xref: /freebsd/sys/kern/kern_ntptime.c (revision b52f49a9a0f22207ad5130ad8faba08de3ed23d8)
1 /***********************************************************************
2  *								       *
3  * Copyright (c) David L. Mills 1993-2001			       *
4  *								       *
5  * Permission to use, copy, modify, and distribute this software and   *
6  * its documentation for any purpose and without fee is hereby	       *
7  * granted, provided that the above copyright notice appears in all    *
8  * copies and that both the copyright notice and this permission       *
9  * notice appear in supporting documentation, and that the name	       *
10  * University of Delaware not be used in advertising or publicity      *
11  * pertaining to distribution of the software without specific,	       *
12  * written prior permission. The University of Delaware makes no       *
13  * representations about the suitability this software for any	       *
14  * purpose. It is provided "as is" without express or implied	       *
15  * warranty.							       *
16  *								       *
17  **********************************************************************/
18 
19 /*
20  * Adapted from the original sources for FreeBSD and timecounters by:
21  * Poul-Henning Kamp <phk@FreeBSD.org>.
22  *
23  * The 32bit version of the "LP" macros seems a bit past its "sell by"
24  * date so I have retained only the 64bit version and included it directly
25  * in this file.
26  *
27  * Only minor changes done to interface with the timecounters over in
28  * sys/kern/kern_clock.c.   Some of the comments below may be (even more)
29  * confusing and/or plain wrong in that context.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_ntp.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/sysproto.h>
40 #include <sys/kernel.h>
41 #include <sys/proc.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/time.h>
45 #include <sys/timex.h>
46 #include <sys/timetc.h>
47 #include <sys/timepps.h>
48 #include <sys/sysctl.h>
49 
50 /*
51  * Single-precision macros for 64-bit machines
52  */
53 typedef long long l_fp;
54 #define L_ADD(v, u)	((v) += (u))
55 #define L_SUB(v, u)	((v) -= (u))
56 #define L_ADDHI(v, a)	((v) += (long long)(a) << 32)
57 #define L_NEG(v)	((v) = -(v))
58 #define L_RSHIFT(v, n) \
59 	do { \
60 		if ((v) < 0) \
61 			(v) = -(-(v) >> (n)); \
62 		else \
63 			(v) = (v) >> (n); \
64 	} while (0)
65 #define L_MPY(v, a)	((v) *= (a))
66 #define L_CLR(v)	((v) = 0)
67 #define L_ISNEG(v)	((v) < 0)
68 #define L_LINT(v, a)	((v) = (long long)(a) << 32)
69 #define L_GINT(v)	((v) < 0 ? -(-(v) >> 32) : (v) >> 32)
70 
71 /*
72  * Generic NTP kernel interface
73  *
74  * These routines constitute the Network Time Protocol (NTP) interfaces
75  * for user and daemon application programs. The ntp_gettime() routine
76  * provides the time, maximum error (synch distance) and estimated error
77  * (dispersion) to client user application programs. The ntp_adjtime()
78  * routine is used by the NTP daemon to adjust the system clock to an
79  * externally derived time. The time offset and related variables set by
80  * this routine are used by other routines in this module to adjust the
81  * phase and frequency of the clock discipline loop which controls the
82  * system clock.
83  *
84  * When the kernel time is reckoned directly in nanoseconds (NTP_NANO
85  * defined), the time at each tick interrupt is derived directly from
86  * the kernel time variable. When the kernel time is reckoned in
87  * microseconds, (NTP_NANO undefined), the time is derived from the
88  * kernel time variable together with a variable representing the
89  * leftover nanoseconds at the last tick interrupt. In either case, the
90  * current nanosecond time is reckoned from these values plus an
91  * interpolated value derived by the clock routines in another
92  * architecture-specific module. The interpolation can use either a
93  * dedicated counter or a processor cycle counter (PCC) implemented in
94  * some architectures.
95  *
96  * Note that all routines must run at priority splclock or higher.
97  */
98 /*
99  * Phase/frequency-lock loop (PLL/FLL) definitions
100  *
101  * The nanosecond clock discipline uses two variable types, time
102  * variables and frequency variables. Both types are represented as 64-
103  * bit fixed-point quantities with the decimal point between two 32-bit
104  * halves. On a 32-bit machine, each half is represented as a single
105  * word and mathematical operations are done using multiple-precision
106  * arithmetic. On a 64-bit machine, ordinary computer arithmetic is
107  * used.
108  *
109  * A time variable is a signed 64-bit fixed-point number in ns and
110  * fraction. It represents the remaining time offset to be amortized
111  * over succeeding tick interrupts. The maximum time offset is about
112  * 0.5 s and the resolution is about 2.3e-10 ns.
113  *
114  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
115  *  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
116  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
117  * |s s s|			 ns				   |
118  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
119  * |			    fraction				   |
120  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
121  *
122  * A frequency variable is a signed 64-bit fixed-point number in ns/s
123  * and fraction. It represents the ns and fraction to be added to the
124  * kernel time variable at each second. The maximum frequency offset is
125  * about +-500000 ns/s and the resolution is about 2.3e-10 ns/s.
