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 __unused, struct intrframe *frame)
403 struct globaldata *gd = mycpu;
406 * Realtime updates are per-cpu. Note that timer corrections as
407 * returned by microtime() and friends make an additional adjustment
408 * using a system-wise 'basetime', but the running time is always
409 * taken from the per-cpu globaldata area. Since the same clock
410 * is distributing (XXX SMP) to all cpus, the per-cpu timebases
413 * Note that we never allow info->time (aka gd->gd_hardclock.time)
414 * to reverse index gd_cpuclock_base, but that it is possible for
415 * it to temporarily get behind in the seconds if something in the
416 * system locks interrupts for a long period of time. Since periodic
417 * timers count events, though everything should resynch again
420 cputicks = info->time - gd->gd_cpuclock_base;
421 if (cputicks >= sys_cputimer->freq) {
422 ++gd->gd_time_seconds;
423 gd->gd_cpuclock_base += sys_cputimer->freq;
424 if (gd->gd_cpuid == 0)
425 ++time_uptime; /* uncorrected monotonic 1-sec gran */
429 * The system-wide ticks counter and NTP related timedelta/tickdelta
430 * adjustments only occur on cpu #0. NTP adjustments are accomplished
431 * by updating basetime.
433 if (gd->gd_cpuid == 0) {
434 struct timespec *nbt;
442 if (tco->tc_poll_pps)
443 tco->tc_poll_pps(tco);
447 * Calculate the new basetime index. We are in a critical section
448 * on cpu #0 and can safely play with basetime_index. Start
449 * with the current basetime and then make adjustments.
451 ni = (basetime_index + 1) & BASETIME_ARYMASK;
453 *nbt = basetime[basetime_index];
456 * Apply adjtime corrections. (adjtime() API)
458 * adjtime() only runs on cpu #0 so our critical section is
459 * sufficient to access these variables.
461 if (ntp_delta != 0) {
462 nbt->tv_nsec += ntp_tick_delta;
463 ntp_delta -= ntp_tick_delta;
464 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
465 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
466 ntp_tick_delta = ntp_delta;
471 * Apply permanent frequency corrections. (sysctl API)
473 if (ntp_tick_permanent != 0) {
474 ntp_tick_acc += ntp_tick_permanent;
475 if (ntp_tick_acc >= (1LL << 32)) {
476 nbt->tv_nsec += ntp_tick_acc >> 32;
477 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
478 } else if (ntp_tick_acc <= -(1LL << 32)) {
479 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
480 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
481 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
485 if (nbt->tv_nsec >= 1000000000) {
487 nbt->tv_nsec -= 1000000000;
488 } else if (nbt->tv_nsec < 0) {
490 nbt->tv_nsec += 1000000000;
494 * Another per-tick compensation. (for ntp_adjtime() API)
497 nsec_acc += nsec_adj;
498 if (nsec_acc >= 0x100000000LL) {
499 nbt->tv_nsec += nsec_acc >> 32;
500 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
501 } else if (nsec_acc <= -0x100000000LL) {
502 nbt->tv_nsec -= -nsec_acc >> 32;
503 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
505 if (nbt->tv_nsec >= 1000000000) {
506 nbt->tv_nsec -= 1000000000;
508 } else if (nbt->tv_nsec < 0) {
509 nbt->tv_nsec += 1000000000;
514 /************************************************************
515 * LEAP SECOND CORRECTION *
516 ************************************************************
518 * Taking into account all the corrections made above, figure
519 * out the new real time. If the seconds field has changed
520 * then apply any pending leap-second corrections.
522 getnanotime_nbt(nbt, &nts);
524 if (time_second != nts.tv_sec) {
526 * Apply leap second (sysctl API). Adjust nts for changes
527 * so we do not have to call getnanotime_nbt again.
529 if (ntp_leap_second) {
530 if (ntp_leap_second == nts.tv_sec) {
531 if (ntp_leap_insert) {
543 * Apply leap second (ntp_adjtime() API), calculate a new
544 * nsec_adj field. ntp_update_second() returns nsec_adj
545 * as a per-second value but we need it as a per-tick value.
547 leap = ntp_update_second(time_second, &nsec_adj);
553 * Update the time_second 'approximate time' global.
