2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
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.
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
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.
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
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 1.52 2006/06/01 16:49:59 dillon Exp $
77 #include "opt_polling.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/callout.h>
82 #include <sys/kernel.h>
83 #include <sys/kinfo.h>
85 #include <sys/malloc.h>
86 #include <sys/resourcevar.h>
87 #include <sys/signalvar.h>
88 #include <sys/timex.h>
89 #include <sys/timepps.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_extern.h>
95 #include <sys/sysctl.h>
96 #include <sys/thread2.h>
98 #include <machine/cpu.h>
99 #include <machine/limits.h>
100 #include <machine/smp.h>
103 #include <sys/gmon.h>
106 #ifdef DEVICE_POLLING
107 extern void init_device_poll(void);
110 static void initclocks (void *dummy);
111 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
114 * Some of these don't belong here, but it's easiest to concentrate them.
115 * Note that cpu_time counts in microseconds, but most userland programs
116 * just compare relative times against the total by delta.
118 struct kinfo_cputime cputime_percpu[MAXCPU];
121 sysctl_cputime(SYSCTL_HANDLER_ARGS)
124 size_t size = sizeof(struct kinfo_cputime);
126 for (cpu = 0; cpu < ncpus; ++cpu) {
127 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
133 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
134 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
136 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
137 "CPU time statistics");
141 * boottime is used to calculate the 'real' uptime. Do not confuse this with
142 * microuptime(). microtime() is not drift compensated. The real uptime
143 * with compensation is nanotime() - bootime. boottime is recalculated
144 * whenever the real time is set based on the compensated elapsed time
145 * in seconds (gd->gd_time_seconds).
147 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
148 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
151 struct timespec boottime; /* boot time (realtime) for reference only */
152 time_t time_second; /* read-only 'passive' uptime in seconds */
155 * basetime is used to calculate the compensated real time of day. The
156 * basetime can be modified on a per-tick basis by the adjtime(),
157 * ntp_adjtime(), and sysctl-based time correction APIs.
159 * Note that frequency corrections can also be made by adjusting
162 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
163 * used on both SMP and UP systems to avoid MP races between cpu's and
164 * interrupt races on UP systems.
166 #define BASETIME_ARYSIZE 16
167 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
168 static struct timespec basetime[BASETIME_ARYSIZE];
169 static volatile int basetime_index;
172 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
179 * Because basetime data and index may be updated by another cpu,
180 * a load fence is required to ensure that the data we read has
181 * not been speculatively read relative to a possibly updated index.
183 index = basetime_index;
185 bt = &basetime[index];
186 error = SYSCTL_OUT(req, bt, sizeof(*bt));
190 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
191 &boottime, timespec, "System boottime");
192 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
193 sysctl_get_basetime, "S,timespec", "System basetime");
195 static void hardclock(systimer_t info, struct intrframe *frame);
196 static void statclock(systimer_t info, struct intrframe *frame);
197 static void schedclock(systimer_t info, struct intrframe *frame);
198 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
200 int ticks; /* system master ticks at hz */
201 int clocks_running; /* tsleep/timeout clocks operational */
202 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
203 int64_t nsec_acc; /* accumulator */
205 /* NTPD time correction fields */
206 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
207 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
208 int64_t ntp_delta; /* one-time correction in nsec */
209 int64_t ntp_big_delta = 1000000000;
210 int32_t ntp_tick_delta; /* current adjustment rate */
211 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
212 time_t ntp_leap_second; /* time of next leap second */
213 int ntp_leap_insert; /* whether to insert or remove a second */
216 * Finish initializing clock frequencies and start all clocks running.
220 initclocks(void *dummy)
223 #ifdef DEVICE_POLLING
226 /*psratio = profhz / stathz;*/
232 * Called on a per-cpu basis
235 initclocks_pcpu(void)
237 struct globaldata *gd = mycpu;
240 if (gd->gd_cpuid == 0) {
241 gd->gd_time_seconds = 1;
242 gd->gd_cpuclock_base = sys_cputimer->count();
245 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
246 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
250 * Use a non-queued periodic systimer to prevent multiple ticks from
251 * building up if the sysclock jumps forward (8254 gets reset). The
252 * sysclock will never jump backwards. Our time sync is based on
253 * the actual sysclock, not the ticks count.
