xref: /netbsd-src/sys/kern/kern_tc.c (revision 0920b4f20b78ab1ccd9f2312fbe10deaf000cbf3)
1 /* $NetBSD: kern_tc.c,v 1.20 2007/08/17 21:20:24 ad Exp $ */
2 
3 /*-
4  * ----------------------------------------------------------------------------
5  * "THE BEER-WARE LICENSE" (Revision 42):
6  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
7  * can do whatever you want with this stuff. If we meet some day, and you think
8  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
9  * ---------------------------------------------------------------------------
10  */
11 
12 #include <sys/cdefs.h>
13 /* __FBSDID("$FreeBSD: src/sys/kern/kern_tc.c,v 1.166 2005/09/19 22:16:31 andre Exp $"); */
14 __KERNEL_RCSID(0, "$NetBSD: kern_tc.c,v 1.20 2007/08/17 21:20:24 ad Exp $");
15 
16 #include "opt_ntp.h"
17 
18 #include <sys/param.h>
19 #ifdef __HAVE_TIMECOUNTER	/* XXX */
20 #include <sys/kernel.h>
21 #include <sys/reboot.h>	/* XXX just to get AB_VERBOSE */
22 #include <sys/sysctl.h>
23 #include <sys/syslog.h>
24 #include <sys/systm.h>
25 #include <sys/timepps.h>
26 #include <sys/timetc.h>
27 #include <sys/timex.h>
28 #include <sys/evcnt.h>
29 #include <sys/kauth.h>
30 
31 /*
32  * A large step happens on boot.  This constant detects such steps.
33  * It is relatively small so that ntp_update_second gets called enough
34  * in the typical 'missed a couple of seconds' case, but doesn't loop
35  * forever when the time step is large.
36  */
37 #define LARGE_STEP	200
38 
39 /*
40  * Implement a dummy timecounter which we can use until we get a real one
41  * in the air.  This allows the console and other early stuff to use
42  * time services.
43  */
44 
45 static u_int
46 dummy_get_timecount(struct timecounter *tc)
47 {
48 	static u_int now;
49 
50 	return (++now);
51 }
52 
53 static struct timecounter dummy_timecounter = {
54 	dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000, NULL, NULL,
55 };
56 
57 struct timehands {
58 	/* These fields must be initialized by the driver. */
59 	struct timecounter	*th_counter;
60 	int64_t			th_adjustment;
61 	u_int64_t		th_scale;
62 	u_int	 		th_offset_count;
63 	struct bintime		th_offset;
64 	struct timeval		th_microtime;
65 	struct timespec		th_nanotime;
66 	/* Fields not to be copied in tc_windup start with th_generation. */
67 	volatile u_int		th_generation;
68 	struct timehands	*th_next;
69 };
70 
71 static struct timehands th0;
72 static struct timehands th9 = { .th_next = &th0, };
73 static struct timehands th8 = { .th_next = &th9, };
74 static struct timehands th7 = { .th_next = &th8, };
75 static struct timehands th6 = { .th_next = &th7, };
76 static struct timehands th5 = { .th_next = &th6, };
77 static struct timehands th4 = { .th_next = &th5, };
78 static struct timehands th3 = { .th_next = &th4, };
79 static struct timehands th2 = { .th_next = &th3, };
80 static struct timehands th1 = { .th_next = &th2, };
81 static struct timehands th0 = {
82 	.th_counter = &dummy_timecounter,
83 	.th_scale = (uint64_t)-1 / 1000000,
84 	.th_offset = { .sec = 1, .frac = 0 },
85 	.th_generation = 1,
86 	.th_next = &th1,
87 };
88 
89 static struct timehands *volatile timehands = &th0;
90 struct timecounter *timecounter = &dummy_timecounter;
91 static struct timecounter *timecounters = &dummy_timecounter;
92 
93 time_t time_second = 1;
94 time_t time_uptime = 1;
95 
