xref: /netbsd-src/sys/kern/kern_tc.c (revision 33881f779a77dce6440bdc44610d94de75bebefe)
1 /* $NetBSD: kern_tc.c,v 1.54 2020/01/02 15:42:27 thorpej Exp $ */
2 
3 /*-
4  * Copyright (c) 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*-
33  * ----------------------------------------------------------------------------
34  * "THE BEER-WARE LICENSE" (Revision 42):
35  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
36  * can do whatever you want with this stuff. If we meet some day, and you think
37  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
38  * ---------------------------------------------------------------------------
39  */
40 
41 #include <sys/cdefs.h>
42 /* __FBSDID("$FreeBSD: src/sys/kern/kern_tc.c,v 1.166 2005/09/19 22:16:31 andre Exp $"); */
43 __KERNEL_RCSID(0, "$NetBSD: kern_tc.c,v 1.54 2020/01/02 15:42:27 thorpej Exp $");
44 
45 #ifdef _KERNEL_OPT
46 #include "opt_ntp.h"
47 #endif
48 
49 #include <sys/param.h>
50 #include <sys/kernel.h>
51 #include <sys/reboot.h>	/* XXX just to get AB_VERBOSE */
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 #include <sys/systm.h>
55 #include <sys/timepps.h>
56 #include <sys/timetc.h>
57 #include <sys/timex.h>
58 #include <sys/evcnt.h>
59 #include <sys/kauth.h>
60 #include <sys/mutex.h>
61 #include <sys/atomic.h>
62 #include <sys/xcall.h>
63 
64 /*
65  * A large step happens on boot.  This constant detects such steps.
66  * It is relatively small so that ntp_update_second gets called enough
67  * in the typical 'missed a couple of seconds' case, but doesn't loop
68  * forever when the time step is large.
69  */
70 #define LARGE_STEP	200
71 
72 /*
73  * Implement a dummy timecounter which we can use until we get a real one
74  * in the air.  This allows the console and other early stuff to use
75  * time services.
76  */
77 
78 static u_int
79 dummy_get_timecount(struct timecounter *tc)
80 {
81 	static u_int now;
82 
83 	return (++now);
84 }
85 
86 static struct timecounter dummy_timecounter = {
87 	.tc_get_timecount	= dummy_get_timecount,
88 	.tc_counter_mask	= ~0u,
89 	.tc_frequency		= 1000000,
90 	.tc_name		= "dummy",
91 	.tc_quality		= -1000000,
92 	.tc_priv		= NULL,
93 };
94 
95 struct timehands {
96 	/* These fields must be initialized by the driver. */
97 	struct timecounter	*th_counter;     /* active timecounter */
98 	int64_t			th_adjustment;   /* frequency adjustment */
99 						 /* (NTP/adjtime) */
100 	u_int64_t		th_scale;        /* scale factor (counter */
101 						 /* tick->time) */
102 	u_int64_t 		th_offset_count; /* offset at last time */
103 						 /* update (tc_windup()) */
104 	struct bintime		th_offset;       /* bin (up)time at windup */
105 	struct timeval		th_microtime;    /* cached microtime */
106 	struct timespec		th_nanotime;     /* cached nanotime */
107 	/* Fields not to be copied in tc_windup start with th_generation. */
108 	volatile u_int		th_generation;   /* current genration */
109 	struct timehands	*th_next;        /* next timehand */
110 };
111 
112 static struct timehands th0;
113 static struct timehands th9 = { .th_next = &th0, };
114 static struct timehands th8 = { .th_next = &th9, };
115 static struct timehands th7 = { .th_next = &th8, };
116 static struct timehands th6 = { .th_next = &th7, };
117 static struct timehands th5 = { .th_next = &th6, };
118 static struct timehands th4 = { .th_next = &th5, };
119 static struct timehands th3 = { .th_next = &th4, };
120 static struct timehands th2 = { .th_next = &th3, };
121 static struct timehands th1 = { .th_next = &th2, };
122 static struct timehands th0 = {
123 	.th_counter = &dummy_timecounter,
124 	.th_scale = (uint64_t)-1 / 1000000,
125 	.th_offset = { .sec = 1, .frac = 0 },
126 	.th_generation = 1,
127 	.th_next = &th1,
128 };
129 
130 static struct timehands *volatile timehands = &th0;
131 struct timecounter *timecounter = &dummy_timecounter;
132 static struct timecounter *timecounters = &dummy_timecounter;
133 
134 volatile time_t time_second __cacheline_aligned = 1;
135 volatile time_t time_uptime __cacheline_aligned = 1;
136 
137 static struct bintime timebasebin;
138 
139 static int timestepwarnings;
140 
141 kmutex_t timecounter_lock;
142 static u_int timecounter_mods;
143 static volatile int timecounter_removals = 1;
144 static u_int timecounter_bad;
145 
146 #ifdef __FreeBSD__
147 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
148     &timestepwarnings, 0, "");
149 #endif /* __FreeBSD__ */
150 
151 /*
152  * sysctl helper routine for kern.timercounter.hardware
153  */
154 static int
155 sysctl_kern_timecounter_hardware(SYSCTLFN_ARGS)
156 {
157 	struct sysctlnode node;
158 	int error;
159 	char newname[MAX_TCNAMELEN];
160 	struct timecounter *newtc, *tc;
161 
162 	tc = timecounter;
163 
164 	strlcpy(newname, tc->tc_name, sizeof(newname));
165 
166 	node = *rnode;
167 	node.sysctl_data = newname;
168 	node.sysctl_size = sizeof(newname);
169 
170 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
171 
172 	if (error ||
173 	    newp == NULL ||
174 	    strncmp(newname, tc->tc_name, sizeof(newname)) == 0)
175 		return error;
176 
177 	if (l != NULL && (error = kauth_authorize_system(l->l_cred,
178 	    KAUTH_SYSTEM_TIME, KAUTH_REQ_SYSTEM_TIME_TIMECOUNTERS, newname,
179 	    NULL, NULL)) != 0)
180 		return (error);
181 
182 	if (!cold)