555 time_second = nts.tv_sec;
559 * Finally, our new basetime is ready to go live!
565 * Update kpmap on each tick. TS updates are integrated with
566 * fences and upticks allowing userland to read the data
572 w = (kpmap->upticks + 1) & 1;
573 getnanouptime(&kpmap->ts_uptime[w]);
574 getnanotime(&kpmap->ts_realtime[w]);
582 * lwkt thread scheduler fair queueing
584 lwkt_schedulerclock(curthread);
587 * softticks are handled for all cpus
589 hardclock_softtick(gd);
592 * ITimer handling is per-tick, per-cpu.
594 * We must acquire the per-process token in order for ksignal()
595 * to be non-blocking. For the moment this requires an AST fault,
596 * the ksignal() cannot be safely issued from this hard interrupt.
598 * XXX Even the trytoken here isn't right, and itimer operation in
599 * a multi threaded environment is going to be weird at the
602 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
605 ++p->p_upmap->runticks;
607 if (frame && CLKF_USERMODE(frame) &&
608 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
609 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
610 p->p_flags |= P_SIGVTALRM;
613 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
614 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
615 p->p_flags |= P_SIGPROF;
619 lwkt_reltoken(&p->p_token);
625 * The statistics clock typically runs at a 125Hz rate, and is intended
626 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
628 * NOTE! systimer! the MP lock might not be held here. We can only safely
629 * manipulate objects owned by the current cpu.
631 * The stats clock is responsible for grabbing a profiling sample.
632 * Most of the statistics are only used by user-level statistics programs.
633 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
636 * Like the other clocks, the stat clock is called from what is effectively
637 * a fast interrupt, so the context should be the thread/process that got
641 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
654 * How big was our timeslice relative to the last time? Calculate
657 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
658 * during early boot. Just use the systimer count to be nice
659 * to e.g. qemu. The systimer has a better chance of being
660 * MPSAFE at early boot.
662 cv = sys_cputimer->count();
663 scv = mycpu->statint.gd_statcv;
667 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32;
673 mycpu->statint.gd_statcv = cv;
676 stv = &mycpu->gd_stattv;
677 if (stv->tv_sec == 0) {
680 bump = tv.tv_usec - stv->tv_usec +
681 (tv.tv_sec - stv->tv_sec) * 1000000;
693 if (frame && CLKF_USERMODE(frame)) {
695 * Came from userland, handle user time and deal with
698 if (p && (p->p_flags & P_PROFIL))
699 addupc_intr(p, CLKF_PC(frame), 1);
700 td->td_uticks += bump;
703 * Charge the time as appropriate
705 if (p && p->p_nice > NZERO)
706 cpu_time.cp_nice += bump;
708 cpu_time.cp_user += bump;
710 int intr_nest = mycpu->gd_intr_nesting_level;
714 * IPI processing code will bump gd_intr_nesting_level
715 * up by one, which breaks following CLKF_INTR testing,
716 * so we subtract it by one here.
722 * Kernel statistics are just like addupc_intr, only easier.
725 if (g->state == GMON_PROF_ON && frame) {
726 i = CLKF_PC(frame) - g->lowpc;
727 if (i < g->textsize) {
728 i /= HISTFRACTION * sizeof(*g->kcount);
734 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
737 * Came from kernel mode, so we were:
738 * - handling an interrupt,
739 * - doing syscall or trap work on behalf of the current
741 * - spinning in the idle loop.
742 * Whichever it is, charge the time as appropriate.
743 * Note that we charge interrupts to the current process,
744 * regardless of whether they are ``for'' that process,
745 * so that we know how much of its real time was spent
746 * in ``non-process'' (i.e., interrupt) work.
748 * XXX assume system if frame is NULL. A NULL frame
749 * can occur if ipi processing is done from a crit_exit().
752 td->td_iticks += bump;
754 td->td_sticks += bump;
756 if (IS_INTR_RUNNING) {
758 * If we interrupted an interrupt thread, well,
759 * count it as interrupt time.