255 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
256 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
257 /* XXX correct the frequency for scheduler / estcpu tests */
258 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
264 * This sets the current real time of day. Timespecs are in seconds and
265 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
266 * instead we adjust basetime so basetime + gd_* results in the current
267 * time of day. This way the gd_* fields are guarenteed to represent
268 * a monotonically increasing 'uptime' value.
270 * When set_timeofday() is called from userland, the system call forces it
271 * onto cpu #0 since only cpu #0 can update basetime_index.
274 set_timeofday(struct timespec *ts)
276 struct timespec *nbt;
280 * XXX SMP / non-atomic basetime updates
283 ni = (basetime_index + 1) & BASETIME_ARYMASK;
286 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
287 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
288 if (nbt->tv_nsec < 0) {
289 nbt->tv_nsec += 1000000000;
294 * Note that basetime diverges from boottime as the clock drift is
295 * compensated for, so we cannot do away with boottime. When setting
296 * the absolute time of day the drift is 0 (for an instant) and we
297 * can simply assign boottime to basetime.
299 * Note that nanouptime() is based on gd_time_seconds which is drift
300 * compensated up to a point (it is guarenteed to remain monotonically
301 * increasing). gd_time_seconds is thus our best uptime guess and
302 * suitable for use in the boottime calculation. It is already taken
303 * into account in the basetime calculation above.
305 boottime.tv_sec = nbt->tv_sec;
309 * We now have a new basetime, make sure all other cpus have it,
310 * then update the index.
319 * Each cpu has its own hardclock, but we only increments ticks and softticks
322 * NOTE! systimer! the MP lock might not be held here. We can only safely
323 * manipulate objects owned by the current cpu.
326 hardclock(systimer_t info, struct intrframe *frame)
330 struct pstats *pstats;
331 struct globaldata *gd = mycpu;
334 * Realtime updates are per-cpu. Note that timer corrections as
335 * returned by microtime() and friends make an additional adjustment
336 * using a system-wise 'basetime', but the running time is always
337 * taken from the per-cpu globaldata area. Since the same clock
338 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
341 * Note that we never allow info->time (aka gd->gd_hardclock.time)
342 * to reverse index gd_cpuclock_base, but that it is possible for
343 * it to temporarily get behind in the seconds if something in the
344 * system locks interrupts for a long period of time. Since periodic
345 * timers count events, though everything should resynch again
348 cputicks = info->time - gd->gd_cpuclock_base;
349 if (cputicks >= sys_cputimer->freq) {
350 ++gd->gd_time_seconds;
351 gd->gd_cpuclock_base += sys_cputimer->freq;
355 * The system-wide ticks counter and NTP related timedelta/tickdelta
356 * adjustments only occur on cpu #0. NTP adjustments are accomplished
357 * by updating basetime.
359 if (gd->gd_cpuid == 0) {
360 struct timespec *nbt;
368 if (tco->tc_poll_pps)
369 tco->tc_poll_pps(tco);
373 * Calculate the new basetime index. We are in a critical section
374 * on cpu #0 and can safely play with basetime_index. Start
375 * with the current basetime and then make adjustments.
377 ni = (basetime_index + 1) & BASETIME_ARYMASK;
379 *nbt = basetime[basetime_index];
382 * Apply adjtime corrections. (adjtime() API)
384 * adjtime() only runs on cpu #0 so our critical section is
385 * sufficient to access these variables.
387 if (ntp_delta != 0) {
388 nbt->tv_nsec += ntp_tick_delta;
389 ntp_delta -= ntp_tick_delta;
390 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
391 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
392 ntp_tick_delta = ntp_delta;
397 * Apply permanent frequency corrections. (sysctl API)
399 if (ntp_tick_permanent != 0) {
400 ntp_tick_acc += ntp_tick_permanent;
401 if (ntp_tick_acc >= (1LL << 32)) {
402 nbt->tv_nsec += ntp_tick_acc >> 32;
403 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
404 } else if (ntp_tick_acc <= -(1LL << 32)) {
405 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
406 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
407 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
411 if (nbt->tv_nsec >= 1000000000) {
413 nbt->tv_nsec -= 1000000000;
414 } else if (nbt->tv_nsec < 0) {
416 nbt->tv_nsec += 1000000000;
420 * Another per-tick compensation. (for ntp_adjtime() API)
423 nsec_acc += nsec_adj;
424 if (nsec_acc >= 0x100000000LL) {
425 nbt->tv_nsec += nsec_acc >> 32;
426 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
427 } else if (nsec_acc <= -0x100000000LL) {
428 nbt->tv_nsec -= -nsec_acc >> 32;
429 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
431 if (nbt->tv_nsec >= 1000000000) {
432 nbt->tv_nsec -= 1000000000;
434 } else if (nbt->tv_nsec < 0) {
435 nbt->tv_nsec += 1000000000;
440 /************************************************************
441 * LEAP SECOND CORRECTION *
442 ************************************************************
444 * Taking into account all the corrections made above, figure
445 * out the new real time. If the seconds field has changed
446 * then apply any pending leap-second corrections.
