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. Neither the name of the University nor the names of its contributors
52 * may be used to endorse or promote products derived from this software
53 * without specific prior written permission.
55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
67 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
68 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $
72 #include "opt_ifpoll.h"
73 #include "opt_pctrack.h"
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/callout.h>
78 #include <sys/kernel.h>
79 #include <sys/kinfo.h>
81 #include <sys/malloc.h>
82 #include <sys/resource.h>
83 #include <sys/resourcevar.h>
84 #include <sys/signalvar.h>
85 #include <sys/timex.h>
86 #include <sys/timepps.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_extern.h>
92 #include <sys/sysctl.h>
94 #include <sys/thread2.h>
95 #include <sys/mplock2.h>
97 #include <machine/cpu.h>
98 #include <machine/limits.h>
99 #include <machine/smp.h>
100 #include <machine/cpufunc.h>
101 #include <machine/specialreg.h>
102 #include <machine/clock.h>
105 #include <sys/gmon.h>
109 extern void ifpoll_init_pcpu(int);
113 static void do_pctrack(struct intrframe
*frame
, int which
);
116 static void initclocks (void *dummy
);
117 SYSINIT(clocks
, SI_BOOT2_CLOCKS
, SI_ORDER_FIRST
, initclocks
, NULL
)
120 * Some of these don't belong here, but it's easiest to concentrate them.
121 * Note that cpu_time counts in microseconds, but most userland programs
122 * just compare relative times against the total by delta.
124 struct kinfo_cputime cputime_percpu
[MAXCPU
];
126 struct kinfo_pcheader cputime_pcheader
= { PCTRACK_SIZE
, PCTRACK_ARYSIZE
};
127 struct kinfo_pctrack cputime_pctrack
[MAXCPU
][PCTRACK_SIZE
];
131 sysctl_cputime(SYSCTL_HANDLER_ARGS
)
134 size_t size
= sizeof(struct kinfo_cputime
);
136 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
137 if ((error
= SYSCTL_OUT(req
, &cputime_percpu
[cpu
], size
)))
143 SYSCTL_PROC(_kern
, OID_AUTO
, cputime
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
144 sysctl_cputime
, "S,kinfo_cputime", "CPU time statistics");
147 sysctl_cp_time(SYSCTL_HANDLER_ARGS
)
149 long cpu_states
[5] = {0};
151 size_t size
= sizeof(cpu_states
);
153 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
154 cpu_states
[CP_USER
] += cputime_percpu
[cpu
].cp_user
;
155 cpu_states
[CP_NICE
] += cputime_percpu
[cpu
].cp_nice
;
156 cpu_states
[CP_SYS
] += cputime_percpu
[cpu
].cp_sys
;
157 cpu_states
[CP_INTR
] += cputime_percpu
[cpu
].cp_intr
;
158 cpu_states
[CP_IDLE
] += cputime_percpu
[cpu
].cp_idle
;
161 error
= SYSCTL_OUT(req
, cpu_states
, size
);
166 SYSCTL_PROC(_kern
, OID_AUTO
, cp_time
, (CTLTYPE_LONG
|CTLFLAG_RD
), 0, 0,
167 sysctl_cp_time
, "LU", "CPU time statistics");
170 * boottime is used to calculate the 'real' uptime. Do not confuse this with
171 * microuptime(). microtime() is not drift compensated. The real uptime
172 * with compensation is nanotime() - bootime. boottime is recalculated
173 * whenever the real time is set based on the compensated elapsed time
174 * in seconds (gd->gd_time_seconds).
176 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
177 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
180 struct timespec boottime
; /* boot time (realtime) for reference only */
181 time_t time_second
; /* read-only 'passive' uptime in seconds */
182 time_t time_uptime
; /* read-only 'passive' uptime in seconds */
185 * basetime is used to calculate the compensated real time of day. The
186 * basetime can be modified on a per-tick basis by the adjtime(),
187 * ntp_adjtime(), and sysctl-based time correction APIs.
189 * Note that frequency corrections can also be made by adjusting
192 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
193 * used on both SMP and UP systems to avoid MP races between cpu's and
194 * interrupt races on UP systems.
196 #define BASETIME_ARYSIZE 16
197 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
198 static struct timespec basetime
[BASETIME_ARYSIZE
];
199 static volatile int basetime_index
;
202 sysctl_get_basetime(SYSCTL_HANDLER_ARGS
)
209 * Because basetime data and index may be updated by another cpu,
210 * a load fence is required to ensure that the data we read has
211 * not been speculatively read relative to a possibly updated index.
