Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.93
      1 /*	$NetBSD: kern_clock.c,v 1.93 2005/02/26 21:34:55 perry 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. Neither the name of the University nor the names of its contributors
     58  *    may be used to endorse or promote products derived from this software
     59  *    without specific prior written permission.
     60  *
     61  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     62  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     63  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     64  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     65  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     66  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     67  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     68  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     69  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     70  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     71  * SUCH DAMAGE.
     72  *
     73  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
     74  */
     75 
     76 #include <sys/cdefs.h>
     77 __KERNEL_RCSID(0, "$NetBSD: kern_clock.c,v 1.93 2005/02/26 21:34:55 perry Exp $");
     78 
     79 #include "opt_ntp.h"
     80 #include "opt_multiprocessor.h"
     81 #include "opt_perfctrs.h"
     82 
     83 #include <sys/param.h>
     84 #include <sys/systm.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 <sys/sysctl.h>
     91 #include <sys/timex.h>
     92 #include <sys/sched.h>
     93 #include <sys/time.h>
     94 
     95 #include <machine/cpu.h>
     96 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
     97 #include <machine/intr.h>
     98 #endif
     99 
    100 #ifdef GPROF
    101 #include <sys/gmon.h>
    102 #endif
    103 
    104 /*
    105  * Clock handling routines.
    106  *
    107  * This code is written to operate with two timers that run independently of
    108  * each other.  The main clock, running hz times per second, is used to keep
    109  * track of real time.  The second timer handles kernel and user profiling,
    110  * and does resource use estimation.  If the second timer is programmable,
    111  * it is randomized to avoid aliasing between the two clocks.  For example,
    112  * the randomization prevents an adversary from always giving up the CPU
    113  * just before its quantum expires.  Otherwise, it would never accumulate
    114  * CPU ticks.  The mean frequency of the second timer is stathz.
    115  *
    116  * If no second timer exists, stathz will be zero; in this case we drive
    117  * profiling and statistics off the main clock.  This WILL NOT be accurate;
    118  * do not do it unless absolutely necessary.
    119  *
    120  * The statistics clock may (or may not) be run at a higher rate while
    121  * profiling.  This profile clock runs at profhz.  We require that profhz
    122  * be an integral multiple of stathz.
    123  *
    124  * If the statistics clock is running fast, it must be divided by the ratio
    125  * profhz/stathz for statistics.  (For profiling, every tick counts.)
    126  */
    127 
    128 #ifdef NTP	/* NTP phase-locked loop in kernel */
    129 /*
    130  * Phase/frequency-lock loop (PLL/FLL) definitions
    131  *
    132  * The following variables are read and set by the ntp_adjtime() system
    133  * call.
    134  *
    135  * time_state shows the state of the system clock, with values defined
    136  * in the timex.h header file.
    137  *
    138  * time_status shows the status of the system clock, with bits defined
    139  * in the timex.h header file.
    140  *
    141  * time_offset is used by the PLL/FLL to adjust the system time in small
    142  * increments.
    143  *
    144  * time_constant determines the bandwidth or "stiffness" of the PLL.
    145  *
    146  * time_tolerance determines maximum frequency error or tolerance of the
    147  * CPU clock oscillator and is a property of the architecture; however,
    148  * in principle it could change as result of the presence of external
    149  * discipline signals, for instance.
    150  *
    151  * time_precision is usually equal to the kernel tick variable; however,
    152  * in cases where a precision clock counter or external clock is
    153  * available, the resolution can be much less than this and depend on
    154  * whether the external clock is working or not.
    155  *
    156  * time_maxerror is initialized by a ntp_adjtime() call and increased by
    157  * the kernel once each second to reflect the maximum error bound
    158  * growth.
    159  *
    160  * time_esterror is set and read by the ntp_adjtime() call, but
    161  * otherwise not used by the kernel.
    162  */
    163 int time_state = TIME_OK;	/* clock state */
    164 int time_status = STA_UNSYNC;	/* clock status bits */
    165 long time_offset = 0;		/* time offset (us) */
    166 long time_constant = 0;		/* pll time constant */
    167 long time_tolerance = MAXFREQ;	/* frequency tolerance (scaled ppm) */
    168 long time_precision = 1;	/* clock precision (us) */
    169 long time_maxerror = MAXPHASE;	/* maximum error (us) */
    170 long time_esterror = MAXPHASE;	/* estimated error (us) */
    171 
    172 /*
    173  * The following variables establish the state of the PLL/FLL and the
    174  * residual time and frequency offset of the local clock. The scale
    175  * factors are defined in the timex.h header file.
    176  *
    177  * time_phase and time_freq are the phase increment and the frequency
    178  * increment, respectively, of the kernel time variable.
    179  *
    180  * time_freq is set via ntp_adjtime() from a value stored in a file when
    181  * the synchronization daemon is first started. Its value is retrieved
    182  * via ntp_adjtime() and written to the file about once per hour by the
    183  * daemon.
    184  *
    185  * time_adj is the adjustment added to the value of tick at each timer
    186  * interrupt and is recomputed from time_phase and time_freq at each
    187  * seconds rollover.
    188  *
    189  * time_reftime is the second's portion of the system time at the last
    190  * call to ntp_adjtime(). It is used to adjust the time_freq variable
    191  * and to increase the time_maxerror as the time since last update
    192  * increases.
    193  */
    194 long time_phase = 0;		/* phase offset (scaled us) */
    195 long time_freq = 0;		/* frequency offset (scaled ppm) */
    196 long time_adj = 0;		/* tick adjust (scaled 1 / hz) */
    197 long time_reftime = 0;		/* time at last adjustment (s) */
    198 
    199 #ifdef PPS_SYNC
    200 /*
    201  * The following variables are used only if the kernel PPS discipline
    202  * code is configured (PPS_SYNC). The scale factors are defined in the
    203  * timex.h header file.
    204  *
    205  * pps_time contains the time at each calibration interval, as read by
    206  * microtime(). pps_count counts the seconds of the calibration
    207  * interval, the duration of which is nominally pps_shift in powers of
    208  * two.
    209  *
    210  * pps_offset is the time offset produced by the time median filter
    211  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
    212  * this filter.
    213  *
    214  * pps_freq is the frequency offset produced by the frequency median
    215  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
    216  * by this filter.
    217  *
    218  * pps_usec is latched from a high resolution counter or external clock
    219  * at pps_time. Here we want the hardware counter contents only, not the
    220  * contents plus the time_tv.usec as usual.
    221  *
    222  * pps_valid counts the number of seconds since the last PPS update. It
    223  * is used as a watchdog timer to disable the PPS discipline should the
    224  * PPS signal be lost.
