
    743 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    744 			} else {
    745 				time_adj += (time_adj >> 2);
    746 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    747 			}
    748 			break;
    749 
    750 		case 100:
    751 			/* A factor of 1.010001111010111 gives about 1ppm
    752 			   error. */
    753 			if (time_adj < 0) {
    754 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    755 				time_adj += (-time_adj >> 10);
    756 			} else {
    757 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    758 				time_adj -= (time_adj >> 10);
    759 			}
    760 			break;
    761 
    762 		case 1000:
    763 			/* A factor of 1.000001100010100001 gives about 50ppm
    764 			   error. */
    765 			if (time_adj < 0) {
    766 				time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
    767 				time_adj -= (-time_adj >> 7);
    768 			} else {
    769 				time_adj += (time_adj >> 6) + (time_adj >> 11);
    770 				time_adj += (time_adj >> 7);
    771 			}
    772 			break;
    773 
    774 		case 1200:
    775 			/* A factor of 1.1011010011100001 gives about 64ppm
    776 			   error. */
    777 			if (time_adj < 0) {
    778 				time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
    779 				time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
    780 			} else {
    781 				time_adj += (time_adj >> 1) + (time_adj >> 6);
    782 				time_adj += (time_adj >> 3) + (time_adj >> 10);
    783 			}
    784 			break;
    785 		}
    786 
    787 #ifdef EXT_CLOCK
    788 		/*
    789 		 * If an external clock is present, it is necessary to
    790 		 * discipline the kernel time variable anyway, since not
    791 		 * all system components use the microtime() interface.
    792 		 * Here, the time offset between the external clock and
    793 		 * kernel time variable is computed every so often.
    794 		 */
    795 		clock_count++;
    796 		if (clock_count > CLOCK_INTERVAL) {
    797 			clock_count = 0;
    798 			microtime(&clock_ext);
    799 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    800 			delta.tv_usec = clock_ext.tv_usec -
    801 			    time.tv_usec;
    802 			if (delta.tv_usec < 0)
    803 				delta.tv_sec--;
    804 			if (delta.tv_usec >= 500000) {
    805 				delta.tv_usec -= 1000000;
    806 				delta.tv_sec++;
    807 			}
    808 			if (delta.tv_usec < -500000) {
    809 				delta.tv_usec += 1000000;
    810 				delta.tv_sec--;
    811 			}
    812 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    813 			    delta.tv_usec > MAXPHASE) ||
    814 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    815 			    delta.tv_usec < -MAXPHASE)) {
    816 				time = clock_ext;
    817 				delta.tv_sec = 0;
    818 				delta.tv_usec = 0;
    819 			}
    820 #ifdef HIGHBALL
    821 			clock_cpu = delta.tv_usec;
    822 #else /* HIGHBALL */
    823 			hardupdate(delta.tv_usec);
    824 #endif /* HIGHBALL */
    825 		}
    826 #endif /* EXT_CLOCK */
    827 	}
    828 
    829 #endif /* NTP */
    830 
    831 	/*
    832 	 * Process callouts at a very low cpu priority, so we don't keep the
    833 	 * relatively high clock interrupt priority any longer than necessary.
    834 	 */
    835 	if (TAILQ_FIRST(&callwheel[hardclock_ticks & callwheelmask]) != NULL) {
    836 		if (CLKF_BASEPRI(frame)) {
    837 			/*
    838 			 * Save the overhead of a software interrupt;
    839 			 * it will happen as soon as we return, so do
    840 			 * it now.
    841 			 *
    842 			 * NOTE: If we're at ``base priority'', softclock()
    843 			 * was not already running.
    844 			 */
    845 			(void)spllowersoftclock();
    846 			softclock();
    847 		} else
    848 			setsoftclock();
    849 	} else if (softclock_running == 0 &&
    850 		   (softclock_ticks + 1) == hardclock_ticks)
    851 		softclock_ticks++;
    852 }
    853 
    854 /*
    855  * Software (low priority) clock interrupt.
    856  * Run periodic events from timeout queue.
