xref: /dragonfly/sys/kern/kern_ntptime.c (revision 36a3d1d6)
1 /***********************************************************************
2  *								       *
3  * Copyright (c) David L. Mills 1993-2001			       *
4  *								       *
5  * Permission to use, copy, modify, and distribute this software and   *
6  * its documentation for any purpose and without fee is hereby	       *
7  * granted, provided that the above copyright notice appears in all    *
8  * copies and that both the copyright notice and this permission       *
9  * notice appear in supporting documentation, and that the name	       *
10  * University of Delaware not be used in advertising or publicity      *
11  * pertaining to distribution of the software without specific,	       *
12  * written prior permission. The University of Delaware makes no       *
13  * representations about the suitability this software for any	       *
14  * purpose. It is provided "as is" without express or implied	       *
15  * warranty.							       *
16  *								       *
17  **********************************************************************/
18 
19 /*
20  * Adapted from the original sources for FreeBSD and timecounters by:
21  * Poul-Henning Kamp <phk@FreeBSD.org>.
22  *
23  * The 32bit version of the "LP" macros seems a bit past its "sell by"
24  * date so I have retained only the 64bit version and included it directly
25  * in this file.
26  *
27  * Only minor changes done to interface with the timecounters over in
28  * sys/kern/kern_clock.c.   Some of the comments below may be (even more)
29  * confusing and/or plain wrong in that context.
30  *
31  * $FreeBSD: src/sys/kern/kern_ntptime.c,v 1.32.2.2 2001/04/22 11:19:46 jhay Exp $
32  * $DragonFly: src/sys/kern/kern_ntptime.c,v 1.13 2007/04/30 07:18:53 dillon Exp $
33  */
34 
35 #include "opt_ntp.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/sysproto.h>
40 #include <sys/kernel.h>
41 #include <sys/proc.h>
42 #include <sys/priv.h>
43 #include <sys/time.h>
44 #include <sys/timex.h>
45 #include <sys/timepps.h>
46 #include <sys/sysctl.h>
47 
48 #include <sys/thread2.h>
49 #include <sys/mplock2.h>
50 
51 /*
52  * Single-precision macros for 64-bit machines
53  */
54 typedef long long l_fp;
55 #define L_ADD(v, u)	((v) += (u))
56 #define L_SUB(v, u)	((v) -= (u))
57 #define L_ADDHI(v, a)	((v) += (long long)(a) << 32)
58 #define L_NEG(v)	((v) = -(v))
59 #define L_RSHIFT(v, n) \
60 	do { \
61 		if ((v) < 0) \
62 			(v) = -(-(v) >> (n)); \
63 		else \
64 			(v) = (v) >> (n); \
65 	} while (0)
66 #define L_MPY(v, a)	((v) *= (a))
67 #define L_CLR(v)	((v) = 0)
68 #define L_ISNEG(v)	((v) < 0)
69 #define L_LINT(v, a)	((v) = (long long)(a) << 32)
70 #define L_GINT(v)	((v) < 0 ? -(-(v) >> 32) : (v) >> 32)
71 
72 /*
73  * Generic NTP kernel interface
74  *
75  * These routines constitute the Network Time Protocol (NTP) interfaces
76  * for user and daemon application programs. The ntp_gettime() routine
77  * provides the time, maximum error (synch distance) and estimated error
78  * (dispersion) to client user application programs. The ntp_adjtime()
79  * routine is used by the NTP daemon to adjust the system clock to an
80  * externally derived time. The time offset and related variables set by
81  * this routine are used by other routines in this module to adjust the
82  * phase and frequency of the clock discipline loop which controls the
83  * system clock.
84  *
85  * When the kernel time is reckoned directly in nanoseconds (NTP_NANO
86  * defined), the time at each tick interrupt is derived directly from
87  * the kernel time variable. When the kernel time is reckoned in
88  * microseconds, (NTP_NANO undefined), the time is derived from the
89  * kernel time variable together with a variable representing the
90  * leftover nanoseconds at the last tick interrupt. In either case, the
91  * current nanosecond time is reckoned from these values plus an
92  * interpolated value derived by the clock routines in another
93  * architecture-specific module. The interpolation can use either a
94  * dedicated counter or a processor cycle counter (PCC) implemented in
95  * some architectures.