    225  *
    226  * pps_glitch counts the number of seconds since the beginning of an
    227  * offset burst more than tick/2 from current nominal offset. It is used
    228  * mainly to suppress error bursts due to priority conflicts between the
    229  * PPS interrupt and timer interrupt.
    230  *
    231  * pps_intcnt counts the calibration intervals for use in the interval-
    232  * adaptation algorithm. It's just too complicated for words.
    233  *
    234  * pps_kc_hardpps_source contains an arbitrary value that uniquely
    235  * identifies the currently bound source of the PPS signal, or NULL
    236  * if no source is bound.
    237  *
    238  * pps_kc_hardpps_mode indicates which transitions, if any, of the PPS
    239  * signal should be reported.
    240  */
    241 struct timeval pps_time;	/* kernel time at last interval */
    242 long pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
    243 long pps_offset = 0;		/* pps time offset (us) */
    244 long pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
    245 long pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
    246 long pps_freq = 0;		/* frequency offset (scaled ppm) */
    247 long pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
    248 long pps_usec = 0;		/* microsec counter at last interval */
    249 long pps_valid = PPS_VALID;	/* pps signal watchdog counter */
    250 int pps_glitch = 0;		/* pps signal glitch counter */
    251 int pps_count = 0;		/* calibration interval counter (s) */
    252 int pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
    253 int pps_intcnt = 0;		/* intervals at current duration */
    254 void *pps_kc_hardpps_source = NULL; /* current PPS supplier's identifier */
    255 int pps_kc_hardpps_mode = 0;	/* interesting edges of PPS signal */
    256 
    257 /*
    258  * PPS signal quality monitors
    259  *
    260  * pps_jitcnt counts the seconds that have been discarded because the
    261  * jitter measured by the time median filter exceeds the limit MAXTIME
    262  * (100 us).
    263  *
    264  * pps_calcnt counts the frequency calibration intervals, which are
    265  * variable from 4 s to 256 s.
    266  *
    267  * pps_errcnt counts the calibration intervals which have been discarded
    268  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
    269  * calibration interval jitter exceeds two ticks.
    270  *
    271  * pps_stbcnt counts the calibration intervals that have been discarded
    272  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
    273  */
    274 long pps_jitcnt = 0;		/* jitter limit exceeded */
    275 long pps_calcnt = 0;		/* calibration intervals */
    276 long pps_errcnt = 0;		/* calibration errors */
    277 long pps_stbcnt = 0;		/* stability limit exceeded */
    278 #endif /* PPS_SYNC */
    279 
    280 #ifdef EXT_CLOCK
    281 /*
    282  * External clock definitions
    283  *
    284  * The following definitions and declarations are used only if an
    285  * external clock is configured on the system.
    286  */
    287 #define CLOCK_INTERVAL 30	/* CPU clock update interval (s) */
    288 
    289 /*
    290  * The clock_count variable is set to CLOCK_INTERVAL at each PPS
    291  * interrupt and decremented once each second.
    292  */
    293 int clock_count = 0;		/* CPU clock counter */
    294 
    295 #ifdef HIGHBALL
    296 /*
    297  * The clock_offset and clock_cpu variables are used by the HIGHBALL
    298  * interface. The clock_offset variable defines the offset between
    299  * system time and the HIGBALL counters. The clock_cpu variable contains
    300  * the offset between the system clock and the HIGHBALL clock for use in
    301  * disciplining the kernel time variable.
    302  */
    303 extern struct timeval clock_offset; /* Highball clock offset */
    304 long clock_cpu = 0;		/* CPU clock adjust */
    305 #endif /* HIGHBALL */
    306 #endif /* EXT_CLOCK */
    307 #endif /* NTP */
    308 
    309 
    310 /*
    311  * Bump a timeval by a small number of usec's.
    312  */
    313 #define BUMPTIME(t, usec) { \
    314 	volatile struct timeval *tp = (t); \
    315 	long us; \
    316  \
    317 	tp->tv_usec = us = tp->tv_usec + (usec); \
    318 	if (us >= 1000000) { \
    319 		tp->tv_usec = us - 1000000; \
    320 		tp->tv_sec++; \
    321 	} \
    322 }
    323 
    324 int	stathz;
    325 int	profhz;
    326 int	profsrc;
    327 int	schedhz;
    328 int	profprocs;
    329 int	hardclock_ticks;
    330 static int statscheddiv; /* stat => sched divider (used if schedhz == 0) */
    331 static int psdiv;			/* prof => stat divider */
    332 int	psratio;			/* ratio: prof / stat */
    333 int	tickfix, tickfixinterval;	/* used if tick not really integral */
    334 #ifndef NTP
    335 static int tickfixcnt;			/* accumulated fractional error */
    336 #else
    337 int	fixtick;			/* used by NTP for same */
    338 int	shifthz;
    339 #endif
    340 
    341 /*
    342  * We might want ldd to load the both words from time at once.
    343  * To succeed we need to be quadword aligned.
    344  * The sparc already does that, and that it has worked so far is a fluke.
    345  */
    346 volatile struct	timeval time  __attribute__((__aligned__(__alignof__(quad_t))));
    347 volatile struct	timeval mono_time;
    348 
    349 void	*softclock_si;
    350 
    351 /*
    352  * Initialize clock frequencies and start both clocks running.
    353  */
    354 void
    355 initclocks(void)
    356 {
    357 	int i;
    358 
    359 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
    360 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
    361 	if (softclock_si == NULL)
    362 		panic("initclocks: unable to register softclock intr");
    363 #endif
    364 
    365 	/*
    366 	 * Set divisors to 1 (normal case) and let the machine-specific
    367 	 * code do its bit.
    368 	 */
    369 	psdiv = 1;
    370 	cpu_initclocks();
    371 
    372 	/*
    373 	 * Compute profhz/stathz/rrticks, and fix profhz if needed.
