xref: /dragonfly/sys/kern/kern_clock.c (revision 611395e5)
1 /*
2  * Copyright (c) 2003,2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org>
35  * Copyright (c) 1982, 1986, 1991, 1993
36  *	The Regents of the University of California.  All rights reserved.
37  * (c) UNIX System Laboratories, Inc.
38  * All or some portions of this file are derived from material licensed
39  * to the University of California by American Telephone and Telegraph
40  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
41  * the permission of UNIX System Laboratories, Inc.
42  *
43  * Redistribution and use in source and binary forms, with or without
44  * modification, are permitted provided that the following conditions
45  * are met:
46  * 1. Redistributions of source code must retain the above copyright
47  *    notice, this list of conditions and the following disclaimer.
48  * 2. Redistributions in binary form must reproduce the above copyright
49  *    notice, this list of conditions and the following disclaimer in the
50  *    documentation and/or other materials provided with the distribution.
51  * 3. All advertising materials mentioning features or use of this software
52  *    must display the following acknowledgement:
53  *	This product includes software developed by the University of
54  *	California, Berkeley and its contributors.
55  * 4. Neither the name of the University nor the names of its contributors
56  *    may be used to endorse or promote products derived from this software
57  *    without specific prior written permission.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  *
71  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
72  * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $
73  * $DragonFly: src/sys/kern/kern_clock.c,v 1.28 2004/12/04 20:38:45 dillon Exp $
74  */
75 
76 #include "opt_ntp.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/dkstat.h>
81 #include <sys/callout.h>
82 #include <sys/kernel.h>
83 #include <sys/proc.h>
84 #include <sys/malloc.h>
85 #include <sys/resourcevar.h>
86 #include <sys/signalvar.h>
87 #include <sys/timex.h>
88 #include <sys/timepps.h>
89 #include <vm/vm.h>
90 #include <sys/lock.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <sys/sysctl.h>
94 #include <sys/thread2.h>
95 
96 #include <machine/cpu.h>
97 #include <machine/limits.h>
98 #include <machine/smp.h>
99 
100 #ifdef GPROF
101 #include <sys/gmon.h>
102 #endif
103 
104 #ifdef DEVICE_POLLING
105 extern void init_device_poll(void);
106 extern void hardclock_device_poll(void);
107 #endif /* DEVICE_POLLING */
108 
109 static void initclocks (void *dummy);
110 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
111 
112 /*
113  * Some of these don't belong here, but it's easiest to concentrate them.
114  * Note that cp_time[] counts in microseconds, but most userland programs
115  * just compare relative times against the total by delta.
116  */
117 long cp_time[CPUSTATES];
118 
119 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time),
120     "LU", "CPU time statistics");
121 
122 long tk_cancc;
123 long tk_nin;
124 long tk_nout;
125 long tk_rawcc;
126 
127 /*
128  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
129  * microuptime().  microtime() is not drift compensated.  The real uptime
130  * with compensation is nanotime() - bootime.  boottime is recalculated
131  * whenever the real time is set based on the compensated elapsed time
132  * in seconds (gd->gd_time_seconds).
133  *
134  * basetime is used to calculate the compensated real time of day.  Chunky
135  * changes to the time, aka settimeofday(), are made by modifying basetime.
136  *
137  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
138  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
139  * the real time.
140  */
141 struct timespec boottime;	/* boot time (realtime) for reference only */
142 struct timespec basetime;	/* base time adjusts uptime -> realtime */
143 time_t time_second;		/* read-only 'passive' uptime in seconds */
144 
145 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
146     &boottime, timeval, "System boottime");
147 SYSCTL_STRUCT(_kern, OID_AUTO, basetime, CTLFLAG_RD,
148     &basetime, timeval, "System basetime");
149 
150 static void hardclock(systimer_t info, struct intrframe *frame);
151 static void statclock(systimer_t info, struct intrframe *frame);
152 static void schedclock(systimer_t info, struct intrframe *frame);
153 
154 int	ticks;			/* system master ticks at hz */
155 int	clocks_running;		/* tsleep/timeout clocks operational */
156 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
157 int64_t	nsec_acc;		/* accumulator */
158 
159 /*
160  * Finish initializing clock frequencies and start all clocks running.
