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>
86 #include <sys/timex.h>
87 #include <sys/timepps.h>
88 #include <sys/upmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_extern.h>
94 #include <sys/sysctl.h>
96 #include <sys/thread2.h>
97 #include <sys/spinlock2.h>
99 #include <machine/cpu.h>
100 #include <machine/limits.h>
101 #include <machine/smp.h>
102 #include <machine/cpufunc.h>
103 #include <machine/specialreg.h>
104 #include <machine/clock.h>
107 #include <sys/gmon.h>
111 extern void ifpoll_init_pcpu(int);
115 static void do_pctrack(struct intrframe *frame, int which);
118 static void initclocks (void *dummy);
119 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
122 * Some of these don't belong here, but it's easiest to concentrate them.
123 * Note that cpu_time counts in microseconds, but most userland programs
124 * just compare relative times against the total by delta.
126 struct kinfo_cputime cputime_percpu[MAXCPU];
128 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
129 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
133 sysctl_cputime(SYSCTL_HANDLER_ARGS)
137 size_t size = sizeof(struct kinfo_cputime);
138 struct kinfo_cputime tmp;
141 * NOTE: For security reasons, only root can sniff %rip
143 root_error = priv_check_cred(curthread->td_ucred, PRIV_ROOT, 0);
145 for (cpu = 0; cpu < ncpus; ++cpu) {
146 tmp = cputime_percpu[cpu];
147 if (root_error == 0) {
149 (int64_t)globaldata_find(cpu)->gd_sample_pc;
151 (int64_t)globaldata_find(cpu)->gd_sample_sp;
153 if ((error = SYSCTL_OUT(req, &tmp, size)) != 0)
162 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
163 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
166 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
168 long cpu_states[5] = {0};
170 size_t size = sizeof(cpu_states);
172 for (cpu = 0; cpu < ncpus; ++cpu) {
173 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
174 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
175 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
176 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
177 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
180 error = SYSCTL_OUT(req, cpu_states, size);
185 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
186 sysctl_cp_time, "LU", "CPU time statistics");
189 * boottime is used to calculate the 'real' uptime. Do not confuse this with
190 * microuptime(). microtime() is not drift compensated. The real uptime
191 * with compensation is nanotime() - bootime. boottime is recalculated
192 * whenever the real time is set based on the compensated elapsed time
193 * in seconds (gd->gd_time_seconds).
195 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
196 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
199 * WARNING! time_second can backstep on time corrections. Also, unlike
200 * time_second, time_uptime is not a "real" time_t (seconds
201 * since the Epoch) but seconds since booting.
203 struct timespec boottime; /* boot time (realtime) for reference only */
204 time_t time_second; /* read-only 'passive' realtime in seconds */
205 time_t time_uptime; /* read-only 'passive' uptime in seconds */
208 * basetime is used to calculate the compensated real time of day. The
209 * basetime can be modified on a per-tick basis by the adjtime(),
210 * ntp_adjtime(), and sysctl-based time correction APIs.
212 * Note that frequency corrections can also be made by adjusting
215 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
216 * used on both SMP and UP systems to avoid MP races between cpu's and
217 * interrupt races on UP systems.
220 __uint32_t time_second;
221 sysclock_t cpuclock_base;
224 #define BASETIME_ARYSIZE 16
225 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
226 static struct timespec basetime[BASETIME_ARYSIZE];
227 static struct hardtime hardtime[BASETIME_ARYSIZE];
228 static volatile int basetime_index;
231 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
238 * Because basetime data and index may be updated by another cpu,
239 * a load fence is required to ensure that the data we read has
240 * not been speculatively read relative to a possibly updated index.
242 index = basetime_index;
244 bt = &basetime[index];
245 error = SYSCTL_OUT(req, bt, sizeof(*bt));
249 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
250 &boottime, timespec, "System boottime");
251 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
252 sysctl_get_basetime, "S,timespec", "System basetime");
254 static void hardclock(systimer_t info, int, struct intrframe *frame);
255 static void statclock(systimer_t info, int, struct intrframe *frame);
256 static void schedclock(systimer_t info, int, struct intrframe *frame);
257 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
259 int ticks; /* system master ticks at hz */
260 int clocks_running; /* tsleep/timeout clocks operational */
261 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
262 int64_t nsec_acc; /* accumulator */
263 int sched_ticks; /* global schedule clock ticks */
265 /* NTPD time correction fields */
266 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
267 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
268 int64_t ntp_delta; /* one-time correction in nsec */
269 int64_t ntp_big_delta = 1000000000;
270 int32_t ntp_tick_delta; /* current adjustment rate */
271 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
272 time_t ntp_leap_second; /* time of next leap second */
273 int ntp_leap_insert; /* whether to insert or remove a second */
274 struct spinlock ntp_spin;
277 * Finish initializing clock frequencies and start all clocks running.