126  *
127  *			1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
128  *  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
129  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
130  * |s s s s s s s s s s s s s|	          ns/s			   |
131  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
132  * |			    fraction				   |
133  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
134  */
135 /*
136  * The following variables establish the state of the PLL/FLL and the
137  * residual time and frequency offset of the local clock.
138  */
139 #define SHIFT_PLL	4		/* PLL loop gain (shift) */
140 #define SHIFT_FLL	2		/* FLL loop gain (shift) */
141 
142 static int time_state = TIME_OK;	/* clock state */
143 static int time_status = STA_UNSYNC;	/* clock status bits */
144 static long time_tai;			/* TAI offset (s) */
145 static long time_monitor;		/* last time offset scaled (ns) */
146 static long time_constant;		/* poll interval (shift) (s) */
147 static long time_precision = 1;		/* clock precision (ns) */
148 static long time_maxerror = MAXPHASE / 1000; /* maximum error (us) */
149 static long time_esterror = MAXPHASE / 1000; /* estimated error (us) */
150 static long time_reftime;		/* time at last adjustment (s) */
151 static l_fp time_offset;		/* time offset (ns) */
152 static l_fp time_freq;			/* frequency offset (ns/s) */
153 static l_fp time_adj;			/* tick adjust (ns/s) */
154 
155 static int64_t time_adjtime;		/* correction from adjtime(2) (usec) */
156 
157 #ifdef PPS_SYNC
158 /*
159  * The following variables are used when a pulse-per-second (PPS) signal
160  * is available and connected via a modem control lead. They establish
161  * the engineering parameters of the clock discipline loop when
162  * controlled by the PPS signal.
163  */
164 #define PPS_FAVG	2		/* min freq avg interval (s) (shift) */
165 #define PPS_FAVGDEF	8		/* default freq avg int (s) (shift) */
166 #define PPS_FAVGMAX	15		/* max freq avg interval (s) (shift) */
167 #define PPS_PAVG	4		/* phase avg interval (s) (shift) */
168 #define PPS_VALID	120		/* PPS signal watchdog max (s) */
169 #define PPS_MAXWANDER	100000		/* max PPS wander (ns/s) */
170 #define PPS_POPCORN	2		/* popcorn spike threshold (shift) */
171 
172 static struct timespec pps_tf[3];	/* phase median filter */
173 static l_fp pps_freq;			/* scaled frequency offset (ns/s) */
174 static long pps_fcount;			/* frequency accumulator */
175 static long pps_jitter;			/* nominal jitter (ns) */
176 static long pps_stabil;			/* nominal stability (scaled ns/s) */
177 static long pps_lastsec;		/* time at last calibration (s) */
178 static int pps_valid;			/* signal watchdog counter */
179 static int pps_shift = PPS_FAVG;	/* interval duration (s) (shift) */
180 static int pps_shiftmax = PPS_FAVGDEF;	/* max interval duration (s) (shift) */
181 static int pps_intcnt;			/* wander counter */
182 
183 /*
184  * PPS signal quality monitors
185  */
186 static long pps_calcnt;			/* calibration intervals */
187 static long pps_jitcnt;			/* jitter limit exceeded */
188 static long pps_stbcnt;			/* stability limit exceeded */
189 static long pps_errcnt;			/* calibration errors */
190 #endif /* PPS_SYNC */
191 /*
192  * End of phase/frequency-lock loop (PLL/FLL) definitions
193  */
194 
195 static void ntp_init(void);
196 static void hardupdate(long offset);
197 
198 /*
199  * ntp_gettime() - NTP user application interface
200  *
201  * See the timex.h header file for synopsis and API description. Note
202  * that the TAI offset is returned in the ntvtimeval.tai structure
203  * member.
204  */
205 static int
206 ntp_sysctl(SYSCTL_HANDLER_ARGS)
207 {
208 	struct ntptimeval ntv;	/* temporary structure */
209 	struct timespec atv;	/* nanosecond time */
210 
211 	nanotime(&atv);
212 	ntv.time.tv_sec = atv.tv_sec;
213 	ntv.time.tv_nsec = atv.tv_nsec;
214 	ntv.maxerror = time_maxerror;
215 	ntv.esterror = time_esterror;
216 	ntv.tai = time_tai;
217 	ntv.time_state = time_state;
218 
219 	/*
220 	 * Status word error decode. If any of these conditions occur,
221 	 * an error is returned, instead of the status word. Most
222 	 * applications will care only about the fact the system clock
223 	 * may not be trusted, not about the details.