448 getnanotime_nbt(nbt, &nts);
450 if (time_second != nts.tv_sec) {
452 * Apply leap second (sysctl API). Adjust nts for changes
453 * so we do not have to call getnanotime_nbt again.
455 if (ntp_leap_second) {
456 if (ntp_leap_second == nts.tv_sec) {
457 if (ntp_leap_insert) {
469 * Apply leap second (ntp_adjtime() API), calculate a new
470 * nsec_adj field. ntp_update_second() returns nsec_adj
471 * as a per-second value but we need it as a per-tick value.
473 leap = ntp_update_second(time_second, &nsec_adj);
479 * Update the time_second 'approximate time' global.
481 time_second = nts.tv_sec;
485 * Finally, our new basetime is ready to go live!
491 * Figure out how badly the system is starved for memory
493 vm_fault_ratecheck();
497 * softticks are handled for all cpus
499 hardclock_softtick(gd);
502 * ITimer handling is per-tick, per-cpu. I don't think psignal()
503 * is mpsafe on curproc, so XXX get the mplock.
505 if ((p = curproc) != NULL && try_mplock()) {
507 if (frame && CLKF_USERMODE(frame) &&
508 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
509 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0)
510 psignal(p, SIGVTALRM);
511 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
512 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0)
520 * The statistics clock typically runs at a 125Hz rate, and is intended
521 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
523 * NOTE! systimer! the MP lock might not be held here. We can only safely
524 * manipulate objects owned by the current cpu.
526 * The stats clock is responsible for grabbing a profiling sample.
527 * Most of the statistics are only used by user-level statistics programs.
528 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
531 * Like the other clocks, the stat clock is called from what is effectively
532 * a fast interrupt, so the context should be the thread/process that got
536 statclock(systimer_t info, struct intrframe *frame)
549 * How big was our timeslice relative to the last time?
551 microuptime(&tv); /* mpsafe */
552 stv = &mycpu->gd_stattv;
553 if (stv->tv_sec == 0) {
556 bump = tv.tv_usec - stv->tv_usec +
557 (tv.tv_sec - stv->tv_sec) * 1000000;
568 if (frame && CLKF_USERMODE(frame)) {
570 * Came from userland, handle user time and deal with
573 if (p && (p->p_flag & P_PROFIL))
574 addupc_intr(p, CLKF_PC(frame), 1);
575 td->td_uticks += bump;
578 * Charge the time as appropriate
580 if (p && p->p_nice > NZERO)
581 cpu_time.cp_nice += bump;
583 cpu_time.cp_user += bump;
587 * Kernel statistics are just like addupc_intr, only easier.
590 if (g->state == GMON_PROF_ON && frame) {
591 i = CLKF_PC(frame) - g->lowpc;
592 if (i < g->textsize) {
593 i /= HISTFRACTION * sizeof(*g->kcount);
599 * Came from kernel mode, so we were:
600 * - handling an interrupt,
601 * - doing syscall or trap work on behalf of the current
603 * - spinning in the idle loop.
604 * Whichever it is, charge the time as appropriate.
605 * Note that we charge interrupts to the current process,
606 * regardless of whether they are ``for'' that process,
607 * so that we know how much of its real time was spent
608 * in ``non-process'' (i.e., interrupt) work.
610 * XXX assume system if frame is NULL. A NULL frame
611 * can occur if ipi processing is done from a crit_exit().
613 if (frame && CLKF_INTR(frame))
614 td->td_iticks += bump;
616 td->td_sticks += bump;
618 if (frame && CLKF_INTR(frame)) {
619 cpu_time.cp_intr += bump;
621 if (td == &mycpu->gd_idlethread)
622 cpu_time.cp_idle += bump;
624 cpu_time.cp_sys += bump;
630 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
631 * the MP lock might not be held. We can safely manipulate parts of curproc
632 * but that's about it.
634 * Each cpu has its own scheduler clock.