96  *
97  * Note that all routines must run at priority splclock or higher.
98  */
99 /*
100  * Phase/frequency-lock loop (PLL/FLL) definitions
101  *
102  * The nanosecond clock discipline uses two variable types, time
103  * variables and frequency variables. Both types are represented as 64-
104  * bit fixed-point quantities with the decimal point between two 32-bit
105  * halves. On a 32-bit machine, each half is represented as a single
106  * word and mathematical operations are done using multiple-precision
107  * arithmetic. On a 64-bit machine, ordinary computer arithmetic is
108  * used.
109  *
110  * A time variable is a signed 64-bit fixed-point number in ns and
111  * fraction. It represents the remaining time offset to be amortized
112  * over succeeding tick interrupts. The maximum time offset is about
113  * 0.5 s and the resolution is about 2.3e-10 ns.
114  *
115  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
116  *  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
117  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
118  * |s s s|			 ns				   |
119  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
120  * |			    fraction				   |
121  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
122  *
123  * A frequency variable is a signed 64-bit fixed-point number in ns/s
124  * and fraction. It represents the ns and fraction to be added to the
125  * kernel time variable at each second. The maximum frequency offset is
126  * about +-500000 ns/s and the resolution is about 2.3e-10 ns/s.
127  *
128  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
129  *  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
130  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
131  * |s s s s s s s s s s s s s|	          ns/s			   |
132  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
133  * |			    fraction				   |
134  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
135  */
136 /*
137  * The following variables establish the state of the PLL/FLL and the
138  * residual time and frequency offset of the local clock.
139  */
140 #define SHIFT_PLL	4		/* PLL loop gain (shift) */
141 #define SHIFT_FLL	2		/* FLL loop gain (shift) */
142 
143 static int time_state = TIME_OK;	/* clock state */
144 static int time_status = STA_UNSYNC;	/* clock status bits */
145 static long time_tai;			/* TAI offset (s) */
146 static long time_monitor;		/* last time offset scaled (ns) */
147 static long time_constant;		/* poll interval (shift) (s) */
148 static long time_precision = 1;		/* clock precision (ns) */
149 static long time_maxerror = MAXPHASE / 1000; /* maximum error (us) */
150 static long time_esterror = MAXPHASE / 1000; /* estimated error (us) */
151 static long time_reftime;		/* time at last adjustment (s) */
152 static long time_tick;			/* nanoseconds per tick (ns) */
153 static l_fp time_offset;		/* time offset (ns) */
154 static l_fp time_freq;			/* frequency offset (ns/s) */
155 static l_fp time_adj;			/* tick adjust (ns/s) */
156 
157 #ifdef PPS_SYNC
158 /*
159  * The following variables are used when a pulse-per-second (PPS) signal
160  * is available and connected via a modem control lead. They establish
161  * the engineering parameters of the clock discipline loop when
162  * controlled by the PPS signal.
163  */
164 #define PPS_FAVG	2		/* min freq avg interval (s) (shift) */
165 #define PPS_FAVGDEF	8		/* default freq avg int (s) (shift) */
166 #define PPS_FAVGMAX	15		/* max freq avg interval (s) (shift) */
167 #define PPS_PAVG	4		/* phase avg interval (s) (shift) */
168 #define PPS_VALID	120		/* PPS signal watchdog max (s) */
169 #define PPS_MAXWANDER	100000		/* max PPS wander (ns/s) */
170 #define PPS_POPCORN	2		/* popcorn spike threshold (shift) */
171 
172 static struct timespec pps_tf[3];	/* phase median filter */
173 static l_fp pps_freq;			/* scaled frequency offset (ns/s) */
174 static long pps_fcount;			/* frequency accumulator */
175 static long pps_jitter;			/* nominal jitter (ns) */
176 static long pps_stabil;			/* nominal stability (scaled ns/s) */
177 static long pps_lastsec;		/* time at last calibration (s) */
178 static int pps_valid;			/* signal watchdog counter */
179 static int pps_shift = PPS_FAVG;	/* interval duration (s) (shift) */
180 static int pps_shiftmax = PPS_FAVGDEF;	/* max interval duration (s) (shift) */
181 static int pps_intcnt;			/* wander counter */
182 
183 /*
184  * PPS signal quality monitors
185  */
186 static long pps_calcnt;			/* calibration intervals */
187 static long pps_jitcnt;			/* jitter limit exceeded */
188 static long pps_stbcnt;			/* stability limit exceeded */
189 static long pps_errcnt;			/* calibration errors */
190 #endif /* PPS_SYNC */
191 /*
192  * End of phase/frequency-lock loop (PLL/FLL) definitions
193  */
194 
195 static void ntp_init(void);
196 static void hardupdate(long offset);
197 
198 /*
199  * ntp_gettime() - NTP user application interface
200  *
201  * See the timex.h header file for synopsis and API description. Note
202  * that the TAI offset is returned in the ntvtimeval.tai structure
203  * member.
