xref: /openbsd-src/sys/kern/kern_tc.c (revision 83dc78394f43f2ee0451af483a410c8cdb2088f0)
1 /*	$OpenBSD: kern_tc.c,v 1.76 2022/07/23 22:58:51 cheloha Exp $ */
2 
3 /*
4  * Copyright (c) 2000 Poul-Henning Kamp <phk@FreeBSD.org>
5  *
6  * Permission to use, copy, modify, and distribute this software for any
7  * purpose with or without fee is hereby granted, provided that the above
8  * copyright notice and this permission notice appear in all copies.
9  *
10  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
11  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
12  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
13  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
14  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
15  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
16  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
17  */
18 
19 /*
20  * If we meet some day, and you think this stuff is worth it, you
21  * can buy me a beer in return. Poul-Henning Kamp
22  */
23 
24 #include <sys/param.h>
25 #include <sys/atomic.h>
26 #include <sys/kernel.h>
27 #include <sys/mutex.h>
28 #include <sys/rwlock.h>
29 #include <sys/stdint.h>
30 #include <sys/timeout.h>
31 #include <sys/sysctl.h>
32 #include <sys/syslog.h>
33 #include <sys/systm.h>
34 #include <sys/timetc.h>
35 #include <sys/queue.h>
36 #include <sys/malloc.h>
37 
38 u_int dummy_get_timecount(struct timecounter *);
39 
40 int sysctl_tc_hardware(void *, size_t *, void *, size_t);
41 int sysctl_tc_choice(void *, size_t *, void *, size_t);
42 
43 /*
44  * Implement a dummy timecounter which we can use until we get a real one
45  * in the air.  This allows the console and other early stuff to use
46  * time services.
47  */
48 
49 u_int
50 dummy_get_timecount(struct timecounter *tc)
51 {
52 	static u_int now;
53 
54 	return atomic_inc_int_nv(&now);
55 }
56 
57 static struct timecounter dummy_timecounter = {
58 	.tc_get_timecount = dummy_get_timecount,
59 	.tc_poll_pps = NULL,
60 	.tc_counter_mask = ~0u,
61 	.tc_frequency = 1000000,
62 	.tc_name = "dummy",
63 	.tc_quality = -1000000,
64 	.tc_priv = NULL,
65 	.tc_user = 0,
66 };
67 
68 /*
69  * Locks used to protect struct members, global variables in this file:
70  *	I	immutable after initialization
71  *	T	tc_lock
72  *	W	windup_mtx
73  */
74 
75 struct timehands {
76 	/* These fields must be initialized by the driver. */
77 	struct timecounter	*th_counter;		/* [W] */
78 	int64_t			th_adjtimedelta;	/* [T,W] */
79 	struct bintime		th_next_ntp_update;	/* [T,W] */
80 	int64_t			th_adjustment;		/* [W] */
81 	u_int64_t		th_scale;		/* [W] */
82 	u_int	 		th_offset_count;	/* [W] */
83 	struct bintime		th_boottime;		/* [T,W] */
84 	struct bintime		th_offset;		/* [W] */
85 	struct bintime		th_naptime;		/* [W] */
86 	struct timeval		th_microtime;		/* [W] */
87 	struct timespec		th_nanotime;		/* [W] */
88 	/* Fields not to be copied in tc_windup start with th_generation. */
89 	volatile u_int		th_generation;		/* [W] */
90 	struct timehands	*th_next;		/* [I] */
91 };
92 
93 static struct timehands th0;
94 static struct timehands th1 = {
95 	.th_next = &th0
96 };
97 static struct timehands th0 = {
98 	.th_counter = &dummy_timecounter,
99 	.th_scale = UINT64_MAX / 1000000,
100 	.th_offset = { .sec = 1, .frac = 0 },
101 	.th_generation = 1,
102 	.th_next = &th1
103 };
104 
105 struct rwlock tc_lock = RWLOCK_INITIALIZER("tc_lock");
106 
107 /*
108  * tc_windup() must be called before leaving this mutex.
