xref: /freebsd/sys/kern/kern_tc.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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  *
13  * Portions of this software were developed by Julien Ridoux at the University
14  * of Melbourne under sponsorship from the FreeBSD Foundation.
15  *
16  * Portions of this software were developed by Konstantin Belousov
17  * under sponsorship from the FreeBSD Foundation.
18  */
19 
20 #include <sys/cdefs.h>
21 #include "opt_ntp.h"
22 #include "opt_ffclock.h"
23 
24 #include <sys/param.h>
25 #include <sys/kernel.h>
26 #include <sys/limits.h>
27 #include <sys/lock.h>
28 #include <sys/mutex.h>
29 #include <sys/proc.h>
30 #include <sys/sbuf.h>
31 #include <sys/sleepqueue.h>
32 #include <sys/sysctl.h>
33 #include <sys/syslog.h>
34 #include <sys/systm.h>
35 #include <sys/timeffc.h>
36 #include <sys/timepps.h>
37 #include <sys/timerfd.h>
38 #include <sys/timetc.h>
39 #include <sys/timex.h>
40 #include <sys/vdso.h>
41 
42 /*
43  * A large step happens on boot.  This constant detects such steps.
44  * It is relatively small so that ntp_update_second gets called enough
45  * in the typical 'missed a couple of seconds' case, but doesn't loop
46  * forever when the time step is large.
47  */
48 #define LARGE_STEP	200
49 
50 /*
51  * Implement a dummy timecounter which we can use until we get a real one
52  * in the air.  This allows the console and other early stuff to use
53  * time services.
54  */
55 
56 static u_int
57 dummy_get_timecount(struct timecounter *tc)
58 {
59 	static u_int now;
60 
61 	return (++now);
62 }
63 
64 static struct timecounter dummy_timecounter = {
65 	dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
66 };
67 
68 struct timehands {
69 	/* These fields must be initialized by the driver. */
70 	struct timecounter	*th_counter;
71 	int64_t			th_adjustment;
72 	uint64_t		th_scale;
73 	u_int			th_large_delta;
74 	u_int	 		th_offset_count;
75 	struct bintime		th_offset;
76 	struct bintime		th_bintime;
77 	struct timeval		th_microtime;
78 	struct timespec		th_nanotime;
79 	struct bintime		th_boottime;
80 	/* Fields not to be copied in tc_windup start with th_generation. */
81 	u_int			th_generation;
82 	struct timehands	*th_next;
83 };
84 
85 static struct timehands ths[16] = {
86     [0] =  {
87 	.th_counter = &dummy_timecounter,
88 	.th_scale = (uint64_t)-1 / 1000000,
89 	.th_large_delta = 1000000,
90 	.th_offset = { .sec = 1 },
91 	.th_generation = 1,
92     },
93 };
94 
95 static struct timehands *volatile timehands = &ths[0];
96 struct timecounter *timecounter = &dummy_timecounter;
97 static struct timecounter *timecounters = &dummy_timecounter;
98 
99 /* Mutex to protect the timecounter list. */
100 static struct mtx tc_lock;
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 /*
108  * The system time is always computed by summing the estimated boot time and the
109  * system uptime. The timehands track boot time, but it changes when the system
110  * time is set by the user, stepped by ntpd or adjusted when resuming. It
111  * is set to new_time - uptime.
112  */
113 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
114 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
115     CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
116     sysctl_kern_boottime, "S,timeval",
117     "Estimated system boottime");
118 
119 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
120     "");
121 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
122     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
123     "");
124 
125 static int timestepwarnings;
126 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
127     &timestepwarnings, 0, "Log time steps");
128 
129 static int timehands_count = 2;
130 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
131     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
132     &timehands_count, 0, "Count of timehands in rotation");
133 
134 struct bintime bt_timethreshold;
135 struct bintime bt_tickthreshold;
136 sbintime_t sbt_timethreshold;
137 sbintime_t sbt_tickthreshold;
138 struct bintime tc_tick_bt;
139 sbintime_t tc_tick_sbt;
140 int tc_precexp;
141 int tc_timepercentage = TC_DEFAULTPERC;
142 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
143 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
144     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
145     sysctl_kern_timecounter_adjprecision, "I",
146     "Allowed time interval deviation in percents");
147 
148 volatile int rtc_generation = 1;
149 
150 static int tc_chosen;	/* Non-zero if a specific tc was chosen via sysctl. */
151 static char tc_from_tunable[16];
152 
153 static void tc_windup(struct bintime *new_boottimebin);
154 static void cpu_tick_calibrate(int);
155 
156 void dtrace_getnanotime(struct timespec *tsp);
157 void dtrace_getnanouptime(struct timespec *tsp);
158 
159 static int
160 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
161 {
162 	struct timeval boottime;
163 
164 	getboottime(&boottime);
165 
166 /* i386 is the only arch which uses a 32bits time_t */
167 #ifdef __amd64__
168 #ifdef SCTL_MASK32
169 	int tv[2];
170 
171 	if (req->flags & SCTL_MASK32) {
172 		tv[0] = boottime.tv_sec;
173 		tv[1] = boottime.tv_usec;
174 		return (SYSCTL_OUT(req, tv, sizeof(tv)));
175 	}
176 #endif
177 #endif
178 	return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
179 }
180 
181 static int
182 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
183 {
184 	u_int ncount;
185 	struct timecounter *tc = arg1;
186 
187 	ncount = tc->tc_get_timecount(tc);
188 	return (sysctl_handle_int(oidp, &ncount, 0, req));
189 }
190 
191 static int
192 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
193 {
194 	uint64_t freq;
195 	struct timecounter *tc = arg1;
196 
197 	freq = tc->tc_frequency;
198 	return (sysctl_handle_64(oidp, &freq, 0, req));
199 }
200 
201 /*
202  * Return the difference between the timehands' counter value now and what
203  * was when we copied it to the timehands' offset_count.
204  */
205 static __inline u_int
206 tc_delta(struct timehands *th)
207 {
208 	struct timecounter *tc;
209 
210 	tc = th->th_counter;
211 	return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
212 	    tc->tc_counter_mask);
213 }
214 
215 static __inline void
216 bintime_add_tc_delta(struct bintime *bt, uint64_t scale,
217     uint64_t large_delta, uint64_t delta)
218 {
219 	uint64_t x;
220 
221 	if (__predict_false(delta >= large_delta)) {
222 		/* Avoid overflow for scale * delta. */
223 		x = (scale >> 32) * delta;
224 		bt->sec += x >> 32;
225 		bintime_addx(bt, x << 32);
226 		bintime_addx(bt, (scale & 0xffffffff) * delta);
227 	} else {
228 		bintime_addx(bt, scale * delta);
229 	}
230 }
231 
232 /*
233  * Functions for reading the time.  We have to loop until we are sure that
234  * the timehands that we operated on was not updated under our feet.  See
235  * the comment in <sys/time.h> for a description of these 12 functions.
