xref: /dflybsd-src/sys/kern/kern_clock.c (revision ee65b806ac08b188bcab21ef0f1efda2cd5bdef7)
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 #include <sys/mplock2.h>
101 
102 #include <machine/cpu.h>
103 #include <machine/limits.h>
104 #include <machine/smp.h>
105 #include <machine/cpufunc.h>
106 #include <machine/specialreg.h>
107 #include <machine/clock.h>
108 
109 #ifdef GPROF
110 #include <sys/gmon.h>
111 #endif
112 
113 #ifdef DEVICE_POLLING
114 extern void init_device_poll_pcpu(int);
115 #endif
116 
117 #ifdef IFPOLL_ENABLE
118 extern void ifpoll_init_pcpu(int);
119 #endif
120 
121 #ifdef DEBUG_PCTRACK
122 static void do_pctrack(struct intrframe *frame, int which);
123 #endif
124 
125 static void initclocks (void *dummy);
126 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
127 
128 /*
129  * Some of these don't belong here, but it's easiest to concentrate them.
130  * Note that cpu_time counts in microseconds, but most userland programs
131  * just compare relative times against the total by delta.
132  */
133 struct kinfo_cputime cputime_percpu[MAXCPU];
134 #ifdef DEBUG_PCTRACK
135 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
136 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
137 #endif
138 
139 #ifdef SMP
140 static int
141 sysctl_cputime(SYSCTL_HANDLER_ARGS)
142 {
143 	int cpu, error = 0;
144 	size_t size = sizeof(struct kinfo_cputime);
145 
146 	for (cpu = 0; cpu < ncpus; ++cpu) {
147 		if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
148 			break;
149 	}
150 
151 	return (error);
152 }
153 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
154 	sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
155 #else
156 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
157     "CPU time statistics");
158 #endif
159 
160 /*
161  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
162  * microuptime().  microtime() is not drift compensated.  The real uptime
163  * with compensation is nanotime() - bootime.  boottime is recalculated
164  * whenever the real time is set based on the compensated elapsed time
165  * in seconds (gd->gd_time_seconds).
166  *
167  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
168  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
169  * the real time.
170  */
171 struct timespec boottime;	/* boot time (realtime) for reference only */
172 time_t time_second;		/* read-only 'passive' uptime in seconds */
173 
174 /*
175  * basetime is used to calculate the compensated real time of day.  The
176  * basetime can be modified on a per-tick basis by the adjtime(),
177  * ntp_adjtime(), and sysctl-based time correction APIs.
178  *
179  * Note that frequency corrections can also be made by adjusting
180  * gd_cpuclock_base.
181  *
182  * basetime is a tail-chasing FIFO, updated only by cpu #0.  The FIFO is
183  * used on both SMP and UP systems to avoid MP races between cpu's and
184  * interrupt races on UP systems.
185  */
186 #define BASETIME_ARYSIZE	16
187 #define BASETIME_ARYMASK	(BASETIME_ARYSIZE - 1)
188 static struct timespec basetime[BASETIME_ARYSIZE];
189 static volatile int basetime_index;
190 
191 static int
192 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
193 {
194 	struct timespec *bt;
195 	int error;
196 	int index;
197 
198 	/*
199 	 * Because basetime data and index may be updated by another cpu,
200 	 * a load fence is required to ensure that the data we read has
201 	 * not been speculatively read relative to a possibly updated index.
202 	 */
203 	index = basetime_index;
204 	cpu_lfence();
205 	bt = &basetime[index];
206 	error = SYSCTL_OUT(req, bt, sizeof(*bt));
207 	return (error);
208 }
209 
210 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
211     &boottime, timespec, "System boottime");
212 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
213     sysctl_get_basetime, "S,timespec", "System basetime");
214 
215 static void hardclock(systimer_t info, struct intrframe *frame);
216 static void statclock(systimer_t info, struct intrframe *frame);
217 static void schedclock(systimer_t info, struct intrframe *frame);
218 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
219 
220 int	ticks;			/* system master ticks at hz */
221 int	clocks_running;		/* tsleep/timeout clocks operational */
222 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
223 int64_t	nsec_acc;		/* accumulator */
224 
225 /* NTPD time correction fields */
226 int64_t	ntp_tick_permanent;	/* per-tick adjustment in nsec << 32 */
227 int64_t	ntp_tick_acc;		/* accumulator for per-tick adjustment */
228 int64_t	ntp_delta;		/* one-time correction in nsec */
229 int64_t ntp_big_delta = 1000000000;
230 int32_t	ntp_tick_delta;		/* current adjustment rate */
231 int32_t	ntp_default_tick_delta;	/* adjustment rate for ntp_delta */
232 time_t	ntp_leap_second;	/* time of next leap second */
233 int	ntp_leap_insert;	/* whether to insert or remove a second */
234 
235 /*
236  * Finish initializing clock frequencies and start all clocks running.
