xref: /freebsd/sys/kern/kern_ntptime.c (revision b13788e396c2b24f88697e7d4a74bab429ef4d0c)
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 ntp_init(void);
211 static void hardupdate(long offset);
212 static void ntp_gettime1(struct ntptimeval *ntvp);
213 static bool ntp_is_time_error(int tsl);
214 
215 static bool
216 ntp_is_time_error(int tsl)
217 {
218 
219 	/*
220 	 * Status word error decode. If any of these conditions occur,
221 	 * an error is returned, instead of the status word. Most
222 	 * applications will care only about the fact the system clock
223 	 * may not be trusted, not about the details.
224 	 *
225 	 * Hardware or software error
226 	 */
227 	if ((tsl & (STA_UNSYNC | STA_CLOCKERR)) ||
228 
229 	/*
230 	 * PPS signal lost when either time or frequency synchronization
231 	 * requested
232 	 */
233 	    (tsl & (STA_PPSFREQ | STA_PPSTIME) &&
234 	    !(tsl & STA_PPSSIGNAL)) ||
235 
236 	/*
237 	 * PPS jitter exceeded when time synchronization requested
238 	 */
239 	    (tsl & STA_PPSTIME && tsl & STA_PPSJITTER) ||
240 
241 	/*
242 	 * PPS wander exceeded or calibration error when frequency
243 	 * synchronization requested
244 	 */
245 	    (tsl & STA_PPSFREQ &&
246 	    tsl & (STA_PPSWANDER | STA_PPSERROR)))
247 		return (true);
248 
249 	return (false);
250 }
251 
252 static void
253 ntp_gettime1(struct ntptimeval *ntvp)
254 {
255 	struct timespec atv;	/* nanosecond time */
256 
257 	NTP_ASSERT_LOCKED();
258 
259 	nanotime(&atv);
260 	ntvp->time.tv_sec = atv.tv_sec;
261 	ntvp->time.tv_nsec = atv.tv_nsec;
262 	ntvp->maxerror = time_maxerror;
263 	ntvp->esterror = time_esterror;
264 	ntvp->tai = time_tai;
265 	ntvp->time_state = time_state;
266 
267 	if (ntp_is_time_error(time_status))
268 		ntvp->time_state = TIME_ERROR;
269 }
270 
271 /*
272  * ntp_gettime() - NTP user application interface
273  *
274  * See the timex.h header file for synopsis and API description.  Note that
275  * the TAI offset is returned in the ntvtimeval.tai structure member.
276  */
277 #ifndef _SYS_SYSPROTO_H_
278 struct ntp_gettime_args {
279 	struct ntptimeval *ntvp;
280 };
281 #endif
282 /* ARGSUSED */
283 int
284 sys_ntp_gettime(struct thread *td, struct ntp_gettime_args *uap)
285 {
286 	struct ntptimeval ntv;
287 
288 	memset(&ntv, 0, sizeof(ntv));
289 
290 	NTP_LOCK();
291 	ntp_gettime1(&ntv);
292 	NTP_UNLOCK();
293 
294 	td->td_retval[0] = ntv.time_state;
295 	return (copyout(&ntv, uap->ntvp, sizeof(ntv)));
296 }
297 
298 static int
299 ntp_sysctl(SYSCTL_HANDLER_ARGS)
300 {
301 	struct ntptimeval ntv;	/* temporary structure */
302 
303 	memset(&ntv, 0, sizeof(ntv));
304 
305 	NTP_LOCK();
306 	ntp_gettime1(&ntv);
307 	NTP_UNLOCK();
308 
309 	return (sysctl_handle_opaque(oidp, &ntv, sizeof(ntv), req));
310 }
311 
312 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
313     "");
314 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE | CTLFLAG_RD |
315     CTLFLAG_MPSAFE, 0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval",
316     "");
317 
318 #ifdef PPS_SYNC
319 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW,
320     &pps_shiftmax, 0, "Max interval duration (sec) (shift)");
321 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW,
322     &pps_shift, 0, "Interval duration (sec) (shift)");
323 SYSCTL_LONG(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD,
324     &time_monitor, 0, "Last time offset scaled (ns)");
325 
326 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
327     &pps_freq, 0,
328     "Scaled frequency offset (ns/sec)");
329 SYSCTL_S64(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD | CTLFLAG_MPSAFE,
330     &time_freq, 0,
331     "Frequency offset (ns/sec)");
332 #endif
333 
334 /*
335  * ntp_adjtime() - NTP daemon application interface
336  *
337  * See the timex.h header file for synopsis and API description.  Note that
338  * the timex.constant structure member has a dual purpose to set the time
339  * constant and to set the TAI offset.
