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 $
76 #include "opt_ifpoll.h"
77 #include "opt_pctrack.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/callout.h>
82 #include <sys/kernel.h>
83 #include <sys/kinfo.h>
85 #include <sys/malloc.h>
86 #include <sys/resource.h>
87 #include <sys/resourcevar.h>
88 #include <sys/signalvar.h>
89 #include <sys/timex.h>
90 #include <sys/timepps.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_extern.h>
96 #include <sys/sysctl.h>
98 #include <sys/thread2.h>
100 #include <machine/cpu.h>
101 #include <machine/limits.h>
102 #include <machine/smp.h>
103 #include <machine/cpufunc.h>
104 #include <machine/specialreg.h>
105 #include <machine/clock.h>
108 #include <sys/gmon.h>
112 extern void ifpoll_init_pcpu(int);
116 static void do_pctrack(struct intrframe *frame, int which);
119 static void initclocks (void *dummy);
120 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
123 * Some of these don't belong here, but it's easiest to concentrate them.
124 * Note that cpu_time counts in microseconds, but most userland programs
125 * just compare relative times against the total by delta.
127 struct kinfo_cputime cputime_percpu[MAXCPU];
129 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
130 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
134 sysctl_cputime(SYSCTL_HANDLER_ARGS)
137 size_t size = sizeof(struct kinfo_cputime);
139 for (cpu = 0; cpu < ncpus; ++cpu) {
140 if ((error = SYSCTL_OUT(req, &cputime_percpu[cpu], size)))
146 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
147 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
150 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
152 long cpu_states[5] = {0};
154 size_t size = sizeof(cpu_states);
156 for (cpu = 0; cpu < ncpus; ++cpu) {
157 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
158 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
159 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
160 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
161 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
164 error = SYSCTL_OUT(req, cpu_states, size);
169 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
170 sysctl_cp_time, "LU", "CPU time statistics");
173 * boottime is used to calculate the 'real' uptime. Do not confuse this with
174 * microuptime(). microtime() is not drift compensated. The real uptime
175 * with compensation is nanotime() - bootime. boottime is recalculated
176 * whenever the real time is set based on the compensated elapsed time
177 * in seconds (gd->gd_time_seconds).
179 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
180 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
183 struct timespec boottime; /* boot time (realtime) for reference only */
184 time_t time_second; /* read-only 'passive' uptime in seconds */
187 * basetime is used to calculate the compensated real time of day. The
188 * basetime can be modified on a per-tick basis by the adjtime(),
189 * ntp_adjtime(), and sysctl-based time correction APIs.
191 * Note that frequency corrections can also be made by adjusting
194 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
195 * used on both SMP and UP systems to avoid MP races between cpu's and
196 * interrupt races on UP systems.
198 #define BASETIME_ARYSIZE 16
199 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
200 static struct timespec basetime[BASETIME_ARYSIZE];
201 static volatile int basetime_index;
204 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
211 * Because basetime data and index may be updated by another cpu,
212 * a load fence is required to ensure that the data we read has
213 * not been speculatively read relative to a possibly updated index.
215 index = basetime_index;
217 bt = &basetime[index];
218 error = SYSCTL_OUT(req, bt, sizeof(*bt));
222 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
223 &boottime, timespec, "System boottime");
224 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
225 sysctl_get_basetime, "S,timespec", "System basetime");
227 static void hardclock(systimer_t info, int, struct intrframe *frame);
228 static void statclock(systimer_t info, int, struct intrframe *frame);
229 static void schedclock(systimer_t info, int, struct intrframe *frame);
230 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
232 int ticks; /* system master ticks at hz */
233 int clocks_running; /* tsleep/timeout clocks operational */
234 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
235 int64_t nsec_acc; /* accumulator */
236 int sched_ticks; /* global schedule clock ticks */
238 /* NTPD time correction fields */
239 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
240 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
241 int64_t ntp_delta; /* one-time correction in nsec */
242 int64_t ntp_big_delta = 1000000000;
243 int32_t ntp_tick_delta; /* current adjustment rate */
244 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
245 time_t ntp_leap_second; /* time of next leap second */
246 int ntp_leap_insert; /* whether to insert or remove a second */
249 * Finish initializing clock frequencies and start all clocks running.