237  */
238 /* ARGSUSED*/
239 static void
240 initclocks(void *dummy)
241 {
242 	/*psratio = profhz / stathz;*/
243 	initclocks_pcpu();
244 	clocks_running = 1;
245 }
246 
247 /*
248  * Called on a per-cpu basis
249  */
250 void
251 initclocks_pcpu(void)
252 {
253 	struct globaldata *gd = mycpu;
254 
255 	crit_enter();
256 	if (gd->gd_cpuid == 0) {
257 	    gd->gd_time_seconds = 1;
258 	    gd->gd_cpuclock_base = sys_cputimer->count();
259 	} else {
260 	    /* XXX */
261 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
262 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
263 	}
264 
265 	systimer_intr_enable();
266 
267 #ifdef DEVICE_POLLING
268 	init_device_poll_pcpu(gd->gd_cpuid);
269 #endif
270 
271 #ifdef IFPOLL_ENABLE
272 	ifpoll_init_pcpu(gd->gd_cpuid);
273 #endif
274 
275 	/*
276 	 * Use a non-queued periodic systimer to prevent multiple ticks from
277 	 * building up if the sysclock jumps forward (8254 gets reset).  The
278 	 * sysclock will never jump backwards.  Our time sync is based on
279 	 * the actual sysclock, not the ticks count.
280 	 */
281 	systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
282 	systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
283 	/* XXX correct the frequency for scheduler / estcpu tests */
284 	systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
285 				NULL, ESTCPUFREQ);
286 	crit_exit();
287 }
288 
289 /*
290  * This sets the current real time of day.  Timespecs are in seconds and
291  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
292  * instead we adjust basetime so basetime + gd_* results in the current
293  * time of day.  This way the gd_* fields are guarenteed to represent
294  * a monotonically increasing 'uptime' value.
295  *
296  * When set_timeofday() is called from userland, the system call forces it
297  * onto cpu #0 since only cpu #0 can update basetime_index.
298  */
299 void
300 set_timeofday(struct timespec *ts)
301 {
302 	struct timespec *nbt;
303 	int ni;
304 
305 	/*
306 	 * XXX SMP / non-atomic basetime updates
307 	 */
308 	crit_enter();
309 	ni = (basetime_index + 1) & BASETIME_ARYMASK;
310 	nbt = &basetime[ni];
311 	nanouptime(nbt);
312 	nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
313 	nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
314 	if (nbt->tv_nsec < 0) {
315 	    nbt->tv_nsec += 1000000000;
316 	    --nbt->tv_sec;
317 	}
318 
319 	/*
320 	 * Note that basetime diverges from boottime as the clock drift is
321 	 * compensated for, so we cannot do away with boottime.  When setting
322 	 * the absolute time of day the drift is 0 (for an instant) and we
323 	 * can simply assign boottime to basetime.
324 	 *
325 	 * Note that nanouptime() is based on gd_time_seconds which is drift
326 	 * compensated up to a point (it is guarenteed to remain monotonically
327 	 * increasing).  gd_time_seconds is thus our best uptime guess and
328 	 * suitable for use in the boottime calculation.  It is already taken
329 	 * into account in the basetime calculation above.
330 	 */
331 	boottime.tv_sec = nbt->tv_sec;
332 	ntp_delta = 0;
333 
334 	/*
335 	 * We now have a new basetime, make sure all other cpus have it,
336 	 * then update the index.
337 	 */
338 	cpu_sfence();
339 	basetime_index = ni;
340 
341 	crit_exit();
342 }
343 
344 /*
345  * Each cpu has its own hardclock, but we only increments ticks and softticks
346  * on cpu #0.
347  *
348  * NOTE! systimer! the MP lock might not be held here.  We can only safely
349  * manipulate objects owned by the current cpu.
350  */
351 static void
352 hardclock(systimer_t info, struct intrframe *frame)
353 {
354 	sysclock_t cputicks;
355 	struct proc *p;
356 	struct globaldata *gd = mycpu;
357 
358 	/*
359 	 * Realtime updates are per-cpu.  Note that timer corrections as
360 	 * returned by microtime() and friends make an additional adjustment
361 	 * using a system-wise 'basetime', but the running time is always
362 	 * taken from the per-cpu globaldata area.  Since the same clock
363 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
364 	 * stay in synch.
365 	 *
366 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
367 	 * to reverse index gd_cpuclock_base, but that it is possible for
368 	 * it to temporarily get behind in the seconds if something in the
369 	 * system locks interrupts for a long period of time.  Since periodic
370 	 * timers count events, though everything should resynch again
371 	 * immediately.
372 	 */
373 	cputicks = info->time - gd->gd_cpuclock_base;
374 	if (cputicks >= sys_cputimer->freq) {
375 		++gd->gd_time_seconds;
376 		gd->gd_cpuclock_base += sys_cputimer->freq;
377 	}
378 
379 	/*
380 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
381 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
382 	 * by updating basetime.
383 	 */
384 	if (gd->gd_cpuid == 0) {
385 	    struct timespec *nbt;
386 	    struct timespec nts;
387 	    int leap;
388 	    int ni;
389 
390 	    ++ticks;
391 
392 #if 0
393 	    if (tco->tc_poll_pps)
394 		tco->tc_poll_pps(tco);
395 #endif
396 
397 	    /*
398 	     * Calculate the new basetime index.  We are in a critical section
399 	     * on cpu #0 and can safely play with basetime_index.  Start
400 	     * with the current basetime and then make adjustments.
401 	     */
402 	    ni = (basetime_index + 1) & BASETIME_ARYMASK;
403 	    nbt = &basetime[ni];
404 	    *nbt = basetime[basetime_index];
405 
406 	    /*
407 	     * Apply adjtime corrections.  (adjtime() API)
408 	     *
409 	     * adjtime() only runs on cpu #0 so our critical section is
410 	     * sufficient to access these variables.
