xref: /dflybsd-src/sys/kern/kern_clock.c (revision d2a4c6204a77aaec984b85e9754b5aa783192476)
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 1.40 2005/04/27 14:31:19 hmp Exp $
74  */
75 
76 #include "opt_ntp.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/callout.h>
81 #include <sys/kernel.h>
82 #include <sys/kinfo.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 cpu_time counts in microseconds, but most userland programs
115  * just compare relative times against the total by delta.
116  */
117 struct kinfo_cputime cputime_percpu[MAXCPU];
118 #ifdef SMP
119 static int
120 sysctl_cputime(SYSCTL_HANDLER_ARGS)
121 {
122 	int cpu, error = 0;
123 	size_t size = sizeof(struct kinfo_cputime);
124 
125 	for (cpu = 0; cpu < ncpus; ++cpu) {
126 		if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
127 			break;
128 	}
129 
130 	return (error);
131 }
132 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
133 	sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
134 #else
135 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
136     "CPU time statistics");
137 #endif
138 
139 /*
140  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
141  * microuptime().  microtime() is not drift compensated.  The real uptime
142  * with compensation is nanotime() - bootime.  boottime is recalculated
143  * whenever the real time is set based on the compensated elapsed time
144  * in seconds (gd->gd_time_seconds).
145  *
146  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
147  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
148  * the real time.
149  */
150 struct timespec boottime;	/* boot time (realtime) for reference only */
151 time_t time_second;		/* read-only 'passive' uptime in seconds */
152 
153 /*
154  * basetime is used to calculate the compensated real time of day.  The
155  * basetime can be modified on a per-tick basis by the adjtime(),
156  * ntp_adjtime(), and sysctl-based time correction APIs.
157  *
158  * Note that frequency corrections can also be made by adjusting
159  * gd_cpuclock_base.
160  *
161  * basetime is a tail-chasing FIFO, updated only by cpu #0.  The FIFO is
162  * used on both SMP and UP systems to avoid MP races between cpu's and
163  * interrupt races on UP systems.
164  */
165 #define BASETIME_ARYSIZE	16
166 #define BASETIME_ARYMASK	(BASETIME_ARYSIZE - 1)
167 static struct timespec basetime[BASETIME_ARYSIZE];
168 static volatile int basetime_index;
169 
170 static int
171 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
172 {
173 	struct timespec *bt;
174 	int error;
175 
176 	bt = &basetime[basetime_index];
177 	error = SYSCTL_OUT(req, bt, sizeof(*bt));
178 	return (error);
179 }
180 
181 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
182     &boottime, timespec, "System boottime");
183 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
184     sysctl_get_basetime, "S,timespec", "System basetime");
185 
186 static void hardclock(systimer_t info, struct intrframe *frame);
187 static void statclock(systimer_t info, struct intrframe *frame);
188 static void schedclock(systimer_t info, struct intrframe *frame);
189 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
190 
191 int	ticks;			/* system master ticks at hz */
192 int	clocks_running;		/* tsleep/timeout clocks operational */
193 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
194 int64_t	nsec_acc;		/* accumulator */
195 
196 /* NTPD time correction fields */
197 int64_t	ntp_tick_permanent;	/* per-tick adjustment in nsec << 32 */
198 int64_t	ntp_tick_acc;		/* accumulator for per-tick adjustment */
199 int64_t	ntp_delta;		/* one-time correction in nsec */
200 int64_t ntp_big_delta = 1000000000;
201 int32_t	ntp_tick_delta;		/* current adjustment rate */
202 int32_t	ntp_default_tick_delta;	/* adjustment rate for ntp_delta */
203 time_t	ntp_leap_second;	/* time of next leap second */
204 int	ntp_leap_insert;	/* whether to insert or remove a second */
205 
206 /*
207  * Finish initializing clock frequencies and start all clocks running.
208  */
209 /* ARGSUSED*/
210 static void
211 initclocks(void *dummy)
212 {
213 	cpu_initclocks();
214 #ifdef DEVICE_POLLING
215 	init_device_poll();
216 #endif
217 	/*psratio = profhz / stathz;*/
218 	initclocks_pcpu();
219 	clocks_running = 1;
220 }
221 
222 /*
223  * Called on a per-cpu basis
224  */
225 void
226 initclocks_pcpu(void)
227 {
228 	struct globaldata *gd = mycpu;
229 
230 	crit_enter();
231 	if (gd->gd_cpuid == 0) {
232 	    gd->gd_time_seconds = 1;
233 	    gd->gd_cpuclock_base = cputimer_count();
234 	} else {
235 	    /* XXX */
236 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
237 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
238 	}
239 
240 	/*
241 	 * Use a non-queued periodic systimer to prevent multiple ticks from
242 	 * building up if the sysclock jumps forward (8254 gets reset).  The
243 	 * sysclock will never jump backwards.  Our time sync is based on
244 	 * the actual sysclock, not the ticks count.