253 initclocks(void *dummy)
255 /*psratio = profhz / stathz;*/
261 * Called on a per-cpu basis
264 initclocks_pcpu(void)
266 struct globaldata *gd = mycpu;
269 if (gd->gd_cpuid == 0) {
270 gd->gd_time_seconds = 1;
271 gd->gd_cpuclock_base = sys_cputimer->count();
274 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
275 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
278 systimer_intr_enable();
281 ifpoll_init_pcpu(gd->gd_cpuid);
285 * Use a non-queued periodic systimer to prevent multiple ticks from
286 * building up if the sysclock jumps forward (8254 gets reset). The
287 * sysclock will never jump backwards. Our time sync is based on
288 * the actual sysclock, not the ticks count.
290 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock, NULL, hz);
291 systimer_init_periodic_nq(&gd->gd_statclock, statclock, NULL, stathz);
292 /* XXX correct the frequency for scheduler / estcpu tests */
293 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
299 * This sets the current real time of day. Timespecs are in seconds and
300 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
301 * instead we adjust basetime so basetime + gd_* results in the current
302 * time of day. This way the gd_* fields are guarenteed to represent
303 * a monotonically increasing 'uptime' value.
305 * When set_timeofday() is called from userland, the system call forces it
306 * onto cpu #0 since only cpu #0 can update basetime_index.
309 set_timeofday(struct timespec *ts)
311 struct timespec *nbt;
315 * XXX SMP / non-atomic basetime updates
318 ni = (basetime_index + 1) & BASETIME_ARYMASK;
321 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
322 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
323 if (nbt->tv_nsec < 0) {
324 nbt->tv_nsec += 1000000000;
329 * Note that basetime diverges from boottime as the clock drift is
330 * compensated for, so we cannot do away with boottime. When setting
331 * the absolute time of day the drift is 0 (for an instant) and we
332 * can simply assign boottime to basetime.
334 * Note that nanouptime() is based on gd_time_seconds which is drift
335 * compensated up to a point (it is guarenteed to remain monotonically
336 * increasing). gd_time_seconds is thus our best uptime guess and
337 * suitable for use in the boottime calculation. It is already taken
338 * into account in the basetime calculation above.
340 boottime.tv_sec = nbt->tv_sec;
344 * We now have a new basetime, make sure all other cpus have it,
345 * then update the index.
354 * Each cpu has its own hardclock, but we only increments ticks and softticks
357 * NOTE! systimer! the MP lock might not be held here. We can only safely
358 * manipulate objects owned by the current cpu.
361 hardclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
365 struct globaldata *gd = mycpu;
368 * Realtime updates are per-cpu. Note that timer corrections as
369 * returned by microtime() and friends make an additional adjustment
370 * using a system-wise 'basetime', but the running time is always
371 * taken from the per-cpu globaldata area. Since the same clock
372 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
375 * Note that we never allow info->time (aka gd->gd_hardclock.time)
376 * to reverse index gd_cpuclock_base, but that it is possible for
377 * it to temporarily get behind in the seconds if something in the
378 * system locks interrupts for a long period of time. Since periodic
379 * timers count events, though everything should resynch again
382 cputicks = info->time - gd->gd_cpuclock_base;
383 if (cputicks >= sys_cputimer->freq) {
384 ++gd->gd_time_seconds;
385 gd->gd_cpuclock_base += sys_cputimer->freq;
389 * The system-wide ticks counter and NTP related timedelta/tickdelta
390 * adjustments only occur on cpu #0. NTP adjustments are accomplished
391 * by updating basetime.
