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>
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>
111 #ifdef DEVICE_POLLING
112 extern void init_device_poll_pcpu(int);
116 extern void ifpoll_init_pcpu(int);
120 static void do_pctrack(struct intrframe
*frame
, int which
);
123 static void initclocks (void *dummy
);
124 SYSINIT(clocks
, SI_BOOT2_CLOCKS
, SI_ORDER_FIRST
, initclocks
, NULL
)
127 * Some of these don't belong here, but it's easiest to concentrate them.
128 * Note that cpu_time counts in microseconds, but most userland programs
129 * just compare relative times against the total by delta.
131 struct kinfo_cputime cputime_percpu
[MAXCPU
];
133 struct kinfo_pcheader cputime_pcheader
= { PCTRACK_SIZE
, PCTRACK_ARYSIZE
};
134 struct kinfo_pctrack cputime_pctrack
[MAXCPU
][PCTRACK_SIZE
];
139 sysctl_cputime(SYSCTL_HANDLER_ARGS
)
142 size_t size
= sizeof(struct kinfo_cputime
);
144 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
145 if ((error
= SYSCTL_OUT(req
, &cputime_percpu
[cpu
], size
)))
151 SYSCTL_PROC(_kern
, OID_AUTO
, cputime
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
152 sysctl_cputime
, "S,kinfo_cputime", "CPU time statistics");
154 SYSCTL_STRUCT(_kern
, OID_AUTO
, cputime
, CTLFLAG_RD
, &cpu_time
, kinfo_cputime
,
155 "CPU time statistics");
159 * boottime is used to calculate the 'real' uptime. Do not confuse this with
160 * microuptime(). microtime() is not drift compensated. The real uptime
161 * with compensation is nanotime() - bootime. boottime is recalculated
162 * whenever the real time is set based on the compensated elapsed time
163 * in seconds (gd->gd_time_seconds).
165 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
166 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
169 struct timespec boottime
; /* boot time (realtime) for reference only */
170 time_t time_second
; /* read-only 'passive' uptime in seconds */
173 * basetime is used to calculate the compensated real time of day. The
174 * basetime can be modified on a per-tick basis by the adjtime(),
175 * ntp_adjtime(), and sysctl-based time correction APIs.
177 * Note that frequency corrections can also be made by adjusting
180 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
181 * used on both SMP and UP systems to avoid MP races between cpu's and
182 * interrupt races on UP systems.
184 #define BASETIME_ARYSIZE 16
185 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
186 static struct timespec basetime
[BASETIME_ARYSIZE
];
187 static volatile int basetime_index
;
190 sysctl_get_basetime(SYSCTL_HANDLER_ARGS
)
197 * Because basetime data and index may be updated by another cpu,
198 * a load fence is required to ensure that the data we read has
199 * not been speculatively read relative to a possibly updated index.
201 index
= basetime_index
;
203 bt
= &basetime
[index
];
204 error
= SYSCTL_OUT(req
, bt
, sizeof(*bt
));
208 SYSCTL_STRUCT(_kern
, KERN_BOOTTIME
, boottime
, CTLFLAG_RD
,
209 &boottime
, timespec
, "System boottime");
210 SYSCTL_PROC(_kern
, OID_AUTO
, basetime
, CTLTYPE_STRUCT
|CTLFLAG_RD
, 0, 0,
211 sysctl_get_basetime
, "S,timespec", "System basetime");
213 static void hardclock(systimer_t info
, struct intrframe
*frame
);
214 static void statclock(systimer_t info
, struct intrframe
*frame
);
215 static void schedclock(systimer_t info
, struct intrframe
*frame
);
216 static void getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
);
218 int ticks
; /* system master ticks at hz */
219 int clocks_running
; /* tsleep/timeout clocks operational */
220 int64_t nsec_adj
; /* ntpd per-tick adjustment in nsec << 32 */
221 int64_t nsec_acc
; /* accumulator */
223 /* NTPD time correction fields */
224 int64_t ntp_tick_permanent
; /* per-tick adjustment in nsec << 32 */
225 int64_t ntp_tick_acc
; /* accumulator for per-tick adjustment */
226 int64_t ntp_delta
; /* one-time correction in nsec */
227 int64_t ntp_big_delta
= 1000000000;
228 int32_t ntp_tick_delta
; /* current adjustment rate */
229 int32_t ntp_default_tick_delta
; /* adjustment rate for ntp_delta */
230 time_t ntp_leap_second
; /* time of next leap second */
231 int ntp_leap_insert
; /* whether to insert or remove a second */
234 * Finish initializing clock frequencies and start all clocks running.
