xref: /dflybsd-src/sys/kern/kern_clock.c (revision ed183f8c2f9bb14cbccb8377f3cdd29e0971d8a0)
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. Neither the name of the University nor the names of its contributors
52  *    may be used to endorse or promote products derived from this software
53  *    without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
56  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
59  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65  * SUCH DAMAGE.
66  *
67  *	@(#)kern_clock.c	8.5 (Berkeley) 1/21/94
68  * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $
69  */
70 
71 #include "opt_ntp.h"
72 #include "opt_pctrack.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/callout.h>
77 #include <sys/kernel.h>
78 #include <sys/kinfo.h>
79 #include <sys/proc.h>
80 #include <sys/malloc.h>
81 #include <sys/resource.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/priv.h>
85 #include <sys/timex.h>
86 #include <sys/timepps.h>
87 #include <sys/upmap.h>
88 #include <sys/lock.h>
89 #include <sys/sysctl.h>
90 #include <sys/kcollect.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_extern.h>
96 
97 #include <sys/thread2.h>
98 #include <sys/spinlock2.h>
99 
100 #include <machine/cpu.h>
101 #include <machine/limits.h>
102 #include <machine/smp.h>
103 #include <machine/cpufunc.h>
104 #include <machine/specialreg.h>
105 #include <machine/clock.h>
106 
107 #ifdef DEBUG_PCTRACK
108 static void do_pctrack(struct intrframe *frame, int which);
109 #endif
110 
111 static void initclocks (void *dummy);
112 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
113 
114 /*
115  * Some of these don't belong here, but it's easiest to concentrate them.
116  * Note that cpu_time counts in microseconds, but most userland programs
117  * just compare relative times against the total by delta.
118  */
119 struct kinfo_cputime cputime_percpu[MAXCPU];
120 #ifdef DEBUG_PCTRACK
121 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
122 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
123 #endif
124 
125 static int sniff_enable = 1;
126 static int sniff_target = -1;
127 SYSCTL_INT(_kern, OID_AUTO, sniff_enable, CTLFLAG_RW, &sniff_enable, 0 , "");
128 SYSCTL_INT(_kern, OID_AUTO, sniff_target, CTLFLAG_RW, &sniff_target, 0 , "");
129 
130 static int
131 sysctl_cputime(SYSCTL_HANDLER_ARGS)
132 {
133 	int cpu, error = 0;
134 	int root_error;
135 	size_t size = sizeof(struct kinfo_cputime);
136 	struct kinfo_cputime tmp;
137 
138 	/*
139 	 * NOTE: For security reasons, only root can sniff %rip
140 	 */
141 	root_error = priv_check_cred(curthread->td_ucred, PRIV_ROOT, 0);
142 
143 	for (cpu = 0; cpu < ncpus; ++cpu) {
144 		tmp = cputime_percpu[cpu];
145 		if (root_error == 0) {
146 			tmp.cp_sample_pc =
147 				(int64_t)globaldata_find(cpu)->gd_sample_pc;
148 			tmp.cp_sample_sp =
149 				(int64_t)globaldata_find(cpu)->gd_sample_sp;
150 		}
151 		if ((error = SYSCTL_OUT(req, &tmp, size)) != 0)
152 			break;
153 	}
154 
155 	if (root_error == 0) {
156 		if (sniff_enable) {
157 			int n = sniff_target;
158 			if (n < 0)
159 				smp_sniff();
160 			else if (n < ncpus)
161 				cpu_sniff(n);
162 		}
163 	}
164 
165 	return (error);
166 }
167 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
168 	sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
169 
170 static int
171 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
172 {
173 	long cpu_states[CPUSTATES] = {0};
174 	int cpu, error = 0;
175 	size_t size = sizeof(cpu_states);
176 
177 	for (cpu = 0; cpu < ncpus; ++cpu) {
178 		cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
179 		cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
180 		cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
181 		cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
182 		cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
183 	}
184 
185 	error = SYSCTL_OUT(req, cpu_states, size);
186 
187 	return (error);
188 }
189 
190 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
191     sysctl_cp_time, "LU", "CPU time statistics");
192 
193 static int
194 sysctl_cp_times(SYSCTL_HANDLER_ARGS)
195 {
196 	long cpu_states[CPUSTATES] = {0};
197 	int cpu, error;
198 	size_t size = sizeof(cpu_states);
199 
200 	for (error = 0, cpu = 0; error == 0 && cpu < ncpus; ++cpu) {
201 		cpu_states[CP_USER] = cputime_percpu[cpu].cp_user;
202 		cpu_states[CP_NICE] = cputime_percpu[cpu].cp_nice;
203 		cpu_states[CP_SYS] = cputime_percpu[cpu].cp_sys;
204 		cpu_states[CP_INTR] = cputime_percpu[cpu].cp_intr;
205 		cpu_states[CP_IDLE] = cputime_percpu[cpu].cp_idle;
206 		error = SYSCTL_OUT(req, cpu_states, size);
207 	}
208 
209 	return (error);
210 }
211 
212 SYSCTL_PROC(_kern, OID_AUTO, cp_times, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
213     sysctl_cp_times, "LU", "per-CPU time statistics");
214 
215 /*
216  * boottime is used to calculate the 'real' uptime.  Do not confuse this with
217  * microuptime().  microtime() is not drift compensated.  The real uptime
218  * with compensation is nanotime() - bootime.  boottime is recalculated
219  * whenever the real time is set based on the compensated elapsed time
220  * in seconds (gd->gd_time_seconds).
221  *
222  * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
223  * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
224  * the real time.
225  *
226  * WARNING! time_second can backstep on time corrections. Also, unlike
227  *          time_second, time_uptime is not a "real" time_t (seconds
228  *          since the Epoch) but seconds since booting.
229  */
230 struct timespec boottime;	/* boot time (realtime) for reference only */
231 time_t time_second;		/* read-only 'passive' realtime in seconds */
232 time_t time_uptime;		/* read-only 'passive' uptime in seconds */
233 
234 /*
235  * basetime is used to calculate the compensated real time of day.  The
236  * basetime can be modified on a per-tick basis by the adjtime(),
237  * ntp_adjtime(), and sysctl-based time correction APIs.
238  *
239  * Note that frequency corrections can also be made by adjusting
240  * gd_cpuclock_base.
241  *
242  * basetime is a tail-chasing FIFO, updated only by cpu #0.  The FIFO is
243  * used on both SMP and UP systems to avoid MP races between cpu's and
244  * interrupt races on UP systems.
245  */
246 struct hardtime {
247 	__uint32_t time_second;
248 	sysclock_t cpuclock_base;
249 };
250 
251 #define BASETIME_ARYSIZE	16
252 #define BASETIME_ARYMASK	(BASETIME_ARYSIZE - 1)
253 static struct timespec basetime[BASETIME_ARYSIZE];
254 static struct hardtime hardtime[BASETIME_ARYSIZE];
255 static volatile int basetime_index;
256 
257 static int
258 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
259 {
260 	struct timespec *bt;
261 	int error;
262 	int index;
263 
264 	/*
265 	 * Because basetime data and index may be updated by another cpu,
266 	 * a load fence is required to ensure that the data we read has
267 	 * not been speculatively read relative to a possibly updated index.
