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 1.45 2005/06/29 01:25:08 dillon Exp $
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/callout.h>
81 #include <sys/kernel.h>
82 #include <sys/kinfo.h>
84 #include <sys/malloc.h>
85 #include <sys/resourcevar.h>
86 #include <sys/signalvar.h>
87 #include <sys/timex.h>
88 #include <sys/timepps.h>
92 #include <vm/vm_map.h>
93 #include <sys/sysctl.h>
94 #include <sys/thread2.h>
96 #include <machine/cpu.h>
97 #include <machine/limits.h>
98 #include <machine/smp.h>
101 #include <sys/gmon.h>
104 #ifdef DEVICE_POLLING
105 extern void init_device_poll(void);
106 extern void hardclock_device_poll(void);
107 #endif /* DEVICE_POLLING */
109 static void initclocks (void *dummy);
110 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
113 * Some of these don't belong here, but it's easiest to concentrate them.
114 * Note that cpu_time counts in microseconds, but most userland programs
115 * just compare relative times against the total by delta.
117 struct kinfo_cputime cputime_percpu[MAXCPU];
120 sysctl_cputime(SYSCTL_HANDLER_ARGS)
123 size_t size = sizeof(struct kinfo_cputime);
125 for (cpu = 0; cpu < ncpus; ++cpu) {
126 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
132 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
133 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
135 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
136 "CPU time statistics");
140 * boottime is used to calculate the 'real' uptime. Do not confuse this with
141 * microuptime(). microtime() is not drift compensated. The real uptime
142 * with compensation is nanotime() - bootime. boottime is recalculated
143 * whenever the real time is set based on the compensated elapsed time
144 * in seconds (gd->gd_time_seconds).
146 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
147 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
150 struct timespec boottime; /* boot time (realtime) for reference only */
151 time_t time_second; /* read-only 'passive' uptime in seconds */
154 * basetime is used to calculate the compensated real time of day. The
155 * basetime can be modified on a per-tick basis by the adjtime(),
156 * ntp_adjtime(), and sysctl-based time correction APIs.
158 * Note that frequency corrections can also be made by adjusting
161 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
162 * used on both SMP and UP systems to avoid MP races between cpu's and
163 * interrupt races on UP systems.
165 #define BASETIME_ARYSIZE 16
166 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
167 static struct timespec basetime[BASETIME_ARYSIZE];
168 static volatile int basetime_index;
171 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
178 * Because basetime data and index may be updated by another cpu,
179 * a load fence is required to ensure that the data we read has
180 * not been speculatively read relative to a possibly updated index.
182 index = basetime_index;
184 bt = &basetime[index];
185 error = SYSCTL_OUT(req, bt, sizeof(*bt));
189 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
190 &boottime, timespec, "System boottime");
191 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
192 sysctl_get_basetime, "S,timespec", "System basetime");
194 static void hardclock(systimer_t info, struct intrframe *frame);
195 static void statclock(systimer_t info, struct intrframe *frame);
196 static void schedclock(systimer_t info, struct intrframe *frame);
197 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
199 int ticks; /* system master ticks at hz */
200 int clocks_running; /* tsleep/timeout clocks operational */
201 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
202 int64_t nsec_acc; /* accumulator */
204 /* NTPD time correction fields */
205 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
206 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
207 int64_t ntp_delta; /* one-time correction in nsec */
208 int64_t ntp_big_delta = 1000000000;
209 int32_t ntp_tick_delta; /* current adjustment rate */
210 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
211 time_t ntp_leap_second; /* time of next leap second */
212 int ntp_leap_insert; /* whether to insert or remove a second */
215 * Finish initializing clock frequencies and start all clocks running.
219 initclocks(void *dummy)
222 #ifdef DEVICE_POLLING
225 /*psratio = profhz / stathz;*/
231 * Called on a per-cpu basis
234 initclocks_pcpu(void)
236 struct globaldata *gd = mycpu;
239 if (gd->gd_cpuid == 0) {
240 gd->gd_time_seconds = 1;
241 gd->gd_cpuclock_base = sys_cputimer->count();
244 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
245 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
249 * Use a non-queued periodic systimer to prevent multiple ticks from
250 * building up if the sysclock jumps forward (8254 gets reset). The
251 * sysclock will never jump backwards. Our time sync is based on
252 * the actual sysclock, not the ticks count.
