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>
101 #include <machine/cpu.h>
102 #include <machine/limits.h>
103 #include <machine/smp.h>
104 #include <machine/cpufunc.h>
105 #include <machine/specialreg.h>
106 #include <machine/clock.h>
109 #include <sys/gmon.h>
112 #ifdef DEVICE_POLLING
113 extern void init_device_poll_pcpu(int);
117 extern void ifpoll_init_pcpu(int);
121 static void do_pctrack(struct intrframe *frame, int which);
124 static void initclocks (void *dummy);
125 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
128 * Some of these don't belong here, but it's easiest to concentrate them.
129 * Note that cpu_time counts in microseconds, but most userland programs
130 * just compare relative times against the total by delta.
132 struct kinfo_cputime cputime_percpu[MAXCPU];
134 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
135 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
140 sysctl_cputime(SYSCTL_HANDLER_ARGS)
143 size_t size = sizeof(struct kinfo_cputime);
145 for (cpu = 0; cpu < ncpus; ++cpu) {
146 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
152 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
153 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
155 SYSCTL_STRUCT(_kern, OID_AUTO, cputime, CTLFLAG_RD, &cpu_time, kinfo_cputime,
156 "CPU time statistics");
160 * boottime is used to calculate the 'real' uptime. Do not confuse this with
161 * microuptime(). microtime() is not drift compensated. The real uptime
162 * with compensation is nanotime() - bootime. boottime is recalculated
163 * whenever the real time is set based on the compensated elapsed time
164 * in seconds (gd->gd_time_seconds).
166 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
167 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
170 struct timespec boottime; /* boot time (realtime) for reference only */
171 time_t time_second; /* read-only 'passive' uptime in seconds */
174 * basetime is used to calculate the compensated real time of day. The
175 * basetime can be modified on a per-tick basis by the adjtime(),
176 * ntp_adjtime(), and sysctl-based time correction APIs.
178 * Note that frequency corrections can also be made by adjusting
181 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
182 * used on both SMP and UP systems to avoid MP races between cpu's and
183 * interrupt races on UP systems.
185 #define BASETIME_ARYSIZE 16
186 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
187 static struct timespec basetime[BASETIME_ARYSIZE];
188 static volatile int basetime_index;
191 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
198 * Because basetime data and index may be updated by another cpu,
199 * a load fence is required to ensure that the data we read has
200 * not been speculatively read relative to a possibly updated index.
202 index = basetime_index;
204 bt = &basetime[index];
205 error = SYSCTL_OUT(req, bt, sizeof(*bt));
209 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
210 &boottime, timespec, "System boottime");
211 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
212 sysctl_get_basetime, "S,timespec", "System basetime");
214 static void hardclock(systimer_t info, struct intrframe *frame);
215 static void statclock(systimer_t info, struct intrframe *frame);
216 static void schedclock(systimer_t info, struct intrframe *frame);
217 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
219 int ticks; /* system master ticks at hz */
220 int clocks_running; /* tsleep/timeout clocks operational */
221 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
222 int64_t nsec_acc; /* accumulator */
224 /* NTPD time correction fields */
225 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
226 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
227 int64_t ntp_delta; /* one-time correction in nsec */
228 int64_t ntp_big_delta = 1000000000;
229 int32_t ntp_tick_delta; /* current adjustment rate */
230 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
231 time_t ntp_leap_second; /* time of next leap second */
232 int ntp_leap_insert; /* whether to insert or remove a second */
235 * Finish initializing clock frequencies and start all clocks running.
239 initclocks(void *dummy)
241 /*psratio = profhz / stathz;*/
247 * Called on a per-cpu basis
250 initclocks_pcpu(void)
252 struct globaldata *gd = mycpu;
255 if (gd->gd_cpuid == 0) {
256 gd->gd_time_seconds = 1;
257 gd->gd_cpuclock_base = sys_cputimer->count();
260 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
261 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
264 systimer_intr_enable();
266 #ifdef DEVICE_POLLING
267 init_device_poll_pcpu(gd->gd_cpuid);
271 ifpoll_init_pcpu(gd->gd_cpuid);
275 * Use a non-queued periodic systimer to prevent multiple ticks from
276 * building up if the sysclock jumps forward (8254 gets reset). The
277 * sysclock will never jump backwards. Our time sync is based on
278 * the actual sysclock, not the ticks count.