411 	     */
412 	    if (ntp_delta != 0) {
413 		nbt->tv_nsec += ntp_tick_delta;
414 		ntp_delta -= ntp_tick_delta;
415 		if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
416 		    (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
417 			ntp_tick_delta = ntp_delta;
418  		}
419  	    }
420 
421 	    /*
422 	     * Apply permanent frequency corrections.  (sysctl API)
423 	     */
424 	    if (ntp_tick_permanent != 0) {
425 		ntp_tick_acc += ntp_tick_permanent;
426 		if (ntp_tick_acc >= (1LL << 32)) {
427 		    nbt->tv_nsec += ntp_tick_acc >> 32;
428 		    ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
429 		} else if (ntp_tick_acc <= -(1LL << 32)) {
430 		    /* Negate ntp_tick_acc to avoid shifting the sign bit. */
431 		    nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
432 		    ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
433 		}
434  	    }
435 
436 	    if (nbt->tv_nsec >= 1000000000) {
437 		    nbt->tv_sec++;
438 		    nbt->tv_nsec -= 1000000000;
439 	    } else if (nbt->tv_nsec < 0) {
440 		    nbt->tv_sec--;
441 		    nbt->tv_nsec += 1000000000;
442 	    }
443 
444 	    /*
445 	     * Another per-tick compensation.  (for ntp_adjtime() API)
446 	     */
447 	    if (nsec_adj != 0) {
448 		nsec_acc += nsec_adj;
449 		if (nsec_acc >= 0x100000000LL) {
450 		    nbt->tv_nsec += nsec_acc >> 32;
451 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
452 		} else if (nsec_acc <= -0x100000000LL) {
453 		    nbt->tv_nsec -= -nsec_acc >> 32;
454 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
455 		}
456 		if (nbt->tv_nsec >= 1000000000) {
457 		    nbt->tv_nsec -= 1000000000;
458 		    ++nbt->tv_sec;
459 		} else if (nbt->tv_nsec < 0) {
460 		    nbt->tv_nsec += 1000000000;
461 		    --nbt->tv_sec;
462 		}
463 	    }
464 
465 	    /************************************************************
466 	     *			LEAP SECOND CORRECTION			*
467 	     ************************************************************
468 	     *
469 	     * Taking into account all the corrections made above, figure
470 	     * out the new real time.  If the seconds field has changed
471 	     * then apply any pending leap-second corrections.
472 	     */
473 	    getnanotime_nbt(nbt, &nts);
474 
475 	    if (time_second != nts.tv_sec) {
476 		/*
477 		 * Apply leap second (sysctl API).  Adjust nts for changes
478 		 * so we do not have to call getnanotime_nbt again.
479 		 */
480 		if (ntp_leap_second) {
481 		    if (ntp_leap_second == nts.tv_sec) {
482 			if (ntp_leap_insert) {
483 			    nbt->tv_sec++;
484 			    nts.tv_sec++;
485 			} else {
486 			    nbt->tv_sec--;
487 			    nts.tv_sec--;
488 			}
489 			ntp_leap_second--;
490 		    }
491 		}
492 
493 		/*
494 		 * Apply leap second (ntp_adjtime() API), calculate a new
495 		 * nsec_adj field.  ntp_update_second() returns nsec_adj
496 		 * as a per-second value but we need it as a per-tick value.
497 		 */
498 		leap = ntp_update_second(time_second, &nsec_adj);
499 		nsec_adj /= hz;
500 		nbt->tv_sec += leap;
501 		nts.tv_sec += leap;
502 
503 		/*
504 		 * Update the time_second 'approximate time' global.
505 		 */
506 		time_second = nts.tv_sec;
507 	    }
508 
509 	    /*
510 	     * Finally, our new basetime is ready to go live!
511 	     */
512 	    cpu_sfence();
513 	    basetime_index = ni;
514 
515 	    /*
516 	     * Figure out how badly the system is starved for memory
517 	     */
518 	    vm_fault_ratecheck();
519 	}
520 
521 	/*
522 	 * softticks are handled for all cpus
523 	 */
524 	hardclock_softtick(gd);
525 
526 	/*
527 	 * The LWKT scheduler will generally allow the current process to
528 	 * return to user mode even if there are other runnable LWKT threads
529 	 * running in kernel mode on behalf of a user process.  This will
530 	 * ensure that those other threads have an opportunity to run in
531 	 * fairly short order (but not instantly).
532 	 */
533 	need_lwkt_resched();
534 
535 	/*
536 	 * ITimer handling is per-tick, per-cpu.  I don't think ksignal()
537 	 * is mpsafe on curproc, so XXX get the mplock.
538 	 */
539 	if ((p = curproc) != NULL && try_mplock()) {
540 		if (frame && CLKF_USERMODE(frame) &&
541 		    timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
542 		    itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0)
543 			ksignal(p, SIGVTALRM);
544 		if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
545 		    itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0)
546 			ksignal(p, SIGPROF);
547 		rel_mplock();
548 	}
549 	setdelayed();
550 }
551 
552 /*
553  * The statistics clock typically runs at a 125Hz rate, and is intended
554  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
555  *
556  * NOTE! systimer! the MP lock might not be held here.  We can only safely
557  * manipulate objects owned by the current cpu.
558  *
559  * The stats clock is responsible for grabbing a profiling sample.
560  * Most of the statistics are only used by user-level statistics programs.
561  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
562  * p->p_estcpu.
