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>
87 #include <sys/upmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_extern.h>
93 #include <sys/sysctl.h>
95 #include <sys/thread2.h>
96 #include <sys/mplock2.h>
98 #include <machine/cpu.h>
99 #include <machine/limits.h>
100 #include <machine/smp.h>
101 #include <machine/cpufunc.h>
102 #include <machine/specialreg.h>
103 #include <machine/clock.h>
106 #include <sys/gmon.h>
110 extern void ifpoll_init_pcpu(int);
114 static void do_pctrack(struct intrframe
*frame
, int which
);
117 static void initclocks (void *dummy
);
118 SYSINIT(clocks
, SI_BOOT2_CLOCKS
, SI_ORDER_FIRST
, initclocks
, NULL
);
121 * Some of these don't belong here, but it's easiest to concentrate them.
122 * Note that cpu_time counts in microseconds, but most userland programs
123 * just compare relative times against the total by delta.
125 struct kinfo_cputime cputime_percpu
[MAXCPU
];
127 struct kinfo_pcheader cputime_pcheader
= { PCTRACK_SIZE
, PCTRACK_ARYSIZE
};
128 struct kinfo_pctrack cputime_pctrack
[MAXCPU
][PCTRACK_SIZE
];
132 sysctl_cputime(SYSCTL_HANDLER_ARGS
)
135 size_t size
= sizeof(struct kinfo_cputime
);
137 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
138 if ((error
= SYSCTL_OUT(req
, &cputime_percpu
[cpu
], size
)))
144 SYSCTL_PROC(_kern
, OID_AUTO
, cputime
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
145 sysctl_cputime
, "S,kinfo_cputime", "CPU time statistics");
148 sysctl_cp_time(SYSCTL_HANDLER_ARGS
)
150 long cpu_states
[5] = {0};
152 size_t size
= sizeof(cpu_states
);
154 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
155 cpu_states
[CP_USER
] += cputime_percpu
[cpu
].cp_user
;
156 cpu_states
[CP_NICE
] += cputime_percpu
[cpu
].cp_nice
;
157 cpu_states
[CP_SYS
] += cputime_percpu
[cpu
].cp_sys
;
158 cpu_states
[CP_INTR
] += cputime_percpu
[cpu
].cp_intr
;
159 cpu_states
[CP_IDLE
] += cputime_percpu
[cpu
].cp_idle
;
162 error
= SYSCTL_OUT(req
, cpu_states
, size
);
167 SYSCTL_PROC(_kern
, OID_AUTO
, cp_time
, (CTLTYPE_LONG
|CTLFLAG_RD
), 0, 0,
168 sysctl_cp_time
, "LU", "CPU time statistics");
171 * boottime is used to calculate the 'real' uptime. Do not confuse this with
172 * microuptime(). microtime() is not drift compensated. The real uptime
173 * with compensation is nanotime() - bootime. boottime is recalculated
174 * whenever the real time is set based on the compensated elapsed time
175 * in seconds (gd->gd_time_seconds).
177 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
178 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
181 * WARNING! time_second can backstep on time corrections. Also, unlike
182 * time second, time_uptime is not a "real" time_t (seconds
183 * since the Epoch) but seconds since booting.
185 struct timespec boottime
; /* boot time (realtime) for reference only */
186 time_t time_second
; /* read-only 'passive' realtime in seconds */
187 time_t time_uptime
; /* read-only 'passive' uptime in seconds */
190 * basetime is used to calculate the compensated real time of day. The
191 * basetime can be modified on a per-tick basis by the adjtime(),
192 * ntp_adjtime(), and sysctl-based time correction APIs.
194 * Note that frequency corrections can also be made by adjusting
197 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
198 * used on both SMP and UP systems to avoid MP races between cpu's and
199 * interrupt races on UP systems.
201 #define BASETIME_ARYSIZE 16
202 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
203 static struct timespec basetime
[BASETIME_ARYSIZE
];
204 static volatile int basetime_index
;
207 sysctl_get_basetime(SYSCTL_HANDLER_ARGS
)
214 * Because basetime data and index may be updated by another cpu,
215 * a load fence is required to ensure that the data we read has
216 * not been speculatively read relative to a possibly updated index.