280 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
281 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
282 /* XXX correct the frequency for scheduler / estcpu tests */
283 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
289 * This sets the current real time of day. Timespecs are in seconds and
290 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
291 * instead we adjust basetime so basetime + gd_* results in the current
292 * time of day. This way the gd_* fields are guarenteed to represent
293 * a monotonically increasing 'uptime' value.
295 * When set_timeofday() is called from userland, the system call forces it
296 * onto cpu #0 since only cpu #0 can update basetime_index.
299 set_timeofday(struct timespec *ts)
301 struct timespec *nbt;
305 * XXX SMP / non-atomic basetime updates
308 ni = (basetime_index + 1) & BASETIME_ARYMASK;
311 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
312 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
313 if (nbt->tv_nsec < 0) {
314 nbt->tv_nsec += 1000000000;
319 * Note that basetime diverges from boottime as the clock drift is
320 * compensated for, so we cannot do away with boottime. When setting
321 * the absolute time of day the drift is 0 (for an instant) and we
322 * can simply assign boottime to basetime.
324 * Note that nanouptime() is based on gd_time_seconds which is drift
325 * compensated up to a point (it is guarenteed to remain monotonically
326 * increasing). gd_time_seconds is thus our best uptime guess and
327 * suitable for use in the boottime calculation. It is already taken
328 * into account in the basetime calculation above.
330 boottime.tv_sec = nbt->tv_sec;
334 * We now have a new basetime, make sure all other cpus have it,
335 * then update the index.
344 * Each cpu has its own hardclock, but we only increments ticks and softticks
347 * NOTE! systimer! the MP lock might not be held here. We can only safely
348 * manipulate objects owned by the current cpu.
351 hardclock(systimer_t info, struct intrframe *frame)
355 struct globaldata *gd = mycpu;
358 * Realtime updates are per-cpu. Note that timer corrections as
359 * returned by microtime() and friends make an additional adjustment
360 * using a system-wise 'basetime', but the running time is always
361 * taken from the per-cpu globaldata area. Since the same clock
362 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
365 * Note that we never allow info->time (aka gd->gd_hardclock.time)
366 * to reverse index gd_cpuclock_base, but that it is possible for
367 * it to temporarily get behind in the seconds if something in the
368 * system locks interrupts for a long period of time. Since periodic
369 * timers count events, though everything should resynch again
372 cputicks = info->time - gd->gd_cpuclock_base;
373 if (cputicks >= sys_cputimer->freq) {
374 ++gd->gd_time_seconds;
375 gd->gd_cpuclock_base += sys_cputimer->freq;
379 * The system-wide ticks counter and NTP related timedelta/tickdelta
380 * adjustments only occur on cpu #0. NTP adjustments are accomplished
381 * by updating basetime.
383 if (gd->gd_cpuid == 0) {
384 struct timespec *nbt;
392 if (tco->tc_poll_pps)
393 tco->tc_poll_pps(tco);
397 * Calculate the new basetime index. We are in a critical section
398 * on cpu #0 and can safely play with basetime_index. Start
399 * with the current basetime and then make adjustments.
401 ni = (basetime_index + 1) & BASETIME_ARYMASK;
403 *nbt = basetime[basetime_index];
406 * Apply adjtime corrections. (adjtime() API)
408 * adjtime() only runs on cpu #0 so our critical section is
409 * sufficient to access these variables.