236  */
237 
238 static __inline void
239 bintime_off(struct bintime *bt, u_int off)
240 {
241 	struct timehands *th;
242 	struct bintime *btp;
243 	uint64_t scale;
244 	u_int delta, gen, large_delta;
245 
246 	do {
247 		th = timehands;
248 		gen = atomic_load_acq_int(&th->th_generation);
249 		btp = (struct bintime *)((vm_offset_t)th + off);
250 		*bt = *btp;
251 		scale = th->th_scale;
252 		delta = tc_delta(th);
253 		large_delta = th->th_large_delta;
254 		atomic_thread_fence_acq();
255 	} while (gen == 0 || gen != th->th_generation);
256 
257 	bintime_add_tc_delta(bt, scale, large_delta, delta);
258 }
259 #define	GETTHBINTIME(dst, member)					\
260 do {									\
261 	_Static_assert(_Generic(((struct timehands *)NULL)->member,	\
262 	    struct bintime: 1, default: 0) == 1,			\
263 	    "struct timehands member is not of struct bintime type");	\
264 	bintime_off(dst, __offsetof(struct timehands, member));		\
265 } while (0)
266 
267 static __inline void
268 getthmember(void *out, size_t out_size, u_int off)
269 {
270 	struct timehands *th;
271 	u_int gen;
272 
273 	do {
274 		th = timehands;
275 		gen = atomic_load_acq_int(&th->th_generation);
276 		memcpy(out, (char *)th + off, out_size);
277 		atomic_thread_fence_acq();
278 	} while (gen == 0 || gen != th->th_generation);
279 }
280 #define	GETTHMEMBER(dst, member)					\
281 do {									\
282 	_Static_assert(_Generic(*dst,					\
283 	    __typeof(((struct timehands *)NULL)->member): 1,		\
284 	    default: 0) == 1,						\
285 	    "*dst and struct timehands member have different types");	\
286 	getthmember(dst, sizeof(*dst), __offsetof(struct timehands,	\
287 	    member));							\
288 } while (0)
289 
290 #ifdef FFCLOCK
291 void
292 fbclock_binuptime(struct bintime *bt)
293 {
294 
295 	GETTHBINTIME(bt, th_offset);
296 }
297 
298 void
299 fbclock_nanouptime(struct timespec *tsp)
300 {
301 	struct bintime bt;
302 
303 	fbclock_binuptime(&bt);
304 	bintime2timespec(&bt, tsp);
305 }
306 
307 void
308 fbclock_microuptime(struct timeval *tvp)
309 {
310 	struct bintime bt;
311 
312 	fbclock_binuptime(&bt);
313 	bintime2timeval(&bt, tvp);
314 }
315 
316 void
317 fbclock_bintime(struct bintime *bt)
318 {
319 
320 	GETTHBINTIME(bt, th_bintime);
321 }
322 
323 void
324 fbclock_nanotime(struct timespec *tsp)
325 {
326 	struct bintime bt;
327 
328 	fbclock_bintime(&bt);
329 	bintime2timespec(&bt, tsp);
330 }
331 
332 void
333 fbclock_microtime(struct timeval *tvp)
334 {
335 	struct bintime bt;
336 
337 	fbclock_bintime(&bt);
338 	bintime2timeval(&bt, tvp);
339 }
340 
341 void
342 fbclock_getbinuptime(struct bintime *bt)
343 {
344 
345 	GETTHMEMBER(bt, th_offset);
346 }
347 
348 void
349 fbclock_getnanouptime(struct timespec *tsp)
350 {
351 	struct bintime bt;
352 
353 	GETTHMEMBER(&bt, th_offset);
354 	bintime2timespec(&bt, tsp);
355 }
356 
357 void
358 fbclock_getmicrouptime(struct timeval *tvp)
359 {
360 	struct bintime bt;
361 
362 	GETTHMEMBER(&bt, th_offset);
363 	bintime2timeval(&bt, tvp);
364 }
365 
366 void
367 fbclock_getbintime(struct bintime *bt)
368 {
369 
370 	GETTHMEMBER(bt, th_bintime);
371 }
372 
373 void
374 fbclock_getnanotime(struct timespec *tsp)
375 {
376 
377 	GETTHMEMBER(tsp, th_nanotime);
378 }
379 
380 void
381 fbclock_getmicrotime(struct timeval *tvp)
382 {
383 
384 	GETTHMEMBER(tvp, th_microtime);
385 }
386 #else /* !FFCLOCK */
387 
388 void
389 binuptime(struct bintime *bt)
390 {
391 
392 	GETTHBINTIME(bt, th_offset);
393 }
394 
395 void
396 nanouptime(struct timespec *tsp)
397 {
398 	struct bintime bt;
399 
400 	binuptime(&bt);
401 	bintime2timespec(&bt, tsp);
402 }
403 
404 void
405 microuptime(struct timeval *tvp)
406 {
407 	struct bintime bt;
408 
409 	binuptime(&bt);
410 	bintime2timeval(&bt, tvp);
411 }
412 
413 void
414 bintime(struct bintime *bt)
415 {
416 
417 	GETTHBINTIME(bt, th_bintime);
418 }
419 
420 void
421 nanotime(struct timespec *tsp)
422 {
423 	struct bintime bt;
424 
425 	bintime(&bt);
426 	bintime2timespec(&bt, tsp);
427 }
428 
429 void
430 microtime(struct timeval *tvp)
431 {
432 	struct bintime bt;
433 
434 	bintime(&bt);
435 	bintime2timeval(&bt, tvp);
436 }
437 
438 void
439 getbinuptime(struct bintime *bt)
440 {
441 
442 	GETTHMEMBER(bt, th_offset);
443 }
444 
445 void
446 getnanouptime(struct timespec *tsp)
447 {
448 	struct bintime bt;
449 
450 	GETTHMEMBER(&bt, th_offset);
451 	bintime2timespec(&bt, tsp);
452 }
453 
454 void
455 getmicrouptime(struct timeval *tvp)
456 {
457 	struct bintime bt;
458 
459 	GETTHMEMBER(&bt, th_offset);
460 	bintime2timeval(&bt, tvp);
461 }
462 
463 void
464 getbintime(struct bintime *bt)
465 {
466 
467 	GETTHMEMBER(bt, th_bintime);
468 }
469 
470 void
471 getnanotime(struct timespec *tsp)
472 {
473 
474 	GETTHMEMBER(tsp, th_nanotime);
475 }
476 
477 void
478 getmicrotime(struct timeval *tvp)
479 {
480 
481 	GETTHMEMBER(tvp, th_microtime);
482 }
483 #endif /* FFCLOCK */
484 
485 void
486 getboottime(struct timeval *boottime)
487 {
488 	struct bintime boottimebin;
489 
490 	getboottimebin(&boottimebin);
491 	bintime2timeval(&boottimebin, boottime);
492 }
493 
494 void
495 getboottimebin(struct bintime *boottimebin)
496 {
497 
498 	GETTHMEMBER(boottimebin, th_boottime);
499 }
500 
501 #ifdef FFCLOCK
502 /*
503  * Support for feed-forward synchronization algorithms. This is heavily inspired
504  * by the timehands mechanism but kept independent from it. *_windup() functions
505  * have some connection to avoid accessing the timecounter hardware more than
506  * necessary.
507  */
508 
509 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
510 struct ffclock_estimate ffclock_estimate;
511 struct bintime ffclock_boottime;	/* Feed-forward boot time estimate. */
512 uint32_t ffclock_status;		/* Feed-forward clock status. */
513 int8_t ffclock_updated;			/* New estimates are available. */
514 struct mtx ffclock_mtx;			/* Mutex on ffclock_estimate. */
515 
516 struct fftimehands {
517 	struct ffclock_estimate	cest;
518 	struct bintime		tick_time;
519 	struct bintime		tick_time_lerp;
520 	ffcounter		tick_ffcount;
521 	uint64_t		period_lerp;
522 	volatile uint8_t	gen;
523 	struct fftimehands	*next;
524 };
525 
526 #define	NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
527 
528 static struct fftimehands ffth[10];
529 static struct fftimehands *volatile fftimehands = ffth;
530 
531 static void
532 ffclock_init(void)
533 {
534 	struct fftimehands *cur;
535 	struct fftimehands *last;
536 
537 	memset(ffth, 0, sizeof(ffth));
538 
539 	last = ffth + NUM_ELEMENTS(ffth) - 1;
540 	for (cur = ffth; cur < last; cur++)
541 		cur->next = cur + 1;
542 	last->next = ffth;
543 
544 	ffclock_updated = 0;
545 	ffclock_status = FFCLOCK_STA_UNSYNC;
546 	mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
547 }
548 
549 /*
550  * Reset the feed-forward clock estimates. Called from inittodr() to get things
551  * kick started and uses the timecounter nominal frequency as a first period
552  * estimate. Note: this function may be called several time just after boot.
553  * Note: this is the only function that sets the value of boot time for the
554  * monotonic (i.e. uptime) version of the feed-forward clock.
555  */
556 void
557 ffclock_reset_clock(struct timespec *ts)
558 {
559 	struct timecounter *tc;
560 	struct ffclock_estimate cest;
561 
562 	tc = timehands->th_counter;
563 	memset(&cest, 0, sizeof(struct ffclock_estimate));
564 
565 	timespec2bintime(ts, &ffclock_boottime);
566 	timespec2bintime(ts, &(cest.update_time));
567 	ffclock_read_counter(&cest.update_ffcount);
568 	cest.leapsec_next = 0;
569 	cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
570 	cest.errb_abs = 0;
571 	cest.errb_rate = 0;
572 	cest.status = FFCLOCK_STA_UNSYNC;
573 	cest.leapsec_total = 0;
574 	cest.leapsec = 0;
575 
576 	mtx_lock(&ffclock_mtx);
577 	bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
578 	ffclock_updated = INT8_MAX;
579 	mtx_unlock(&ffclock_mtx);
580 
581 	printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
582 	    (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
583 	    (unsigned long)ts->tv_nsec);
584 }
585 
586 /*
587  * Sub-routine to convert a time interval measured in RAW counter units to time
588  * in seconds stored in bintime format.
589  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
590  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
591  * extra cycles.
592  */
593 static void
594 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
595 {
596 	struct bintime bt2;
597 	ffcounter delta, delta_max;
598 
599 	delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
600 	bintime_clear(bt);
601 	do {
602 		if (ffdelta > delta_max)
603 			delta = delta_max;
604 		else
605 			delta = ffdelta;
606 		bt2.sec = 0;
607 		bt2.frac = period;
608 		bintime_mul(&bt2, (unsigned int)delta);
609 		bintime_add(bt, &bt2);
610 		ffdelta -= delta;
611 	} while (ffdelta > 0);
612 }
613 
614 /*
615  * Update the fftimehands.
616  * Push the tick ffcount and time(s) forward based on current clock estimate.
617  * The conversion from ffcounter to bintime relies on the difference clock
618  * principle, whose accuracy relies on computing small time intervals. If a new
619  * clock estimate has been passed by the synchronisation daemon, make it
620  * current, and compute the linear interpolation for monotonic time if needed.
621  */
622 static void
623 ffclock_windup(unsigned int delta)
624 {
625 	struct ffclock_estimate *cest;
626 	struct fftimehands *ffth;
627 	struct bintime bt, gap_lerp;
628 	ffcounter ffdelta;
629 	uint64_t frac;
630 	unsigned int polling;
631 	uint8_t forward_jump, ogen;
632 
633 	/*
634 	 * Pick the next timehand, copy current ffclock estimates and move tick
635 	 * times and counter forward.