218 index
= basetime_index
;
220 bt
= &basetime
[index
];
221 error
= SYSCTL_OUT(req
, bt
, sizeof(*bt
));
225 SYSCTL_STRUCT(_kern
, KERN_BOOTTIME
, boottime
, CTLFLAG_RD
,
226 &boottime
, timespec
, "System boottime");
227 SYSCTL_PROC(_kern
, OID_AUTO
, basetime
, CTLTYPE_STRUCT
|CTLFLAG_RD
, 0, 0,
228 sysctl_get_basetime
, "S,timespec", "System basetime");
230 static void hardclock(systimer_t info
, int, struct intrframe
*frame
);
231 static void statclock(systimer_t info
, int, struct intrframe
*frame
);
232 static void schedclock(systimer_t info
, int, struct intrframe
*frame
);
233 static void getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
);
235 int ticks
; /* system master ticks at hz */
236 int clocks_running
; /* tsleep/timeout clocks operational */
237 int64_t nsec_adj
; /* ntpd per-tick adjustment in nsec << 32 */
238 int64_t nsec_acc
; /* accumulator */
239 int sched_ticks
; /* global schedule clock ticks */
241 /* NTPD time correction fields */
242 int64_t ntp_tick_permanent
; /* per-tick adjustment in nsec << 32 */
243 int64_t ntp_tick_acc
; /* accumulator for per-tick adjustment */
244 int64_t ntp_delta
; /* one-time correction in nsec */
245 int64_t ntp_big_delta
= 1000000000;
246 int32_t ntp_tick_delta
; /* current adjustment rate */
247 int32_t ntp_default_tick_delta
; /* adjustment rate for ntp_delta */
248 time_t ntp_leap_second
; /* time of next leap second */
249 int ntp_leap_insert
; /* whether to insert or remove a second */
252 * Finish initializing clock frequencies and start all clocks running.
256 initclocks(void *dummy
)
258 /*psratio = profhz / stathz;*/
262 kpmap
->tsc_freq
= (uint64_t)tsc_frequency
;
263 kpmap
->tick_freq
= hz
;
268 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
269 * during SMP initialization.
271 * This routine is called concurrently during low-level SMP initialization
272 * and may not block in any way. Meaning, among other things, we can't
273 * acquire any tokens.
276 initclocks_pcpu(void)
278 struct globaldata
*gd
= mycpu
;
281 if (gd
->gd_cpuid
== 0) {
282 gd
->gd_time_seconds
= 1;
283 gd
->gd_cpuclock_base
= sys_cputimer
->count();
286 gd
->gd_time_seconds
= globaldata_find(0)->gd_time_seconds
;
287 gd
->gd_cpuclock_base
= globaldata_find(0)->gd_cpuclock_base
;
290 systimer_intr_enable();
296 * This routine is called on just the BSP, just after SMP initialization
297 * completes to * finish initializing any clocks that might contend/block
298 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
299 * that function is called from the idle thread bootstrap for each cpu and
300 * not allowed to block at all.
304 initclocks_other(void *dummy
)
306 struct globaldata
*ogd
= mycpu
;
307 struct globaldata
*gd
;
310 for (n
= 0; n
< ncpus
; ++n
) {
311 lwkt_setcpu_self(globaldata_find(n
));
315 * Use a non-queued periodic systimer to prevent multiple
316 * ticks from building up if the sysclock jumps forward
317 * (8254 gets reset). The sysclock will never jump backwards.
318 * Our time sync is based on the actual sysclock, not the
321 systimer_init_periodic_nq(&gd
->gd_hardclock
, hardclock
,
323 systimer_init_periodic_nq(&gd
->gd_statclock
, statclock
,
325 /* XXX correct the frequency for scheduler / estcpu tests */
326 systimer_init_periodic_nq(&gd
->gd_schedclock
, schedclock
,
329 ifpoll_init_pcpu(gd
->gd_cpuid
);
332 lwkt_setcpu_self(ogd
);
334 SYSINIT(clocks2
, SI_BOOT2_POST_SMP
, SI_ORDER_ANY
, initclocks_other
, NULL
);
337 * This sets the current real time of day. Timespecs are in seconds and
338 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
339 * instead we adjust basetime so basetime + gd_* results in the current
340 * time of day. This way the gd_* fields are guaranteed to represent
341 * a monotonically increasing 'uptime' value.
343 * When set_timeofday() is called from userland, the system call forces it
344 * onto cpu #0 since only cpu #0 can update basetime_index.
347 set_timeofday(struct timespec
*ts
)
349 struct timespec
*nbt
;
353 * XXX SMP / non-atomic basetime updates
356 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
359 nbt
->tv_sec
= ts
->tv_sec
- nbt
->tv_sec
;
360 nbt
->tv_nsec
= ts
->tv_nsec
- nbt
->tv_nsec
;
361 if (nbt
->tv_nsec
< 0) {
362 nbt
->tv_nsec
+= 1000000000;
367 * Note that basetime diverges from boottime as the clock drift is
368 * compensated for, so we cannot do away with boottime. When setting
369 * the absolute time of day the drift is 0 (for an instant) and we
370 * can simply assign boottime to basetime.
372 * Note that nanouptime() is based on gd_time_seconds which is drift
373 * compensated up to a point (it is guaranteed to remain monotonically
374 * increasing). gd_time_seconds is thus our best uptime guess and
375 * suitable for use in the boottime calculation. It is already taken
376 * into account in the basetime calculation above.
378 boottime
.tv_sec
= nbt
->tv_sec
;
382 * We now have a new basetime, make sure all other cpus have it,
383 * then update the index.
392 * Each cpu has its own hardclock, but we only increments ticks and softticks
395 * NOTE! systimer! the MP lock might not be held here. We can only safely
396 * manipulate objects owned by the current cpu.