204  */
205 static int
206 ntp_sysctl(SYSCTL_HANDLER_ARGS)
207 {
208 	struct ntptimeval ntv;	/* temporary structure */
209 	struct timespec atv;	/* nanosecond time */
210 
211 	nanotime(&atv);
212 	ntv.time.tv_sec = atv.tv_sec;
213 	ntv.time.tv_nsec = atv.tv_nsec;
214 	ntv.maxerror = time_maxerror;
215 	ntv.esterror = time_esterror;
216 	ntv.tai = time_tai;
217 	ntv.time_state = time_state;
218 
219 	/*
220 	 * Status word error decode. If any of these conditions occur,
221 	 * an error is returned, instead of the status word. Most
222 	 * applications will care only about the fact the system clock
223 	 * may not be trusted, not about the details.
224 	 *
225 	 * Hardware or software error
226 	 */
227 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
228 
229 	/*
230 	 * PPS signal lost when either time or frequency synchronization
231 	 * requested
232 	 */
233 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
234 	    !(time_status & STA_PPSSIGNAL)) ||
235 
236 	/*
237 	 * PPS jitter exceeded when time synchronization requested
238 	 */
239 	    (time_status & STA_PPSTIME &&
240 	    time_status & STA_PPSJITTER) ||
241 
242 	/*
243 	 * PPS wander exceeded or calibration error when frequency
244 	 * synchronization requested
245 	 */
246 	    (time_status & STA_PPSFREQ &&
247 	    time_status & (STA_PPSWANDER | STA_PPSERROR)))
248 		ntv.time_state = TIME_ERROR;
249 	return (sysctl_handle_opaque(oidp, &ntv, sizeof ntv, req));
250 }
251 
252 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW, 0, "");
253 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE|CTLFLAG_RD,
254 	0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval", "");
255 
256 #ifdef PPS_SYNC
257 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW, &pps_shiftmax, 0, "");
258 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW, &pps_shift, 0, "");
259 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD, &time_monitor, 0, "");
260 
261 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD, &pps_freq, sizeof(pps_freq), "I", "");
262 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD, &time_freq, sizeof(time_freq), "I", "");
263 #endif
264 /*
265  * ntp_adjtime() - NTP daemon application interface
266  *
267  * See the timex.h header file for synopsis and API description. Note
268  * that the timex.constant structure member has a dual purpose to set
269  * the time constant and to set the TAI offset.
270  *
271  * MPALMOSTSAFE
272  */
273 int
274 sys_ntp_adjtime(struct ntp_adjtime_args *uap)
275 {
276 	struct thread *td = curthread;
277 	struct timex ntv;	/* temporary structure */
278 	long freq;		/* frequency ns/s) */
279 	int modes;		/* mode bits from structure */
280 	int error;
281 
282 	error = copyin((caddr_t)uap->tp, (caddr_t)&ntv, sizeof(ntv));
283 	if (error)
284 		return(error);
285 
286 	/*
287 	 * Update selected clock variables - only the superuser can
288 	 * change anything. Note that there is no error checking here on
289 	 * the assumption the superuser should know what it is doing.
290 	 * Note that either the time constant or TAI offset are loaded
291 	 * from the ntv.constant member, depending on the mode bits. If
292 	 * the STA_PLL bit in the status word is cleared, the state and
293 	 * status words are reset to the initial values at boot.