254 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
255 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
256 /* XXX correct the frequency for scheduler / estcpu tests */
257 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
263 * This sets the current real time of day. Timespecs are in seconds and
264 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
265 * instead we adjust basetime so basetime + gd_* results in the current
266 * time of day. This way the gd_* fields are guarenteed to represent
267 * a monotonically increasing 'uptime' value.
269 * When set_timeofday() is called from userland, the system call forces it
270 * onto cpu #0 since only cpu #0 can update basetime_index.
273 set_timeofday(struct timespec *ts)
275 struct timespec *nbt;
279 * XXX SMP / non-atomic basetime updates
282 ni = (basetime_index + 1) & BASETIME_ARYMASK;
285 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
286 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
287 if (nbt->tv_nsec < 0) {
288 nbt->tv_nsec += 1000000000;
293 * Note that basetime diverges from boottime as the clock drift is
294 * compensated for, so we cannot do away with boottime. When setting
295 * the absolute time of day the drift is 0 (for an instant) and we
296 * can simply assign boottime to basetime.
298 * Note that nanouptime() is based on gd_time_seconds which is drift
299 * compensated up to a point (it is guarenteed to remain monotonically
300 * increasing). gd_time_seconds is thus our best uptime guess and
301 * suitable for use in the boottime calculation. It is already taken
302 * into account in the basetime calculation above.
304 boottime.tv_sec = nbt->tv_sec;
308 * We now have a new basetime, make sure all other cpus have it,
309 * then update the index.
318 * Each cpu has its own hardclock, but we only increments ticks and softticks
321 * NOTE! systimer! the MP lock might not be held here. We can only safely
322 * manipulate objects owned by the current cpu.
325 hardclock(systimer_t info, struct intrframe *frame)
329 struct pstats *pstats;
330 struct globaldata *gd = mycpu;
333 * Realtime updates are per-cpu. Note that timer corrections as
334 * returned by microtime() and friends make an additional adjustment
335 * using a system-wise 'basetime', but the running time is always
336 * taken from the per-cpu globaldata area. Since the same clock
337 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
340 * Note that we never allow info->time (aka gd->gd_hardclock.time)
341 * to reverse index gd_cpuclock_base, but that it is possible for
342 * it to temporarily get behind in the seconds if something in the
343 * system locks interrupts for a long period of time. Since periodic
344 * timers count events, though everything should resynch again
347 cputicks = info->time - gd->gd_cpuclock_base;
348 if (cputicks >= sys_cputimer->freq) {
349 ++gd->gd_time_seconds;
350 gd->gd_cpuclock_base += sys_cputimer->freq;
354 * The system-wide ticks counter and NTP related timedelta/tickdelta
355 * adjustments only occur on cpu #0. NTP adjustments are accomplished
356 * by updating basetime.
358 if (gd->gd_cpuid == 0) {
359 struct timespec *nbt;
366 #ifdef DEVICE_POLLING
367 hardclock_device_poll(); /* mpsafe, short and quick */
368 #endif /* DEVICE_POLLING */
371 if (tco->tc_poll_pps)
372 tco->tc_poll_pps(tco);
376 * Calculate the new basetime index. We are in a critical section
377 * on cpu #0 and can safely play with basetime_index. Start
378 * with the current basetime and then make adjustments.
380 ni = (basetime_index + 1) & BASETIME_ARYMASK;
382 *nbt = basetime[basetime_index];
385 * Apply adjtime corrections. (adjtime() API)
387 * adjtime() only runs on cpu #0 so our critical section is
388 * sufficient to access these variables.
390 if (ntp_delta != 0) {
391 nbt->tv_nsec += ntp_tick_delta;
392 ntp_delta -= ntp_tick_delta;
393 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
394 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
395 ntp_tick_delta = ntp_delta;
400 * Apply permanent frequency corrections. (sysctl API)
402 if (ntp_tick_permanent != 0) {
403 ntp_tick_acc += ntp_tick_permanent;
404 if (ntp_tick_acc >= (1LL << 32)) {
405 nbt->tv_nsec += ntp_tick_acc >> 32;
406 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
407 } else if (ntp_tick_acc <= -(1LL << 32)) {
408 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
409 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
410 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
414 if (nbt->tv_nsec >= 1000000000) {
416 nbt->tv_nsec -= 1000000000;
417 } else if (nbt->tv_nsec < 0) {
419 nbt->tv_nsec += 1000000000;
423 * Another per-tick compensation. (for ntp_adjtime() API)
426 nsec_acc += nsec_adj;
427 if (nsec_acc >= 0x100000000LL) {
428 nbt->tv_nsec += nsec_acc >> 32;
429 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
430 } else if (nsec_acc <= -0x100000000LL) {
431 nbt->tv_nsec -= -nsec_acc >> 32;
432 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
434 if (nbt->tv_nsec >= 1000000000) {
435 nbt->tv_nsec -= 1000000000;
437 } else if (nbt->tv_nsec < 0) {
438 nbt->tv_nsec += 1000000000;
443 /************************************************************
444 * LEAP SECOND CORRECTION *
445 ************************************************************
447 * Taking into account all the corrections made above, figure
448 * out the new real time. If the seconds field has changed
449 * then apply any pending leap-second corrections.
