xref: /dragonfly/sys/kern/kern_clock.c (revision 6e278935)
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 1.62 2008/09/09 04:06:13 dillon Exp $
74  */
75 
76 #include "opt_ntp.h"
77 #include "opt_polling.h"
78 #include "opt_ifpoll.h"
79 #include "opt_pctrack.h"
80 
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/callout.h>
84 #include <sys/kernel.h>
85 #include <sys/kinfo.h>
86 #include <sys/proc.h>
87 #include <sys/malloc.h>
88 #include <sys/resourcevar.h>
89 #include <sys/signalvar.h>
90 #include <sys/timex.h>
91 #include <sys/timepps.h>
92 #include <vm/vm.h>
93 #include <sys/lock.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_extern.h>
97 #include <sys/sysctl.h>
98 
99 #include <sys/thread2.h>
100 
101 #include <machine/cpu.h>
102 #include <machine/limits.h>
103 #include <machine/smp.h>
104 #include <machine/cpufunc.h>
105 #include <machine/specialreg.h>
106 #include <machine/clock.h>
107 
108 #ifdef GPROF
109 #include <sys/gmon.h>
110 #endif
111 
112 #ifdef DEVICE_POLLING
113 extern void init_device_poll_pcpu(int);
114 #endif
115 
116 #ifdef IFPOLL_ENABLE
117 extern void ifpoll_init_pcpu(int);
118 #endif
119 
120 #ifdef DEBUG_PCTRACK
121 static void do_pctrack(struct intrframe *frame, int which);
122 #endif
123 
124 static void initclocks (void *dummy);
125 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
126 
127 /*
128  * Some of these don't belong here, but it's easiest to concentrate them.
129  * Note that cpu_time counts in microseconds, but most userland programs
130  * just compare relative times against the total by delta.
131  */
132 struct kinfo_cputime cputime_percpu[MAXCPU];
133 #ifdef DEBUG_PCTRACK
134 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
135 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
136 #endif
137 
138 #ifdef SMP
139 static int
140 sysctl_cputime(SYSCTL_HANDLER_ARGS)
141 {
142 	int cpu, error = 0;
143 	size_t size = sizeof(struct kinfo_cputime);
144 
145 	for (cpu = 0; cpu < ncpus; ++cpu) {
146 		if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
147 			break;
148 	}
149 
150 	return (error);
151 }
152 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
153 	sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
154 #else
155 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
156     "CPU time statistics");
157 #endif
158 
159 static int
160 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
161 {
162 	long cpu_states[5] = {0};
163 	int cpu, error = 0;
164 	size_t size = sizeof(cpu_states);
165 
166 	for (cpu = 0; cpu < ncpus; ++cpu) {
167 		cpu_states[0] += cputime_percpu[cpu].cp_user;
168 		cpu_states[1] += cputime_percpu[cpu].cp_nice;
169 		cpu_states[2] += cputime_percpu[cpu].cp_sys;
170 		cpu_states[3] += cputime_percpu[cpu].cp_intr;
171 		cpu_states[4] += cputime_percpu[cpu].cp_idle;
172 	}
173 
174 	error = SYSCTL_OUT(req, cpu_states, size);
175 
176 	return (error);
177 }
178 
179 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
180 	sysctl_cp_time, "LU", "CPU time statistics");
181 
182 /*
183  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
184  * microuptime().  microtime() is not drift compensated.  The real uptime
185  * with compensation is nanotime() - bootime.  boottime is recalculated
186  * whenever the real time is set based on the compensated elapsed time
187  * in seconds (gd->gd_time_seconds).
188  *
189  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
190  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
191  * the real time.
192  */
193 struct timespec boottime;	/* boot time (realtime) for reference only */
194 time_t time_second;		/* read-only 'passive' uptime in seconds */
195 
196 /*
197  * basetime is used to calculate the compensated real time of day.  The
198  * basetime can be modified on a per-tick basis by the adjtime(),
199  * ntp_adjtime(), and sysctl-based time correction APIs.
200  *
201  * Note that frequency corrections can also be made by adjusting
202  * gd_cpuclock_base.
203  *
204  * basetime is a tail-chasing FIFO, updated only by cpu #0.  The FIFO is
205  * used on both SMP and UP systems to avoid MP races between cpu's and
206  * interrupt races on UP systems.
207  */
208 #define BASETIME_ARYSIZE	16
209 #define BASETIME_ARYMASK	(BASETIME_ARYSIZE - 1)
210 static struct timespec basetime[BASETIME_ARYSIZE];
211 static volatile int basetime_index;
212 
213 static int
214 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
215 {
216 	struct timespec *bt;
217 	int error;
218 	int index;
219 
220 	/*
221 	 * Because basetime data and index may be updated by another cpu,
222 	 * a load fence is required to ensure that the data we read has
223 	 * not been speculatively read relative to a possibly updated index.
224 	 */
225 	index = basetime_index;
226 	cpu_lfence();
227 	bt = &basetime[index];
228 	error = SYSCTL_OUT(req, bt, sizeof(*bt));
229 	return (error);
230 }
231 
232 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
233     &boottime, timespec, "System boottime");
234 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
235     sysctl_get_basetime, "S,timespec", "System basetime");
236 
237 static void hardclock(systimer_t info, int, struct intrframe *frame);
238 static void statclock(systimer_t info, int, struct intrframe *frame);
239 static void schedclock(systimer_t info, int, struct intrframe *frame);
240 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
241 
242 int	ticks;			/* system master ticks at hz */
243 int	clocks_running;		/* tsleep/timeout clocks operational */
244 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
245 int64_t	nsec_acc;		/* accumulator */
246 
247 /* NTPD time correction fields */
248 int64_t	ntp_tick_permanent;	/* per-tick adjustment in nsec << 32 */
249 int64_t	ntp_tick_acc;		/* accumulator for per-tick adjustment */
250 int64_t	ntp_delta;		/* one-time correction in nsec */
251 int64_t ntp_big_delta = 1000000000;
252 int32_t	ntp_tick_delta;		/* current adjustment rate */
253 int32_t	ntp_default_tick_delta;	/* adjustment rate for ntp_delta */
254 time_t	ntp_leap_second;	/* time of next leap second */
255 int	ntp_leap_insert;	/* whether to insert or remove a second */
256 
257 /*
258  * Finish initializing clock frequencies and start all clocks running.
