xref: /dragonfly/sys/kern/kern_ntptime.c (revision 2cf4b273)
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  */
33 
34 #include "opt_ntp.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/sysmsg.h>
39 #include <sys/kernel.h>
40 #include <sys/proc.h>
41 #include <sys/caps.h>
42 #include <sys/time.h>
43 #include <sys/timex.h>
44 #include <sys/timepps.h>
45 #include <sys/sysctl.h>
46 
47 #include <sys/thread2.h>
48 
49 /*
50  * Single-precision macros for 64-bit machines
51  */
52 typedef long long l_fp;
53 #define L_ADD(v, u)	((v) += (u))
54 #define L_SUB(v, u)	((v) -= (u))
55 #define L_ADDHI(v, a)	((v) += (long long)(a) << 32)
56 #define L_NEG(v)	((v) = -(v))
57 #define L_RSHIFT(v, n) \
58 	do { \
59 		if ((v) < 0) \
60 			(v) = -(-(v) >> (n)); \
61 		else \
62 			(v) = (v) >> (n); \
63 	} while (0)
64 #define L_MPY(v, a)	((v) *= (a))
65 #define L_CLR(v)	((v) = 0)
66 #define L_ISNEG(v)	((v) < 0)
67 #define L_LINT(v, a)	((v) = (long long)(a) << 32)
68 #define L_GINT(v)	((v) < 0 ? -(-(v) >> 32) : (v) >> 32)
69 
70 /*
71  * Generic NTP kernel interface
72  *
73  * These routines constitute the Network Time Protocol (NTP) interfaces
74  * for user and daemon application programs. The ntp_gettime() routine
75  * provides the time, maximum error (synch distance) and estimated error
76  * (dispersion) to client user application programs. The ntp_adjtime()
77  * routine is used by the NTP daemon to adjust the system clock to an
78  * externally derived time. The time offset and related variables set by
79  * this routine are used by other routines in this module to adjust the
80  * phase and frequency of the clock discipline loop which controls the
81  * system clock.
82  *
83  * When the kernel time is reckoned directly in nanoseconds (NTP_NANO
84  * defined), the time at each tick interrupt is derived directly from
85  * the kernel time variable. When the kernel time is reckoned in
86  * microseconds, (NTP_NANO undefined), the time is derived from the
87  * kernel time variable together with a variable representing the
88  * leftover nanoseconds at the last tick interrupt. In either case, the
89  * current nanosecond time is reckoned from these values plus an
90  * interpolated value derived by the clock routines in another
91  * architecture-specific module. The interpolation can use either a
92  * dedicated counter or a processor cycle counter (PCC) implemented in
93  * some architectures.
94  *
95  * Note that all routines must run at priority splclock or higher.
96  */
97 /*
98  * Phase/frequency-lock loop (PLL/FLL) definitions
99  *
100  * The nanosecond clock discipline uses two variable types, time
101  * variables and frequency variables. Both types are represented as 64-
102  * bit fixed-point quantities with the decimal point between two 32-bit
103  * halves. On a 32-bit machine, each half is represented as a single
104  * word and mathematical operations are done using multiple-precision
105  * arithmetic. On a 64-bit machine, ordinary computer arithmetic is
106  * used.
107  *
108  * A time variable is a signed 64-bit fixed-point number in ns and
109  * fraction. It represents the remaining time offset to be amortized
110  * over succeeding tick interrupts. The maximum time offset is about
111  * 0.5 s and the resolution is about 2.3e-10 ns.
112  *
113  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
114  *  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
115  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
116  * |s s s|			 ns				   |
117  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
118  * |			    fraction				   |
119  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
120  *
121  * A frequency variable is a signed 64-bit fixed-point number in ns/s
122  * and fraction. It represents the ns and fraction to be added to the
123  * kernel time variable at each second. The maximum frequency offset is
124  * about +-500000 ns/s and the resolution is about 2.3e-10 ns/s.
125  *
126  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
127  *  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
128  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
129  * |s s s s s s s s s s s s s|	          ns/s			   |
130  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
131  * |			    fraction				   |
132  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
133  */
134 /*
135  * The following variables establish the state of the PLL/FLL and the
136  * residual time and frequency offset of the local clock.