109  */
110 struct mutex windup_mtx = MUTEX_INITIALIZER(IPL_CLOCK);
111 
112 static struct timehands *volatile timehands = &th0;		/* [W] */
113 struct timecounter *timecounter = &dummy_timecounter;		/* [T] */
114 static SLIST_HEAD(, timecounter) tc_list = SLIST_HEAD_INITIALIZER(tc_list);
115 
116 /*
117  * These are updated from tc_windup().  They are useful when
118  * examining kernel core dumps.
119  */
120 volatile time_t naptime = 0;
121 volatile time_t time_second = 1;
122 volatile time_t time_uptime = 0;
123 
124 static int timestepwarnings;
125 
126 void ntp_update_second(struct timehands *);
127 void tc_windup(struct bintime *, struct bintime *, int64_t *);
128 
129 /*
130  * Return the difference between the timehands' counter value now and what
131  * was when we copied it to the timehands' offset_count.
132  */
133 static __inline u_int
134 tc_delta(struct timehands *th)
135 {
136 	struct timecounter *tc;
137 
138 	tc = th->th_counter;
139 	return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
140 	    tc->tc_counter_mask);
141 }
142 
143 /*
144  * Functions for reading the time.  We have to loop until we are sure that
145  * the timehands that we operated on was not updated under our feet.  See
146  * the comment in <sys/time.h> for a description of these functions.
147  */
148 
149 void
150 binboottime(struct bintime *bt)
151 {
152 	struct timehands *th;
153 	u_int gen;
154 
155 	do {
156 		th = timehands;
157 		gen = th->th_generation;
158 		membar_consumer();
159 		*bt = th->th_boottime;
160 		membar_consumer();
161 	} while (gen == 0 || gen != th->th_generation);
162 }
163 
164 void
165 microboottime(struct timeval *tvp)
166 {
167 	struct bintime bt;
168 
169 	binboottime(&bt);
170 	BINTIME_TO_TIMEVAL(&bt, tvp);
171 }
172 
173 void
174 nanoboottime(struct timespec *tsp)
175 {
176 	struct bintime bt;
177 
178 	binboottime(&bt);
179 	BINTIME_TO_TIMESPEC(&bt, tsp);
180 }
181 
182 void
183 binuptime(struct bintime *bt)
184 {
185 	struct timehands *th;
186 	u_int gen;
187 
188 	do {
189 		th = timehands;
190 		gen = th->th_generation;
191 		membar_consumer();
192 		TIMECOUNT_TO_BINTIME(tc_delta(th), th->th_scale, bt);
193 		bintimeadd(bt, &th->th_offset, bt);
194 		membar_consumer();
195 	} while (gen == 0 || gen != th->th_generation);
196 }
197 
198 void
199 getbinuptime(struct bintime *bt)
200 {
201 	struct timehands *th;
202 	u_int gen;
203 
204 	do {
205 		th = timehands;
206 		gen = th->th_generation;
207 		membar_consumer();
208 		*bt = th->th_offset;
209 		membar_consumer();
210 	} while (gen == 0 || gen != th->th_generation);
211 }
212 
213 void
214 nanouptime(struct timespec *tsp)
215 {
216 	struct bintime bt;
217 
218 	binuptime(&bt);
219 	BINTIME_TO_TIMESPEC(&bt, tsp);
220 }
221 
222 void
223 microuptime(struct timeval *tvp)
224 {
225 	struct bintime bt;
226 
227 	binuptime(&bt);
228 	BINTIME_TO_TIMEVAL(&bt, tvp);
229 }
230 
231 time_t
232 getuptime(void)
233 {
234 #if defined(__LP64__)
235 	return time_uptime;	/* atomic */
236 #else
237 	time_t now;
238 	struct timehands *th;
239 	u_int gen;
240 
241 	do {
242 		th = timehands;
243 		gen = th->th_generation;
244 		membar_consumer();
245 		now = th->th_offset.sec;
246 		membar_consumer();
247 	} while (gen == 0 || gen != th->th_generation);
248 
249 	return now;
250 #endif
251 }
