xref: /dflybsd-src/sys/kern/kern_clock.c (revision bd4539cc23771f3c0b3fae4ecf80e725b613b305)
1 /*-
2  * Copyright (c) 2004, Matthew Dillon <dillon@backplane.com>
3  * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org>
4  * Copyright (c) 1982, 1986, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
41  * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $
42  * $DragonFly: src/sys/kern/kern_clock.c,v 1.17 2004/03/20 19:21:08 dillon Exp $
43  */
44 
45 #include "opt_ntp.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/dkstat.h>
50 #include <sys/callout.h>
51 #include <sys/kernel.h>
52 #include <sys/proc.h>
53 #include <sys/malloc.h>
54 #include <sys/resourcevar.h>
55 #include <sys/signalvar.h>
56 #include <sys/timex.h>
57 #include <sys/timepps.h>
58 #include <vm/vm.h>
59 #include <sys/lock.h>
60 #include <vm/pmap.h>
61 #include <vm/vm_map.h>
62 #include <sys/sysctl.h>
63 #include <sys/thread2.h>
64 
65 #include <machine/cpu.h>
66 #include <machine/limits.h>
67 #include <machine/smp.h>
68 
69 #ifdef GPROF
70 #include <sys/gmon.h>
71 #endif
72 
73 #ifdef DEVICE_POLLING
74 extern void init_device_poll(void);
75 extern void hardclock_device_poll(void);
76 #endif /* DEVICE_POLLING */
77 
78 static void initclocks (void *dummy);
79 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
80 
81 /*
82  * Some of these don't belong here, but it's easiest to concentrate them.
83  * Note that cp_time[] counts in microseconds, but most userland programs
84  * just compare relative times against the total by delta.
85  */
86 long cp_time[CPUSTATES];
87 
88 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time),
89     "LU", "CPU time statistics");
90 
91 long tk_cancc;
92 long tk_nin;
93 long tk_nout;
94 long tk_rawcc;
95 
96 /*
97  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
98  * microuptime().  microtime() is not drift compensated.  The real uptime
99  * with compensation is nanotime() - bootime.
100  *
101  * basetime is used to calculate the compensated real time of day.  Chunky
102  * changes to the time, aka settimeofday(), are made by modifying basetime.
103  *
104  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
105  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
106  * the real time.
107  */
108 struct timespec boottime;	/* boot time (realtime) for reference only */
109 struct timespec basetime;	/* base time adjusts uptime -> realtime */
110 time_t time_second;		/* read-only 'passive' uptime in seconds */
111 
112 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
113     &boottime, timeval, "System boottime");
114 SYSCTL_STRUCT(_kern, OID_AUTO, basetime, CTLFLAG_RD,
115     &basetime, timeval, "System basetime");
116 
117 static void hardclock(systimer_t info, struct intrframe *frame);
118 static void statclock(systimer_t info, struct intrframe *frame);
119 static void schedclock(systimer_t info, struct intrframe *frame);
120 
121 int	ticks;			/* system master ticks at hz */
122 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
123 int64_t	nsec_acc;		/* accumulator */
124 
125 /*
126  * Finish initializing clock frequencies and start all clocks running.
127  */
128 /* ARGSUSED*/
129 static void
130 initclocks(void *dummy)
131 {
132 	cpu_initclocks();
133 #ifdef DEVICE_POLLING
134 	init_device_poll();
135 #endif
136 	/*psratio = profhz / stathz;*/
137 	initclocks_pcpu();
138 }
139 
140 /*
141  * Called on a per-cpu basis
142  */
143 void
144 initclocks_pcpu(void)
145 {
146 	struct globaldata *gd = mycpu;
147 
148 	crit_enter();
149 	if (gd->gd_cpuid == 0) {
150 	    gd->gd_time_seconds = 1;
151 	    gd->gd_cpuclock_base = cputimer_count();
152 	} else {
153 	    /* XXX */
154 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
155 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
156 	}
157 	systimer_init_periodic(&gd->gd_hardclock, hardclock, NULL, hz);
158 	systimer_init_periodic(&gd->gd_statclock, statclock, NULL, stathz);
159 	/* XXX correct the frequency for scheduler / estcpu tests */
160 	systimer_init_periodic(&gd->gd_schedclock, schedclock,
161 				NULL, ESTCPUFREQ);
162 	crit_exit();
163 }
164 
165 /*
166  * This sets the current real time of day.  Timespecs are in seconds and
167  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
168  * instead we adjust basetime so basetime + gd_* results in the current
169  * time of day.  This way the gd_* fields are guarenteed to represent
170  * a monotonically increasing 'uptime' value.