399 hardclock(systimer_t info
, int in_ipi
, struct intrframe
*frame
)
403 struct globaldata
*gd
= mycpu
;
405 if ((gd
->gd_reqflags
& RQF_IPIQ
) == 0 && lwkt_need_ipiq_process(gd
)) {
406 /* Defer to doreti on passive IPIQ processing */
411 * Realtime updates are per-cpu. Note that timer corrections as
412 * returned by microtime() and friends make an additional adjustment
413 * using a system-wise 'basetime', but the running time is always
414 * taken from the per-cpu globaldata area. Since the same clock
415 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
418 * Note that we never allow info->time (aka gd->gd_hardclock.time)
419 * to reverse index gd_cpuclock_base, but that it is possible for
420 * it to temporarily get behind in the seconds if something in the
421 * system locks interrupts for a long period of time. Since periodic
422 * timers count events, though everything should resynch again
425 cputicks
= info
->time
- gd
->gd_cpuclock_base
;
426 if (cputicks
>= sys_cputimer
->freq
) {
427 ++gd
->gd_time_seconds
;
428 gd
->gd_cpuclock_base
+= sys_cputimer
->freq
;
429 if (gd
->gd_cpuid
== 0)
430 ++time_uptime
; /* uncorrected monotonic 1-sec gran */
434 * The system-wide ticks counter and NTP related timedelta/tickdelta
435 * adjustments only occur on cpu #0. NTP adjustments are accomplished
436 * by updating basetime.
438 if (gd
->gd_cpuid
== 0) {
439 struct timespec
*nbt
;
447 if (tco
->tc_poll_pps
)
448 tco
->tc_poll_pps(tco
);
452 * Calculate the new basetime index. We are in a critical section
453 * on cpu #0 and can safely play with basetime_index. Start
454 * with the current basetime and then make adjustments.
456 ni
= (basetime_index
+ 1) & BASETIME_ARYMASK
;
458 *nbt
= basetime
[basetime_index
];
461 * Apply adjtime corrections. (adjtime() API)
463 * adjtime() only runs on cpu #0 so our critical section is
464 * sufficient to access these variables.
466 if (ntp_delta
!= 0) {
467 nbt
->tv_nsec
+= ntp_tick_delta
;
468 ntp_delta
-= ntp_tick_delta
;
469 if ((ntp_delta
> 0 && ntp_delta
< ntp_tick_delta
) ||
470 (ntp_delta
< 0 && ntp_delta
> ntp_tick_delta
)) {
471 ntp_tick_delta
= ntp_delta
;
476 * Apply permanent frequency corrections. (sysctl API)
478 if (ntp_tick_permanent
!= 0) {
479 ntp_tick_acc
+= ntp_tick_permanent
;
480 if (ntp_tick_acc
>= (1LL << 32)) {
481 nbt
->tv_nsec
+= ntp_tick_acc
>> 32;
482 ntp_tick_acc
-= (ntp_tick_acc
>> 32) << 32;
483 } else if (ntp_tick_acc
<= -(1LL << 32)) {
484 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
485 nbt
->tv_nsec
-= (-ntp_tick_acc
) >> 32;
486 ntp_tick_acc
+= ((-ntp_tick_acc
) >> 32) << 32;
490 if (nbt
->tv_nsec
>= 1000000000) {
492 nbt
->tv_nsec
-= 1000000000;
493 } else if (nbt
->tv_nsec
< 0) {
495 nbt
->tv_nsec
+= 1000000000;
499 * Another per-tick compensation. (for ntp_adjtime() API)
502 nsec_acc
+= nsec_adj
;
503 if (nsec_acc
>= 0x100000000LL
) {
504 nbt
->tv_nsec
+= nsec_acc
>> 32;
505 nsec_acc
= (nsec_acc
& 0xFFFFFFFFLL
);
506 } else if (nsec_acc
<= -0x100000000LL
) {
507 nbt
->tv_nsec
-= -nsec_acc
>> 32;
508 nsec_acc
= -(-nsec_acc
& 0xFFFFFFFFLL
);
510 if (nbt
->tv_nsec
>= 1000000000) {
511 nbt
->tv_nsec
-= 1000000000;
513 } else if (nbt
->tv_nsec
< 0) {
514 nbt
->tv_nsec
+= 1000000000;
519 /************************************************************
520 * LEAP SECOND CORRECTION *
521 ************************************************************
523 * Taking into account all the corrections made above, figure
524 * out the new real time. If the seconds field has changed
525 * then apply any pending leap-second corrections.
527 getnanotime_nbt(nbt
, &nts
);
529 if (time_second
!= nts
.tv_sec
) {
531 * Apply leap second (sysctl API). Adjust nts for changes
532 * so we do not have to call getnanotime_nbt again.
534 if (ntp_leap_second
) {
535 if (ntp_leap_second
== nts
.tv_sec
) {
536 if (ntp_leap_insert
) {
548 * Apply leap second (ntp_adjtime() API), calculate a new
549 * nsec_adj field. ntp_update_second() returns nsec_adj
550 * as a per-second value but we need it as a per-tick value.
552 leap
= ntp_update_second(time_second
, &nsec_adj
);
558 * Update the time_second 'approximate time' global.