393 if (gd->gd_cpuid == 0) {
394 struct timespec *nbt;
402 if (tco->tc_poll_pps)
403 tco->tc_poll_pps(tco);
407 * Calculate the new basetime index. We are in a critical section
408 * on cpu #0 and can safely play with basetime_index. Start
409 * with the current basetime and then make adjustments.
411 ni = (basetime_index + 1) & BASETIME_ARYMASK;
413 *nbt = basetime[basetime_index];
416 * Apply adjtime corrections. (adjtime() API)
418 * adjtime() only runs on cpu #0 so our critical section is
419 * sufficient to access these variables.
421 if (ntp_delta != 0) {
422 nbt->tv_nsec += ntp_tick_delta;
423 ntp_delta -= ntp_tick_delta;
424 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
425 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
426 ntp_tick_delta = ntp_delta;
431 * Apply permanent frequency corrections. (sysctl API)
433 if (ntp_tick_permanent != 0) {
434 ntp_tick_acc += ntp_tick_permanent;
435 if (ntp_tick_acc >= (1LL << 32)) {
436 nbt->tv_nsec += ntp_tick_acc >> 32;
437 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
438 } else if (ntp_tick_acc <= -(1LL << 32)) {
439 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
440 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
441 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
445 if (nbt->tv_nsec >= 1000000000) {
447 nbt->tv_nsec -= 1000000000;
448 } else if (nbt->tv_nsec < 0) {
450 nbt->tv_nsec += 1000000000;
454 * Another per-tick compensation. (for ntp_adjtime() API)
457 nsec_acc += nsec_adj;
458 if (nsec_acc >= 0x100000000LL) {
459 nbt->tv_nsec += nsec_acc >> 32;
460 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
461 } else if (nsec_acc <= -0x100000000LL) {
462 nbt->tv_nsec -= -nsec_acc >> 32;
463 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
465 if (nbt->tv_nsec >= 1000000000) {
466 nbt->tv_nsec -= 1000000000;
468 } else if (nbt->tv_nsec < 0) {
469 nbt->tv_nsec += 1000000000;
474 /************************************************************
475 * LEAP SECOND CORRECTION *
476 ************************************************************
478 * Taking into account all the corrections made above, figure
479 * out the new real time. If the seconds field has changed
480 * then apply any pending leap-second corrections.
482 getnanotime_nbt(nbt, &nts);
484 if (time_second != nts.tv_sec) {
486 * Apply leap second (sysctl API). Adjust nts for changes
487 * so we do not have to call getnanotime_nbt again.
489 if (ntp_leap_second) {
490 if (ntp_leap_second == nts.tv_sec) {
491 if (ntp_leap_insert) {
503 * Apply leap second (ntp_adjtime() API), calculate a new
504 * nsec_adj field. ntp_update_second() returns nsec_adj
505 * as a per-second value but we need it as a per-tick value.
507 leap = ntp_update_second(time_second, &nsec_adj);
513 * Update the time_second 'approximate time' global.
515 time_second = nts.tv_sec;
519 * Finally, our new basetime is ready to go live!
526 * lwkt thread scheduler fair queueing
528 lwkt_schedulerclock(curthread);
531 * softticks are handled for all cpus
533 hardclock_softtick(gd);
536 * ITimer handling is per-tick, per-cpu.
538 * We must acquire the per-process token in order for ksignal()
539 * to be non-blocking. For the moment this requires an AST fault,
540 * the ksignal() cannot be safely issued from this hard interrupt.
542 * XXX Even the trytoken here isn't right, and itimer operation in
543 * a multi threaded environment is going to be weird at the
546 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
548 if (frame && CLKF_USERMODE(frame) &&
549 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
550 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
551 p->p_flags |= P_SIGVTALRM;
554 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
555 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
556 p->p_flags |= P_SIGPROF;
560 lwkt_reltoken(&p->p_token);
566 * The statistics clock typically runs at a 125Hz rate, and is intended
567 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
569 * NOTE! systimer! the MP lock might not be held here. We can only safely
570 * manipulate objects owned by the current cpu.