171  */
172 void
173 set_timeofday(struct timespec *ts)
174 {
175 	struct timespec ts2;
176 
177 	/*
178 	 * XXX SMP / non-atomic basetime updates
179 	 */
180 	crit_enter();
181 	nanouptime(&ts2);
182 	basetime.tv_sec = ts->tv_sec - ts2.tv_sec;
183 	basetime.tv_nsec = ts->tv_nsec - ts2.tv_nsec;
184 	if (basetime.tv_nsec < 0) {
185 	    basetime.tv_nsec += 1000000000;
186 	    --basetime.tv_sec;
187 	}
188 	if (boottime.tv_sec == 0)
189 		boottime = basetime;
190 	timedelta = 0;
191 	crit_exit();
192 }
193 
194 /*
195  * Each cpu has its own hardclock, but we only increments ticks and softticks
196  * on cpu #0.
197  *
198  * NOTE! systimer! the MP lock might not be held here.  We can only safely
199  * manipulate objects owned by the current cpu.
200  */
201 static void
202 hardclock(systimer_t info, struct intrframe *frame)
203 {
204 	sysclock_t cputicks;
205 	struct proc *p;
206 	struct pstats *pstats;
207 	struct globaldata *gd = mycpu;
208 
209 	/*
210 	 * Realtime updates are per-cpu.  Note that timer corrections as
211 	 * returned by microtime() and friends make an additional adjustment
212 	 * using a system-wise 'basetime', but the running time is always
213 	 * taken from the per-cpu globaldata area.  Since the same clock
214 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
215 	 * stay in synch.
216 	 *
217 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
218 	 * to reverse index gd_cpuclock_base.
219 	 */
220 	cputicks = info->time - gd->gd_cpuclock_base;
221 	if (cputicks > cputimer_freq) {
222 		++gd->gd_time_seconds;
223 		gd->gd_cpuclock_base += cputimer_freq;
224 	}
225 
226 	/*
227 	 * The system-wide ticks and softticks are only updated by cpu #0.
228 	 * Callwheel actions are also (at the moment) only handled by cpu #0.
229 	 * Finally, we also do NTP related timedelta/tickdelta adjustments
230 	 * by adjusting basetime.
231 	 */
232 	if (gd->gd_cpuid == 0) {
233 	    struct timespec nts;
234 	    int leap;
235 
236 	    ++ticks;
237 
238 #ifdef DEVICE_POLLING
239 	    hardclock_device_poll();	/* mpsafe, short and quick */
240 #endif /* DEVICE_POLLING */
241 
242 	    if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) {
243 		setsoftclock();
244 	    } else if (softticks + 1 == ticks) {
245 		++softticks;
246 	    }
247 
248 #if 0
249 	    if (tco->tc_poll_pps)
250 		tco->tc_poll_pps(tco);
251 #endif
252 	    /*
253 	     * Apply adjtime corrections.  At the moment only do this if
254 	     * we can get the MP lock to interlock with adjtime's modification
255 	     * of these variables.  Note that basetime adjustments are not
256 	     * MP safe either XXX.
257 	     */
258 	    if (timedelta != 0 && try_mplock()) {
259 		basetime.tv_nsec += tickdelta * 1000;
260 		if (basetime.tv_nsec >= 1000000000) {
261 		    basetime.tv_nsec -= 1000000000;
262 		    ++basetime.tv_sec;
263 		} else if (basetime.tv_nsec < 0) {
264 		    basetime.tv_nsec += 1000000000;
265 		    --basetime.tv_sec;
266 		}
267 		timedelta -= tickdelta;
268 		rel_mplock();
269 	    }
270 
271 	    /*
272 	     * Apply per-tick compensation.  ticks_adj adjusts for both
273 	     * offset and frequency, and could be negative.
