xref: /freebsd/sys/kern/kern_ntptime.c (revision 02e9120893770924227138ba49df1edb3896112a)
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 #include "opt_ntp.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/sysproto.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/priv.h>
42 #include <sys/proc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/time.h>
46 #include <sys/timex.h>
47 #include <sys/timetc.h>
48 #include <sys/timepps.h>
49 #include <sys/syscallsubr.h>
50 #include <sys/sysctl.h>
51 
52 #ifdef PPS_SYNC
53 FEATURE(pps_sync, "Support usage of external PPS signal by kernel PLL");
54 #endif
55 
56 /*
57  * Single-precision macros for 64-bit machines
58  */
59 typedef int64_t l_fp;
60 #define L_ADD(v, u)	((v) += (u))
61 #define L_SUB(v, u)	((v) -= (u))
62 #define L_ADDHI(v, a)	((v) += (int64_t)(a) << 32)
63 #define L_NEG(v)	((v) = -(v))
64 #define L_RSHIFT(v, n) \
65 	do { \
66 		if ((v) < 0) \
67 			(v) = -(-(v) >> (n)); \
68 		else \
69 			(v) = (v) >> (n); \
70 	} while (0)
71 #define L_MPY(v, a)	((v) *= (a))
72 #define L_CLR(v)	((v) = 0)
73 #define L_ISNEG(v)	((v) < 0)
74 #define L_LINT(v, a) \
75 	do { \
76 		if ((a) < 0) \
77 			((v) = -((int64_t)(-(a)) << 32)); \
78 		else \
79 			((v) = (int64_t)(a) << 32); \
80 	} while (0)
81 #define L_GINT(v)	((v) < 0 ? -(-(v) >> 32) : (v) >> 32)
82 
83 /*
84  * Generic NTP kernel interface
85  *
86  * These routines constitute the Network Time Protocol (NTP) interfaces
87  * for user and daemon application programs. The ntp_gettime() routine
88  * provides the time, maximum error (synch distance) and estimated error
89  * (dispersion) to client user application programs. The ntp_adjtime()
90  * routine is used by the NTP daemon to adjust the system clock to an
91  * externally derived time. The time offset and related variables set by
92  * this routine are used by other routines in this module to adjust the
93  * phase and frequency of the clock discipline loop which controls the
94  * system clock.
95  *
96  * When the kernel time is reckoned directly in nanoseconds (NTP_NANO
97  * defined), the time at each tick interrupt is derived directly from
98  * the kernel time variable. When the kernel time is reckoned in
99  * microseconds, (NTP_NANO undefined), the time is derived from the
100  * kernel time variable together with a variable representing the
101  * leftover nanoseconds at the last tick interrupt. In either case, the
102  * current nanosecond time is reckoned from these values plus an
103  * interpolated value derived by the clock routines in another
104  * architecture-specific module. The interpolation can use either a
105  * dedicated counter or a processor cycle counter (PCC) implemented in
106  * some architectures.
107  *
108  * Note that all routines must run at priority splclock or higher.
109  */
110 /*
111  * Phase/frequency-lock loop (PLL/FLL) definitions
112  *
113  * The nanosecond clock discipline uses two variable types, time
114  * variables and frequency variables. Both types are represented as 64-
115  * bit fixed-point quantities with the decimal point between two 32-bit
116  * halves. On a 32-bit machine, each half is represented as a single
117  * word and mathematical operations are done using multiple-precision
118  * arithmetic. On a 64-bit machine, ordinary computer arithmetic is
119  * used.
120  *
121  * A time variable is a signed 64-bit fixed-point number in ns and
122  * fraction. It represents the remaining time offset to be amortized
123  * over succeeding tick interrupts. The maximum time offset is about
124  * 0.5 s and the resolution is about 2.3e-10 ns.
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|			 ns				   |
130  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
131  * |			    fraction				   |
132  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
133  *
134  * A frequency variable is a signed 64-bit fixed-point number in ns/s
135  * and fraction. It represents the ns and fraction to be added to the
136  * kernel time variable at each second. The maximum frequency offset is
137  * about +-500000 ns/s and the resolution is about 2.3e-10 ns/s.
138  *
139  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
140  *  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
141  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
142  * |s s s s s s s s s s s s s|	          ns/s			   |
143  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
144  * |			    fraction				   |
145  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
146  */
147 /*
148  * The following variables establish the state of the PLL/FLL and the
149  * residual time and frequency offset of the local clock.