    857  */
    858 /*ARGSUSED*/
    859 void
    860 softclock()
    861 {
    862 	struct callout_queue *bucket;
    863 	struct callout *c;
    864 	void (*func) __P((void *));
    865 	void *arg;
    866 	int s, idx;
    867 	int steps = 0;
    868 
    869 	s = splhigh();
    870 	softclock_running = 1;
    871 
    872 #ifdef CALLWHEEL_STATS
    873 	callwheel_softclocks++;
    874 #endif
    875 
    876 	while (softclock_ticks != hardclock_ticks) {
    877 		softclock_ticks++;
    878 		idx = (int)(softclock_ticks & callwheelmask);
    879 		bucket = &callwheel[idx];
    880 		c = TAILQ_FIRST(bucket);
    881 #ifdef CALLWHEEL_STATS
    882 		if (c == NULL)
    883 			callwheel_softempty++;
    884 #endif
    885 		while (c != NULL) {
    886 #ifdef CALLWHEEL_STATS
    887 			callwheel_softchecks++;
    888 #endif
    889 			if (c->c_time != softclock_ticks) {
    890 				c = TAILQ_NEXT(c, c_link);
    891 				if (++steps >= MAX_SOFTCLOCK_STEPS) {
    892 					nextsoftcheck = c;
    893 					/* Give interrupts a chance. */
    894 					splx(s);
    895 					(void) splhigh();
    896 					c = nextsoftcheck;
    897 					steps = 0;
    898 				}
    899 			} else {
    900 				nextsoftcheck = TAILQ_NEXT(c, c_link);
    901 				TAILQ_REMOVE(bucket, c, c_link);
    902 #ifdef CALLWHEEL_STATS
    903 				callwheel_sizes[idx]--;
    904 				callwheel_fired++;
    905 				callwheel_count--;
    906 #endif
    907 				func = c->c_func;
    908 				arg = c->c_arg;
    909 				c->c_func = NULL;
    910 				c->c_flags &= ~CALLOUT_PENDING;
    911 				splx(s);
    912 				(*func)(arg);
    913 				(void) splhigh();
    914 				steps = 0;
    915 				c = nextsoftcheck;
    916 			}
    917 		}
    918 	}
    919 	nextsoftcheck = NULL;
    920 	softclock_running = 0;
    921 	splx(s);
    922 }
    923 
    924 /*
    925  * callout_setsize:
    926  *
    927  *	Determine how many callwheels are necessary and
    928  *	set hash mask.  Called from allocsys().
    929  */
    930 void
    931 callout_setsize()
    932 {
    933 
    934 	for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1)
    935 		/* loop */ ;
    936 	callwheelmask = callwheelsize - 1;
    937 }
    938 
    939 /*
    940  * callout_startup:
    941  *
    942  *	Initialize the callwheel buckets.
    943  */
    944 void
    945 callout_startup()
    946 {
    947 	int i;
    948 
    949 	for (i = 0; i < callwheelsize; i++)
    950 		TAILQ_INIT(&callwheel[i]);
    951 }
    952 
    953 /*
    954  * callout_init:
    955  *
    956  *	Initialize a callout structure so that it can be used
    957  *	by callout_reset() and callout_stop().
    958  */
    959 void
    960 callout_init(c)
    961 	struct callout *c;
    962 {
    963 
    964 	memset(c, 0, sizeof(*c));
    965 }
    966 
    967 /*
    968  * callout_reset:
    969  *
    970  *	Establish or change a timeout.
    971  */
    972 void
    973 callout_reset(c, ticks, func, arg)
    974 	struct callout *c;
    975 	int ticks;
    976 	void (*func) __P((void *));
    977 	void *arg;
    978 {
    979 	struct callout_queue *bucket;
    980 	int s;
    981 
    982 	if (ticks <= 0)
    983 		ticks = 1;
    984 
    985 	/* Lock out the clock. */
    986 	s = splhigh();
    987 
    988 	/*
    989 	 * If this callout's timer is already running, cancel it
    990 	 * before we modify it.