411 if (ntp_delta != 0) {
412 nbt->tv_nsec += ntp_tick_delta;
413 ntp_delta -= ntp_tick_delta;
414 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
415 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
416 ntp_tick_delta = ntp_delta;
421 * Apply permanent frequency corrections. (sysctl API)
423 if (ntp_tick_permanent != 0) {
424 ntp_tick_acc += ntp_tick_permanent;
425 if (ntp_tick_acc >= (1LL << 32)) {
426 nbt->tv_nsec += ntp_tick_acc >> 32;
427 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
428 } else if (ntp_tick_acc <= -(1LL << 32)) {
429 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
430 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
431 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
435 if (nbt->tv_nsec >= 1000000000) {
437 nbt->tv_nsec -= 1000000000;
438 } else if (nbt->tv_nsec < 0) {
440 nbt->tv_nsec += 1000000000;
444 * Another per-tick compensation. (for ntp_adjtime() API)
447 nsec_acc += nsec_adj;
448 if (nsec_acc >= 0x100000000LL) {
449 nbt->tv_nsec += nsec_acc >> 32;
450 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
451 } else if (nsec_acc <= -0x100000000LL) {
452 nbt->tv_nsec -= -nsec_acc >> 32;
453 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
455 if (nbt->tv_nsec >= 1000000000) {
456 nbt->tv_nsec -= 1000000000;
458 } else if (nbt->tv_nsec < 0) {
459 nbt->tv_nsec += 1000000000;
464 /************************************************************
465 * LEAP SECOND CORRECTION *
466 ************************************************************
468 * Taking into account all the corrections made above, figure
469 * out the new real time. If the seconds field has changed
470 * then apply any pending leap-second corrections.
472 getnanotime_nbt(nbt, &nts);
474 if (time_second != nts.tv_sec) {
476 * Apply leap second (sysctl API). Adjust nts for changes
477 * so we do not have to call getnanotime_nbt again.
479 if (ntp_leap_second) {
480 if (ntp_leap_second == nts.tv_sec) {
481 if (ntp_leap_insert) {
493 * Apply leap second (ntp_adjtime() API), calculate a new
494 * nsec_adj field. ntp_update_second() returns nsec_adj
495 * as a per-second value but we need it as a per-tick value.
497 leap = ntp_update_second(time_second, &nsec_adj);
503 * Update the time_second 'approximate time' global.
505 time_second = nts.tv_sec;
509 * Finally, our new basetime is ready to go live!
515 * Figure out how badly the system is starved for memory
517 vm_fault_ratecheck();
521 * lwkt thread scheduler fair queueing
523 lwkt_fairq_schedulerclock(curthread);
526 * softticks are handled for all cpus
528 hardclock_softtick(gd);
531 * The LWKT scheduler will generally allow the current process to
532 * return to user mode even if there are other runnable LWKT threads
533 * running in kernel mode on behalf of a user process. This will
534 * ensure that those other threads have an opportunity to run in
535 * fairly short order (but not instantly).
540 * ITimer handling is per-tick, per-cpu.
542 * We must acquire the per-process token in order for ksignal()
543 * to be non-blocking.
545 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
547 if (frame && CLKF_USERMODE(frame) &&
548 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
549 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0)
550 ksignal(p, SIGVTALRM);
551 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
552 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0)
555 lwkt_reltoken(&p->p_token);
561 * The statistics clock typically runs at a 125Hz rate, and is intended
562 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
564 * NOTE! systimer! the MP lock might not be held here. We can only safely
565 * manipulate objects owned by the current cpu.
567 * The stats clock is responsible for grabbing a profiling sample.
568 * Most of the statistics are only used by user-level statistics programs.
569 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
572 * Like the other clocks, the stat clock is called from what is effectively
573 * a fast interrupt, so the context should be the thread/process that got
577 statclock(systimer_t info, struct intrframe *frame)
590 * How big was our timeslice relative to the last time?
592 microuptime(&tv); /* mpsafe */
593 stv = &mycpu->gd_stattv;
594 if (stv->tv_sec == 0) {
597 bump = tv.tv_usec - stv->tv_usec +
598 (tv.tv_sec - stv->tv_sec) * 1000000;
609 if (frame && CLKF_USERMODE(frame)) {
611 * Came from userland, handle user time and deal with
614 if (p && (p->p_flag & P_PROFIL))
615 addupc_intr(p, CLKF_PC(frame), 1);
616 td->td_uticks += bump;
619 * Charge the time as appropriate
621 if (p && p->p_nice > NZERO)
622 cpu_time.cp_nice += bump;
624 cpu_time.cp_user += bump;
628 * Kernel statistics are just like addupc_intr, only easier.
