xref: /freebsd/sys/kern/kern_ntptime.c (revision 4d3fc8b0570b29fb0d6ee9525f104d52176ff0d4)
1 /*-
2  ***********************************************************************
3  *								       *
4  * Copyright (c) David L. Mills 1993-2001			       *
5  *								       *
6  * Permission to use, copy, modify, and distribute this software and   *
7  * its documentation for any purpose and without fee is hereby	       *
8  * granted, provided that the above copyright notice appears in all    *
9  * copies and that both the copyright notice and this permission       *
10  * notice appear in supporting documentation, and that the name	       *
11  * University of Delaware not be used in advertising or publicity      *
12  * pertaining to distribution of the software without specific,	       *
13  * written prior permission. The University of Delaware makes no       *
14  * representations about the suitability this software for any	       *
15  * purpose. It is provided "as is" without express or implied	       *
16  * warranty.							       *
17  *								       *
18  **********************************************************************/
19 
20 /*
21  * Adapted from the original sources for FreeBSD and timecounters by:
22  * Poul-Henning Kamp <phk@FreeBSD.org>.
23  *
24  * The 32bit version of the "LP" macros seems a bit past its "sell by"
25  * date so I have retained only the 64bit version and included it directly
26  * in this file.
27  *
28  * Only minor changes done to interface with the timecounters over in
29  * sys/kern/kern_clock.c.   Some of the comments below may be (even more)
30  * confusing and/or plain wrong in that context.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_ntp.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/sysproto.h>
41 #include <sys/eventhandler.h>
42 #include <sys/kernel.h>
43 #include <sys/priv.h>
44 #include <sys/proc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/time.h>
48 #include <sys/timex.h>
49 #include <sys/timetc.h>
50 #include <sys/timepps.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/sysctl.h>
53 
54 #ifdef PPS_SYNC
55 FEATURE(pps_sync, "Support usage of external PPS signal by kernel PLL");
56 #endif
57 
58 /*
59  * Single-precision macros for 64-bit machines
60  */
61 typedef int64_t l_fp;
62 #define L_ADD(v, u)	((v) += (u))
63 #define L_SUB(v, u)	((v) -= (u))
64 #define L_ADDHI(v, a)	((v) += (int64_t)(a) << 32)
65 #define L_NEG(v)	((v) = -(v))
66 #define L_RSHIFT(v, n) \
67 	do { \
68 		if ((v) < 0) \
69 			(v) = -(-(v) >> (n)); \
70 		else \
71 			(v) = (v) >> (n); \
72 	} while (0)
73 #define L_MPY(v, a)	((v) *= (a))
74 #define L_CLR(v)	((v) = 0)
75 #define L_ISNEG(v)	((v) < 0)
76 #define L_LINT(v, a)	((v) = (int64_t)(a) << 32)
77 #define L_GINT(v)	((v) < 0 ? -(-(v) >> 32) : (v) >> 32)
78 
79 /*
80  * Generic NTP kernel interface
81  *
82  * These routines constitute the Network Time Protocol (NTP) interfaces
83  * for user and daemon application programs. The ntp_gettime() routine
84  * provides the time, maximum error (synch distance) and estimated error
85  * (dispersion) to client user application programs. The ntp_adjtime()
86  * routine is used by the NTP daemon to adjust the system clock to an
87  * externally derived time. The time offset and related variables set by
88  * this routine are used by other routines in this module to adjust the
89  * phase and frequency of the clock discipline loop which controls the
90  * system clock.
91  *
92  * When the kernel time is reckoned directly in nanoseconds (NTP_NANO
93  * defined), the time at each tick interrupt is derived directly from
94  * the kernel time variable. When the kernel time is reckoned in
95  * microseconds, (NTP_NANO undefined), the time is derived from the
96  * kernel time variable together with a variable representing the
97  * leftover nanoseconds at the last tick interrupt. In either case, the
98  * current nanosecond time is reckoned from these values plus an
99  * interpolated value derived by the clock routines in another
100  * architecture-specific module. The interpolation can use either a
101  * dedicated counter or a processor cycle counter (PCC) implemented in
102  * some architectures.
103  *
104  * Note that all routines must run at priority splclock or higher.
105  */
106 /*
107  * Phase/frequency-lock loop (PLL/FLL) definitions
108  *
109  * The nanosecond clock discipline uses two variable types, time
110  * variables and frequency variables. Both types are represented as 64-
111  * bit fixed-point quantities with the decimal point between two 32-bit
112  * halves. On a 32-bit machine, each half is represented as a single
113  * word and mathematical operations are done using multiple-precision
114  * arithmetic. On a 64-bit machine, ordinary computer arithmetic is
115  * used.
116  *
117  * A time variable is a signed 64-bit fixed-point number in ns and
118  * fraction. It represents the remaining time offset to be amortized
119  * over succeeding tick interrupts. The maximum time offset is about
120  * 0.5 s and the resolution is about 2.3e-10 ns.