183 		mutex_spin_enter(&timecounter_lock);
184 	error = EINVAL;
185 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
186 		if (strcmp(newname, newtc->tc_name) != 0)
187 			continue;
188 		/* Warm up new timecounter. */
189 		(void)newtc->tc_get_timecount(newtc);
190 		(void)newtc->tc_get_timecount(newtc);
191 		timecounter = newtc;
192 		error = 0;
193 		break;
194 	}
195 	if (!cold)
196 		mutex_spin_exit(&timecounter_lock);
197 	return error;
198 }
199 
200 static int
201 sysctl_kern_timecounter_choice(SYSCTLFN_ARGS)
202 {
203 	char buf[MAX_TCNAMELEN+48];
204 	char *where;
205 	const char *spc;
206 	struct timecounter *tc;
207 	size_t needed, left, slen;
208 	int error, mods;
209 
210 	if (newp != NULL)
211 		return (EPERM);
212 	if (namelen != 0)
213 		return (EINVAL);
214 
215 	mutex_spin_enter(&timecounter_lock);
216  retry:
217 	spc = "";
218 	error = 0;
219 	needed = 0;
220 	left = *oldlenp;
221 	where = oldp;
222 	for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
223 		if (where == NULL) {
224 			needed += sizeof(buf);  /* be conservative */
225 		} else {
226 			slen = snprintf(buf, sizeof(buf), "%s%s(q=%d, f=%" PRId64
227 					" Hz)", spc, tc->tc_name, tc->tc_quality,
228 					tc->tc_frequency);
229 			if (left < slen + 1)
230 				break;
231 		 	mods = timecounter_mods;
232 			mutex_spin_exit(&timecounter_lock);
233 			error = copyout(buf, where, slen + 1);
234 			mutex_spin_enter(&timecounter_lock);
235 			if (mods != timecounter_mods) {
236 				goto retry;
237 			}
238 			spc = " ";
239 			where += slen;
240 			needed += slen;
241 			left -= slen;
242 		}
243 	}
244 	mutex_spin_exit(&timecounter_lock);
245 
246 	*oldlenp = needed;
247 	return (error);
248 }
249 
250 SYSCTL_SETUP(sysctl_timecounter_setup, "sysctl timecounter setup")
251 {
252 	const struct sysctlnode *node;
253 
254 	sysctl_createv(clog, 0, NULL, &node,
255 		       CTLFLAG_PERMANENT,
256 		       CTLTYPE_NODE, "timecounter",
257 		       SYSCTL_DESCR("time counter information"),
258 		       NULL, 0, NULL, 0,
259 		       CTL_KERN, CTL_CREATE, CTL_EOL);
260 
261 	if (node != NULL) {
262 		sysctl_createv(clog, 0, NULL, NULL,
263 			       CTLFLAG_PERMANENT,
264 			       CTLTYPE_STRING, "choice",
265 			       SYSCTL_DESCR("available counters"),
266 			       sysctl_kern_timecounter_choice, 0, NULL, 0,
267 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
268 
269 		sysctl_createv(clog, 0, NULL, NULL,
270 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
271 			       CTLTYPE_STRING, "hardware",
272 			       SYSCTL_DESCR("currently active time counter"),
273 			       sysctl_kern_timecounter_hardware, 0, NULL, MAX_TCNAMELEN,
274 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
275 
276 		sysctl_createv(clog, 0, NULL, NULL,
277 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
278 			       CTLTYPE_INT, "timestepwarnings",
279 			       SYSCTL_DESCR("log time steps"),
280 			       NULL, 0, &timestepwarnings, 0,
281 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
282 	}
283 }
284 
285 #ifdef TC_COUNTERS
286 #define	TC_STATS(name)							\
287 static struct evcnt n##name =						\
288     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "timecounter", #name);	\
289 EVCNT_ATTACH_STATIC(n##name)
290 TC_STATS(binuptime);    TC_STATS(nanouptime);    TC_STATS(microuptime);
291 TC_STATS(bintime);      TC_STATS(nanotime);      TC_STATS(microtime);
292 TC_STATS(getbinuptime); TC_STATS(getnanouptime); TC_STATS(getmicrouptime);
293 TC_STATS(getbintime);   TC_STATS(getnanotime);   TC_STATS(getmicrotime);
294 TC_STATS(setclock);
295 #define	TC_COUNT(var)	var.ev_count++
296 #undef TC_STATS
297 #else
298 #define	TC_COUNT(var)	/* nothing */
299 #endif	/* TC_COUNTERS */
300 
301 static void tc_windup(void);
302 
303 /*
304  * Return the difference between the timehands' counter value now and what
305  * was when we copied it to the timehands' offset_count.
306  */
307 static inline u_int
308 tc_delta(struct timehands *th)
309 {
310 	struct timecounter *tc;
311 
312 	tc = th->th_counter;
313 	return ((tc->tc_get_timecount(tc) -
314 		 th->th_offset_count) & tc->tc_counter_mask);
315 }
316 
317 /*
318  * Functions for reading the time.  We have to loop until we are sure that
319  * the timehands that we operated on was not updated under our feet.  See
320  * the comment in <sys/timevar.h> for a description of these 12 functions.
321  */
322 
323 void
324 binuptime(struct bintime *bt)
325 {
326 	struct timehands *th;
327 	lwp_t *l;
328 	u_int lgen, gen;
329 
330 	TC_COUNT(nbinuptime);
331 
332 	/*
333 	 * Provide exclusion against tc_detach().
334 	 *
335 	 * We record the number of timecounter removals before accessing
336 	 * timecounter state.  Note that the LWP can be using multiple
337 	 * "generations" at once, due to interrupts (interrupted while in
338 	 * this function).  Hardware interrupts will borrow the interrupted
339 	 * LWP's l_tcgen value for this purpose, and can themselves be
340 	 * interrupted by higher priority interrupts.  In this case we need
341 	 * to ensure that the oldest generation in use is recorded.
342 	 *
343 	 * splsched() is too expensive to use, so we take care to structure
344 	 * this code in such a way that it is not required.  Likewise, we
345 	 * do not disable preemption.
