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 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
162 long cpu_states[5] = {0};
164 size_t size = sizeof(cpu_states);
166 for (cpu = 0; cpu < ncpus; ++cpu) {
167 cpu_states[0] += cputime_percpu[cpu].cp_user;
168 cpu_states[1] += cputime_percpu[cpu].cp_nice;
169 cpu_states[2] += cputime_percpu[cpu].cp_sys;
170 cpu_states[3] += cputime_percpu[cpu].cp_intr;
171 cpu_states[4] += cputime_percpu[cpu].cp_idle;
174 error = SYSCTL_OUT(req, cpu_states, size);
179 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
180 sysctl_cp_time, "LU", "CPU time statistics");
183 * boottime is used to calculate the 'real' uptime. Do not confuse this with
184 * microuptime(). microtime() is not drift compensated. The real uptime
185 * with compensation is nanotime() - bootime. boottime is recalculated
186 * whenever the real time is set based on the compensated elapsed time
187 * in seconds (gd->gd_time_seconds).
189 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
190 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
193 struct timespec boottime; /* boot time (realtime) for reference only */
194 time_t time_second; /* read-only 'passive' uptime in seconds */
197 * basetime is used to calculate the compensated real time of day. The
198 * basetime can be modified on a per-tick basis by the adjtime(),
199 * ntp_adjtime(), and sysctl-based time correction APIs.
201 * Note that frequency corrections can also be made by adjusting
204 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
205 * used on both SMP and UP systems to avoid MP races between cpu's and
206 * interrupt races on UP systems.
208 #define BASETIME_ARYSIZE 16
209 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
210 static struct timespec basetime[BASETIME_ARYSIZE];
211 static volatile int basetime_index;
214 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
221 * Because basetime data and index may be updated by another cpu,
222 * a load fence is required to ensure that the data we read has
223 * not been speculatively read relative to a possibly updated index.
225 index = basetime_index;
227 bt = &basetime[index];
228 error = SYSCTL_OUT(req, bt, sizeof(*bt));
232 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
233 &boottime, timespec, "System boottime");
234 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
235 sysctl_get_basetime, "S,timespec", "System basetime");
237 static void hardclock(systimer_t info, int, struct intrframe *frame);
238 static void statclock(systimer_t info, int, struct intrframe *frame);
239 static void schedclock(systimer_t info, int, struct intrframe *frame);
240 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
242 int ticks; /* system master ticks at hz */
243 int clocks_running; /* tsleep/timeout clocks operational */
244 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
245 int64_t nsec_acc; /* accumulator */
247 /* NTPD time correction fields */
248 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
249 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
250 int64_t ntp_delta; /* one-time correction in nsec */
251 int64_t ntp_big_delta = 1000000000;
252 int32_t ntp_tick_delta; /* current adjustment rate */
253 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
254 time_t ntp_leap_second; /* time of next leap second */
255 int ntp_leap_insert; /* whether to insert or remove a second */
258 * Finish initializing clock frequencies and start all clocks running.
262 initclocks(void *dummy)
264 /*psratio = profhz / stathz;*/
270 * Called on a per-cpu basis
273 initclocks_pcpu(void)
275 struct globaldata *gd = mycpu;
278 if (gd->gd_cpuid == 0) {
279 gd->gd_time_seconds = 1;
280 gd->gd_cpuclock_base = sys_cputimer->count();
283 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
284 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
287 systimer_intr_enable();
289 #ifdef DEVICE_POLLING
290 init_device_poll_pcpu(gd->gd_cpuid);
294 ifpoll_init_pcpu(gd->gd_cpuid);
298 * Use a non-queued periodic systimer to prevent multiple ticks from
299 * building up if the sysclock jumps forward (8254 gets reset). The
300 * sysclock will never jump backwards. Our time sync is based on
301 * the actual sysclock, not the ticks count.
303 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
304 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
305 /* XXX correct the frequency for scheduler / estcpu tests */
306 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
312 * This sets the current real time of day. Timespecs are in seconds and
313 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
314 * instead we adjust basetime so basetime + gd_* results in the current
315 * time of day. This way the gd_* fields are guarenteed to represent
316 * a monotonically increasing 'uptime' value.