    991 	 */
    992 	if (c->c_flags & CALLOUT_PENDING) {
    993 		callout_stop(c);
    994 #ifdef CALLWHEEL_STATS
    995 		callwheel_changed++;
    996 #endif
    997 	}
    998 
    999 	c->c_arg = arg;
   1000 	c->c_func = func;
   1001 	c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING;
   1002 	c->c_time = hardclock_ticks + ticks;
   1003 
   1004 	bucket = &callwheel[c->c_time & callwheelmask];
   1005 
   1006 #ifdef CALLWHEEL_STATS
   1007 	if (TAILQ_FIRST(bucket) != NULL)
   1008 		callwheel_collisions++;
   1009 #endif
   1010 
   1011 	TAILQ_INSERT_TAIL(bucket, c, c_link);
   1012 
   1013 #ifdef CALLWHEEL_STATS
   1014 	callwheel_count++;
   1015 	callwheel_established++;
   1016 	if (++callwheel_sizes[c->c_time & callwheelmask] > callwheel_maxlength)
   1017 		callwheel_maxlength =
   1018 		    callwheel_sizes[c->c_time & callwheelmask];
   1019 #endif
   1020 
   1021 	splx(s);
   1022 }
   1023 
   1024 /*
   1025  * callout_stop:
   1026  *
   1027  *	Disestablish a timeout.
   1028  */
   1029 void
   1030 callout_stop(c)
   1031 	struct callout *c;
   1032 {
   1033 	int s;
   1034 
   1035 	/* Lock out the clock. */
   1036 	s = splhigh();
   1037 
   1038 	/*
   1039 	 * Don't attempt to delete a callout that's not on the queue.
   1040 	 */
   1041 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
   1042 		c->c_flags &= ~CALLOUT_ACTIVE;
   1043 		splx(s);
   1044 		return;
   1045 	}
   1046 
   1047 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
   1048 
   1049 	if (nextsoftcheck == c)
   1050 		nextsoftcheck = TAILQ_NEXT(c, c_link);
   1051 
   1052 	TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_link);
   1053 #ifdef CALLWHEEL_STATS
   1054 	callwheel_count--;
   1055 	callwheel_disestablished++;
   1056 	callwheel_sizes[c->c_time & callwheelmask]--;
   1057 #endif
   1058 
   1059 	c->c_func = NULL;
   1060 
   1061 	splx(s);
   1062 }
   1063 
   1064 #ifdef CALLWHEEL_STATS
   1065 /*
   1066  * callout_showstats:
   1067  *
   1068  *	Display callout statistics.  Call it from DDB.
   1069  */
   1070 void
   1071 callout_showstats()
   1072 {
   1073 	u_int64_t curticks;
   1074 	int s;
   1075 
   1076 	s = splclock();
   1077 	curticks = softclock_ticks;
   1078 	splx(s);
   1079 
   1080 	printf("Callwheel statistics:\n");
   1081 	printf("\tCallouts currently queued: %llu\n", callwheel_count);
   1082 	printf("\tCallouts established: %llu\n", callwheel_established);
   1083 	printf("\tCallouts disestablished: %llu\n", callwheel_disestablished);
   1084 	if (callwheel_changed != 0)
   1085 		printf("\t\tOf those, %llu were changes\n", callwheel_changed);
   1086 	printf("\tCallouts that fired: %llu\n", callwheel_fired);
   1087 	printf("\tNumber of buckets: %d\n", callwheelsize);
   1088 	printf("\tNumber of hash collisions: %d\n", callwheel_collisions);
   1089 	printf("\tMaximum hash chain length: %d\n", callwheel_maxlength);
   1090 	printf("\tSoftclocks: %llu, Softchecks: %llu\n",
   1091 	    callwheel_softclocks, callwheel_softchecks);
   1092 	printf("\t\tEmpty buckets seen: %llu\n", callwheel_softempty);
   1093 }
   1094 #endif
   1095 
   1096 /*
   1097  * Compute number of hz until specified time.  Used to compute second
   1098  * argument to callout_reset() from an absolute time.
   1099  */
   1100 int
   1101 hzto(tv)
   1102 	struct timeval *tv;
   1103 {
   1104 	long ticks, sec;
   1105 	int s;
   1106 
   1107 	/*
   1108 	 * If number of microseconds will fit in 32 bit arithmetic,
   1109 	 * then compute number of microseconds to time and scale to
   1110 	 * ticks.  Otherwise just compute number of hz in time, rounding
   1111 	 * times greater than representible to maximum value.  (We must
   1112 	 * compute in microseconds, because hz can be greater than 1000,
   1113 	 * and thus tick can be less than one millisecond).