252 
253 uint64_t
254 nsecuptime(void)
255 {
256 	struct bintime bt;
257 
258 	binuptime(&bt);
259 	return BINTIME_TO_NSEC(&bt);
260 }
261 
262 uint64_t
263 getnsecuptime(void)
264 {
265 	struct bintime bt;
266 
267 	getbinuptime(&bt);
268 	return BINTIME_TO_NSEC(&bt);
269 }
270 
271 void
272 binruntime(struct bintime *bt)
273 {
274 	struct timehands *th;
275 	u_int gen;
276 
277 	do {
278 		th = timehands;
279 		gen = th->th_generation;
280 		membar_consumer();
281 		TIMECOUNT_TO_BINTIME(tc_delta(th), th->th_scale, bt);
282 		bintimeadd(bt, &th->th_offset, bt);
283 		bintimesub(bt, &th->th_naptime, bt);
284 		membar_consumer();
285 	} while (gen == 0 || gen != th->th_generation);
286 }
287 
288 void
289 nanoruntime(struct timespec *ts)
290 {
291 	struct bintime bt;
292 
293 	binruntime(&bt);
294 	BINTIME_TO_TIMESPEC(&bt, ts);
295 }
296 
297 void
298 bintime(struct bintime *bt)
299 {
300 	struct timehands *th;
301 	u_int gen;
302 
303 	do {
304 		th = timehands;
305 		gen = th->th_generation;
306 		membar_consumer();
307 		TIMECOUNT_TO_BINTIME(tc_delta(th), th->th_scale, bt);
308 		bintimeadd(bt, &th->th_offset, bt);
309 		bintimeadd(bt, &th->th_boottime, bt);
310 		membar_consumer();
311 	} while (gen == 0 || gen != th->th_generation);
312 }
313 
314 void
315 nanotime(struct timespec *tsp)
316 {
317 	struct bintime bt;
318 
319 	bintime(&bt);
320 	BINTIME_TO_TIMESPEC(&bt, tsp);
321 }
322 
323 void
324 microtime(struct timeval *tvp)
325 {
326 	struct bintime bt;
327 
328 	bintime(&bt);
329 	BINTIME_TO_TIMEVAL(&bt, tvp);
330 }
331 
332 time_t
333 gettime(void)
334 {
335 #if defined(__LP64__)
336 	return time_second;	/* atomic */
337 #else
338 	time_t now;
339 	struct timehands *th;
340 	u_int gen;
341 
342 	do {
343 		th = timehands;
344 		gen = th->th_generation;
345 		membar_consumer();
346 		now = th->th_microtime.tv_sec;
347 		membar_consumer();
348 	} while (gen == 0 || gen != th->th_generation);
349 
350 	return now;
351 #endif
352 }
353 
354 void
355 getnanouptime(struct timespec *tsp)
356 {
357 	struct timehands *th;
358 	u_int gen;
359 
360 	do {
361 		th = timehands;
362 		gen = th->th_generation;
363 		membar_consumer();
364 		BINTIME_TO_TIMESPEC(&th->th_offset, tsp);
365 		membar_consumer();
366 	} while (gen == 0 || gen != th->th_generation);
367 }
368 
369 void
370 getmicrouptime(struct timeval *tvp)
371 {
372 	struct timehands *th;
373 	u_int gen;
374 
375 	do {
376 		th = timehands;
377 		gen = th->th_generation;
378 		membar_consumer();
379 		BINTIME_TO_TIMEVAL(&th->th_offset, tvp);
380 		membar_consumer();
381 	} while (gen == 0 || gen != th->th_generation);
382 }
383 
384 void
385 getnanotime(struct timespec *tsp)
386 {
387 	struct timehands *th;
388 	u_int gen;
389 
390 	do {
391 		th = timehands;
392 		gen = th->th_generation;
393 		membar_consumer();
394 		*tsp = th->th_nanotime;
395 		membar_consumer();
396 	} while (gen == 0 || gen != th->th_generation);
397 }
398 
399 void
400 getmicrotime(struct timeval *tvp)
401 {
402 	struct timehands *th;
403 	u_int gen;
404 
405 	do {
406 		th = timehands;
407 		gen = th->th_generation;
408 		membar_consumer();
409 		*tvp = th->th_microtime;
410 		membar_consumer();
411 	} while (gen == 0 || gen != th->th_generation);
412 }
413 
414 /*
415  * Initialize a new timecounter and possibly use it.