340  */
341 #ifndef _SYS_SYSPROTO_H_
342 struct ntp_adjtime_args {
343 	struct timex *tp;
344 };
345 #endif
346 
347 int
348 sys_ntp_adjtime(struct thread *td, struct ntp_adjtime_args *uap)
349 {
350 	struct timex ntv;	/* temporary structure */
351 	long freq;		/* frequency ns/s) */
352 	int modes;		/* mode bits from structure */
353 	int error, retval;
354 
355 	error = copyin((caddr_t)uap->tp, (caddr_t)&ntv, sizeof(ntv));
356 	if (error)
357 		return (error);
358 
359 	/*
360 	 * Update selected clock variables - only the superuser can
361 	 * change anything. Note that there is no error checking here on
362 	 * the assumption the superuser should know what it is doing.
363 	 * Note that either the time constant or TAI offset are loaded
364 	 * from the ntv.constant member, depending on the mode bits. If
365 	 * the STA_PLL bit in the status word is cleared, the state and
366 	 * status words are reset to the initial values at boot.
367 	 */
368 	modes = ntv.modes;
369 	if (modes)
370 		error = priv_check(td, PRIV_NTP_ADJTIME);
371 	if (error != 0)
372 		return (error);
373 	NTP_LOCK();
374 	if (modes & MOD_MAXERROR)
375 		time_maxerror = ntv.maxerror;
376 	if (modes & MOD_ESTERROR)
377 		time_esterror = ntv.esterror;
378 	if (modes & MOD_STATUS) {
379 		if (time_status & STA_PLL && !(ntv.status & STA_PLL)) {
380 			time_state = TIME_OK;
381 			time_status = STA_UNSYNC;
382 #ifdef PPS_SYNC
383 			pps_shift = PPS_FAVG;
384 #endif /* PPS_SYNC */
385 		}
386 		time_status &= STA_RONLY;
387 		time_status |= ntv.status & ~STA_RONLY;
388 	}
389 	if (modes & MOD_TIMECONST) {
390 		if (ntv.constant < 0)
391 			time_constant = 0;
392 		else if (ntv.constant > MAXTC)
393 			time_constant = MAXTC;
394 		else
395 			time_constant = ntv.constant;
396 	}
397 	if (modes & MOD_TAI) {
398 		if (ntv.constant > 0) /* XXX zero & negative numbers ? */
399 			time_tai = ntv.constant;
400 	}
401 #ifdef PPS_SYNC
402 	if (modes & MOD_PPSMAX) {
403 		if (ntv.shift < PPS_FAVG)
404 			pps_shiftmax = PPS_FAVG;
405 		else if (ntv.shift > PPS_FAVGMAX)
406 			pps_shiftmax = PPS_FAVGMAX;
407 		else
408 			pps_shiftmax = ntv.shift;
409 	}
410 #endif /* PPS_SYNC */
411 	if (modes & MOD_NANO)
412 		time_status |= STA_NANO;
413 	if (modes & MOD_MICRO)
414 		time_status &= ~STA_NANO;
415 	if (modes & MOD_CLKB)
416 		time_status |= STA_CLK;
417 	if (modes & MOD_CLKA)
418 		time_status &= ~STA_CLK;
419 	if (modes & MOD_FREQUENCY) {
420 		freq = (ntv.freq * 1000LL) >> 16;
421 		if (freq > MAXFREQ)
422 			L_LINT(time_freq, MAXFREQ);
423 		else if (freq < -MAXFREQ)
424 			L_LINT(time_freq, -MAXFREQ);
425 		else {
426 			/*
427 			 * ntv.freq is [PPM * 2^16] = [us/s * 2^16]
428 			 * time_freq is [ns/s * 2^32]
429 			 */
430 			time_freq = ntv.freq * 1000LL * 65536LL;
431 		}
432 #ifdef PPS_SYNC
433 		pps_freq = time_freq;
434 #endif /* PPS_SYNC */
435 	}
436 	if (modes & MOD_OFFSET) {
