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 1.61 2007/09/30 04:37:27 sephe Exp $
77 #include "opt_polling.h"
78 #include "opt_pctrack.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/callout.h>
83 #include <sys/kernel.h>
84 #include <sys/kinfo.h>
86 #include <sys/malloc.h>
87 #include <sys/resourcevar.h>
88 #include <sys/signalvar.h>
89 #include <sys/timex.h>
90 #include <sys/timepps.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_extern.h>
96 #include <sys/sysctl.h>
97 #include <sys/thread2.h>
99 #include <machine/cpu.h>
100 #include <machine/limits.h>
101 #include <machine/smp.h>
104 #include <sys/gmon.h>
107 #ifdef DEVICE_POLLING
108 extern void init_device_poll_pcpu(int);
112 static void do_pctrack(struct intrframe *frame, int which);
115 static void initclocks (void *dummy);
116 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
119 * Some of these don't belong here, but it's easiest to concentrate them.
120 * Note that cpu_time counts in microseconds, but most userland programs
121 * just compare relative times against the total by delta.
123 struct kinfo_cputime cputime_percpu[MAXCPU];
125 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
126 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
131 sysctl_cputime(SYSCTL_HANDLER_ARGS)
134 size_t size = sizeof(struct kinfo_cputime);
136 for (cpu = 0; cpu < ncpus; ++cpu) {
137 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
143 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
144 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
146 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
147 "CPU time statistics");
151 * boottime is used to calculate the 'real' uptime. Do not confuse this with
152 * microuptime(). microtime() is not drift compensated. The real uptime
153 * with compensation is nanotime() - bootime. boottime is recalculated
154 * whenever the real time is set based on the compensated elapsed time
155 * in seconds (gd->gd_time_seconds).
157 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
158 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
161 struct timespec boottime; /* boot time (realtime) for reference only */
162 time_t time_second; /* read-only 'passive' uptime in seconds */
165 * basetime is used to calculate the compensated real time of day. The
166 * basetime can be modified on a per-tick basis by the adjtime(),
167 * ntp_adjtime(), and sysctl-based time correction APIs.
169 * Note that frequency corrections can also be made by adjusting
172 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
173 * used on both SMP and UP systems to avoid MP races between cpu's and
174 * interrupt races on UP systems.
176 #define BASETIME_ARYSIZE 16
177 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
178 static struct timespec basetime[BASETIME_ARYSIZE];
179 static volatile int basetime_index;
182 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
189 * Because basetime data and index may be updated by another cpu,
190 * a load fence is required to ensure that the data we read has
191 * not been speculatively read relative to a possibly updated index.
193 index = basetime_index;
195 bt = &basetime[index];
196 error = SYSCTL_OUT(req, bt, sizeof(*bt));
200 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
201 &boottime, timespec, "System boottime");
202 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
203 sysctl_get_basetime, "S,timespec", "System basetime");
205 static void hardclock(systimer_t info, struct intrframe *frame);
206 static void statclock(systimer_t info, struct intrframe *frame);
207 static void schedclock(systimer_t info, struct intrframe *frame);
208 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
210 int ticks; /* system master ticks at hz */
211 int clocks_running; /* tsleep/timeout clocks operational */
212 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
213 int64_t nsec_acc; /* accumulator */
215 /* NTPD time correction fields */
216 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
217 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
218 int64_t ntp_delta; /* one-time correction in nsec */
219 int64_t ntp_big_delta = 1000000000;
220 int32_t ntp_tick_delta; /* current adjustment rate */
221 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
222 time_t ntp_leap_second; /* time of next leap second */
223 int ntp_leap_insert; /* whether to insert or remove a second */
226 * Finish initializing clock frequencies and start all clocks running.