137  */
138 #define SHIFT_PLL	4		/* PLL loop gain (shift) */
139 #define SHIFT_FLL	2		/* FLL loop gain (shift) */
140 
141 static int time_state = TIME_OK;	/* clock state */
142 static int time_status = STA_UNSYNC;	/* clock status bits */
143 static long time_tai;			/* TAI offset (s) */
144 static long time_monitor;		/* last time offset scaled (ns) */
145 static long time_constant;		/* poll interval (shift) (s) */
146 static long time_precision = 1;		/* clock precision (ns) */
147 static long time_maxerror = MAXPHASE / 1000; /* maximum error (us) */
148 static long time_esterror = MAXPHASE / 1000; /* estimated error (us) */
149 static time_t time_reftime;		/* time at last adjustment (s) */
150 static long time_tick;			/* nanoseconds per tick (ns) */
151 static l_fp time_offset;		/* time offset (ns) */
152 static l_fp time_freq;			/* frequency offset (ns/s) */
153 static l_fp time_adj;			/* tick adjust (ns/s) */
154 
155 static struct lock ntp_lock = LOCK_INITIALIZER("ntplk", 0, 0);
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 	int error;
211 
212 	lockmgr(&ntp_lock, LK_EXCLUSIVE);
213 
214 	nanotime(&atv);
215 	ntv.time.tv_sec = atv.tv_sec;
216 	ntv.time.tv_nsec = atv.tv_nsec;
217 	ntv.maxerror = time_maxerror;
218 	ntv.esterror = time_esterror;
219 	ntv.tai = time_tai;
220 	ntv.time_state = time_state;
221 
222 	/*
223 	 * Status word error decode. If any of these conditions occur,
224 	 * an error is returned, instead of the status word. Most
225 	 * applications will care only about the fact the system clock
226 	 * may not be trusted, not about the details.
227 	 *
228 	 * Hardware or software error
229 	 */
230 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
231 
232 	/*
233 	 * PPS signal lost when either time or frequency synchronization
234 	 * requested
235 	 */
236 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
237 	    !(time_status & STA_PPSSIGNAL)) ||
238 
239 	/*
240 	 * PPS jitter exceeded when time synchronization requested
241 	 */
242 	    (time_status & STA_PPSTIME &&
243 	    time_status & STA_PPSJITTER) ||
244 
245 	/*
246 	 * PPS wander exceeded or calibration error when frequency
247 	 * synchronization requested
248 	 */
249 	    (time_status & STA_PPSFREQ &&
250 	    time_status & (STA_PPSWANDER | STA_PPSERROR))) {
251 		ntv.time_state = TIME_ERROR;
252 	}
253 
254 	error = sysctl_handle_opaque(oidp, &ntv, sizeof ntv, req);
255 	lockmgr(&ntp_lock, LK_RELEASE);
256 
257 	return error;
258 }
259 
260 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW, 0, "");
261 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE|CTLFLAG_RD,
262 	0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval", "");
263 
264 #ifdef PPS_SYNC
265 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW, &pps_shiftmax, 0, "");
266 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW, &pps_shift, 0, "");
267 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD, &time_monitor, 0, "");
268 
269 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD, &pps_freq, sizeof(pps_freq), "I", "");
270 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD, &time_freq, sizeof(time_freq), "I", "");
271 #endif
272 /*
273  * ntp_adjtime() - NTP daemon application interface
274  *
275  * See the timex.h header file for synopsis and API description. Note
276  * that the timex.constant structure member has a dual purpose to set
277  * the time constant and to set the TAI offset.
278  *
279  * MPALMOSTSAFE
280  */
281 int
282 sys_ntp_adjtime(struct sysmsg *sysmsg, const struct ntp_adjtime_args *uap)
283 {
284 	struct timex ntv;	/* temporary structure */
285 	long freq;		/* frequency ns/s) */
286 	int modes;		/* mode bits from structure */
287 	int error;
288 
289 	error = copyin((caddr_t)uap->tp, (caddr_t)&ntv, sizeof(ntv));
290 	if (error)
291 		return(error);
292 
293 	/*
294 	 * Update selected clock variables - only the superuser can
295 	 * change anything. Note that there is no error checking here on
296 	 * the assumption the superuser should know what it is doing.
297 	 * Note that either the time constant or TAI offset are loaded
298 	 * from the ntv.constant member, depending on the mode bits. If
299 	 * the STA_PLL bit in the status word is cleared, the state and
300 	 * status words are reset to the initial values at boot.