563  *
564  * Like the other clocks, the stat clock is called from what is effectively
565  * a fast interrupt, so the context should be the thread/process that got
566  * interrupted.
567  */
568 static void
569 statclock(systimer_t info, struct intrframe *frame)
570 {
571 #ifdef GPROF
572 	struct gmonparam *g;
573 	int i;
574 #endif
575 	thread_t td;
576 	struct proc *p;
577 	int bump;
578 	struct timeval tv;
579 	struct timeval *stv;
580 
581 	/*
582 	 * How big was our timeslice relative to the last time?
583 	 */
584 	microuptime(&tv);	/* mpsafe */
585 	stv = &mycpu->gd_stattv;
586 	if (stv->tv_sec == 0) {
587 	    bump = 1;
588 	} else {
589 	    bump = tv.tv_usec - stv->tv_usec +
590 		(tv.tv_sec - stv->tv_sec) * 1000000;
591 	    if (bump < 0)
592 		bump = 0;
593 	    if (bump > 1000000)
594 		bump = 1000000;
595 	}
596 	*stv = tv;
597 
598 	td = curthread;
599 	p = td->td_proc;
600 
601 	if (frame && CLKF_USERMODE(frame)) {
602 		/*
603 		 * Came from userland, handle user time and deal with
604 		 * possible process.
605 		 */
606 		if (p && (p->p_flag & P_PROFIL))
607 			addupc_intr(p, CLKF_PC(frame), 1);
608 		td->td_uticks += bump;
609 
610 		/*
611 		 * Charge the time as appropriate
612 		 */
613 		if (p && p->p_nice > NZERO)
614 			cpu_time.cp_nice += bump;
615 		else
616 			cpu_time.cp_user += bump;
617 	} else {
618 #ifdef GPROF
619 		/*
620 		 * Kernel statistics are just like addupc_intr, only easier.
621 		 */
622 		g = &_gmonparam;
623 		if (g->state == GMON_PROF_ON && frame) {
624 			i = CLKF_PC(frame) - g->lowpc;
625 			if (i < g->textsize) {
626 				i /= HISTFRACTION * sizeof(*g->kcount);
627 				g->kcount[i]++;
628 			}
629 		}
630 #endif
631 		/*
632 		 * Came from kernel mode, so we were:
633 		 * - handling an interrupt,
634 		 * - doing syscall or trap work on behalf of the current
635 		 *   user process, or
636 		 * - spinning in the idle loop.
637 		 * Whichever it is, charge the time as appropriate.
638 		 * Note that we charge interrupts to the current process,
639 		 * regardless of whether they are ``for'' that process,
640 		 * so that we know how much of its real time was spent
641 		 * in ``non-process'' (i.e., interrupt) work.
642 		 *
643 		 * XXX assume system if frame is NULL.  A NULL frame
644 		 * can occur if ipi processing is done from a crit_exit().
645 		 */
646 		if (frame && CLKF_INTR(frame))
647 			td->td_iticks += bump;
648 		else
649 			td->td_sticks += bump;
650 
651 		if (frame && CLKF_INTR(frame)) {
652 #ifdef DEBUG_PCTRACK
653 			do_pctrack(frame, PCTRACK_INT);
654 #endif
655 			cpu_time.cp_intr += bump;
656 		} else {
657 			if (td == &mycpu->gd_idlethread) {
658 				cpu_time.cp_idle += bump;
659 			} else {
660 #ifdef DEBUG_PCTRACK
661 				if (frame)
662 					do_pctrack(frame, PCTRACK_SYS);
663 #endif
664 				cpu_time.cp_sys += bump;
665 			}
666 		}
667 	}
668 }
669 
670 #ifdef DEBUG_PCTRACK
671 /*
672  * Sample the PC when in the kernel or in an interrupt.  User code can
673  * retrieve the information and generate a histogram or other output.
674  */
675 
676 static void
677 do_pctrack(struct intrframe *frame, int which)
678 {
679 	struct kinfo_pctrack *pctrack;
680 
681 	pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
682 	pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
683 		(void *)CLKF_PC(frame);
684 	++pctrack->pc_index;
685 }
686 
687 static int
688 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
689 {
690 	struct kinfo_pcheader head;
691 	int error;
692 	int cpu;
693 	int ntrack;
694 
695 	head.pc_ntrack = PCTRACK_SIZE;
696 	head.pc_arysize = PCTRACK_ARYSIZE;
697 
698 	if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
699 		return (error);
700 
701 	for (cpu = 0; cpu < ncpus; ++cpu) {
702 		for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
703 			error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
704 					   sizeof(struct kinfo_pctrack));
705 			if (error)
706 				break;
707 		}
708 		if (error)
709 			break;
710 	}
711 	return (error);
712 }
713 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
714 	sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
715 
716 #endif
717 
718 /*
719  * The scheduler clock typically runs at a 50Hz rate.  NOTE! systimer,
720  * the MP lock might not be held.  We can safely manipulate parts of curproc
721  * but that's about it.
722  *
723  * Each cpu has its own scheduler clock.
724  */
725 static void
726 schedclock(systimer_t info, struct intrframe *frame)
727 {
728 	struct lwp *lp;
729 	struct rusage *ru;
730 	struct vmspace *vm;
731 	long rss;
732 
733 	if ((lp = lwkt_preempted_proc()) != NULL) {
734 		/*
735 		 * Account for cpu time used and hit the scheduler.  Note
736 		 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
737 		 * HERE.