294 	 */
295 	modes = ntv.modes;
296 	if (modes)
297 		error = priv_check(td, PRIV_NTP_ADJTIME);
298 	if (error)
299 		return (error);
300 
301 	get_mplock();
302 	crit_enter();
303 	if (modes & MOD_MAXERROR)
304 		time_maxerror = ntv.maxerror;
305 	if (modes & MOD_ESTERROR)
306 		time_esterror = ntv.esterror;
307 	if (modes & MOD_STATUS) {
308 		if (time_status & STA_PLL && !(ntv.status & STA_PLL)) {
309 			time_state = TIME_OK;
310 			time_status = STA_UNSYNC;
311 #ifdef PPS_SYNC
312 			pps_shift = PPS_FAVG;
313 #endif /* PPS_SYNC */
314 		}
315 		time_status &= STA_RONLY;
316 		time_status |= ntv.status & ~STA_RONLY;
317 	}
318 	if (modes & MOD_TIMECONST) {
319 		if (ntv.constant < 0)
320 			time_constant = 0;
321 		else if (ntv.constant > MAXTC)
322 			time_constant = MAXTC;
323 		else
324 			time_constant = ntv.constant;
325 	}
326 	if (modes & MOD_TAI) {
327 		if (ntv.constant > 0) /* XXX zero & negative numbers ? */
328 			time_tai = ntv.constant;
329 	}
330 #ifdef PPS_SYNC
331 	if (modes & MOD_PPSMAX) {
332 		if (ntv.shift < PPS_FAVG)
333 			pps_shiftmax = PPS_FAVG;
334 		else if (ntv.shift > PPS_FAVGMAX)
335 			pps_shiftmax = PPS_FAVGMAX;
336 		else
337 			pps_shiftmax = ntv.shift;
338 	}
339 #endif /* PPS_SYNC */
340 	if (modes & MOD_NANO)
341 		time_status |= STA_NANO;
342 	if (modes & MOD_MICRO)
343 		time_status &= ~STA_NANO;
344 	if (modes & MOD_CLKB)
345 		time_status |= STA_CLK;
346 	if (modes & MOD_CLKA)
347 		time_status &= ~STA_CLK;
348 	if (modes & MOD_OFFSET) {
349 		if (time_status & STA_NANO)
350 			hardupdate(ntv.offset);
351 		else
352 			hardupdate(ntv.offset * 1000);
353 	}
354 	/*
355 	 * Note: the userland specified frequency is in seconds per second
356 	 * times 65536e+6.  Multiply by a thousand and divide by 65336 to
357 	 * get nanoseconds.
358 	 */
359 	if (modes & MOD_FREQUENCY) {
360 		freq = (ntv.freq * 1000LL) >> 16;
361 		if (freq > MAXFREQ)
362 			L_LINT(time_freq, MAXFREQ);
363 		else if (freq < -MAXFREQ)
364 			L_LINT(time_freq, -MAXFREQ);
365 		else
366 			L_LINT(time_freq, freq);
367 #ifdef PPS_SYNC
368 		pps_freq = time_freq;
369 #endif /* PPS_SYNC */
370 	}
371 
372 	/*
373 	 * Retrieve all clock variables. Note that the TAI offset is
374 	 * returned only by ntp_gettime();
375 	 */
376 	if (time_status & STA_NANO)
377 		ntv.offset = time_monitor;
378 	else
379 		ntv.offset = time_monitor / 1000; /* XXX rounding ? */
380 	ntv.freq = L_GINT((time_freq / 1000LL) << 16);
381 	ntv.maxerror = time_maxerror;
382 	ntv.esterror = time_esterror;
383 	ntv.status = time_status;
384 	ntv.constant = time_constant;
385 	if (time_status & STA_NANO)
386 		ntv.precision = time_precision;
387 	else
388 		ntv.precision = time_precision / 1000;
389 	ntv.tolerance = MAXFREQ * SCALE_PPM;
390 #ifdef PPS_SYNC
391 	ntv.shift = pps_shift;
392 	ntv.ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
393 	if (time_status & STA_NANO)
394 		ntv.jitter = pps_jitter;
395 	else
396 		ntv.jitter = pps_jitter / 1000;
397 	ntv.stabil = pps_stabil;
398 	ntv.calcnt = pps_calcnt;
399 	ntv.errcnt = pps_errcnt;
400 	ntv.jitcnt = pps_jitcnt;
401 	ntv.stbcnt = pps_stbcnt;
402 #endif /* PPS_SYNC */
403 	crit_exit();
404 	rel_mplock();
405 
406 	error = copyout((caddr_t)&ntv, (caddr_t)uap->tp, sizeof(ntv));
407 	if (error)
408 		return (error);
409 
410 	/*
411 	 * Status word error decode. See comments in
412 	 * ntp_gettime() routine.