121  *
122  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
123  *  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
124  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
125  * |s s s|			 ns				   |
126  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
127  * |			    fraction				   |
128  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
129  *
130  * A frequency variable is a signed 64-bit fixed-point number in ns/s
131  * and fraction. It represents the ns and fraction to be added to the
132  * kernel time variable at each second. The maximum frequency offset is
133  * about +-500000 ns/s and the resolution is about 2.3e-10 ns/s.
134  *
135  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
136  *  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
137  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
138  * |s s s s s s s s s s s s s|	          ns/s			   |
139  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
140  * |			    fraction				   |
141  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
142  */
143 /*
144  * The following variables establish the state of the PLL/FLL and the
145  * residual time and frequency offset of the local clock.
146  */
147 #define SHIFT_PLL	4		/* PLL loop gain (shift) */
148 #define SHIFT_FLL	2		/* FLL loop gain (shift) */
149 
150 static int time_state = TIME_OK;	/* clock state */
151 int time_status = STA_UNSYNC;	/* clock status bits */
152 static long time_tai;			/* TAI offset (s) */
153 static long time_monitor;		/* last time offset scaled (ns) */
154 static long time_constant;		/* poll interval (shift) (s) */
155 static long time_precision = 1;		/* clock precision (ns) */
156 static long time_maxerror = MAXPHASE / 1000; /* maximum error (us) */
157 long time_esterror = MAXPHASE / 1000; /* estimated error (us) */
158 static long time_reftime;		/* uptime at last adjustment (s) */
159 static l_fp time_offset;		/* time offset (ns) */
160 static l_fp time_freq;			/* frequency offset (ns/s) */
161 static l_fp time_adj;			/* tick adjust (ns/s) */
162 
163 static int64_t time_adjtime;		/* correction from adjtime(2) (usec) */
164 
165 static struct mtx ntp_lock;
166 MTX_SYSINIT(ntp, &ntp_lock, "ntp", MTX_SPIN);
167 
168 #define	NTP_LOCK()		mtx_lock_spin(&ntp_lock)
169 #define	NTP_UNLOCK()		mtx_unlock_spin(&ntp_lock)
170 #define	NTP_ASSERT_LOCKED()	mtx_assert(&ntp_lock, MA_OWNED)
171 
172 #ifdef PPS_SYNC
173 /*
174  * The following variables are used when a pulse-per-second (PPS) signal
175  * is available and connected via a modem control lead. They establish
176  * the engineering parameters of the clock discipline loop when
177  * controlled by the PPS signal.
178  */
179 #define PPS_FAVG	2		/* min freq avg interval (s) (shift) */
180 #define PPS_FAVGDEF	8		/* default freq avg int (s) (shift) */
181 #define PPS_FAVGMAX	15		/* max freq avg interval (s) (shift) */
182 #define PPS_PAVG	4		/* phase avg interval (s) (shift) */
183 #define PPS_VALID	120		/* PPS signal watchdog max (s) */
184 #define PPS_MAXWANDER	100000		/* max PPS wander (ns/s) */
185 #define PPS_POPCORN	2		/* popcorn spike threshold (shift) */
186 
187 static struct timespec pps_tf[3];	/* phase median filter */
188 static l_fp pps_freq;			/* scaled frequency offset (ns/s) */
189 static long pps_fcount;			/* frequency accumulator */
190 static long pps_jitter;			/* nominal jitter (ns) */
191 static long pps_stabil;			/* nominal stability (scaled ns/s) */
192 static long pps_lastsec;		/* time at last calibration (s) */
193 static int pps_valid;			/* signal watchdog counter */
194 static int pps_shift = PPS_FAVG;	/* interval duration (s) (shift) */
195 static int pps_shiftmax = PPS_FAVGDEF;	/* max interval duration (s) (shift) */
196 static int pps_intcnt;			/* wander counter */
197 
198 /*
199  * PPS signal quality monitors
200  */
201 static long pps_calcnt;			/* calibration intervals */
202 static long pps_jitcnt;			/* jitter limit exceeded */
203 static long pps_stbcnt;			/* stability limit exceeded */
204 static long pps_errcnt;			/* calibration errors */
205 #endif /* PPS_SYNC */
206 /*
207  * End of phase/frequency-lock loop (PLL/FLL) definitions
208  */
209 
210 static void hardupdate(long offset);
211 static void ntp_gettime1(struct ntptimeval *ntvp);
212 static bool ntp_is_time_error(int tsl);
213 
214 static bool
215 ntp_is_time_error(int tsl)
216 {
217 
218 	/*
219 	 * Status word error decode. If any of these conditions occur,
220 	 * an error is returned, instead of the status word. Most
221 	 * applications will care only about the fact the system clock
222 	 * may not be trusted, not about the details.
