xref: /freebsd/sys/kern/kern_tc.c (revision 0957b409)
1 /*-
2  * SPDX-License-Identifier: Beerware
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  * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
12  * All rights reserved.
13  *
14  * Portions of this software were developed by Julien Ridoux at the University
15  * of Melbourne under sponsorship from the FreeBSD Foundation.
16  *
17  * Portions of this software were developed by Konstantin Belousov
18  * under sponsorship from the FreeBSD Foundation.
19  */
20 
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
23 
24 #include "opt_ntp.h"
25 #include "opt_ffclock.h"
26 
27 #include <sys/param.h>
28 #include <sys/kernel.h>
29 #include <sys/limits.h>
30 #include <sys/lock.h>
31 #include <sys/mutex.h>
32 #include <sys/proc.h>
33 #include <sys/sbuf.h>
34 #include <sys/sleepqueue.h>
35 #include <sys/sysctl.h>
36 #include <sys/syslog.h>
37 #include <sys/systm.h>
38 #include <sys/timeffc.h>
39 #include <sys/timepps.h>
40 #include <sys/timetc.h>
41 #include <sys/timex.h>
42 #include <sys/vdso.h>
43 
44 /*
45  * A large step happens on boot.  This constant detects such steps.
46  * It is relatively small so that ntp_update_second gets called enough
47  * in the typical 'missed a couple of seconds' case, but doesn't loop
48  * forever when the time step is large.
49  */
50 #define LARGE_STEP	200
51 
52 /*
53  * Implement a dummy timecounter which we can use until we get a real one
54  * in the air.  This allows the console and other early stuff to use
55  * time services.
56  */
57 
58 static u_int
59 dummy_get_timecount(struct timecounter *tc)
60 {
61 	static u_int now;
62 
63 	return (++now);
64 }
65 
66 static struct timecounter dummy_timecounter = {
67 	dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
68 };
69 
70 struct timehands {
71 	/* These fields must be initialized by the driver. */
72 	struct timecounter	*th_counter;
73 	int64_t			th_adjustment;
74 	uint64_t		th_scale;
75 	u_int	 		th_offset_count;
76 	struct bintime		th_offset;
77 	struct bintime		th_bintime;
78 	struct timeval		th_microtime;
79 	struct timespec		th_nanotime;
80 	struct bintime		th_boottime;
81 	/* Fields not to be copied in tc_windup start with th_generation. */
82 	u_int			th_generation;
83 	struct timehands	*th_next;
84 };
85 
86 static struct timehands th0;
87 static struct timehands th1 = {
88 	.th_next = &th0
89 };
90 static struct timehands th0 = {
91 	.th_counter = &dummy_timecounter,
92 	.th_scale = (uint64_t)-1 / 1000000,
93 	.th_offset = { .sec = 1 },
94 	.th_generation = 1,
95 	.th_next = &th1
96 };
97 
98 static struct timehands *volatile timehands = &th0;
99 struct timecounter *timecounter = &dummy_timecounter;
100 static struct timecounter *timecounters = &dummy_timecounter;
101 
102 int tc_min_ticktock_freq = 1;
103 
104 volatile time_t time_second = 1;
105 volatile time_t time_uptime = 1;
106 
107 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
108 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
109     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
110 
111 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
112 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
113 
114 static int timestepwarnings;
115 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
116     &timestepwarnings, 0, "Log time steps");
117 
118 struct bintime bt_timethreshold;
119 struct bintime bt_tickthreshold;
120 sbintime_t sbt_timethreshold;
121 sbintime_t sbt_tickthreshold;
122 struct bintime tc_tick_bt;
123 sbintime_t tc_tick_sbt;
124 int tc_precexp;
125 int tc_timepercentage = TC_DEFAULTPERC;
126 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
127 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
128     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
129     sysctl_kern_timecounter_adjprecision, "I",
130     "Allowed time interval deviation in percents");
131 
132 volatile int rtc_generation = 1;
133 
134 static int tc_chosen;	/* Non-zero if a specific tc was chosen via sysctl. */
135 
136 static void tc_windup(struct bintime *new_boottimebin);
137 static void cpu_tick_calibrate(int);
138 
139 void dtrace_getnanotime(struct timespec *tsp);
140 
141 static int
142 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
143 {
144 	struct timeval boottime;
145 
146 	getboottime(&boottime);
147 
148 /* i386 is the only arch which uses a 32bits time_t */
149 #ifdef __amd64__
150 #ifdef SCTL_MASK32
151 	int tv[2];
152 
153 	if (req->flags & SCTL_MASK32) {
154 		tv[0] = boottime.tv_sec;
155 		tv[1] = boottime.tv_usec;
156 		return (SYSCTL_OUT(req, tv, sizeof(tv)));
157 	}
158 #endif
159 #endif
160 	return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
161 }
162 
163 static int
164 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
165 {
166 	u_int ncount;
167 	struct timecounter *tc = arg1;
168 
169 	ncount = tc->tc_get_timecount(tc);
170 	return (sysctl_handle_int(oidp, &ncount, 0, req));
171 }
172 
173 static int
174 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
175 {
176 	uint64_t freq;
177 	struct timecounter *tc = arg1;
178 
179 	freq = tc->tc_frequency;
180 	return (sysctl_handle_64(oidp, &freq, 0, req));
181 }
182 
183 /*
184  * Return the difference between the timehands' counter value now and what
185  * was when we copied it to the timehands' offset_count.
186  */
187 static __inline u_int
188 tc_delta(struct timehands *th)
189 {
190 	struct timecounter *tc;
191 
192 	tc = th->th_counter;
193 	return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
194 	    tc->tc_counter_mask);
195 }
196 
197 /*
198  * Functions for reading the time.  We have to loop until we are sure that
199  * the timehands that we operated on was not updated under our feet.  See
200  * the comment in <sys/time.h> for a description of these 12 functions.
201  */
202 
203 #ifdef FFCLOCK
204 void
205 fbclock_binuptime(struct bintime *bt)
206 {
207 	struct timehands *th;
208 	unsigned int gen;
209 
210 	do {
211 		th = timehands;
212 		gen = atomic_load_acq_int(&th->th_generation);
213 		*bt = th->th_offset;
214 		bintime_addx(bt, th->th_scale * tc_delta(th));
215 		atomic_thread_fence_acq();
216 	} while (gen == 0 || gen != th->th_generation);
217 }
218 
219 void
220 fbclock_nanouptime(struct timespec *tsp)
221 {
222 	struct bintime bt;
223 
224 	fbclock_binuptime(&bt);
225 	bintime2timespec(&bt, tsp);
226 }
227 
228 void
229 fbclock_microuptime(struct timeval *tvp)
230 {
231 	struct bintime bt;
232 
233 	fbclock_binuptime(&bt);
234 	bintime2timeval(&bt, tvp);
235 }
236 
237 void
238 fbclock_bintime(struct bintime *bt)
239 {
240 	struct timehands *th;
241 	unsigned int gen;
242 
243 	do {
244 		th = timehands;
245 		gen = atomic_load_acq_int(&th->th_generation);
246 		*bt = th->th_bintime;
247 		bintime_addx(bt, th->th_scale * tc_delta(th));
248 		atomic_thread_fence_acq();
249 	} while (gen == 0 || gen != th->th_generation);
250 }
251 
252 void
253 fbclock_nanotime(struct timespec *tsp)
254 {
255 	struct bintime bt;
256 
257 	fbclock_bintime(&bt);
258 	bintime2timespec(&bt, tsp);
259 }
260 
261 void
262 fbclock_microtime(struct timeval *tvp)
263 {
264 	struct bintime bt;
265 
266 	fbclock_bintime(&bt);
267 	bintime2timeval(&bt, tvp);
268 }
269 
270 void
271 fbclock_getbinuptime(struct bintime *bt)
272 {
273 	struct timehands *th;
274 	unsigned int gen;
275 
276 	do {
277 		th = timehands;
278 		gen = atomic_load_acq_int(&th->th_generation);
279 		*bt = th->th_offset;
280 		atomic_thread_fence_acq();
281 	} while (gen == 0 || gen != th->th_generation);
282 }
283 
284 void
285 fbclock_getnanouptime(struct timespec *tsp)
286 {
287 	struct timehands *th;
288 	unsigned int gen;
289 
290 	do {
291 		th = timehands;
292 		gen = atomic_load_acq_int(&th->th_generation);
293 		bintime2timespec(&th->th_offset, tsp);
294 		atomic_thread_fence_acq();
295 	} while (gen == 0 || gen != th->th_generation);
296 }
297 
298 void
299 fbclock_getmicrouptime(struct timeval *tvp)
300 {
301 	struct timehands *th;