150  */
151 #define SHIFT_PLL	4		/* PLL loop gain (shift) */
152 #define SHIFT_FLL	2		/* FLL loop gain (shift) */
153 
154 static int time_state = TIME_OK;	/* clock state */
155 int time_status = STA_UNSYNC;	/* clock status bits */
156 static long time_tai;			/* TAI offset (s) */
157 static long time_monitor;		/* last time offset scaled (ns) */
158 static long time_constant;		/* poll interval (shift) (s) */
159 static long time_precision = 1;		/* clock precision (ns) */
160 static long time_maxerror = MAXPHASE / 1000; /* maximum error (us) */
161 long time_esterror = MAXPHASE / 1000; /* estimated error (us) */
162 static long time_reftime;		/* uptime at last adjustment (s) */
163 static l_fp time_offset;		/* time offset (ns) */
164 static l_fp time_freq;			/* frequency offset (ns/s) */
165 static l_fp time_adj;			/* tick adjust (ns/s) */
166 
167 static int64_t time_adjtime;		/* correction from adjtime(2) (usec) */
168 
169 static struct mtx ntp_lock;
170 MTX_SYSINIT(ntp, &ntp_lock, "ntp", MTX_SPIN);
171 
172 #define	NTP_LOCK()		mtx_lock_spin(&ntp_lock)
173 #define	NTP_UNLOCK()		mtx_unlock_spin(&ntp_lock)
174 #define	NTP_ASSERT_LOCKED()	mtx_assert(&ntp_lock, MA_OWNED)
175 
176 #ifdef PPS_SYNC
177 /*
178  * The following variables are used when a pulse-per-second (PPS) signal
179  * is available and connected via a modem control lead. They establish
180  * the engineering parameters of the clock discipline loop when
181  * controlled by the PPS signal.
182  */
183 #define PPS_FAVG	2		/* min freq avg interval (s) (shift) */
184 #define PPS_FAVGDEF	8		/* default freq avg int (s) (shift) */
185 #define PPS_FAVGMAX	15		/* max freq avg interval (s) (shift) */
186 #define PPS_PAVG	4		/* phase avg interval (s) (shift) */
187 #define PPS_VALID	120		/* PPS signal watchdog max (s) */
188 #define PPS_MAXWANDER	100000		/* max PPS wander (ns/s) */
189 #define PPS_POPCORN	2		/* popcorn spike threshold (shift) */
190 
191 static struct timespec pps_tf[3];	/* phase median filter */
192 static l_fp pps_freq;			/* scaled frequency offset (ns/s) */
193 static long pps_fcount;			/* frequency accumulator */
194 static long pps_jitter;			/* nominal jitter (ns) */
195 static long pps_stabil;			/* nominal stability (scaled ns/s) */
196 static long pps_lastsec;		/* time at last calibration (s) */
197 static int pps_valid;			/* signal watchdog counter */
198 static int pps_shift = PPS_FAVG;	/* interval duration (s) (shift) */
199 static int pps_shiftmax = PPS_FAVGDEF;	/* max interval duration (s) (shift) */
200 static int pps_intcnt;			/* wander counter */
201 
202 /*
203  * PPS signal quality monitors
204  */
205 static long pps_calcnt;			/* calibration intervals */
206 static long pps_jitcnt;			/* jitter limit exceeded */
207 static long pps_stbcnt;			/* stability limit exceeded */
208 static long pps_errcnt;			/* calibration errors */
209 #endif /* PPS_SYNC */
210 /*
211  * End of phase/frequency-lock loop (PLL/FLL) definitions
212  */
213 
214 static void hardupdate(long offset);
215 static void ntp_gettime1(struct ntptimeval *ntvp);
216 static bool ntp_is_time_error(int tsl);
217 
218 static bool
219 ntp_is_time_error(int tsl)
220 {
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 ((tsl & (STA_UNSYNC | STA_CLOCKERR)) ||
231 
232 	/*
233 	 * PPS signal lost when either time or frequency synchronization
234 	 * requested
235 	 */
236 	    (tsl & (STA_PPSFREQ | STA_PPSTIME) &&
237 	    !(tsl & STA_PPSSIGNAL)) ||
238 
239 	/*
240 	 * PPS jitter exceeded when time synchronization requested
241 	 */
242 	    (tsl & STA_PPSTIME && tsl & STA_PPSJITTER) ||
243 
244 	/*
245 	 * PPS wander exceeded or calibration error when frequency
246 	 * synchronization requested
247 	 */
248 	    (tsl & STA_PPSFREQ &&
249 	    tsl & (STA_PPSWANDER | STA_PPSERROR)))
250 		return (true);
251 
252 	return (false);
253 }
254 
255 static void
256 ntp_gettime1(struct ntptimeval *ntvp)
257 {
258 	struct timespec atv;	/* nanosecond time */
259 
260 	NTP_ASSERT_LOCKED();
261 
262 	nanotime(&atv);
263 	ntvp->time.tv_sec = atv.tv_sec;
264 	ntvp->time.tv_nsec = atv.tv_nsec;
265 	ntvp->maxerror = time_maxerror;
266 	ntvp->esterror = time_esterror;
267 	ntvp->tai = time_tai;
268 	ntvp->time_state = time_state;
269 
270 	if (ntp_is_time_error(time_status))
271 		ntvp->time_state = TIME_ERROR;
272 }
273 
274 /*
275  * ntp_gettime() - NTP user application interface
276  *
277  * See the timex.h header file for synopsis and API description.  Note that
278  * the TAI offset is returned in the ntvtimeval.tai structure member.