161  */
162 /* ARGSUSED*/
163 static void
164 initclocks(void *dummy)
165 {
166 	cpu_initclocks();
167 #ifdef DEVICE_POLLING
168 	init_device_poll();
169 #endif
170 	/*psratio = profhz / stathz;*/
171 	initclocks_pcpu();
172 	clocks_running = 1;
173 }
174 
175 /*
176  * Called on a per-cpu basis
177  */
178 void
179 initclocks_pcpu(void)
180 {
181 	struct globaldata *gd = mycpu;
182 
183 	crit_enter();
184 	if (gd->gd_cpuid == 0) {
185 	    gd->gd_time_seconds = 1;
186 	    gd->gd_cpuclock_base = cputimer_count();
187 	} else {
188 	    /* XXX */
189 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
190 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
191 	}
192 
193 	/*
194 	 * Use a non-queued periodic systimer to prevent multiple ticks from
195 	 * building up if the sysclock jumps forward (8254 gets reset).  The
196 	 * sysclock will never jump backwards.  Our time sync is based on
197 	 * the actual sysclock, not the ticks count.
198 	 */
199 	systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
200 	systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
201 	/* XXX correct the frequency for scheduler / estcpu tests */
202 	systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
203 				NULL, ESTCPUFREQ);
204 	crit_exit();
205 }
206 
207 /*
208  * This sets the current real time of day.  Timespecs are in seconds and
209  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
210  * instead we adjust basetime so basetime + gd_* results in the current
211  * time of day.  This way the gd_* fields are guarenteed to represent
212  * a monotonically increasing 'uptime' value.
213  */
214 void
215 set_timeofday(struct timespec *ts)
216 {
217 	struct timespec ts2;
218 
219 	/*
220 	 * XXX SMP / non-atomic basetime updates
221 	 */
222 	crit_enter();
223 	nanouptime(&ts2);
224 	basetime.tv_sec = ts->tv_sec - ts2.tv_sec;
225 	basetime.tv_nsec = ts->tv_nsec - ts2.tv_nsec;
226 	if (basetime.tv_nsec < 0) {
227 	    basetime.tv_nsec += 1000000000;
228 	    --basetime.tv_sec;
229 	}
230 
231 	/*
232 	 * Note that basetime diverges from boottime as the clock drift is
233 	 * compensated for, so we cannot do away with boottime.  When setting
234 	 * the absolute time of day the drift is 0 (for an instant) and we
235 	 * can simply assign boottime to basetime.
236 	 *
237 	 * Note that nanouptime() is based on gd_time_seconds which is drift
238 	 * compensated up to a point (it is guarenteed to remain monotonically
239 	 * increasing).  gd_time_seconds is thus our best uptime guess and
240 	 * suitable for use in the boottime calculation.  It is already taken
241 	 * into account in the basetime calculation above.
242 	 */
243 	boottime.tv_sec = basetime.tv_sec;
244 	timedelta = 0;
245 	crit_exit();
246 }
247 
248 /*
249  * Each cpu has its own hardclock, but we only increments ticks and softticks
250  * on cpu #0.
251  *
252  * NOTE! systimer! the MP lock might not be held here.  We can only safely
253  * manipulate objects owned by the current cpu.
254  */
255 static void
256 hardclock(systimer_t info, struct intrframe *frame)
257 {
258 	sysclock_t cputicks;
259 	struct proc *p;
260 	struct pstats *pstats;
261 	struct globaldata *gd = mycpu;
262 
263 	/*
264 	 * Realtime updates are per-cpu.  Note that timer corrections as
265 	 * returned by microtime() and friends make an additional adjustment
266 	 * using a system-wise 'basetime', but the running time is always
267 	 * taken from the per-cpu globaldata area.  Since the same clock
268 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
269 	 * stay in synch.
270 	 *
271 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
272 	 * to reverse index gd_cpuclock_base, but that it is possible for
273 	 * it to temporarily get behind in the seconds if something in the
274 	 * system locks interrupts for a long period of time.  Since periodic
275 	 * timers count events, though everything should resynch again
276 	 * immediately.