302 	unsigned int gen;
303 
304 	do {
305 		th = timehands;
306 		gen = atomic_load_acq_int(&th->th_generation);
307 		bintime2timeval(&th->th_offset, tvp);
308 		atomic_thread_fence_acq();
309 	} while (gen == 0 || gen != th->th_generation);
310 }
311 
312 void
313 fbclock_getbintime(struct bintime *bt)
314 {
315 	struct timehands *th;
316 	unsigned int gen;
317 
318 	do {
319 		th = timehands;
320 		gen = atomic_load_acq_int(&th->th_generation);
321 		*bt = th->th_bintime;
322 		atomic_thread_fence_acq();
323 	} while (gen == 0 || gen != th->th_generation);
324 }
325 
326 void
327 fbclock_getnanotime(struct timespec *tsp)
328 {
329 	struct timehands *th;
330 	unsigned int gen;
331 
332 	do {
333 		th = timehands;
334 		gen = atomic_load_acq_int(&th->th_generation);
335 		*tsp = th->th_nanotime;
336 		atomic_thread_fence_acq();
337 	} while (gen == 0 || gen != th->th_generation);
338 }
339 
340 void
341 fbclock_getmicrotime(struct timeval *tvp)
342 {
343 	struct timehands *th;
344 	unsigned int gen;
345 
346 	do {
347 		th = timehands;
348 		gen = atomic_load_acq_int(&th->th_generation);
349 		*tvp = th->th_microtime;
350 		atomic_thread_fence_acq();
351 	} while (gen == 0 || gen != th->th_generation);
352 }
353 #else /* !FFCLOCK */
354 void
355 binuptime(struct bintime *bt)
356 {
357 	struct timehands *th;
358 	u_int gen;
359 
360 	do {
361 		th = timehands;
362 		gen = atomic_load_acq_int(&th->th_generation);
363 		*bt = th->th_offset;
364 		bintime_addx(bt, th->th_scale * tc_delta(th));
365 		atomic_thread_fence_acq();
366 	} while (gen == 0 || gen != th->th_generation);
367 }
368 
369 void
370 nanouptime(struct timespec *tsp)
371 {
372 	struct bintime bt;
373 
374 	binuptime(&bt);
375 	bintime2timespec(&bt, tsp);
376 }
377 
378 void
379 microuptime(struct timeval *tvp)
380 {
381 	struct bintime bt;
382 
383 	binuptime(&bt);
384 	bintime2timeval(&bt, tvp);
385 }
386 
387 void
388 bintime(struct bintime *bt)
389 {
390 	struct timehands *th;
391 	u_int gen;
392 
393 	do {
394 		th = timehands;
395 		gen = atomic_load_acq_int(&th->th_generation);
396 		*bt = th->th_bintime;
397 		bintime_addx(bt, th->th_scale * tc_delta(th));
398 		atomic_thread_fence_acq();
399 	} while (gen == 0 || gen != th->th_generation);
400 }
401 
402 void
403 nanotime(struct timespec *tsp)
404 {
405 	struct bintime bt;
406 
407 	bintime(&bt);
408 	bintime2timespec(&bt, tsp);
409 }
410 
411 void
412 microtime(struct timeval *tvp)
413 {
414 	struct bintime bt;
415 
416 	bintime(&bt);
417 	bintime2timeval(&bt, tvp);
418 }
419 
420 void
421 getbinuptime(struct bintime *bt)
422 {
423 	struct timehands *th;
424 	u_int gen;
425 
426 	do {
427 		th = timehands;
428 		gen = atomic_load_acq_int(&th->th_generation);
429 		*bt = th->th_offset;
430 		atomic_thread_fence_acq();
431 	} while (gen == 0 || gen != th->th_generation);
432 }
433 
434 void
435 getnanouptime(struct timespec *tsp)
436 {
437 	struct timehands *th;
438 	u_int gen;
439 
440 	do {
441 		th = timehands;
442 		gen = atomic_load_acq_int(&th->th_generation);
443 		bintime2timespec(&th->th_offset, tsp);
444 		atomic_thread_fence_acq();
445 	} while (gen == 0 || gen != th->th_generation);
446 }
447 
448 void
449 getmicrouptime(struct timeval *tvp)
450 {
451 	struct timehands *th;
452 	u_int gen;
453 
454 	do {
455 		th = timehands;
456 		gen = atomic_load_acq_int(&th->th_generation);
457 		bintime2timeval(&th->th_offset, tvp);
458 		atomic_thread_fence_acq();
459 	} while (gen == 0 || gen != th->th_generation);
460 }
461 
462 void
463 getbintime(struct bintime *bt)
464 {
465 	struct timehands *th;
466 	u_int gen;
467 
468 	do {
469 		th = timehands;
470 		gen = atomic_load_acq_int(&th->th_generation);
471 		*bt = th->th_bintime;
472 		atomic_thread_fence_acq();
473 	} while (gen == 0 || gen != th->th_generation);
474 }
475 
476 void
477 getnanotime(struct timespec *tsp)
478 {
479 	struct timehands *th;
480 	u_int gen;
481 
482 	do {
483 		th = timehands;
484 		gen = atomic_load_acq_int(&th->th_generation);
485 		*tsp = th->th_nanotime;
486 		atomic_thread_fence_acq();
487 	} while (gen == 0 || gen != th->th_generation);
488 }
489 
490 void
491 getmicrotime(struct timeval *tvp)
492 {
493 	struct timehands *th;
494 	u_int gen;
495 
496 	do {
497 		th = timehands;
498 		gen = atomic_load_acq_int(&th->th_generation);
499 		*tvp = th->th_microtime;
500 		atomic_thread_fence_acq();
501 	} while (gen == 0 || gen != th->th_generation);
502 }
503 #endif /* FFCLOCK */
504 
505 void
506 getboottime(struct timeval *boottime)
507 {
508 	struct bintime boottimebin;
509 
510 	getboottimebin(&boottimebin);
511 	bintime2timeval(&boottimebin, boottime);
512 }
513 
514 void
515 getboottimebin(struct bintime *boottimebin)
516 {
517 	struct timehands *th;
518 	u_int gen;
519 
520 	do {
521 		th = timehands;
522 		gen = atomic_load_acq_int(&th->th_generation);
523 		*boottimebin = th->th_boottime;
524 		atomic_thread_fence_acq();
525 	} while (gen == 0 || gen != th->th_generation);
526 }
527 
528 #ifdef FFCLOCK
529 /*
530  * Support for feed-forward synchronization algorithms. This is heavily inspired
531  * by the timehands mechanism but kept independent from it. *_windup() functions
532  * have some connection to avoid accessing the timecounter hardware more than
533  * necessary.
534  */
535 
536 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
537 struct ffclock_estimate ffclock_estimate;
538 struct bintime ffclock_boottime;	/* Feed-forward boot time estimate. */
539 uint32_t ffclock_status;		/* Feed-forward clock status. */
540 int8_t ffclock_updated;			/* New estimates are available. */
541 struct mtx ffclock_mtx;			/* Mutex on ffclock_estimate. */
542 
543 struct fftimehands {
544 	struct ffclock_estimate	cest;
545 	struct bintime		tick_time;
546 	struct bintime		tick_time_lerp;
547 	ffcounter		tick_ffcount;
548 	uint64_t		period_lerp;
549 	volatile uint8_t	gen;
550 	struct fftimehands	*next;
551 };
552 
553 #define	NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
554 
555 static struct fftimehands ffth[10];
556 static struct fftimehands *volatile fftimehands = ffth;
557 
558 static void
559 ffclock_init(void)
560 {
561 	struct fftimehands *cur;
562 	struct fftimehands *last;
563 
564 	memset(ffth, 0, sizeof(ffth));
565 
566 	last = ffth + NUM_ELEMENTS(ffth) - 1;
567 	for (cur = ffth; cur < last; cur++)
568 		cur->next = cur + 1;
569 	last->next = ffth;
570 
571 	ffclock_updated = 0;
572 	ffclock_status = FFCLOCK_STA_UNSYNC;
573 	mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
574 }
575 
576 /*
577  * Reset the feed-forward clock estimates. Called from inittodr() to get things
578  * kick started and uses the timecounter nominal frequency as a first period
579  * estimate. Note: this function may be called several time just after boot.
580  * Note: this is the only function that sets the value of boot time for the
581  * monotonic (i.e. uptime) version of the feed-forward clock.
582  */
583 void
584 ffclock_reset_clock(struct timespec *ts)
585 {
586 	struct timecounter *tc;
587 	struct ffclock_estimate cest;
588 
589 	tc = timehands->th_counter;
590 	memset(&cest, 0, sizeof(struct ffclock_estimate));
591 
592 	timespec2bintime(ts, &ffclock_boottime);
593 	timespec2bintime(ts, &(cest.update_time));
594 	ffclock_read_counter(&cest.update_ffcount);
595 	cest.leapsec_next = 0;
596 	cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
597 	cest.errb_abs = 0;
598 	cest.errb_rate = 0;
599 	cest.status = FFCLOCK_STA_UNSYNC;
600 	cest.leapsec_total = 0;
601 	cest.leapsec = 0;
602 
603 	mtx_lock(&ffclock_mtx);
604 	bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
605 	ffclock_updated = INT8_MAX;
606 	mtx_unlock(&ffclock_mtx);
607 
608 	printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
609 	    (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
610 	    (unsigned long)ts->tv_nsec);
611 }
612 
613 /*
614  * Sub-routine to convert a time interval measured in RAW counter units to time
615  * in seconds stored in bintime format.