560 time_second
= nts
.tv_sec
;
564 * Finally, our new basetime is ready to go live!
570 * Update kpmap on each tick. TS updates are integrated with
571 * fences and upticks allowing userland to read the data
577 w
= (kpmap
->upticks
+ 1) & 1;
578 getnanouptime(&kpmap
->ts_uptime
[w
]);
579 getnanotime(&kpmap
->ts_realtime
[w
]);
587 * lwkt thread scheduler fair queueing
589 lwkt_schedulerclock(curthread
);
592 * softticks are handled for all cpus
594 hardclock_softtick(gd
);
597 * ITimer handling is per-tick, per-cpu.
599 * We must acquire the per-process token in order for ksignal()
600 * to be non-blocking. For the moment this requires an AST fault,
601 * the ksignal() cannot be safely issued from this hard interrupt.
603 * XXX Even the trytoken here isn't right, and itimer operation in
604 * a multi threaded environment is going to be weird at the
607 if ((p
= curproc
) != NULL
&& lwkt_trytoken(&p
->p_token
)) {
610 ++p
->p_upmap
->runticks
;
612 if (frame
&& CLKF_USERMODE(frame
) &&
613 timevalisset(&p
->p_timer
[ITIMER_VIRTUAL
].it_value
) &&
614 itimerdecr(&p
->p_timer
[ITIMER_VIRTUAL
], ustick
) == 0) {
615 p
->p_flags
|= P_SIGVTALRM
;
618 if (timevalisset(&p
->p_timer
[ITIMER_PROF
].it_value
) &&
619 itimerdecr(&p
->p_timer
[ITIMER_PROF
], ustick
) == 0) {
620 p
->p_flags
|= P_SIGPROF
;
624 lwkt_reltoken(&p
->p_token
);
630 * The statistics clock typically runs at a 125Hz rate, and is intended
631 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
633 * NOTE! systimer! the MP lock might not be held here. We can only safely
634 * manipulate objects owned by the current cpu.
636 * The stats clock is responsible for grabbing a profiling sample.
637 * Most of the statistics are only used by user-level statistics programs.
638 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
641 * Like the other clocks, the stat clock is called from what is effectively
642 * a fast interrupt, so the context should be the thread/process that got
646 statclock(systimer_t info
, int in_ipi
, struct intrframe
*frame
)
659 * How big was our timeslice relative to the last time? Calculate
662 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
663 * during early boot. Just use the systimer count to be nice
664 * to e.g. qemu. The systimer has a better chance of being
665 * MPSAFE at early boot.
667 cv
= sys_cputimer
->count();
668 scv
= mycpu
->statint
.gd_statcv
;
672 bump
= (sys_cputimer
->freq64_usec
* (cv
- scv
)) >> 32;
678 mycpu
->statint
.gd_statcv
= cv
;
681 stv
= &mycpu
->gd_stattv
;
682 if (stv
->tv_sec
== 0) {
685 bump
= tv
.tv_usec
- stv
->tv_usec
+
686 (tv
.tv_sec
- stv
->tv_sec
) * 1000000;
698 if (frame
&& CLKF_USERMODE(frame
)) {
700 * Came from userland, handle user time and deal with
703 if (p
&& (p
->p_flags
& P_PROFIL
))
704 addupc_intr(p
, CLKF_PC(frame
), 1);
705 td
->td_uticks
+= bump
;
708 * Charge the time as appropriate
710 if (p
&& p
->p_nice
> NZERO
)
711 cpu_time
.cp_nice
+= bump
;
713 cpu_time
.cp_user
+= bump
;
715 int intr_nest
= mycpu
->gd_intr_nesting_level
;
719 * IPI processing code will bump gd_intr_nesting_level
720 * up by one, which breaks following CLKF_INTR testing,
721 * so we subtract it by one here.
727 * Kernel statistics are just like addupc_intr, only easier.
730 if (g
->state
== GMON_PROF_ON
&& frame
) {
731 i
= CLKF_PC(frame
) - g
->lowpc
;
732 if (i
< g
->textsize
) {
733 i
/= HISTFRACTION
* sizeof(*g
->kcount
);
739 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
742 * Came from kernel mode, so we were:
743 * - handling an interrupt,
744 * - doing syscall or trap work on behalf of the current
746 * - spinning in the idle loop.
747 * Whichever it is, charge the time as appropriate.
748 * Note that we charge interrupts to the current process,
749 * regardless of whether they are ``for'' that process,
750 * so that we know how much of its real time was spent
751 * in ``non-process'' (i.e., interrupt) work.
753 * XXX assume system if frame is NULL. A NULL frame
754 * can occur if ipi processing is done from a crit_exit().
757 td
->td_iticks
+= bump
;
759 td
->td_sticks
+= bump
;
761 if (IS_INTR_RUNNING
) {
763 * If we interrupted an interrupt thread, well,
764 * count it as interrupt time.