636 	 */
637 	forward_jump = 0;
638 	ffth = fftimehands->next;
639 	ogen = ffth->gen;
640 	ffth->gen = 0;
641 	cest = &ffth->cest;
642 	bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
643 	ffdelta = (ffcounter)delta;
644 	ffth->period_lerp = fftimehands->period_lerp;
645 
646 	ffth->tick_time = fftimehands->tick_time;
647 	ffclock_convert_delta(ffdelta, cest->period, &bt);
648 	bintime_add(&ffth->tick_time, &bt);
649 
650 	ffth->tick_time_lerp = fftimehands->tick_time_lerp;
651 	ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
652 	bintime_add(&ffth->tick_time_lerp, &bt);
653 
654 	ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
655 
656 	/*
657 	 * Assess the status of the clock, if the last update is too old, it is
658 	 * likely the synchronisation daemon is dead and the clock is free
659 	 * running.
660 	 */
661 	if (ffclock_updated == 0) {
662 		ffdelta = ffth->tick_ffcount - cest->update_ffcount;
663 		ffclock_convert_delta(ffdelta, cest->period, &bt);
664 		if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
665 			ffclock_status |= FFCLOCK_STA_UNSYNC;
666 	}
667 
668 	/*
669 	 * If available, grab updated clock estimates and make them current.
670 	 * Recompute time at this tick using the updated estimates. The clock
671 	 * estimates passed the feed-forward synchronisation daemon may result
672 	 * in time conversion that is not monotonically increasing (just after
673 	 * the update). time_lerp is a particular linear interpolation over the
674 	 * synchronisation algo polling period that ensures monotonicity for the
675 	 * clock ids requesting it.
676 	 */
677 	if (ffclock_updated > 0) {
678 		bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
679 		ffdelta = ffth->tick_ffcount - cest->update_ffcount;
680 		ffth->tick_time = cest->update_time;
681 		ffclock_convert_delta(ffdelta, cest->period, &bt);
682 		bintime_add(&ffth->tick_time, &bt);
683 
684 		/* ffclock_reset sets ffclock_updated to INT8_MAX */
685 		if (ffclock_updated == INT8_MAX)
686 			ffth->tick_time_lerp = ffth->tick_time;
687 
688 		if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
689 			forward_jump = 1;
690 		else
691 			forward_jump = 0;
692 
693 		bintime_clear(&gap_lerp);
694 		if (forward_jump) {
695 			gap_lerp = ffth->tick_time;
696 			bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
697 		} else {
698 			gap_lerp = ffth->tick_time_lerp;
699 			bintime_sub(&gap_lerp, &ffth->tick_time);
700 		}
701 
702 		/*
703 		 * The reset from the RTC clock may be far from accurate, and
704 		 * reducing the gap between real time and interpolated time
705 		 * could take a very long time if the interpolated clock insists
706 		 * on strict monotonicity. The clock is reset under very strict
707 		 * conditions (kernel time is known to be wrong and
708 		 * synchronization daemon has been restarted recently.
709 		 * ffclock_boottime absorbs the jump to ensure boot time is
710 		 * correct and uptime functions stay consistent.
711 		 */
712 		if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
713 		    ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
714 		    ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
715 			if (forward_jump)
716 				bintime_add(&ffclock_boottime, &gap_lerp);
717 			else
718 				bintime_sub(&ffclock_boottime, &gap_lerp);
719 			ffth->tick_time_lerp = ffth->tick_time;
720 			bintime_clear(&gap_lerp);
721 		}
722 
723 		ffclock_status = cest->status;
724 		ffth->period_lerp = cest->period;
725 
726 		/*
727 		 * Compute corrected period used for the linear interpolation of
728 		 * time. The rate of linear interpolation is capped to 5000PPM
729 		 * (5ms/s).
730 		 */
731 		if (bintime_isset(&gap_lerp)) {
732 			ffdelta = cest->update_ffcount;
733 			ffdelta -= fftimehands->cest.update_ffcount;
734 			ffclock_convert_delta(ffdelta, cest->period, &bt);
735 			polling = bt.sec;
736 			bt.sec = 0;
737 			bt.frac = 5000000 * (uint64_t)18446744073LL;
738 			bintime_mul(&bt, polling);
739 			if (bintime_cmp(&gap_lerp, &bt, >))
740 				gap_lerp = bt;
741 
742 			/* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
743 			frac = 0;
744 			if (gap_lerp.sec > 0) {
745 				frac -= 1;
746 				frac /= ffdelta / gap_lerp.sec;
747 			}
748 			frac += gap_lerp.frac / ffdelta;
749 
750 			if (forward_jump)
751 				ffth->period_lerp += frac;
752 			else
753 				ffth->period_lerp -= frac;
754 		}
755 
756 		ffclock_updated = 0;
757 	}
758 	if (++ogen == 0)
759 		ogen = 1;
760 	ffth->gen = ogen;
761 	fftimehands = ffth;
762 }
763 
764 /*
765  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
766  * the old and new hardware counter cannot be read simultaneously. tc_windup()
767  * does read the two counters 'back to back', but a few cycles are effectively
768  * lost, and not accumulated in tick_ffcount. This is a fairly radical
769  * operation for a feed-forward synchronization daemon, and it is its job to not
770  * pushing irrelevant data to the kernel. Because there is no locking here,
771  * simply force to ignore pending or next update to give daemon a chance to
772  * realize the counter has changed.
773  */
774 static void
775 ffclock_change_tc(struct timehands *th)
776 {
777 	struct fftimehands *ffth;
778 	struct ffclock_estimate *cest;
779 	struct timecounter *tc;
780 	uint8_t ogen;
781 
782 	tc = th->th_counter;
783 	ffth = fftimehands->next;
784 	ogen = ffth->gen;
785 	ffth->gen = 0;
786 
787 	cest = &ffth->cest;
788 	bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
789 	cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
790 	cest->errb_abs = 0;
791 	cest->errb_rate = 0;
792 	cest->status |= FFCLOCK_STA_UNSYNC;
793 
794 	ffth->tick_ffcount = fftimehands->tick_ffcount;
795 	ffth->tick_time_lerp = fftimehands->tick_time_lerp;
796 	ffth->tick_time = fftimehands->tick_time;
797 	ffth->period_lerp = cest->period;
798 
799 	/* Do not lock but ignore next update from synchronization daemon. */
800 	ffclock_updated--;
801 
802 	if (++ogen == 0)
803 		ogen = 1;
804 	ffth->gen = ogen;
805 	fftimehands = ffth;
806 }
807 
808 /*
809  * Retrieve feed-forward counter and time of last kernel tick.
810  */
811 void
812 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
813 {
814 	struct fftimehands *ffth;
815 	uint8_t gen;
816 
817 	/*
818 	 * No locking but check generation has not changed. Also need to make
819 	 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
820 	 */
821 	do {
822 		ffth = fftimehands;
823 		gen = ffth->gen;
824 		if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
825 			*bt = ffth->tick_time_lerp;
826 		else
827 			*bt = ffth->tick_time;
828 		*ffcount = ffth->tick_ffcount;
829 	} while (gen == 0 || gen != ffth->gen);
830 }
831 
832 /*
833  * Absolute clock conversion. Low level function to convert ffcounter to
834  * bintime. The ffcounter is converted using the current ffclock period estimate
835  * or the "interpolated period" to ensure monotonicity.
836  * NOTE: this conversion may have been deferred, and the clock updated since the
837  * hardware counter has been read.
838  */
839 void
840 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
841 {
842 	struct fftimehands *ffth;
843 	struct bintime bt2;
844 	ffcounter ffdelta;
845 	uint8_t gen;
846 
847 	/*
848 	 * No locking but check generation has not changed. Also need to make
849 	 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
850 	 */
851 	do {
852 		ffth = fftimehands;
853 		gen = ffth->gen;
854 		if (ffcount > ffth->tick_ffcount)
855 			ffdelta = ffcount - ffth->tick_ffcount;
856 		else
857 			ffdelta = ffth->tick_ffcount - ffcount;
858 
859 		if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
860 			*bt = ffth->tick_time_lerp;
861 			ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
862 		} else {
863 			*bt = ffth->tick_time;
864 			ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
865 		}
866 
867 		if (ffcount > ffth->tick_ffcount)
868 			bintime_add(bt, &bt2);
869 		else
870 			bintime_sub(bt, &bt2);
871 	} while (gen == 0 || gen != ffth->gen);
872 }
873 
874 /*
875  * Difference clock conversion.
876  * Low level function to Convert a time interval measured in RAW counter units
877  * into bintime. The difference clock allows measuring small intervals much more
878  * reliably than the absolute clock.
879  */
880 void
881 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
882 {
883 	struct fftimehands *ffth;
884 	uint8_t gen;
885 
886 	/* No locking but check generation has not changed. */
887 	do {
888 		ffth = fftimehands;
889 		gen = ffth->gen;
890 		ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
891 	} while (gen == 0 || gen != ffth->gen);
892 }
893 
894 /*
895  * Access to current ffcounter value.