   1114 	 *
   1115 	 * Delta times less than 14 hours can be computed ``exactly''.
   1116 	 * (Note that if hz would yeild a non-integral number of us per
   1117 	 * tick, i.e. tickfix is nonzero, timouts can be a tick longer
   1118 	 * than they should be.)  Maximum value for any timeout in 10ms
   1119 	 * ticks is 250 days.
   1120 	 */
   1121 	s = splclock();
   1122 	sec = tv->tv_sec - time.tv_sec;
   1123 	if (sec <= 0x7fffffff / 1000000 - 1)
   1124 		ticks = ((tv->tv_sec - time.tv_sec) * 1000000 +
   1125 			(tv->tv_usec - time.tv_usec)) / tick;
   1126 	else if (sec <= 0x7fffffff / hz)
   1127 		ticks = sec * hz;
   1128 	else
   1129 		ticks = 0x7fffffff;
   1130 	splx(s);
   1131 	return (ticks);
   1132 }
   1133 
   1134 /*
   1135  * Start profiling on a process.
   1136  *
   1137  * Kernel profiling passes proc0 which never exits and hence
   1138  * keeps the profile clock running constantly.
   1139  */
   1140 void
   1141 startprofclock(p)
   1142 	struct proc *p;
   1143 {
   1144 	int s;
   1145 
   1146 	if ((p->p_flag & P_PROFIL) == 0) {
   1147 		p->p_flag |= P_PROFIL;
   1148 		if (++profprocs == 1 && stathz != 0) {
   1149 			s = splstatclock();
   1150 			psdiv = pscnt = psratio;
   1151 			setstatclockrate(profhz);
   1152 			splx(s);
   1153 		}
   1154 	}
   1155 }
   1156 
   1157 /*
   1158  * Stop profiling on a process.
   1159  */
   1160 void
   1161 stopprofclock(p)
   1162 	struct proc *p;
   1163 {
   1164 	int s;
   1165 
   1166 	if (p->p_flag & P_PROFIL) {
   1167 		p->p_flag &= ~P_PROFIL;
   1168 		if (--profprocs == 0 && stathz != 0) {
   1169 			s = splstatclock();
   1170 			psdiv = pscnt = 1;
   1171 			setstatclockrate(stathz);
   1172 			splx(s);
   1173 		}
   1174 	}
   1175 }
   1176 
   1177 /*
   1178  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1179  * do process and kernel statistics.
   1180  */
   1181 void
   1182 statclock(frame)
   1183 	struct clockframe *frame;
   1184 {
   1185 #ifdef GPROF
   1186 	struct gmonparam *g;
   1187 	int i;
   1188 #endif
   1189 	static int schedclk;
   1190 	struct proc *p;
   1191 
   1192 	if (CLKF_USERMODE(frame)) {
   1193 		p = curproc;
   1194 		if (p->p_flag & P_PROFIL)
   1195 			addupc_intr(p, CLKF_PC(frame), 1);
   1196 		if (--pscnt > 0)
   1197 			return;
   1198 		/*
   1199 		 * Came from user mode; CPU was in user state.
   1200 		 * If this process is being profiled record the tick.
   1201 		 */
   1202 		p->p_uticks++;
   1203 		if (p->p_nice > NZERO)
   1204 			cp_time[CP_NICE]++;
   1205 		else
   1206 			cp_time[CP_USER]++;
   1207 	} else {
   1208 #ifdef GPROF
   1209 		/*
   1210 		 * Kernel statistics are just like addupc_intr, only easier.
   1211 		 */
   1212 		g = &_gmonparam;
   1213 		if (g->state == GMON_PROF_ON) {
   1214 			i = CLKF_PC(frame) - g->lowpc;
   1215 			if (i < g->textsize) {
   1216 				i /= HISTFRACTION * sizeof(*g->kcount);
   1217 				g->kcount[i]++;
   1218 			}
   1219 		}
   1220 #endif
   1221 		if (--pscnt > 0)
   1222 			return;
   1223 		/*
   1224 		 * Came from kernel mode, so we were:
   1225 		 * - handling an interrupt,
   1226 		 * - doing syscall or trap work on behalf of the current
   1227 		 *   user process, or
   1228 		 * - spinning in the idle loop.