213 index
= basetime_index
;
215 bt
= &basetime
[index
];
216 error
= SYSCTL_OUT(req
, bt
, sizeof(*bt
));
220 SYSCTL_STRUCT(_kern
, KERN_BOOTTIME
, boottime
, CTLFLAG_RD
,
221 &boottime
, timespec
, "System boottime");
222 SYSCTL_PROC(_kern
, OID_AUTO
, basetime
, CTLTYPE_STRUCT
|CTLFLAG_RD
, 0, 0,
223 sysctl_get_basetime
, "S,timespec", "System basetime");
225 static void hardclock(systimer_t info
, int, struct intrframe
*frame
);
226 static void statclock(systimer_t info
, int, struct intrframe
*frame
);
227 static void schedclock(systimer_t info
, int, struct intrframe
*frame
);
228 static void getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
);
230 int ticks
; /* system master ticks at hz */
231 int clocks_running
; /* tsleep/timeout clocks operational */
232 int64_t nsec_adj
; /* ntpd per-tick adjustment in nsec << 32 */
233 int64_t nsec_acc
; /* accumulator */
234 int sched_ticks
; /* global schedule clock ticks */
236 /* NTPD time correction fields */
237 int64_t ntp_tick_permanent
; /* per-tick adjustment in nsec << 32 */
238 int64_t ntp_tick_acc
; /* accumulator for per-tick adjustment */
239 int64_t ntp_delta
; /* one-time correction in nsec */
240 int64_t ntp_big_delta
= 1000000000;
241 int32_t ntp_tick_delta
; /* current adjustment rate */
242 int32_t ntp_default_tick_delta
; /* adjustment rate for ntp_delta */
243 time_t ntp_leap_second
; /* time of next leap second */
244 int ntp_leap_insert
; /* whether to insert or remove a second */
247 * Finish initializing clock frequencies and start all clocks running.
251 initclocks(void *dummy
)
253 /*psratio = profhz / stathz;*/
259 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
260 * during SMP initialization.
262 * This routine is called concurrently during low-level SMP initialization
263 * and may not block in any way. Meaning, among other things, we can't
264 * acquire any tokens.
267 initclocks_pcpu(void)
269 struct globaldata
*gd
= mycpu
;
272 if (gd
->gd_cpuid
== 0) {
273 gd
->gd_time_seconds
= 1;
274 gd
->gd_cpuclock_base
= sys_cputimer
->count();
277 gd
->gd_time_seconds
= globaldata_find(0)->gd_time_seconds
;
278 gd
->gd_cpuclock_base
= globaldata_find(0)->gd_cpuclock_base
;
281 systimer_intr_enable();
287 * This routine is called on just the BSP, just after SMP initialization
288 * completes to * finish initializing any clocks that might contend/block
289 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
290 * that function is called from the idle thread bootstrap for each cpu and
291 * not allowed to block at all.
295 initclocks_other(void *dummy
)
297 struct globaldata
*ogd
= mycpu
;
298 struct globaldata
*gd
;
301 for (n
= 0; n
< ncpus
; ++n
) {
302 lwkt_setcpu_self(globaldata_find(n
));
306 * Use a non-queued periodic systimer to prevent multiple
307 * ticks from building up if the sysclock jumps forward
308 * (8254 gets reset). The sysclock will never jump backwards.
309 * Our time sync is based on the actual sysclock, not the
312 systimer_init_periodic_nq(&gd
->gd_hardclock
, hardclock
,
314 systimer_init_periodic_nq(&gd
->gd_statclock
, statclock
,
316 /* XXX correct the frequency for scheduler / estcpu tests */
317 systimer_init_periodic_nq(&gd
->gd_schedclock
, schedclock
,
320 ifpoll_init_pcpu(gd
->gd_cpuid
);
323 lwkt_setcpu_self(ogd
);
325 SYSINIT(clocks2
, SI_BOOT2_POST_SMP
, SI_ORDER_ANY
, initclocks_other
, NULL
)
328 * This sets the current real time of day. Timespecs are in seconds and
329 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
330 * instead we adjust basetime so basetime + gd_* results in the current
331 * time of day. This way the gd_* fields are guarenteed to represent
332 * a monotonically increasing 'uptime' value.
334 * When set_timeofday() is called from userland, the system call forces it
335 * onto cpu #0 since only cpu #0 can update basetime_index.
338 set_timeofday(struct timespec
*ts
)
340 struct timespec
*nbt
;
344 * XXX SMP / non-atomic basetime updates
347 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
350 nbt
->tv_sec
= ts
->tv_sec
- nbt
->tv_sec
;
351 nbt
->tv_nsec
= ts
->tv_nsec
- nbt
->tv_nsec
;
352 if (nbt
->tv_nsec
< 0) {
353 nbt
->tv_nsec
+= 1000000000;
358 * Note that basetime diverges from boottime as the clock drift is
359 * compensated for, so we cannot do away with boottime. When setting
360 * the absolute time of day the drift is 0 (for an instant) and we
361 * can simply assign boottime to basetime.
363 * Note that nanouptime() is based on gd_time_seconds which is drift
364 * compensated up to a point (it is guarenteed to remain monotonically
365 * increasing). gd_time_seconds is thus our best uptime guess and
366 * suitable for use in the boottime calculation. It is already taken
367 * into account in the basetime calculation above.