277 	 */
278 	cputicks = info->time - gd->gd_cpuclock_base;
279 	if (cputicks >= cputimer_freq) {
280 		++gd->gd_time_seconds;
281 		gd->gd_cpuclock_base += cputimer_freq;
282 	}
283 
284 	/*
285 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
286 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
287 	 * by updating basetime.
288 	 */
289 	if (gd->gd_cpuid == 0) {
290 	    struct timespec nts;
291 	    int leap;
292 
293 	    ++ticks;
294 
295 #ifdef DEVICE_POLLING
296 	    hardclock_device_poll();	/* mpsafe, short and quick */
297 #endif /* DEVICE_POLLING */
298 
299 #if 0
300 	    if (tco->tc_poll_pps)
301 		tco->tc_poll_pps(tco);
302 #endif
303 	    /*
304 	     * Apply adjtime corrections.  At the moment only do this if
305 	     * we can get the MP lock to interlock with adjtime's modification
306 	     * of these variables.  Note that basetime adjustments are not
307 	     * MP safe either XXX.
308 	     */
309 	    if (timedelta != 0 && try_mplock()) {
310 		basetime.tv_nsec += tickdelta * 1000;
311 		if (basetime.tv_nsec >= 1000000000) {
312 		    basetime.tv_nsec -= 1000000000;
313 		    ++basetime.tv_sec;
314 		} else if (basetime.tv_nsec < 0) {
315 		    basetime.tv_nsec += 1000000000;
316 		    --basetime.tv_sec;
317 		}
318 		timedelta -= tickdelta;
319 		rel_mplock();
320 	    }
321 
322 	    /*
323 	     * Apply per-tick compensation.  ticks_adj adjusts for both
324 	     * offset and frequency, and could be negative.
325 	     */
326 	    if (nsec_adj != 0 && try_mplock()) {
327 		nsec_acc += nsec_adj;
328 		if (nsec_acc >= 0x100000000LL) {
329 		    basetime.tv_nsec += nsec_acc >> 32;
330 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
331 		} else if (nsec_acc <= -0x100000000LL) {
332 		    basetime.tv_nsec -= -nsec_acc >> 32;
333 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
334 		}
335 		if (basetime.tv_nsec >= 1000000000) {
336 		    basetime.tv_nsec -= 1000000000;
337 		    ++basetime.tv_sec;
338 		} else if (basetime.tv_nsec < 0) {
339 		    basetime.tv_nsec += 1000000000;
340 		    --basetime.tv_sec;
341 		}
342 		rel_mplock();
343 	    }
344 
345 	    /*
346 	     * If the realtime-adjusted seconds hand rolls over then tell
347 	     * ntp_update_second() what we did in the last second so it can
348 	     * calculate what to do in the next second.  It may also add
349 	     * or subtract a leap second.
350 	     */
351 	    getnanotime(&nts);
352 	    if (time_second != nts.tv_sec) {
353 		leap = ntp_update_second(time_second, &nsec_adj);
354 		basetime.tv_sec += leap;
355 		time_second = nts.tv_sec + leap;
356 		nsec_adj /= hz;
357 	    }
358 	}
359 
360 	/*
361 	 * softticks are handled for all cpus
362 	 */
363 	hardclock_softtick(gd);
364 
365 	/*
366 	 * ITimer handling is per-tick, per-cpu.  I don't think psignal()
367 	 * is mpsafe on curproc, so XXX get the mplock.
368 	 */
369 	if ((p = curproc) != NULL && try_mplock()) {
370 		pstats = p->p_stats;
371 		if (frame && CLKF_USERMODE(frame) &&
372 		    timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
373 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
374 			psignal(p, SIGVTALRM);
375 		if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
376 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
377 			psignal(p, SIGPROF);
378 		rel_mplock();
379 	}
380 	setdelayed();
381 }
382 
383 /*
384  * The statistics clock typically runs at a 125Hz rate, and is intended
385  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
386  *
387  * NOTE! systimer! the MP lock might not be held here.  We can only safely
388  * manipulate objects owned by the current cpu.
389  *
390  * The stats clock is responsible for grabbing a profiling sample.