238 initclocks(void *dummy
)
240 /*psratio = profhz / stathz;*/
246 * Called on a per-cpu basis
249 initclocks_pcpu(void)
251 struct globaldata
*gd
= mycpu
;
254 if (gd
->gd_cpuid
== 0) {
255 gd
->gd_time_seconds
= 1;
256 gd
->gd_cpuclock_base
= sys_cputimer
->count();
259 gd
->gd_time_seconds
= globaldata_find(0)->gd_time_seconds
;
260 gd
->gd_cpuclock_base
= globaldata_find(0)->gd_cpuclock_base
;
263 systimer_intr_enable();
265 #ifdef DEVICE_POLLING
266 init_device_poll_pcpu(gd
->gd_cpuid
);
270 ifpoll_init_pcpu(gd
->gd_cpuid
);
274 * Use a non-queued periodic systimer to prevent multiple ticks from
275 * building up if the sysclock jumps forward (8254 gets reset). The
276 * sysclock will never jump backwards. Our time sync is based on
277 * the actual sysclock, not the ticks count.
279 systimer_init_periodic_nq(&gd
->gd_hardclock
, hardclock
, NULL
, hz
);
280 systimer_init_periodic_nq(&gd
->gd_statclock
, statclock
, NULL
, stathz
);
281 /* XXX correct the frequency for scheduler / estcpu tests */
282 systimer_init_periodic_nq(&gd
->gd_schedclock
, schedclock
,
288 * This sets the current real time of day. Timespecs are in seconds and
289 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
290 * instead we adjust basetime so basetime + gd_* results in the current
291 * time of day. This way the gd_* fields are guarenteed to represent
292 * a monotonically increasing 'uptime' value.
294 * When set_timeofday() is called from userland, the system call forces it
295 * onto cpu #0 since only cpu #0 can update basetime_index.
298 set_timeofday(struct timespec
*ts
)
300 struct timespec
*nbt
;
304 * XXX SMP / non-atomic basetime updates
307 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
310 nbt
->tv_sec
= ts
->tv_sec
- nbt
->tv_sec
;
311 nbt
->tv_nsec
= ts
->tv_nsec
- nbt
->tv_nsec
;
312 if (nbt
->tv_nsec
< 0) {
313 nbt
->tv_nsec
+= 1000000000;
318 * Note that basetime diverges from boottime as the clock drift is
319 * compensated for, so we cannot do away with boottime. When setting
320 * the absolute time of day the drift is 0 (for an instant) and we
321 * can simply assign boottime to basetime.
323 * Note that nanouptime() is based on gd_time_seconds which is drift
324 * compensated up to a point (it is guarenteed to remain monotonically
325 * increasing). gd_time_seconds is thus our best uptime guess and
326 * suitable for use in the boottime calculation. It is already taken
327 * into account in the basetime calculation above.
329 boottime
.tv_sec
= nbt
->tv_sec
;
333 * We now have a new basetime, make sure all other cpus have it,
334 * then update the index.
343 * Each cpu has its own hardclock, but we only increments ticks and softticks
346 * NOTE! systimer! the MP lock might not be held here. We can only safely
347 * manipulate objects owned by the current cpu.
350 hardclock(systimer_t info
, struct intrframe
*frame
)
354 struct globaldata
*gd
= mycpu
;
357 * Realtime updates are per-cpu. Note that timer corrections as
358 * returned by microtime() and friends make an additional adjustment
359 * using a system-wise 'basetime', but the running time is always
360 * taken from the per-cpu globaldata area. Since the same clock
361 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
364 * Note that we never allow info->time (aka gd->gd_hardclock.time)
365 * to reverse index gd_cpuclock_base, but that it is possible for
366 * it to temporarily get behind in the seconds if something in the
367 * system locks interrupts for a long period of time. Since periodic
368 * timers count events, though everything should resynch again
371 cputicks
= info
->time
- gd
->gd_cpuclock_base
;
372 if (cputicks
>= sys_cputimer
->freq
) {
373 ++gd
->gd_time_seconds
;
374 gd
->gd_cpuclock_base
+= sys_cputimer
->freq
;
378 * The system-wide ticks counter and NTP related timedelta/tickdelta
379 * adjustments only occur on cpu #0. NTP adjustments are accomplished
380 * by updating basetime.