451 getnanotime_nbt(nbt, &nts);
453 if (time_second != nts.tv_sec) {
455 * Apply leap second (sysctl API). Adjust nts for changes
456 * so we do not have to call getnanotime_nbt again.
458 if (ntp_leap_second) {
459 if (ntp_leap_second == nts.tv_sec) {
460 if (ntp_leap_insert) {
472 * Apply leap second (ntp_adjtime() API), calculate a new
473 * nsec_adj field. ntp_update_second() returns nsec_adj
474 * as a per-second value but we need it as a per-tick value.
476 leap = ntp_update_second(time_second, &nsec_adj);
482 * Update the time_second 'approximate time' global.
484 time_second = nts.tv_sec;
488 * Finally, our new basetime is ready to go live!
495 * softticks are handled for all cpus
497 hardclock_softtick(gd);
500 * ITimer handling is per-tick, per-cpu. I don't think psignal()
501 * is mpsafe on curproc, so XXX get the mplock.
503 if ((p = curproc) != NULL && try_mplock()) {
505 if (frame && CLKF_USERMODE(frame) &&
506 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
507 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
508 psignal(p, SIGVTALRM);
509 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
510 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
518 * The statistics clock typically runs at a 125Hz rate, and is intended
519 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
521 * NOTE! systimer! the MP lock might not be held here. We can only safely
522 * manipulate objects owned by the current cpu.
524 * The stats clock is responsible for grabbing a profiling sample.
525 * Most of the statistics are only used by user-level statistics programs.
526 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
529 * Like the other clocks, the stat clock is called from what is effectively
530 * a fast interrupt, so the context should be the thread/process that got
534 statclock(systimer_t info, struct intrframe *frame)
547 * How big was our timeslice relative to the last time?
549 microuptime(&tv); /* mpsafe */
550 stv = &mycpu->gd_stattv;
551 if (stv->tv_sec == 0) {
554 bump = tv.tv_usec - stv->tv_usec +
555 (tv.tv_sec - stv->tv_sec) * 1000000;
566 if (frame && CLKF_USERMODE(frame)) {
568 * Came from userland, handle user time and deal with
571 if (p && (p->p_flag & P_PROFIL))
572 addupc_intr(p, CLKF_PC(frame), 1);
573 td->td_uticks += bump;
576 * Charge the time as appropriate
578 if (p && p->p_nice > NZERO)
579 cpu_time.cp_nice += bump;
581 cpu_time.cp_user += bump;
585 * Kernel statistics are just like addupc_intr, only easier.
588 if (g->state == GMON_PROF_ON && frame) {
589 i = CLKF_PC(frame) - g->lowpc;
590 if (i < g->textsize) {
591 i /= HISTFRACTION * sizeof(*g->kcount);
597 * Came from kernel mode, so we were:
598 * - handling an interrupt,
599 * - doing syscall or trap work on behalf of the current
601 * - spinning in the idle loop.
602 * Whichever it is, charge the time as appropriate.
603 * Note that we charge interrupts to the current process,
604 * regardless of whether they are ``for'' that process,
605 * so that we know how much of its real time was spent
606 * in ``non-process'' (i.e., interrupt) work.
608 * XXX assume system if frame is NULL. A NULL frame
609 * can occur if ipi processing is done from a crit_exit().
611 if (frame && CLKF_INTR(frame))
612 td->td_iticks += bump;
614 td->td_sticks += bump;
616 if (frame && CLKF_INTR(frame)) {
617 cpu_time.cp_intr += bump;
619 if (td == &mycpu->gd_idlethread)
620 cpu_time.cp_idle += bump;
622 cpu_time.cp_sys += bump;
628 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
629 * the MP lock might not be held. We can safely manipulate parts of curproc
630 * but that's about it.
632 * Each cpu has its own scheduler clock.
