xref: /dflybsd-src/sys/kern/kern_clock.c (revision 10f84ad9e1cb939878f6c1ec7e73c30dd133a0f2)
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 1.24 2004/09/17 00:18:09 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 	systimer_init_periodic(&gd->gd_hardclock, hardclock, NULL, hz);
193 	systimer_init_periodic(&gd->gd_statclock, statclock, NULL, stathz);
194 	/* XXX correct the frequency for scheduler / estcpu tests */
195 	systimer_init_periodic(&gd->gd_schedclock, schedclock,
196 				NULL, ESTCPUFREQ);
197 	crit_exit();
198 }
199 
200 /*
201  * Resynchronize gd_cpuclock_base after the system has been woken up from
202  * a sleep.  It is absolutely essential that all the cpus be properly
203  * synchronized.  Resynching is required because nanouptime() and friends
204  * will overflow intermediate multiplications if more then 2 seconds
205  * worth of cputimer_cont() delta has built up.
206  */
207 #ifdef SMP
208 
209 static
210 void
211 restoreclocks_remote(lwkt_cpusync_t poll)
212 {
213 	mycpu->gd_cpuclock_base = *(sysclock_t *)poll->cs_data;
214 	mycpu->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
215 }
216 
217 #endif
218 
219 void
220 restoreclocks(void)
221 {
222 	sysclock_t base = cputimer_count();
223 #ifdef SMP
224 	lwkt_cpusync_simple(-1, restoreclocks_remote, &base);
225 #else
226 	mycpu->gd_cpuclock_base = base;
227 #endif
228 }
229 
230 /*
231  * This sets the current real time of day.  Timespecs are in seconds and
232  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
233  * instead we adjust basetime so basetime + gd_* results in the current
234  * time of day.  This way the gd_* fields are guarenteed to represent
235  * a monotonically increasing 'uptime' value.
236  */
237 void
238 set_timeofday(struct timespec *ts)
239 {
240 	struct timespec ts2;
241 
242 	/*
243 	 * XXX SMP / non-atomic basetime updates
244 	 */
245 	crit_enter();
246 	nanouptime(&ts2);
247 	basetime.tv_sec = ts->tv_sec - ts2.tv_sec;
248 	basetime.tv_nsec = ts->tv_nsec - ts2.tv_nsec;
249 	if (basetime.tv_nsec < 0) {
250 	    basetime.tv_nsec += 1000000000;
251 	    --basetime.tv_sec;
252 	}
253 	boottime.tv_sec = basetime.tv_sec - mycpu->gd_time_seconds;
254 	timedelta = 0;
255 	crit_exit();
256 }
257 
258 /*
259  * Each cpu has its own hardclock, but we only increments ticks and softticks
260  * on cpu #0.
261  *
262  * NOTE! systimer! the MP lock might not be held here.  We can only safely
263  * manipulate objects owned by the current cpu.
264  */
265 static void
266 hardclock(systimer_t info, struct intrframe *frame)
267 {
268 	sysclock_t cputicks;
269 	struct proc *p;
270 	struct pstats *pstats;
271 	struct globaldata *gd = mycpu;
272 
273 	/*
274 	 * Realtime updates are per-cpu.  Note that timer corrections as
275 	 * returned by microtime() and friends make an additional adjustment
276 	 * using a system-wise 'basetime', but the running time is always
277 	 * taken from the per-cpu globaldata area.  Since the same clock
278 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
279 	 * stay in synch.
280 	 *
281 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
282 	 * to reverse index gd_cpuclock_base.
283 	 */
284 	cputicks = info->time - gd->gd_cpuclock_base;
285 	if (cputicks > cputimer_freq) {
286 		++gd->gd_time_seconds;
287 		gd->gd_cpuclock_base += cputimer_freq;
288 	}
289 
290 	/*
291 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
292 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
293 	 * by updating basetime.
294 	 */
295 	if (gd->gd_cpuid == 0) {
296 	    struct timespec nts;
297 	    int leap;
298 
299 	    ++ticks;
300 
301 #ifdef DEVICE_POLLING
302 	    hardclock_device_poll();	/* mpsafe, short and quick */
303 #endif /* DEVICE_POLLING */
304 
305 #if 0
306 	    if (tco->tc_poll_pps)
307 		tco->tc_poll_pps(tco);
308 #endif
309 	    /*
310 	     * Apply adjtime corrections.  At the moment only do this if
311 	     * we can get the MP lock to interlock with adjtime's modification
312 	     * of these variables.  Note that basetime adjustments are not
313 	     * MP safe either XXX.