382 if (gd
->gd_cpuid
== 0) {
383 struct timespec
*nbt
;
391 if (tco
->tc_poll_pps
)
392 tco
->tc_poll_pps(tco
);
396 * Calculate the new basetime index. We are in a critical section
397 * on cpu #0 and can safely play with basetime_index. Start
398 * with the current basetime and then make adjustments.
400 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
402 *nbt
= basetime
[basetime_index
];
405 * Apply adjtime corrections. (adjtime() API)
407 * adjtime() only runs on cpu #0 so our critical section is
408 * sufficient to access these variables.
410 if (ntp_delta
!= 0) {
411 nbt
->tv_nsec
+= ntp_tick_delta
;
412 ntp_delta
-= ntp_tick_delta
;
413 if ((ntp_delta
> 0 && ntp_delta
< ntp_tick_delta
) ||
414 (ntp_delta
< 0 && ntp_delta
> ntp_tick_delta
)) {
415 ntp_tick_delta
= ntp_delta
;
420 * Apply permanent frequency corrections. (sysctl API)
422 if (ntp_tick_permanent
!= 0) {
423 ntp_tick_acc
+= ntp_tick_permanent
;
424 if (ntp_tick_acc
>= (1LL << 32)) {
425 nbt
->tv_nsec
+= ntp_tick_acc
>> 32;
426 ntp_tick_acc
-= (ntp_tick_acc
>> 32) << 32;
427 } else if (ntp_tick_acc
<= -(1LL << 32)) {
428 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
429 nbt
->tv_nsec
-= (-ntp_tick_acc
) >> 32;
430 ntp_tick_acc
+= ((-ntp_tick_acc
) >> 32) << 32;
434 if (nbt
->tv_nsec
>= 1000000000) {
436 nbt
->tv_nsec
-= 1000000000;
437 } else if (nbt
->tv_nsec
< 0) {
439 nbt
->tv_nsec
+= 1000000000;
443 * Another per-tick compensation. (for ntp_adjtime() API)
446 nsec_acc
+= nsec_adj
;
447 if (nsec_acc
>= 0x100000000LL
) {
448 nbt
->tv_nsec
+= nsec_acc
>> 32;
449 nsec_acc
= (nsec_acc
& 0xFFFFFFFFLL
);
450 } else if (nsec_acc
<= -0x100000000LL
) {
451 nbt
->tv_nsec
-= -nsec_acc
>> 32;
452 nsec_acc
= -(-nsec_acc
& 0xFFFFFFFFLL
);
454 if (nbt
->tv_nsec
>= 1000000000) {
455 nbt
->tv_nsec
-= 1000000000;
457 } else if (nbt
->tv_nsec
< 0) {
458 nbt
->tv_nsec
+= 1000000000;
463 /************************************************************
464 * LEAP SECOND CORRECTION *
465 ************************************************************
467 * Taking into account all the corrections made above, figure
468 * out the new real time. If the seconds field has changed
469 * then apply any pending leap-second corrections.
471 getnanotime_nbt(nbt
, &nts
);
473 if (time_second
!= nts
.tv_sec
) {
475 * Apply leap second (sysctl API). Adjust nts for changes
476 * so we do not have to call getnanotime_nbt again.
478 if (ntp_leap_second
) {
479 if (ntp_leap_second
== nts
.tv_sec
) {
480 if (ntp_leap_insert
) {
492 * Apply leap second (ntp_adjtime() API), calculate a new
493 * nsec_adj field. ntp_update_second() returns nsec_adj
494 * as a per-second value but we need it as a per-tick value.
496 leap
= ntp_update_second(time_second
, &nsec_adj
);
502 * Update the time_second 'approximate time' global.
504 time_second
= nts
.tv_sec
;
508 * Finally, our new basetime is ready to go live!