259  */
260 /* ARGSUSED*/
261 static void
262 initclocks(void *dummy)
263 {
264 	/*psratio = profhz / stathz;*/
265 	initclocks_pcpu();
266 	clocks_running = 1;
267 }
268 
269 /*
270  * Called on a per-cpu basis
271  */
272 void
273 initclocks_pcpu(void)
274 {
275 	struct globaldata *gd = mycpu;
276 
277 	crit_enter();
278 	if (gd->gd_cpuid == 0) {
279 	    gd->gd_time_seconds = 1;
280 	    gd->gd_cpuclock_base = sys_cputimer->count();
281 	} else {
282 	    /* XXX */
283 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
284 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
285 	}
286 
287 	systimer_intr_enable();
288 
289 #ifdef DEVICE_POLLING
290 	init_device_poll_pcpu(gd->gd_cpuid);
291 #endif
292 
293 #ifdef IFPOLL_ENABLE
294 	ifpoll_init_pcpu(gd->gd_cpuid);
295 #endif
296 
297 	/*
298 	 * Use a non-queued periodic systimer to prevent multiple ticks from
299 	 * building up if the sysclock jumps forward (8254 gets reset).  The
300 	 * sysclock will never jump backwards.  Our time sync is based on
301 	 * the actual sysclock, not the ticks count.
302 	 */
303 	systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
304 	systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
305 	/* XXX correct the frequency for scheduler / estcpu tests */
306 	systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
307 				NULL, ESTCPUFREQ);
308 	crit_exit();
309 }
310 
311 /*
312  * This sets the current real time of day.  Timespecs are in seconds and
313  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
314  * instead we adjust basetime so basetime + gd_* results in the current
315  * time of day.  This way the gd_* fields are guarenteed to represent
316  * a monotonically increasing 'uptime' value.
317  *
318  * When set_timeofday() is called from userland, the system call forces it
319  * onto cpu #0 since only cpu #0 can update basetime_index.
320  */
321 void
322 set_timeofday(struct timespec *ts)
323 {
324 	struct timespec *nbt;
325 	int ni;
326 
327 	/*
328 	 * XXX SMP / non-atomic basetime updates
329 	 */
330 	crit_enter();
331 	ni = (basetime_index + 1) & BASETIME_ARYMASK;
332 	nbt = &basetime[ni];
333 	nanouptime(nbt);
334 	nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
335 	nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
336 	if (nbt->tv_nsec < 0) {
337 	    nbt->tv_nsec += 1000000000;
338 	    --nbt->tv_sec;
339 	}
340 
341 	/*
342 	 * Note that basetime diverges from boottime as the clock drift is
343 	 * compensated for, so we cannot do away with boottime.  When setting
344 	 * the absolute time of day the drift is 0 (for an instant) and we
345 	 * can simply assign boottime to basetime.
346 	 *
347 	 * Note that nanouptime() is based on gd_time_seconds which is drift
348 	 * compensated up to a point (it is guarenteed to remain monotonically
349 	 * increasing).  gd_time_seconds is thus our best uptime guess and
350 	 * suitable for use in the boottime calculation.  It is already taken
351 	 * into account in the basetime calculation above.
352 	 */
353 	boottime.tv_sec = nbt->tv_sec;
354 	ntp_delta = 0;
355 
356 	/*
357 	 * We now have a new basetime, make sure all other cpus have it,
358 	 * then update the index.
359 	 */
360 	cpu_sfence();
361 	basetime_index = ni;
362 
363 	crit_exit();
364 }
365 
366 /*
367  * Each cpu has its own hardclock, but we only increments ticks and softticks
368  * on cpu #0.
369  *
370  * NOTE! systimer! the MP lock might not be held here.  We can only safely
371  * manipulate objects owned by the current cpu.
372  */
373 static void
374 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
375 {
376 	sysclock_t cputicks;
377 	struct proc *p;
378 	struct globaldata *gd = mycpu;
379 
380 	/*
381 	 * Realtime updates are per-cpu.  Note that timer corrections as
382 	 * returned by microtime() and friends make an additional adjustment
383 	 * using a system-wise 'basetime', but the running time is always
384 	 * taken from the per-cpu globaldata area.  Since the same clock
385 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
386 	 * stay in synch.
387 	 *
388 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
389 	 * to reverse index gd_cpuclock_base, but that it is possible for
390 	 * it to temporarily get behind in the seconds if something in the
391 	 * system locks interrupts for a long period of time.  Since periodic
392 	 * timers count events, though everything should resynch again
393 	 * immediately.
394 	 */
395 	cputicks = info->time - gd->gd_cpuclock_base;
396 	if (cputicks >= sys_cputimer->freq) {
397 		++gd->gd_time_seconds;
398 		gd->gd_cpuclock_base += sys_cputimer->freq;
399 	}
400 
401 	/*
402 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
403 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
404 	 * by updating basetime.
