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 1.62 2008/09/09 04:06:13 dillon Exp $
77 #include "opt_polling.h"
78 #include "opt_ifpoll.h"
79 #include "opt_pctrack.h"
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/callout.h>
84 #include <sys/kernel.h>
85 #include <sys/kinfo.h>
87 #include <sys/malloc.h>
88 #include <sys/resourcevar.h>
89 #include <sys/signalvar.h>
90 #include <sys/timex.h>
91 #include <sys/timepps.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_extern.h>
97 #include <sys/sysctl.h>
99 #include <sys/thread2.h>
100 #include <sys/mplock2.h>
102 #include <machine/cpu.h>
103 #include <machine/limits.h>
104 #include <machine/smp.h>
105 #include <machine/cpufunc.h>
106 #include <machine/specialreg.h>
107 #include <machine/clock.h>
110 #include <sys/gmon.h>
113 #ifdef DEVICE_POLLING
114 extern void init_device_poll_pcpu(int);
118 extern void ifpoll_init_pcpu(int);
122 static void do_pctrack(struct intrframe *frame, int which);
125 static void initclocks (void *dummy);
126 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
129 * Some of these don't belong here, but it's easiest to concentrate them.
130 * Note that cpu_time counts in microseconds, but most userland programs
131 * just compare relative times against the total by delta.
133 struct kinfo_cputime cputime_percpu[MAXCPU];
135 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
136 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
141 sysctl_cputime(SYSCTL_HANDLER_ARGS)
144 size_t size = sizeof(struct kinfo_cputime);
146 for (cpu = 0; cpu < ncpus; ++cpu) {
147 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
153 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
154 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
156 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
157 "CPU time statistics");
161 * boottime is used to calculate the 'real' uptime. Do not confuse this with
162 * microuptime(). microtime() is not drift compensated. The real uptime
163 * with compensation is nanotime() - bootime. boottime is recalculated
164 * whenever the real time is set based on the compensated elapsed time
165 * in seconds (gd->gd_time_seconds).
167 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
168 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
171 struct timespec boottime; /* boot time (realtime) for reference only */
172 time_t time_second; /* read-only 'passive' uptime in seconds */
175 * basetime is used to calculate the compensated real time of day. The
176 * basetime can be modified on a per-tick basis by the adjtime(),
177 * ntp_adjtime(), and sysctl-based time correction APIs.
179 * Note that frequency corrections can also be made by adjusting
182 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
183 * used on both SMP and UP systems to avoid MP races between cpu's and
184 * interrupt races on UP systems.
186 #define BASETIME_ARYSIZE 16
187 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
188 static struct timespec basetime[BASETIME_ARYSIZE];
189 static volatile int basetime_index;
192 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
199 * Because basetime data and index may be updated by another cpu,
200 * a load fence is required to ensure that the data we read has
201 * not been speculatively read relative to a possibly updated index.
203 index = basetime_index;
205 bt = &basetime[index];
206 error = SYSCTL_OUT(req, bt, sizeof(*bt));
210 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
211 &boottime, timespec, "System boottime");
212 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
213 sysctl_get_basetime, "S,timespec", "System basetime");
215 static void hardclock(systimer_t info, struct intrframe *frame);
216 static void statclock(systimer_t info, struct intrframe *frame);
217 static void schedclock(systimer_t info, struct intrframe *frame);
218 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
220 int ticks; /* system master ticks at hz */
221 int clocks_running; /* tsleep/timeout clocks operational */
222 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
223 int64_t nsec_acc; /* accumulator */
225 /* NTPD time correction fields */
226 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
227 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
228 int64_t ntp_delta; /* one-time correction in nsec */
229 int64_t ntp_big_delta = 1000000000;
230 int32_t ntp_tick_delta; /* current adjustment rate */
231 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
232 time_t ntp_leap_second; /* time of next leap second */
233 int ntp_leap_insert; /* whether to insert or remove a second */
236 * Finish initializing clock frequencies and start all clocks running.
