Home | History | Annotate | Line # | Download | only in kern
kern_clock.c revision 1.74
      1 /*	$NetBSD: kern_clock.c,v 1.74 2001/01/17 18:21:41 thorpej Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2000 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
      9  * NASA Ames Research Center.
     10  *
     11  * Redistribution and use in source and binary forms, with or without
     12  * modification, are permitted provided that the following conditions
     13  * are met:
     14  * 1. Redistributions of source code must retain the above copyright
     15  *    notice, this list of conditions and the following disclaimer.
     16  * 2. Redistributions in binary form must reproduce the above copyright
     17  *    notice, this list of conditions and the following disclaimer in the
     18  *    documentation and/or other materials provided with the distribution.
     19  * 3. All advertising materials mentioning features or use of this software
     20  *    must display the following acknowledgement:
     21  *	This product includes software developed by the NetBSD
     22  *	Foundation, Inc. and its contributors.
     23  * 4. Neither the name of The NetBSD Foundation nor the names of its
     24  *    contributors may be used to endorse or promote products derived
     25  *    from this software without specific prior written permission.
     26  *
     27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     37  * POSSIBILITY OF SUCH DAMAGE.
     38  */
     39 
     40 /*-
     41  * Copyright (c) 1982, 1986, 1991, 1993
     42  *	The Regents of the University of California.  All rights reserved.
     43  * (c) UNIX System Laboratories, Inc.
     44  * All or some portions of this file are derived from material licensed
     45  * to the University of California by American Telephone and Telegraph
     46  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     47  * the permission of UNIX System Laboratories, Inc.
     48  *
     49  * Redistribution and use in source and binary forms, with or without
     50  * modification, are permitted provided that the following conditions
     51  * are met:
     52  * 1. Redistributions of source code must retain the above copyright
     53  *    notice, this list of conditions and the following disclaimer.
     54  * 2. Redistributions in binary form must reproduce the above copyright
     55  *    notice, this list of conditions and the following disclaimer in the
     56  *    documentation and/or other materials provided with the distribution.
     57  * 3. All advertising materials mentioning features or use of this software
     58  *    must display the following acknowledgement:
     59  *	This product includes software developed by the University of
     60  *	California, Berkeley and its contributors.
     61  * 4. Neither the name of the University nor the names of its contributors
     62  *    may be used to endorse or promote products derived from this software
     63  *    without specific prior written permission.
     64  *
     65  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     66  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     67  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     68  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     69  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     70  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     71  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     72  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     73  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     74  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     75  * SUCH DAMAGE.
     76  *
     77  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
     78  */
     79 
     80 #include "opt_ntp.h"
     81 
     82 #include <sys/param.h>
     83 #include <sys/systm.h>
     84 #include <sys/dkstat.h>
     85 #include <sys/callout.h>
     86 #include <sys/kernel.h>
     87 #include <sys/proc.h>
     88 #include <sys/resourcevar.h>
     89 #include <sys/signalvar.h>
     90 #include <uvm/uvm_extern.h>
     91 #include <sys/sysctl.h>
     92 #include <sys/timex.h>
     93 #include <sys/sched.h>
     94 
     95 #include <machine/cpu.h>
     96 #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 struct timeval pps_time;	/* kernel time at last interval */
    235 long pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
    236 long pps_offset = 0;		/* pps time offset (us) */
    237 long pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
    238 long pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
    239 long pps_freq = 0;		/* frequency offset (scaled ppm) */
    240 long pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
    241 long pps_usec = 0;		/* microsec counter at last interval */
    242 long pps_valid = PPS_VALID;	/* pps signal watchdog counter */
    243 int pps_glitch = 0;		/* pps signal glitch counter */
    244 int pps_count = 0;		/* calibration interval counter (s) */
    245 int pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
    246 int pps_intcnt = 0;		/* intervals at current duration */
    247 
    248 /*
    249  * PPS signal quality monitors
    250  *
    251  * pps_jitcnt counts the seconds that have been discarded because the
    252  * jitter measured by the time median filter exceeds the limit MAXTIME
    253  * (100 us).
    254  *
    255  * pps_calcnt counts the frequency calibration intervals, which are
    256  * variable from 4 s to 256 s.
    257  *
    258  * pps_errcnt counts the calibration intervals which have been discarded
    259  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
    260  * calibration interval jitter exceeds two ticks.
    261  *
    262  * pps_stbcnt counts the calibration intervals that have been discarded
    263  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
    264  */
    265 long pps_jitcnt = 0;		/* jitter limit exceeded */
    266 long pps_calcnt = 0;		/* calibration intervals */
    267 long pps_errcnt = 0;		/* calibration errors */
    268 long pps_stbcnt = 0;		/* stability limit exceeded */
    269 #endif /* PPS_SYNC */
    270 
    271 #ifdef EXT_CLOCK
    272 /*
    273  * External clock definitions
    274  *
    275  * The following definitions and declarations are used only if an
    276  * external clock is configured on the system.
    277  */
    278 #define CLOCK_INTERVAL 30	/* CPU clock update interval (s) */
    279 
    280 /*
    281  * The clock_count variable is set to CLOCK_INTERVAL at each PPS
    282  * interrupt and decremented once each second.
    283  */
    284 int clock_count = 0;		/* CPU clock counter */
    285 
    286 #ifdef HIGHBALL
    287 /*
    288  * The clock_offset and clock_cpu variables are used by the HIGHBALL
    289  * interface. The clock_offset variable defines the offset between
    290  * system time and the HIGBALL counters. The clock_cpu variable contains
    291  * the offset between the system clock and the HIGHBALL clock for use in
    292  * disciplining the kernel time variable.
    293  */
    294 extern struct timeval clock_offset; /* Highball clock offset */
    295 long clock_cpu = 0;		/* CPU clock adjust */
    296 #endif /* HIGHBALL */
    297 #endif /* EXT_CLOCK */
    298 #endif /* NTP */
    299 
    300 
    301 /*
    302  * Bump a timeval by a small number of usec's.
    303  */
    304 #define BUMPTIME(t, usec) { \
    305 	volatile struct timeval *tp = (t); \
    306 	long us; \
    307  \
    308 	tp->tv_usec = us = tp->tv_usec + (usec); \
    309 	if (us >= 1000000) { \
    310 		tp->tv_usec = us - 1000000; \
    311 		tp->tv_sec++; \
    312 	} \
    313 }
    314 
    315 int	stathz;
    316 int	profhz;
    317 int	profprocs;
    318 int	softclock_running;		/* 1 => softclock() is running */
    319 static int psdiv;			/* prof => stat divider */
    320 int	psratio;			/* ratio: prof / stat */
    321 int	tickfix, tickfixinterval;	/* used if tick not really integral */
    322 #ifndef NTP
    323 static int tickfixcnt;			/* accumulated fractional error */
    324 #else
    325 int	fixtick;			/* used by NTP for same */
    326 int	shifthz;
    327 #endif
    328 
    329 /*
    330  * We might want ldd to load the both words from time at once.
    331  * To succeed we need to be quadword aligned.
    332  * The sparc already does that, and that it has worked so far is a fluke.
    333  */
    334 volatile struct	timeval time  __attribute__((__aligned__(__alignof__(quad_t))));
    335 volatile struct	timeval mono_time;
    336 
    337 /*
    338  * The callout mechanism is based on the work of Adam M. Costello and
    339  * George Varghese, published in a technical report entitled "Redesigning
    340  * the BSD Callout and Timer Facilities", and Justin Gibbs's subsequent
    341  * integration into FreeBSD, modified for NetBSD by Jason R. Thorpe.