224 	 *
225 	 * Hardware or software error
226 	 */
227 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
228 
229 	/*
230 	 * PPS signal lost when either time or frequency synchronization
231 	 * requested
232 	 */
233 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
234 	    !(time_status & STA_PPSSIGNAL)) ||
235 
236 	/*
237 	 * PPS jitter exceeded when time synchronization requested
238 	 */
239 	    (time_status & STA_PPSTIME &&
240 	    time_status & STA_PPSJITTER) ||
241 
242 	/*
243 	 * PPS wander exceeded or calibration error when frequency
244 	 * synchronization requested
245 	 */
246 	    (time_status & STA_PPSFREQ &&
247 	    time_status & (STA_PPSWANDER | STA_PPSERROR)))
248 		ntv.time_state = TIME_ERROR;
249 	return (sysctl_handle_opaque(oidp, &ntv, sizeof ntv, req));
250 }
251 
252 SYSCTL_NODE(_kern, OID_AUTO, ntp_pll, CTLFLAG_RW, 0, "");
253 SYSCTL_PROC(_kern_ntp_pll, OID_AUTO, gettime, CTLTYPE_OPAQUE|CTLFLAG_RD,
254 	0, sizeof(struct ntptimeval) , ntp_sysctl, "S,ntptimeval", "");
255 
256 #ifdef PPS_SYNC
257 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shiftmax, CTLFLAG_RW, &pps_shiftmax, 0, "");
258 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, pps_shift, CTLFLAG_RW, &pps_shift, 0, "");
259 SYSCTL_INT(_kern_ntp_pll, OID_AUTO, time_monitor, CTLFLAG_RD, &time_monitor, 0, "");
260 
261 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, pps_freq, CTLFLAG_RD, &pps_freq, sizeof(pps_freq), "I", "");
262 SYSCTL_OPAQUE(_kern_ntp_pll, OID_AUTO, time_freq, CTLFLAG_RD, &time_freq, sizeof(time_freq), "I", "");
263 #endif
264 /*
265  * ntp_adjtime() - NTP daemon application interface
266  *
267  * See the timex.h header file for synopsis and API description. Note
268  * that the timex.constant structure member has a dual purpose to set
269  * the time constant and to set the TAI offset.
270  */
271 #ifndef _SYS_SYSPROTO_H_
272 struct ntp_adjtime_args {
273 	struct timex *tp;
274 };
275 #endif
276 
277 /*
278  * MPSAFE
279  */
280 int
281 ntp_adjtime(struct thread *td, struct ntp_adjtime_args *uap)
282 {
283 	struct timex ntv;	/* temporary structure */
284 	long freq;		/* frequency ns/s) */
285 	int modes;		/* mode bits from structure */
286 	int s;			/* caller priority */
287 	int error;
288 
289 	error = copyin((caddr_t)uap->tp, (caddr_t)&ntv, sizeof(ntv));
290 	if (error)
291 		return(error);
292 
293 	/*
294 	 * Update selected clock variables - only the superuser can
295 	 * change anything. Note that there is no error checking here on
296 	 * the assumption the superuser should know what it is doing.
297 	 * Note that either the time constant or TAI offset are loaded
298 	 * from the ntv.constant member, depending on the mode bits. If
299 	 * the STA_PLL bit in the status word is cleared, the state and
300 	 * status words are reset to the initial values at boot.
301 	 */
302 	mtx_lock(&Giant);
303 	modes = ntv.modes;
304 	if (modes)
305 		error = suser(td);
306 	if (error)
307 		goto done2;
308 	s = splclock();
309 	if (modes & MOD_MAXERROR)
310 		time_maxerror = ntv.maxerror;
311 	if (modes & MOD_ESTERROR)
312 		time_esterror = ntv.esterror;
313 	if (modes & MOD_STATUS) {
314 		if (time_status & STA_PLL && !(ntv.status & STA_PLL)) {
315 			time_state = TIME_OK;
316 			time_status = STA_UNSYNC;
317 #ifdef PPS_SYNC
318 			pps_shift = PPS_FAVG;
319 #endif /* PPS_SYNC */
320 		}
321 		time_status &= STA_RONLY;
322 		time_status |= ntv.status & ~STA_RONLY;
323 	}
324 	if (modes & MOD_TIMECONST) {
325 		if (ntv.constant < 0)
326 			time_constant = 0;
327 		else if (ntv.constant > MAXTC)
328 			time_constant = MAXTC;
329 		else
330 			time_constant = ntv.constant;
331 	}
332 	if (modes & MOD_TAI) {
333 		if (ntv.constant > 0) /* XXX zero & negative numbers ? */
334 			time_tai = ntv.constant;
335 	}
336 #ifdef PPS_SYNC
337 	if (modes & MOD_PPSMAX) {