437 		if (time_status & STA_NANO)
438 			hardupdate(ntv.offset);
439 		else
440 			hardupdate(ntv.offset * 1000);
441 	}
442 
443 	/*
444 	 * Retrieve all clock variables. Note that the TAI offset is
445 	 * returned only by ntp_gettime();
446 	 */
447 	if (time_status & STA_NANO)
448 		ntv.offset = L_GINT(time_offset);
449 	else
450 		ntv.offset = L_GINT(time_offset) / 1000; /* XXX rounding ? */
451 	ntv.freq = L_GINT((time_freq / 1000LL) << 16);
452 	ntv.maxerror = time_maxerror;
453 	ntv.esterror = time_esterror;
454 	ntv.status = time_status;
455 	ntv.constant = time_constant;
456 	if (time_status & STA_NANO)
457 		ntv.precision = time_precision;
458 	else
459 		ntv.precision = time_precision / 1000;
460 	ntv.tolerance = MAXFREQ * SCALE_PPM;
461 #ifdef PPS_SYNC
462 	ntv.shift = pps_shift;
463 	ntv.ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
464 	if (time_status & STA_NANO)
465 		ntv.jitter = pps_jitter;
466 	else
467 		ntv.jitter = pps_jitter / 1000;
468 	ntv.stabil = pps_stabil;
469 	ntv.calcnt = pps_calcnt;
470 	ntv.errcnt = pps_errcnt;
471 	ntv.jitcnt = pps_jitcnt;
472 	ntv.stbcnt = pps_stbcnt;
473 #endif /* PPS_SYNC */
474 	retval = ntp_is_time_error(time_status) ? TIME_ERROR : time_state;
475 	NTP_UNLOCK();
476 
477 	error = copyout((caddr_t)&ntv, (caddr_t)uap->tp, sizeof(ntv));
478 	if (error == 0)
479 		td->td_retval[0] = retval;
480 	return (error);
481 }
482 
483 /*
484  * second_overflow() - called after ntp_tick_adjust()
485  *
486  * This routine is ordinarily called immediately following the above
487  * routine ntp_tick_adjust(). While these two routines are normally
488  * combined, they are separated here only for the purposes of
489  * simulation.
490  */
491 void
492 ntp_update_second(int64_t *adjustment, time_t *newsec)
493 {
494 	int tickrate;
495 	l_fp ftemp;		/* 32/64-bit temporary */
496 
497 	NTP_LOCK();
498 
499 	/*
500 	 * On rollover of the second both the nanosecond and microsecond
501 	 * clocks are updated and the state machine cranked as
502 	 * necessary. The phase adjustment to be used for the next
503 	 * second is calculated and the maximum error is increased by
504 	 * the tolerance.
505 	 */
506 	time_maxerror += MAXFREQ / 1000;
507 
508 	/*
509 	 * Leap second processing. If in leap-insert state at
510 	 * the end of the day, the system clock is set back one
511 	 * second; if in leap-delete state, the system clock is
512 	 * set ahead one second. The nano_time() routine or
513 	 * external clock driver will insure that reported time
514 	 * is always monotonic.
515 	 */
516 	switch (time_state) {
517 
518 		/*
519 		 * No warning.
520 		 */
521 		case TIME_OK:
522 		if (time_status & STA_INS)
523 			time_state = TIME_INS;
524 		else if (time_status & STA_DEL)
525 			time_state = TIME_DEL;
526 		break;
527 
528 		/*
529 		 * Insert second 23:59:60 following second
530 		 * 23:59:59.