    374 	 */
    375 	i = stathz ? stathz : hz;
    376 	if (profhz == 0)
    377 		profhz = i;
    378 	psratio = profhz / i;
    379 	rrticks = hz / 10;
    380 	if (schedhz == 0) {
    381 		/* 16Hz is best */
    382 		statscheddiv = i / 16;
    383 		if (statscheddiv <= 0)
    384 			panic("statscheddiv");
    385 	}
    386 
    387 #ifdef NTP
    388 	switch (hz) {
    389 	case 1:
    390 		shifthz = SHIFT_SCALE - 0;
    391 		break;
    392 	case 2:
    393 		shifthz = SHIFT_SCALE - 1;
    394 		break;
    395 	case 4:
    396 		shifthz = SHIFT_SCALE - 2;
    397 		break;
    398 	case 8:
    399 		shifthz = SHIFT_SCALE - 3;
    400 		break;
    401 	case 16:
    402 		shifthz = SHIFT_SCALE - 4;
    403 		break;
    404 	case 32:
    405 		shifthz = SHIFT_SCALE - 5;
    406 		break;
    407 	case 50:
    408 	case 60:
    409 	case 64:
    410 		shifthz = SHIFT_SCALE - 6;
    411 		break;
    412 	case 96:
    413 	case 100:
    414 	case 128:
    415 		shifthz = SHIFT_SCALE - 7;
    416 		break;
    417 	case 256:
    418 		shifthz = SHIFT_SCALE - 8;
    419 		break;
    420 	case 512:
    421 		shifthz = SHIFT_SCALE - 9;
    422 		break;
    423 	case 1000:
    424 	case 1024:
    425 		shifthz = SHIFT_SCALE - 10;
    426 		break;
    427 	case 1200:
    428 	case 2048:
    429 		shifthz = SHIFT_SCALE - 11;
    430 		break;
    431 	case 4096:
    432 		shifthz = SHIFT_SCALE - 12;
    433 		break;
    434 	case 8192:
    435 		shifthz = SHIFT_SCALE - 13;
    436 		break;
    437 	case 16384:
    438 		shifthz = SHIFT_SCALE - 14;
    439 		break;
    440 	case 32768:
    441 		shifthz = SHIFT_SCALE - 15;
    442 		break;
    443 	case 65536:
    444 		shifthz = SHIFT_SCALE - 16;
    445 		break;
    446 	default:
    447 		panic("weird hz");
    448 	}
    449 	if (fixtick == 0) {
    450 		/*
    451 		 * Give MD code a chance to set this to a better
    452 		 * value; but, if it doesn't, we should.
    453 		 */
    454 		fixtick = (1000000 - (hz*tick));
    455 	}
    456 #endif
    457 }
    458 
    459 /*
    460  * The real-time timer, interrupting hz times per second.
    461  */
    462 void
    463 hardclock(struct clockframe *frame)
    464 {
    465 	struct lwp *l;
    466 	struct proc *p;
    467 	int delta;
    468 	extern int tickdelta;
    469 	extern long timedelta;
    470 	struct cpu_info *ci = curcpu();
    471 	struct ptimer *pt;
    472 #ifdef NTP
    473 	int time_update;
    474 	int ltemp;
    475 #endif
    476 
    477 	l = curlwp;
    478 	if (l) {
    479 		p = l->l_proc;
    480 		/*
    481 		 * Run current process's virtual and profile time, as needed.
    482 		 */
    483 		if (CLKF_USERMODE(frame) && p->p_timers &&
    484 		    (pt = LIST_FIRST(&p->p_timers->pts_virtual)) != NULL)
    485 			if (itimerdecr(pt, tick) == 0)
    486 				itimerfire(pt);
    487 		if (p->p_timers &&
    488 		    (pt = LIST_FIRST(&p->p_timers->pts_prof)) != NULL)
    489 			if (itimerdecr(pt, tick) == 0)
    490 				itimerfire(pt);
    491 	}
    492 
    493 	/*
    494 	 * If no separate statistics clock is available, run it from here.
    495 	 */
    496 	if (stathz == 0)
    497 		statclock(frame);
    498 	if ((--ci->ci_schedstate.spc_rrticks) <= 0)
    499 		roundrobin(ci);
    500 
    501 #if defined(MULTIPROCESSOR)
    502 	/*
    503 	 * If we are not the primary CPU, we're not allowed to do
    504 	 * any more work.
    505 	 */
    506 	if (CPU_IS_PRIMARY(ci) == 0)
    507 		return;
    508 #endif
    509 
    510 	/*
    511 	 * Increment the time-of-day.  The increment is normally just
    512 	 * ``tick''.  If the machine is one which has a clock frequency
    513 	 * such that ``hz'' would not divide the second evenly into
    514 	 * milliseconds, a periodic adjustment must be applied.  Finally,
    515 	 * if we are still adjusting the time (see adjtime()),
    516 	 * ``tickdelta'' may also be added in.
    517 	 */
    518 	hardclock_ticks++;
    519 	delta = tick;
    520 
    521 #ifndef NTP
    522 	if (tickfix) {
    523 		tickfixcnt += tickfix;
    524 		if (tickfixcnt >= tickfixinterval) {
    525 			delta++;
    526 			tickfixcnt -= tickfixinterval;
    527 		}
    528 	}
    529 #endif /* !NTP */
    530 	/* Imprecise 4bsd adjtime() handling */
    531 	if (timedelta != 0) {
    532 		delta += tickdelta;
    533 		timedelta -= tickdelta;
    534 	}
    535 
    536 #ifdef notyet
    537 	microset();
    538 #endif
    539 
    540 #ifndef NTP
    541 	BUMPTIME(&time, delta);		/* XXX Now done using NTP code below */
    542 #endif
    543 	BUMPTIME(&mono_time, delta);
    544 
    545 #ifdef NTP
    546 	time_update = delta;
    547 
    548 	/*
    549 	 * Compute the phase adjustment. If the low-order bits
    550 	 * (time_phase) of the update overflow, bump the high-order bits
    551 	 * (time_update).
    552 	 */
    553 	time_phase += time_adj;
    554 	if (time_phase <= -FINEUSEC) {
    555 		ltemp = -time_phase >> SHIFT_SCALE;
    556 		time_phase += ltemp << SHIFT_SCALE;
    557 		time_update -= ltemp;
    558 	} else if (time_phase >= FINEUSEC) {
    559 		ltemp = time_phase >> SHIFT_SCALE;
    560 		time_phase -= ltemp << SHIFT_SCALE;
    561 		time_update += ltemp;
    562 	}
    563 
    564 #ifdef HIGHBALL
    565 	/*
    566 	 * If the HIGHBALL board is installed, we need to adjust the
    567 	 * external clock offset in order to close the hardware feedback
    568 	 * loop. This will adjust the external clock phase and frequency
    569 	 * in small amounts. The additional phase noise and frequency
    570 	 * wander this causes should be minimal. We also need to
    571 	 * discipline the kernel time variable, since the PLL is used to
    572 	 * discipline the external clock. If the Highball board is not
    573 	 * present, we discipline kernel time with the PLL as usual. We
    574 	 * assume that the external clock phase adjustment (time_update)
    575 	 * and kernel phase adjustment (clock_cpu) are less than the
    576 	 * value of tick.