240 initclocks(void *dummy)
242 /*psratio = profhz / stathz;*/
248 * Called on a per-cpu basis
251 initclocks_pcpu(void)
253 struct globaldata *gd = mycpu;
256 if (gd->gd_cpuid == 0) {
257 gd->gd_time_seconds = 1;
258 gd->gd_cpuclock_base = sys_cputimer->count();
261 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
262 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
265 systimer_intr_enable();
267 #ifdef DEVICE_POLLING
268 init_device_poll_pcpu(gd->gd_cpuid);
272 ifpoll_init_pcpu(gd->gd_cpuid);
276 * Use a non-queued periodic systimer to prevent multiple ticks from
277 * building up if the sysclock jumps forward (8254 gets reset). The
278 * sysclock will never jump backwards. Our time sync is based on
279 * the actual sysclock, not the ticks count.
281 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
282 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
283 /* XXX correct the frequency for scheduler / estcpu tests */
284 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
290 * This sets the current real time of day. Timespecs are in seconds and
291 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
292 * instead we adjust basetime so basetime + gd_* results in the current
293 * time of day. This way the gd_* fields are guarenteed to represent
294 * a monotonically increasing 'uptime' value.
296 * When set_timeofday() is called from userland, the system call forces it
297 * onto cpu #0 since only cpu #0 can update basetime_index.
300 set_timeofday(struct timespec *ts)
302 struct timespec *nbt;
306 * XXX SMP / non-atomic basetime updates
309 ni = (basetime_index + 1) & BASETIME_ARYMASK;
312 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
313 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
314 if (nbt->tv_nsec < 0) {
315 nbt->tv_nsec += 1000000000;
320 * Note that basetime diverges from boottime as the clock drift is
321 * compensated for, so we cannot do away with boottime. When setting
322 * the absolute time of day the drift is 0 (for an instant) and we
323 * can simply assign boottime to basetime.
325 * Note that nanouptime() is based on gd_time_seconds which is drift
326 * compensated up to a point (it is guarenteed to remain monotonically
327 * increasing). gd_time_seconds is thus our best uptime guess and
328 * suitable for use in the boottime calculation. It is already taken
329 * into account in the basetime calculation above.
331 boottime.tv_sec = nbt->tv_sec;
335 * We now have a new basetime, make sure all other cpus have it,
336 * then update the index.
345 * Each cpu has its own hardclock, but we only increments ticks and softticks
348 * NOTE! systimer! the MP lock might not be held here. We can only safely
349 * manipulate objects owned by the current cpu.
352 hardclock(systimer_t info, struct intrframe *frame)
356 struct globaldata *gd = mycpu;
359 * Realtime updates are per-cpu. Note that timer corrections as
360 * returned by microtime() and friends make an additional adjustment
361 * using a system-wise 'basetime', but the running time is always
362 * taken from the per-cpu globaldata area. Since the same clock
363 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
366 * Note that we never allow info->time (aka gd->gd_hardclock.time)
367 * to reverse index gd_cpuclock_base, but that it is possible for
368 * it to temporarily get behind in the seconds if something in the
369 * system locks interrupts for a long period of time. Since periodic
370 * timers count events, though everything should resynch again
373 cputicks = info->time - gd->gd_cpuclock_base;
374 if (cputicks >= sys_cputimer->freq) {
375 ++gd->gd_time_seconds;
376 gd->gd_cpuclock_base += sys_cputimer->freq;
380 * The system-wide ticks counter and NTP related timedelta/tickdelta
381 * adjustments only occur on cpu #0. NTP adjustments are accomplished
382 * by updating basetime.
384 if (gd->gd_cpuid == 0) {
385 struct timespec *nbt;
393 if (tco->tc_poll_pps)
394 tco->tc_poll_pps(tco);
398 * Calculate the new basetime index. We are in a critical section
399 * on cpu #0 and can safely play with basetime_index. Start
400 * with the current basetime and then make adjustments.
402 ni = (basetime_index + 1) & BASETIME_ARYMASK;
404 *nbt = basetime[basetime_index];
407 * Apply adjtime corrections. (adjtime() API)
409 * adjtime() only runs on cpu #0 so our critical section is
410 * sufficient to access these variables.