738 		 */
739 		++lp->lwp_cpticks;
740 		lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
741 						       info->time);
742 	}
743 	if ((lp = curthread->td_lwp) != NULL) {
744 		/*
745 		 * Update resource usage integrals and maximums.
746 		 */
747 		if ((ru = &lp->lwp_proc->p_ru) &&
748 		    (vm = lp->lwp_proc->p_vmspace) != NULL) {
749 			ru->ru_ixrss += pgtok(vm->vm_tsize);
750 			ru->ru_idrss += pgtok(vm->vm_dsize);
751 			ru->ru_isrss += pgtok(vm->vm_ssize);
752 			rss = pgtok(vmspace_resident_count(vm));
753 			if (ru->ru_maxrss < rss)
754 				ru->ru_maxrss = rss;
755 		}
756 	}
757 }
758 
759 /*
760  * Compute number of ticks for the specified amount of time.  The
761  * return value is intended to be used in a clock interrupt timed
762  * operation and guarenteed to meet or exceed the requested time.
763  * If the representation overflows, return INT_MAX.  The minimum return
764  * value is 1 ticks and the function will average the calculation up.
765  * If any value greater then 0 microseconds is supplied, a value
766  * of at least 2 will be returned to ensure that a near-term clock
767  * interrupt does not cause the timeout to occur (degenerately) early.
768  *
769  * Note that limit checks must take into account microseconds, which is
770  * done simply by using the smaller signed long maximum instead of
771  * the unsigned long maximum.
772  *
773  * If ints have 32 bits, then the maximum value for any timeout in
774  * 10ms ticks is 248 days.
775  */
776 int
777 tvtohz_high(struct timeval *tv)
778 {
779 	int ticks;
780 	long sec, usec;
781 
782 	sec = tv->tv_sec;
783 	usec = tv->tv_usec;
784 	if (usec < 0) {
785 		sec--;
786 		usec += 1000000;
787 	}
788 	if (sec < 0) {
789 #ifdef DIAGNOSTIC
790 		if (usec > 0) {
791 			sec++;
792 			usec -= 1000000;
793 		}
794 		kprintf("tvtohz_high: negative time difference "
795 			"%ld sec %ld usec\n",
796 			sec, usec);
797 #endif
798 		ticks = 1;
799 	} else if (sec <= INT_MAX / hz) {
800 		ticks = (int)(sec * hz +
801 			    ((u_long)usec + (ustick - 1)) / ustick) + 1;
802 	} else {
803 		ticks = INT_MAX;
804 	}
805 	return (ticks);
806 }
807 
808 int
809 tstohz_high(struct timespec *ts)
810 {
811 	int ticks;
812 	long sec, nsec;
813 
814 	sec = ts->tv_sec;
815 	nsec = ts->tv_nsec;
816 	if (nsec < 0) {
817 		sec--;
818 		nsec += 1000000000;
819 	}
820 	if (sec < 0) {
821 #ifdef DIAGNOSTIC
822 		if (nsec > 0) {
823 			sec++;
824 			nsec -= 1000000000;
825 		}
826 		kprintf("tstohz_high: negative time difference "
827 			"%ld sec %ld nsec\n",
828 			sec, nsec);
829 #endif
830 		ticks = 1;
831 	} else if (sec <= INT_MAX / hz) {
832 		ticks = (int)(sec * hz +
833 			    ((u_long)nsec + (nstick - 1)) / nstick) + 1;
834 	} else {
835 		ticks = INT_MAX;
836 	}
837 	return (ticks);
838 }
839 
840 
841 /*
842  * Compute number of ticks for the specified amount of time, erroring on
843  * the side of it being too low to ensure that sleeping the returned number
844  * of ticks will not result in a late return.
845  *
846  * The supplied timeval may not be negative and should be normalized.  A
847  * return value of 0 is possible if the timeval converts to less then
848  * 1 tick.
849  *
850  * If ints have 32 bits, then the maximum value for any timeout in
851  * 10ms ticks is 248 days.
852  */
853 int
854 tvtohz_low(struct timeval *tv)
855 {
856 	int ticks;
857 	long sec;
858 
859 	sec = tv->tv_sec;
860 	if (sec <= INT_MAX / hz)
861 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
862 	else
863 		ticks = INT_MAX;
864 	return (ticks);
865 }
866 
867 int
868 tstohz_low(struct timespec *ts)
869 {
870 	int ticks;
871 	long sec;
872 
873 	sec = ts->tv_sec;
874 	if (sec <= INT_MAX / hz)
875 		ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
876 	else
877 		ticks = INT_MAX;
878 	return (ticks);
879 }
880 
881 /*
882  * Start profiling on a process.
883  *
884  * Kernel profiling passes proc0 which never exits and hence
885  * keeps the profile clock running constantly.
886  */
887 void
888 startprofclock(struct proc *p)
889 {
890 	if ((p->p_flag & P_PROFIL) == 0) {
891 		p->p_flag |= P_PROFIL;
892 #if 0	/* XXX */
893 		if (++profprocs == 1 && stathz != 0) {
894 			crit_enter();
895 			psdiv = psratio;
896 			setstatclockrate(profhz);
897 			crit_exit();
898 		}
899 #endif
900 	}
901 }
902 
903 /*
904  * Stop profiling on a process.