405 	 */
406 	if (gd->gd_cpuid == 0) {
407 	    struct timespec *nbt;
408 	    struct timespec nts;
409 	    int leap;
410 	    int ni;
411 
412 	    ++ticks;
413 
414 #if 0
415 	    if (tco->tc_poll_pps)
416 		tco->tc_poll_pps(tco);
417 #endif
418 
419 	    /*
420 	     * Calculate the new basetime index.  We are in a critical section
421 	     * on cpu #0 and can safely play with basetime_index.  Start
422 	     * with the current basetime and then make adjustments.
423 	     */
424 	    ni = (basetime_index + 1) & BASETIME_ARYMASK;
425 	    nbt = &basetime[ni];
426 	    *nbt = basetime[basetime_index];
427 
428 	    /*
429 	     * Apply adjtime corrections.  (adjtime() API)
430 	     *
431 	     * adjtime() only runs on cpu #0 so our critical section is
432 	     * sufficient to access these variables.
433 	     */
434 	    if (ntp_delta != 0) {
435 		nbt->tv_nsec += ntp_tick_delta;
436 		ntp_delta -= ntp_tick_delta;
437 		if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
438 		    (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
439 			ntp_tick_delta = ntp_delta;
440  		}
441  	    }
442 
443 	    /*
444 	     * Apply permanent frequency corrections.  (sysctl API)
445 	     */
446 	    if (ntp_tick_permanent != 0) {
447 		ntp_tick_acc += ntp_tick_permanent;
448 		if (ntp_tick_acc >= (1LL << 32)) {
449 		    nbt->tv_nsec += ntp_tick_acc >> 32;
450 		    ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
451 		} else if (ntp_tick_acc <= -(1LL << 32)) {
452 		    /* Negate ntp_tick_acc to avoid shifting the sign bit. */
453 		    nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
454 		    ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
455 		}
456  	    }
457 
458 	    if (nbt->tv_nsec >= 1000000000) {
459 		    nbt->tv_sec++;
460 		    nbt->tv_nsec -= 1000000000;
461 	    } else if (nbt->tv_nsec < 0) {
462 		    nbt->tv_sec--;
463 		    nbt->tv_nsec += 1000000000;
464 	    }
465 
466 	    /*
467 	     * Another per-tick compensation.  (for ntp_adjtime() API)
468 	     */
469 	    if (nsec_adj != 0) {
470 		nsec_acc += nsec_adj;
471 		if (nsec_acc >= 0x100000000LL) {
472 		    nbt->tv_nsec += nsec_acc >> 32;
473 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
474 		} else if (nsec_acc <= -0x100000000LL) {
475 		    nbt->tv_nsec -= -nsec_acc >> 32;
476 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
477 		}
478 		if (nbt->tv_nsec >= 1000000000) {
479 		    nbt->tv_nsec -= 1000000000;
480 		    ++nbt->tv_sec;
481 		} else if (nbt->tv_nsec < 0) {
482 		    nbt->tv_nsec += 1000000000;
483 		    --nbt->tv_sec;
484 		}
485 	    }
486 
487 	    /************************************************************
488 	     *			LEAP SECOND CORRECTION			*
489 	     ************************************************************
490 	     *
491 	     * Taking into account all the corrections made above, figure
492 	     * out the new real time.  If the seconds field has changed
493 	     * then apply any pending leap-second corrections.
494 	     */
495 	    getnanotime_nbt(nbt, &nts);
496 
497 	    if (time_second != nts.tv_sec) {
498 		/*
499 		 * Apply leap second (sysctl API).  Adjust nts for changes
500 		 * so we do not have to call getnanotime_nbt again.
501 		 */
502 		if (ntp_leap_second) {
503 		    if (ntp_leap_second == nts.tv_sec) {
504 			if (ntp_leap_insert) {
505 			    nbt->tv_sec++;
506 			    nts.tv_sec++;
507 			} else {
508 			    nbt->tv_sec--;
509 			    nts.tv_sec--;
510 			}
511 			ntp_leap_second--;
512 		    }
513 		}
514 
515 		/*
516 		 * Apply leap second (ntp_adjtime() API), calculate a new
517 		 * nsec_adj field.  ntp_update_second() returns nsec_adj
518 		 * as a per-second value but we need it as a per-tick value.
519 		 */
520 		leap = ntp_update_second(time_second, &nsec_adj);
521 		nsec_adj /= hz;
522 		nbt->tv_sec += leap;
523 		nts.tv_sec += leap;
524 
525 		/*
526 		 * Update the time_second 'approximate time' global.
527 		 */
528 		time_second = nts.tv_sec;
529 	    }
530 
531 	    /*
532 	     * Finally, our new basetime is ready to go live!
533 	     */
534 	    cpu_sfence();
535 	    basetime_index = ni;
536 	}
537 
538 	/*
539 	 * lwkt thread scheduler fair queueing
540 	 */
541 	lwkt_schedulerclock(curthread);
542 
543 	/*
544 	 * softticks are handled for all cpus
545 	 */
546 	hardclock_softtick(gd);
547 
548 	/*
549 	 * ITimer handling is per-tick, per-cpu.
550 	 *
551 	 * We must acquire the per-process token in order for ksignal()
552 	 * to be non-blocking.  For the moment this requires an AST fault,
553 	 * the ksignal() cannot be safely issued from this hard interrupt.
554 	 *
555 	 * XXX Even the trytoken here isn't right, and itimer operation in
556 	 *     a multi threaded environment is going to be weird at the
557 	 *     very least.
558 	 */
559 	if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
560 		crit_enter_hard();
561 		if (frame && CLKF_USERMODE(frame) &&
562 		    timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
563 		    itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
564 			p->p_flags |= P_SIGVTALRM;
565 			need_user_resched();
566 		}
567 		if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
568 		    itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
569 			p->p_flags |= P_SIGPROF;
570 			need_user_resched();
571 		}
572 		crit_exit_hard();
573 		lwkt_reltoken(&p->p_token);
574 	}
575 	setdelayed();
576 }
577 
578 /*
579  * The statistics clock typically runs at a 125Hz rate, and is intended
580  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
581  *
582  * NOTE! systimer! the MP lock might not be held here.  We can only safely
583  * manipulate objects owned by the current cpu.