245 	 */
246 	systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
247 	systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
248 	/* XXX correct the frequency for scheduler / estcpu tests */
249 	systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
250 				NULL, ESTCPUFREQ);
251 	crit_exit();
252 }
253 
254 /*
255  * This sets the current real time of day.  Timespecs are in seconds and
256  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
257  * instead we adjust basetime so basetime + gd_* results in the current
258  * time of day.  This way the gd_* fields are guarenteed to represent
259  * a monotonically increasing 'uptime' value.
260  *
261  * When set_timeofday() is called from userland, the system call forces it
262  * onto cpu #0 since only cpu #0 can update basetime_index.
263  */
264 void
265 set_timeofday(struct timespec *ts)
266 {
267 	struct timespec *nbt;
268 	int ni;
269 
270 	/*
271 	 * XXX SMP / non-atomic basetime updates
272 	 */
273 	crit_enter();
274 	ni = (basetime_index + 1) & BASETIME_ARYMASK;
275 	nbt = &basetime[ni];
276 	nanouptime(nbt);
277 	nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
278 	nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
279 	if (nbt->tv_nsec < 0) {
280 	    nbt->tv_nsec += 1000000000;
281 	    --nbt->tv_sec;
282 	}
283 
284 	/*
285 	 * Note that basetime diverges from boottime as the clock drift is
286 	 * compensated for, so we cannot do away with boottime.  When setting
287 	 * the absolute time of day the drift is 0 (for an instant) and we
288 	 * can simply assign boottime to basetime.
289 	 *
290 	 * Note that nanouptime() is based on gd_time_seconds which is drift
291 	 * compensated up to a point (it is guarenteed to remain monotonically
292 	 * increasing).  gd_time_seconds is thus our best uptime guess and
293 	 * suitable for use in the boottime calculation.  It is already taken
294 	 * into account in the basetime calculation above.
295 	 */
296 	boottime.tv_sec = nbt->tv_sec;
297 	ntp_delta = 0;
298 
299 	/*
300 	 * We now have a new basetime, update the index.
301 	 */
302 	cpu_mb1();
303 	basetime_index = ni;
304 
305 	crit_exit();
306 }
307 
308 /*
309  * Each cpu has its own hardclock, but we only increments ticks and softticks
310  * on cpu #0.
311  *
312  * NOTE! systimer! the MP lock might not be held here.  We can only safely
313  * manipulate objects owned by the current cpu.
314  */
315 static void
316 hardclock(systimer_t info, struct intrframe *frame)
317 {
318 	sysclock_t cputicks;
319 	struct proc *p;
320 	struct pstats *pstats;
321 	struct globaldata *gd = mycpu;
322 
323 	/*
324 	 * Realtime updates are per-cpu.  Note that timer corrections as
325 	 * returned by microtime() and friends make an additional adjustment
326 	 * using a system-wise 'basetime', but the running time is always
327 	 * taken from the per-cpu globaldata area.  Since the same clock
328 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
329 	 * stay in synch.
330 	 *
331 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
332 	 * to reverse index gd_cpuclock_base, but that it is possible for
333 	 * it to temporarily get behind in the seconds if something in the
334 	 * system locks interrupts for a long period of time.  Since periodic
335 	 * timers count events, though everything should resynch again
336 	 * immediately.
337 	 */
338 	cputicks = info->time - gd->gd_cpuclock_base;
339 	if (cputicks >= cputimer_freq) {
340 		++gd->gd_time_seconds;
341 		gd->gd_cpuclock_base += cputimer_freq;
342 	}
343 
344 	/*
345 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
346 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
347 	 * by updating basetime.
348 	 */
349 	if (gd->gd_cpuid == 0) {
350 	    struct timespec *nbt;
351 	    struct timespec nts;
352 	    int leap;
353 	    int ni;
354 
355 	    ++ticks;
356 
357 #ifdef DEVICE_POLLING
358 	    hardclock_device_poll();	/* mpsafe, short and quick */
359 #endif /* DEVICE_POLLING */
360 
361 #if 0
362 	    if (tco->tc_poll_pps)
363 		tco->tc_poll_pps(tco);
364 #endif
365 
366 	    /*
367 	     * Calculate the new basetime index.  We are in a critical section
368 	     * on cpu #0 and can safely play with basetime_index.  Start
369 	     * with the current basetime and then make adjustments.