346 	 *
347 	 * Memory barriers are also too expensive to use for such a
348 	 * performance critical function.  The good news is that we do not
349 	 * need memory barriers for this type of exclusion, as the thread
350 	 * updating timecounter_removals will issue a broadcast cross call
351 	 * before inspecting our l_tcgen value (this elides memory ordering
352 	 * issues).
353 	 */
354 	l = curlwp;
355 	lgen = l->l_tcgen;
356 	if (__predict_true(lgen == 0)) {
357 		l->l_tcgen = timecounter_removals;
358 	}
359 	__insn_barrier();
360 
361 	do {
362 		th = timehands;
363 		gen = th->th_generation;
364 		*bt = th->th_offset;
365 		bintime_addx(bt, th->th_scale * tc_delta(th));
366 	} while (gen == 0 || gen != th->th_generation);
367 
368 	__insn_barrier();
369 	l->l_tcgen = lgen;
370 }
371 
372 void
373 nanouptime(struct timespec *tsp)
374 {
375 	struct bintime bt;
376 
377 	TC_COUNT(nnanouptime);
378 	binuptime(&bt);
379 	bintime2timespec(&bt, tsp);
380 }
381 
382 void
383 microuptime(struct timeval *tvp)
384 {
385 	struct bintime bt;
386 
387 	TC_COUNT(nmicrouptime);
388 	binuptime(&bt);
389 	bintime2timeval(&bt, tvp);
390 }
391 
392 void
393 bintime(struct bintime *bt)
394 {
395 
396 	TC_COUNT(nbintime);
397 	binuptime(bt);
398 	bintime_add(bt, &timebasebin);
399 }
400 
401 void
402 nanotime(struct timespec *tsp)
403 {
404 	struct bintime bt;
405 
406 	TC_COUNT(nnanotime);
407 	bintime(&bt);
408 	bintime2timespec(&bt, tsp);
409 }
410 
411 void
412 microtime(struct timeval *tvp)
413 {
414 	struct bintime bt;
415 
416 	TC_COUNT(nmicrotime);
417 	bintime(&bt);
418 	bintime2timeval(&bt, tvp);
419 }
420 
421 void
422 getbinuptime(struct bintime *bt)
423 {
424 	struct timehands *th;
425 	u_int gen;
426 
427 	TC_COUNT(ngetbinuptime);
428 	do {
429 		th = timehands;
430 		gen = th->th_generation;
431 		*bt = th->th_offset;
432 	} while (gen == 0 || gen != th->th_generation);
433 }
434 
435 void
436 getnanouptime(struct timespec *tsp)
437 {
438 	struct timehands *th;
439 	u_int gen;
440 
441 	TC_COUNT(ngetnanouptime);
442 	do {
443 		th = timehands;
444 		gen = th->th_generation;
445 		bintime2timespec(&th->th_offset, tsp);
446 	} while (gen == 0 || gen != th->th_generation);
447 }
448 
449 void
450 getmicrouptime(struct timeval *tvp)
451 {
452 	struct timehands *th;
453 	u_int gen;
454 
455 	TC_COUNT(ngetmicrouptime);
456 	do {
457 		th = timehands;
458 		gen = th->th_generation;
459 		bintime2timeval(&th->th_offset, tvp);
460 	} while (gen == 0 || gen != th->th_generation);
461 }
462 
463 void
464 getbintime(struct bintime *bt)
465 {
466 	struct timehands *th;
467 	u_int gen;
468 
469 	TC_COUNT(ngetbintime);
470 	do {
471 		th = timehands;
472 		gen = th->th_generation;
473 		*bt = th->th_offset;
474 	} while (gen == 0 || gen != th->th_generation);
475 	bintime_add(bt, &timebasebin);
476 }
477 
478 static inline void
479 dogetnanotime(struct timespec *tsp)
480 {
481 	struct timehands *th;
482 	u_int gen;
483 
484 	TC_COUNT(ngetnanotime);
485 	do {
486 		th = timehands;
487 		gen = th->th_generation;
488 		*tsp = th->th_nanotime;
489 	} while (gen == 0 || gen != th->th_generation);
490 }
491 
492 void
493 getnanotime(struct timespec *tsp)
494 {
495 
496 	dogetnanotime(tsp);
497 }
498 
499 void dtrace_getnanotime(struct timespec *tsp);
500 
501 void
502 dtrace_getnanotime(struct timespec *tsp)
503 {
504 
505 	dogetnanotime(tsp);
506 }
507 
508 void
509 getmicrotime(struct timeval *tvp)
510 {
511 	struct timehands *th;
512 	u_int gen;
513 
514 	TC_COUNT(ngetmicrotime);
515 	do {
516 		th = timehands;
517 		gen = th->th_generation;
518 		*tvp = th->th_microtime;
519 	} while (gen == 0 || gen != th->th_generation);
520 }
521 
522 void
523 getnanoboottime(struct timespec *tsp)
524 {
525 	struct bintime bt;
526 
527 	getbinboottime(&bt);
528 	bintime2timespec(&bt, tsp);
529 }
530 
531 void
532 getmicroboottime(struct timeval *tvp)
533 {
534 	struct bintime bt;
535 
536 	getbinboottime(&bt);
537 	bintime2timeval(&bt, tvp);
538 }
539 
540 void
541 getbinboottime(struct bintime *bt)
542 {
543 
544 	/*
545 	 * XXX Need lockless read synchronization around timebasebin
546 	 * (and not just here).
547 	 */
548 	*bt = timebasebin;
549 }
550 
551 /*
552  * Initialize a new timecounter and possibly use it.
