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/resource.h>
89 #include <sys/resourcevar.h>
90 #include <sys/signalvar.h>
91 #include <sys/timex.h>
92 #include <sys/timepps.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_extern.h>
98 #include <sys/sysctl.h>
100 #include <sys/thread2.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 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
163 long cpu_states[5] = {0};
165 size_t size = sizeof(cpu_states);
167 for (cpu = 0; cpu < ncpus; ++cpu) {
168 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
169 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
170 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
171 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
172 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
175 error = SYSCTL_OUT(req, cpu_states, size);
180 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
181 sysctl_cp_time, "LU", "CPU time statistics");
184 * boottime is used to calculate the 'real' uptime. Do not confuse this with
185 * microuptime(). microtime() is not drift compensated. The real uptime
186 * with compensation is nanotime() - bootime. boottime is recalculated
187 * whenever the real time is set based on the compensated elapsed time
188 * in seconds (gd->gd_time_seconds).
190 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
191 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
194 struct timespec boottime; /* boot time (realtime) for reference only */
195 time_t time_second; /* read-only 'passive' uptime in seconds */
198 * basetime is used to calculate the compensated real time of day. The
199 * basetime can be modified on a per-tick basis by the adjtime(),
200 * ntp_adjtime(), and sysctl-based time correction APIs.
202 * Note that frequency corrections can also be made by adjusting
205 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
206 * used on both SMP and UP systems to avoid MP races between cpu's and
207 * interrupt races on UP systems.
209 #define BASETIME_ARYSIZE 16
210 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
211 static struct timespec basetime[BASETIME_ARYSIZE];
212 static volatile int basetime_index;
215 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
222 * Because basetime data and index may be updated by another cpu,
223 * a load fence is required to ensure that the data we read has
224 * not been speculatively read relative to a possibly updated index.
226 index = basetime_index;
228 bt = &basetime[index];
229 error = SYSCTL_OUT(req, bt, sizeof(*bt));
233 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
234 &boottime, timespec, "System boottime");
235 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
236 sysctl_get_basetime, "S,timespec", "System basetime");
238 static void hardclock(systimer_t info, int, struct intrframe *frame);
239 static void statclock(systimer_t info, int, struct intrframe *frame);
240 static void schedclock(systimer_t info, int, struct intrframe *frame);
241 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
243 int ticks; /* system master ticks at hz */
244 int clocks_running; /* tsleep/timeout clocks operational */
245 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
246 int64_t nsec_acc; /* accumulator */
248 /* NTPD time correction fields */
249 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
250 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
251 int64_t ntp_delta; /* one-time correction in nsec */
252 int64_t ntp_big_delta = 1000000000;
253 int32_t ntp_tick_delta; /* current adjustment rate */
254 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
255 time_t ntp_leap_second; /* time of next leap second */
256 int ntp_leap_insert; /* whether to insert or remove a second */
259 * Finish initializing clock frequencies and start all clocks running.
263 initclocks(void *dummy)
265 /*psratio = profhz / stathz;*/
271 * Called on a per-cpu basis
274 initclocks_pcpu(void)
276 struct globaldata *gd = mycpu;
279 if (gd->gd_cpuid == 0) {
280 gd->gd_time_seconds = 1;
281 gd->gd_cpuclock_base = sys_cputimer->count();
284 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
285 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
288 systimer_intr_enable();
290 #ifdef DEVICE_POLLING
291 init_device_poll_pcpu(gd->gd_cpuid);
295 ifpoll_init_pcpu(gd->gd_cpuid);
299 * Use a non-queued periodic systimer to prevent multiple ticks from
300 * building up if the sysclock jumps forward (8254 gets reset). The
301 * sysclock will never jump backwards. Our time sync is based on
302 * the actual sysclock, not the ticks count.
304 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
305 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
306 /* XXX correct the frequency for scheduler / estcpu tests */
307 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
313 * This sets the current real time of day. Timespecs are in seconds and
314 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
315 * instead we adjust basetime so basetime + gd_* results in the current
316 * time of day. This way the gd_* fields are guarenteed to represent
317 * a monotonically increasing 'uptime' value.