223 	 *
224 	 * Hardware or software error
225 	 */
226 	if ((tsl & (STA_UNSYNC | STA_CLOCKERR)) ||
227 
228 	/*
229 	 * PPS signal lost when either time or frequency synchronization
230 	 * requested
231 	 */
232 	    (tsl & (STA_PPSFREQ | STA_PPSTIME) &&
233 	    !(tsl & STA_PPSSIGNAL)) ||
234 
235 	/*
236 	 * PPS jitter exceeded when time synchronization requested
237 	 */
238 	    (tsl & STA_PPSTIME && tsl & STA_PPSJITTER) ||
239 
240 	/*
241 	 * PPS wander exceeded or calibration error when frequency
242 	 * synchronization requested
243 	 */
244 	    (tsl & STA_PPSFREQ &&
245 	    tsl & (STA_PPSWANDER | STA_PPSERROR)))
246 		return (true);
247 
248 	return (false);
249 }
250 
251 static void
252 ntp_gettime1(struct ntptimeval *ntvp)
253 {
254 	struct timespec atv;	/* nanosecond time */
255 
256 	NTP_ASSERT_LOCKED();
257 
258 	nanotime(&atv);
259 	ntvp->time.tv_sec = atv.tv_sec;
260 	ntvp->time.tv_nsec = atv.tv_nsec;
261 	ntvp->maxerror = time_maxerror;
262 	ntvp->esterror = time_esterror;
263 	ntvp->tai = time_tai;
264 	ntvp->time_state = time_state;
265 
266 	if (ntp_is_time_error(time_status))
267 		ntvp->time_state = TIME_ERROR;
268 }
269 
270 /*
271  * ntp_gettime() - NTP user application interface
272  *
273  * See the timex.h header file for synopsis and API description.  Note that
274  * the TAI offset is returned in the ntvtimeval.tai structure member.
275  */
276 #ifndef _SYS_SYSPROTO_H_
277 struct ntp_gettime_args {
278 	struct ntptimeval *ntvp;
279 };
280 #endif
281 /* ARGSUSED */
282 int
283 sys_ntp_gettime(struct thread *td, struct ntp_gettime_args *uap)
284 {
285 	struct ntptimeval ntv;
286 
287 	memset(&ntv, 0, sizeof(ntv));
288 
289 	NTP_LOCK();
290 	ntp_gettime1(&ntv);
291 	NTP_UNLOCK();
292 
293 	td->td_retval[0] = ntv.time_state;
294 	return (copyout(&ntv, uap->ntvp, sizeof(ntv)));
295 }
296 
297 static int
298 ntp_sysctl(SYSCTL_HANDLER_ARGS)
299 {
300 	struct ntptimeval ntv;	/* temporary structure */
301 
302 	memset(&ntv, 0, sizeof(ntv));
303 
304 	NTP_LOCK();
305 	ntp_gettime1(&ntv);
306 	NTP_UNLOCK();
307 
308 	return (sysctl_handle_opaque(oidp, &ntv, sizeof(ntv), req));
309 }
310 
311 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
312     "");
313 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE | CTLFLAG_RD |
314     CTLFLAG_MPSAFE, 0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval",
315     "");
316 
317 #ifdef PPS_SYNC
318 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW,
319     &pps_shiftmax, 0, "Max interval duration (sec) (shift)");
320 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW,
321     &pps_shift, 0, "Interval duration (sec) (shift)");
322 SYSCTL_LONG(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD,
323     &time_monitor, 0, "Last time offset scaled (ns)");
324 
325 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
326     &pps_freq, 0,
327     "Scaled frequency offset (ns/sec)");
328 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
329     &time_freq, 0,
330     "Frequency offset (ns/sec)");
331 #endif
332 
333 /*
334  * ntp_adjtime() - NTP daemon application interface
335  *
336  * See the timex.h header file for synopsis and API description.  Note that
337  * the timex.constant structure member has a dual purpose to set the time
338  * constant and to set the TAI offset.
339  */
340 int
341 kern_ntp_adjtime(struct thread *td, struct timex *ntv, int *retvalp)
342 {
343 	long freq;		/* frequency ns/s) */
344 	int modes;		/* mode bits from structure */
345 	int error, retval;
346 
347 	/*
348 	 * Update selected clock variables - only the superuser can
349 	 * change anything. Note that there is no error checking here on
350 	 * the assumption the superuser should know what it is doing.
351 	 * Note that either the time constant or TAI offset are loaded
352 	 * from the ntv.constant member, depending on the mode bits. If
353 	 * the STA_PLL bit in the status word is cleared, the state and
354 	 * status words are reset to the initial values at boot.