96 static struct bintime timebasebin;
97 
98 static int timestepwarnings;
99 
100 #ifdef __FreeBSD__
101 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
102     &timestepwarnings, 0, "");
103 #endif /* __FreeBSD__ */
104 
105 /*
106  * sysctl helper routine for kern.timercounter.current
107  */
108 static int
109 sysctl_kern_timecounter_hardware(SYSCTLFN_ARGS)
110 {
111 	struct sysctlnode node;
112 	int error;
113 	char newname[MAX_TCNAMELEN];
114 	struct timecounter *newtc, *tc;
115 
116 	tc = timecounter;
117 
118 	strlcpy(newname, tc->tc_name, sizeof(newname));
119 
120 	node = *rnode;
121 	node.sysctl_data = newname;
122 	node.sysctl_size = sizeof(newname);
123 
124 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
125 
126 	if (error ||
127 	    newp == NULL ||
128 	    strncmp(newname, tc->tc_name, sizeof(newname)) == 0)
129 		return error;
130 
131 	if (l != NULL && (error = kauth_authorize_generic(l->l_cred,
132 	    KAUTH_GENERIC_ISSUSER, NULL)) != 0)
133 		return (error);
134 
135 	/* XXX locking */
136 
137 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
138 		if (strcmp(newname, newtc->tc_name) != 0)
139 			continue;
140 
141 		/* Warm up new timecounter. */
142 		(void)newtc->tc_get_timecount(newtc);
143 		(void)newtc->tc_get_timecount(newtc);
144 
145 		timecounter = newtc;
146 
147 		/* XXX unlock */
148 
149 		return (0);
150 	}
151 
152 	/* XXX unlock */
153 
154 	return (EINVAL);
155 }
156 
157 static int
158 sysctl_kern_timecounter_choice(SYSCTLFN_ARGS)
159 {
160 	char buf[MAX_TCNAMELEN+48];
161 	char *where = oldp;
162 	const char *spc;
163 	struct timecounter *tc;
164 	size_t needed, left, slen;
165 	int error;
166 
167 	if (newp != NULL)
168 		return (EPERM);
169 	if (namelen != 0)
170 		return (EINVAL);
171 
172 	spc = "";
173 	error = 0;
174 	needed = 0;
175 	left = *oldlenp;
176 
177 	/* XXX locking */
178 
179 	for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
180 		if (where == NULL) {
181 			needed += sizeof(buf);  /* be conservative */
182 		} else {
183 			slen = snprintf(buf, sizeof(buf), "%s%s(q=%d, f=%" PRId64
184 					" Hz)", spc, tc->tc_name, tc->tc_quality,
185 					tc->tc_frequency);
186 			if (left < slen + 1)
187 				break;
188 			/* XXX use sysctl_copyout? (from sysctl_hw_disknames) */
189 			error = copyout(buf, where, slen + 1);
190 			spc = " ";
191 			where += slen;
192 			needed += slen;
193 			left -= slen;
194 		}
195 	}
196 
197 	/* XXX unlock */
198 
199 	*oldlenp = needed;
200 	return (error);
201 }
202 
203 SYSCTL_SETUP(sysctl_timecounter_setup, "sysctl timecounter setup")
204 {
205 	const struct sysctlnode *node;
206 
207 	sysctl_createv(clog, 0, NULL, &node,
208 		       CTLFLAG_PERMANENT,
209 		       CTLTYPE_NODE, "timecounter",
210 		       SYSCTL_DESCR("time counter information"),
211 		       NULL, 0, NULL, 0,
212 		       CTL_KERN, CTL_CREATE, CTL_EOL);
213 
214 	if (node != NULL) {
215 		sysctl_createv(clog, 0, NULL, NULL,
216 			       CTLFLAG_PERMANENT,
217 			       CTLTYPE_STRING, "choice",
218 			       SYSCTL_DESCR("available counters"),
219 			       sysctl_kern_timecounter_choice, 0, NULL, 0,
220 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
221 
222 		sysctl_createv(clog, 0, NULL, NULL,
223 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
224 			       CTLTYPE_STRING, "hardware",