319 * When set_timeofday() is called from userland, the system call forces it
320 * onto cpu #0 since only cpu #0 can update basetime_index.
323 set_timeofday(struct timespec *ts)
325 struct timespec *nbt;
329 * XXX SMP / non-atomic basetime updates
332 ni = (basetime_index + 1) & BASETIME_ARYMASK;
335 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
336 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
337 if (nbt->tv_nsec < 0) {
338 nbt->tv_nsec += 1000000000;
343 * Note that basetime diverges from boottime as the clock drift is
344 * compensated for, so we cannot do away with boottime. When setting
345 * the absolute time of day the drift is 0 (for an instant) and we
346 * can simply assign boottime to basetime.
348 * Note that nanouptime() is based on gd_time_seconds which is drift
349 * compensated up to a point (it is guarenteed to remain monotonically
350 * increasing). gd_time_seconds is thus our best uptime guess and
351 * suitable for use in the boottime calculation. It is already taken
352 * into account in the basetime calculation above.
354 boottime.tv_sec = nbt->tv_sec;
358 * We now have a new basetime, make sure all other cpus have it,
359 * then update the index.
368 * Each cpu has its own hardclock, but we only increments ticks and softticks
371 * NOTE! systimer! the MP lock might not be held here. We can only safely
372 * manipulate objects owned by the current cpu.
375 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
379 struct globaldata *gd = mycpu;
382 * Realtime updates are per-cpu. Note that timer corrections as
383 * returned by microtime() and friends make an additional adjustment
384 * using a system-wise 'basetime', but the running time is always
385 * taken from the per-cpu globaldata area. Since the same clock
386 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
389 * Note that we never allow info->time (aka gd->gd_hardclock.time)
390 * to reverse index gd_cpuclock_base, but that it is possible for
391 * it to temporarily get behind in the seconds if something in the
392 * system locks interrupts for a long period of time. Since periodic
393 * timers count events, though everything should resynch again
396 cputicks = info->time - gd->gd_cpuclock_base;
397 if (cputicks >= sys_cputimer->freq) {
398 ++gd->gd_time_seconds;
399 gd->gd_cpuclock_base += sys_cputimer->freq;
403 * The system-wide ticks counter and NTP related timedelta/tickdelta
404 * adjustments only occur on cpu #0. NTP adjustments are accomplished
405 * by updating basetime.
407 if (gd->gd_cpuid == 0) {
408 struct timespec *nbt;
416 if (tco->tc_poll_pps)
417 tco->tc_poll_pps(tco);
421 * Calculate the new basetime index. We are in a critical section
422 * on cpu #0 and can safely play with basetime_index. Start
423 * with the current basetime and then make adjustments.
425 ni = (basetime_index + 1) & BASETIME_ARYMASK;
427 *nbt = basetime[basetime_index];
430 * Apply adjtime corrections. (adjtime() API)
432 * adjtime() only runs on cpu #0 so our critical section is
433 * sufficient to access these variables.
435 if (ntp_delta != 0) {
436 nbt->tv_nsec += ntp_tick_delta;
437 ntp_delta -= ntp_tick_delta;
438 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
439 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
440 ntp_tick_delta = ntp_delta;
445 * Apply permanent frequency corrections. (sysctl API)
447 if (ntp_tick_permanent != 0) {
448 ntp_tick_acc += ntp_tick_permanent;
449 if (ntp_tick_acc >= (1LL << 32)) {
450 nbt->tv_nsec += ntp_tick_acc >> 32;
451 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
452 } else if (ntp_tick_acc <= -(1LL << 32)) {
453 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
454 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
455 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
459 if (nbt->tv_nsec >= 1000000000) {
461 nbt->tv_nsec -= 1000000000;
462 } else if (nbt->tv_nsec < 0) {
464 nbt->tv_nsec += 1000000000;
468 * Another per-tick compensation. (for ntp_adjtime() API)
471 nsec_acc += nsec_adj;
472 if (nsec_acc >= 0x100000000LL) {
473 nbt->tv_nsec += nsec_acc >> 32;
474 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
475 } else if (nsec_acc <= -0x100000000LL) {
476 nbt->tv_nsec -= -nsec_acc >> 32;
477 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
479 if (nbt->tv_nsec >= 1000000000) {
480 nbt->tv_nsec -= 1000000000;
482 } else if (nbt->tv_nsec < 0) {
483 nbt->tv_nsec += 1000000000;
488 /************************************************************
489 * LEAP SECOND CORRECTION *
490 ************************************************************
492 * Taking into account all the corrections made above, figure
493 * out the new real time. If the seconds field has changed
494 * then apply any pending leap-second corrections.