370 	     */
371 	    ni = (basetime_index + 1) & BASETIME_ARYMASK;
372 	    nbt = &basetime[ni];
373 	    *nbt = basetime[basetime_index];
374 
375 	    /*
376 	     * Apply adjtime corrections.  (adjtime() API)
377 	     *
378 	     * adjtime() only runs on cpu #0 so our critical section is
379 	     * sufficient to access these variables.
380 	     */
381 	    if (ntp_delta != 0) {
382 		nbt->tv_nsec += ntp_tick_delta;
383 		ntp_delta -= ntp_tick_delta;
384 		if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
385 		    (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
386 			ntp_tick_delta = ntp_delta;
387  		}
388  	    }
389 
390 	    /*
391 	     * Apply permanent frequency corrections.  (sysctl API)
392 	     */
393 	    if (ntp_tick_permanent != 0) {
394 		ntp_tick_acc += ntp_tick_permanent;
395 		if (ntp_tick_acc >= (1LL << 32)) {
396 		    nbt->tv_nsec += ntp_tick_acc >> 32;
397 		    ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
398 		} else if (ntp_tick_acc <= -(1LL << 32)) {
399 		    /* Negate ntp_tick_acc to avoid shifting the sign bit. */
400 		    nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
401 		    ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
402 		}
403  	    }
404 
405 	    if (nbt->tv_nsec >= 1000000000) {
406 		    nbt->tv_sec++;
407 		    nbt->tv_nsec -= 1000000000;
408 	    } else if (nbt->tv_nsec < 0) {
409 		    nbt->tv_sec--;
410 		    nbt->tv_nsec += 1000000000;
411 	    }
412 
413 	    /*
414 	     * Another per-tick compensation.  (for ntp_adjtime() API)
415 	     */
416 	    if (nsec_adj != 0) {
417 		nsec_acc += nsec_adj;
418 		if (nsec_acc >= 0x100000000LL) {
419 		    nbt->tv_nsec += nsec_acc >> 32;
420 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
421 		} else if (nsec_acc <= -0x100000000LL) {
422 		    nbt->tv_nsec -= -nsec_acc >> 32;
423 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
424 		}
425 		if (nbt->tv_nsec >= 1000000000) {
426 		    nbt->tv_nsec -= 1000000000;
427 		    ++nbt->tv_sec;
428 		} else if (nbt->tv_nsec < 0) {
429 		    nbt->tv_nsec += 1000000000;
430 		    --nbt->tv_sec;
431 		}
432 	    }
433 
434 	    /************************************************************
435 	     *			LEAP SECOND CORRECTION			*
436 	     ************************************************************
437 	     *
438 	     * Taking into account all the corrections made above, figure
439 	     * out the new real time.  If the seconds field has changed
440 	     * then apply any pending leap-second corrections.
441 	     */
442 	    getnanotime_nbt(nbt, &nts);
443 
444 	    if (time_second != nts.tv_sec) {
445 		/*
446 		 * Apply leap second (sysctl API).  Adjust nts for changes
447 		 * so we do not have to call getnanotime_nbt again.
448 		 */
449 		if (ntp_leap_second) {
450 		    if (ntp_leap_second == nts.tv_sec) {
451 			if (ntp_leap_insert) {
452 			    nbt->tv_sec++;
453 			    nts.tv_sec++;
454 			} else {
455 			    nbt->tv_sec--;
456 			    nts.tv_sec--;
457 			}
458 			ntp_leap_second--;
459 		    }
460 		}
461 
462 		/*
463 		 * Apply leap second (ntp_adjtime() API), calculate a new
464 		 * nsec_adj field.  ntp_update_second() returns nsec_adj
465 		 * as a per-second value but we need it as a per-tick value.
466 		 */
467 		leap = ntp_update_second(time_second, &nsec_adj);
468 		nsec_adj /= hz;
469 		nbt->tv_sec += leap;
470 		nts.tv_sec += leap;
471 
472 		/*
473 		 * Update the time_second 'approximate time' global.
474 		 */
475 		time_second = nts.tv_sec;
476 	    }
477 
478 	    /*
479 	     * Finally, our new basetime is ready to go live!
480 	     */
481 	    cpu_mb1();
482 	    basetime_index = ni;
483 	}
484 
485 	/*
486 	 * softticks are handled for all cpus
487 	 */
488 	hardclock_softtick(gd);
489 
490 	/*
491 	 * ITimer handling is per-tick, per-cpu.  I don't think psignal()
492 	 * is mpsafe on curproc, so XXX get the mplock.