896  */
897 void
898 ffclock_read_counter(ffcounter *ffcount)
899 {
900 	struct timehands *th;
901 	struct fftimehands *ffth;
902 	unsigned int gen, delta;
903 
904 	/*
905 	 * ffclock_windup() called from tc_windup(), safe to rely on
906 	 * th->th_generation only, for correct delta and ffcounter.
907 	 */
908 	do {
909 		th = timehands;
910 		gen = atomic_load_acq_int(&th->th_generation);
911 		ffth = fftimehands;
912 		delta = tc_delta(th);
913 		*ffcount = ffth->tick_ffcount;
914 		atomic_thread_fence_acq();
915 	} while (gen == 0 || gen != th->th_generation);
916 
917 	*ffcount += delta;
918 }
919 
920 void
921 binuptime(struct bintime *bt)
922 {
923 
924 	binuptime_fromclock(bt, sysclock_active);
925 }
926 
927 void
928 nanouptime(struct timespec *tsp)
929 {
930 
931 	nanouptime_fromclock(tsp, sysclock_active);
932 }
933 
934 void
935 microuptime(struct timeval *tvp)
936 {
937 
938 	microuptime_fromclock(tvp, sysclock_active);
939 }
940 
941 void
942 bintime(struct bintime *bt)
943 {
944 
945 	bintime_fromclock(bt, sysclock_active);
946 }
947 
948 void
949 nanotime(struct timespec *tsp)
950 {
951 
952 	nanotime_fromclock(tsp, sysclock_active);
953 }
954 
955 void
956 microtime(struct timeval *tvp)
957 {
958 
959 	microtime_fromclock(tvp, sysclock_active);
960 }
961 
962 void
963 getbinuptime(struct bintime *bt)
964 {
965 
966 	getbinuptime_fromclock(bt, sysclock_active);
967 }
968 
969 void
970 getnanouptime(struct timespec *tsp)
971 {
972 
973 	getnanouptime_fromclock(tsp, sysclock_active);
974 }
975 
976 void
977 getmicrouptime(struct timeval *tvp)
978 {
979 
980 	getmicrouptime_fromclock(tvp, sysclock_active);
981 }
982 
983 void
984 getbintime(struct bintime *bt)
985 {
986 
987 	getbintime_fromclock(bt, sysclock_active);
988 }
989 
990 void
991 getnanotime(struct timespec *tsp)
992 {
993 
994 	getnanotime_fromclock(tsp, sysclock_active);
995 }
996 
997 void
998 getmicrotime(struct timeval *tvp)
999 {
1000 
1001 	getmicrouptime_fromclock(tvp, sysclock_active);
1002 }
1003 
1004 #endif /* FFCLOCK */
1005 
1006 /*
1007  * This is a clone of getnanotime and used for walltimestamps.
1008  * The dtrace_ prefix prevents fbt from creating probes for
1009  * it so walltimestamp can be safely used in all fbt probes.
1010  */
1011 void
1012 dtrace_getnanotime(struct timespec *tsp)
1013 {
1014 
1015 	GETTHMEMBER(tsp, th_nanotime);
1016 }
1017 
1018 /*
1019  * This is a clone of getnanouptime used for time since boot.
1020  * The dtrace_ prefix prevents fbt from creating probes for
1021  * it so an uptime that can be safely used in all fbt probes.
1022  */
1023 void
1024 dtrace_getnanouptime(struct timespec *tsp)
1025 {
1026 	struct bintime bt;
1027 
1028 	GETTHMEMBER(&bt, th_offset);
1029 	bintime2timespec(&bt, tsp);
1030 }
1031 
1032 /*
1033  * System clock currently providing time to the system. Modifiable via sysctl
1034  * when the FFCLOCK option is defined.
1035  */
1036 int sysclock_active = SYSCLOCK_FBCK;
1037 
1038 /* Internal NTP status and error estimates. */
1039 extern int time_status;
1040 extern long time_esterror;
1041 
1042 /*
1043  * Take a snapshot of sysclock data which can be used to compare system clocks
1044  * and generate timestamps after the fact.
1045  */
1046 void
1047 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1048 {
1049 	struct fbclock_info *fbi;
1050 	struct timehands *th;
1051 	struct bintime bt;
1052 	unsigned int delta, gen;
1053 #ifdef FFCLOCK
1054 	ffcounter ffcount;
1055 	struct fftimehands *ffth;
1056 	struct ffclock_info *ffi;
1057 	struct ffclock_estimate cest;
1058 
1059 	ffi = &clock_snap->ff_info;
1060 #endif
1061 
1062 	fbi = &clock_snap->fb_info;
1063 	delta = 0;
1064 
1065 	do {
1066 		th = timehands;
1067 		gen = atomic_load_acq_int(&th->th_generation);
1068 		fbi->th_scale = th->th_scale;
1069 		fbi->tick_time = th->th_offset;
1070 #ifdef FFCLOCK
1071 		ffth = fftimehands;
1072 		ffi->tick_time = ffth->tick_time_lerp;
1073 		ffi->tick_time_lerp = ffth->tick_time_lerp;
1074 		ffi->period = ffth->cest.period;
1075 		ffi->period_lerp = ffth->period_lerp;
1076 		clock_snap->ffcount = ffth->tick_ffcount;
1077 		cest = ffth->cest;
1078 #endif
1079 		if (!fast)
1080 			delta = tc_delta(th);
1081 		atomic_thread_fence_acq();
1082 	} while (gen == 0 || gen != th->th_generation);
1083 
1084 	clock_snap->delta = delta;
1085 	clock_snap->sysclock_active = sysclock_active;
1086 
1087 	/* Record feedback clock status and error. */
1088 	clock_snap->fb_info.status = time_status;
1089 	/* XXX: Very crude estimate of feedback clock error. */
1090 	bt.sec = time_esterror / 1000000;
1091 	bt.frac = ((time_esterror - bt.sec) * 1000000) *
1092 	    (uint64_t)18446744073709ULL;
1093 	clock_snap->fb_info.error = bt;
1094 
1095 #ifdef FFCLOCK
1096 	if (!fast)
1097 		clock_snap->ffcount += delta;
1098 
1099 	/* Record feed-forward clock leap second adjustment. */
1100 	ffi->leapsec_adjustment = cest.leapsec_total;
1101 	if (clock_snap->ffcount > cest.leapsec_next)
1102 		ffi->leapsec_adjustment -= cest.leapsec;
1103 
1104 	/* Record feed-forward clock status and error. */
1105 	clock_snap->ff_info.status = cest.status;
1106 	ffcount = clock_snap->ffcount - cest.update_ffcount;
1107 	ffclock_convert_delta(ffcount, cest.period, &bt);
1108 	/* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1109 	bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1110 	/* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1111 	bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1112 	clock_snap->ff_info.error = bt;
1113 #endif
1114 }
1115 
1116 /*
1117  * Convert a sysclock snapshot into a struct bintime based on the specified
1118  * clock source and flags.
1119  */
1120 int
1121 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1122     int whichclock, uint32_t flags)
1123 {
1124 	struct bintime boottimebin;
1125 #ifdef FFCLOCK
1126 	struct bintime bt2;
1127 	uint64_t period;
1128 #endif
1129 
1130 	switch (whichclock) {
1131 	case SYSCLOCK_FBCK:
1132 		*bt = cs->fb_info.tick_time;
1133 
1134 		/* If snapshot was created with !fast, delta will be >0. */
1135 		if (cs->delta > 0)
1136 			bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1137 
1138 		if ((flags & FBCLOCK_UPTIME) == 0) {
1139 			getboottimebin(&boottimebin);
1140 			bintime_add(bt, &boottimebin);
1141 		}
1142 		break;
1143 #ifdef FFCLOCK
1144 	case SYSCLOCK_FFWD:
1145 		if (flags & FFCLOCK_LERP) {
1146 			*bt = cs->ff_info.tick_time_lerp;
1147 			period = cs->ff_info.period_lerp;
1148 		} else {
1149 			*bt = cs->ff_info.tick_time;
1150 			period = cs->ff_info.period;
1151 		}
1152 
1153 		/* If snapshot was created with !fast, delta will be >0. */
1154 		if (cs->delta > 0) {
1155 			ffclock_convert_delta(cs->delta, period, &bt2);
1156 			bintime_add(bt, &bt2);
1157 		}
1158 
1159 		/* Leap second adjustment. */
1160 		if (flags & FFCLOCK_LEAPSEC)
1161 			bt->sec -= cs->ff_info.leapsec_adjustment;
1162 
1163 		/* Boot time adjustment, for uptime/monotonic clocks. */
1164 		if (flags & FFCLOCK_UPTIME)
1165 			bintime_sub(bt, &ffclock_boottime);
1166 		break;
1167 #endif
1168 	default:
1169 		return (EINVAL);
1170 		break;
1171 	}
1172 
1173 	return (0);
1174 }
1175 
1176 /*
1177  * Initialize a new timecounter and possibly use it.