301 	 */
302 	modes = ntv.modes;
303 	if (modes)
304 		error = caps_priv_check_self(SYSCAP_NOSETTIME);
305 	if (error)
306 		return (error);
307 
308 	lockmgr(&ntp_lock, LK_EXCLUSIVE);
309 	crit_enter();
310 	if (modes & MOD_MAXERROR)
311 		time_maxerror = ntv.maxerror;
312 	if (modes & MOD_ESTERROR)
313 		time_esterror = ntv.esterror;
314 	if (modes & MOD_STATUS) {
315 		if (time_status & STA_PLL && !(ntv.status & STA_PLL)) {
316 			time_state = TIME_OK;
317 			time_status = STA_UNSYNC;
318 #ifdef PPS_SYNC
319 			pps_shift = PPS_FAVG;
320 #endif /* PPS_SYNC */
321 		}
322 		time_status &= STA_RONLY;
323 		time_status |= ntv.status & ~STA_RONLY;
324 	}
325 	if (modes & MOD_TIMECONST) {
326 		if (ntv.constant < 0)
327 			time_constant = 0;
328 		else if (ntv.constant > MAXTC)
329 			time_constant = MAXTC;
330 		else
331 			time_constant = ntv.constant;
332 	}
333 	if (modes & MOD_TAI) {
334 		if (ntv.constant > 0) /* XXX zero & negative numbers ? */
335 			time_tai = ntv.constant;
336 	}
337 #ifdef PPS_SYNC
338 	if (modes & MOD_PPSMAX) {
339 		if (ntv.shift < PPS_FAVG)
340 			pps_shiftmax = PPS_FAVG;
341 		else if (ntv.shift > PPS_FAVGMAX)
342 			pps_shiftmax = PPS_FAVGMAX;
343 		else
344 			pps_shiftmax = ntv.shift;
345 	}
346 #endif /* PPS_SYNC */
347 	if (modes & MOD_NANO)
348 		time_status |= STA_NANO;
349 	if (modes & MOD_MICRO)
350 		time_status &= ~STA_NANO;
351 	if (modes & MOD_CLKB)
352 		time_status |= STA_CLK;
353 	if (modes & MOD_CLKA)
354 		time_status &= ~STA_CLK;
355 	if (modes & MOD_OFFSET) {
356 		if (time_status & STA_NANO)
357 			hardupdate(ntv.offset);
358 		else
359 			hardupdate(ntv.offset * 1000);
360 	}
361 	/*
362 	 * Note: the userland specified frequency is in seconds per second
363 	 * times 65536e+6.  Multiply by a thousand and divide by 65336 to
364 	 * get nanoseconds.
365 	 */
366 	if (modes & MOD_FREQUENCY) {
367 		freq = (ntv.freq * 1000LL) >> 16;
368 		if (freq > MAXFREQ)
369 			L_LINT(time_freq, MAXFREQ);
370 		else if (freq < -MAXFREQ)
371 			L_LINT(time_freq, -MAXFREQ);
372 		else
373 			L_LINT(time_freq, freq);
374 #ifdef PPS_SYNC
375 		pps_freq = time_freq;
376 #endif /* PPS_SYNC */
377 	}
378 
379 	/*
380 	 * Retrieve all clock variables. Note that the TAI offset is
381 	 * returned only by ntp_gettime();
382 	 */
383 	if (time_status & STA_NANO)
384 		ntv.offset = time_monitor;
385 	else
386 		ntv.offset = time_monitor / 1000; /* XXX rounding ? */
387 	ntv.freq = L_GINT((time_freq / 1000LL) << 16);
388 	ntv.maxerror = time_maxerror;
389 	ntv.esterror = time_esterror;
390 	ntv.status = time_status;
391 	ntv.constant = time_constant;
392 	if (time_status & STA_NANO)
393 		ntv.precision = time_precision;
394 	else
395 		ntv.precision = time_precision / 1000;
396 	ntv.tolerance = MAXFREQ * SCALE_PPM;
397 #ifdef PPS_SYNC
398 	ntv.shift = pps_shift;
399 	ntv.ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
400 	if (time_status & STA_NANO)
401 		ntv.jitter = pps_jitter;
402 	else
403 		ntv.jitter = pps_jitter / 1000;
404 	ntv.stabil = pps_stabil;
405 	ntv.calcnt = pps_calcnt;
406 	ntv.errcnt = pps_errcnt;
407 	ntv.jitcnt = pps_jitcnt;
408 	ntv.stbcnt = pps_stbcnt;
409 #endif /* PPS_SYNC */
410 	crit_exit();
411 	lockmgr(&ntp_lock, LK_RELEASE);
412 
413 	error = copyout((caddr_t)&ntv, (caddr_t)uap->tp, sizeof(ntv));
414 	if (error)
415 		return (error);
416 
417 	/*
418 	 * Status word error decode. See comments in
419 	 * ntp_gettime() routine.
