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 && lwkt_trytoken(&proc_token)) {
546 if (frame && CLKF_USERMODE(frame) &&
547 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
548 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0)
549 ksignal(p, SIGVTALRM);
550 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
551 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0)
554 lwkt_reltoken(&proc_token);
560 * The statistics clock typically runs at a 125Hz rate, and is intended
561 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
563 * NOTE! systimer! the MP lock might not be held here. We can only safely
564 * manipulate objects owned by the current cpu.
566 * The stats clock is responsible for grabbing a profiling sample.
567 * Most of the statistics are only used by user-level statistics programs.
568 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
571 * Like the other clocks, the stat clock is called from what is effectively
572 * a fast interrupt, so the context should be the thread/process that got
576 statclock(systimer_t info, struct intrframe *frame)
589 * How big was our timeslice relative to the last time?
591 microuptime(&tv); /* mpsafe */
592 stv = &mycpu->gd_stattv;
593 if (stv->tv_sec == 0) {
596 bump = tv.tv_usec - stv->tv_usec +
597 (tv.tv_sec - stv->tv_sec) * 1000000;
608 if (frame && CLKF_USERMODE(frame)) {
610 * Came from userland, handle user time and deal with
613 if (p && (p->p_flag & P_PROFIL))
614 addupc_intr(p, CLKF_PC(frame), 1);
615 td->td_uticks += bump;
618 * Charge the time as appropriate
620 if (p && p->p_nice > NZERO)
621 cpu_time.cp_nice += bump;
623 cpu_time.cp_user += bump;
627 * Kernel statistics are just like addupc_intr, only easier.
630 if (g->state == GMON_PROF_ON && frame) {
631 i = CLKF_PC(frame) - g->lowpc;
632 if (i < g->textsize) {
633 i /= HISTFRACTION * sizeof(*g->kcount);
639 * Came from kernel mode, so we were:
640 * - handling an interrupt,
641 * - doing syscall or trap work on behalf of the current
643 * - spinning in the idle loop.
644 * Whichever it is, charge the time as appropriate.
645 * Note that we charge interrupts to the current process,
646 * regardless of whether they are ``for'' that process,
647 * so that we know how much of its real time was spent
648 * in ``non-process'' (i.e., interrupt) work.
650 * XXX assume system if frame is NULL. A NULL frame
651 * can occur if ipi processing is done from a crit_exit().
653 if (frame && CLKF_INTR(frame))
654 td->td_iticks += bump;
656 td->td_sticks += bump;
658 if (frame && CLKF_INTR(frame)) {
660 do_pctrack(frame, PCTRACK_INT);
662 cpu_time.cp_intr += bump;
664 if (td == &mycpu->gd_idlethread) {
665 cpu_time.cp_idle += bump;
669 do_pctrack(frame, PCTRACK_SYS);
671 cpu_time.cp_sys += bump;
679 * Sample the PC when in the kernel or in an interrupt. User code can
680 * retrieve the information and generate a histogram or other output.
684 do_pctrack(struct intrframe *frame, int which)
686 struct kinfo_pctrack *pctrack;
688 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
689 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
690 (void *)CLKF_PC(frame);
695 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
697 struct kinfo_pcheader head;
702 head.pc_ntrack = PCTRACK_SIZE;
703 head.pc_arysize = PCTRACK_ARYSIZE;
705 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
708 for (cpu = 0; cpu < ncpus; ++cpu) {
709 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
710 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
711 sizeof(struct kinfo_pctrack));
720 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
721 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
726 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
727 * the MP lock might not be held. We can safely manipulate parts of curproc
728 * but that's about it.
730 * Each cpu has its own scheduler clock.
733 schedclock(systimer_t info, struct intrframe *frame)
740 if ((lp = lwkt_preempted_proc()) != NULL) {
742 * Account for cpu time used and hit the scheduler. Note
743 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
747 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
750 if ((lp = curthread->td_lwp) != NULL) {
752 * Update resource usage integrals and maximums.