279  */
280 #ifndef _SYS_SYSPROTO_H_
281 struct ntp_gettime_args {
282 	struct ntptimeval *ntvp;
283 };
284 #endif
285 /* ARGSUSED */
286 int
287 sys_ntp_gettime(struct thread *td, struct ntp_gettime_args *uap)
288 {
289 	struct ntptimeval ntv;
290 
291 	memset(&ntv, 0, sizeof(ntv));
292 
293 	NTP_LOCK();
294 	ntp_gettime1(&ntv);
295 	NTP_UNLOCK();
296 
297 	td->td_retval[0] = ntv.time_state;
298 	return (copyout(&ntv, uap->ntvp, sizeof(ntv)));
299 }
300 
301 static int
302 ntp_sysctl(SYSCTL_HANDLER_ARGS)
303 {
304 	struct ntptimeval ntv;	/* temporary structure */
305 
306 	memset(&ntv, 0, sizeof(ntv));
307 
308 	NTP_LOCK();
309 	ntp_gettime1(&ntv);
310 	NTP_UNLOCK();
311 
312 	return (sysctl_handle_opaque(oidp, &ntv, sizeof(ntv), req));
313 }
314 
315 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
316     "");
317 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE | CTLFLAG_RD |
318     CTLFLAG_MPSAFE, 0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval",
319     "");
320 
321 #ifdef PPS_SYNC
322 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW,
323     &pps_shiftmax, 0, "Max interval duration (sec) (shift)");
324 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW,
325     &pps_shift, 0, "Interval duration (sec) (shift)");
326 SYSCTL_LONG(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD,
327     &time_monitor, 0, "Last time offset scaled (ns)");
328 
329 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
330     &pps_freq, 0,
331     "Scaled frequency offset (ns/sec)");
332 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
333     &time_freq, 0,
334     "Frequency offset (ns/sec)");
335 #endif
336 
337 /*
338  * ntp_adjtime() - NTP daemon application interface
339  *
340  * See the timex.h header file for synopsis and API description.  Note that
341  * the timex.constant structure member has a dual purpose to set the time
342  * constant and to set the TAI offset.
343  */
344 int
345 kern_ntp_adjtime(struct thread *td, struct timex *ntv, int *retvalp)
346 {
347 	long freq;		/* frequency ns/s) */
348 	int modes;		/* mode bits from structure */
349 	int error, retval;
350 
351 	/*
352 	 * Update selected clock variables - only the superuser can
353 	 * change anything. Note that there is no error checking here on
354 	 * the assumption the superuser should know what it is doing.
355 	 * Note that either the time constant or TAI offset are loaded
356 	 * from the ntv.constant member, depending on the mode bits. If
357 	 * the STA_PLL bit in the status word is cleared, the state and
358 	 * status words are reset to the initial values at boot.
359 	 */
360 	modes = ntv->modes;
361 	error = 0;
362 	if (modes)
363 		error = priv_check(td, PRIV_NTP_ADJTIME);
364 	if (error != 0)
365 		return (error);
366 	NTP_LOCK();
367 	if (modes & MOD_MAXERROR)
368 		time_maxerror = ntv->maxerror;
369 	if (modes & MOD_ESTERROR)
370 		time_esterror = ntv->esterror;
371 	if (modes & MOD_STATUS) {
372 		if (time_status & STA_PLL && !(ntv->status & STA_PLL)) {
373 			time_state = TIME_OK;
374 			time_status = STA_UNSYNC;
375 #ifdef PPS_SYNC
376 			pps_shift = PPS_FAVG;
377 #endif /* PPS_SYNC */
378 		}
379 		time_status &= STA_RONLY;
380 		time_status |= ntv->status & ~STA_RONLY;
381 	}
382 	if (modes & MOD_TIMECONST) {
383 		if (ntv->constant < 0)
384 			time_constant = 0;
385 		else if (ntv->constant > MAXTC)
386 			time_constant = MAXTC;
387 		else
388 			time_constant = ntv->constant;
389 	}
390 	if (modes & MOD_TAI) {
391 		if (ntv->constant > 0) /* XXX zero & negative numbers ? */
392 			time_tai = ntv->constant;
393 	}
394 #ifdef PPS_SYNC