369 boottime
.tv_sec
= nbt
->tv_sec
;
373 * We now have a new basetime, make sure all other cpus have it,
374 * then update the index.
383 * Each cpu has its own hardclock, but we only increments ticks and softticks
386 * NOTE! systimer! the MP lock might not be held here. We can only safely
387 * manipulate objects owned by the current cpu.
390 hardclock(systimer_t info
, int in_ipi __unused
, struct intrframe
*frame
)
394 struct globaldata
*gd
= mycpu
;
397 * Realtime updates are per-cpu. Note that timer corrections as
398 * returned by microtime() and friends make an additional adjustment
399 * using a system-wise 'basetime', but the running time is always
400 * taken from the per-cpu globaldata area. Since the same clock
401 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
404 * Note that we never allow info->time (aka gd->gd_hardclock.time)
405 * to reverse index gd_cpuclock_base, but that it is possible for
406 * it to temporarily get behind in the seconds if something in the
407 * system locks interrupts for a long period of time. Since periodic
408 * timers count events, though everything should resynch again
411 cputicks
= info
->time
- gd
->gd_cpuclock_base
;
412 if (cputicks
>= sys_cputimer
->freq
) {
413 ++gd
->gd_time_seconds
;
414 gd
->gd_cpuclock_base
+= sys_cputimer
->freq
;
415 if (gd
->gd_cpuid
== 0)
416 ++time_uptime
; /* uncorrected monotonic 1-sec gran */
420 * The system-wide ticks counter and NTP related timedelta/tickdelta
421 * adjustments only occur on cpu #0. NTP adjustments are accomplished
422 * by updating basetime.
424 if (gd
->gd_cpuid
== 0) {
425 struct timespec
*nbt
;
433 if (tco
->tc_poll_pps
)
434 tco
->tc_poll_pps(tco
);
438 * Calculate the new basetime index. We are in a critical section
439 * on cpu #0 and can safely play with basetime_index. Start
440 * with the current basetime and then make adjustments.
442 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
444 *nbt
= basetime
[basetime_index
];
447 * Apply adjtime corrections. (adjtime() API)
449 * adjtime() only runs on cpu #0 so our critical section is
450 * sufficient to access these variables.
452 if (ntp_delta
!= 0) {
453 nbt
->tv_nsec
+= ntp_tick_delta
;
454 ntp_delta
-= ntp_tick_delta
;
455 if ((ntp_delta
> 0 && ntp_delta
< ntp_tick_delta
) ||
456 (ntp_delta
< 0 && ntp_delta
> ntp_tick_delta
)) {
457 ntp_tick_delta
= ntp_delta
;
462 * Apply permanent frequency corrections. (sysctl API)
464 if (ntp_tick_permanent
!= 0) {
465 ntp_tick_acc
+= ntp_tick_permanent
;
466 if (ntp_tick_acc
>= (1LL << 32)) {
467 nbt
->tv_nsec
+= ntp_tick_acc
>> 32;
468 ntp_tick_acc
-= (ntp_tick_acc
>> 32) << 32;
469 } else if (ntp_tick_acc
<= -(1LL << 32)) {
470 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
471 nbt
->tv_nsec
-= (-ntp_tick_acc
) >> 32;
472 ntp_tick_acc
+= ((-ntp_tick_acc
) >> 32) << 32;
476 if (nbt
->tv_nsec
>= 1000000000) {
478 nbt
->tv_nsec
-= 1000000000;
479 } else if (nbt
->tv_nsec
< 0) {
481 nbt
->tv_nsec
+= 1000000000;
485 * Another per-tick compensation. (for ntp_adjtime() API)
488 nsec_acc
+= nsec_adj
;
489 if (nsec_acc
>= 0x100000000LL
) {
490 nbt
->tv_nsec
+= nsec_acc
>> 32;
491 nsec_acc
= (nsec_acc
& 0xFFFFFFFFLL
);
492 } else if (nsec_acc
<= -0x100000000LL
) {
493 nbt
->tv_nsec
-= -nsec_acc
>> 32;
494 nsec_acc
= -(-nsec_acc
& 0xFFFFFFFFLL
);
496 if (nbt
->tv_nsec
>= 1000000000) {
497 nbt
->tv_nsec
-= 1000000000;
499 } else if (nbt
->tv_nsec
< 0) {
500 nbt
->tv_nsec
+= 1000000000;
505 /************************************************************
506 * LEAP SECOND CORRECTION *
507 ************************************************************
509 * Taking into account all the corrections made above, figure
510 * out the new real time. If the seconds field has changed
511 * then apply any pending leap-second corrections.
513 getnanotime_nbt(nbt
, &nts
);
515 if (time_second
!= nts
.tv_sec
) {
517 * Apply leap second (sysctl API). Adjust nts for changes
518 * so we do not have to call getnanotime_nbt again.