553  */
554 void
555 tc_init(struct timecounter *tc)
556 {
557 	u_int u;
558 
559 	u = tc->tc_frequency / tc->tc_counter_mask;
560 	/* XXX: We need some margin here, 10% is a guess */
561 	u *= 11;
562 	u /= 10;
563 	if (u > hz && tc->tc_quality >= 0) {
564 		tc->tc_quality = -2000;
565 		aprint_verbose(
566 		    "timecounter: Timecounter \"%s\" frequency %ju Hz",
567 			    tc->tc_name, (uintmax_t)tc->tc_frequency);
568 		aprint_verbose(" -- Insufficient hz, needs at least %u\n", u);
569 	} else if (tc->tc_quality >= 0 || bootverbose) {
570 		aprint_verbose(
571 		    "timecounter: Timecounter \"%s\" frequency %ju Hz "
572 		    "quality %d\n", tc->tc_name, (uintmax_t)tc->tc_frequency,
573 		    tc->tc_quality);
574 	}
575 
576 	mutex_spin_enter(&timecounter_lock);
577 	tc->tc_next = timecounters;
578 	timecounters = tc;
579 	timecounter_mods++;
580 	/*
581 	 * Never automatically use a timecounter with negative quality.
582 	 * Even though we run on the dummy counter, switching here may be
583 	 * worse since this timecounter may not be monotonous.
584 	 */
585 	if (tc->tc_quality >= 0 && (tc->tc_quality > timecounter->tc_quality ||
586 	    (tc->tc_quality == timecounter->tc_quality &&
587 	    tc->tc_frequency > timecounter->tc_frequency))) {
588 		(void)tc->tc_get_timecount(tc);
589 		(void)tc->tc_get_timecount(tc);
590 		timecounter = tc;
591 		tc_windup();
592 	}
593 	mutex_spin_exit(&timecounter_lock);
594 }
595 
596 /*
597  * Pick a new timecounter due to the existing counter going bad.
598  */
599 static void
600 tc_pick(void)
601 {
602 	struct timecounter *best, *tc;
603 
604 	KASSERT(mutex_owned(&timecounter_lock));
605 
606 	for (best = tc = timecounters; tc != NULL; tc = tc->tc_next) {
607 		if (tc->tc_quality > best->tc_quality)
608 			best = tc;
609 		else if (tc->tc_quality < best->tc_quality)
610 			continue;
611 		else if (tc->tc_frequency > best->tc_frequency)
612 			best = tc;
613 	}
614 	(void)best->tc_get_timecount(best);
615 	(void)best->tc_get_timecount(best);
616 	timecounter = best;
617 }
618 
619 /*
620  * A timecounter has gone bad, arrange to pick a new one at the next
621  * clock tick.
622  */
623 void
624 tc_gonebad(struct timecounter *tc)
625 {
626 
627 	tc->tc_quality = -100;
628 	membar_producer();
629 	atomic_inc_uint(&timecounter_bad);
630 }
631 
632 /*
633  * Stop using a timecounter and remove it from the timecounters list.
634  */
635 int
636 tc_detach(struct timecounter *target)
637 {
638 	struct timecounter *tc;
639 	struct timecounter **tcp = NULL;
640 	int removals;
641 	lwp_t *l;
642 
643 	/* First, find the timecounter. */
644 	mutex_spin_enter(&timecounter_lock);
645 	for (tcp = &timecounters, tc = timecounters;
646 	     tc != NULL;
647 	     tcp = &tc->tc_next, tc = tc->tc_next) {
648 		if (tc == target)
649 			break;
650 	}
651 	if (tc == NULL) {
652 		mutex_spin_exit(&timecounter_lock);
653 		return ESRCH;
654 	}
655 
656 	/* And now, remove it. */
657 	*tcp = tc->tc_next;
658 	if (timecounter == target) {
659 		tc_pick();
660 		tc_windup();
661 	}
662 	timecounter_mods++;
663 	removals = timecounter_removals++;
664 	mutex_spin_exit(&timecounter_lock);
665 
666 	/*
667 	 * We now have to determine if any threads in the system are still
668 	 * making use of this timecounter.
669 	 *
670 	 * We issue a broadcast cross call to elide memory ordering issues,
671 	 * then scan all LWPs in the system looking at each's timecounter
672 	 * generation number.  We need to see a value of zero (not actively
673 	 * using a timecounter) or a value greater than our removal value.
674 	 *
675 	 * We may race with threads that read `timecounter_removals' and
676 	 * and then get preempted before updating `l_tcgen'.  This is not
677 	 * a problem, since it means that these threads have not yet started
678 	 * accessing timecounter state.  All we do need is one clean
679 	 * snapshot of the system where every thread appears not to be using
680 	 * old timecounter state.
681 	 */
682 	for (;;) {
683 		xc_barrier(0);
684 
685 		mutex_enter(proc_lock);
686 		LIST_FOREACH(l, &alllwp, l_list) {
687 			if (l->l_tcgen == 0 || l->l_tcgen > removals) {
688 				/*
689 				 * Not using timecounter or old timecounter
690 				 * state at time of our xcall or later.
691 				 */
692 				continue;
693 			}
694 			break;
695 		}
696 		mutex_exit(proc_lock);
697 
698 		/*
699 		 * If the timecounter is still in use, wait at least 10ms
700 		 * before retrying.
701 		 */
702 		if (l == NULL) {
703 			return 0;
704 		}
705 		(void)kpause("tcdetach", false, mstohz(10), NULL);
706 	}
707 }
708 
709 /* Report the frequency of the current timecounter. */
710 u_int64_t
711 tc_getfrequency(void)
712 {
713 
714 	return (timehands->th_counter->tc_frequency);
715 }
716 
717 /*
718  * Step our concept of UTC.  This is done by modifying our estimate of
719  * when we booted.
720  */
721 void
722 tc_setclock(const struct timespec *ts)
723 {
724 	struct timespec ts2;
725 	struct bintime bt, bt2;
726 
727 	mutex_spin_enter(&timecounter_lock);
728 	TC_COUNT(nsetclock);
729 	binuptime(&bt2);
730 	timespec2bintime(ts, &bt);
731 	bintime_sub(&bt, &bt2);
732 	bintime_add(&bt2, &timebasebin);
733 	timebasebin = bt;
734 	tc_windup();
735 	mutex_spin_exit(&timecounter_lock);
736 
737 	if (timestepwarnings) {
738 		bintime2timespec(&bt2, &ts2);
739 		log(LOG_INFO,
740 		    "Time stepped from %lld.%09ld to %lld.%09ld\n",
741 		    (long long)ts2.tv_sec, ts2.tv_nsec,
742 		    (long long)ts->tv_sec, ts->tv_nsec);
743 	}
744 }
745 
746 /*
747  * Initialize the next struct timehands in the ring and make
748  * it the active timehands.  Along the way we might switch to a different
749  * timecounter and/or do seconds processing in NTP.  Slightly magic.