    342  *
    343  * The original work on the data structures used in this implementation
    344  * was published by G. Varghese and A. Lauck in the paper "Hashed and
    345  * Hierarchical Timing Wheels: Data Structures for the Efficient
    346  * Implementation of a Timer Facility" in the Proceedings of the 11th
    347  * ACM Annual Symposium on Operating System Principles, Austin, Texas,
    348  * November 1987.
    349  */
    350 struct callout_queue *callwheel;
    351 int	callwheelsize, callwheelbits, callwheelmask;
    352 
    353 static struct callout *nextsoftcheck;	/* next callout to be checked */
    354 
    355 #ifdef CALLWHEEL_STATS
    356 int	callwheel_collisions;		/* number of hash collisions */
    357 int	callwheel_maxlength;		/* length of the longest hash chain */
    358 int	*callwheel_sizes;		/* per-bucket length count */
    359 u_int64_t callwheel_count;		/* # callouts currently */
    360 u_int64_t callwheel_established;	/* # callouts established */
    361 u_int64_t callwheel_fired;		/* # callouts that fired */
    362 u_int64_t callwheel_disestablished;	/* # callouts disestablished */
    363 u_int64_t callwheel_changed;		/* # callouts changed */
    364 u_int64_t callwheel_softclocks;		/* # times softclock() called */
    365 u_int64_t callwheel_softchecks;		/* # checks per softclock() */
    366 u_int64_t callwheel_softempty;		/* # empty buckets seen */
    367 #endif /* CALLWHEEL_STATS */
    368 
    369 /*
    370  * This value indicates the number of consecutive callouts that
    371  * will be checked before we allow interrupts to have a chance
    372  * again.
    373  */
    374 #ifndef MAX_SOFTCLOCK_STEPS
    375 #define	MAX_SOFTCLOCK_STEPS	100
    376 #endif
    377 
    378 struct simplelock callwheel_slock;
    379 
    380 #define	CALLWHEEL_LOCK(s)						\
    381 do {									\
    382 	s = splclock();							\
    383 	simple_lock(&callwheel_slock);					\
    384 } while (0)
    385 
    386 #define	CALLWHEEL_UNLOCK(s)						\
    387 do {									\
    388 	simple_unlock(&callwheel_slock);				\
    389 	splx(s);							\
    390 } while (0)
    391 
    392 static void callout_stop_locked(struct callout *);
    393 
    394 /*
    395  * These are both protected by callwheel_lock.
    396  * XXX SHOULD BE STATIC!!
    397  */
    398 u_int64_t hardclock_ticks, softclock_ticks;
    399 
    400 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
    401 void	softclock(void *);
    402 void	*softclock_si;
    403 #endif
    404 
    405 /*
    406  * Initialize clock frequencies and start both clocks running.
    407  */
    408 void
    409 initclocks(void)
    410 {
    411 	int i;
    412 
    413 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
    414 	softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
    415 	if (softclock_si == NULL)
    416 		panic("initclocks: unable to register softclock intr");
    417 #endif
    418 
    419 	/*
    420 	 * Set divisors to 1 (normal case) and let the machine-specific
    421 	 * code do its bit.
    422 	 */
    423 	psdiv = 1;
    424 	cpu_initclocks();
    425 
    426 	/*
    427 	 * Compute profhz/stathz/rrticks, and fix profhz if needed.
    428 	 */
    429 	i = stathz ? stathz : hz;
    430 	if (profhz == 0)
    431 		profhz = i;
    432 	psratio = profhz / i;
    433 	rrticks = hz / 10;
    434 
    435 #ifdef NTP
    436 	switch (hz) {
    437 	case 1:
    438 		shifthz = SHIFT_SCALE - 0;
    439 		break;
    440 	case 2:
    441 		shifthz = SHIFT_SCALE - 1;
    442 		break;
    443 	case 4:
    444 		shifthz = SHIFT_SCALE - 2;
    445 		break;
    446 	case 8:
    447 		shifthz = SHIFT_SCALE - 3;
    448 		break;
    449 	case 16:
    450 		shifthz = SHIFT_SCALE - 4;
    451 		break;
    452 	case 32:
    453 		shifthz = SHIFT_SCALE - 5;
    454 		break;
    455 	case 60:
    456 	case 64:
    457 		shifthz = SHIFT_SCALE - 6;
    458 		break;
    459 	case 96:
    460 	case 100:
    461 	case 128:
    462 		shifthz = SHIFT_SCALE - 7;
    463 		break;
    464 	case 256:
    465 		shifthz = SHIFT_SCALE - 8;
    466 		break;
    467 	case 512:
    468 		shifthz = SHIFT_SCALE - 9;
    469 		break;
    470 	case 1000:
    471 	case 1024:
    472 		shifthz = SHIFT_SCALE - 10;
    473 		break;
    474 	case 1200:
    475 	case 2048:
    476 		shifthz = SHIFT_SCALE - 11;
    477 		break;
    478 	case 4096:
    479 		shifthz = SHIFT_SCALE - 12;
    480 		break;
    481 	case 8192:
    482 		shifthz = SHIFT_SCALE - 13;
    483 		break;
    484 	case 16384:
    485 		shifthz = SHIFT_SCALE - 14;
    486 		break;
    487 	case 32768:
    488 		shifthz = SHIFT_SCALE - 15;
    489 		break;
    490 	case 65536:
    491 		shifthz = SHIFT_SCALE - 16;
    492 		break;
    493 	default:
    494 		panic("weird hz");
    495 	}
    496 	if (fixtick == 0) {
    497 		/*
    498 		 * Give MD code a chance to set this to a better
    499 		 * value; but, if it doesn't, we should.
    500 		 */
    501 		fixtick = (1000000 - (hz*tick));
    502 	}
    503 #endif
    504 }
    505 
    506 /*
    507  * The real-time timer, interrupting hz times per second.
    508  */
    509 void
    510 hardclock(struct clockframe *frame)
    511 {
    512 	struct proc *p;
    513 	int delta;
    514 	extern int tickdelta;
    515 	extern long timedelta;
    516 	struct cpu_info *ci = curcpu();
    517 #ifdef NTP
    518 	int time_update;
    519 	int ltemp;
    520 #endif
    521 
    522 	p = curproc;
    523 	if (p) {
    524 		struct pstats *pstats;
    525 
    526 		/*
    527 		 * Run current process's virtual and profile time, as needed.
    528 		 */
    529 		pstats = p->p_stats;
    530 		if (CLKF_USERMODE(frame) &&
    531 		    timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
    532 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
    533 			psignal(p, SIGVTALRM);
    534 		if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
    535 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
    536 			psignal(p, SIGPROF);
    537 	}
    538 
    539 	/*
    540 	 * If no separate statistics clock is available, run it from here.
    541 	 */
    542 	if (stathz == 0)
    543 		statclock(frame);
    544 	if ((--ci->ci_schedstate.spc_rrticks) <= 0)
    545 		roundrobin(ci);
    546 
    547 #if defined(MULTIPROCESSOR)
    548 	/*
    549 	 * If we are not the primary CPU, we're not allowed to do
    550 	 * any more work.
    551 	 */
    552 	if (CPU_IS_PRIMARY(ci) == 0)
    553 		return;
    554 #endif
    555 
    556 	/*
    557 	 * Increment the time-of-day.  The increment is normally just
    558 	 * ``tick''.  If the machine is one which has a clock frequency
    559 	 * such that ``hz'' would not divide the second evenly into
    560 	 * milliseconds, a periodic adjustment must be applied.  Finally,
    561 	 * if we are still adjusting the time (see adjtime()),
    562 	 * ``tickdelta'' may also be added in.