391  * Most of the statistics are only used by user-level statistics programs.
392  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
393  * p->p_estcpu.
394  *
395  * Like the other clocks, the stat clock is called from what is effectively
396  * a fast interrupt, so the context should be the thread/process that got
397  * interrupted.
398  */
399 static void
400 statclock(systimer_t info, struct intrframe *frame)
401 {
402 #ifdef GPROF
403 	struct gmonparam *g;
404 	int i;
405 #endif
406 	thread_t td;
407 	struct proc *p;
408 	int bump;
409 	struct timeval tv;
410 	struct timeval *stv;
411 
412 	/*
413 	 * How big was our timeslice relative to the last time?
414 	 */
415 	microuptime(&tv);	/* mpsafe */
416 	stv = &mycpu->gd_stattv;
417 	if (stv->tv_sec == 0) {
418 	    bump = 1;
419 	} else {
420 	    bump = tv.tv_usec - stv->tv_usec +
421 		(tv.tv_sec - stv->tv_sec) * 1000000;
422 	    if (bump < 0)
423 		bump = 0;
424 	    if (bump > 1000000)
425 		bump = 1000000;
426 	}
427 	*stv = tv;
428 
429 	td = curthread;
430 	p = td->td_proc;
431 
432 	if (frame && CLKF_USERMODE(frame)) {
433 		/*
434 		 * Came from userland, handle user time and deal with
435 		 * possible process.
436 		 */
437 		if (p && (p->p_flag & P_PROFIL))
438 			addupc_intr(p, CLKF_PC(frame), 1);
439 		td->td_uticks += bump;
440 
441 		/*
442 		 * Charge the time as appropriate
443 		 */
444 		if (p && p->p_nice > NZERO)
445 			cp_time[CP_NICE] += bump;
446 		else
447 			cp_time[CP_USER] += bump;
448 	} else {
449 #ifdef GPROF
450 		/*
451 		 * Kernel statistics are just like addupc_intr, only easier.
452 		 */
453 		g = &_gmonparam;
454 		if (g->state == GMON_PROF_ON && frame) {
455 			i = CLKF_PC(frame) - g->lowpc;
456 			if (i < g->textsize) {
457 				i /= HISTFRACTION * sizeof(*g->kcount);
458 				g->kcount[i]++;
459 			}
460 		}
461 #endif
462 		/*
463 		 * Came from kernel mode, so we were:
464 		 * - handling an interrupt,
465 		 * - doing syscall or trap work on behalf of the current
466 		 *   user process, or
467 		 * - spinning in the idle loop.
468 		 * Whichever it is, charge the time as appropriate.
469 		 * Note that we charge interrupts to the current process,
470 		 * regardless of whether they are ``for'' that process,
471 		 * so that we know how much of its real time was spent
472 		 * in ``non-process'' (i.e., interrupt) work.
473 		 *
474 		 * XXX assume system if frame is NULL.  A NULL frame
475 		 * can occur if ipi processing is done from an splx().
476 		 */
477 		if (frame && CLKF_INTR(frame))
478 			td->td_iticks += bump;
479 		else
480 			td->td_sticks += bump;
481 
482 		if (frame && CLKF_INTR(frame)) {
483 			cp_time[CP_INTR] += bump;
484 		} else {
485 			if (td == &mycpu->gd_idlethread)
486 				cp_time[CP_IDLE] += bump;
487 			else
488 				cp_time[CP_SYS] += bump;
489 		}
490 	}
491 }
492 
493 /*
494  * The scheduler clock typically runs at a 20Hz rate.  NOTE! systimer,
495  * the MP lock might not be held.  We can safely manipulate parts of curproc
496  * but that's about it.