318 * When set_timeofday() is called from userland, the system call forces it
319 * onto cpu #0 since only cpu #0 can update basetime_index.
322 set_timeofday(struct timespec *ts)
324 struct timespec *nbt;
328 * XXX SMP / non-atomic basetime updates
331 ni = (basetime_index + 1) & BASETIME_ARYMASK;
334 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
335 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
336 if (nbt->tv_nsec < 0) {
337 nbt->tv_nsec += 1000000000;
342 * Note that basetime diverges from boottime as the clock drift is
343 * compensated for, so we cannot do away with boottime. When setting
344 * the absolute time of day the drift is 0 (for an instant) and we
345 * can simply assign boottime to basetime.
347 * Note that nanouptime() is based on gd_time_seconds which is drift
348 * compensated up to a point (it is guarenteed to remain monotonically
349 * increasing). gd_time_seconds is thus our best uptime guess and
350 * suitable for use in the boottime calculation. It is already taken
351 * into account in the basetime calculation above.
353 boottime.tv_sec = nbt->tv_sec;
357 * We now have a new basetime, make sure all other cpus have it,
358 * then update the index.
367 * Each cpu has its own hardclock, but we only increments ticks and softticks
370 * NOTE! systimer! the MP lock might not be held here. We can only safely
371 * manipulate objects owned by the current cpu.
374 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
378 struct globaldata *gd = mycpu;
381 * Realtime updates are per-cpu. Note that timer corrections as
382 * returned by microtime() and friends make an additional adjustment
383 * using a system-wise 'basetime', but the running time is always
384 * taken from the per-cpu globaldata area. Since the same clock
385 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
388 * Note that we never allow info->time (aka gd->gd_hardclock.time)
389 * to reverse index gd_cpuclock_base, but that it is possible for
390 * it to temporarily get behind in the seconds if something in the
391 * system locks interrupts for a long period of time. Since periodic
392 * timers count events, though everything should resynch again
395 cputicks = info->time - gd->gd_cpuclock_base;
396 if (cputicks >= sys_cputimer->freq) {
397 ++gd->gd_time_seconds;
398 gd->gd_cpuclock_base += sys_cputimer->freq;
402 * The system-wide ticks counter and NTP related timedelta/tickdelta
403 * adjustments only occur on cpu #0. NTP adjustments are accomplished
404 * by updating basetime.
406 if (gd->gd_cpuid == 0) {
407 struct timespec *nbt;
415 if (tco->tc_poll_pps)
416 tco->tc_poll_pps(tco);
420 * Calculate the new basetime index. We are in a critical section
421 * on cpu #0 and can safely play with basetime_index. Start
422 * with the current basetime and then make adjustments.
424 ni = (basetime_index + 1) & BASETIME_ARYMASK;
426 *nbt = basetime[basetime_index];
429 * Apply adjtime corrections. (adjtime() API)
431 * adjtime() only runs on cpu #0 so our critical section is
432 * sufficient to access these variables.
434 if (ntp_delta != 0) {
435 nbt->tv_nsec += ntp_tick_delta;
436 ntp_delta -= ntp_tick_delta;
437 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
438 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
439 ntp_tick_delta = ntp_delta;
444 * Apply permanent frequency corrections. (sysctl API)
446 if (ntp_tick_permanent != 0) {
447 ntp_tick_acc += ntp_tick_permanent;
448 if (ntp_tick_acc >= (1LL << 32)) {
449 nbt->tv_nsec += ntp_tick_acc >> 32;
450 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
451 } else if (ntp_tick_acc <= -(1LL << 32)) {
452 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
453 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
454 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
458 if (nbt->tv_nsec >= 1000000000) {
460 nbt->tv_nsec -= 1000000000;
461 } else if (nbt->tv_nsec < 0) {
463 nbt->tv_nsec += 1000000000;
467 * Another per-tick compensation. (for ntp_adjtime() API)
470 nsec_acc += nsec_adj;
471 if (nsec_acc >= 0x100000000LL) {
472 nbt->tv_nsec += nsec_acc >> 32;
473 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
474 } else if (nsec_acc <= -0x100000000LL) {
475 nbt->tv_nsec -= -nsec_acc >> 32;
476 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
478 if (nbt->tv_nsec >= 1000000000) {
479 nbt->tv_nsec -= 1000000000;
481 } else if (nbt->tv_nsec < 0) {
482 nbt->tv_nsec += 1000000000;
487 /************************************************************
488 * LEAP SECOND CORRECTION *
489 ************************************************************
491 * Taking into account all the corrections made above, figure
492 * out the new real time. If the seconds field has changed
493 * then apply any pending leap-second corrections.