338 		if (ntv.shift < PPS_FAVG)
339 			pps_shiftmax = PPS_FAVG;
340 		else if (ntv.shift > PPS_FAVGMAX)
341 			pps_shiftmax = PPS_FAVGMAX;
342 		else
343 			pps_shiftmax = ntv.shift;
344 	}
345 #endif /* PPS_SYNC */
346 	if (modes & MOD_NANO)
347 		time_status |= STA_NANO;
348 	if (modes & MOD_MICRO)
349 		time_status &= ~STA_NANO;
350 	if (modes & MOD_CLKB)
351 		time_status |= STA_CLK;
352 	if (modes & MOD_CLKA)
353 		time_status &= ~STA_CLK;
354 	if (modes & MOD_OFFSET) {
355 		if (time_status & STA_NANO)
356 			hardupdate(ntv.offset);
357 		else
358 			hardupdate(ntv.offset * 1000);
359 	}
360 	if (modes & MOD_FREQUENCY) {
361 		freq = (ntv.freq * 1000LL) >> 16;
362 		if (freq > MAXFREQ)
363 			L_LINT(time_freq, MAXFREQ);
364 		else if (freq < -MAXFREQ)
365 			L_LINT(time_freq, -MAXFREQ);
366 		else
367 			L_LINT(time_freq, freq);
368 #ifdef PPS_SYNC
369 		pps_freq = time_freq;
370 #endif /* PPS_SYNC */
371 	}
372 
373 	/*
374 	 * Retrieve all clock variables. Note that the TAI offset is
375 	 * returned only by ntp_gettime();
376 	 */
377 	if (time_status & STA_NANO)
378 		ntv.offset = L_GINT(time_offset);
379 	else
380 		ntv.offset = L_GINT(time_offset) / 1000; /* XXX rounding ? */
381 	ntv.freq = L_GINT((time_freq / 1000LL) << 16);
382 	ntv.maxerror = time_maxerror;
383 	ntv.esterror = time_esterror;
384 	ntv.status = time_status;
385 	ntv.constant = time_constant;
386 	if (time_status & STA_NANO)
387 		ntv.precision = time_precision;
388 	else
389 		ntv.precision = time_precision / 1000;
390 	ntv.tolerance = MAXFREQ * SCALE_PPM;
391 #ifdef PPS_SYNC
392 	ntv.shift = pps_shift;
393 	ntv.ppsfreq = L_GINT((pps_freq / 1000LL) << 16);
394 	if (time_status & STA_NANO)
395 		ntv.jitter = pps_jitter;
396 	else
397 		ntv.jitter = pps_jitter / 1000;
398 	ntv.stabil = pps_stabil;
399 	ntv.calcnt = pps_calcnt;
400 	ntv.errcnt = pps_errcnt;
401 	ntv.jitcnt = pps_jitcnt;
402 	ntv.stbcnt = pps_stbcnt;
403 #endif /* PPS_SYNC */
404 	splx(s);
405 
406 	error = copyout((caddr_t)&ntv, (caddr_t)uap->tp, sizeof(ntv));
407 	if (error)
408 		goto done2;
409 
410 	/*
411 	 * Status word error decode. See comments in
412 	 * ntp_gettime() routine.
413 	 */
414 	if ((time_status & (STA_UNSYNC | STA_CLOCKERR)) ||
415 	    (time_status & (STA_PPSFREQ | STA_PPSTIME) &&
416 	    !(time_status & STA_PPSSIGNAL)) ||
417 	    (time_status & STA_PPSTIME &&
418 	    time_status & STA_PPSJITTER) ||
419 	    (time_status & STA_PPSFREQ &&
420 	    time_status & (STA_PPSWANDER | STA_PPSERROR))) {
421 		td->td_retval[0] = TIME_ERROR;
422 	} else {
423 		td->td_retval[0] = time_state;
424 	}
425 done2:
426 	mtx_unlock(&Giant);
427 	return (error);
428 }
429 
430 /*
431  * second_overflow() - called after ntp_tick_adjust()
432  *
433  * This routine is ordinarily called immediately following the above
434  * routine ntp_tick_adjust(). While these two routines are normally
435  * combined, they are separated here only for the purposes of
436  * simulation.
437  */
438 void
439 ntp_update_second(int64_t *adjustment, time_t *newsec)
440 {
441 	int tickrate;
442 	l_fp ftemp;		/* 32/64-bit temporary */
443 
444 	/*
445 	 * On rollover of the second both the nanosecond and microsecond
446 	 * clocks are updated and the state machine cranked as
447 	 * necessary. The phase adjustment to be used for the next
448 	 * second is calculated and the maximum error is increased by
449 	 * the tolerance.
450 	 */
451 	time_maxerror += MAXFREQ / 1000;
452 
453 	/*
454 	 * Leap second processing. If in leap-insert state at
455 	 * the end of the day, the system clock is set back one
456 	 * second; if in leap-delete state, the system clock is
457 	 * set ahead one second. The nano_time() routine or
458 	 * external clock driver will insure that reported time
459 	 * is always monotonic.
460 	 */
461 	switch (time_state) {
462 
463 		/*
464 		 * No warning.