616  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
617  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
618  * extra cycles.
619  */
620 static void
621 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
622 {
623 	struct bintime bt2;
624 	ffcounter delta, delta_max;
625 
626 	delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
627 	bintime_clear(bt);
628 	do {
629 		if (ffdelta > delta_max)
630 			delta = delta_max;
631 		else
632 			delta = ffdelta;
633 		bt2.sec = 0;
634 		bt2.frac = period;
635 		bintime_mul(&bt2, (unsigned int)delta);
636 		bintime_add(bt, &bt2);
637 		ffdelta -= delta;
638 	} while (ffdelta > 0);
639 }
640 
641 /*
642  * Update the fftimehands.
643  * Push the tick ffcount and time(s) forward based on current clock estimate.
644  * The conversion from ffcounter to bintime relies on the difference clock
645  * principle, whose accuracy relies on computing small time intervals. If a new
646  * clock estimate has been passed by the synchronisation daemon, make it
647  * current, and compute the linear interpolation for monotonic time if needed.
648  */
649 static void
650 ffclock_windup(unsigned int delta)
651 {
652 	struct ffclock_estimate *cest;
653 	struct fftimehands *ffth;
654 	struct bintime bt, gap_lerp;
655 	ffcounter ffdelta;
656 	uint64_t frac;
657 	unsigned int polling;
658 	uint8_t forward_jump, ogen;
659 
660 	/*
661 	 * Pick the next timehand, copy current ffclock estimates and move tick
662 	 * times and counter forward.
663 	 */
664 	forward_jump = 0;
665 	ffth = fftimehands->next;
666 	ogen = ffth->gen;
667 	ffth->gen = 0;
668 	cest = &ffth->cest;
669 	bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
670 	ffdelta = (ffcounter)delta;
671 	ffth->period_lerp = fftimehands->period_lerp;
672 
673 	ffth->tick_time = fftimehands->tick_time;
674 	ffclock_convert_delta(ffdelta, cest->period, &bt);
675 	bintime_add(&ffth->tick_time, &bt);
676 
677 	ffth->tick_time_lerp = fftimehands->tick_time_lerp;
678 	ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
679 	bintime_add(&ffth->tick_time_lerp, &bt);
680 
681 	ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
682 
683 	/*
684 	 * Assess the status of the clock, if the last update is too old, it is
685 	 * likely the synchronisation daemon is dead and the clock is free
686 	 * running.
687 	 */
688 	if (ffclock_updated == 0) {
689 		ffdelta = ffth->tick_ffcount - cest->update_ffcount;
690 		ffclock_convert_delta(ffdelta, cest->period, &bt);
691 		if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
692 			ffclock_status |= FFCLOCK_STA_UNSYNC;
693 	}
694 
695 	/*
696 	 * If available, grab updated clock estimates and make them current.
697 	 * Recompute time at this tick using the updated estimates. The clock
698 	 * estimates passed the feed-forward synchronisation daemon may result
699 	 * in time conversion that is not monotonically increasing (just after
700 	 * the update). time_lerp is a particular linear interpolation over the
701 	 * synchronisation algo polling period that ensures monotonicity for the
702 	 * clock ids requesting it.
703 	 */
704 	if (ffclock_updated > 0) {
705 		bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
706 		ffdelta = ffth->tick_ffcount - cest->update_ffcount;
707 		ffth->tick_time = cest->update_time;
708 		ffclock_convert_delta(ffdelta, cest->period, &bt);
709 		bintime_add(&ffth->tick_time, &bt);
710 
711 		/* ffclock_reset sets ffclock_updated to INT8_MAX */
712 		if (ffclock_updated == INT8_MAX)
713 			ffth->tick_time_lerp = ffth->tick_time;
714 
715 		if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
716 			forward_jump = 1;
717 		else
718 			forward_jump = 0;
719 
720 		bintime_clear(&gap_lerp);
721 		if (forward_jump) {
722 			gap_lerp = ffth->tick_time;
723 			bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
724 		} else {
725 			gap_lerp = ffth->tick_time_lerp;
726 			bintime_sub(&gap_lerp, &ffth->tick_time);
727 		}
728 
729 		/*
730 		 * The reset from the RTC clock may be far from accurate, and
731 		 * reducing the gap between real time and interpolated time
732 		 * could take a very long time if the interpolated clock insists
733 		 * on strict monotonicity. The clock is reset under very strict
734 		 * conditions (kernel time is known to be wrong and
735 		 * synchronization daemon has been restarted recently.
736 		 * ffclock_boottime absorbs the jump to ensure boot time is
737 		 * correct and uptime functions stay consistent.
738 		 */
739 		if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
740 		    ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
741 		    ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
742 			if (forward_jump)
743 				bintime_add(&ffclock_boottime, &gap_lerp);
744 			else
745 				bintime_sub(&ffclock_boottime, &gap_lerp);
746 			ffth->tick_time_lerp = ffth->tick_time;
747 			bintime_clear(&gap_lerp);
748 		}
749 
750 		ffclock_status = cest->status;
751 		ffth->period_lerp = cest->period;
752 
753 		/*
754 		 * Compute corrected period used for the linear interpolation of
755 		 * time. The rate of linear interpolation is capped to 5000PPM
756 		 * (5ms/s).
757 		 */
758 		if (bintime_isset(&gap_lerp)) {
759 			ffdelta = cest->update_ffcount;
760 			ffdelta -= fftimehands->cest.update_ffcount;
761 			ffclock_convert_delta(ffdelta, cest->period, &bt);
762 			polling = bt.sec;
763 			bt.sec = 0;
764 			bt.frac = 5000000 * (uint64_t)18446744073LL;
765 			bintime_mul(&bt, polling);
766 			if (bintime_cmp(&gap_lerp, &bt, >))
767 				gap_lerp = bt;
768 
769 			/* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
770 			frac = 0;
771 			if (gap_lerp.sec > 0) {
772 				frac -= 1;
773 				frac /= ffdelta / gap_lerp.sec;
774 			}
775 			frac += gap_lerp.frac / ffdelta;
776 
777 			if (forward_jump)
778 				ffth->period_lerp += frac;
779 			else
780 				ffth->period_lerp -= frac;
781 		}
782 
783 		ffclock_updated = 0;
784 	}
785 	if (++ogen == 0)
786 		ogen = 1;
787 	ffth->gen = ogen;
788 	fftimehands = ffth;
789 }
790 
791 /*
792  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
793  * the old and new hardware counter cannot be read simultaneously. tc_windup()
794  * does read the two counters 'back to back', but a few cycles are effectively
795  * lost, and not accumulated in tick_ffcount. This is a fairly radical
796  * operation for a feed-forward synchronization daemon, and it is its job to not
797  * pushing irrelevant data to the kernel. Because there is no locking here,
798  * simply force to ignore pending or next update to give daemon a chance to
799  * realize the counter has changed.
800  */
801 static void
802 ffclock_change_tc(struct timehands *th)
803 {
804 	struct fftimehands *ffth;
805 	struct ffclock_estimate *cest;
806 	struct timecounter *tc;
807 	uint8_t ogen;
808 
809 	tc = th->th_counter;
810 	ffth = fftimehands->next;
811 	ogen = ffth->gen;
812 	ffth->gen = 0;
813 
814 	cest = &ffth->cest;
815 	bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
816 	cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
817 	cest->errb_abs = 0;
818 	cest->errb_rate = 0;
819 	cest->status |= FFCLOCK_STA_UNSYNC;
820 
821 	ffth->tick_ffcount = fftimehands->tick_ffcount;
822 	ffth->tick_time_lerp = fftimehands->tick_time_lerp;
823 	ffth->tick_time = fftimehands->tick_time;
824 	ffth->period_lerp = cest->period;
825 
826 	/* Do not lock but ignore next update from synchronization daemon. */
827 	ffclock_updated--;
828 
829 	if (++ogen == 0)
830 		ogen = 1;
831 	ffth->gen = ogen;
832 	fftimehands = ffth;
833 }
834 
835 /*
836  * Retrieve feed-forward counter and time of last kernel tick.
837  */
838 void
839 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
840 {
841 	struct fftimehands *ffth;
842 	uint8_t gen;
843 
844 	/*
845 	 * No locking but check generation has not changed. Also need to make
846 	 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
847 	 */
848 	do {
849 		ffth = fftimehands;
850 		gen = ffth->gen;
851 		if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
852 			*bt = ffth->tick_time_lerp;
853 		else
854 			*bt = ffth->tick_time;
855 		*ffcount = ffth->tick_ffcount;
856 	} while (gen == 0 || gen != ffth->gen);
857 }
858 
859 /*
860  * Absolute clock conversion. Low level function to convert ffcounter to
861  * bintime. The ffcounter is converted using the current ffclock period estimate
862  * or the "interpolated period" to ensure monotonicity.
863  * NOTE: this conversion may have been deferred, and the clock updated since the
864  * hardware counter has been read.