230 initclocks(void *dummy)
232 /*psratio = profhz / stathz;*/
238 * Called on a per-cpu basis
241 initclocks_pcpu(void)
243 struct globaldata *gd = mycpu;
246 if (gd->gd_cpuid == 0) {
247 gd->gd_time_seconds = 1;
248 gd->gd_cpuclock_base = sys_cputimer->count();
251 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
252 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
255 #ifdef DEVICE_POLLING
256 init_device_poll_pcpu(gd->gd_cpuid);
260 * Use a non-queued periodic systimer to prevent multiple ticks from
261 * building up if the sysclock jumps forward (8254 gets reset). The
262 * sysclock will never jump backwards. Our time sync is based on
263 * the actual sysclock, not the ticks count.
265 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
266 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
267 /* XXX correct the frequency for scheduler / estcpu tests */
268 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
274 * This sets the current real time of day. Timespecs are in seconds and
275 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
276 * instead we adjust basetime so basetime + gd_* results in the current
277 * time of day. This way the gd_* fields are guarenteed to represent
278 * a monotonically increasing 'uptime' value.
280 * When set_timeofday() is called from userland, the system call forces it
281 * onto cpu #0 since only cpu #0 can update basetime_index.
284 set_timeofday(struct timespec *ts)
286 struct timespec *nbt;
290 * XXX SMP / non-atomic basetime updates
293 ni = (basetime_index + 1) & BASETIME_ARYMASK;
296 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
297 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
298 if (nbt->tv_nsec < 0) {
299 nbt->tv_nsec += 1000000000;
304 * Note that basetime diverges from boottime as the clock drift is
305 * compensated for, so we cannot do away with boottime. When setting
306 * the absolute time of day the drift is 0 (for an instant) and we
307 * can simply assign boottime to basetime.
309 * Note that nanouptime() is based on gd_time_seconds which is drift
310 * compensated up to a point (it is guarenteed to remain monotonically
311 * increasing). gd_time_seconds is thus our best uptime guess and
312 * suitable for use in the boottime calculation. It is already taken
313 * into account in the basetime calculation above.
315 boottime.tv_sec = nbt->tv_sec;
319 * We now have a new basetime, make sure all other cpus have it,
320 * then update the index.
329 * Each cpu has its own hardclock, but we only increments ticks and softticks
332 * NOTE! systimer! the MP lock might not be held here. We can only safely
333 * manipulate objects owned by the current cpu.
336 hardclock(systimer_t info, struct intrframe *frame)
340 struct globaldata *gd = mycpu;
343 * Realtime updates are per-cpu. Note that timer corrections as
344 * returned by microtime() and friends make an additional adjustment
345 * using a system-wise 'basetime', but the running time is always
346 * taken from the per-cpu globaldata area. Since the same clock
347 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
350 * Note that we never allow info->time (aka gd->gd_hardclock.time)
351 * to reverse index gd_cpuclock_base, but that it is possible for
352 * it to temporarily get behind in the seconds if something in the
353 * system locks interrupts for a long period of time. Since periodic
354 * timers count events, though everything should resynch again
357 cputicks = info->time - gd->gd_cpuclock_base;
358 if (cputicks >= sys_cputimer->freq) {
359 ++gd->gd_time_seconds;
360 gd->gd_cpuclock_base += sys_cputimer->freq;
364 * The system-wide ticks counter and NTP related timedelta/tickdelta
365 * adjustments only occur on cpu #0. NTP adjustments are accomplished
366 * by updating basetime.
368 if (gd->gd_cpuid == 0) {
369 struct timespec *nbt;
377 if (tco->tc_poll_pps)
378 tco->tc_poll_pps(tco);
382 * Calculate the new basetime index. We are in a critical section
383 * on cpu #0 and can safely play with basetime_index. Start
384 * with the current basetime and then make adjustments.
386 ni = (basetime_index + 1) & BASETIME_ARYMASK;
388 *nbt = basetime[basetime_index];
391 * Apply adjtime corrections. (adjtime() API)
393 * adjtime() only runs on cpu #0 so our critical section is
394 * sufficient to access these variables.
