xref: /freebsd/sys/kern/kern_clock.c (revision 2ad872c5794e4c26fdf6ed219ad3f09ca0d5304a)
1 /*-
2  * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org>
3  * Copyright (c) 1982, 1986, 1991, 1993
4  *	The Regents of the University of California.  All rights reserved.
5  * (c) UNIX System Laboratories, Inc.
6  * All or some portions of this file are derived from material licensed
7  * to the University of California by American Telephone and Telegraph
8  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
9  * the permission of UNIX System Laboratories, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. All advertising materials mentioning features or use of this software
20  *    must display the following acknowledgement:
21  *	This product includes software developed by the University of
22  *	California, Berkeley and its contributors.
23  * 4. Neither the name of the University nor the names of its contributors
24  *    may be used to endorse or promote products derived from this software
25  *    without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37  * SUCH DAMAGE.
38  *
39  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
40  * $Id: kern_clock.c,v 1.85 1998/11/23 09:58:53 phk Exp $
41  */
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/dkstat.h>
46 #include <sys/callout.h>
47 #include <sys/kernel.h>
48 #include <sys/proc.h>
49 #include <sys/malloc.h>
50 #include <sys/resourcevar.h>
51 #include <sys/signalvar.h>
52 #include <sys/timex.h>
53 #include <vm/vm.h>
54 #include <sys/lock.h>
55 #include <vm/pmap.h>
56 #include <vm/vm_map.h>
57 #include <sys/sysctl.h>
58 
59 #include <machine/cpu.h>
60 #include <machine/limits.h>
61 
62 #ifdef GPROF
63 #include <sys/gmon.h>
64 #endif
65 
66 #if defined(SMP) && defined(BETTER_CLOCK)
67 #include <machine/smp.h>
68 #endif
69 
70 /* This is where the NTIMECOUNTER option hangs out */
71 #include "opt_ntp.h"
72 
73 /*
74  * Number of timecounters used to implement stable storage
75  */
76 #ifndef NTIMECOUNTER
77 #define NTIMECOUNTER	5
78 #endif
79 
80 static MALLOC_DEFINE(M_TIMECOUNTER, "timecounter",
81 	"Timecounter stable storage");
82 
83 static void initclocks __P((void *dummy));
84 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
85 
86 static void tco_forward __P((int force));
87 static void tco_setscales __P((struct timecounter *tc));
88 static __inline unsigned tco_delta __P((struct timecounter *tc));
89 
90 /* Some of these don't belong here, but it's easiest to concentrate them. */
91 #if defined(SMP) && defined(BETTER_CLOCK)
92 long cp_time[CPUSTATES];
93 #else
94 static long cp_time[CPUSTATES];
95 #endif
96 
97 long tk_cancc;
98 long tk_nin;
99 long tk_nout;
100 long tk_rawcc;
101 
102 time_t time_second;
103 
104 /*
105  * Which update policy to use.
106  *   0 - every tick, bad hardware may fail with "calcru negative..."
107  *   1 - more resistent to the above hardware, but less efficient.
108  */
109 static int tco_method;
110 
111 /*
112  * Implement a dummy timecounter which we can use until we get a real one
113  * in the air.  This allows the console and other early stuff to use
114  * timeservices.
115  */
116 
117 static unsigned
118 dummy_get_timecount(struct timecounter *tc)
119 {
120 	static unsigned now;
121 	return (++now);
122 }
123 
124 static struct timecounter dummy_timecounter = {
125 	dummy_get_timecount,
126 	0,
127 	~0u,
128 	1000000,
129 	"dummy"
130 };
131 
132 struct timecounter *timecounter = &dummy_timecounter;
133 
134 /*
135  * Clock handling routines.
136  *
137  * This code is written to operate with two timers that run independently of
138  * each other.
139  *
140  * The main timer, running hz times per second, is used to trigger interval
141  * timers, timeouts and rescheduling as needed.
142  *
143  * The second timer handles kernel and user profiling,
144  * and does resource use estimation.  If the second timer is programmable,
145  * it is randomized to avoid aliasing between the two clocks.  For example,
146  * the randomization prevents an adversary from always giving up the cpu
147  * just before its quantum expires.  Otherwise, it would never accumulate
148  * cpu ticks.  The mean frequency of the second timer is stathz.