395 	if (modes & MOD_PPSMAX) {
396 		if (ntv->shift < PPS_FAVG)
397 			pps_shiftmax = PPS_FAVG;
398 		else if (ntv->shift > PPS_FAVGMAX)
399 			pps_shiftmax = PPS_FAVGMAX;
400 		else
401 			pps_shiftmax = ntv->shift;
402 	}
403 #endif /* PPS_SYNC */
404 	if (modes & MOD_NANO)
405 		time_status |= STA_NANO;
406 	if (modes & MOD_MICRO)
407 		time_status &= ~STA_NANO;
408 	if (modes & MOD_CLKB)
409 		time_status |= STA_CLK;
410 	if (modes & MOD_CLKA)
411 		time_status &= ~STA_CLK;
412 	if (modes & MOD_FREQUENCY) {
413 		freq = (ntv->freq * 1000LL) >> 16;
414 		if (freq > MAXFREQ)
415 			L_LINT(time_freq, MAXFREQ);
416 		else if (freq < -MAXFREQ)
417 			L_LINT(time_freq, -MAXFREQ);
418 		else {
419 			/*
420 			 * ntv->freq is [PPM * 2^16] = [us/s * 2^16]
421 			 * time_freq is [ns/s * 2^32]
422 			 */
423 			time_freq = ntv->freq * 1000LL * 65536LL;
424 		}
425 #ifdef PPS_SYNC
426 		pps_freq = time_freq;
427 #endif /* PPS_SYNC */
428 	}
429 	if (modes & MOD_OFFSET) {
430 		if (time_status & STA_NANO)
431 			hardupdate(ntv->offset);
432 		else
433 			hardupdate(ntv->offset * 1000);
434 	}
435 
436 	/*
437 	 * Retrieve all clock variables. Note that the TAI offset is
438 	 * returned only by ntp_gettime();
439 	 */
440 	if (time_status & STA_NANO)
441 		ntv->offset = L_GINT(time_offset);
442 	else
443 		ntv->offset = L_GINT(time_offset) / 1000; /* XXX rounding ? */
444 	ntv->freq = L_GINT((time_freq / 1000LL) << 16);
445 	ntv->maxerror = time_maxerror;
446 	ntv->esterror = time_esterror;
447 	ntv->status = time_status;
448 	ntv->constant = time_constant;
449 	if (time_status & STA_NANO)
450 		ntv->precision = time_precision;
451 	else
452 		ntv->precision = time_precision / 1000;
453 	ntv->tolerance = MAXFREQ * SCALE_PPM;
454 #ifdef PPS_SYNC
455 	ntv->shift = pps_shift;
456 	ntv->ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
457 	if (time_status & STA_NANO)
458 		ntv->jitter = pps_jitter;
459 	else
460 		ntv->jitter = pps_jitter / 1000;
461 	ntv->stabil = pps_stabil;
462 	ntv->calcnt = pps_calcnt;
463 	ntv->errcnt = pps_errcnt;
464 	ntv->jitcnt = pps_jitcnt;
465 	ntv->stbcnt = pps_stbcnt;
466 #endif /* PPS_SYNC */
467 	retval = ntp_is_time_error(time_status) ? TIME_ERROR : time_state;
468 	NTP_UNLOCK();
469 
470 	*retvalp = retval;
471 	return (0);
472 }
473 
474 #ifndef _SYS_SYSPROTO_H_
475 struct ntp_adjtime_args {
476 	struct timex *tp;
477 };
478 #endif
479 
480 int
481 sys_ntp_adjtime(struct thread *td, struct ntp_adjtime_args *uap)
482 {
483 	struct timex ntv;
484 	int error, retval;
485 
486 	error = copyin(uap->tp, &ntv, sizeof(ntv));
487 	if (error == 0) {
488 		error = kern_ntp_adjtime(td, &ntv, &retval);
489 		if (error == 0) {
490 			error = copyout(&ntv, uap->tp, sizeof(ntv));
491 			if (error == 0)
492 				td->td_retval[0] = retval;
493 		}
494 	}
495 	return (error);
496 }
497 
498 /*
499  * second_overflow() - called after ntp_tick_adjust()
500  *
501  * This routine is ordinarily called immediately following the above
502  * routine ntp_tick_adjust(). While these two routines are normally
503  * combined, they are separated here only for the purposes of
504  * simulation.
505  */
506 void
507 ntp_update_second(int64_t *adjustment, time_t *newsec)
508 {
509 	int tickrate;
510 	l_fp ftemp;		/* 32/64-bit temporary */
511 
512 	NTP_LOCK();
513 
514 	/*
515 	 * On rollover of the second both the nanosecond and microsecond
516 	 * clocks are updated and the state machine cranked as
517 	 * necessary. The phase adjustment to be used for the next
518 	 * second is calculated and the maximum error is increased by
519 	 * the tolerance.