531 		 */
532 		case TIME_INS:
533 		if (!(time_status & STA_INS))
534 			time_state = TIME_OK;
535 		else if ((*newsec) % 86400 == 0) {
536 			(*newsec)--;
537 			time_state = TIME_OOP;
538 			time_tai++;
539 		}
540 		break;
541 
542 		/*
543 		 * Delete second 23:59:59.
544 		 */
545 		case TIME_DEL:
546 		if (!(time_status & STA_DEL))
547 			time_state = TIME_OK;
548 		else if (((*newsec) + 1) % 86400 == 0) {
549 			(*newsec)++;
550 			time_tai--;
551 			time_state = TIME_WAIT;
552 		}
553 		break;
554 
555 		/*
556 		 * Insert second in progress.
557 		 */
558 		case TIME_OOP:
559 			time_state = TIME_WAIT;
560 		break;
561 
562 		/*
563 		 * Wait for status bits to clear.
564 		 */
565 		case TIME_WAIT:
566 		if (!(time_status & (STA_INS | STA_DEL)))
567 			time_state = TIME_OK;
568 	}
569 
570 	/*
571 	 * Compute the total time adjustment for the next second
572 	 * in ns. The offset is reduced by a factor depending on
573 	 * whether the PPS signal is operating. Note that the
574 	 * value is in effect scaled by the clock frequency,
575 	 * since the adjustment is added at each tick interrupt.
576 	 */
577 	ftemp = time_offset;
578 #ifdef PPS_SYNC
579 	/* XXX even if PPS signal dies we should finish adjustment ? */
580 	if (time_status & STA_PPSTIME && time_status &
581 	    STA_PPSSIGNAL)
582 		L_RSHIFT(ftemp, pps_shift);
583 	else
584 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
585 #else
586 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
587 #endif /* PPS_SYNC */
588 	time_adj = ftemp;
589 	L_SUB(time_offset, ftemp);
590 	L_ADD(time_adj, time_freq);
591 
592 	/*
593 	 * Apply any correction from adjtime(2).  If more than one second
594 	 * off we slew at a rate of 5ms/s (5000 PPM) else 500us/s (500PPM)
595 	 * until the last second is slewed the final < 500 usecs.
596 	 */
597 	if (time_adjtime != 0) {
598 		if (time_adjtime > 1000000)
599 			tickrate = 5000;
600 		else if (time_adjtime < -1000000)
601 			tickrate = -5000;
602 		else if (time_adjtime > 500)
603 			tickrate = 500;
604 		else if (time_adjtime < -500)
605 			tickrate = -500;
606 		else
607 			tickrate = time_adjtime;
608 		time_adjtime -= tickrate;
609 		L_LINT(ftemp, tickrate * 1000);
610 		L_ADD(time_adj, ftemp);
611 	}
612 	*adjustment = time_adj;
613 
614 #ifdef PPS_SYNC
615 	if (pps_valid > 0)
616 		pps_valid--;
617 	else
618 		time_status &= ~STA_PPSSIGNAL;
619 #endif /* PPS_SYNC */
620 
621 	NTP_UNLOCK();
622 }
623 
624 /*
625  * ntp_init() - initialize variables and structures
626  *
627  * This routine must be called after the kernel variables hz and tick
628  * are set or changed and before the next tick interrupt. In this
629  * particular implementation, these values are assumed set elsewhere in
630  * the kernel. The design allows the clock frequency and tick interval
631  * to be changed while the system is running. So, this routine should
632  * probably be integrated with the code that does that.
633  */
634 static void
635 ntp_init(void)
636 {
637 
638 	/*
639 	 * The following variables are initialized only at startup. Only
640 	 * those structures not cleared by the compiler need to be
641 	 * initialized, and these only in the simulator. In the actual
642 	 * kernel, any nonzero values here will quickly evaporate.