497  */
498 static void
499 schedclock(systimer_t info, struct intrframe *frame)
500 {
501 	struct proc *p;
502 	struct pstats *pstats;
503 	struct rusage *ru;
504 	struct vmspace *vm;
505 	long rss;
506 
507 	schedulerclock(NULL);	/* mpsafe */
508 	if ((p = curproc) != NULL) {
509 		/* Update resource usage integrals and maximums. */
510 		if ((pstats = p->p_stats) != NULL &&
511 		    (ru = &pstats->p_ru) != NULL &&
512 		    (vm = p->p_vmspace) != NULL) {
513 			ru->ru_ixrss += pgtok(vm->vm_tsize);
514 			ru->ru_idrss += pgtok(vm->vm_dsize);
515 			ru->ru_isrss += pgtok(vm->vm_ssize);
516 			rss = pgtok(vmspace_resident_count(vm));
517 			if (ru->ru_maxrss < rss)
518 				ru->ru_maxrss = rss;
519 		}
520 	}
521 }
522 
523 /*
524  * Compute number of ticks for the specified amount of time.  The
525  * return value is intended to be used in a clock interrupt timed
526  * operation and guarenteed to meet or exceed the requested time.
527  * If the representation overflows, return INT_MAX.  The minimum return
528  * value is 1 ticks and the function will average the calculation up.
529  * If any value greater then 0 microseconds is supplied, a value
530  * of at least 2 will be returned to ensure that a near-term clock
531  * interrupt does not cause the timeout to occur (degenerately) early.
532  *
533  * Note that limit checks must take into account microseconds, which is
534  * done simply by using the smaller signed long maximum instead of
535  * the unsigned long maximum.
536  *
537  * If ints have 32 bits, then the maximum value for any timeout in
538  * 10ms ticks is 248 days.
539  */
540 int
541 tvtohz_high(struct timeval *tv)
542 {
543 	int ticks;
544 	long sec, usec;
545 
546 	sec = tv->tv_sec;
547 	usec = tv->tv_usec;
548 	if (usec < 0) {
549 		sec--;
550 		usec += 1000000;
551 	}
552 	if (sec < 0) {
553 #ifdef DIAGNOSTIC
554 		if (usec > 0) {
555 			sec++;
556 			usec -= 1000000;
557 		}
558 		printf("tvotohz: negative time difference %ld sec %ld usec\n",
559 		       sec, usec);
560 #endif
561 		ticks = 1;
562 	} else if (sec <= INT_MAX / hz) {
563 		ticks = (int)(sec * hz +
564 			    ((u_long)usec + (tick - 1)) / tick) + 1;
565 	} else {
566 		ticks = INT_MAX;
567 	}
568 	return (ticks);
569 }
570 
571 /*
572  * Compute number of ticks for the specified amount of time, erroring on
573  * the side of it being too low to ensure that sleeping the returned number
574  * of ticks will not result in a late return.
575  *
576  * The supplied timeval may not be negative and should be normalized.  A
577  * return value of 0 is possible if the timeval converts to less then
578  * 1 tick.
579  *
580  * If ints have 32 bits, then the maximum value for any timeout in
581  * 10ms ticks is 248 days.
582  */
583 int
584 tvtohz_low(struct timeval *tv)
585 {
586 	int ticks;
587 	long sec;
588 
589 	sec = tv->tv_sec;
590 	if (sec <= INT_MAX / hz)
591 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
592 	else
593 		ticks = INT_MAX;
594 	return (ticks);
595 }
596 
597 
598 /*
599  * Start profiling on a process.
600  *
601  * Kernel profiling passes proc0 which never exits and hence
602  * keeps the profile clock running constantly.
603  */
604 void
605 startprofclock(struct proc *p)
606 {
607 	if ((p->p_flag & P_PROFIL) == 0) {
608 		p->p_flag |= P_PROFIL;
609 #if 0	/* XXX */
610 		if (++profprocs == 1 && stathz != 0) {
611 			s = splstatclock();
612 			psdiv = psratio;
613 			setstatclockrate(profhz);
614 			splx(s);
615 		}
616 #endif
617 	}
618 }
619 
620 /*
621  * Stop profiling on a process.
622  */
623 void
624 stopprofclock(struct proc *p)
625 {
626 	if (p->p_flag & P_PROFIL) {
627 		p->p_flag &= ~P_PROFIL;
628 #if 0	/* XXX */
629 		if (--profprocs == 0 && stathz != 0) {
630 			s = splstatclock();
631 			psdiv = 1;
632 			setstatclockrate(stathz);
633 			splx(s);
634 		}
635 #endif
636 	}
637 }
638 
639 /*
640  * Return information about system clocks.