865  */
866 void
867 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
868 {
869 	struct fftimehands *ffth;
870 	struct bintime bt2;
871 	ffcounter ffdelta;
872 	uint8_t gen;
873 
874 	/*
875 	 * No locking but check generation has not changed. Also need to make
876 	 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
877 	 */
878 	do {
879 		ffth = fftimehands;
880 		gen = ffth->gen;
881 		if (ffcount > ffth->tick_ffcount)
882 			ffdelta = ffcount - ffth->tick_ffcount;
883 		else
884 			ffdelta = ffth->tick_ffcount - ffcount;
885 
886 		if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
887 			*bt = ffth->tick_time_lerp;
888 			ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
889 		} else {
890 			*bt = ffth->tick_time;
891 			ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
892 		}
893 
894 		if (ffcount > ffth->tick_ffcount)
895 			bintime_add(bt, &bt2);
896 		else
897 			bintime_sub(bt, &bt2);
898 	} while (gen == 0 || gen != ffth->gen);
899 }
900 
901 /*
902  * Difference clock conversion.
903  * Low level function to Convert a time interval measured in RAW counter units
904  * into bintime. The difference clock allows measuring small intervals much more
905  * reliably than the absolute clock.
906  */
907 void
908 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
909 {
910 	struct fftimehands *ffth;
911 	uint8_t gen;
912 
913 	/* No locking but check generation has not changed. */
914 	do {
915 		ffth = fftimehands;
916 		gen = ffth->gen;
917 		ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
918 	} while (gen == 0 || gen != ffth->gen);
919 }
920 
921 /*
922  * Access to current ffcounter value.
923  */
924 void
925 ffclock_read_counter(ffcounter *ffcount)
926 {
927 	struct timehands *th;
928 	struct fftimehands *ffth;
929 	unsigned int gen, delta;
930 
931 	/*
932 	 * ffclock_windup() called from tc_windup(), safe to rely on
933 	 * th->th_generation only, for correct delta and ffcounter.
934 	 */
935 	do {
936 		th = timehands;
937 		gen = atomic_load_acq_int(&th->th_generation);
938 		ffth = fftimehands;
939 		delta = tc_delta(th);
940 		*ffcount = ffth->tick_ffcount;
941 		atomic_thread_fence_acq();
942 	} while (gen == 0 || gen != th->th_generation);
943 
944 	*ffcount += delta;
945 }
946 
947 void
948 binuptime(struct bintime *bt)
949 {
950 
951 	binuptime_fromclock(bt, sysclock_active);
952 }
953 
954 void
955 nanouptime(struct timespec *tsp)
956 {
957 
958 	nanouptime_fromclock(tsp, sysclock_active);
959 }
960 
961 void
962 microuptime(struct timeval *tvp)
963 {
964 
965 	microuptime_fromclock(tvp, sysclock_active);
966 }
967 
968 void
969 bintime(struct bintime *bt)
970 {
971 
972 	bintime_fromclock(bt, sysclock_active);
973 }
974 
975 void
976 nanotime(struct timespec *tsp)
977 {
978 
979 	nanotime_fromclock(tsp, sysclock_active);
980 }
981 
982 void
983 microtime(struct timeval *tvp)
984 {
985 
986 	microtime_fromclock(tvp, sysclock_active);
987 }
988 
989 void
990 getbinuptime(struct bintime *bt)
991 {
992 
993 	getbinuptime_fromclock(bt, sysclock_active);
994 }
995 
996 void
997 getnanouptime(struct timespec *tsp)
998 {
999 
1000 	getnanouptime_fromclock(tsp, sysclock_active);
1001 }
1002 
1003 void
1004 getmicrouptime(struct timeval *tvp)
1005 {
1006 
1007 	getmicrouptime_fromclock(tvp, sysclock_active);
1008 }
1009 
1010 void
1011 getbintime(struct bintime *bt)
1012 {
1013 
1014 	getbintime_fromclock(bt, sysclock_active);
1015 }
1016 
1017 void
1018 getnanotime(struct timespec *tsp)
1019 {
1020 
1021 	getnanotime_fromclock(tsp, sysclock_active);
1022 }
1023 
1024 void
1025 getmicrotime(struct timeval *tvp)
1026 {
1027 
1028 	getmicrouptime_fromclock(tvp, sysclock_active);
1029 }
1030 
1031 #endif /* FFCLOCK */
1032 
1033 /*
1034  * This is a clone of getnanotime and used for walltimestamps.
1035  * The dtrace_ prefix prevents fbt from creating probes for
1036  * it so walltimestamp can be safely used in all fbt probes.
1037  */
1038 void
1039 dtrace_getnanotime(struct timespec *tsp)
1040 {
1041 	struct timehands *th;
1042 	u_int gen;
1043 
1044 	do {
1045 		th = timehands;
1046 		gen = atomic_load_acq_int(&th->th_generation);
1047 		*tsp = th->th_nanotime;
1048 		atomic_thread_fence_acq();
1049 	} while (gen == 0 || gen != th->th_generation);
1050 }
1051 
1052 /*
1053  * System clock currently providing time to the system. Modifiable via sysctl
1054  * when the FFCLOCK option is defined.
1055  */
1056 int sysclock_active = SYSCLOCK_FBCK;
1057 
1058 /* Internal NTP status and error estimates. */
1059 extern int time_status;
1060 extern long time_esterror;
1061 
1062 /*
1063  * Take a snapshot of sysclock data which can be used to compare system clocks
1064  * and generate timestamps after the fact.
1065  */
1066 void
1067 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1068 {
1069 	struct fbclock_info *fbi;
1070 	struct timehands *th;
1071 	struct bintime bt;
1072 	unsigned int delta, gen;
1073 #ifdef FFCLOCK
1074 	ffcounter ffcount;
1075 	struct fftimehands *ffth;
1076 	struct ffclock_info *ffi;
1077 	struct ffclock_estimate cest;
1078 
1079 	ffi = &clock_snap->ff_info;
1080 #endif
1081 
1082 	fbi = &clock_snap->fb_info;
1083 	delta = 0;
1084 
1085 	do {
1086 		th = timehands;
1087 		gen = atomic_load_acq_int(&th->th_generation);
1088 		fbi->th_scale = th->th_scale;
1089 		fbi->tick_time = th->th_offset;
1090 #ifdef FFCLOCK
1091 		ffth = fftimehands;
1092 		ffi->tick_time = ffth->tick_time_lerp;
1093 		ffi->tick_time_lerp = ffth->tick_time_lerp;
1094 		ffi->period = ffth->cest.period;
1095 		ffi->period_lerp = ffth->period_lerp;
1096 		clock_snap->ffcount = ffth->tick_ffcount;
1097 		cest = ffth->cest;
1098 #endif
1099 		if (!fast)
1100 			delta = tc_delta(th);
1101 		atomic_thread_fence_acq();
1102 	} while (gen == 0 || gen != th->th_generation);
1103 
1104 	clock_snap->delta = delta;
1105 	clock_snap->sysclock_active = sysclock_active;
1106 
1107 	/* Record feedback clock status and error. */
1108 	clock_snap->fb_info.status = time_status;
1109 	/* XXX: Very crude estimate of feedback clock error. */
1110 	bt.sec = time_esterror / 1000000;
1111 	bt.frac = ((time_esterror - bt.sec) * 1000000) *
1112 	    (uint64_t)18446744073709ULL;
1113 	clock_snap->fb_info.error = bt;
1114 
1115 #ifdef FFCLOCK
1116 	if (!fast)
1117 		clock_snap->ffcount += delta;
1118 
1119 	/* Record feed-forward clock leap second adjustment. */
1120 	ffi->leapsec_adjustment = cest.leapsec_total;
1121 	if (clock_snap->ffcount > cest.leapsec_next)
1122 		ffi->leapsec_adjustment -= cest.leapsec;
1123 
1124 	/* Record feed-forward clock status and error. */
1125 	clock_snap->ff_info.status = cest.status;
1126 	ffcount = clock_snap->ffcount - cest.update_ffcount;
1127 	ffclock_convert_delta(ffcount, cest.period, &bt);
1128 	/* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1129 	bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1130 	/* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1131 	bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1132 	clock_snap->ff_info.error = bt;
1133 #endif
1134 }
1135 
1136 /*
1137  * Convert a sysclock snapshot into a struct bintime based on the specified
1138  * clock source and flags.