643 	 */
644 	L_CLR(time_offset);
645 	L_CLR(time_freq);
646 #ifdef PPS_SYNC
647 	pps_tf[0].tv_sec = pps_tf[0].tv_nsec = 0;
648 	pps_tf[1].tv_sec = pps_tf[1].tv_nsec = 0;
649 	pps_tf[2].tv_sec = pps_tf[2].tv_nsec = 0;
650 	pps_fcount = 0;
651 	L_CLR(pps_freq);
652 #endif /* PPS_SYNC */
653 }
654 
655 SYSINIT(ntpclocks, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, ntp_init, NULL);
656 
657 /*
658  * hardupdate() - local clock update
659  *
660  * This routine is called by ntp_adjtime() to update the local clock
661  * phase and frequency. The implementation is of an adaptive-parameter,
662  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
663  * time and frequency offset estimates for each call. If the kernel PPS
664  * discipline code is configured (PPS_SYNC), the PPS signal itself
665  * determines the new time offset, instead of the calling argument.
666  * Presumably, calls to ntp_adjtime() occur only when the caller
667  * believes the local clock is valid within some bound (+-128 ms with
668  * NTP). If the caller's time is far different than the PPS time, an
669  * argument will ensue, and it's not clear who will lose.
670  *
671  * For uncompensated quartz crystal oscillators and nominal update
672  * intervals less than 256 s, operation should be in phase-lock mode,
673  * where the loop is disciplined to phase. For update intervals greater
674  * than 1024 s, operation should be in frequency-lock mode, where the
675  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
676  * is selected by the STA_MODE status bit.
677  */
678 static void
679 hardupdate(offset)
680 	long offset;		/* clock offset (ns) */
681 {
682 	long mtemp;
683 	l_fp ftemp;
684 
685 	NTP_ASSERT_LOCKED();
686 
687 	/*
688 	 * Select how the phase is to be controlled and from which
689 	 * source. If the PPS signal is present and enabled to
690 	 * discipline the time, the PPS offset is used; otherwise, the
691 	 * argument offset is used.
692 	 */
693 	if (!(time_status & STA_PLL))
694 		return;
695 	if (!(time_status & STA_PPSTIME && time_status &
696 	    STA_PPSSIGNAL)) {
697 		if (offset > MAXPHASE)
698 			time_monitor = MAXPHASE;
699 		else if (offset < -MAXPHASE)
700 			time_monitor = -MAXPHASE;
701 		else
702 			time_monitor = offset;
703 		L_LINT(time_offset, time_monitor);
704 	}
705 
706 	/*
707 	 * Select how the frequency is to be controlled and in which
708 	 * mode (PLL or FLL). If the PPS signal is present and enabled
709 	 * to discipline the frequency, the PPS frequency is used;
710 	 * otherwise, the argument offset is used to compute it.
711 	 */
712 	if (time_status & STA_PPSFREQ && time_status & STA_PPSSIGNAL) {
713 		time_reftime = time_uptime;
714 		return;
715 	}
716 	if (time_status & STA_FREQHOLD || time_reftime == 0)
717 		time_reftime = time_uptime;
718 	mtemp = time_uptime - time_reftime;
719 	L_LINT(ftemp, time_monitor);
720 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
721 	L_MPY(ftemp, mtemp);
722 	L_ADD(time_freq, ftemp);
723 	time_status &= ~STA_MODE;
724 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp >
725 	    MAXSEC)) {
726 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
727 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
728 		L_ADD(time_freq, ftemp);
729 		time_status |= STA_MODE;
730 	}
731 	time_reftime = time_uptime;
732 	if (L_GINT(time_freq) > MAXFREQ)
733 		L_LINT(time_freq, MAXFREQ);
734 	else if (L_GINT(time_freq) < -MAXFREQ)
735 		L_LINT(time_freq, -MAXFREQ);
736 }
737 
738 #ifdef PPS_SYNC
739 /*
740  * hardpps() - discipline CPU clock oscillator to external PPS signal
741  *
742  * This routine is called at each PPS interrupt in order to discipline
743  * the CPU clock oscillator to the PPS signal. There are two independent
744  * first-order feedback loops, one for the phase, the other for the
745  * frequency. The phase loop measures and grooms the PPS phase offset
746  * and leaves it in a handy spot for the seconds overflow routine. The
747  * frequency loop averages successive PPS phase differences and
748  * calculates the PPS frequency offset, which is also processed by the
749  * seconds overflow routine. The code requires the caller to capture the
750  * time and architecture-dependent hardware counter values in
751  * nanoseconds at the on-time PPS signal transition.