750  */
751 static void
752 tc_windup(void)
753 {
754 	struct bintime bt;
755 	struct timehands *th, *tho;
756 	u_int64_t scale;
757 	u_int delta, ncount, ogen;
758 	int i, s_update;
759 	time_t t;
760 
761 	KASSERT(mutex_owned(&timecounter_lock));
762 
763 	s_update = 0;
764 
765 	/*
766 	 * Make the next timehands a copy of the current one, but do not
767 	 * overwrite the generation or next pointer.  While we update
768 	 * the contents, the generation must be zero.  Ensure global
769 	 * visibility of the generation before proceeding.
770 	 */
771 	tho = timehands;
772 	th = tho->th_next;
773 	ogen = th->th_generation;
774 	th->th_generation = 0;
775 	membar_producer();
776 	bcopy(tho, th, offsetof(struct timehands, th_generation));
777 
778 	/*
779 	 * Capture a timecounter delta on the current timecounter and if
780 	 * changing timecounters, a counter value from the new timecounter.
781 	 * Update the offset fields accordingly.
782 	 */
783 	delta = tc_delta(th);
784 	if (th->th_counter != timecounter)
785 		ncount = timecounter->tc_get_timecount(timecounter);
786 	else
787 		ncount = 0;
788 	th->th_offset_count += delta;
789 	bintime_addx(&th->th_offset, th->th_scale * delta);
790 
791 	/*
792 	 * Hardware latching timecounters may not generate interrupts on
793 	 * PPS events, so instead we poll them.  There is a finite risk that
794 	 * the hardware might capture a count which is later than the one we
795 	 * got above, and therefore possibly in the next NTP second which might
796 	 * have a different rate than the current NTP second.  It doesn't
797 	 * matter in practice.
798 	 */
799 	if (tho->th_counter->tc_poll_pps)
800 		tho->th_counter->tc_poll_pps(tho->th_counter);
801 
802 	/*
803 	 * Deal with NTP second processing.  The for loop normally
804 	 * iterates at most once, but in extreme situations it might
805 	 * keep NTP sane if timeouts are not run for several seconds.
806 	 * At boot, the time step can be large when the TOD hardware
807 	 * has been read, so on really large steps, we call
808 	 * ntp_update_second only twice.  We need to call it twice in
809 	 * case we missed a leap second.
810 	 * If NTP is not compiled in ntp_update_second still calculates
811 	 * the adjustment resulting from adjtime() calls.
812 	 */
813 	bt = th->th_offset;
814 	bintime_add(&bt, &timebasebin);
815 	i = bt.sec - tho->th_microtime.tv_sec;
816 	if (i > LARGE_STEP)
817 		i = 2;
818 	for (; i > 0; i--) {
819 		t = bt.sec;
820 		ntp_update_second(&th->th_adjustment, &bt.sec);
821 		s_update = 1;
822 		if (bt.sec != t)
823 			timebasebin.sec += bt.sec - t;
824 	}
825 
826 	/* Update the UTC timestamps used by the get*() functions. */
827 	/* XXX shouldn't do this here.  Should force non-`get' versions. */
828 	bintime2timeval(&bt, &th->th_microtime);
829 	bintime2timespec(&bt, &th->th_nanotime);
830 	/* Now is a good time to change timecounters. */
831 	if (th->th_counter != timecounter) {
832 		th->th_counter = timecounter;
833 		th->th_offset_count = ncount;
834 		s_update = 1;
835 	}
836 
837 	/*-
838 	 * Recalculate the scaling factor.  We want the number of 1/2^64
839 	 * fractions of a second per period of the hardware counter, taking
840 	 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
841 	 * processing provides us with.
842 	 *
843 	 * The th_adjustment is nanoseconds per second with 32 bit binary
844 	 * fraction and we want 64 bit binary fraction of second:
845 	 *
846 	 *	 x = a * 2^32 / 10^9 = a * 4.294967296
847 	 *
848 	 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
849 	 * we can only multiply by about 850 without overflowing, but that
850 	 * leaves suitably precise fractions for multiply before divide.
851 	 *
852 	 * Divide before multiply with a fraction of 2199/512 results in a
853 	 * systematic undercompensation of 10PPM of th_adjustment.  On a
854 	 * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
855  	 *
856 	 * We happily sacrifice the lowest of the 64 bits of our result
857 	 * to the goddess of code clarity.
858 	 *
859 	 */
860 	if (s_update) {
861 		scale = (u_int64_t)1 << 63;
862 		scale += (th->th_adjustment / 1024) * 2199;
863 		scale /= th->th_counter->tc_frequency;
864 		th->th_scale = scale * 2;
865 	}
866 	/*
867 	 * Now that the struct timehands is again consistent, set the new
868 	 * generation number, making sure to not make it zero.  Ensure
869 	 * changes are globally visible before changing.
870 	 */
871 	if (++ogen == 0)
872 		ogen = 1;
873 	membar_producer();
874 	th->th_generation = ogen;
875 
876 	/*
877 	 * Go live with the new struct timehands.  Ensure changes are
878 	 * globally visible before changing.
879 	 */
880 	time_second = th->th_microtime.tv_sec;
881 	time_uptime = th->th_offset.sec;
882 	membar_producer();
883 	timehands = th;
884 
885 	/*
886 	 * Force users of the old timehand to move on.  This is
887 	 * necessary for MP systems; we need to ensure that the
888 	 * consumers will move away from the old timehand before
889 	 * we begin updating it again when we eventually wrap
890 	 * around.