584  *
585  * The stats clock is responsible for grabbing a profiling sample.
586  * Most of the statistics are only used by user-level statistics programs.
587  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
588  * p->p_estcpu.
589  *
590  * Like the other clocks, the stat clock is called from what is effectively
591  * a fast interrupt, so the context should be the thread/process that got
592  * interrupted.
593  */
594 static void
595 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
596 {
597 #ifdef GPROF
598 	struct gmonparam *g;
599 	int i;
600 #endif
601 	thread_t td;
602 	struct proc *p;
603 	int bump;
604 	struct timeval tv;
605 	struct timeval *stv;
606 
607 	/*
608 	 * How big was our timeslice relative to the last time?
609 	 */
610 	microuptime(&tv);	/* mpsafe */
611 	stv = &mycpu->gd_stattv;
612 	if (stv->tv_sec == 0) {
613 	    bump = 1;
614 	} else {
615 	    bump = tv.tv_usec - stv->tv_usec +
616 		(tv.tv_sec - stv->tv_sec) * 1000000;
617 	    if (bump < 0)
618 		bump = 0;
619 	    if (bump > 1000000)
620 		bump = 1000000;
621 	}
622 	*stv = tv;
623 
624 	td = curthread;
625 	p = td->td_proc;
626 
627 	if (frame && CLKF_USERMODE(frame)) {
628 		/*
629 		 * Came from userland, handle user time and deal with
630 		 * possible process.
631 		 */
632 		if (p && (p->p_flags & P_PROFIL))
633 			addupc_intr(p, CLKF_PC(frame), 1);
634 		td->td_uticks += bump;
635 
636 		/*
637 		 * Charge the time as appropriate
638 		 */
639 		if (p && p->p_nice > NZERO)
640 			cpu_time.cp_nice += bump;
641 		else
642 			cpu_time.cp_user += bump;
643 	} else {
644 		int intr_nest = mycpu->gd_intr_nesting_level;
645 
646 		if (in_ipi) {
647 			/*
648 			 * IPI processing code will bump gd_intr_nesting_level
649 			 * up by one, which breaks following CLKF_INTR testing,
650 			 * so we substract it by one here.
651 			 */
652 			--intr_nest;
653 		}
654 #ifdef GPROF
655 		/*
656 		 * Kernel statistics are just like addupc_intr, only easier.
657 		 */
658 		g = &_gmonparam;
659 		if (g->state == GMON_PROF_ON && frame) {
660 			i = CLKF_PC(frame) - g->lowpc;
661 			if (i < g->textsize) {
662 				i /= HISTFRACTION * sizeof(*g->kcount);
663 				g->kcount[i]++;
664 			}
665 		}
666 #endif
667 		/*
668 		 * Came from kernel mode, so we were:
669 		 * - handling an interrupt,
670 		 * - doing syscall or trap work on behalf of the current
671 		 *   user process, or
672 		 * - spinning in the idle loop.
673 		 * Whichever it is, charge the time as appropriate.
674 		 * Note that we charge interrupts to the current process,
675 		 * regardless of whether they are ``for'' that process,
676 		 * so that we know how much of its real time was spent
677 		 * in ``non-process'' (i.e., interrupt) work.
678 		 *
679 		 * XXX assume system if frame is NULL.  A NULL frame
680 		 * can occur if ipi processing is done from a crit_exit().
681 		 */
682 		if (frame && CLKF_INTR(intr_nest))
683 			td->td_iticks += bump;
684 		else
685 			td->td_sticks += bump;
686 
687 		if (frame && CLKF_INTR(intr_nest)) {
688 #ifdef DEBUG_PCTRACK
689 			do_pctrack(frame, PCTRACK_INT);
690 #endif
691 			cpu_time.cp_intr += bump;
692 		} else {
693 			if (td == &mycpu->gd_idlethread) {
694 				cpu_time.cp_idle += bump;
695 			} else {
696 #ifdef DEBUG_PCTRACK
697 				if (frame)
698 					do_pctrack(frame, PCTRACK_SYS);
699 #endif
700 				cpu_time.cp_sys += bump;
701 			}
702 		}
703 	}
704 }
705 
706 #ifdef DEBUG_PCTRACK
707 /*
708  * Sample the PC when in the kernel or in an interrupt.  User code can
709  * retrieve the information and generate a histogram or other output.
710  */
711 
712 static void
713 do_pctrack(struct intrframe *frame, int which)
714 {
715 	struct kinfo_pctrack *pctrack;
716 
717 	pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
718 	pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
719 		(void *)CLKF_PC(frame);
720 	++pctrack->pc_index;
721 }
722 
723 static int
724 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
725 {
726 	struct kinfo_pcheader head;
727 	int error;
728 	int cpu;
729 	int ntrack;
730 
731 	head.pc_ntrack = PCTRACK_SIZE;
732 	head.pc_arysize = PCTRACK_ARYSIZE;
733 
734 	if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
735 		return (error);
736 
737 	for (cpu = 0; cpu < ncpus; ++cpu) {
738 		for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
739 			error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
740 					   sizeof(struct kinfo_pctrack));
741 			if (error)
742 				break;
743 		}
744 		if (error)
745 			break;
746 	}
747 	return (error);
748 }
749 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
750 	sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
751 
752 #endif
753 
754 /*
755  * The scheduler clock typically runs at a 50Hz rate.  NOTE! systimer,
756  * the MP lock might not be held.  We can safely manipulate parts of curproc
757  * but that's about it.