   1229 		 * Whichever it is, charge the time as appropriate.
   1230 		 * Note that we charge interrupts to the current process,
   1231 		 * regardless of whether they are ``for'' that process,
   1232 		 * so that we know how much of its real time was spent
   1233 		 * in ``non-process'' (i.e., interrupt) work.
   1234 		 */
   1235 		p = curproc;
   1236 		if (CLKF_INTR(frame)) {
   1237 			if (p != NULL)
   1238 				p->p_iticks++;
   1239 			cp_time[CP_INTR]++;
   1240 		} else if (p != NULL) {
   1241 			p->p_sticks++;
   1242 			cp_time[CP_SYS]++;
   1243 		} else
   1244 			cp_time[CP_IDLE]++;
   1245 	}
   1246 	pscnt = psdiv;
   1247 
   1248 	if (p != NULL) {
   1249 		++p->p_cpticks;
   1250 		/*
   1251 		 * If no schedclock is provided, call it here at ~~12-25 Hz,
   1252 		 * ~~16 Hz is best
   1253 		 */
   1254 		if(schedhz == 0)
   1255 			if ((++schedclk & 3) == 0)
   1256 				schedclock(p);
   1257 	}
   1258 }
   1259 
   1260 
   1261 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1262 
   1263 /*
   1264  * hardupdate() - local clock update
   1265  *
   1266  * This routine is called by ntp_adjtime() to update the local clock
   1267  * phase and frequency. The implementation is of an adaptive-parameter,
   1268  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1269  * time and frequency offset estimates for each call. If the kernel PPS
   1270  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1271  * determines the new time offset, instead of the calling argument.
   1272  * Presumably, calls to ntp_adjtime() occur only when the caller
   1273  * believes the local clock is valid within some bound (+-128 ms with
   1274  * NTP). If the caller's time is far different than the PPS time, an
   1275  * argument will ensue, and it's not clear who will lose.
   1276  *
   1277  * For uncompensated quartz crystal oscillatores and nominal update
   1278  * intervals less than 1024 s, operation should be in phase-lock mode
   1279  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1280  * intervals greater than thiss, operation should be in frequency-lock
   1281  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1282  *
   1283  * Note: splclock() is in effect.
   1284  */
   1285 void
   1286 hardupdate(offset)
   1287 	long offset;
   1288 {
   1289 	long ltemp, mtemp;
   1290 
   1291 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1292 		return;
   1293 	ltemp = offset;
   1294 #ifdef PPS_SYNC
   1295 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1296 		ltemp = pps_offset;
   1297 #endif /* PPS_SYNC */
   1298 
   1299 	/*
   1300 	 * Scale the phase adjustment and clamp to the operating range.
   1301 	 */
   1302 	if (ltemp > MAXPHASE)
   1303 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1304 	else if (ltemp < -MAXPHASE)
   1305 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1306 	else
   1307 		time_offset = ltemp << SHIFT_UPDATE;
   1308 
   1309 	/*
   1310 	 * Select whether the frequency is to be controlled and in which
   1311 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1312 	 * multiply/divide should be replaced someday.