572 * The stats clock is responsible for grabbing a profiling sample.
573 * Most of the statistics are only used by user-level statistics programs.
574 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
577 * Like the other clocks, the stat clock is called from what is effectively
578 * a fast interrupt, so the context should be the thread/process that got
582 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
595 * How big was our timeslice relative to the last time?
597 microuptime(&tv); /* mpsafe */
598 stv = &mycpu->gd_stattv;
599 if (stv->tv_sec == 0) {
602 bump = tv.tv_usec - stv->tv_usec +
603 (tv.tv_sec - stv->tv_sec) * 1000000;
614 if (frame && CLKF_USERMODE(frame)) {
616 * Came from userland, handle user time and deal with
619 if (p && (p->p_flags & P_PROFIL))
620 addupc_intr(p, CLKF_PC(frame), 1);
621 td->td_uticks += bump;
624 * Charge the time as appropriate
626 if (p && p->p_nice > NZERO)
627 cpu_time.cp_nice += bump;
629 cpu_time.cp_user += bump;
631 int intr_nest = mycpu->gd_intr_nesting_level;
635 * IPI processing code will bump gd_intr_nesting_level
636 * up by one, which breaks following CLKF_INTR testing,
637 * so we substract it by one here.
643 * Kernel statistics are just like addupc_intr, only easier.
646 if (g->state == GMON_PROF_ON && frame) {
647 i = CLKF_PC(frame) - g->lowpc;
648 if (i < g->textsize) {
649 i /= HISTFRACTION * sizeof(*g->kcount);
655 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
658 * Came from kernel mode, so we were:
659 * - handling an interrupt,
660 * - doing syscall or trap work on behalf of the current
662 * - spinning in the idle loop.
663 * Whichever it is, charge the time as appropriate.
664 * Note that we charge interrupts to the current process,
665 * regardless of whether they are ``for'' that process,
666 * so that we know how much of its real time was spent
667 * in ``non-process'' (i.e., interrupt) work.
669 * XXX assume system if frame is NULL. A NULL frame
670 * can occur if ipi processing is done from a crit_exit().
673 td->td_iticks += bump;
675 td->td_sticks += bump;
677 if (IS_INTR_RUNNING) {
679 * If we interrupted an interrupt thread, well,
680 * count it as interrupt time.
684 do_pctrack(frame, PCTRACK_INT);
686 cpu_time.cp_intr += bump;
688 if (td == &mycpu->gd_idlethread) {
690 * Even if the current thread is the idle
691 * thread it could be due to token contention
692 * in the LWKT scheduler. Count such as
695 if (mycpu->gd_reqflags & RQF_AST_LWKT_RESCHED)
696 cpu_time.cp_sys += bump;
698 cpu_time.cp_idle += bump;
701 * System thread was running.
705 do_pctrack(frame, PCTRACK_SYS);
707 cpu_time.cp_sys += bump;
711 #undef IS_INTR_RUNNING
717 * Sample the PC when in the kernel or in an interrupt. User code can
718 * retrieve the information and generate a histogram or other output.
722 do_pctrack(struct intrframe *frame, int which)
724 struct kinfo_pctrack *pctrack;
726 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
727 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
728 (void *)CLKF_PC(frame);
733 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
735 struct kinfo_pcheader head;
740 head.pc_ntrack = PCTRACK_SIZE;
741 head.pc_arysize = PCTRACK_ARYSIZE;
743 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
746 for (cpu = 0; cpu < ncpus; ++cpu) {
747 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
748 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
749 sizeof(struct kinfo_pctrack));
758 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
759 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
764 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
765 * the MP lock might not be held. We can safely manipulate parts of curproc
766 * but that's about it.