758  *
759  * Each cpu has its own scheduler clock.
760  */
761 static void
762 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
763 {
764 	struct lwp *lp;
765 	struct rusage *ru;
766 	struct vmspace *vm;
767 	long rss;
768 
769 	if ((lp = lwkt_preempted_proc()) != NULL) {
770 		/*
771 		 * Account for cpu time used and hit the scheduler.  Note
772 		 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
773 		 * HERE.
774 		 */
775 		++lp->lwp_cpticks;
776 		lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
777 						       info->time);
778 	}
779 	if ((lp = curthread->td_lwp) != NULL) {
780 		/*
781 		 * Update resource usage integrals and maximums.
782 		 */
783 		if ((ru = &lp->lwp_proc->p_ru) &&
784 		    (vm = lp->lwp_proc->p_vmspace) != NULL) {
785 			ru->ru_ixrss += pgtok(vm->vm_tsize);
786 			ru->ru_idrss += pgtok(vm->vm_dsize);
787 			ru->ru_isrss += pgtok(vm->vm_ssize);
788 			if (lwkt_trytoken(&vm->vm_map.token)) {
789 				rss = pgtok(vmspace_resident_count(vm));
790 				if (ru->ru_maxrss < rss)
791 					ru->ru_maxrss = rss;
792 				lwkt_reltoken(&vm->vm_map.token);
793 			}
794 		}
795 	}
796 }
797 
798 /*
799  * Compute number of ticks for the specified amount of time.  The
800  * return value is intended to be used in a clock interrupt timed
801  * operation and guarenteed to meet or exceed the requested time.
802  * If the representation overflows, return INT_MAX.  The minimum return
803  * value is 1 ticks and the function will average the calculation up.
804  * If any value greater then 0 microseconds is supplied, a value
805  * of at least 2 will be returned to ensure that a near-term clock
806  * interrupt does not cause the timeout to occur (degenerately) early.
807  *
808  * Note that limit checks must take into account microseconds, which is
809  * done simply by using the smaller signed long maximum instead of
810  * the unsigned long maximum.
811  *
812  * If ints have 32 bits, then the maximum value for any timeout in
813  * 10ms ticks is 248 days.
814  */
815 int
816 tvtohz_high(struct timeval *tv)
817 {
818 	int ticks;
819 	long sec, usec;
820 
821 	sec = tv->tv_sec;
822 	usec = tv->tv_usec;
823 	if (usec < 0) {
824 		sec--;
825 		usec += 1000000;
826 	}
827 	if (sec < 0) {
828 #ifdef DIAGNOSTIC
829 		if (usec > 0) {
830 			sec++;
831 			usec -= 1000000;
832 		}
833 		kprintf("tvtohz_high: negative time difference "
834 			"%ld sec %ld usec\n",
835 			sec, usec);
836 #endif
837 		ticks = 1;
838 	} else if (sec <= INT_MAX / hz) {
839 		ticks = (int)(sec * hz +
840 			    ((u_long)usec + (ustick - 1)) / ustick) + 1;
841 	} else {
842 		ticks = INT_MAX;
843 	}
844 	return (ticks);
845 }
846 
847 int
848 tstohz_high(struct timespec *ts)
849 {
850 	int ticks;
851 	long sec, nsec;
852 
853 	sec = ts->tv_sec;
854 	nsec = ts->tv_nsec;
855 	if (nsec < 0) {
856 		sec--;
857 		nsec += 1000000000;
858 	}
859 	if (sec < 0) {
860 #ifdef DIAGNOSTIC
861 		if (nsec > 0) {
862 			sec++;
863 			nsec -= 1000000000;
864 		}
865 		kprintf("tstohz_high: negative time difference "
866 			"%ld sec %ld nsec\n",
867 			sec, nsec);
868 #endif
869 		ticks = 1;
870 	} else if (sec <= INT_MAX / hz) {
871 		ticks = (int)(sec * hz +
872 			    ((u_long)nsec + (nstick - 1)) / nstick) + 1;
873 	} else {
874 		ticks = INT_MAX;
875 	}
876 	return (ticks);
877 }
878 
879 
880 /*
881  * Compute number of ticks for the specified amount of time, erroring on
882  * the side of it being too low to ensure that sleeping the returned number
883  * of ticks will not result in a late return.
884  *
885  * The supplied timeval may not be negative and should be normalized.  A
886  * return value of 0 is possible if the timeval converts to less then
887  * 1 tick.
888  *
889  * If ints have 32 bits, then the maximum value for any timeout in
890  * 10ms ticks is 248 days.
891  */
892 int
893 tvtohz_low(struct timeval *tv)
894 {
895 	int ticks;
896 	long sec;
897 
898 	sec = tv->tv_sec;
899 	if (sec <= INT_MAX / hz)
900 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
901 	else
902 		ticks = INT_MAX;
903 	return (ticks);
904 }
905 
906 int
907 tstohz_low(struct timespec *ts)
908 {
909 	int ticks;
910 	long sec;
911 
912 	sec = ts->tv_sec;
913 	if (sec <= INT_MAX / hz)
914 		ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
915 	else
916 		ticks = INT_MAX;
917 	return (ticks);
918 }
919 
920 /*
921  * Start profiling on a process.
922  *
923  * Kernel profiling passes proc0 which never exits and hence
924  * keeps the profile clock running constantly.
925  */
926 void
927 startprofclock(struct proc *p)
928 {
929 	if ((p->p_flags & P_PROFIL) == 0) {
930 		p->p_flags |= P_PROFIL;
931 #if 0	/* XXX */
932 		if (++profprocs == 1 && stathz != 0) {
933 			crit_enter();
934 			psdiv = psratio;
935 			setstatclockrate(profhz);
936 			crit_exit();
937 		}
938 #endif
939 	}
940 }
941 
942 /*
943  * Stop profiling on a process.