396 if (ntp_delta != 0) {
397 nbt->tv_nsec += ntp_tick_delta;
398 ntp_delta -= ntp_tick_delta;
399 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
400 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
401 ntp_tick_delta = ntp_delta;
406 * Apply permanent frequency corrections. (sysctl API)
408 if (ntp_tick_permanent != 0) {
409 ntp_tick_acc += ntp_tick_permanent;
410 if (ntp_tick_acc >= (1LL << 32)) {
411 nbt->tv_nsec += ntp_tick_acc >> 32;
412 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
413 } else if (ntp_tick_acc <= -(1LL << 32)) {
414 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
415 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
416 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
420 if (nbt->tv_nsec >= 1000000000) {
422 nbt->tv_nsec -= 1000000000;
423 } else if (nbt->tv_nsec < 0) {
425 nbt->tv_nsec += 1000000000;
429 * Another per-tick compensation. (for ntp_adjtime() API)
432 nsec_acc += nsec_adj;
433 if (nsec_acc >= 0x100000000LL) {
434 nbt->tv_nsec += nsec_acc >> 32;
435 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
436 } else if (nsec_acc <= -0x100000000LL) {
437 nbt->tv_nsec -= -nsec_acc >> 32;
438 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
440 if (nbt->tv_nsec >= 1000000000) {
441 nbt->tv_nsec -= 1000000000;
443 } else if (nbt->tv_nsec < 0) {
444 nbt->tv_nsec += 1000000000;
449 /************************************************************
450 * LEAP SECOND CORRECTION *
451 ************************************************************
453 * Taking into account all the corrections made above, figure
454 * out the new real time. If the seconds field has changed
455 * then apply any pending leap-second corrections.
457 getnanotime_nbt(nbt, &nts);
459 if (time_second != nts.tv_sec) {
461 * Apply leap second (sysctl API). Adjust nts for changes
462 * so we do not have to call getnanotime_nbt again.
464 if (ntp_leap_second) {
465 if (ntp_leap_second == nts.tv_sec) {
466 if (ntp_leap_insert) {
478 * Apply leap second (ntp_adjtime() API), calculate a new
479 * nsec_adj field. ntp_update_second() returns nsec_adj
480 * as a per-second value but we need it as a per-tick value.
482 leap = ntp_update_second(time_second, &nsec_adj);
488 * Update the time_second 'approximate time' global.
490 time_second = nts.tv_sec;
494 * Finally, our new basetime is ready to go live!
500 * Figure out how badly the system is starved for memory
502 vm_fault_ratecheck();
506 * softticks are handled for all cpus
508 hardclock_softtick(gd);
511 * ITimer handling is per-tick, per-cpu. I don't think ksignal()
512 * is mpsafe on curproc, so XXX get the mplock.
514 if ((p = curproc) != NULL && try_mplock()) {
515 if (frame && CLKF_USERMODE(frame) &&
516 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
517 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], tick) == 0)
518 ksignal(p, SIGVTALRM);
519 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
520 itimerdecr(&p->p_timer[ITIMER_PROF], tick) == 0)
528 * The statistics clock typically runs at a 125Hz rate, and is intended
529 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
531 * NOTE! systimer! the MP lock might not be held here. We can only safely
532 * manipulate objects owned by the current cpu.
534 * The stats clock is responsible for grabbing a profiling sample.
535 * Most of the statistics are only used by user-level statistics programs.
536 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
539 * Like the other clocks, the stat clock is called from what is effectively
540 * a fast interrupt, so the context should be the thread/process that got
544 statclock(systimer_t info, struct intrframe *frame)
557 * How big was our timeslice relative to the last time?
559 microuptime(&tv); /* mpsafe */
560 stv = &mycpu->gd_stattv;
561 if (stv->tv_sec == 0) {
564 bump = tv.tv_usec - stv->tv_usec +
565 (tv.tv_sec - stv->tv_sec) * 1000000;
576 if (frame && CLKF_USERMODE(frame)) {
578 * Came from userland, handle user time and deal with
581 if (p && (p->p_flag & P_PROFIL))
582 addupc_intr(p, CLKF_PC(frame), 1);
583 td->td_uticks += bump;
586 * Charge the time as appropriate
588 if (p && p->p_nice > NZERO)
589 cpu_time.cp_nice += bump;
591 cpu_time.cp_user += bump;
595 * Kernel statistics are just like addupc_intr, only easier.