754 if ((ru = &lp->lwp_proc->p_ru) &&
755 (vm = lp->lwp_proc->p_vmspace) != NULL) {
756 ru->ru_ixrss += pgtok(vm->vm_tsize);
757 ru->ru_idrss += pgtok(vm->vm_dsize);
758 ru->ru_isrss += pgtok(vm->vm_ssize);
759 rss = pgtok(vmspace_resident_count(vm));
760 if (ru->ru_maxrss < rss)
767 * Compute number of ticks for the specified amount of time. The
768 * return value is intended to be used in a clock interrupt timed
769 * operation and guarenteed to meet or exceed the requested time.
770 * If the representation overflows, return INT_MAX. The minimum return
771 * value is 1 ticks and the function will average the calculation up.
772 * If any value greater then 0 microseconds is supplied, a value
773 * of at least 2 will be returned to ensure that a near-term clock
774 * interrupt does not cause the timeout to occur (degenerately) early.
776 * Note that limit checks must take into account microseconds, which is
777 * done simply by using the smaller signed long maximum instead of
778 * the unsigned long maximum.
780 * If ints have 32 bits, then the maximum value for any timeout in
781 * 10ms ticks is 248 days.
784 tvtohz_high(struct timeval *tv)
801 kprintf("tvtohz_high: negative time difference "
802 "%ld sec %ld usec\n",
806 } else if (sec <= INT_MAX / hz) {
807 ticks = (int)(sec * hz +
808 ((u_long)usec + (ustick - 1)) / ustick) + 1;
816 tstohz_high(struct timespec *ts)
833 kprintf("tstohz_high: negative time difference "
834 "%ld sec %ld nsec\n",
838 } else if (sec <= INT_MAX / hz) {
839 ticks = (int)(sec * hz +
840 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
849 * Compute number of ticks for the specified amount of time, erroring on
850 * the side of it being too low to ensure that sleeping the returned number
851 * of ticks will not result in a late return.
853 * The supplied timeval may not be negative and should be normalized. A
854 * return value of 0 is possible if the timeval converts to less then
857 * If ints have 32 bits, then the maximum value for any timeout in
858 * 10ms ticks is 248 days.
861 tvtohz_low(struct timeval *tv)
867 if (sec <= INT_MAX / hz)
868 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
875 tstohz_low(struct timespec *ts)
881 if (sec <= INT_MAX / hz)
882 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
889 * Start profiling on a process.
891 * Kernel profiling passes proc0 which never exits and hence
892 * keeps the profile clock running constantly.
895 startprofclock(struct proc *p)
897 if ((p->p_flag & P_PROFIL) == 0) {
898 p->p_flag |= P_PROFIL;
900 if (++profprocs == 1 && stathz != 0) {
903 setstatclockrate(profhz);
911 * Stop profiling on a process.
914 stopprofclock(struct proc *p)
916 if (p->p_flag & P_PROFIL) {
917 p->p_flag &= ~P_PROFIL;
919 if (--profprocs == 0 && stathz != 0) {
922 setstatclockrate(stathz);
930 * Return information about system clocks.
933 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
935 struct kinfo_clockinfo clkinfo;
937 * Construct clockinfo structure.
940 clkinfo.ci_tick = ustick;
941 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
942 clkinfo.ci_profhz = profhz;
943 clkinfo.ci_stathz = stathz ? stathz : hz;
944 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
947 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
948 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
951 * We have eight functions for looking at the clock, four for
952 * microseconds and four for nanoseconds. For each there is fast
953 * but less precise version "get{nano|micro}[up]time" which will
954 * return a time which is up to 1/HZ previous to the call, whereas
955 * the raw version "{nano|micro}[up]time" will return a timestamp
956 * which is as precise as possible. The "up" variants return the
957 * time relative to system boot, these are well suited for time
958 * interval measurements.
960 * Each cpu independantly maintains the current time of day, so all
961 * we need to do to protect ourselves from changes is to do a loop
962 * check on the seconds field changing out from under us.
964 * The system timer maintains a 32 bit count and due to various issues
965 * it is possible for the calculated delta to occassionally exceed
966 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
967 * multiplication can easily overflow, so we deal with the case. For
968 * uniformity we deal with the case in the usec case too.