520 if (ntp_leap_second
) {
521 if (ntp_leap_second
== nts
.tv_sec
) {
522 if (ntp_leap_insert
) {
534 * Apply leap second (ntp_adjtime() API), calculate a new
535 * nsec_adj field. ntp_update_second() returns nsec_adj
536 * as a per-second value but we need it as a per-tick value.
538 leap
= ntp_update_second(time_second
, &nsec_adj
);
544 * Update the time_second 'approximate time' global.
546 time_second
= nts
.tv_sec
;
550 * Finally, our new basetime is ready to go live!
557 * lwkt thread scheduler fair queueing
559 lwkt_schedulerclock(curthread
);
562 * softticks are handled for all cpus
564 hardclock_softtick(gd
);
567 * ITimer handling is per-tick, per-cpu.
569 * We must acquire the per-process token in order for ksignal()
570 * to be non-blocking. For the moment this requires an AST fault,
571 * the ksignal() cannot be safely issued from this hard interrupt.
573 * XXX Even the trytoken here isn't right, and itimer operation in
574 * a multi threaded environment is going to be weird at the
577 if ((p
= curproc
) != NULL
&& lwkt_trytoken(&p
->p_token
)) {
579 if (frame
&& CLKF_USERMODE(frame
) &&
580 timevalisset(&p
->p_timer
[ITIMER_VIRTUAL
].it_value
) &&
581 itimerdecr(&p
->p_timer
[ITIMER_VIRTUAL
], ustick
) == 0) {
582 p
->p_flags
|= P_SIGVTALRM
;
585 if (timevalisset(&p
->p_timer
[ITIMER_PROF
].it_value
) &&
586 itimerdecr(&p
->p_timer
[ITIMER_PROF
], ustick
) == 0) {
587 p
->p_flags
|= P_SIGPROF
;
591 lwkt_reltoken(&p
->p_token
);
597 * The statistics clock typically runs at a 125Hz rate, and is intended
598 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
600 * NOTE! systimer! the MP lock might not be held here. We can only safely
601 * manipulate objects owned by the current cpu.
603 * The stats clock is responsible for grabbing a profiling sample.
604 * Most of the statistics are only used by user-level statistics programs.
605 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
608 * Like the other clocks, the stat clock is called from what is effectively
609 * a fast interrupt, so the context should be the thread/process that got
613 statclock(systimer_t info
, int in_ipi
, struct intrframe
*frame
)
626 * How big was our timeslice relative to the last time? Calculate
629 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
630 * during early boot. Just use the systimer count to be nice
631 * to e.g. qemu. The systimer has a better chance of being
632 * MPSAFE at early boot.
634 cv
= sys_cputimer
->count();
635 scv
= mycpu
->statint
.gd_statcv
;
639 bump
= (sys_cputimer
->freq64_usec
* (cv
- scv
)) >> 32;
645 mycpu
->statint
.gd_statcv
= cv
;
648 stv
= &mycpu
->gd_stattv
;
649 if (stv
->tv_sec
== 0) {
652 bump
= tv
.tv_usec
- stv
->tv_usec
+
653 (tv
.tv_sec
- stv
->tv_sec
) * 1000000;
665 if (frame
&& CLKF_USERMODE(frame
)) {
667 * Came from userland, handle user time and deal with
670 if (p
&& (p
->p_flags
& P_PROFIL
))
671 addupc_intr(p
, CLKF_PC(frame
), 1);
672 td
->td_uticks
+= bump
;
675 * Charge the time as appropriate
677 if (p
&& p
->p_nice
> NZERO
)
678 cpu_time
.cp_nice
+= bump
;
680 cpu_time
.cp_user
+= bump
;
682 int intr_nest
= mycpu
->gd_intr_nesting_level
;
686 * IPI processing code will bump gd_intr_nesting_level
687 * up by one, which breaks following CLKF_INTR testing,
688 * so we substract it by one here.
694 * Kernel statistics are just like addupc_intr, only easier.
697 if (g
->state
== GMON_PROF_ON
&& frame
) {
698 i
= CLKF_PC(frame
) - g
->lowpc
;
699 if (i
< g
->textsize
) {
700 i
/= HISTFRACTION
* sizeof(*g
->kcount
);
706 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
709 * Came from kernel mode, so we were:
710 * - handling an interrupt,
711 * - doing syscall or trap work on behalf of the current
713 * - spinning in the idle loop.
714 * Whichever it is, charge the time as appropriate.
715 * Note that we charge interrupts to the current process,
716 * regardless of whether they are ``for'' that process,
717 * so that we know how much of its real time was spent
718 * in ``non-process'' (i.e., interrupt) work.
720 * XXX assume system if frame is NULL. A NULL frame
721 * can occur if ipi processing is done from a crit_exit().
724 td
->td_iticks
+= bump
;
726 td
->td_sticks
+= bump
;
728 if (IS_INTR_RUNNING
) {
730 * If we interrupted an interrupt thread, well,
731 * count it as interrupt time.