768 * Each cpu has its own scheduler clock.
771 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
778 if ((lp = lwkt_preempted_proc()) != NULL) {
780 * Account for cpu time used and hit the scheduler. Note
781 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
785 usched_schedulerclock(lp, info->periodic, info->time);
787 usched_schedulerclock(NULL, info->periodic, info->time);
789 if ((lp = curthread->td_lwp) != NULL) {
791 * Update resource usage integrals and maximums.
793 if ((ru = &lp->lwp_proc->p_ru) &&
794 (vm = lp->lwp_proc->p_vmspace) != NULL) {
795 ru->ru_ixrss += pgtok(vm->vm_tsize);
796 ru->ru_idrss += pgtok(vm->vm_dsize);
797 ru->ru_isrss += pgtok(vm->vm_ssize);
798 if (lwkt_trytoken(&vm->vm_map.token)) {
799 rss = pgtok(vmspace_resident_count(vm));
800 if (ru->ru_maxrss < rss)
802 lwkt_reltoken(&vm->vm_map.token);
806 /* Increment the global sched_ticks */
807 if (mycpu->gd_cpuid == 0)
812 * Compute number of ticks for the specified amount of time. The
813 * return value is intended to be used in a clock interrupt timed
814 * operation and guarenteed to meet or exceed the requested time.
815 * If the representation overflows, return INT_MAX. The minimum return
816 * value is 1 ticks and the function will average the calculation up.
817 * If any value greater then 0 microseconds is supplied, a value
818 * of at least 2 will be returned to ensure that a near-term clock
819 * interrupt does not cause the timeout to occur (degenerately) early.
821 * Note that limit checks must take into account microseconds, which is
822 * done simply by using the smaller signed long maximum instead of
823 * the unsigned long maximum.
825 * If ints have 32 bits, then the maximum value for any timeout in
826 * 10ms ticks is 248 days.
829 tvtohz_high(struct timeval *tv)
846 kprintf("tvtohz_high: negative time difference "
847 "%ld sec %ld usec\n",
851 } else if (sec <= INT_MAX / hz) {
852 ticks = (int)(sec * hz +
853 ((u_long)usec + (ustick - 1)) / ustick) + 1;
861 tstohz_high(struct timespec *ts)
878 kprintf("tstohz_high: negative time difference "
879 "%ld sec %ld nsec\n",
883 } else if (sec <= INT_MAX / hz) {
884 ticks = (int)(sec * hz +
885 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
894 * Compute number of ticks for the specified amount of time, erroring on
895 * the side of it being too low to ensure that sleeping the returned number
896 * of ticks will not result in a late return.
898 * The supplied timeval may not be negative and should be normalized. A
899 * return value of 0 is possible if the timeval converts to less then
902 * If ints have 32 bits, then the maximum value for any timeout in
903 * 10ms ticks is 248 days.
906 tvtohz_low(struct timeval *tv)
912 if (sec <= INT_MAX / hz)
913 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
920 tstohz_low(struct timespec *ts)
926 if (sec <= INT_MAX / hz)
927 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
934 * Start profiling on a process.
936 * Kernel profiling passes proc0 which never exits and hence
937 * keeps the profile clock running constantly.
940 startprofclock(struct proc *p)
942 if ((p->p_flags & P_PROFIL) == 0) {
943 p->p_flags |= P_PROFIL;
945 if (++profprocs == 1 && stathz != 0) {
948 setstatclockrate(profhz);
956 * Stop profiling on a process.
958 * caller must hold p->p_token
961 stopprofclock(struct proc *p)
963 if (p->p_flags & P_PROFIL) {
964 p->p_flags &= ~P_PROFIL;
966 if (--profprocs == 0 && stathz != 0) {
969 setstatclockrate(stathz);
977 * Return information about system clocks.