355 	 */
356 	modes = ntv->modes;
357 	error = 0;
358 	if (modes)
359 		error = priv_check(td, PRIV_NTP_ADJTIME);
360 	if (error != 0)
361 		return (error);
362 	NTP_LOCK();
363 	if (modes & MOD_MAXERROR)
364 		time_maxerror = ntv->maxerror;
365 	if (modes & MOD_ESTERROR)
366 		time_esterror = ntv->esterror;
367 	if (modes & MOD_STATUS) {
368 		if (time_status & STA_PLL && !(ntv->status & STA_PLL)) {
369 			time_state = TIME_OK;
370 			time_status = STA_UNSYNC;
371 #ifdef PPS_SYNC
372 			pps_shift = PPS_FAVG;
373 #endif /* PPS_SYNC */
374 		}
375 		time_status &= STA_RONLY;
376 		time_status |= ntv->status & ~STA_RONLY;
377 	}
378 	if (modes & MOD_TIMECONST) {
379 		if (ntv->constant < 0)
380 			time_constant = 0;
381 		else if (ntv->constant > MAXTC)
382 			time_constant = MAXTC;
383 		else
384 			time_constant = ntv->constant;
385 	}
386 	if (modes & MOD_TAI) {
387 		if (ntv->constant > 0) /* XXX zero & negative numbers ? */
388 			time_tai = ntv->constant;
389 	}
390 #ifdef PPS_SYNC
391 	if (modes & MOD_PPSMAX) {
392 		if (ntv->shift < PPS_FAVG)
393 			pps_shiftmax = PPS_FAVG;
394 		else if (ntv->shift > PPS_FAVGMAX)
395 			pps_shiftmax = PPS_FAVGMAX;
396 		else
397 			pps_shiftmax = ntv->shift;
398 	}
399 #endif /* PPS_SYNC */
400 	if (modes & MOD_NANO)
401 		time_status |= STA_NANO;
402 	if (modes & MOD_MICRO)
403 		time_status &= ~STA_NANO;
404 	if (modes & MOD_CLKB)
405 		time_status |= STA_CLK;
406 	if (modes & MOD_CLKA)
407 		time_status &= ~STA_CLK;
408 	if (modes & MOD_FREQUENCY) {
409 		freq = (ntv->freq * 1000LL) >> 16;
410 		if (freq > MAXFREQ)
411 			L_LINT(time_freq, MAXFREQ);
412 		else if (freq < -MAXFREQ)
413 			L_LINT(time_freq, -MAXFREQ);
414 		else {
415 			/*
416 			 * ntv->freq is [PPM * 2^16] = [us/s * 2^16]
417 			 * time_freq is [ns/s * 2^32]
418 			 */
419 			time_freq = ntv->freq * 1000LL * 65536LL;
420 		}
421 #ifdef PPS_SYNC
422 		pps_freq = time_freq;
423 #endif /* PPS_SYNC */
424 	}
425 	if (modes & MOD_OFFSET) {
426 		if (time_status & STA_NANO)
427 			hardupdate(ntv->offset);
428 		else
429 			hardupdate(ntv->offset * 1000);
430 	}
431 
432 	/*
433 	 * Retrieve all clock variables. Note that the TAI offset is
434 	 * returned only by ntp_gettime();
435 	 */
436 	if (time_status & STA_NANO)
437 		ntv->offset = L_GINT(time_offset);
438 	else
439 		ntv->offset = L_GINT(time_offset) / 1000; /* XXX rounding ? */
440 	ntv->freq = L_GINT((time_freq / 1000LL) << 16);
441 	ntv->maxerror = time_maxerror;
442 	ntv->esterror = time_esterror;
443 	ntv->status = time_status;
444 	ntv->constant = time_constant;
445 	if (time_status & STA_NANO)
446 		ntv->precision = time_precision;
447 	else
448 		ntv->precision = time_precision / 1000;
449 	ntv->tolerance = MAXFREQ * SCALE_PPM;
450 #ifdef PPS_SYNC
451 	ntv->shift = pps_shift;
452 	ntv->ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
453 	if (time_status & STA_NANO)
454 		ntv->jitter = pps_jitter;
455 	else
456 		ntv->jitter = pps_jitter / 1000;
457 	ntv->stabil = pps_stabil;
458 	ntv->calcnt = pps_calcnt;
459 	ntv->errcnt = pps_errcnt;
460 	ntv->jitcnt = pps_jitcnt;
461 	ntv->stbcnt = pps_stbcnt;
462 #endif /* PPS_SYNC */
463 	retval = ntp_is_time_error(time_status) ? TIME_ERROR : time_state;
464 	NTP_UNLOCK();
465 
466 	*retvalp = retval;
467 	return (0);
468 }
469 
470 #ifndef _SYS_SYSPROTO_H_
471 struct ntp_adjtime_args {
472 	struct timex *tp;
473 };
474 #endif
475 
476 int
477 sys_ntp_adjtime(struct thread *td, struct ntp_adjtime_args *uap)
478 {
479 	struct timex ntv;
480 	int error, retval;
481 
482 	error = copyin(uap->tp, &ntv, sizeof(ntv));
483 	if (error == 0) {
484 		error = kern_ntp_adjtime(td, &ntv, &retval);
485 		if (error == 0) {
486 			error = copyout(&ntv, uap->tp, sizeof(ntv));
487 			if (error == 0)
488 				td->td_retval[0] = retval;
489 		}
490 	}
491 	return (error);
492 }
493 
494 /*
495  * second_overflow() - called after ntp_tick_adjust()
496  *
497  * This routine is ordinarily called immediately following the above
498  * routine ntp_tick_adjust(). While these two routines are normally
499  * combined, they are separated here only for the purposes of
500  * simulation.