496 getnanotime_nbt(nbt, &nts);
498 if (time_second != nts.tv_sec) {
500 * Apply leap second (sysctl API). Adjust nts for changes
501 * so we do not have to call getnanotime_nbt again.
503 if (ntp_leap_second) {
504 if (ntp_leap_second == nts.tv_sec) {
505 if (ntp_leap_insert) {
517 * Apply leap second (ntp_adjtime() API), calculate a new
518 * nsec_adj field. ntp_update_second() returns nsec_adj
519 * as a per-second value but we need it as a per-tick value.
521 leap = ntp_update_second(time_second, &nsec_adj);
527 * Update the time_second 'approximate time' global.
529 time_second = nts.tv_sec;
533 * Finally, our new basetime is ready to go live!
540 * lwkt thread scheduler fair queueing
542 lwkt_schedulerclock(curthread);
545 * softticks are handled for all cpus
547 hardclock_softtick(gd);
550 * ITimer handling is per-tick, per-cpu.
552 * We must acquire the per-process token in order for ksignal()
553 * to be non-blocking. For the moment this requires an AST fault,
554 * the ksignal() cannot be safely issued from this hard interrupt.
556 * XXX Even the trytoken here isn't right, and itimer operation in
557 * a multi threaded environment is going to be weird at the
560 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
562 if (frame && CLKF_USERMODE(frame) &&
563 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
564 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
565 p->p_flags |= P_SIGVTALRM;
568 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
569 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
570 p->p_flags |= P_SIGPROF;
574 lwkt_reltoken(&p->p_token);
580 * The statistics clock typically runs at a 125Hz rate, and is intended
581 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
583 * NOTE! systimer! the MP lock might not be held here. We can only safely
584 * manipulate objects owned by the current cpu.
586 * The stats clock is responsible for grabbing a profiling sample.
587 * Most of the statistics are only used by user-level statistics programs.
588 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
591 * Like the other clocks, the stat clock is called from what is effectively
592 * a fast interrupt, so the context should be the thread/process that got
596 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
609 * How big was our timeslice relative to the last time?
611 microuptime(&tv); /* mpsafe */
612 stv = &mycpu->gd_stattv;
613 if (stv->tv_sec == 0) {
616 bump = tv.tv_usec - stv->tv_usec +
617 (tv.tv_sec - stv->tv_sec) * 1000000;
628 if (frame && CLKF_USERMODE(frame)) {
630 * Came from userland, handle user time and deal with
633 if (p && (p->p_flags & P_PROFIL))
634 addupc_intr(p, CLKF_PC(frame), 1);
635 td->td_uticks += bump;
638 * Charge the time as appropriate
640 if (p && p->p_nice > NZERO)
641 cpu_time.cp_nice += bump;
643 cpu_time.cp_user += bump;
645 int intr_nest = mycpu->gd_intr_nesting_level;
649 * IPI processing code will bump gd_intr_nesting_level
650 * up by one, which breaks following CLKF_INTR testing,
651 * so we substract it by one here.
657 * Kernel statistics are just like addupc_intr, only easier.
660 if (g->state == GMON_PROF_ON && frame) {
661 i = CLKF_PC(frame) - g->lowpc;
662 if (i < g->textsize) {
663 i /= HISTFRACTION * sizeof(*g->kcount);
669 * Came from kernel mode, so we were:
670 * - handling an interrupt,
671 * - doing syscall or trap work on behalf of the current
673 * - spinning in the idle loop.
674 * Whichever it is, charge the time as appropriate.
675 * Note that we charge interrupts to the current process,
676 * regardless of whether they are ``for'' that process,
677 * so that we know how much of its real time was spent
678 * in ``non-process'' (i.e., interrupt) work.