1139  */
1140 int
1141 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1142     int whichclock, uint32_t flags)
1143 {
1144 	struct bintime boottimebin;
1145 #ifdef FFCLOCK
1146 	struct bintime bt2;
1147 	uint64_t period;
1148 #endif
1149 
1150 	switch (whichclock) {
1151 	case SYSCLOCK_FBCK:
1152 		*bt = cs->fb_info.tick_time;
1153 
1154 		/* If snapshot was created with !fast, delta will be >0. */
1155 		if (cs->delta > 0)
1156 			bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1157 
1158 		if ((flags & FBCLOCK_UPTIME) == 0) {
1159 			getboottimebin(&boottimebin);
1160 			bintime_add(bt, &boottimebin);
1161 		}
1162 		break;
1163 #ifdef FFCLOCK
1164 	case SYSCLOCK_FFWD:
1165 		if (flags & FFCLOCK_LERP) {
1166 			*bt = cs->ff_info.tick_time_lerp;
1167 			period = cs->ff_info.period_lerp;
1168 		} else {
1169 			*bt = cs->ff_info.tick_time;
1170 			period = cs->ff_info.period;
1171 		}
1172 
1173 		/* If snapshot was created with !fast, delta will be >0. */
1174 		if (cs->delta > 0) {
1175 			ffclock_convert_delta(cs->delta, period, &bt2);
1176 			bintime_add(bt, &bt2);
1177 		}
1178 
1179 		/* Leap second adjustment. */
1180 		if (flags & FFCLOCK_LEAPSEC)
1181 			bt->sec -= cs->ff_info.leapsec_adjustment;
1182 
1183 		/* Boot time adjustment, for uptime/monotonic clocks. */
1184 		if (flags & FFCLOCK_UPTIME)
1185 			bintime_sub(bt, &ffclock_boottime);
1186 		break;
1187 #endif
1188 	default:
1189 		return (EINVAL);
1190 		break;
1191 	}
1192 
1193 	return (0);
1194 }
1195 
1196 /*
1197  * Initialize a new timecounter and possibly use it.
1198  */
1199 void
1200 tc_init(struct timecounter *tc)
1201 {
1202 	u_int u;
1203 	struct sysctl_oid *tc_root;
1204 
1205 	u = tc->tc_frequency / tc->tc_counter_mask;
1206 	/* XXX: We need some margin here, 10% is a guess */
1207 	u *= 11;
1208 	u /= 10;
1209 	if (u > hz && tc->tc_quality >= 0) {
1210 		tc->tc_quality = -2000;
1211 		if (bootverbose) {
1212 			printf("Timecounter \"%s\" frequency %ju Hz",
1213 			    tc->tc_name, (uintmax_t)tc->tc_frequency);
1214 			printf(" -- Insufficient hz, needs at least %u\n", u);
1215 		}
1216 	} else if (tc->tc_quality >= 0 || bootverbose) {
1217 		printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1218 		    tc->tc_name, (uintmax_t)tc->tc_frequency,
1219 		    tc->tc_quality);
1220 	}
1221 
1222 	tc->tc_next = timecounters;
1223 	timecounters = tc;
1224 	/*
1225 	 * Set up sysctl tree for this counter.
1226 	 */
1227 	tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1228 	    SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1229 	    CTLFLAG_RW, 0, "timecounter description", "timecounter");
1230 	SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1231 	    "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1232 	    "mask for implemented bits");
1233 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1234 	    "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1235 	    sysctl_kern_timecounter_get, "IU", "current timecounter value");
1236 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1237 	    "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1238 	     sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1239 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1240 	    "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1241 	    "goodness of time counter");
1242 	/*
1243 	 * Do not automatically switch if the current tc was specifically
1244 	 * chosen.  Never automatically use a timecounter with negative quality.
1245 	 * Even though we run on the dummy counter, switching here may be
1246 	 * worse since this timecounter may not be monotonic.
1247 	 */
1248 	if (tc_chosen)
1249 		return;
1250 	if (tc->tc_quality < 0)
1251 		return;
1252 	if (tc->tc_quality < timecounter->tc_quality)
1253 		return;
1254 	if (tc->tc_quality == timecounter->tc_quality &&
1255 	    tc->tc_frequency < timecounter->tc_frequency)
1256 		return;
1257 	(void)tc->tc_get_timecount(tc);
1258 	(void)tc->tc_get_timecount(tc);
1259 	timecounter = tc;
1260 }
1261 
1262 /* Report the frequency of the current timecounter. */
1263 uint64_t
1264 tc_getfrequency(void)
1265 {
1266 
1267 	return (timehands->th_counter->tc_frequency);
1268 }
1269 
1270 static bool
1271 sleeping_on_old_rtc(struct thread *td)
1272 {
1273 
1274 	/*
1275 	 * td_rtcgen is modified by curthread when it is running,
1276 	 * and by other threads in this function.  By finding the thread
1277 	 * on a sleepqueue and holding the lock on the sleepqueue
1278 	 * chain, we guarantee that the thread is not running and that
1279 	 * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
1280 	 * the thread that it was woken due to a real-time clock adjustment.
1281 	 * (The declaration of td_rtcgen refers to this comment.)
1282 	 */
1283 	if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1284 		td->td_rtcgen = 0;
1285 		return (true);
1286 	}
1287 	return (false);
1288 }
1289 
1290 static struct mtx tc_setclock_mtx;
1291 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1292 
1293 /*
1294  * Step our concept of UTC.  This is done by modifying our estimate of
1295  * when we booted.
1296  */
1297 void
1298 tc_setclock(struct timespec *ts)
1299 {
1300 	struct timespec tbef, taft;
1301 	struct bintime bt, bt2;
1302 
1303 	timespec2bintime(ts, &bt);
1304 	nanotime(&tbef);
1305 	mtx_lock_spin(&tc_setclock_mtx);
1306 	cpu_tick_calibrate(1);
1307 	binuptime(&bt2);
1308 	bintime_sub(&bt, &bt2);
1309 
1310 	/* XXX fiddle all the little crinkly bits around the fiords... */
1311 	tc_windup(&bt);
1312 	mtx_unlock_spin(&tc_setclock_mtx);
1313 
1314 	/* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1315 	atomic_add_rel_int(&rtc_generation, 2);
1316 	sleepq_chains_remove_matching(sleeping_on_old_rtc);
1317 	if (timestepwarnings) {
1318 		nanotime(&taft);
1319 		log(LOG_INFO,
1320 		    "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1321 		    (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1322 		    (intmax_t)taft.tv_sec, taft.tv_nsec,
1323 		    (intmax_t)ts->tv_sec, ts->tv_nsec);
1324 	}
1325 }
1326 
1327 /*
1328  * Initialize the next struct timehands in the ring and make
1329  * it the active timehands.  Along the way we might switch to a different
1330  * timecounter and/or do seconds processing in NTP.  Slightly magic.
1331  */
1332 static void
1333 tc_windup(struct bintime *new_boottimebin)
1334 {
1335 	struct bintime bt;
1336 	struct timehands *th, *tho;
1337 	uint64_t scale;
1338 	u_int delta, ncount, ogen;
1339 	int i;
1340 	time_t t;
1341 
1342 	/*
1343 	 * Make the next timehands a copy of the current one, but do
1344 	 * not overwrite the generation or next pointer.  While we
1345 	 * update the contents, the generation must be zero.  We need
1346 	 * to ensure that the zero generation is visible before the
1347 	 * data updates become visible, which requires release fence.
1348 	 * For similar reasons, re-reading of the generation after the
1349 	 * data is read should use acquire fence.
1350 	 */
1351 	tho = timehands;
1352 	th = tho->th_next;
1353 	ogen = th->th_generation;
1354 	th->th_generation = 0;
1355 	atomic_thread_fence_rel();
1356 	memcpy(th, tho, offsetof(struct timehands, th_generation));
1357 	if (new_boottimebin != NULL)
1358 		th->th_boottime = *new_boottimebin;
1359 
1360 	/*
1361 	 * Capture a timecounter delta on the current timecounter and if
1362 	 * changing timecounters, a counter value from the new timecounter.
1363 	 * Update the offset fields accordingly.
1364 	 */
1365 	delta = tc_delta(th);
1366 	if (th->th_counter != timecounter)
1367 		ncount = timecounter->tc_get_timecount(timecounter);
1368 	else
1369 		ncount = 0;
1370 #ifdef FFCLOCK
1371 	ffclock_windup(delta);
1372 #endif
1373 	th->th_offset_count += delta;
1374 	th->th_offset_count &= th->th_counter->tc_counter_mask;
1375 	while (delta > th->th_counter->tc_frequency) {
1376 		/* Eat complete unadjusted seconds. */
1377 		delta -= th->th_counter->tc_frequency;
1378 		th->th_offset.sec++;
1379 	}
1380 	if ((delta > th->th_counter->tc_frequency / 2) &&
1381 	    (th->th_scale * delta < ((uint64_t)1 << 63))) {
1382 		/* The product th_scale * delta just barely overflows. */
1383 		th->th_offset.sec++;
1384 	}
1385 	bintime_addx(&th->th_offset, th->th_scale * delta);
1386 
1387 	/*
1388 	 * Hardware latching timecounters may not generate interrupts on
1389 	 * PPS events, so instead we poll them.  There is a finite risk that
1390 	 * the hardware might capture a count which is later than the one we
1391 	 * got above, and therefore possibly in the next NTP second which might
1392 	 * have a different rate than the current NTP second.  It doesn't
1393 	 * matter in practice.
1394 	 */
1395 	if (tho->th_counter->tc_poll_pps)
1396 		tho->th_counter->tc_poll_pps(tho->th_counter);
1397 
1398 	/*
1399 	 * Deal with NTP second processing.  The for loop normally
1400 	 * iterates at most once, but in extreme situations it might
1401 	 * keep NTP sane if timeouts are not run for several seconds.
1402 	 * At boot, the time step can be large when the TOD hardware
1403 	 * has been read, so on really large steps, we call
1404 	 * ntp_update_second only twice.  We need to call it twice in
1405 	 * case we missed a leap second.