768 do_pctrack(frame
, PCTRACK_INT
);
770 cpu_time
.cp_intr
+= bump
;
772 if (td
== &mycpu
->gd_idlethread
) {
774 * Even if the current thread is the idle
775 * thread it could be due to token contention
776 * in the LWKT scheduler. Count such as
779 if (mycpu
->gd_reqflags
& RQF_IDLECHECK_WK_MASK
)
780 cpu_time
.cp_sys
+= bump
;
782 cpu_time
.cp_idle
+= bump
;
785 * System thread was running.
789 do_pctrack(frame
, PCTRACK_SYS
);
791 cpu_time
.cp_sys
+= bump
;
795 #undef IS_INTR_RUNNING
801 * Sample the PC when in the kernel or in an interrupt. User code can
802 * retrieve the information and generate a histogram or other output.
806 do_pctrack(struct intrframe
*frame
, int which
)
808 struct kinfo_pctrack
*pctrack
;
810 pctrack
= &cputime_pctrack
[mycpu
->gd_cpuid
][which
];
811 pctrack
->pc_array
[pctrack
->pc_index
& PCTRACK_ARYMASK
] =
812 (void *)CLKF_PC(frame
);
817 sysctl_pctrack(SYSCTL_HANDLER_ARGS
)
819 struct kinfo_pcheader head
;
824 head
.pc_ntrack
= PCTRACK_SIZE
;
825 head
.pc_arysize
= PCTRACK_ARYSIZE
;
827 if ((error
= SYSCTL_OUT(req
, &head
, sizeof(head
))) != 0)
830 for (cpu
= 0; cpu
< ncpus
; ++cpu
) {
831 for (ntrack
= 0; ntrack
< PCTRACK_SIZE
; ++ntrack
) {
832 error
= SYSCTL_OUT(req
, &cputime_pctrack
[cpu
][ntrack
],
833 sizeof(struct kinfo_pctrack
));
842 SYSCTL_PROC(_kern
, OID_AUTO
, pctrack
, (CTLTYPE_OPAQUE
|CTLFLAG_RD
), 0, 0,
843 sysctl_pctrack
, "S,kinfo_pcheader", "CPU PC tracking");
848 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
849 * the MP lock might not be held. We can safely manipulate parts of curproc
850 * but that's about it.
852 * Each cpu has its own scheduler clock.
855 schedclock(systimer_t info
, int in_ipi __unused
, struct intrframe
*frame
)
862 if ((lp
= lwkt_preempted_proc()) != NULL
) {
864 * Account for cpu time used and hit the scheduler. Note
865 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
869 usched_schedulerclock(lp
, info
->periodic
, info
->time
);
871 usched_schedulerclock(NULL
, info
->periodic
, info
->time
);
873 if ((lp
= curthread
->td_lwp
) != NULL
) {
875 * Update resource usage integrals and maximums.
877 if ((ru
= &lp
->lwp_proc
->p_ru
) &&
878 (vm
= lp
->lwp_proc
->p_vmspace
) != NULL
) {
879 ru
->ru_ixrss
+= pgtok(vm
->vm_tsize
);
880 ru
->ru_idrss
+= pgtok(vm
->vm_dsize
);
881 ru
->ru_isrss
+= pgtok(vm
->vm_ssize
);
882 if (lwkt_trytoken(&vm
->vm_map
.token
)) {
883 rss
= pgtok(vmspace_resident_count(vm
));
884 if (ru
->ru_maxrss
< rss
)
886 lwkt_reltoken(&vm
->vm_map
.token
);
890 /* Increment the global sched_ticks */
891 if (mycpu
->gd_cpuid
== 0)
896 * Compute number of ticks for the specified amount of time. The
897 * return value is intended to be used in a clock interrupt timed
898 * operation and guaranteed to meet or exceed the requested time.
899 * If the representation overflows, return INT_MAX. The minimum return
900 * value is 1 ticks and the function will average the calculation up.
901 * If any value greater then 0 microseconds is supplied, a value
902 * of at least 2 will be returned to ensure that a near-term clock
903 * interrupt does not cause the timeout to occur (degenerately) early.
905 * Note that limit checks must take into account microseconds, which is
906 * done simply by using the smaller signed long maximum instead of
907 * the unsigned long maximum.
909 * If ints have 32 bits, then the maximum value for any timeout in
910 * 10ms ticks is 248 days.
913 tvtohz_high(struct timeval
*tv
)
930 kprintf("tvtohz_high: negative time difference "
931 "%ld sec %ld usec\n",
935 } else if (sec
<= INT_MAX
/ hz
) {
936 ticks
= (int)(sec
* hz
+
937 ((u_long
)usec
+ (ustick
- 1)) / ustick
) + 1;
945 tstohz_high(struct timespec
*ts
)
962 kprintf("tstohz_high: negative time difference "
963 "%ld sec %ld nsec\n",
967 } else if (sec
<= INT_MAX
/ hz
) {
968 ticks
= (int)(sec
* hz
+
969 ((u_long
)nsec
+ (nstick
- 1)) / nstick
) + 1;
978 * Compute number of ticks for the specified amount of time, erroring on
979 * the side of it being too low to ensure that sleeping the returned number
980 * of ticks will not result in a late return.