    577 	 */
    578 	clock_offset.tv_usec += time_update;
    579 	if (clock_offset.tv_usec >= 1000000) {
    580 		clock_offset.tv_sec++;
    581 		clock_offset.tv_usec -= 1000000;
    582 	}
    583 	if (clock_offset.tv_usec < 0) {
    584 		clock_offset.tv_sec--;
    585 		clock_offset.tv_usec += 1000000;
    586 	}
    587 	time.tv_usec += clock_cpu;
    588 	clock_cpu = 0;
    589 #else
    590 	time.tv_usec += time_update;
    591 #endif /* HIGHBALL */
    592 
    593 	/*
    594 	 * On rollover of the second the phase adjustment to be used for
    595 	 * the next second is calculated. Also, the maximum error is
    596 	 * increased by the tolerance. If the PPS frequency discipline
    597 	 * code is present, the phase is increased to compensate for the
    598 	 * CPU clock oscillator frequency error.
    599 	 *
    600  	 * On a 32-bit machine and given parameters in the timex.h
    601 	 * header file, the maximum phase adjustment is +-512 ms and
    602 	 * maximum frequency offset is a tad less than) +-512 ppm. On a
    603 	 * 64-bit machine, you shouldn't need to ask.
    604 	 */
    605 	if (time.tv_usec >= 1000000) {
    606 		time.tv_usec -= 1000000;
    607 		time.tv_sec++;
    608 		time_maxerror += time_tolerance >> SHIFT_USEC;
    609 
    610 		/*
    611 		 * Leap second processing. If in leap-insert state at
    612 		 * the end of the day, the system clock is set back one
    613 		 * second; if in leap-delete state, the system clock is
    614 		 * set ahead one second. The microtime() routine or
    615 		 * external clock driver will insure that reported time
    616 		 * is always monotonic. The ugly divides should be
    617 		 * replaced.
    618 		 */
    619 		switch (time_state) {
    620 		case TIME_OK:
    621 			if (time_status & STA_INS)
    622 				time_state = TIME_INS;
    623 			else if (time_status & STA_DEL)
    624 				time_state = TIME_DEL;
    625 			break;
    626 
    627 		case TIME_INS:
    628 			if (time.tv_sec % 86400 == 0) {
    629 				time.tv_sec--;
    630 				time_state = TIME_OOP;
    631 			}
    632 			break;
    633 
    634 		case TIME_DEL:
    635 			if ((time.tv_sec + 1) % 86400 == 0) {
    636 				time.tv_sec++;
    637 				time_state = TIME_WAIT;
    638 			}
    639 			break;
    640 
    641 		case TIME_OOP:
    642 			time_state = TIME_WAIT;
    643 			break;
    644 
    645 		case TIME_WAIT:
    646 			if (!(time_status & (STA_INS | STA_DEL)))
    647 				time_state = TIME_OK;
    648 			break;
    649 		}
    650 
    651 		/*
    652 		 * Compute the phase adjustment for the next second. In
    653 		 * PLL mode, the offset is reduced by a fixed factor
    654 		 * times the time constant. In FLL mode the offset is
    655 		 * used directly. In either mode, the maximum phase
    656 		 * adjustment for each second is clamped so as to spread
    657 		 * the adjustment over not more than the number of
    658 		 * seconds between updates.
    659 		 */
    660 		if (time_offset < 0) {
    661 			ltemp = -time_offset;
    662 			if (!(time_status & STA_FLL))
    663 				ltemp >>= SHIFT_KG + time_constant;
    664 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    665 				ltemp = (MAXPHASE / MINSEC) <<
    666 				    SHIFT_UPDATE;
    667 			time_offset += ltemp;
    668 			time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
    669 		} else if (time_offset > 0) {
    670 			ltemp = time_offset;
    671 			if (!(time_status & STA_FLL))
    672 				ltemp >>= SHIFT_KG + time_constant;
    673 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    674 				ltemp = (MAXPHASE / MINSEC) <<
    675 				    SHIFT_UPDATE;
    676 			time_offset -= ltemp;
    677 			time_adj = ltemp << (shifthz - SHIFT_UPDATE);
    678 		} else
    679 			time_adj = 0;
    680 
    681 		/*
    682 		 * Compute the frequency estimate and additional phase
    683 		 * adjustment due to frequency error for the next
    684 		 * second. When the PPS signal is engaged, gnaw on the
    685 		 * watchdog counter and update the frequency computed by
    686 		 * the pll and the PPS signal.
    687 		 */
    688 #ifdef PPS_SYNC
    689 		pps_valid++;
    690 		if (pps_valid == PPS_VALID) {
    691 			pps_jitter = MAXTIME;
    692 			pps_stabil = MAXFREQ;
    693 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
    694 			    STA_PPSWANDER | STA_PPSERROR);
    695 		}
    696 		ltemp = time_freq + pps_freq;
    697 #else
    698 		ltemp = time_freq;
    699 #endif /* PPS_SYNC */
    700 
    701 		if (ltemp < 0)
    702 			time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
    703 		else
    704 			time_adj += ltemp >> (SHIFT_USEC - shifthz);
    705 		time_adj += (long)fixtick << shifthz;
    706 
    707 		/*
    708 		 * When the CPU clock oscillator frequency is not a
    709 		 * power of 2 in Hz, shifthz is only an approximate
    710 		 * scale factor.
    711 		 *
    712 		 * To determine the adjustment, you can do the following:
    713 		 *   bc -q
    714 		 *   scale=24
    715 		 *   obase=2
    716 		 *   idealhz/realhz
    717 		 * where `idealhz' is the next higher power of 2, and `realhz'
    718 		 * is the actual value.  You may need to factor this result
    719 		 * into a sequence of 2 multipliers to get better precision.
    720 		 *
    721 		 * Likewise, the error can be calculated with (e.g. for 100Hz):
    722 		 *   bc -q
    723 		 *   scale=24
    724 		 *   ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz
    725 		 * (and then multiply by 1000000 to get ppm).