465 		 */
466 		case TIME_OK:
467 		if (time_status & STA_INS)
468 			time_state = TIME_INS;
469 		else if (time_status & STA_DEL)
470 			time_state = TIME_DEL;
471 		break;
472 
473 		/*
474 		 * Insert second 23:59:60 following second
475 		 * 23:59:59.
476 		 */
477 		case TIME_INS:
478 		if (!(time_status & STA_INS))
479 			time_state = TIME_OK;
480 		else if ((*newsec) % 86400 == 0) {
481 			(*newsec)--;
482 			time_state = TIME_OOP;
483 		}
484 		break;
485 
486 		/*
487 		 * Delete second 23:59:59.
488 		 */
489 		case TIME_DEL:
490 		if (!(time_status & STA_DEL))
491 			time_state = TIME_OK;
492 		else if (((*newsec) + 1) % 86400 == 0) {
493 			(*newsec)++;
494 			time_tai--;
495 			time_state = TIME_WAIT;
496 		}
497 		break;
498 
499 		/*
500 		 * Insert second in progress.
501 		 */
502 		case TIME_OOP:
503 			time_tai++;
504 			time_state = TIME_WAIT;
505 		break;
506 
507 		/*
508 		 * Wait for status bits to clear.
509 		 */
510 		case TIME_WAIT:
511 		if (!(time_status & (STA_INS | STA_DEL)))
512 			time_state = TIME_OK;
513 	}
514 
515 	/*
516 	 * Compute the total time adjustment for the next second
517 	 * in ns. The offset is reduced by a factor depending on
518 	 * whether the PPS signal is operating. Note that the
519 	 * value is in effect scaled by the clock frequency,
520 	 * since the adjustment is added at each tick interrupt.
521 	 */
522 	ftemp = time_offset;
523 #ifdef PPS_SYNC
524 	/* XXX even if PPS signal dies we should finish adjustment ? */
525 	if (time_status & STA_PPSTIME && time_status &
526 	    STA_PPSSIGNAL)
527 		L_RSHIFT(ftemp, pps_shift);
528 	else
529 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
530 #else
531 		L_RSHIFT(ftemp, SHIFT_PLL + time_constant);
532 #endif /* PPS_SYNC */
533 	time_adj = ftemp;
534 	L_SUB(time_offset, ftemp);
535 	L_ADD(time_adj, time_freq);
536 
537 	/*
538 	 * Apply any correction from adjtime(2).  If more than one second
539 	 * off we slew at a rate of 5ms/s (5000 PPM) else 500us/s (500PPM)
540 	 * until the last second is slewed the final < 500 usecs.
541 	 */
542 	if (time_adjtime != 0) {
543 		if (time_adjtime > 1000000)
544 			tickrate = 5000;
545 		else if (time_adjtime < -1000000)
546 			tickrate = -5000;
547 		else if (time_adjtime > 500)
548 			tickrate = 500;
549 		else if (time_adjtime < -500)
550 			tickrate = -500;
551 		else if (time_adjtime != 0)
552 			tickrate = time_adjtime;
553 		else
554 			tickrate = 0;	/* GCC sucks! */
555 		time_adjtime -= tickrate;
556 		L_LINT(ftemp, tickrate * 1000);
557 		L_ADD(time_adj, ftemp);
558 	}
559 	*adjustment = time_adj;
560 
561 #ifdef PPS_SYNC
562 	if (pps_valid > 0)
563 		pps_valid--;
564 	else
565 		time_status &= ~STA_PPSSIGNAL;
566 #endif /* PPS_SYNC */
567 }
568 
569 /*
570  * ntp_init() - initialize variables and structures
571  *
572  * This routine must be called after the kernel variables hz and tick
573  * are set or changed and before the next tick interrupt. In this
574  * particular implementation, these values are assumed set elsewhere in
575  * the kernel. The design allows the clock frequency and tick interval
576  * to be changed while the system is running. So, this routine should
577  * probably be integrated with the code that does that.
578  */
579 static void
580 ntp_init()
581 {
582 
583 	/*
584 	 * The following variables are initialized only at startup. Only
585 	 * those structures not cleared by the compiler need to be
586 	 * initialized, and these only in the simulator. In the actual
587 	 * kernel, any nonzero values here will quickly evaporate.
588 	 */
589 	L_CLR(time_offset);
590 	L_CLR(time_freq);
591 #ifdef PPS_SYNC
592 	pps_tf[0].tv_sec = pps_tf[0].tv_nsec = 0;
593 	pps_tf[1].tv_sec = pps_tf[1].tv_nsec = 0;
594 	pps_tf[2].tv_sec = pps_tf[2].tv_nsec = 0;
595 	pps_fcount = 0;
596 	L_CLR(pps_freq);
597 #endif /* PPS_SYNC */
598 }
599 
600 SYSINIT(ntpclocks, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, ntp_init, NULL)
601 
602 /*
603  * hardupdate() - local clock update
604  *
605  * This routine is called by ntp_adjtime() to update the local clock
606  * phase and frequency. The implementation is of an adaptive-parameter,
607  * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
608  * time and frequency offset estimates for each call. If the kernel PPS
609  * discipline code is configured (PPS_SYNC), the PPS signal itself
610  * determines the new time offset, instead of the calling argument.