268 	 */
269 	index = basetime_index;
270 	cpu_lfence();
271 	bt = &basetime[index];
272 	error = SYSCTL_OUT(req, bt, sizeof(*bt));
273 	return (error);
274 }
275 
276 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
277     &boottime, timespec, "System boottime");
278 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
279     sysctl_get_basetime, "S,timespec", "System basetime");
280 
281 static void hardclock(systimer_t info, int, struct intrframe *frame);
282 static void statclock(systimer_t info, int, struct intrframe *frame);
283 static void schedclock(systimer_t info, int, struct intrframe *frame);
284 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
285 
286 int	ticks;			/* system master ticks at hz */
287 int	clocks_running;		/* tsleep/timeout clocks operational */
288 int64_t	nsec_adj;		/* ntpd per-tick adjustment in nsec << 32 */
289 int64_t	nsec_acc;		/* accumulator */
290 int	sched_ticks;		/* global schedule clock ticks */
291 
292 /* NTPD time correction fields */
293 int64_t	ntp_tick_permanent;	/* per-tick adjustment in nsec << 32 */
294 int64_t	ntp_tick_acc;		/* accumulator for per-tick adjustment */
295 int64_t	ntp_delta;		/* one-time correction in nsec */
296 int64_t ntp_big_delta = 1000000000;
297 int32_t	ntp_tick_delta;		/* current adjustment rate */
298 int32_t	ntp_default_tick_delta;	/* adjustment rate for ntp_delta */
299 time_t	ntp_leap_second;	/* time of next leap second */
300 int	ntp_leap_insert;	/* whether to insert or remove a second */
301 struct spinlock ntp_spin;
302 
303 /*
304  * Finish initializing clock frequencies and start all clocks running.
305  */
306 /* ARGSUSED*/
307 static void
308 initclocks(void *dummy)
309 {
310 	/*psratio = profhz / stathz;*/
311 	spin_init(&ntp_spin, "ntp");
312 	initclocks_pcpu();
313 	clocks_running = 1;
314 	if (kpmap) {
315 	    kpmap->tsc_freq = tsc_frequency;
316 	    kpmap->tick_freq = hz;
317 	}
318 }
319 
320 /*
321  * Called on a per-cpu basis from the idle thread bootstrap on each cpu
322  * during SMP initialization.
323  *
324  * This routine is called concurrently during low-level SMP initialization
325  * and may not block in any way.  Meaning, among other things, we can't
326  * acquire any tokens.
327  */
328 void
329 initclocks_pcpu(void)
330 {
331 	struct globaldata *gd = mycpu;
332 
333 	crit_enter();
334 	if (gd->gd_cpuid == 0) {
335 	    gd->gd_time_seconds = 1;
336 	    gd->gd_cpuclock_base = sys_cputimer->count();
337 	    hardtime[0].time_second = gd->gd_time_seconds;
338 	    hardtime[0].cpuclock_base = gd->gd_cpuclock_base;
339 	} else {
340 	    gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
341 	    gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
342 	}
343 
344 	systimer_intr_enable();
345 
346 	crit_exit();
347 }
348 
349 /*
350  * Called on a 10-second interval after the system is operational.
351  * Return the collection data for USERPCT and install the data for
352  * SYSTPCT and IDLEPCT.
353  */
354 static
355 uint64_t
356 collect_cputime_callback(int n)
357 {
358 	static long cpu_base[CPUSTATES];
359 	long cpu_states[CPUSTATES];
360 	long total;
361 	long acc;
362 	long lsb;
363 
364 	bzero(cpu_states, sizeof(cpu_states));
365 	for (n = 0; n < ncpus; ++n) {
366 		cpu_states[CP_USER] += cputime_percpu[n].cp_user;
367 		cpu_states[CP_NICE] += cputime_percpu[n].cp_nice;
368 		cpu_states[CP_SYS] += cputime_percpu[n].cp_sys;
369 		cpu_states[CP_INTR] += cputime_percpu[n].cp_intr;
370 		cpu_states[CP_IDLE] += cputime_percpu[n].cp_idle;
371 	}
372 
373 	acc = 0;
374 	for (n = 0; n < CPUSTATES; ++n) {
375 		total = cpu_states[n] - cpu_base[n];
376 		cpu_base[n] = cpu_states[n];
377 		cpu_states[n] = total;
378 		acc += total;
379 	}
380 	if (acc == 0)		/* prevent degenerate divide by 0 */
381 		acc = 1;
382 	lsb = acc / (10000 * 2);
383 	kcollect_setvalue(KCOLLECT_SYSTPCT,
384 			  (cpu_states[CP_SYS] + lsb) * 10000 / acc);
385 	kcollect_setvalue(KCOLLECT_IDLEPCT,
386 			  (cpu_states[CP_IDLE] + lsb) * 10000 / acc);
387 	kcollect_setvalue(KCOLLECT_INTRPCT,
388 			  (cpu_states[CP_INTR] + lsb) * 10000 / acc);
389 	return((cpu_states[CP_USER] + cpu_states[CP_NICE] + lsb) * 10000 / acc);
390 }
391 
392 /*
393  * This routine is called on just the BSP, just after SMP initialization
394  * completes to * finish initializing any clocks that might contend/block
395  * (e.g. like on a token).  We can't do this in initclocks_pcpu() because
396  * that function is called from the idle thread bootstrap for each cpu and
397  * not allowed to block at all.
398  */
399 static
400 void
401 initclocks_other(void *dummy)
402 {
403 	struct globaldata *ogd = mycpu;
404 	struct globaldata *gd;
405 	int n;
406 
407 	for (n = 0; n < ncpus; ++n) {
408 		lwkt_setcpu_self(globaldata_find(n));
409 		gd = mycpu;
410 
411 		/*
412 		 * Use a non-queued periodic systimer to prevent multiple
413 		 * ticks from building up if the sysclock jumps forward
414 		 * (8254 gets reset).  The sysclock will never jump backwards.
415 		 * Our time sync is based on the actual sysclock, not the
416 		 * ticks count.
417 		 *
418 		 * Install statclock before hardclock to prevent statclock
419 		 * from misinterpreting gd_flags for tick assignment when
420 		 * they overlap.
421 		 */
422 		systimer_init_periodic_flags(&gd->gd_statclock, statclock,
423 					  NULL, stathz,
424 					  SYSTF_MSSYNC | SYSTF_FIRST);
425 		systimer_init_periodic_flags(&gd->gd_hardclock, hardclock,
426 					  NULL, hz, SYSTF_MSSYNC);
427 	}
428 	lwkt_setcpu_self(ogd);
429 
430 	/*
431 	 * Regular data collection
432 	 */
433 	kcollect_register(KCOLLECT_USERPCT, "user", collect_cputime_callback,
434 			  KCOLLECT_SCALE(KCOLLECT_USERPCT_FORMAT, 0));
435 	kcollect_register(KCOLLECT_SYSTPCT, "syst", NULL,
436 			  KCOLLECT_SCALE(KCOLLECT_SYSTPCT_FORMAT, 0));
437 	kcollect_register(KCOLLECT_IDLEPCT, "idle", NULL,
438 			  KCOLLECT_SCALE(KCOLLECT_IDLEPCT_FORMAT, 0));
439 }
440 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL);
441 
442 /*
443  * This method is called on just the BSP, after all the usched implementations
444  * are initialized. This avoids races between usched initialization functions
445  * and usched_schedulerclock().