225 			       SYSCTL_DESCR("currently active time counter"),
226 			       sysctl_kern_timecounter_hardware, 0, NULL, MAX_TCNAMELEN,
227 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
228 
229 		sysctl_createv(clog, 0, NULL, NULL,
230 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
231 			       CTLTYPE_INT, "timestepwarnings",
232 			       SYSCTL_DESCR("log time steps"),
233 			       NULL, 0, &timestepwarnings, 0,
234 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
235 	}
236 }
237 
238 #define	TC_STATS(name)							\
239 static struct evcnt n##name =						\
240     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "timecounter", #name);	\
241 EVCNT_ATTACH_STATIC(n##name)
242 
243 TC_STATS(binuptime);    TC_STATS(nanouptime);    TC_STATS(microuptime);
244 TC_STATS(bintime);      TC_STATS(nanotime);      TC_STATS(microtime);
245 TC_STATS(getbinuptime); TC_STATS(getnanouptime); TC_STATS(getmicrouptime);
246 TC_STATS(getbintime);   TC_STATS(getnanotime);   TC_STATS(getmicrotime);
247 TC_STATS(setclock);
248 
249 #undef TC_STATS
250 
251 static void tc_windup(void);
252 
253 /*
254  * Return the difference between the timehands' counter value now and what
255  * was when we copied it to the timehands' offset_count.
256  */
257 static __inline u_int
258 tc_delta(struct timehands *th)
259 {
260 	struct timecounter *tc;
261 
262 	tc = th->th_counter;
263 	return ((tc->tc_get_timecount(tc) -
264 		 th->th_offset_count) & tc->tc_counter_mask);
265 }
266 
267 /*
268  * Functions for reading the time.  We have to loop until we are sure that
269  * the timehands that we operated on was not updated under our feet.  See
270  * the comment in <sys/time.h> for a description of these 12 functions.
271  */
272 
273 void
274 binuptime(struct bintime *bt)
275 {
276 	struct timehands *th;
277 	u_int gen;
278 
279 	nbinuptime.ev_count++;
280 	do {
281 		th = timehands;
282 		gen = th->th_generation;
283 		mb_read();
284 		*bt = th->th_offset;
285 		bintime_addx(bt, th->th_scale * tc_delta(th));
286 		mb_read();
287 	} while (gen == 0 || gen != th->th_generation);
288 }
289 
290 void
291 nanouptime(struct timespec *tsp)
292 {
293 	struct bintime bt;
294 
295 	nnanouptime.ev_count++;
296 	binuptime(&bt);
297 	bintime2timespec(&bt, tsp);
298 }
299 
300 void
301 microuptime(struct timeval *tvp)
302 {
303 	struct bintime bt;
304 
305 	nmicrouptime.ev_count++;
306 	binuptime(&bt);
307 	bintime2timeval(&bt, tvp);
308 }
309 
310 void
311 bintime(struct bintime *bt)
312 {
313 
314 	nbintime.ev_count++;
315 	binuptime(bt);
316 	bintime_add(bt, &timebasebin);
317 }
318 
319 void
320 nanotime(struct timespec *tsp)
321 {
322 	struct bintime bt;
323 
324 	nnanotime.ev_count++;
325 	bintime(&bt);
326 	bintime2timespec(&bt, tsp);
327 }
328 
329 void
330 microtime(struct timeval *tvp)
331 {
332 	struct bintime bt;
333 
334 	nmicrotime.ev_count++;
335 	bintime(&bt);
336 	bintime2timeval(&bt, tvp);
337 }
338 
339 void
340 getbinuptime(struct bintime *bt)
341 {
342 	struct timehands *th;
343 	u_int gen;
344 
345 	ngetbinuptime.ev_count++;
346 	do {
347 		th = timehands;
348 		gen = th->th_generation;
349 		mb_read();
350 		*bt = th->th_offset;
351 		mb_read();
352 	} while (gen == 0 || gen != th->th_generation);
353 }
354 
355 void
356 getnanouptime(struct timespec *tsp)
357 {
358 	struct timehands *th;
359 	u_int gen;
360 