    563 	 */
    564 	delta = tick;
    565 
    566 #ifndef NTP
    567 	if (tickfix) {
    568 		tickfixcnt += tickfix;
    569 		if (tickfixcnt >= tickfixinterval) {
    570 			delta++;
    571 			tickfixcnt -= tickfixinterval;
    572 		}
    573 	}
    574 #endif /* !NTP */
    575 	/* Imprecise 4bsd adjtime() handling */
    576 	if (timedelta != 0) {
    577 		delta += tickdelta;
    578 		timedelta -= tickdelta;
    579 	}
    580 
    581 #ifdef notyet
    582 	microset();
    583 #endif
    584 
    585 #ifndef NTP
    586 	BUMPTIME(&time, delta);		/* XXX Now done using NTP code below */
    587 #endif
    588 	BUMPTIME(&mono_time, delta);
    589 
    590 #ifdef NTP
    591 	time_update = delta;
    592 
    593 	/*
    594 	 * Compute the phase adjustment. If the low-order bits
    595 	 * (time_phase) of the update overflow, bump the high-order bits
    596 	 * (time_update).
    597 	 */
    598 	time_phase += time_adj;
    599 	if (time_phase <= -FINEUSEC) {
    600 		ltemp = -time_phase >> SHIFT_SCALE;
    601 		time_phase += ltemp << SHIFT_SCALE;
    602 		time_update -= ltemp;
    603 	} else if (time_phase >= FINEUSEC) {
    604 		ltemp = time_phase >> SHIFT_SCALE;
    605 		time_phase -= ltemp << SHIFT_SCALE;
    606 		time_update += ltemp;
    607 	}
    608 
    609 #ifdef HIGHBALL
    610 	/*
    611 	 * If the HIGHBALL board is installed, we need to adjust the
    612 	 * external clock offset in order to close the hardware feedback
    613 	 * loop. This will adjust the external clock phase and frequency
    614 	 * in small amounts. The additional phase noise and frequency
    615 	 * wander this causes should be minimal. We also need to
    616 	 * discipline the kernel time variable, since the PLL is used to
    617 	 * discipline the external clock. If the Highball board is not
    618 	 * present, we discipline kernel time with the PLL as usual. We
    619 	 * assume that the external clock phase adjustment (time_update)
    620 	 * and kernel phase adjustment (clock_cpu) are less than the
    621 	 * value of tick.
    622 	 */
    623 	clock_offset.tv_usec += time_update;
    624 	if (clock_offset.tv_usec >= 1000000) {
    625 		clock_offset.tv_sec++;
    626 		clock_offset.tv_usec -= 1000000;
    627 	}
    628 	if (clock_offset.tv_usec < 0) {
    629 		clock_offset.tv_sec--;
    630 		clock_offset.tv_usec += 1000000;
    631 	}
    632 	time.tv_usec += clock_cpu;
    633 	clock_cpu = 0;
    634 #else
    635 	time.tv_usec += time_update;
    636 #endif /* HIGHBALL */
    637 
    638 	/*
    639 	 * On rollover of the second the phase adjustment to be used for
    640 	 * the next second is calculated. Also, the maximum error is
    641 	 * increased by the tolerance. If the PPS frequency discipline
    642 	 * code is present, the phase is increased to compensate for the
    643 	 * CPU clock oscillator frequency error.
    644 	 *
    645  	 * On a 32-bit machine and given parameters in the timex.h
    646 	 * header file, the maximum phase adjustment is +-512 ms and
    647 	 * maximum frequency offset is a tad less than) +-512 ppm. On a
    648 	 * 64-bit machine, you shouldn't need to ask.
    649 	 */
    650 	if (time.tv_usec >= 1000000) {
    651 		time.tv_usec -= 1000000;
    652 		time.tv_sec++;
    653 		time_maxerror += time_tolerance >> SHIFT_USEC;
    654 
    655 		/*
    656 		 * Leap second processing. If in leap-insert state at
    657 		 * the end of the day, the system clock is set back one
    658 		 * second; if in leap-delete state, the system clock is
    659 		 * set ahead one second. The microtime() routine or
    660 		 * external clock driver will insure that reported time
    661 		 * is always monotonic. The ugly divides should be
    662 		 * replaced.
    663 		 */
    664 		switch (time_state) {
    665 		case TIME_OK:
    666 			if (time_status & STA_INS)
    667 				time_state = TIME_INS;
    668 			else if (time_status & STA_DEL)
    669 				time_state = TIME_DEL;
    670 			break;
    671 
    672 		case TIME_INS:
    673 			if (time.tv_sec % 86400 == 0) {
    674 				time.tv_sec--;
    675 				time_state = TIME_OOP;
    676 			}
    677 			break;
    678 
    679 		case TIME_DEL:
    680 			if ((time.tv_sec + 1) % 86400 == 0) {
    681 				time.tv_sec++;
    682 				time_state = TIME_WAIT;
    683 			}
    684 			break;
    685 
    686 		case TIME_OOP:
    687 			time_state = TIME_WAIT;
    688 			break;
    689 
    690 		case TIME_WAIT:
    691 			if (!(time_status & (STA_INS | STA_DEL)))
    692 				time_state = TIME_OK;
    693 			break;
    694 		}
    695 
    696 		/*
    697 		 * Compute the phase adjustment for the next second. In
    698 		 * PLL mode, the offset is reduced by a fixed factor
    699 		 * times the time constant. In FLL mode the offset is
    700 		 * used directly. In either mode, the maximum phase
    701 		 * adjustment for each second is clamped so as to spread
    702 		 * the adjustment over not more than the number of
    703 		 * seconds between updates.
    704 		 */
    705 		if (time_offset < 0) {
    706 			ltemp = -time_offset;
    707 			if (!(time_status & STA_FLL))
    708 				ltemp >>= SHIFT_KG + time_constant;
    709 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    710 				ltemp = (MAXPHASE / MINSEC) <<
    711 				    SHIFT_UPDATE;
    712 			time_offset += ltemp;
    713 			time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
    714 		} else if (time_offset > 0) {
    715 			ltemp = time_offset;
    716 			if (!(time_status & STA_FLL))
    717 				ltemp >>= SHIFT_KG + time_constant;
    718 			if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
    719 				ltemp = (MAXPHASE / MINSEC) <<
    720 				    SHIFT_UPDATE;
    721 			time_offset -= ltemp;
    722 			time_adj = ltemp << (shifthz - SHIFT_UPDATE);
    723 		} else
    724 			time_adj = 0;
    725 
    726 		/*
    727 		 * Compute the frequency estimate and additional phase
    728 		 * adjustment due to frequency error for the next
    729 		 * second. When the PPS signal is engaged, gnaw on the
    730 		 * watchdog counter and update the frequency computed by
    731 		 * the pll and the PPS signal.
    732 		 */
    733 #ifdef PPS_SYNC
    734 		pps_valid++;
    735 		if (pps_valid == PPS_VALID) {
    736 			pps_jitter = MAXTIME;
    737 			pps_stabil = MAXFREQ;
    738 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
    739 			    STA_PPSWANDER | STA_PPSERROR);
    740 		}
    741 		ltemp = time_freq + pps_freq;
    742 #else
    743 		ltemp = time_freq;
    744 #endif /* PPS_SYNC */
    745 
    746 		if (ltemp < 0)
    747 			time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
    748 		else
    749 			time_adj += ltemp >> (SHIFT_USEC - shifthz);
    750 		time_adj += (long)fixtick << shifthz;
    751 
    752 		/*
    753 		 * When the CPU clock oscillator frequency is not a
    754 		 * power of 2 in Hz, shifthz is only an approximate
    755 		 * scale factor.