495 getnanotime_nbt(nbt, &nts);
497 if (time_second != nts.tv_sec) {
499 * Apply leap second (sysctl API). Adjust nts for changes
500 * so we do not have to call getnanotime_nbt again.
502 if (ntp_leap_second) {
503 if (ntp_leap_second == nts.tv_sec) {
504 if (ntp_leap_insert) {
516 * Apply leap second (ntp_adjtime() API), calculate a new
517 * nsec_adj field. ntp_update_second() returns nsec_adj
518 * as a per-second value but we need it as a per-tick value.
520 leap = ntp_update_second(time_second, &nsec_adj);
526 * Update the time_second 'approximate time' global.
528 time_second = nts.tv_sec;
532 * Finally, our new basetime is ready to go live!
538 * Figure out how badly the system is starved for memory
540 vm_fault_ratecheck();
544 * lwkt thread scheduler fair queueing
546 lwkt_fairq_schedulerclock(curthread);
549 * softticks are handled for all cpus
551 hardclock_softtick(gd);
554 * The LWKT scheduler will generally allow the current process to
555 * return to user mode even if there are other runnable LWKT threads
556 * running in kernel mode on behalf of a user process. This will
557 * ensure that those other threads have an opportunity to run in
558 * fairly short order (but not instantly).
563 * ITimer handling is per-tick, per-cpu.
565 * We must acquire the per-process token in order for ksignal()
566 * to be non-blocking.
568 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
570 if (frame && CLKF_USERMODE(frame) &&
571 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
572 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0)
573 ksignal(p, SIGVTALRM);
574 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
575 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0)
578 lwkt_reltoken(&p->p_token);
584 * The statistics clock typically runs at a 125Hz rate, and is intended
585 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
587 * NOTE! systimer! the MP lock might not be held here. We can only safely
588 * manipulate objects owned by the current cpu.
590 * The stats clock is responsible for grabbing a profiling sample.
591 * Most of the statistics are only used by user-level statistics programs.
592 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
595 * Like the other clocks, the stat clock is called from what is effectively
596 * a fast interrupt, so the context should be the thread/process that got
600 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
613 * How big was our timeslice relative to the last time?
615 microuptime(&tv); /* mpsafe */
616 stv = &mycpu->gd_stattv;
617 if (stv->tv_sec == 0) {
620 bump = tv.tv_usec - stv->tv_usec +
621 (tv.tv_sec - stv->tv_sec) * 1000000;
632 if (frame && CLKF_USERMODE(frame)) {
634 * Came from userland, handle user time and deal with
637 if (p && (p->p_flag & P_PROFIL))
638 addupc_intr(p, CLKF_PC(frame), 1);
639 td->td_uticks += bump;
642 * Charge the time as appropriate
644 if (p && p->p_nice > NZERO)
645 cpu_time.cp_nice += bump;
647 cpu_time.cp_user += bump;
649 int intr_nest = mycpu->gd_intr_nesting_level;
653 * IPI processing code will bump gd_intr_nesting_level
654 * up by one, which breaks following CLKF_INTR testing,
655 * so we substract it by one here.
661 * Kernel statistics are just like addupc_intr, only easier.
664 if (g->state == GMON_PROF_ON && frame) {
665 i = CLKF_PC(frame) - g->lowpc;
666 if (i < g->textsize) {
667 i /= HISTFRACTION * sizeof(*g->kcount);
673 * Came from kernel mode, so we were:
674 * - handling an interrupt,
675 * - doing syscall or trap work on behalf of the current
677 * - spinning in the idle loop.