361 	ngetnanouptime.ev_count++;
362 	do {
363 		th = timehands;
364 		gen = th->th_generation;
365 		mb_read();
366 		bintime2timespec(&th->th_offset, tsp);
367 		mb_read();
368 	} while (gen == 0 || gen != th->th_generation);
369 }
370 
371 void
372 getmicrouptime(struct timeval *tvp)
373 {
374 	struct timehands *th;
375 	u_int gen;
376 
377 	ngetmicrouptime.ev_count++;
378 	do {
379 		th = timehands;
380 		gen = th->th_generation;
381 		mb_read();
382 		bintime2timeval(&th->th_offset, tvp);
383 		mb_read();
384 	} while (gen == 0 || gen != th->th_generation);
385 }
386 
387 void
388 getbintime(struct bintime *bt)
389 {
390 	struct timehands *th;
391 	u_int gen;
392 
393 	ngetbintime.ev_count++;
394 	do {
395 		th = timehands;
396 		gen = th->th_generation;
397 		mb_read();
398 		*bt = th->th_offset;
399 		mb_read();
400 	} while (gen == 0 || gen != th->th_generation);
401 	bintime_add(bt, &timebasebin);
402 }
403 
404 void
405 getnanotime(struct timespec *tsp)
406 {
407 	struct timehands *th;
408 	u_int gen;
409 
410 	ngetnanotime.ev_count++;
411 	do {
412 		th = timehands;
413 		gen = th->th_generation;
414 		mb_read();
415 		*tsp = th->th_nanotime;
416 		mb_read();
417 	} while (gen == 0 || gen != th->th_generation);
418 }
419 
420 void
421 getmicrotime(struct timeval *tvp)
422 {
423 	struct timehands *th;
424 	u_int gen;
425 
426 	ngetmicrotime.ev_count++;
427 	do {
428 		th = timehands;
429 		gen = th->th_generation;
430 		mb_read();
431 		*tvp = th->th_microtime;
432 		mb_read();
433 	} while (gen == 0 || gen != th->th_generation);
434 }
435 
436 /*
437  * Initialize a new timecounter and possibly use it.
438  */
439 void
440 tc_init(struct timecounter *tc)
441 {
442 	u_int u;
443 	int s;
444 
445 	u = tc->tc_frequency / tc->tc_counter_mask;
446 	/* XXX: We need some margin here, 10% is a guess */
447 	u *= 11;
448 	u /= 10;
449 	if (u > hz && tc->tc_quality >= 0) {
450 		tc->tc_quality = -2000;
451 		aprint_verbose(
452 		    "timecounter: Timecounter \"%s\" frequency %ju Hz",
453 			    tc->tc_name, (uintmax_t)tc->tc_frequency);
454 		aprint_verbose(" -- Insufficient hz, needs at least %u\n", u);
455 	} else if (tc->tc_quality >= 0 || bootverbose) {
456 		aprint_verbose(
457 		    "timecounter: Timecounter \"%s\" frequency %ju Hz "
458 		    "quality %d\n", tc->tc_name, (uintmax_t)tc->tc_frequency,
459 		    tc->tc_quality);
460 	}
461 
462 	s = splclock();
463 
464 	tc->tc_next = timecounters;
465 	timecounters = tc;
466 	/*
467 	 * Never automatically use a timecounter with negative quality.
468 	 * Even though we run on the dummy counter, switching here may be
469 	 * worse since this timecounter may not be monotonous.
470 	 */
471 	if (tc->tc_quality < 0)
472 		goto out;
473 	if (tc->tc_quality < timecounter->tc_quality)
474 		goto out;
475 	if (tc->tc_quality == timecounter->tc_quality &&
476 	    tc->tc_frequency < timecounter->tc_frequency)
477 		goto out;
478 	(void)tc->tc_get_timecount(tc);
479 	(void)tc->tc_get_timecount(tc);
480 	timecounter = tc;
481 	tc_windup();
482 
483  out:
484 	splx(s);
485 }
486 
487 /* Report the frequency of the current timecounter. */
488 u_int64_t
489 tc_getfrequency(void)
490 {
491 
492 	return (timehands->th_counter->tc_frequency);
493 }
494 
495 /*
496  * Step our concept of UTC.  This is done by modifying our estimate of
497  * when we booted.