420 	 */
421 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
422 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
423 	    !(time_status & STA_PPSSIGNAL)) ||
424 	    (time_status & STA_PPSTIME &&
425 	    time_status & STA_PPSJITTER) ||
426 	    (time_status & STA_PPSFREQ &&
427 	    time_status & (STA_PPSWANDER | STA_PPSERROR))) {
428 		sysmsg->sysmsg_result = TIME_ERROR;
429 	} else {
430 		sysmsg->sysmsg_result = time_state;
431 	}
432 	return (0);
433 }
434 
435 /*
436  * second_overflow() - called after ntp_tick_adjust()
437  *
438  * This routine is ordinarily called from hardclock() whenever the seconds
439  * hand rolls over.  It returns leap seconds to add or drop, and sets nsec_adj
440  * to the total adjustment to make over the next second in (ns << 32).
441  *
442  * This routine is only called by cpu #0.
443  */
444 int
445 ntp_update_second(time_t newsec, int64_t *nsec_adj)
446 {
447 	l_fp ftemp;		/* 32/64-bit temporary */
448 	int  adjsec = 0;
449 
450 	/*
451 	 * On rollover of the second both the nanosecond and microsecond
452 	 * clocks are updated and the state machine cranked as
453 	 * necessary. The phase adjustment to be used for the next
454 	 * second is calculated and the maximum error is increased by
455 	 * the tolerance.
456 	 */
457 	time_maxerror += MAXFREQ / 1000;
458 
459 	/*
460 	 * Leap second processing. If in leap-insert state at
461 	 * the end of the day, the system clock is set back one
462 	 * second; if in leap-delete state, the system clock is
463 	 * set ahead one second. The nano_time() routine or
464 	 * external clock driver will insure that reported time
465 	 * is always monotonic.
466 	 */
467 	switch (time_state) {
468 
469 		/*
470 		 * No warning.
471 		 */
472 		case TIME_OK:
473 		if (time_status & STA_INS)
474 			time_state = TIME_INS;
475 		else if (time_status & STA_DEL)
476 			time_state = TIME_DEL;
477 		break;
478 
479 		/*
480 		 * Insert second 23:59:60 following second
481 		 * 23:59:59.
482 		 */
483 		case TIME_INS:
484 		if (!(time_status & STA_INS))
485 			time_state = TIME_OK;
486 		else if ((newsec) % 86400 == 0) {
487 			--adjsec;
488 			time_state = TIME_OOP;
489 		}
490 		break;
491 
492 		/*
493 		 * Delete second 23:59:59.
494 		 */
495 		case TIME_DEL:
496 		if (!(time_status & STA_DEL))
497 			time_state = TIME_OK;
498 		else if (((newsec) + 1) % 86400 == 0) {
499 			++adjsec;
500 			time_tai--;
501 			time_state = TIME_WAIT;
502 		}
503 		break;
504 
505 		/*
506 		 * Insert second in progress.
507 		 */
508 		case TIME_OOP:
509 			time_tai++;
510 			time_state = TIME_WAIT;
511 		break;
512 
513 		/*
514 		 * Wait for status bits to clear.
515 		 */
516 		case TIME_WAIT:
517 		if (!(time_status & (STA_INS | STA_DEL)))
518 			time_state = TIME_OK;
519 	}
520 
521 	/*
522 	 * time_offset represents the total time adjustment we wish to
523 	 * make (over no particular period of time).  time_freq represents
524 	 * the frequency compensation we wish to apply.
525 	 *
526 	 * time_adj represents the total adjustment we wish to make over
527 	 * one full second.  hardclock usually applies this adjustment in
528 	 * time_adj / hz jumps, hz times a second.