763 do_pctrack(frame, PCTRACK_INT);
765 cpu_time.cp_intr += bump;
767 if (td == &mycpu->gd_idlethread) {
769 * Even if the current thread is the idle
770 * thread it could be due to token contention
771 * in the LWKT scheduler. Count such as
774 if (mycpu->gd_reqflags & RQF_IDLECHECK_WK_MASK)
775 cpu_time.cp_sys += bump;
777 cpu_time.cp_idle += bump;
780 * System thread was running.
784 do_pctrack(frame, PCTRACK_SYS);
786 cpu_time.cp_sys += bump;
790 #undef IS_INTR_RUNNING
796 * Sample the PC when in the kernel or in an interrupt. User code can
797 * retrieve the information and generate a histogram or other output.
801 do_pctrack(struct intrframe *frame, int which)
803 struct kinfo_pctrack *pctrack;
805 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
806 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
807 (void *)CLKF_PC(frame);
812 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
814 struct kinfo_pcheader head;
819 head.pc_ntrack = PCTRACK_SIZE;
820 head.pc_arysize = PCTRACK_ARYSIZE;
822 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
825 for (cpu = 0; cpu < ncpus; ++cpu) {
826 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
827 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
828 sizeof(struct kinfo_pctrack));
837 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
838 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
843 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
844 * the MP lock might not be held. We can safely manipulate parts of curproc
845 * but that's about it.
847 * Each cpu has its own scheduler clock.
850 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
857 if ((lp = lwkt_preempted_proc()) != NULL) {
859 * Account for cpu time used and hit the scheduler. Note
860 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
864 usched_schedulerclock(lp, info->periodic, info->time);
866 usched_schedulerclock(NULL, info->periodic, info->time);
868 if ((lp = curthread->td_lwp) != NULL) {
870 * Update resource usage integrals and maximums.
872 if ((ru = &lp->lwp_proc->p_ru) &&
873 (vm = lp->lwp_proc->p_vmspace) != NULL) {
874 ru->ru_ixrss += pgtok(vm->vm_tsize);
875 ru->ru_idrss += pgtok(vm->vm_dsize);
876 ru->ru_isrss += pgtok(vm->vm_ssize);
877 if (lwkt_trytoken(&vm->vm_map.token)) {
878 rss = pgtok(vmspace_resident_count(vm));
879 if (ru->ru_maxrss < rss)
881 lwkt_reltoken(&vm->vm_map.token);
885 /* Increment the global sched_ticks */
886 if (mycpu->gd_cpuid == 0)
891 * Compute number of ticks for the specified amount of time. The
892 * return value is intended to be used in a clock interrupt timed
893 * operation and guaranteed to meet or exceed the requested time.
894 * If the representation overflows, return INT_MAX. The minimum return
895 * value is 1 ticks and the function will average the calculation up.
896 * If any value greater then 0 microseconds is supplied, a value
897 * of at least 2 will be returned to ensure that a near-term clock
898 * interrupt does not cause the timeout to occur (degenerately) early.
900 * Note that limit checks must take into account microseconds, which is
901 * done simply by using the smaller signed long maximum instead of
902 * the unsigned long maximum.
904 * If ints have 32 bits, then the maximum value for any timeout in
905 * 10ms ticks is 248 days.
908 tvtohz_high(struct timeval *tv)
925 kprintf("tvtohz_high: negative time difference "
926 "%ld sec %ld usec\n",
930 } else if (sec <= INT_MAX / hz) {
931 ticks = (int)(sec * hz +
932 ((u_long)usec + (ustick - 1)) / ustick) + 1;
940 tstohz_high(struct timespec *ts)
957 kprintf("tstohz_high: negative time difference "
958 "%ld sec %ld nsec\n",
962 } else if (sec <= INT_MAX / hz) {
963 ticks = (int)(sec * hz +
964 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
973 * Compute number of ticks for the specified amount of time, erroring on
974 * the side of it being too low to ensure that sleeping the returned number
975 * of ticks will not result in a late return.
977 * The supplied timeval may not be negative and should be normalized. A
978 * return value of 0 is possible if the timeval converts to less then
981 * If ints have 32 bits, then the maximum value for any timeout in
982 * 10ms ticks is 248 days.