412 if (ntp_delta != 0) {
413 nbt->tv_nsec += ntp_tick_delta;
414 ntp_delta -= ntp_tick_delta;
415 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
416 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
417 ntp_tick_delta = ntp_delta;
422 * Apply permanent frequency corrections. (sysctl API)
424 if (ntp_tick_permanent != 0) {
425 ntp_tick_acc += ntp_tick_permanent;
426 if (ntp_tick_acc >= (1LL << 32)) {
427 nbt->tv_nsec += ntp_tick_acc >> 32;
428 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
429 } else if (ntp_tick_acc <= -(1LL << 32)) {
430 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
431 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
432 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
436 if (nbt->tv_nsec >= 1000000000) {
438 nbt->tv_nsec -= 1000000000;
439 } else if (nbt->tv_nsec < 0) {
441 nbt->tv_nsec += 1000000000;
445 * Another per-tick compensation. (for ntp_adjtime() API)
448 nsec_acc += nsec_adj;
449 if (nsec_acc >= 0x100000000LL) {
450 nbt->tv_nsec += nsec_acc >> 32;
451 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
452 } else if (nsec_acc <= -0x100000000LL) {
453 nbt->tv_nsec -= -nsec_acc >> 32;
454 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
456 if (nbt->tv_nsec >= 1000000000) {
457 nbt->tv_nsec -= 1000000000;
459 } else if (nbt->tv_nsec < 0) {
460 nbt->tv_nsec += 1000000000;
465 /************************************************************
466 * LEAP SECOND CORRECTION *
467 ************************************************************
469 * Taking into account all the corrections made above, figure
470 * out the new real time. If the seconds field has changed
471 * then apply any pending leap-second corrections.
473 getnanotime_nbt(nbt, &nts);
475 if (time_second != nts.tv_sec) {
477 * Apply leap second (sysctl API). Adjust nts for changes
478 * so we do not have to call getnanotime_nbt again.
480 if (ntp_leap_second) {
481 if (ntp_leap_second == nts.tv_sec) {
482 if (ntp_leap_insert) {
494 * Apply leap second (ntp_adjtime() API), calculate a new
495 * nsec_adj field. ntp_update_second() returns nsec_adj
496 * as a per-second value but we need it as a per-tick value.
498 leap = ntp_update_second(time_second, &nsec_adj);
504 * Update the time_second 'approximate time' global.
506 time_second = nts.tv_sec;
510 * Finally, our new basetime is ready to go live!
516 * Figure out how badly the system is starved for memory
518 vm_fault_ratecheck();
522 * lwkt thread scheduler fair queueing
524 lwkt_fairq_schedulerclock(curthread);
527 * softticks are handled for all cpus
529 hardclock_softtick(gd);
532 * The LWKT scheduler will generally allow the current process to
533 * return to user mode even if there are other runnable LWKT threads
534 * running in kernel mode on behalf of a user process. This will
535 * ensure that those other threads have an opportunity to run in
536 * fairly short order (but not instantly).
541 * ITimer handling is per-tick, per-cpu. I don't think ksignal()
542 * is mpsafe on curproc, so XXX get the mplock.
544 if ((p = curproc) != NULL && try_mplock()) {
545 if (frame && CLKF_USERMODE(frame) &&
546 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
547 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0)
548 ksignal(p, SIGVTALRM);
549 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
550 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0)
558 * The statistics clock typically runs at a 125Hz rate, and is intended
559 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
561 * NOTE! systimer! the MP lock might not be held here. We can only safely
562 * manipulate objects owned by the current cpu.
564 * The stats clock is responsible for grabbing a profiling sample.
565 * Most of the statistics are only used by user-level statistics programs.
566 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
569 * Like the other clocks, the stat clock is called from what is effectively
570 * a fast interrupt, so the context should be the thread/process that got
574 statclock(systimer_t info, struct intrframe *frame)
587 * How big was our timeslice relative to the last time?
589 microuptime(&tv); /* mpsafe */
590 stv = &mycpu->gd_stattv;
591 if (stv->tv_sec == 0) {
594 bump = tv.tv_usec - stv->tv_usec +
595 (tv.tv_sec - stv->tv_sec) * 1000000;
606 if (frame && CLKF_USERMODE(frame)) {
608 * Came from userland, handle user time and deal with
611 if (p && (p->p_flag & P_PROFIL))
612 addupc_intr(p, CLKF_PC(frame), 1);
613 td->td_uticks += bump;
616 * Charge the time as appropriate
618 if (p && p->p_nice > NZERO)
619 cpu_time.cp_nice += bump;
621 cpu_time.cp_user += bump;
625 * Kernel statistics are just like addupc_intr, only easier.