529 	 */
530 	ftemp = time_offset;
531 #ifdef PPS_SYNC
532 	/* XXX even if PPS signal dies we should finish adjustment ? */
533 	if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
534 		L_RSHIFT(ftemp, pps_shift);
535 	else
536 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
537 #else
538 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
539 #endif /* PPS_SYNC */
540 	time_adj = ftemp;		/* adjustment for part of the offset */
541 	L_SUB(time_offset, ftemp);
542 	L_ADD(time_adj, time_freq);	/* add frequency correction */
543 	*nsec_adj = time_adj;
544 #ifdef PPS_SYNC
545 	if (pps_valid > 0)
546 		pps_valid--;
547 	else
548 		time_status &= ~STA_PPSSIGNAL;
549 #endif /* PPS_SYNC */
550 	return(adjsec);
551 }
552 
553 /*
554  * ntp_init() - initialize variables and structures
555  *
556  * This routine must be called after the kernel variables hz and tick
557  * are set or changed and before the next tick interrupt. In this
558  * particular implementation, these values are assumed set elsewhere in
559  * the kernel. The design allows the clock frequency and tick interval
560  * to be changed while the system is running. So, this routine should
561  * probably be integrated with the code that does that.
562  */
563 static void
564 ntp_init(void)
565 {
566 
567 	/*
568 	 * The following variable must be initialized any time the
569 	 * kernel variable hz is changed.
570 	 */
571 	time_tick = NANOSECOND / hz;
572 
573 	/*
574 	 * The following variables are initialized only at startup. Only
575 	 * those structures not cleared by the compiler need to be
576 	 * initialized, and these only in the simulator. In the actual
577 	 * kernel, any nonzero values here will quickly evaporate.
578 	 */
579 	L_CLR(time_offset);
580 	L_CLR(time_freq);
581 #ifdef PPS_SYNC
582 	pps_tf[0].tv_sec = pps_tf[0].tv_nsec = 0;
583 	pps_tf[1].tv_sec = pps_tf[1].tv_nsec = 0;
584 	pps_tf[2].tv_sec = pps_tf[2].tv_nsec = 0;
585 	pps_fcount = 0;
586 	L_CLR(pps_freq);
587 #endif /* PPS_SYNC */
588 }
589 
590 SYSINIT(ntpclocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, ntp_init, NULL);
591 
592 /*
593  * hardupdate() - local clock update
594  *
595  * This routine is called by ntp_adjtime() to update the local clock
596  * phase and frequency. The implementation is of an adaptive-parameter,
597  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
598  * time and frequency offset estimates for each call. If the kernel PPS
599  * discipline code is configured (PPS_SYNC), the PPS signal itself
600  * determines the new time offset, instead of the calling argument.
601  * Presumably, calls to ntp_adjtime() occur only when the caller
602  * believes the local clock is valid within some bound (+-128 ms with
603  * NTP). If the caller's time is far different than the PPS time, an
604  * argument will ensue, and it's not clear who will lose.
605  *
606  * For uncompensated quartz crystal oscillators and nominal update
607  * intervals less than 256 s, operation should be in phase-lock mode,
608  * where the loop is disciplined to phase. For update intervals greater
609  * than 1024 s, operation should be in frequency-lock mode, where the
610  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
611  * is selected by the STA_MODE status bit.
612  */
613 static void
614 hardupdate(long offset)
615 {
616 	long mtemp;
617 	l_fp ftemp;
618 
619 	/*
620 	 * Select how the phase is to be controlled and from which
621 	 * source. If the PPS signal is present and enabled to
622 	 * discipline the time, the PPS offset is used; otherwise, the
623 	 * argument offset is used.
624 	 */
625 	if (!(time_status & STA_PLL))
626 		return;
627 	if (!((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))) {
628 		if (offset > MAXPHASE)
629 			time_monitor = MAXPHASE;
630 		else if (offset < -MAXPHASE)
631 			time_monitor = -MAXPHASE;
632 		else
633 			time_monitor = offset;
634 		L_LINT(time_offset, time_monitor);
635 	}
636 
637 	/*
638 	 * Select how the frequency is to be controlled and in which
639 	 * mode (PLL or FLL). If the PPS signal is present and enabled
640 	 * to discipline the frequency, the PPS frequency is used;
641 	 * otherwise, the argument offset is used to compute it.