985 tvtohz_low(struct timeval *tv)
991 if (sec <= INT_MAX / hz)
992 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
999 tstohz_low(struct timespec *ts)
1005 if (sec <= INT_MAX / hz)
1006 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1013 * Start profiling on a process.
1015 * Kernel profiling passes proc0 which never exits and hence
1016 * keeps the profile clock running constantly.
1019 startprofclock(struct proc *p)
1021 if ((p->p_flags & P_PROFIL) == 0) {
1022 p->p_flags |= P_PROFIL;
1024 if (++profprocs == 1 && stathz != 0) {
1027 setstatclockrate(profhz);
1035 * Stop profiling on a process.
1037 * caller must hold p->p_token
1040 stopprofclock(struct proc *p)
1042 if (p->p_flags & P_PROFIL) {
1043 p->p_flags &= ~P_PROFIL;
1045 if (--profprocs == 0 && stathz != 0) {
1048 setstatclockrate(stathz);
1056 * Return information about system clocks.
1059 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1061 struct kinfo_clockinfo clkinfo;
1063 * Construct clockinfo structure.
1066 clkinfo.ci_tick = ustick;
1067 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1068 clkinfo.ci_profhz = profhz;
1069 clkinfo.ci_stathz = stathz ? stathz : hz;
1070 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1073 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1074 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1077 * We have eight functions for looking at the clock, four for
1078 * microseconds and four for nanoseconds. For each there is fast
1079 * but less precise version "get{nano|micro}[up]time" which will
1080 * return a time which is up to 1/HZ previous to the call, whereas
1081 * the raw version "{nano|micro}[up]time" will return a timestamp
1082 * which is as precise as possible. The "up" variants return the
1083 * time relative to system boot, these are well suited for time
1084 * interval measurements.
1086 * Each cpu independently maintains the current time of day, so all
1087 * we need to do to protect ourselves from changes is to do a loop
1088 * check on the seconds field changing out from under us.
1090 * The system timer maintains a 32 bit count and due to various issues
1091 * it is possible for the calculated delta to occasionally exceed
1092 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1093 * multiplication can easily overflow, so we deal with the case. For
1094 * uniformity we deal with the case in the usec case too.
1096 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1099 getmicrouptime(struct timeval *tvp)
1101 struct globaldata *gd = mycpu;
1105 tvp->tv_sec = gd->gd_time_seconds;
1106 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1107 } while (tvp->tv_sec != gd->gd_time_seconds);
1109 if (delta >= sys_cputimer->freq) {
1110 tvp->tv_sec += delta / sys_cputimer->freq;
1111 delta %= sys_cputimer->freq;
1113 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1114 if (tvp->tv_usec >= 1000000) {
1115 tvp->tv_usec -= 1000000;
1121 getnanouptime(struct timespec *tsp)
1123 struct globaldata *gd = mycpu;
1127 tsp->tv_sec = gd->gd_time_seconds;
1128 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1129 } while (tsp->tv_sec != gd->gd_time_seconds);
1131 if (delta >= sys_cputimer->freq) {
1132 tsp->tv_sec += delta / sys_cputimer->freq;
1133 delta %= sys_cputimer->freq;
1135 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1139 microuptime(struct timeval *tvp)
1141 struct globaldata *gd = mycpu;
1145 tvp->tv_sec = gd->gd_time_seconds;
1146 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1147 } while (tvp->tv_sec != gd->gd_time_seconds);
1149 if (delta >= sys_cputimer->freq) {
1150 tvp->tv_sec += delta / sys_cputimer->freq;