416  */
417 void
418 tc_init(struct timecounter *tc)
419 {
420 	u_int64_t tmp;
421 	u_int u;
422 
423 	u = tc->tc_frequency / tc->tc_counter_mask;
424 	/* XXX: We need some margin here, 10% is a guess */
425 	u *= 11;
426 	u /= 10;
427 	if (tc->tc_quality >= 0) {
428 		if (u > hz) {
429 			tc->tc_quality = -2000;
430 			printf("Timecounter \"%s\" frequency %lu Hz",
431 			    tc->tc_name, (unsigned long)tc->tc_frequency);
432 			printf(" -- Insufficient hz, needs at least %u\n", u);
433 		}
434 	}
435 
436 	/* Determine the counter's precision. */
437 	for (tmp = 1; (tmp & tc->tc_counter_mask) == 0; tmp <<= 1)
438 		continue;
439 	tc->tc_precision = tmp;
440 
441 	SLIST_INSERT_HEAD(&tc_list, tc, tc_next);
442 
443 	/*
444 	 * Never automatically use a timecounter with negative quality.
445 	 * Even though we run on the dummy counter, switching here may be
446 	 * worse since this timecounter may not be monotonic.
447 	 */
448 	if (tc->tc_quality < 0)
449 		return;
450 	if (tc->tc_quality < timecounter->tc_quality)
451 		return;
452 	if (tc->tc_quality == timecounter->tc_quality &&
453 	    tc->tc_frequency < timecounter->tc_frequency)
454 		return;
455 	(void)tc->tc_get_timecount(tc);
456 	enqueue_randomness(tc->tc_get_timecount(tc));
457 
458 	timecounter = tc;
459 }
460 
461 /* Report the frequency of the current timecounter. */
462 u_int64_t
463 tc_getfrequency(void)
464 {
465 	return (timehands->th_counter->tc_frequency);
466 }
467 
468 /* Report the precision of the current timecounter. */
469 u_int64_t
470 tc_getprecision(void)
471 {
472 	return (timehands->th_counter->tc_precision);
473 }
474 
475 /*
476  * Step our concept of UTC, aka the realtime clock.
477  * This is done by modifying our estimate of when we booted.
478  *
479  * Any ongoing adjustment is meaningless after a clock jump,
480  * so we zero adjtimedelta here as well.
481  */
482 void
483 tc_setrealtimeclock(const struct timespec *ts)
484 {
485 	struct bintime boottime, old_utc, uptime, utc;
486 	struct timespec tmp;
487 	int64_t zero = 0;
488 
489 	TIMESPEC_TO_BINTIME(ts, &utc);
490 
491 	rw_enter_write(&tc_lock);
492 	mtx_enter(&windup_mtx);
493 
494 	binuptime(&uptime);
495 	bintimesub(&utc, &uptime, &boottime);
496 	bintimeadd(&timehands->th_boottime, &uptime, &old_utc);
497 	/* XXX fiddle all the little crinkly bits around the fiords... */
498 	tc_windup(&boottime, NULL, &zero);
499 
500 	mtx_leave(&windup_mtx);
501 	rw_exit_write(&tc_lock);
502 
503 	enqueue_randomness(ts->tv_sec);
504 
505 	if (timestepwarnings) {
506 		BINTIME_TO_TIMESPEC(&old_utc, &tmp);
507 		log(LOG_INFO, "Time stepped from %lld.%09ld to %lld.%09ld\n",
508 		    (long long)tmp.tv_sec, tmp.tv_nsec,
509 		    (long long)ts->tv_sec, ts->tv_nsec);
510 	}
511 }
512 
513 /*
514  * Step the monotonic and realtime clocks, triggering any timeouts that
515  * should have occurred across the interval.