413 	 */
414 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
415 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
416 	    !(time_status & STA_PPSSIGNAL)) ||
417 	    (time_status & STA_PPSTIME &&
418 	    time_status & STA_PPSJITTER) ||
419 	    (time_status & STA_PPSFREQ &&
420 	    time_status & (STA_PPSWANDER | STA_PPSERROR))) {
421 		uap->sysmsg_result = TIME_ERROR;
422 	} else {
423 		uap->sysmsg_result = time_state;
424 	}
425 	return (error);
426 }
427 
428 /*
429  * second_overflow() - called after ntp_tick_adjust()
430  *
431  * This routine is ordinarily called from hardclock() whenever the seconds
432  * hand rolls over.  It returns leap seconds to add or drop, and sets nsec_adj
433  * to the total adjustment to make over the next second in (ns << 32).
434  *
435  * This routine is only called by cpu #0.
436  */
437 int
438 ntp_update_second(time_t newsec, int64_t *nsec_adj)
439 {
440 	l_fp ftemp;		/* 32/64-bit temporary */
441 	int  adjsec = 0;
442 
443 	/*
444 	 * On rollover of the second both the nanosecond and microsecond
445 	 * clocks are updated and the state machine cranked as
446 	 * necessary. The phase adjustment to be used for the next
447 	 * second is calculated and the maximum error is increased by
448 	 * the tolerance.
449 	 */
450 	time_maxerror += MAXFREQ / 1000;
451 
452 	/*
453 	 * Leap second processing. If in leap-insert state at
454 	 * the end of the day, the system clock is set back one
455 	 * second; if in leap-delete state, the system clock is
456 	 * set ahead one second. The nano_time() routine or
457 	 * external clock driver will insure that reported time
458 	 * is always monotonic.
459 	 */
460 	switch (time_state) {
461 
462 		/*
463 		 * No warning.
464 		 */
465 		case TIME_OK:
466 		if (time_status & STA_INS)
467 			time_state = TIME_INS;
468 		else if (time_status & STA_DEL)
469 			time_state = TIME_DEL;
470 		break;
471 
472 		/*
473 		 * Insert second 23:59:60 following second
474 		 * 23:59:59.
475 		 */
476 		case TIME_INS:
477 		if (!(time_status & STA_INS))
478 			time_state = TIME_OK;
479 		else if ((newsec) % 86400 == 0) {
480 			--adjsec;
481 			time_state = TIME_OOP;
482 		}
483 		break;
484 
485 		/*
486 		 * Delete second 23:59:59.
487 		 */
488 		case TIME_DEL:
489 		if (!(time_status & STA_DEL))
490 			time_state = TIME_OK;
491 		else if (((newsec) + 1) % 86400 == 0) {
492 			++adjsec;
493 			time_tai--;
494 			time_state = TIME_WAIT;
495 		}
496 		break;
497 
498 		/*
499 		 * Insert second in progress.
500 		 */
501 		case TIME_OOP:
502 			time_tai++;
503 			time_state = TIME_WAIT;
504 		break;
505 
506 		/*
507 		 * Wait for status bits to clear.
508 		 */
509 		case TIME_WAIT:
510 		if (!(time_status & (STA_INS | STA_DEL)))
511 			time_state = TIME_OK;
512 	}
513 
514 	/*
515 	 * time_offset represents the total time adjustment we wish to
516 	 * make (over no particular period of time).  time_freq represents
517 	 * the frequency compensation we wish to apply.
518 	 *
519 	 * time_adj represents the total adjustment we wish to make over
520 	 * one full second.  hardclock usually applies this adjustment in
521 	 * time_adj / hz jumps, hz times a second.