493 	 */
494 	if ((p = curproc) != NULL && try_mplock()) {
495 		pstats = p->p_stats;
496 		if (frame && CLKF_USERMODE(frame) &&
497 		    timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
498 		    itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
499 			psignal(p, SIGVTALRM);
500 		if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
501 		    itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
502 			psignal(p, SIGPROF);
503 		rel_mplock();
504 	}
505 	setdelayed();
506 }
507 
508 /*
509  * The statistics clock typically runs at a 125Hz rate, and is intended
510  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
511  *
512  * NOTE! systimer! the MP lock might not be held here.  We can only safely
513  * manipulate objects owned by the current cpu.
514  *
515  * The stats clock is responsible for grabbing a profiling sample.
516  * Most of the statistics are only used by user-level statistics programs.
517  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
518  * p->p_estcpu.
519  *
520  * Like the other clocks, the stat clock is called from what is effectively
521  * a fast interrupt, so the context should be the thread/process that got
522  * interrupted.
523  */
524 static void
525 statclock(systimer_t info, struct intrframe *frame)
526 {
527 #ifdef GPROF
528 	struct gmonparam *g;
529 	int i;
530 #endif
531 	thread_t td;
532 	struct proc *p;
533 	int bump;
534 	struct timeval tv;
535 	struct timeval *stv;
536 
537 	/*
538 	 * How big was our timeslice relative to the last time?
539 	 */
540 	microuptime(&tv);	/* mpsafe */
541 	stv = &mycpu->gd_stattv;
542 	if (stv->tv_sec == 0) {
543 	    bump = 1;
544 	} else {
545 	    bump = tv.tv_usec - stv->tv_usec +
546 		(tv.tv_sec - stv->tv_sec) * 1000000;
547 	    if (bump < 0)
548 		bump = 0;
549 	    if (bump > 1000000)
550 		bump = 1000000;
551 	}
552 	*stv = tv;
553 
554 	td = curthread;
555 	p = td->td_proc;
556 
557 	if (frame && CLKF_USERMODE(frame)) {
558 		/*
559 		 * Came from userland, handle user time and deal with
560 		 * possible process.
561 		 */
562 		if (p && (p->p_flag & P_PROFIL))
563 			addupc_intr(p, CLKF_PC(frame), 1);
564 		td->td_uticks += bump;
565 
566 		/*
567 		 * Charge the time as appropriate
568 		 */
569 		if (p && p->p_nice > NZERO)
570 			cpu_time.cp_nice += bump;
571 		else
572 			cpu_time.cp_user += bump;
573 	} else {
574 #ifdef GPROF
575 		/*
576 		 * Kernel statistics are just like addupc_intr, only easier.
577 		 */
578 		g = &_gmonparam;
579 		if (g->state == GMON_PROF_ON && frame) {
580 			i = CLKF_PC(frame) - g->lowpc;
581 			if (i < g->textsize) {
582 				i /= HISTFRACTION * sizeof(*g->kcount);
583 				g->kcount[i]++;
584 			}
585 		}
586 #endif
587 		/*
588 		 * Came from kernel mode, so we were:
589 		 * - handling an interrupt,
590 		 * - doing syscall or trap work on behalf of the current
591 		 *   user process, or
592 		 * - spinning in the idle loop.
593 		 * Whichever it is, charge the time as appropriate.
594 		 * Note that we charge interrupts to the current process,
595 		 * regardless of whether they are ``for'' that process,
596 		 * so that we know how much of its real time was spent
597 		 * in ``non-process'' (i.e., interrupt) work.
598 		 *
599 		 * XXX assume system if frame is NULL.  A NULL frame
600 		 * can occur if ipi processing is done from an splx().
601 		 */
602 		if (frame && CLKF_INTR(frame))
603 			td->td_iticks += bump;
604 		else
605 			td->td_sticks += bump;
606 
607 		if (frame && CLKF_INTR(frame)) {
608 			cpu_time.cp_intr += bump;
609 		} else {
610 			if (td == &mycpu->gd_idlethread)
611 				cpu_time.cp_idle += bump;
612 			else
613 				cpu_time.cp_sys += bump;
614 		}
615 	}
616 }
617 
618 /*
619  * The scheduler clock typically runs at a 20Hz rate.  NOTE! systimer,
620  * the MP lock might not be held.  We can safely manipulate parts of curproc
621  * but that's about it.