982 * The supplied timeval may not be negative and should be normalized. A
983 * return value of 0 is possible if the timeval converts to less then
986 * If ints have 32 bits, then the maximum value for any timeout in
987 * 10ms ticks is 248 days.
990 tvtohz_low(struct timeval
*tv
)
996 if (sec
<= INT_MAX
/ hz
)
997 ticks
= (int)(sec
* hz
+ (u_long
)tv
->tv_usec
/ ustick
);
1004 tstohz_low(struct timespec
*ts
)
1010 if (sec
<= INT_MAX
/ hz
)
1011 ticks
= (int)(sec
* hz
+ (u_long
)ts
->tv_nsec
/ nstick
);
1018 * Start profiling on a process.
1020 * Kernel profiling passes proc0 which never exits and hence
1021 * keeps the profile clock running constantly.
1024 startprofclock(struct proc
*p
)
1026 if ((p
->p_flags
& P_PROFIL
) == 0) {
1027 p
->p_flags
|= P_PROFIL
;
1029 if (++profprocs
== 1 && stathz
!= 0) {
1032 setstatclockrate(profhz
);
1040 * Stop profiling on a process.
1042 * caller must hold p->p_token
1045 stopprofclock(struct proc
*p
)
1047 if (p
->p_flags
& P_PROFIL
) {
1048 p
->p_flags
&= ~P_PROFIL
;
1050 if (--profprocs
== 0 && stathz
!= 0) {
1053 setstatclockrate(stathz
);
1061 * Return information about system clocks.
1064 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS
)
1066 struct kinfo_clockinfo clkinfo
;
1068 * Construct clockinfo structure.
1071 clkinfo
.ci_tick
= ustick
;
1072 clkinfo
.ci_tickadj
= ntp_default_tick_delta
/ 1000;
1073 clkinfo
.ci_profhz
= profhz
;
1074 clkinfo
.ci_stathz
= stathz
? stathz
: hz
;
1075 return (sysctl_handle_opaque(oidp
, &clkinfo
, sizeof clkinfo
, req
));
1078 SYSCTL_PROC(_kern
, KERN_CLOCKRATE
, clockrate
, CTLTYPE_STRUCT
|CTLFLAG_RD
,
1079 0, 0, sysctl_kern_clockrate
, "S,clockinfo","");
1082 * We have eight functions for looking at the clock, four for
1083 * microseconds and four for nanoseconds. For each there is fast
1084 * but less precise version "get{nano|micro}[up]time" which will
1085 * return a time which is up to 1/HZ previous to the call, whereas
1086 * the raw version "{nano|micro}[up]time" will return a timestamp
1087 * which is as precise as possible. The "up" variants return the
1088 * time relative to system boot, these are well suited for time
1089 * interval measurements.
1091 * Each cpu independently maintains the current time of day, so all
1092 * we need to do to protect ourselves from changes is to do a loop
1093 * check on the seconds field changing out from under us.
1095 * The system timer maintains a 32 bit count and due to various issues
1096 * it is possible for the calculated delta to occasionally exceed
1097 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1098 * multiplication can easily overflow, so we deal with the case. For
1099 * uniformity we deal with the case in the usec case too.
1101 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1104 getmicrouptime(struct timeval
*tvp
)
1106 struct globaldata
*gd
= mycpu
;
1110 tvp
->tv_sec
= gd
->gd_time_seconds
;
1111 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1112 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1114 if (delta
>= sys_cputimer
->freq
) {
1115 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1116 delta
%= sys_cputimer
->freq
;
1118 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1119 if (tvp
->tv_usec
>= 1000000) {
1120 tvp
->tv_usec
-= 1000000;
1126 getnanouptime(struct timespec
*tsp
)
1128 struct globaldata
*gd
= mycpu
;
1132 tsp
->tv_sec
= gd
->gd_time_seconds
;
1133 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1134 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1136 if (delta
>= sys_cputimer
->freq
) {
1137 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1138 delta
%= sys_cputimer
->freq
;
1140 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1144 microuptime(struct timeval
*tvp
)
1146 struct globaldata
*gd
= mycpu
;
1150 tvp
->tv_sec
= gd
->gd_time_seconds
;
1151 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1152 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1154 if (delta