281 initclocks(void *dummy)
283 /*psratio = profhz / stathz;*/
284 spin_init(&ntp_spin, "ntp");
288 kpmap->tsc_freq = (uint64_t)tsc_frequency;
289 kpmap->tick_freq = hz;
294 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
295 * during SMP initialization.
297 * This routine is called concurrently during low-level SMP initialization
298 * and may not block in any way. Meaning, among other things, we can't
299 * acquire any tokens.
302 initclocks_pcpu(void)
304 struct globaldata *gd = mycpu;
307 if (gd->gd_cpuid == 0) {
308 gd->gd_time_seconds = 1;
309 gd->gd_cpuclock_base = sys_cputimer->count();
310 hardtime[0].time_second = gd->gd_time_seconds;
311 hardtime[0].cpuclock_base = gd->gd_cpuclock_base;
313 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
314 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
317 systimer_intr_enable();
323 * This routine is called on just the BSP, just after SMP initialization
324 * completes to * finish initializing any clocks that might contend/block
325 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
326 * that function is called from the idle thread bootstrap for each cpu and
327 * not allowed to block at all.
331 initclocks_other(void *dummy)
333 struct globaldata *ogd = mycpu;
334 struct globaldata *gd;
337 for (n = 0; n < ncpus; ++n) {
338 lwkt_setcpu_self(globaldata_find(n));
342 * Use a non-queued periodic systimer to prevent multiple
343 * ticks from building up if the sysclock jumps forward
344 * (8254 gets reset). The sysclock will never jump backwards.
345 * Our time sync is based on the actual sysclock, not the
348 * Install statclock before hardclock to prevent statclock
349 * from misinterpreting gd_flags for tick assignment when
352 systimer_init_periodic_nq(&gd->gd_statclock, statclock,
354 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock,
356 /* XXX correct the frequency for scheduler / estcpu tests */
357 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
360 ifpoll_init_pcpu(gd->gd_cpuid);
363 lwkt_setcpu_self(ogd);
365 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL);
368 * This sets the current real time of day. Timespecs are in seconds and
369 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
370 * instead we adjust basetime so basetime + gd_* results in the current
371 * time of day. This way the gd_* fields are guaranteed to represent
372 * a monotonically increasing 'uptime' value.
374 * When set_timeofday() is called from userland, the system call forces it
375 * onto cpu #0 since only cpu #0 can update basetime_index.
378 set_timeofday(struct timespec *ts)
380 struct timespec *nbt;
384 * XXX SMP / non-atomic basetime updates
387 ni = (basetime_index + 1) & BASETIME_ARYMASK;
391 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
392 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
393 if (nbt->tv_nsec < 0) {
394 nbt->tv_nsec += 1000000000;
399 * Note that basetime diverges from boottime as the clock drift is
400 * compensated for, so we cannot do away with boottime. When setting
401 * the absolute time of day the drift is 0 (for an instant) and we
402 * can simply assign boottime to basetime.
404 * Note that nanouptime() is based on gd_time_seconds which is drift
405 * compensated up to a point (it is guaranteed to remain monotonically
406 * increasing). gd_time_seconds is thus our best uptime guess and
407 * suitable for use in the boottime calculation. It is already taken
408 * into account in the basetime calculation above.
410 spin_lock(&ntp_spin);
411 boottime.tv_sec = nbt->tv_sec;
415 * We now have a new basetime, make sure all other cpus have it,
416 * then update the index.
420 spin_unlock(&ntp_spin);
426 * Each cpu has its own hardclock, but we only increments ticks and softticks
429 * NOTE! systimer! the MP lock might not be held here. We can only safely
430 * manipulate objects owned by the current cpu.