735 do_pctrack(frame
, PCTRACK_INT
);
737 cpu_time
.cp_intr
+= bump
;
739 if (td
== &mycpu
->gd_idlethread
) {
741 * Even if the current thread is the idle
742 * thread it could be due to token contention
743 * in the LWKT scheduler. Count such as
746 if (mycpu
->gd_reqflags
& RQF_IDLECHECK_WK_MASK
)
747 cpu_time
.cp_sys
+= bump
;
749 cpu_time
.cp_idle
+= bump
;
752 * System thread was running.
756 do_pctrack(frame
, PCTRACK_SYS
);
758 cpu_time
.cp_sys
+= bump
;
762 #undef IS_INTR_RUNNING
768 * Sample the PC when in the kernel or in an interrupt. User code can
769 * retrieve the information and generate a histogram or other output.
773 do_pctrack(struct intrframe
*frame
, int which
)
775 struct kinfo_pctrack
*pctrack
;
777 pctrack
= &cputime_pctrack
[mycpu
->gd_cpuid
][which
];
778 pctrack
->pc_array
[pctrack
->pc_index
& PCTRACK_ARYMASK
] =
779 (void *)CLKF_PC(frame
);
784 sysctl_pctrack(SYSCTL_HANDLER_ARGS
)
786 struct kinfo_pcheader head
;
791 head
.pc_ntrack
= PCTRACK_SIZE
;
792 head
.pc_arysize
= PCTRACK_ARYSIZE
;
794 if ((error
= SYSCTL_OUT(req
, &head
, sizeof(head
))) != 0)
797 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
798 for (ntrack
= 0; ntrack
< PCTRACK_SIZE
; ++ntrack
) {
799 error
= SYSCTL_OUT(req
, &cputime_pctrack
[cpu
][ntrack
],
800 sizeof(struct kinfo_pctrack
));
809 SYSCTL_PROC(_kern
, OID_AUTO
, pctrack
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
810 sysctl_pctrack
, "S,kinfo_pcheader", "CPU PC tracking");
815 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
816 * the MP lock might not be held. We can safely manipulate parts of curproc
817 * but that's about it.
819 * Each cpu has its own scheduler clock.
822 schedclock(systimer_t info
, int in_ipi __unused
, struct intrframe
*frame
)
829 if ((lp
= lwkt_preempted_proc()) != NULL
) {
831 * Account for cpu time used and hit the scheduler. Note
832 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
836 usched_schedulerclock(lp
, info
->periodic
, info
->time
);
838 usched_schedulerclock(NULL
, info
->periodic
, info
->time
);
840 if ((lp
= curthread
->td_lwp
) != NULL
) {
842 * Update resource usage integrals and maximums.
844 if ((ru
= &lp
->lwp_proc
->p_ru
) &&
845 (vm
= lp
->lwp_proc
->p_vmspace
) != NULL
) {
846 ru
->ru_ixrss
+= pgtok(vm
->vm_tsize
);
847 ru
->ru_idrss
+= pgtok(vm
->vm_dsize
);
848 ru
->ru_isrss
+= pgtok(vm
->vm_ssize
);
849 if (lwkt_trytoken(&vm
->vm_map
.token
)) {
850 rss
= pgtok(vmspace_resident_count(vm
));
851 if (ru
->ru_maxrss
< rss
)
853 lwkt_reltoken(&vm
->vm_map
.token
);
857 /* Increment the global sched_ticks */
858 if (mycpu
->gd_cpuid
== 0)
863 * Compute number of ticks for the specified amount of time. The
864 * return value is intended to be used in a clock interrupt timed
865 * operation and guarenteed to meet or exceed the requested time.
866 * If the representation overflows, return INT_MAX. The minimum return
867 * value is 1 ticks and the function will average the calculation up.
868 * If any value greater then 0 microseconds is supplied, a value
869 * of at least 2 will be returned to ensure that a near-term clock
870 * interrupt does not cause the timeout to occur (degenerately) early.
872 * Note that limit checks must take into account microseconds, which is
873 * done simply by using the smaller signed long maximum instead of
874 * the unsigned long maximum.
876 * If ints have 32 bits, then the maximum value for any timeout in
877 * 10ms ticks is 248 days.
880 tvtohz_high(struct timeval
*tv
)
897 kprintf("tvtohz_high: negative time difference "
898 "%ld sec %ld usec\n",
902 } else if (sec
<= INT_MAX
/ hz
) {
903 ticks
= (int)(sec
* hz
+
904 ((u_long
)usec
+ (ustick
- 1)) / ustick
) + 1;
912 tstohz_high(struct timespec
*ts
)
929 kprintf("tstohz_high: negative time difference "
930 "%ld sec %ld nsec\n",
934 } else if (sec
<= INT_MAX
/ hz
) {
935 ticks
= (int)(sec
* hz
+
936 ((u_long
)nsec
+ (nstick
- 1)) / nstick
) + 1;
945 * Compute number of ticks for the specified amount of time, erroring on
946 * the side of it being too low to ensure that sleeping the returned number
947 * of ticks will not result in a late return.