631 if (g->state == GMON_PROF_ON && frame) {
632 i = CLKF_PC(frame) - g->lowpc;
633 if (i < g->textsize) {
634 i /= HISTFRACTION * sizeof(*g->kcount);
640 * Came from kernel mode, so we were:
641 * - handling an interrupt,
642 * - doing syscall or trap work on behalf of the current
644 * - spinning in the idle loop.
645 * Whichever it is, charge the time as appropriate.
646 * Note that we charge interrupts to the current process,
647 * regardless of whether they are ``for'' that process,
648 * so that we know how much of its real time was spent
649 * in ``non-process'' (i.e., interrupt) work.
651 * XXX assume system if frame is NULL. A NULL frame
652 * can occur if ipi processing is done from a crit_exit().
654 if (frame && CLKF_INTR(frame))
655 td->td_iticks += bump;
657 td->td_sticks += bump;
659 if (frame && CLKF_INTR(frame)) {
661 do_pctrack(frame, PCTRACK_INT);
663 cpu_time.cp_intr += bump;
665 if (td == &mycpu->gd_idlethread) {
666 cpu_time.cp_idle += bump;
670 do_pctrack(frame, PCTRACK_SYS);
672 cpu_time.cp_sys += bump;
680 * Sample the PC when in the kernel or in an interrupt. User code can
681 * retrieve the information and generate a histogram or other output.
685 do_pctrack(struct intrframe *frame, int which)
687 struct kinfo_pctrack *pctrack;
689 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
690 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
691 (void *)CLKF_PC(frame);
696 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
698 struct kinfo_pcheader head;
703 head.pc_ntrack = PCTRACK_SIZE;
704 head.pc_arysize = PCTRACK_ARYSIZE;
706 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
709 for (cpu = 0; cpu < ncpus; ++cpu) {
710 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
711 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
712 sizeof(struct kinfo_pctrack));
721 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
722 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
727 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
728 * the MP lock might not be held. We can safely manipulate parts of curproc
729 * but that's about it.
731 * Each cpu has its own scheduler clock.
734 schedclock(systimer_t info, struct intrframe *frame)
741 if ((lp = lwkt_preempted_proc()) != NULL) {
743 * Account for cpu time used and hit the scheduler. Note
744 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
748 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
751 if ((lp = curthread->td_lwp) != NULL) {
753 * Update resource usage integrals and maximums.
755 if ((ru = &lp->lwp_proc->p_ru) &&
756 (vm = lp->lwp_proc->p_vmspace) != NULL) {
757 ru->ru_ixrss += pgtok(vm->vm_tsize);
758 ru->ru_idrss += pgtok(vm->vm_dsize);
759 ru->ru_isrss += pgtok(vm->vm_ssize);
760 rss = pgtok(vmspace_resident_count(vm));
761 if (ru->ru_maxrss < rss)
768 * Compute number of ticks for the specified amount of time. The
769 * return value is intended to be used in a clock interrupt timed
770 * operation and guarenteed to meet or exceed the requested time.
771 * If the representation overflows, return INT_MAX. The minimum return
772 * value is 1 ticks and the function will average the calculation up.
773 * If any value greater then 0 microseconds is supplied, a value
774 * of at least 2 will be returned to ensure that a near-term clock
775 * interrupt does not cause the timeout to occur (degenerately) early.
777 * Note that limit checks must take into account microseconds, which is
778 * done simply by using the smaller signed long maximum instead of
779 * the unsigned long maximum.
781 * If ints have 32 bits, then the maximum value for any timeout in
782 * 10ms ticks is 248 days.
785 tvtohz_high(struct timeval *tv)
802 kprintf("tvtohz_high: negative time difference "
803 "%ld sec %ld usec\n",
807 } else if (sec <= INT_MAX / hz) {
808 ticks = (int)(sec * hz +
809 ((u_long)usec + (ustick - 1)) / ustick) + 1;
817 tstohz_high(struct timespec *ts)
834 kprintf("tstohz_high: negative time difference "
835 "%ld sec %ld nsec\n",
839 } else if (sec <= INT_MAX / hz) {
840 ticks = (int)(sec * hz +
841 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
850 * Compute number of ticks for the specified amount of time, erroring on
851 * the side of it being too low to ensure that sleeping the returned number
852 * of ticks will not result in a late return.