446  */
447 static
448 void
449 initclocks_usched(void *dummy)
450 {
451 	struct globaldata *ogd = mycpu;
452 	struct globaldata *gd;
453 	int n;
454 
455 	for (n = 0; n < ncpus; ++n) {
456 		lwkt_setcpu_self(globaldata_find(n));
457 		gd = mycpu;
458 
459 		/* XXX correct the frequency for scheduler / estcpu tests */
460 		systimer_init_periodic_flags(&gd->gd_schedclock, schedclock,
461 					  NULL, ESTCPUFREQ, SYSTF_MSSYNC);
462 	}
463 	lwkt_setcpu_self(ogd);
464 }
465 SYSINIT(clocks3, SI_BOOT2_USCHED, SI_ORDER_ANY, initclocks_usched, NULL);
466 
467 /*
468  * This sets the current real time of day.  Timespecs are in seconds and
469  * nanoseconds.  We do not mess with gd_time_seconds and gd_cpuclock_base,
470  * instead we adjust basetime so basetime + gd_* results in the current
471  * time of day.  This way the gd_* fields are guaranteed to represent
472  * a monotonically increasing 'uptime' value.
473  *
474  * When set_timeofday() is called from userland, the system call forces it
475  * onto cpu #0 since only cpu #0 can update basetime_index.
476  */
477 void
478 set_timeofday(struct timespec *ts)
479 {
480 	struct timespec *nbt;
481 	int ni;
482 
483 	/*
484 	 * XXX SMP / non-atomic basetime updates
485 	 */
486 	crit_enter();
487 	ni = (basetime_index + 1) & BASETIME_ARYMASK;
488 	cpu_lfence();
489 	nbt = &basetime[ni];
490 	nanouptime(nbt);
491 	nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
492 	nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
493 	if (nbt->tv_nsec < 0) {
494 	    nbt->tv_nsec += 1000000000;
495 	    --nbt->tv_sec;
496 	}
497 
498 	/*
499 	 * Note that basetime diverges from boottime as the clock drift is
500 	 * compensated for, so we cannot do away with boottime.  When setting
501 	 * the absolute time of day the drift is 0 (for an instant) and we
502 	 * can simply assign boottime to basetime.
503 	 *
504 	 * Note that nanouptime() is based on gd_time_seconds which is drift
505 	 * compensated up to a point (it is guaranteed to remain monotonically
506 	 * increasing).  gd_time_seconds is thus our best uptime guess and
507 	 * suitable for use in the boottime calculation.  It is already taken
508 	 * into account in the basetime calculation above.
509 	 */
510 	spin_lock(&ntp_spin);
511 	boottime.tv_sec = nbt->tv_sec;
512 	ntp_delta = 0;
513 
514 	/*
515 	 * We now have a new basetime, make sure all other cpus have it,
516 	 * then update the index.
517 	 */
518 	cpu_sfence();
519 	basetime_index = ni;
520 	spin_unlock(&ntp_spin);
521 
522 	crit_exit();
523 }
524 
525 /*
526  * Each cpu has its own hardclock, but we only increment ticks and softticks
527  * on cpu #0.
528  *
529  * NOTE! systimer! the MP lock might not be held here.  We can only safely
530  * manipulate objects owned by the current cpu.
531  */
532 static void
533 hardclock(systimer_t info, int in_ipi, struct intrframe *frame)
534 {
535 	sysclock_t cputicks;
536 	struct proc *p;
537 	struct globaldata *gd = mycpu;
538 
539 	if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) {
540 		/* Defer to doreti on passive IPIQ processing */
541 		need_ipiq();
542 	}
543 
544 	/*
545 	 * We update the compensation base to calculate fine-grained time
546 	 * from the sys_cputimer on a per-cpu basis in order to avoid
547 	 * having to mess around with locks.  sys_cputimer is assumed to
548 	 * be consistent across all cpus.  CPU N copies the base state from
549 	 * CPU 0 using the same FIFO trick that we use for basetime (so we
550 	 * don't catch a CPU 0 update in the middle).
551 	 *
552 	 * Note that we never allow info->time (aka gd->gd_hardclock.time)
553 	 * to reverse index gd_cpuclock_base, but that it is possible for
554 	 * it to temporarily get behind in the seconds if something in the
555 	 * system locks interrupts for a long period of time.  Since periodic
556 	 * timers count events, though everything should resynch again
557 	 * immediately.
558 	 */
559 	if (gd->gd_cpuid == 0) {
560 		int ni;
561 
562 		cputicks = info->time - gd->gd_cpuclock_base;
563 		if (cputicks >= sys_cputimer->freq) {
564 			cputicks /= sys_cputimer->freq;
565 			if (cputicks != 0 && cputicks != 1)
566 				kprintf("Warning: hardclock missed > 1 sec\n");
567 			gd->gd_time_seconds += cputicks;
568 			gd->gd_cpuclock_base += sys_cputimer->freq * cputicks;
569 			/* uncorrected monotonic 1-sec gran */
570 			time_uptime += cputicks;
571 		}
572 		ni = (basetime_index + 1) & BASETIME_ARYMASK;
573 		hardtime[ni].time_second = gd->gd_time_seconds;
574 		hardtime[ni].cpuclock_base = gd->gd_cpuclock_base;
575 	} else {
576 		int ni;
577 
578 		ni = basetime_index;
579 		cpu_lfence();
580 		gd->gd_time_seconds = hardtime[ni].time_second;
581 		gd->gd_cpuclock_base = hardtime[ni].cpuclock_base;
582 	}
583 
584 	/*
585 	 * The system-wide ticks counter and NTP related timedelta/tickdelta
586 	 * adjustments only occur on cpu #0.  NTP adjustments are accomplished
587 	 * by updating basetime.
588 	 */
589 	if (gd->gd_cpuid == 0) {
590 	    struct timespec *nbt;
591 	    struct timespec nts;
592 	    int leap;
593 	    int ni;
594 
595 	    ++ticks;
596 
597 #if 0
598 	    if (tco->tc_poll_pps)
599 		tco->tc_poll_pps(tco);
600 #endif
601 
602 	    /*
603 	     * Calculate the new basetime index.  We are in a critical section
604 	     * on cpu #0 and can safely play with basetime_index.  Start
605 	     * with the current basetime and then make adjustments.
606 	     */
607 	    ni = (basetime_index + 1) & BASETIME_ARYMASK;
608 	    nbt = &basetime[ni];
609 	    *nbt = basetime[basetime_index];
610 
611 	    /*
612 	     * ntp adjustments only occur on cpu 0 and are protected by
613 	     * ntp_spin.  This spinlock virtually never conflicts.
614 	     */
615 	    spin_lock(&ntp_spin);
616 
617 	    /*
618 	     * Apply adjtime corrections.  (adjtime() API)
619 	     *
620 	     * adjtime() only runs on cpu #0 so our critical section is
621 	     * sufficient to access these variables.
