xref: /dflybsd-src/sys/kern/kern_clock.c (revision c7c26905e41d70fe0ea9d02c5af1fc0a688408d4)
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/resource.h>
89 #include <sys/resourcevar.h>
90 #include <sys/signalvar.h>
91 #include <sys/timex.h>
92 #include <sys/timepps.h>
93 #include <vm/vm.h>
94 #include <sys/lock.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_extern.h>
98 #include <sys/sysctl.h>
99 
100 #include <sys/thread2.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 static int
161 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
162 {
163 	long cpu_states[5] = {0};
164 	int cpu, error = 0;
165 	size_t size = sizeof(cpu_states);
166 
167 	for (cpu = 0; cpu < ncpus; ++cpu) {
168 		cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
169 		cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
170 		cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
171 		cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
172 		cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
173 	}
174 
175 	error = SYSCTL_OUT(req, cpu_states, size);
176 
177 	return (error);
178 }
179 
180 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
181 	sysctl_cp_time, "LU", "CPU time statistics");
182 
183 /*
184  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
185  * microuptime().  microtime() is not drift compensated.  The real uptime
186  * with compensation is nanotime() - bootime.  boottime is recalculated
187  * whenever the real time is set based on the compensated elapsed time
188  * in seconds (gd->gd_time_seconds).
189  *
190  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
191  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
192  * the real time.
193  */
194 struct timespec boottime;	/* boot time (realtime) for reference only */
195 time_t time_second;		/* read-only 'passive' uptime in seconds */
196 
197 /*
198  * basetime is used to calculate the compensated real time of day.  The
199  * basetime can be modified on a per-tick basis by the adjtime(),
200  * ntp_adjtime(), and sysctl-based time correction APIs.
201  *
202  * Note that frequency corrections can also be made by adjusting
203  * gd_cpuclock_base.
204  *
205  * basetime is a tail-chasing FIFO, updated only by cpu #0.  The FIFO is
206  * used on both SMP and UP systems to avoid MP races between cpu's and
207  * interrupt races on UP systems.
208  */
209 #define BASETIME_ARYSIZE	16
210 #define BASETIME_ARYMASK	(BASETIME_ARYSIZE - 1)
211 static struct timespec basetime[BASETIME_ARYSIZE];
212 static volatile int basetime_index;
213 
214 static int
215 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
216 {
217 	struct timespec *bt;
218 	int error;
219 	int index;
220 
221 	/*
222 	 * Because basetime data and index may be updated by another cpu,
223 	 * a load fence is required to ensure that the data we read has
224 	 * not been speculatively read relative to a possibly updated index.
225 	 */
226 	index = basetime_index;
227 	cpu_lfence();
228 	bt = &basetime[index];
229 	error = SYSCTL_OUT(req, bt, sizeof(*bt));
230 	return (error);
231 }
232 
233 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
234     &boottime, timespec, "System boottime");
235 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
236     sysctl_get_basetime, "S,timespec", "System basetime");
237 
238 static void hardclock(systimer_t info, int, struct intrframe *frame);
239 static void statclock(systimer_t info, int, struct intrframe *frame);
240 static void schedclock(systimer_t info, int, struct intrframe *frame);
241 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
242 
243 int	ticks;			/* system master ticks at hz */
244 int	clocks_running;		/* tsleep/timeout clocks operational */
245 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
246 int64_t	nsec_acc;		/* accumulator */
247 
248 /* NTPD time correction fields */
249 int64_t	ntp_tick_permanent;	/* per-tick adjustment in nsec << 32 */
250 int64_t	ntp_tick_acc;		/* accumulator for per-tick adjustment */
251 int64_t	ntp_delta;		/* one-time correction in nsec */
252 int64_t ntp_big_delta = 1000000000;
253 int32_t	ntp_tick_delta;		/* current adjustment rate */
254 int32_t	ntp_default_tick_delta;	/* adjustment rate for ntp_delta */
255 time_t	ntp_leap_second;	/* time of next leap second */
256 int	ntp_leap_insert;	/* whether to insert or remove a second */
257 
258 /*
259  * Finish initializing clock frequencies and start all clocks running.
260  */
261 /* ARGSUSED*/
262 static void
263 initclocks(void *dummy)
264 {
265 	/*psratio = profhz / stathz;*/
266 	initclocks_pcpu();
267 	clocks_running = 1;
268 }
269 
270 /*
271  * Called on a per-cpu basis
272  */
273 void
274 initclocks_pcpu(void)
275 {
276 	struct globaldata *gd = mycpu;
277 
278 	crit_enter();
279 	if (gd->gd_cpuid == 0) {
280 	    gd->gd_time_seconds = 1;
281 	    gd->gd_cpuclock_base = sys_cputimer->count();
282 	} else {
283 	    /* XXX */
284 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
285 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
286 	}
287 
288 	systimer_intr_enable();
289 
290 #ifdef DEVICE_POLLING
291 	init_device_poll_pcpu(gd->gd_cpuid);
292 #endif
293 
294 #ifdef IFPOLL_ENABLE
295 	ifpoll_init_pcpu(gd->gd_cpuid);
296 #endif
297 
298 	/*
299 	 * Use a non-queued periodic systimer to prevent multiple ticks from
300 	 * building up if the sysclock jumps forward (8254 gets reset).  The
301 	 * sysclock will never jump backwards.  Our time sync is based on
302 	 * the actual sysclock, not the ticks count.