498  * XXX: not locked.
499  */
500 void
501 tc_setclock(struct timespec *ts)
502 {
503 	struct timespec ts2;
504 	struct bintime bt, bt2;
505 
506 	nsetclock.ev_count++;
507 	binuptime(&bt2);
508 	timespec2bintime(ts, &bt);
509 	bintime_sub(&bt, &bt2);
510 	bintime_add(&bt2, &timebasebin);
511 	timebasebin = bt;
512 
513 	/* XXX fiddle all the little crinkly bits around the fiords... */
514 	tc_windup();
515 	if (timestepwarnings) {
516 		bintime2timespec(&bt2, &ts2);
517 		log(LOG_INFO, "Time stepped from %jd.%09ld to %jd.%09ld\n",
518 		    (intmax_t)ts2.tv_sec, ts2.tv_nsec,
519 		    (intmax_t)ts->tv_sec, ts->tv_nsec);
520 	}
521 }
522 
523 /*
524  * Initialize the next struct timehands in the ring and make
525  * it the active timehands.  Along the way we might switch to a different
526  * timecounter and/or do seconds processing in NTP.  Slightly magic.
527  */
528 static void
529 tc_windup(void)
530 {
531 	struct bintime bt;
532 	struct timehands *th, *tho;
533 	u_int64_t scale;
534 	u_int delta, ncount, ogen;
535 	int i, s_update;
536 	time_t t;
537 
538 	s_update = 0;
539 
540 	/*
541 	 * Make the next timehands a copy of the current one, but do not
542 	 * overwrite the generation or next pointer.  While we update
543 	 * the contents, the generation must be zero.  Ensure global
544 	 * visibility of the generation before proceeding.
545 	 */
546 	tho = timehands;
547 	th = tho->th_next;
548 	ogen = th->th_generation;
549 	th->th_generation = 0;
550 	mb_write();
551 	bcopy(tho, th, offsetof(struct timehands, th_generation));
552 
553 	/*
554 	 * Capture a timecounter delta on the current timecounter and if
555 	 * changing timecounters, a counter value from the new timecounter.
556 	 * Update the offset fields accordingly.
557 	 */
558 	delta = tc_delta(th);
559 	if (th->th_counter != timecounter)
560 		ncount = timecounter->tc_get_timecount(timecounter);
561 	else
562 		ncount = 0;
563 	th->th_offset_count += delta;
564 	th->th_offset_count &= th->th_counter->tc_counter_mask;
565 	bintime_addx(&th->th_offset, th->th_scale * delta);
566 
567 	/*
568 	 * Hardware latching timecounters may not generate interrupts on
569 	 * PPS events, so instead we poll them.  There is a finite risk that
570 	 * the hardware might capture a count which is later than the one we
571 	 * got above, and therefore possibly in the next NTP second which might
572 	 * have a different rate than the current NTP second.  It doesn't
573 	 * matter in practice.
574 	 */
575 	if (tho->th_counter->tc_poll_pps)
576 		tho->th_counter->tc_poll_pps(tho->th_counter);
577 
578 	/*
579 	 * Deal with NTP second processing.  The for loop normally
580 	 * iterates at most once, but in extreme situations it might
581 	 * keep NTP sane if timeouts are not run for several seconds.
582 	 * At boot, the time step can be large when the TOD hardware
583 	 * has been read, so on really large steps, we call
584 	 * ntp_update_second only twice.  We need to call it twice in
585 	 * case we missed a leap second.
586 	 * If NTP is not compiled in ntp_update_second still calculates
587 	 * the adjustment resulting from adjtime() calls.