622  */
623 static void
624 schedclock(systimer_t info, struct intrframe *frame)
625 {
626 	struct proc *p;
627 	struct pstats *pstats;
628 	struct rusage *ru;
629 	struct vmspace *vm;
630 	long rss;
631 
632 	schedulerclock(NULL);	/* mpsafe */
633 	if ((p = curproc) != NULL) {
634 		/* Update resource usage integrals and maximums. */
635 		if ((pstats = p->p_stats) != NULL &&
636 		    (ru = &pstats->p_ru) != NULL &&
637 		    (vm = p->p_vmspace) != NULL) {
638 			ru->ru_ixrss += pgtok(vm->vm_tsize);
639 			ru->ru_idrss += pgtok(vm->vm_dsize);
640 			ru->ru_isrss += pgtok(vm->vm_ssize);
641 			rss = pgtok(vmspace_resident_count(vm));
642 			if (ru->ru_maxrss < rss)
643 				ru->ru_maxrss = rss;
644 		}
645 	}
646 }
647 
648 /*
649  * Compute number of ticks for the specified amount of time.  The
650  * return value is intended to be used in a clock interrupt timed
651  * operation and guarenteed to meet or exceed the requested time.
652  * If the representation overflows, return INT_MAX.  The minimum return
653  * value is 1 ticks and the function will average the calculation up.
654  * If any value greater then 0 microseconds is supplied, a value
655  * of at least 2 will be returned to ensure that a near-term clock
656  * interrupt does not cause the timeout to occur (degenerately) early.
657  *
658  * Note that limit checks must take into account microseconds, which is
659  * done simply by using the smaller signed long maximum instead of
660  * the unsigned long maximum.
661  *
662  * If ints have 32 bits, then the maximum value for any timeout in
663  * 10ms ticks is 248 days.
664  */
665 int
666 tvtohz_high(struct timeval *tv)
667 {
668 	int ticks;
669 	long sec, usec;
670 
671 	sec = tv->tv_sec;
672 	usec = tv->tv_usec;
673 	if (usec < 0) {
674 		sec--;
675 		usec += 1000000;
676 	}
677 	if (sec < 0) {
678 #ifdef DIAGNOSTIC
679 		if (usec > 0) {
680 			sec++;
681 			usec -= 1000000;
682 		}
683 		printf("tvotohz: negative time difference %ld sec %ld usec\n",
684 		       sec, usec);
685 #endif
686 		ticks = 1;
687 	} else if (sec <= INT_MAX / hz) {
688 		ticks = (int)(sec * hz +
689 			    ((u_long)usec + (tick - 1)) / tick) + 1;
690 	} else {
691 		ticks = INT_MAX;
692 	}
693 	return (ticks);
694 }
695 
696 /*
697  * Compute number of ticks for the specified amount of time, erroring on
698  * the side of it being too low to ensure that sleeping the returned number
699  * of ticks will not result in a late return.
700  *
701  * The supplied timeval may not be negative and should be normalized.  A
702  * return value of 0 is possible if the timeval converts to less then
703  * 1 tick.
704  *
705  * If ints have 32 bits, then the maximum value for any timeout in
706  * 10ms ticks is 248 days.
707  */
708 int
709 tvtohz_low(struct timeval *tv)
710 {
711 	int ticks;
712 	long sec;
713 
714 	sec = tv->tv_sec;
715 	if (sec <= INT_MAX / hz)
716 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
717 	else
718 		ticks = INT_MAX;
719 	return (ticks);
720 }
721 
722 
723 /*
724  * Start profiling on a process.
725  *
726  * Kernel profiling passes proc0 which never exits and hence
727  * keeps the profile clock running constantly.
728  */
729 void
730 startprofclock(struct proc *p)
731 {
732 	if ((p->p_flag & P_PROFIL) == 0) {
733 		p->p_flag |= P_PROFIL;
734 #if 0	/* XXX */
735 		if (++profprocs == 1 && stathz != 0) {
736 			s = splstatclock();
737 			psdiv = psratio;
738 			setstatclockrate(profhz);
739 			splx(s);
740 		}
741 #endif
742 	}
743 }
744 
745 /*
746  * Stop profiling on a process.
747  */
748 void
749 stopprofclock(struct proc *p)
750 {
751 	if (p->p_flag & P_PROFIL) {
752 		p->p_flag &= ~P_PROFIL;
753 #if 0	/* XXX */
754 		if (--profprocs == 0 && stathz != 0) {
755 			s = splstatclock();
756 			psdiv = 1;
757 			setstatclockrate(stathz);
758 			splx(s);
759 		}
760 #endif
761 	}
762 }
763 
764 /*
765  * Return information about system clocks.
766  */
767 static int
768 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
769 {
770 	struct kinfo_clockinfo clkinfo;
771 	/*
772 	 * Construct clockinfo structure.
773 	 */
774 	clkinfo.ci_hz = hz;
775 	clkinfo.ci_tick = tick;
776 	clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
777 	clkinfo.ci_profhz = profhz;
778 	clkinfo.ci_stathz = stathz ? stathz : hz;
779 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
780 }
781 
782 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
783 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
784 
785 /*
786  * We have eight functions for looking at the clock, four for
787  * microseconds and four for nanoseconds.  For each there is fast
788  * but less precise version "get{nano|micro}[up]time" which will
789  * return a time which is up to 1/HZ previous to the call, whereas
790  * the raw version "{nano|micro}[up]time" will return a timestamp
791  * which is as precise as possible.  The "up" variants return the
792  * time relative to system boot, these are well suited for time
793  * interval measurements.