    756 		 *
    757 		 * To determine the adjustment, you can do the following:
    758 		 *   bc -q
    759 		 *   scale=24
    760 		 *   obase=2
    761 		 *   idealhz/realhz
    762 		 * where `idealhz' is the next higher power of 2, and `realhz'
    763 		 * is the actual value.  You may need to factor this result
    764 		 * into a sequence of 2 multipliers to get better precision.
    765 		 *
    766 		 * Likewise, the error can be calculated with (e.g. for 100Hz):
    767 		 *   bc -q
    768 		 *   scale=24
    769 		 *   ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz
    770 		 * (and then multiply by 1000000 to get ppm).
    771 		 */
    772 		switch (hz) {
    773 		case 60:
    774 			/* A factor of 1.000100010001 gives about 15ppm
    775 			   error. */
    776 			if (time_adj < 0) {
    777 				time_adj -= (-time_adj >> 4);
    778 				time_adj -= (-time_adj >> 8);
    779 			} else {
    780 				time_adj += (time_adj >> 4);
    781 				time_adj += (time_adj >> 8);
    782 			}
    783 			break;
    784 
    785 		case 96:
    786 			/* A factor of 1.0101010101 gives about 244ppm error. */
    787 			if (time_adj < 0) {
    788 				time_adj -= (-time_adj >> 2);
    789 				time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
    790 			} else {
    791 				time_adj += (time_adj >> 2);
    792 				time_adj += (time_adj >> 4) + (time_adj >> 8);
    793 			}
    794 			break;
    795 
    796 		case 100:
    797 			/* A factor of 1.010001111010111 gives about 1ppm
    798 			   error. */
    799 			if (time_adj < 0) {
    800 				time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
    801 				time_adj += (-time_adj >> 10);
    802 			} else {
    803 				time_adj += (time_adj >> 2) + (time_adj >> 5);
    804 				time_adj -= (time_adj >> 10);
    805 			}
    806 			break;
    807 
    808 		case 1000:
    809 			/* A factor of 1.000001100010100001 gives about 50ppm
    810 			   error. */
    811 			if (time_adj < 0) {
    812 				time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
    813 				time_adj -= (-time_adj >> 7);
    814 			} else {
    815 				time_adj += (time_adj >> 6) + (time_adj >> 11);
    816 				time_adj += (time_adj >> 7);
    817 			}
    818 			break;
    819 
    820 		case 1200:
    821 			/* A factor of 1.1011010011100001 gives about 64ppm
    822 			   error. */
    823 			if (time_adj < 0) {
    824 				time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
    825 				time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
    826 			} else {
    827 				time_adj += (time_adj >> 1) + (time_adj >> 6);
    828 				time_adj += (time_adj >> 3) + (time_adj >> 10);
    829 			}
    830 			break;
    831 		}
    832 
    833 #ifdef EXT_CLOCK
    834 		/*
    835 		 * If an external clock is present, it is necessary to
    836 		 * discipline the kernel time variable anyway, since not
    837 		 * all system components use the microtime() interface.
    838 		 * Here, the time offset between the external clock and
    839 		 * kernel time variable is computed every so often.
    840 		 */
    841 		clock_count++;
    842 		if (clock_count > CLOCK_INTERVAL) {
    843 			clock_count = 0;
    844 			microtime(&clock_ext);
    845 			delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
    846 			delta.tv_usec = clock_ext.tv_usec -
    847 			    time.tv_usec;
    848 			if (delta.tv_usec < 0)
    849 				delta.tv_sec--;
    850 			if (delta.tv_usec >= 500000) {
    851 				delta.tv_usec -= 1000000;
    852 				delta.tv_sec++;
    853 			}
    854 			if (delta.tv_usec < -500000) {
    855 				delta.tv_usec += 1000000;
    856 				delta.tv_sec--;
    857 			}
    858 			if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
    859 			    delta.tv_usec > MAXPHASE) ||
    860 			    delta.tv_sec < -1 || (delta.tv_sec == -1 &&
    861 			    delta.tv_usec < -MAXPHASE)) {
    862 				time = clock_ext;
    863 				delta.tv_sec = 0;
    864 				delta.tv_usec = 0;
    865 			}
    866 #ifdef HIGHBALL
    867 			clock_cpu = delta.tv_usec;
    868 #else /* HIGHBALL */
    869 			hardupdate(delta.tv_usec);
    870 #endif /* HIGHBALL */
    871 		}
    872 #endif /* EXT_CLOCK */
    873 	}
    874 
    875 #endif /* NTP */
    876 
    877 	/*
    878 	 * Process callouts at a very low cpu priority, so we don't keep the
    879 	 * relatively high clock interrupt priority any longer than necessary.
    880 	 */
    881 	simple_lock(&callwheel_slock);	/* already at splclock() */
    882 	hardclock_ticks++;
    883 	if (TAILQ_FIRST(&callwheel[hardclock_ticks & callwheelmask]) != NULL) {
    884 		simple_unlock(&callwheel_slock);
    885 		if (CLKF_BASEPRI(frame)) {
    886 			/*
    887 			 * Save the overhead of a software interrupt;
    888 			 * it will happen as soon as we return, so do
    889 			 * it now.
    890 			 *
    891 			 * NOTE: If we're at ``base priority'', softclock()
    892 			 * was not already running.
    893 			 */
    894 			spllowersoftclock();
    895 			KERNEL_LOCK(LK_CANRECURSE|LK_EXCLUSIVE);
    896 			softclock(NULL);
    897 			KERNEL_UNLOCK();
    898 		} else {
    899 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS
    900 			softintr_schedule(softclock_si);
    901 #else
    902 			setsoftclock();
    903 #endif
    904 		}
    905 		return;
    906 	} else if (softclock_running == 0 &&
    907 		   (softclock_ticks + 1) == hardclock_ticks) {
    908 		softclock_ticks++;
    909 	}
    910 	simple_unlock(&callwheel_slock);
    911 }
    912 
    913 /*
    914  * Software (low priority) clock interrupt.
    915  * Run periodic events from timeout queue.
    916  */
    917 /*ARGSUSED*/
    918 void
    919 softclock(void *v)
    920 {
    921 	struct callout_queue *bucket;
    922 	struct callout *c;
    923 	void (*func)(void *);
    924 	void *arg;
    925 	int s, idx;
    926 	int steps = 0;
    927 
    928 	CALLWHEEL_LOCK(s);
    929 
    930 	softclock_running = 1;
    931 
    932 #ifdef CALLWHEEL_STATS
    933 	callwheel_softclocks++;
    934 #endif
    935 
    936 	while (softclock_ticks != hardclock_ticks) {
    937 		softclock_ticks++;
    938 		idx = (int)(softclock_ticks & callwheelmask);
    939 		bucket = &callwheel[idx];
    940 		c = TAILQ_FIRST(bucket);
    941 #ifdef CALLWHEEL_STATS
    942 		if (c == NULL)
    943 			callwheel_softempty++;
    944 #endif
    945 		while (c != NULL) {
    946 #ifdef CALLWHEEL_STATS
    947 			callwheel_softchecks++;
    948 #endif
    949 			if (c->c_time != softclock_ticks) {
    950 				c = TAILQ_NEXT(c, c_link);
    951 				if (++steps >= MAX_SOFTCLOCK_STEPS) {
    952 					nextsoftcheck = c;
    953 					/* Give interrupts a chance. */
    954 					CALLWHEEL_UNLOCK(s);
    955 					CALLWHEEL_LOCK(s);
    956 					c = nextsoftcheck;
    957 					steps = 0;
    958 				}
    959 			} else {
    960 				nextsoftcheck = TAILQ_NEXT(c, c_link);
    961 				TAILQ_REMOVE(bucket, c, c_link);
    962 #ifdef CALLWHEEL_STATS
    963 				callwheel_sizes[idx]--;
    964 				callwheel_fired++;
    965 				callwheel_count--;
    966 #endif
    967 				func = c->c_func;
    968 				arg = c->c_arg;
    969 				c->c_func = NULL;
    970 				c->c_flags &= ~CALLOUT_PENDING;
    971 				CALLWHEEL_UNLOCK(s);
    972 				(*func)(arg);
    973 				CALLWHEEL_LOCK(s);
    974 				steps = 0;
    975 				c = nextsoftcheck;
    976 			}
    977 		}
    978 	}
    979 	nextsoftcheck = NULL;
    980 	softclock_running = 0;
    981 	CALLWHEEL_UNLOCK(s);
    982 }
    983 
    984 /*
    985  * callout_setsize:
    986  *
    987  *	Determine how many callwheels are necessary and
    988  *	set hash mask.  Called from allocsys().