314 	     */
315 	    if (timedelta != 0 && try_mplock()) {
316 		basetime.tv_nsec += tickdelta * 1000;
317 		if (basetime.tv_nsec >= 1000000000) {
318 		    basetime.tv_nsec -= 1000000000;
319 		    ++basetime.tv_sec;
320 		} else if (basetime.tv_nsec < 0) {
321 		    basetime.tv_nsec += 1000000000;
322 		    --basetime.tv_sec;
323 		}
324 		timedelta -= tickdelta;
325 		rel_mplock();
326 	    }
327 
328 	    /*
329 	     * Apply per-tick compensation.  ticks_adj adjusts for both
330 	     * offset and frequency, and could be negative.
331 	     */
332 	    if (nsec_adj != 0 && try_mplock()) {
333 		nsec_acc += nsec_adj;
334 		if (nsec_acc >= 0x100000000LL) {
335 		    basetime.tv_nsec += nsec_acc >> 32;
336 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
337 		} else if (nsec_acc <= -0x100000000LL) {
338 		    basetime.tv_nsec -= -nsec_acc >> 32;
339 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
340 		}
341 		if (basetime.tv_nsec >= 1000000000) {
342 		    basetime.tv_nsec -= 1000000000;
343 		    ++basetime.tv_sec;
344 		} else if (basetime.tv_nsec < 0) {
345 		    basetime.tv_nsec += 1000000000;
346 		    --basetime.tv_sec;
347 		}
348 		rel_mplock();
349 	    }
350 
351 	    /*
352 	     * If the realtime-adjusted seconds hand rolls over then tell
353 	     * ntp_update_second() what we did in the last second so it can
354 	     * calculate what to do in the next second.  It may also add
355 	     * or subtract a leap second.
356 	     */
357 	    getnanotime(&nts);
358 	    if (time_second != nts.tv_sec) {
359 		leap = ntp_update_second(time_second, &nsec_adj);
360 		basetime.tv_sec += leap;
361 		time_second = nts.tv_sec + leap;
362 		nsec_adj /= hz;
363 	    }
364 	}
365 
366 	/*
367 	 * softticks are handled for all cpus
368 	 */
369 	hardclock_softtick(gd);
370 
371 	/*
372 	 * ITimer handling is per-tick, per-cpu.  I don't think psignal()
373 	 * is mpsafe on curproc, so XXX get the mplock.
374 	 */
375 	if ((p = curproc) != NULL && try_mplock()) {
376 		pstats = p->p_stats;
377 		if (frame && CLKF_USERMODE(frame) &&
378 		    timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
379 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
380 			psignal(p, SIGVTALRM);
381 		if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
382 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
383 			psignal(p, SIGPROF);
384 		rel_mplock();
385 	}
386 	setdelayed();
387 }
388 
389 /*
390  * The statistics clock typically runs at a 125Hz rate, and is intended
391  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
392  *
393  * NOTE! systimer! the MP lock might not be held here.  We can only safely
394  * manipulate objects owned by the current cpu.
395  *
396  * The stats clock is responsible for grabbing a profiling sample.
397  * Most of the statistics are only used by user-level statistics programs.
398  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
399  * p->p_estcpu.
400  *
401  * Like the other clocks, the stat clock is called from what is effectively
402  * a fast interrupt, so the context should be the thread/process that got
403  * interrupted.
404  */
405 static void
406 statclock(systimer_t info, struct intrframe *frame)
407 {
408 #ifdef GPROF
409 	struct gmonparam *g;
410 	int i;
411 #endif
412 	thread_t td;
413 	struct proc *p;
414 	int bump;
415 	struct timeval tv;
416 	struct timeval *stv;
417 
418 	/*
419 	 * How big was our timeslice relative to the last time?
420 	 */
421 	microuptime(&tv);	/* mpsafe */
422 	stv = &mycpu->gd_stattv;
423 	if (stv->tv_sec == 0) {
424 	    bump = 1;
425 	} else {
426 	    bump = tv.tv_usec - stv->tv_usec +
427 		(tv.tv_sec - stv->tv_sec) * 1000000;
428 	    if (bump < 0)
429 		bump = 0;
430 	    if (bump > 1000000)
431 		bump = 1000000;
432 	}
433 	*stv = tv;
434 
435 	td = curthread;
436 	p = td->td_proc;
437 
438 	if (frame && CLKF_USERMODE(frame)) {
439 		/*
440 		 * Came from userland, handle user time and deal with
441 		 * possible process.