1406 	 */
1407 	bt = th->th_offset;
1408 	bintime_add(&bt, &th->th_boottime);
1409 	i = bt.sec - tho->th_microtime.tv_sec;
1410 	if (i > LARGE_STEP)
1411 		i = 2;
1412 	for (; i > 0; i--) {
1413 		t = bt.sec;
1414 		ntp_update_second(&th->th_adjustment, &bt.sec);
1415 		if (bt.sec != t)
1416 			th->th_boottime.sec += bt.sec - t;
1417 	}
1418 	/* Update the UTC timestamps used by the get*() functions. */
1419 	th->th_bintime = bt;
1420 	bintime2timeval(&bt, &th->th_microtime);
1421 	bintime2timespec(&bt, &th->th_nanotime);
1422 
1423 	/* Now is a good time to change timecounters. */
1424 	if (th->th_counter != timecounter) {
1425 #ifndef __arm__
1426 		if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1427 			cpu_disable_c2_sleep++;
1428 		if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1429 			cpu_disable_c2_sleep--;
1430 #endif
1431 		th->th_counter = timecounter;
1432 		th->th_offset_count = ncount;
1433 		tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1434 		    (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1435 #ifdef FFCLOCK
1436 		ffclock_change_tc(th);
1437 #endif
1438 	}
1439 
1440 	/*-
1441 	 * Recalculate the scaling factor.  We want the number of 1/2^64
1442 	 * fractions of a second per period of the hardware counter, taking
1443 	 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1444 	 * processing provides us with.
1445 	 *
1446 	 * The th_adjustment is nanoseconds per second with 32 bit binary
1447 	 * fraction and we want 64 bit binary fraction of second:
1448 	 *
1449 	 *	 x = a * 2^32 / 10^9 = a * 4.294967296
1450 	 *
1451 	 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1452 	 * we can only multiply by about 850 without overflowing, that
1453 	 * leaves no suitably precise fractions for multiply before divide.
1454 	 *
1455 	 * Divide before multiply with a fraction of 2199/512 results in a
1456 	 * systematic undercompensation of 10PPM of th_adjustment.  On a
1457 	 * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
1458  	 *
1459 	 * We happily sacrifice the lowest of the 64 bits of our result
1460 	 * to the goddess of code clarity.
1461 	 *
1462 	 */
1463 	scale = (uint64_t)1 << 63;
1464 	scale += (th->th_adjustment / 1024) * 2199;
1465 	scale /= th->th_counter->tc_frequency;
1466 	th->th_scale = scale * 2;
1467 
1468 	/*
1469 	 * Now that the struct timehands is again consistent, set the new
1470 	 * generation number, making sure to not make it zero.
1471 	 */
1472 	if (++ogen == 0)
1473 		ogen = 1;
1474 	atomic_store_rel_int(&th->th_generation, ogen);
1475 
1476 	/* Go live with the new struct timehands. */
1477 #ifdef FFCLOCK
1478 	switch (sysclock_active) {
1479 	case SYSCLOCK_FBCK:
1480 #endif
1481 		time_second = th->th_microtime.tv_sec;
1482 		time_uptime = th->th_offset.sec;
1483 #ifdef FFCLOCK
1484 		break;
1485 	case SYSCLOCK_FFWD:
1486 		time_second = fftimehands->tick_time_lerp.sec;
1487 		time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1488 		break;
1489 	}
1490 #endif
1491 
1492 	timehands = th;
1493 	timekeep_push_vdso();
1494 }
1495 
1496 /* Report or change the active timecounter hardware. */
1497 static int
1498 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1499 {
1500 	char newname[32];
1501 	struct timecounter *newtc, *tc;
1502 	int error;
1503 
1504 	tc = timecounter;
1505 	strlcpy(newname, tc->tc_name, sizeof(newname));
1506 
1507 	error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1508 	if (error != 0 || req->newptr == NULL)
1509 		return (error);
1510 	/* Record that the tc in use now was specifically chosen. */
1511 	tc_chosen = 1;
1512 	if (strcmp(newname, tc->tc_name) == 0)
1513 		return (0);
1514 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1515 		if (strcmp(newname, newtc->tc_name) != 0)
1516 			continue;
1517 
1518 		/* Warm up new timecounter. */
1519 		(void)newtc->tc_get_timecount(newtc);
1520 		(void)newtc->tc_get_timecount(newtc);
1521 
1522 		timecounter = newtc;
1523 
1524 		/*
1525 		 * The vdso timehands update is deferred until the next
1526 		 * 'tc_windup()'.
1527 		 *
1528 		 * This is prudent given that 'timekeep_push_vdso()' does not
1529 		 * use any locking and that it can be called in hard interrupt
1530 		 * context via 'tc_windup()'.
1531 		 */
1532 		return (0);
1533 	}
1534 	return (EINVAL);
1535 }
1536 
1537 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1538     0, 0, sysctl_kern_timecounter_hardware, "A",
1539     "Timecounter hardware selected");
1540 
1541 
1542 /* Report the available timecounter hardware. */
1543 static int
1544 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1545 {
1546 	struct sbuf sb;
1547 	struct timecounter *tc;
1548 	int error;
1549 
1550 	sbuf_new_for_sysctl(&sb, NULL, 0, req);
1551 	for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1552 		if (tc != timecounters)
1553 			sbuf_putc(&sb, ' ');
1554 		sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1555 	}
1556 	error = sbuf_finish(&sb);
1557 	sbuf_delete(&sb);
1558 	return (error);
1559 }
1560 
1561 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1562     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1563 
1564 /*
1565  * RFC 2783 PPS-API implementation.
1566  */
1567 
1568 /*
1569  *  Return true if the driver is aware of the abi version extensions in the
1570  *  pps_state structure, and it supports at least the given abi version number.
1571  */
1572 static inline int
1573 abi_aware(struct pps_state *pps, int vers)
1574 {
1575 
1576 	return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1577 }
1578 
1579 static int
1580 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1581 {
1582 	int err, timo;
1583 	pps_seq_t aseq, cseq;
1584 	struct timeval tv;
1585 
1586 	if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1587 		return (EINVAL);
1588 
1589 	/*
1590 	 * If no timeout is requested, immediately return whatever values were
1591 	 * most recently captured.  If timeout seconds is -1, that's a request
1592 	 * to block without a timeout.  WITNESS won't let us sleep forever
1593 	 * without a lock (we really don't need a lock), so just repeatedly
1594 	 * sleep a long time.