949 * The supplied timeval may not be negative and should be normalized. A
950 * return value of 0 is possible if the timeval converts to less then
953 * If ints have 32 bits, then the maximum value for any timeout in
954 * 10ms ticks is 248 days.
957 tvtohz_low(struct timeval
*tv
)
963 if (sec
<= INT_MAX
/ hz
)
964 ticks
= (int)(sec
* hz
+ (u_long
)tv
->tv_usec
/ ustick
);
971 tstohz_low(struct timespec
*ts
)
977 if (sec
<= INT_MAX
/ hz
)
978 ticks
= (int)(sec
* hz
+ (u_long
)ts
->tv_nsec
/ nstick
);
985 * Start profiling on a process.
987 * Kernel profiling passes proc0 which never exits and hence
988 * keeps the profile clock running constantly.
991 startprofclock(struct proc
*p
)
993 if ((p
->p_flags
& P_PROFIL
) == 0) {
994 p
->p_flags
|= P_PROFIL
;
996 if (++profprocs
== 1 && stathz
!= 0) {
999 setstatclockrate(profhz
);
1007 * Stop profiling on a process.
1009 * caller must hold p->p_token
1012 stopprofclock(struct proc
*p
)
1014 if (p
->p_flags
& P_PROFIL
) {
1015 p
->p_flags
&= ~P_PROFIL
;
1017 if (--profprocs
== 0 && stathz
!= 0) {
1020 setstatclockrate(stathz
);
1028 * Return information about system clocks.
1031 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS
)
1033 struct kinfo_clockinfo clkinfo
;
1035 * Construct clockinfo structure.
1038 clkinfo
.ci_tick
= ustick
;
1039 clkinfo
.ci_tickadj
= ntp_default_tick_delta
/ 1000;
1040 clkinfo
.ci_profhz
= profhz
;
1041 clkinfo
.ci_stathz
= stathz
? stathz
: hz
;
1042 return (sysctl_handle_opaque(oidp
, &clkinfo
, sizeof clkinfo
, req
));
1045 SYSCTL_PROC(_kern
, KERN_CLOCKRATE
, clockrate
, CTLTYPE_STRUCT
|CTLFLAG_RD
,
1046 0, 0, sysctl_kern_clockrate
, "S,clockinfo","");
1049 * We have eight functions for looking at the clock, four for
1050 * microseconds and four for nanoseconds. For each there is fast
1051 * but less precise version "get{nano|micro}[up]time" which will
1052 * return a time which is up to 1/HZ previous to the call, whereas
1053 * the raw version "{nano|micro}[up]time" will return a timestamp
1054 * which is as precise as possible. The "up" variants return the
1055 * time relative to system boot, these are well suited for time
1056 * interval measurements.
1058 * Each cpu independantly maintains the current time of day, so all
1059 * we need to do to protect ourselves from changes is to do a loop
1060 * check on the seconds field changing out from under us.
1062 * The system timer maintains a 32 bit count and due to various issues
1063 * it is possible for the calculated delta to occassionally exceed
1064 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1065 * multiplication can easily overflow, so we deal with the case. For
1066 * uniformity we deal with the case in the usec case too.
1068 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1071 getmicrouptime(struct timeval
*tvp
)
1073 struct globaldata
*gd
= mycpu
;
1077 tvp
->tv_sec
= gd
->gd_time_seconds
;
1078 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1079 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1081 if (delta
>= sys_cputimer
->freq
) {
1082 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1083 delta
%= sys_cputimer
->freq
;
1085 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1086 if (tvp
->tv_usec
>= 1000000) {
1087 tvp
->tv_usec
-= 1000000;
1093 getnanouptime(struct timespec
*tsp
)
1095 struct globaldata
*gd
= mycpu
;
1099 tsp
->tv_sec
= gd
->gd_time_seconds
;
1100 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1101 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1103 if (delta
>= sys_cputimer
->freq
) {
1104 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1105 delta
%= sys_cputimer
->freq
;
1107 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1111 microuptime(struct timeval
*tvp
)
1113 struct globaldata
*gd
= mycpu
;
1117 tvp
->tv_sec
= gd
->gd_time_seconds
;
1118 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1119 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1121 if (delta
>= sys_cputimer
->freq
) {
1122 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1123 delta
%= sys_cputimer
->freq
;