514 * Figure out how badly the system is starved for memory
516 vm_fault_ratecheck();
520 * softticks are handled for all cpus
522 hardclock_softtick(gd
);
525 * The LWKT scheduler will generally allow the current process to
526 * return to user mode even if there are other runnable LWKT threads
527 * running in kernel mode on behalf of a user process. This will
528 * ensure that those other threads have an opportunity to run in
529 * fairly short order (but not instantly).
534 * ITimer handling is per-tick, per-cpu. I don't think ksignal()
535 * is mpsafe on curproc, so XXX get the mplock.
537 if ((p
= curproc
) != NULL
&& try_mplock()) {
538 if (frame
&& CLKF_USERMODE(frame
) &&
539 timevalisset(&p
->p_timer
[ITIMER_VIRTUAL
].it_value
) &&
540 itimerdecr(&p
->p_timer
[ITIMER_VIRTUAL
], tick
) == 0)
541 ksignal(p
, SIGVTALRM
);
542 if (timevalisset(&p
->p_timer
[ITIMER_PROF
].it_value
) &&
543 itimerdecr(&p
->p_timer
[ITIMER_PROF
], tick
) == 0)
551 * The statistics clock typically runs at a 125Hz rate, and is intended
552 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
554 * NOTE! systimer! the MP lock might not be held here. We can only safely
555 * manipulate objects owned by the current cpu.
557 * The stats clock is responsible for grabbing a profiling sample.
558 * Most of the statistics are only used by user-level statistics programs.
559 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
562 * Like the other clocks, the stat clock is called from what is effectively
563 * a fast interrupt, so the context should be the thread/process that got
567 statclock(systimer_t info
, struct intrframe
*frame
)
580 * How big was our timeslice relative to the last time?
582 microuptime(&tv
); /* mpsafe */
583 stv
= &mycpu
->gd_stattv
;
584 if (stv
->tv_sec
== 0) {
587 bump
= tv
.tv_usec
- stv
->tv_usec
+
588 (tv
.tv_sec
- stv
->tv_sec
) * 1000000;
599 if (frame
&& CLKF_USERMODE(frame
)) {
601 * Came from userland, handle user time and deal with
604 if (p
&& (p
->p_flag
& P_PROFIL
))
605 addupc_intr(p
, CLKF_PC(frame
), 1);
606 td
->td_uticks
+= bump
;
609 * Charge the time as appropriate
611 if (p
&& p
->p_nice
> NZERO
)
612 cpu_time
.cp_nice
+= bump
;
614 cpu_time
.cp_user
+= bump
;
618 * Kernel statistics are just like addupc_intr, only easier.
621 if (g
->state
== GMON_PROF_ON
&& frame
) {
622 i
= CLKF_PC(frame
) - g
->lowpc
;
623 if (i
< g
->textsize
) {
624 i
/= HISTFRACTION
* sizeof(*g
->kcount
);
630 * Came from kernel mode, so we were:
631 * - handling an interrupt,
632 * - doing syscall or trap work on behalf of the current
634 * - spinning in the idle loop.
635 * Whichever it is, charge the time as appropriate.
636 * Note that we charge interrupts to the current process,
637 * regardless of whether they are ``for'' that process,
638 * so that we know how much of its real time was spent
639 * in ``non-process'' (i.e., interrupt) work.
641 * XXX assume system if frame is NULL. A NULL frame
642 * can occur if ipi processing is done from a crit_exit().
644 if (frame
&& CLKF_INTR(frame
))
645 td
->td_iticks
+= bump
;
647 td
->td_sticks
+= bump
;
649 if (frame
&& CLKF_INTR(frame
)) {
651 do_pctrack(frame
, PCTRACK_INT
);
653 cpu_time
.cp_intr
+= bump
;
655 if (td
== &mycpu
->gd_idlethread
) {
656 cpu_time
.cp_idle
+= bump
;
660 do_pctrack(frame
, PCTRACK_SYS
);
662 cpu_time
.cp_sys
+= bump
;
670 * Sample the PC when in the kernel or in an interrupt. User code can
671 * retrieve the information and generate a histogram or other output.