520 	 */
521 	time_maxerror += MAXFREQ / 1000;
522 
523 	/*
524 	 * Leap second processing. If in leap-insert state at
525 	 * the end of the day, the system clock is set back one
526 	 * second; if in leap-delete state, the system clock is
527 	 * set ahead one second. The nano_time() routine or
528 	 * external clock driver will insure that reported time
529 	 * is always monotonic.
530 	 */
531 	switch (time_state) {
532 		/*
533 		 * No warning.
534 		 */
535 		case TIME_OK:
536 		if (time_status & STA_INS)
537 			time_state = TIME_INS;
538 		else if (time_status & STA_DEL)
539 			time_state = TIME_DEL;
540 		break;
541 
542 		/*
543 		 * Insert second 23:59:60 following second
544 		 * 23:59:59.
545 		 */
546 		case TIME_INS:
547 		if (!(time_status & STA_INS))
548 			time_state = TIME_OK;
549 		else if ((*newsec) % 86400 == 0) {
550 			(*newsec)--;
551 			time_state = TIME_OOP;
552 			time_tai++;
553 		}
554 		break;
555 
556 		/*
557 		 * Delete second 23:59:59.
558 		 */
559 		case TIME_DEL:
560 		if (!(time_status & STA_DEL))
561 			time_state = TIME_OK;
562 		else if (((*newsec) + 1) % 86400 == 0) {
563 			(*newsec)++;
564 			time_tai--;
565 			time_state = TIME_WAIT;
566 		}
567 		break;
568 
569 		/*
570 		 * Insert second in progress.
571 		 */
572 		case TIME_OOP:
573 			time_state = TIME_WAIT;
574 		break;
575 
576 		/*
577 		 * Wait for status bits to clear.
578 		 */
579 		case TIME_WAIT:
580 		if (!(time_status & (STA_INS | STA_DEL)))
581 			time_state = TIME_OK;
582 	}
583 
584 	/*
585 	 * Compute the total time adjustment for the next second
586 	 * in ns. The offset is reduced by a factor depending on
587 	 * whether the PPS signal is operating. Note that the
588 	 * value is in effect scaled by the clock frequency,
589 	 * since the adjustment is added at each tick interrupt.
590 	 */
591 	ftemp = time_offset;
592 #ifdef PPS_SYNC
593 	/* XXX even if PPS signal dies we should finish adjustment ? */
594 	if (time_status & STA_PPSTIME && time_status &
595 	    STA_PPSSIGNAL)
596 		L_RSHIFT(ftemp, pps_shift);
597 	else
598 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
599 #else
600 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
601 #endif /* PPS_SYNC */
602 	time_adj = ftemp;
603 	L_SUB(time_offset, ftemp);
604 	L_ADD(time_adj, time_freq);
605 
606 	/*
607 	 * Apply any correction from adjtime(2).  If more than one second
608 	 * off we slew at a rate of 5ms/s (5000 PPM) else 500us/s (500 PPM)
609 	 * until the last second is slewed the final < 500 usecs.
610 	 */
611 	if (time_adjtime != 0) {
612 		if (time_adjtime > 1000000)
613 			tickrate = 5000;
614 		else if (time_adjtime < -1000000)
615 			tickrate = -5000;
616 		else if (time_adjtime > 500)
617 			tickrate = 500;
618 		else if (time_adjtime < -500)
619 			tickrate = -500;
620 		else
621 			tickrate = time_adjtime;
622 		time_adjtime -= tickrate;
623 		L_LINT(ftemp, tickrate * 1000);
624 		L_ADD(time_adj, ftemp);
625 	}
626 	*adjustment = time_adj;
627 
628 #ifdef PPS_SYNC
629 	if (pps_valid > 0)
630 		pps_valid--;
631 	else
632 		time_status &= ~STA_PPSSIGNAL;
633 #endif /* PPS_SYNC */
634 
635 	NTP_UNLOCK();
636 }
637 
638 /*
639  * hardupdate() - local clock update
640  *
641  * This routine is called by ntp_adjtime() to update the local clock
642  * phase and frequency. The implementation is of an adaptive-parameter,
643  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
644  * time and frequency offset estimates for each call. If the kernel PPS
645  * discipline code is configured (PPS_SYNC), the PPS signal itself
646  * determines the new time offset, instead of the calling argument.
647  * Presumably, calls to ntp_adjtime() occur only when the caller
648  * believes the local clock is valid within some bound (+-128 ms with
649  * NTP). If the caller's time is far different than the PPS time, an
650  * argument will ensue, and it's not clear who will lose.
651  *
652  * For uncompensated quartz crystal oscillators and nominal update
653  * intervals less than 256 s, operation should be in phase-lock mode,
654  * where the loop is disciplined to phase. For update intervals greater
655  * than 1024 s, operation should be in frequency-lock mode, where the
656  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
657  * is selected by the STA_MODE status bit.