588 	 */
589 	bt = th->th_offset;
590 	bintime_add(&bt, &timebasebin);
591 	i = bt.sec - tho->th_microtime.tv_sec;
592 	if (i > LARGE_STEP)
593 		i = 2;
594 	for (; i > 0; i--) {
595 		t = bt.sec;
596 		ntp_update_second(&th->th_adjustment, &bt.sec);
597 		s_update = 1;
598 		if (bt.sec != t)
599 			timebasebin.sec += bt.sec - t;
600 	}
601 
602 	/* Update the UTC timestamps used by the get*() functions. */
603 	/* XXX shouldn't do this here.  Should force non-`get' versions. */
604 	bintime2timeval(&bt, &th->th_microtime);
605 	bintime2timespec(&bt, &th->th_nanotime);
606 
607 	/* Now is a good time to change timecounters. */
608 	if (th->th_counter != timecounter) {
609 		th->th_counter = timecounter;
610 		th->th_offset_count = ncount;
611 		s_update = 1;
612 	}
613 
614 	/*-
615 	 * Recalculate the scaling factor.  We want the number of 1/2^64
616 	 * fractions of a second per period of the hardware counter, taking
617 	 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
618 	 * processing provides us with.
619 	 *
620 	 * The th_adjustment is nanoseconds per second with 32 bit binary
621 	 * fraction and we want 64 bit binary fraction of second:
622 	 *
623 	 *	 x = a * 2^32 / 10^9 = a * 4.294967296
624 	 *
625 	 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
626 	 * we can only multiply by about 850 without overflowing, but that
627 	 * leaves suitably precise fractions for multiply before divide.
628 	 *
629 	 * Divide before multiply with a fraction of 2199/512 results in a
630 	 * systematic undercompensation of 10PPM of th_adjustment.  On a
631 	 * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
632  	 *
633 	 * We happily sacrifice the lowest of the 64 bits of our result
634 	 * to the goddess of code clarity.
635 	 *
636 	 */
637 	if (s_update) {
638 		scale = (u_int64_t)1 << 63;
639 		scale += (th->th_adjustment / 1024) * 2199;
640 		scale /= th->th_counter->tc_frequency;
641 		th->th_scale = scale * 2;
642 	}
643 	/*
644 	 * Now that the struct timehands is again consistent, set the new
645 	 * generation number, making sure to not make it zero.  Ensure
646 	 * changes are globally visible before changing.
647 	 */
648 	if (++ogen == 0)
649 		ogen = 1;
650 	mb_write();
651 	th->th_generation = ogen;
652 
653 	/*
654 	 * Go live with the new struct timehands.  Ensure changes are
655 	 * globally visible before changing.
656 	 */
657 	time_second = th->th_microtime.tv_sec;
658 	time_uptime = th->th_offset.sec;
659 	mb_write();
660 	timehands = th;
661 }
662 
663 #ifdef __FreeBSD__
664 /* Report or change the active timecounter hardware. */
665 static int
666 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
667 {
668 	char newname[32];
669 	struct timecounter *newtc, *tc;
670 	int error;
671 
672 	tc = timecounter;
673 	strlcpy(newname, tc->tc_name, sizeof(newname));
674 
675 	error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
676 	if (error != 0 || req->newptr == NULL ||
677 	    strcmp(newname, tc->tc_name) == 0)
678 		return (error);
679 
680 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
681 		if (strcmp(newname, newtc->tc_name) != 0)
682 			continue;
683 
684 		/* Warm up new timecounter. */
685 		(void)newtc->tc_get_timecount(newtc);
686 		(void)newtc->tc_get_timecount(newtc);
687 
688 		timecounter = newtc;
689 		return (0);
690 	}
691 	return (EINVAL);
692 }
693 
694 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
695     0, 0, sysctl_kern_timecounter_hardware, "A", "");
696 
697 
698 /* Report or change the active timecounter hardware. */
699 static int
700 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
701 {
702 	char buf[32], *spc;
703 	struct timecounter *tc;
704 	int error;
705 
706 	spc = "";
707 	error = 0;
708 	for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
709 		sprintf(buf, "%s%s(%d)",
710 		    spc, tc->tc_name, tc->tc_quality);
711 		error = SYSCTL_OUT(req, buf, strlen(buf));
712 		spc = " ";
713 	}
714 	return (error);
715 }
716 
717 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
718     0, 0, sysctl_kern_timecounter_choice, "A", "");
719 #endif /* __FreeBSD__ */
720 
721 /*
722  * RFC 2783 PPS-API implementation.