891 	 */
892 	if (++tho->th_generation == 0)
893 		tho->th_generation = 1;
894 }
895 
896 /*
897  * RFC 2783 PPS-API implementation.
898  */
899 
900 int
901 pps_ioctl(u_long cmd, void *data, struct pps_state *pps)
902 {
903 	pps_params_t *app;
904 	pps_info_t *pipi;
905 #ifdef PPS_SYNC
906 	int *epi;
907 #endif
908 
909 	KASSERT(mutex_owned(&timecounter_lock));
910 
911 	KASSERT(pps != NULL);
912 
913 	switch (cmd) {
914 	case PPS_IOC_CREATE:
915 		return (0);
916 	case PPS_IOC_DESTROY:
917 		return (0);
918 	case PPS_IOC_SETPARAMS:
919 		app = (pps_params_t *)data;
920 		if (app->mode & ~pps->ppscap)
921 			return (EINVAL);
922 		pps->ppsparam = *app;
923 		return (0);
924 	case PPS_IOC_GETPARAMS:
925 		app = (pps_params_t *)data;
926 		*app = pps->ppsparam;
927 		app->api_version = PPS_API_VERS_1;
928 		return (0);
929 	case PPS_IOC_GETCAP:
930 		*(int*)data = pps->ppscap;
931 		return (0);
932 	case PPS_IOC_FETCH:
933 		pipi = (pps_info_t *)data;
934 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
935 		*pipi = pps->ppsinfo;
936 		return (0);
937 	case PPS_IOC_KCBIND:
938 #ifdef PPS_SYNC
939 		epi = (int *)data;
940 		/* XXX Only root should be able to do this */
941 		if (*epi & ~pps->ppscap)
942 			return (EINVAL);
943 		pps->kcmode = *epi;
944 		return (0);
945 #else
946 		return (EOPNOTSUPP);
947 #endif
948 	default:
949 		return (EPASSTHROUGH);
950 	}
951 }
952 
953 void
954 pps_init(struct pps_state *pps)
955 {
956 
957 	KASSERT(mutex_owned(&timecounter_lock));
958 
959 	pps->ppscap |= PPS_TSFMT_TSPEC;
960 	if (pps->ppscap & PPS_CAPTUREASSERT)
961 		pps->ppscap |= PPS_OFFSETASSERT;
962 	if (pps->ppscap & PPS_CAPTURECLEAR)
963 		pps->ppscap |= PPS_OFFSETCLEAR;
964 }
965 
966 /*
967  * capture a timetamp in the pps structure
968  */
969 void
970 pps_capture(struct pps_state *pps)
971 {
972 	struct timehands *th;
973 
974 	KASSERT(mutex_owned(&timecounter_lock));
975 	KASSERT(pps != NULL);
976 
977 	th = timehands;
978 	pps->capgen = th->th_generation;
979 	pps->capth = th;
980 	pps->capcount = (u_int64_t)tc_delta(th) + th->th_offset_count;
981 	if (pps->capgen != th->th_generation)
982 		pps->capgen = 0;
983 }
984 
985 #ifdef PPS_DEBUG
986 int ppsdebug = 0;
987 #endif
988 
989 /*
990  * process a pps_capture()ed event
991  */
992 void
993 pps_event(struct pps_state *pps, int event)
994 {
995 	pps_ref_event(pps, event, NULL, PPS_REFEVNT_PPS|PPS_REFEVNT_CAPTURE);
996 }
997 
998 /*
999  * extended pps api /  kernel pll/fll entry point
1000  *
1001  * feed reference time stamps to PPS engine
1002  *
1003  * will simulate a PPS event and feed
1004  * the NTP PLL/FLL if requested.
1005  *
1006  * the ref time stamps should be roughly once
1007  * a second but do not need to be exactly in phase
1008  * with the UTC second but should be close to it.
1009  * this relaxation of requirements allows callout
1010  * driven timestamping mechanisms to feed to pps
1011  * capture/kernel pll logic.
1012  *
1013  * calling pattern is:
1014  *  pps_capture() (for PPS_REFEVNT_{CAPTURE|CAPCUR})
1015  *  read timestamp from reference source
1016  *  pps_ref_event()
1017  *
1018  * supported refmodes:
1019  *  PPS_REFEVNT_CAPTURE
1020  *    use system timestamp of pps_capture()
1021  *  PPS_REFEVNT_CURRENT
1022  *    use system timestamp of this call
1023  *  PPS_REFEVNT_CAPCUR
1024  *    use average of read capture and current system time stamp
1025  *  PPS_REFEVNT_PPS
1026  *    assume timestamp on second mark - ref_ts is ignored
1027  *
1028  */
1029 
1030 void
1031 pps_ref_event(struct pps_state *pps,
1032 	      int event,
1033 	      struct bintime *ref_ts,
1034 	      int refmode
1035 	)
1036 {
1037 	struct bintime bt;	/* current time */
1038 	struct bintime btd;	/* time difference */
1039 	struct bintime bt_ref;	/* reference time */
1040 	struct timespec ts, *tsp, *osp;
1041 	struct timehands *th;
1042 	u_int64_t tcount, acount, dcount, *pcount;
1043 	int foff, gen;
1044 #ifdef PPS_SYNC
1045 	int fhard;
1046 #endif
1047 	pps_seq_t *pseq;
1048 
1049 	KASSERT(mutex_owned(&timecounter_lock));
1050 
1051 	KASSERT(pps != NULL);
1052 
1053         /* pick up current time stamp if needed */
1054 	if (refmode & (PPS_REFEVNT_CURRENT|PPS_REFEVNT_CAPCUR)) {
1055 		/* pick up current time stamp */
1056 		th = timehands;
1057 		gen = th->th_generation;
1058 		tcount = (u_int64_t)tc_delta(th) + th->th_offset_count;
1059 		if (gen != th->th_generation)
1060 			gen = 0;
1061 
1062 		/* If the timecounter was wound up underneath us, bail out. */
1063 		if (pps->capgen == 0 ||
1064 		    pps->capgen != pps->capth->th_generation ||
1065 		    gen == 0 ||
1066 		    gen != pps->capgen) {
1067 #ifdef PPS_DEBUG
1068 			if (ppsdebug & 0x1) {
1069 				log(LOG_DEBUG,
1070 				    "pps_ref_event(pps=%p, event=%d, ...): DROP (wind-up)\n",
1071 				    pps, event);
1072 			}
1073 #endif