680 * XXX assume system if frame is NULL. A NULL frame
681 * can occur if ipi processing is done from a crit_exit().
683 if (frame && CLKF_INTR(intr_nest))
684 td->td_iticks += bump;
686 td->td_sticks += bump;
688 if (frame && CLKF_INTR(intr_nest)) {
690 do_pctrack(frame, PCTRACK_INT);
692 cpu_time.cp_intr += bump;
694 if (td == &mycpu->gd_idlethread) {
695 cpu_time.cp_idle += bump;
699 do_pctrack(frame, PCTRACK_SYS);
701 cpu_time.cp_sys += bump;
709 * Sample the PC when in the kernel or in an interrupt. User code can
710 * retrieve the information and generate a histogram or other output.
714 do_pctrack(struct intrframe *frame, int which)
716 struct kinfo_pctrack *pctrack;
718 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
719 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
720 (void *)CLKF_PC(frame);
725 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
727 struct kinfo_pcheader head;
732 head.pc_ntrack = PCTRACK_SIZE;
733 head.pc_arysize = PCTRACK_ARYSIZE;
735 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
738 for (cpu = 0; cpu < ncpus; ++cpu) {
739 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
740 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
741 sizeof(struct kinfo_pctrack));
750 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
751 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
756 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
757 * the MP lock might not be held. We can safely manipulate parts of curproc
758 * but that's about it.
760 * Each cpu has its own scheduler clock.
763 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
770 if ((lp = lwkt_preempted_proc()) != NULL) {
772 * Account for cpu time used and hit the scheduler. Note
773 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
777 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
780 if ((lp = curthread->td_lwp) != NULL) {
782 * Update resource usage integrals and maximums.
784 if ((ru = &lp->lwp_proc->p_ru) &&
785 (vm = lp->lwp_proc->p_vmspace) != NULL) {
786 ru->ru_ixrss += pgtok(vm->vm_tsize);
787 ru->ru_idrss += pgtok(vm->vm_dsize);
788 ru->ru_isrss += pgtok(vm->vm_ssize);
789 if (lwkt_trytoken(&vm->vm_map.token)) {
790 rss = pgtok(vmspace_resident_count(vm));
791 if (ru->ru_maxrss < rss)
793 lwkt_reltoken(&vm->vm_map.token);
800 * Compute number of ticks for the specified amount of time. The
801 * return value is intended to be used in a clock interrupt timed
802 * operation and guarenteed to meet or exceed the requested time.
803 * If the representation overflows, return INT_MAX. The minimum return
804 * value is 1 ticks and the function will average the calculation up.
805 * If any value greater then 0 microseconds is supplied, a value
806 * of at least 2 will be returned to ensure that a near-term clock
807 * interrupt does not cause the timeout to occur (degenerately) early.
809 * Note that limit checks must take into account microseconds, which is
810 * done simply by using the smaller signed long maximum instead of
811 * the unsigned long maximum.
813 * If ints have 32 bits, then the maximum value for any timeout in
814 * 10ms ticks is 248 days.
817 tvtohz_high(struct timeval *tv)
834 kprintf("tvtohz_high: negative time difference "
835 "%ld sec %ld usec\n",
839 } else if (sec <= INT_MAX / hz) {
840 ticks = (int)(sec * hz +
841 ((u_long)usec + (ustick - 1)) / ustick) + 1;
849 tstohz_high(struct timespec *ts)
866 kprintf("tstohz_high: negative time difference "
867 "%ld sec %ld nsec\n",
871 } else if (sec <= INT_MAX / hz) {
872 ticks = (int)(sec * hz +
873 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
882 * Compute number of ticks for the specified amount of time, erroring on
883 * the side of it being too low to ensure that sleeping the returned number
884 * of ticks will not result in a late return.
886 * The supplied timeval may not be negative and should be normalized. A
887 * return value of 0 is possible if the timeval converts to less then
890 * If ints have 32 bits, then the maximum value for any timeout in
891 * 10ms ticks is 248 days.