635 schedclock(systimer_t info, struct intrframe *frame)
638 struct pstats *pstats;
643 if ((p = lwkt_preempted_proc()) != NULL) {
645 * Account for cpu time used and hit the scheduler. Note
646 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
650 p->p_usched->schedulerclock(p, info->periodic, info->time);
652 if ((p = curproc) != NULL) {
654 * Update resource usage integrals and maximums.
656 if ((pstats = p->p_stats) != NULL &&
657 (ru = &pstats->p_ru) != NULL &&
658 (vm = p->p_vmspace) != NULL) {
659 ru->ru_ixrss += pgtok(vm->vm_tsize);
660 ru->ru_idrss += pgtok(vm->vm_dsize);
661 ru->ru_isrss += pgtok(vm->vm_ssize);
662 rss = pgtok(vmspace_resident_count(vm));
663 if (ru->ru_maxrss < rss)
670 * Compute number of ticks for the specified amount of time. The
671 * return value is intended to be used in a clock interrupt timed
672 * operation and guarenteed to meet or exceed the requested time.
673 * If the representation overflows, return INT_MAX. The minimum return
674 * value is 1 ticks and the function will average the calculation up.
675 * If any value greater then 0 microseconds is supplied, a value
676 * of at least 2 will be returned to ensure that a near-term clock
677 * interrupt does not cause the timeout to occur (degenerately) early.
679 * Note that limit checks must take into account microseconds, which is
680 * done simply by using the smaller signed long maximum instead of
681 * the unsigned long maximum.
683 * If ints have 32 bits, then the maximum value for any timeout in
684 * 10ms ticks is 248 days.
687 tvtohz_high(struct timeval *tv)
704 printf("tvotohz: negative time difference %ld sec %ld usec\n",
708 } else if (sec <= INT_MAX / hz) {
709 ticks = (int)(sec * hz +
710 ((u_long)usec + (tick - 1)) / tick) + 1;
718 * Compute number of ticks for the specified amount of time, erroring on
719 * the side of it being too low to ensure that sleeping the returned number
720 * of ticks will not result in a late return.
722 * The supplied timeval may not be negative and should be normalized. A
723 * return value of 0 is possible if the timeval converts to less then
726 * If ints have 32 bits, then the maximum value for any timeout in
727 * 10ms ticks is 248 days.
730 tvtohz_low(struct timeval *tv)
736 if (sec <= INT_MAX / hz)
737 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
745 * Start profiling on a process.
747 * Kernel profiling passes proc0 which never exits and hence
748 * keeps the profile clock running constantly.
751 startprofclock(struct proc *p)
753 if ((p->p_flag & P_PROFIL) == 0) {
754 p->p_flag |= P_PROFIL;
756 if (++profprocs == 1 && stathz != 0) {
759 setstatclockrate(profhz);
767 * Stop profiling on a process.
770 stopprofclock(struct proc *p)
772 if (p->p_flag & P_PROFIL) {
773 p->p_flag &= ~P_PROFIL;
775 if (--profprocs == 0 && stathz != 0) {
778 setstatclockrate(stathz);
786 * Return information about system clocks.
789 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
791 struct kinfo_clockinfo clkinfo;
793 * Construct clockinfo structure.
796 clkinfo.ci_tick = tick;
797 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
798 clkinfo.ci_profhz = profhz;
799 clkinfo.ci_stathz = stathz ? stathz : hz;
800 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
803 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
804 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
807 * We have eight functions for looking at the clock, four for
808 * microseconds and four for nanoseconds. For each there is fast
809 * but less precise version "get{nano|micro}[up]time" which will
810 * return a time which is up to 1/HZ previous to the call, whereas
811 * the raw version "{nano|micro}[up]time" will return a timestamp
812 * which is as precise as possible. The "up" variants return the
813 * time relative to system boot, these are well suited for time
814 * interval measurements.
816 * Each cpu independantly maintains the current time of day, so all
817 * we need to do to protect ourselves from changes is to do a loop
818 * check on the seconds field changing out from under us.
820 * The system timer maintains a 32 bit count and due to various issues
821 * it is possible for the calculated delta to occassionally exceed
822 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
823 * multiplication can easily overflow, so we deal with the case. For
824 * uniformity we deal with the case in the usec case too.