944  *
945  * caller must hold p->p_token
946  */
947 void
948 stopprofclock(struct proc *p)
949 {
950 	if (p->p_flags & P_PROFIL) {
951 		p->p_flags &= ~P_PROFIL;
952 #if 0	/* XXX */
953 		if (--profprocs == 0 && stathz != 0) {
954 			crit_enter();
955 			psdiv = 1;
956 			setstatclockrate(stathz);
957 			crit_exit();
958 		}
959 #endif
960 	}
961 }
962 
963 /*
964  * Return information about system clocks.
965  */
966 static int
967 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
968 {
969 	struct kinfo_clockinfo clkinfo;
970 	/*
971 	 * Construct clockinfo structure.
972 	 */
973 	clkinfo.ci_hz = hz;
974 	clkinfo.ci_tick = ustick;
975 	clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
976 	clkinfo.ci_profhz = profhz;
977 	clkinfo.ci_stathz = stathz ? stathz : hz;
978 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
979 }
980 
981 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
982 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
983 
984 /*
985  * We have eight functions for looking at the clock, four for
986  * microseconds and four for nanoseconds.  For each there is fast
987  * but less precise version "get{nano|micro}[up]time" which will
988  * return a time which is up to 1/HZ previous to the call, whereas
989  * the raw version "{nano|micro}[up]time" will return a timestamp
990  * which is as precise as possible.  The "up" variants return the
991  * time relative to system boot, these are well suited for time
992  * interval measurements.
993  *
994  * Each cpu independantly maintains the current time of day, so all
995  * we need to do to protect ourselves from changes is to do a loop
996  * check on the seconds field changing out from under us.
997  *
998  * The system timer maintains a 32 bit count and due to various issues
999  * it is possible for the calculated delta to occassionally exceed
1000  * sys_cputimer->freq.  If this occurs the sys_cputimer->freq64_nsec
1001  * multiplication can easily overflow, so we deal with the case.  For
1002  * uniformity we deal with the case in the usec case too.
1003  *
1004  * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1005  */
1006 void
1007 getmicrouptime(struct timeval *tvp)
1008 {
1009 	struct globaldata *gd = mycpu;
1010 	sysclock_t delta;
1011 
1012 	do {
1013 		tvp->tv_sec = gd->gd_time_seconds;
1014 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1015 	} while (tvp->tv_sec != gd->gd_time_seconds);
1016 
1017 	if (delta >= sys_cputimer->freq) {
1018 		tvp->tv_sec += delta / sys_cputimer->freq;
1019 		delta %= sys_cputimer->freq;
1020 	}
1021 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1022 	if (tvp->tv_usec >= 1000000) {
1023 		tvp->tv_usec -= 1000000;
1024 		++tvp->tv_sec;
1025 	}
1026 }
1027 
1028 void
1029 getnanouptime(struct timespec *tsp)
1030 {
1031 	struct globaldata *gd = mycpu;
1032 	sysclock_t delta;
1033 
1034 	do {
1035 		tsp->tv_sec = gd->gd_time_seconds;
1036 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1037 	} while (tsp->tv_sec != gd->gd_time_seconds);
1038 
1039 	if (delta >= sys_cputimer->freq) {
1040 		tsp->tv_sec += delta / sys_cputimer->freq;
1041 		delta %= sys_cputimer->freq;
1042 	}
1043 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1044 }
1045 
1046 void
1047 microuptime(struct timeval *tvp)
1048 {
1049 	struct globaldata *gd = mycpu;
1050 	sysclock_t delta;
1051 
1052 	do {
1053 		tvp->tv_sec = gd->gd_time_seconds;
1054 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1055 	} while (tvp->tv_sec != gd->gd_time_seconds);
1056 
1057 	if (delta >= sys_cputimer->freq) {
1058 		tvp->tv_sec += delta / sys_cputimer->freq;
1059 		delta %= sys_cputimer->freq;
1060 	}
1061 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1062 }
1063 
1064 void
1065 nanouptime(struct timespec *tsp)
1066 {
1067 	struct globaldata *gd = mycpu;
1068 	sysclock_t delta;
1069 
1070 	do {
1071 		tsp->tv_sec = gd->gd_time_seconds;
1072 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1073 	} while (tsp->tv_sec != gd->gd_time_seconds);
1074 
1075 	if (delta >= sys_cputimer->freq) {
1076 		tsp->tv_sec += delta / sys_cputimer->freq;
1077 		delta %= sys_cputimer->freq;
1078 	}
1079 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1080 }
1081 
1082 /*
1083  * realtime routines
1084  */
1085 void
1086 getmicrotime(struct timeval *tvp)
1087 {
1088 	struct globaldata *gd = mycpu;
1089 	struct timespec *bt;
1090 	sysclock_t delta;
1091 
1092 	do {
1093 		tvp->tv_sec = gd->gd_time_seconds;
1094 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1095 	} while (tvp->tv_sec != gd->gd_time_seconds);
1096 
1097 	if (delta >= sys_cputimer->freq) {
1098 		tvp->tv_sec += delta / sys_cputimer->freq;