    989  */
    990 void
    991 callout_setsize(void)
    992 {
    993 
    994 	for (callwheelsize = 1; callwheelsize < ncallout; callwheelsize <<= 1)
    995 		/* loop */ ;
    996 	callwheelmask = callwheelsize - 1;
    997 }
    998 
    999 /*
   1000  * callout_startup:
   1001  *
   1002  *	Initialize the callwheel buckets.
   1003  */
   1004 void
   1005 callout_startup(void)
   1006 {
   1007 	int i;
   1008 
   1009 	for (i = 0; i < callwheelsize; i++)
   1010 		TAILQ_INIT(&callwheel[i]);
   1011 
   1012 	simple_lock_init(&callwheel_slock);
   1013 }
   1014 
   1015 /*
   1016  * callout_init:
   1017  *
   1018  *	Initialize a callout structure so that it can be used
   1019  *	by callout_reset() and callout_stop().
   1020  */
   1021 void
   1022 callout_init(struct callout *c)
   1023 {
   1024 
   1025 	memset(c, 0, sizeof(*c));
   1026 }
   1027 
   1028 /*
   1029  * callout_reset:
   1030  *
   1031  *	Establish or change a timeout.
   1032  */
   1033 void
   1034 callout_reset(struct callout *c, int ticks, void (*func)(void *), void *arg)
   1035 {
   1036 	struct callout_queue *bucket;
   1037 	int s;
   1038 
   1039 	if (ticks <= 0)
   1040 		ticks = 1;
   1041 
   1042 	CALLWHEEL_LOCK(s);
   1043 
   1044 	/*
   1045 	 * If this callout's timer is already running, cancel it
   1046 	 * before we modify it.
   1047 	 */
   1048 	if (c->c_flags & CALLOUT_PENDING) {
   1049 		callout_stop_locked(c);	/* Already locked */
   1050 #ifdef CALLWHEEL_STATS
   1051 		callwheel_changed++;
   1052 #endif
   1053 	}
   1054 
   1055 	c->c_arg = arg;
   1056 	c->c_func = func;
   1057 	c->c_flags = CALLOUT_ACTIVE | CALLOUT_PENDING;
   1058 	c->c_time = hardclock_ticks + ticks;
   1059 
   1060 	bucket = &callwheel[c->c_time & callwheelmask];
   1061 
   1062 #ifdef CALLWHEEL_STATS
   1063 	if (TAILQ_FIRST(bucket) != NULL)
   1064 		callwheel_collisions++;
   1065 #endif
   1066 
   1067 	TAILQ_INSERT_TAIL(bucket, c, c_link);
   1068 
   1069 #ifdef CALLWHEEL_STATS
   1070 	callwheel_count++;
   1071 	callwheel_established++;
   1072 	if (++callwheel_sizes[c->c_time & callwheelmask] > callwheel_maxlength)
   1073 		callwheel_maxlength =
   1074 		    callwheel_sizes[c->c_time & callwheelmask];
   1075 #endif
   1076 
   1077 	CALLWHEEL_UNLOCK(s);
   1078 }
   1079 
   1080 /*
   1081  * callout_stop_locked:
   1082  *
   1083  *	Disestablish a timeout.  Callwheel is locked.
   1084  */
   1085 static void
   1086 callout_stop_locked(struct callout *c)
   1087 {
   1088 
   1089 	/*
   1090 	 * Don't attempt to delete a callout that's not on the queue.
   1091 	 */
   1092 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
   1093 		c->c_flags &= ~CALLOUT_ACTIVE;
   1094 		return;
   1095 	}
   1096 
   1097 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
   1098 
   1099 	if (nextsoftcheck == c)
   1100 		nextsoftcheck = TAILQ_NEXT(c, c_link);
   1101 
   1102 	TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_link);
   1103 #ifdef CALLWHEEL_STATS
   1104 	callwheel_count--;
   1105 	callwheel_disestablished++;
   1106 	callwheel_sizes[c->c_time & callwheelmask]--;
   1107 #endif
   1108 
   1109 	c->c_func = NULL;
   1110 }
   1111 
   1112 /*
   1113  * callout_stop:
   1114  *
   1115  *	Disestablish a timeout.  Callwheel is unlocked.  This is
   1116  *	the standard entry point.
   1117  */
   1118 void
   1119 callout_stop(struct callout *c)
   1120 {
   1121 	int s;
   1122 
   1123 	CALLWHEEL_LOCK(s);
   1124 	callout_stop_locked(c);
   1125 	CALLWHEEL_UNLOCK(s);
   1126 }
   1127 
   1128 #ifdef CALLWHEEL_STATS
   1129 /*
   1130  * callout_showstats:
   1131  *
   1132  *	Display callout statistics.  Call it from DDB.
   1133  */
   1134 void
   1135 callout_showstats(void)
   1136 {
   1137 	u_int64_t curticks;
   1138 	int s;
   1139 
   1140 	s = splclock();
   1141 	curticks = softclock_ticks;
   1142 	splx(s);
   1143 
   1144 	printf("Callwheel statistics:\n");
   1145 	printf("\tCallouts currently queued: %llu\n", callwheel_count);
   1146 	printf("\tCallouts established: %llu\n", callwheel_established);
   1147 	printf("\tCallouts disestablished: %llu\n", callwheel_disestablished);
   1148 	if (callwheel_changed != 0)
   1149 		printf("\t\tOf those, %llu were changes\n", callwheel_changed);
   1150 	printf("\tCallouts that fired: %llu\n", callwheel_fired);
   1151 	printf("\tNumber of buckets: %d\n", callwheelsize);
   1152 	printf("\tNumber of hash collisions: %d\n", callwheel_collisions);
   1153 	printf("\tMaximum hash chain length: %d\n", callwheel_maxlength);
   1154 	printf("\tSoftclocks: %llu, Softchecks: %llu\n",
   1155 	    callwheel_softclocks, callwheel_softchecks);
   1156 	printf("\t\tEmpty buckets seen: %llu\n", callwheel_softempty);
   1157 }
   1158 #endif
   1159 
   1160 /*
   1161  * Compute number of hz until specified time.  Used to compute second
   1162  * argument to callout_reset() from an absolute time.
   1163  */
   1164 int
   1165 hzto(struct timeval *tv)
   1166 {
   1167 	unsigned long ticks;
   1168 	long sec, usec;
   1169 	int s;
   1170 
   1171 	/*
   1172 	 * If the number of usecs in the whole seconds part of the time
   1173 	 * difference fits in a long, then the total number of usecs will
   1174 	 * fit in an unsigned long.  Compute the total and convert it to
   1175 	 * ticks, rounding up and adding 1 to allow for the current tick
   1176 	 * to expire.  Rounding also depends on unsigned long arithmetic
   1177 	 * to avoid overflow.