827 getmicrouptime(struct timeval *tvp)
829 struct globaldata *gd = mycpu;
833 tvp->tv_sec = gd->gd_time_seconds;
834 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
835 } while (tvp->tv_sec != gd->gd_time_seconds);
837 if (delta >= sys_cputimer->freq) {
838 tvp->tv_sec += delta / sys_cputimer->freq;
839 delta %= sys_cputimer->freq;
841 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
842 if (tvp->tv_usec >= 1000000) {
843 tvp->tv_usec -= 1000000;
849 getnanouptime(struct timespec *tsp)
851 struct globaldata *gd = mycpu;
855 tsp->tv_sec = gd->gd_time_seconds;
856 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
857 } while (tsp->tv_sec != gd->gd_time_seconds);
859 if (delta >= sys_cputimer->freq) {
860 tsp->tv_sec += delta / sys_cputimer->freq;
861 delta %= sys_cputimer->freq;
863 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
867 microuptime(struct timeval *tvp)
869 struct globaldata *gd = mycpu;
873 tvp->tv_sec = gd->gd_time_seconds;
874 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
875 } while (tvp->tv_sec != gd->gd_time_seconds);
877 if (delta >= sys_cputimer->freq) {
878 tvp->tv_sec += delta / sys_cputimer->freq;
879 delta %= sys_cputimer->freq;
881 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
885 nanouptime(struct timespec *tsp)
887 struct globaldata *gd = mycpu;
891 tsp->tv_sec = gd->gd_time_seconds;
892 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
893 } while (tsp->tv_sec != gd->gd_time_seconds);
895 if (delta >= sys_cputimer->freq) {
896 tsp->tv_sec += delta / sys_cputimer->freq;
897 delta %= sys_cputimer->freq;
899 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
907 getmicrotime(struct timeval *tvp)
909 struct globaldata *gd = mycpu;
914 tvp->tv_sec = gd->gd_time_seconds;
915 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
916 } while (tvp->tv_sec != gd->gd_time_seconds);
918 if (delta >= sys_cputimer->freq) {
919 tvp->tv_sec += delta / sys_cputimer->freq;
920 delta %= sys_cputimer->freq;
922 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
924 bt = &basetime[basetime_index];
925 tvp->tv_sec += bt->tv_sec;
926 tvp->tv_usec += bt->tv_nsec / 1000;
927 while (tvp->tv_usec >= 1000000) {
928 tvp->tv_usec -= 1000000;
934 getnanotime(struct timespec *tsp)
936 struct globaldata *gd = mycpu;
941 tsp->tv_sec = gd->gd_time_seconds;
942 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
943 } while (tsp->tv_sec != gd->gd_time_seconds);
945 if (delta >= sys_cputimer->freq) {
946 tsp->tv_sec += delta / sys_cputimer->freq;
947 delta %= sys_cputimer->freq;
949 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
951 bt = &basetime[basetime_index];
952 tsp->tv_sec += bt->tv_sec;
953 tsp->tv_nsec += bt->tv_nsec;
954 while (tsp->tv_nsec >= 1000000000) {
955 tsp->tv_nsec -= 1000000000;
961 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
963 struct globaldata *gd = mycpu;
967 tsp->tv_sec = gd->gd_time_seconds;
968 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
969 } while (tsp->tv_sec != gd->gd_time_seconds);
971 if (delta >= sys_cputimer->freq) {
972 tsp->tv_sec += delta / sys_cputimer->freq;
973 delta %= sys_cputimer->freq;
975 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
977 tsp->tv_sec += nbt->tv_sec;
978 tsp->tv_nsec += nbt->tv_nsec;
979 while (tsp->tv_nsec >= 1000000000) {
980 tsp->tv_nsec -= 1000000000;
987 microtime(struct timeval *tvp)
989 struct globaldata *gd = mycpu;
994 tvp->tv_sec = gd->gd_time_seconds;
995 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
996 } while (tvp->tv_sec != gd->gd_time_seconds);
998 if (delta >= sys_cputimer->freq) {
999 tvp->tv_sec += delta / sys_cputimer->freq;
1000 delta %= sys_cputimer->freq;
1002 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1004 bt = &basetime[basetime_index];
1005 tvp->tv_sec += bt->tv_sec;
1006 tvp->tv_usec += bt->tv_nsec / 1000;
1007 while (tvp->tv_usec >= 1000000) {
1008 tvp->tv_usec -= 1000000;
1014 nanotime(struct timespec *tsp)
1016 struct globaldata *gd = mycpu;
1017 struct timespec *bt;
1021 tsp->tv_sec = gd->gd_time_seconds;
1022 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1023 } while (tsp->tv_sec != gd->gd_time_seconds);
1025 if (delta >= sys_cputimer->freq) {
1026 tsp->tv_sec += delta / sys_cputimer->freq;
1027 delta %= sys_cputimer->freq;
1029 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1031 bt = &basetime[basetime_index];
1032 tsp->tv_sec += bt->tv_sec;
1033 tsp->tv_nsec += bt->tv_nsec;
1034 while (tsp->tv_nsec >= 1000000000) {
1035 tsp->tv_nsec -= 1000000000;
1041 * note: this is not exactly synchronized with real time. To do that we
1042 * would have to do what microtime does and check for a nanoseconds overflow.