641  */
642 static int
643 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
644 {
645 	struct clockinfo clkinfo;
646 	/*
647 	 * Construct clockinfo structure.
648 	 */
649 	clkinfo.hz = hz;
650 	clkinfo.tick = tick;
651 	clkinfo.tickadj = tickadj;
652 	clkinfo.profhz = profhz;
653 	clkinfo.stathz = stathz ? stathz : hz;
654 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
655 }
656 
657 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
658 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
659 
660 /*
661  * We have eight functions for looking at the clock, four for
662  * microseconds and four for nanoseconds.  For each there is fast
663  * but less precise version "get{nano|micro}[up]time" which will
664  * return a time which is up to 1/HZ previous to the call, whereas
665  * the raw version "{nano|micro}[up]time" will return a timestamp
666  * which is as precise as possible.  The "up" variants return the
667  * time relative to system boot, these are well suited for time
668  * interval measurements.
669  *
670  * Each cpu independantly maintains the current time of day, so all
671  * we need to do to protect ourselves from changes is to do a loop
672  * check on the seconds field changing out from under us.
673  *
674  * The system timer maintains a 32 bit count and due to various issues
675  * it is possible for the calculated delta to occassionally exceed
676  * cputimer_freq.  If this occurs the cputimer_freq64_nsec multiplication
677  * can easily overflow, so we deal with the case.  For uniformity we deal
678  * with the case in the usec case too.
679  */
680 void
681 getmicrouptime(struct timeval *tvp)
682 {
683 	struct globaldata *gd = mycpu;
684 	sysclock_t delta;
685 
686 	do {
687 		tvp->tv_sec = gd->gd_time_seconds;
688 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
689 	} while (tvp->tv_sec != gd->gd_time_seconds);
690 
691 	if (delta >= cputimer_freq) {
692 		tvp->tv_sec += delta / cputimer_freq;
693 		delta %= cputimer_freq;
694 	}
695 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
696 	if (tvp->tv_usec >= 1000000) {
697 		tvp->tv_usec -= 1000000;
698 		++tvp->tv_sec;
699 	}
700 }
701 
702 void
703 getnanouptime(struct timespec *tsp)
704 {
705 	struct globaldata *gd = mycpu;
706 	sysclock_t delta;
707 
708 	do {
709 		tsp->tv_sec = gd->gd_time_seconds;
710 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
711 	} while (tsp->tv_sec != gd->gd_time_seconds);
712 
713 	if (delta >= cputimer_freq) {
714 		tsp->tv_sec += delta / cputimer_freq;
715 		delta %= cputimer_freq;
716 	}
717 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
718 }
719 
720 void
721 microuptime(struct timeval *tvp)
722 {
723 	struct globaldata *gd = mycpu;
724 	sysclock_t delta;
725 
726 	do {
727 		tvp->tv_sec = gd->gd_time_seconds;
728 		delta = cputimer_count() - gd->gd_cpuclock_base;
729 	} while (tvp->tv_sec != gd->gd_time_seconds);
730 
731 	if (delta >= cputimer_freq) {
732 		tvp->tv_sec += delta / cputimer_freq;
733 		delta %= cputimer_freq;
734 	}
735 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
736 }
737 
738 void
739 nanouptime(struct timespec *tsp)
740 {
741 	struct globaldata *gd = mycpu;
742 	sysclock_t delta;
743 
744 	do {
745 		tsp->tv_sec = gd->gd_time_seconds;
746 		delta = cputimer_count() - gd->gd_cpuclock_base;
747 	} while (tsp->tv_sec != gd->gd_time_seconds);
748 
749 	if (delta >= cputimer_freq) {
750 		tsp->tv_sec += delta / cputimer_freq;