   1313 	 */
   1314 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1315 		time_reftime = time.tv_sec;
   1316 	mtemp = time.tv_sec - time_reftime;
   1317 	time_reftime = time.tv_sec;
   1318 	if (time_status & STA_FLL) {
   1319 		if (mtemp >= MINSEC) {
   1320 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1321 			    SHIFT_UPDATE));
   1322 			if (ltemp < 0)
   1323 				time_freq -= -ltemp >> SHIFT_KH;
   1324 			else
   1325 				time_freq += ltemp >> SHIFT_KH;
   1326 		}
   1327 	} else {
   1328 		if (mtemp < MAXSEC) {
   1329 			ltemp *= mtemp;
   1330 			if (ltemp < 0)
   1331 				time_freq -= -ltemp >> (time_constant +
   1332 				    time_constant + SHIFT_KF -
   1333 				    SHIFT_USEC);
   1334 			else
   1335 				time_freq += ltemp >> (time_constant +
   1336 				    time_constant + SHIFT_KF -
   1337 				    SHIFT_USEC);
   1338 		}
   1339 	}
   1340 	if (time_freq > time_tolerance)
   1341 		time_freq = time_tolerance;
   1342 	else if (time_freq < -time_tolerance)
   1343 		time_freq = -time_tolerance;
   1344 }
   1345 
   1346 #ifdef PPS_SYNC
   1347 /*
   1348  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1349  *
   1350  * This routine is called at each PPS interrupt in order to discipline
   1351  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1352  * and leaves it in a handy spot for the hardclock() routine. It
   1353  * integrates successive PPS phase differences and calculates the
   1354  * frequency offset. This is used in hardclock() to discipline the CPU
   1355  * clock oscillator so that intrinsic frequency error is cancelled out.
   1356  * The code requires the caller to capture the time and hardware counter
   1357  * value at the on-time PPS signal transition.
   1358  *
   1359  * Note that, on some Unix systems, this routine runs at an interrupt
   1360  * priority level higher than the timer interrupt routine hardclock().
   1361  * Therefore, the variables used are distinct from the hardclock()
   1362  * variables, except for certain exceptions: The PPS frequency pps_freq
   1363  * and phase pps_offset variables are determined by this routine and
   1364  * updated atomically. The time_tolerance variable can be considered a
   1365  * constant, since it is infrequently changed, and then only when the
   1366  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1367  * once per second by hardclock() and is atomically cleared in this
   1368  * routine.
   1369  */
   1370 void
   1371 hardpps(tvp, usec)
   1372 	struct timeval *tvp;		/* time at PPS */
   1373 	long usec;			/* hardware counter at PPS */
   1374 {
   1375 	long u_usec, v_usec, bigtick;
   1376 	long cal_sec, cal_usec;
   1377 
   1378 	/*
   1379 	 * An occasional glitch can be produced when the PPS interrupt
   1380 	 * occurs in the hardclock() routine before the time variable is
   1381 	 * updated. Here the offset is discarded when the difference
   1382 	 * between it and the last one is greater than tick/2, but not
   1383 	 * if the interval since the first discard exceeds 30 s.
   1384 	 */
   1385 	time_status |= STA_PPSSIGNAL;
   1386 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1387 	pps_valid = 0;
   1388 	u_usec = -tvp->tv_usec;
   1389 	if (u_usec < -500000)
   1390 		u_usec += 1000000;
   1391 	v_usec = pps_offset - u_usec;
   1392 	if (v_usec < 0)
   1393 		v_usec = -v_usec;
   1394 	if (v_usec > (tick >> 1)) {
   1395 		if (pps_glitch > MAXGLITCH) {
   1396 			pps_glitch = 0;
   1397 			pps_tf[2] = u_usec;
   1398 			pps_tf[1] = u_usec;
   1399 		} else {
   1400 			pps_glitch++;
   1401 			u_usec = pps_offset;
   1402 		}
   1403 	} else
   1404 		pps_glitch = 0;
   1405 
   1406 	/*
   1407 	 * A three-stage median filter is used to help deglitch the pps
   1408 	 * time. The median sample becomes the time offset estimate; the
   1409 	 * difference between the other two samples becomes the time
   1410 	 * dispersion (jitter) estimate.