598 if (g->state == GMON_PROF_ON && frame) {
599 i = CLKF_PC(frame) - g->lowpc;
600 if (i < g->textsize) {
601 i /= HISTFRACTION * sizeof(*g->kcount);
607 * Came from kernel mode, so we were:
608 * - handling an interrupt,
609 * - doing syscall or trap work on behalf of the current
611 * - spinning in the idle loop.
612 * Whichever it is, charge the time as appropriate.
613 * Note that we charge interrupts to the current process,
614 * regardless of whether they are ``for'' that process,
615 * so that we know how much of its real time was spent
616 * in ``non-process'' (i.e., interrupt) work.
618 * XXX assume system if frame is NULL. A NULL frame
619 * can occur if ipi processing is done from a crit_exit().
621 if (frame && CLKF_INTR(frame))
622 td->td_iticks += bump;
624 td->td_sticks += bump;
626 if (frame && CLKF_INTR(frame)) {
628 do_pctrack(frame, PCTRACK_INT);
630 cpu_time.cp_intr += bump;
632 if (td == &mycpu->gd_idlethread) {
633 cpu_time.cp_idle += bump;
637 do_pctrack(frame, PCTRACK_SYS);
639 cpu_time.cp_sys += bump;
647 * Sample the PC when in the kernel or in an interrupt. User code can
648 * retrieve the information and generate a histogram or other output.
652 do_pctrack(struct intrframe *frame, int which)
654 struct kinfo_pctrack *pctrack;
656 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
657 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
658 (void *)CLKF_PC(frame);
663 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
665 struct kinfo_pcheader head;
670 head.pc_ntrack = PCTRACK_SIZE;
671 head.pc_arysize = PCTRACK_ARYSIZE;
673 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
676 for (cpu = 0; cpu < ncpus; ++cpu) {
677 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
678 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
679 sizeof(struct kinfo_pctrack));
688 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
689 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
694 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
695 * the MP lock might not be held. We can safely manipulate parts of curproc
696 * but that's about it.
698 * Each cpu has its own scheduler clock.
701 schedclock(systimer_t info, struct intrframe *frame)
708 if ((lp = lwkt_preempted_proc()) != NULL) {
710 * Account for cpu time used and hit the scheduler. Note
711 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
715 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
718 if ((lp = curthread->td_lwp) != NULL) {
720 * Update resource usage integrals and maximums.
722 if ((ru = &lp->lwp_proc->p_ru) &&
723 (vm = lp->lwp_proc->p_vmspace) != NULL) {
724 ru->ru_ixrss += pgtok(vm->vm_tsize);
725 ru->ru_idrss += pgtok(vm->vm_dsize);
726 ru->ru_isrss += pgtok(vm->vm_ssize);
727 rss = pgtok(vmspace_resident_count(vm));
728 if (ru->ru_maxrss < rss)
735 * Compute number of ticks for the specified amount of time. The
736 * return value is intended to be used in a clock interrupt timed
737 * operation and guarenteed to meet or exceed the requested time.
738 * If the representation overflows, return INT_MAX. The minimum return
739 * value is 1 ticks and the function will average the calculation up.
740 * If any value greater then 0 microseconds is supplied, a value
741 * of at least 2 will be returned to ensure that a near-term clock
742 * interrupt does not cause the timeout to occur (degenerately) early.
744 * Note that limit checks must take into account microseconds, which is
745 * done simply by using the smaller signed long maximum instead of
746 * the unsigned long maximum.
748 * If ints have 32 bits, then the maximum value for any timeout in
749 * 10ms ticks is 248 days.