442 		 */
443 		if (p && (p->p_flag & P_PROFIL))
444 			addupc_intr(p, CLKF_PC(frame), 1);
445 		td->td_uticks += bump;
446 
447 		/*
448 		 * Charge the time as appropriate
449 		 */
450 		if (p && p->p_nice > NZERO)
451 			cp_time[CP_NICE] += bump;
452 		else
453 			cp_time[CP_USER] += bump;
454 	} else {
455 #ifdef GPROF
456 		/*
457 		 * Kernel statistics are just like addupc_intr, only easier.
458 		 */
459 		g = &_gmonparam;
460 		if (g->state == GMON_PROF_ON && frame) {
461 			i = CLKF_PC(frame) - g->lowpc;
462 			if (i < g->textsize) {
463 				i /= HISTFRACTION * sizeof(*g->kcount);
464 				g->kcount[i]++;
465 			}
466 		}
467 #endif
468 		/*
469 		 * Came from kernel mode, so we were:
470 		 * - handling an interrupt,
471 		 * - doing syscall or trap work on behalf of the current
472 		 *   user process, or
473 		 * - spinning in the idle loop.
474 		 * Whichever it is, charge the time as appropriate.
475 		 * Note that we charge interrupts to the current process,
476 		 * regardless of whether they are ``for'' that process,
477 		 * so that we know how much of its real time was spent
478 		 * in ``non-process'' (i.e., interrupt) work.
479 		 *
480 		 * XXX assume system if frame is NULL.  A NULL frame
481 		 * can occur if ipi processing is done from an splx().
482 		 */
483 		if (frame && CLKF_INTR(frame))
484 			td->td_iticks += bump;
485 		else
486 			td->td_sticks += bump;
487 
488 		if (frame && CLKF_INTR(frame)) {
489 			cp_time[CP_INTR] += bump;
490 		} else {
491 			if (td == &mycpu->gd_idlethread)
492 				cp_time[CP_IDLE] += bump;
493 			else
494 				cp_time[CP_SYS] += bump;
495 		}
496 	}
497 }
498 
499 /*
500  * The scheduler clock typically runs at a 20Hz rate.  NOTE! systimer,
501  * the MP lock might not be held.  We can safely manipulate parts of curproc
502  * but that's about it.
503  */
504 static void
505 schedclock(systimer_t info, struct intrframe *frame)
506 {
507 	struct proc *p;
508 	struct pstats *pstats;
509 	struct rusage *ru;
510 	struct vmspace *vm;
511 	long rss;
512 
513 	schedulerclock(NULL);	/* mpsafe */
514 	if ((p = curproc) != NULL) {
515 		/* Update resource usage integrals and maximums. */
516 		if ((pstats = p->p_stats) != NULL &&
517 		    (ru = &pstats->p_ru) != NULL &&
518 		    (vm = p->p_vmspace) != NULL) {
519 			ru->ru_ixrss += pgtok(vm->vm_tsize);
520 			ru->ru_idrss += pgtok(vm->vm_dsize);
521 			ru->ru_isrss += pgtok(vm->vm_ssize);
522 			rss = pgtok(vmspace_resident_count(vm));
523 			if (ru->ru_maxrss < rss)
524 				ru->ru_maxrss = rss;
525 		}
526 	}
527 }
528 
529 /*
530  * Compute number of ticks for the specified amount of time.  The
531  * return value is intended to be used in a clock interrupt timed
532  * operation and guarenteed to meet or exceed the requested time.
533  * If the representation overflows, return INT_MAX.  The minimum return
534  * value is 1 ticks and the function will average the calculation up.
535  * If any value greater then 0 microseconds is supplied, a value
536  * of at least 2 will be returned to ensure that a near-term clock
537  * interrupt does not cause the timeout to occur (degenerately) early.
538  *
539  * Note that limit checks must take into account microseconds, which is
540  * done simply by using the smaller signed long maximum instead of
541  * the unsigned long maximum.
542  *
543  * If ints have 32 bits, then the maximum value for any timeout in
544  * 10ms ticks is 248 days.