1151 delta %= sys_cputimer->freq;
1153 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1157 nanouptime(struct timespec *tsp)
1159 struct globaldata *gd = mycpu;
1163 tsp->tv_sec = gd->gd_time_seconds;
1164 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1165 } while (tsp->tv_sec != gd->gd_time_seconds);
1167 if (delta >= sys_cputimer->freq) {
1168 tsp->tv_sec += delta / sys_cputimer->freq;
1169 delta %= sys_cputimer->freq;
1171 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1178 getmicrotime(struct timeval *tvp)
1180 struct globaldata *gd = mycpu;
1181 struct timespec *bt;
1185 tvp->tv_sec = gd->gd_time_seconds;
1186 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1187 } while (tvp->tv_sec != gd->gd_time_seconds);
1189 if (delta >= sys_cputimer->freq) {
1190 tvp->tv_sec += delta / sys_cputimer->freq;
1191 delta %= sys_cputimer->freq;
1193 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1195 bt = &basetime[basetime_index];
1196 tvp->tv_sec += bt->tv_sec;
1197 tvp->tv_usec += bt->tv_nsec / 1000;
1198 while (tvp->tv_usec >= 1000000) {
1199 tvp->tv_usec -= 1000000;
1205 getnanotime(struct timespec *tsp)
1207 struct globaldata *gd = mycpu;
1208 struct timespec *bt;
1212 tsp->tv_sec = gd->gd_time_seconds;
1213 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1214 } while (tsp->tv_sec != gd->gd_time_seconds);
1216 if (delta >= sys_cputimer->freq) {
1217 tsp->tv_sec += delta / sys_cputimer->freq;
1218 delta %= sys_cputimer->freq;
1220 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1222 bt = &basetime[basetime_index];
1223 tsp->tv_sec += bt->tv_sec;
1224 tsp->tv_nsec += bt->tv_nsec;
1225 while (tsp->tv_nsec >= 1000000000) {
1226 tsp->tv_nsec -= 1000000000;
1232 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1234 struct globaldata *gd = mycpu;
1238 tsp->tv_sec = gd->gd_time_seconds;
1239 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1240 } while (tsp->tv_sec != gd->gd_time_seconds);
1242 if (delta >= sys_cputimer->freq) {
1243 tsp->tv_sec += delta / sys_cputimer->freq;
1244 delta %= sys_cputimer->freq;
1246 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1248 tsp->tv_sec += nbt->tv_sec;
1249 tsp->tv_nsec += nbt->tv_nsec;
1250 while (tsp->tv_nsec >= 1000000000) {
1251 tsp->tv_nsec -= 1000000000;
1258 microtime(struct timeval *tvp)
1260 struct globaldata *gd = mycpu;
1261 struct timespec *bt;
1265 tvp->tv_sec = gd->gd_time_seconds;
1266 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1267 } while (tvp->tv_sec != gd->gd_time_seconds);
1269 if (delta >= sys_cputimer->freq) {
1270 tvp->tv_sec += delta / sys_cputimer->freq;
1271 delta %= sys_cputimer->freq;
1273 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1275 bt = &basetime[basetime_index];
1276 tvp->tv_sec += bt->tv_sec;
1277 tvp->tv_usec += bt->tv_nsec / 1000;
1278 while (tvp->tv_usec >= 1000000) {
1279 tvp->tv_usec -= 1000000;
1285 nanotime(struct timespec *tsp)
1287 struct globaldata *gd = mycpu;
1288 struct timespec *bt;
1292 tsp->tv_sec = gd->gd_time_seconds;
1293 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1294 } while (tsp->tv_sec != gd->gd_time_seconds);
1296 if (delta >= sys_cputimer->freq) {
1297 tsp->tv_sec += delta / sys_cputimer->freq;
1298 delta %= sys_cputimer->freq;
1300 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1302 bt = &basetime[basetime_index];
1303 tsp->tv_sec += bt->tv_sec;
1304 tsp->tv_nsec += bt->tv_nsec;
1305 while (tsp->tv_nsec >= 1000000000) {
1306 tsp->tv_nsec -= 1000000000;
1312 * note: this is not exactly synchronized with real time. To do that we
1313 * would have to do what microtime does and check for a nanoseconds overflow.