905  */
906 void
907 stopprofclock(struct proc *p)
908 {
909 	if (p->p_flag & P_PROFIL) {
910 		p->p_flag &= ~P_PROFIL;
911 #if 0	/* XXX */
912 		if (--profprocs == 0 && stathz != 0) {
913 			crit_enter();
914 			psdiv = 1;
915 			setstatclockrate(stathz);
916 			crit_exit();
917 		}
918 #endif
919 	}
920 }
921 
922 /*
923  * Return information about system clocks.
924  */
925 static int
926 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
927 {
928 	struct kinfo_clockinfo clkinfo;
929 	/*
930 	 * Construct clockinfo structure.
931 	 */
932 	clkinfo.ci_hz = hz;
933 	clkinfo.ci_tick = ustick;
934 	clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
935 	clkinfo.ci_profhz = profhz;
936 	clkinfo.ci_stathz = stathz ? stathz : hz;
937 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
938 }
939 
940 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
941 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
942 
943 /*
944  * We have eight functions for looking at the clock, four for
945  * microseconds and four for nanoseconds.  For each there is fast
946  * but less precise version "get{nano|micro}[up]time" which will
947  * return a time which is up to 1/HZ previous to the call, whereas
948  * the raw version "{nano|micro}[up]time" will return a timestamp
949  * which is as precise as possible.  The "up" variants return the
950  * time relative to system boot, these are well suited for time
951  * interval measurements.
952  *
953  * Each cpu independantly maintains the current time of day, so all
954  * we need to do to protect ourselves from changes is to do a loop
955  * check on the seconds field changing out from under us.
956  *
957  * The system timer maintains a 32 bit count and due to various issues
958  * it is possible for the calculated delta to occassionally exceed
959  * sys_cputimer->freq.  If this occurs the sys_cputimer->freq64_nsec
960  * multiplication can easily overflow, so we deal with the case.  For
961  * uniformity we deal with the case in the usec case too.
962  *
963  * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
964  */
965 void
966 getmicrouptime(struct timeval *tvp)
967 {
968 	struct globaldata *gd = mycpu;
969 	sysclock_t delta;
970 
971 	do {
972 		tvp->tv_sec = gd->gd_time_seconds;
973 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
974 	} while (tvp->tv_sec != gd->gd_time_seconds);
975 
976 	if (delta >= sys_cputimer->freq) {
977 		tvp->tv_sec += delta / sys_cputimer->freq;
978 		delta %= sys_cputimer->freq;
979 	}
980 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
981 	if (tvp->tv_usec >= 1000000) {
982 		tvp->tv_usec -= 1000000;
983 		++tvp->tv_sec;
984 	}
985 }
986 
987 void
988 getnanouptime(struct timespec *tsp)
989 {
990 	struct globaldata *gd = mycpu;
991 	sysclock_t delta;
992 
993 	do {
994 		tsp->tv_sec = gd->gd_time_seconds;
995 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
996 	} while (tsp->tv_sec != gd->gd_time_seconds);
997 
998 	if (delta >= sys_cputimer->freq) {
999 		tsp->tv_sec += delta / sys_cputimer->freq;
1000 		delta %= sys_cputimer->freq;
1001 	}
1002 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1003 }
1004 
1005 void
1006 microuptime(struct timeval *tvp)
1007 {
1008 	struct globaldata *gd = mycpu;
1009 	sysclock_t delta;
1010 
1011 	do {
1012 		tvp->tv_sec = gd->gd_time_seconds;
1013 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1014 	} while (tvp->tv_sec != gd->gd_time_seconds);
1015 
1016 	if (delta >= sys_cputimer->freq) {
1017 		tvp->tv_sec += delta / sys_cputimer->freq;
1018 		delta %= sys_cputimer->freq;
1019 	}
1020 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1021 }
1022 
1023 void
1024 nanouptime(struct timespec *tsp)
1025 {
1026 	struct globaldata *gd = mycpu;
1027 	sysclock_t delta;
1028 
1029 	do {
1030 		tsp->tv_sec = gd->gd_time_seconds;
1031 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1032 	} while (tsp->tv_sec != gd->gd_time_seconds);
1033 
1034 	if (delta >= sys_cputimer->freq) {
1035 		tsp->tv_sec += delta / sys_cputimer->freq;
1036 		delta %= sys_cputimer->freq;
1037 	}
1038 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1039 }
1040 
1041 /*
1042  * realtime routines
1043  */
1044 void
1045 getmicrotime(struct timeval *tvp)
1046 {
1047 	struct globaldata *gd = mycpu;
1048 	struct timespec *bt;
1049 	sysclock_t delta;
1050 
1051 	do {
1052 		tvp->tv_sec = gd->gd_time_seconds;
1053 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1054 	} while (tvp->tv_sec != gd->gd_time_seconds);
1055 