622 	     */
623 	    if (ntp_delta != 0) {
624 		nbt->tv_nsec += ntp_tick_delta;
625 		ntp_delta -= ntp_tick_delta;
626 		if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
627 		    (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
628 			ntp_tick_delta = ntp_delta;
629  		}
630  	    }
631 
632 	    /*
633 	     * Apply permanent frequency corrections.  (sysctl API)
634 	     */
635 	    if (ntp_tick_permanent != 0) {
636 		ntp_tick_acc += ntp_tick_permanent;
637 		if (ntp_tick_acc >= (1LL << 32)) {
638 		    nbt->tv_nsec += ntp_tick_acc >> 32;
639 		    ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
640 		} else if (ntp_tick_acc <= -(1LL << 32)) {
641 		    /* Negate ntp_tick_acc to avoid shifting the sign bit. */
642 		    nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
643 		    ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
644 		}
645  	    }
646 
647 	    if (nbt->tv_nsec >= 1000000000) {
648 		    nbt->tv_sec++;
649 		    nbt->tv_nsec -= 1000000000;
650 	    } else if (nbt->tv_nsec < 0) {
651 		    nbt->tv_sec--;
652 		    nbt->tv_nsec += 1000000000;
653 	    }
654 
655 	    /*
656 	     * Another per-tick compensation.  (for ntp_adjtime() API)
657 	     */
658 	    if (nsec_adj != 0) {
659 		nsec_acc += nsec_adj;
660 		if (nsec_acc >= 0x100000000LL) {
661 		    nbt->tv_nsec += nsec_acc >> 32;
662 		    nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
663 		} else if (nsec_acc <= -0x100000000LL) {
664 		    nbt->tv_nsec -= -nsec_acc >> 32;
665 		    nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
666 		}
667 		if (nbt->tv_nsec >= 1000000000) {
668 		    nbt->tv_nsec -= 1000000000;
669 		    ++nbt->tv_sec;
670 		} else if (nbt->tv_nsec < 0) {
671 		    nbt->tv_nsec += 1000000000;
672 		    --nbt->tv_sec;
673 		}
674 	    }
675 	    spin_unlock(&ntp_spin);
676 
677 	    /************************************************************
678 	     *			LEAP SECOND CORRECTION			*
679 	     ************************************************************
680 	     *
681 	     * Taking into account all the corrections made above, figure
682 	     * out the new real time.  If the seconds field has changed
683 	     * then apply any pending leap-second corrections.
684 	     */
685 	    getnanotime_nbt(nbt, &nts);
686 
687 	    if (time_second != nts.tv_sec) {
688 		/*
689 		 * Apply leap second (sysctl API).  Adjust nts for changes
690 		 * so we do not have to call getnanotime_nbt again.
691 		 */
692 		if (ntp_leap_second) {
693 		    if (ntp_leap_second == nts.tv_sec) {
694 			if (ntp_leap_insert) {
695 			    nbt->tv_sec++;
696 			    nts.tv_sec++;
697 			} else {
698 			    nbt->tv_sec--;
699 			    nts.tv_sec--;
700 			}
701 			ntp_leap_second--;
702 		    }
703 		}
704 
705 		/*
706 		 * Apply leap second (ntp_adjtime() API), calculate a new
707 		 * nsec_adj field.  ntp_update_second() returns nsec_adj
708 		 * as a per-second value but we need it as a per-tick value.
709 		 */
710 		leap = ntp_update_second(time_second, &nsec_adj);
711 		nsec_adj /= hz;
712 		nbt->tv_sec += leap;
713 		nts.tv_sec += leap;
714 
715 		/*
716 		 * Update the time_second 'approximate time' global.
717 		 */
718 		time_second = nts.tv_sec;
719 
720 		/*
721 		 * Clear the IPC hint for the currently running thread once
722 		 * per second, allowing us to disconnect the hint from a
723 		 * thread which may no longer care.
724 		 */
725 		curthread->td_wakefromcpu = -1;
726 
727 	    }
728 
729 	    /*
730 	     * Finally, our new basetime is ready to go live!
731 	     */
732 	    cpu_sfence();
733 	    basetime_index = ni;
734 
735 	    /*
736 	     * Update kpmap on each tick.  TS updates are integrated with
737 	     * fences and upticks allowing userland to read the data
738 	     * deterministically.
739 	     */
740 	    if (kpmap) {
741 		int w;
742 
743 		w = (kpmap->upticks + 1) & 1;
744 		getnanouptime(&kpmap->ts_uptime[w]);
745 		getnanotime(&kpmap->ts_realtime[w]);
746 		cpu_sfence();
747 		++kpmap->upticks;
748 		cpu_sfence();
749 	    }
750 	}
751 
752 	/*
753 	 * lwkt thread scheduler fair queueing
754 	 */
755 	lwkt_schedulerclock(curthread);
756 
757 	/*
758 	 * softticks are handled for all cpus
759 	 */
760 	hardclock_softtick(gd);
761 
762 	/*
763 	 * Rollup accumulated vmstats, copy-back for critical path checks.
764 	 */
765 	vmstats_rollup_cpu(gd);
766 	vfscache_rollup_cpu(gd);
767 	mycpu->gd_vmstats = vmstats;
768 
769 	/*
770 	 * ITimer handling is per-tick, per-cpu.
771 	 *
772 	 * We must acquire the per-process token in order for ksignal()
773 	 * to be non-blocking.  For the moment this requires an AST fault,
774 	 * the ksignal() cannot be safely issued from this hard interrupt.
775 	 *
776 	 * XXX Even the trytoken here isn't right, and itimer operation in
777 	 *     a multi threaded environment is going to be weird at the
778 	 *     very least.
779 	 */
780 	if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
781 		crit_enter_hard();
782 		if (p->p_upmap)
783 			++p->p_upmap->runticks;
784 
785 		if (frame && CLKF_USERMODE(frame) &&
786 		    timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
787 		    itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
788 			p->p_flags |= P_SIGVTALRM;
789 			need_user_resched();
790 		}
791 		if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
792 		    itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
793 			p->p_flags |= P_SIGPROF;
794 			need_user_resched();
795 		}
796 		crit_exit_hard();
797 		lwkt_reltoken(&p->p_token);
798 	}
799 	setdelayed();
800 }
801 
802 /*
803  * The statistics clock typically runs at a 125Hz rate, and is intended
804  * to be frequency offset from the hardclock (typ 100Hz).  It is per-cpu.
805  *
806  * NOTE! systimer! the MP lock might not be held here.  We can only safely
807  * manipulate objects owned by the current cpu.
808  *
809  * The stats clock is responsible for grabbing a profiling sample.
810  * Most of the statistics are only used by user-level statistics programs.
811  * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
812  * p->p_estcpu.
813  *
814  * Like the other clocks, the stat clock is called from what is effectively
815  * a fast interrupt, so the context should be the thread/process that got
816  * interrupted.
817  */
818 static void
819 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
820 {
821 	globaldata_t gd = mycpu;
822 	thread_t td;
823 	struct proc *p;
824 	int bump;
825 	sysclock_t cv;
826 	sysclock_t scv;
827 
828 	/*
829 	 * How big was our timeslice relative to the last time?  Calculate
830 	 * in microseconds.
831 	 *
832 	 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
833 	 *	 during early boot.  Just use the systimer count to be nice
834 	 *	 to e.g. qemu.  The systimer has a better chance of being
835 	 *	 MPSAFE at early boot.
836 	 */
837 	cv = sys_cputimer->count();
838 	scv = gd->statint.gd_statcv;
839 	if (scv == 0) {
840 		bump = 1;
841 	} else {
842 		bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32;
843 		if (bump < 0)
844 			bump = 0;
845 		if (bump > 1000000)
846 			bump = 1000000;
847 	}
848 	gd->statint.gd_statcv = cv;
849 
850 #if 0
851 	stv = &gd->gd_stattv;
852 	if (stv->tv_sec == 0) {
853 	    bump = 1;
854 	} else {
855 	    bump = tv.tv_usec - stv->tv_usec +
856 		(tv.tv_sec - stv->tv_sec) * 1000000;
857 	    if (bump < 0)
858 		bump = 0;
859 	    if (bump > 1000000)
860 		bump = 1000000;
861 	}
862 	*stv = tv;
863 #endif
864 
865 	td = curthread;
866 	p = td->td_proc;
867 
868 	if (frame && CLKF_USERMODE(frame)) {
869 		/*
870 		 * Came from userland, handle user time and deal with
871 		 * possible process.