>= sys_cputimer
->freq
) {
1155 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1156 delta
%= sys_cputimer
->freq
;
1158 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1162 nanouptime(struct timespec
*tsp
)
1164 struct globaldata
*gd
= mycpu
;
1168 tsp
->tv_sec
= gd
->gd_time_seconds
;
1169 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1170 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1172 if (delta
>= sys_cputimer
->freq
) {
1173 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1174 delta
%= sys_cputimer
->freq
;
1176 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1183 getmicrotime(struct timeval
*tvp
)
1185 struct globaldata
*gd
= mycpu
;
1186 struct timespec
*bt
;
1190 tvp
->tv_sec
= gd
->gd_time_seconds
;
1191 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1192 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1194 if (delta
>= sys_cputimer
->freq
) {
1195 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1196 delta
%= sys_cputimer
->freq
;
1198 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1200 bt
= &basetime
[basetime_index
];
1201 tvp
->tv_sec
+= bt
->tv_sec
;
1202 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1203 while (tvp
->tv_usec
>= 1000000) {
1204 tvp
->tv_usec
-= 1000000;
1210 getnanotime(struct timespec
*tsp
)
1212 struct globaldata
*gd
= mycpu
;
1213 struct timespec
*bt
;
1217 tsp
->tv_sec
= gd
->gd_time_seconds
;
1218 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1219 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1221 if (delta
>= sys_cputimer
->freq
) {
1222 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1223 delta
%= sys_cputimer
->freq
;
1225 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1227 bt
= &basetime
[basetime_index
];
1228 tsp
->tv_sec
+= bt
->tv_sec
;
1229 tsp
->tv_nsec
+= bt
->tv_nsec
;
1230 while (tsp
->tv_nsec
>= 1000000000) {
1231 tsp
->tv_nsec
-= 1000000000;
1237 getnanotime_nbt(struct timespec
*nbt
, struct timespec
*tsp
)
1239 struct globaldata
*gd
= mycpu
;
1243 tsp
->tv_sec
= gd
->gd_time_seconds
;
1244 delta
= gd
->gd_hardclock
.time
- gd
->gd_cpuclock_base
;
1245 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1247 if (delta
>= sys_cputimer
->freq
) {
1248 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1249 delta
%= sys_cputimer
->freq
;
1251 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1253 tsp
->tv_sec
+= nbt
->tv_sec
;
1254 tsp
->tv_nsec
+= nbt
->tv_nsec
;
1255 while (tsp
->tv_nsec
>= 1000000000) {
1256 tsp
->tv_nsec
-= 1000000000;
1263 microtime(struct timeval
*tvp
)
1265 struct globaldata
*gd
= mycpu
;
1266 struct timespec
*bt
;
1270 tvp
->tv_sec
= gd
->gd_time_seconds
;
1271 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1272 } while (tvp
->tv_sec
!= gd
->gd_time_seconds
);
1274 if (delta
>= sys_cputimer
->freq
) {
1275 tvp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1276 delta
%= sys_cputimer
->freq
;
1278 tvp
->tv_usec
= (sys_cputimer
->freq64_usec
* delta
) >> 32;
1280 bt
= &basetime
[basetime_index
];
1281 tvp
->tv_sec
+= bt
->tv_sec
;
1282 tvp
->tv_usec
+= bt
->tv_nsec
/ 1000;
1283 while (tvp
->tv_usec
>= 1000000) {
1284 tvp
->tv_usec
-= 1000000;
1290 nanotime(struct timespec
*tsp
)
1292 struct globaldata
*gd
= mycpu
;
1293 struct timespec
*bt
;
1297 tsp
->tv_sec
= gd
->gd_time_seconds
;
1298 delta
= sys_cputimer
->count() - gd
->gd_cpuclock_base
;
1299 } while (tsp
->tv_sec
!= gd
->gd_time_seconds
);
1301 if (delta
>= sys_cputimer
->freq
) {
1302 tsp
->tv_sec
+= delta
/ sys_cputimer
->freq
;
1303 delta
%= sys_cputimer
->freq
;
1305 tsp
->tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1307 bt
= &basetime
[basetime_index
];
1308 tsp
->tv_sec
+= bt
->tv_sec
;
1309 tsp
->tv_nsec
+= bt
->tv_nsec
;
1310 while (tsp
->tv_nsec
>= 1000000000) {
1311 tsp
->tv_nsec
-= 1000000000;
1317 * note: this is not exactly synchronized with real time. To do that we
1318 * would have to do what microtime does and check for a nanoseconds overflow.