    726 		 */
    727 		switch (hz) {
    728 		case 60:
    729 			/* A factor of 1.000100010001 gives about 15ppm
    730 			   error. */
    731 			if (time_adj < 0) {
    732 				time_adj -= (-time_adj >> 4);
    733 				time_adj -= (-time_adj >> 8);
    734 			} else {
    735 				time_adj += (time_adj >> 4);
    736 				time_adj += (time_adj >> 8);
    737 			}
    738 			break;
    739 
    740 		case 96:
    741 			/* A factor of 1.0101010101 gives about 244ppm error. */
    742 			if (time_adj < 0) {
    743 				time_adj -= (-time_adj >> 2);
    744 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    745 			} else {
    746 				time_adj += (time_adj >> 2);
    747 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    748 			}
    749 			break;
    750 
    751 		case 50:
    752 		case 100:
    753 			/* A factor of 1.010001111010111 gives about 1ppm
    754 			   error. */
    755 			if (time_adj < 0) {
    756 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    757 				time_adj += (-time_adj >> 10);
    758 			} else {
    759 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    760 				time_adj -= (time_adj >> 10);
    761 			}
    762 			break;
    763 
    764 		case 1000:
    765 			/* A factor of 1.000001100010100001 gives about 50ppm
    766 			   error. */
    767 			if (time_adj < 0) {
    768 				time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
    769 				time_adj -= (-time_adj >> 7);
    770 			} else {
    771 				time_adj += (time_adj >> 6) + (time_adj >> 11);
    772 				time_adj += (time_adj >> 7);
    773 			}
    774 			break;
    775 
    776 		case 1200:
    777 			/* A factor of 1.1011010011100001 gives about 64ppm
    778 			   error. */
    779 			if (time_adj < 0) {
    780 				time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
    781 				time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
    782 			} else {
    783 				time_adj += (time_adj >> 1) + (time_adj >> 6);
    784 				time_adj += (time_adj >> 3) + (time_adj >> 10);
    785 			}
    786 			break;
    787 		}
    788 
    789 #ifdef EXT_CLOCK
    790 		/*
    791 		 * If an external clock is present, it is necessary to
    792 		 * discipline the kernel time variable anyway, since not
    793 		 * all system components use the microtime() interface.
    794 		 * Here, the time offset between the external clock and
    795 		 * kernel time variable is computed every so often.
    796 		 */
    797 		clock_count++;
    798 		if (clock_count > CLOCK_INTERVAL) {
    799 			clock_count = 0;
    800 			microtime(&clock_ext);
    801 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    802 			delta.tv_usec = clock_ext.tv_usec -
    803 			    time.tv_usec;
    804 			if (delta.tv_usec < 0)
    805 				delta.tv_sec--;
    806 			if (delta.tv_usec >= 500000) {
    807 				delta.tv_usec -= 1000000;
    808 				delta.tv_sec++;
    809 			}
    810 			if (delta.tv_usec < -500000) {
    811 				delta.tv_usec += 1000000;
    812 				delta.tv_sec--;
    813 			}
    814 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    815 			    delta.tv_usec > MAXPHASE) ||
    816 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    817 			    delta.tv_usec < -MAXPHASE)) {
    818 				time = clock_ext;
    819 				delta.tv_sec = 0;
    820 				delta.tv_usec = 0;
    821 			}
    822 #ifdef HIGHBALL
    823 			clock_cpu = delta.tv_usec;
    824 #else /* HIGHBALL */
    825 			hardupdate(delta.tv_usec);
    826 #endif /* HIGHBALL */
    827 		}
    828 #endif /* EXT_CLOCK */
    829 	}
    830 
    831 #endif /* NTP */
    832 
    833 	/*
    834 	 * Update real-time timeout queue.
    835 	 * Process callouts at a very low CPU priority, so we don't keep the
    836 	 * relatively high clock interrupt priority any longer than necessary.
    837 	 */
    838 	if (callout_hardclock()) {
    839 		if (CLKF_BASEPRI(frame)) {
    840 			/*
    841 			 * Save the overhead of a software interrupt;
    842 			 * it will happen as soon as we return, so do
    843 			 * it now.
    844 			 */
    845 			spllowersoftclock();
    846 			KERNEL_LOCK(LK_CANRECURSE|LK_EXCLUSIVE);
    847 			softclock(NULL);
    848 			KERNEL_UNLOCK();
    849 		} else {
    850 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
    851 			softintr_schedule(softclock_si);
    852 #else
    853 			setsoftclock();
    854 #endif
    855 		}
    856 	}
    857 }
    858 
    859 /*
    860  * Compute number of hz until specified time.  Used to compute second
    861  * argument to callout_reset() from an absolute time.
    862  */
    863 int
    864 hzto(struct timeval *tv)
    865 {
    866 	unsigned long ticks;
    867 	long sec, usec;
    868 	int s;
    869 
    870 	/*
    871 	 * If the number of usecs in the whole seconds part of the time
    872 	 * difference fits in a long, then the total number of usecs will
    873 	 * fit in an unsigned long.  Compute the total and convert it to
    874 	 * ticks, rounding up and adding 1 to allow for the current tick
    875 	 * to expire.  Rounding also depends on unsigned long arithmetic
    876 	 * to avoid overflow.
    877 	 *
    878 	 * Otherwise, if the number of ticks in the whole seconds part of
    879 	 * the time difference fits in a long, then convert the parts to
    880 	 * ticks separately and add, using similar rounding methods and
    881 	 * overflow avoidance.  This method would work in the previous
    882 	 * case, but it is slightly slower and assume that hz is integral.
    883 	 *
    884 	 * Otherwise, round the time difference down to the maximum
    885 	 * representable value.
    886 	 *
    887 	 * If ints are 32-bit, then the maximum value for any timeout in
    888 	 * 10ms ticks is 248 days.
    889 	 */
    890 	s = splclock();
    891 	sec = tv->tv_sec - time.tv_sec;
    892 	usec = tv->tv_usec - time.tv_usec;
    893 	splx(s);
    894 
    895 	if (usec < 0) {
    896 		sec--;
    897 		usec += 1000000;
    898 	}
    899 
    900 	if (sec < 0 || (sec == 0 && usec <= 0)) {
    901 		/*
    902 		 * Would expire now or in the past.  Return 0 ticks.
    903 		 * This is different from the legacy hzto() interface,
    904 		 * and callers need to check for it.
    905 		 */
    906 		ticks = 0;
    907 	} else if (sec <= (LONG_MAX / 1000000))
    908 		ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1))
    909 		    / tick) + 1;
    910 	else if (sec <= (LONG_MAX / hz))
    911 		ticks = (sec * hz) +
    912 		    (((unsigned long)usec + (tick - 1)) / tick) + 1;
    913 	else
    914 		ticks = LONG_MAX;
    915 
    916 	if (ticks > INT_MAX)
    917 		ticks = INT_MAX;
    918 
    919 	return ((int)ticks);
    920 }
    921 
    922 /*
    923  * Start profiling on a process.
    924  *
    925  * Kernel profiling passes proc0 which never exits and hence
    926  * keeps the profile clock running constantly.
    927  */
    928 void
    929 startprofclock(struct proc *p)
    930 {
    931 
    932 	if ((p->p_flag & P_PROFIL) == 0) {
    933 		p->p_flag |= P_PROFIL;
    934 		/*
    935 		 * This is only necessary if using the clock as the
    936 		 * profiling source.
    937 		 */
    938 		if (++profprocs == 1 && stathz != 0)
    939 			psdiv = psratio;
    940 	}
    941 }
    942 
    943 /*
    944  * Stop profiling on a process.