516  */
517 void
518 tc_setclock(const struct timespec *ts)
519 {
520 	struct bintime new_naptime, old_naptime, uptime, utc;
521 	struct timespec tmp;
522 	static int first = 1;
523 #ifndef SMALL_KERNEL
524 	struct bintime elapsed;
525 	long long adj_ticks;
526 #endif
527 
528 	/*
529 	 * When we're called for the first time, during boot when
530 	 * the root partition is mounted, we need to set boottime.
531 	 */
532 	if (first) {
533 		tc_setrealtimeclock(ts);
534 		first = 0;
535 		return;
536 	}
537 
538 	enqueue_randomness(ts->tv_sec);
539 
540 	TIMESPEC_TO_BINTIME(ts, &utc);
541 
542 	mtx_enter(&windup_mtx);
543 
544 	bintimesub(&utc, &timehands->th_boottime, &uptime);
545 	old_naptime = timehands->th_naptime;
546 	/* XXX fiddle all the little crinkly bits around the fiords... */
547 	tc_windup(NULL, &uptime, NULL);
548 	new_naptime = timehands->th_naptime;
549 
550 	mtx_leave(&windup_mtx);
551 
552 	if (bintimecmp(&old_naptime, &new_naptime, ==)) {
553 		BINTIME_TO_TIMESPEC(&uptime, &tmp);
554 		printf("%s: cannot rewind uptime to %lld.%09ld\n",
555 		    __func__, (long long)tmp.tv_sec, tmp.tv_nsec);
556 	}
557 
558 #ifndef SMALL_KERNEL
559 	/* convert the bintime to ticks */
560 	bintimesub(&new_naptime, &old_naptime, &elapsed);
561 	adj_ticks = BINTIME_TO_NSEC(&elapsed) / tick_nsec;
562 	if (adj_ticks > 0) {
563 		if (adj_ticks > INT_MAX)
564 			adj_ticks = INT_MAX;
565 		timeout_adjust_ticks(adj_ticks);
566 	}
567 #endif
568 }
569 
570 void
571 tc_update_timekeep(void)
572 {
573 	static struct timecounter *last_tc = NULL;
574 	struct timehands *th;
575 
576 	MUTEX_ASSERT_LOCKED(&windup_mtx);
577 
578 	if (timekeep == NULL)
579 		return;
580 
581 	th = timehands;
582 	timekeep->tk_generation = 0;
583 	membar_producer();
584 	timekeep->tk_scale = th->th_scale;
585 	timekeep->tk_offset_count = th->th_offset_count;
586 	timekeep->tk_offset = th->th_offset;
587 	timekeep->tk_naptime = th->th_naptime;
588 	timekeep->tk_boottime = th->th_boottime;
589 	if (last_tc != th->th_counter) {
590 		timekeep->tk_counter_mask = th->th_counter->tc_counter_mask;
591 		timekeep->tk_user = th->th_counter->tc_user;
592 		last_tc = th->th_counter;
593 	}
594 	membar_producer();
595 	timekeep->tk_generation = th->th_generation;
596 
597 	return;
598 }
599 
600 /*
601  * Initialize the next struct timehands in the ring and make
602  * it the active timehands.  Along the way we might switch to a different
603  * timecounter and/or do seconds processing in NTP.  Slightly magic.
604  */
605 void
606 tc_windup(struct bintime *new_boottime, struct bintime *new_offset,
607     int64_t *new_adjtimedelta)
608 {
609 	struct bintime bt;
610 	struct timecounter *active_tc;
611 	struct timehands *th, *tho;
612 	u_int64_t scale;
613 	u_int delta, ncount, ogen;
614 
615 	if (new_boottime != NULL || new_adjtimedelta != NULL)
616 		rw_assert_wrlock(&tc_lock);
617 	MUTEX_ASSERT_LOCKED(&windup_mtx);
618 
619 	active_tc = timecounter;
620 
621 	/*
622 	 * Make the next timehands a copy of the current one, but do not
623 	 * overwrite the generation or next pointer.  While we update
624 	 * the contents, the generation must be zero.