522 	 */
523 	ftemp = time_offset;
524 #ifdef PPS_SYNC
525 	/* XXX even if PPS signal dies we should finish adjustment ? */
526 	if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
527 		L_RSHIFT(ftemp, pps_shift);
528 	else
529 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
530 #else
531 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
532 #endif /* PPS_SYNC */
533 	time_adj = ftemp;		/* adjustment for part of the offset */
534 	L_SUB(time_offset, ftemp);
535 	L_ADD(time_adj, time_freq);	/* add frequency correction */
536 	*nsec_adj = time_adj;
537 #ifdef PPS_SYNC
538 	if (pps_valid > 0)
539 		pps_valid--;
540 	else
541 		time_status &= ~STA_PPSSIGNAL;
542 #endif /* PPS_SYNC */
543 	return(adjsec);
544 }
545 
546 /*
547  * ntp_init() - initialize variables and structures
548  *
549  * This routine must be called after the kernel variables hz and tick
550  * are set or changed and before the next tick interrupt. In this
551  * particular implementation, these values are assumed set elsewhere in
552  * the kernel. The design allows the clock frequency and tick interval
553  * to be changed while the system is running. So, this routine should
554  * probably be integrated with the code that does that.
555  */
556 static void
557 ntp_init(void)
558 {
559 
560 	/*
561 	 * The following variable must be initialized any time the
562 	 * kernel variable hz is changed.
563 	 */
564 	time_tick = NANOSECOND / hz;
565 
566 	/*
567 	 * The following variables are initialized only at startup. Only
568 	 * those structures not cleared by the compiler need to be
569 	 * initialized, and these only in the simulator. In the actual
570 	 * kernel, any nonzero values here will quickly evaporate.
571 	 */
572 	L_CLR(time_offset);
573 	L_CLR(time_freq);
574 #ifdef PPS_SYNC
575 	pps_tf[0].tv_sec = pps_tf[0].tv_nsec = 0;
576 	pps_tf[1].tv_sec = pps_tf[1].tv_nsec = 0;
577 	pps_tf[2].tv_sec = pps_tf[2].tv_nsec = 0;
578 	pps_fcount = 0;
579 	L_CLR(pps_freq);
580 #endif /* PPS_SYNC */
581 }
582 
583 SYSINIT(ntpclocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, ntp_init, NULL)
584 
585 /*
586  * hardupdate() - local clock update
587  *
588  * This routine is called by ntp_adjtime() to update the local clock
589  * phase and frequency. The implementation is of an adaptive-parameter,
590  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
591  * time and frequency offset estimates for each call. If the kernel PPS
592  * discipline code is configured (PPS_SYNC), the PPS signal itself
593  * determines the new time offset, instead of the calling argument.
594  * Presumably, calls to ntp_adjtime() occur only when the caller
595  * believes the local clock is valid within some bound (+-128 ms with
596  * NTP). If the caller's time is far different than the PPS time, an
597  * argument will ensue, and it's not clear who will lose.
598  *
599  * For uncompensated quartz crystal oscillators and nominal update
600  * intervals less than 256 s, operation should be in phase-lock mode,
601  * where the loop is disciplined to phase. For update intervals greater
602  * than 1024 s, operation should be in frequency-lock mode, where the
603  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
604  * is selected by the STA_MODE status bit.
605  */
606 static void
607 hardupdate(long offset)
608 {
609 	long mtemp;
610 	l_fp ftemp;
611 	globaldata_t gd;
612 
613 	gd = mycpu;
614 
615 	/*
616 	 * Select how the phase is to be controlled and from which
617 	 * source. If the PPS signal is present and enabled to
618 	 * discipline the time, the PPS offset is used; otherwise, the
619 	 * argument offset is used.
620 	 */
621 	if (!(time_status & STA_PLL))
622 		return;
623 	if (!((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))) {
624 		if (offset > MAXPHASE)
625 			time_monitor = MAXPHASE;
626 		else if (offset < -MAXPHASE)
627 			time_monitor = -MAXPHASE;
628 		else
629 			time_monitor = offset;
630 		L_LINT(time_offset, time_monitor);
631 	}
632 
633 	/*
634 	 * Select how the frequency is to be controlled and in which
635 	 * mode (PLL or FLL). If the PPS signal is present and enabled
636 	 * to discipline the frequency, the PPS frequency is used;
637 	 * otherwise, the argument offset is used to compute it.
638 	 *
639 	 * gd_time_seconds is basically an uncompensated uptime.  We use
640 	 * this for consistency.