   1411 	 */
   1412 	pps_tf[2] = pps_tf[1];
   1413 	pps_tf[1] = pps_tf[0];
   1414 	pps_tf[0] = u_usec;
   1415 	if (pps_tf[0] > pps_tf[1]) {
   1416 		if (pps_tf[1] > pps_tf[2]) {
   1417 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1418 			v_usec = pps_tf[0] - pps_tf[2];
   1419 		} else if (pps_tf[2] > pps_tf[0]) {
   1420 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1421 			v_usec = pps_tf[2] - pps_tf[1];
   1422 		} else {
   1423 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1424 			v_usec = pps_tf[0] - pps_tf[1];
   1425 		}
   1426 	} else {
   1427 		if (pps_tf[1] < pps_tf[2]) {
   1428 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1429 			v_usec = pps_tf[2] - pps_tf[0];
   1430 		} else  if (pps_tf[2] < pps_tf[0]) {
   1431 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1432 			v_usec = pps_tf[1] - pps_tf[2];
   1433 		} else {
   1434 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1435 			v_usec = pps_tf[1] - pps_tf[0];
   1436 		}
   1437 	}
   1438 	if (v_usec > MAXTIME)
   1439 		pps_jitcnt++;
   1440 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1441 	if (v_usec < 0)
   1442 		pps_jitter -= -v_usec >> PPS_AVG;
   1443 	else
   1444 		pps_jitter += v_usec >> PPS_AVG;
   1445 	if (pps_jitter > (MAXTIME >> 1))
   1446 		time_status |= STA_PPSJITTER;
   1447 
   1448 	/*
   1449 	 * During the calibration interval adjust the starting time when
   1450 	 * the tick overflows. At the end of the interval compute the
   1451 	 * duration of the interval and the difference of the hardware
   1452 	 * counters at the beginning and end of the interval. This code
   1453 	 * is deliciously complicated by the fact valid differences may
   1454 	 * exceed the value of tick when using long calibration
   1455 	 * intervals and small ticks. Note that the counter can be
   1456 	 * greater than tick if caught at just the wrong instant, but
   1457 	 * the values returned and used here are correct.
   1458 	 */
   1459 	bigtick = (long)tick << SHIFT_USEC;
   1460 	pps_usec -= pps_freq;
   1461 	if (pps_usec >= bigtick)
   1462 		pps_usec -= bigtick;
   1463 	if (pps_usec < 0)
   1464 		pps_usec += bigtick;
   1465 	pps_time.tv_sec++;
   1466 	pps_count++;
   1467 	if (pps_count < (1 << pps_shift))
   1468 		return;
   1469 	pps_count = 0;
   1470 	pps_calcnt++;
   1471 	u_usec = usec << SHIFT_USEC;
   1472 	v_usec = pps_usec - u_usec;
   1473 	if (v_usec >= bigtick >> 1)
   1474 		v_usec -= bigtick;
   1475 	if (v_usec < -(bigtick >> 1))
   1476 		v_usec += bigtick;
   1477 	if (v_usec < 0)
   1478 		v_usec = -(-v_usec >> pps_shift);
   1479 	else
   1480 		v_usec = v_usec >> pps_shift;
   1481 	pps_usec = u_usec;
   1482 	cal_sec = tvp->tv_sec;
   1483 	cal_usec = tvp->tv_usec;
   1484 	cal_sec -= pps_time.tv_sec;
   1485 	cal_usec -= pps_time.tv_usec;
   1486 	if (cal_usec < 0) {
   1487 		cal_usec += 1000000;
   1488 		cal_sec--;
   1489 	}
   1490 	pps_time = *tvp;
   1491 
   1492 	/*
   1493 	 * Check for lost interrupts, noise, excessive jitter and
   1494 	 * excessive frequency error. The number of timer ticks during
   1495 	 * the interval may vary +-1 tick. Add to this a margin of one
   1496 	 * tick for the PPS signal jitter and maximum frequency
   1497 	 * deviation. If the limits are exceeded, the calibration
   1498 	 * interval is reset to the minimum and we start over.
   1499 	 */
   1500 	u_usec = (long)tick << 1;
   1501 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1502 	    || (cal_sec == 0 && cal_usec < u_usec))
   1503 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1504 		pps_errcnt++;
   1505 		pps_shift = PPS_SHIFT;
   1506 		pps_intcnt = 0;
   1507 		time_status |= STA_PPSERROR;
   1508 		return;
   1509 	}
   1510 
   1511 	/*
   1512 	 * A three-stage median filter is used to help deglitch the pps
   1513 	 * frequency. The median sample becomes the frequency offset
   1514 	 * estimate; the difference between the other two samples
   1515 	 * becomes the frequency dispersion (stability) estimate.