872 		 */
873 		if (p && (p->p_flags & P_PROFIL))
874 			addupc_intr(p, CLKF_PC(frame), 1);
875 		td->td_uticks += bump;
876 
877 		/*
878 		 * Charge the time as appropriate
879 		 */
880 		if (p && p->p_nice > NZERO)
881 			cpu_time.cp_nice += bump;
882 		else
883 			cpu_time.cp_user += bump;
884 	} else {
885 		int intr_nest = gd->gd_intr_nesting_level;
886 
887 		if (in_ipi) {
888 			/*
889 			 * IPI processing code will bump gd_intr_nesting_level
890 			 * up by one, which breaks following CLKF_INTR testing,
891 			 * so we subtract it by one here.
892 			 */
893 			--intr_nest;
894 		}
895 
896 #define IS_INTR_RUNNING	((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
897 
898 		/*
899 		 * Came from kernel mode, so we were:
900 		 * - handling an interrupt,
901 		 * - doing syscall or trap work on behalf of the current
902 		 *   user process, or
903 		 * - spinning in the idle loop.
904 		 * Whichever it is, charge the time as appropriate.
905 		 * Note that we charge interrupts to the current process,
906 		 * regardless of whether they are ``for'' that process,
907 		 * so that we know how much of its real time was spent
908 		 * in ``non-process'' (i.e., interrupt) work.
909 		 *
910 		 * XXX assume system if frame is NULL.  A NULL frame
911 		 * can occur if ipi processing is done from a crit_exit().
912 		 */
913 		if (IS_INTR_RUNNING ||
914 		    (gd->gd_reqflags & RQF_INTPEND)) {
915 			/*
916 			 * If we interrupted an interrupt thread, well,
917 			 * count it as interrupt time.
918 			 */
919 			td->td_iticks += bump;
920 #ifdef DEBUG_PCTRACK
921 			if (frame)
922 				do_pctrack(frame, PCTRACK_INT);
923 #endif
924 			cpu_time.cp_intr += bump;
925 		} else if (gd->gd_flags & GDF_VIRTUSER) {
926 			/*
927 			 * The vkernel doesn't do a good job providing trap
928 			 * frames that we can test.  If the GDF_VIRTUSER
929 			 * flag is set we probably interrupted user mode.
930 			 *
931 			 * We also use this flag on the host when entering
932 			 * VMM mode.
933 			 */
934 			td->td_uticks += bump;
935 
936 			/*
937 			 * Charge the time as appropriate
938 			 */
939 			if (p && p->p_nice > NZERO)
940 				cpu_time.cp_nice += bump;
941 			else
942 				cpu_time.cp_user += bump;
943 		} else {
944 			td->td_sticks += bump;
945 			if (td == &gd->gd_idlethread) {
946 				/*
947 				 * We want to count token contention as
948 				 * system time.  When token contention occurs
949 				 * the cpu may only be outside its critical
950 				 * section while switching through the idle
951 				 * thread.  In this situation, various flags
952 				 * will be set in gd_reqflags.
953 				 */
954 				if (gd->gd_reqflags & RQF_IDLECHECK_WK_MASK)
955 					cpu_time.cp_sys += bump;
956 				else
957 					cpu_time.cp_idle += bump;
958 			} else {
959 				/*
960 				 * System thread was running.
961 				 */
962 #ifdef DEBUG_PCTRACK
963 				if (frame)
964 					do_pctrack(frame, PCTRACK_SYS);
965 #endif
966 				cpu_time.cp_sys += bump;
967 			}
968 		}
969 
970 #undef IS_INTR_RUNNING
971 	}
972 }
973 
974 #ifdef DEBUG_PCTRACK
975 /*
976  * Sample the PC when in the kernel or in an interrupt.  User code can
977  * retrieve the information and generate a histogram or other output.
978  */
979 
980 static void
981 do_pctrack(struct intrframe *frame, int which)
982 {
983 	struct kinfo_pctrack *pctrack;
984 
985 	pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
986 	pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
987 		(void *)CLKF_PC(frame);
988 	++pctrack->pc_index;
989 }
990 
991 static int
992 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
993 {
994 	struct kinfo_pcheader head;
995 	int error;
996 	int cpu;
997 	int ntrack;
998 
999 	head.pc_ntrack = PCTRACK_SIZE;
1000 	head.pc_arysize = PCTRACK_ARYSIZE;
1001 
1002 	if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
1003 		return (error);
1004 
1005 	for (cpu = 0; cpu < ncpus; ++cpu) {
1006 		for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
1007 			error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
1008 					   sizeof(struct kinfo_pctrack));
1009 			if (error)
1010 				break;
1011 		}
1012 		if (error)
1013 			break;
1014 	}
1015 	return (error);
1016 }
1017 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
1018 	sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
1019 
1020 #endif
1021 
1022 /*
1023  * The scheduler clock typically runs at a 50Hz rate.  NOTE! systimer,
1024  * the MP lock might not be held.  We can safely manipulate parts of curproc
1025  * but that's about it.
1026  *
1027  * Each cpu has its own scheduler clock.
1028  */
1029 static void
1030 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
1031 {
1032 	struct lwp *lp;
1033 	struct rusage *ru;
1034 	struct vmspace *vm;
1035 	long rss;
1036 
1037 	if ((lp = lwkt_preempted_proc()) != NULL) {
1038 		/*
1039 		 * Account for cpu time used and hit the scheduler.  Note
1040 		 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
1041 		 * HERE.
1042 		 */
1043 		++lp->lwp_cpticks;
1044 		usched_schedulerclock(lp, info->periodic, info->time);
1045 	} else {
1046 		usched_schedulerclock(NULL, info->periodic, info->time);
1047 	}
1048 	if ((lp = curthread->td_lwp) != NULL) {
1049 		/*
1050 		 * Update resource usage integrals and maximums.
1051 		 */
1052 		if ((ru = &lp->lwp_proc->p_ru) &&
1053 		    (vm = lp->lwp_proc->p_vmspace) != NULL) {
1054 			ru->ru_ixrss += pgtok(btoc(vm->vm_tsize));
1055 			ru->ru_idrss += pgtok(btoc(vm->vm_dsize));
1056 			ru->ru_isrss += pgtok(btoc(vm->vm_ssize));
1057 			if (lwkt_trytoken(&vm->vm_map.token)) {
1058 				rss = pgtok(vmspace_resident_count(vm));
1059 				if (ru->ru_maxrss < rss)
1060 					ru->ru_maxrss = rss;
1061 				lwkt_reltoken(&vm->vm_map.token);
1062 			}
1063 		}
1064 	}
1065 	/* Increment the global sched_ticks */
1066 	if (mycpu->gd_cpuid == 0)
1067 		++sched_ticks;
1068 }
1069 
1070 /*
1071  * Compute number of ticks for the specified amount of time.  The
1072  * return value is intended to be used in a clock interrupt timed
1073  * operation and guaranteed to meet or exceed the requested time.
1074  * If the representation overflows, return INT_MAX.  The minimum return
1075  * value is 1 ticks and the function will average the calculation up.