980 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
982 struct kinfo_clockinfo clkinfo;
984 * Construct clockinfo structure.
987 clkinfo.ci_tick = ustick;
988 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
989 clkinfo.ci_profhz = profhz;
990 clkinfo.ci_stathz = stathz ? stathz : hz;
991 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
994 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
995 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
998 * We have eight functions for looking at the clock, four for
999 * microseconds and four for nanoseconds. For each there is fast
1000 * but less precise version "get{nano|micro}[up]time" which will
1001 * return a time which is up to 1/HZ previous to the call, whereas
1002 * the raw version "{nano|micro}[up]time" will return a timestamp
1003 * which is as precise as possible. The "up" variants return the
1004 * time relative to system boot, these are well suited for time
1005 * interval measurements.
1007 * Each cpu independantly maintains the current time of day, so all
1008 * we need to do to protect ourselves from changes is to do a loop
1009 * check on the seconds field changing out from under us.
1011 * The system timer maintains a 32 bit count and due to various issues
1012 * it is possible for the calculated delta to occassionally exceed
1013 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1014 * multiplication can easily overflow, so we deal with the case. For
1015 * uniformity we deal with the case in the usec case too.
1017 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1020 getmicrouptime(struct timeval *tvp)
1022 struct globaldata *gd = mycpu;
1026 tvp->tv_sec = gd->gd_time_seconds;
1027 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1028 } while (tvp->tv_sec != gd->gd_time_seconds);
1030 if (delta >= sys_cputimer->freq) {
1031 tvp->tv_sec += delta / sys_cputimer->freq;
1032 delta %= sys_cputimer->freq;
1034 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1035 if (tvp->tv_usec >= 1000000) {
1036 tvp->tv_usec -= 1000000;
1042 getnanouptime(struct timespec *tsp)
1044 struct globaldata *gd = mycpu;
1048 tsp->tv_sec = gd->gd_time_seconds;
1049 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1050 } while (tsp->tv_sec != gd->gd_time_seconds);
1052 if (delta >= sys_cputimer->freq) {
1053 tsp->tv_sec += delta / sys_cputimer->freq;
1054 delta %= sys_cputimer->freq;
1056 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1060 microuptime(struct timeval *tvp)
1062 struct globaldata *gd = mycpu;
1066 tvp->tv_sec = gd->gd_time_seconds;
1067 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1068 } while (tvp->tv_sec != gd->gd_time_seconds);
1070 if (delta >= sys_cputimer->freq) {
1071 tvp->tv_sec += delta / sys_cputimer->freq;
1072 delta %= sys_cputimer->freq;
1074 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1078 nanouptime(struct timespec *tsp)
1080 struct globaldata *gd = mycpu;
1084 tsp->tv_sec = gd->gd_time_seconds;
1085 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1086 } while (tsp->tv_sec != gd->gd_time_seconds);
1088 if (delta >= sys_cputimer->freq) {
1089 tsp->tv_sec += delta / sys_cputimer->freq;
1090 delta %= sys_cputimer->freq;