675 do_pctrack(struct intrframe
*frame
, int which
)
677 struct kinfo_pctrack
*pctrack
;
679 pctrack
= &cputime_pctrack
[mycpu
->gd_cpuid
][which
];
680 pctrack
->pc_array
[pctrack
->pc_index
& PCTRACK_ARYMASK
] =
681 (void *)CLKF_PC(frame
);
686 sysctl_pctrack(SYSCTL_HANDLER_ARGS
)
688 struct kinfo_pcheader head
;
693 head
.pc_ntrack
= PCTRACK_SIZE
;
694 head
.pc_arysize
= PCTRACK_ARYSIZE
;
696 if ((error
= SYSCTL_OUT(req
, &head
, sizeof(head
))) != 0)
699 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
700 for (ntrack
= 0; ntrack
< PCTRACK_SIZE
; ++ntrack
) {
701 error
= SYSCTL_OUT(req
, &cputime_pctrack
[cpu
][ntrack
],
702 sizeof(struct kinfo_pctrack
));
711 SYSCTL_PROC(_kern
, OID_AUTO
, pctrack
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
712 sysctl_pctrack
, "S,kinfo_pcheader", "CPU PC tracking");
717 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
718 * the MP lock might not be held. We can safely manipulate parts of curproc
719 * but that's about it.
721 * Each cpu has its own scheduler clock.
724 schedclock(systimer_t info
, struct intrframe
*frame
)
731 if ((lp
= lwkt_preempted_proc()) != NULL
) {
733 * Account for cpu time used and hit the scheduler. Note
734 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
738 lp
->lwp_proc
->p_usched
->schedulerclock(lp
, info
->periodic
,
741 if ((lp
= curthread
->td_lwp
) != NULL
) {
743 * Update resource usage integrals and maximums.
745 if ((ru
= &lp
->lwp_proc
->p_ru
) &&
746 (vm
= lp
->lwp_proc
->p_vmspace
) != NULL
) {
747 ru
->ru_ixrss
+= pgtok(vm
->vm_tsize
);
748 ru
->ru_idrss
+= pgtok(vm
->vm_dsize
);
749 ru
->ru_isrss
+= pgtok(vm
->vm_ssize
);
750 rss
= pgtok(vmspace_resident_count(vm
));
751 if (ru
->ru_maxrss
< rss
)
758 * Compute number of ticks for the specified amount of time. The
759 * return value is intended to be used in a clock interrupt timed
760 * operation and guarenteed to meet or exceed the requested time.
761 * If the representation overflows, return INT_MAX. The minimum return
762 * value is 1 ticks and the function will average the calculation up.
763 * If any value greater then 0 microseconds is supplied, a value
764 * of at least 2 will be returned to ensure that a near-term clock
765 * interrupt does not cause the timeout to occur (degenerately) early.
767 * Note that limit checks must take into account microseconds, which is
768 * done simply by using the smaller signed long maximum instead of
769 * the unsigned long maximum.
771 * If ints have 32 bits, then the maximum value for any timeout in
772 * 10ms ticks is 248 days.
775 tvtohz_high(struct timeval
*tv
)
792 kprintf("tvtohz_high: negative time difference %ld sec %ld usec\n",
796 } else if (sec
<= INT_MAX
/ hz
) {
797 ticks
= (int)(sec
* hz
+
798 ((u_long
)usec
+ (tick
- 1)) / tick
) + 1;
806 * Compute number of ticks for the specified amount of time, erroring on
807 * the side of it being too low to ensure that sleeping the returned number
808 * of ticks will not result in a late return.
810 * The supplied timeval may not be negative and should be normalized. A
811 * return value of 0 is possible if the timeval converts to less then
814 * If ints have 32 bits, then the maximum value for any timeout in
815 * 10ms ticks is 248 days.
818 tvtohz_low(struct timeval
*tv
)
824 if (sec
<= INT_MAX
/ hz
)
825 ticks
= (int)(sec
* hz
+ (u_long
)tv
->tv_usec
/ tick
);
833 * Start profiling on a process.
835 * Kernel profiling passes proc0 which never exits and hence
836 * keeps the profile clock running constantly.
839 startprofclock(struct proc
*p
)
841 if ((p
->p_flag
& P_PROFIL
) == 0) {
842 p
->p_flag
|= P_PROFIL
;
844 if (++profprocs
== 1 && stathz
!= 0) {
847 setstatclockrate(profhz
);
855 * Stop profiling on a process.