303 	 */
304 	systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
305 	systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
306 	/* XXX correct the frequency for scheduler / estcpu tests */
307 	systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
308 				NULL, ESTCPUFREQ);
309 	crit_exit();
310 }
311 
312 /*
313  * This sets the current real time of day.  Timespecs are in seconds and
314  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
315  * instead we adjust basetime so basetime + gd_* results in the current
316  * time of day.  This way the gd_* fields are guarenteed to represent
317  * a monotonically increasing 'uptime' value.
318  *
319  * When set_timeofday() is called from userland, the system call forces it
320  * onto cpu #0 since only cpu #0 can update basetime_index.
321  */
322 void
323 set_timeofday(struct timespec *ts)
324 {
325 	struct timespec *nbt;
326 	int ni;
327 
328 	/*
329 	 * XXX SMP / non-atomic basetime updates
330 	 */
331 	crit_enter();
332 	ni = (basetime_index + 1) & BASETIME_ARYMASK;
333 	nbt = &basetime[ni];
334 	nanouptime(nbt);
335 	nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
336 	nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
337 	if (nbt->tv_nsec < 0) {
338 	    nbt->tv_nsec += 1000000000;
339 	    --nbt->tv_sec;
340 	}
341 
342 	/*
343 	 * Note that basetime diverges from boottime as the clock drift is
344 	 * compensated for, so we cannot do away with boottime.  When setting
345 	 * the absolute time of day the drift is 0 (for an instant) and we
346 	 * can simply assign boottime to basetime.
347 	 *
348 	 * Note that nanouptime() is based on gd_time_seconds which is drift
349 	 * compensated up to a point (it is guarenteed to remain monotonically
350 	 * increasing).  gd_time_seconds is thus our best uptime guess and
351 	 * suitable for use in the boottime calculation.  It is already taken
352 	 * into account in the basetime calculation above.
353 	 */
354 	boottime.tv_sec = nbt->tv_sec;
355 	ntp_delta = 0;
356 
357 	/*
358 	 * We now have a new basetime, make sure all other cpus have it,
359 	 * then update the index.
360 	 */
361 	cpu_sfence();
362 	basetime_index = ni;
363 
364 	crit_exit();
365 }
366 
367 /*
368  * Each cpu has its own hardclock, but we only increments ticks and softticks
369  * on cpu #0.
370  *
371  * NOTE! systimer! the MP lock might not be held here.  We can only safely
372  * manipulate objects owned by the current cpu.
373  */
374 static void
375 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
376 {
377 	sysclock_t cputicks;
378 	struct proc *p;
379 	struct globaldata *gd = mycpu;
380 
381 	/*
382 	 * Realtime updates are per-cpu.  Note that timer corrections as
383 	 * returned by microtime() and friends make an additional adjustment
384 	 * using a system-wise 'basetime', but the running time is always
385 	 * taken from the per-cpu globaldata area.  Since the same clock
386 	 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
387 	 * stay in synch.
388 	 *
389 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
390 	 * to reverse index gd_cpuclock_base, but that it is possible for
391 	 * it to temporarily get behind in the seconds if something in the
392 	 * system locks interrupts for a long period of time.  Since periodic
393 	 * timers count events, though everything should resynch again
394 	 * immediately.
395 	 */
396 	cputicks = info->time - gd->gd_cpuclock_base;
397 	if (cputicks >= sys_cputimer->freq) {
398 		++gd->gd_time_seconds;
399 		gd->gd_cpuclock_base += sys_cputimer->freq;
400 	}
401 
402 	/*
403 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
404 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
405 	 * by updating basetime.
406 	 */
407 	if (gd->gd_cpuid == 0) {
408 	    struct timespec *nbt;
409 	    struct timespec nts;
410 	    int leap;
411 	    int ni;
412 
413 	    ++ticks;
414 
415 #if 0
416 	    if (tco->tc_poll_pps)
417 		tco->tc_poll_pps(tco);
418 #endif
419 
420 	    /*
421 	     * Calculate the new basetime index.  We are in a critical section
422 	     * on cpu #0 and can safely play with basetime_index.  Start
423 	     * with the current basetime and then make adjustments.
424 	     */
425 	    ni = (basetime_index + 1) & BASETIME_ARYMASK;
426 	    nbt = &basetime[ni];
427 	    *nbt = basetime[basetime_index];
428 
429 	    /*
430 	     * Apply adjtime corrections.  (adjtime() API)
431 	     *
432 	     * adjtime() only runs on cpu #0 so our critical section is
433 	     * sufficient to access these variables.