637 schedclock(systimer_t info, struct intrframe *frame)
640 struct pstats *pstats;
645 if ((lp = lwkt_preempted_proc()) != NULL) {
647 * Account for cpu time used and hit the scheduler. Note
648 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
652 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
655 if ((lp = curthread->td_lwp) != NULL) {
657 * Update resource usage integrals and maximums.
659 if ((pstats = lp->lwp_stats) != NULL &&
660 (ru = &pstats->p_ru) != NULL &&
661 (vm = lp->lwp_proc->p_vmspace) != NULL) {
662 ru->ru_ixrss += pgtok(vm->vm_tsize);
663 ru->ru_idrss += pgtok(vm->vm_dsize);
664 ru->ru_isrss += pgtok(vm->vm_ssize);
665 rss = pgtok(vmspace_resident_count(vm));
666 if (ru->ru_maxrss < rss)
673 * Compute number of ticks for the specified amount of time. The
674 * return value is intended to be used in a clock interrupt timed
675 * operation and guarenteed to meet or exceed the requested time.
676 * If the representation overflows, return INT_MAX. The minimum return
677 * value is 1 ticks and the function will average the calculation up.
678 * If any value greater then 0 microseconds is supplied, a value
679 * of at least 2 will be returned to ensure that a near-term clock
680 * interrupt does not cause the timeout to occur (degenerately) early.
682 * Note that limit checks must take into account microseconds, which is
683 * done simply by using the smaller signed long maximum instead of
684 * the unsigned long maximum.
686 * If ints have 32 bits, then the maximum value for any timeout in
687 * 10ms ticks is 248 days.
690 tvtohz_high(struct timeval *tv)
707 printf("tvotohz: negative time difference %ld sec %ld usec\n",
711 } else if (sec <= INT_MAX / hz) {
712 ticks = (int)(sec * hz +
713 ((u_long)usec + (tick - 1)) / tick) + 1;
721 * Compute number of ticks for the specified amount of time, erroring on
722 * the side of it being too low to ensure that sleeping the returned number
723 * of ticks will not result in a late return.
725 * The supplied timeval may not be negative and should be normalized. A
726 * return value of 0 is possible if the timeval converts to less then
729 * If ints have 32 bits, then the maximum value for any timeout in
730 * 10ms ticks is 248 days.
733 tvtohz_low(struct timeval *tv)
739 if (sec <= INT_MAX / hz)
740 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
748 * Start profiling on a process.
750 * Kernel profiling passes proc0 which never exits and hence
751 * keeps the profile clock running constantly.
754 startprofclock(struct proc *p)
756 if ((p->p_flag & P_PROFIL) == 0) {
757 p->p_flag |= P_PROFIL;
759 if (++profprocs == 1 && stathz != 0) {
762 setstatclockrate(profhz);
770 * Stop profiling on a process.
773 stopprofclock(struct proc *p)
775 if (p->p_flag & P_PROFIL) {
776 p->p_flag &= ~P_PROFIL;
778 if (--profprocs == 0 && stathz != 0) {
781 setstatclockrate(stathz);
789 * Return information about system clocks.
792 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
794 struct kinfo_clockinfo clkinfo;
796 * Construct clockinfo structure.
799 clkinfo.ci_tick = tick;
800 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
801 clkinfo.ci_profhz = profhz;
802 clkinfo.ci_stathz = stathz ? stathz : hz;
803 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
806 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
807 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
810 * We have eight functions for looking at the clock, four for
811 * microseconds and four for nanoseconds. For each there is fast
812 * but less precise version "get{nano|micro}[up]time" which will
813 * return a time which is up to 1/HZ previous to the call, whereas
814 * the raw version "{nano|micro}[up]time" will return a timestamp
815 * which is as precise as possible. The "up" variants return the
816 * time relative to system boot, these are well suited for time
817 * interval measurements.
819 * Each cpu independantly maintains the current time of day, so all
820 * we need to do to protect ourselves from changes is to do a loop
821 * check on the seconds field changing out from under us.
823 * The system timer maintains a 32 bit count and due to various issues
824 * it is possible for the calculated delta to occassionally exceed
825 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
826 * multiplication can easily overflow, so we deal with the case. For
827 * uniformity we deal with the case in the usec case too.