751 		delta %= cputimer_freq;
752 	}
753 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
754 }
755 
756 /*
757  * realtime routines
758  */
759 
760 void
761 getmicrotime(struct timeval *tvp)
762 {
763 	struct globaldata *gd = mycpu;
764 	sysclock_t delta;
765 
766 	do {
767 		tvp->tv_sec = gd->gd_time_seconds;
768 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
769 	} while (tvp->tv_sec != gd->gd_time_seconds);
770 
771 	if (delta >= cputimer_freq) {
772 		tvp->tv_sec += delta / cputimer_freq;
773 		delta %= cputimer_freq;
774 	}
775 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
776 
777 	tvp->tv_sec += basetime.tv_sec;
778 	tvp->tv_usec += basetime.tv_nsec / 1000;
779 	while (tvp->tv_usec >= 1000000) {
780 		tvp->tv_usec -= 1000000;
781 		++tvp->tv_sec;
782 	}
783 }
784 
785 void
786 getnanotime(struct timespec *tsp)
787 {
788 	struct globaldata *gd = mycpu;
789 	sysclock_t delta;
790 
791 	do {
792 		tsp->tv_sec = gd->gd_time_seconds;
793 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
794 	} while (tsp->tv_sec != gd->gd_time_seconds);
795 
796 	if (delta >= cputimer_freq) {
797 		tsp->tv_sec += delta / cputimer_freq;
798 		delta %= cputimer_freq;
799 	}
800 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
801 
802 	tsp->tv_sec += basetime.tv_sec;
803 	tsp->tv_nsec += basetime.tv_nsec;
804 	while (tsp->tv_nsec >= 1000000000) {
805 		tsp->tv_nsec -= 1000000000;
806 		++tsp->tv_sec;
807 	}
808 }
809 
810 void
811 microtime(struct timeval *tvp)
812 {
813 	struct globaldata *gd = mycpu;
814 	sysclock_t delta;
815 
816 	do {
817 		tvp->tv_sec = gd->gd_time_seconds;
818 		delta = cputimer_count() - gd->gd_cpuclock_base;
819 	} while (tvp->tv_sec != gd->gd_time_seconds);
820 
821 	if (delta >= cputimer_freq) {
822 		tvp->tv_sec += delta / cputimer_freq;
823 		delta %= cputimer_freq;
824 	}
825 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
826 
827 	tvp->tv_sec += basetime.tv_sec;
828 	tvp->tv_usec += basetime.tv_nsec / 1000;
829 	while (tvp->tv_usec >= 1000000) {
830 		tvp->tv_usec -= 1000000;
831 		++tvp->tv_sec;
832 	}
833 }
834 
835 void
836 nanotime(struct timespec *tsp)
837 {
838 	struct globaldata *gd = mycpu;
839 	sysclock_t delta;
840 
841 	do {
842 		tsp->tv_sec = gd->gd_time_seconds;
843 		delta = cputimer_count() - gd->gd_cpuclock_base;
844 	} while (tsp->tv_sec != gd->gd_time_seconds);
845 
846 	if (delta >= cputimer_freq) {
847 		tsp->tv_sec += delta / cputimer_freq;
848 		delta %= cputimer_freq;
849 	}
850 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
851 
852 	tsp->tv_sec += basetime.tv_sec;
853 	tsp->tv_nsec += basetime.tv_nsec;
854 	while (tsp->tv_nsec >= 1000000000) {
855 		tsp->tv_nsec -= 1000000000;
856 		++tsp->tv_sec;
857 	}
858 }
859 
860 int
861 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
862 {
863 	pps_params_t *app;
864 	struct pps_fetch_args *fapi;
865 #ifdef PPS_SYNC
866 	struct pps_kcbind_args *kapi;
867 #endif
868 
869 	switch (cmd) {
870 	case PPS_IOC_CREATE:
871 		return (0);
872 	case PPS_IOC_DESTROY:
873 		return (0);
874 	case PPS_IOC_SETPARAMS:
875 		app = (pps_params_t *)data;
876 		if (app->mode & ~pps->ppscap)
877 			return (EINVAL);
878 		pps->ppsparam = *app;
879 		return (0);
880 	case PPS_IOC_GETPARAMS:
881 		app = (pps_params_t *)data;
882 		*app = pps->ppsparam;
883 		app->api_version = PPS_API_VERS_1;
884 		return (0);
885 	case PPS_IOC_GETCAP:
886 		*(int*)data = pps->ppscap;