1076  * If any value greater then 0 microseconds is supplied, a value
1077  * of at least 2 will be returned to ensure that a near-term clock
1078  * interrupt does not cause the timeout to occur (degenerately) early.
1079  *
1080  * Note that limit checks must take into account microseconds, which is
1081  * done simply by using the smaller signed long maximum instead of
1082  * the unsigned long maximum.
1083  *
1084  * If ints have 32 bits, then the maximum value for any timeout in
1085  * 10ms ticks is 248 days.
1086  */
1087 int
1088 tvtohz_high(struct timeval *tv)
1089 {
1090 	int ticks;
1091 	long sec, usec;
1092 
1093 	sec = tv->tv_sec;
1094 	usec = tv->tv_usec;
1095 	if (usec < 0) {
1096 		sec--;
1097 		usec += 1000000;
1098 	}
1099 	if (sec < 0) {
1100 #ifdef DIAGNOSTIC
1101 		if (usec > 0) {
1102 			sec++;
1103 			usec -= 1000000;
1104 		}
1105 		kprintf("tvtohz_high: negative time difference "
1106 			"%ld sec %ld usec\n",
1107 			sec, usec);
1108 #endif
1109 		ticks = 1;
1110 	} else if (sec <= INT_MAX / hz) {
1111 		ticks = (int)(sec * hz + howmany((u_long)usec, ustick)) + 1;
1112 	} else {
1113 		ticks = INT_MAX;
1114 	}
1115 	return (ticks);
1116 }
1117 
1118 int
1119 tstohz_high(struct timespec *ts)
1120 {
1121 	int ticks;
1122 	long sec, nsec;
1123 
1124 	sec = ts->tv_sec;
1125 	nsec = ts->tv_nsec;
1126 	if (nsec < 0) {
1127 		sec--;
1128 		nsec += 1000000000;
1129 	}
1130 	if (sec < 0) {
1131 #ifdef DIAGNOSTIC
1132 		if (nsec > 0) {
1133 			sec++;
1134 			nsec -= 1000000000;
1135 		}
1136 		kprintf("tstohz_high: negative time difference "
1137 			"%ld sec %ld nsec\n",
1138 			sec, nsec);
1139 #endif
1140 		ticks = 1;
1141 	} else if (sec <= INT_MAX / hz) {
1142 		ticks = (int)(sec * hz + howmany((u_long)nsec, nstick)) + 1;
1143 	} else {
1144 		ticks = INT_MAX;
1145 	}
1146 	return (ticks);
1147 }
1148 
1149 
1150 /*
1151  * Compute number of ticks for the specified amount of time, erroring on
1152  * the side of it being too low to ensure that sleeping the returned number
1153  * of ticks will not result in a late return.
1154  *
1155  * The supplied timeval may not be negative and should be normalized.  A
1156  * return value of 0 is possible if the timeval converts to less then
1157  * 1 tick.
1158  *
1159  * If ints have 32 bits, then the maximum value for any timeout in
1160  * 10ms ticks is 248 days.
1161  */
1162 int
1163 tvtohz_low(struct timeval *tv)
1164 {
1165 	int ticks;
1166 	long sec;
1167 
1168 	sec = tv->tv_sec;
1169 	if (sec <= INT_MAX / hz)
1170 		ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
1171 	else
1172 		ticks = INT_MAX;
1173 	return (ticks);
1174 }
1175 
1176 int
1177 tstohz_low(struct timespec *ts)
1178 {
1179 	int ticks;
1180 	long sec;
1181 
1182 	sec = ts->tv_sec;
1183 	if (sec <= INT_MAX / hz)
1184 		ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1185 	else
1186 		ticks = INT_MAX;
1187 	return (ticks);
1188 }
1189 
1190 /*
1191  * Start profiling on a process.
1192  *
1193  * Caller must hold p->p_token();
1194  *
1195  * Kernel profiling passes proc0 which never exits and hence
1196  * keeps the profile clock running constantly.
1197  */
1198 void
1199 startprofclock(struct proc *p)
1200 {
1201 	if ((p->p_flags & P_PROFIL) == 0) {
1202 		p->p_flags |= P_PROFIL;
1203 #if 0	/* XXX */
1204 		if (++profprocs == 1 && stathz != 0) {
1205 			crit_enter();
1206 			psdiv = psratio;
1207 			setstatclockrate(profhz);
1208 			crit_exit();
1209 		}
1210 #endif
1211 	}
1212 }
1213 
1214 /*
1215  * Stop profiling on a process.
1216  *
1217  * caller must hold p->p_token
1218  */
1219 void
1220 stopprofclock(struct proc *p)
1221 {
1222 	if (p->p_flags & P_PROFIL) {
1223 		p->p_flags &= ~P_PROFIL;
1224 #if 0	/* XXX */
1225 		if (--profprocs == 0 && stathz != 0) {
1226 			crit_enter();
1227 			psdiv = 1;
1228 			setstatclockrate(stathz);
1229 			crit_exit();
1230 		}
1231 #endif
1232 	}
1233 }
1234 
1235 /*
1236  * Return information about system clocks.
1237  */
1238 static int
1239 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1240 {
1241 	struct kinfo_clockinfo clkinfo;
1242 	/*
1243 	 * Construct clockinfo structure.
1244 	 */
1245 	clkinfo.ci_hz = hz;
1246 	clkinfo.ci_tick = ustick;
1247 	clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1248 	clkinfo.ci_profhz = profhz;
1249 	clkinfo.ci_stathz = stathz ? stathz : hz;
1250 	return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1251 }
1252 
1253 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1254 	0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1255 
1256 /*
1257  * We have eight functions for looking at the clock, four for
1258  * microseconds and four for nanoseconds.  For each there is fast
1259  * but less precise version "get{nano|micro}[up]time" which will
1260  * return a time which is up to 1/HZ previous to the call, whereas
1261  * the raw version "{nano|micro}[up]time" will return a timestamp
1262  * which is as precise as possible.  The "up" variants return the
1263  * time relative to system boot, these are well suited for time
1264  * interval measurements.
1265  *
1266  * Each cpu independently maintains the current time of day, so all
1267  * we need to do to protect ourselves from changes is to do a loop
1268  * check on the seconds field changing out from under us.
1269  *
1270  * The system timer maintains a 32 bit count and due to various issues
1271  * it is possible for the calculated delta to occasionally exceed
1272  * sys_cputimer->freq.  If this occurs the sys_cputimer->freq64_nsec
1273  * multiplication can easily overflow, so we deal with the case.  For
1274  * uniformity we deal with the case in the usec case too.