854 * The supplied timeval may not be negative and should be normalized. A
855 * return value of 0 is possible if the timeval converts to less then
858 * If ints have 32 bits, then the maximum value for any timeout in
859 * 10ms ticks is 248 days.
862 tvtohz_low(struct timeval *tv)
868 if (sec <= INT_MAX / hz)
869 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
876 tstohz_low(struct timespec *ts)
882 if (sec <= INT_MAX / hz)
883 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
890 * Start profiling on a process.
892 * Kernel profiling passes proc0 which never exits and hence
893 * keeps the profile clock running constantly.
896 startprofclock(struct proc *p)
898 if ((p->p_flag & P_PROFIL) == 0) {
899 p->p_flag |= P_PROFIL;
901 if (++profprocs == 1 && stathz != 0) {
904 setstatclockrate(profhz);
912 * Stop profiling on a process.
915 stopprofclock(struct proc *p)
917 if (p->p_flag & P_PROFIL) {
918 p->p_flag &= ~P_PROFIL;
920 if (--profprocs == 0 && stathz != 0) {
923 setstatclockrate(stathz);
931 * Return information about system clocks.
934 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
936 struct kinfo_clockinfo clkinfo;
938 * Construct clockinfo structure.
941 clkinfo.ci_tick = ustick;
942 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
943 clkinfo.ci_profhz = profhz;
944 clkinfo.ci_stathz = stathz ? stathz : hz;
945 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
948 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
949 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
952 * We have eight functions for looking at the clock, four for
953 * microseconds and four for nanoseconds. For each there is fast
954 * but less precise version "get{nano|micro}[up]time" which will
955 * return a time which is up to 1/HZ previous to the call, whereas
956 * the raw version "{nano|micro}[up]time" will return a timestamp
957 * which is as precise as possible. The "up" variants return the
958 * time relative to system boot, these are well suited for time
959 * interval measurements.
961 * Each cpu independantly maintains the current time of day, so all
962 * we need to do to protect ourselves from changes is to do a loop
963 * check on the seconds field changing out from under us.
965 * The system timer maintains a 32 bit count and due to various issues
966 * it is possible for the calculated delta to occassionally exceed
967 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
968 * multiplication can easily overflow, so we deal with the case. For
969 * uniformity we deal with the case in the usec case too.
971 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
974 getmicrouptime(struct timeval *tvp)
976 struct globaldata *gd = mycpu;
980 tvp->tv_sec = gd->gd_time_seconds;
981 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
982 } while (tvp->tv_sec != gd->gd_time_seconds);
984 if (delta >= sys_cputimer->freq) {
985 tvp->tv_sec += delta / sys_cputimer->freq;
986 delta %= sys_cputimer->freq;
988 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
989 if (tvp->tv_usec >= 1000000) {
990 tvp->tv_usec -= 1000000;
996 getnanouptime(struct timespec *tsp)
998 struct globaldata *gd = mycpu;
1002 tsp->tv_sec = gd->gd_time_seconds;
1003 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1004 } while (tsp->tv_sec != gd->gd_time_seconds);
1006 if (delta >= sys_cputimer->freq) {
1007 tsp->tv_sec += delta / sys_cputimer->freq;
1008 delta %= sys_cputimer->freq;
1010 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1014 microuptime(struct timeval *tvp)
1016 struct globaldata *gd = mycpu;
1020 tvp->tv_sec = gd->gd_time_seconds;
1021 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1022 } while (tvp->tv_sec != gd->gd_time_seconds);