1178  */
1179 void
1180 tc_init(struct timecounter *tc)
1181 {
1182 	u_int u;
1183 	struct sysctl_oid *tc_root;
1184 
1185 	u = tc->tc_frequency / tc->tc_counter_mask;
1186 	/* XXX: We need some margin here, 10% is a guess */
1187 	u *= 11;
1188 	u /= 10;
1189 	if (u > hz && tc->tc_quality >= 0) {
1190 		tc->tc_quality = -2000;
1191 		if (bootverbose) {
1192 			printf("Timecounter \"%s\" frequency %ju Hz",
1193 			    tc->tc_name, (uintmax_t)tc->tc_frequency);
1194 			printf(" -- Insufficient hz, needs at least %u\n", u);
1195 		}
1196 	} else if (tc->tc_quality >= 0 || bootverbose) {
1197 		printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1198 		    tc->tc_name, (uintmax_t)tc->tc_frequency,
1199 		    tc->tc_quality);
1200 	}
1201 
1202 	/*
1203 	 * Set up sysctl tree for this counter.
1204 	 */
1205 	tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1206 	    SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1207 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1208 	    "timecounter description", "timecounter");
1209 	SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1210 	    "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1211 	    "mask for implemented bits");
1212 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1213 	    "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1214 	    sizeof(*tc), sysctl_kern_timecounter_get, "IU",
1215 	    "current timecounter value");
1216 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1217 	    "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1218 	    sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
1219 	    "timecounter frequency");
1220 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1221 	    "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1222 	    "goodness of time counter");
1223 
1224 	mtx_lock(&tc_lock);
1225 	tc->tc_next = timecounters;
1226 	timecounters = tc;
1227 
1228 	/*
1229 	 * Do not automatically switch if the current tc was specifically
1230 	 * chosen.  Never automatically use a timecounter with negative quality.
1231 	 * Even though we run on the dummy counter, switching here may be
1232 	 * worse since this timecounter may not be monotonic.
1233 	 */
1234 	if (tc_chosen)
1235 		goto unlock;
1236 	if (tc->tc_quality < 0)
1237 		goto unlock;
1238 	if (tc_from_tunable[0] != '\0' &&
1239 	    strcmp(tc->tc_name, tc_from_tunable) == 0) {
1240 		tc_chosen = 1;
1241 		tc_from_tunable[0] = '\0';
1242 	} else {
1243 		if (tc->tc_quality < timecounter->tc_quality)
1244 			goto unlock;
1245 		if (tc->tc_quality == timecounter->tc_quality &&
1246 		    tc->tc_frequency < timecounter->tc_frequency)
1247 			goto unlock;
1248 	}
1249 	(void)tc->tc_get_timecount(tc);
1250 	timecounter = tc;
1251 unlock:
1252 	mtx_unlock(&tc_lock);
1253 }
1254 
1255 /* Report the frequency of the current timecounter. */
1256 uint64_t
1257 tc_getfrequency(void)
1258 {
1259 
1260 	return (timehands->th_counter->tc_frequency);
1261 }
1262 
1263 static bool
1264 sleeping_on_old_rtc(struct thread *td)
1265 {
1266 
1267 	/*
1268 	 * td_rtcgen is modified by curthread when it is running,
1269 	 * and by other threads in this function.  By finding the thread
1270 	 * on a sleepqueue and holding the lock on the sleepqueue
1271 	 * chain, we guarantee that the thread is not running and that
1272 	 * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
1273 	 * the thread that it was woken due to a real-time clock adjustment.
1274 	 * (The declaration of td_rtcgen refers to this comment.)
1275 	 */
1276 	if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1277 		td->td_rtcgen = 0;
1278 		return (true);
1279 	}
1280 	return (false);
1281 }
1282 
1283 static struct mtx tc_setclock_mtx;
1284 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1285 
1286 /*
1287  * Step our concept of UTC.  This is done by modifying our estimate of
1288  * when we booted.
1289  */
1290 void
1291 tc_setclock(struct timespec *ts)
1292 {
1293 	struct timespec tbef, taft;
1294 	struct bintime bt, bt2;
1295 
1296 	timespec2bintime(ts, &bt);
1297 	nanotime(&tbef);
1298 	mtx_lock_spin(&tc_setclock_mtx);
1299 	cpu_tick_calibrate(1);
1300 	binuptime(&bt2);
1301 	bintime_sub(&bt, &bt2);
1302 
1303 	/* XXX fiddle all the little crinkly bits around the fiords... */
1304 	tc_windup(&bt);
1305 	mtx_unlock_spin(&tc_setclock_mtx);
1306 
1307 	/* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1308 	atomic_add_rel_int(&rtc_generation, 2);
1309 	timerfd_jumped();
1310 	sleepq_chains_remove_matching(sleeping_on_old_rtc);
1311 	if (timestepwarnings) {
1312 		nanotime(&taft);
1313 		log(LOG_INFO,
1314 		    "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1315 		    (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1316 		    (intmax_t)taft.tv_sec, taft.tv_nsec,
1317 		    (intmax_t)ts->tv_sec, ts->tv_nsec);
1318 	}
1319 }
1320 
1321 /*
1322  * Recalculate the scaling factor.  We want the number of 1/2^64
1323  * fractions of a second per period of the hardware counter, taking
1324  * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1325  * processing provides us with.
1326  *
1327  * The th_adjustment is nanoseconds per second with 32 bit binary
1328  * fraction and we want 64 bit binary fraction of second:
1329  *
1330  *	 x = a * 2^32 / 10^9 = a * 4.294967296
1331  *
1332  * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1333  * we can only multiply by about 850 without overflowing, that
1334  * leaves no suitably precise fractions for multiply before divide.
1335  *
1336  * Divide before multiply with a fraction of 2199/512 results in a
1337  * systematic undercompensation of 10PPM of th_adjustment.  On a
1338  * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
1339  *
1340  * We happily sacrifice the lowest of the 64 bits of our result
1341  * to the goddess of code clarity.
1342  */
1343 static void
1344 recalculate_scaling_factor_and_large_delta(struct timehands *th)
1345 {
1346 	uint64_t scale;
1347 
1348 	scale = (uint64_t)1 << 63;
1349 	scale += (th->th_adjustment / 1024) * 2199;
1350 	scale /= th->th_counter->tc_frequency;
1351 	th->th_scale = scale * 2;
1352 	th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1353 }
1354 
1355 /*
1356  * Initialize the next struct timehands in the ring and make
1357  * it the active timehands.  Along the way we might switch to a different
1358  * timecounter and/or do seconds processing in NTP.  Slightly magic.
1359  */
1360 static void
1361 tc_windup(struct bintime *new_boottimebin)
1362 {
1363 	struct bintime bt;
1364 	struct timecounter *tc;
1365 	struct timehands *th, *tho;
1366 	u_int delta, ncount, ogen;
1367 	int i;
1368 	time_t t;
1369 
1370 	/*
1371 	 * Make the next timehands a copy of the current one, but do
1372 	 * not overwrite the generation or next pointer.  While we
1373 	 * update the contents, the generation must be zero.  We need
1374 	 * to ensure that the zero generation is visible before the
1375 	 * data updates become visible, which requires release fence.
1376 	 * For similar reasons, re-reading of the generation after the
1377 	 * data is read should use acquire fence.
1378 	 */
1379 	tho = timehands;
1380 	th = tho->th_next;
1381 	ogen = th->th_generation;
1382 	th->th_generation = 0;
1383 	atomic_thread_fence_rel();
1384 	memcpy(th, tho, offsetof(struct timehands, th_generation));
1385 	if (new_boottimebin != NULL)
1386 		th->th_boottime = *new_boottimebin;
1387 
1388 	/*
1389 	 * Capture a timecounter delta on the current timecounter and if
1390 	 * changing timecounters, a counter value from the new timecounter.
1391 	 * Update the offset fields accordingly.
1392 	 */
1393 	tc = atomic_load_ptr(&timecounter);
1394 	delta = tc_delta(th);
1395 	if (th->th_counter != tc)
1396 		ncount = tc->tc_get_timecount(tc);
1397 	else
1398 		ncount = 0;
1399 #ifdef FFCLOCK
1400 	ffclock_windup(delta);
1401 #endif
1402 	th->th_offset_count += delta;
1403 	th->th_offset_count &= th->th_counter->tc_counter_mask;
1404 	bintime_add_tc_delta(&th->th_offset, th->th_scale,
1405 	    th->th_large_delta, delta);
1406 
1407 	/*
1408 	 * Hardware latching timecounters may not generate interrupts on
1409 	 * PPS events, so instead we poll them.  There is a finite risk that
1410 	 * the hardware might capture a count which is later than the one we
1411 	 * got above, and therefore possibly in the next NTP second which might
1412 	 * have a different rate than the current NTP second.  It doesn't
1413 	 * matter in practice.
1414 	 */
1415 	if (tho->th_counter->tc_poll_pps)
1416 		tho->th_counter->tc_poll_pps(tho->th_counter);
1417 
1418 	/*
1419 	 * Deal with NTP second processing.  The loop normally
1420 	 * iterates at most once, but in extreme situations it might
1421 	 * keep NTP sane if timeouts are not run for several seconds.