625 	 */
626 	tho = timehands;
627 	ogen = tho->th_generation;
628 	th = tho->th_next;
629 	th->th_generation = 0;
630 	membar_producer();
631 	memcpy(th, tho, offsetof(struct timehands, th_generation));
632 
633 	/*
634 	 * Capture a timecounter delta on the current timecounter and if
635 	 * changing timecounters, a counter value from the new timecounter.
636 	 * Update the offset fields accordingly.
637 	 */
638 	delta = tc_delta(th);
639 	if (th->th_counter != active_tc)
640 		ncount = active_tc->tc_get_timecount(active_tc);
641 	else
642 		ncount = 0;
643 	th->th_offset_count += delta;
644 	th->th_offset_count &= th->th_counter->tc_counter_mask;
645 	TIMECOUNT_TO_BINTIME(delta, th->th_scale, &bt);
646 	bintimeadd(&th->th_offset, &bt, &th->th_offset);
647 
648 	/*
649 	 * Ignore new offsets that predate the current offset.
650 	 * If changing the offset, first increase the naptime
651 	 * accordingly.
652 	 */
653 	if (new_offset != NULL && bintimecmp(&th->th_offset, new_offset, <)) {
654 		bintimesub(new_offset, &th->th_offset, &bt);
655 		bintimeadd(&th->th_naptime, &bt, &th->th_naptime);
656 		naptime = th->th_naptime.sec;
657 		th->th_offset = *new_offset;
658 	}
659 
660 #ifdef notyet
661 	/*
662 	 * Hardware latching timecounters may not generate interrupts on
663 	 * PPS events, so instead we poll them.  There is a finite risk that
664 	 * the hardware might capture a count which is later than the one we
665 	 * got above, and therefore possibly in the next NTP second which might
666 	 * have a different rate than the current NTP second.  It doesn't
667 	 * matter in practice.
668 	 */
669 	if (tho->th_counter->tc_poll_pps)
670 		tho->th_counter->tc_poll_pps(tho->th_counter);
671 #endif
672 
673 	/*
674 	 * If changing the boot time or clock adjustment, do so before
675 	 * NTP processing.
676 	 */
677 	if (new_boottime != NULL)
678 		th->th_boottime = *new_boottime;
679 	if (new_adjtimedelta != NULL) {
680 		th->th_adjtimedelta = *new_adjtimedelta;
681 		/* Reset the NTP update period. */
682 		bintimesub(&th->th_offset, &th->th_naptime,
683 		    &th->th_next_ntp_update);
684 	}
685 
686 	/*
687 	 * Deal with NTP second processing.  The while-loop normally
688 	 * iterates at most once, but in extreme situations it might
689 	 * keep NTP sane if tc_windup() is not run for several seconds.
690 	 */
691 	bintimesub(&th->th_offset, &th->th_naptime, &bt);
692 	while (bintimecmp(&th->th_next_ntp_update, &bt, <=)) {
693 		ntp_update_second(th);
694 		th->th_next_ntp_update.sec++;
695 	}
696 
697 	/* Update the UTC timestamps used by the get*() functions. */
698 	bintimeadd(&th->th_boottime, &th->th_offset, &bt);
699 	BINTIME_TO_TIMEVAL(&bt, &th->th_microtime);
700 	BINTIME_TO_TIMESPEC(&bt, &th->th_nanotime);
701 
702 	/* Now is a good time to change timecounters. */
703 	if (th->th_counter != active_tc) {
704 		th->th_counter = active_tc;
705 		th->th_offset_count = ncount;
706 	}
707 
708 	/*-
709 	 * Recalculate the scaling factor.  We want the number of 1/2^64
710 	 * fractions of a second per period of the hardware counter, taking
711 	 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
712 	 * processing provides us with.
713 	 *
714 	 * The th_adjustment is nanoseconds per second with 32 bit binary
715 	 * fraction and we want 64 bit binary fraction of second:
716 	 *
717 	 *	 x = a * 2^32 / 10^9 = a * 4.294967296
718 	 *
719 	 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
720 	 * we can only multiply by about 850 without overflowing, but that
721 	 * leaves suitably precise fractions for multiply before divide.