678 * Whichever it is, charge the time as appropriate.
679 * Note that we charge interrupts to the current process,
680 * regardless of whether they are ``for'' that process,
681 * so that we know how much of its real time was spent
682 * in ``non-process'' (i.e., interrupt) work.
684 * XXX assume system if frame is NULL. A NULL frame
685 * can occur if ipi processing is done from a crit_exit().
687 if (frame && CLKF_INTR(intr_nest))
688 td->td_iticks += bump;
690 td->td_sticks += bump;
692 if (frame && CLKF_INTR(intr_nest)) {
694 do_pctrack(frame, PCTRACK_INT);
696 cpu_time.cp_intr += bump;
698 if (td == &mycpu->gd_idlethread) {
699 cpu_time.cp_idle += bump;
703 do_pctrack(frame, PCTRACK_SYS);
705 cpu_time.cp_sys += bump;
713 * Sample the PC when in the kernel or in an interrupt. User code can
714 * retrieve the information and generate a histogram or other output.
718 do_pctrack(struct intrframe *frame, int which)
720 struct kinfo_pctrack *pctrack;
722 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
723 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
724 (void *)CLKF_PC(frame);
729 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
731 struct kinfo_pcheader head;
736 head.pc_ntrack = PCTRACK_SIZE;
737 head.pc_arysize = PCTRACK_ARYSIZE;
739 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
742 for (cpu = 0; cpu < ncpus; ++cpu) {
743 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
744 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
745 sizeof(struct kinfo_pctrack));
754 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
755 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
760 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
761 * the MP lock might not be held. We can safely manipulate parts of curproc
762 * but that's about it.
764 * Each cpu has its own scheduler clock.
767 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
774 if ((lp = lwkt_preempted_proc()) != NULL) {
776 * Account for cpu time used and hit the scheduler. Note
777 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
781 lp->lwp_proc->p_usched->schedulerclock(lp, info->periodic,
784 if ((lp = curthread->td_lwp) != NULL) {
786 * Update resource usage integrals and maximums.
788 if ((ru = &lp->lwp_proc->p_ru) &&
789 (vm = lp->lwp_proc->p_vmspace) != NULL) {
790 ru->ru_ixrss += pgtok(vm->vm_tsize);
791 ru->ru_idrss += pgtok(vm->vm_dsize);
792 ru->ru_isrss += pgtok(vm->vm_ssize);
793 rss = pgtok(vmspace_resident_count(vm));
794 if (ru->ru_maxrss < rss)
801 * Compute number of ticks for the specified amount of time. The
802 * return value is intended to be used in a clock interrupt timed
803 * operation and guarenteed to meet or exceed the requested time.
804 * If the representation overflows, return INT_MAX. The minimum return
805 * value is 1 ticks and the function will average the calculation up.
806 * If any value greater then 0 microseconds is supplied, a value
807 * of at least 2 will be returned to ensure that a near-term clock
808 * interrupt does not cause the timeout to occur (degenerately) early.
810 * Note that limit checks must take into account microseconds, which is
811 * done simply by using the smaller signed long maximum instead of
812 * the unsigned long maximum.
814 * If ints have 32 bits, then the maximum value for any timeout in
815 * 10ms ticks is 248 days.
818 tvtohz_high(struct timeval *tv)
835 kprintf("tvtohz_high: negative time difference "
836 "%ld sec %ld usec\n",
840 } else if (sec <= INT_MAX / hz) {
841 ticks = (int)(sec * hz +
842 ((u_long)usec + (ustick - 1)) / ustick) + 1;
850 tstohz_high(struct timespec *ts)
867 kprintf("tstohz_high: negative time difference "
868 "%ld sec %ld nsec\n",
872 } else if (sec <= INT_MAX / hz) {
873 ticks = (int)(sec * hz +
874 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
883 * Compute number of ticks for the specified amount of time, erroring on
884 * the side of it being too low to ensure that sleeping the returned number
885 * of ticks will not result in a late return.
887 * The supplied timeval may not be negative and should be normalized. A
888 * return value of 0 is possible if the timeval converts to less then
891 * If ints have 32 bits, then the maximum value for any timeout in
892 * 10ms ticks is 248 days.