1056 	if (delta >= sys_cputimer->freq) {
1057 		tvp->tv_sec += delta / sys_cputimer->freq;
1058 		delta %= sys_cputimer->freq;
1059 	}
1060 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1061 
1062 	bt = &basetime[basetime_index];
1063 	tvp->tv_sec += bt->tv_sec;
1064 	tvp->tv_usec += bt->tv_nsec / 1000;
1065 	while (tvp->tv_usec >= 1000000) {
1066 		tvp->tv_usec -= 1000000;
1067 		++tvp->tv_sec;
1068 	}
1069 }
1070 
1071 void
1072 getnanotime(struct timespec *tsp)
1073 {
1074 	struct globaldata *gd = mycpu;
1075 	struct timespec *bt;
1076 	sysclock_t delta;
1077 
1078 	do {
1079 		tsp->tv_sec = gd->gd_time_seconds;
1080 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1081 	} while (tsp->tv_sec != gd->gd_time_seconds);
1082 
1083 	if (delta >= sys_cputimer->freq) {
1084 		tsp->tv_sec += delta / sys_cputimer->freq;
1085 		delta %= sys_cputimer->freq;
1086 	}
1087 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1088 
1089 	bt = &basetime[basetime_index];
1090 	tsp->tv_sec += bt->tv_sec;
1091 	tsp->tv_nsec += bt->tv_nsec;
1092 	while (tsp->tv_nsec >= 1000000000) {
1093 		tsp->tv_nsec -= 1000000000;
1094 		++tsp->tv_sec;
1095 	}
1096 }
1097 
1098 static void
1099 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1100 {
1101 	struct globaldata *gd = mycpu;
1102 	sysclock_t delta;
1103 
1104 	do {
1105 		tsp->tv_sec = gd->gd_time_seconds;
1106 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1107 	} while (tsp->tv_sec != gd->gd_time_seconds);
1108 
1109 	if (delta >= sys_cputimer->freq) {
1110 		tsp->tv_sec += delta / sys_cputimer->freq;
1111 		delta %= sys_cputimer->freq;
1112 	}
1113 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1114 
1115 	tsp->tv_sec += nbt->tv_sec;
1116 	tsp->tv_nsec += nbt->tv_nsec;
1117 	while (tsp->tv_nsec >= 1000000000) {
1118 		tsp->tv_nsec -= 1000000000;
1119 		++tsp->tv_sec;
1120 	}
1121 }
1122 
1123 
1124 void
1125 microtime(struct timeval *tvp)
1126 {
1127 	struct globaldata *gd = mycpu;
1128 	struct timespec *bt;
1129 	sysclock_t delta;
1130 
1131 	do {
1132 		tvp->tv_sec = gd->gd_time_seconds;
1133 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1134 	} while (tvp->tv_sec != gd->gd_time_seconds);
1135 
1136 	if (delta >= sys_cputimer->freq) {
1137 		tvp->tv_sec += delta / sys_cputimer->freq;
1138 		delta %= sys_cputimer->freq;
1139 	}
1140 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1141 
1142 	bt = &basetime[basetime_index];
1143 	tvp->tv_sec += bt->tv_sec;
1144 	tvp->tv_usec += bt->tv_nsec / 1000;
1145 	while (tvp->tv_usec >= 1000000) {
1146 		tvp->tv_usec -= 1000000;
1147 		++tvp->tv_sec;
1148 	}
1149 }
1150 
1151 void
1152 nanotime(struct timespec *tsp)
1153 {
1154 	struct globaldata *gd = mycpu;
1155 	struct timespec *bt;
1156 	sysclock_t delta;
1157 
1158 	do {
1159 		tsp->tv_sec = gd->gd_time_seconds;
1160 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1161 	} while (tsp->tv_sec != gd->gd_time_seconds);
1162 
1163 	if (delta >= sys_cputimer->freq) {
1164 		tsp->tv_sec += delta / sys_cputimer->freq;
1165 		delta %= sys_cputimer->freq;
1166 	}
1167 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1168 
1169 	bt = &basetime[basetime_index];
1170 	tsp->tv_sec += bt->tv_sec;
1171 	tsp->tv_nsec += bt->tv_nsec;
1172 	while (tsp->tv_nsec >= 1000000000) {
1173 		tsp->tv_nsec -= 1000000000;
1174 		++tsp->tv_sec;
1175 	}
1176 }
1177 
1178 /*
1179  * note: this is not exactly synchronized with real time.  To do that we
1180  * would have to do what microtime does and check for a nanoseconds overflow.
1181  */
1182 time_t
1183 get_approximate_time_t(void)
1184 {
1185 	struct globaldata *gd = mycpu;
1186 	struct timespec *bt;
1187 
1188 	bt = &basetime[basetime_index];
1189 	return(gd->gd_time_seconds + bt->tv_sec);
1190 }
1191 
1192 int
1193 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1194 {
1195 	pps_params_t *app;
1196 	struct pps_fetch_args *fapi;
1197 #ifdef PPS_SYNC
1198 	struct pps_kcbind_args *kapi;
1199 #endif
1200 
1201 	switch (cmd) {
1202 	case PPS_IOC_CREATE:
1203 		return (0);
1204 	case PPS_IOC_DESTROY:
1205 		return (0);
1206 	case PPS_IOC_SETPARAMS:
1207 		app = (pps_params_t *)data;
1208 		if (app->mode & ~pps->ppscap)
1209 			return (EINVAL);
1210 		pps->ppsparam = *app;
1211 		return (0);
1212 	case PPS_IOC_GETPARAMS:
1213 		app = (pps_params_t *)data;
1214 		*app = pps->ppsparam;
1215 		app->api_version = PPS_API_VERS_1;
1216 		return (0);
1217 	case PPS_IOC_GETCAP:
1218 		*(int*)data = pps->ppscap;
1219 		return (0);
1220 	case PPS_IOC_FETCH:
1221 		fapi = (struct pps_fetch_args *)data;
1222 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1223 			return (EINVAL);
1224 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1225 			return (EOPNOTSUPP);
1226 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
1227 		fapi->pps_info_buf = pps->ppsinfo;
1228 		return (0);
1229 	case PPS_IOC_KCBIND:
1230 #ifdef PPS_SYNC
1231 		kapi = (struct pps_kcbind_args *)data;
1232 		/* XXX Only root should be able to do this */
1233 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1234 			return (EINVAL);
1235 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1236 			return (EINVAL);
1237 		if (kapi->edge & ~pps->ppscap)
1238 			return (EINVAL);
1239 		pps->kcmode = kapi->edge;
1240 		return (0);
1241 #else
1242 		return (EOPNOTSUPP);
1243 #endif
1244 	default:
1245 		return (ENOTTY);
1246 	}
1247 }
1248 
1249 void
1250 pps_init(struct pps_state *pps)
1251 {
1252 	pps->ppscap |= PPS_TSFMT_TSPEC;
1253 	if (pps->ppscap & PPS_CAPTUREASSERT)
1254 		pps->ppscap |= PPS_OFFSETASSERT;
1255 	if (pps->ppscap & PPS_CAPTURECLEAR)
1256 		pps->ppscap |= PPS_OFFSETCLEAR;
1257 }
1258 
1259 void
1260 pps_event(struct pps_state *pps, sysclock_t count, int event)
1261 {
1262 	struct globaldata *gd;
1263 	struct timespec *tsp;
1264 	struct timespec *osp;
1265 	struct timespec *bt;
1266 	struct timespec ts;
1267 	sysclock_t *pcount;
1268 #ifdef PPS_SYNC
1269 	sysclock_t tcount;
1270 #endif
1271 	sysclock_t delta;
1272 	pps_seq_t *pseq;
1273 	int foff;
1274 	int fhard;
1275 
1276 	gd = mycpu;
1277 
1278 	/* Things would be easier with arrays... */
1279 	if (event == PPS_CAPTUREASSERT) {
1280 		tsp = &pps->ppsinfo.assert_timestamp;
1281 		osp = &pps->ppsparam.assert_offset;
1282 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1283 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1284 		pcount = &pps->ppscount[0];
1285 		pseq = &pps->ppsinfo.assert_sequence;
1286 	} else {
1287 		tsp = &pps->ppsinfo.clear_timestamp;
1288 		osp = &pps->ppsparam.clear_offset;
1289 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1290 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1291 		pcount = &pps->ppscount[1];
1292 		pseq = &pps->ppsinfo.clear_sequence;
1293 	}
1294 
1295 	/* Nothing really happened */
1296 	if (*pcount == count)
1297 		return;
1298 
1299 	*pcount = count;
1300 
1301 	do {
1302 		ts.tv_sec = gd->gd_time_seconds;
1303 		delta = count - gd->gd_cpuclock_base;
1304 	} while (ts.tv_sec != gd->gd_time_seconds);
1305 
1306 	if (delta >= sys_cputimer->freq) {
1307 		ts.tv_sec += delta / sys_cputimer->freq;
1308 		delta %= sys_cputimer->freq;
1309 	}
1310 	ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1311 	bt = &basetime[basetime_index];
1312 	ts.tv_sec += bt->tv_sec;
1313 	ts.tv_nsec += bt->tv_nsec;
1314 	while (ts.tv_nsec >= 1000000000) {
1315 		ts.tv_nsec -= 1000000000;
1316 		++ts.tv_sec;
1317 	}
1318 
1319 	(*pseq)++;
1320 	*tsp = ts;
1321 
1322 	if (foff) {
1323 		timespecadd(tsp, osp);
1324 		if (tsp->tv_nsec < 0) {
1325 			tsp->tv_nsec += 1000000000;
1326 			tsp->tv_sec -= 1;
1327 		}
1328 	}
1329 #ifdef PPS_SYNC
1330 	if (fhard) {
1331 		/* magic, at its best... */
1332 		tcount = count - pps->ppscount[2];
1333 		pps->ppscount[2] = count;
1334 		if (tcount >= sys_cputimer->freq) {
1335 			delta = (1000000000 * (tcount / sys_cputimer->freq) +
1336 				 sys_cputimer->freq64_nsec *
1337 				 (tcount % sys_cputimer->freq)) >> 32;
1338 		} else {
1339 			delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1340 		}
1341 		hardpps(tsp, delta);
1342 	}
1343 #endif
1344 }
1345 
1346 /*
1347  * Return the tsc target value for a delay of (ns).
1348  *
1349  * Returns -1 if the TSC is not supported.
1350  */
1351 int64_t
1352 tsc_get_target(int ns)
1353 {
1354 #if defined(_RDTSC_SUPPORTED_)
1355 	if (cpu_feature & CPUID_TSC) {
1356 		return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1357 	}
1358 #endif
1359 	return(-1);
1360 }
1361 
1362 /*
1363  * Compare the tsc against the passed target
1364  *
1365  * Returns +1 if the target has been reached
1366  * Returns  0 if the target has not yet been reached
1367  * Returns -1 if the TSC is not supported.
1368  *
1369  * Typical use:		while (tsc_test_target(target) == 0) { ...poll... }
1370  */
1371 int
1372 tsc_test_target(int64_t target)
1373 {
1374 #if defined(_RDTSC_SUPPORTED_)
1375 	if (cpu_feature & CPUID_TSC) {
1376 		if ((int64_t)(target - rdtsc()) <= 0)
1377 			return(1);
1378 		return(0);
1379 	}
1380 #endif
1381 	return(-1);
1382 }
1383