1092 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1099 getmicrotime(struct timeval *tvp)
1101 struct globaldata *gd = mycpu;
1102 struct timespec *bt;
1106 tvp->tv_sec = gd->gd_time_seconds;
1107 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1108 } while (tvp->tv_sec != gd->gd_time_seconds);
1110 if (delta >= sys_cputimer->freq) {
1111 tvp->tv_sec += delta / sys_cputimer->freq;
1112 delta %= sys_cputimer->freq;
1114 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1116 bt = &basetime[basetime_index];
1117 tvp->tv_sec += bt->tv_sec;
1118 tvp->tv_usec += bt->tv_nsec / 1000;
1119 while (tvp->tv_usec >= 1000000) {
1120 tvp->tv_usec -= 1000000;
1126 getnanotime(struct timespec *tsp)
1128 struct globaldata *gd = mycpu;
1129 struct timespec *bt;
1133 tsp->tv_sec = gd->gd_time_seconds;
1134 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1135 } while (tsp->tv_sec != gd->gd_time_seconds);
1137 if (delta >= sys_cputimer->freq) {
1138 tsp->tv_sec += delta / sys_cputimer->freq;
1139 delta %= sys_cputimer->freq;
1141 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1143 bt = &basetime[basetime_index];
1144 tsp->tv_sec += bt->tv_sec;
1145 tsp->tv_nsec += bt->tv_nsec;
1146 while (tsp->tv_nsec >= 1000000000) {
1147 tsp->tv_nsec -= 1000000000;
1153 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1155 struct globaldata *gd = mycpu;
1159 tsp->tv_sec = gd->gd_time_seconds;
1160 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1161 } while (tsp->tv_sec != gd->gd_time_seconds);
1163 if (delta >= sys_cputimer->freq) {
1164 tsp->tv_sec += delta / sys_cputimer->freq;
1165 delta %= sys_cputimer->freq;
1167 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1169 tsp->tv_sec += nbt->tv_sec;
1170 tsp->tv_nsec += nbt->tv_nsec;
1171 while (tsp->tv_nsec >= 1000000000) {
1172 tsp->tv_nsec -= 1000000000;
1179 microtime(struct timeval *tvp)
1181 struct globaldata *gd = mycpu;
1182 struct timespec *bt;
1186 tvp->tv_sec = gd->gd_time_seconds;
1187 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1188 } while (tvp->tv_sec != gd->gd_time_seconds);
1190 if (delta >= sys_cputimer->freq) {
1191 tvp->tv_sec += delta / sys_cputimer->freq;
1192 delta %= sys_cputimer->freq;
1194 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1196 bt = &basetime[basetime_index];
1197 tvp->tv_sec += bt->tv_sec;
1198 tvp->tv_usec += bt->tv_nsec / 1000;
1199 while (tvp->tv_usec >= 1000000) {
1200 tvp->tv_usec -= 1000000;
1206 nanotime(struct timespec *tsp)
1208 struct globaldata *gd = mycpu;
1209 struct timespec *bt;
1213 tsp->tv_sec = gd->gd_time_seconds;
1214 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1215 } while (tsp->tv_sec != gd->gd_time_seconds);
1217 if (delta >= sys_cputimer->freq) {
1218 tsp->tv_sec += delta / sys_cputimer->freq;
1219 delta %= sys_cputimer->freq;
1221 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1223 bt = &basetime[basetime_index];
1224 tsp->tv_sec += bt->tv_sec;
1225 tsp->tv_nsec += bt->tv_nsec;
1226 while (tsp->tv_nsec >= 1000000000) {
1227 tsp->tv_nsec -= 1000000000;
1233 * note: this is not exactly synchronized with real time. To do that we
1234 * would have to do what microtime does and check for a nanoseconds overflow.