433 hardclock(systimer_t info, int in_ipi, struct intrframe *frame)
437 struct globaldata *gd = mycpu;
439 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) {
440 /* Defer to doreti on passive IPIQ processing */
445 * We update the compensation base to calculate fine-grained time
446 * from the sys_cputimer on a per-cpu basis in order to avoid
447 * having to mess around with locks. sys_cputimer is assumed to
448 * be consistent across all cpus. CPU N copies the base state from
449 * CPU 0 using the same FIFO trick that we use for basetime (so we
450 * don't catch a CPU 0 update in the middle).
452 * Note that we never allow info->time (aka gd->gd_hardclock.time)
453 * to reverse index gd_cpuclock_base, but that it is possible for
454 * it to temporarily get behind in the seconds if something in the
455 * system locks interrupts for a long period of time. Since periodic
456 * timers count events, though everything should resynch again
459 if (gd->gd_cpuid == 0) {
462 cputicks = info->time - gd->gd_cpuclock_base;
463 if (cputicks >= sys_cputimer->freq) {
464 cputicks /= sys_cputimer->freq;
465 if (cputicks != 0 && cputicks != 1)
466 kprintf("Warning: hardclock missed > 1 sec\n");
467 gd->gd_time_seconds += cputicks;
468 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks;
469 /* uncorrected monotonic 1-sec gran */
470 time_uptime += cputicks;
472 ni = (basetime_index + 1) & BASETIME_ARYMASK;
473 hardtime[ni].time_second = gd->gd_time_seconds;
474 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base;
480 gd->gd_time_seconds = hardtime[ni].time_second;
481 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base;
485 * The system-wide ticks counter and NTP related timedelta/tickdelta
486 * adjustments only occur on cpu #0. NTP adjustments are accomplished
487 * by updating basetime.
489 if (gd->gd_cpuid == 0) {
490 struct timespec *nbt;
498 if (tco->tc_poll_pps)
499 tco->tc_poll_pps(tco);
503 * Calculate the new basetime index. We are in a critical section
504 * on cpu #0 and can safely play with basetime_index. Start
505 * with the current basetime and then make adjustments.
507 ni = (basetime_index + 1) & BASETIME_ARYMASK;
509 *nbt = basetime[basetime_index];
512 * ntp adjustments only occur on cpu 0 and are protected by
513 * ntp_spin. This spinlock virtually never conflicts.
515 spin_lock(&ntp_spin);
518 * Apply adjtime corrections. (adjtime() API)
520 * adjtime() only runs on cpu #0 so our critical section is
521 * sufficient to access these variables.
523 if (ntp_delta != 0) {
524 nbt->tv_nsec += ntp_tick_delta;
525 ntp_delta -= ntp_tick_delta;
526 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
527 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
528 ntp_tick_delta = ntp_delta;
533 * Apply permanent frequency corrections. (sysctl API)
535 if (ntp_tick_permanent != 0) {
536 ntp_tick_acc += ntp_tick_permanent;
537 if (ntp_tick_acc >= (1LL << 32)) {
538 nbt->tv_nsec += ntp_tick_acc >> 32;
539 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
540 } else if (ntp_tick_acc <= -(1LL << 32)) {
541 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
542 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
543 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
547 if (nbt->tv_nsec >= 1000000000) {
549 nbt->tv_nsec -= 1000000000;
550 } else if (nbt->tv_nsec < 0) {
552 nbt->tv_nsec += 1000000000;
556 * Another per-tick compensation. (for ntp_adjtime() API)
559 nsec_acc += nsec_adj;
560 if (nsec_acc >= 0x100000000LL) {
561 nbt->tv_nsec += nsec_acc >> 32;
562 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
563 } else if (nsec_acc <= -0x100000000LL) {
564 nbt->tv_nsec -= -nsec_acc >> 32;
565 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
567 if (nbt->tv_nsec >= 1000000000) {
568 nbt->tv_nsec -= 1000000000;
570 } else if (nbt->tv_nsec < 0) {
571 nbt->tv_nsec += 1000000000;
575 spin_unlock(&ntp_spin);
577 /************************************************************
578 * LEAP SECOND CORRECTION *
579 ************************************************************
581 * Taking into account all the corrections made above, figure
582 * out the new real time. If the seconds field has changed
583 * then apply any pending leap-second corrections.
585 getnanotime_nbt(nbt, &nts);
587 if (time_second != nts.tv_sec) {
589 * Apply leap second (sysctl API). Adjust nts for changes
590 * so we do not have to call getnanotime_nbt again.