1099 		delta %= sys_cputimer->freq;
1100 	}
1101 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1102 
1103 	bt = &basetime[basetime_index];
1104 	tvp->tv_sec += bt->tv_sec;
1105 	tvp->tv_usec += bt->tv_nsec / 1000;
1106 	while (tvp->tv_usec >= 1000000) {
1107 		tvp->tv_usec -= 1000000;
1108 		++tvp->tv_sec;
1109 	}
1110 }
1111 
1112 void
1113 getnanotime(struct timespec *tsp)
1114 {
1115 	struct globaldata *gd = mycpu;
1116 	struct timespec *bt;
1117 	sysclock_t delta;
1118 
1119 	do {
1120 		tsp->tv_sec = gd->gd_time_seconds;
1121 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1122 	} while (tsp->tv_sec != gd->gd_time_seconds);
1123 
1124 	if (delta >= sys_cputimer->freq) {
1125 		tsp->tv_sec += delta / sys_cputimer->freq;
1126 		delta %= sys_cputimer->freq;
1127 	}
1128 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1129 
1130 	bt = &basetime[basetime_index];
1131 	tsp->tv_sec += bt->tv_sec;
1132 	tsp->tv_nsec += bt->tv_nsec;
1133 	while (tsp->tv_nsec >= 1000000000) {
1134 		tsp->tv_nsec -= 1000000000;
1135 		++tsp->tv_sec;
1136 	}
1137 }
1138 
1139 static void
1140 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1141 {
1142 	struct globaldata *gd = mycpu;
1143 	sysclock_t delta;
1144 
1145 	do {
1146 		tsp->tv_sec = gd->gd_time_seconds;
1147 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1148 	} while (tsp->tv_sec != gd->gd_time_seconds);
1149 
1150 	if (delta >= sys_cputimer->freq) {
1151 		tsp->tv_sec += delta / sys_cputimer->freq;
1152 		delta %= sys_cputimer->freq;
1153 	}
1154 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1155 
1156 	tsp->tv_sec += nbt->tv_sec;
1157 	tsp->tv_nsec += nbt->tv_nsec;
1158 	while (tsp->tv_nsec >= 1000000000) {
1159 		tsp->tv_nsec -= 1000000000;
1160 		++tsp->tv_sec;
1161 	}
1162 }
1163 
1164 
1165 void
1166 microtime(struct timeval *tvp)
1167 {
1168 	struct globaldata *gd = mycpu;
1169 	struct timespec *bt;
1170 	sysclock_t delta;
1171 
1172 	do {
1173 		tvp->tv_sec = gd->gd_time_seconds;
1174 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1175 	} while (tvp->tv_sec != gd->gd_time_seconds);
1176 
1177 	if (delta >= sys_cputimer->freq) {
1178 		tvp->tv_sec += delta / sys_cputimer->freq;
1179 		delta %= sys_cputimer->freq;
1180 	}
1181 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1182 
1183 	bt = &basetime[basetime_index];
1184 	tvp->tv_sec += bt->tv_sec;
1185 	tvp->tv_usec += bt->tv_nsec / 1000;
1186 	while (tvp->tv_usec >= 1000000) {
1187 		tvp->tv_usec -= 1000000;
1188 		++tvp->tv_sec;
1189 	}
1190 }
1191 
1192 void
1193 nanotime(struct timespec *tsp)
1194 {
1195 	struct globaldata *gd = mycpu;
1196 	struct timespec *bt;
1197 	sysclock_t delta;
1198 
1199 	do {
1200 		tsp->tv_sec = gd->gd_time_seconds;
1201 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1202 	} while (tsp->tv_sec != gd->gd_time_seconds);
1203 
1204 	if (delta >= sys_cputimer->freq) {
1205 		tsp->tv_sec += delta / sys_cputimer->freq;
1206 		delta %= sys_cputimer->freq;
1207 	}
1208 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1209 
1210 	bt = &basetime[basetime_index];
1211 	tsp->tv_sec += bt->tv_sec;
1212 	tsp->tv_nsec += bt->tv_nsec;
1213 	while (tsp->tv_nsec >= 1000000000) {
1214 		tsp->tv_nsec -= 1000000000;
1215 		++tsp->tv_sec;
1216 	}
1217 }
1218 
1219 /*
1220  * note: this is not exactly synchronized with real time.  To do that we
1221  * would have to do what microtime does and check for a nanoseconds overflow.
1222  */
1223 time_t
1224 get_approximate_time_t(void)
1225 {
1226 	struct globaldata *gd = mycpu;
1227 	struct timespec *bt;
1228 
1229 	bt = &basetime[basetime_index];
1230 	return(gd->gd_time_seconds + bt->tv_sec);
1231 }
1232 
1233 int
1234 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1235 {
1236 	pps_params_t *app;
1237 	struct pps_fetch_args *fapi;
1238 #ifdef PPS_SYNC
1239 	struct pps_kcbind_args *kapi;
1240 #endif
1241 
1242 	switch (cmd) {
1243 	case PPS_IOC_CREATE:
1244 		return (0);
1245 	case PPS_IOC_DESTROY:
1246 		return (0);
1247 	case PPS_IOC_SETPARAMS:
1248 		app = (pps_params_t *)data;
1249 		if (app->mode & ~pps->ppscap)
1250 			return (EINVAL);
1251 		pps->ppsparam = *app;
1252 		return (0);
1253 	case PPS_IOC_GETPARAMS:
1254 		app = (pps_params_t *)data;
1255 		*app = pps->ppsparam;
1256 		app->api_version = PPS_API_VERS_1;
1257 		return (0);
1258 	case PPS_IOC_GETCAP:
1259 		*(int*)data = pps->ppscap;
1260 		return (0);
1261 	case PPS_IOC_FETCH:
1262 		fapi = (struct pps_fetch_args *)data;
1263 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1264 			return (EINVAL);
1265 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1266 			return (EOPNOTSUPP);