642 	 */
643 	if ((time_status & STA_PPSFREQ) && time_status & STA_PPSSIGNAL) {
644 		time_reftime = time_uptime;
645 		return;
646 	}
647 	if ((time_status & STA_FREQHOLD) || time_reftime == 0)
648 		time_reftime = time_uptime;
649 	mtemp = time_uptime - time_reftime;
650 	L_LINT(ftemp, time_monitor);
651 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
652 	L_MPY(ftemp, mtemp);
653 	L_ADD(time_freq, ftemp);
654 	time_status &= ~STA_MODE;
655 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp > MAXSEC)) {
656 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
657 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
658 		L_ADD(time_freq, ftemp);
659 		time_status |= STA_MODE;
660 	}
661 	time_reftime = time_uptime;
662 	if (L_GINT(time_freq) > MAXFREQ)
663 		L_LINT(time_freq, MAXFREQ);
664 	else if (L_GINT(time_freq) < -MAXFREQ)
665 		L_LINT(time_freq, -MAXFREQ);
666 }
667 
668 #ifdef PPS_SYNC
669 /*
670  * hardpps() - discipline CPU clock oscillator to external PPS signal
671  *
672  * This routine is called at each PPS interrupt in order to discipline
673  * the CPU clock oscillator to the PPS signal. There are two independent
674  * first-order feedback loops, one for the phase, the other for the
675  * frequency. The phase loop measures and grooms the PPS phase offset
676  * and leaves it in a handy spot for the seconds overflow routine. The
677  * frequency loop averages successive PPS phase differences and
678  * calculates the PPS frequency offset, which is also processed by the
679  * seconds overflow routine. The code requires the caller to capture the
680  * time and architecture-dependent hardware counter values in
681  * nanoseconds at the on-time PPS signal transition.
682  *
683  * Note that, on some Unix systems this routine runs at an interrupt
684  * priority level higher than the timer interrupt routine hardclock().
685  * Therefore, the variables used are distinct from the hardclock()
686  * variables, except for the actual time and frequency variables, which
687  * are determined by this routine and updated atomically.
688  */
689 void
690 hardpps(struct timespec *tsp, long nsec)
691 {
692 	long u_sec, u_nsec, v_nsec; /* temps */
693 	l_fp ftemp;
694 
695 	/*
696 	 * The signal is first processed by a range gate and frequency
697 	 * discriminator. The range gate rejects noise spikes outside
698 	 * the range +-500 us. The frequency discriminator rejects input
699 	 * signals with apparent frequency outside the range 1 +-500
700 	 * PPM. If two hits occur in the same second, we ignore the
701 	 * later hit; if not and a hit occurs outside the range gate,
702 	 * keep the later hit for later comparison, but do not process
703 	 * it.
704 	 */
705 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
706 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
707 	pps_valid = PPS_VALID;
708 	u_sec = tsp->tv_sec;
709 	u_nsec = tsp->tv_nsec;
710 	if (u_nsec >= (NANOSECOND >> 1)) {
711 		u_nsec -= NANOSECOND;
712 		u_sec++;
713 	}
714 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
715 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND -
716 	    MAXFREQ)
717 		return;
718 	pps_tf[2] = pps_tf[1];
719 	pps_tf[1] = pps_tf[0];
720 	pps_tf[0].tv_sec = u_sec;
721 	pps_tf[0].tv_nsec = u_nsec;
722 
723 	/*
724 	 * Compute the difference between the current and previous
725 	 * counter values. If the difference exceeds 0.5 s, assume it
726 	 * has wrapped around, so correct 1.0 s. If the result exceeds
727 	 * the tick interval, the sample point has crossed a tick
728 	 * boundary during the last second, so correct the tick. Very
729 	 * intricate.
730 	 */
731 	u_nsec = nsec;
732 	if (u_nsec > (NANOSECOND >> 1))
733 		u_nsec -= NANOSECOND;
734 	else if (u_nsec < -(NANOSECOND >> 1))
735 		u_nsec += NANOSECOND;
736 	pps_fcount += u_nsec;
737 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
738 		return;
739 	time_status &= ~STA_PPSJITTER;
740 
741 	/*
742 	 * A three-stage median filter is used to help denoise the PPS
743 	 * time. The median sample becomes the time offset estimate; the
744 	 * difference between the other two samples becomes the time
745 	 * dispersion (jitter) estimate.