592 if (ntp_leap_second) {
593 if (ntp_leap_second == nts.tv_sec) {
594 if (ntp_leap_insert) {
606 * Apply leap second (ntp_adjtime() API), calculate a new
607 * nsec_adj field. ntp_update_second() returns nsec_adj
608 * as a per-second value but we need it as a per-tick value.
610 leap = ntp_update_second(time_second, &nsec_adj);
616 * Update the time_second 'approximate time' global.
618 time_second = nts.tv_sec;
622 * Finally, our new basetime is ready to go live!
628 * Update kpmap on each tick. TS updates are integrated with
629 * fences and upticks allowing userland to read the data
635 w = (kpmap->upticks + 1) & 1;
636 getnanouptime(&kpmap->ts_uptime[w]);
637 getnanotime(&kpmap->ts_realtime[w]);
645 * lwkt thread scheduler fair queueing
647 lwkt_schedulerclock(curthread);
650 * softticks are handled for all cpus
652 hardclock_softtick(gd);
655 * Rollup accumulated vmstats, copy-back for critical path checks.
657 vmstats_rollup_cpu(gd);
658 mycpu->gd_vmstats = vmstats;
661 * ITimer handling is per-tick, per-cpu.
663 * We must acquire the per-process token in order for ksignal()
664 * to be non-blocking. For the moment this requires an AST fault,
665 * the ksignal() cannot be safely issued from this hard interrupt.
667 * XXX Even the trytoken here isn't right, and itimer operation in
668 * a multi threaded environment is going to be weird at the
671 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
674 ++p->p_upmap->runticks;
676 if (frame && CLKF_USERMODE(frame) &&
677 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
678 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
679 p->p_flags |= P_SIGVTALRM;
682 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
683 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
684 p->p_flags |= P_SIGPROF;
688 lwkt_reltoken(&p->p_token);
694 * The statistics clock typically runs at a 125Hz rate, and is intended
695 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
697 * NOTE! systimer! the MP lock might not be held here. We can only safely
698 * manipulate objects owned by the current cpu.
700 * The stats clock is responsible for grabbing a profiling sample.
701 * Most of the statistics are only used by user-level statistics programs.
702 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
705 * Like the other clocks, the stat clock is called from what is effectively
706 * a fast interrupt, so the context should be the thread/process that got
710 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
716 globaldata_t gd = mycpu;
724 * How big was our timeslice relative to the last time? Calculate
727 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
728 * during early boot. Just use the systimer count to be nice
729 * to e.g. qemu. The systimer has a better chance of being
730 * MPSAFE at early boot.
732 cv = sys_cputimer->count();
733 scv = gd->statint.gd_statcv;
737 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32;
743 gd->statint.gd_statcv = cv;
746 stv = &gd->gd_stattv;
747 if (stv->tv_sec == 0) {
750 bump = tv.tv_usec - stv->tv_usec +
751 (tv.tv_sec - stv->tv_sec) * 1000000;
763 if (frame && CLKF_USERMODE(frame)) {
765 * Came from userland, handle user time and deal with
768 if (p && (p->p_flags & P_PROFIL))
769 addupc_intr(p, CLKF_PC(frame), 1);
770 td->td_uticks += bump;
773 * Charge the time as appropriate
775 if (p && p->p_nice > NZERO)
776 cpu_time.cp_nice += bump;
778 cpu_time.cp_user += bump;
780 int intr_nest = gd->gd_intr_nesting_level;
784 * IPI processing code will bump gd_intr_nesting_level
785 * up by one, which breaks following CLKF_INTR testing,
786 * so we subtract it by one here.
792 * Kernel statistics are just like addupc_intr, only easier.
795 if (g->state == GMON_PROF_ON && frame) {
796 i = CLKF_PC(frame) - g->lowpc;
797 if (i < g->textsize) {
798 i /= HISTFRACTION * sizeof(*g->kcount);
804 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
807 * Came from kernel mode, so we were:
808 * - handling an interrupt,
809 * - doing syscall or trap work on behalf of the current
811 * - spinning in the idle loop.
812 * Whichever it is, charge the time as appropriate.
813 * Note that we charge interrupts to the current process,
814 * regardless of whether they are ``for'' that process,
815 * so that we know how much of its real time was spent
816 * in ``non-process'' (i.e., interrupt) work.
818 * XXX assume system if frame is NULL. A NULL frame
819 * can occur if ipi processing is done from a crit_exit().