723  */
724 
725 int
726 pps_ioctl(u_long cmd, void *data, struct pps_state *pps)
727 {
728 	pps_params_t *app;
729 	pps_info_t *pipi;
730 #ifdef PPS_SYNC
731 	int *epi;
732 #endif
733 
734 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_ioctl") */
735 	switch (cmd) {
736 	case PPS_IOC_CREATE:
737 		return (0);
738 	case PPS_IOC_DESTROY:
739 		return (0);
740 	case PPS_IOC_SETPARAMS:
741 		app = (pps_params_t *)data;
742 		if (app->mode & ~pps->ppscap)
743 			return (EINVAL);
744 		pps->ppsparam = *app;
745 		return (0);
746 	case PPS_IOC_GETPARAMS:
747 		app = (pps_params_t *)data;
748 		*app = pps->ppsparam;
749 		app->api_version = PPS_API_VERS_1;
750 		return (0);
751 	case PPS_IOC_GETCAP:
752 		*(int*)data = pps->ppscap;
753 		return (0);
754 	case PPS_IOC_FETCH:
755 		pipi = (pps_info_t *)data;
756 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
757 		*pipi = pps->ppsinfo;
758 		return (0);
759 	case PPS_IOC_KCBIND:
760 #ifdef PPS_SYNC
761 		epi = (int *)data;
762 		/* XXX Only root should be able to do this */
763 		if (*epi & ~pps->ppscap)
764 			return (EINVAL);
765 		pps->kcmode = *epi;
766 		return (0);
767 #else
768 		return (EOPNOTSUPP);
769 #endif
770 	default:
771 		return (EPASSTHROUGH);
772 	}
773 }
774 
775 void
776 pps_init(struct pps_state *pps)
777 {
778 	pps->ppscap |= PPS_TSFMT_TSPEC;
779 	if (pps->ppscap & PPS_CAPTUREASSERT)
780 		pps->ppscap |= PPS_OFFSETASSERT;
781 	if (pps->ppscap & PPS_CAPTURECLEAR)
782 		pps->ppscap |= PPS_OFFSETCLEAR;
783 }
784 
785 void
786 pps_capture(struct pps_state *pps)
787 {
788 	struct timehands *th;
789 
790 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_capture") */
791 	th = timehands;
792 	pps->capgen = th->th_generation;
793 	pps->capth = th;
794 	pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
795 	if (pps->capgen != th->th_generation)
796 		pps->capgen = 0;
797 }
798 
799 void
800 pps_event(struct pps_state *pps, int event)
801 {
802 	struct bintime bt;
803 	struct timespec ts, *tsp, *osp;
804 	u_int tcount, *pcount;
805 	int foff, fhard;
806 	pps_seq_t *pseq;
807 
808 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_event") */
809 	/* If the timecounter was wound up underneath us, bail out. */
810 	if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
811 		return;
812 
813 	/* Things would be easier with arrays. */
814 	if (event == PPS_CAPTUREASSERT) {
815 		tsp = &pps->ppsinfo.assert_timestamp;
816 		osp = &pps->ppsparam.assert_offset;
817 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
818 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
819 		pcount = &pps->ppscount[0];
820 		pseq = &pps->ppsinfo.assert_sequence;
821 	} else {
822 		tsp = &pps->ppsinfo.clear_timestamp;
823 		osp = &pps->ppsparam.clear_offset;
824 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
825 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
826 		pcount = &pps->ppscount[1];
827 		pseq = &pps->ppsinfo.clear_sequence;
828 	}
829 
830 	/*
831 	 * If the timecounter changed, we cannot compare the count values, so
832 	 * we have to drop the rest of the PPS-stuff until the next event.