894 tvtohz_low(struct timeval *tv)
900 if (sec <= INT_MAX / hz)
901 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
908 tstohz_low(struct timespec *ts)
914 if (sec <= INT_MAX / hz)
915 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
922 * Start profiling on a process.
924 * Kernel profiling passes proc0 which never exits and hence
925 * keeps the profile clock running constantly.
928 startprofclock(struct proc *p)
930 if ((p->p_flags & P_PROFIL) == 0) {
931 p->p_flags |= P_PROFIL;
933 if (++profprocs == 1 && stathz != 0) {
936 setstatclockrate(profhz);
944 * Stop profiling on a process.
946 * caller must hold p->p_token
949 stopprofclock(struct proc *p)
951 if (p->p_flags & P_PROFIL) {
952 p->p_flags &= ~P_PROFIL;
954 if (--profprocs == 0 && stathz != 0) {
957 setstatclockrate(stathz);
965 * Return information about system clocks.
968 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
970 struct kinfo_clockinfo clkinfo;
972 * Construct clockinfo structure.
975 clkinfo.ci_tick = ustick;
976 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
977 clkinfo.ci_profhz = profhz;
978 clkinfo.ci_stathz = stathz ? stathz : hz;
979 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
982 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
983 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
986 * We have eight functions for looking at the clock, four for
987 * microseconds and four for nanoseconds. For each there is fast
988 * but less precise version "get{nano|micro}[up]time" which will
989 * return a time which is up to 1/HZ previous to the call, whereas
990 * the raw version "{nano|micro}[up]time" will return a timestamp
991 * which is as precise as possible. The "up" variants return the
992 * time relative to system boot, these are well suited for time
993 * interval measurements.
995 * Each cpu independantly maintains the current time of day, so all
996 * we need to do to protect ourselves from changes is to do a loop
997 * check on the seconds field changing out from under us.
999 * The system timer maintains a 32 bit count and due to various issues
1000 * it is possible for the calculated delta to occassionally exceed
1001 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1002 * multiplication can easily overflow, so we deal with the case. For
1003 * uniformity we deal with the case in the usec case too.
1005 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1008 getmicrouptime(struct timeval *tvp)
1010 struct globaldata *gd = mycpu;
1014 tvp->tv_sec = gd->gd_time_seconds;
1015 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1016 } while (tvp->tv_sec != gd->gd_time_seconds);
1018 if (delta >= sys_cputimer->freq) {
1019 tvp->tv_sec += delta / sys_cputimer->freq;
1020 delta %= sys_cputimer->freq;
1022 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1023 if (tvp->tv_usec >= 1000000) {
1024 tvp->tv_usec -= 1000000;
1030 getnanouptime(struct timespec *tsp)
1032 struct globaldata *gd = mycpu;
1036 tsp->tv_sec = gd->gd_time_seconds;
1037 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1038 } while (tsp->tv_sec != gd->gd_time_seconds);
1040 if (delta >= sys_cputimer->freq) {
1041 tsp->tv_sec += delta / sys_cputimer->freq;
1042 delta %= sys_cputimer->freq;
1044 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1048 microuptime(struct timeval *tvp)
1050 struct globaldata *gd = mycpu;
1054 tvp->tv_sec = gd->gd_time_seconds;
1055 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1056 } while (tvp->tv_sec != gd->gd_time_seconds);
1058 if (delta >= sys_cputimer->freq) {
1059 tvp->tv_sec += delta / sys_cputimer->freq;