641 	 */
642 	if (time_status & STA_PPSFREQ && time_status & STA_PPSSIGNAL) {
643 		time_reftime = time_second;
644 		return;
645 	}
646 	if (time_status & STA_FREQHOLD || time_reftime == 0)
647 		time_reftime = time_second;
648 	mtemp = time_second - time_reftime;
649 	L_LINT(ftemp, time_monitor);
650 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
651 	L_MPY(ftemp, mtemp);
652 	L_ADD(time_freq, ftemp);
653 	time_status &= ~STA_MODE;
654 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp > MAXSEC)) {
655 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
656 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
657 		L_ADD(time_freq, ftemp);
658 		time_status |= STA_MODE;
659 	}
660 	time_reftime = time_second;
661 	if (L_GINT(time_freq) > MAXFREQ)
662 		L_LINT(time_freq, MAXFREQ);
663 	else if (L_GINT(time_freq) < -MAXFREQ)
664 		L_LINT(time_freq, -MAXFREQ);
665 }
666 
667 #ifdef PPS_SYNC
668 /*
669  * hardpps() - discipline CPU clock oscillator to external PPS signal
670  *
671  * This routine is called at each PPS interrupt in order to discipline
672  * the CPU clock oscillator to the PPS signal. There are two independent
673  * first-order feedback loops, one for the phase, the other for the
674  * frequency. The phase loop measures and grooms the PPS phase offset
675  * and leaves it in a handy spot for the seconds overflow routine. The
676  * frequency loop averages successive PPS phase differences and
677  * calculates the PPS frequency offset, which is also processed by the
678  * seconds overflow routine. The code requires the caller to capture the
679  * time and architecture-dependent hardware counter values in
680  * nanoseconds at the on-time PPS signal transition.
681  *
682  * Note that, on some Unix systems this routine runs at an interrupt
683  * priority level higher than the timer interrupt routine hardclock().
684  * Therefore, the variables used are distinct from the hardclock()
685  * variables, except for the actual time and frequency variables, which
686  * are determined by this routine and updated atomically.
687  */
688 void
689 hardpps(struct timespec *tsp, long nsec)
690 {
691 	long u_sec, u_nsec, v_nsec; /* temps */
692 	l_fp ftemp;
693 
694 	/*
695 	 * The signal is first processed by a range gate and frequency
696 	 * discriminator. The range gate rejects noise spikes outside
697 	 * the range +-500 us. The frequency discriminator rejects input
698 	 * signals with apparent frequency outside the range 1 +-500
699 	 * PPM. If two hits occur in the same second, we ignore the
700 	 * later hit; if not and a hit occurs outside the range gate,
701 	 * keep the later hit for later comparison, but do not process
702 	 * it.
703 	 */
704 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
705 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
706 	pps_valid = PPS_VALID;
707 	u_sec = tsp->tv_sec;
708 	u_nsec = tsp->tv_nsec;
709 	if (u_nsec >= (NANOSECOND >> 1)) {
710 		u_nsec -= NANOSECOND;
711 		u_sec++;
712 	}
713 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
714 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND -
715 	    MAXFREQ)
716 		return;
717 	pps_tf[2] = pps_tf[1];
718 	pps_tf[1] = pps_tf[0];
719 	pps_tf[0].tv_sec = u_sec;
720 	pps_tf[0].tv_nsec = u_nsec;
721 
722 	/*
723 	 * Compute the difference between the current and previous
724 	 * counter values. If the difference exceeds 0.5 s, assume it
725 	 * has wrapped around, so correct 1.0 s. If the result exceeds
726 	 * the tick interval, the sample point has crossed a tick
727 	 * boundary during the last second, so correct the tick. Very
728 	 * intricate.
729 	 */
730 	u_nsec = nsec;
731 	if (u_nsec > (NANOSECOND >> 1))
732 		u_nsec -= NANOSECOND;
733 	else if (u_nsec < -(NANOSECOND >> 1))
734 		u_nsec += NANOSECOND;
735 	pps_fcount += u_nsec;
736 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
737 		return;
738 	time_status &= ~STA_PPSJITTER;
739 
740 	/*
741 	 * A three-stage median filter is used to help denoise the PPS
742 	 * time. The median sample becomes the time offset estimate; the
743 	 * difference between the other two samples becomes the time
744 	 * dispersion (jitter) estimate.