   1178 	 *
   1179 	 * Otherwise, if the number of ticks in the whole seconds part of
   1180 	 * the time difference fits in a long, then convert the parts to
   1181 	 * ticks separately and add, using similar rounding methods and
   1182 	 * overflow avoidance.  This method would work in the previous
   1183 	 * case, but it is slightly slower and assume that hz is integral.
   1184 	 *
   1185 	 * Otherwise, round the time difference down to the maximum
   1186 	 * representable value.
   1187 	 *
   1188 	 * If ints are 32-bit, then the maximum value for any timeout in
   1189 	 * 10ms ticks is 248 days.
   1190 	 */
   1191 	s = splclock();
   1192 	sec = tv->tv_sec - time.tv_sec;
   1193 	usec = tv->tv_usec - time.tv_usec;
   1194 	splx(s);
   1195 
   1196 	if (usec < 0) {
   1197 		sec--;
   1198 		usec += 1000000;
   1199 	}
   1200 
   1201 	if (sec < 0 || (sec == 0 && usec <= 0)) {
   1202 		/*
   1203 		 * Would expire now or in the past.  Return 0 ticks.
   1204 		 * This is different from the legacy hzto() interface,
   1205 		 * and callers need to check for it.
   1206 		 */
   1207 		ticks = 0;
   1208 	} else if (sec <= (LONG_MAX / 1000000))
   1209 		ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1))
   1210 		    / tick) + 1;
   1211 	else if (sec <= (LONG_MAX / hz))
   1212 		ticks = (sec * hz) +
   1213 		    (((unsigned long)usec + (tick - 1)) / tick) + 1;
   1214 	else
   1215 		ticks = LONG_MAX;
   1216 
   1217 	if (ticks > INT_MAX)
   1218 		ticks = INT_MAX;
   1219 
   1220 	return ((int)ticks);
   1221 }
   1222 
   1223 /*
   1224  * Start profiling on a process.
   1225  *
   1226  * Kernel profiling passes proc0 which never exits and hence
   1227  * keeps the profile clock running constantly.
   1228  */
   1229 void
   1230 startprofclock(struct proc *p)
   1231 {
   1232 
   1233 	if ((p->p_flag & P_PROFIL) == 0) {
   1234 		p->p_flag |= P_PROFIL;
   1235 		if (++profprocs == 1 && stathz != 0)
   1236 			psdiv = psratio;
   1237 	}
   1238 }
   1239 
   1240 /*
   1241  * Stop profiling on a process.
   1242  */
   1243 void
   1244 stopprofclock(struct proc *p)
   1245 {
   1246 
   1247 	if (p->p_flag & P_PROFIL) {
   1248 		p->p_flag &= ~P_PROFIL;
   1249 		if (--profprocs == 0 && stathz != 0)
   1250 			psdiv = 1;
   1251 	}
   1252 }
   1253 
   1254 /*
   1255  * Statistics clock.  Grab profile sample, and if divider reaches 0,
   1256  * do process and kernel statistics.
   1257  */
   1258 void
   1259 statclock(struct clockframe *frame)
   1260 {
   1261 #ifdef GPROF
   1262 	struct gmonparam *g;
   1263 	intptr_t i;
   1264 #endif
   1265 	struct cpu_info *ci = curcpu();
   1266 	struct schedstate_percpu *spc = &ci->ci_schedstate;
   1267 	struct proc *p;
   1268 
   1269 	/*
   1270 	 * Notice changes in divisor frequency, and adjust clock
   1271 	 * frequency accordingly.
   1272 	 */
   1273 	if (spc->spc_psdiv != psdiv) {
   1274 		spc->spc_psdiv = psdiv;
   1275 		spc->spc_pscnt = psdiv;
   1276 		if (psdiv == 1) {
   1277 			setstatclockrate(stathz);
   1278 		} else {
   1279 			setstatclockrate(profhz);
   1280 		}
   1281 	}
   1282 	p = curproc;
   1283 	if (CLKF_USERMODE(frame)) {
   1284 		if (p->p_flag & P_PROFIL)
   1285 			addupc_intr(p, CLKF_PC(frame));
   1286 		if (--spc->spc_pscnt > 0)
   1287 			return;
   1288 		/*
   1289 		 * Came from user mode; CPU was in user state.
   1290 		 * If this process is being profiled record the tick.
   1291 		 */
   1292 		p->p_uticks++;
   1293 		if (p->p_nice > NZERO)
   1294 			spc->spc_cp_time[CP_NICE]++;
   1295 		else
   1296 			spc->spc_cp_time[CP_USER]++;
   1297 	} else {
   1298 #ifdef GPROF
   1299 		/*
   1300 		 * Kernel statistics are just like addupc_intr, only easier.
   1301 		 */
   1302 		g = &_gmonparam;
   1303 		if (g->state == GMON_PROF_ON) {
   1304 			i = CLKF_PC(frame) - g->lowpc;
   1305 			if (i < g->textsize) {
   1306 				i /= HISTFRACTION * sizeof(*g->kcount);
   1307 				g->kcount[i]++;
   1308 			}
   1309 		}
   1310 #endif
   1311 #ifdef PROC_PC
   1312 		if (p && p->p_flag & P_PROFIL)
   1313 			addupc_intr(p, PROC_PC(p));
   1314 #endif
   1315 		if (--spc->spc_pscnt > 0)
   1316 			return;
   1317 		/*
   1318 		 * Came from kernel mode, so we were:
   1319 		 * - handling an interrupt,
   1320 		 * - doing syscall or trap work on behalf of the current
   1321 		 *   user process, or
   1322 		 * - spinning in the idle loop.
   1323 		 * Whichever it is, charge the time as appropriate.
   1324 		 * Note that we charge interrupts to the current process,
   1325 		 * regardless of whether they are ``for'' that process,
   1326 		 * so that we know how much of its real time was spent
   1327 		 * in ``non-process'' (i.e., interrupt) work.
   1328 		 */
   1329 		if (CLKF_INTR(frame)) {
   1330 			if (p != NULL)
   1331 				p->p_iticks++;
   1332 			spc->spc_cp_time[CP_INTR]++;
   1333 		} else if (p != NULL) {
   1334 			p->p_sticks++;
   1335 			spc->spc_cp_time[CP_SYS]++;
   1336 		} else
   1337 			spc->spc_cp_time[CP_IDLE]++;
   1338 	}
   1339 	spc->spc_pscnt = psdiv;
   1340 
   1341 	if (p != NULL) {
   1342 		++p->p_cpticks;
   1343 		/*
   1344 		 * If no separate schedclock is provided, call it here
   1345 		 * at ~~12-25 Hz, ~~16 Hz is best
   1346 		 */
   1347 		if (schedhz == 0)
   1348 			if ((++ci->ci_schedstate.spc_schedticks & 3) == 0)
   1349 				schedclock(p);
   1350 	}
   1351 }
   1352 
   1353 
   1354 #ifdef NTP	/* NTP phase-locked loop in kernel */
   1355 
   1356 /*
   1357  * hardupdate() - local clock update
   1358  *
   1359  * This routine is called by ntp_adjtime() to update the local clock
   1360  * phase and frequency. The implementation is of an adaptive-parameter,
   1361  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
   1362  * time and frequency offset estimates for each call. If the kernel PPS
   1363  * discipline code is configured (PPS_SYNC), the PPS signal itself
   1364  * determines the new time offset, instead of the calling argument.
   1365  * Presumably, calls to ntp_adjtime() occur only when the caller
   1366  * believes the local clock is valid within some bound (+-128 ms with
   1367  * NTP). If the caller's time is far different than the PPS time, an
   1368  * argument will ensue, and it's not clear who will lose.