895 tvtohz_low(struct timeval *tv)
901 if (sec <= INT_MAX / hz)
902 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
909 tstohz_low(struct timespec *ts)
915 if (sec <= INT_MAX / hz)
916 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
923 * Start profiling on a process.
925 * Kernel profiling passes proc0 which never exits and hence
926 * keeps the profile clock running constantly.
929 startprofclock(struct proc *p)
931 if ((p->p_flag & P_PROFIL) == 0) {
932 p->p_flag |= P_PROFIL;
934 if (++profprocs == 1 && stathz != 0) {
937 setstatclockrate(profhz);
945 * Stop profiling on a process.
948 stopprofclock(struct proc *p)
950 if (p->p_flag & P_PROFIL) {
951 p->p_flag &= ~P_PROFIL;
953 if (--profprocs == 0 && stathz != 0) {
956 setstatclockrate(stathz);
964 * Return information about system clocks.
967 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
969 struct kinfo_clockinfo clkinfo;
971 * Construct clockinfo structure.
974 clkinfo.ci_tick = ustick;
975 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
976 clkinfo.ci_profhz = profhz;
977 clkinfo.ci_stathz = stathz ? stathz : hz;
978 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
981 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
982 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
985 * We have eight functions for looking at the clock, four for
986 * microseconds and four for nanoseconds. For each there is fast
987 * but less precise version "get{nano|micro}[up]time" which will
988 * return a time which is up to 1/HZ previous to the call, whereas
989 * the raw version "{nano|micro}[up]time" will return a timestamp
990 * which is as precise as possible. The "up" variants return the
991 * time relative to system boot, these are well suited for time
992 * interval measurements.
994 * Each cpu independantly maintains the current time of day, so all
995 * we need to do to protect ourselves from changes is to do a loop
996 * check on the seconds field changing out from under us.
998 * The system timer maintains a 32 bit count and due to various issues
999 * it is possible for the calculated delta to occassionally exceed
1000 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1001 * multiplication can easily overflow, so we deal with the case. For
1002 * uniformity we deal with the case in the usec case too.
1004 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1007 getmicrouptime(struct timeval *tvp)
1009 struct globaldata *gd = mycpu;
1013 tvp->tv_sec = gd->gd_time_seconds;
1014 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1015 } while (tvp->tv_sec != gd->gd_time_seconds);
1017 if (delta >= sys_cputimer->freq) {
1018 tvp->tv_sec += delta / sys_cputimer->freq;
1019 delta %= sys_cputimer->freq;
1021 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1022 if (tvp->tv_usec >= 1000000) {
1023 tvp->tv_usec -= 1000000;
1029 getnanouptime(struct timespec *tsp)
1031 struct globaldata *gd = mycpu;
1035 tsp->tv_sec = gd->gd_time_seconds;
1036 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1037 } while (tsp->tv_sec != gd->gd_time_seconds);
1039 if (delta >= sys_cputimer->freq) {
1040 tsp->tv_sec += delta / sys_cputimer->freq;
1041 delta %= sys_cputimer->freq;
1043 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1047 microuptime(struct timeval *tvp)
1049 struct globaldata *gd = mycpu;
1053 tvp->tv_sec = gd->gd_time_seconds;
1054 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1055 } while (tvp->tv_sec != gd->gd_time_seconds);
1057 if (delta >= sys_cputimer->freq) {
1058 tvp->tv_sec += delta / sys_cputimer->freq;
1059 delta %= sys_cputimer->freq;