658  */
659 static void
660 hardupdate(long offset /* clock offset (ns) */)
661 {
662 	long mtemp;
663 	l_fp ftemp;
664 
665 	NTP_ASSERT_LOCKED();
666 
667 	/*
668 	 * Select how the phase is to be controlled and from which
669 	 * source. If the PPS signal is present and enabled to
670 	 * discipline the time, the PPS offset is used; otherwise, the
671 	 * argument offset is used.
672 	 */
673 	if (!(time_status & STA_PLL))
674 		return;
675 	if (!(time_status & STA_PPSTIME && time_status &
676 	    STA_PPSSIGNAL)) {
677 		if (offset > MAXPHASE)
678 			time_monitor = MAXPHASE;
679 		else if (offset < -MAXPHASE)
680 			time_monitor = -MAXPHASE;
681 		else
682 			time_monitor = offset;
683 		L_LINT(time_offset, time_monitor);
684 	}
685 
686 	/*
687 	 * Select how the frequency is to be controlled and in which
688 	 * mode (PLL or FLL). If the PPS signal is present and enabled
689 	 * to discipline the frequency, the PPS frequency is used;
690 	 * otherwise, the argument offset is used to compute it.
691 	 */
692 	if (time_status & STA_PPSFREQ && time_status & STA_PPSSIGNAL) {
693 		time_reftime = time_uptime;
694 		return;
695 	}
696 	if (time_status & STA_FREQHOLD || time_reftime == 0)
697 		time_reftime = time_uptime;
698 	mtemp = time_uptime - time_reftime;
699 	L_LINT(ftemp, time_monitor);
700 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
701 	L_MPY(ftemp, mtemp);
702 	L_ADD(time_freq, ftemp);
703 	time_status &= ~STA_MODE;
704 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp >
705 	    MAXSEC)) {
706 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
707 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
708 		L_ADD(time_freq, ftemp);
709 		time_status |= STA_MODE;
710 	}
711 	time_reftime = time_uptime;
712 	if (L_GINT(time_freq) > MAXFREQ)
713 		L_LINT(time_freq, MAXFREQ);
714 	else if (L_GINT(time_freq) < -MAXFREQ)
715 		L_LINT(time_freq, -MAXFREQ);
716 }
717 
718 #ifdef PPS_SYNC
719 /*
720  * hardpps() - discipline CPU clock oscillator to external PPS signal
721  *
722  * This routine is called at each PPS interrupt in order to discipline
723  * the CPU clock oscillator to the PPS signal. There are two independent
724  * first-order feedback loops, one for the phase, the other for the
725  * frequency. The phase loop measures and grooms the PPS phase offset
726  * and leaves it in a handy spot for the seconds overflow routine. The
727  * frequency loop averages successive PPS phase differences and
728  * calculates the PPS frequency offset, which is also processed by the
729  * seconds overflow routine. The code requires the caller to capture the
730  * time and architecture-dependent hardware counter values in
731  * nanoseconds at the on-time PPS signal transition.
732  *
733  * Note that, on some Unix systems this routine runs at an interrupt
734  * priority level higher than the timer interrupt routine hardclock().
735  * Therefore, the variables used are distinct from the hardclock()
736  * variables, except for the actual time and frequency variables, which
737  * are determined by this routine and updated atomically.
738  *
739  * tsp  - time at current PPS event
740  * delta_nsec - time elapsed between the previous and current PPS event
741  */
742 void
743 hardpps(struct timespec *tsp, long delta_nsec)
744 {
745 	long u_sec, u_nsec, v_nsec; /* temps */
746 	l_fp ftemp;
747 
748 	NTP_LOCK();
749 
750 	/*
751 	 * The signal is first processed by a range gate and frequency
752 	 * discriminator. The range gate rejects noise spikes outside
753 	 * the range +-500 us. The frequency discriminator rejects input
754 	 * signals with apparent frequency outside the range 1 +-500
755 	 * PPM. If two hits occur in the same second, we ignore the
756 	 * later hit; if not and a hit occurs outside the range gate,
757 	 * keep the later hit for later comparison, but do not process
758 	 * it.
759 	 */
760 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
761 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
762 	pps_valid = PPS_VALID;
763 	u_sec = tsp->tv_sec;
764 	u_nsec = tsp->tv_nsec;
765 	if (u_nsec >= (NANOSECOND >> 1)) {
766 		u_nsec -= NANOSECOND;
767 		u_sec++;
768 	}
769 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
770 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND - MAXFREQ)
771 		goto out;
772 	pps_tf[2] = pps_tf[1];
773 	pps_tf[1] = pps_tf[0];
774 	pps_tf[0].tv_sec = u_sec;
775 	pps_tf[0].tv_nsec = u_nsec;
776 
777 	/*
778 	 * Update the frequency accumulator using the difference between the
779 	 * current and previous PPS event measured directly by the timecounter.
