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.
202 __uint32_t time_second;
203 sysclock_t cpuclock_base;
206 #define BASETIME_ARYSIZE 16
207 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
208 static struct timespec basetime[BASETIME_ARYSIZE];
209 static struct hardtime hardtime[BASETIME_ARYSIZE];
210 static volatile int basetime_index;
213 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
220 * Because basetime data and index may be updated by another cpu,
221 * a load fence is required to ensure that the data we read has
222 * not been speculatively read relative to a possibly updated index.
224 index = basetime_index;
226 bt = &basetime[index];
227 error = SYSCTL_OUT(req, bt, sizeof(*bt));
231 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
232 &boottime, timespec, "System boottime");
233 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
234 sysctl_get_basetime, "S,timespec", "System basetime");
236 static void hardclock(systimer_t info, int, struct intrframe *frame);
237 static void statclock(systimer_t info, int, struct intrframe *frame);
238 static void schedclock(systimer_t info, int, struct intrframe *frame);
239 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
241 int ticks; /* system master ticks at hz */
242 int clocks_running; /* tsleep/timeout clocks operational */
243 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
244 int64_t nsec_acc; /* accumulator */
245 int sched_ticks; /* global schedule clock ticks */
247 /* NTPD time correction fields */
248 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
249 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
250 int64_t ntp_delta; /* one-time correction in nsec */
251 int64_t ntp_big_delta = 1000000000;
252 int32_t ntp_tick_delta; /* current adjustment rate */
253 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
254 time_t ntp_leap_second; /* time of next leap second */
255 int ntp_leap_insert; /* whether to insert or remove a second */
258 * Finish initializing clock frequencies and start all clocks running.
262 initclocks(void *dummy)
264 /*psratio = profhz / stathz;*/
268 kpmap->tsc_freq = (uint64_t)tsc_frequency;
269 kpmap->tick_freq = hz;
274 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
275 * during SMP initialization.
277 * This routine is called concurrently during low-level SMP initialization
278 * and may not block in any way. Meaning, among other things, we can't
279 * acquire any tokens.
282 initclocks_pcpu(void)
284 struct globaldata *gd = mycpu;
287 if (gd->gd_cpuid == 0) {
288 gd->gd_time_seconds = 1;
289 gd->gd_cpuclock_base = sys_cputimer->count();
290 hardtime[0].time_second = gd->gd_time_seconds;
291 hardtime[0].cpuclock_base = gd->gd_cpuclock_base;
293 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
294 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
297 systimer_intr_enable();
303 * This routine is called on just the BSP, just after SMP initialization
304 * completes to * finish initializing any clocks that might contend/block
305 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
306 * that function is called from the idle thread bootstrap for each cpu and
307 * not allowed to block at all.
311 initclocks_other(void *dummy)
313 struct globaldata *ogd = mycpu;
314 struct globaldata *gd;
317 for (n = 0; n < ncpus; ++n) {
318 lwkt_setcpu_self(globaldata_find(n));
322 * Use a non-queued periodic systimer to prevent multiple
323 * ticks from building up if the sysclock jumps forward
324 * (8254 gets reset). The sysclock will never jump backwards.
325 * Our time sync is based on the actual sysclock, not the
328 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock,
330 systimer_init_periodic_nq(&gd->gd_statclock, statclock,
332 /* XXX correct the frequency for scheduler / estcpu tests */
333 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
336 ifpoll_init_pcpu(gd->gd_cpuid);
339 lwkt_setcpu_self(ogd);
341 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL);
344 * This sets the current real time of day. Timespecs are in seconds and
345 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
346 * instead we adjust basetime so basetime + gd_* results in the current
347 * time of day. This way the gd_* fields are guaranteed to represent
348 * a monotonically increasing 'uptime' value.
350 * When set_timeofday() is called from userland, the system call forces it
351 * onto cpu #0 since only cpu #0 can update basetime_index.
354 set_timeofday(struct timespec *ts)
356 struct timespec *nbt;
360 * XXX SMP / non-atomic basetime updates
363 ni = (basetime_index + 1) & BASETIME_ARYMASK;
367 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
368 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
369 if (nbt->tv_nsec < 0) {
370 nbt->tv_nsec += 1000000000;
375 * Note that basetime diverges from boottime as the clock drift is
376 * compensated for, so we cannot do away with boottime. When setting
377 * the absolute time of day the drift is 0 (for an instant) and we
378 * can simply assign boottime to basetime.