1595 	 */
1596 	if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1597 		if (fapi->timeout.tv_sec == -1)
1598 			timo = 0x7fffffff;
1599 		else {
1600 			tv.tv_sec = fapi->timeout.tv_sec;
1601 			tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1602 			timo = tvtohz(&tv);
1603 		}
1604 		aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1605 		cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1606 		while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1607 		    cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1608 			if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1609 				if (pps->flags & PPSFLAG_MTX_SPIN) {
1610 					err = msleep_spin(pps, pps->driver_mtx,
1611 					    "ppsfch", timo);
1612 				} else {
1613 					err = msleep(pps, pps->driver_mtx, PCATCH,
1614 					    "ppsfch", timo);
1615 				}
1616 			} else {
1617 				err = tsleep(pps, PCATCH, "ppsfch", timo);
1618 			}
1619 			if (err == EWOULDBLOCK) {
1620 				if (fapi->timeout.tv_sec == -1) {
1621 					continue;
1622 				} else {
1623 					return (ETIMEDOUT);
1624 				}
1625 			} else if (err != 0) {
1626 				return (err);
1627 			}
1628 		}
1629 	}
1630 
1631 	pps->ppsinfo.current_mode = pps->ppsparam.mode;
1632 	fapi->pps_info_buf = pps->ppsinfo;
1633 
1634 	return (0);
1635 }
1636 
1637 int
1638 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1639 {
1640 	pps_params_t *app;
1641 	struct pps_fetch_args *fapi;
1642 #ifdef FFCLOCK
1643 	struct pps_fetch_ffc_args *fapi_ffc;
1644 #endif
1645 #ifdef PPS_SYNC
1646 	struct pps_kcbind_args *kapi;
1647 #endif
1648 
1649 	KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1650 	switch (cmd) {
1651 	case PPS_IOC_CREATE:
1652 		return (0);
1653 	case PPS_IOC_DESTROY:
1654 		return (0);
1655 	case PPS_IOC_SETPARAMS:
1656 		app = (pps_params_t *)data;
1657 		if (app->mode & ~pps->ppscap)
1658 			return (EINVAL);
1659 #ifdef FFCLOCK
1660 		/* Ensure only a single clock is selected for ffc timestamp. */
1661 		if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1662 			return (EINVAL);
1663 #endif
1664 		pps->ppsparam = *app;
1665 		return (0);
1666 	case PPS_IOC_GETPARAMS:
1667 		app = (pps_params_t *)data;
1668 		*app = pps->ppsparam;
1669 		app->api_version = PPS_API_VERS_1;
1670 		return (0);
1671 	case PPS_IOC_GETCAP:
1672 		*(int*)data = pps->ppscap;
1673 		return (0);
1674 	case PPS_IOC_FETCH:
1675 		fapi = (struct pps_fetch_args *)data;
1676 		return (pps_fetch(fapi, pps));
1677 #ifdef FFCLOCK
1678 	case PPS_IOC_FETCH_FFCOUNTER:
1679 		fapi_ffc = (struct pps_fetch_ffc_args *)data;
1680 		if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1681 		    PPS_TSFMT_TSPEC)
1682 			return (EINVAL);
1683 		if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1684 			return (EOPNOTSUPP);
1685 		pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1686 		fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1687 		/* Overwrite timestamps if feedback clock selected. */
1688 		switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1689 		case PPS_TSCLK_FBCK:
1690 			fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1691 			    pps->ppsinfo.assert_timestamp;
1692 			fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1693 			    pps->ppsinfo.clear_timestamp;
1694 			break;
1695 		case PPS_TSCLK_FFWD:
1696 			break;
1697 		default:
1698 			break;
1699 		}
1700 		return (0);
1701 #endif /* FFCLOCK */
1702 	case PPS_IOC_KCBIND:
1703 #ifdef PPS_SYNC
1704 		kapi = (struct pps_kcbind_args *)data;
1705 		/* XXX Only root should be able to do this */
1706 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1707 			return (EINVAL);
1708 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1709 			return (EINVAL);
1710 		if (kapi->edge & ~pps->ppscap)
1711 			return (EINVAL);
1712 		pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1713 		    (pps->kcmode & KCMODE_ABIFLAG);
1714 		return (0);
1715 #else
1716 		return (EOPNOTSUPP);
1717 #endif
1718 	default:
1719 		return (ENOIOCTL);
1720 	}
1721 }
1722 
1723 void
1724 pps_init(struct pps_state *pps)
1725 {
1726 	pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1727 	if (pps->ppscap & PPS_CAPTUREASSERT)
1728 		pps->ppscap |= PPS_OFFSETASSERT;
1729 	if (pps->ppscap & PPS_CAPTURECLEAR)
1730 		pps->ppscap |= PPS_OFFSETCLEAR;
1731 #ifdef FFCLOCK
1732 	pps->ppscap |= PPS_TSCLK_MASK;
1733 #endif
1734 	pps->kcmode &= ~KCMODE_ABIFLAG;
1735 }
1736 
1737 void
1738 pps_init_abi(struct pps_state *pps)
1739 {
1740 
1741 	pps_init(pps);
1742 	if (pps->driver_abi > 0) {
1743 		pps->kcmode |= KCMODE_ABIFLAG;
1744 		pps->kernel_abi = PPS_ABI_VERSION;
1745 	}
1746 }
1747 
1748 void
1749 pps_capture(struct pps_state *pps)
1750 {
1751 	struct timehands *th;
1752 
1753 	KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1754 	th = timehands;
1755 	pps->capgen = atomic_load_acq_int(&th->th_generation);
1756 	pps->capth = th;
1757 #ifdef FFCLOCK
1758 	pps->capffth = fftimehands;
1759 #endif
1760 	pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1761 	atomic_thread_fence_acq();
1762 	if (pps->capgen != th->th_generation)
1763 		pps->capgen = 0;
1764 }
1765 
1766 void
1767 pps_event(struct pps_state *pps, int event)
1768 {
1769 	struct bintime bt;
1770 	struct timespec ts, *tsp, *osp;
1771 	u_int tcount, *pcount;
1772 	int foff;
1773 	pps_seq_t *pseq;
1774 #ifdef FFCLOCK
1775 	struct timespec *tsp_ffc;
1776 	pps_seq_t *pseq_ffc;
1777 	ffcounter *ffcount;
1778 #endif
1779 #ifdef PPS_SYNC
1780 	int fhard;
1781 #endif
1782 
1783 	KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1784 	/* Nothing to do if not currently set to capture this event type. */
1785 	if ((event & pps->ppsparam.mode) == 0)
1786 		return;
1787 	/* If the timecounter was wound up underneath us, bail out. */
1788 	if (pps->capgen == 0 || pps->capgen !=
1789 	    atomic_load_acq_int(&pps->capth->th_generation))
1790 		return;
1791 
1792 	/* Things would be easier with arrays. */
1793 	if (event == PPS_CAPTUREASSERT) {
1794 		tsp = &pps->ppsinfo.assert_timestamp;
1795 		osp = &pps->ppsparam.assert_offset;
1796 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1797 #ifdef PPS_SYNC
1798 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1799 #endif
1800 		pcount = &pps->ppscount[0];
1801 		pseq = &pps->ppsinfo.assert_sequence;
1802 #ifdef FFCLOCK
1803 		ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1804 		tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1805 		pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1806 #endif
1807 	} else {
1808 		tsp = &pps->ppsinfo.clear_timestamp;
1809 		osp = &pps->ppsparam.clear_offset;
1810 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1811 #ifdef PPS_SYNC
1812 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1813 #endif
1814 		pcount = &pps->ppscount[1];
1815 		pseq = &pps->ppsinfo.clear_sequence;
1816 #ifdef FFCLOCK
1817 		ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1818 		tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1819 		pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1820 #endif
1821 	}
1822 
1823 	/*
1824 	 * If the timecounter changed, we cannot compare the count values, so
1825 	 * we have to drop the rest of the PPS-stuff until the next event.
1826 	 */
1827 	if (pps->ppstc != pps->capth->th_counter) {
1828 		pps->ppstc = pps->capth->th_counter;
1829 		*pcount = pps->capcount;
1830 		pps->ppscount[2] = pps->capcount;
1831 		return;
1832 	}
1833 
1834 	/* Convert the count to a timespec. */
1835 	tcount = pps->capcount - pps->capth->th_offset_count;
1836 	tcount &= pps->capth->th_counter->tc_counter_mask;
1837 	bt = pps->capth->th_bintime;
1838 	bintime_addx(&bt, pps->capth->th_scale * tcount);
1839 	bintime2timespec(&bt, &ts);
1840 
1841 	/* If the timecounter was wound up underneath us, bail out. */
1842 	atomic_thread_fence_acq();
1843 	if (pps->capgen != pps->capth->th_generation)
1844 		return;
1845 
1846 	*pcount = pps->capcount;
1847 	(*pseq)++;
1848 	*tsp = ts;
1849 
1850 	if (foff) {
1851 		timespecadd(tsp, osp, tsp);
1852 		if (tsp->tv_nsec < 0) {
1853 			tsp->tv_nsec += 1000000000;
1854 			tsp->tv_sec -= 1;
1855 		}
1856 	}
1857 
1858 #ifdef FFCLOCK
1859 	*ffcount = pps->capffth->tick_ffcount + tcount;
1860 	bt = pps->capffth->tick_time;
1861 	ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1862 	bintime_add(&bt, &pps->capffth->tick_time);
1863 	bintime2timespec(&bt, &ts);
1864 	(*pseq_ffc)++;
1865 	*tsp_ffc = ts;
1866 #endif
1867 
1868 #ifdef PPS_SYNC
1869 	if (fhard) {
1870 		uint64_t scale;
1871 
1872 		/*
1873 		 * Feed the NTP PLL/FLL.
1874 		 * The FLL wants to know how many (hardware) nanoseconds
1875 		 * elapsed since the previous event.
1876 		 */
1877 		tcount = pps->capcount - pps->ppscount[2];
1878 		pps->ppscount[2] = pps->capcount;
1879 		tcount &= pps->capth->th_counter->tc_counter_mask;
1880 		scale = (uint64_t)1 << 63;
1881 		scale /= pps->capth->th_counter->tc_frequency;
1882 		scale *= 2;
1883 		bt.sec = 0;
1884 		bt.frac = 0;
1885 		bintime_addx(&bt, scale * tcount);
1886 		bintime2timespec(&bt, &ts);
1887 		hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1888 	}
1889 #endif
1890 
1891 	/* Wakeup anyone sleeping in pps_fetch().  */
1892 	wakeup(pps);
1893 }
1894 
1895 /*
1896  * Timecounters need to be updated every so often to prevent the hardware
1897  * counter from overflowing.  Updating also recalculates the cached values
1898  * used by the get*() family of functions, so their precision depends on
1899  * the update frequency.