149  *
150  * If no second timer exists, stathz will be zero; in this case we drive
151  * profiling and statistics off the main clock.  This WILL NOT be accurate;
152  * do not do it unless absolutely necessary.
153  *
154  * The statistics clock may (or may not) be run at a higher rate while
155  * profiling.  This profile clock runs at profhz.  We require that profhz
156  * be an integral multiple of stathz.
157  *
158  * If the statistics clock is running fast, it must be divided by the ratio
159  * profhz/stathz for statistics.  (For profiling, every tick counts.)
160  *
161  * Time-of-day is maintained using a "timecounter", which may or may
162  * not be related to the hardware generating the above mentioned
163  * interrupts.
164  */
165 
166 int	stathz;
167 int	profhz;
168 static int profprocs;
169 int	ticks;
170 static int psdiv, pscnt;		/* prof => stat divider */
171 int	psratio;			/* ratio: prof / stat */
172 
173 /*
174  * Initialize clock frequencies and start both clocks running.
175  */
176 /* ARGSUSED*/
177 static void
178 initclocks(dummy)
179 	void *dummy;
180 {
181 	register int i;
182 
183 	/*
184 	 * Set divisors to 1 (normal case) and let the machine-specific
185 	 * code do its bit.
186 	 */
187 	psdiv = pscnt = 1;
188 	cpu_initclocks();
189 
190 	/*
191 	 * Compute profhz/stathz, and fix profhz if needed.
192 	 */
193 	i = stathz ? stathz : hz;
194 	if (profhz == 0)
195 		profhz = i;
196 	psratio = profhz / i;
197 }
198 
199 /*
200  * The real-time timer, interrupting hz times per second.
201  */
202 void
203 hardclock(frame)
204 	register struct clockframe *frame;
205 {
206 	register struct proc *p;
207 
208 	p = curproc;
209 	if (p) {
210 		register struct pstats *pstats;
211 
212 		/*
213 		 * Run current process's virtual and profile time, as needed.
214 		 */
215 		pstats = p->p_stats;
216 		if (CLKF_USERMODE(frame) &&
217 		    timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
218 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
219 			psignal(p, SIGVTALRM);
220 		if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
221 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
222 			psignal(p, SIGPROF);
223 	}
224 
225 #if defined(SMP) && defined(BETTER_CLOCK)
226 	forward_hardclock(pscnt);
227 #endif
228 
229 	/*
230 	 * If no separate statistics clock is available, run it from here.
231 	 */
232 	if (stathz == 0)
233 		statclock(frame);
234 
235 	tco_forward(0);
236 	ticks++;
237 
238 	/*
239 	 * Process callouts at a very low cpu priority, so we don't keep the
240 	 * relatively high clock interrupt priority any longer than necessary.
241 	 */
242 	if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) {
243 		if (CLKF_BASEPRI(frame)) {
244 			/*
245 			 * Save the overhead of a software interrupt;
246 			 * it will happen as soon as we return, so do it now.
247 			 */
248 			(void)splsoftclock();
249 			softclock();
250 		} else
251 			setsoftclock();
252 	} else if (softticks + 1 == ticks)
253 		++softticks;
254 }
255 
256 /*
257  * Compute number of ticks in the specified amount of time.
258  */
259 int
260 tvtohz(tv)
261 	struct timeval *tv;
262 {
263 	register unsigned long ticks;
264 	register long sec, usec;
265 
266 	/*
267 	 * If the number of usecs in the whole seconds part of the time
268 	 * difference fits in a long, then the total number of usecs will
269 	 * fit in an unsigned long.  Compute the total and convert it to
270 	 * ticks, rounding up and adding 1 to allow for the current tick
271 	 * to expire.  Rounding also depends on unsigned long arithmetic
272 	 * to avoid overflow.
273 	 *
274 	 * Otherwise, if the number of ticks in the whole seconds part of
275 	 * the time difference fits in a long, then convert the parts to
276 	 * ticks separately and add, using similar rounding methods and
277 	 * overflow avoidance.  This method would work in the previous
278 	 * case but it is slightly slower and assumes that hz is integral.
279 	 *
280 	 * Otherwise, round the time difference down to the maximum
281 	 * representable value.