628 if (g->state == GMON_PROF_ON && frame) {
629 i = CLKF_PC(frame) - g->lowpc;
630 if (i < g->textsize) {
631 i /= HISTFRACTION * sizeof(*g->kcount);
637 * Came from kernel mode, so we were:
638 * - handling an interrupt,
639 * - doing syscall or trap work on behalf of the current
641 * - spinning in the idle loop.
642 * Whichever it is, charge the time as appropriate.
643 * Note that we charge interrupts to the current process,
644 * regardless of whether they are ``for'' that process,
645 * so that we know how much of its real time was spent
646 * in ``non-process'' (i.e., interrupt) work.
648 * XXX assume system if frame is NULL. A NULL frame
649 * can occur if ipi processing is done from a crit_exit().
651 if (frame && CLKF_INTR(frame))
652 td->td_iticks += bump;
654 td->td_sticks += bump;
656 if (frame && CLKF_INTR(frame)) {
658 do_pctrack(frame, PCTRACK_INT);
660 cpu_time.cp_intr += bump;
662 if (td == &mycpu->gd_idlethread) {
663 cpu_time.cp_idle += bump;
667 do_pctrack(frame, PCTRACK_SYS);
669 cpu_time.cp_sys += bump;
677 * Sample the PC when in the kernel or in an interrupt. User code can
678 * retrieve the information and generate a histogram or other output.
682 do_pctrack(struct intrframe *frame, int which)
684 struct kinfo_pctrack *pctrack;
686 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
687 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
688 (void *)CLKF_PC(frame);
693 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
695 struct kinfo_pcheader head;
700 head.pc_ntrack = PCTRACK_SIZE;
701 head.pc_arysize = PCTRACK_ARYSIZE;
703 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
706 for (cpu = 0; cpu < ncpus; ++cpu) {
707 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
708 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
709 sizeof(struct kinfo_pctrack));
718 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
719 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
724 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
725 * the MP lock might not be held. We can safely manipulate parts of curproc
726 * but that's about it.
728 * Each cpu has its own scheduler clock.
731 schedclock(systimer_t info, struct intrframe *frame)
738 if ((lp = lwkt_preempted_proc()) != NULL) {
740 * Account for cpu time used and hit the scheduler. Note
741 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
745 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
748 if ((lp = curthread->td_lwp) != NULL) {
750 * Update resource usage integrals and maximums.
752 if ((ru = &lp->lwp_proc->p_ru) &&
753 (vm = lp->lwp_proc->p_vmspace) != NULL) {
754 ru->ru_ixrss += pgtok(vm->vm_tsize);
755 ru->ru_idrss += pgtok(vm->vm_dsize);
756 ru->ru_isrss += pgtok(vm->vm_ssize);
757 rss = pgtok(vmspace_resident_count(vm));
758 if (ru->ru_maxrss < rss)
765 * Compute number of ticks for the specified amount of time. The
766 * return value is intended to be used in a clock interrupt timed
767 * operation and guarenteed to meet or exceed the requested time.
768 * If the representation overflows, return INT_MAX. The minimum return
769 * value is 1 ticks and the function will average the calculation up.
770 * If any value greater then 0 microseconds is supplied, a value
771 * of at least 2 will be returned to ensure that a near-term clock
772 * interrupt does not cause the timeout to occur (degenerately) early.
774 * Note that limit checks must take into account microseconds, which is
775 * done simply by using the smaller signed long maximum instead of
776 * the unsigned long maximum.
778 * If ints have 32 bits, then the maximum value for any timeout in
779 * 10ms ticks is 248 days.
782 tvtohz_high(struct timeval *tv)
799 kprintf("tvtohz_high: negative time difference "
800 "%ld sec %ld usec\n",
804 } else if (sec <= INT_MAX / hz) {
805 ticks = (int)(sec * hz +
806 ((u_long)usec + (ustick - 1)) / ustick) + 1;
814 tstohz_high(struct timespec *ts)
831 kprintf("tstohz_high: negative time difference "
832 "%ld sec %ld nsec\n",
836 } else if (sec <= INT_MAX / hz) {
837 ticks = (int)(sec * hz +
838 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
847 * Compute number of ticks for the specified amount of time, erroring on
848 * the side of it being too low to ensure that sleeping the returned number
849 * of ticks will not result in a late return.