794  *
795  * Each cpu independantly maintains the current time of day, so all
796  * we need to do to protect ourselves from changes is to do a loop
797  * check on the seconds field changing out from under us.
798  *
799  * The system timer maintains a 32 bit count and due to various issues
800  * it is possible for the calculated delta to occassionally exceed
801  * cputimer_freq.  If this occurs the cputimer_freq64_nsec multiplication
802  * can easily overflow, so we deal with the case.  For uniformity we deal
803  * with the case in the usec case too.
804  */
805 void
806 getmicrouptime(struct timeval *tvp)
807 {
808 	struct globaldata *gd = mycpu;
809 	sysclock_t delta;
810 
811 	do {
812 		tvp->tv_sec = gd->gd_time_seconds;
813 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
814 	} while (tvp->tv_sec != gd->gd_time_seconds);
815 
816 	if (delta >= cputimer_freq) {
817 		tvp->tv_sec += delta / cputimer_freq;
818 		delta %= cputimer_freq;
819 	}
820 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
821 	if (tvp->tv_usec >= 1000000) {
822 		tvp->tv_usec -= 1000000;
823 		++tvp->tv_sec;
824 	}
825 }
826 
827 void
828 getnanouptime(struct timespec *tsp)
829 {
830 	struct globaldata *gd = mycpu;
831 	sysclock_t delta;
832 
833 	do {
834 		tsp->tv_sec = gd->gd_time_seconds;
835 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
836 	} while (tsp->tv_sec != gd->gd_time_seconds);
837 
838 	if (delta >= cputimer_freq) {
839 		tsp->tv_sec += delta / cputimer_freq;
840 		delta %= cputimer_freq;
841 	}
842 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
843 }
844 
845 void
846 microuptime(struct timeval *tvp)
847 {
848 	struct globaldata *gd = mycpu;
849 	sysclock_t delta;
850 
851 	do {
852 		tvp->tv_sec = gd->gd_time_seconds;
853 		delta = cputimer_count() - gd->gd_cpuclock_base;
854 	} while (tvp->tv_sec != gd->gd_time_seconds);
855 
856 	if (delta >= cputimer_freq) {
857 		tvp->tv_sec += delta / cputimer_freq;
858 		delta %= cputimer_freq;
859 	}
860 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
861 }
862 
863 void
864 nanouptime(struct timespec *tsp)
865 {
866 	struct globaldata *gd = mycpu;
867 	sysclock_t delta;
868 
869 	do {
870 		tsp->tv_sec = gd->gd_time_seconds;
871 		delta = cputimer_count() - gd->gd_cpuclock_base;
872 	} while (tsp->tv_sec != gd->gd_time_seconds);
873 
874 	if (delta >= cputimer_freq) {
875 		tsp->tv_sec += delta / cputimer_freq;
876 		delta %= cputimer_freq;
877 	}
878 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
879 }
880 
881 /*
882  * realtime routines
883  */
884 
885 void
886 getmicrotime(struct timeval *tvp)
887 {
888 	struct globaldata *gd = mycpu;
889 	struct timespec *bt;
890 	sysclock_t delta;
891 
892 	do {
893 		tvp->tv_sec = gd->gd_time_seconds;
894 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
895 	} while (tvp->tv_sec != gd->gd_time_seconds);
896 
897 	if (delta >= cputimer_freq) {
898 		tvp->tv_sec += delta / cputimer_freq;
899 		delta %= cputimer_freq;
900 	}
901 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
902 
903 	bt = &basetime[basetime_index];
904 	tvp->tv_sec += bt->tv_sec;