821 if (IS_INTR_RUNNING) {
823 * If we interrupted an interrupt thread, well,
824 * count it as interrupt time.
826 td->td_iticks += bump;
829 do_pctrack(frame, PCTRACK_INT);
831 cpu_time.cp_intr += bump;
832 #ifdef _KERNEL_VIRTUAL
833 } else if (gd->gd_flags & GDF_VIRTUSER) {
835 * The vkernel doesn't do a good job providing trap
836 * frames that we can test. If the GDF_VIRTUSER
837 * flag is set we probably interrupted user mode.
839 td->td_uticks += bump;
842 * Charge the time as appropriate
844 if (p && p->p_nice > NZERO)
845 cpu_time.cp_nice += bump;
847 cpu_time.cp_user += bump;
850 td->td_sticks += bump;
851 if (td == &gd->gd_idlethread) {
853 * Token contention can cause us to mis-count
854 * a contended as idle, but it doesn't work
855 * properly for VKERNELs so just test on a
858 #ifdef _KERNEL_VIRTUAL
859 cpu_time.cp_idle += bump;
861 if (mycpu->gd_reqflags & RQF_IDLECHECK_WK_MASK)
862 cpu_time.cp_sys += bump;
864 cpu_time.cp_idle += bump;
868 * System thread was running.
872 do_pctrack(frame, PCTRACK_SYS);
874 cpu_time.cp_sys += bump;
878 #undef IS_INTR_RUNNING
884 * Sample the PC when in the kernel or in an interrupt. User code can
885 * retrieve the information and generate a histogram or other output.
889 do_pctrack(struct intrframe *frame, int which)
891 struct kinfo_pctrack *pctrack;
893 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
894 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
895 (void *)CLKF_PC(frame);
900 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
902 struct kinfo_pcheader head;
907 head.pc_ntrack = PCTRACK_SIZE;
908 head.pc_arysize = PCTRACK_ARYSIZE;
910 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
913 for (cpu = 0; cpu < ncpus; ++cpu) {
914 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
915 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
916 sizeof(struct kinfo_pctrack));
925 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
926 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
931 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
932 * the MP lock might not be held. We can safely manipulate parts of curproc
933 * but that's about it.
935 * Each cpu has its own scheduler clock.
938 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
945 if ((lp = lwkt_preempted_proc()) != NULL) {
947 * Account for cpu time used and hit the scheduler. Note
948 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
952 usched_schedulerclock(lp, info->periodic, info->time);
954 usched_schedulerclock(NULL, info->periodic, info->time);
956 if ((lp = curthread->td_lwp) != NULL) {
958 * Update resource usage integrals and maximums.
960 if ((ru = &lp->lwp_proc->p_ru) &&
961 (vm = lp->lwp_proc->p_vmspace) != NULL) {
962 ru->ru_ixrss += pgtok(vm->vm_tsize);
963 ru->ru_idrss += pgtok(vm->vm_dsize);
964 ru->ru_isrss += pgtok(vm->vm_ssize);
965 if (lwkt_trytoken(&vm->vm_map.token)) {
966 rss = pgtok(vmspace_resident_count(vm));
967 if (ru->ru_maxrss < rss)
969 lwkt_reltoken(&vm->vm_map.token);
973 /* Increment the global sched_ticks */
974 if (mycpu->gd_cpuid == 0)
979 * Compute number of ticks for the specified amount of time. The
980 * return value is intended to be used in a clock interrupt timed
981 * operation and guaranteed to meet or exceed the requested time.
982 * If the representation overflows, return INT_MAX. The minimum return
983 * value is 1 ticks and the function will average the calculation up.
984 * If any value greater then 0 microseconds is supplied, a value
985 * of at least 2 will be returned to ensure that a near-term clock
986 * interrupt does not cause the timeout to occur (degenerately) early.
988 * Note that limit checks must take into account microseconds, which is
989 * done simply by using the smaller signed long maximum instead of
990 * the unsigned long maximum.
992 * If ints have 32 bits, then the maximum value for any timeout in
993 * 10ms ticks is 248 days.