752  *
753  * Note that, on some Unix systems this routine runs at an interrupt
754  * priority level higher than the timer interrupt routine hardclock().
755  * Therefore, the variables used are distinct from the hardclock()
756  * variables, except for the actual time and frequency variables, which
757  * are determined by this routine and updated atomically.
758  *
759  * tsp  - time at PPS
760  * nsec - hardware counter at PPS
761  */
762 void
763 hardpps(struct timespec *tsp, long nsec)
764 {
765 	long u_sec, u_nsec, v_nsec; /* temps */
766 	l_fp ftemp;
767 
768 	NTP_LOCK();
769 
770 	/*
771 	 * The signal is first processed by a range gate and frequency
772 	 * discriminator. The range gate rejects noise spikes outside
773 	 * the range +-500 us. The frequency discriminator rejects input
774 	 * signals with apparent frequency outside the range 1 +-500
775 	 * PPM. If two hits occur in the same second, we ignore the
776 	 * later hit; if not and a hit occurs outside the range gate,
777 	 * keep the later hit for later comparison, but do not process
778 	 * it.
779 	 */
780 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
781 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
782 	pps_valid = PPS_VALID;
783 	u_sec = tsp->tv_sec;
784 	u_nsec = tsp->tv_nsec;
785 	if (u_nsec >= (NANOSECOND >> 1)) {
786 		u_nsec -= NANOSECOND;
787 		u_sec++;
788 	}
789 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
790 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND - MAXFREQ)
791 		goto out;
792 	pps_tf[2] = pps_tf[1];
793 	pps_tf[1] = pps_tf[0];
794 	pps_tf[0].tv_sec = u_sec;
795 	pps_tf[0].tv_nsec = u_nsec;
796 
797 	/*
798 	 * Compute the difference between the current and previous
799 	 * counter values. If the difference exceeds 0.5 s, assume it
800 	 * has wrapped around, so correct 1.0 s. If the result exceeds
801 	 * the tick interval, the sample point has crossed a tick
802 	 * boundary during the last second, so correct the tick. Very
803 	 * intricate.
804 	 */
805 	u_nsec = nsec;
806 	if (u_nsec > (NANOSECOND >> 1))
807 		u_nsec -= NANOSECOND;
808 	else if (u_nsec < -(NANOSECOND >> 1))
809 		u_nsec += NANOSECOND;
810 	pps_fcount += u_nsec;
811 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
812 		goto out;
813 	time_status &= ~STA_PPSJITTER;
814 
815 	/*
816 	 * A three-stage median filter is used to help denoise the PPS
817 	 * time. The median sample becomes the time offset estimate; the
818 	 * difference between the other two samples becomes the time
819 	 * dispersion (jitter) estimate.
820 	 */
821 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
822 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
823 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
824 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
825 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
826 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
827 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
828 		} else {
829 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
830 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
831 		}
832 	} else {
833 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
834 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
835 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
836 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
837 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
838 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
839 		} else {
840 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
841 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
842 		}
843 	}
844 
845 	/*
846 	 * Nominal jitter is due to PPS signal noise and interrupt
847 	 * latency. If it exceeds the popcorn threshold, the sample is
848 	 * discarded. otherwise, if so enabled, the time offset is
849 	 * updated. We can tolerate a modest loss of data here without
850 	 * much degrading time accuracy.
851 	 *
852 	 * The measurements being checked here were made with the system
853 	 * timecounter, so the popcorn threshold is not allowed to fall below
854 	 * the number of nanoseconds in two ticks of the timecounter.  For a
855 	 * timecounter running faster than 1 GHz the lower bound is 2ns, just
856 	 * to avoid a nonsensical threshold of zero.