380 * Note that nanouptime() is based on gd_time_seconds which is drift
381 * compensated up to a point (it is guaranteed to remain monotonically
382 * increasing). gd_time_seconds is thus our best uptime guess and
383 * suitable for use in the boottime calculation. It is already taken
384 * into account in the basetime calculation above.
386 boottime.tv_sec = nbt->tv_sec;
390 * We now have a new basetime, make sure all other cpus have it,
391 * then update the index.
400 * Each cpu has its own hardclock, but we only increments ticks and softticks
403 * NOTE! systimer! the MP lock might not be held here. We can only safely
404 * manipulate objects owned by the current cpu.
407 hardclock(systimer_t info, int in_ipi, struct intrframe *frame)
411 struct globaldata *gd = mycpu;
413 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) {
414 /* Defer to doreti on passive IPIQ processing */
419 * We update the compensation base to calculate fine-grained time
420 * from the sys_cputimer on a per-cpu basis in order to avoid
421 * having to mess around with locks. sys_cputimer is assumed to
422 * be consistent across all cpus. CPU N copies the base state from
423 * CPU 0 using the same FIFO trick that we use for basetime (so we
424 * don't catch a CPU 0 update in the middle).
426 * Note that we never allow info->time (aka gd->gd_hardclock.time)
427 * to reverse index gd_cpuclock_base, but that it is possible for
428 * it to temporarily get behind in the seconds if something in the
429 * system locks interrupts for a long period of time. Since periodic
430 * timers count events, though everything should resynch again
433 if (gd->gd_cpuid == 0) {
436 cputicks = info->time - gd->gd_cpuclock_base;
437 if (cputicks >= sys_cputimer->freq) {
438 cputicks /= sys_cputimer->freq;
439 if (cputicks != 0 && cputicks != 1)
440 kprintf("Warning: hardclock missed > 1 sec\n");
441 gd->gd_time_seconds += cputicks;
442 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks;
443 /* uncorrected monotonic 1-sec gran */
444 time_uptime += cputicks;
446 ni = (basetime_index + 1) & BASETIME_ARYMASK;
447 hardtime[ni].time_second = gd->gd_time_seconds;
448 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base;
454 gd->gd_time_seconds = hardtime[ni].time_second;
455 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base;
459 * The system-wide ticks counter and NTP related timedelta/tickdelta
460 * adjustments only occur on cpu #0. NTP adjustments are accomplished
461 * by updating basetime.
463 if (gd->gd_cpuid == 0) {
464 struct timespec *nbt;
472 if (tco->tc_poll_pps)
473 tco->tc_poll_pps(tco);
477 * Calculate the new basetime index. We are in a critical section
478 * on cpu #0 and can safely play with basetime_index. Start
479 * with the current basetime and then make adjustments.
481 ni = (basetime_index + 1) & BASETIME_ARYMASK;
483 *nbt = basetime[basetime_index];
486 * Apply adjtime corrections. (adjtime() API)
488 * adjtime() only runs on cpu #0 so our critical section is
489 * sufficient to access these variables.
491 if (ntp_delta != 0) {
492 nbt->tv_nsec += ntp_tick_delta;
493 ntp_delta -= ntp_tick_delta;
494 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
495 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
496 ntp_tick_delta = ntp_delta;
501 * Apply permanent frequency corrections. (sysctl API)
503 if (ntp_tick_permanent != 0) {
504 ntp_tick_acc += ntp_tick_permanent;
505 if (ntp_tick_acc >= (1LL << 32)) {
506 nbt->tv_nsec += ntp_tick_acc >> 32;
507 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
508 } else if (ntp_tick_acc <= -(1LL << 32)) {
509 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
510 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
511 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
515 if (nbt->tv_nsec >= 1000000000) {
517 nbt->tv_nsec -= 1000000000;
518 } else if (nbt->tv_nsec < 0) {
520 nbt->tv_nsec += 1000000000;
524 * Another per-tick compensation. (for ntp_adjtime() API)
527 nsec_acc += nsec_adj;
528 if (nsec_acc >= 0x100000000LL) {
529 nbt->tv_nsec += nsec_acc >> 32;
530 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
531 } else if (nsec_acc <= -0x100000000LL) {
532 nbt->tv_nsec -= -nsec_acc >> 32;
533 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
535 if (nbt->tv_nsec >= 1000000000) {
536 nbt->tv_nsec -= 1000000000;
538 } else if (nbt->tv_nsec < 0) {
539 nbt->tv_nsec += 1000000000;
544 /************************************************************
545 * LEAP SECOND CORRECTION *
546 ************************************************************
548 * Taking into account all the corrections made above, figure
549 * out the new real time. If the seconds field has changed
550 * then apply any pending leap-second corrections.