   1516 	 */
   1517 	pps_ff[2] = pps_ff[1];
   1518 	pps_ff[1] = pps_ff[0];
   1519 	pps_ff[0] = v_usec;
   1520 	if (pps_ff[0] > pps_ff[1]) {
   1521 		if (pps_ff[1] > pps_ff[2]) {
   1522 			u_usec = pps_ff[1];		/* 0 1 2 */
   1523 			v_usec = pps_ff[0] - pps_ff[2];
   1524 		} else if (pps_ff[2] > pps_ff[0]) {
   1525 			u_usec = pps_ff[0];		/* 2 0 1 */
   1526 			v_usec = pps_ff[2] - pps_ff[1];
   1527 		} else {
   1528 			u_usec = pps_ff[2];		/* 0 2 1 */
   1529 			v_usec = pps_ff[0] - pps_ff[1];
   1530 		}
   1531 	} else {
   1532 		if (pps_ff[1] < pps_ff[2]) {
   1533 			u_usec = pps_ff[1];		/* 2 1 0 */
   1534 			v_usec = pps_ff[2] - pps_ff[0];
   1535 		} else  if (pps_ff[2] < pps_ff[0]) {
   1536 			u_usec = pps_ff[0];		/* 1 0 2 */
   1537 			v_usec = pps_ff[1] - pps_ff[2];
   1538 		} else {
   1539 			u_usec = pps_ff[2];		/* 1 2 0 */
   1540 			v_usec = pps_ff[1] - pps_ff[0];
   1541 		}
   1542 	}
   1543 
   1544 	/*
   1545 	 * Here the frequency dispersion (stability) is updated. If it
   1546 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1547 	 * offset is updated as well, but clamped to the tolerance. It
   1548 	 * will be processed later by the hardclock() routine.
   1549 	 */
   1550 	v_usec = (v_usec >> 1) - pps_stabil;
   1551 	if (v_usec < 0)
   1552 		pps_stabil -= -v_usec >> PPS_AVG;
   1553 	else
   1554 		pps_stabil += v_usec >> PPS_AVG;
   1555 	if (pps_stabil > MAXFREQ >> 2) {
   1556 		pps_stbcnt++;
   1557 		time_status |= STA_PPSWANDER;
   1558 		return;
   1559 	}
   1560 	if (time_status & STA_PPSFREQ) {
   1561 		if (u_usec < 0) {
   1562 			pps_freq -= -u_usec >> PPS_AVG;
   1563 			if (pps_freq < -time_tolerance)
   1564 				pps_freq = -time_tolerance;
   1565 			u_usec = -u_usec;
   1566 		} else {
   1567 			pps_freq += u_usec >> PPS_AVG;
   1568 			if (pps_freq > time_tolerance)
   1569 				pps_freq = time_tolerance;
   1570 		}
   1571 	}
   1572 
   1573 	/*
   1574 	 * Here the calibration interval is adjusted. If the maximum
   1575 	 * time difference is greater than tick / 4, reduce the interval
   1576 	 * by half. If this is not the case for four consecutive
   1577 	 * intervals, double the interval.
   1578 	 */
   1579 	if (u_usec << pps_shift > bigtick >> 2) {
   1580 		pps_intcnt = 0;
   1581 		if (pps_shift > PPS_SHIFT)
   1582 			pps_shift--;
   1583 	} else if (pps_intcnt >= 4) {
   1584 		pps_intcnt = 0;
   1585 		if (pps_shift < PPS_SHIFTMAX)
   1586 			pps_shift++;
   1587 	} else
   1588 		pps_intcnt++;
   1589 }
   1590 #endif /* PPS_SYNC */
   1591 #endif /* NTP  */
   1592 
   1593 
   1594 /*
   1595  * Return information about system clocks.
   1596  */
   1597 int
   1598 sysctl_clockrate(where, sizep)
   1599 	char *where;
   1600 	size_t *sizep;
   1601 {
   1602 	struct clockinfo clkinfo;
   1603 
   1604 	/*
   1605 	 * Construct clockinfo structure.
   1606 	 */
   1607 	clkinfo.tick = tick;
   1608 	clkinfo.tickadj = tickadj;
   1609 	clkinfo.hz = hz;
   1610 	clkinfo.profhz = profhz;
   1611 	clkinfo.stathz = stathz ? stathz : hz;
   1612 	return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
   1613 }
   1614