996 tvtohz_high(struct timeval *tv)
1013 kprintf("tvtohz_high: negative time difference "
1014 "%ld sec %ld usec\n",
1018 } else if (sec <= INT_MAX / hz) {
1019 ticks = (int)(sec * hz +
1020 ((u_long)usec + (ustick - 1)) / ustick) + 1;
1028 tstohz_high(struct timespec *ts)
1045 kprintf("tstohz_high: negative time difference "
1046 "%ld sec %ld nsec\n",
1050 } else if (sec <= INT_MAX / hz) {
1051 ticks = (int)(sec * hz +
1052 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
1061 * Compute number of ticks for the specified amount of time, erroring on
1062 * the side of it being too low to ensure that sleeping the returned number
1063 * of ticks will not result in a late return.
1065 * The supplied timeval may not be negative and should be normalized. A
1066 * return value of 0 is possible if the timeval converts to less then
1069 * If ints have 32 bits, then the maximum value for any timeout in
1070 * 10ms ticks is 248 days.
1073 tvtohz_low(struct timeval *tv)
1079 if (sec <= INT_MAX / hz)
1080 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
1087 tstohz_low(struct timespec *ts)
1093 if (sec <= INT_MAX / hz)
1094 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1101 * Start profiling on a process.
1103 * Caller must hold p->p_token();
1105 * Kernel profiling passes proc0 which never exits and hence
1106 * keeps the profile clock running constantly.
1109 startprofclock(struct proc *p)
1111 if ((p->p_flags & P_PROFIL) == 0) {
1112 p->p_flags |= P_PROFIL;
1114 if (++profprocs == 1 && stathz != 0) {
1117 setstatclockrate(profhz);
1125 * Stop profiling on a process.
1127 * caller must hold p->p_token
1130 stopprofclock(struct proc *p)
1132 if (p->p_flags & P_PROFIL) {
1133 p->p_flags &= ~P_PROFIL;
1135 if (--profprocs == 0 && stathz != 0) {
1138 setstatclockrate(stathz);
1146 * Return information about system clocks.
1149 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1151 struct kinfo_clockinfo clkinfo;
1153 * Construct clockinfo structure.
1156 clkinfo.ci_tick = ustick;
1157 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1158 clkinfo.ci_profhz = profhz;
1159 clkinfo.ci_stathz = stathz ? stathz : hz;
1160 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1163 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1164 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1167 * We have eight functions for looking at the clock, four for
1168 * microseconds and four for nanoseconds. For each there is fast
1169 * but less precise version "get{nano|micro}[up]time" which will
1170 * return a time which is up to 1/HZ previous to the call, whereas
1171 * the raw version "{nano|micro}[up]time" will return a timestamp
1172 * which is as precise as possible. The "up" variants return the
1173 * time relative to system boot, these are well suited for time
1174 * interval measurements.
1176 * Each cpu independently maintains the current time of day, so all
1177 * we need to do to protect ourselves from changes is to do a loop
1178 * check on the seconds field changing out from under us.
1180 * The system timer maintains a 32 bit count and due to various issues
1181 * it is possible for the calculated delta to occasionally exceed
1182 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1183 * multiplication can easily overflow, so we deal with the case. For
1184 * uniformity we deal with the case in the usec case too.
1186 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1189 getmicrouptime(struct timeval *tvp)
1191 struct globaldata *gd = mycpu;
1195 tvp->tv_sec = gd->gd_time_seconds;
1196 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1197 } while (tvp->tv_sec != gd->gd_time_seconds);
1199 if (delta >= sys_cputimer->freq) {
1200 tvp->tv_sec += delta / sys_cputimer->freq;
1201 delta %= sys_cputimer->freq;
1203 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1204 if (tvp->tv_usec >= 1000000) {
1205 tvp->tv_usec -= 1000000;
1211 getnanouptime(struct timespec *tsp)
1213 struct globaldata *gd = mycpu;
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;
1229 microuptime(struct timeval *tvp)
1231 struct globaldata *gd = mycpu;
1235 tvp->tv_sec = gd->gd_time_seconds;