552 getnanotime_nbt(nbt, &nts);
554 if (time_second != nts.tv_sec) {
556 * Apply leap second (sysctl API). Adjust nts for changes
557 * so we do not have to call getnanotime_nbt again.
559 if (ntp_leap_second) {
560 if (ntp_leap_second == nts.tv_sec) {
561 if (ntp_leap_insert) {
573 * Apply leap second (ntp_adjtime() API), calculate a new
574 * nsec_adj field. ntp_update_second() returns nsec_adj
575 * as a per-second value but we need it as a per-tick value.
577 leap = ntp_update_second(time_second, &nsec_adj);
583 * Update the time_second 'approximate time' global.
585 time_second = nts.tv_sec;
589 * Finally, our new basetime is ready to go live!
595 * Update kpmap on each tick. TS updates are integrated with
596 * fences and upticks allowing userland to read the data
602 w = (kpmap->upticks + 1) & 1;
603 getnanouptime(&kpmap->ts_uptime[w]);
604 getnanotime(&kpmap->ts_realtime[w]);
612 * lwkt thread scheduler fair queueing
614 lwkt_schedulerclock(curthread);
617 * softticks are handled for all cpus
619 hardclock_softtick(gd);
622 * ITimer handling is per-tick, per-cpu.
624 * We must acquire the per-process token in order for ksignal()
625 * to be non-blocking. For the moment this requires an AST fault,
626 * the ksignal() cannot be safely issued from this hard interrupt.
628 * XXX Even the trytoken here isn't right, and itimer operation in
629 * a multi threaded environment is going to be weird at the
632 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
635 ++p->p_upmap->runticks;
637 if (frame && CLKF_USERMODE(frame) &&
638 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
639 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
640 p->p_flags |= P_SIGVTALRM;
643 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
644 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
645 p->p_flags |= P_SIGPROF;
649 lwkt_reltoken(&p->p_token);
655 * The statistics clock typically runs at a 125Hz rate, and is intended
656 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
658 * NOTE! systimer! the MP lock might not be held here. We can only safely
659 * manipulate objects owned by the current cpu.
661 * The stats clock is responsible for grabbing a profiling sample.
662 * Most of the statistics are only used by user-level statistics programs.
663 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
666 * Like the other clocks, the stat clock is called from what is effectively
667 * a fast interrupt, so the context should be the thread/process that got
671 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
684 * How big was our timeslice relative to the last time? Calculate
687 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
688 * during early boot. Just use the systimer count to be nice
689 * to e.g. qemu. The systimer has a better chance of being
690 * MPSAFE at early boot.
692 cv = sys_cputimer->count();
693 scv = mycpu->statint.gd_statcv;
697 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32;
703 mycpu->statint.gd_statcv = cv;
706 stv = &mycpu->gd_stattv;
707 if (stv->tv_sec == 0) {
710 bump = tv.tv_usec - stv->tv_usec +
711 (tv.tv_sec - stv->tv_sec) * 1000000;
723 if (frame && CLKF_USERMODE(frame)) {
725 * Came from userland, handle user time and deal with
728 if (p && (p->p_flags & P_PROFIL))
729 addupc_intr(p, CLKF_PC(frame), 1);
730 td->td_uticks += bump;
733 * Charge the time as appropriate
735 if (p && p->p_nice > NZERO)
736 cpu_time.cp_nice += bump;
738 cpu_time.cp_user += bump;
740 int intr_nest = mycpu->gd_intr_nesting_level;
744 * IPI processing code will bump gd_intr_nesting_level
745 * up by one, which breaks following CLKF_INTR testing,
746 * so we subtract it by one here.
752 * Kernel statistics are just like addupc_intr, only easier.
755 if (g->state == GMON_PROF_ON && frame) {
756 i = CLKF_PC(frame) - g->lowpc;
757 if (i < g->textsize) {
758 i /= HISTFRACTION * sizeof(*g->kcount);
764 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
767 * Came from kernel mode, so we were:
768 * - handling an interrupt,
769 * - doing syscall or trap work on behalf of the current
771 * - spinning in the idle loop.