434 	     */
435 	    if (ntp_delta != 0) {
436 		nbt->tv_nsec += ntp_tick_delta;
437 		ntp_delta -= ntp_tick_delta;
438 		if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
439 		    (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
440 			ntp_tick_delta = ntp_delta;
441  		}
442  	    }
443 
444 	    /*
445 	     * Apply permanent frequency corrections.  (sysctl API)
446 	     */
447 	    if (ntp_tick_permanent != 0) {
448 		ntp_tick_acc += ntp_tick_permanent;
449 		if (ntp_tick_acc >= (1LL << 32)) {
450 		    nbt->tv_nsec += ntp_tick_acc >> 32;
451 		    ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
452 		} else if (ntp_tick_acc <= -(1LL << 32)) {
453 		    /* Negate ntp_tick_acc to avoid shifting the sign bit. */
454 		    nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
455 		    ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
456 		}
457  	    }
458 
459 	    if (nbt->tv_nsec >= 1000000000) {
460 		    nbt->tv_sec++;
461 		    nbt->tv_nsec -= 1000000000;
462 	    } else if (nbt->tv_nsec < 0) {
463 		    nbt->tv_sec--;
464 		    nbt->tv_nsec += 1000000000;
465 	    }
466 
467 	    /*
468 	     * Another per-tick compensation.  (for ntp_adjtime() API)
469 	     */
470 	    if (nsec_adj != 0) {
471 		nsec_acc += nsec_adj;
472 		if (nsec_acc >= 0x100000000LL) {
473 		    nbt->tv_nsec += nsec_acc >> 32;
474 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
475 		} else if (nsec_acc <= -0x100000000LL) {
476 		    nbt->tv_nsec -= -nsec_acc >> 32;
477 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
478 		}
479 		if (nbt->tv_nsec >= 1000000000) {
480 		    nbt->tv_nsec -= 1000000000;
481 		    ++nbt->tv_sec;
482 		} else if (nbt->tv_nsec < 0) {
483 		    nbt->tv_nsec += 1000000000;
484 		    --nbt->tv_sec;
485 		}
486 	    }
487 
488 	    /************************************************************
489 	     *			LEAP SECOND CORRECTION			*
490 	     ************************************************************
491 	     *
492 	     * Taking into account all the corrections made above, figure
493 	     * out the new real time.  If the seconds field has changed
494 	     * then apply any pending leap-second corrections.
495 	     */
496 	    getnanotime_nbt(nbt, &nts);
497 
498 	    if (time_second != nts.tv_sec) {
499 		/*
500 		 * Apply leap second (sysctl API).  Adjust nts for changes
501 		 * so we do not have to call getnanotime_nbt again.
502 		 */
503 		if (ntp_leap_second) {
504 		    if (ntp_leap_second == nts.tv_sec) {
505 			if (ntp_leap_insert) {
506 			    nbt->tv_sec++;
507 			    nts.tv_sec++;
508 			} else {
509 			    nbt->tv_sec--;
510 			    nts.tv_sec--;
511 			}
512 			ntp_leap_second--;
513 		    }
514 		}
515 
516 		/*
517 		 * Apply leap second (ntp_adjtime() API), calculate a new
518 		 * nsec_adj field.  ntp_update_second() returns nsec_adj
519 		 * as a per-second value but we need it as a per-tick value.
520 		 */
521 		leap = ntp_update_second(time_second, &nsec_adj);
522 		nsec_adj /= hz;
523 		nbt->tv_sec += leap;
524 		nts.tv_sec += leap;
525 
526 		/*
527 		 * Update the time_second 'approximate time' global.
528 		 */
529 		time_second = nts.tv_sec;
530 	    }
531 
532 	    /*
533 	     * Finally, our new basetime is ready to go live!
534 	     */
535 	    cpu_sfence();
536 	    basetime_index = ni;
537 	}
538 
539 	/*
540 	 * lwkt thread scheduler fair queueing
541 	 */
542 	lwkt_schedulerclock(curthread);
543 
544 	/*
545 	 * softticks are handled for all cpus
546 	 */
547 	hardclock_softtick(gd);
548 
549 	/*
550 	 * ITimer handling is per-tick, per-cpu.
551 	 *
552 	 * We must acquire the per-process token in order for ksignal()
553 	 * to be non-blocking.  For the moment this requires an AST fault,
554 	 * the ksignal() cannot be safely issued from this hard interrupt.
555 	 *
556 	 * XXX Even the trytoken here isn't right, and itimer operation in
557 	 *     a multi threaded environment is going to be weird at the
558 	 *     very least.
559 	 */
560 	if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
561 		crit_enter_hard();
562 		if (frame && CLKF_USERMODE(frame) &&
563 		    timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
564 		    itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
565 			p->p_flags |= P_SIGVTALRM;
566 			need_user_resched();
567 		}
568 		if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
569 		    itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
570 			p->p_flags |= P_SIGPROF;
571 			need_user_resched();
572 		}
573 		crit_exit_hard();
574 		lwkt_reltoken(&p->p_token);
575 	}
576 	setdelayed();
577 }
578 
579 /*
580  * The statistics clock typically runs at a 125Hz rate, and is intended
581  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
582  *
583  * NOTE! systimer! the MP lock might not be held here.  We can only safely
584  * manipulate objects owned by the current cpu.
585  *
586  * The stats clock is responsible for grabbing a profiling sample.
587  * Most of the statistics are only used by user-level statistics programs.
588  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
589  * p->p_estcpu.
590  *
591  * Like the other clocks, the stat clock is called from what is effectively
592  * a fast interrupt, so the context should be the thread/process that got
593  * interrupted.
594  */
595 static void
596 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
597 {
598 #ifdef GPROF
599 	struct gmonparam *g;
600 	int i;
601 #endif
602 	thread_t td;
603 	struct proc *p;
604 	int bump;
605 	struct timeval tv;
606 	struct timeval *stv;
607 
608 	/*
609 	 * How big was our timeslice relative to the last time?
610 	 */
611 	microuptime(&tv);	/* mpsafe */
612 	stv = &mycpu->gd_stattv;
613 	if (stv->tv_sec == 0) {
614 	    bump = 1;
615 	} else {
616 	    bump = tv.tv_usec - stv->tv_usec +
617 		(tv.tv_sec - stv->tv_sec) * 1000000;
618 	    if (bump < 0)
619 		bump = 0;
620 	    if (bump > 1000000)
621 		bump = 1000000;
622 	}
623 	*stv = tv;
624 
625 	td = curthread;
626 	p = td->td_proc;
627 
628 	if (frame && CLKF_USERMODE(frame)) {
629 		/*
630 		 * Came from userland, handle user time and deal with
631 		 * possible process.