    945  */
    946 void
    947 stopprofclock(struct proc *p)
    948 {
    949 
    950 	if (p->p_flag & P_PROFIL) {
    951 		p->p_flag &= ~P_PROFIL;
    952 		/*
    953 		 * This is only necessary if using the clock as the
    954 		 * profiling source.
    955 		 */
    956 		if (--profprocs == 0 && stathz != 0)
    957 			psdiv = 1;
    958 	}
    959 }
    960 
    961 #if defined(PERFCTRS)
    962 /*
    963  * Independent profiling "tick" in case we're using a separate
    964  * clock or profiling event source.  Currently, that's just
    965  * performance counters--hence the wrapper.
    966  */
    967 void
    968 proftick(struct clockframe *frame)
    969 {
    970 #ifdef GPROF
    971         struct gmonparam *g;
    972         intptr_t i;
    973 #endif
    974 	struct proc *p;
    975 
    976 	p = curproc;
    977 	if (CLKF_USERMODE(frame)) {
    978 		if (p->p_flag & P_PROFIL)
    979 			addupc_intr(p, CLKF_PC(frame));
    980 	} else {
    981 #ifdef GPROF
    982 		g = &_gmonparam;
    983 		if (g->state == GMON_PROF_ON) {
    984 			i = CLKF_PC(frame) - g->lowpc;
    985 			if (i < g->textsize) {
    986 				i /= HISTFRACTION * sizeof(*g->kcount);
    987 				g->kcount[i]++;
    988 			}
    989 		}
    990 #endif
    991 #ifdef PROC_PC
    992                 if (p && p->p_flag & P_PROFIL)
    993                         addupc_intr(p, PROC_PC(p));
    994 #endif
    995 	}
    996 }
    997 #endif
    998 
    999 /*
   1000  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1001  * do process and kernel statistics.
   1002  */
   1003 void
   1004 statclock(struct clockframe *frame)
   1005 {
   1006 #ifdef GPROF
   1007 	struct gmonparam *g;
   1008 	intptr_t i;
   1009 #endif
   1010 	struct cpu_info *ci = curcpu();
   1011 	struct schedstate_percpu *spc = &ci->ci_schedstate;
   1012 	struct lwp *l;
   1013 	struct proc *p;
   1014 
   1015 	/*
   1016 	 * Notice changes in divisor frequency, and adjust clock
   1017 	 * frequency accordingly.
   1018 	 */
   1019 	if (spc->spc_psdiv != psdiv) {
   1020 		spc->spc_psdiv = psdiv;
   1021 		spc->spc_pscnt = psdiv;
   1022 		if (psdiv == 1) {
   1023 			setstatclockrate(stathz);
   1024 		} else {
   1025 			setstatclockrate(profhz);
   1026 		}
   1027 	}
   1028 	l = curlwp;
   1029 	p = (l ? l->l_proc : 0);
   1030 	if (CLKF_USERMODE(frame)) {
   1031 		if (p->p_flag & P_PROFIL && profsrc == PROFSRC_CLOCK)
   1032 			addupc_intr(p, CLKF_PC(frame));
   1033 		if (--spc->spc_pscnt > 0)
   1034 			return;
   1035 		/*
   1036 		 * Came from user mode; CPU was in user state.
   1037 		 * If this process is being profiled record the tick.
   1038 		 */
   1039 		p->p_uticks++;
   1040 		if (p->p_nice > NZERO)
   1041 			spc->spc_cp_time[CP_NICE]++;
   1042 		else
   1043 			spc->spc_cp_time[CP_USER]++;
   1044 	} else {
   1045 #ifdef GPROF
   1046 		/*
   1047 		 * Kernel statistics are just like addupc_intr, only easier.
   1048 		 */
   1049 		g = &_gmonparam;
   1050 		if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) {
   1051 			i = CLKF_PC(frame) - g->lowpc;
   1052 			if (i < g->textsize) {
   1053 				i /= HISTFRACTION * sizeof(*g->kcount);
   1054 				g->kcount[i]++;
   1055 			}
   1056 		}
   1057 #endif
   1058 #ifdef LWP_PC
   1059 		if (p && profsrc == PROFSRC_CLOCK && p->p_flag & P_PROFIL)
   1060 			addupc_intr(p, LWP_PC(l));
   1061 #endif
   1062 		if (--spc->spc_pscnt > 0)
   1063 			return;
   1064 		/*
   1065 		 * Came from kernel mode, so we were:
   1066 		 * - handling an interrupt,
   1067 		 * - doing syscall or trap work on behalf of the current
   1068 		 *   user process, or
   1069 		 * - spinning in the idle loop.
   1070 		 * Whichever it is, charge the time as appropriate.
   1071 		 * Note that we charge interrupts to the current process,
   1072 		 * regardless of whether they are ``for'' that process,
   1073 		 * so that we know how much of its real time was spent
   1074 		 * in ``non-process'' (i.e., interrupt) work.
   1075 		 */
   1076 		if (CLKF_INTR(frame)) {
   1077 			if (p != NULL)
   1078 				p->p_iticks++;
   1079 			spc->spc_cp_time[CP_INTR]++;
   1080 		} else if (p != NULL) {
   1081 			p->p_sticks++;
   1082 			spc->spc_cp_time[CP_SYS]++;
   1083 		} else
   1084 			spc->spc_cp_time[CP_IDLE]++;
   1085 	}
   1086 	spc->spc_pscnt = psdiv;
   1087 
   1088 	if (l != NULL) {
   1089 		++p->p_cpticks;
   1090 		/*
   1091 		 * If no separate schedclock is provided, call it here
   1092 		 * at about 16 Hz.
   1093 		 */
   1094 		if (schedhz == 0)
   1095 			if ((int)(--ci->ci_schedstate.spc_schedticks) <= 0) {
   1096 				schedclock(l);
   1097 				ci->ci_schedstate.spc_schedticks = statscheddiv;
   1098 			}
   1099 	}
   1100 }
   1101 
   1102 
   1103 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1104 
   1105 /*
   1106  * hardupdate() - local clock update
   1107  *
   1108  * This routine is called by ntp_adjtime() to update the local clock
   1109  * phase and frequency. The implementation is of an adaptive-parameter,
   1110  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1111  * time and frequency offset estimates for each call. If the kernel PPS
   1112  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1113  * determines the new time offset, instead of the calling argument.
   1114  * Presumably, calls to ntp_adjtime() occur only when the caller
   1115  * believes the local clock is valid within some bound (+-128 ms with
   1116  * NTP). If the caller's time is far different than the PPS time, an
   1117  * argument will ensue, and it's not clear who will lose.
   1118  *
   1119  * For uncompensated quartz crystal oscillatores and nominal update
   1120  * intervals less than 1024 s, operation should be in phase-lock mode
   1121  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1122  * intervals greater than thiss, operation should be in frequency-lock
   1123  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1124  *
   1125  * Note: splclock() is in effect.