887 		return (0);
888 	case PPS_IOC_FETCH:
889 		fapi = (struct pps_fetch_args *)data;
890 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
891 			return (EINVAL);
892 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
893 			return (EOPNOTSUPP);
894 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
895 		fapi->pps_info_buf = pps->ppsinfo;
896 		return (0);
897 	case PPS_IOC_KCBIND:
898 #ifdef PPS_SYNC
899 		kapi = (struct pps_kcbind_args *)data;
900 		/* XXX Only root should be able to do this */
901 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
902 			return (EINVAL);
903 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
904 			return (EINVAL);
905 		if (kapi->edge & ~pps->ppscap)
906 			return (EINVAL);
907 		pps->kcmode = kapi->edge;
908 		return (0);
909 #else
910 		return (EOPNOTSUPP);
911 #endif
912 	default:
913 		return (ENOTTY);
914 	}
915 }
916 
917 void
918 pps_init(struct pps_state *pps)
919 {
920 	pps->ppscap |= PPS_TSFMT_TSPEC;
921 	if (pps->ppscap & PPS_CAPTUREASSERT)
922 		pps->ppscap |= PPS_OFFSETASSERT;
923 	if (pps->ppscap & PPS_CAPTURECLEAR)
924 		pps->ppscap |= PPS_OFFSETCLEAR;
925 }
926 
927 void
928 pps_event(struct pps_state *pps, sysclock_t count, int event)
929 {
930 	struct globaldata *gd;
931 	struct timespec *tsp;
932 	struct timespec *osp;
933 	struct timespec ts;
934 	sysclock_t *pcount;
935 #ifdef PPS_SYNC
936 	sysclock_t tcount;
937 #endif
938 	sysclock_t delta;
939 	pps_seq_t *pseq;
940 	int foff;
941 	int fhard;
942 
943 	gd = mycpu;
944 
945 	/* Things would be easier with arrays... */
946 	if (event == PPS_CAPTUREASSERT) {
947 		tsp = &pps->ppsinfo.assert_timestamp;
948 		osp = &pps->ppsparam.assert_offset;
949 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
950 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
951 		pcount = &pps->ppscount[0];
952 		pseq = &pps->ppsinfo.assert_sequence;
953 	} else {
954 		tsp = &pps->ppsinfo.clear_timestamp;
955 		osp = &pps->ppsparam.clear_offset;
956 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
957 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
958 		pcount = &pps->ppscount[1];
959 		pseq = &pps->ppsinfo.clear_sequence;
960 	}
961 
962 	/* Nothing really happened */
963 	if (*pcount == count)
964 		return;
965 
966 	*pcount = count;
967 
968 	do {
969 		ts.tv_sec = gd->gd_time_seconds;
970 		delta = count - gd->gd_cpuclock_base;
971 	} while (ts.tv_sec != gd->gd_time_seconds);
972 
973 	if (delta >= cputimer_freq) {
974 		ts.tv_sec += delta / cputimer_freq;
975 		delta %= cputimer_freq;
976 	}
977 	ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
978 	ts.tv_sec += basetime.tv_sec;
979 	ts.tv_nsec += basetime.tv_nsec;
980 	while (ts.tv_nsec >= 1000000000) {
981 		ts.tv_nsec -= 1000000000;
982 		++ts.tv_sec;
983 	}
984 
985 	(*pseq)++;
986 	*tsp = ts;
987 
988 	if (foff) {
989 		timespecadd(tsp, osp);
990 		if (tsp->tv_nsec < 0) {
991 			tsp->tv_nsec += 1000000000;
992 			tsp->tv_sec -= 1;
993 		}
994 	}
995 #ifdef PPS_SYNC
996 	if (fhard) {
997 		/* magic, at its best... */
998 		tcount = count - pps->ppscount[2];
999 		pps->ppscount[2] = count;
1000 		if (tcount >= cputimer_freq) {
1001 			delta = 1000000000 * (tcount / cputimer_freq) +
1002 				(cputimer_freq64_nsec *
1003 				 (tcount % cputimer_freq)) >> 32;
1004 		} else {
1005 			delta = (cputimer_freq64_nsec * tcount) >> 32;
1006 		}
1007 		hardpps(tsp, delta);
1008 	}
1009 #endif
1010 }
1011 
1012