632 		 */
633 		if (p && (p->p_flags & P_PROFIL))
634 			addupc_intr(p, CLKF_PC(frame), 1);
635 		td->td_uticks += bump;
636 
637 		/*
638 		 * Charge the time as appropriate
639 		 */
640 		if (p && p->p_nice > NZERO)
641 			cpu_time.cp_nice += bump;
642 		else
643 			cpu_time.cp_user += bump;
644 	} else {
645 		int intr_nest = mycpu->gd_intr_nesting_level;
646 
647 		if (in_ipi) {
648 			/*
649 			 * IPI processing code will bump gd_intr_nesting_level
650 			 * up by one, which breaks following CLKF_INTR testing,
651 			 * so we substract it by one here.
652 			 */
653 			--intr_nest;
654 		}
655 #ifdef GPROF
656 		/*
657 		 * Kernel statistics are just like addupc_intr, only easier.
658 		 */
659 		g = &_gmonparam;
660 		if (g->state == GMON_PROF_ON && frame) {
661 			i = CLKF_PC(frame) - g->lowpc;
662 			if (i < g->textsize) {
663 				i /= HISTFRACTION * sizeof(*g->kcount);
664 				g->kcount[i]++;
665 			}
666 		}
667 #endif
668 
669 #define IS_INTR_RUNNING	((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
670 
671 		/*
672 		 * Came from kernel mode, so we were:
673 		 * - handling an interrupt,
674 		 * - doing syscall or trap work on behalf of the current
675 		 *   user process, or
676 		 * - spinning in the idle loop.
677 		 * Whichever it is, charge the time as appropriate.
678 		 * Note that we charge interrupts to the current process,
679 		 * regardless of whether they are ``for'' that process,
680 		 * so that we know how much of its real time was spent
681 		 * in ``non-process'' (i.e., interrupt) work.
682 		 *
683 		 * XXX assume system if frame is NULL.  A NULL frame
684 		 * can occur if ipi processing is done from a crit_exit().
685 		 */
686 		if (IS_INTR_RUNNING)
687 			td->td_iticks += bump;
688 		else
689 			td->td_sticks += bump;
690 
691 		if (IS_INTR_RUNNING) {
692 #ifdef DEBUG_PCTRACK
693 			if (frame)
694 				do_pctrack(frame, PCTRACK_INT);
695 #endif
696 			cpu_time.cp_intr += bump;
697 		} else {
698 			if (td == &mycpu->gd_idlethread) {
699 				cpu_time.cp_idle += bump;
700 			} else {
701 #ifdef DEBUG_PCTRACK
702 				if (frame)
703 					do_pctrack(frame, PCTRACK_SYS);
704 #endif
705 				cpu_time.cp_sys += bump;
706 			}
707 		}
708 
709 #undef IS_INTR_RUNNING
710 	}
711 }
712 
713 #ifdef DEBUG_PCTRACK
714 /*
715  * Sample the PC when in the kernel or in an interrupt.  User code can
716  * retrieve the information and generate a histogram or other output.
717  */
718 
719 static void
720 do_pctrack(struct intrframe *frame, int which)
721 {
722 	struct kinfo_pctrack *pctrack;
723 
724 	pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
725 	pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
726 		(void *)CLKF_PC(frame);
727 	++pctrack->pc_index;
728 }
729 
730 static int
731 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
732 {
733 	struct kinfo_pcheader head;
734 	int error;
735 	int cpu;
736 	int ntrack;
737 
738 	head.pc_ntrack = PCTRACK_SIZE;
739 	head.pc_arysize = PCTRACK_ARYSIZE;
740 
741 	if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
742 		return (error);
743 
744 	for (cpu = 0; cpu < ncpus; ++cpu) {
745 		for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
746 			error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
747 					   sizeof(struct kinfo_pctrack));
748 			if (error)
749 				break;
750 		}
751 		if (error)
752 			break;
753 	}
754 	return (error);
755 }
756 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
757 	sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
758 
759 #endif
760 
761 /*
762  * The scheduler clock typically runs at a 50Hz rate.  NOTE! systimer,
763  * the MP lock might not be held.  We can safely manipulate parts of curproc
764  * but that's about it.
765  *
766  * Each cpu has its own scheduler clock.
767  */
768 static void
769 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
770 {
771 	struct lwp *lp;
772 	struct rusage *ru;
773 	struct vmspace *vm;
774 	long rss;
775 
776 	if ((lp = lwkt_preempted_proc()) != NULL) {
777 		/*
778 		 * Account for cpu time used and hit the scheduler.  Note
779 		 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
780 		 * HERE.
781 		 */
782 		++lp->lwp_cpticks;
783 		lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
784 						       info->time);
785 	}
786 	if ((lp = curthread->td_lwp) != NULL) {
787 		/*
788 		 * Update resource usage integrals and maximums.