830 getmicrouptime(struct timeval *tvp)
832 struct globaldata *gd = mycpu;
836 tvp->tv_sec = gd->gd_time_seconds;
837 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
838 } while (tvp->tv_sec != gd->gd_time_seconds);
840 if (delta >= sys_cputimer->freq) {
841 tvp->tv_sec += delta / sys_cputimer->freq;
842 delta %= sys_cputimer->freq;
844 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
845 if (tvp->tv_usec >= 1000000) {
846 tvp->tv_usec -= 1000000;
852 getnanouptime(struct timespec *tsp)
854 struct globaldata *gd = mycpu;
858 tsp->tv_sec = gd->gd_time_seconds;
859 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
860 } while (tsp->tv_sec != gd->gd_time_seconds);
862 if (delta >= sys_cputimer->freq) {
863 tsp->tv_sec += delta / sys_cputimer->freq;
864 delta %= sys_cputimer->freq;
866 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
870 microuptime(struct timeval *tvp)
872 struct globaldata *gd = mycpu;
876 tvp->tv_sec = gd->gd_time_seconds;
877 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
878 } while (tvp->tv_sec != gd->gd_time_seconds);
880 if (delta >= sys_cputimer->freq) {
881 tvp->tv_sec += delta / sys_cputimer->freq;
882 delta %= sys_cputimer->freq;
884 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
888 nanouptime(struct timespec *tsp)
890 struct globaldata *gd = mycpu;
894 tsp->tv_sec = gd->gd_time_seconds;
895 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
896 } while (tsp->tv_sec != gd->gd_time_seconds);
898 if (delta >= sys_cputimer->freq) {
899 tsp->tv_sec += delta / sys_cputimer->freq;
900 delta %= sys_cputimer->freq;
902 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
910 getmicrotime(struct timeval *tvp)
912 struct globaldata *gd = mycpu;
917 tvp->tv_sec = gd->gd_time_seconds;
918 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
919 } while (tvp->tv_sec != gd->gd_time_seconds);
921 if (delta >= sys_cputimer->freq) {
922 tvp->tv_sec += delta / sys_cputimer->freq;
923 delta %= sys_cputimer->freq;
925 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
927 bt = &basetime[basetime_index];
928 tvp->tv_sec += bt->tv_sec;
929 tvp->tv_usec += bt->tv_nsec / 1000;
930 while (tvp->tv_usec >= 1000000) {
931 tvp->tv_usec -= 1000000;
937 getnanotime(struct timespec *tsp)
939 struct globaldata *gd = mycpu;
944 tsp->tv_sec = gd->gd_time_seconds;
945 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
946 } while (tsp->tv_sec != gd->gd_time_seconds);
948 if (delta >= sys_cputimer->freq) {
949 tsp->tv_sec += delta / sys_cputimer->freq;
950 delta %= sys_cputimer->freq;
952 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
954 bt = &basetime[basetime_index];
955 tsp->tv_sec += bt->tv_sec;
956 tsp->tv_nsec += bt->tv_nsec;
957 while (tsp->tv_nsec >= 1000000000) {
958 tsp->tv_nsec -= 1000000000;
964 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
966 struct globaldata *gd = mycpu;
970 tsp->tv_sec = gd->gd_time_seconds;
971 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
972 } while (tsp->tv_sec != gd->gd_time_seconds);
974 if (delta >= sys_cputimer->freq) {
975 tsp->tv_sec += delta / sys_cputimer->freq;
976 delta %= sys_cputimer->freq;
978 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
980 tsp->tv_sec += nbt->tv_sec;
981 tsp->tv_nsec += nbt->tv_nsec;
982 while (tsp->tv_nsec >= 1000000000) {
983 tsp->tv_nsec -= 1000000000;
990 microtime(struct timeval *tvp)
992 struct globaldata *gd = mycpu;
997 tvp->tv_sec = gd->gd_time_seconds;
998 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
999 } while (tvp->tv_sec != gd->gd_time_seconds);
1001 if (delta >= sys_cputimer->freq) {
1002 tvp->tv_sec += delta / sys_cputimer->freq;
1003 delta %= sys_cputimer->freq;
1005 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1007 bt = &basetime[basetime_index];
1008 tvp->tv_sec += bt->tv_sec;
1009 tvp->tv_usec += bt->tv_nsec / 1000;
1010 while (tvp->tv_usec >= 1000000) {
1011 tvp->tv_usec -= 1000000;
1017 nanotime(struct timespec *tsp)
1019 struct globaldata *gd = mycpu;
1020 struct timespec *bt;
1024 tsp->tv_sec = gd->gd_time_seconds;
1025 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1026 } while (tsp->tv_sec != gd->gd_time_seconds);
1028 if (delta >= sys_cputimer->freq) {
1029 tsp->tv_sec += delta / sys_cputimer->freq;
1030 delta %= sys_cputimer->freq;
1032 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1034 bt = &basetime[basetime_index];
1035 tsp->tv_sec += bt->tv_sec;
1036 tsp->tv_nsec += bt->tv_nsec;
1037 while (tsp->tv_nsec >= 1000000000) {
1038 tsp->tv_nsec -= 1000000000;
1044 * note: this is not exactly synchronized with real time. To do that we
1045 * would have to do what microtime does and check for a nanoseconds overflow.