1125 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1129 nanouptime(struct timespec
*tsp
)
1131 struct globaldata
*gd
= mycpu
;
1135 tsp
->tv_sec
= gd
->gd_time_seconds
;
1136 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1137 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1139 if (delta
>= sys_cputimer
->freq
) {
1140 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1141 delta
%= sys_cputimer
->freq
;
1143 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1150 getmicrotime(struct timeval
*tvp
)
1152 struct globaldata
*gd
= mycpu
;
1153 struct timespec
*bt
;
1157 tvp
->tv_sec
= gd
->gd_time_seconds
;
1158 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1159 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1161 if (delta
>= sys_cputimer
->freq
) {
1162 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1163 delta
%= sys_cputimer
->freq
;
1165 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1167 bt
= &basetime
[basetime_index
];
1168 tvp
->tv_sec
+= bt
->tv_sec
;
1169 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1170 while (tvp
->tv_usec
>= 1000000) {
1171 tvp
->tv_usec
-= 1000000;
1177 getnanotime(struct timespec
*tsp
)
1179 struct globaldata
*gd
= mycpu
;
1180 struct timespec
*bt
;
1184 tsp
->tv_sec
= gd
->gd_time_seconds
;
1185 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1186 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1188 if (delta
>= sys_cputimer
->freq
) {
1189 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1190 delta
%= sys_cputimer
->freq
;
1192 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1194 bt
= &basetime
[basetime_index
];
1195 tsp
->tv_sec
+= bt
->tv_sec
;
1196 tsp
->tv_nsec
+= bt
->tv_nsec
;
1197 while (tsp
->tv_nsec
>= 1000000000) {
1198 tsp
->tv_nsec
-= 1000000000;
1204 getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
)
1206 struct globaldata
*gd
= mycpu
;
1210 tsp
->tv_sec
= gd
->gd_time_seconds
;
1211 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1212 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1214 if (delta
>= sys_cputimer
->freq
) {
1215 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1216 delta
%= sys_cputimer
->freq
;
1218 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1220 tsp
->tv_sec
+= nbt
->tv_sec
;
1221 tsp
->tv_nsec
+= nbt
->tv_nsec
;
1222 while (tsp
->tv_nsec
>= 1000000000) {
1223 tsp
->tv_nsec
-= 1000000000;
1230 microtime(struct timeval
*tvp
)
1232 struct globaldata
*gd
= mycpu
;
1233 struct timespec
*bt
;
1237 tvp
->tv_sec
= gd
->gd_time_seconds
;
1238 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1239 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1241 if (delta
>= sys_cputimer
->freq
) {
1242 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1243 delta
%= sys_cputimer
->freq
;
1245 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1247 bt
= &basetime
[basetime_index
];
1248 tvp
->tv_sec
+= bt
->tv_sec
;
1249 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1250 while (tvp
->tv_usec
>= 1000000) {
1251 tvp
->tv_usec
-= 1000000;
1257 nanotime(struct timespec
*tsp
)
1259 struct globaldata
*gd
= mycpu
;
1260 struct timespec
*bt
;
1264 tsp
->tv_sec
= gd
->gd_time_seconds
;
1265 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1266 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1268 if (delta
>= sys_cputimer
->freq
) {
1269 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1270 delta
%= sys_cputimer
->freq
;
1272 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1274 bt
= &basetime
[basetime_index
];
1275 tsp
->tv_sec
+= bt
->tv_sec
;
1276 tsp
->tv_nsec
+= bt
->tv_nsec
;
1277 while (tsp
->tv_nsec
>= 1000000000) {
1278 tsp
->tv_nsec
-= 1000000000;
1284 * note: this is not exactly synchronized with real time. To do that we
1285 * would have to do what microtime does and check for a nanoseconds overflow.
1288 get_approximate_time_t(void)
1290 struct globaldata
*gd
= mycpu
;
1291 struct timespec
*bt
;
1293 bt
= &basetime
[basetime_index
];
1294 return(gd
->gd_time_seconds
+ bt
->tv_sec
);
1298 pps_ioctl(u_long cmd
, caddr_t data
, struct pps_state
*pps
)
1301 struct pps_fetch_args
*fapi