858 stopprofclock(struct proc
*p
)
860 if (p
->p_flag
& P_PROFIL
) {
861 p
->p_flag
&= ~P_PROFIL
;
863 if (--profprocs
== 0 && stathz
!= 0) {
866 setstatclockrate(stathz
);
874 * Return information about system clocks.
877 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS
)
879 struct kinfo_clockinfo clkinfo
;
881 * Construct clockinfo structure.
884 clkinfo
.ci_tick
= tick
;
885 clkinfo
.ci_tickadj
= ntp_default_tick_delta
/ 1000;
886 clkinfo
.ci_profhz
= profhz
;
887 clkinfo
.ci_stathz
= stathz
? stathz
: hz
;
888 return (sysctl_handle_opaque(oidp
, &clkinfo
, sizeof clkinfo
, req
));
891 SYSCTL_PROC(_kern
, KERN_CLOCKRATE
, clockrate
, CTLTYPE_STRUCT
|CTLFLAG_RD
,
892 0, 0, sysctl_kern_clockrate
, "S,clockinfo","");
895 * We have eight functions for looking at the clock, four for
896 * microseconds and four for nanoseconds. For each there is fast
897 * but less precise version "get{nano|micro}[up]time" which will
898 * return a time which is up to 1/HZ previous to the call, whereas
899 * the raw version "{nano|micro}[up]time" will return a timestamp
900 * which is as precise as possible. The "up" variants return the
901 * time relative to system boot, these are well suited for time
902 * interval measurements.
904 * Each cpu independantly maintains the current time of day, so all
905 * we need to do to protect ourselves from changes is to do a loop
906 * check on the seconds field changing out from under us.
908 * The system timer maintains a 32 bit count and due to various issues
909 * it is possible for the calculated delta to occassionally exceed
910 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
911 * multiplication can easily overflow, so we deal with the case. For
912 * uniformity we deal with the case in the usec case too.
914 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
917 getmicrouptime(struct timeval
*tvp
)
919 struct globaldata
*gd
= mycpu
;
923 tvp
->tv_sec
= gd
->gd_time_seconds
;
924 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
925 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
927 if (delta
>= sys_cputimer
->freq
) {
928 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
929 delta
%= sys_cputimer
->freq
;
931 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
932 if (tvp
->tv_usec
>= 1000000) {
933 tvp
->tv_usec
-= 1000000;
939 getnanouptime(struct timespec
*tsp
)
941 struct globaldata
*gd
= mycpu
;
945 tsp
->tv_sec
= gd
->gd_time_seconds
;
946 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
947 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
949 if (delta
>= sys_cputimer
->freq
) {
950 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
951 delta
%= sys_cputimer
->freq
;
953 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
957 microuptime(struct timeval
*tvp
)
959 struct globaldata
*gd
= mycpu
;
963 tvp
->tv_sec
= gd
->gd_time_seconds
;
964 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
965 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
967 if (delta
>= sys_cputimer
->freq
) {
968 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
969 delta
%= sys_cputimer
->freq
;
971 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
975 nanouptime(struct timespec
*tsp
)
977 struct globaldata
*gd
= mycpu
;
981 tsp
->tv_sec
= gd
->gd_time_seconds
;
982 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
983 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
985 if (delta
>= sys_cputimer
->freq
) {
986 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
987 delta
%= sys_cputimer
->freq
;
989 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
996 getmicrotime(struct timeval
*tvp
)
998 struct globaldata
*gd
= mycpu
;
1003 tvp
->tv_sec
= gd