789 		 */
790 		if ((ru = &lp->lwp_proc->p_ru) &&
791 		    (vm = lp->lwp_proc->p_vmspace) != NULL) {
792 			ru->ru_ixrss += pgtok(vm->vm_tsize);
793 			ru->ru_idrss += pgtok(vm->vm_dsize);
794 			ru->ru_isrss += pgtok(vm->vm_ssize);
795 			if (lwkt_trytoken(&vm->vm_map.token)) {
796 				rss = pgtok(vmspace_resident_count(vm));
797 				if (ru->ru_maxrss < rss)
798 					ru->ru_maxrss = rss;
799 				lwkt_reltoken(&vm->vm_map.token);
800 			}
801 		}
802 	}
803 }
804 
805 /*
806  * Compute number of ticks for the specified amount of time.  The
807  * return value is intended to be used in a clock interrupt timed
808  * operation and guarenteed to meet or exceed the requested time.
809  * If the representation overflows, return INT_MAX.  The minimum return
810  * value is 1 ticks and the function will average the calculation up.
811  * If any value greater then 0 microseconds is supplied, a value
812  * of at least 2 will be returned to ensure that a near-term clock
813  * interrupt does not cause the timeout to occur (degenerately) early.
814  *
815  * Note that limit checks must take into account microseconds, which is
816  * done simply by using the smaller signed long maximum instead of
817  * the unsigned long maximum.
818  *
819  * If ints have 32 bits, then the maximum value for any timeout in
820  * 10ms ticks is 248 days.
821  */
822 int
823 tvtohz_high(struct timeval *tv)
824 {
825 	int ticks;
826 	long sec, usec;
827 
828 	sec = tv->tv_sec;
829 	usec = tv->tv_usec;
830 	if (usec < 0) {
831 		sec--;
832 		usec += 1000000;
833 	}
834 	if (sec < 0) {
835 #ifdef DIAGNOSTIC
836 		if (usec > 0) {
837 			sec++;
838 			usec -= 1000000;
839 		}
840 		kprintf("tvtohz_high: negative time difference "
841 			"%ld sec %ld usec\n",
842 			sec, usec);
843 #endif
844 		ticks = 1;
845 	} else if (sec <= INT_MAX / hz) {
846 		ticks = (int)(sec * hz +
847 			    ((u_long)usec + (ustick - 1)) / ustick) + 1;
848 	} else {
849 		ticks = INT_MAX;
850 	}
851 	return (ticks);
852 }
853 
854 int
855 tstohz_high(struct timespec *ts)
856 {
857 	int ticks;
858 	long sec, nsec;
859 
860 	sec = ts->tv_sec;
861 	nsec = ts->tv_nsec;
862 	if (nsec < 0) {
863 		sec--;
864 		nsec += 1000000000;
865 	}
866 	if (sec < 0) {
867 #ifdef DIAGNOSTIC
868 		if (nsec > 0) {
869 			sec++;
870 			nsec -= 1000000000;
871 		}
872 		kprintf("tstohz_high: negative time difference "
873 			"%ld sec %ld nsec\n",
874 			sec, nsec);
875 #endif
876 		ticks = 1;
877 	} else if (sec <= INT_MAX / hz) {
878 		ticks = (int)(sec * hz +
879 			    ((u_long)nsec + (nstick - 1)) / nstick) + 1;
880 	} else {
881 		ticks = INT_MAX;
882 	}
883 	return (ticks);
884 }
885 
886 
887 /*
888  * Compute number of ticks for the specified amount of time, erroring on
889  * the side of it being too low to ensure that sleeping the returned number
890  * of ticks will not result in a late return.
891  *
892  * The supplied timeval may not be negative and should be normalized.  A
893  * return value of 0 is possible if the timeval converts to less then
894  * 1 tick.
895  *
896  * If ints have 32 bits, then the maximum value for any timeout in
897  * 10ms ticks is 248 days.
898  */
899 int
900 tvtohz_low(struct timeval *tv)
901 {
902 	int ticks;
903 	long sec;
904 
905 	sec = tv->tv_sec;
906 	if (sec <= INT_MAX / hz)
907 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
908 	else
909 		ticks = INT_MAX;
910 	return (ticks);
911 }
912 
913 int
914 tstohz_low(struct timespec *ts)
915 {
916 	int ticks;
917 	long sec;
918 
919 	sec = ts->tv_sec;
920 	if (sec <= INT_MAX / hz)
921 		ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
922 	else
923 		ticks = INT_MAX;
924 	return (ticks);
925 }
926 
927 /*
928  * Start profiling on a process.
929  *
930  * Kernel profiling passes proc0 which never exits and hence
931  * keeps the profile clock running constantly.
932  */
933 void
934 startprofclock(struct proc *p)
935 {
936 	if ((p->p_flags & P_PROFIL) == 0) {
937 		p->p_flags |= P_PROFIL;
938 #if 0	/* XXX */
939 		if (++profprocs == 1 && stathz != 0) {
940 			crit_enter();
941 			psdiv = psratio;
942 			setstatclockrate(profhz);
943 			crit_exit();
944 		}
945 #endif
946 	}
947 }
948 
949 /*
950  * Stop profiling on a process.
951  *
952  * caller must hold p->p_token
953  */
954 void
955 stopprofclock(struct proc *p)
956 {
957 	if (p->p_flags & P_PROFIL) {
958 		p->p_flags &= ~P_PROFIL;
959 #if 0	/* XXX */
960 		if (--profprocs == 0 && stathz != 0) {
961 			crit_enter();
962 			psdiv = 1;
963 			setstatclockrate(stathz);
964 			crit_exit();
965 		}
966 #endif
967 	}
968 }
969 
970 /*
971  * Return information about system clocks.
972  */
973 static int
974 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
975 {
976 	struct kinfo_clockinfo clkinfo;
977 	/*
978 	 * Construct clockinfo structure.