;
1303 struct pps_kcbind_args
*kapi
;
1307 case PPS_IOC_CREATE
:
1309 case PPS_IOC_DESTROY
:
1311 case PPS_IOC_SETPARAMS
:
1312 app
= (pps_params_t
*)data
;
1313 if (app
->mode
& ~pps
->ppscap
)
1315 pps
->ppsparam
= *app
;
1317 case PPS_IOC_GETPARAMS
:
1318 app
= (pps_params_t
*)data
;
1319 *app
= pps
->ppsparam
;
1320 app
->api_version
= PPS_API_VERS_1
;
1322 case PPS_IOC_GETCAP
:
1323 *(int*)data
= pps
->ppscap
;
1326 fapi
= (struct pps_fetch_args
*)data
;
1327 if (fapi
->tsformat
&& fapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1329 if (fapi
->timeout
.tv_sec
|| fapi
->timeout
.tv_nsec
)
1330 return (EOPNOTSUPP
);
1331 pps
->ppsinfo
.current_mode
= pps
->ppsparam
.mode
;
1332 fapi
->pps_info_buf
= pps
->ppsinfo
;
1334 case PPS_IOC_KCBIND
:
1336 kapi
= (struct pps_kcbind_args
*)data
;
1337 /* XXX Only root should be able to do this */
1338 if (kapi
->tsformat
&& kapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1340 if (kapi
->kernel_consumer
!= PPS_KC_HARDPPS
)
1342 if (kapi
->edge
& ~pps
->ppscap
)
1344 pps
->kcmode
= kapi
->edge
;
1347 return (EOPNOTSUPP
);
1355 pps_init(struct pps_state
*pps
)
1357 pps
->ppscap
|= PPS_TSFMT_TSPEC
;
1358 if (pps
->ppscap
& PPS_CAPTUREASSERT
)
1359 pps
->ppscap
|= PPS_OFFSETASSERT
;
1360 if (pps
->ppscap
& PPS_CAPTURECLEAR
)
1361 pps
->ppscap
|= PPS_OFFSETCLEAR
;
1365 pps_event(struct pps_state
*pps
, sysclock_t count
, int event
)
1367 struct globaldata
*gd
;
1368 struct timespec
*tsp
;
1369 struct timespec
*osp
;
1370 struct timespec
*bt
;
1383 /* Things would be easier with arrays... */
1384 if (event
== PPS_CAPTUREASSERT
) {
1385 tsp
= &pps
->ppsinfo
.assert_timestamp
;
1386 osp
= &pps
->ppsparam
.assert_offset
;
1387 foff
= pps
->ppsparam
.mode
& PPS_OFFSETASSERT
;
1388 fhard
= pps
->kcmode
& PPS_CAPTUREASSERT
;
1389 pcount
= &pps
->ppscount
[0];
1390 pseq
= &pps
->ppsinfo
.assert_sequence
;
1392 tsp
= &pps
->ppsinfo
.clear_timestamp
;
1393 osp
= &pps
->ppsparam
.clear_offset
;
1394 foff
= pps
->ppsparam
.mode
& PPS_OFFSETCLEAR
;
1395 fhard
= pps
->kcmode
& PPS_CAPTURECLEAR
;
1396 pcount
= &pps
->ppscount
[1];
1397 pseq
= &pps
->ppsinfo
.clear_sequence
;
1400 /* Nothing really happened */
1401 if (*pcount
== count
)
1407 ts
.tv_sec
= gd
->gd_time_seconds
;
1408 delta
= count
- gd
->gd_cpuclock_base
;
1409 } while (ts
.tv_sec
!= gd
->gd_time_seconds
);
1411 if (delta
>= sys_cputimer
->freq
) {
1412 ts
.tv_sec
+= delta
/ sys_cputimer
->freq
;
1413 delta
%= sys_cputimer
->freq
;
1415 ts
.tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1416 bt
= &basetime
[basetime_index
];
1417 ts
.tv_sec
+= bt
->tv_sec
;
1418 ts
.tv_nsec
+= bt
->tv_nsec
;
1419 while (ts
.tv_nsec
>= 1000000000) {
1420 ts
.tv_nsec
-= 1000000000;
1428 timespecadd(tsp
, osp
);
1429 if (tsp
->tv_nsec
< 0) {
1430 tsp
->tv_nsec
+= 1000000000;
1436 /* magic, at its best... */
1437 tcount
= count
- pps
->ppscount
[2];
1438 pps
->ppscount
[2] = count
;
1439 if (tcount
>= sys_cputimer
->freq
) {
1440 delta
= (1000000000 * (tcount
/ sys_cputimer
->freq
) +
1441 sys_cputimer
->freq64_nsec
*
1442 (tcount
% sys_cputimer
->freq
)) >> 32;
1444 delta
= (sys_cputimer
->freq64_nsec
* tcount
) >> 32;
1446 hardpps(tsp
, delta
);
1452 * Return the tsc target value for a delay of (ns).
1454 * Returns -1 if the TSC is not supported.
1457 tsc_get_target(int ns
)
1459 #if defined(_RDTSC_SUPPORTED_)
1460 if (cpu_feature
& CPUID_TSC
) {
1461 return (rdtsc() + tsc_frequency
* ns
/ (int64_t)1000000000);
1468 * Compare the tsc against the passed target
1470 * Returns +1 if the target has been reached
1471 * Returns 0 if the target has not yet been reached
1472 * Returns -1 if the TSC is not supported.
1474 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1477 tsc_test_target(int64_t target
)
1479 #if defined(_RDTSC_SUPPORTED_)
1480 if (cpu_feature
& CPUID_TSC
) {
1481 if ((int64_t)(target
- rdtsc()) <= 0)
1490 * Delay the specified number of nanoseconds using the tsc. This function
1491 * returns immediately if the TSC is not supported. At least one cpu_pause()
1499 clk
= tsc_get_target(ns
);
1501 while (tsc_test_target(clk
) == 0)