282 	 *
283 	 * If ints have 32 bits, then the maximum value for any timeout in
284 	 * 10ms ticks is 248 days.
285 	 */
286 	sec = tv->tv_sec;
287 	usec = tv->tv_usec;
288 	if (usec < 0) {
289 		sec--;
290 		usec += 1000000;
291 	}
292 	if (sec < 0) {
293 #ifdef DIAGNOSTIC
294 		if (usec > 0) {
295 			sec++;
296 			usec -= 1000000;
297 		}
298 		printf("tvotohz: negative time difference %ld sec %ld usec\n",
299 		       sec, usec);
300 #endif
301 		ticks = 1;
302 	} else if (sec <= LONG_MAX / 1000000)
303 		ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
304 			/ tick + 1;
305 	else if (sec <= LONG_MAX / hz)
306 		ticks = sec * hz
307 			+ ((unsigned long)usec + (tick - 1)) / tick + 1;
308 	else
309 		ticks = LONG_MAX;
310 	if (ticks > INT_MAX)
311 		ticks = INT_MAX;
312 	return ((int)ticks);
313 }
314 
315 /*
316  * Start profiling on a process.
317  *
318  * Kernel profiling passes proc0 which never exits and hence
319  * keeps the profile clock running constantly.
320  */
321 void
322 startprofclock(p)
323 	register struct proc *p;
324 {
325 	int s;
326 
327 	if ((p->p_flag & P_PROFIL) == 0) {
328 		p->p_flag |= P_PROFIL;
329 		if (++profprocs == 1 && stathz != 0) {
330 			s = splstatclock();
331 			psdiv = pscnt = psratio;
332 			setstatclockrate(profhz);
333 			splx(s);
334 		}
335 	}
336 }
337 
338 /*
339  * Stop profiling on a process.
340  */
341 void
342 stopprofclock(p)
343 	register struct proc *p;
344 {
345 	int s;
346 
347 	if (p->p_flag & P_PROFIL) {
348 		p->p_flag &= ~P_PROFIL;
349 		if (--profprocs == 0 && stathz != 0) {
350 			s = splstatclock();
351 			psdiv = pscnt = 1;
352 			setstatclockrate(stathz);
353 			splx(s);
354 		}
355 	}
356 }
357 
358 /*
359  * Statistics clock.  Grab profile sample, and if divider reaches 0,
360  * do process and kernel statistics.
361  */
362 void
363 statclock(frame)
364 	register struct clockframe *frame;
365 {
366 #ifdef GPROF
367 	register struct gmonparam *g;
368 	int i;
369 #endif
370 	register struct proc *p;
371 	struct pstats *pstats;
372 	long rss;
373 	struct rusage *ru;
374 	struct vmspace *vm;
375 
376 	if (curproc != NULL && CLKF_USERMODE(frame)) {
377 		p = curproc;
378 		if (p->p_flag & P_PROFIL)
379 			addupc_intr(p, CLKF_PC(frame), 1);
380 #if defined(SMP) && defined(BETTER_CLOCK)
381 		if (stathz != 0)
382 			forward_statclock(pscnt);
383 #endif
384 		if (--pscnt > 0)
385 			return;
386 		/*
387 		 * Came from user mode; CPU was in user state.
388 		 * If this process is being profiled record the tick.
389 		 */
390 		p->p_uticks++;
391 		if (p->p_nice > NZERO)
392 			cp_time[CP_NICE]++;
393 		else
394 			cp_time[CP_USER]++;
395 	} else {
396 #ifdef GPROF
397 		/*
398 		 * Kernel statistics are just like addupc_intr, only easier.
399 		 */
400 		g = &_gmonparam;
401 		if (g->state == GMON_PROF_ON) {
402 			i = CLKF_PC(frame) - g->lowpc;
403 			if (i < g->textsize) {
404 				i /= HISTFRACTION * sizeof(*g->kcount);
405 				g->kcount[i]++;
406 			}
407 		}
408 #endif
409 #if defined(SMP) && defined(BETTER_CLOCK)
410 		if (stathz != 0)
411 			forward_statclock(pscnt);
412 #endif
413 		if (--pscnt > 0)
414 			return;
415 		/*
416 		 * Came from kernel mode, so we were:
417 		 * - handling an interrupt,
418 		 * - doing syscall or trap work on behalf of the current
419 		 *   user process, or
420 		 * - spinning in the idle loop.