1321 get_approximate_time_t(void)
1323 struct globaldata
*gd
= mycpu
;
1324 struct timespec
*bt
;
1326 bt
= &basetime
[basetime_index
];
1327 return(gd
->gd_time_seconds
+ bt
->tv_sec
);
1331 pps_ioctl(u_long cmd
, caddr_t data
, struct pps_state
*pps
)
1334 struct pps_fetch_args
*fapi
;
1336 struct pps_kcbind_args
*kapi
;
1340 case PPS_IOC_CREATE
:
1342 case PPS_IOC_DESTROY
:
1344 case PPS_IOC_SETPARAMS
:
1345 app
= (pps_params_t
*)data
;
1346 if (app
->mode
& ~pps
->ppscap
)
1348 pps
->ppsparam
= *app
;
1350 case PPS_IOC_GETPARAMS
:
1351 app
= (pps_params_t
*)data
;
1352 *app
= pps
->ppsparam
;
1353 app
->api_version
= PPS_API_VERS_1
;
1355 case PPS_IOC_GETCAP
:
1356 *(int*)data
= pps
->ppscap
;
1359 fapi
= (struct pps_fetch_args
*)data
;
1360 if (fapi
->tsformat
&& fapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1362 if (fapi
->timeout
.tv_sec
|| fapi
->timeout
.tv_nsec
)
1363 return (EOPNOTSUPP
);
1364 pps
->ppsinfo
.current_mode
= pps
->ppsparam
.mode
;
1365 fapi
->pps_info_buf
= pps
->ppsinfo
;
1367 case PPS_IOC_KCBIND
:
1369 kapi
= (struct pps_kcbind_args
*)data
;
1370 /* XXX Only root should be able to do this */
1371 if (kapi
->tsformat
&& kapi
->tsformat
!= PPS_TSFMT_TSPEC
)
1373 if (kapi
->kernel_consumer
!= PPS_KC_HARDPPS
)
1375 if (kapi
->edge
& ~pps
->ppscap
)
1377 pps
->kcmode
= kapi
->edge
;
1380 return (EOPNOTSUPP
);
1388 pps_init(struct pps_state
*pps
)
1390 pps
->ppscap
|= PPS_TSFMT_TSPEC
;
1391 if (pps
->ppscap
& PPS_CAPTUREASSERT
)
1392 pps
->ppscap
|= PPS_OFFSETASSERT
;
1393 if (pps
->ppscap
& PPS_CAPTURECLEAR
)
1394 pps
->ppscap
|= PPS_OFFSETCLEAR
;
1398 pps_event(struct pps_state
*pps
, sysclock_t count
, int event
)
1400 struct globaldata
*gd
;
1401 struct timespec
*tsp
;
1402 struct timespec
*osp
;
1403 struct timespec
*bt
;
1420 /* Things would be easier with arrays... */
1421 if (event
== PPS_CAPTUREASSERT
) {
1422 tsp
= &pps
->ppsinfo
.assert_timestamp
;
1423 osp
= &pps
->ppsparam
.assert_offset
;
1424 foff
= pps
->ppsparam
.mode
& PPS_OFFSETASSERT
;
1425 fhard
= pps
->kcmode
& PPS_CAPTUREASSERT
;
1426 pcount
= &pps
->ppscount
[0];
1427 pseq
= &pps
->ppsinfo
.assert_sequence
;
1429 tsp
= &pps
->ppsinfo
.clear_timestamp
;
1430 osp
= &pps
->ppsparam
.clear_offset
;
1431 foff
= pps
->ppsparam
.mode
& PPS_OFFSETCLEAR
;
1432 fhard
= pps
->kcmode
& PPS_CAPTURECLEAR
;
1433 pcount
= &pps
->ppscount
[1];
1434 pseq
= &pps
->ppsinfo
.clear_sequence
;
1437 /* Nothing really happened */
1438 if (*pcount
== count
)
1444 ts
.tv_sec
= gd
->gd_time_seconds
;
1445 delta
= count
- gd
->gd_cpuclock_base
;
1446 } while (ts
.tv_sec
!= gd
->gd_time_seconds
);
1448 if (delta
>= sys_cputimer
->freq
) {
1449 ts
.tv_sec
+= delta
/ sys_cputimer
->freq
;
1450 delta
%= sys_cputimer
->freq
;
1452 ts
.tv_nsec
= (sys_cputimer
->freq64_nsec
* delta
) >> 32;
1453 bt
= &basetime
[basetime_index
];
1454 ts
.tv_sec
+= bt
->tv_sec
;
1455 ts
.tv_nsec
+= bt
->tv_nsec
;
1456 while (ts
.tv_nsec
>= 1000000000) {
1457 ts
.tv_nsec
-= 1000000000;
1465 timespecadd(tsp
, osp
);
1466 if (tsp
->tv_nsec
< 0) {
1467 tsp
->tv_nsec
+= 1000000000;
1473 /* magic, at its best... */
1474 tcount
= count
- pps
->ppscount
[2];
1475 pps
->ppscount
[2] = count
;
1476 if (tcount
>= sys_cputimer
->freq
) {
1477 delta
= (1000000000 * (tcount
/ sys_cputimer
->freq
) +
1478 sys_cputimer
->freq64_nsec
*
1479 (tcount
% sys_cputimer
->freq
)) >> 32;
1481 delta
= (sys_cputimer
->freq64_nsec
* tcount
) >> 32;
1483 hardpps(tsp
, delta
);
1489 * Return the tsc target value for a delay of (ns).
1491 * Returns -1 if the TSC is not supported.
1494 tsc_get_target(int ns
)
1496 #if defined(_RDTSC_SUPPORTED_)
1497 if (cpu_feature
& CPUID_TSC
) {
1498 return (rdtsc() + tsc_frequency
* ns
/ (int64_t)1000000000);
1505 * Compare the tsc against the passed target
1507 * Returns +1 if the target has been reached
1508 * Returns 0 if the target has not yet been reached
1509 * Returns -1 if the TSC is not supported.
1511 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1514 tsc_test_target(int64_t target
)
1516 #if defined(_RDTSC_SUPPORTED_)
1517 if (cpu_feature
& CPUID_TSC
) {
1518 if ((int64_t)(target
- rdtsc()) <= 0)
1527 * Delay the specified number of nanoseconds using the tsc. This function
1528 * returns immediately if the TSC is not supported. At least one cpu_pause()
1536 clk
= tsc_get_target(ns
);
1538 while (tsc_test_target(clk
) == 0)