746 	 */
747 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
748 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
749 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
750 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
751 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
752 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
753 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
754 		} else {
755 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
756 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
757 		}
758 	} else {
759 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
760 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
761 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
762 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
763 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
764 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
765 		} else {
766 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
767 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
768 		}
769 	}
770 
771 	/*
772 	 * Nominal jitter is due to PPS signal noise and interrupt
773 	 * latency. If it exceeds the popcorn threshold, the sample is
774 	 * discarded. otherwise, if so enabled, the time offset is
775 	 * updated. We can tolerate a modest loss of data here without
776 	 * much degrading time accuracy.
777 	 */
778 	if (u_nsec > (pps_jitter << PPS_POPCORN)) {
779 		time_status |= STA_PPSJITTER;
780 		pps_jitcnt++;
781 	} else if (time_status & STA_PPSTIME) {
782 		time_monitor = -v_nsec;
783 		L_LINT(time_offset, time_monitor);
784 	}
785 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
786 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
787 	if (u_sec < (1 << pps_shift))
788 		return;
789 
790 	/*
791 	 * At the end of the calibration interval the difference between
792 	 * the first and last counter values becomes the scaled
793 	 * frequency. It will later be divided by the length of the
794 	 * interval to determine the frequency update. If the frequency
795 	 * exceeds a sanity threshold, or if the actual calibration
796 	 * interval is not equal to the expected length, the data are
797 	 * discarded. We can tolerate a modest loss of data here without
798 	 * much degrading frequency accuracy.
799 	 */
800 	pps_calcnt++;
801 	v_nsec = -pps_fcount;
802 	pps_lastsec = pps_tf[0].tv_sec;
803 	pps_fcount = 0;
804 	u_nsec = MAXFREQ << pps_shift;
805 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 <<
806 	    pps_shift)) {
807 		time_status |= STA_PPSERROR;
808 		pps_errcnt++;
809 		return;
810 	}
811 
812 	/*
813 	 * Here the raw frequency offset and wander (stability) is
814 	 * calculated. If the wander is less than the wander threshold
815 	 * for four consecutive averaging intervals, the interval is
816 	 * doubled; if it is greater than the threshold for four
817 	 * consecutive intervals, the interval is halved. The scaled
818 	 * frequency offset is converted to frequency offset. The
819 	 * stability metric is calculated as the average of recent
820 	 * frequency changes, but is used only for performance
821 	 * monitoring.
822 	 */
823 	L_LINT(ftemp, v_nsec);
824 	L_RSHIFT(ftemp, pps_shift);
825 	L_SUB(ftemp, pps_freq);
826 	u_nsec = L_GINT(ftemp);
827 	if (u_nsec > PPS_MAXWANDER) {
828 		L_LINT(ftemp, PPS_MAXWANDER);
829 		pps_intcnt--;
830 		time_status |= STA_PPSWANDER;
831 		pps_stbcnt++;
832 	} else if (u_nsec < -PPS_MAXWANDER) {
833 		L_LINT(ftemp, -PPS_MAXWANDER);
834 		pps_intcnt--;
835 		time_status |= STA_PPSWANDER;
836 		pps_stbcnt++;
837 	} else {
838 		pps_intcnt++;
839 	}
840 	if (pps_intcnt >= 4) {
841 		pps_intcnt = 4;
842 		if (pps_shift < pps_shiftmax) {
843 			pps_shift++;
844 			pps_intcnt = 0;
845 		}
846 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
847 		pps_intcnt = -4;
848 		if (pps_shift > PPS_FAVG) {
849 			pps_shift--;
850 			pps_intcnt = 0;
851 		}
852 	}
853 	if (u_nsec < 0)
854 		u_nsec = -u_nsec;
855 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
856 
857 	/*
858 	 * The PPS frequency is recalculated and clamped to the maximum
859 	 * MAXFREQ. If enabled, the system clock frequency is updated as
860 	 * well.
861 	 */
862 	L_ADD(pps_freq, ftemp);
863 	u_nsec = L_GINT(pps_freq);
864 	if (u_nsec > MAXFREQ)
865 		L_LINT(pps_freq, MAXFREQ);
866 	else if (u_nsec < -MAXFREQ)
867 		L_LINT(pps_freq, -MAXFREQ);
868 	if (time_status & STA_PPSFREQ)
869 		time_freq = pps_freq;
870 }
871 #endif /* PPS_SYNC */
872