421 		 * Whichever it is, charge the time as appropriate.
422 		 * Note that we charge interrupts to the current process,
423 		 * regardless of whether they are ``for'' that process,
424 		 * so that we know how much of its real time was spent
425 		 * in ``non-process'' (i.e., interrupt) work.
426 		 */
427 		p = curproc;
428 		if (CLKF_INTR(frame)) {
429 			if (p != NULL)
430 				p->p_iticks++;
431 			cp_time[CP_INTR]++;
432 		} else if (p != NULL) {
433 			p->p_sticks++;
434 			cp_time[CP_SYS]++;
435 		} else
436 			cp_time[CP_IDLE]++;
437 	}
438 	pscnt = psdiv;
439 
440 	/*
441 	 * We maintain statistics shown by user-level statistics
442 	 * programs:  the amount of time in each cpu state.
443 	 */
444 
445 	/*
446 	 * We adjust the priority of the current process.  The priority of
447 	 * a process gets worse as it accumulates CPU time.  The cpu usage
448 	 * estimator (p_estcpu) is increased here.  The formula for computing
449 	 * priorities (in kern_synch.c) will compute a different value each
450 	 * time p_estcpu increases by 4.  The cpu usage estimator ramps up
451 	 * quite quickly when the process is running (linearly), and decays
452 	 * away exponentially, at a rate which is proportionally slower when
453 	 * the system is busy.  The basic principal is that the system will
454 	 * 90% forget that the process used a lot of CPU time in 5 * loadav
455 	 * seconds.  This causes the system to favor processes which haven't
456 	 * run much recently, and to round-robin among other processes.
457 	 */
458 	if (p != NULL) {
459 		p->p_cpticks++;
460 		if (++p->p_estcpu == 0)
461 			p->p_estcpu--;
462 		if ((p->p_estcpu & 3) == 0) {
463 			resetpriority(p);
464 			if (p->p_priority >= PUSER)
465 				p->p_priority = p->p_usrpri;
466 		}
467 
468 		/* Update resource usage integrals and maximums. */
469 		if ((pstats = p->p_stats) != NULL &&
470 		    (ru = &pstats->p_ru) != NULL &&
471 		    (vm = p->p_vmspace) != NULL) {
472 			ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024;
473 			ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024;
474 			ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024;
475 			rss = vm->vm_pmap.pm_stats.resident_count *
476 			      PAGE_SIZE / 1024;
477 			if (ru->ru_maxrss < rss)
478 				ru->ru_maxrss = rss;
479         	}
480 	}
481 }
482 
483 /*
484  * Return information about system clocks.
485  */
486 static int
487 sysctl_kern_clockrate SYSCTL_HANDLER_ARGS
488 {
489 	struct clockinfo clkinfo;
490 	/*
491 	 * Construct clockinfo structure.
492 	 */
493 	clkinfo.hz = hz;
494 	clkinfo.tick = tick;
495 	clkinfo.tickadj = tickadj;
496 	clkinfo.profhz = profhz;
497 	clkinfo.stathz = stathz ? stathz : hz;
498 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
499 }
500 
501 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
502 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
503 
504 static __inline unsigned
505 tco_delta(struct timecounter *tc)
506 {
507 
508 	return ((tc->tc_get_timecount(tc) - tc->tc_offset_count) &
509 	    tc->tc_counter_mask);
510 }
511 
512 /*
513  * We have four functions for looking at the clock, two for microseconds
514  * and two for nanoseconds.  For each there is fast but less precise
515  * version "get{nano|micro}time" which will return a time which is up
516  * to 1/HZ previous to the call, whereas the raw version "{nano|micro}time"
517  * will return a timestamp which is as precise as possible.