772 * Whichever it is, charge the time as appropriate.
773 * Note that we charge interrupts to the current process,
774 * regardless of whether they are ``for'' that process,
775 * so that we know how much of its real time was spent
776 * in ``non-process'' (i.e., interrupt) work.
778 * XXX assume system if frame is NULL. A NULL frame
779 * can occur if ipi processing is done from a crit_exit().
782 td->td_iticks += bump;
784 td->td_sticks += bump;
786 if (IS_INTR_RUNNING) {
788 * If we interrupted an interrupt thread, well,
789 * count it as interrupt time.
793 do_pctrack(frame, PCTRACK_INT);
795 cpu_time.cp_intr += bump;
797 if (td == &mycpu->gd_idlethread) {
799 * Even if the current thread is the idle
800 * thread it could be due to token contention
801 * in the LWKT scheduler. Count such as
804 if (mycpu->gd_reqflags & RQF_IDLECHECK_WK_MASK)
805 cpu_time.cp_sys += bump;
807 cpu_time.cp_idle += bump;
810 * System thread was running.
814 do_pctrack(frame, PCTRACK_SYS);
816 cpu_time.cp_sys += bump;
820 #undef IS_INTR_RUNNING
826 * Sample the PC when in the kernel or in an interrupt. User code can
827 * retrieve the information and generate a histogram or other output.
831 do_pctrack(struct intrframe *frame, int which)
833 struct kinfo_pctrack *pctrack;
835 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
836 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
837 (void *)CLKF_PC(frame);
842 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
844 struct kinfo_pcheader head;
849 head.pc_ntrack = PCTRACK_SIZE;
850 head.pc_arysize = PCTRACK_ARYSIZE;
852 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
855 for (cpu = 0; cpu < ncpus; ++cpu) {
856 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
857 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
858 sizeof(struct kinfo_pctrack));
867 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
868 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
873 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
874 * the MP lock might not be held. We can safely manipulate parts of curproc
875 * but that's about it.
877 * Each cpu has its own scheduler clock.
880 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
887 if ((lp = lwkt_preempted_proc()) != NULL) {
889 * Account for cpu time used and hit the scheduler. Note
890 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
894 usched_schedulerclock(lp, info->periodic, info->time);
896 usched_schedulerclock(NULL, info->periodic, info->time);
898 if ((lp = curthread->td_lwp) != NULL) {
900 * Update resource usage integrals and maximums.
902 if ((ru = &lp->lwp_proc->p_ru) &&
903 (vm = lp->lwp_proc->p_vmspace) != NULL) {
904 ru->ru_ixrss += pgtok(vm->vm_tsize);
905 ru->ru_idrss += pgtok(vm->vm_dsize);
906 ru->ru_isrss += pgtok(vm->vm_ssize);
907 if (lwkt_trytoken(&vm->vm_map.token)) {
908 rss = pgtok(vmspace_resident_count(vm));
909 if (ru->ru_maxrss < rss)
911 lwkt_reltoken(&vm->vm_map.token);
915 /* Increment the global sched_ticks */
916 if (mycpu->gd_cpuid == 0)
921 * Compute number of ticks for the specified amount of time. The
922 * return value is intended to be used in a clock interrupt timed
923 * operation and guaranteed to meet or exceed the requested time.
924 * If the representation overflows, return INT_MAX. The minimum return
925 * value is 1 ticks and the function will average the calculation up.
926 * If any value greater then 0 microseconds is supplied, a value
927 * of at least 2 will be returned to ensure that a near-term clock
928 * interrupt does not cause the timeout to occur (degenerately) early.
930 * Note that limit checks must take into account microseconds, which is
931 * done simply by using the smaller signed long maximum instead of
932 * the unsigned long maximum.
934 * If ints have 32 bits, then the maximum value for any timeout in
935 * 10ms ticks is 248 days.