545  */
546 int
547 tvtohz_high(struct timeval *tv)
548 {
549 	int ticks;
550 	long sec, usec;
551 
552 	sec = tv->tv_sec;
553 	usec = tv->tv_usec;
554 	if (usec < 0) {
555 		sec--;
556 		usec += 1000000;
557 	}
558 	if (sec < 0) {
559 #ifdef DIAGNOSTIC
560 		if (usec > 0) {
561 			sec++;
562 			usec -= 1000000;
563 		}
564 		printf("tvotohz: negative time difference %ld sec %ld usec\n",
565 		       sec, usec);
566 #endif
567 		ticks = 1;
568 	} else if (sec <= INT_MAX / hz) {
569 		ticks = (int)(sec * hz +
570 			    ((u_long)usec + (tick - 1)) / tick) + 1;
571 	} else {
572 		ticks = INT_MAX;
573 	}
574 	return (ticks);
575 }
576 
577 /*
578  * Compute number of ticks for the specified amount of time, erroring on
579  * the side of it being too low to ensure that sleeping the returned number
580  * of ticks will not result in a late return.
581  *
582  * The supplied timeval may not be negative and should be normalized.  A
583  * return value of 0 is possible if the timeval converts to less then
584  * 1 tick.
585  *
586  * If ints have 32 bits, then the maximum value for any timeout in
587  * 10ms ticks is 248 days.
588  */
589 int
590 tvtohz_low(struct timeval *tv)
591 {
592 	int ticks;
593 	long sec;
594 
595 	sec = tv->tv_sec;
596 	if (sec <= INT_MAX / hz)
597 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
598 	else
599 		ticks = INT_MAX;
600 	return (ticks);
601 }
602 
603 
604 /*
605  * Start profiling on a process.
606  *
607  * Kernel profiling passes proc0 which never exits and hence
608  * keeps the profile clock running constantly.
609  */
610 void
611 startprofclock(struct proc *p)
612 {
613 	if ((p->p_flag & P_PROFIL) == 0) {
614 		p->p_flag |= P_PROFIL;
615 #if 0	/* XXX */
616 		if (++profprocs == 1 && stathz != 0) {
617 			s = splstatclock();
618 			psdiv = psratio;
619 			setstatclockrate(profhz);
620 			splx(s);
621 		}
622 #endif
623 	}
624 }
625 
626 /*
627  * Stop profiling on a process.
628  */
629 void
630 stopprofclock(struct proc *p)
631 {
632 	if (p->p_flag & P_PROFIL) {
633 		p->p_flag &= ~P_PROFIL;
634 #if 0	/* XXX */
635 		if (--profprocs == 0 && stathz != 0) {
636 			s = splstatclock();
637 			psdiv = 1;
638 			setstatclockrate(stathz);
639 			splx(s);
640 		}
641 #endif
642 	}
643 }
644 
645 /*
646  * Return information about system clocks.
647  */
648 static int
649 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
650 {
651 	struct clockinfo clkinfo;
652 	/*
653 	 * Construct clockinfo structure.
654 	 */
655 	clkinfo.hz = hz;
656 	clkinfo.tick = tick;
657 	clkinfo.tickadj = tickadj;
658 	clkinfo.profhz = profhz;
659 	clkinfo.stathz = stathz ? stathz : hz;
660 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
661 }
662 
663 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
664 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
665 
666 /*
667  * We have eight functions for looking at the clock, four for
668  * microseconds and four for nanoseconds.  For each there is fast
669  * but less precise version "get{nano|micro}[up]time" which will
670  * return a time which is up to 1/HZ previous to the call, whereas
671  * the raw version "{nano|micro}[up]time" will return a timestamp
672  * which is as precise as possible.  The "up" variants return the
673  * time relative to system boot, these are well suited for time
674  * interval measurements.
675  *
676  * Each cpu independantly maintains the current time of day, so all
677  * we need to do to protect ourselves from changes is to do a loop
678  * check on the seconds field changing out from under us.