274 	     */
275 	    if (nsec_adj != 0 && try_mplock()) {
276 		nsec_acc += nsec_adj;
277 		if (nsec_acc >= 0x100000000LL) {
278 		    basetime.tv_nsec += nsec_acc >> 32;
279 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
280 		} else if (nsec_acc <= -0x100000000LL) {
281 		    basetime.tv_nsec -= -nsec_acc >> 32;
282 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
283 		}
284 		if (basetime.tv_nsec >= 1000000000) {
285 		    basetime.tv_nsec -= 1000000000;
286 		    ++basetime.tv_sec;
287 		} else if (basetime.tv_nsec < 0) {
288 		    basetime.tv_nsec += 1000000000;
289 		    --basetime.tv_sec;
290 		}
291 		rel_mplock();
292 	    }
293 
294 	    /*
295 	     * If the realtime-adjusted seconds hand rolls over then tell
296 	     * ntp_update_second() what we did in the last second so it can
297 	     * calculate what to do in the next second.  It may also add
298 	     * or subtract a leap second.
299 	     */
300 	    getnanotime(&nts);
301 	    if (time_second != nts.tv_sec) {
302 		leap = ntp_update_second(time_second, &nsec_adj);
303 		basetime.tv_sec += leap;
304 		time_second = nts.tv_sec + leap;
305 		nsec_adj /= hz;
306 	    }
307 	}
308 
309 	/*
310 	 * ITimer handling is per-tick, per-cpu.  I don't think psignal()
311 	 * is mpsafe on curproc, so XXX get the mplock.
312 	 */
313 	if ((p = curproc) != NULL && try_mplock()) {
314 		pstats = p->p_stats;
315 		if (frame && CLKF_USERMODE(frame) &&
316 		    timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
317 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
318 			psignal(p, SIGVTALRM);
319 		if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
320 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
321 			psignal(p, SIGPROF);
322 		rel_mplock();
323 	}
324 	setdelayed();
325 }
326 
327 /*
328  * The statistics clock typically runs at a 125Hz rate, and is intended
329  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
330  *
331  * NOTE! systimer! the MP lock might not be held here.  We can only safely
332  * manipulate objects owned by the current cpu.
333  *
334  * The stats clock is responsible for grabbing a profiling sample.
335  * Most of the statistics are only used by user-level statistics programs.
336  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
337  * p->p_estcpu.
338  *
339  * Like the other clocks, the stat clock is called from what is effectively
340  * a fast interrupt, so the context should be the thread/process that got
341  * interrupted.
342  */
343 static void
344 statclock(systimer_t info, struct intrframe *frame)
345 {
346 #ifdef GPROF
347 	struct gmonparam *g;
348 	int i;
349 #endif
350 	thread_t td;
351 	struct proc *p;
352 	int bump;
353 	struct timeval tv;
354 	struct timeval *stv;
355 
356 	/*
357 	 * How big was our timeslice relative to the last time?
358 	 */
359 	microuptime(&tv);	/* mpsafe */
360 	stv = &mycpu->gd_stattv;
361 	if (stv->tv_sec == 0) {
362 	    bump = 1;
363 	} else {
364 	    bump = tv.tv_usec - stv->tv_usec +
365 		(tv.tv_sec - stv->tv_sec) * 1000000;
366 	    if (bump < 0)
367 		bump = 0;
368 	    if (bump > 1000000)
369 		bump = 1000000;
370 	}
371 	*stv = tv;
372 
373 	td = curthread;
374 	p = td->td_proc;
375 
376 	if (frame && CLKF_USERMODE(frame)) {
377 		/*
378 		 * Came from userland, handle user time and deal with
379 		 * possible process.
380 		 */
381 		if (p && (p->p_flag & P_PROFIL))
382 			addupc_intr(p, CLKF_PC(frame), 1);
383 		td->td_uticks += bump;
384 
385 		/*
386 		 * Charge the time as appropriate
387 		 */
388 		if (p && p->p_nice > NZERO)
389 			cp_time[CP_NICE] += bump;
390 		else
391 			cp_time[CP_USER] += bump;
392 	} else {
393 #ifdef GPROF
394 		/*
395 		 * Kernel statistics are just like addupc_intr, only easier.