1074 			return;
1075 		}
1076 	} else {
1077 		tcount = 0;	/* keep GCC happy */
1078 	}
1079 
1080 #ifdef PPS_DEBUG
1081 	if (ppsdebug & 0x1) {
1082 		struct timespec tmsp;
1083 
1084 		if (ref_ts == NULL) {
1085 			tmsp.tv_sec = 0;
1086 			tmsp.tv_nsec = 0;
1087 		} else {
1088 			bintime2timespec(ref_ts, &tmsp);
1089 		}
1090 
1091 		log(LOG_DEBUG,
1092 		    "pps_ref_event(pps=%p, event=%d, ref_ts=%"PRIi64
1093 		    ".%09"PRIi32", refmode=0x%1x)\n",
1094 		    pps, event, tmsp.tv_sec, (int32_t)tmsp.tv_nsec, refmode);
1095 	}
1096 #endif
1097 
1098 	/* setup correct event references */
1099 	if (event == PPS_CAPTUREASSERT) {
1100 		tsp = &pps->ppsinfo.assert_timestamp;
1101 		osp = &pps->ppsparam.assert_offset;
1102 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1103 #ifdef PPS_SYNC
1104 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1105 #endif
1106 		pcount = &pps->ppscount[0];
1107 		pseq = &pps->ppsinfo.assert_sequence;
1108 	} else {
1109 		tsp = &pps->ppsinfo.clear_timestamp;
1110 		osp = &pps->ppsparam.clear_offset;
1111 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1112 #ifdef PPS_SYNC
1113 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1114 #endif
1115 		pcount = &pps->ppscount[1];
1116 		pseq = &pps->ppsinfo.clear_sequence;
1117 	}
1118 
1119 	/* determine system time stamp according to refmode */
1120 	dcount = 0;		/* keep GCC happy */
1121 	switch (refmode & PPS_REFEVNT_RMASK) {
1122 	case PPS_REFEVNT_CAPTURE:
1123 		acount = pps->capcount;	/* use capture timestamp */
1124 		break;
1125 
1126 	case PPS_REFEVNT_CURRENT:
1127 		acount = tcount; /* use current timestamp */
1128 		break;
1129 
1130 	case PPS_REFEVNT_CAPCUR:
1131 		/*
1132 		 * calculate counter value between pps_capture() and
1133 		 * pps_ref_event()
1134 		 */
1135 		dcount = tcount - pps->capcount;
1136 		acount = (dcount / 2) + pps->capcount;
1137 		break;
1138 
1139 	default:		/* ignore call error silently */
1140 		return;
1141 	}
1142 
1143 	/*
1144 	 * If the timecounter changed, we cannot compare the count values, so
1145 	 * we have to drop the rest of the PPS-stuff until the next event.
1146 	 */
1147 	if (pps->ppstc != pps->capth->th_counter) {
1148 		pps->ppstc = pps->capth->th_counter;
1149 		pps->capcount = acount;
1150 		*pcount = acount;
1151 		pps->ppscount[2] = acount;
1152 #ifdef PPS_DEBUG
1153 		if (ppsdebug & 0x1) {
1154 			log(LOG_DEBUG,
1155 			    "pps_ref_event(pps=%p, event=%d, ...): DROP (time-counter change)\n",
1156 			    pps, event);
1157 		}
1158 #endif
1159 		return;
1160 	}
1161 
1162 	pps->capcount = acount;
1163 
1164 	/* Convert the count to a bintime. */
1165 	bt = pps->capth->th_offset;
1166 	bintime_addx(&bt, pps->capth->th_scale * (acount - pps->capth->th_offset_count));
1167 	bintime_add(&bt, &timebasebin);
1168 
1169 	if ((refmode & PPS_REFEVNT_PPS) == 0) {
1170 		/* determine difference to reference time stamp */
1171 		bt_ref = *ref_ts;
1172 
1173 		btd = bt;
1174 		bintime_sub(&btd, &bt_ref);
1175 
1176 		/*
1177 		 * simulate a PPS timestamp by dropping the fraction
1178 		 * and applying the offset
1179 		 */
1180 		if (bt.frac >= (uint64_t)1<<63)	/* skip to nearest second */
1181 			bt.sec++;
1182 		bt.frac = 0;
1183 		bintime_add(&bt, &btd);
1184 	} else {
1185 		/*
1186 		 * create ref_ts from current time -
1187 		 * we are supposed to be called on
1188 		 * the second mark
1189 		 */
1190 		bt_ref = bt;
1191 		if (bt_ref.frac >= (uint64_t)1<<63)	/* skip to nearest second */
1192 			bt_ref.sec++;
1193 		bt_ref.frac = 0;
1194 	}
1195 
1196 	/* convert bintime to timestamp */
1197 	bintime2timespec(&bt, &ts);
1198 
1199 	/* If the timecounter was wound up underneath us, bail out. */
1200 	if (pps->capgen != pps->capth->th_generation)
1201 		return;
1202 
1203 	/* store time stamp */
1204 	*pcount = pps->capcount;
1205 	(*pseq)++;
1206 	*tsp = ts;
1207 
1208 	/* add offset correction */
1209 	if (foff) {
1210 		timespecadd(tsp, osp, tsp);
1211 		if (tsp->tv_nsec < 0) {
1212 			tsp->tv_nsec += 1000000000;
1213 			tsp->tv_sec -= 1;
1214 		}
1215 	}
1216 
1217 #ifdef PPS_DEBUG
1218 	if (ppsdebug & 0x2) {
1219 		struct timespec ts2;
1220 		struct timespec ts3;
1221 
1222 		bintime2timespec(&bt_ref, &ts2);
1223 
1224 		bt.sec = 0;
1225 		bt.frac = 0;
1226 
1227 		if (refmode & PPS_REFEVNT_CAPCUR) {
1228 			    bintime_addx(&bt, pps->capth->th_scale * dcount);
1229 		}
1230 		bintime2timespec(&bt, &ts3);
1231 
1232 		log(LOG_DEBUG, "ref_ts=%"PRIi64".%09"PRIi32
1233 		    ", ts=%"PRIi64".%09"PRIi32", read latency=%"PRIi64" ns\n",
1234 		    ts2.tv_sec, (int32_t)ts2.tv_nsec,
1235 		    tsp->tv_sec, (int32_t)tsp->tv_nsec,
1236 		    timespec2ns(&ts3));
1237 	}
1238 #endif
1239 
1240 #ifdef PPS_SYNC
1241 	if (fhard) {
1242 		uint64_t scale;
1243 		uint64_t div;
1244 
1245 		/*
1246 		 * Feed the NTP PLL/FLL.