970 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
973 getmicrouptime(struct timeval *tvp)
975 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;
988 if (tvp->tv_usec >= 1000000) {
989 tvp->tv_usec -= 1000000;
995 getnanouptime(struct timespec *tsp)
997 struct globaldata *gd = mycpu;
1001 tsp->tv_sec = gd->gd_time_seconds;
1002 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1003 } while (tsp->tv_sec != gd->gd_time_seconds);
1005 if (delta >= sys_cputimer->freq) {
1006 tsp->tv_sec += delta / sys_cputimer->freq;
1007 delta %= sys_cputimer->freq;
1009 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1013 microuptime(struct timeval *tvp)
1015 struct globaldata *gd = mycpu;
1019 tvp->tv_sec = gd->gd_time_seconds;
1020 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1021 } while (tvp->tv_sec != gd->gd_time_seconds);
1023 if (delta >= sys_cputimer->freq) {
1024 tvp->tv_sec += delta / sys_cputimer->freq;
1025 delta %= sys_cputimer->freq;
1027 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1031 nanouptime(struct timespec *tsp)
1033 struct globaldata *gd = mycpu;
1037 tsp->tv_sec = gd->gd_time_seconds;
1038 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1039 } while (tsp->tv_sec != gd->gd_time_seconds);
1041 if (delta >= sys_cputimer->freq) {
1042 tsp->tv_sec += delta / sys_cputimer->freq;
1043 delta %= sys_cputimer->freq;
1045 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1052 getmicrotime(struct timeval *tvp)
1054 struct globaldata *gd = mycpu;
1055 struct timespec *bt;
1059 tvp->tv_sec = gd->gd_time_seconds;
1060 delta = gd->gd_hardclock.time - 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 getnanotime(struct timespec *tsp)
1081 struct globaldata *gd = mycpu;
1082 struct timespec *bt;
1086 tsp->tv_sec = gd->gd_time_seconds;
1087 delta = gd->gd_hardclock.time - 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 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1108 struct globaldata *gd = mycpu;
1112 tsp->tv_sec = gd->gd_time_seconds;
1113 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1114 } while (tsp->tv_sec != gd->gd_time_seconds);
1116 if (delta >= sys_cputimer->freq) {
1117 tsp->tv_sec += delta / sys_cputimer->freq;
1118 delta %= sys_cputimer->freq;
1120 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1122 tsp->tv_sec += nbt->tv_sec;
1123 tsp->tv_nsec += nbt->tv_nsec;
1124 while (tsp->tv_nsec >= 1000000000) {
1125 tsp->tv_nsec -= 1000000000;
1132 microtime(struct timeval *tvp)
1134 struct globaldata *gd = mycpu;
1135 struct timespec *bt;
1139 tvp->tv_sec = gd->gd_time_seconds;
1140 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1141 } while (tvp->tv_sec != gd->gd_time_seconds);
1143 if (delta >= sys_cputimer->freq) {
1144 tvp->tv_sec += delta / sys_cputimer->freq;
1145 delta %= sys_cputimer->freq;
1147 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1149 bt = &basetime[basetime_index];
1150 tvp->tv_sec += bt->tv_sec;
1151 tvp->tv_usec += bt->tv_nsec / 1000;
1152 while (tvp->tv_usec >= 1000000) {
1153 tvp->tv_usec -= 1000000;
1159 nanotime(struct timespec *tsp)
1161 struct globaldata *gd = mycpu;
1162 struct timespec *bt;
1166 tsp->tv_sec = gd->gd_time_seconds;
1167 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1168 } while (tsp->tv_sec != gd->gd_time_seconds);
1170 if (delta >= sys_cputimer->freq) {
1171 tsp->tv_sec += delta / sys_cputimer->freq;
1172 delta %= sys_cputimer->freq;
1174 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1176 bt = &basetime[basetime_index];
1177 tsp->tv_sec += bt->tv_sec;
1178 tsp->tv_nsec += bt->tv_nsec;
1179 while (tsp->tv_nsec >= 1000000000) {
1180 tsp->tv_nsec -= 1000000000;
1186 * note: this is not exactly synchronized with real time. To do that we
1187 * would have to do what microtime does and check for a nanoseconds overflow.