518  */
519 
520 void
521 getmicrotime(struct timeval *tvp)
522 {
523 	struct timecounter *tc;
524 
525 	if (!tco_method) {
526 		tc = timecounter;
527 		*tvp = tc->tc_microtime;
528 	} else {
529 		microtime(tvp);
530 	}
531 }
532 
533 void
534 getnanotime(struct timespec *tsp)
535 {
536 	struct timecounter *tc;
537 
538 	if (!tco_method) {
539 		tc = timecounter;
540 		*tsp = tc->tc_nanotime;
541 	} else {
542 		nanotime(tsp);
543 	}
544 }
545 
546 void
547 microtime(struct timeval *tv)
548 {
549 	struct timecounter *tc;
550 
551 	tc = (struct timecounter *)timecounter;
552 	tv->tv_sec = tc->tc_offset_sec;
553 	tv->tv_usec = tc->tc_offset_micro;
554 	tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32;
555 	tv->tv_usec += boottime.tv_usec;
556 	tv->tv_sec += boottime.tv_sec;
557 	while (tv->tv_usec >= 1000000) {
558 		tv->tv_usec -= 1000000;
559 		tv->tv_sec++;
560 	}
561 }
562 
563 void
564 nanotime(struct timespec *ts)
565 {
566 	unsigned count;
567 	u_int64_t delta;
568 	struct timecounter *tc;
569 
570 	tc = (struct timecounter *)timecounter;
571 	ts->tv_sec = tc->tc_offset_sec;
572 	count = tco_delta(tc);
573 	delta = tc->tc_offset_nano;
574 	delta += ((u_int64_t)count * tc->tc_scale_nano_f);
575 	delta >>= 32;
576 	delta += ((u_int64_t)count * tc->tc_scale_nano_i);
577 	delta += boottime.tv_usec * 1000;
578 	ts->tv_sec += boottime.tv_sec;
579 	while (delta >= 1000000000) {
580 		delta -= 1000000000;
581 		ts->tv_sec++;
582 	}
583 	ts->tv_nsec = delta;
584 }
585 
586 void
587 timecounter_timespec(unsigned count, struct timespec *ts)
588 {
589 	u_int64_t delta;
590 	struct timecounter *tc;
591 
592 	tc = (struct timecounter *)timecounter;
593 	ts->tv_sec = tc->tc_offset_sec;
594 	count -= tc->tc_offset_count;
595 	count &= tc->tc_counter_mask;
596 	delta = tc->tc_offset_nano;
597 	delta += ((u_int64_t)count * tc->tc_scale_nano_f);
598 	delta >>= 32;
599 	delta += ((u_int64_t)count * tc->tc_scale_nano_i);
600 	delta += boottime.tv_usec * 1000;
601 	ts->tv_sec += boottime.tv_sec;
602 	while (delta >= 1000000000) {
603 		delta -= 1000000000;
604 		ts->tv_sec++;
605 	}
606 	ts->tv_nsec = delta;
607 }
608 
609 void
610 getmicrouptime(struct timeval *tvp)
611 {
612 	struct timecounter *tc;
613 
614 	if (!tco_method) {
615 		tc = timecounter;
616 		tvp->tv_sec = tc->tc_offset_sec;
617 		tvp->tv_usec = tc->tc_offset_micro;
618 	} else {
619 		microuptime(tvp);
620 	}
621 }
622 
623 void
624 getnanouptime(struct timespec *tsp)
625 {
626 	struct timecounter *tc;
627 
628 	if (!tco_method) {
629 		tc = timecounter;
630 		tsp->tv_sec = tc->tc_offset_sec;
631 		tsp->tv_nsec = tc->tc_offset_nano >> 32;
632 	} else {
633 		nanouptime(tsp);
634 	}
635 }
636 
637 void
638 microuptime(struct timeval *tv)
639 {
640 	struct timecounter *tc;
641 
642 	tc = (struct timecounter *)timecounter;
643 	tv->tv_sec = tc->tc_offset_sec;
644 	tv->tv_usec = tc->tc_offset_micro;
645 	tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32;
646 	if (tv->tv_usec >= 1000000) {
647 		tv->tv_usec -= 1000000;
648 		tv->tv_sec++;
649 	}
650 }
651 
652 void
653 nanouptime(struct timespec *ts)
654 {
655 	unsigned count;
656 	u_int64_t delta;
657 	struct timecounter *tc;
658 
659 	tc = (struct timecounter *)timecounter;
660 	ts->tv_sec = tc->tc_offset_sec;
661 	count = tco_delta(tc);
662 	delta = tc->tc_offset_nano;
663 	delta += ((u_int64_t)count * tc->tc_scale_nano_f);
664 	delta >>= 32;
665 	delta += ((u_int64_t)count * tc->tc_scale_nano_i);
666 	if (delta >= 1000000000) {