1900  */
1901 
1902 static int tc_tick;
1903 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1904     "Approximate number of hardclock ticks in a millisecond");
1905 
1906 void
1907 tc_ticktock(int cnt)
1908 {
1909 	static int count;
1910 
1911 	if (mtx_trylock_spin(&tc_setclock_mtx)) {
1912 		count += cnt;
1913 		if (count >= tc_tick) {
1914 			count = 0;
1915 			tc_windup(NULL);
1916 		}
1917 		mtx_unlock_spin(&tc_setclock_mtx);
1918 	}
1919 }
1920 
1921 static void __inline
1922 tc_adjprecision(void)
1923 {
1924 	int t;
1925 
1926 	if (tc_timepercentage > 0) {
1927 		t = (99 + tc_timepercentage) / tc_timepercentage;
1928 		tc_precexp = fls(t + (t >> 1)) - 1;
1929 		FREQ2BT(hz / tc_tick, &bt_timethreshold);
1930 		FREQ2BT(hz, &bt_tickthreshold);
1931 		bintime_shift(&bt_timethreshold, tc_precexp);
1932 		bintime_shift(&bt_tickthreshold, tc_precexp);
1933 	} else {
1934 		tc_precexp = 31;
1935 		bt_timethreshold.sec = INT_MAX;
1936 		bt_timethreshold.frac = ~(uint64_t)0;
1937 		bt_tickthreshold = bt_timethreshold;
1938 	}
1939 	sbt_timethreshold = bttosbt(bt_timethreshold);
1940 	sbt_tickthreshold = bttosbt(bt_tickthreshold);
1941 }
1942 
1943 static int
1944 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1945 {
1946 	int error, val;
1947 
1948 	val = tc_timepercentage;
1949 	error = sysctl_handle_int(oidp, &val, 0, req);
1950 	if (error != 0 || req->newptr == NULL)
1951 		return (error);
1952 	tc_timepercentage = val;
1953 	if (cold)
1954 		goto done;
1955 	tc_adjprecision();
1956 done:
1957 	return (0);
1958 }
1959 
1960 static void
1961 inittimecounter(void *dummy)
1962 {
1963 	u_int p;
1964 	int tick_rate;
1965 
1966 	/*
1967 	 * Set the initial timeout to
1968 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
1969 	 * People should probably not use the sysctl to set the timeout
1970 	 * to smaller than its initial value, since that value is the
1971 	 * smallest reasonable one.  If they want better timestamps they
1972 	 * should use the non-"get"* functions.
1973 	 */
1974 	if (hz > 1000)
1975 		tc_tick = (hz + 500) / 1000;
1976 	else
1977 		tc_tick = 1;
1978 	tc_adjprecision();
1979 	FREQ2BT(hz, &tick_bt);
1980 	tick_sbt = bttosbt(tick_bt);
1981 	tick_rate = hz / tc_tick;
1982 	FREQ2BT(tick_rate, &tc_tick_bt);
1983 	tc_tick_sbt = bttosbt(tc_tick_bt);
1984 	p = (tc_tick * 1000000) / hz;
1985 	printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1986 
1987 #ifdef FFCLOCK
1988 	ffclock_init();
1989 #endif
1990 	/* warm up new timecounter (again) and get rolling. */
1991 	(void)timecounter->tc_get_timecount(timecounter);
1992 	(void)timecounter->tc_get_timecount(timecounter);
1993 	mtx_lock_spin(&tc_setclock_mtx);
1994 	tc_windup(NULL);
1995 	mtx_unlock_spin(&tc_setclock_mtx);
1996 }
1997 
1998 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
1999 
2000 /* Cpu tick handling -------------------------------------------------*/
2001 
2002 static int cpu_tick_variable;
2003 static uint64_t	cpu_tick_frequency;
2004 
2005 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2006 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2007 
2008 static uint64_t
2009 tc_cpu_ticks(void)
2010 {
2011 	struct timecounter *tc;
2012 	uint64_t res, *base;
2013 	unsigned u, *last;
2014 
2015 	critical_enter();
2016 	base = DPCPU_PTR(tc_cpu_ticks_base);
2017 	last = DPCPU_PTR(tc_cpu_ticks_last);
2018 	tc = timehands->th_counter;
2019 	u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2020 	if (u < *last)
2021 		*base += (uint64_t)tc->tc_counter_mask + 1;
2022 	*last = u;
2023 	res = u + *base;
2024 	critical_exit();
2025 	return (res);
2026 }
2027 
2028 void
2029 cpu_tick_calibration(void)
2030 {
2031 	static time_t last_calib;
2032 
2033 	if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2034 		cpu_tick_calibrate(0);
2035 		last_calib = time_uptime;
2036 	}
2037 }
2038 
2039 /*
2040  * This function gets called every 16 seconds on only one designated
2041  * CPU in the system from hardclock() via cpu_tick_calibration()().
2042  *
2043  * Whenever the real time clock is stepped we get called with reset=1
2044  * to make sure we handle suspend/resume and similar events correctly.
2045  */
2046 
2047 static void
2048 cpu_tick_calibrate(int reset)
2049 {
2050 	static uint64_t c_last;
2051 	uint64_t c_this, c_delta;
2052 	static struct bintime  t_last;
2053 	struct bintime t_this, t_delta;
2054 	uint32_t divi;
2055 
2056 	if (reset) {
2057 		/* The clock was stepped, abort & reset */
2058 		t_last.sec = 0;
2059 		return;
2060 	}
2061 
2062 	/* we don't calibrate fixed rate cputicks */
2063 	if (!cpu_tick_variable)
2064 		return;
2065 
2066 	getbinuptime(&t_this);
2067 	c_this = cpu_ticks();
2068 	if (t_last.sec != 0) {
2069 		c_delta = c_this - c_last;
2070 		t_delta = t_this;
2071 		bintime_sub(&t_delta, &t_last);
2072 		/*
2073 		 * Headroom:
2074 		 * 	2^(64-20) / 16[s] =
2075 		 * 	2^(44) / 16[s] =
2076 		 * 	17.592.186.044.416 / 16 =
2077 		 * 	1.099.511.627.776 [Hz]
2078 		 */
2079 		divi = t_delta.sec << 20;
2080 		divi |= t_delta.frac >> (64 - 20);
2081 		c_delta <<= 20;
2082 		c_delta /= divi;
2083 		if (c_delta > cpu_tick_frequency) {
2084 			if (0 && bootverbose)
2085 				printf("cpu_tick increased to %ju Hz\n",
2086 				    c_delta);
2087 			cpu_tick_frequency = c_delta;
2088 		}
2089 	}
2090 	c_last = c_this;
2091 	t_last = t_this;
2092 }
2093 
2094 void
2095 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2096 {
2097 
2098 	if (func == NULL) {
2099 		cpu_ticks = tc_cpu_ticks;
2100 	} else {
2101 		cpu_tick_frequency = freq;
2102 		cpu_tick_variable = var;
2103 		cpu_ticks = func;
2104 	}
2105 }
2106 
2107 uint64_t
2108 cpu_tickrate(void)
2109 {
2110 
2111 	if (cpu_ticks == tc_cpu_ticks)
2112 		return (tc_getfrequency());
2113 	return (cpu_tick_frequency);
2114 }
2115 
2116 /*
2117  * We need to be slightly careful converting cputicks to microseconds.
2118  * There is plenty of margin in 64 bits of microseconds (half a million
2119  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2120  * before divide conversion (to retain precision) we find that the
2121  * margin shrinks to 1.5 hours (one millionth of 146y).
2122  * With a three prong approach we never lose significant bits, no
2123  * matter what the cputick rate and length of timeinterval is.
2124  */
2125 
2126 uint64_t
2127 cputick2usec(uint64_t tick)
2128 {
2129 
2130 	if (tick > 18446744073709551LL)		/* floor(2^64 / 1000) */
2131 		return (tick / (cpu_tickrate() / 1000000LL));
2132 	else if (tick > 18446744073709LL)	/* floor(2^64 / 1000000) */
2133 		return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2134 	else
2135 		return ((tick * 1000000LL) / cpu_tickrate());
2136 }
2137 
2138 cpu_tick_f	*cpu_ticks = tc_cpu_ticks;
2139 
2140 static int vdso_th_enable = 1;
2141 static int
2142 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2143 {
2144 	int old_vdso_th_enable, error;
2145 
2146 	old_vdso_th_enable = vdso_th_enable;
2147 	error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2148 	if (error != 0)
2149 		return (error);
2150 	vdso_th_enable = old_vdso_th_enable;
2151 	return (0);
2152 }
2153 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2154     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2155     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2156 
2157 uint32_t
2158 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2159 {
2160 	struct timehands *th;
2161 	uint32_t enabled;
2162 
2163 	th = timehands;
2164 	vdso_th->th_scale = th->th_scale;
2165 	vdso_th->th_offset_count = th->th_offset_count;
2166 	vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2167 	vdso_th->th_offset = th->th_offset;
2168 	vdso_th->th_boottime = th->th_boottime;
2169 	if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2170 		enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2171 		    th->th_counter);
2172 	} else
2173 		enabled = 0;
2174 	if (!vdso_th_enable)
2175 		enabled = 0;
2176 	return (enabled);
2177 }
2178 
2179 #ifdef COMPAT_FREEBSD32
2180 uint32_t
2181 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2182 {
2183 	struct timehands *th;
2184 	uint32_t enabled;
2185 
2186 	th = timehands;
2187 	*(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2188 	vdso_th32->th_offset_count = th->th_offset_count;
2189 	vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2190 	vdso_th32->th_offset.sec = th->th_offset.sec;
2191 	*(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2192 	vdso_th32->th_boottime.sec = th->th_boottime.sec;
2193 	*(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2194 	if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2195 		enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2196 		    th->th_counter);
2197 	} else
2198 		enabled = 0;
2199 	if (!vdso_th_enable)
2200 		enabled = 0;
2201 	return (enabled);
2202 }
2203 #endif
2204