1247 		 * The FLL wants to know how many (hardware) nanoseconds
1248 		 * elapsed since the previous event (mod 1 second) thus
1249 		 * we are actually looking at the frequency difference scaled
1250 		 * in nsec.
1251 		 * As the counter time stamps are not truly at 1Hz
1252 		 * we need to scale the count by the elapsed
1253 		 * reference time.
1254 		 * valid sampling interval: [0.5..2[ sec
1255 		 */
1256 
1257 		/* calculate elapsed raw count */
1258 		tcount = pps->capcount - pps->ppscount[2];
1259 		pps->ppscount[2] = pps->capcount;
1260 		tcount &= pps->capth->th_counter->tc_counter_mask;
1261 
1262 		/* calculate elapsed ref time */
1263 		btd = bt_ref;
1264 		bintime_sub(&btd, &pps->ref_time);
1265 		pps->ref_time = bt_ref;
1266 
1267 		/* check that we stay below 2 sec */
1268 		if (btd.sec < 0 || btd.sec > 1)
1269 			return;
1270 
1271 		/* we want at least 0.5 sec between samples */
1272 		if (btd.sec == 0 && btd.frac < (uint64_t)1<<63)
1273 			return;
1274 
1275 		/*
1276 		 * calculate cycles per period by multiplying
1277 		 * the frequency with the elapsed period
1278 		 * we pick a fraction of 30 bits
1279 		 * ~1ns resolution for elapsed time
1280 		 */
1281 		div   = (uint64_t)btd.sec << 30;
1282 		div  |= (btd.frac >> 34) & (((uint64_t)1 << 30) - 1);
1283 		div  *= pps->capth->th_counter->tc_frequency;
1284 		div >>= 30;
1285 
1286 		if (div == 0)	/* safeguard */
1287 			return;
1288 
1289 		scale = (uint64_t)1 << 63;
1290 		scale /= div;
1291 		scale *= 2;
1292 
1293 		bt.sec = 0;
1294 		bt.frac = 0;
1295 		bintime_addx(&bt, scale * tcount);
1296 		bintime2timespec(&bt, &ts);
1297 
1298 #ifdef PPS_DEBUG
1299 		if (ppsdebug & 0x4) {
1300 			struct timespec ts2;
1301 			int64_t df;
1302 
1303 			bintime2timespec(&bt_ref, &ts2);
1304 			df = timespec2ns(&ts);
1305 			if (df > 500000000)
1306 				df -= 1000000000;
1307 			log(LOG_DEBUG, "hardpps: ref_ts=%"PRIi64
1308 			    ".%09"PRIi32", ts=%"PRIi64".%09"PRIi32
1309 			    ", freqdiff=%"PRIi64" ns/s\n",
1310 			    ts2.tv_sec, (int32_t)ts2.tv_nsec,
1311 			    tsp->tv_sec, (int32_t)tsp->tv_nsec,
1312 			    df);
1313 		}
1314 #endif
1315 
1316 		hardpps(tsp, timespec2ns(&ts));
1317 	}
1318 #endif
1319 }
1320 
1321 /*
1322  * Timecounters need to be updated every so often to prevent the hardware
1323  * counter from overflowing.  Updating also recalculates the cached values
1324  * used by the get*() family of functions, so their precision depends on
1325  * the update frequency.
1326  */
1327 
1328 static int tc_tick;
1329 
1330 void
1331 tc_ticktock(void)
1332 {
1333 	static int count;
1334 
1335 	if (++count < tc_tick)
1336 		return;
1337 	count = 0;
1338 	mutex_spin_enter(&timecounter_lock);
1339 	if (timecounter_bad != 0) {
1340 		/* An existing timecounter has gone bad, pick a new one. */
1341 		(void)atomic_swap_uint(&timecounter_bad, 0);
1342 		if (timecounter->tc_quality < 0) {
1343 			tc_pick();
1344 		}
1345 	}
1346 	tc_windup();
1347 	mutex_spin_exit(&timecounter_lock);
1348 }
1349 
1350 void
1351 inittimecounter(void)
1352 {
1353 	u_int p;
1354 
1355 	mutex_init(&timecounter_lock, MUTEX_DEFAULT, IPL_HIGH);
1356 
1357 	/*
1358 	 * Set the initial timeout to
1359 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
1360 	 * People should probably not use the sysctl to set the timeout
1361 	 * to smaller than its initial value, since that value is the
1362 	 * smallest reasonable one.  If they want better timestamps they
1363 	 * should use the non-"get"* functions.
1364 	 */
1365 	if (hz > 1000)
1366 		tc_tick = (hz + 500) / 1000;
1367 	else
1368 		tc_tick = 1;
1369 	p = (tc_tick * 1000000) / hz;
1370 	aprint_verbose("timecounter: Timecounters tick every %d.%03u msec\n",
1371 	    p / 1000, p % 1000);
1372 
1373 	/* warm up new timecounter (again) and get rolling. */
1374 	(void)timecounter->tc_get_timecount(timecounter);
1375 	(void)timecounter->tc_get_timecount(timecounter);
1376 }
1377