611  * Presumably, calls to ntp_adjtime() occur only when the caller
612  * believes the local clock is valid within some bound (+-128 ms with
613  * NTP). If the caller's time is far different than the PPS time, an
614  * argument will ensue, and it's not clear who will lose.
615  *
616  * For uncompensated quartz crystal oscillators and nominal update
617  * intervals less than 256 s, operation should be in phase-lock mode,
618  * where the loop is disciplined to phase. For update intervals greater
619  * than 1024 s, operation should be in frequency-lock mode, where the
620  * loop is disciplined to frequency. Between 256 s and 1024 s, the mode
621  * is selected by the STA_MODE status bit.
622  */
623 static void
624 hardupdate(offset)
625 	long offset;		/* clock offset (ns) */
626 {
627 	long mtemp;
628 	l_fp ftemp;
629 
630 	/*
631 	 * Select how the phase is to be controlled and from which
632 	 * source. If the PPS signal is present and enabled to
633 	 * discipline the time, the PPS offset is used; otherwise, the
634 	 * argument offset is used.
635 	 */
636 	if (!(time_status & STA_PLL))
637 		return;
638 	if (!(time_status & STA_PPSTIME && time_status &
639 	    STA_PPSSIGNAL)) {
640 		if (offset > MAXPHASE)
641 			time_monitor = MAXPHASE;
642 		else if (offset < -MAXPHASE)
643 			time_monitor = -MAXPHASE;
644 		else
645 			time_monitor = offset;
646 		L_LINT(time_offset, time_monitor);
647 	}
648 
649 	/*
650 	 * Select how the frequency is to be controlled and in which
651 	 * mode (PLL or FLL). If the PPS signal is present and enabled
652 	 * to discipline the frequency, the PPS frequency is used;
653 	 * otherwise, the argument offset is used to compute it.
654 	 */
655 	if (time_status & STA_PPSFREQ && time_status & STA_PPSSIGNAL) {
656 		time_reftime = time_second;
657 		return;
658 	}
659 	if (time_status & STA_FREQHOLD || time_reftime == 0)
660 		time_reftime = time_second;
661 	mtemp = time_second - time_reftime;
662 	L_LINT(ftemp, time_monitor);
663 	L_RSHIFT(ftemp, (SHIFT_PLL + 2 + time_constant) << 1);
664 	L_MPY(ftemp, mtemp);
665 	L_ADD(time_freq, ftemp);
666 	time_status &= ~STA_MODE;
667 	if (mtemp >= MINSEC && (time_status & STA_FLL || mtemp >
668 	    MAXSEC)) {
669 		L_LINT(ftemp, (time_monitor << 4) / mtemp);
670 		L_RSHIFT(ftemp, SHIFT_FLL + 4);
671 		L_ADD(time_freq, ftemp);
672 		time_status |= STA_MODE;
673 	}
674 	time_reftime = time_second;
675 	if (L_GINT(time_freq) > MAXFREQ)
676 		L_LINT(time_freq, MAXFREQ);
677 	else if (L_GINT(time_freq) < -MAXFREQ)
678 		L_LINT(time_freq, -MAXFREQ);
679 }
680 
681 #ifdef PPS_SYNC
682 /*
683  * hardpps() - discipline CPU clock oscillator to external PPS signal
684  *
685  * This routine is called at each PPS interrupt in order to discipline
686  * the CPU clock oscillator to the PPS signal. There are two independent
687  * first-order feedback loops, one for the phase, the other for the
688  * frequency. The phase loop measures and grooms the PPS phase offset
689  * and leaves it in a handy spot for the seconds overflow routine. The
690  * frequency loop averages successive PPS phase differences and
691  * calculates the PPS frequency offset, which is also processed by the
692  * seconds overflow routine. The code requires the caller to capture the
693  * time and architecture-dependent hardware counter values in
694  * nanoseconds at the on-time PPS signal transition.
695  *
696  * Note that, on some Unix systems this routine runs at an interrupt
697  * priority level higher than the timer interrupt routine hardclock().
698  * Therefore, the variables used are distinct from the hardclock()
699  * variables, except for the actual time and frequency variables, which
700  * are determined by this routine and updated atomically.
701  */
702 void
703 hardpps(tsp, nsec)
704 	struct timespec *tsp;	/* time at PPS */
705 	long nsec;		/* hardware counter at PPS */
706 {
707 	long u_sec, u_nsec, v_nsec; /* temps */
708 	l_fp ftemp;
709 
710 	/*
711 	 * The signal is first processed by a range gate and frequency
712 	 * discriminator. The range gate rejects noise spikes outside
713 	 * the range +-500 us. The frequency discriminator rejects input
714 	 * signals with apparent frequency outside the range 1 +-500
715 	 * PPM. If two hits occur in the same second, we ignore the
716 	 * later hit; if not and a hit occurs outside the range gate,
717 	 * keep the later hit for later comparison, but do not process
718 	 * it.