780 	 */
781 	pps_fcount += delta_nsec - NANOSECOND;
782 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
783 		goto out;
784 	time_status &= ~STA_PPSJITTER;
785 
786 	/*
787 	 * A three-stage median filter is used to help denoise the PPS
788 	 * time. The median sample becomes the time offset estimate; the
789 	 * difference between the other two samples becomes the time
790 	 * dispersion (jitter) estimate.
791 	 */
792 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
793 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
794 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
795 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
796 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
797 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
798 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
799 		} else {
800 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
801 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
802 		}
803 	} else {
804 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
805 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
806 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
807 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
808 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
809 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
810 		} else {
811 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
812 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
813 		}
814 	}
815 
816 	/*
817 	 * Nominal jitter is due to PPS signal noise and interrupt
818 	 * latency. If it exceeds the popcorn threshold, the sample is
819 	 * discarded. otherwise, if so enabled, the time offset is
820 	 * updated. We can tolerate a modest loss of data here without
821 	 * much degrading time accuracy.
822 	 *
823 	 * The measurements being checked here were made with the system
824 	 * timecounter, so the popcorn threshold is not allowed to fall below
825 	 * the number of nanoseconds in two ticks of the timecounter.  For a
826 	 * timecounter running faster than 1 GHz the lower bound is 2ns, just
827 	 * to avoid a nonsensical threshold of zero.
828 	*/
829 	if (u_nsec > lmax(pps_jitter << PPS_POPCORN,
830 	    2 * (NANOSECOND / (long)qmin(NANOSECOND, tc_getfrequency())))) {
831 		time_status |= STA_PPSJITTER;
832 		pps_jitcnt++;
833 	} else if (time_status & STA_PPSTIME) {
834 		time_monitor = -v_nsec;
835 		L_LINT(time_offset, time_monitor);
836 	}
837 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
838 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
839 	if (u_sec < (1 << pps_shift))
840 		goto out;
841 
842 	/*
843 	 * At the end of the calibration interval the difference between
844 	 * the first and last counter values becomes the scaled
845 	 * frequency. It will later be divided by the length of the
846 	 * interval to determine the frequency update. If the frequency
847 	 * exceeds a sanity threshold, or if the actual calibration
848 	 * interval is not equal to the expected length, the data are
849 	 * discarded. We can tolerate a modest loss of data here without
850 	 * much degrading frequency accuracy.
851 	 */
852 	pps_calcnt++;
853 	v_nsec = -pps_fcount;
854 	pps_lastsec = pps_tf[0].tv_sec;
855 	pps_fcount = 0;
856 	u_nsec = MAXFREQ << pps_shift;
857 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 << pps_shift)) {
858 		time_status |= STA_PPSERROR;
859 		pps_errcnt++;
860 		goto out;
861 	}
862 
863 	/*
864 	 * Here the raw frequency offset and wander (stability) is
865 	 * calculated. If the wander is less than the wander threshold
866 	 * for four consecutive averaging intervals, the interval is
867 	 * doubled; if it is greater than the threshold for four
868 	 * consecutive intervals, the interval is halved. The scaled
869 	 * frequency offset is converted to frequency offset. The
870 	 * stability metric is calculated as the average of recent
871 	 * frequency changes, but is used only for performance
872 	 * monitoring.
873 	 */
874 	L_LINT(ftemp, v_nsec);
875 	L_RSHIFT(ftemp, pps_shift);
876 	L_SUB(ftemp, pps_freq);
877 	u_nsec = L_GINT(ftemp);
878 	if (u_nsec > PPS_MAXWANDER) {
879 		L_LINT(ftemp, PPS_MAXWANDER);
880 		pps_intcnt--;
881 		time_status |= STA_PPSWANDER;
882 		pps_stbcnt++;
883 	} else if (u_nsec < -PPS_MAXWANDER) {
884 		L_LINT(ftemp, -PPS_MAXWANDER);
885 		pps_intcnt--;
886 		time_status |= STA_PPSWANDER;
887 		pps_stbcnt++;
888 	} else {
889 		pps_intcnt++;
890 	}
891 	if (pps_intcnt >= 4) {
892 		pps_intcnt = 4;
893 		if (pps_shift < pps_shiftmax) {
894 			pps_shift++;
895 			pps_intcnt = 0;
896 		}
897 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
898 		pps_intcnt = -4;
899 		if (pps_shift > PPS_FAVG) {
900 			pps_shift--;
901 			pps_intcnt = 0;
902 		}
903 	}
904 	if (u_nsec < 0)
905 		u_nsec = -u_nsec;
906 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
907 
908 	/*
909 	 * The PPS frequency is recalculated and clamped to the maximum
910 	 * MAXFREQ. If enabled, the system clock frequency is updated as
911 	 * well.