396 		 */
397 		g = &_gmonparam;
398 		if (g->state == GMON_PROF_ON && frame) {
399 			i = CLKF_PC(frame) - g->lowpc;
400 			if (i < g->textsize) {
401 				i /= HISTFRACTION * sizeof(*g->kcount);
402 				g->kcount[i]++;
403 			}
404 		}
405 #endif
406 		/*
407 		 * Came from kernel mode, so we were:
408 		 * - handling an interrupt,
409 		 * - doing syscall or trap work on behalf of the current
410 		 *   user process, or
411 		 * - spinning in the idle loop.
412 		 * Whichever it is, charge the time as appropriate.
413 		 * Note that we charge interrupts to the current process,
414 		 * regardless of whether they are ``for'' that process,
415 		 * so that we know how much of its real time was spent
416 		 * in ``non-process'' (i.e., interrupt) work.
417 		 *
418 		 * XXX assume system if frame is NULL.  A NULL frame
419 		 * can occur if ipi processing is done from an splx().
420 		 */
421 		if (frame && CLKF_INTR(frame))
422 			td->td_iticks += bump;
423 		else
424 			td->td_sticks += bump;
425 
426 		if (frame && CLKF_INTR(frame)) {
427 			cp_time[CP_INTR] += bump;
428 		} else {
429 			if (td == &mycpu->gd_idlethread)
430 				cp_time[CP_IDLE] += bump;
431 			else
432 				cp_time[CP_SYS] += bump;
433 		}
434 	}
435 }
436 
437 /*
438  * The scheduler clock typically runs at a 10Hz rate.  NOTE! systimer,
439  * the MP lock might not be held.  We can safely manipulate parts of curproc
440  * but that's about it.
441  */
442 static void
443 schedclock(systimer_t info, struct intrframe *frame)
444 {
445 	struct proc *p;
446 	struct pstats *pstats;
447 	struct rusage *ru;
448 	struct vmspace *vm;
449 	long rss;
450 
451 	schedulerclock(NULL);	/* mpsafe */
452 	if ((p = curproc) != NULL) {
453 		/* Update resource usage integrals and maximums. */
454 		if ((pstats = p->p_stats) != NULL &&
455 		    (ru = &pstats->p_ru) != NULL &&
456 		    (vm = p->p_vmspace) != NULL) {
457 			ru->ru_ixrss += pgtok(vm->vm_tsize);
458 			ru->ru_idrss += pgtok(vm->vm_dsize);
459 			ru->ru_isrss += pgtok(vm->vm_ssize);
460 			rss = pgtok(vmspace_resident_count(vm));
461 			if (ru->ru_maxrss < rss)
462 				ru->ru_maxrss = rss;
463 		}
464 	}
465 }
466 
467 /*
468  * Compute number of ticks for the specified amount of time.  The
469  * return value is intended to be used in a clock interrupt timed
470  * operation and guarenteed to meet or exceed the requested time.
471  * If the representation overflows, return INT_MAX.  The minimum return
472  * value is 1 ticks and the function will average the calculation up.
473  * If any value greater then 0 microseconds is supplied, a value
474  * of at least 2 will be returned to ensure that a near-term clock
475  * interrupt does not cause the timeout to occur (degenerately) early.
476  *
477  * Note that limit checks must take into account microseconds, which is
478  * done simply by using the smaller signed long maximum instead of
479  * the unsigned long maximum.
480  *
481  * If ints have 32 bits, then the maximum value for any timeout in
482  * 10ms ticks is 248 days.
483  */
484 int
485 tvtohz_high(struct timeval *tv)
486 {
487 	int ticks;
488 	long sec, usec;
489 
490 	sec = tv->tv_sec;
491 	usec = tv->tv_usec;
492 	if (usec < 0) {
493 		sec--;
494 		usec += 1000000;
495 	}
496 	if (sec < 0) {
497 #ifdef DIAGNOSTIC
498 		if (usec > 0) {
499 			sec++;
500 			usec -= 1000000;
501 		}
502 		printf("tvotohz: negative time difference %ld sec %ld usec\n",
503 		       sec, usec);
504 #endif
505 		ticks = 1;
506 	} else if (sec <= INT_MAX / hz) {
507 		ticks = (int)(sec * hz +
508 			    ((u_long)usec + (tick - 1)) / tick) + 1;
509 	} else {
510 		ticks = INT_MAX;
511 	}
512 	return (ticks);
513 }
514 
515 /*
516  * Compute number of ticks for the specified amount of time, erroring on
517  * the side of it being too low to ensure that sleeping the returned number
518  * of ticks will not result in a late return.