745 	 */
746 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
747 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
748 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
749 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
750 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
751 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
752 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
753 		} else {
754 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
755 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
756 		}
757 	} else {
758 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
759 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
760 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
761 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
762 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
763 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
764 		} else {
765 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
766 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
767 		}
768 	}
769 
770 	/*
771 	 * Nominal jitter is due to PPS signal noise and interrupt
772 	 * latency. If it exceeds the popcorn threshold, the sample is
773 	 * discarded. otherwise, if so enabled, the time offset is
774 	 * updated. We can tolerate a modest loss of data here without
775 	 * much degrading time accuracy.
776 	 */
777 	if (u_nsec > (pps_jitter << PPS_POPCORN)) {
778 		time_status |= STA_PPSJITTER;
779 		pps_jitcnt++;
780 	} else if (time_status & STA_PPSTIME) {
781 		time_monitor = -v_nsec;
782 		L_LINT(time_offset, time_monitor);
783 	}
784 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
785 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
786 	if (u_sec < (1 << pps_shift))
787 		return;
788 
789 	/*
790 	 * At the end of the calibration interval the difference between
791 	 * the first and last counter values becomes the scaled
792 	 * frequency. It will later be divided by the length of the
793 	 * interval to determine the frequency update. If the frequency
794 	 * exceeds a sanity threshold, or if the actual calibration
795 	 * interval is not equal to the expected length, the data are
796 	 * discarded. We can tolerate a modest loss of data here without
797 	 * much degrading frequency accuracy.
798 	 */
799 	pps_calcnt++;
800 	v_nsec = -pps_fcount;
801 	pps_lastsec = pps_tf[0].tv_sec;
802 	pps_fcount = 0;
803 	u_nsec = MAXFREQ << pps_shift;
804 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 <<
805 	    pps_shift)) {
806 		time_status |= STA_PPSERROR;
807 		pps_errcnt++;
808 		return;
809 	}
810 
811 	/*
812 	 * Here the raw frequency offset and wander (stability) is
813 	 * calculated. If the wander is less than the wander threshold
814 	 * for four consecutive averaging intervals, the interval is
815 	 * doubled; if it is greater than the threshold for four
816 	 * consecutive intervals, the interval is halved. The scaled
817 	 * frequency offset is converted to frequency offset. The
818 	 * stability metric is calculated as the average of recent
819 	 * frequency changes, but is used only for performance
820 	 * monitoring.
821 	 */
822 	L_LINT(ftemp, v_nsec);
823 	L_RSHIFT(ftemp, pps_shift);
824 	L_SUB(ftemp, pps_freq);
825 	u_nsec = L_GINT(ftemp);
826 	if (u_nsec > PPS_MAXWANDER) {
827 		L_LINT(ftemp, PPS_MAXWANDER);
828 		pps_intcnt--;
829 		time_status |= STA_PPSWANDER;
830 		pps_stbcnt++;
831 	} else if (u_nsec < -PPS_MAXWANDER) {
832 		L_LINT(ftemp, -PPS_MAXWANDER);
833 		pps_intcnt--;
834 		time_status |= STA_PPSWANDER;
835 		pps_stbcnt++;
836 	} else {
837 		pps_intcnt++;
838 	}
839 	if (pps_intcnt >= 4) {
840 		pps_intcnt = 4;
841 		if (pps_shift < pps_shiftmax) {
842 			pps_shift++;
843 			pps_intcnt = 0;
844 		}
845 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
846 		pps_intcnt = -4;
847 		if (pps_shift > PPS_FAVG) {
848 			pps_shift--;
849 			pps_intcnt = 0;
850 		}
851 	}
852 	if (u_nsec < 0)
853 		u_nsec = -u_nsec;
854 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
855 
856 	/*
857 	 * The PPS frequency is recalculated and clamped to the maximum
858 	 * MAXFREQ. If enabled, the system clock frequency is updated as
859 	 * well.
860 	 */
861 	L_ADD(pps_freq, ftemp);
862 	u_nsec = L_GINT(pps_freq);
863 	if (u_nsec > MAXFREQ)
864 		L_LINT(pps_freq, MAXFREQ);
865 	else if (u_nsec < -MAXFREQ)
866 		L_LINT(pps_freq, -MAXFREQ);
867 	if (time_status & STA_PPSFREQ)
868 		time_freq = pps_freq;
869 }
870 #endif /* PPS_SYNC */
871