1061 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1065 nanouptime(struct timespec *tsp)
1067 struct globaldata *gd = mycpu;
1071 tsp->tv_sec = gd->gd_time_seconds;
1072 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1073 } while (tsp->tv_sec != gd->gd_time_seconds);
1075 if (delta >= sys_cputimer->freq) {
1076 tsp->tv_sec += delta / sys_cputimer->freq;
1077 delta %= sys_cputimer->freq;
1079 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1086 getmicrotime(struct timeval *tvp)
1088 struct globaldata *gd = mycpu;
1089 struct timespec *bt;
1093 tvp->tv_sec = gd->gd_time_seconds;
1094 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1095 } while (tvp->tv_sec != gd->gd_time_seconds);
1097 if (delta >= sys_cputimer->freq) {
1098 tvp->tv_sec += delta / sys_cputimer->freq;
1099 delta %= sys_cputimer->freq;
1101 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1103 bt = &basetime[basetime_index];
1104 tvp->tv_sec += bt->tv_sec;
1105 tvp->tv_usec += bt->tv_nsec / 1000;
1106 while (tvp->tv_usec >= 1000000) {
1107 tvp->tv_usec -= 1000000;
1113 getnanotime(struct timespec *tsp)
1115 struct globaldata *gd = mycpu;
1116 struct timespec *bt;
1120 tsp->tv_sec = gd->gd_time_seconds;
1121 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1122 } while (tsp->tv_sec != gd->gd_time_seconds);
1124 if (delta >= sys_cputimer->freq) {
1125 tsp->tv_sec += delta / sys_cputimer->freq;
1126 delta %= sys_cputimer->freq;
1128 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1130 bt = &basetime[basetime_index];
1131 tsp->tv_sec += bt->tv_sec;
1132 tsp->tv_nsec += bt->tv_nsec;
1133 while (tsp->tv_nsec >= 1000000000) {
1134 tsp->tv_nsec -= 1000000000;
1140 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1142 struct globaldata *gd = mycpu;
1146 tsp->tv_sec = gd->gd_time_seconds;
1147 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1148 } while (tsp->tv_sec != gd->gd_time_seconds);
1150 if (delta >= sys_cputimer->freq) {
1151 tsp->tv_sec += delta / sys_cputimer->freq;
1152 delta %= sys_cputimer->freq;
1154 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1156 tsp->tv_sec += nbt->tv_sec;
1157 tsp->tv_nsec += nbt->tv_nsec;
1158 while (tsp->tv_nsec >= 1000000000) {
1159 tsp->tv_nsec -= 1000000000;
1166 microtime(struct timeval *tvp)
1168 struct globaldata *gd = mycpu;
1169 struct timespec *bt;
1173 tvp->tv_sec = gd->gd_time_seconds;
1174 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1175 } while (tvp->tv_sec != gd->gd_time_seconds);
1177 if (delta >= sys_cputimer->freq) {
1178 tvp->tv_sec += delta / sys_cputimer->freq;
1179 delta %= sys_cputimer->freq;
1181 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1183 bt = &basetime[basetime_index];
1184 tvp->tv_sec += bt->tv_sec;
1185 tvp->tv_usec += bt->tv_nsec / 1000;
1186 while (tvp->tv_usec >= 1000000) {
1187 tvp->tv_usec -= 1000000;
1193 nanotime(struct timespec *tsp)
1195 struct globaldata *gd = mycpu;
1196 struct timespec *bt;
1200 tsp->tv_sec = gd->gd_time_seconds;
1201 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1202 } while (tsp->tv_sec != gd->gd_time_seconds);
1204 if (delta >= sys_cputimer->freq) {
1205 tsp->tv_sec += delta / sys_cputimer->freq;
1206 delta %= sys_cputimer->freq;
1208 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1210 bt = &basetime[basetime_index];
1211 tsp->tv_sec += bt->tv_sec;
1212 tsp->tv_nsec += bt->tv_nsec;
1213 while (tsp->tv_nsec >= 1000000000) {
1214 tsp->tv_nsec -= 1000000000;
1220 * note: this is not exactly synchronized with real time. To do that we
1221 * would have to do what microtime does and check for a nanoseconds overflow.