1060 delta %= sys_cputimer->freq;
1062 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1066 nanouptime(struct timespec *tsp)
1068 struct globaldata *gd = mycpu;
1072 tsp->tv_sec = gd->gd_time_seconds;
1073 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1074 } while (tsp->tv_sec != gd->gd_time_seconds);
1076 if (delta >= sys_cputimer->freq) {
1077 tsp->tv_sec += delta / sys_cputimer->freq;
1078 delta %= sys_cputimer->freq;
1080 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1087 getmicrotime(struct timeval *tvp)
1089 struct globaldata *gd = mycpu;
1090 struct timespec *bt;
1094 tvp->tv_sec = gd->gd_time_seconds;
1095 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1096 } while (tvp->tv_sec != gd->gd_time_seconds);
1098 if (delta >= sys_cputimer->freq) {
1099 tvp->tv_sec += delta / sys_cputimer->freq;
1100 delta %= sys_cputimer->freq;
1102 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1104 bt = &basetime[basetime_index];
1105 tvp->tv_sec += bt->tv_sec;
1106 tvp->tv_usec += bt->tv_nsec / 1000;
1107 while (tvp->tv_usec >= 1000000) {
1108 tvp->tv_usec -= 1000000;
1114 getnanotime(struct timespec *tsp)
1116 struct globaldata *gd = mycpu;
1117 struct timespec *bt;
1121 tsp->tv_sec = gd->gd_time_seconds;
1122 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1123 } while (tsp->tv_sec != gd->gd_time_seconds);
1125 if (delta >= sys_cputimer->freq) {
1126 tsp->tv_sec += delta / sys_cputimer->freq;
1127 delta %= sys_cputimer->freq;
1129 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1131 bt = &basetime[basetime_index];
1132 tsp->tv_sec += bt->tv_sec;
1133 tsp->tv_nsec += bt->tv_nsec;
1134 while (tsp->tv_nsec >= 1000000000) {
1135 tsp->tv_nsec -= 1000000000;
1141 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1143 struct globaldata *gd = mycpu;
1147 tsp->tv_sec = gd->gd_time_seconds;
1148 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1149 } while (tsp->tv_sec != gd->gd_time_seconds);
1151 if (delta >= sys_cputimer->freq) {
1152 tsp->tv_sec += delta / sys_cputimer->freq;
1153 delta %= sys_cputimer->freq;
1155 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1157 tsp->tv_sec += nbt->tv_sec;
1158 tsp->tv_nsec += nbt->tv_nsec;
1159 while (tsp->tv_nsec >= 1000000000) {
1160 tsp->tv_nsec -= 1000000000;
1167 microtime(struct timeval *tvp)
1169 struct globaldata *gd = mycpu;
1170 struct timespec *bt;
1174 tvp->tv_sec = gd->gd_time_seconds;
1175 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1176 } while (tvp->tv_sec != gd->gd_time_seconds);
1178 if (delta >= sys_cputimer->freq) {
1179 tvp->tv_sec += delta / sys_cputimer->freq;
1180 delta %= sys_cputimer->freq;
1182 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1184 bt = &basetime[basetime_index];
1185 tvp->tv_sec += bt->tv_sec;
1186 tvp->tv_usec += bt->tv_nsec / 1000;
1187 while (tvp->tv_usec >= 1000000) {
1188 tvp->tv_usec -= 1000000;
1194 nanotime(struct timespec *tsp)
1196 struct globaldata *gd = mycpu;
1197 struct timespec *bt;
1201 tsp->tv_sec = gd->gd_time_seconds;
1202 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1203 } while (tsp->tv_sec != gd->gd_time_seconds);
1205 if (delta >= sys_cputimer->freq) {
1206 tsp->tv_sec += delta / sys_cputimer->freq;
1207 delta %= sys_cputimer->freq;
1209 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1211 bt = &basetime[basetime_index];
1212 tsp->tv_sec += bt->tv_sec;
1213 tsp->tv_nsec += bt->tv_nsec;
1214 while (tsp->tv_nsec >= 1000000000) {
1215 tsp->tv_nsec -= 1000000000;
1221 * note: this is not exactly synchronized with real time. To do that we
1222 * would have to do what microtime does and check for a nanoseconds overflow.