501  */
502 void
503 ntp_update_second(int64_t *adjustment, time_t *newsec)
504 {
505 	int tickrate;
506 	l_fp ftemp;		/* 32/64-bit temporary */
507 
508 	NTP_LOCK();
509 
510 	/*
511 	 * On rollover of the second both the nanosecond and microsecond
512 	 * clocks are updated and the state machine cranked as
513 	 * necessary. The phase adjustment to be used for the next
514 	 * second is calculated and the maximum error is increased by
515 	 * the tolerance.
516 	 */
517 	time_maxerror += MAXFREQ / 1000;
518 
519 	/*
520 	 * Leap second processing. If in leap-insert state at
521 	 * the end of the day, the system clock is set back one
522 	 * second; if in leap-delete state, the system clock is
523 	 * set ahead one second. The nano_time() routine or
524 	 * external clock driver will insure that reported time
525 	 * is always monotonic.
526 	 */
527 	switch (time_state) {
528 		/*
529 		 * No warning.
530 		 */
531 		case TIME_OK:
532 		if (time_status & STA_INS)
533 			time_state = TIME_INS;
534 		else if (time_status & STA_DEL)
535 			time_state = TIME_DEL;
536 		break;
537 
538 		/*
539 		 * Insert second 23:59:60 following second
540 		 * 23:59:59.
541 		 */
542 		case TIME_INS:
543 		if (!(time_status & STA_INS))
544 			time_state = TIME_OK;
545 		else if ((*newsec) % 86400 == 0) {
546 			(*newsec)--;
547 			time_state = TIME_OOP;
548 			time_tai++;
549 		}
550 		break;
551 
552 		/*
553 		 * Delete second 23:59:59.
554 		 */
555 		case TIME_DEL:
556 		if (!(time_status & STA_DEL))
557 			time_state = TIME_OK;
558 		else if (((*newsec) + 1) % 86400 == 0) {
559 			(*newsec)++;
560 			time_tai--;
561 			time_state = TIME_WAIT;
562 		}
563 		break;
564 
565 		/*
566 		 * Insert second in progress.
567 		 */
568 		case TIME_OOP:
569 			time_state = TIME_WAIT;
570 		break;
571 
572 		/*
573 		 * Wait for status bits to clear.
574 		 */
575 		case TIME_WAIT:
576 		if (!(time_status & (STA_INS | STA_DEL)))
577 			time_state = TIME_OK;
578 	}
579 
580 	/*
581 	 * Compute the total time adjustment for the next second
582 	 * in ns. The offset is reduced by a factor depending on
583 	 * whether the PPS signal is operating. Note that the
584 	 * value is in effect scaled by the clock frequency,
585 	 * since the adjustment is added at each tick interrupt.
586 	 */
587 	ftemp = time_offset;
588 #ifdef PPS_SYNC
589 	/* XXX even if PPS signal dies we should finish adjustment ? */
590 	if (time_status & STA_PPSTIME && time_status &
591 	    STA_PPSSIGNAL)
592 		L_RSHIFT(ftemp, pps_shift);
593 	else
594 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
595 #else
596 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
597 #endif /* PPS_SYNC */
598 	time_adj = ftemp;
599 	L_SUB(time_offset, ftemp);
600 	L_ADD(time_adj, time_freq);
601 
602 	/*
603 	 * Apply any correction from adjtime(2).  If more than one second
604 	 * off we slew at a rate of 5ms/s (5000 PPM) else 500us/s (500 PPM)
605 	 * until the last second is slewed the final < 500 usecs.
606 	 */
607 	if (time_adjtime != 0) {
608 		if (time_adjtime > 1000000)
609 			tickrate = 5000;
610 		else if (time_adjtime < -1000000)
611 			tickrate = -5000;
612 		else if (time_adjtime > 500)
613 			tickrate = 500;
614 		else if (time_adjtime < -500)
615 			tickrate = -500;
616 		else
617 			tickrate = time_adjtime;
618 		time_adjtime -= tickrate;
619 		L_LINT(ftemp, tickrate * 1000);
620 		L_ADD(time_adj, ftemp);
621 	}
622 	*adjustment = time_adj;
623 
624 #ifdef PPS_SYNC
625 	if (pps_valid > 0)
626 		pps_valid--;
627 	else
628 		time_status &= ~STA_PPSSIGNAL;
629 #endif /* PPS_SYNC */
630 
631 	NTP_UNLOCK();
632 }
633 
634 /*
635  * hardupdate() - local clock update
636  *
637  * This routine is called by ntp_adjtime() to update the local clock
638  * phase and frequency. The implementation is of an adaptive-parameter,
639  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
640  * time and frequency offset estimates for each call. If the kernel PPS
641  * discipline code is configured (PPS_SYNC), the PPS signal itself
642  * determines the new time offset, instead of the calling argument.