938 tvtohz_high(struct timeval *tv)
955 kprintf("tvtohz_high: negative time difference "
956 "%ld sec %ld usec\n",
960 } else if (sec <= INT_MAX / hz) {
961 ticks = (int)(sec * hz +
962 ((u_long)usec + (ustick - 1)) / ustick) + 1;
970 tstohz_high(struct timespec *ts)
987 kprintf("tstohz_high: negative time difference "
988 "%ld sec %ld nsec\n",
992 } else if (sec <= INT_MAX / hz) {
993 ticks = (int)(sec * hz +
994 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
1003 * Compute number of ticks for the specified amount of time, erroring on
1004 * the side of it being too low to ensure that sleeping the returned number
1005 * of ticks will not result in a late return.
1007 * The supplied timeval may not be negative and should be normalized. A
1008 * return value of 0 is possible if the timeval converts to less then
1011 * If ints have 32 bits, then the maximum value for any timeout in
1012 * 10ms ticks is 248 days.
1015 tvtohz_low(struct timeval *tv)
1021 if (sec <= INT_MAX / hz)
1022 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
1029 tstohz_low(struct timespec *ts)
1035 if (sec <= INT_MAX / hz)
1036 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1043 * Start profiling on a process.
1045 * Kernel profiling passes proc0 which never exits and hence
1046 * keeps the profile clock running constantly.
1049 startprofclock(struct proc *p)
1051 if ((p->p_flags & P_PROFIL) == 0) {
1052 p->p_flags |= P_PROFIL;
1054 if (++profprocs == 1 && stathz != 0) {
1057 setstatclockrate(profhz);
1065 * Stop profiling on a process.
1067 * caller must hold p->p_token
1070 stopprofclock(struct proc *p)
1072 if (p->p_flags & P_PROFIL) {
1073 p->p_flags &= ~P_PROFIL;
1075 if (--profprocs == 0 && stathz != 0) {
1078 setstatclockrate(stathz);
1086 * Return information about system clocks.
1089 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1091 struct kinfo_clockinfo clkinfo;
1093 * Construct clockinfo structure.
1096 clkinfo.ci_tick = ustick;
1097 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1098 clkinfo.ci_profhz = profhz;
1099 clkinfo.ci_stathz = stathz ? stathz : hz;
1100 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1103 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1104 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1107 * We have eight functions for looking at the clock, four for
1108 * microseconds and four for nanoseconds. For each there is fast
1109 * but less precise version "get{nano|micro}[up]time" which will
1110 * return a time which is up to 1/HZ previous to the call, whereas
1111 * the raw version "{nano|micro}[up]time" will return a timestamp
1112 * which is as precise as possible. The "up" variants return the
1113 * time relative to system boot, these are well suited for time
1114 * interval measurements.
1116 * Each cpu independently maintains the current time of day, so all
1117 * we need to do to protect ourselves from changes is to do a loop
1118 * check on the seconds field changing out from under us.
1120 * The system timer maintains a 32 bit count and due to various issues
1121 * it is possible for the calculated delta to occasionally exceed
1122 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1123 * multiplication can easily overflow, so we deal with the case. For
1124 * uniformity we deal with the case in the usec case too.
1126 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1129 getmicrouptime(struct timeval *tvp)
1131 struct globaldata *gd = mycpu;
1135 tvp->tv_sec = gd->gd_time_seconds;
1136 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1137 } while (tvp->tv_sec != gd->gd_time_seconds);
1139 if (delta >= sys_cputimer->freq) {
1140 tvp->tv_sec += delta / sys_cputimer->freq;
1141 delta %= sys_cputimer->freq;
1143 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1144 if (tvp->tv_usec >= 1000000) {
1145 tvp->tv_usec -= 1000000;
1151 getnanouptime(struct timespec *tsp)
1153 struct globaldata *gd = mycpu;
1157 tsp->tv_sec = gd->gd_time_seconds;
1158 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1159 } while (tsp->tv_sec != gd->gd_time_seconds);
1161 if (delta >= sys_cputimer->freq) {