979 	 */
980 	clkinfo.ci_hz = hz;
981 	clkinfo.ci_tick = ustick;
982 	clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
983 	clkinfo.ci_profhz = profhz;
984 	clkinfo.ci_stathz = stathz ? stathz : hz;
985 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
986 }
987 
988 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
989 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
990 
991 /*
992  * We have eight functions for looking at the clock, four for
993  * microseconds and four for nanoseconds.  For each there is fast
994  * but less precise version "get{nano|micro}[up]time" which will
995  * return a time which is up to 1/HZ previous to the call, whereas
996  * the raw version "{nano|micro}[up]time" will return a timestamp
997  * which is as precise as possible.  The "up" variants return the
998  * time relative to system boot, these are well suited for time
999  * interval measurements.
1000  *
1001  * Each cpu independantly maintains the current time of day, so all
1002  * we need to do to protect ourselves from changes is to do a loop
1003  * check on the seconds field changing out from under us.
1004  *
1005  * The system timer maintains a 32 bit count and due to various issues
1006  * it is possible for the calculated delta to occassionally exceed
1007  * sys_cputimer->freq.  If this occurs the sys_cputimer->freq64_nsec
1008  * multiplication can easily overflow, so we deal with the case.  For
1009  * uniformity we deal with the case in the usec case too.
1010  *
1011  * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1012  */
1013 void
1014 getmicrouptime(struct timeval *tvp)
1015 {
1016 	struct globaldata *gd = mycpu;
1017 	sysclock_t delta;
1018 
1019 	do {
1020 		tvp->tv_sec = gd->gd_time_seconds;
1021 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1022 	} while (tvp->tv_sec != gd->gd_time_seconds);
1023 
1024 	if (delta >= sys_cputimer->freq) {
1025 		tvp->tv_sec += delta / sys_cputimer->freq;
1026 		delta %= sys_cputimer->freq;
1027 	}
1028 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1029 	if (tvp->tv_usec >= 1000000) {
1030 		tvp->tv_usec -= 1000000;
1031 		++tvp->tv_sec;
1032 	}
1033 }
1034 
1035 void
1036 getnanouptime(struct timespec *tsp)
1037 {
1038 	struct globaldata *gd = mycpu;
1039 	sysclock_t delta;
1040 
1041 	do {
1042 		tsp->tv_sec = gd->gd_time_seconds;
1043 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1044 	} while (tsp->tv_sec != gd->gd_time_seconds);
1045 
1046 	if (delta >= sys_cputimer->freq) {
1047 		tsp->tv_sec += delta / sys_cputimer->freq;
1048 		delta %= sys_cputimer->freq;
1049 	}
1050 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1051 }
1052 
1053 void
1054 microuptime(struct timeval *tvp)
1055 {
1056 	struct globaldata *gd = mycpu;
1057 	sysclock_t delta;
1058 
1059 	do {
1060 		tvp->tv_sec = gd->gd_time_seconds;
1061 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1062 	} while (tvp->tv_sec != gd->gd_time_seconds);
1063 
1064 	if (delta >= sys_cputimer->freq) {
1065 		tvp->tv_sec += delta / sys_cputimer->freq;
1066 		delta %= sys_cputimer->freq;
1067 	}
1068 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1069 }
1070 
1071 void
1072 nanouptime(struct timespec *tsp)
1073 {
1074 	struct globaldata *gd = mycpu;
1075 	sysclock_t delta;
1076 
1077 	do {
1078 		tsp->tv_sec = gd->gd_time_seconds;
1079 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1080 	} while (tsp->tv_sec != gd->gd_time_seconds);
1081 
1082 	if (delta >= sys_cputimer->freq) {
1083 		tsp->tv_sec += delta / sys_cputimer->freq;
1084 		delta %= sys_cputimer->freq;
1085 	}
1086 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1087 }
1088 
1089 /*
1090  * realtime routines
1091  */
1092 void
1093 getmicrotime(struct timeval *tvp)
1094 {
1095 	struct globaldata *gd = mycpu;
1096 	struct timespec *bt;
1097 	sysclock_t delta;
1098 
1099 	do {
1100 		tvp->tv_sec = gd->gd_time_seconds;
1101 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1102 	} while (tvp->tv_sec != gd->gd_time_seconds);
1103 
1104 	if (delta >= sys_cputimer->freq) {
1105 		tvp->tv_sec += delta / sys_cputimer->freq;
1106 		delta %= sys_cputimer->freq;
1107 	}
1108 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1109 
1110 	bt = &basetime[basetime_index];
1111 	tvp->tv_sec += bt->tv_sec;
1112 	tvp->tv_usec += bt->tv_nsec / 1000;
1113 	while (tvp->tv_usec >= 1000000) {
1114 		tvp->tv_usec -= 1000000;
1115 		++tvp->tv_sec;
1116 	}
1117 }
1118 
1119 void
1120 getnanotime(struct timespec *tsp)