857 	*/
858 	if (u_nsec > lmax(pps_jitter << PPS_POPCORN,
859 	    2 * (NANOSECOND / (long)qmin(NANOSECOND, tc_getfrequency())))) {
860 		time_status |= STA_PPSJITTER;
861 		pps_jitcnt++;
862 	} else if (time_status & STA_PPSTIME) {
863 		time_monitor = -v_nsec;
864 		L_LINT(time_offset, time_monitor);
865 	}
866 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
867 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
868 	if (u_sec < (1 << pps_shift))
869 		goto out;
870 
871 	/*
872 	 * At the end of the calibration interval the difference between
873 	 * the first and last counter values becomes the scaled
874 	 * frequency. It will later be divided by the length of the
875 	 * interval to determine the frequency update. If the frequency
876 	 * exceeds a sanity threshold, or if the actual calibration
877 	 * interval is not equal to the expected length, the data are
878 	 * discarded. We can tolerate a modest loss of data here without
879 	 * much degrading frequency accuracy.
880 	 */
881 	pps_calcnt++;
882 	v_nsec = -pps_fcount;
883 	pps_lastsec = pps_tf[0].tv_sec;
884 	pps_fcount = 0;
885 	u_nsec = MAXFREQ << pps_shift;
886 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 << pps_shift)) {
887 		time_status |= STA_PPSERROR;
888 		pps_errcnt++;
889 		goto out;
890 	}
891 
892 	/*
893 	 * Here the raw frequency offset and wander (stability) is
894 	 * calculated. If the wander is less than the wander threshold
895 	 * for four consecutive averaging intervals, the interval is
896 	 * doubled; if it is greater than the threshold for four
897 	 * consecutive intervals, the interval is halved. The scaled
898 	 * frequency offset is converted to frequency offset. The
899 	 * stability metric is calculated as the average of recent
900 	 * frequency changes, but is used only for performance
901 	 * monitoring.
902 	 */
903 	L_LINT(ftemp, v_nsec);
904 	L_RSHIFT(ftemp, pps_shift);
905 	L_SUB(ftemp, pps_freq);
906 	u_nsec = L_GINT(ftemp);
907 	if (u_nsec > PPS_MAXWANDER) {
908 		L_LINT(ftemp, PPS_MAXWANDER);
909 		pps_intcnt--;
910 		time_status |= STA_PPSWANDER;
911 		pps_stbcnt++;
912 	} else if (u_nsec < -PPS_MAXWANDER) {
913 		L_LINT(ftemp, -PPS_MAXWANDER);
914 		pps_intcnt--;
915 		time_status |= STA_PPSWANDER;
916 		pps_stbcnt++;
917 	} else {
918 		pps_intcnt++;
919 	}
920 	if (pps_intcnt >= 4) {
921 		pps_intcnt = 4;
922 		if (pps_shift < pps_shiftmax) {
923 			pps_shift++;
924 			pps_intcnt = 0;
925 		}
926 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
927 		pps_intcnt = -4;
928 		if (pps_shift > PPS_FAVG) {
929 			pps_shift--;
930 			pps_intcnt = 0;
931 		}
932 	}
933 	if (u_nsec < 0)
934 		u_nsec = -u_nsec;
935 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
936 
937 	/*
938 	 * The PPS frequency is recalculated and clamped to the maximum
939 	 * MAXFREQ. If enabled, the system clock frequency is updated as
940 	 * well.