   1369  *
   1370  * For uncompensated quartz crystal oscillatores and nominal update
   1371  * intervals less than 1024 s, operation should be in phase-lock mode
   1372  * (STA_FLL = 0), where the loop is disciplined to phase. For update
   1373  * intervals greater than thiss, operation should be in frequency-lock
   1374  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
   1375  *
   1376  * Note: splclock() is in effect.
   1377  */
   1378 void
   1379 hardupdate(long offset)
   1380 {
   1381 	long ltemp, mtemp;
   1382 
   1383 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
   1384 		return;
   1385 	ltemp = offset;
   1386 #ifdef PPS_SYNC
   1387 	if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
   1388 		ltemp = pps_offset;
   1389 #endif /* PPS_SYNC */
   1390 
   1391 	/*
   1392 	 * Scale the phase adjustment and clamp to the operating range.
   1393 	 */
   1394 	if (ltemp > MAXPHASE)
   1395 		time_offset = MAXPHASE << SHIFT_UPDATE;
   1396 	else if (ltemp < -MAXPHASE)
   1397 		time_offset = -(MAXPHASE << SHIFT_UPDATE);
   1398 	else
   1399 		time_offset = ltemp << SHIFT_UPDATE;
   1400 
   1401 	/*
   1402 	 * Select whether the frequency is to be controlled and in which
   1403 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
   1404 	 * multiply/divide should be replaced someday.
   1405 	 */
   1406 	if (time_status & STA_FREQHOLD || time_reftime == 0)
   1407 		time_reftime = time.tv_sec;
   1408 	mtemp = time.tv_sec - time_reftime;
   1409 	time_reftime = time.tv_sec;
   1410 	if (time_status & STA_FLL) {
   1411 		if (mtemp >= MINSEC) {
   1412 			ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
   1413 			    SHIFT_UPDATE));
   1414 			if (ltemp < 0)
   1415 				time_freq -= -ltemp >> SHIFT_KH;
   1416 			else
   1417 				time_freq += ltemp >> SHIFT_KH;
   1418 		}
   1419 	} else {
   1420 		if (mtemp < MAXSEC) {
   1421 			ltemp *= mtemp;
   1422 			if (ltemp < 0)
   1423 				time_freq -= -ltemp >> (time_constant +
   1424 				    time_constant + SHIFT_KF -
   1425 				    SHIFT_USEC);
   1426 			else
   1427 				time_freq += ltemp >> (time_constant +
   1428 				    time_constant + SHIFT_KF -
   1429 				    SHIFT_USEC);
   1430 		}
   1431 	}
   1432 	if (time_freq > time_tolerance)
   1433 		time_freq = time_tolerance;
   1434 	else if (time_freq < -time_tolerance)
   1435 		time_freq = -time_tolerance;
   1436 }
   1437 
   1438 #ifdef PPS_SYNC
   1439 /*
   1440  * hardpps() - discipline CPU clock oscillator to external PPS signal
   1441  *
   1442  * This routine is called at each PPS interrupt in order to discipline
   1443  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
   1444  * and leaves it in a handy spot for the hardclock() routine. It
   1445  * integrates successive PPS phase differences and calculates the
   1446  * frequency offset. This is used in hardclock() to discipline the CPU
   1447  * clock oscillator so that intrinsic frequency error is cancelled out.
   1448  * The code requires the caller to capture the time and hardware counter
   1449  * value at the on-time PPS signal transition.
   1450  *
   1451  * Note that, on some Unix systems, this routine runs at an interrupt
   1452  * priority level higher than the timer interrupt routine hardclock().
   1453  * Therefore, the variables used are distinct from the hardclock()
   1454  * variables, except for certain exceptions: The PPS frequency pps_freq
   1455  * and phase pps_offset variables are determined by this routine and
   1456  * updated atomically. The time_tolerance variable can be considered a
   1457  * constant, since it is infrequently changed, and then only when the
   1458  * PPS signal is disabled. The watchdog counter pps_valid is updated
   1459  * once per second by hardclock() and is atomically cleared in this
   1460  * routine.
   1461  */
   1462 void
   1463 hardpps(struct timeval *tvp,		/* time at PPS */
   1464 	long usec			/* hardware counter at PPS */)
   1465 {
   1466 	long u_usec, v_usec, bigtick;
   1467 	long cal_sec, cal_usec;
   1468 
   1469 	/*
   1470 	 * An occasional glitch can be produced when the PPS interrupt
   1471 	 * occurs in the hardclock() routine before the time variable is
   1472 	 * updated. Here the offset is discarded when the difference
   1473 	 * between it and the last one is greater than tick/2, but not
   1474 	 * if the interval since the first discard exceeds 30 s.
   1475 	 */
   1476 	time_status |= STA_PPSSIGNAL;
   1477 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
   1478 	pps_valid = 0;
   1479 	u_usec = -tvp->tv_usec;
   1480 	if (u_usec < -500000)
   1481 		u_usec += 1000000;
   1482 	v_usec = pps_offset - u_usec;
   1483 	if (v_usec < 0)
   1484 		v_usec = -v_usec;
   1485 	if (v_usec > (tick >> 1)) {
   1486 		if (pps_glitch > MAXGLITCH) {
   1487 			pps_glitch = 0;
   1488 			pps_tf[2] = u_usec;
   1489 			pps_tf[1] = u_usec;
   1490 		} else {
   1491 			pps_glitch++;
   1492 			u_usec = pps_offset;
   1493 		}
   1494 	} else
   1495 		pps_glitch = 0;
   1496 
   1497 	/*
   1498 	 * A three-stage median filter is used to help deglitch the pps
   1499 	 * time. The median sample becomes the time offset estimate; the
   1500 	 * difference between the other two samples becomes the time
   1501 	 * dispersion (jitter) estimate.
   1502 	 */
   1503 	pps_tf[2] = pps_tf[1];
   1504 	pps_tf[1] = pps_tf[0];
   1505 	pps_tf[0] = u_usec;
   1506 	if (pps_tf[0] > pps_tf[1]) {
   1507 		if (pps_tf[1] > pps_tf[2]) {
   1508 			pps_offset = pps_tf[1];		/* 0 1 2 */
   1509 			v_usec = pps_tf[0] - pps_tf[2];
   1510 		} else if (pps_tf[2] > pps_tf[0]) {
   1511 			pps_offset = pps_tf[0];		/* 2 0 1 */
   1512 			v_usec = pps_tf[2] - pps_tf[1];
   1513 		} else {
   1514 			pps_offset = pps_tf[2];		/* 0 2 1 */
   1515 			v_usec = pps_tf[0] - pps_tf[1];
   1516 		}
   1517 	} else {
   1518 		if (pps_tf[1] < pps_tf[2]) {
   1519 			pps_offset = pps_tf[1];		/* 2 1 0 */
   1520 			v_usec = pps_tf[2] - pps_tf[0];
   1521 		} else  if (pps_tf[2] < pps_tf[0]) {
   1522 			pps_offset = pps_tf[0];		/* 1 0 2 */
   1523 			v_usec = pps_tf[1] - pps_tf[2];
   1524 		} else {
   1525 			pps_offset = pps_tf[2];		/* 1 2 0 */
   1526 			v_usec = pps_tf[1] - pps_tf[0];
   1527 		}
   1528 	}
   1529 	if (v_usec > MAXTIME)
   1530 		pps_jitcnt++;
   1531 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
   1532 	if (v_usec < 0)
   1533 		pps_jitter -= -v_usec >> PPS_AVG;
   1534 	else
   1535 		pps_jitter += v_usec >> PPS_AVG;
   1536 	if (pps_jitter > (MAXTIME >> 1))
   1537 		time_status |= STA_PPSJITTER;
   1538 
   1539 	/*
   1540 	 * During the calibration interval adjust the starting time when
   1541 	 * the tick overflows. At the end of the interval compute the
   1542 	 * duration of the interval and the difference of the hardware
   1543 	 * counters at the beginning and end of the interval. This code
   1544 	 * is deliciously complicated by the fact valid differences may
   1545 	 * exceed the value of tick when using long calibration
   1546 	 * intervals and small ticks. Note that the counter can be
   1547 	 * greater than tick if caught at just the wrong instant, but
   1548 	 * the values returned and used here are correct.