719 	 */
720 	time_status |= STA_PPSSIGNAL | STA_PPSJITTER;
721 	time_status &= ~(STA_PPSWANDER | STA_PPSERROR);
722 	pps_valid = PPS_VALID;
723 	u_sec = tsp->tv_sec;
724 	u_nsec = tsp->tv_nsec;
725 	if (u_nsec >= (NANOSECOND >> 1)) {
726 		u_nsec -= NANOSECOND;
727 		u_sec++;
728 	}
729 	v_nsec = u_nsec - pps_tf[0].tv_nsec;
730 	if (u_sec == pps_tf[0].tv_sec && v_nsec < NANOSECOND -
731 	    MAXFREQ)
732 		return;
733 	pps_tf[2] = pps_tf[1];
734 	pps_tf[1] = pps_tf[0];
735 	pps_tf[0].tv_sec = u_sec;
736 	pps_tf[0].tv_nsec = u_nsec;
737 
738 	/*
739 	 * Compute the difference between the current and previous
740 	 * counter values. If the difference exceeds 0.5 s, assume it
741 	 * has wrapped around, so correct 1.0 s. If the result exceeds
742 	 * the tick interval, the sample point has crossed a tick
743 	 * boundary during the last second, so correct the tick. Very
744 	 * intricate.
745 	 */
746 	u_nsec = nsec;
747 	if (u_nsec > (NANOSECOND >> 1))
748 		u_nsec -= NANOSECOND;
749 	else if (u_nsec < -(NANOSECOND >> 1))
750 		u_nsec += NANOSECOND;
751 	pps_fcount += u_nsec;
752 	if (v_nsec > MAXFREQ || v_nsec < -MAXFREQ)
753 		return;
754 	time_status &= ~STA_PPSJITTER;
755 
756 	/*
757 	 * A three-stage median filter is used to help denoise the PPS
758 	 * time. The median sample becomes the time offset estimate; the
759 	 * difference between the other two samples becomes the time
760 	 * dispersion (jitter) estimate.
761 	 */
762 	if (pps_tf[0].tv_nsec > pps_tf[1].tv_nsec) {
763 		if (pps_tf[1].tv_nsec > pps_tf[2].tv_nsec) {
764 			v_nsec = pps_tf[1].tv_nsec;	/* 0 1 2 */
765 			u_nsec = pps_tf[0].tv_nsec - pps_tf[2].tv_nsec;
766 		} else if (pps_tf[2].tv_nsec > pps_tf[0].tv_nsec) {
767 			v_nsec = pps_tf[0].tv_nsec;	/* 2 0 1 */
768 			u_nsec = pps_tf[2].tv_nsec - pps_tf[1].tv_nsec;
769 		} else {
770 			v_nsec = pps_tf[2].tv_nsec;	/* 0 2 1 */
771 			u_nsec = pps_tf[0].tv_nsec - pps_tf[1].tv_nsec;
772 		}
773 	} else {
774 		if (pps_tf[1].tv_nsec < pps_tf[2].tv_nsec) {
775 			v_nsec = pps_tf[1].tv_nsec;	/* 2 1 0 */
776 			u_nsec = pps_tf[2].tv_nsec - pps_tf[0].tv_nsec;
777 		} else if (pps_tf[2].tv_nsec < pps_tf[0].tv_nsec) {
778 			v_nsec = pps_tf[0].tv_nsec;	/* 1 0 2 */
779 			u_nsec = pps_tf[1].tv_nsec - pps_tf[2].tv_nsec;
780 		} else {
781 			v_nsec = pps_tf[2].tv_nsec;	/* 1 2 0 */
782 			u_nsec = pps_tf[1].tv_nsec - pps_tf[0].tv_nsec;
783 		}
784 	}
785 
786 	/*
787 	 * Nominal jitter is due to PPS signal noise and interrupt
788 	 * latency. If it exceeds the popcorn threshold, the sample is
789 	 * discarded. otherwise, if so enabled, the time offset is
790 	 * updated. We can tolerate a modest loss of data here without
791 	 * much degrading time accuracy.
792 	 */
793 	if (u_nsec > (pps_jitter << PPS_POPCORN)) {
794 		time_status |= STA_PPSJITTER;
795 		pps_jitcnt++;
796 	} else if (time_status & STA_PPSTIME) {
797 		time_monitor = -v_nsec;
798 		L_LINT(time_offset, time_monitor);
799 	}
800 	pps_jitter += (u_nsec - pps_jitter) >> PPS_FAVG;
801 	u_sec = pps_tf[0].tv_sec - pps_lastsec;
802 	if (u_sec < (1 << pps_shift))
803 		return;
804 
805 	/*
806 	 * At the end of the calibration interval the difference between
807 	 * the first and last counter values becomes the scaled
808 	 * frequency. It will later be divided by the length of the
809 	 * interval to determine the frequency update. If the frequency
810 	 * exceeds a sanity threshold, or if the actual calibration
811 	 * interval is not equal to the expected length, the data are
812 	 * discarded. We can tolerate a modest loss of data here without
813 	 * much degrading frequency accuracy.