851 * The supplied timeval may not be negative and should be normalized. A
852 * return value of 0 is possible if the timeval converts to less then
855 * If ints have 32 bits, then the maximum value for any timeout in
856 * 10ms ticks is 248 days.
859 tvtohz_low(struct timeval *tv)
865 if (sec <= INT_MAX / hz)
866 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
873 tstohz_low(struct timespec *ts)
879 if (sec <= INT_MAX / hz)
880 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
887 * Start profiling on a process.
889 * Kernel profiling passes proc0 which never exits and hence
890 * keeps the profile clock running constantly.
893 startprofclock(struct proc *p)
895 if ((p->p_flag & P_PROFIL) == 0) {
896 p->p_flag |= P_PROFIL;
898 if (++profprocs == 1 && stathz != 0) {
901 setstatclockrate(profhz);
909 * Stop profiling on a process.
912 stopprofclock(struct proc *p)
914 if (p->p_flag & P_PROFIL) {
915 p->p_flag &= ~P_PROFIL;
917 if (--profprocs == 0 && stathz != 0) {
920 setstatclockrate(stathz);
928 * Return information about system clocks.
931 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
933 struct kinfo_clockinfo clkinfo;
935 * Construct clockinfo structure.
938 clkinfo.ci_tick = ustick;
939 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
940 clkinfo.ci_profhz = profhz;
941 clkinfo.ci_stathz = stathz ? stathz : hz;
942 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
945 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
946 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
949 * We have eight functions for looking at the clock, four for
950 * microseconds and four for nanoseconds. For each there is fast
951 * but less precise version "get{nano|micro}[up]time" which will
952 * return a time which is up to 1/HZ previous to the call, whereas
953 * the raw version "{nano|micro}[up]time" will return a timestamp
954 * which is as precise as possible. The "up" variants return the
955 * time relative to system boot, these are well suited for time
956 * interval measurements.
958 * Each cpu independantly maintains the current time of day, so all
959 * we need to do to protect ourselves from changes is to do a loop
960 * check on the seconds field changing out from under us.
962 * The system timer maintains a 32 bit count and due to various issues
963 * it is possible for the calculated delta to occassionally exceed
964 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
965 * multiplication can easily overflow, so we deal with the case. For
966 * uniformity we deal with the case in the usec case too.
968 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
971 getmicrouptime(struct timeval *tvp)
973 struct globaldata *gd = mycpu;
977 tvp->tv_sec = gd->gd_time_seconds;
978 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
979 } while (tvp->tv_sec != gd->gd_time_seconds);
981 if (delta >= sys_cputimer->freq) {
982 tvp->tv_sec += delta / sys_cputimer->freq;
983 delta %= sys_cputimer->freq;
985 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
986 if (tvp->tv_usec >= 1000000) {
987 tvp->tv_usec -= 1000000;
993 getnanouptime(struct timespec *tsp)
995 struct globaldata *gd = mycpu;
999 tsp->tv_sec = gd->gd_time_seconds;
1000 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1001 } while (tsp->tv_sec != gd->gd_time_seconds);
1003 if (delta >= sys_cputimer->freq) {
1004 tsp->tv_sec += delta / sys_cputimer->freq;
1005 delta %= sys_cputimer->freq;
1007 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1011 microuptime(struct timeval *tvp)
1013 struct globaldata *gd = mycpu;
1017 tvp->tv_sec = gd->gd_time_seconds;
1018 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1019 } while (tvp->tv_sec != gd->gd_time_seconds);