905 	tvp->tv_usec += bt->tv_nsec / 1000;
906 	while (tvp->tv_usec >= 1000000) {
907 		tvp->tv_usec -= 1000000;
908 		++tvp->tv_sec;
909 	}
910 }
911 
912 void
913 getnanotime(struct timespec *tsp)
914 {
915 	struct globaldata *gd = mycpu;
916 	struct timespec *bt;
917 	sysclock_t delta;
918 
919 	do {
920 		tsp->tv_sec = gd->gd_time_seconds;
921 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
922 	} while (tsp->tv_sec != gd->gd_time_seconds);
923 
924 	if (delta >= cputimer_freq) {
925 		tsp->tv_sec += delta / cputimer_freq;
926 		delta %= cputimer_freq;
927 	}
928 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
929 
930 	bt = &basetime[basetime_index];
931 	tsp->tv_sec += bt->tv_sec;
932 	tsp->tv_nsec += bt->tv_nsec;
933 	while (tsp->tv_nsec >= 1000000000) {
934 		tsp->tv_nsec -= 1000000000;
935 		++tsp->tv_sec;
936 	}
937 }
938 
939 static void
940 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
941 {
942 	struct globaldata *gd = mycpu;
943 	sysclock_t delta;
944 
945 	do {
946 		tsp->tv_sec = gd->gd_time_seconds;
947 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
948 	} while (tsp->tv_sec != gd->gd_time_seconds);
949 
950 	if (delta >= cputimer_freq) {
951 		tsp->tv_sec += delta / cputimer_freq;
952 		delta %= cputimer_freq;
953 	}
954 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
955 
956 	tsp->tv_sec += nbt->tv_sec;
957 	tsp->tv_nsec += nbt->tv_nsec;
958 	while (tsp->tv_nsec >= 1000000000) {
959 		tsp->tv_nsec -= 1000000000;
960 		++tsp->tv_sec;
961 	}
962 }
963 
964 
965 void
966 microtime(struct timeval *tvp)
967 {
968 	struct globaldata *gd = mycpu;
969 	struct timespec *bt;
970 	sysclock_t delta;
971 
972 	do {
973 		tvp->tv_sec = gd->gd_time_seconds;
974 		delta = cputimer_count() - gd->gd_cpuclock_base;
975 	} while (tvp->tv_sec != gd->gd_time_seconds);
976 
977 	if (delta >= cputimer_freq) {
978 		tvp->tv_sec += delta / cputimer_freq;
979 		delta %= cputimer_freq;
980 	}
981 	tvp->tv_usec = (cputimer_freq64_usec * delta) >> 32;
982 
983 	bt = &basetime[basetime_index];
984 	tvp->tv_sec += bt->tv_sec;
985 	tvp->tv_usec += bt->tv_nsec / 1000;
986 	while (tvp->tv_usec >= 1000000) {
987 		tvp->tv_usec -= 1000000;
988 		++tvp->tv_sec;
989 	}
990 }
991 
992 void
993 nanotime(struct timespec *tsp)
994 {
995 	struct globaldata *gd = mycpu;
996 	struct timespec *bt;
997 	sysclock_t delta;
998 
999 	do {
1000 		tsp->tv_sec = gd->gd_time_seconds;
1001 		delta = cputimer_count() - gd->gd_cpuclock_base;
1002 	} while (tsp->tv_sec != gd->gd_time_seconds);
1003 
1004 	if (delta >= cputimer_freq) {
1005 		tsp->tv_sec += delta / cputimer_freq;
1006 		delta %= cputimer_freq;
1007 	}
1008 	tsp->tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
1009 
1010 	bt = &basetime[basetime_index];
1011 	tsp->tv_sec += bt->tv_sec;
1012 	tsp->tv_nsec += bt->tv_nsec;
1013 	while (tsp->tv_nsec >= 1000000000) {
1014 		tsp->tv_nsec -= 1000000000;
1015 		++tsp->tv_sec;
1016 	}
1017 }
1018 
1019 /*
1020  * note: this is not exactly synchronized with real time.  To do that we
1021  * would have to do what microtime does and check for a nanoseconds overflow.