1224 get_approximate_time_t(void)
1226 struct globaldata *gd = mycpu;
1227 struct timespec *bt;
1229 bt = &basetime[basetime_index];
1230 return(gd->gd_time_seconds + bt->tv_sec);
1234 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1237 struct pps_fetch_args *fapi;
1239 struct pps_kcbind_args *kapi;
1243 case PPS_IOC_CREATE:
1245 case PPS_IOC_DESTROY:
1247 case PPS_IOC_SETPARAMS:
1248 app = (pps_params_t *)data;
1249 if (app->mode & ~pps->ppscap)
1251 pps->ppsparam = *app;
1253 case PPS_IOC_GETPARAMS:
1254 app = (pps_params_t *)data;
1255 *app = pps->ppsparam;
1256 app->api_version = PPS_API_VERS_1;
1258 case PPS_IOC_GETCAP:
1259 *(int*)data = pps->ppscap;
1262 fapi = (struct pps_fetch_args *)data;
1263 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1265 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1266 return (EOPNOTSUPP);
1267 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1268 fapi->pps_info_buf = pps->ppsinfo;
1270 case PPS_IOC_KCBIND:
1272 kapi = (struct pps_kcbind_args *)data;
1273 /* XXX Only root should be able to do this */
1274 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1276 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1278 if (kapi->edge & ~pps->ppscap)
1280 pps->kcmode = kapi->edge;
1283 return (EOPNOTSUPP);
1291 pps_init(struct pps_state *pps)
1293 pps->ppscap |= PPS_TSFMT_TSPEC;
1294 if (pps->ppscap & PPS_CAPTUREASSERT)
1295 pps->ppscap |= PPS_OFFSETASSERT;
1296 if (pps->ppscap & PPS_CAPTURECLEAR)
1297 pps->ppscap |= PPS_OFFSETCLEAR;
1301 pps_event(struct pps_state *pps, sysclock_t count, int event)
1303 struct globaldata *gd;
1304 struct timespec *tsp;
1305 struct timespec *osp;
1306 struct timespec *bt;
1319 /* Things would be easier with arrays... */
1320 if (event == PPS_CAPTUREASSERT) {
1321 tsp = &pps->ppsinfo.assert_timestamp;
1322 osp = &pps->ppsparam.assert_offset;
1323 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1324 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1325 pcount = &pps->ppscount[0];
1326 pseq = &pps->ppsinfo.assert_sequence;
1328 tsp = &pps->ppsinfo.clear_timestamp;
1329 osp = &pps->ppsparam.clear_offset;
1330 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1331 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1332 pcount = &pps->ppscount[1];
1333 pseq = &pps->ppsinfo.clear_sequence;
1336 /* Nothing really happened */
1337 if (*pcount == count)
1343 ts.tv_sec = gd->gd_time_seconds;
1344 delta = count - gd->gd_cpuclock_base;
1345 } while (ts.tv_sec != gd->gd_time_seconds);
1347 if (delta >= sys_cputimer->freq) {
1348 ts.tv_sec += delta / sys_cputimer->freq;
1349 delta %= sys_cputimer->freq;
1351 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1352 bt = &basetime[basetime_index];
1353 ts.tv_sec += bt->tv_sec;
1354 ts.tv_nsec += bt->tv_nsec;
1355 while (ts.tv_nsec >= 1000000000) {
1356 ts.tv_nsec -= 1000000000;
1364 timespecadd(tsp, osp);
1365 if (tsp->tv_nsec < 0) {
1366 tsp->tv_nsec += 1000000000;
1372 /* magic, at its best... */
1373 tcount = count - pps->ppscount[2];
1374 pps->ppscount[2] = count;
1375 if (tcount >= sys_cputimer->freq) {
1376 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1377 sys_cputimer->freq64_nsec *
1378 (tcount % sys_cputimer->freq)) >> 32;
1380 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1382 hardpps(tsp, delta);
1388 * Return the tsc target value for a delay of (ns).
1390 * Returns -1 if the TSC is not supported.
1393 tsc_get_target(int ns)
1395 #if defined(_RDTSC_SUPPORTED_)
1396 if (cpu_feature & CPUID_TSC) {
1397 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1404 * Compare the tsc against the passed target
1406 * Returns +1 if the target has been reached
1407 * Returns 0 if the target has not yet been reached
1408 * Returns -1 if the TSC is not supported.
1410 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1413 tsc_test_target(int64_t target)
1415 #if defined(_RDTSC_SUPPORTED_)
1416 if (cpu_feature & CPUID_TSC) {
1417 if ((int64_t)(target - rdtsc()) <= 0)