519  *
520  * The supplied timeval may not be negative and should be normalized.  A
521  * return value of 0 is possible if the timeval converts to less then
522  * 1 tick.
523  *
524  * If ints have 32 bits, then the maximum value for any timeout in
525  * 10ms ticks is 248 days.
526  */
527 int
528 tvtohz_low(struct timeval *tv)
529 {
530 	int ticks;
531 	long sec;
532 
533 	sec = tv->tv_sec;
534 	if (sec <= INT_MAX / hz)
535 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
536 	else
537 		ticks = INT_MAX;
538 	return (ticks);
539 }
540 
541 
542 /*
543  * Start profiling on a process.
544  *
545  * Kernel profiling passes proc0 which never exits and hence
546  * keeps the profile clock running constantly.
547  */
548 void
549 startprofclock(struct proc *p)
550 {
551 	if ((p->p_flag & P_PROFIL) == 0) {
552 		p->p_flag |= P_PROFIL;
553 #if 0	/* XXX */
554 		if (++profprocs == 1 && stathz != 0) {
555 			s = splstatclock();
556 			psdiv = psratio;
557 			setstatclockrate(profhz);
558 			splx(s);
559 		}
560 #endif
561 	}
562 }
563 
564 /*
565  * Stop profiling on a process.
566  */
567 void
568 stopprofclock(struct proc *p)
569 {
570 	if (p->p_flag & P_PROFIL) {
571 		p->p_flag &= ~P_PROFIL;
572 #if 0	/* XXX */
573 		if (--profprocs == 0 && stathz != 0) {
574 			s = splstatclock();
575 			psdiv = 1;
576 			setstatclockrate(stathz);
577 			splx(s);
578 		}
579 #endif
580 	}
581 }
582 
583 /*
584  * Return information about system clocks.
585  */
586 static int
587 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
588 {
589 	struct clockinfo clkinfo;
590 	/*
591 	 * Construct clockinfo structure.
592 	 */
593 	clkinfo.hz = hz;
594 	clkinfo.tick = tick;
595 	clkinfo.tickadj = tickadj;
596 	clkinfo.profhz = profhz;
597 	clkinfo.stathz = stathz ? stathz : hz;
598 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
599 }
600 
601 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
602 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
603 
604 /*
605  * We have eight functions for looking at the clock, four for
606  * microseconds and four for nanoseconds.  For each there is fast
607  * but less precise version "get{nano|micro}[up]time" which will
608  * return a time which is up to 1/HZ previous to the call, whereas
609  * the raw version "{nano|micro}[up]time" will return a timestamp
610  * which is as precise as possible.  The "up" variants return the
611  * time relative to system boot, these are well suited for time
612  * interval measurements.
613  *
614  * Each cpu independantly maintains the current time of day, so all
615  * we need to do to protect ourselves from changes is to do a loop
616  * check on the seconds field changing out from under us.
617  */
618 void
619 getmicrouptime(struct timeval *tvp)
620 {
621 	struct globaldata *gd = mycpu;
622 	sysclock_t delta;
623 
624 	do {
625 		tvp->tv_sec = gd->gd_time_seconds;
626 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
627 	} while (tvp->tv_sec != gd->gd_time_seconds);
628 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
629 	if (tvp->tv_usec >= 1000000) {
630 		tvp->tv_usec -= 1000000;
631 		++tvp->tv_sec;
632 	}
633 }
634 
635 void
636 getnanouptime(struct timespec *tsp)
637 {
638 	struct globaldata *gd = mycpu;
639 	sysclock_t delta;
640 
641 	do {
642 		tsp->tv_sec = gd->gd_time_seconds;
643 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
644 	} while (tsp->tv_sec != gd->gd_time_seconds);
645 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
646 	if (tsp->tv_nsec >= 1000000000) {