1022  */
1023 time_t
1024 get_approximate_time_t(void)
1025 {
1026 	struct globaldata *gd = mycpu;
1027 	struct timespec *bt;
1028 
1029 	bt = &basetime[basetime_index];
1030 	return(gd->gd_time_seconds + bt->tv_sec);
1031 }
1032 
1033 int
1034 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1035 {
1036 	pps_params_t *app;
1037 	struct pps_fetch_args *fapi;
1038 #ifdef PPS_SYNC
1039 	struct pps_kcbind_args *kapi;
1040 #endif
1041 
1042 	switch (cmd) {
1043 	case PPS_IOC_CREATE:
1044 		return (0);
1045 	case PPS_IOC_DESTROY:
1046 		return (0);
1047 	case PPS_IOC_SETPARAMS:
1048 		app = (pps_params_t *)data;
1049 		if (app->mode & ~pps->ppscap)
1050 			return (EINVAL);
1051 		pps->ppsparam = *app;
1052 		return (0);
1053 	case PPS_IOC_GETPARAMS:
1054 		app = (pps_params_t *)data;
1055 		*app = pps->ppsparam;
1056 		app->api_version = PPS_API_VERS_1;
1057 		return (0);
1058 	case PPS_IOC_GETCAP:
1059 		*(int*)data = pps->ppscap;
1060 		return (0);
1061 	case PPS_IOC_FETCH:
1062 		fapi = (struct pps_fetch_args *)data;
1063 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1064 			return (EINVAL);
1065 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1066 			return (EOPNOTSUPP);
1067 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
1068 		fapi->pps_info_buf = pps->ppsinfo;
1069 		return (0);
1070 	case PPS_IOC_KCBIND:
1071 #ifdef PPS_SYNC
1072 		kapi = (struct pps_kcbind_args *)data;
1073 		/* XXX Only root should be able to do this */
1074 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1075 			return (EINVAL);
1076 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1077 			return (EINVAL);
1078 		if (kapi->edge & ~pps->ppscap)
1079 			return (EINVAL);
1080 		pps->kcmode = kapi->edge;
1081 		return (0);
1082 #else
1083 		return (EOPNOTSUPP);
1084 #endif
1085 	default:
1086 		return (ENOTTY);
1087 	}
1088 }
1089 
1090 void
1091 pps_init(struct pps_state *pps)
1092 {
1093 	pps->ppscap |= PPS_TSFMT_TSPEC;
1094 	if (pps->ppscap & PPS_CAPTUREASSERT)
1095 		pps->ppscap |= PPS_OFFSETASSERT;
1096 	if (pps->ppscap & PPS_CAPTURECLEAR)
1097 		pps->ppscap |= PPS_OFFSETCLEAR;
1098 }
1099 
1100 void
1101 pps_event(struct pps_state *pps, sysclock_t count, int event)
1102 {
1103 	struct globaldata *gd;
1104 	struct timespec *tsp;
1105 	struct timespec *osp;
1106 	struct timespec *bt;
1107 	struct timespec ts;
1108 	sysclock_t *pcount;
1109 #ifdef PPS_SYNC
1110 	sysclock_t tcount;
1111 #endif
1112 	sysclock_t delta;
1113 	pps_seq_t *pseq;
1114 	int foff;
1115 	int fhard;
1116 
1117 	gd = mycpu;
1118 
1119 	/* Things would be easier with arrays... */
1120 	if (event == PPS_CAPTUREASSERT) {
1121 		tsp = &pps->ppsinfo.assert_timestamp;
1122 		osp = &pps->ppsparam.assert_offset;
1123 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1124 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1125 		pcount = &pps->ppscount[0];
1126 		pseq = &pps->ppsinfo.assert_sequence;
1127 	} else {
1128 		tsp = &pps->ppsinfo.clear_timestamp;
1129 		osp = &pps->ppsparam.clear_offset;
1130 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1131 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1132 		pcount = &pps->ppscount[1];
1133 		pseq = &pps->ppsinfo.clear_sequence;
1134 	}
1135 
1136 	/* Nothing really happened */
1137 	if (*pcount == count)
1138 		return;
1139 
1140 	*pcount = count;
1141 
1142 	do {
1143 		ts.tv_sec = gd->gd_time_seconds;
1144 		delta = count - gd->gd_cpuclock_base;
1145 	} while (ts.tv_sec != gd->gd_time_seconds);
1146 
1147 	if (delta >= cputimer_freq) {
1148 		ts.tv_sec += delta / cputimer_freq;
1149 		delta %= cputimer_freq;
1150 	}
1151 	ts.tv_nsec = (cputimer_freq64_nsec * delta) >> 32;
1152 	bt = &basetime[basetime_index];
1153 	ts.tv_sec += bt->tv_sec;
1154 	ts.tv_nsec += bt->tv_nsec;
1155 	while (ts.tv_nsec >= 1000000000) {
1156 		ts.tv_nsec -= 1000000000;
1157 		++ts.tv_sec;
1158 	}
1159 
1160 	(*pseq)++;
1161 	*tsp = ts;
1162 
1163 	if (foff) {
1164 		timespecadd(tsp, osp);
1165 		if (tsp->tv_nsec < 0) {
1166 			tsp->tv_nsec += 1000000000;
1167 			tsp->tv_sec -= 1;
1168 		}
1169 	}
1170 #ifdef PPS_SYNC
1171 	if (fhard) {
1172 		/* magic, at its best... */
1173 		tcount = count - pps->ppscount[2];
1174 		pps->ppscount[2] = count;
1175 		if (tcount >= cputimer_freq) {
1176 			delta = (1000000000 * (tcount / cputimer_freq) +
1177 				 cputimer_freq64_nsec *
1178 				 (tcount % cputimer_freq)) >> 32;
1179 		} else {
1180 			delta = (cputimer_freq64_nsec * tcount) >> 32;
1181 		}
1182 		hardpps(tsp, delta);
1183 	}
1184 #endif
1185 }
1186 
1187