1021 if (delta >= sys_cputimer->freq) {
1022 tvp->tv_sec += delta / sys_cputimer->freq;
1023 delta %= sys_cputimer->freq;
1025 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1029 nanouptime(struct timespec *tsp)
1031 struct globaldata *gd = mycpu;
1035 tsp->tv_sec = gd->gd_time_seconds;
1036 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1037 } while (tsp->tv_sec != gd->gd_time_seconds);
1039 if (delta >= sys_cputimer->freq) {
1040 tsp->tv_sec += delta / sys_cputimer->freq;
1041 delta %= sys_cputimer->freq;
1043 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1050 getmicrotime(struct timeval *tvp)
1052 struct globaldata *gd = mycpu;
1053 struct timespec *bt;
1057 tvp->tv_sec = gd->gd_time_seconds;
1058 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1059 } while (tvp->tv_sec != gd->gd_time_seconds);
1061 if (delta >= sys_cputimer->freq) {
1062 tvp->tv_sec += delta / sys_cputimer->freq;
1063 delta %= sys_cputimer->freq;
1065 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1067 bt = &basetime[basetime_index];
1068 tvp->tv_sec += bt->tv_sec;
1069 tvp->tv_usec += bt->tv_nsec / 1000;
1070 while (tvp->tv_usec >= 1000000) {
1071 tvp->tv_usec -= 1000000;
1077 getnanotime(struct timespec *tsp)
1079 struct globaldata *gd = mycpu;
1080 struct timespec *bt;
1084 tsp->tv_sec = gd->gd_time_seconds;
1085 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1086 } while (tsp->tv_sec != gd->gd_time_seconds);
1088 if (delta >= sys_cputimer->freq) {
1089 tsp->tv_sec += delta / sys_cputimer->freq;
1090 delta %= sys_cputimer->freq;
1092 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1094 bt = &basetime[basetime_index];
1095 tsp->tv_sec += bt->tv_sec;
1096 tsp->tv_nsec += bt->tv_nsec;
1097 while (tsp->tv_nsec >= 1000000000) {
1098 tsp->tv_nsec -= 1000000000;
1104 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1106 struct globaldata *gd = mycpu;
1110 tsp->tv_sec = gd->gd_time_seconds;
1111 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1112 } while (tsp->tv_sec != gd->gd_time_seconds);
1114 if (delta >= sys_cputimer->freq) {
1115 tsp->tv_sec += delta / sys_cputimer->freq;
1116 delta %= sys_cputimer->freq;
1118 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1120 tsp->tv_sec += nbt->tv_sec;
1121 tsp->tv_nsec += nbt->tv_nsec;
1122 while (tsp->tv_nsec >= 1000000000) {
1123 tsp->tv_nsec -= 1000000000;
1130 microtime(struct timeval *tvp)
1132 struct globaldata *gd = mycpu;
1133 struct timespec *bt;
1137 tvp->tv_sec = gd->gd_time_seconds;
1138 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1139 } while (tvp->tv_sec != gd->gd_time_seconds);
1141 if (delta >= sys_cputimer->freq) {
1142 tvp->tv_sec += delta / sys_cputimer->freq;
1143 delta %= sys_cputimer->freq;
1145 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1147 bt = &basetime[basetime_index];
1148 tvp->tv_sec += bt->tv_sec;
1149 tvp->tv_usec += bt->tv_nsec / 1000;
1150 while (tvp->tv_usec >= 1000000) {
1151 tvp->tv_usec -= 1000000;
1157 nanotime(struct timespec *tsp)
1159 struct globaldata *gd = mycpu;
1160 struct timespec *bt;
1164 tsp->tv_sec = gd->gd_time_seconds;
1165 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1166 } while (tsp->tv_sec != gd->gd_time_seconds);
1168 if (delta >= sys_cputimer->freq) {
1169 tsp->tv_sec += delta / sys_cputimer->freq;
1170 delta %= sys_cputimer->freq;
1172 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1174 bt = &basetime[basetime_index];
1175 tsp->tv_sec += bt->tv_sec;
1176 tsp->tv_nsec += bt->tv_nsec;
1177 while (tsp->tv_nsec >= 1000000000) {
1178 tsp->tv_nsec -= 1000000000;
1184 * note: this is not exactly synchronized with real time. To do that we
1185 * would have to do what microtime does and check for a nanoseconds overflow.