643  * Presumably, calls to ntp_adjtime() occur only when the caller
644  * believes the local clock is valid within some bound (+-128 ms with
645  * NTP). If the caller's time is far different than the PPS time, an
646  * argument will ensue, and it's not clear who will lose.
647  *
648  * For uncompensated quartz crystal oscillators and nominal update
649  * intervals less than 256 s, operation should be in phase-lock mode,
650  * where the loop is disciplined to phase. For update intervals greater
651  * than 1024 s, operation should be in frequency-lock mode, where the
652  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
653  * is selected by the STA_MODE status bit.
654  */
655 static void
656 hardupdate(long offset /* clock offset (ns) */)
657 {
658 	long mtemp;
659 	l_fp ftemp;
660 
661 	NTP_ASSERT_LOCKED();
662 
663 	/*
664 	 * Select how the phase is to be controlled and from which
665 	 * source. If the PPS signal is present and enabled to
666 	 * discipline the time, the PPS offset is used; otherwise, the
667 	 * argument offset is used.
668 	 */
669 	if (!(time_status & STA_PLL))
670 		return;
671 	if (!(time_status & STA_PPSTIME && time_status &
672 	    STA_PPSSIGNAL)) {
673 		if (offset > MAXPHASE)
674 			time_monitor = MAXPHASE;
675 		else if (offset < -MAXPHASE)
676 			time_monitor = -MAXPHASE;
677 		else
678 			time_monitor = offset;
679 		L_LINT(time_offset, time_monitor);
680 	}
681 
682 	/*
683 	 * Select how the frequency is to be controlled and in which
684 	 * mode (PLL or FLL). If the PPS signal is present and enabled
685 	 * to discipline the frequency, the PPS frequency is used;
686 	 * otherwise, the argument offset is used to compute it.
687 	 */
688 	if (time_status & STA_PPSFREQ && time_status & STA_PPSSIGNAL) {
689 		time_reftime = time_uptime;
690 		return;
691 	}
692 	if (time_status & STA_FREQHOLD || time_reftime == 0)
693 		time_reftime = time_uptime;
694 	mtemp = time_uptime - time_reftime;
695 	L_LINT(ftemp, time_monitor);
696 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
697 	L_MPY(ftemp, mtemp);
698 	L_ADD(time_freq, ftemp);
699 	time_status &= ~STA_MODE;
700 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp >
701 	    MAXSEC)) {
702 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
703 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
704 		L_ADD(time_freq, ftemp);
705 		time_status |= STA_MODE;
706 	}
707 	time_reftime = time_uptime;
708 	if (L_GINT(time_freq) > MAXFREQ)
709 		L_LINT(time_freq, MAXFREQ);
710 	else if (L_GINT(time_freq) < -MAXFREQ)
711 		L_LINT(time_freq, -MAXFREQ);
712 }
713 
714 #ifdef PPS_SYNC
715 /*
716  * hardpps() - discipline CPU clock oscillator to external PPS signal
717  *
718  * This routine is called at each PPS interrupt in order to discipline
719  * the CPU clock oscillator to the PPS signal. There are two independent
720  * first-order feedback loops, one for the phase, the other for the
721  * frequency. The phase loop measures and grooms the PPS phase offset
722  * and leaves it in a handy spot for the seconds overflow routine. The
723  * frequency loop averages successive PPS phase differences and
724  * calculates the PPS frequency offset, which is also processed by the
725  * seconds overflow routine. The code requires the caller to capture the
726  * time and architecture-dependent hardware counter values in
727  * nanoseconds at the on-time PPS signal transition.
728  *
729  * Note that, on some Unix systems this routine runs at an interrupt
730  * priority level higher than the timer interrupt routine hardclock().
731  * Therefore, the variables used are distinct from the hardclock()
732  * variables, except for the actual time and frequency variables, which
733  * are determined by this routine and updated atomically.
734  *
735  * tsp  - time at current PPS event
736  * delta_nsec - time elapsed between the previous and current PPS event
737  */
738 void
739 hardpps(struct timespec *tsp, long delta_nsec)
740 {
741 	long u_sec, u_nsec, v_nsec; /* temps */
742 	l_fp ftemp;
743 
744 	NTP_LOCK();
745 
746 	/*
747 	 * The signal is first processed by a range gate and frequency
748 	 * discriminator. The range gate rejects noise spikes outside
749 	 * the range +-500 us. The frequency discriminator rejects input
750 	 * signals with apparent frequency outside the range 1 +-500
751 	 * PPM. If two hits occur in the same second, we ignore the
752 	 * later hit; if not and a hit occurs outside the range gate,
753 	 * keep the later hit for later comparison, but do not process
754 	 * it.