1275  *
1276  * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1277  */
1278 void
1279 getmicrouptime(struct timeval *tvp)
1280 {
1281 	struct globaldata *gd = mycpu;
1282 	sysclock_t delta;
1283 
1284 	do {
1285 		tvp->tv_sec = gd->gd_time_seconds;
1286 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1287 	} while (tvp->tv_sec != gd->gd_time_seconds);
1288 
1289 	if (delta >= sys_cputimer->freq) {
1290 		tvp->tv_sec += delta / sys_cputimer->freq;
1291 		delta %= sys_cputimer->freq;
1292 	}
1293 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1294 	if (tvp->tv_usec >= 1000000) {
1295 		tvp->tv_usec -= 1000000;
1296 		++tvp->tv_sec;
1297 	}
1298 }
1299 
1300 void
1301 getnanouptime(struct timespec *tsp)
1302 {
1303 	struct globaldata *gd = mycpu;
1304 	sysclock_t delta;
1305 
1306 	do {
1307 		tsp->tv_sec = gd->gd_time_seconds;
1308 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1309 	} while (tsp->tv_sec != gd->gd_time_seconds);
1310 
1311 	if (delta >= sys_cputimer->freq) {
1312 		tsp->tv_sec += delta / sys_cputimer->freq;
1313 		delta %= sys_cputimer->freq;
1314 	}
1315 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1316 }
1317 
1318 void
1319 microuptime(struct timeval *tvp)
1320 {
1321 	struct globaldata *gd = mycpu;
1322 	sysclock_t delta;
1323 
1324 	do {
1325 		tvp->tv_sec = gd->gd_time_seconds;
1326 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1327 	} while (tvp->tv_sec != gd->gd_time_seconds);
1328 
1329 	if (delta >= sys_cputimer->freq) {
1330 		tvp->tv_sec += delta / sys_cputimer->freq;
1331 		delta %= sys_cputimer->freq;
1332 	}
1333 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1334 }
1335 
1336 void
1337 nanouptime(struct timespec *tsp)
1338 {
1339 	struct globaldata *gd = mycpu;
1340 	sysclock_t delta;
1341 
1342 	do {
1343 		tsp->tv_sec = gd->gd_time_seconds;
1344 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1345 	} while (tsp->tv_sec != gd->gd_time_seconds);
1346 
1347 	if (delta >= sys_cputimer->freq) {
1348 		tsp->tv_sec += delta / sys_cputimer->freq;
1349 		delta %= sys_cputimer->freq;
1350 	}
1351 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1352 }
1353 
1354 /*
1355  * realtime routines
1356  */
1357 void
1358 getmicrotime(struct timeval *tvp)
1359 {
1360 	struct globaldata *gd = mycpu;
1361 	struct timespec *bt;
1362 	sysclock_t delta;
1363 
1364 	do {
1365 		tvp->tv_sec = gd->gd_time_seconds;
1366 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1367 	} while (tvp->tv_sec != gd->gd_time_seconds);
1368 
1369 	if (delta >= sys_cputimer->freq) {
1370 		tvp->tv_sec += delta / sys_cputimer->freq;
1371 		delta %= sys_cputimer->freq;
1372 	}
1373 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1374 
1375 	bt = &basetime[basetime_index];
1376 	cpu_lfence();
1377 	tvp->tv_sec += bt->tv_sec;
1378 	tvp->tv_usec += bt->tv_nsec / 1000;
1379 	while (tvp->tv_usec >= 1000000) {
1380 		tvp->tv_usec -= 1000000;
1381 		++tvp->tv_sec;
1382 	}
1383 }
1384 
1385 void
1386 getnanotime(struct timespec *tsp)
1387 {
1388 	struct globaldata *gd = mycpu;
1389 	struct timespec *bt;
1390 	sysclock_t delta;
1391 
1392 	do {
1393 		tsp->tv_sec = gd->gd_time_seconds;
1394 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1395 	} while (tsp->tv_sec != gd->gd_time_seconds);
1396 
1397 	if (delta >= sys_cputimer->freq) {
1398 		tsp->tv_sec += delta / sys_cputimer->freq;
1399 		delta %= sys_cputimer->freq;
1400 	}
1401 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1402 
1403 	bt = &basetime[basetime_index];
1404 	cpu_lfence();
1405 	tsp->tv_sec += bt->tv_sec;
1406 	tsp->tv_nsec += bt->tv_nsec;
1407 	while (tsp->tv_nsec >= 1000000000) {
1408 		tsp->tv_nsec -= 1000000000;
1409 		++tsp->tv_sec;
1410 	}
1411 }
1412 
1413 static void
1414 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1415 {
1416 	struct globaldata *gd = mycpu;
1417 	sysclock_t delta;
1418 
1419 	do {
1420 		tsp->tv_sec = gd->gd_time_seconds;
1421 		delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1422 	} while (tsp->tv_sec != gd->gd_time_seconds);
1423 
1424 	if (delta >= sys_cputimer->freq) {
1425 		tsp->tv_sec += delta / sys_cputimer->freq;
1426 		delta %= sys_cputimer->freq;
1427 	}
1428 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1429 
1430 	tsp->tv_sec += nbt->tv_sec;
1431 	tsp->tv_nsec += nbt->tv_nsec;
1432 	while (tsp->tv_nsec >= 1000000000) {
1433 		tsp->tv_nsec -= 1000000000;
1434 		++tsp->tv_sec;
1435 	}
1436 }
1437 
1438 
1439 void
1440 microtime(struct timeval *tvp)
1441 {
1442 	struct globaldata *gd = mycpu;
1443 	struct timespec *bt;
1444 	sysclock_t delta;
1445 
1446 	do {
1447 		tvp->tv_sec = gd->gd_time_seconds;
1448 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1449 	} while (tvp->tv_sec != gd->gd_time_seconds);
1450 
1451 	if (delta >= sys_cputimer->freq) {
1452 		tvp->tv_sec += delta / sys_cputimer->freq;
1453 		delta %= sys_cputimer->freq;
1454 	}
1455 	tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1456 
1457 	bt = &basetime[basetime_index];
1458 	cpu_lfence();
1459 	tvp->tv_sec += bt->tv_sec;
1460 	tvp->tv_usec += bt->tv_nsec / 1000;
1461 	while (tvp->tv_usec >= 1000000) {
1462 		tvp->tv_usec -= 1000000;
1463 		++tvp->tv_sec;
1464 	}
1465 }
1466 
1467 void
1468 nanotime(struct timespec *tsp)
1469 {
1470 	struct globaldata *gd = mycpu;
1471 	struct timespec *bt;
1472 	sysclock_t delta;
1473 
1474 	do {
1475 		tsp->tv_sec = gd->gd_time_seconds;
1476 		delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1477 	} while (tsp->tv_sec != gd->gd_time_seconds);
1478 
1479 	if (delta >= sys_cputimer->freq) {
1480 		tsp->tv_sec += delta / sys_cputimer->freq;
1481 		delta %= sys_cputimer->freq;
1482 	}
1483 	tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1484 
1485 	bt = &basetime[basetime_index];
1486 	cpu_lfence();
1487 	tsp->tv_sec += bt->tv_sec;
1488 	tsp->tv_nsec += bt->tv_nsec;
1489 	while (tsp->tv_nsec >= 1000000000) {
1490 		tsp->tv_nsec -= 1000000000;
1491 		++tsp->tv_sec;
1492 	}
1493 }
1494 
1495 /*
1496  * Get an approximate time_t.  It does not have to be accurate.  This
1497  * function is called only from KTR and can be called with the system in
1498  * any state so do not use a critical section or other complex operation
1499  * here.
1500  *
1501  * NOTE: This is not exactly synchronized with real time.  To do that we
1502  *	 would have to do what microtime does and check for a nanoseconds
1503  *	 overflow.