752 tvtohz_high(struct timeval *tv)
769 kprintf("tvtohz_high: negative time difference %ld sec %ld usec\n",
773 } else if (sec <= INT_MAX / hz) {
774 ticks = (int)(sec * hz +
775 ((u_long)usec + (tick - 1)) / tick) + 1;
783 * Compute number of ticks for the specified amount of time, erroring on
784 * the side of it being too low to ensure that sleeping the returned number
785 * of ticks will not result in a late return.
787 * The supplied timeval may not be negative and should be normalized. A
788 * return value of 0 is possible if the timeval converts to less then
791 * If ints have 32 bits, then the maximum value for any timeout in
792 * 10ms ticks is 248 days.
795 tvtohz_low(struct timeval *tv)
801 if (sec <= INT_MAX / hz)
802 ticks = (int)(sec * hz + (u_long)tv->tv_usec / tick);
810 * Start profiling on a process.
812 * Kernel profiling passes proc0 which never exits and hence
813 * keeps the profile clock running constantly.
816 startprofclock(struct proc *p)
818 if ((p->p_flag & P_PROFIL) == 0) {
819 p->p_flag |= P_PROFIL;
821 if (++profprocs == 1 && stathz != 0) {
824 setstatclockrate(profhz);
832 * Stop profiling on a process.
835 stopprofclock(struct proc *p)
837 if (p->p_flag & P_PROFIL) {
838 p->p_flag &= ~P_PROFIL;
840 if (--profprocs == 0 && stathz != 0) {
843 setstatclockrate(stathz);
851 * Return information about system clocks.
854 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
856 struct kinfo_clockinfo clkinfo;
858 * Construct clockinfo structure.
861 clkinfo.ci_tick = tick;
862 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
863 clkinfo.ci_profhz = profhz;
864 clkinfo.ci_stathz = stathz ? stathz : hz;
865 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
868 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
869 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
872 * We have eight functions for looking at the clock, four for
873 * microseconds and four for nanoseconds. For each there is fast
874 * but less precise version "get{nano|micro}[up]time" which will
875 * return a time which is up to 1/HZ previous to the call, whereas
876 * the raw version "{nano|micro}[up]time" will return a timestamp
877 * which is as precise as possible. The "up" variants return the
878 * time relative to system boot, these are well suited for time
879 * interval measurements.
881 * Each cpu independantly maintains the current time of day, so all
882 * we need to do to protect ourselves from changes is to do a loop
883 * check on the seconds field changing out from under us.
885 * The system timer maintains a 32 bit count and due to various issues
886 * it is possible for the calculated delta to occassionally exceed
887 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
888 * multiplication can easily overflow, so we deal with the case. For
889 * uniformity we deal with the case in the usec case too.
892 getmicrouptime(struct timeval *tvp)
894 struct globaldata *gd = mycpu;
898 tvp->tv_sec = gd->gd_time_seconds;
899 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
900 } while (tvp->tv_sec != gd->gd_time_seconds);
902 if (delta >= sys_cputimer->freq) {
903 tvp->tv_sec += delta / sys_cputimer->freq;
904 delta %= sys_cputimer->freq;
906 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
907 if (tvp->tv_usec >= 1000000) {
908 tvp->tv_usec -= 1000000;
914 getnanouptime(struct timespec *tsp)
916 struct globaldata *gd = mycpu;
920 tsp->tv_sec = gd->gd_time_seconds;
921 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
922 } while (tsp->tv_sec != gd->gd_time_seconds);
924 if (delta >= sys_cputimer->freq) {
925 tsp->tv_sec += delta / sys_cputimer->freq;
926 delta %= sys_cputimer->freq;
928 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
932 microuptime(struct timeval *tvp)
934 struct globaldata *gd = mycpu;