1162 tsp->tv_sec += delta / sys_cputimer->freq;
1163 delta %= sys_cputimer->freq;
1165 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1169 microuptime(struct timeval *tvp)
1171 struct globaldata *gd = mycpu;
1175 tvp->tv_sec = gd->gd_time_seconds;
1176 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1177 } while (tvp->tv_sec != gd->gd_time_seconds);
1179 if (delta >= sys_cputimer->freq) {
1180 tvp->tv_sec += delta / sys_cputimer->freq;
1181 delta %= sys_cputimer->freq;
1183 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1187 nanouptime(struct timespec *tsp)
1189 struct globaldata *gd = mycpu;
1193 tsp->tv_sec = gd->gd_time_seconds;
1194 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1195 } while (tsp->tv_sec != gd->gd_time_seconds);
1197 if (delta >= sys_cputimer->freq) {
1198 tsp->tv_sec += delta / sys_cputimer->freq;
1199 delta %= sys_cputimer->freq;
1201 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1208 getmicrotime(struct timeval *tvp)
1210 struct globaldata *gd = mycpu;
1211 struct timespec *bt;
1215 tvp->tv_sec = gd->gd_time_seconds;
1216 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1217 } while (tvp->tv_sec != gd->gd_time_seconds);
1219 if (delta >= sys_cputimer->freq) {
1220 tvp->tv_sec += delta / sys_cputimer->freq;
1221 delta %= sys_cputimer->freq;
1223 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1225 bt = &basetime[basetime_index];
1227 tvp->tv_sec += bt->tv_sec;
1228 tvp->tv_usec += bt->tv_nsec / 1000;
1229 while (tvp->tv_usec >= 1000000) {
1230 tvp->tv_usec -= 1000000;
1236 getnanotime(struct timespec *tsp)
1238 struct globaldata *gd = mycpu;
1239 struct timespec *bt;
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 bt = &basetime[basetime_index];
1255 tsp->tv_sec += bt->tv_sec;
1256 tsp->tv_nsec += bt->tv_nsec;
1257 while (tsp->tv_nsec >= 1000000000) {
1258 tsp->tv_nsec -= 1000000000;
1264 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1266 struct globaldata *gd = mycpu;
1270 tsp->tv_sec = gd->gd_time_seconds;
1271 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1272 } while (tsp->tv_sec != gd->gd_time_seconds);
1274 if (delta >= sys_cputimer->freq) {
1275 tsp->tv_sec += delta / sys_cputimer->freq;
1276 delta %= sys_cputimer->freq;
1278 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1280 tsp->tv_sec += nbt->tv_sec;
1281 tsp->tv_nsec += nbt->tv_nsec;
1282 while (tsp->tv_nsec >= 1000000000) {
1283 tsp->tv_nsec -= 1000000000;
1290 microtime(struct timeval *tvp)
1292 struct globaldata *gd = mycpu;
1293 struct timespec *bt;
1297 tvp->tv_sec = gd->gd_time_seconds;
1298 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1299 } while (tvp->tv_sec != gd->gd_time_seconds);
1301 if (delta >= sys_cputimer->freq) {
1302 tvp->tv_sec += delta / sys_cputimer->freq;
1303 delta %= sys_cputimer->freq;
1305 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1307 bt = &basetime[basetime_index];
1309 tvp->tv_sec += bt->tv_sec;
1310 tvp->tv_usec += bt->tv_nsec / 1000;
1311 while (tvp->tv_usec >= 1000000) {
1312 tvp->tv_usec -= 1000000;
1318 nanotime(struct timespec *tsp)
1320 struct globaldata *gd = mycpu;
1321 struct timespec *bt;
1325 tsp->tv_sec = gd->gd_time_seconds;
1326 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1327 } while (tsp->tv_sec != gd->gd_time_seconds);
1329 if (delta >= sys_cputimer->freq) {
1330 tsp->tv_sec += delta / sys_cputimer->freq;
1331 delta %= sys_cputimer->freq;
1333 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1335 bt = &basetime[basetime_index];
1337 tsp->tv_sec += bt->tv_sec;
1338 tsp->tv_nsec += bt->tv_nsec;
1339 while (tsp->tv_nsec >= 1000000000) {
1340 tsp->tv_nsec -= 1000000000;
1346 * Get an approximate time_t. It does not have to be accurate. This
1347 * function is called only from KTR and can be called with the system in
1348 * any state so do not use a critical section or other complex operation
1351 * NOTE: This is not exactly synchronized with real time. To do that we
1352 * would have to do what microtime does and check for a nanoseconds
1356 get_approximate_time_t(void)
1358 struct globaldata *gd = mycpu;
1359 struct timespec *bt;