   1126  */
   1127 void
   1128 hardupdate(long offset)
   1129 {
   1130 	long ltemp, mtemp;
   1131 
   1132 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1133 		return;
   1134 	ltemp = offset;
   1135 #ifdef PPS_SYNC
   1136 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1137 		ltemp = pps_offset;
   1138 #endif /* PPS_SYNC */
   1139 
   1140 	/*
   1141 	 * Scale the phase adjustment and clamp to the operating range.
   1142 	 */
   1143 	if (ltemp > MAXPHASE)
   1144 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1145 	else if (ltemp < -MAXPHASE)
   1146 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1147 	else
   1148 		time_offset = ltemp << SHIFT_UPDATE;
   1149 
   1150 	/*
   1151 	 * Select whether the frequency is to be controlled and in which
   1152 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1153 	 * multiply/divide should be replaced someday.
   1154 	 */
   1155 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1156 		time_reftime = time.tv_sec;
   1157 	mtemp = time.tv_sec - time_reftime;
   1158 	time_reftime = time.tv_sec;
   1159 	if (time_status & STA_FLL) {
   1160 		if (mtemp >= MINSEC) {
   1161 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1162 			    SHIFT_UPDATE));
   1163 			if (ltemp < 0)
   1164 				time_freq -= -ltemp >> SHIFT_KH;
   1165 			else
   1166 				time_freq += ltemp >> SHIFT_KH;
   1167 		}
   1168 	} else {
   1169 		if (mtemp < MAXSEC) {
   1170 			ltemp *= mtemp;
   1171 			if (ltemp < 0)
   1172 				time_freq -= -ltemp >> (time_constant +
   1173 				    time_constant + SHIFT_KF -
   1174 				    SHIFT_USEC);
   1175 			else
   1176 				time_freq += ltemp >> (time_constant +
   1177 				    time_constant + SHIFT_KF -
   1178 				    SHIFT_USEC);
   1179 		}
   1180 	}
   1181 	if (time_freq > time_tolerance)
   1182 		time_freq = time_tolerance;
   1183 	else if (time_freq < -time_tolerance)
   1184 		time_freq = -time_tolerance;
   1185 }
   1186 
   1187 #ifdef PPS_SYNC
   1188 /*
   1189  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1190  *
   1191  * This routine is called at each PPS interrupt in order to discipline
   1192  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1193  * and leaves it in a handy spot for the hardclock() routine. It
   1194  * integrates successive PPS phase differences and calculates the
   1195  * frequency offset. This is used in hardclock() to discipline the CPU
   1196  * clock oscillator so that intrinsic frequency error is cancelled out.
   1197  * The code requires the caller to capture the time and hardware counter
   1198  * value at the on-time PPS signal transition.
   1199  *
   1200  * Note that, on some Unix systems, this routine runs at an interrupt
   1201  * priority level higher than the timer interrupt routine hardclock().
   1202  * Therefore, the variables used are distinct from the hardclock()
   1203  * variables, except for certain exceptions: The PPS frequency pps_freq
   1204  * and phase pps_offset variables are determined by this routine and
   1205  * updated atomically. The time_tolerance variable can be considered a
   1206  * constant, since it is infrequently changed, and then only when the
   1207  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1208  * once per second by hardclock() and is atomically cleared in this
   1209  * routine.
   1210  */
   1211 void
   1212 hardpps(struct timeval *tvp,		/* time at PPS */
   1213 	long usec			/* hardware counter at PPS */)
   1214 {
   1215 	long u_usec, v_usec, bigtick;
   1216 	long cal_sec, cal_usec;
   1217 
   1218 	/*
   1219 	 * An occasional glitch can be produced when the PPS interrupt
   1220 	 * occurs in the hardclock() routine before the time variable is
   1221 	 * updated. Here the offset is discarded when the difference
   1222 	 * between it and the last one is greater than tick/2, but not
   1223 	 * if the interval since the first discard exceeds 30 s.
   1224 	 */
   1225 	time_status |= STA_PPSSIGNAL;
   1226 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1227 	pps_valid = 0;
   1228 	u_usec = -tvp->tv_usec;
   1229 	if (u_usec < -500000)
   1230 		u_usec += 1000000;
   1231 	v_usec = pps_offset - u_usec;
   1232 	if (v_usec < 0)
   1233 		v_usec = -v_usec;
   1234 	if (v_usec > (tick >> 1)) {
   1235 		if (pps_glitch > MAXGLITCH) {
   1236 			pps_glitch = 0;
   1237 			pps_tf[2] = u_usec;
   1238 			pps_tf[1] = u_usec;
   1239 		} else {
   1240 			pps_glitch++;
   1241 			u_usec = pps_offset;
   1242 		}
   1243 	} else
   1244 		pps_glitch = 0;
   1245 
   1246 	/*
   1247 	 * A three-stage median filter is used to help deglitch the pps
   1248 	 * time. The median sample becomes the time offset estimate; the
   1249 	 * difference between the other two samples becomes the time
   1250 	 * dispersion (jitter) estimate.
   1251 	 */
   1252 	pps_tf[2] = pps_tf[1];
   1253 	pps_tf[1] = pps_tf[0];
   1254 	pps_tf[0] = u_usec;
   1255 	if (pps_tf[0] > pps_tf[1]) {
   1256 		if (pps_tf[1] > pps_tf[2]) {
   1257 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1258 			v_usec = pps_tf[0] - pps_tf[2];
   1259 		} else if (pps_tf[2] > pps_tf[0]) {
   1260 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1261 			v_usec = pps_tf[2] - pps_tf[1];
   1262 		} else {
   1263 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1264 			v_usec = pps_tf[0] - pps_tf[1];
   1265 		}
   1266 	} else {
   1267 		if (pps_tf[1] < pps_tf[2]) {
   1268 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1269 			v_usec = pps_tf[2] - pps_tf[0];
   1270 		} else  if (pps_tf[2] < pps_tf[0]) {
   1271 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1272 			v_usec = pps_tf[1] - pps_tf[2];
   1273 		} else {
   1274 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1275 			v_usec = pps_tf[1] - pps_tf[0];
   1276 		}
   1277 	}
   1278 	if (v_usec > MAXTIME)
   1279 		pps_jitcnt++;
   1280 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1281 	if (v_usec < 0)
   1282 		pps_jitter -= -v_usec >> PPS_AVG;
   1283 	else
   1284 		pps_jitter += v_usec >> PPS_AVG;
   1285 	if (pps_jitter > (MAXTIME >> 1))
   1286 		time_status |= STA_PPSJITTER;
   1287 
   1288 	/*
   1289 	 * During the calibration interval adjust the starting time when
   1290 	 * the tick overflows. At the end of the interval compute the
   1291 	 * duration of the interval and the difference of the hardware
   1292 	 * counters at the beginning and end of the interval. This code
   1293 	 * is deliciously complicated by the fact valid differences may
   1294 	 * exceed the value of tick when using long calibration
   1295 	 * intervals and small ticks. Note that the counter can be
   1296 	 * greater than tick if caught at just the wrong instant, but
   1297 	 * the values returned and used here are correct.