1024 if (delta >= sys_cputimer->freq) {
1025 tvp->tv_sec += delta / sys_cputimer->freq;
1026 delta %= sys_cputimer->freq;
1028 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1032 nanouptime(struct timespec *tsp)
1034 struct globaldata *gd = mycpu;
1038 tsp->tv_sec = gd->gd_time_seconds;
1039 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1040 } while (tsp->tv_sec != gd->gd_time_seconds);
1042 if (delta >= sys_cputimer->freq) {
1043 tsp->tv_sec += delta / sys_cputimer->freq;
1044 delta %= sys_cputimer->freq;
1046 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1053 getmicrotime(struct timeval *tvp)
1055 struct globaldata *gd = mycpu;
1056 struct timespec *bt;
1060 tvp->tv_sec = gd->gd_time_seconds;
1061 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1062 } while (tvp->tv_sec != gd->gd_time_seconds);
1064 if (delta >= sys_cputimer->freq) {
1065 tvp->tv_sec += delta / sys_cputimer->freq;
1066 delta %= sys_cputimer->freq;
1068 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1070 bt = &basetime[basetime_index];
1071 tvp->tv_sec += bt->tv_sec;
1072 tvp->tv_usec += bt->tv_nsec / 1000;
1073 while (tvp->tv_usec >= 1000000) {
1074 tvp->tv_usec -= 1000000;
1080 getnanotime(struct timespec *tsp)
1082 struct globaldata *gd = mycpu;
1083 struct timespec *bt;
1087 tsp->tv_sec = gd->gd_time_seconds;
1088 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1089 } while (tsp->tv_sec != gd->gd_time_seconds);
1091 if (delta >= sys_cputimer->freq) {
1092 tsp->tv_sec += delta / sys_cputimer->freq;
1093 delta %= sys_cputimer->freq;
1095 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1097 bt = &basetime[basetime_index];
1098 tsp->tv_sec += bt->tv_sec;
1099 tsp->tv_nsec += bt->tv_nsec;
1100 while (tsp->tv_nsec >= 1000000000) {
1101 tsp->tv_nsec -= 1000000000;
1107 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1109 struct globaldata *gd = mycpu;
1113 tsp->tv_sec = gd->gd_time_seconds;
1114 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1115 } while (tsp->tv_sec != gd->gd_time_seconds);
1117 if (delta >= sys_cputimer->freq) {
1118 tsp->tv_sec += delta / sys_cputimer->freq;
1119 delta %= sys_cputimer->freq;
1121 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1123 tsp->tv_sec += nbt->tv_sec;
1124 tsp->tv_nsec += nbt->tv_nsec;
1125 while (tsp->tv_nsec >= 1000000000) {
1126 tsp->tv_nsec -= 1000000000;
1133 microtime(struct timeval *tvp)
1135 struct globaldata *gd = mycpu;
1136 struct timespec *bt;
1140 tvp->tv_sec = gd->gd_time_seconds;
1141 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1142 } while (tvp->tv_sec != gd->gd_time_seconds);
1144 if (delta >= sys_cputimer->freq) {
1145 tvp->tv_sec += delta / sys_cputimer->freq;
1146 delta %= sys_cputimer->freq;
1148 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1150 bt = &basetime[basetime_index];
1151 tvp->tv_sec += bt->tv_sec;
1152 tvp->tv_usec += bt->tv_nsec / 1000;
1153 while (tvp->tv_usec >= 1000000) {
1154 tvp->tv_usec -= 1000000;
1160 nanotime(struct timespec *tsp)
1162 struct globaldata *gd = mycpu;
1163 struct timespec *bt;
1167 tsp->tv_sec = gd->gd_time_seconds;
1168 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1169 } while (tsp->tv_sec != gd->gd_time_seconds);
1171 if (delta >= sys_cputimer->freq) {
1172 tsp->tv_sec += delta / sys_cputimer->freq;
1173 delta %= sys_cputimer->freq;
1175 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1177 bt = &basetime[basetime_index];
1178 tsp->tv_sec += bt->tv_sec;
1179 tsp->tv_nsec += bt->tv_nsec;
1180 while (tsp->tv_nsec >= 1000000000) {
1181 tsp->tv_nsec -= 1000000000;
1187 * note: this is not exactly synchronized with real time. To do that we
1188 * would have to do what microtime does and check for a nanoseconds overflow.