1422 	 * At boot, the time step can be large when the TOD hardware
1423 	 * has been read, so on really large steps, we call
1424 	 * ntp_update_second only twice.  We need to call it twice in
1425 	 * case we missed a leap second.
1426 	 */
1427 	bt = th->th_offset;
1428 	bintime_add(&bt, &th->th_boottime);
1429 	i = bt.sec - tho->th_microtime.tv_sec;
1430 	if (i > 0) {
1431 		if (i > LARGE_STEP)
1432 			i = 2;
1433 
1434 		do {
1435 			t = bt.sec;
1436 			ntp_update_second(&th->th_adjustment, &bt.sec);
1437 			if (bt.sec != t)
1438 				th->th_boottime.sec += bt.sec - t;
1439 			--i;
1440 		} while (i > 0);
1441 
1442 		recalculate_scaling_factor_and_large_delta(th);
1443 	}
1444 
1445 	/* Update the UTC timestamps used by the get*() functions. */
1446 	th->th_bintime = bt;
1447 	bintime2timeval(&bt, &th->th_microtime);
1448 	bintime2timespec(&bt, &th->th_nanotime);
1449 
1450 	/* Now is a good time to change timecounters. */
1451 	if (th->th_counter != tc) {
1452 #ifndef __arm__
1453 		if ((tc->tc_flags & TC_FLAGS_C2STOP) != 0)
1454 			cpu_disable_c2_sleep++;
1455 		if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1456 			cpu_disable_c2_sleep--;
1457 #endif
1458 		th->th_counter = tc;
1459 		th->th_offset_count = ncount;
1460 		tc_min_ticktock_freq = max(1, tc->tc_frequency /
1461 		    (((uint64_t)tc->tc_counter_mask + 1) / 3));
1462 		recalculate_scaling_factor_and_large_delta(th);
1463 #ifdef FFCLOCK
1464 		ffclock_change_tc(th);
1465 #endif
1466 	}
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 	mtx_lock(&tc_lock);
1505 	tc = timecounter;
1506 	strlcpy(newname, tc->tc_name, sizeof(newname));
1507 	mtx_unlock(&tc_lock);
1508 
1509 	error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1510 	if (error != 0 || req->newptr == NULL)
1511 		return (error);
1512 
1513 	mtx_lock(&tc_lock);
1514 	/* Record that the tc in use now was specifically chosen. */
1515 	tc_chosen = 1;
1516 	if (strcmp(newname, tc->tc_name) == 0) {
1517 		mtx_unlock(&tc_lock);
1518 		return (0);
1519 	}
1520 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1521 		if (strcmp(newname, newtc->tc_name) != 0)
1522 			continue;
1523 
1524 		/* Warm up new timecounter. */
1525 		(void)newtc->tc_get_timecount(newtc);
1526 
1527 		timecounter = newtc;
1528 
1529 		/*
1530 		 * The vdso timehands update is deferred until the next
1531 		 * 'tc_windup()'.
1532 		 *
1533 		 * This is prudent given that 'timekeep_push_vdso()' does not
1534 		 * use any locking and that it can be called in hard interrupt
1535 		 * context via 'tc_windup()'.
1536 		 */
1537 		break;
1538 	}
1539 	mtx_unlock(&tc_lock);
1540 	return (newtc != NULL ? 0 : EINVAL);
1541 }
1542 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1543     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1544     sysctl_kern_timecounter_hardware, "A",
1545     "Timecounter hardware selected");
1546 
1547 /* Report the available timecounter hardware. */
1548 static int
1549 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1550 {
1551 	struct sbuf sb;
1552 	struct timecounter *tc;
1553 	int error;
1554 
1555 	error = sysctl_wire_old_buffer(req, 0);
1556 	if (error != 0)
1557 		return (error);
1558 	sbuf_new_for_sysctl(&sb, NULL, 0, req);
1559 	mtx_lock(&tc_lock);
1560 	for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1561 		if (tc != timecounters)
1562 			sbuf_putc(&sb, ' ');
1563 		sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1564 	}
1565 	mtx_unlock(&tc_lock);
1566 	error = sbuf_finish(&sb);
1567 	sbuf_delete(&sb);
1568 	return (error);
1569 }
1570 
1571 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
1572     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
1573     sysctl_kern_timecounter_choice, "A",
1574     "Timecounter hardware detected");
1575 
1576 /*
1577  * RFC 2783 PPS-API implementation.
1578  */
1579 
1580 /*
1581  *  Return true if the driver is aware of the abi version extensions in the
1582  *  pps_state structure, and it supports at least the given abi version number.
1583  */
1584 static inline int
1585 abi_aware(struct pps_state *pps, int vers)
1586 {
1587 
1588 	return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1589 }
1590 
1591 static int
1592 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1593 {
1594 	int err, timo;
1595 	pps_seq_t aseq, cseq;
1596 	struct timeval tv;
1597 
1598 	if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1599 		return (EINVAL);
1600 
1601 	/*
1602 	 * If no timeout is requested, immediately return whatever values were
1603 	 * most recently captured.  If timeout seconds is -1, that's a request
1604 	 * to block without a timeout.  WITNESS won't let us sleep forever
1605 	 * without a lock (we really don't need a lock), so just repeatedly
1606 	 * sleep a long time.
1607 	 */
1608 	if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1609 		if (fapi->timeout.tv_sec == -1)
1610 			timo = 0x7fffffff;
1611 		else {
1612 			tv.tv_sec = fapi->timeout.tv_sec;
1613 			tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1614 			timo = tvtohz(&tv);
1615 		}
1616 		aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1617 		cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1618 		while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1619 		    cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1620 			if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1621 				if (pps->flags & PPSFLAG_MTX_SPIN) {
1622 					err = msleep_spin(pps, pps->driver_mtx,
1623 					    "ppsfch", timo);
1624 				} else {
1625 					err = msleep(pps, pps->driver_mtx, PCATCH,
1626 					    "ppsfch", timo);
1627 				}
1628 			} else {
1629 				err = tsleep(pps, PCATCH, "ppsfch", timo);
1630 			}
1631 			if (err == EWOULDBLOCK) {
1632 				if (fapi->timeout.tv_sec == -1) {
1633 					continue;
1634 				} else {
1635 					return (ETIMEDOUT);
1636 				}
1637 			} else if (err != 0) {
1638 				return (err);
1639 			}
1640 		}
1641 	}
1642 
1643 	pps->ppsinfo.current_mode = pps->ppsparam.mode;
1644 	fapi->pps_info_buf = pps->ppsinfo;
1645 
1646 	return (0);
1647 }
1648 
1649 int
1650 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1651 {
1652 	pps_params_t *app;
1653 	struct pps_fetch_args *fapi;
1654 #ifdef FFCLOCK
1655 	struct pps_fetch_ffc_args *fapi_ffc;
1656 #endif
1657 #ifdef PPS_SYNC
1658 	struct pps_kcbind_args *kapi;
1659 #endif
1660 
1661 	KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1662 	switch (cmd) {
1663 	case PPS_IOC_CREATE:
1664 		return (0);
1665 	case PPS_IOC_DESTROY:
1666 		return (0);
1667 	case PPS_IOC_SETPARAMS:
1668 		app = (pps_params_t *)data;
1669 		if (app->mode & ~pps->ppscap)
1670 			return (EINVAL);
1671 #ifdef FFCLOCK
1672 		/* Ensure only a single clock is selected for ffc timestamp. */
1673 		if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1674 			return (EINVAL);
1675 #endif
1676 		pps->ppsparam = *app;
1677 		return (0);
1678 	case PPS_IOC_GETPARAMS:
1679 		app = (pps_params_t *)data;
1680 		*app = pps->ppsparam;
1681 		app->api_version = PPS_API_VERS_1;
1682 		return (0);
1683 	case PPS_IOC_GETCAP:
1684 		*(int*)data = pps->ppscap;
1685 		return (0);
1686 	case PPS_IOC_FETCH:
1687 		fapi = (struct pps_fetch_args *)data;
1688 		return (pps_fetch(fapi, pps));
1689 #ifdef FFCLOCK
1690 	case PPS_IOC_FETCH_FFCOUNTER:
1691 		fapi_ffc = (struct pps_fetch_ffc_args *)data;
1692 		if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1693 		    PPS_TSFMT_TSPEC)
1694 			return (EINVAL);
1695 		if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1696 			return (EOPNOTSUPP);
1697 		pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1698 		fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1699 		/* Overwrite timestamps if feedback clock selected. */
1700 		switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1701 		case PPS_TSCLK_FBCK:
1702 			fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1703 			    pps->ppsinfo.assert_timestamp;
1704 			fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1705 			    pps->ppsinfo.clear_timestamp;
1706 			break;
1707 		case PPS_TSCLK_FFWD:
1708 			break;
1709 		default:
1710 			break;
1711 		}
1712 		return (0);
1713 #endif /* FFCLOCK */
1714 	case PPS_IOC_KCBIND:
1715 #ifdef PPS_SYNC
1716 		kapi = (struct pps_kcbind_args *)data;
1717 		/* XXX Only root should be able to do this */
1718 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1719 			return (EINVAL);