   1298 	 */
   1299 	bigtick = (long)tick << SHIFT_USEC;
   1300 	pps_usec -= pps_freq;
   1301 	if (pps_usec >= bigtick)
   1302 		pps_usec -= bigtick;
   1303 	if (pps_usec < 0)
   1304 		pps_usec += bigtick;
   1305 	pps_time.tv_sec++;
   1306 	pps_count++;
   1307 	if (pps_count < (1 << pps_shift))
   1308 		return;
   1309 	pps_count = 0;
   1310 	pps_calcnt++;
   1311 	u_usec = usec << SHIFT_USEC;
   1312 	v_usec = pps_usec - u_usec;
   1313 	if (v_usec >= bigtick >> 1)
   1314 		v_usec -= bigtick;
   1315 	if (v_usec < -(bigtick >> 1))
   1316 		v_usec += bigtick;
   1317 	if (v_usec < 0)
   1318 		v_usec = -(-v_usec >> pps_shift);
   1319 	else
   1320 		v_usec = v_usec >> pps_shift;
   1321 	pps_usec = u_usec;
   1322 	cal_sec = tvp->tv_sec;
   1323 	cal_usec = tvp->tv_usec;
   1324 	cal_sec -= pps_time.tv_sec;
   1325 	cal_usec -= pps_time.tv_usec;
   1326 	if (cal_usec < 0) {
   1327 		cal_usec += 1000000;
   1328 		cal_sec--;
   1329 	}
   1330 	pps_time = *tvp;
   1331 
   1332 	/*
   1333 	 * Check for lost interrupts, noise, excessive jitter and
   1334 	 * excessive frequency error. The number of timer ticks during
   1335 	 * the interval may vary +-1 tick. Add to this a margin of one
   1336 	 * tick for the PPS signal jitter and maximum frequency
   1337 	 * deviation. If the limits are exceeded, the calibration
   1338 	 * interval is reset to the minimum and we start over.
   1339 	 */
   1340 	u_usec = (long)tick << 1;
   1341 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1342 	    || (cal_sec == 0 && cal_usec < u_usec))
   1343 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1344 		pps_errcnt++;
   1345 		pps_shift = PPS_SHIFT;
   1346 		pps_intcnt = 0;
   1347 		time_status |= STA_PPSERROR;
   1348 		return;
   1349 	}
   1350 
   1351 	/*
   1352 	 * A three-stage median filter is used to help deglitch the pps
   1353 	 * frequency. The median sample becomes the frequency offset
   1354 	 * estimate; the difference between the other two samples
   1355 	 * becomes the frequency dispersion (stability) estimate.
   1356 	 */
   1357 	pps_ff[2] = pps_ff[1];
   1358 	pps_ff[1] = pps_ff[0];
   1359 	pps_ff[0] = v_usec;
   1360 	if (pps_ff[0] > pps_ff[1]) {
   1361 		if (pps_ff[1] > pps_ff[2]) {
   1362 			u_usec = pps_ff[1];		/* 0 1 2 */
   1363 			v_usec = pps_ff[0] - pps_ff[2];
   1364 		} else if (pps_ff[2] > pps_ff[0]) {
   1365 			u_usec = pps_ff[0];		/* 2 0 1 */
   1366 			v_usec = pps_ff[2] - pps_ff[1];
   1367 		} else {
   1368 			u_usec = pps_ff[2];		/* 0 2 1 */
   1369 			v_usec = pps_ff[0] - pps_ff[1];
   1370 		}
   1371 	} else {
   1372 		if (pps_ff[1] < pps_ff[2]) {
   1373 			u_usec = pps_ff[1];		/* 2 1 0 */
   1374 			v_usec = pps_ff[2] - pps_ff[0];
   1375 		} else  if (pps_ff[2] < pps_ff[0]) {
   1376 			u_usec = pps_ff[0];		/* 1 0 2 */
   1377 			v_usec = pps_ff[1] - pps_ff[2];
   1378 		} else {
   1379 			u_usec = pps_ff[2];		/* 1 2 0 */
   1380 			v_usec = pps_ff[1] - pps_ff[0];
   1381 		}
   1382 	}
   1383 
   1384 	/*
   1385 	 * Here the frequency dispersion (stability) is updated. If it
   1386 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1387 	 * offset is updated as well, but clamped to the tolerance. It
   1388 	 * will be processed later by the hardclock() routine.
   1389 	 */
   1390 	v_usec = (v_usec >> 1) - pps_stabil;
   1391 	if (v_usec < 0)
   1392 		pps_stabil -= -v_usec >> PPS_AVG;
   1393 	else
   1394 		pps_stabil += v_usec >> PPS_AVG;
   1395 	if (pps_stabil > MAXFREQ >> 2) {
   1396 		pps_stbcnt++;
   1397 		time_status |= STA_PPSWANDER;
   1398 		return;
   1399 	}
   1400 	if (time_status & STA_PPSFREQ) {
   1401 		if (u_usec < 0) {
   1402 			pps_freq -= -u_usec >> PPS_AVG;
   1403 			if (pps_freq < -time_tolerance)
   1404 				pps_freq = -time_tolerance;
   1405 			u_usec = -u_usec;
   1406 		} else {
   1407 			pps_freq += u_usec >> PPS_AVG;
   1408 			if (pps_freq > time_tolerance)
   1409 				pps_freq = time_tolerance;
   1410 		}
   1411 	}
   1412 
   1413 	/*
   1414 	 * Here the calibration interval is adjusted. If the maximum
   1415 	 * time difference is greater than tick / 4, reduce the interval
   1416 	 * by half. If this is not the case for four consecutive
   1417 	 * intervals, double the interval.
   1418 	 */
   1419 	if (u_usec << pps_shift > bigtick >> 2) {
   1420 		pps_intcnt = 0;
   1421 		if (pps_shift > PPS_SHIFT)
   1422 			pps_shift--;
   1423 	} else if (pps_intcnt >= 4) {
   1424 		pps_intcnt = 0;
   1425 		if (pps_shift < PPS_SHIFTMAX)
   1426 			pps_shift++;
   1427 	} else
   1428 		pps_intcnt++;
   1429 }
   1430 #endif /* PPS_SYNC */
   1431 #endif /* NTP  */
   1432