1191 get_approximate_time_t(void)
1193 struct globaldata *gd = mycpu;
1194 struct timespec *bt;
1196 bt = &basetime[basetime_index];
1197 return(gd->gd_time_seconds + bt->tv_sec);
1201 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1204 struct pps_fetch_args *fapi;
1206 struct pps_kcbind_args *kapi;
1210 case PPS_IOC_CREATE:
1212 case PPS_IOC_DESTROY:
1214 case PPS_IOC_SETPARAMS:
1215 app = (pps_params_t *)data;
1216 if (app->mode & ~pps->ppscap)
1218 pps->ppsparam = *app;
1220 case PPS_IOC_GETPARAMS:
1221 app = (pps_params_t *)data;
1222 *app = pps->ppsparam;
1223 app->api_version = PPS_API_VERS_1;
1225 case PPS_IOC_GETCAP:
1226 *(int*)data = pps->ppscap;
1229 fapi = (struct pps_fetch_args *)data;
1230 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1232 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1233 return (EOPNOTSUPP);
1234 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1235 fapi->pps_info_buf = pps->ppsinfo;
1237 case PPS_IOC_KCBIND:
1239 kapi = (struct pps_kcbind_args *)data;
1240 /* XXX Only root should be able to do this */
1241 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1243 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1245 if (kapi->edge & ~pps->ppscap)
1247 pps->kcmode = kapi->edge;
1250 return (EOPNOTSUPP);
1258 pps_init(struct pps_state *pps)
1260 pps->ppscap |= PPS_TSFMT_TSPEC;
1261 if (pps->ppscap & PPS_CAPTUREASSERT)
1262 pps->ppscap |= PPS_OFFSETASSERT;
1263 if (pps->ppscap & PPS_CAPTURECLEAR)
1264 pps->ppscap |= PPS_OFFSETCLEAR;
1268 pps_event(struct pps_state *pps, sysclock_t count, int event)
1270 struct globaldata *gd;
1271 struct timespec *tsp;
1272 struct timespec *osp;
1273 struct timespec *bt;
1286 /* Things would be easier with arrays... */
1287 if (event == PPS_CAPTUREASSERT) {
1288 tsp = &pps->ppsinfo.assert_timestamp;
1289 osp = &pps->ppsparam.assert_offset;
1290 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1291 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1292 pcount = &pps->ppscount[0];
1293 pseq = &pps->ppsinfo.assert_sequence;
1295 tsp = &pps->ppsinfo.clear_timestamp;
1296 osp = &pps->ppsparam.clear_offset;
1297 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1298 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1299 pcount = &pps->ppscount[1];
1300 pseq = &pps->ppsinfo.clear_sequence;
1303 /* Nothing really happened */
1304 if (*pcount == count)
1310 ts.tv_sec = gd->gd_time_seconds;
1311 delta = count - gd->gd_cpuclock_base;
1312 } while (ts.tv_sec != gd->gd_time_seconds);
1314 if (delta >= sys_cputimer->freq) {
1315 ts.tv_sec += delta / sys_cputimer->freq;
1316 delta %= sys_cputimer->freq;
1318 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1319 bt = &basetime[basetime_index];
1320 ts.tv_sec += bt->tv_sec;
1321 ts.tv_nsec += bt->tv_nsec;
1322 while (ts.tv_nsec >= 1000000000) {
1323 ts.tv_nsec -= 1000000000;
1331 timespecadd(tsp, osp);
1332 if (tsp->tv_nsec < 0) {
1333 tsp->tv_nsec += 1000000000;
1339 /* magic, at its best... */
1340 tcount = count - pps->ppscount[2];
1341 pps->ppscount[2] = count;
1342 if (tcount >= sys_cputimer->freq) {
1343 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1344 sys_cputimer->freq64_nsec *
1345 (tcount % sys_cputimer->freq)) >> 32;
1347 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1349 hardpps(tsp, delta);
1355 * Return the tsc target value for a delay of (ns).
1357 * Returns -1 if the TSC is not supported.
1360 tsc_get_target(int ns)
1362 #if defined(_RDTSC_SUPPORTED_)
1363 if (cpu_feature & CPUID_TSC) {
1364 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1371 * Compare the tsc against the passed target
1373 * Returns +1 if the target has been reached
1374 * Returns 0 if the target has not yet been reached
1375 * Returns -1 if the TSC is not supported.
1377 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1380 tsc_test_target(int64_t target)
1382 #if defined(_RDTSC_SUPPORTED_)
1383 if (cpu_feature & CPUID_TSC) {
1384 if ((int64_t)(target - rdtsc()) <= 0)