938 tvp->tv_sec = gd->gd_time_seconds;
939 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
940 } while (tvp->tv_sec != gd->gd_time_seconds);
942 if (delta >= sys_cputimer->freq) {
943 tvp->tv_sec += delta / sys_cputimer->freq;
944 delta %= sys_cputimer->freq;
946 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
950 nanouptime(struct timespec *tsp)
952 struct globaldata *gd = mycpu;
956 tsp->tv_sec = gd->gd_time_seconds;
957 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
958 } while (tsp->tv_sec != gd->gd_time_seconds);
960 if (delta >= sys_cputimer->freq) {
961 tsp->tv_sec += delta / sys_cputimer->freq;
962 delta %= sys_cputimer->freq;
964 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
972 getmicrotime(struct timeval *tvp)
974 struct globaldata *gd = mycpu;
979 tvp->tv_sec = gd->gd_time_seconds;
980 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
981 } while (tvp->tv_sec != gd->gd_time_seconds);
983 if (delta >= sys_cputimer->freq) {
984 tvp->tv_sec += delta / sys_cputimer->freq;
985 delta %= sys_cputimer->freq;
987 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
989 bt = &basetime[basetime_index];
990 tvp->tv_sec += bt->tv_sec;
991 tvp->tv_usec += bt->tv_nsec / 1000;
992 while (tvp->tv_usec >= 1000000) {
993 tvp->tv_usec -= 1000000;
999 getnanotime(struct timespec *tsp)
1001 struct globaldata *gd = mycpu;
1002 struct timespec *bt;
1006 tsp->tv_sec = gd->gd_time_seconds;
1007 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1008 } while (tsp->tv_sec != gd->gd_time_seconds);
1010 if (delta >= sys_cputimer->freq) {
1011 tsp->tv_sec += delta / sys_cputimer->freq;
1012 delta %= sys_cputimer->freq;
1014 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1016 bt = &basetime[basetime_index];
1017 tsp->tv_sec += bt->tv_sec;
1018 tsp->tv_nsec += bt->tv_nsec;
1019 while (tsp->tv_nsec >= 1000000000) {
1020 tsp->tv_nsec -= 1000000000;
1026 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1028 struct globaldata *gd = mycpu;
1032 tsp->tv_sec = gd->gd_time_seconds;
1033 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1034 } while (tsp->tv_sec != gd->gd_time_seconds);
1036 if (delta >= sys_cputimer->freq) {
1037 tsp->tv_sec += delta / sys_cputimer->freq;
1038 delta %= sys_cputimer->freq;
1040 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1042 tsp->tv_sec += nbt->tv_sec;
1043 tsp->tv_nsec += nbt->tv_nsec;
1044 while (tsp->tv_nsec >= 1000000000) {
1045 tsp->tv_nsec -= 1000000000;
1052 microtime(struct timeval *tvp)
1054 struct globaldata *gd = mycpu;
1055 struct timespec *bt;
1059 tvp->tv_sec = gd->gd_time_seconds;
1060 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1061 } while (tvp->tv_sec != gd->gd_time_seconds);
1063 if (delta >= sys_cputimer->freq) {
1064 tvp->tv_sec += delta / sys_cputimer->freq;
1065 delta %= sys_cputimer->freq;
1067 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1069 bt = &basetime[basetime_index];
1070 tvp->tv_sec += bt->tv_sec;
1071 tvp->tv_usec += bt->tv_nsec / 1000;
1072 while (tvp->tv_usec >= 1000000) {
1073 tvp->tv_usec -= 1000000;
1079 nanotime(struct timespec *tsp)
1081 struct globaldata *gd = mycpu;
1082 struct timespec *bt;
1086 tsp->tv_sec = gd->gd_time_seconds;
1087 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1088 } while (tsp->tv_sec != gd->gd_time_seconds);
1090 if (delta >= sys_cputimer->freq) {
1091 tsp->tv_sec += delta / sys_cputimer->freq;
1092 delta %= sys_cputimer->freq;
1094 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1096 bt = &basetime[basetime_index];
1097 tsp->tv_sec += bt->tv_sec;
1098 tsp->tv_nsec += bt->tv_nsec;
1099 while (tsp->tv_nsec >= 1000000000) {
1100 tsp->tv_nsec -= 1000000000;
1106 * note: this is not exactly synchronized with real time. To do that we
1107 * would have to do what microtime does and check for a nanoseconds overflow.