1190 get_approximate_time_t(void)
1192 struct globaldata *gd = mycpu;
1193 struct timespec *bt;
1195 bt = &basetime[basetime_index];
1196 return(gd->gd_time_seconds + bt->tv_sec);
1200 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1203 struct pps_fetch_args *fapi;
1205 struct pps_kcbind_args *kapi;
1209 case PPS_IOC_CREATE:
1211 case PPS_IOC_DESTROY:
1213 case PPS_IOC_SETPARAMS:
1214 app = (pps_params_t *)data;
1215 if (app->mode & ~pps->ppscap)
1217 pps->ppsparam = *app;
1219 case PPS_IOC_GETPARAMS:
1220 app = (pps_params_t *)data;
1221 *app = pps->ppsparam;
1222 app->api_version = PPS_API_VERS_1;
1224 case PPS_IOC_GETCAP:
1225 *(int*)data = pps->ppscap;
1228 fapi = (struct pps_fetch_args *)data;
1229 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1231 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1232 return (EOPNOTSUPP);
1233 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1234 fapi->pps_info_buf = pps->ppsinfo;
1236 case PPS_IOC_KCBIND:
1238 kapi = (struct pps_kcbind_args *)data;
1239 /* XXX Only root should be able to do this */
1240 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1242 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1244 if (kapi->edge & ~pps->ppscap)
1246 pps->kcmode = kapi->edge;
1249 return (EOPNOTSUPP);
1257 pps_init(struct pps_state *pps)
1259 pps->ppscap |= PPS_TSFMT_TSPEC;
1260 if (pps->ppscap & PPS_CAPTUREASSERT)
1261 pps->ppscap |= PPS_OFFSETASSERT;
1262 if (pps->ppscap & PPS_CAPTURECLEAR)
1263 pps->ppscap |= PPS_OFFSETCLEAR;
1267 pps_event(struct pps_state *pps, sysclock_t count, int event)
1269 struct globaldata *gd;
1270 struct timespec *tsp;
1271 struct timespec *osp;
1272 struct timespec *bt;
1285 /* Things would be easier with arrays... */
1286 if (event == PPS_CAPTUREASSERT) {
1287 tsp = &pps->ppsinfo.assert_timestamp;
1288 osp = &pps->ppsparam.assert_offset;
1289 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1290 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1291 pcount = &pps->ppscount[0];
1292 pseq = &pps->ppsinfo.assert_sequence;
1294 tsp = &pps->ppsinfo.clear_timestamp;
1295 osp = &pps->ppsparam.clear_offset;
1296 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1297 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1298 pcount = &pps->ppscount[1];
1299 pseq = &pps->ppsinfo.clear_sequence;
1302 /* Nothing really happened */
1303 if (*pcount == count)
1309 ts.tv_sec = gd->gd_time_seconds;
1310 delta = count - gd->gd_cpuclock_base;
1311 } while (ts.tv_sec != gd->gd_time_seconds);
1313 if (delta >= sys_cputimer->freq) {
1314 ts.tv_sec += delta / sys_cputimer->freq;
1315 delta %= sys_cputimer->freq;
1317 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1318 bt = &basetime[basetime_index];
1319 ts.tv_sec += bt->tv_sec;
1320 ts.tv_nsec += bt->tv_nsec;
1321 while (ts.tv_nsec >= 1000000000) {
1322 ts.tv_nsec -= 1000000000;
1330 timespecadd(tsp, osp);
1331 if (tsp->tv_nsec < 0) {
1332 tsp->tv_nsec += 1000000000;
1338 /* magic, at its best... */
1339 tcount = count - pps->ppscount[2];
1340 pps->ppscount[2] = count;
1341 if (tcount >= sys_cputimer->freq) {
1342 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1343 sys_cputimer->freq64_nsec *
1344 (tcount % sys_cputimer->freq)) >> 32;
1346 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1348 hardpps(tsp, delta);
1354 * Return the tsc target value for a delay of (ns).
1356 * Returns -1 if the TSC is not supported.
1359 tsc_get_target(int ns)
1361 #if defined(_RDTSC_SUPPORTED_)
1362 if (cpu_feature & CPUID_TSC) {
1363 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1370 * Compare the tsc against the passed target
1372 * Returns +1 if the target has been reached
1373 * Returns 0 if the target has not yet been reached
1374 * Returns -1 if the TSC is not supported.
1376 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1379 tsc_test_target(int64_t target)
1381 #if defined(_RDTSC_SUPPORTED_)
1382 if (cpu_feature & CPUID_TSC) {
1383 if ((int64_t)(target - rdtsc()) <= 0)