941 	 */
942 	L_ADD(pps_freq, ftemp);
943 	u_nsec = L_GINT(pps_freq);
944 	if (u_nsec > MAXFREQ)
945 		L_LINT(pps_freq, MAXFREQ);
946 	else if (u_nsec < -MAXFREQ)
947 		L_LINT(pps_freq, -MAXFREQ);
948 	if (time_status & STA_PPSFREQ)
949 		time_freq = pps_freq;
950 
951 out:
952 	NTP_UNLOCK();
953 }
954 #endif /* PPS_SYNC */
955 
956 #ifndef _SYS_SYSPROTO_H_
957 struct adjtime_args {
958 	struct timeval *delta;
959 	struct timeval *olddelta;
960 };
961 #endif
962 /* ARGSUSED */
963 int
964 sys_adjtime(struct thread *td, struct adjtime_args *uap)
965 {
966 	struct timeval delta, olddelta, *deltap;
967 	int error;
968 
969 	if (uap->delta) {
970 		error = copyin(uap->delta, &delta, sizeof(delta));
971 		if (error)
972 			return (error);
973 		deltap = &delta;
974 	} else
975 		deltap = NULL;
976 	error = kern_adjtime(td, deltap, &olddelta);
977 	if (uap->olddelta && error == 0)
978 		error = copyout(&olddelta, uap->olddelta, sizeof(olddelta));
979 	return (error);
980 }
981 
982 int
983 kern_adjtime(struct thread *td, struct timeval *delta, struct timeval *olddelta)
984 {
985 	struct timeval atv;
986 	int64_t ltr, ltw;
987 	int error;
988 
989 	if (delta != NULL) {
990 		error = priv_check(td, PRIV_ADJTIME);
991 		if (error != 0)
992 			return (error);
993 		ltw = (int64_t)delta->tv_sec * 1000000 + delta->tv_usec;
994 	}
995 	NTP_LOCK();
996 	ltr = time_adjtime;
997 	if (delta != NULL)
998 		time_adjtime = ltw;
999 	NTP_UNLOCK();
1000 	if (olddelta != NULL) {
1001 		atv.tv_sec = ltr / 1000000;
1002 		atv.tv_usec = ltr % 1000000;
1003 		if (atv.tv_usec < 0) {
1004 			atv.tv_usec += 1000000;
1005 			atv.tv_sec--;
1006 		}
1007 		*olddelta = atv;
1008 	}
1009 	return (0);
1010 }
1011 
1012 static struct callout resettodr_callout;
1013 static int resettodr_period = 1800;
1014 
1015 static void
1016 periodic_resettodr(void *arg __unused)
1017 {
1018 
1019 	/*
1020 	 * Read of time_status is lock-less, which is fine since
1021 	 * ntp_is_time_error() operates on the consistent read value.
1022 	 */
1023 	if (!ntp_is_time_error(time_status))
1024 		resettodr();
1025 	if (resettodr_period > 0)
1026 		callout_schedule(&resettodr_callout, resettodr_period * hz);
1027 }
1028 
1029 static void
1030 shutdown_resettodr(void *arg __unused, int howto __unused)
1031 {
1032 
1033 	callout_drain(&resettodr_callout);
1034 	/* Another unlocked read of time_status */
1035 	if (resettodr_period > 0 && !ntp_is_time_error(time_status))
1036 		resettodr();
1037 }
1038 
1039 static int
1040 sysctl_resettodr_period(SYSCTL_HANDLER_ARGS)
1041 {
1042 	int error;
1043 
1044 	error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
1045 	if (error || !req->newptr)
1046 		return (error);
1047 	if (cold)
1048 		goto done;
1049 	if (resettodr_period == 0)
1050 		callout_stop(&resettodr_callout);
1051 	else
1052 		callout_reset(&resettodr_callout, resettodr_period * hz,
1053 		    periodic_resettodr, NULL);
1054 done:
1055 	return (0);
1056 }
1057 
1058 SYSCTL_PROC(_machdep, OID_AUTO, rtc_save_period, CTLTYPE_INT | CTLFLAG_RWTUN |
1059     CTLFLAG_MPSAFE, &resettodr_period, 1800, sysctl_resettodr_period, "I",
1060     "Save system time to RTC with this period (in seconds)");
1061 
1062 static void
1063 start_periodic_resettodr(void *arg __unused)
1064 {
1065 
1066 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_resettodr, NULL,
1067 	    SHUTDOWN_PRI_FIRST);
1068 	callout_init(&resettodr_callout, 1);
1069 	if (resettodr_period == 0)
1070 		return;
1071 	callout_reset(&resettodr_callout, resettodr_period * hz,
1072 	    periodic_resettodr, NULL);
1073 }
1074 
1075 SYSINIT(periodic_resettodr, SI_SUB_LAST, SI_ORDER_MIDDLE,
1076 	start_periodic_resettodr, NULL);
1077