1720 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1721 			return (EINVAL);
1722 		if (kapi->edge & ~pps->ppscap)
1723 			return (EINVAL);
1724 		pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1725 		    (pps->kcmode & KCMODE_ABIFLAG);
1726 		return (0);
1727 #else
1728 		return (EOPNOTSUPP);
1729 #endif
1730 	default:
1731 		return (ENOIOCTL);
1732 	}
1733 }
1734 
1735 void
1736 pps_init(struct pps_state *pps)
1737 {
1738 	pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1739 	if (pps->ppscap & PPS_CAPTUREASSERT)
1740 		pps->ppscap |= PPS_OFFSETASSERT;
1741 	if (pps->ppscap & PPS_CAPTURECLEAR)
1742 		pps->ppscap |= PPS_OFFSETCLEAR;
1743 #ifdef FFCLOCK
1744 	pps->ppscap |= PPS_TSCLK_MASK;
1745 #endif
1746 	pps->kcmode &= ~KCMODE_ABIFLAG;
1747 }
1748 
1749 void
1750 pps_init_abi(struct pps_state *pps)
1751 {
1752 
1753 	pps_init(pps);
1754 	if (pps->driver_abi > 0) {
1755 		pps->kcmode |= KCMODE_ABIFLAG;
1756 		pps->kernel_abi = PPS_ABI_VERSION;
1757 	}
1758 }
1759 
1760 void
1761 pps_capture(struct pps_state *pps)
1762 {
1763 	struct timehands *th;
1764 	struct timecounter *tc;
1765 
1766 	KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1767 	th = timehands;
1768 	pps->capgen = atomic_load_acq_int(&th->th_generation);
1769 	pps->capth = th;
1770 #ifdef FFCLOCK
1771 	pps->capffth = fftimehands;
1772 #endif
1773 	tc = th->th_counter;
1774 	pps->capcount = tc->tc_get_timecount(tc);
1775 }
1776 
1777 void
1778 pps_event(struct pps_state *pps, int event)
1779 {
1780 	struct timehands *capth;
1781 	struct timecounter *captc;
1782 	uint64_t capth_scale;
1783 	struct bintime bt;
1784 	struct timespec *tsp, *osp;
1785 	u_int tcount, *pcount;
1786 	int foff;
1787 	pps_seq_t *pseq;
1788 #ifdef FFCLOCK
1789 	struct timespec *tsp_ffc;
1790 	pps_seq_t *pseq_ffc;
1791 	ffcounter *ffcount;
1792 #endif
1793 #ifdef PPS_SYNC
1794 	int fhard;
1795 #endif
1796 
1797 	KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1798 	/* Nothing to do if not currently set to capture this event type. */
1799 	if ((event & pps->ppsparam.mode) == 0)
1800 		return;
1801 
1802 	/* Make a snapshot of the captured timehand */
1803 	capth = pps->capth;
1804 	captc = capth->th_counter;
1805 	capth_scale = capth->th_scale;
1806 	tcount = capth->th_offset_count;
1807 	bt = capth->th_bintime;
1808 
1809 	/* If the timecounter was wound up underneath us, bail out. */
1810 	atomic_thread_fence_acq();
1811 	if (pps->capgen == 0 || pps->capgen != capth->th_generation)
1812 		return;
1813 
1814 	/* Things would be easier with arrays. */
1815 	if (event == PPS_CAPTUREASSERT) {
1816 		tsp = &pps->ppsinfo.assert_timestamp;
1817 		osp = &pps->ppsparam.assert_offset;
1818 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1819 #ifdef PPS_SYNC
1820 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1821 #endif
1822 		pcount = &pps->ppscount[0];
1823 		pseq = &pps->ppsinfo.assert_sequence;
1824 #ifdef FFCLOCK
1825 		ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1826 		tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1827 		pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1828 #endif
1829 	} else {
1830 		tsp = &pps->ppsinfo.clear_timestamp;
1831 		osp = &pps->ppsparam.clear_offset;
1832 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1833 #ifdef PPS_SYNC
1834 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1835 #endif
1836 		pcount = &pps->ppscount[1];
1837 		pseq = &pps->ppsinfo.clear_sequence;
1838 #ifdef FFCLOCK
1839 		ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1840 		tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1841 		pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1842 #endif
1843 	}
1844 
1845 	*pcount = pps->capcount;
1846 
1847 	/*
1848 	 * If the timecounter changed, we cannot compare the count values, so
1849 	 * we have to drop the rest of the PPS-stuff until the next event.
1850 	 */
1851 	if (__predict_false(pps->ppstc != captc)) {
1852 		pps->ppstc = captc;
1853 		pps->ppscount[2] = pps->capcount;
1854 		return;
1855 	}
1856 
1857 	(*pseq)++;
1858 
1859 	/* Convert the count to a timespec. */
1860 	tcount = pps->capcount - tcount;
1861 	tcount &= captc->tc_counter_mask;
1862 	bintime_addx(&bt, capth_scale * tcount);
1863 	bintime2timespec(&bt, tsp);
1864 
1865 	if (foff) {
1866 		timespecadd(tsp, osp, tsp);
1867 		if (tsp->tv_nsec < 0) {
1868 			tsp->tv_nsec += 1000000000;
1869 			tsp->tv_sec -= 1;
1870 		}
1871 	}
1872 
1873 #ifdef FFCLOCK
1874 	*ffcount = pps->capffth->tick_ffcount + tcount;
1875 	bt = pps->capffth->tick_time;
1876 	ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1877 	bintime_add(&bt, &pps->capffth->tick_time);
1878 	(*pseq_ffc)++;
1879 	bintime2timespec(&bt, tsp_ffc);
1880 #endif
1881 
1882 #ifdef PPS_SYNC
1883 	if (fhard) {
1884 		uint64_t delta_nsec;
1885 		uint64_t freq;
1886 
1887 		/*
1888 		 * Feed the NTP PLL/FLL.
1889 		 * The FLL wants to know how many (hardware) nanoseconds
1890 		 * elapsed since the previous event.
1891 		 */
1892 		tcount = pps->capcount - pps->ppscount[2];
1893 		pps->ppscount[2] = pps->capcount;
1894 		tcount &= captc->tc_counter_mask;
1895 		delta_nsec = 1000000000;
1896 		delta_nsec *= tcount;
1897 		freq = captc->tc_frequency;
1898 		delta_nsec = (delta_nsec + freq / 2) / freq;
1899 		hardpps(tsp, (long)delta_nsec);
1900 	}
1901 #endif
1902 
1903 	/* Wakeup anyone sleeping in pps_fetch().  */
1904 	wakeup(pps);
1905 }
1906 
1907 /*
1908  * Timecounters need to be updated every so often to prevent the hardware
1909  * counter from overflowing.  Updating also recalculates the cached values
1910  * used by the get*() family of functions, so their precision depends on
1911  * the update frequency.
1912  */
1913 
1914 static int tc_tick;
1915 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1916     "Approximate number of hardclock ticks in a millisecond");
1917 
1918 void
1919 tc_ticktock(int cnt)
1920 {
1921 	static int count;
1922 
1923 	if (mtx_trylock_spin(&tc_setclock_mtx)) {
1924 		count += cnt;
1925 		if (count >= tc_tick) {
1926 			count = 0;
1927 			tc_windup(NULL);
1928 		}
1929 		mtx_unlock_spin(&tc_setclock_mtx);
1930 	}
1931 }
1932 
1933 static void __inline
1934 tc_adjprecision(void)
1935 {
1936 	int t;
1937 
1938 	if (tc_timepercentage > 0) {
1939 		t = (99 + tc_timepercentage) / tc_timepercentage;
1940 		tc_precexp = fls(t + (t >> 1)) - 1;
1941 		FREQ2BT(hz / tc_tick, &bt_timethreshold);
1942 		FREQ2BT(hz, &bt_tickthreshold);
1943 		bintime_shift(&bt_timethreshold, tc_precexp);
1944 		bintime_shift(&bt_tickthreshold, tc_precexp);
1945 	} else {
1946 		tc_precexp = 31;
1947 		bt_timethreshold.sec = INT_MAX;
1948 		bt_timethreshold.frac = ~(uint64_t)0;
1949 		bt_tickthreshold = bt_timethreshold;
1950 	}
1951 	sbt_timethreshold = bttosbt(bt_timethreshold);
1952 	sbt_tickthreshold = bttosbt(bt_tickthreshold);
1953 }
1954 
1955 static int
1956 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1957 {
1958 	int error, val;
1959 
1960 	val = tc_timepercentage;
1961 	error = sysctl_handle_int(oidp, &val, 0, req);
1962 	if (error != 0 || req->newptr == NULL)
1963 		return (error);
1964 	tc_timepercentage = val;
1965 	if (cold)
1966 		goto done;
1967 	tc_adjprecision();
1968 done:
1969 	return (0);
1970 }
1971 
1972 /* Set up the requested number of timehands. */
1973 static void
1974 inittimehands(void *dummy)
1975 {
1976 	struct timehands *thp;
1977 	int i;
1978 
1979 	TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1980 	    &timehands_count);
1981 	if (timehands_count < 1)
1982 		timehands_count = 1;
1983 	if (timehands_count > nitems(ths))
1984 		timehands_count = nitems(ths);
1985 	for (i = 1, thp = &ths[0]; i < timehands_count;  thp = &ths[i++])
1986 		thp->th_next = &ths[i];
1987 	thp->th_next = &ths[0];
1988 
1989 	TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1990 	    sizeof(tc_from_tunable));
1991 
1992 	mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
1993 }
1994 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1995 
1996 static void
1997 inittimecounter(void *dummy)
1998 {
1999 	u_int p;
2000 	int tick_rate;
2001 
2002 	/*
2003 	 * Set the initial timeout to
2004 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
2005 	 * People should probably not use the sysctl to set the timeout
2006 	 * to smaller than its initial value, since that value is the
2007 	 * smallest reasonable one.  If they want better timestamps they
2008 	 * should use the non-"get"* functions.