1237 get_approximate_time_t(void)
1239 struct globaldata *gd = mycpu;
1240 struct timespec *bt;
1242 bt = &basetime[basetime_index];
1243 return(gd->gd_time_seconds + bt->tv_sec);
1247 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1250 struct pps_fetch_args *fapi;
1252 struct pps_kcbind_args *kapi;
1256 case PPS_IOC_CREATE:
1258 case PPS_IOC_DESTROY:
1260 case PPS_IOC_SETPARAMS:
1261 app = (pps_params_t *)data;
1262 if (app->mode & ~pps->ppscap)
1264 pps->ppsparam = *app;
1266 case PPS_IOC_GETPARAMS:
1267 app = (pps_params_t *)data;
1268 *app = pps->ppsparam;
1269 app->api_version = PPS_API_VERS_1;
1271 case PPS_IOC_GETCAP:
1272 *(int*)data = pps->ppscap;
1275 fapi = (struct pps_fetch_args *)data;
1276 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1278 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1279 return (EOPNOTSUPP);
1280 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1281 fapi->pps_info_buf = pps->ppsinfo;
1283 case PPS_IOC_KCBIND:
1285 kapi = (struct pps_kcbind_args *)data;
1286 /* XXX Only root should be able to do this */
1287 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1289 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1291 if (kapi->edge & ~pps->ppscap)
1293 pps->kcmode = kapi->edge;
1296 return (EOPNOTSUPP);
1304 pps_init(struct pps_state *pps)
1306 pps->ppscap |= PPS_TSFMT_TSPEC;
1307 if (pps->ppscap & PPS_CAPTUREASSERT)
1308 pps->ppscap |= PPS_OFFSETASSERT;
1309 if (pps->ppscap & PPS_CAPTURECLEAR)
1310 pps->ppscap |= PPS_OFFSETCLEAR;
1314 pps_event(struct pps_state *pps, sysclock_t count, int event)
1316 struct globaldata *gd;
1317 struct timespec *tsp;
1318 struct timespec *osp;
1319 struct timespec *bt;
1332 /* Things would be easier with arrays... */
1333 if (event == PPS_CAPTUREASSERT) {
1334 tsp = &pps->ppsinfo.assert_timestamp;
1335 osp = &pps->ppsparam.assert_offset;
1336 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1337 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1338 pcount = &pps->ppscount[0];
1339 pseq = &pps->ppsinfo.assert_sequence;
1341 tsp = &pps->ppsinfo.clear_timestamp;
1342 osp = &pps->ppsparam.clear_offset;
1343 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1344 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1345 pcount = &pps->ppscount[1];
1346 pseq = &pps->ppsinfo.clear_sequence;
1349 /* Nothing really happened */
1350 if (*pcount == count)
1356 ts.tv_sec = gd->gd_time_seconds;
1357 delta = count - gd->gd_cpuclock_base;
1358 } while (ts.tv_sec != gd->gd_time_seconds);
1360 if (delta >= sys_cputimer->freq) {
1361 ts.tv_sec += delta / sys_cputimer->freq;
1362 delta %= sys_cputimer->freq;
1364 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1365 bt = &basetime[basetime_index];
1366 ts.tv_sec += bt->tv_sec;
1367 ts.tv_nsec += bt->tv_nsec;
1368 while (ts.tv_nsec >= 1000000000) {
1369 ts.tv_nsec -= 1000000000;
1377 timespecadd(tsp, osp);
1378 if (tsp->tv_nsec < 0) {
1379 tsp->tv_nsec += 1000000000;
1385 /* magic, at its best... */
1386 tcount = count - pps->ppscount[2];
1387 pps->ppscount[2] = count;
1388 if (tcount >= sys_cputimer->freq) {
1389 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1390 sys_cputimer->freq64_nsec *
1391 (tcount % sys_cputimer->freq)) >> 32;
1393 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1395 hardpps(tsp, delta);
1401 * Return the tsc target value for a delay of (ns).
1403 * Returns -1 if the TSC is not supported.
1406 tsc_get_target(int ns)
1408 #if defined(_RDTSC_SUPPORTED_)
1409 if (cpu_feature & CPUID_TSC) {
1410 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1417 * Compare the tsc against the passed target
1419 * Returns +1 if the target has been reached
1420 * Returns 0 if the target has not yet been reached
1421 * Returns -1 if the TSC is not supported.
1423 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1426 tsc_test_target(int64_t target)
1428 #if defined(_RDTSC_SUPPORTED_)
1429 if (cpu_feature & CPUID_TSC) {
1430 if ((int64_t)(target - rdtsc()) <= 0)
1439 * Delay the specified number of nanoseconds using the tsc. This function
1440 * returns immediately if the TSC is not supported. At least one cpu_pause()
1448 clk = tsc_get_target(ns);
1450 while (tsc_test_target(clk) == 0)