1048 get_approximate_time_t(void)
1050 struct globaldata *gd = mycpu;
1051 struct timespec *bt;
1053 bt = &basetime[basetime_index];
1054 return(gd->gd_time_seconds + bt->tv_sec);
1058 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1061 struct pps_fetch_args *fapi;
1063 struct pps_kcbind_args *kapi;
1067 case PPS_IOC_CREATE:
1069 case PPS_IOC_DESTROY:
1071 case PPS_IOC_SETPARAMS:
1072 app = (pps_params_t *)data;
1073 if (app->mode & ~pps->ppscap)
1075 pps->ppsparam = *app;
1077 case PPS_IOC_GETPARAMS:
1078 app = (pps_params_t *)data;
1079 *app = pps->ppsparam;
1080 app->api_version = PPS_API_VERS_1;
1082 case PPS_IOC_GETCAP:
1083 *(int*)data = pps->ppscap;
1086 fapi = (struct pps_fetch_args *)data;
1087 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1089 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1090 return (EOPNOTSUPP);
1091 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1092 fapi->pps_info_buf = pps->ppsinfo;
1094 case PPS_IOC_KCBIND:
1096 kapi = (struct pps_kcbind_args *)data;
1097 /* XXX Only root should be able to do this */
1098 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1100 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1102 if (kapi->edge & ~pps->ppscap)
1104 pps->kcmode = kapi->edge;
1107 return (EOPNOTSUPP);
1115 pps_init(struct pps_state *pps)
1117 pps->ppscap |= PPS_TSFMT_TSPEC;
1118 if (pps->ppscap & PPS_CAPTUREASSERT)
1119 pps->ppscap |= PPS_OFFSETASSERT;
1120 if (pps->ppscap & PPS_CAPTURECLEAR)
1121 pps->ppscap |= PPS_OFFSETCLEAR;
1125 pps_event(struct pps_state *pps, sysclock_t count, int event)
1127 struct globaldata *gd;
1128 struct timespec *tsp;
1129 struct timespec *osp;
1130 struct timespec *bt;
1143 /* Things would be easier with arrays... */
1144 if (event == PPS_CAPTUREASSERT) {
1145 tsp = &pps->ppsinfo.assert_timestamp;
1146 osp = &pps->ppsparam.assert_offset;
1147 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1148 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1149 pcount = &pps->ppscount[0];
1150 pseq = &pps->ppsinfo.assert_sequence;
1152 tsp = &pps->ppsinfo.clear_timestamp;
1153 osp = &pps->ppsparam.clear_offset;
1154 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1155 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1156 pcount = &pps->ppscount[1];
1157 pseq = &pps->ppsinfo.clear_sequence;
1160 /* Nothing really happened */
1161 if (*pcount == count)
1167 ts.tv_sec = gd->gd_time_seconds;
1168 delta = count - gd->gd_cpuclock_base;
1169 } while (ts.tv_sec != gd->gd_time_seconds);
1171 if (delta >= sys_cputimer->freq) {
1172 ts.tv_sec += delta / sys_cputimer->freq;
1173 delta %= sys_cputimer->freq;
1175 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1176 bt = &basetime[basetime_index];
1177 ts.tv_sec += bt->tv_sec;
1178 ts.tv_nsec += bt->tv_nsec;
1179 while (ts.tv_nsec >= 1000000000) {
1180 ts.tv_nsec -= 1000000000;
1188 timespecadd(tsp, osp);
1189 if (tsp->tv_nsec < 0) {
1190 tsp->tv_nsec += 1000000000;
1196 /* magic, at its best... */
1197 tcount = count - pps->ppscount[2];
1198 pps->ppscount[2] = count;
1199 if (tcount >= sys_cputimer->freq) {
1200 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1201 sys_cputimer->freq64_nsec *
1202 (tcount % sys_cputimer->freq)) >> 32;
1204 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1206 hardpps(tsp, delta);