667 		delta -= 1000000000;
668 		ts->tv_sec++;
669 	}
670 	ts->tv_nsec = delta;
671 }
672 
673 static void
674 tco_setscales(struct timecounter *tc)
675 {
676 	u_int64_t scale;
677 
678 	scale = 1000000000LL << 32;
679 	if (tc->tc_adjustment > 0)
680 		scale += (tc->tc_adjustment * 1000LL) << 10;
681 	else
682 		scale -= (-tc->tc_adjustment * 1000LL) << 10;
683 	scale /= tc->tc_frequency;
684 	tc->tc_scale_micro = scale / 1000;
685 	tc->tc_scale_nano_f = scale & 0xffffffff;
686 	tc->tc_scale_nano_i = scale >> 32;
687 }
688 
689 void
690 init_timecounter(struct timecounter *tc)
691 {
692 	struct timespec ts1;
693 	struct timecounter *t1, *t2, *t3;
694 	int i;
695 
696 	tc->tc_adjustment = 0;
697 	tco_setscales(tc);
698 	tc->tc_offset_count = tc->tc_get_timecount(tc);
699 	tc->tc_tweak = tc;
700 	MALLOC(t1, struct timecounter *, sizeof *t1, M_TIMECOUNTER, M_WAITOK);
701 	*t1 = *tc;
702 	t2 = t1;
703 	for (i = 1; i < NTIMECOUNTER; i++) {
704 		MALLOC(t3, struct timecounter *, sizeof *t3,
705 		    M_TIMECOUNTER, M_WAITOK);
706 		*t3 = *tc;
707 		t3->tc_other = t2;
708 		t2 = t3;
709 	}
710 	t1->tc_other = t3;
711 	tc = t1;
712 
713 	printf("Timecounter \"%s\"  frequency %lu Hz\n",
714 	    tc->tc_name, (u_long)tc->tc_frequency);
715 
716 	/* XXX: For now always start using the counter. */
717 	tc->tc_offset_count = tc->tc_get_timecount(tc);
718 	nanouptime(&ts1);
719 	tc->tc_offset_nano = (u_int64_t)ts1.tv_nsec << 32;
720 	tc->tc_offset_micro = ts1.tv_nsec / 1000;
721 	tc->tc_offset_sec = ts1.tv_sec;
722 	timecounter = tc;
723 }
724 
725 void
726 set_timecounter(struct timespec *ts)
727 {
728 	struct timespec ts2;
729 
730 	nanouptime(&ts2);
731 	boottime.tv_sec = ts->tv_sec - ts2.tv_sec;
732 	boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000;
733 	if (boottime.tv_usec < 0) {
734 		boottime.tv_usec += 1000000;
735 		boottime.tv_sec--;
736 	}
737 	/* fiddle all the little crinkly bits around the fiords... */
738 	tco_forward(1);
739 }
740 
741 
742 #if 0 /* Currently unused */
743 void
744 switch_timecounter(struct timecounter *newtc)
745 {
746 	int s;
747 	struct timecounter *tc;
748 	struct timespec ts;
749 
750 	s = splclock();
751 	tc = timecounter;
752 	if (newtc == tc || newtc == tc->tc_other) {
753 		splx(s);
754 		return;
755 	}
756 	nanouptime(&ts);
757 	newtc->tc_offset_sec = ts.tv_sec;
758 	newtc->tc_offset_nano = (u_int64_t)ts.tv_nsec << 32;
759 	newtc->tc_offset_micro = ts.tv_nsec / 1000;
760 	newtc->tc_offset_count = newtc->tc_get_timecount(newtc);
761 	timecounter = newtc;
762 	splx(s);
763 }
764 #endif
765 
766 static struct timecounter *
767 sync_other_counter(void)
768 {
769 	struct timecounter *tc, *tcn, *tco;
770 	unsigned delta;
771 
772 	tco = timecounter;
773 	tc = tco->tc_other;
774 	tcn = tc->tc_other;
775 	*tc = *tco;
776 	tc->tc_other = tcn;
777 	delta = tco_delta(tc);
778 	tc->tc_offset_count += delta;
779 	tc->tc_offset_count &= tc->tc_counter_mask;
780 	tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_f;
781 	tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_i << 32;
782 	return (tc);
783 }
784 
785 static void
786 tco_forward(int force)
787 {
788 	struct timecounter *tc, *tco;
789 
790 	tco = timecounter;
791 	tc = sync_other_counter();
792 	/*
793 	 * We may be inducing a tiny error here, the tc_poll_pps() may
794 	 * process a latched count which happens after the tco_delta()
795 	 * in sync_other_counter(), which would extend the previous
796 	 * counters parameters into the domain of this new one.