722 	 *
723 	 * Divide before multiply with a fraction of 2199/512 results in a
724 	 * systematic undercompensation of 10PPM of th_adjustment.  On a
725 	 * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
726  	 *
727 	 * We happily sacrifice the lowest of the 64 bits of our result
728 	 * to the goddess of code clarity.
729 	 *
730 	 */
731 	scale = (u_int64_t)1 << 63;
732 	scale += \
733 	    ((th->th_adjustment + th->th_counter->tc_freq_adj) / 1024) * 2199;
734 	scale /= th->th_counter->tc_frequency;
735 	th->th_scale = scale * 2;
736 
737 	/*
738 	 * Now that the struct timehands is again consistent, set the new
739 	 * generation number, making sure to not make it zero.
740 	 */
741 	if (++ogen == 0)
742 		ogen = 1;
743 	membar_producer();
744 	th->th_generation = ogen;
745 
746 	/* Go live with the new struct timehands. */
747 	time_second = th->th_microtime.tv_sec;
748 	time_uptime = th->th_offset.sec;
749 	membar_producer();
750 	timehands = th;
751 
752 	tc_update_timekeep();
753 }
754 
755 /* Report or change the active timecounter hardware. */
756 int
757 sysctl_tc_hardware(void *oldp, size_t *oldlenp, void *newp, size_t newlen)
758 {
759 	char newname[32];
760 	struct timecounter *newtc, *tc;
761 	int error;
762 
763 	tc = timecounter;
764 	strlcpy(newname, tc->tc_name, sizeof(newname));
765 
766 	error = sysctl_string(oldp, oldlenp, newp, newlen, newname, sizeof(newname));
767 	if (error != 0 || strcmp(newname, tc->tc_name) == 0)
768 		return (error);
769 	SLIST_FOREACH(newtc, &tc_list, tc_next) {
770 		if (strcmp(newname, newtc->tc_name) != 0)
771 			continue;
772 
773 		/* Warm up new timecounter. */
774 		(void)newtc->tc_get_timecount(newtc);
775 		(void)newtc->tc_get_timecount(newtc);
776 
777 		rw_enter_write(&tc_lock);
778 		timecounter = newtc;
779 		rw_exit_write(&tc_lock);
780 
781 		return (0);
782 	}
783 	return (EINVAL);
784 }
785 
786 /* Report or change the active timecounter hardware. */
787 int
788 sysctl_tc_choice(void *oldp, size_t *oldlenp, void *newp, size_t newlen)
789 {
790 	char buf[32], *spc, *choices;
791 	struct timecounter *tc;
792 	int error, maxlen;
793 
794 	if (SLIST_EMPTY(&tc_list))
795 		return (sysctl_rdstring(oldp, oldlenp, newp, ""));
796 
797 	spc = "";
798 	maxlen = 0;
799 	SLIST_FOREACH(tc, &tc_list, tc_next)
800 		maxlen += sizeof(buf);
801 	choices = malloc(maxlen, M_TEMP, M_WAITOK);
802 	*choices = '\0';
803 	SLIST_FOREACH(tc, &tc_list, tc_next) {
804 		snprintf(buf, sizeof(buf), "%s%s(%d)",
805 		    spc, tc->tc_name, tc->tc_quality);
806 		spc = " ";
807 		strlcat(choices, buf, maxlen);
808 	}
809 	error = sysctl_rdstring(oldp, oldlenp, newp, choices);
810 	free(choices, M_TEMP, maxlen);
811 	return (error);
812 }
813 
814 /*
815  * Timecounters need to be updated every so often to prevent the hardware
816  * counter from overflowing.  Updating also recalculates the cached values
817  * used by the get*() family of functions, so their precision depends on
818  * the update frequency.