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 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
691 	if (tvp->tv_usec >= 1000000) {
692 		tvp->tv_usec -= 1000000;
693 		++tvp->tv_sec;
694 	}
695 }
696 
697 void
698 getnanouptime(struct timespec *tsp)
699 {
700 	struct globaldata *gd = mycpu;
701 	sysclock_t delta;
702 
703 	do {
704 		tsp->tv_sec = gd->gd_time_seconds;
705 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
706 	} while (tsp->tv_sec != gd->gd_time_seconds);
707 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
708 	if (tsp->tv_nsec >= 1000000000) {
709 		tsp->tv_nsec -= 1000000000;
710 		++tsp->tv_sec;
711 	}
712 }
713 
714 void
715 microuptime(struct timeval *tvp)
716 {
717 	struct globaldata *gd = mycpu;
718 	sysclock_t delta;
719 
720 	do {
721 		tvp->tv_sec = gd->gd_time_seconds;
722 		delta = cputimer_count() - gd->gd_cpuclock_base;
723 	} while (tvp->tv_sec != gd->gd_time_seconds);
724 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
725 	if (tvp->tv_usec >= 1000000) {
726 		tvp->tv_usec -= 1000000;
727 		++tvp->tv_sec;
728 	}
729 }
730 
731 void
732 nanouptime(struct timespec *tsp)
733 {
734 	struct globaldata *gd = mycpu;
735 	sysclock_t delta;
736 
737 	do {
738 		tsp->tv_sec = gd->gd_time_seconds;
739 		delta = cputimer_count() - gd->gd_cpuclock_base;
740 	} while (tsp->tv_sec != gd->gd_time_seconds);
741 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
742 	if (tsp->tv_nsec >= 1000000000) {
743 		tsp->tv_nsec -= 1000000000;
744 		++tsp->tv_sec;
745 	}
746 }
747 
748 /*
749  * realtime routines
750  */
751 
752 void
753 getmicrotime(struct timeval *tvp)
754 {
755 	struct globaldata *gd = mycpu;
756 	sysclock_t delta;
757 
758 	do {
759 		tvp->tv_sec = gd->gd_time_seconds;
760 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
761 	} while (tvp->tv_sec != gd->gd_time_seconds);
762 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
763 
764 	tvp->tv_sec += basetime.tv_sec;
765 	tvp->tv_usec += basetime.tv_nsec / 1000;
766 	while (tvp->tv_usec >= 1000000) {
767 		tvp->tv_usec -= 1000000;
768 		++tvp->tv_sec;
769 	}
770 }
771 
772 void
773 getnanotime(struct timespec *tsp)
774 {
775 	struct globaldata *gd = mycpu;
776 	sysclock_t delta;
777 
778 	do {
779 		tsp->tv_sec = gd->gd_time_seconds;
780 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
781 	} while (tsp->tv_sec != gd->gd_time_seconds);
782 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
783 
784 	tsp->tv_sec += basetime.tv_sec;
785 	tsp->tv_nsec += basetime.tv_nsec;
786 	while (tsp->tv_nsec >= 1000000000) {
787 		tsp->tv_nsec -= 1000000000;
788 		++tsp->tv_sec;
789 	}
790 }
791 
792 void
793 microtime(struct timeval *tvp)
794 {
795 	struct globaldata *gd = mycpu;
796 	sysclock_t delta;
797 
798 	do {
799 		tvp->tv_sec = gd->gd_time_seconds;
800 		delta = cputimer_count() - gd->gd_cpuclock_base;
801 	} while (tvp->tv_sec != gd->gd_time_seconds);
802 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
803 
804 	tvp->tv_sec += basetime.tv_sec;
805 	tvp->tv_usec += basetime.tv_nsec / 1000;
806 	while (tvp->tv_usec >= 1000000) {
807 		tvp->tv_usec -= 1000000;
808 		++tvp->tv_sec;
809 	}
810 }
811 
812 void
813 nanotime(struct timespec *tsp)
814 {
815 	struct globaldata *gd = mycpu;
816 	sysclock_t delta;
817 
818 	do {
819 		tsp->tv_sec = gd->gd_time_seconds;
820 		delta = cputimer_count() - gd->gd_cpuclock_base;
821 	} while (tsp->tv_sec != gd->gd_time_seconds);
822 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
823 
824 	tsp->tv_sec += basetime.tv_sec;
825 	tsp->tv_nsec += basetime.tv_nsec;
826 	while (tsp->tv_nsec >= 1000000000) {
827 		tsp->tv_nsec -= 1000000000;
828 		++tsp->tv_sec;
829 	}
830 }
831 
832 int
833 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
834 {
835 	pps_params_t *app;