1121 {
1122 	struct globaldata *gd = mycpu;
1123 	struct timespec *bt;
1124 	sysclock_t delta;
1125 
1126 	do {
1127 		tsp->tv_sec = gd->gd_time_seconds;
1128 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1129 	} while (tsp->tv_sec != gd->gd_time_seconds);
1130 
1131 	if (delta >= sys_cputimer->freq) {
1132 		tsp->tv_sec += delta / sys_cputimer->freq;
1133 		delta %= sys_cputimer->freq;
1134 	}
1135 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1136 
1137 	bt = &basetime[basetime_index];
1138 	tsp->tv_sec += bt->tv_sec;
1139 	tsp->tv_nsec += bt->tv_nsec;
1140 	while (tsp->tv_nsec >= 1000000000) {
1141 		tsp->tv_nsec -= 1000000000;
1142 		++tsp->tv_sec;
1143 	}
1144 }
1145 
1146 static void
1147 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1148 {
1149 	struct globaldata *gd = mycpu;
1150 	sysclock_t delta;
1151 
1152 	do {
1153 		tsp->tv_sec = gd->gd_time_seconds;
1154 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1155 	} while (tsp->tv_sec != gd->gd_time_seconds);
1156 
1157 	if (delta >= sys_cputimer->freq) {
1158 		tsp->tv_sec += delta / sys_cputimer->freq;
1159 		delta %= sys_cputimer->freq;
1160 	}
1161 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1162 
1163 	tsp->tv_sec += nbt->tv_sec;
1164 	tsp->tv_nsec += nbt->tv_nsec;
1165 	while (tsp->tv_nsec >= 1000000000) {
1166 		tsp->tv_nsec -= 1000000000;
1167 		++tsp->tv_sec;
1168 	}
1169 }
1170 
1171 
1172 void
1173 microtime(struct timeval *tvp)
1174 {
1175 	struct globaldata *gd = mycpu;
1176 	struct timespec *bt;
1177 	sysclock_t delta;
1178 
1179 	do {
1180 		tvp->tv_sec = gd->gd_time_seconds;
1181 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1182 	} while (tvp->tv_sec != gd->gd_time_seconds);
1183 
1184 	if (delta >= sys_cputimer->freq) {
1185 		tvp->tv_sec += delta / sys_cputimer->freq;
1186 		delta %= sys_cputimer->freq;
1187 	}
1188 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1189 
1190 	bt = &basetime[basetime_index];
1191 	tvp->tv_sec += bt->tv_sec;
1192 	tvp->tv_usec += bt->tv_nsec / 1000;
1193 	while (tvp->tv_usec >= 1000000) {
1194 		tvp->tv_usec -= 1000000;
1195 		++tvp->tv_sec;
1196 	}
1197 }
1198 
1199 void
1200 nanotime(struct timespec *tsp)
1201 {
1202 	struct globaldata *gd = mycpu;
1203 	struct timespec *bt;
1204 	sysclock_t delta;
1205 
1206 	do {
1207 		tsp->tv_sec = gd->gd_time_seconds;
1208 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1209 	} while (tsp->tv_sec != gd->gd_time_seconds);
1210 
1211 	if (delta >= sys_cputimer->freq) {
1212 		tsp->tv_sec += delta / sys_cputimer->freq;
1213 		delta %= sys_cputimer->freq;
1214 	}
1215 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1216 
1217 	bt = &basetime[basetime_index];
1218 	tsp->tv_sec += bt->tv_sec;
1219 	tsp->tv_nsec += bt->tv_nsec;
1220 	while (tsp->tv_nsec >= 1000000000) {
1221 		tsp->tv_nsec -= 1000000000;
1222 		++tsp->tv_sec;
1223 	}
1224 }
1225 
1226 /*
1227  * note: this is not exactly synchronized with real time.  To do that we
1228  * would have to do what microtime does and check for a nanoseconds overflow.
1229  */
1230 time_t
1231 get_approximate_time_t(void)
1232 {
1233 	struct globaldata *gd = mycpu;
1234 	struct timespec *bt;
1235 
1236 	bt = &basetime[basetime_index];
1237 	return(gd->gd_time_seconds + bt->tv_sec);
1238 }
1239 
1240 int
1241 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1242 {
1243 	pps_params_t *app;
1244 	struct pps_fetch_args *fapi;
1245 #ifdef PPS_SYNC
1246 	struct pps_kcbind_args *kapi;
1247 #endif
1248 
1249 	switch (cmd) {
1250 	case PPS_IOC_CREATE:
1251 		return (0);
1252 	case PPS_IOC_DESTROY:
1253 		return (0);
1254 	case PPS_IOC_SETPARAMS:
1255 		app = (pps_params_t *)data;
1256 		if (app->mode & ~pps->ppscap)
1257 			return (EINVAL);
1258 		pps->ppsparam = *app;
1259 		return (0);
1260 	case PPS_IOC_GETPARAMS:
1261 		app = (pps_params_t *)data;
1262 		*app = pps->ppsparam;
1263 		app->api_version = PPS_API_VERS_1;
1264 		return (0);
1265 	case PPS_IOC_GETCAP:
1266 		*(int*)data = pps->ppscap;
1267 		return (0);
1268 	case PPS_IOC_FETCH:
1269 		fapi = (struct pps_fetch_args *)data;
1270 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1271 			return (EINVAL);
1272 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1273 			return (EOPNOTSUPP);
1274 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
1275 		fapi->pps_info_buf = pps->ppsinfo;
1276 		return (0);
1277 	case PPS_IOC_KCBIND:
1278 #ifdef PPS_SYNC
1279 		kapi = (struct pps_kcbind_args *)data;