->gd_time_seconds
;
1004 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1005 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1007 if (delta
>= sys_cputimer
->freq
) {
1008 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1009 delta
%= sys_cputimer
->freq
;
1011 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1013 bt
= &basetime
[basetime_index
];
1014 tvp
->tv_sec
+= bt
->tv_sec
;
1015 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1016 while (tvp
->tv_usec
>= 1000000) {
1017 tvp
->tv_usec
-= 1000000;
1023 getnanotime(struct timespec
*tsp
)
1025 struct globaldata
*gd
= mycpu
;
1026 struct timespec
*bt
;
1030 tsp
->tv_sec
= gd
->gd_time_seconds
;
1031 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1032 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1034 if (delta
>= sys_cputimer
->freq
) {
1035 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1036 delta
%= sys_cputimer
->freq
;
1038 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1040 bt
= &basetime
[basetime_index
];
1041 tsp
->tv_sec
+= bt
->tv_sec
;
1042 tsp
->tv_nsec
+= bt
->tv_nsec
;
1043 while (tsp
->tv_nsec
>= 1000000000) {
1044 tsp
->tv_nsec
-= 1000000000;
1050 getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
)
1052 struct globaldata
*gd
= mycpu
;
1056 tsp
->tv_sec
= gd
->gd_time_seconds
;
1057 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1058 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1060 if (delta
>= sys_cputimer
->freq
) {
1061 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1062 delta
%= sys_cputimer
->freq
;
1064 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1066 tsp
->tv_sec
+= nbt
->tv_sec
;
1067 tsp
->tv_nsec
+= nbt
->tv_nsec
;
1068 while (tsp
->tv_nsec
>= 1000000000) {
1069 tsp
->tv_nsec
-= 1000000000;
1076 microtime(struct timeval
*tvp
)
1078 struct globaldata
*gd
= mycpu
;
1079 struct timespec
*bt
;
1083 tvp
->tv_sec
= gd
->gd_time_seconds
;
1084 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1085 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1087 if (delta
>= sys_cputimer
->freq
) {
1088 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1089 delta
%= sys_cputimer
->freq
;
1091 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1093 bt
= &basetime
[basetime_index
];
1094 tvp
->tv_sec
+= bt
->tv_sec
;
1095 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1096 while (tvp
->tv_usec
>= 1000000) {
1097 tvp
->tv_usec
-= 1000000;
1103 nanotime(struct timespec
*tsp
)
1105 struct globaldata
*gd
= mycpu
;
1106 struct timespec
*bt
;
1110 tsp
->tv_sec
= gd
->gd_time_seconds
;
1111 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1112 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1114 if (delta
>= sys_cputimer
->freq
) {
1115 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1116 delta
%= sys_cputimer
->freq
;
1118 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1120 bt
= &basetime
[basetime_index
];
1121 tsp
->tv_sec
+= bt
->tv_sec
;
1122 tsp
->tv_nsec
+= bt
->tv_nsec
;
1123 while (tsp
->tv_nsec
>= 1000000000) {
1124 tsp
->tv_nsec
-= 1000000000;
1130 * note: this is not exactly synchronized with real time. To do that we
1131 * would have to do what microtime does and check for a nanoseconds overflow.
1134 get_approximate_time_t(void)
1136 struct globaldata
*gd
= mycpu
;
1137 struct timespec
*bt
;
1139 bt
= &basetime
[basetime_index
];
1140 return(gd
->gd_time_seconds
+ bt
->tv_sec
);
1144 pps_ioctl(u_long cmd
, caddr_t data
, struct pps_state
*pps
)
1147 struct pps_fetch_args
*fapi
;
1149 struct pps_kcbind_args
*kapi
;
1153 case PPS_IOC_CREATE
:
1155 case PPS_IOC_DESTROY
:
1157 case PPS_IOC_SETPARAMS
:
1158 app
= (pps_params_t
*)data
;
1159 if (app
->mode