1316 get_approximate_time_t(void)
1318 struct globaldata *gd = mycpu;
1319 struct timespec *bt;
1321 bt = &basetime[basetime_index];
1322 return(gd->gd_time_seconds + bt->tv_sec);
1326 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1329 struct pps_fetch_args *fapi;
1331 struct pps_kcbind_args *kapi;
1335 case PPS_IOC_CREATE:
1337 case PPS_IOC_DESTROY:
1339 case PPS_IOC_SETPARAMS:
1340 app = (pps_params_t *)data;
1341 if (app->mode & ~pps->ppscap)
1343 pps->ppsparam = *app;
1345 case PPS_IOC_GETPARAMS:
1346 app = (pps_params_t *)data;
1347 *app = pps->ppsparam;
1348 app->api_version = PPS_API_VERS_1;
1350 case PPS_IOC_GETCAP:
1351 *(int*)data = pps->ppscap;
1354 fapi = (struct pps_fetch_args *)data;
1355 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1357 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1358 return (EOPNOTSUPP);
1359 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1360 fapi->pps_info_buf = pps->ppsinfo;
1362 case PPS_IOC_KCBIND:
1364 kapi = (struct pps_kcbind_args *)data;
1365 /* XXX Only root should be able to do this */
1366 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1368 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1370 if (kapi->edge & ~pps->ppscap)
1372 pps->kcmode = kapi->edge;
1375 return (EOPNOTSUPP);
1383 pps_init(struct pps_state *pps)
1385 pps->ppscap |= PPS_TSFMT_TSPEC;
1386 if (pps->ppscap & PPS_CAPTUREASSERT)
1387 pps->ppscap |= PPS_OFFSETASSERT;
1388 if (pps->ppscap & PPS_CAPTURECLEAR)
1389 pps->ppscap |= PPS_OFFSETCLEAR;
1393 pps_event(struct pps_state *pps, sysclock_t count, int event)
1395 struct globaldata *gd;
1396 struct timespec *tsp;
1397 struct timespec *osp;
1398 struct timespec *bt;
1415 /* Things would be easier with arrays... */
1416 if (event == PPS_CAPTUREASSERT) {
1417 tsp = &pps->ppsinfo.assert_timestamp;
1418 osp = &pps->ppsparam.assert_offset;
1419 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1420 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1421 pcount = &pps->ppscount[0];
1422 pseq = &pps->ppsinfo.assert_sequence;
1424 tsp = &pps->ppsinfo.clear_timestamp;
1425 osp = &pps->ppsparam.clear_offset;
1426 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1427 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1428 pcount = &pps->ppscount[1];
1429 pseq = &pps->ppsinfo.clear_sequence;
1432 /* Nothing really happened */
1433 if (*pcount == count)
1439 ts.tv_sec = gd->gd_time_seconds;
1440 delta = count - gd->gd_cpuclock_base;
1441 } while (ts.tv_sec != gd->gd_time_seconds);
1443 if (delta >= sys_cputimer->freq) {
1444 ts.tv_sec += delta / sys_cputimer->freq;
1445 delta %= sys_cputimer->freq;
1447 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1448 bt = &basetime[basetime_index];
1449 ts.tv_sec += bt->tv_sec;
1450 ts.tv_nsec += bt->tv_nsec;
1451 while (ts.tv_nsec >= 1000000000) {
1452 ts.tv_nsec -= 1000000000;
1460 timespecadd(tsp, osp);
1461 if (tsp->tv_nsec < 0) {
1462 tsp->tv_nsec += 1000000000;
1468 /* magic, at its best... */
1469 tcount = count - pps->ppscount[2];
1470 pps->ppscount[2] = count;
1471 if (tcount >= sys_cputimer->freq) {
1472 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1473 sys_cputimer->freq64_nsec *
1474 (tcount % sys_cputimer->freq)) >> 32;
1476 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1478 hardpps(tsp, delta);
1484 * Return the tsc target value for a delay of (ns).
1486 * Returns -1 if the TSC is not supported.
1489 tsc_get_target(int ns)
1491 #if defined(_RDTSC_SUPPORTED_)
1492 if (cpu_feature & CPUID_TSC) {
1493 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1500 * Compare the tsc against the passed target
1502 * Returns +1 if the target has been reached
1503 * Returns 0 if the target has not yet been reached
1504 * Returns -1 if the TSC is not supported.
1506 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1509 tsc_test_target(int64_t target)
1511 #if defined(_RDTSC_SUPPORTED_)
1512 if (cpu_feature & CPUID_TSC) {
1513 if ((int64_t)(target - rdtsc()) <= 0)
1522 * Delay the specified number of nanoseconds using the tsc. This function
1523 * returns immediately if the TSC is not supported. At least one cpu_pause()
1531 clk = tsc_get_target(ns);
1533 while (tsc_test_target(clk) == 0)