755 	 */
756 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
757 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
758 	pps_valid = PPS_VALID;
759 	u_sec = tsp->tv_sec;
760 	u_nsec = tsp->tv_nsec;
761 	if (u_nsec >= (NANOSECOND >> 1)) {
762 		u_nsec -= NANOSECOND;
763 		u_sec++;
764 	}
765 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
766 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND - MAXFREQ)
767 		goto out;
768 	pps_tf[2] = pps_tf[1];
769 	pps_tf[1] = pps_tf[0];
770 	pps_tf[0].tv_sec = u_sec;
771 	pps_tf[0].tv_nsec = u_nsec;
772 
773 	/*
774 	 * Update the frequency accumulator using the difference between the
775 	 * current and previous PPS event measured directly by the timecounter.
776 	 */
777 	pps_fcount += delta_nsec - NANOSECOND;
778 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
779 		goto out;
780 	time_status &= ~STA_PPSJITTER;
781 
782 	/*
783 	 * A three-stage median filter is used to help denoise the PPS
784 	 * time. The median sample becomes the time offset estimate; the
785 	 * difference between the other two samples becomes the time
786 	 * dispersion (jitter) estimate.
787 	 */
788 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
789 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
790 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
791 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
792 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
793 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
794 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
795 		} else {
796 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
797 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
798 		}
799 	} else {
800 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
801 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
802 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
803 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
804 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
805 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
806 		} else {
807 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
808 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
809 		}
810 	}
811 
812 	/*
813 	 * Nominal jitter is due to PPS signal noise and interrupt
814 	 * latency. If it exceeds the popcorn threshold, the sample is
815 	 * discarded. otherwise, if so enabled, the time offset is
816 	 * updated. We can tolerate a modest loss of data here without
817 	 * much degrading time accuracy.
818 	 *
819 	 * The measurements being checked here were made with the system
820 	 * timecounter, so the popcorn threshold is not allowed to fall below
821 	 * the number of nanoseconds in two ticks of the timecounter.  For a
822 	 * timecounter running faster than 1 GHz the lower bound is 2ns, just
823 	 * to avoid a nonsensical threshold of zero.
824 	*/
825 	if (u_nsec > lmax(pps_jitter << PPS_POPCORN,
826 	    2 * (NANOSECOND / (long)qmin(NANOSECOND, tc_getfrequency())))) {
827 		time_status |= STA_PPSJITTER;
828 		pps_jitcnt++;
829 	} else if (time_status & STA_PPSTIME) {
830 		time_monitor = -v_nsec;
831 		L_LINT(time_offset, time_monitor);
832 	}
833 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
834 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
835 	if (u_sec < (1 << pps_shift))
836 		goto out;
837 
838 	/*
839 	 * At the end of the calibration interval the difference between
840 	 * the first and last counter values becomes the scaled
841 	 * frequency. It will later be divided by the length of the
842 	 * interval to determine the frequency update. If the frequency
843 	 * exceeds a sanity threshold, or if the actual calibration
844 	 * interval is not equal to the expected length, the data are
845 	 * discarded. We can tolerate a modest loss of data here without
846 	 * much degrading frequency accuracy.
847 	 */
848 	pps_calcnt++;
849 	v_nsec = -pps_fcount;
850 	pps_lastsec = pps_tf[0].tv_sec;
851 	pps_fcount = 0;
852 	u_nsec = MAXFREQ << pps_shift;
853 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 << pps_shift)) {
854 		time_status |= STA_PPSERROR;
855 		pps_errcnt++;
856 		goto out;
857 	}
858 
859 	/*
860 	 * Here the raw frequency offset and wander (stability) is
861 	 * calculated. If the wander is less than the wander threshold
862 	 * for four consecutive averaging intervals, the interval is
863 	 * doubled; if it is greater than the threshold for four
864 	 * consecutive intervals, the interval is halved. The scaled
865 	 * frequency offset is converted to frequency offset. The
866 	 * stability metric is calculated as the average of recent
867 	 * frequency changes, but is used only for performance
868 	 * monitoring.
869 	 */
870 	L_LINT(ftemp, v_nsec);
871 	L_RSHIFT(ftemp, pps_shift);
872 	L_SUB(ftemp, pps_freq);
873 	u_nsec = L_GINT(ftemp);
874 	if (u_nsec > PPS_MAXWANDER) {
875 		L_LINT(ftemp, PPS_MAXWANDER);
876 		pps_intcnt--;
877 		time_status |= STA_PPSWANDER;
878 		pps_stbcnt++;
879 	} else if (u_nsec < -PPS_MAXWANDER) {
880 		L_LINT(ftemp, -PPS_MAXWANDER);
881 		pps_intcnt--;
882 		time_status |= STA_PPSWANDER;
883 		pps_stbcnt++;
884 	} else {
885 		pps_intcnt++;
886 	}
887 	if (pps_intcnt >= 4) {
888 		pps_intcnt = 4;
889 		if (pps_shift < pps_shiftmax) {
890 			pps_shift++;
891 			pps_intcnt = 0;
892 		}
893 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
894 		pps_intcnt = -4;
895 		if (pps_shift > PPS_FAVG) {
896 			pps_shift--;
897 			pps_intcnt = 0;
898 		}
899 	}
900 	if (u_nsec < 0)
901 		u_nsec = -u_nsec;
902 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
903 
904 	/*
905 	 * The PPS frequency is recalculated and clamped to the maximum
906 	 * MAXFREQ. If enabled, the system clock frequency is updated as
907 	 * well.