1236 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1237 } while (tvp->tv_sec != gd->gd_time_seconds);
1239 if (delta >= sys_cputimer->freq) {
1240 tvp->tv_sec += delta / sys_cputimer->freq;
1241 delta %= sys_cputimer->freq;
1243 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1247 nanouptime(struct timespec *tsp)
1249 struct globaldata *gd = mycpu;
1253 tsp->tv_sec = gd->gd_time_seconds;
1254 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1255 } while (tsp->tv_sec != gd->gd_time_seconds);
1257 if (delta >= sys_cputimer->freq) {
1258 tsp->tv_sec += delta / sys_cputimer->freq;
1259 delta %= sys_cputimer->freq;
1261 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1268 getmicrotime(struct timeval *tvp)
1270 struct globaldata *gd = mycpu;
1271 struct timespec *bt;
1275 tvp->tv_sec = gd->gd_time_seconds;
1276 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1277 } while (tvp->tv_sec != gd->gd_time_seconds);
1279 if (delta >= sys_cputimer->freq) {
1280 tvp->tv_sec += delta / sys_cputimer->freq;
1281 delta %= sys_cputimer->freq;
1283 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1285 bt = &basetime[basetime_index];
1287 tvp->tv_sec += bt->tv_sec;
1288 tvp->tv_usec += bt->tv_nsec / 1000;
1289 while (tvp->tv_usec >= 1000000) {
1290 tvp->tv_usec -= 1000000;
1296 getnanotime(struct timespec *tsp)
1298 struct globaldata *gd = mycpu;
1299 struct timespec *bt;
1303 tsp->tv_sec = gd->gd_time_seconds;
1304 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1305 } while (tsp->tv_sec != gd->gd_time_seconds);
1307 if (delta >= sys_cputimer->freq) {
1308 tsp->tv_sec += delta / sys_cputimer->freq;
1309 delta %= sys_cputimer->freq;
1311 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1313 bt = &basetime[basetime_index];
1315 tsp->tv_sec += bt->tv_sec;
1316 tsp->tv_nsec += bt->tv_nsec;
1317 while (tsp->tv_nsec >= 1000000000) {
1318 tsp->tv_nsec -= 1000000000;
1324 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1326 struct globaldata *gd = mycpu;
1330 tsp->tv_sec = gd->gd_time_seconds;
1331 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1332 } while (tsp->tv_sec != gd->gd_time_seconds);
1334 if (delta >= sys_cputimer->freq) {
1335 tsp->tv_sec += delta / sys_cputimer->freq;
1336 delta %= sys_cputimer->freq;
1338 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1340 tsp->tv_sec += nbt->tv_sec;
1341 tsp->tv_nsec += nbt->tv_nsec;
1342 while (tsp->tv_nsec >= 1000000000) {
1343 tsp->tv_nsec -= 1000000000;
1350 microtime(struct timeval *tvp)
1352 struct globaldata *gd = mycpu;
1353 struct timespec *bt;
1357 tvp->tv_sec = gd->gd_time_seconds;
1358 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1359 } while (tvp->tv_sec != gd->gd_time_seconds);
1361 if (delta >= sys_cputimer->freq) {
1362 tvp->tv_sec += delta / sys_cputimer->freq;
1363 delta %= sys_cputimer->freq;
1365 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1367 bt = &basetime[basetime_index];
1369 tvp->tv_sec += bt->tv_sec;
1370 tvp->tv_usec += bt->tv_nsec / 1000;
1371 while (tvp->tv_usec >= 1000000) {
1372 tvp->tv_usec -= 1000000;
1378 nanotime(struct timespec *tsp)
1380 struct globaldata *gd = mycpu;
1381 struct timespec *bt;
1385 tsp->tv_sec = gd->gd_time_seconds;
1386 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1387 } while (tsp->tv_sec != gd->gd_time_seconds);
1389 if (delta >= sys_cputimer->freq) {
1390 tsp->tv_sec += delta / sys_cputimer->freq;
1391 delta %= sys_cputimer->freq;
1393 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1395 bt = &basetime[basetime_index];
1397 tsp->tv_sec += bt->tv_sec;
1398 tsp->tv_nsec += bt->tv_nsec;
1399 while (tsp->tv_nsec >= 1000000000) {
1400 tsp->tv_nsec -= 1000000000;
1406 * Get an approximate time_t. It does not have to be accurate. This
1407 * function is called only from KTR and can be called with the system in
1408 * any state so do not use a critical section or other complex operation
1411 * NOTE: This is not exactly synchronized with real time. To do that we
1412 * would have to do what microtime does and check for a nanoseconds
1416 get_approximate_time_t(void)
1418 struct globaldata *gd = mycpu;