912 	 */
913 	L_ADD(pps_freq, ftemp);
914 	u_nsec = L_GINT(pps_freq);
915 	if (u_nsec > MAXFREQ)
916 		L_LINT(pps_freq, MAXFREQ);
917 	else if (u_nsec < -MAXFREQ)
918 		L_LINT(pps_freq, -MAXFREQ);
919 	if (time_status & STA_PPSFREQ)
920 		time_freq = pps_freq;
921 
922 out:
923 	NTP_UNLOCK();
924 }
925 #endif /* PPS_SYNC */
926 
927 #ifndef _SYS_SYSPROTO_H_
928 struct adjtime_args {
929 	struct timeval *delta;
930 	struct timeval *olddelta;
931 };
932 #endif
933 /* ARGSUSED */
934 int
935 sys_adjtime(struct thread *td, struct adjtime_args *uap)
936 {
937 	struct timeval delta, olddelta, *deltap;
938 	int error;
939 
940 	if (uap->delta) {
941 		error = copyin(uap->delta, &delta, sizeof(delta));
942 		if (error)
943 			return (error);
944 		deltap = &delta;
945 	} else
946 		deltap = NULL;
947 	error = kern_adjtime(td, deltap, &olddelta);
948 	if (uap->olddelta && error == 0)
949 		error = copyout(&olddelta, uap->olddelta, sizeof(olddelta));
950 	return (error);
951 }
952 
953 int
954 kern_adjtime(struct thread *td, struct timeval *delta, struct timeval *olddelta)
955 {
956 	struct timeval atv;
957 	int64_t ltr, ltw;
958 	int error;
959 
960 	if (delta != NULL) {
961 		error = priv_check(td, PRIV_ADJTIME);
962 		if (error != 0)
963 			return (error);
964 		ltw = (int64_t)delta->tv_sec * 1000000 + delta->tv_usec;
965 	}
966 	NTP_LOCK();
967 	ltr = time_adjtime;
968 	if (delta != NULL)
969 		time_adjtime = ltw;
970 	NTP_UNLOCK();
971 	if (olddelta != NULL) {
972 		atv.tv_sec = ltr / 1000000;
973 		atv.tv_usec = ltr % 1000000;
974 		if (atv.tv_usec < 0) {
975 			atv.tv_usec += 1000000;
976 			atv.tv_sec--;
977 		}
978 		*olddelta = atv;
979 	}
980 	return (0);
981 }
982 
983 static struct callout resettodr_callout;
984 static int resettodr_period = 1800;
985 
986 static void
987 periodic_resettodr(void *arg __unused)
988 {
989 
990 	/*
991 	 * Read of time_status is lock-less, which is fine since
992 	 * ntp_is_time_error() operates on the consistent read value.
993 	 */
994 	if (!ntp_is_time_error(time_status))
995 		resettodr();
996 	if (resettodr_period > 0)
997 		callout_schedule(&resettodr_callout, resettodr_period * hz);
998 }
999 
1000 static void
1001 shutdown_resettodr(void *arg __unused, int howto __unused)
1002 {
1003 
1004 	callout_drain(&resettodr_callout);
1005 	/* Another unlocked read of time_status */
1006 	if (resettodr_period > 0 && !ntp_is_time_error(time_status))
1007 		resettodr();
1008 }
1009 
1010 static int
1011 sysctl_resettodr_period(SYSCTL_HANDLER_ARGS)
1012 {
1013 	int error;
1014 
1015 	error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
1016 	if (error || !req->newptr)
1017 		return (error);
1018 	if (cold)
1019 		goto done;
1020 	if (resettodr_period == 0)
1021 		callout_stop(&resettodr_callout);
1022 	else
1023 		callout_reset(&resettodr_callout, resettodr_period * hz,
1024 		    periodic_resettodr, NULL);
1025 done:
1026 	return (0);
1027 }
1028 
1029 SYSCTL_PROC(_machdep, OID_AUTO, rtc_save_period, CTLTYPE_INT | CTLFLAG_RWTUN |
1030     CTLFLAG_MPSAFE, &resettodr_period, 1800, sysctl_resettodr_period, "I",
1031     "Save system time to RTC with this period (in seconds)");
1032 
1033 static void
1034 start_periodic_resettodr(void *arg __unused)
1035 {
1036 
1037 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_resettodr, NULL,
1038 	    SHUTDOWN_PRI_FIRST);
1039 	callout_init(&resettodr_callout, 1);
1040 	if (resettodr_period == 0)
1041 		return;
1042 	callout_reset(&resettodr_callout, resettodr_period * hz,
1043 	    periodic_resettodr, NULL);
1044 }
1045 
1046 SYSINIT(periodic_resettodr, SI_SUB_LAST, SI_ORDER_MIDDLE,
1047 	start_periodic_resettodr, NULL);
1048