1110 get_approximate_time_t(void)
1112 struct globaldata *gd = mycpu;
1113 struct timespec *bt;
1115 bt = &basetime[basetime_index];
1116 return(gd->gd_time_seconds + bt->tv_sec);
1120 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1123 struct pps_fetch_args *fapi;
1125 struct pps_kcbind_args *kapi;
1129 case PPS_IOC_CREATE:
1131 case PPS_IOC_DESTROY:
1133 case PPS_IOC_SETPARAMS:
1134 app = (pps_params_t *)data;
1135 if (app->mode & ~pps->ppscap)
1137 pps->ppsparam = *app;
1139 case PPS_IOC_GETPARAMS:
1140 app = (pps_params_t *)data;
1141 *app = pps->ppsparam;
1142 app->api_version = PPS_API_VERS_1;
1144 case PPS_IOC_GETCAP:
1145 *(int*)data = pps->ppscap;
1148 fapi = (struct pps_fetch_args *)data;
1149 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1151 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1152 return (EOPNOTSUPP);
1153 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1154 fapi->pps_info_buf = pps->ppsinfo;
1156 case PPS_IOC_KCBIND:
1158 kapi = (struct pps_kcbind_args *)data;
1159 /* XXX Only root should be able to do this */
1160 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1162 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1164 if (kapi->edge & ~pps->ppscap)
1166 pps->kcmode = kapi->edge;
1169 return (EOPNOTSUPP);
1177 pps_init(struct pps_state *pps)
1179 pps->ppscap |= PPS_TSFMT_TSPEC;
1180 if (pps->ppscap & PPS_CAPTUREASSERT)
1181 pps->ppscap |= PPS_OFFSETASSERT;
1182 if (pps->ppscap & PPS_CAPTURECLEAR)
1183 pps->ppscap |= PPS_OFFSETCLEAR;
1187 pps_event(struct pps_state *pps, sysclock_t count, int event)
1189 struct globaldata *gd;
1190 struct timespec *tsp;
1191 struct timespec *osp;
1192 struct timespec *bt;
1205 /* Things would be easier with arrays... */
1206 if (event == PPS_CAPTUREASSERT) {
1207 tsp = &pps->ppsinfo.assert_timestamp;
1208 osp = &pps->ppsparam.assert_offset;
1209 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1210 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1211 pcount = &pps->ppscount[0];
1212 pseq = &pps->ppsinfo.assert_sequence;
1214 tsp = &pps->ppsinfo.clear_timestamp;
1215 osp = &pps->ppsparam.clear_offset;
1216 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1217 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1218 pcount = &pps->ppscount[1];
1219 pseq = &pps->ppsinfo.clear_sequence;
1222 /* Nothing really happened */
1223 if (*pcount == count)
1229 ts.tv_sec = gd->gd_time_seconds;
1230 delta = count - gd->gd_cpuclock_base;
1231 } while (ts.tv_sec != gd->gd_time_seconds);
1233 if (delta >= sys_cputimer->freq) {
1234 ts.tv_sec += delta / sys_cputimer->freq;
1235 delta %= sys_cputimer->freq;
1237 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1238 bt = &basetime[basetime_index];
1239 ts.tv_sec += bt->tv_sec;
1240 ts.tv_nsec += bt->tv_nsec;
1241 while (ts.tv_nsec >= 1000000000) {
1242 ts.tv_nsec -= 1000000000;
1250 timespecadd(tsp, osp);
1251 if (tsp->tv_nsec < 0) {
1252 tsp->tv_nsec += 1000000000;
1258 /* magic, at its best... */
1259 tcount = count - pps->ppscount[2];
1260 pps->ppscount[2] = count;
1261 if (tcount >= sys_cputimer->freq) {
1262 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1263 sys_cputimer->freq64_nsec *
1264 (tcount % sys_cputimer->freq)) >> 32;
1266 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1268 hardpps(tsp, delta);