908 	 */
909 	L_ADD(pps_freq, ftemp);
910 	u_nsec = L_GINT(pps_freq);
911 	if (u_nsec > MAXFREQ)
912 		L_LINT(pps_freq, MAXFREQ);
913 	else if (u_nsec < -MAXFREQ)
914 		L_LINT(pps_freq, -MAXFREQ);
915 	if (time_status & STA_PPSFREQ)
916 		time_freq = pps_freq;
917 
918 out:
919 	NTP_UNLOCK();
920 }
921 #endif /* PPS_SYNC */
922 
923 #ifndef _SYS_SYSPROTO_H_
924 struct adjtime_args {
925 	struct timeval *delta;
926 	struct timeval *olddelta;
927 };
928 #endif
929 /* ARGSUSED */
930 int
931 sys_adjtime(struct thread *td, struct adjtime_args *uap)
932 {
933 	struct timeval delta, olddelta, *deltap;
934 	int error;
935 
936 	if (uap->delta) {
937 		error = copyin(uap->delta, &delta, sizeof(delta));
938 		if (error)
939 			return (error);
940 		deltap = &delta;
941 	} else
942 		deltap = NULL;
943 	error = kern_adjtime(td, deltap, &olddelta);
944 	if (uap->olddelta && error == 0)
945 		error = copyout(&olddelta, uap->olddelta, sizeof(olddelta));
946 	return (error);
947 }
948 
949 int
950 kern_adjtime(struct thread *td, struct timeval *delta, struct timeval *olddelta)
951 {
952 	struct timeval atv;
953 	int64_t ltr, ltw;
954 	int error;
955 
956 	if (delta != NULL) {
957 		error = priv_check(td, PRIV_ADJTIME);
958 		if (error != 0)
959 			return (error);
960 		ltw = (int64_t)delta->tv_sec * 1000000 + delta->tv_usec;
961 	}
962 	NTP_LOCK();
963 	ltr = time_adjtime;
964 	if (delta != NULL)
965 		time_adjtime = ltw;
966 	NTP_UNLOCK();
967 	if (olddelta != NULL) {
968 		atv.tv_sec = ltr / 1000000;
969 		atv.tv_usec = ltr % 1000000;
970 		if (atv.tv_usec < 0) {
971 			atv.tv_usec += 1000000;
972 			atv.tv_sec--;
973 		}
974 		*olddelta = atv;
975 	}
976 	return (0);
977 }
978 
979 static struct callout resettodr_callout;
980 static int resettodr_period = 1800;
981 
982 static void
983 periodic_resettodr(void *arg __unused)
984 {
985 
986 	/*
987 	 * Read of time_status is lock-less, which is fine since
988 	 * ntp_is_time_error() operates on the consistent read value.
989 	 */
990 	if (!ntp_is_time_error(time_status))
991 		resettodr();
992 	if (resettodr_period > 0)
993 		callout_schedule(&resettodr_callout, resettodr_period * hz);
994 }
995 
996 static void
997 shutdown_resettodr(void *arg __unused, int howto __unused)
998 {
999 
1000 	callout_drain(&resettodr_callout);
1001 	/* Another unlocked read of time_status */
1002 	if (resettodr_period > 0 && !ntp_is_time_error(time_status))
1003 		resettodr();
1004 }
1005 
1006 static int
1007 sysctl_resettodr_period(SYSCTL_HANDLER_ARGS)
1008 {
1009 	int error;
1010 
1011 	error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
1012 	if (error || !req->newptr)
1013 		return (error);
1014 	if (cold)
1015 		goto done;
1016 	if (resettodr_period == 0)
1017 		callout_stop(&resettodr_callout);
1018 	else
1019 		callout_reset(&resettodr_callout, resettodr_period * hz,
1020 		    periodic_resettodr, NULL);
1021 done:
1022 	return (0);
1023 }
1024 
1025 SYSCTL_PROC(_machdep, OID_AUTO, rtc_save_period, CTLTYPE_INT | CTLFLAG_RWTUN |
1026     CTLFLAG_MPSAFE, &resettodr_period, 1800, sysctl_resettodr_period, "I",
1027     "Save system time to RTC with this period (in seconds)");
1028 
1029 static void
1030 start_periodic_resettodr(void *arg __unused)
1031 {
1032 
1033 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_resettodr, NULL,
1034 	    SHUTDOWN_PRI_FIRST);
1035 	callout_init(&resettodr_callout, 1);
1036 	if (resettodr_period == 0)
1037 		return;
1038 	callout_reset(&resettodr_callout, resettodr_period * hz,
1039 	    periodic_resettodr, NULL);
1040 }
1041 
1042 SYSINIT(periodic_resettodr, SI_SUB_LAST, SI_ORDER_MIDDLE,
1043 	start_periodic_resettodr, NULL);
1044