1188 get_approximate_time_t(void)
1190 struct globaldata *gd = mycpu;
1191 struct timespec *bt;
1193 bt = &basetime[basetime_index];
1194 return(gd->gd_time_seconds + bt->tv_sec);
1198 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1201 struct pps_fetch_args *fapi;
1203 struct pps_kcbind_args *kapi;
1207 case PPS_IOC_CREATE:
1209 case PPS_IOC_DESTROY:
1211 case PPS_IOC_SETPARAMS:
1212 app = (pps_params_t *)data;
1213 if (app->mode & ~pps->ppscap)
1215 pps->ppsparam = *app;
1217 case PPS_IOC_GETPARAMS:
1218 app = (pps_params_t *)data;
1219 *app = pps->ppsparam;
1220 app->api_version = PPS_API_VERS_1;
1222 case PPS_IOC_GETCAP:
1223 *(int*)data = pps->ppscap;
1226 fapi = (struct pps_fetch_args *)data;
1227 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1229 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1230 return (EOPNOTSUPP);
1231 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1232 fapi->pps_info_buf = pps->ppsinfo;
1234 case PPS_IOC_KCBIND:
1236 kapi = (struct pps_kcbind_args *)data;
1237 /* XXX Only root should be able to do this */
1238 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1240 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1242 if (kapi->edge & ~pps->ppscap)
1244 pps->kcmode = kapi->edge;
1247 return (EOPNOTSUPP);
1255 pps_init(struct pps_state *pps)
1257 pps->ppscap |= PPS_TSFMT_TSPEC;
1258 if (pps->ppscap & PPS_CAPTUREASSERT)
1259 pps->ppscap |= PPS_OFFSETASSERT;
1260 if (pps->ppscap & PPS_CAPTURECLEAR)
1261 pps->ppscap |= PPS_OFFSETCLEAR;
1265 pps_event(struct pps_state *pps, sysclock_t count, int event)
1267 struct globaldata *gd;
1268 struct timespec *tsp;
1269 struct timespec *osp;
1270 struct timespec *bt;
1283 /* Things would be easier with arrays... */
1284 if (event == PPS_CAPTUREASSERT) {
1285 tsp = &pps->ppsinfo.assert_timestamp;
1286 osp = &pps->ppsparam.assert_offset;
1287 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1288 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1289 pcount = &pps->ppscount[0];
1290 pseq = &pps->ppsinfo.assert_sequence;
1292 tsp = &pps->ppsinfo.clear_timestamp;
1293 osp = &pps->ppsparam.clear_offset;
1294 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1295 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1296 pcount = &pps->ppscount[1];
1297 pseq = &pps->ppsinfo.clear_sequence;
1300 /* Nothing really happened */
1301 if (*pcount == count)
1307 ts.tv_sec = gd->gd_time_seconds;
1308 delta = count - gd->gd_cpuclock_base;
1309 } while (ts.tv_sec != gd->gd_time_seconds);
1311 if (delta >= sys_cputimer->freq) {
1312 ts.tv_sec += delta / sys_cputimer->freq;
1313 delta %= sys_cputimer->freq;
1315 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1316 bt = &basetime[basetime_index];
1317 ts.tv_sec += bt->tv_sec;
1318 ts.tv_nsec += bt->tv_nsec;
1319 while (ts.tv_nsec >= 1000000000) {
1320 ts.tv_nsec -= 1000000000;
1328 timespecadd(tsp, osp);
1329 if (tsp->tv_nsec < 0) {
1330 tsp->tv_nsec += 1000000000;
1336 /* magic, at its best... */
1337 tcount = count - pps->ppscount[2];
1338 pps->ppscount[2] = count;
1339 if (tcount >= sys_cputimer->freq) {
1340 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1341 sys_cputimer->freq64_nsec *
1342 (tcount % sys_cputimer->freq)) >> 32;
1344 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1346 hardpps(tsp, delta);
1352 * Return the tsc target value for a delay of (ns).
1354 * Returns -1 if the TSC is not supported.
1357 tsc_get_target(int ns)
1359 #if defined(_RDTSC_SUPPORTED_)
1360 if (cpu_feature & CPUID_TSC) {
1361 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1368 * Compare the tsc against the passed target
1370 * Returns +1 if the target has been reached
1371 * Returns 0 if the target has not yet been reached
1372 * Returns -1 if the TSC is not supported.
1374 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1377 tsc_test_target(int64_t target)
1379 #if defined(_RDTSC_SUPPORTED_)
1380 if (cpu_feature & CPUID_TSC) {
1381 if ((int64_t)(target - rdtsc()) <= 0)