647 		tsp->tv_nsec -= 1000000000;
648 		++tsp->tv_sec;
649 	}
650 }
651 
652 void
653 microuptime(struct timeval *tvp)
654 {
655 	struct globaldata *gd = mycpu;
656 	sysclock_t delta;
657 
658 	do {
659 		tvp->tv_sec = gd->gd_time_seconds;
660 		delta = cputimer_count() - gd->gd_cpuclock_base;
661 	} while (tvp->tv_sec != gd->gd_time_seconds);
662 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
663 	if (tvp->tv_usec >= 1000000) {
664 		tvp->tv_usec -= 1000000;
665 		++tvp->tv_sec;
666 	}
667 }
668 
669 void
670 nanouptime(struct timespec *tsp)
671 {
672 	struct globaldata *gd = mycpu;
673 	sysclock_t delta;
674 
675 	do {
676 		tsp->tv_sec = gd->gd_time_seconds;
677 		delta = cputimer_count() - gd->gd_cpuclock_base;
678 	} while (tsp->tv_sec != gd->gd_time_seconds);
679 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
680 	if (tsp->tv_nsec >= 1000000000) {
681 		tsp->tv_nsec -= 1000000000;
682 		++tsp->tv_sec;
683 	}
684 }
685 
686 /*
687  * realtime routines
688  */
689 
690 void
691 getmicrotime(struct timeval *tvp)
692 {
693 	struct globaldata *gd = mycpu;
694 	sysclock_t delta;
695 
696 	do {
697 		tvp->tv_sec = gd->gd_time_seconds;
698 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
699 	} while (tvp->tv_sec != gd->gd_time_seconds);
700 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
701 
702 	tvp->tv_sec += basetime.tv_sec;
703 	tvp->tv_usec += basetime.tv_nsec / 1000;
704 	while (tvp->tv_usec >= 1000000) {
705 		tvp->tv_usec -= 1000000;
706 		++tvp->tv_sec;
707 	}
708 }
709 
710 void
711 getnanotime(struct timespec *tsp)
712 {
713 	struct globaldata *gd = mycpu;
714 	sysclock_t delta;
715 
716 	do {
717 		tsp->tv_sec = gd->gd_time_seconds;
718 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
719 	} while (tsp->tv_sec != gd->gd_time_seconds);
720 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
721 
722 	tsp->tv_sec += basetime.tv_sec;
723 	tsp->tv_nsec += basetime.tv_nsec;
724 	while (tsp->tv_nsec >= 1000000000) {
725 		tsp->tv_nsec -= 1000000000;
726 		++tsp->tv_sec;
727 	}
728 }
729 
730 void
731 microtime(struct timeval *tvp)
732 {
733 	struct globaldata *gd = mycpu;
734 	sysclock_t delta;
735 
736 	do {
737 		tvp->tv_sec = gd->gd_time_seconds;
738 		delta = cputimer_count() - gd->gd_cpuclock_base;
739 	} while (tvp->tv_sec != gd->gd_time_seconds);
740 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
741 
742 	tvp->tv_sec += basetime.tv_sec;
743 	tvp->tv_usec += basetime.tv_nsec / 1000;
744 	while (tvp->tv_usec >= 1000000) {
745 		tvp->tv_usec -= 1000000;
746 		++tvp->tv_sec;
747 	}
748 }
749 
750 void
751 nanotime(struct timespec *tsp)
752 {
753 	struct globaldata *gd = mycpu;
754 	sysclock_t delta;
755 
756 	do {
757 		tsp->tv_sec = gd->gd_time_seconds;
758 		delta = cputimer_count() - gd->gd_cpuclock_base;
759 	} while (tsp->tv_sec != gd->gd_time_seconds);
760 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
761 
762 	tsp->tv_sec += basetime.tv_sec;
763 	tsp->tv_nsec += basetime.tv_nsec;
764 	while (tsp->tv_nsec >= 1000000000) {
765 		tsp->tv_nsec -= 1000000000;
766 		++tsp->tv_sec;
767 	}
768 }
769 
770 int
771 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
772 {
773 	pps_params_t *app;
774 	struct pps_fetch_args *fapi;
775 #ifdef PPS_SYNC
776 	struct pps_kcbind_args *kapi;
777 #endif
778 
779 	switch (cmd) {
780 	case PPS_IOC_CREATE:
781 		return (0);
782 	case PPS_IOC_DESTROY:
783 		return (0);
784 	case PPS_IOC_SETPARAMS:
785 		app = (pps_params_t *)data;
786 		if (app->mode & ~pps->ppscap)