1504  */
1505 time_t
1506 get_approximate_time_t(void)
1507 {
1508 	struct globaldata *gd = mycpu;
1509 	struct timespec *bt;
1510 
1511 	bt = &basetime[basetime_index];
1512 	return(gd->gd_time_seconds + bt->tv_sec);
1513 }
1514 
1515 int
1516 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1517 {
1518 	pps_params_t *app;
1519 	struct pps_fetch_args *fapi;
1520 #ifdef PPS_SYNC
1521 	struct pps_kcbind_args *kapi;
1522 #endif
1523 
1524 	switch (cmd) {
1525 	case PPS_IOC_CREATE:
1526 		return (0);
1527 	case PPS_IOC_DESTROY:
1528 		return (0);
1529 	case PPS_IOC_SETPARAMS:
1530 		app = (pps_params_t *)data;
1531 		if (app->mode & ~pps->ppscap)
1532 			return (EINVAL);
1533 		pps->ppsparam = *app;
1534 		return (0);
1535 	case PPS_IOC_GETPARAMS:
1536 		app = (pps_params_t *)data;
1537 		*app = pps->ppsparam;
1538 		app->api_version = PPS_API_VERS_1;
1539 		return (0);
1540 	case PPS_IOC_GETCAP:
1541 		*(int*)data = pps->ppscap;
1542 		return (0);
1543 	case PPS_IOC_FETCH:
1544 		fapi = (struct pps_fetch_args *)data;
1545 		if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1546 			return (EINVAL);
1547 		if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1548 			return (EOPNOTSUPP);
1549 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
1550 		fapi->pps_info_buf = pps->ppsinfo;
1551 		return (0);
1552 	case PPS_IOC_KCBIND:
1553 #ifdef PPS_SYNC
1554 		kapi = (struct pps_kcbind_args *)data;
1555 		/* XXX Only root should be able to do this */
1556 		if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1557 			return (EINVAL);
1558 		if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1559 			return (EINVAL);
1560 		if (kapi->edge & ~pps->ppscap)
1561 			return (EINVAL);
1562 		pps->kcmode = kapi->edge;
1563 		return (0);
1564 #else
1565 		return (EOPNOTSUPP);
1566 #endif
1567 	default:
1568 		return (ENOTTY);
1569 	}
1570 }
1571 
1572 void
1573 pps_init(struct pps_state *pps)
1574 {
1575 	pps->ppscap |= PPS_TSFMT_TSPEC;
1576 	if (pps->ppscap & PPS_CAPTUREASSERT)
1577 		pps->ppscap |= PPS_OFFSETASSERT;
1578 	if (pps->ppscap & PPS_CAPTURECLEAR)
1579 		pps->ppscap |= PPS_OFFSETCLEAR;
1580 }
1581 
1582 void
1583 pps_event(struct pps_state *pps, sysclock_t count, int event)
1584 {
1585 	struct globaldata *gd;
1586 	struct timespec *tsp;
1587 	struct timespec *osp;
1588 	struct timespec *bt;
1589 	struct timespec ts;
1590 	sysclock_t *pcount;
1591 #ifdef PPS_SYNC
1592 	sysclock_t tcount;
1593 #endif
1594 	sysclock_t delta;
1595 	pps_seq_t *pseq;
1596 	int foff;
1597 #ifdef PPS_SYNC
1598 	int fhard;
1599 #endif
1600 	int ni;
1601 
1602 	gd = mycpu;
1603 
1604 	/* Things would be easier with arrays... */
1605 	if (event == PPS_CAPTUREASSERT) {
1606 		tsp = &pps->ppsinfo.assert_timestamp;
1607 		osp = &pps->ppsparam.assert_offset;
1608 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1609 #ifdef PPS_SYNC
1610 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
1611 #endif
1612 		pcount = &pps->ppscount[0];
1613 		pseq = &pps->ppsinfo.assert_sequence;
1614 	} else {
1615 		tsp = &pps->ppsinfo.clear_timestamp;
1616 		osp = &pps->ppsparam.clear_offset;
1617 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1618 #ifdef PPS_SYNC
1619 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
1620 #endif
1621 		pcount = &pps->ppscount[1];
1622 		pseq = &pps->ppsinfo.clear_sequence;
1623 	}
1624 
1625 	/* Nothing really happened */
1626 	if (*pcount == count)
1627 		return;
1628 
1629 	*pcount = count;
1630 
1631 	do {
1632 		ts.tv_sec = gd->gd_time_seconds;
1633 		delta = count - gd->gd_cpuclock_base;
1634 	} while (ts.tv_sec != gd->gd_time_seconds);
1635 
1636 	if (delta >= sys_cputimer->freq) {
1637 		ts.tv_sec += delta / sys_cputimer->freq;
1638 		delta %= sys_cputimer->freq;
1639 	}
1640 	ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1641 	ni = basetime_index;
1642 	cpu_lfence();
1643 	bt = &basetime[ni];
1644 	ts.tv_sec += bt->tv_sec;
1645 	ts.tv_nsec += bt->tv_nsec;
1646 	while (ts.tv_nsec >= 1000000000) {
1647 		ts.tv_nsec -= 1000000000;
1648 		++ts.tv_sec;
1649 	}
1650 
1651 	(*pseq)++;
1652 	*tsp = ts;
1653 
1654 	if (foff) {
1655 		timespecadd(tsp, osp, tsp);
1656 		if (tsp->tv_nsec < 0) {
1657 			tsp->tv_nsec += 1000000000;
1658 			tsp->tv_sec -= 1;
1659 		}
1660 	}
1661 #ifdef PPS_SYNC
1662 	if (fhard) {
1663 		/* magic, at its best... */
1664 		tcount = count - pps->ppscount[2];
1665 		pps->ppscount[2] = count;
1666 		if (tcount >= sys_cputimer->freq) {
1667 			delta = (1000000000 * (tcount / sys_cputimer->freq) +
1668 				 sys_cputimer->freq64_nsec *
1669 				 (tcount % sys_cputimer->freq)) >> 32;
1670 		} else {
1671 			delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1672 		}
1673 		hardpps(tsp, delta);
1674 	}
1675 #endif
1676 }
1677 
1678 /*
1679  * Return the tsc target value for a delay of (ns).
1680  *
1681  * Returns -1 if the TSC is not supported.
1682  */
1683 tsc_uclock_t
1684 tsc_get_target(int ns)
1685 {
1686 #if defined(_RDTSC_SUPPORTED_)
1687 	if (cpu_feature & CPUID_TSC) {
1688 		return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1689 	}
1690 #endif
1691 	return(-1);
1692 }
1693 
1694 /*
1695  * Compare the tsc against the passed target
1696  *
1697  * Returns +1 if the target has been reached
1698  * Returns  0 if the target has not yet been reached
1699  * Returns -1 if the TSC is not supported.
1700  *
1701  * Typical use:		while (tsc_test_target(target) == 0) { ...poll... }
1702  */
1703 int
1704 tsc_test_target(int64_t target)
1705 {
1706 #if defined(_RDTSC_SUPPORTED_)
1707 	if (cpu_feature & CPUID_TSC) {
1708 		if ((int64_t)(target - rdtsc()) <= 0)
1709 			return(1);
1710 		return(0);
1711 	}
1712 #endif
1713 	return(-1);
1714 }
1715 
1716 /*
1717  * Delay the specified number of nanoseconds using the tsc.  This function
1718  * returns immediately if the TSC is not supported.  At least one cpu_pause()
1719  * will be issued.
1720  */
1721 void
1722 tsc_delay(int ns)
1723 {
1724 	int64_t clk;
1725 
1726 	clk = tsc_get_target(ns);
1727 	cpu_pause();
1728 	cpu_pause();
1729 	while (tsc_test_target(clk) == 0) {
1730 		cpu_pause();
1731 		cpu_pause();
1732 		cpu_pause();
1733 		cpu_pause();
1734 	}
1735 }
1736