1267 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
1268 		fapi->pps_info_buf = pps->ppsinfo;
1269 		return (0);
1270 	case PPS_IOC_KCBIND:
1271 #ifdef PPS_SYNC
1272 		kapi = (struct pps_kcbind_args *)data;
1273 		/* XXX Only root should be able to do this */
1274 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1275 			return (EINVAL);
1276 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1277 			return (EINVAL);
1278 		if (kapi->edge & ~pps->ppscap)
1279 			return (EINVAL);
1280 		pps->kcmode = kapi->edge;
1281 		return (0);
1282 #else
1283 		return (EOPNOTSUPP);
1284 #endif
1285 	default:
1286 		return (ENOTTY);
1287 	}
1288 }
1289 
1290 void
1291 pps_init(struct pps_state *pps)
1292 {
1293 	pps->ppscap |= PPS_TSFMT_TSPEC;
1294 	if (pps->ppscap & PPS_CAPTUREASSERT)
1295 		pps->ppscap |= PPS_OFFSETASSERT;
1296 	if (pps->ppscap & PPS_CAPTURECLEAR)
1297 		pps->ppscap |= PPS_OFFSETCLEAR;
1298 }
1299 
1300 void
1301 pps_event(struct pps_state *pps, sysclock_t count, int event)
1302 {
1303 	struct globaldata *gd;
1304 	struct timespec *tsp;
1305 	struct timespec *osp;
1306 	struct timespec *bt;
1307 	struct timespec ts;
1308 	sysclock_t *pcount;
1309 #ifdef PPS_SYNC
1310 	sysclock_t tcount;
1311 #endif
1312 	sysclock_t delta;
1313 	pps_seq_t *pseq;
1314 	int foff;
1315 	int fhard;
1316 
1317 	gd = mycpu;
1318 
1319 	/* Things would be easier with arrays... */
1320 	if (event == PPS_CAPTUREASSERT) {
1321 		tsp = &pps->ppsinfo.assert_timestamp;
1322 		osp = &pps->ppsparam.assert_offset;
1323 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1324 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1325 		pcount = &pps->ppscount[0];
1326 		pseq = &pps->ppsinfo.assert_sequence;
1327 	} else {
1328 		tsp = &pps->ppsinfo.clear_timestamp;
1329 		osp = &pps->ppsparam.clear_offset;
1330 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1331 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1332 		pcount = &pps->ppscount[1];
1333 		pseq = &pps->ppsinfo.clear_sequence;
1334 	}
1335 
1336 	/* Nothing really happened */
1337 	if (*pcount == count)
1338 		return;
1339 
1340 	*pcount = count;
1341 
1342 	do {
1343 		ts.tv_sec = gd->gd_time_seconds;
1344 		delta = count - gd->gd_cpuclock_base;
1345 	} while (ts.tv_sec != gd->gd_time_seconds);
1346 
1347 	if (delta >= sys_cputimer->freq) {
1348 		ts.tv_sec += delta / sys_cputimer->freq;
1349 		delta %= sys_cputimer->freq;
1350 	}
1351 	ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1352 	bt = &basetime[basetime_index];
1353 	ts.tv_sec += bt->tv_sec;
1354 	ts.tv_nsec += bt->tv_nsec;
1355 	while (ts.tv_nsec >= 1000000000) {
1356 		ts.tv_nsec -= 1000000000;
1357 		++ts.tv_sec;
1358 	}
1359 
1360 	(*pseq)++;
1361 	*tsp = ts;
1362 
1363 	if (foff) {
1364 		timespecadd(tsp, osp);
1365 		if (tsp->tv_nsec < 0) {
1366 			tsp->tv_nsec += 1000000000;
1367 			tsp->tv_sec -= 1;
1368 		}
1369 	}
1370 #ifdef PPS_SYNC
1371 	if (fhard) {
1372 		/* magic, at its best... */
1373 		tcount = count - pps->ppscount[2];
1374 		pps->ppscount[2] = count;
1375 		if (tcount >= sys_cputimer->freq) {
1376 			delta = (1000000000 * (tcount / sys_cputimer->freq) +
1377 				 sys_cputimer->freq64_nsec *
1378 				 (tcount % sys_cputimer->freq)) >> 32;
1379 		} else {
1380 			delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1381 		}
1382 		hardpps(tsp, delta);
1383 	}
1384 #endif
1385 }
1386 
1387 /*
1388  * Return the tsc target value for a delay of (ns).
1389  *
1390  * Returns -1 if the TSC is not supported.
1391  */
1392 int64_t
1393 tsc_get_target(int ns)
1394 {
1395 #if defined(_RDTSC_SUPPORTED_)
1396 	if (cpu_feature & CPUID_TSC) {
1397 		return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1398 	}
1399 #endif
1400 	return(-1);
1401 }
1402 
1403 /*
1404  * Compare the tsc against the passed target
1405  *
1406  * Returns +1 if the target has been reached
1407  * Returns  0 if the target has not yet been reached
1408  * Returns -1 if the TSC is not supported.
1409  *
1410  * Typical use:		while (tsc_test_target(target) == 0) { ...poll... }
1411  */
1412 int
1413 tsc_test_target(int64_t target)
1414 {
1415 #if defined(_RDTSC_SUPPORTED_)
1416 	if (cpu_feature & CPUID_TSC) {
1417 		if ((int64_t)(target - rdtsc()) <= 0)
1418 			return(1);
1419 		return(0);
1420 	}
1421 #endif
1422 	return(-1);
1423 }
1424 
1425 /*
1426  * Delay the specified number of nanoseconds using the tsc.  This function
1427  * returns immediately if the TSC is not supported.  At least one cpu_pause()
1428  * will be issued.
1429  */
1430 void
1431 tsc_delay(int ns)
1432 {
1433 	int64_t clk;
1434 
1435 	clk = tsc_get_target(ns);
1436 	cpu_pause();
1437 	while (tsc_test_target(clk) == 0)
1438 		cpu_pause();
1439 }
1440