   1549 	 */
   1550 	bigtick = (long)tick << SHIFT_USEC;
   1551 	pps_usec -= pps_freq;
   1552 	if (pps_usec >= bigtick)
   1553 		pps_usec -= bigtick;
   1554 	if (pps_usec < 0)
   1555 		pps_usec += bigtick;
   1556 	pps_time.tv_sec++;
   1557 	pps_count++;
   1558 	if (pps_count < (1 << pps_shift))
   1559 		return;
   1560 	pps_count = 0;
   1561 	pps_calcnt++;
   1562 	u_usec = usec << SHIFT_USEC;
   1563 	v_usec = pps_usec - u_usec;
   1564 	if (v_usec >= bigtick >> 1)
   1565 		v_usec -= bigtick;
   1566 	if (v_usec < -(bigtick >> 1))
   1567 		v_usec += bigtick;
   1568 	if (v_usec < 0)
   1569 		v_usec = -(-v_usec >> pps_shift);
   1570 	else
   1571 		v_usec = v_usec >> pps_shift;
   1572 	pps_usec = u_usec;
   1573 	cal_sec = tvp->tv_sec;
   1574 	cal_usec = tvp->tv_usec;
   1575 	cal_sec -= pps_time.tv_sec;
   1576 	cal_usec -= pps_time.tv_usec;
   1577 	if (cal_usec < 0) {
   1578 		cal_usec += 1000000;
   1579 		cal_sec--;
   1580 	}
   1581 	pps_time = *tvp;
   1582 
   1583 	/*
   1584 	 * Check for lost interrupts, noise, excessive jitter and
   1585 	 * excessive frequency error. The number of timer ticks during
   1586 	 * the interval may vary +-1 tick. Add to this a margin of one
   1587 	 * tick for the PPS signal jitter and maximum frequency
   1588 	 * deviation. If the limits are exceeded, the calibration
   1589 	 * interval is reset to the minimum and we start over.
   1590 	 */
   1591 	u_usec = (long)tick << 1;
   1592 	if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
   1593 	    || (cal_sec == 0 && cal_usec < u_usec))
   1594 	    || v_usec > time_tolerance || v_usec < -time_tolerance) {
   1595 		pps_errcnt++;
   1596 		pps_shift = PPS_SHIFT;
   1597 		pps_intcnt = 0;
   1598 		time_status |= STA_PPSERROR;
   1599 		return;
   1600 	}
   1601 
   1602 	/*
   1603 	 * A three-stage median filter is used to help deglitch the pps
   1604 	 * frequency. The median sample becomes the frequency offset
   1605 	 * estimate; the difference between the other two samples
   1606 	 * becomes the frequency dispersion (stability) estimate.
   1607 	 */
   1608 	pps_ff[2] = pps_ff[1];
   1609 	pps_ff[1] = pps_ff[0];
   1610 	pps_ff[0] = v_usec;
   1611 	if (pps_ff[0] > pps_ff[1]) {
   1612 		if (pps_ff[1] > pps_ff[2]) {
   1613 			u_usec = pps_ff[1];		/* 0 1 2 */
   1614 			v_usec = pps_ff[0] - pps_ff[2];
   1615 		} else if (pps_ff[2] > pps_ff[0]) {
   1616 			u_usec = pps_ff[0];		/* 2 0 1 */
   1617 			v_usec = pps_ff[2] - pps_ff[1];
   1618 		} else {
   1619 			u_usec = pps_ff[2];		/* 0 2 1 */
   1620 			v_usec = pps_ff[0] - pps_ff[1];
   1621 		}
   1622 	} else {
   1623 		if (pps_ff[1] < pps_ff[2]) {
   1624 			u_usec = pps_ff[1];		/* 2 1 0 */
   1625 			v_usec = pps_ff[2] - pps_ff[0];
   1626 		} else  if (pps_ff[2] < pps_ff[0]) {
   1627 			u_usec = pps_ff[0];		/* 1 0 2 */
   1628 			v_usec = pps_ff[1] - pps_ff[2];
   1629 		} else {
   1630 			u_usec = pps_ff[2];		/* 1 2 0 */
   1631 			v_usec = pps_ff[1] - pps_ff[0];
   1632 		}
   1633 	}
   1634 
   1635 	/*
   1636 	 * Here the frequency dispersion (stability) is updated. If it
   1637 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
   1638 	 * offset is updated as well, but clamped to the tolerance. It
   1639 	 * will be processed later by the hardclock() routine.
   1640 	 */
   1641 	v_usec = (v_usec >> 1) - pps_stabil;
   1642 	if (v_usec < 0)
   1643 		pps_stabil -= -v_usec >> PPS_AVG;
   1644 	else
   1645 		pps_stabil += v_usec >> PPS_AVG;
   1646 	if (pps_stabil > MAXFREQ >> 2) {
   1647 		pps_stbcnt++;
   1648 		time_status |= STA_PPSWANDER;
   1649 		return;
   1650 	}
   1651 	if (time_status & STA_PPSFREQ) {
   1652 		if (u_usec < 0) {
   1653 			pps_freq -= -u_usec >> PPS_AVG;
   1654 			if (pps_freq < -time_tolerance)
   1655 				pps_freq = -time_tolerance;
   1656 			u_usec = -u_usec;
   1657 		} else {
   1658 			pps_freq += u_usec >> PPS_AVG;
   1659 			if (pps_freq > time_tolerance)
   1660 				pps_freq = time_tolerance;
   1661 		}
   1662 	}
   1663 
   1664 	/*
   1665 	 * Here the calibration interval is adjusted. If the maximum
   1666 	 * time difference is greater than tick / 4, reduce the interval
   1667 	 * by half. If this is not the case for four consecutive
   1668 	 * intervals, double the interval.
   1669 	 */
   1670 	if (u_usec << pps_shift > bigtick >> 2) {
   1671 		pps_intcnt = 0;
   1672 		if (pps_shift > PPS_SHIFT)
   1673 			pps_shift--;
   1674 	} else if (pps_intcnt >= 4) {
   1675 		pps_intcnt = 0;
   1676 		if (pps_shift < PPS_SHIFTMAX)
   1677 			pps_shift++;
   1678 	} else
   1679 		pps_intcnt++;
   1680 }
   1681 #endif /* PPS_SYNC */
   1682 #endif /* NTP  */
   1683 
   1684 /*
   1685  * Return information about system clocks.
   1686  */
   1687 int
   1688 sysctl_clockrate(void *where, size_t *sizep)
   1689 {
   1690 	struct clockinfo clkinfo;
   1691 
   1692 	/*
   1693 	 * Construct clockinfo structure.
   1694 	 */
   1695 	clkinfo.tick = tick;
   1696 	clkinfo.tickadj = tickadj;
   1697 	clkinfo.hz = hz;
   1698 	clkinfo.profhz = profhz;
   1699 	clkinfo.stathz = stathz ? stathz : hz;
   1700 	return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
   1701 }
   1702