1045 get_approximate_time_t(void)
1047 struct globaldata *gd = mycpu;
1048 struct timespec *bt;
1050 bt = &basetime[basetime_index];
1051 return(gd->gd_time_seconds + bt->tv_sec);
1055 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1058 struct pps_fetch_args *fapi;
1060 struct pps_kcbind_args *kapi;
1064 case PPS_IOC_CREATE:
1066 case PPS_IOC_DESTROY:
1068 case PPS_IOC_SETPARAMS:
1069 app = (pps_params_t *)data;
1070 if (app->mode & ~pps->ppscap)
1072 pps->ppsparam = *app;
1074 case PPS_IOC_GETPARAMS:
1075 app = (pps_params_t *)data;
1076 *app = pps->ppsparam;
1077 app->api_version = PPS_API_VERS_1;
1079 case PPS_IOC_GETCAP:
1080 *(int*)data = pps->ppscap;
1083 fapi = (struct pps_fetch_args *)data;
1084 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1086 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1087 return (EOPNOTSUPP);
1088 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1089 fapi->pps_info_buf = pps->ppsinfo;
1091 case PPS_IOC_KCBIND:
1093 kapi = (struct pps_kcbind_args *)data;
1094 /* XXX Only root should be able to do this */
1095 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1097 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1099 if (kapi->edge & ~pps->ppscap)
1101 pps->kcmode = kapi->edge;
1104 return (EOPNOTSUPP);
1112 pps_init(struct pps_state *pps)
1114 pps->ppscap |= PPS_TSFMT_TSPEC;
1115 if (pps->ppscap & PPS_CAPTUREASSERT)
1116 pps->ppscap |= PPS_OFFSETASSERT;
1117 if (pps->ppscap & PPS_CAPTURECLEAR)
1118 pps->ppscap |= PPS_OFFSETCLEAR;
1122 pps_event(struct pps_state *pps, sysclock_t count, int event)
1124 struct globaldata *gd;
1125 struct timespec *tsp;
1126 struct timespec *osp;
1127 struct timespec *bt;
1140 /* Things would be easier with arrays... */
1141 if (event == PPS_CAPTUREASSERT) {
1142 tsp = &pps->ppsinfo.assert_timestamp;
1143 osp = &pps->ppsparam.assert_offset;
1144 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1145 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1146 pcount = &pps->ppscount[0];
1147 pseq = &pps->ppsinfo.assert_sequence;
1149 tsp = &pps->ppsinfo.clear_timestamp;
1150 osp = &pps->ppsparam.clear_offset;
1151 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1152 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1153 pcount = &pps->ppscount[1];
1154 pseq = &pps->ppsinfo.clear_sequence;
1157 /* Nothing really happened */
1158 if (*pcount == count)
1164 ts.tv_sec = gd->gd_time_seconds;
1165 delta = count - gd->gd_cpuclock_base;
1166 } while (ts.tv_sec != gd->gd_time_seconds);
1168 if (delta >= sys_cputimer->freq) {
1169 ts.tv_sec += delta / sys_cputimer->freq;
1170 delta %= sys_cputimer->freq;
1172 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1173 bt = &basetime[basetime_index];
1174 ts.tv_sec += bt->tv_sec;
1175 ts.tv_nsec += bt->tv_nsec;
1176 while (ts.tv_nsec >= 1000000000) {
1177 ts.tv_nsec -= 1000000000;
1185 timespecadd(tsp, osp);
1186 if (tsp->tv_nsec < 0) {
1187 tsp->tv_nsec += 1000000000;
1193 /* magic, at its best... */
1194 tcount = count - pps->ppscount[2];
1195 pps->ppscount[2] = count;
1196 if (tcount >= sys_cputimer->freq) {
1197 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1198 sys_cputimer->freq64_nsec *
1199 (tcount % sys_cputimer->freq)) >> 32;
1201 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1203 hardpps(tsp, delta);