1225 get_approximate_time_t(void)
1227 struct globaldata *gd = mycpu;
1228 struct timespec *bt;
1230 bt = &basetime[basetime_index];
1231 return(gd->gd_time_seconds + bt->tv_sec);
1235 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1238 struct pps_fetch_args *fapi;
1240 struct pps_kcbind_args *kapi;
1244 case PPS_IOC_CREATE:
1246 case PPS_IOC_DESTROY:
1248 case PPS_IOC_SETPARAMS:
1249 app = (pps_params_t *)data;
1250 if (app->mode & ~pps->ppscap)
1252 pps->ppsparam = *app;
1254 case PPS_IOC_GETPARAMS:
1255 app = (pps_params_t *)data;
1256 *app = pps->ppsparam;
1257 app->api_version = PPS_API_VERS_1;
1259 case PPS_IOC_GETCAP:
1260 *(int*)data = pps->ppscap;
1263 fapi = (struct pps_fetch_args *)data;
1264 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1266 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1267 return (EOPNOTSUPP);
1268 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1269 fapi->pps_info_buf = pps->ppsinfo;
1271 case PPS_IOC_KCBIND:
1273 kapi = (struct pps_kcbind_args *)data;
1274 /* XXX Only root should be able to do this */
1275 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1277 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1279 if (kapi->edge & ~pps->ppscap)
1281 pps->kcmode = kapi->edge;
1284 return (EOPNOTSUPP);
1292 pps_init(struct pps_state *pps)
1294 pps->ppscap |= PPS_TSFMT_TSPEC;
1295 if (pps->ppscap & PPS_CAPTUREASSERT)
1296 pps->ppscap |= PPS_OFFSETASSERT;
1297 if (pps->ppscap & PPS_CAPTURECLEAR)
1298 pps->ppscap |= PPS_OFFSETCLEAR;
1302 pps_event(struct pps_state *pps, sysclock_t count, int event)
1304 struct globaldata *gd;
1305 struct timespec *tsp;
1306 struct timespec *osp;
1307 struct timespec *bt;
1320 /* Things would be easier with arrays... */
1321 if (event == PPS_CAPTUREASSERT) {
1322 tsp = &pps->ppsinfo.assert_timestamp;
1323 osp = &pps->ppsparam.assert_offset;
1324 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1325 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1326 pcount = &pps->ppscount[0];
1327 pseq = &pps->ppsinfo.assert_sequence;
1329 tsp = &pps->ppsinfo.clear_timestamp;
1330 osp = &pps->ppsparam.clear_offset;
1331 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1332 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1333 pcount = &pps->ppscount[1];
1334 pseq = &pps->ppsinfo.clear_sequence;
1337 /* Nothing really happened */
1338 if (*pcount == count)
1344 ts.tv_sec = gd->gd_time_seconds;
1345 delta = count - gd->gd_cpuclock_base;
1346 } while (ts.tv_sec != gd->gd_time_seconds);
1348 if (delta >= sys_cputimer->freq) {
1349 ts.tv_sec += delta / sys_cputimer->freq;
1350 delta %= sys_cputimer->freq;
1352 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1353 bt = &basetime[basetime_index];
1354 ts.tv_sec += bt->tv_sec;
1355 ts.tv_nsec += bt->tv_nsec;
1356 while (ts.tv_nsec >= 1000000000) {
1357 ts.tv_nsec -= 1000000000;
1365 timespecadd(tsp, osp);
1366 if (tsp->tv_nsec < 0) {
1367 tsp->tv_nsec += 1000000000;
1373 /* magic, at its best... */
1374 tcount = count - pps->ppscount[2];
1375 pps->ppscount[2] = count;
1376 if (tcount >= sys_cputimer->freq) {
1377 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1378 sys_cputimer->freq64_nsec *
1379 (tcount % sys_cputimer->freq)) >> 32;
1381 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1383 hardpps(tsp, delta);
1389 * Return the tsc target value for a delay of (ns).
1391 * Returns -1 if the TSC is not supported.
1394 tsc_get_target(int ns)
1396 #if defined(_RDTSC_SUPPORTED_)
1397 if (cpu_feature & CPUID_TSC) {
1398 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1405 * Compare the tsc against the passed target
1407 * Returns +1 if the target has been reached
1408 * Returns 0 if the target has not yet been reached
1409 * Returns -1 if the TSC is not supported.
1411 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1414 tsc_test_target(int64_t target)
1416 #if defined(_RDTSC_SUPPORTED_)
1417 if (cpu_feature & CPUID_TSC) {
1418 if ((int64_t)(target - rdtsc()) <= 0)
1427 * Delay the specified number of nanoseconds using the tsc. This function
1428 * returns immediately if the TSC is not supported. At least one cpu_pause()
1436 clk = tsc_get_target(ns);
1438 while (tsc_test_target(clk) == 0)