787 			return (EINVAL);
788 		pps->ppsparam = *app;
789 		return (0);
790 	case PPS_IOC_GETPARAMS:
791 		app = (pps_params_t *)data;
792 		*app = pps->ppsparam;
793 		app->api_version = PPS_API_VERS_1;
794 		return (0);
795 	case PPS_IOC_GETCAP:
796 		*(int*)data = pps->ppscap;
797 		return (0);
798 	case PPS_IOC_FETCH:
799 		fapi = (struct pps_fetch_args *)data;
800 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
801 			return (EINVAL);
802 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
803 			return (EOPNOTSUPP);
804 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
805 		fapi->pps_info_buf = pps->ppsinfo;
806 		return (0);
807 	case PPS_IOC_KCBIND:
808 #ifdef PPS_SYNC
809 		kapi = (struct pps_kcbind_args *)data;
810 		/* XXX Only root should be able to do this */
811 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
812 			return (EINVAL);
813 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
814 			return (EINVAL);
815 		if (kapi->edge & ~pps->ppscap)
816 			return (EINVAL);
817 		pps->kcmode = kapi->edge;
818 		return (0);
819 #else
820 		return (EOPNOTSUPP);
821 #endif
822 	default:
823 		return (ENOTTY);
824 	}
825 }
826 
827 void
828 pps_init(struct pps_state *pps)
829 {
830 	pps->ppscap |= PPS_TSFMT_TSPEC;
831 	if (pps->ppscap & PPS_CAPTUREASSERT)
832 		pps->ppscap |= PPS_OFFSETASSERT;
833 	if (pps->ppscap & PPS_CAPTURECLEAR)
834 		pps->ppscap |= PPS_OFFSETCLEAR;
835 }
836 
837 void
838 pps_event(struct pps_state *pps, sysclock_t count, int event)
839 {
840 	struct globaldata *gd;
841 	struct timespec *tsp;
842 	struct timespec *osp;
843 	struct timespec ts;
844 	sysclock_t *pcount;
845 #ifdef PPS_SYNC
846 	sysclock_t tcount;
847 #endif
848 	sysclock_t delta;
849 	pps_seq_t *pseq;
850 	int foff;
851 	int fhard;
852 
853 	gd = mycpu;
854 
855 	/* Things would be easier with arrays... */
856 	if (event == PPS_CAPTUREASSERT) {
857 		tsp = &pps->ppsinfo.assert_timestamp;
858 		osp = &pps->ppsparam.assert_offset;
859 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
860 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
861 		pcount = &pps->ppscount[0];
862 		pseq = &pps->ppsinfo.assert_sequence;
863 	} else {
864 		tsp = &pps->ppsinfo.clear_timestamp;
865 		osp = &pps->ppsparam.clear_offset;
866 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
867 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
868 		pcount = &pps->ppscount[1];
869 		pseq = &pps->ppsinfo.clear_sequence;
870 	}
871 
872 	/* Nothing really happened */
873 	if (*pcount == count)
874 		return;
875 
876 	*pcount = count;
877 
878 	do {
879 		ts.tv_sec = gd->gd_time_seconds;
880 		delta = count - gd->gd_cpuclock_base;
881 	} while (ts.tv_sec != gd->gd_time_seconds);
882 	if (delta > cputimer_freq) {
883 		ts.tv_sec += delta / cputimer_freq;
884 		delta %= cputimer_freq;
885 	}
886 	ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
887 	ts.tv_sec += basetime.tv_sec;
888 	ts.tv_nsec += basetime.tv_nsec;
889 	while (ts.tv_nsec >= 1000000000) {
890 		ts.tv_nsec -= 1000000000;
891 		++ts.tv_sec;
892 	}
893 
894 	(*pseq)++;
895 	*tsp = ts;
896 
897 	if (foff) {
898 		timespecadd(tsp, osp);
899 		if (tsp->tv_nsec < 0) {
900 			tsp->tv_nsec += 1000000000;
901 			tsp->tv_sec -= 1;
902 		}
903 	}
904 #ifdef PPS_SYNC
905 	if (fhard) {
906 		/* magic, at its best... */
907 		tcount = count - pps->ppscount[2];
908 		pps->ppscount[2] = count;
909 		delta = (cputimer_freq64_nsec * tcount) >> 32;
910 		hardpps(tsp, delta);
911 	}
912 #endif
913 }
914 
915