1361 bt = &basetime[basetime_index];
1362 return(gd->gd_time_seconds + bt->tv_sec);
1366 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1369 struct pps_fetch_args *fapi;
1371 struct pps_kcbind_args *kapi;
1375 case PPS_IOC_CREATE:
1377 case PPS_IOC_DESTROY:
1379 case PPS_IOC_SETPARAMS:
1380 app = (pps_params_t *)data;
1381 if (app->mode & ~pps->ppscap)
1383 pps->ppsparam = *app;
1385 case PPS_IOC_GETPARAMS:
1386 app = (pps_params_t *)data;
1387 *app = pps->ppsparam;
1388 app->api_version = PPS_API_VERS_1;
1390 case PPS_IOC_GETCAP:
1391 *(int*)data = pps->ppscap;
1394 fapi = (struct pps_fetch_args *)data;
1395 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1397 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1398 return (EOPNOTSUPP);
1399 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1400 fapi->pps_info_buf = pps->ppsinfo;
1402 case PPS_IOC_KCBIND:
1404 kapi = (struct pps_kcbind_args *)data;
1405 /* XXX Only root should be able to do this */
1406 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1408 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1410 if (kapi->edge & ~pps->ppscap)
1412 pps->kcmode = kapi->edge;
1415 return (EOPNOTSUPP);
1423 pps_init(struct pps_state *pps)
1425 pps->ppscap |= PPS_TSFMT_TSPEC;
1426 if (pps->ppscap & PPS_CAPTUREASSERT)
1427 pps->ppscap |= PPS_OFFSETASSERT;
1428 if (pps->ppscap & PPS_CAPTURECLEAR)
1429 pps->ppscap |= PPS_OFFSETCLEAR;
1433 pps_event(struct pps_state *pps, sysclock_t count, int event)
1435 struct globaldata *gd;
1436 struct timespec *tsp;
1437 struct timespec *osp;
1438 struct timespec *bt;
1456 /* Things would be easier with arrays... */
1457 if (event == PPS_CAPTUREASSERT) {
1458 tsp = &pps->ppsinfo.assert_timestamp;
1459 osp = &pps->ppsparam.assert_offset;
1460 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1461 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1462 pcount = &pps->ppscount[0];
1463 pseq = &pps->ppsinfo.assert_sequence;
1465 tsp = &pps->ppsinfo.clear_timestamp;
1466 osp = &pps->ppsparam.clear_offset;
1467 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1468 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1469 pcount = &pps->ppscount[1];
1470 pseq = &pps->ppsinfo.clear_sequence;
1473 /* Nothing really happened */
1474 if (*pcount == count)
1480 ts.tv_sec = gd->gd_time_seconds;
1481 delta = count - gd->gd_cpuclock_base;
1482 } while (ts.tv_sec != gd->gd_time_seconds);
1484 if (delta >= sys_cputimer->freq) {
1485 ts.tv_sec += delta / sys_cputimer->freq;
1486 delta %= sys_cputimer->freq;
1488 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1489 ni = basetime_index;
1492 ts.tv_sec += bt->tv_sec;
1493 ts.tv_nsec += bt->tv_nsec;
1494 while (ts.tv_nsec >= 1000000000) {
1495 ts.tv_nsec -= 1000000000;
1503 timespecadd(tsp, osp);
1504 if (tsp->tv_nsec < 0) {
1505 tsp->tv_nsec += 1000000000;
1511 /* magic, at its best... */
1512 tcount = count - pps->ppscount[2];
1513 pps->ppscount[2] = count;
1514 if (tcount >= sys_cputimer->freq) {
1515 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1516 sys_cputimer->freq64_nsec *
1517 (tcount % sys_cputimer->freq)) >> 32;
1519 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1521 hardpps(tsp, delta);
1527 * Return the tsc target value for a delay of (ns).
1529 * Returns -1 if the TSC is not supported.
1532 tsc_get_target(int ns)
1534 #if defined(_RDTSC_SUPPORTED_)
1535 if (cpu_feature & CPUID_TSC) {
1536 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1543 * Compare the tsc against the passed target
1545 * Returns +1 if the target has been reached
1546 * Returns 0 if the target has not yet been reached
1547 * Returns -1 if the TSC is not supported.
1549 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1552 tsc_test_target(int64_t target)
1554 #if defined(_RDTSC_SUPPORTED_)
1555 if (cpu_feature & CPUID_TSC) {
1556 if ((int64_t)(target - rdtsc()) <= 0)
1565 * Delay the specified number of nanoseconds using the tsc. This function
1566 * returns immediately if the TSC is not supported. At least one cpu_pause()
1574 clk = tsc_get_target(ns);
1576 while (tsc_test_target(clk) == 0)