1419 struct timespec *bt;
1421 bt = &basetime[basetime_index];
1422 return(gd->gd_time_seconds + bt->tv_sec);
1426 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1429 struct pps_fetch_args *fapi;
1431 struct pps_kcbind_args *kapi;
1435 case PPS_IOC_CREATE:
1437 case PPS_IOC_DESTROY:
1439 case PPS_IOC_SETPARAMS:
1440 app = (pps_params_t *)data;
1441 if (app->mode & ~pps->ppscap)
1443 pps->ppsparam = *app;
1445 case PPS_IOC_GETPARAMS:
1446 app = (pps_params_t *)data;
1447 *app = pps->ppsparam;
1448 app->api_version = PPS_API_VERS_1;
1450 case PPS_IOC_GETCAP:
1451 *(int*)data = pps->ppscap;
1454 fapi = (struct pps_fetch_args *)data;
1455 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1457 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1458 return (EOPNOTSUPP);
1459 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1460 fapi->pps_info_buf = pps->ppsinfo;
1462 case PPS_IOC_KCBIND:
1464 kapi = (struct pps_kcbind_args *)data;
1465 /* XXX Only root should be able to do this */
1466 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1468 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1470 if (kapi->edge & ~pps->ppscap)
1472 pps->kcmode = kapi->edge;
1475 return (EOPNOTSUPP);
1483 pps_init(struct pps_state *pps)
1485 pps->ppscap |= PPS_TSFMT_TSPEC;
1486 if (pps->ppscap & PPS_CAPTUREASSERT)
1487 pps->ppscap |= PPS_OFFSETASSERT;
1488 if (pps->ppscap & PPS_CAPTURECLEAR)
1489 pps->ppscap |= PPS_OFFSETCLEAR;
1493 pps_event(struct pps_state *pps, sysclock_t count, int event)
1495 struct globaldata *gd;
1496 struct timespec *tsp;
1497 struct timespec *osp;
1498 struct timespec *bt;
1514 /* Things would be easier with arrays... */
1515 if (event == PPS_CAPTUREASSERT) {
1516 tsp = &pps->ppsinfo.assert_timestamp;
1517 osp = &pps->ppsparam.assert_offset;
1518 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1520 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1522 pcount = &pps->ppscount[0];
1523 pseq = &pps->ppsinfo.assert_sequence;
1525 tsp = &pps->ppsinfo.clear_timestamp;
1526 osp = &pps->ppsparam.clear_offset;
1527 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1529 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1531 pcount = &pps->ppscount[1];
1532 pseq = &pps->ppsinfo.clear_sequence;
1535 /* Nothing really happened */
1536 if (*pcount == count)
1542 ts.tv_sec = gd->gd_time_seconds;
1543 delta = count - gd->gd_cpuclock_base;
1544 } while (ts.tv_sec != gd->gd_time_seconds);
1546 if (delta >= sys_cputimer->freq) {
1547 ts.tv_sec += delta / sys_cputimer->freq;
1548 delta %= sys_cputimer->freq;
1550 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1551 ni = basetime_index;
1554 ts.tv_sec += bt->tv_sec;
1555 ts.tv_nsec += bt->tv_nsec;
1556 while (ts.tv_nsec >= 1000000000) {
1557 ts.tv_nsec -= 1000000000;
1565 timespecadd(tsp, osp);
1566 if (tsp->tv_nsec < 0) {
1567 tsp->tv_nsec += 1000000000;
1573 /* magic, at its best... */
1574 tcount = count - pps->ppscount[2];
1575 pps->ppscount[2] = count;
1576 if (tcount >= sys_cputimer->freq) {
1577 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1578 sys_cputimer->freq64_nsec *
1579 (tcount % sys_cputimer->freq)) >> 32;
1581 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1583 hardpps(tsp, delta);
1589 * Return the tsc target value for a delay of (ns).
1591 * Returns -1 if the TSC is not supported.
1594 tsc_get_target(int ns)
1596 #if defined(_RDTSC_SUPPORTED_)
1597 if (cpu_feature & CPUID_TSC) {
1598 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1605 * Compare the tsc against the passed target
1607 * Returns +1 if the target has been reached
1608 * Returns 0 if the target has not yet been reached
1609 * Returns -1 if the TSC is not supported.
1611 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1614 tsc_test_target(int64_t target)
1616 #if defined(_RDTSC_SUPPORTED_)
1617 if (cpu_feature & CPUID_TSC) {
1618 if ((int64_t)(target - rdtsc()) <= 0)
1627 * Delay the specified number of nanoseconds using the tsc. This function
1628 * returns immediately if the TSC is not supported. At least one cpu_pause()
1636 clk = tsc_get_target(ns);
1638 while (tsc_test_target(clk) == 0)