833 	 */
834 	if (pps->ppstc != pps->capth->th_counter) {
835 		pps->ppstc = pps->capth->th_counter;
836 		*pcount = pps->capcount;
837 		pps->ppscount[2] = pps->capcount;
838 		return;
839 	}
840 
841 	/* Convert the count to a timespec. */
842 	tcount = pps->capcount - pps->capth->th_offset_count;
843 	tcount &= pps->capth->th_counter->tc_counter_mask;
844 	bt = pps->capth->th_offset;
845 	bintime_addx(&bt, pps->capth->th_scale * tcount);
846 	bintime_add(&bt, &timebasebin);
847 	bintime2timespec(&bt, &ts);
848 
849 	/* If the timecounter was wound up underneath us, bail out. */
850 	if (pps->capgen != pps->capth->th_generation)
851 		return;
852 
853 	*pcount = pps->capcount;
854 	(*pseq)++;
855 	*tsp = ts;
856 
857 	if (foff) {
858 		timespecadd(tsp, osp, tsp);
859 		if (tsp->tv_nsec < 0) {
860 			tsp->tv_nsec += 1000000000;
861 			tsp->tv_sec -= 1;
862 		}
863 	}
864 #ifdef PPS_SYNC
865 	if (fhard) {
866 		u_int64_t scale;
867 
868 		/*
869 		 * Feed the NTP PLL/FLL.
870 		 * The FLL wants to know how many (hardware) nanoseconds
871 		 * elapsed since the previous event.
872 		 */
873 		tcount = pps->capcount - pps->ppscount[2];
874 		pps->ppscount[2] = pps->capcount;
875 		tcount &= pps->capth->th_counter->tc_counter_mask;
876 		scale = (u_int64_t)1 << 63;
877 		scale /= pps->capth->th_counter->tc_frequency;
878 		scale *= 2;
879 		bt.sec = 0;
880 		bt.frac = 0;
881 		bintime_addx(&bt, scale * tcount);
882 		bintime2timespec(&bt, &ts);
883 		hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
884 	}
885 #endif
886 }
887 
888 /*
889  * Timecounters need to be updated every so often to prevent the hardware
890  * counter from overflowing.  Updating also recalculates the cached values
891  * used by the get*() family of functions, so their precision depends on
892  * the update frequency.
893  */
894 
895 static int tc_tick;
896 #ifdef __FreeBSD__
897 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
898 #endif /* __FreeBSD__ */
899 
900 void
901 tc_ticktock(void)
902 {
903 	static int count;
904 
905 	if (++count < tc_tick)
906 		return;
907 	count = 0;
908 	tc_windup();
909 }
910 
911 void
912 inittimecounter(void)
913 {
914 	u_int p;
915 
916 	/*
917 	 * Set the initial timeout to
918 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
919 	 * People should probably not use the sysctl to set the timeout
920 	 * to smaller than its inital value, since that value is the
921 	 * smallest reasonable one.  If they want better timestamps they
922 	 * should use the non-"get"* functions.
923 	 */
924 	if (hz > 1000)
925 		tc_tick = (hz + 500) / 1000;
926 	else
927 		tc_tick = 1;
928 	p = (tc_tick * 1000000) / hz;
929 	aprint_verbose("timecounter: Timecounters tick every %d.%03u msec\n",
930 	    p / 1000, p % 1000);
931 
932 	/* warm up new timecounter (again) and get rolling. */
933 	(void)timecounter->tc_get_timecount(timecounter);
934 	(void)timecounter->tc_get_timecount(timecounter);
935 }
936 
937 #ifdef __FreeBSD__
938 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL)
939 #endif /* __FreeBSD__ */
940 #endif /* __HAVE_TIMECOUNTER */
941