2009 	 */
2010 	if (hz > 1000)
2011 		tc_tick = (hz + 500) / 1000;
2012 	else
2013 		tc_tick = 1;
2014 	tc_adjprecision();
2015 	FREQ2BT(hz, &tick_bt);
2016 	tick_sbt = bttosbt(tick_bt);
2017 	tick_rate = hz / tc_tick;
2018 	FREQ2BT(tick_rate, &tc_tick_bt);
2019 	tc_tick_sbt = bttosbt(tc_tick_bt);
2020 	p = (tc_tick * 1000000) / hz;
2021 	printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
2022 
2023 #ifdef FFCLOCK
2024 	ffclock_init();
2025 #endif
2026 
2027 	/* warm up new timecounter (again) and get rolling. */
2028 	(void)timecounter->tc_get_timecount(timecounter);
2029 	mtx_lock_spin(&tc_setclock_mtx);
2030 	tc_windup(NULL);
2031 	mtx_unlock_spin(&tc_setclock_mtx);
2032 }
2033 
2034 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2035 
2036 /* Cpu tick handling -------------------------------------------------*/
2037 
2038 static bool cpu_tick_variable;
2039 static uint64_t	cpu_tick_frequency;
2040 
2041 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2042 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2043 
2044 static uint64_t
2045 tc_cpu_ticks(void)
2046 {
2047 	struct timecounter *tc;
2048 	uint64_t res, *base;
2049 	unsigned u, *last;
2050 
2051 	critical_enter();
2052 	base = DPCPU_PTR(tc_cpu_ticks_base);
2053 	last = DPCPU_PTR(tc_cpu_ticks_last);
2054 	tc = timehands->th_counter;
2055 	u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2056 	if (u < *last)
2057 		*base += (uint64_t)tc->tc_counter_mask + 1;
2058 	*last = u;
2059 	res = u + *base;
2060 	critical_exit();
2061 	return (res);
2062 }
2063 
2064 void
2065 cpu_tick_calibration(void)
2066 {
2067 	static time_t last_calib;
2068 
2069 	if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2070 		cpu_tick_calibrate(0);
2071 		last_calib = time_uptime;
2072 	}
2073 }
2074 
2075 /*
2076  * This function gets called every 16 seconds on only one designated
2077  * CPU in the system from hardclock() via cpu_tick_calibration()().
2078  *
2079  * Whenever the real time clock is stepped we get called with reset=1
2080  * to make sure we handle suspend/resume and similar events correctly.
2081  */
2082 
2083 static void
2084 cpu_tick_calibrate(int reset)
2085 {
2086 	static uint64_t c_last;
2087 	uint64_t c_this, c_delta;
2088 	static struct bintime  t_last;
2089 	struct bintime t_this, t_delta;
2090 	uint32_t divi;
2091 
2092 	if (reset) {
2093 		/* The clock was stepped, abort & reset */
2094 		t_last.sec = 0;
2095 		return;
2096 	}
2097 
2098 	/* we don't calibrate fixed rate cputicks */
2099 	if (!cpu_tick_variable)
2100 		return;
2101 
2102 	getbinuptime(&t_this);
2103 	c_this = cpu_ticks();
2104 	if (t_last.sec != 0) {
2105 		c_delta = c_this - c_last;
2106 		t_delta = t_this;
2107 		bintime_sub(&t_delta, &t_last);
2108 		/*
2109 		 * Headroom:
2110 		 * 	2^(64-20) / 16[s] =
2111 		 * 	2^(44) / 16[s] =
2112 		 * 	17.592.186.044.416 / 16 =
2113 		 * 	1.099.511.627.776 [Hz]
2114 		 */
2115 		divi = t_delta.sec << 20;
2116 		divi |= t_delta.frac >> (64 - 20);
2117 		c_delta <<= 20;
2118 		c_delta /= divi;
2119 		if (c_delta > cpu_tick_frequency) {
2120 			if (0 && bootverbose)
2121 				printf("cpu_tick increased to %ju Hz\n",
2122 				    c_delta);
2123 			cpu_tick_frequency = c_delta;
2124 		}
2125 	}
2126 	c_last = c_this;
2127 	t_last = t_this;
2128 }
2129 
2130 void
2131 set_cputicker(cpu_tick_f *func, uint64_t freq, bool isvariable)
2132 {
2133 
2134 	if (func == NULL) {
2135 		cpu_ticks = tc_cpu_ticks;
2136 	} else {
2137 		cpu_tick_frequency = freq;
2138 		cpu_tick_variable = isvariable;
2139 		cpu_ticks = func;
2140 	}
2141 }
2142 
2143 uint64_t
2144 cpu_tickrate(void)
2145 {
2146 
2147 	if (cpu_ticks == tc_cpu_ticks)
2148 		return (tc_getfrequency());
2149 	return (cpu_tick_frequency);
2150 }
2151 
2152 /*
2153  * We need to be slightly careful converting cputicks to microseconds.
2154  * There is plenty of margin in 64 bits of microseconds (half a million
2155  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2156  * before divide conversion (to retain precision) we find that the
2157  * margin shrinks to 1.5 hours (one millionth of 146y).
2158  */
2159 
2160 uint64_t
2161 cputick2usec(uint64_t tick)
2162 {
2163 	uint64_t tr;
2164 	tr = cpu_tickrate();
2165 	return ((tick / tr) * 1000000ULL) + ((tick % tr) * 1000000ULL) / tr;
2166 }
2167 
2168 cpu_tick_f	*cpu_ticks = tc_cpu_ticks;
2169 
2170 static int vdso_th_enable = 1;
2171 static int
2172 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2173 {
2174 	int old_vdso_th_enable, error;
2175 
2176 	old_vdso_th_enable = vdso_th_enable;
2177 	error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2178 	if (error != 0)
2179 		return (error);
2180 	vdso_th_enable = old_vdso_th_enable;
2181 	return (0);
2182 }
2183 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2184     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2185     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2186 
2187 uint32_t
2188 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2189 {
2190 	struct timehands *th;
2191 	uint32_t enabled;
2192 
2193 	th = timehands;
2194 	vdso_th->th_scale = th->th_scale;
2195 	vdso_th->th_offset_count = th->th_offset_count;
2196 	vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2197 	vdso_th->th_offset = th->th_offset;
2198 	vdso_th->th_boottime = th->th_boottime;
2199 	if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2200 		enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2201 		    th->th_counter);
2202 	} else
2203 		enabled = 0;
2204 	if (!vdso_th_enable)
2205 		enabled = 0;
2206 	return (enabled);
2207 }
2208 
2209 #ifdef COMPAT_FREEBSD32
2210 uint32_t
2211 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2212 {
2213 	struct timehands *th;
2214 	uint32_t enabled;
2215 
2216 	th = timehands;
2217 	*(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2218 	vdso_th32->th_offset_count = th->th_offset_count;
2219 	vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2220 	vdso_th32->th_offset.sec = th->th_offset.sec;
2221 	*(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2222 	vdso_th32->th_boottime.sec = th->th_boottime.sec;
2223 	*(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2224 	if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2225 		enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2226 		    th->th_counter);
2227 	} else
2228 		enabled = 0;
2229 	if (!vdso_th_enable)
2230 		enabled = 0;
2231 	return (enabled);
2232 }
2233 #endif
2234 
2235 #include "opt_ddb.h"
2236 #ifdef DDB
2237 #include <ddb/ddb.h>
2238 
2239 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2240 {
2241 	struct timehands *th;
2242 	struct timecounter *tc;
2243 	u_int val1, val2;
2244 
2245 	th = timehands;
2246 	tc = th->th_counter;
2247 	val1 = tc->tc_get_timecount(tc);
2248 	__compiler_membar();
2249 	val2 = tc->tc_get_timecount(tc);
2250 
2251 	db_printf("timecounter %p %s\n", tc, tc->tc_name);
2252 	db_printf("  mask %#x freq %ju qual %d flags %#x priv %p\n",
2253 	    tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2254 	    tc->tc_flags, tc->tc_priv);
2255 	db_printf("  val %#x %#x\n", val1, val2);
2256 	db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2257 	    (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2258 	    th->th_large_delta, th->th_offset_count, th->th_generation);
2259 	db_printf("  offset %jd %jd boottime %jd %jd\n",
2260 	    (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2261 	    (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);
2262 }
2263 #endif
2264