819  */
820 static int tc_tick;
821 
822 void
823 tc_ticktock(void)
824 {
825 	static int count;
826 
827 	if (++count < tc_tick)
828 		return;
829 	if (!mtx_enter_try(&windup_mtx))
830 		return;
831 	count = 0;
832 	tc_windup(NULL, NULL, NULL);
833 	mtx_leave(&windup_mtx);
834 }
835 
836 void
837 inittimecounter(void)
838 {
839 #ifdef DEBUG
840 	u_int p;
841 #endif
842 
843 	/*
844 	 * Set the initial timeout to
845 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
846 	 * People should probably not use the sysctl to set the timeout
847 	 * to smaller than its initial value, since that value is the
848 	 * smallest reasonable one.  If they want better timestamps they
849 	 * should use the non-"get"* functions.
850 	 */
851 	if (hz > 1000)
852 		tc_tick = (hz + 500) / 1000;
853 	else
854 		tc_tick = 1;
855 #ifdef DEBUG
856 	p = (tc_tick * 1000000) / hz;
857 	printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
858 #endif
859 
860 	/* warm up new timecounter (again) and get rolling. */
861 	(void)timecounter->tc_get_timecount(timecounter);
862 	(void)timecounter->tc_get_timecount(timecounter);
863 }
864 
865 const struct sysctl_bounded_args tc_vars[] = {
866 	{ KERN_TIMECOUNTER_TICK, &tc_tick, SYSCTL_INT_READONLY },
867 	{ KERN_TIMECOUNTER_TIMESTEPWARNINGS, &timestepwarnings, 0, 1 },
868 };
869 
870 /*
871  * Return timecounter-related information.
872  */
873 int
874 sysctl_tc(int *name, u_int namelen, void *oldp, size_t *oldlenp,
875     void *newp, size_t newlen)
876 {
877 	if (namelen != 1)
878 		return (ENOTDIR);
879 
880 	switch (name[0]) {
881 	case KERN_TIMECOUNTER_HARDWARE:
882 		return (sysctl_tc_hardware(oldp, oldlenp, newp, newlen));
883 	case KERN_TIMECOUNTER_CHOICE:
884 		return (sysctl_tc_choice(oldp, oldlenp, newp, newlen));
885 	default:
886 		return (sysctl_bounded_arr(tc_vars, nitems(tc_vars), name,
887 		    namelen, oldp, oldlenp, newp, newlen));
888 	}
889 	/* NOTREACHED */
890 }
891 
892 /*
893  * Skew the timehands according to any adjtime(2) adjustment.
894  */
895 void
896 ntp_update_second(struct timehands *th)
897 {
898 	int64_t adj;
899 
900 	MUTEX_ASSERT_LOCKED(&windup_mtx);
901 
902 	if (th->th_adjtimedelta > 0)
903 		adj = MIN(5000, th->th_adjtimedelta);
904 	else
905 		adj = MAX(-5000, th->th_adjtimedelta);
906 	th->th_adjtimedelta -= adj;
907 	th->th_adjustment = (adj * 1000) << 32;
908 }
909 
910 void
911 tc_adjfreq(int64_t *old, int64_t *new)
912 {
913 	if (old != NULL) {
914 		rw_assert_anylock(&tc_lock);
915 		*old = timecounter->tc_freq_adj;
916 	}
917 	if (new != NULL) {
918 		rw_assert_wrlock(&tc_lock);
919 		mtx_enter(&windup_mtx);
920 		timecounter->tc_freq_adj = *new;
921 		tc_windup(NULL, NULL, NULL);
922 		mtx_leave(&windup_mtx);
923 	}
924 }
925 
926 void
927 tc_adjtime(int64_t *old, int64_t *new)
928 {
929 	struct timehands *th;
930 	u_int gen;
931 
932 	if (old != NULL) {
933 		do {
934 			th = timehands;
935 			gen = th->th_generation;
936 			membar_consumer();
937 			*old = th->th_adjtimedelta;
938 			membar_consumer();
939 		} while (gen == 0 || gen != th->th_generation);
940 	}
941 	if (new != NULL) {
942 		rw_assert_wrlock(&tc_lock);
943 		mtx_enter(&windup_mtx);
944 		tc_windup(NULL, NULL, new);
945 		mtx_leave(&windup_mtx);
946 	}
947 }
948