814 	 */
815 	pps_calcnt++;
816 	v_nsec = -pps_fcount;
817 	pps_lastsec = pps_tf[0].tv_sec;
818 	pps_fcount = 0;
819 	u_nsec = MAXFREQ << pps_shift;
820 	if (v_nsec > u_nsec || v_nsec < -u_nsec || u_sec != (1 <<
821 	    pps_shift)) {
822 		time_status |= STA_PPSERROR;
823 		pps_errcnt++;
824 		return;
825 	}
826 
827 	/*
828 	 * Here the raw frequency offset and wander (stability) is
829 	 * calculated. If the wander is less than the wander threshold
830 	 * for four consecutive averaging intervals, the interval is
831 	 * doubled; if it is greater than the threshold for four
832 	 * consecutive intervals, the interval is halved. The scaled
833 	 * frequency offset is converted to frequency offset. The
834 	 * stability metric is calculated as the average of recent
835 	 * frequency changes, but is used only for performance
836 	 * monitoring.
837 	 */
838 	L_LINT(ftemp, v_nsec);
839 	L_RSHIFT(ftemp, pps_shift);
840 	L_SUB(ftemp, pps_freq);
841 	u_nsec = L_GINT(ftemp);
842 	if (u_nsec > PPS_MAXWANDER) {
843 		L_LINT(ftemp, PPS_MAXWANDER);
844 		pps_intcnt--;
845 		time_status |= STA_PPSWANDER;
846 		pps_stbcnt++;
847 	} else if (u_nsec < -PPS_MAXWANDER) {
848 		L_LINT(ftemp, -PPS_MAXWANDER);
849 		pps_intcnt--;
850 		time_status |= STA_PPSWANDER;
851 		pps_stbcnt++;
852 	} else {
853 		pps_intcnt++;
854 	}
855 	if (pps_intcnt >= 4) {
856 		pps_intcnt = 4;
857 		if (pps_shift < pps_shiftmax) {
858 			pps_shift++;
859 			pps_intcnt = 0;
860 		}
861 	} else if (pps_intcnt <= -4 || pps_shift > pps_shiftmax) {
862 		pps_intcnt = -4;
863 		if (pps_shift > PPS_FAVG) {
864 			pps_shift--;
865 			pps_intcnt = 0;
866 		}
867 	}
868 	if (u_nsec < 0)
869 		u_nsec = -u_nsec;
870 	pps_stabil += (u_nsec * SCALE_PPM - pps_stabil) >> PPS_FAVG;
871 
872 	/*
873 	 * The PPS frequency is recalculated and clamped to the maximum
874 	 * MAXFREQ. If enabled, the system clock frequency is updated as
875 	 * well.
876 	 */
877 	L_ADD(pps_freq, ftemp);
878 	u_nsec = L_GINT(pps_freq);
879 	if (u_nsec > MAXFREQ)
880 		L_LINT(pps_freq, MAXFREQ);
881 	else if (u_nsec < -MAXFREQ)
882 		L_LINT(pps_freq, -MAXFREQ);
883 	if (time_status & STA_PPSFREQ)
884 		time_freq = pps_freq;
885 }
886 #endif /* PPS_SYNC */
887 
888 #ifndef _SYS_SYSPROTO_H_
889 struct adjtime_args {
890 	struct timeval *delta;
891 	struct timeval *olddelta;
892 };
893 #endif
894 /*
895  * MPSAFE
896  */
897 /* ARGSUSED */
898 int
899 adjtime(struct thread *td, struct adjtime_args *uap)
900 {
901 	struct timeval atv;
902 	int error;
903 
904 	if ((error = suser(td)))
905 		return (error);
906 
907 	mtx_lock(&Giant);
908 	if (uap->olddelta) {
909 		atv.tv_sec = time_adjtime / 1000000;
910 		atv.tv_usec = time_adjtime % 1000000;
911 		if (atv.tv_usec < 0) {
912 			atv.tv_usec += 1000000;
913 			atv.tv_sec--;
914 		}
915 		error = copyout(&atv, uap->olddelta, sizeof(atv));
916 		if (error)
917 			goto done2;
918 	}
919 	if (uap->delta) {
920 		error = copyin(uap->delta, &atv, sizeof(atv));
921 		if (error)
922 			goto done2;
923 		time_adjtime = (int64_t)atv.tv_sec * 1000000 + atv.tv_usec;
924 	}
925 done2:
926 	mtx_unlock(&Giant);
927 	return (error);
928 }
929 
930