797 	 * Since the timewindow is very small for this, the error is
798 	 * going to be only a few weenieseconds (as Dave Mills would
799 	 * say), so lets just not talk more about it, OK ?
800 	 */
801 	if (tco->tc_poll_pps)
802 		tco->tc_poll_pps(tco);
803 	if (timedelta != 0) {
804 		tc->tc_offset_nano += (u_int64_t)(tickdelta * 1000) << 32;
805 		timedelta -= tickdelta;
806 		force++;
807 	}
808 
809 	while (tc->tc_offset_nano >= 1000000000ULL << 32) {
810 		tc->tc_offset_nano -= 1000000000ULL << 32;
811 		tc->tc_offset_sec++;
812 		tc->tc_frequency = tc->tc_tweak->tc_frequency;
813 		tc->tc_adjustment = tc->tc_tweak->tc_adjustment;
814 		ntp_update_second(tc);	/* XXX only needed if xntpd runs */
815 		tco_setscales(tc);
816 		force++;
817 	}
818 
819 	if (tco_method && !force)
820 		return;
821 
822 	tc->tc_offset_micro = (tc->tc_offset_nano / 1000) >> 32;
823 
824 	/* Figure out the wall-clock time */
825 	tc->tc_nanotime.tv_sec = tc->tc_offset_sec + boottime.tv_sec;
826 	tc->tc_nanotime.tv_nsec =
827 	    (tc->tc_offset_nano >> 32) + boottime.tv_usec * 1000;
828 	tc->tc_microtime.tv_usec = tc->tc_offset_micro + boottime.tv_usec;
829 	if (tc->tc_nanotime.tv_nsec >= 1000000000) {
830 		tc->tc_nanotime.tv_nsec -= 1000000000;
831 		tc->tc_microtime.tv_usec -= 1000000;
832 		tc->tc_nanotime.tv_sec++;
833 	}
834 	time_second = tc->tc_microtime.tv_sec = tc->tc_nanotime.tv_sec;
835 
836 	timecounter = tc;
837 }
838 
839 static int
840 sysctl_kern_timecounter_frequency SYSCTL_HANDLER_ARGS
841 {
842 
843 	return (sysctl_handle_opaque(oidp,
844 	    &timecounter->tc_tweak->tc_frequency,
845 	    sizeof(timecounter->tc_tweak->tc_frequency), req));
846 }
847 
848 static int
849 sysctl_kern_timecounter_adjustment SYSCTL_HANDLER_ARGS
850 {
851 
852 	return (sysctl_handle_opaque(oidp,
853 	    &timecounter->tc_tweak->tc_adjustment,
854 	    sizeof(timecounter->tc_tweak->tc_adjustment), req));
855 }
856 
857 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
858 
859 SYSCTL_INT(_kern_timecounter, KERN_ARGMAX, method, CTLFLAG_RW, &tco_method, 0,
860     "This variable determines the method used for updating timecounters. "
861     "If the default algorithm (0) fails with \"calcru negative...\" messages "
862     "try the alternate algorithm (1) which handles bad hardware better."
863 
864 );
865 
866 SYSCTL_PROC(_kern_timecounter, OID_AUTO, frequency, CTLTYPE_INT | CTLFLAG_RW,
867     0, sizeof(u_int), sysctl_kern_timecounter_frequency, "I", "");
868 
869 SYSCTL_PROC(_kern_timecounter, OID_AUTO, adjustment, CTLTYPE_INT | CTLFLAG_RW,
870     0, sizeof(int), sysctl_kern_timecounter_adjustment, "I", "");
871