1280 		/* XXX Only root should be able to do this */
1281 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1282 			return (EINVAL);
1283 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1284 			return (EINVAL);
1285 		if (kapi->edge & ~pps->ppscap)
1286 			return (EINVAL);
1287 		pps->kcmode = kapi->edge;
1288 		return (0);
1289 #else
1290 		return (EOPNOTSUPP);
1291 #endif
1292 	default:
1293 		return (ENOTTY);
1294 	}
1295 }
1296 
1297 void
1298 pps_init(struct pps_state *pps)
1299 {
1300 	pps->ppscap |= PPS_TSFMT_TSPEC;
1301 	if (pps->ppscap & PPS_CAPTUREASSERT)
1302 		pps->ppscap |= PPS_OFFSETASSERT;
1303 	if (pps->ppscap & PPS_CAPTURECLEAR)
1304 		pps->ppscap |= PPS_OFFSETCLEAR;
1305 }
1306 
1307 void
1308 pps_event(struct pps_state *pps, sysclock_t count, int event)
1309 {
1310 	struct globaldata *gd;
1311 	struct timespec *tsp;
1312 	struct timespec *osp;
1313 	struct timespec *bt;
1314 	struct timespec ts;
1315 	sysclock_t *pcount;
1316 #ifdef PPS_SYNC
1317 	sysclock_t tcount;
1318 #endif
1319 	sysclock_t delta;
1320 	pps_seq_t *pseq;
1321 	int foff;
1322 	int fhard;
1323 
1324 	gd = mycpu;
1325 
1326 	/* Things would be easier with arrays... */
1327 	if (event == PPS_CAPTUREASSERT) {
1328 		tsp = &pps->ppsinfo.assert_timestamp;
1329 		osp = &pps->ppsparam.assert_offset;
1330 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1331 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1332 		pcount = &pps->ppscount[0];
1333 		pseq = &pps->ppsinfo.assert_sequence;
1334 	} else {
1335 		tsp = &pps->ppsinfo.clear_timestamp;
1336 		osp = &pps->ppsparam.clear_offset;
1337 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1338 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1339 		pcount = &pps->ppscount[1];
1340 		pseq = &pps->ppsinfo.clear_sequence;
1341 	}
1342 
1343 	/* Nothing really happened */
1344 	if (*pcount == count)
1345 		return;
1346 
1347 	*pcount = count;
1348 
1349 	do {
1350 		ts.tv_sec = gd->gd_time_seconds;
1351 		delta = count - gd->gd_cpuclock_base;
1352 	} while (ts.tv_sec != gd->gd_time_seconds);
1353 
1354 	if (delta >= sys_cputimer->freq) {
1355 		ts.tv_sec += delta / sys_cputimer->freq;
1356 		delta %= sys_cputimer->freq;
1357 	}
1358 	ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1359 	bt = &basetime[basetime_index];
1360 	ts.tv_sec += bt->tv_sec;
1361 	ts.tv_nsec += bt->tv_nsec;
1362 	while (ts.tv_nsec >= 1000000000) {
1363 		ts.tv_nsec -= 1000000000;
1364 		++ts.tv_sec;
1365 	}
1366 
1367 	(*pseq)++;
1368 	*tsp = ts;
1369 
1370 	if (foff) {
1371 		timespecadd(tsp, osp);
1372 		if (tsp->tv_nsec < 0) {
1373 			tsp->tv_nsec += 1000000000;
1374 			tsp->tv_sec -= 1;
1375 		}
1376 	}
1377 #ifdef PPS_SYNC
1378 	if (fhard) {
1379 		/* magic, at its best... */
1380 		tcount = count - pps->ppscount[2];
1381 		pps->ppscount[2] = count;
1382 		if (tcount >= sys_cputimer->freq) {
1383 			delta = (1000000000 * (tcount / sys_cputimer->freq) +
1384 				 sys_cputimer->freq64_nsec *
1385 				 (tcount % sys_cputimer->freq)) >> 32;
1386 		} else {
1387 			delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1388 		}
1389 		hardpps(tsp, delta);
1390 	}
1391 #endif
1392 }
1393 
1394 /*
1395  * Return the tsc target value for a delay of (ns).
1396  *
1397  * Returns -1 if the TSC is not supported.
1398  */
1399 int64_t
1400 tsc_get_target(int ns)
1401 {
1402 #if defined(_RDTSC_SUPPORTED_)
1403 	if (cpu_feature & CPUID_TSC) {
1404 		return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1405 	}
1406 #endif
1407 	return(-1);
1408 }
1409 
1410 /*
1411  * Compare the tsc against the passed target
1412  *
1413  * Returns +1 if the target has been reached
1414  * Returns  0 if the target has not yet been reached
1415  * Returns -1 if the TSC is not supported.
1416  *
1417  * Typical use:		while (tsc_test_target(target) == 0) { ...poll... }
1418  */
1419 int
1420 tsc_test_target(int64_t target)
1421 {
1422 #if defined(_RDTSC_SUPPORTED_)
1423 	if (cpu_feature & CPUID_TSC) {
1424 		if ((int64_t)(target - rdtsc()) <= 0)
1425 			return(1);
1426 		return(0);
1427 	}
1428 #endif
1429 	return(-1);
1430 }
1431 
1432 /*
1433  * Delay the specified number of nanoseconds using the tsc.  This function
1434  * returns immediately if the TSC is not supported.  At least one cpu_pause()
1435  * will be issued.
1436  */
1437 void
1438 tsc_delay(int ns)
1439 {
1440 	int64_t clk;
1441 
1442 	clk = tsc_get_target(ns);
1443 	cpu_pause();
1444 	while (tsc_test_target(clk) == 0)
1445 		cpu_pause();
1446 }
1447