836 	struct pps_fetch_args *fapi;
837 #ifdef PPS_SYNC
838 	struct pps_kcbind_args *kapi;
839 #endif
840 
841 	switch (cmd) {
842 	case PPS_IOC_CREATE:
843 		return (0);
844 	case PPS_IOC_DESTROY:
845 		return (0);
846 	case PPS_IOC_SETPARAMS:
847 		app = (pps_params_t *)data;
848 		if (app->mode & ~pps->ppscap)
849 			return (EINVAL);
850 		pps->ppsparam = *app;
851 		return (0);
852 	case PPS_IOC_GETPARAMS:
853 		app = (pps_params_t *)data;
854 		*app = pps->ppsparam;
855 		app->api_version = PPS_API_VERS_1;
856 		return (0);
857 	case PPS_IOC_GETCAP:
858 		*(int*)data = pps->ppscap;
859 		return (0);
860 	case PPS_IOC_FETCH:
861 		fapi = (struct pps_fetch_args *)data;
862 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
863 			return (EINVAL);
864 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
865 			return (EOPNOTSUPP);
866 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
867 		fapi->pps_info_buf = pps->ppsinfo;
868 		return (0);
869 	case PPS_IOC_KCBIND:
870 #ifdef PPS_SYNC
871 		kapi = (struct pps_kcbind_args *)data;
872 		/* XXX Only root should be able to do this */
873 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
874 			return (EINVAL);
875 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
876 			return (EINVAL);
877 		if (kapi->edge & ~pps->ppscap)
878 			return (EINVAL);
879 		pps->kcmode = kapi->edge;
880 		return (0);
881 #else
882 		return (EOPNOTSUPP);
883 #endif
884 	default:
885 		return (ENOTTY);
886 	}
887 }
888 
889 void
890 pps_init(struct pps_state *pps)
891 {
892 	pps->ppscap |= PPS_TSFMT_TSPEC;
893 	if (pps->ppscap & PPS_CAPTUREASSERT)
894 		pps->ppscap |= PPS_OFFSETASSERT;
895 	if (pps->ppscap & PPS_CAPTURECLEAR)
896 		pps->ppscap |= PPS_OFFSETCLEAR;
897 }
898 
899 void
900 pps_event(struct pps_state *pps, sysclock_t count, int event)
901 {
902 	struct globaldata *gd;
903 	struct timespec *tsp;
904 	struct timespec *osp;
905 	struct timespec ts;
906 	sysclock_t *pcount;
907 #ifdef PPS_SYNC
908 	sysclock_t tcount;
909 #endif
910 	sysclock_t delta;
911 	pps_seq_t *pseq;
912 	int foff;
913 	int fhard;
914 
915 	gd = mycpu;
916 
917 	/* Things would be easier with arrays... */
918 	if (event == PPS_CAPTUREASSERT) {
919 		tsp = &pps->ppsinfo.assert_timestamp;
920 		osp = &pps->ppsparam.assert_offset;
921 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
922 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
923 		pcount = &pps->ppscount[0];
924 		pseq = &pps->ppsinfo.assert_sequence;
925 	} else {
926 		tsp = &pps->ppsinfo.clear_timestamp;
927 		osp = &pps->ppsparam.clear_offset;
928 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
929 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
930 		pcount = &pps->ppscount[1];
931 		pseq = &pps->ppsinfo.clear_sequence;
932 	}
933 
934 	/* Nothing really happened */
935 	if (*pcount == count)
936 		return;
937 
938 	*pcount = count;
939 
940 	do {
941 		ts.tv_sec = gd->gd_time_seconds;
942 		delta = count - gd->gd_cpuclock_base;
943 	} while (ts.tv_sec != gd->gd_time_seconds);
944 	if (delta > cputimer_freq) {
945 		ts.tv_sec += delta / cputimer_freq;
946 		delta %= cputimer_freq;
947 	}
948 	ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
949 	ts.tv_sec += basetime.tv_sec;
950 	ts.tv_nsec += basetime.tv_nsec;
951 	while (ts.tv_nsec >= 1000000000) {
952 		ts.tv_nsec -= 1000000000;
953 		++ts.tv_sec;
954 	}
955 
956 	(*pseq)++;
957 	*tsp = ts;
958 
959 	if (foff) {
960 		timespecadd(tsp, osp);
961 		if (tsp->tv_nsec < 0) {
962 			tsp->tv_nsec += 1000000000;
963 			tsp->tv_sec -= 1;
964 		}
965 	}
966 #ifdef PPS_SYNC
967 	if (fhard) {
968 		/* magic, at its best... */
969 		tcount = count - pps->ppscount[2];
970 		pps->ppscount[2] = count;
971 		delta = (cputimer_freq64_nsec * tcount) >> 32;
972 		hardpps(tsp, delta);
973 	}
974 #endif
975 }
976 
977