& ~pps
->ppscap
)
1161 pps
->ppsparam
= *app
;
1163 case PPS_IOC_GETPARAMS
:
1164 app
= (pps_params_t
*)data
;
1165 *app
= pps
->ppsparam
;
1166 app
->api_version
= PPS_API_VERS_1
;
1168 case PPS_IOC_GETCAP
:
1169 *(int*)data
= pps
->ppscap
;
1172 fapi
= (struct pps_fetch_args
*)data
;
1173 if (fapi
->tsformat
&& fapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1175 if (fapi
->timeout
.tv_sec
|| fapi
->timeout
.tv_nsec
)
1176 return (EOPNOTSUPP
);
1177 pps
->ppsinfo
.current_mode
= pps
->ppsparam
.mode
;
1178 fapi
->pps_info_buf
= pps
->ppsinfo
;
1180 case PPS_IOC_KCBIND
:
1182 kapi
= (struct pps_kcbind_args
*)data
;
1183 /* XXX Only root should be able to do this */
1184 if (kapi
->tsformat
&& kapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1186 if (kapi
->kernel_consumer
!= PPS_KC_HARDPPS
)
1188 if (kapi
->edge
& ~pps
->ppscap
)
1190 pps
->kcmode
= kapi
->edge
;
1193 return (EOPNOTSUPP
);
1201 pps_init(struct pps_state
*pps
)
1203 pps
->ppscap
|= PPS_TSFMT_TSPEC
;
1204 if (pps
->ppscap
& PPS_CAPTUREASSERT
)
1205 pps
->ppscap
|= PPS_OFFSETASSERT
;
1206 if (pps
->ppscap
& PPS_CAPTURECLEAR
)
1207 pps
->ppscap
|= PPS_OFFSETCLEAR
;
1211 pps_event(struct pps_state
*pps
, sysclock_t count
, int event
)
1213 struct globaldata
*gd
;
1214 struct timespec
*tsp
;
1215 struct timespec
*osp
;
1216 struct timespec
*bt
;
1229 /* Things would be easier with arrays... */
1230 if (event
== PPS_CAPTUREASSERT
) {
1231 tsp
= &pps
->ppsinfo
.assert_timestamp
;
1232 osp
= &pps
->ppsparam
.assert_offset
;
1233 foff
= pps
->ppsparam
.mode
& PPS_OFFSETASSERT
;
1234 fhard
= pps
->kcmode
& PPS_CAPTUREASSERT
;
1235 pcount
= &pps
->ppscount
[0];
1236 pseq
= &pps
->ppsinfo
.assert_sequence
;
1238 tsp
= &pps
->ppsinfo
.clear_timestamp
;
1239 osp
= &pps
->ppsparam
.clear_offset
;
1240 foff
= pps
->ppsparam
.mode
& PPS_OFFSETCLEAR
;
1241 fhard
= pps
->kcmode
& PPS_CAPTURECLEAR
;
1242 pcount
= &pps
->ppscount
[1];
1243 pseq
= &pps
->ppsinfo
.clear_sequence
;
1246 /* Nothing really happened */
1247 if (*pcount
== count
)
1253 ts
.tv_sec
= gd
->gd_time_seconds
;
1254 delta
= count
- gd
->gd_cpuclock_base
;
1255 } while (ts
.tv_sec
!= gd
->gd_time_seconds
);
1257 if (delta
>= sys_cputimer
->freq
) {
1258 ts
.tv_sec
+= delta
/ sys_cputimer
->freq
;
1259 delta
%= sys_cputimer
->freq
;
1261 ts
.tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1262 bt
= &basetime
[basetime_index
];
1263 ts
.tv_sec
+= bt
->tv_sec
;
1264 ts
.tv_nsec
+= bt
->tv_nsec
;
1265 while (ts
.tv_nsec
>= 1000000000) {
1266 ts
.tv_nsec
-= 1000000000;
1274 timespecadd(tsp
, osp
);
1275 if (tsp
->tv_nsec
< 0) {
1276 tsp
->tv_nsec
+= 1000000000;
1282 /* magic, at its best... */
1283 tcount
= count
- pps
->ppscount
[2];
1284 pps
->ppscount
[2] = count
;
1285 if (tcount
>= sys_cputimer
->freq
) {
1286 delta
= (1000000000 * (tcount
/ sys_cputimer
->freq
) +
1287 sys_cputimer
->freq64_nsec
*
1288 (tcount
% sys_cputimer
->freq
)) >> 32;
1290 delta
= (sys_cputimer
->freq64_nsec
* tcount
) >> 32;
1292 hardpps(tsp
, delta
);
1298 * Return the tsc target value for a delay of (ns).
1300 * Returns -1 if the TSC is not supported.
1303 tsc_get_target(int ns
)
1305 #if defined(_RDTSC_SUPPORTED_)
1306 if (cpu_feature
& CPUID_TSC
) {
1307 return (rdtsc() + tsc_frequency
* ns
/ (int64_t)1000000000);
1314 * Compare the tsc against the passed target
1316 * Returns +1 if the target has been reached
1317 * Returns 0 if the target has not yet been reached
1318 * Returns -1 if the TSC is not supported.
1320 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1323 tsc_test_target(int64_t target
)
1325 #if defined(_RDTSC_SUPPORTED_)
1326 if (cpu_feature
& CPUID_TSC
) {
1327 if ((int64_t)(target
- rdtsc()) <= 0)