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];
132 static int sniff_enable = 1;
133 static int sniff_target = -1;
134 SYSCTL_INT(_kern, OID_AUTO, sniff_enable, CTLFLAG_RW, &sniff_enable, 0 , "");
135 SYSCTL_INT(_kern, OID_AUTO, sniff_target, CTLFLAG_RW, &sniff_target, 0 , "");
138 sysctl_cputime(SYSCTL_HANDLER_ARGS)
142 size_t size = sizeof(struct kinfo_cputime);
143 struct kinfo_cputime tmp;
146 * NOTE: For security reasons, only root can sniff %rip
148 root_error = priv_check_cred(curthread->td_ucred, PRIV_ROOT, 0);
150 for (cpu = 0; cpu < ncpus; ++cpu) {
151 tmp = cputime_percpu[cpu];
152 if (root_error == 0) {
154 (int64_t)globaldata_find(cpu)->gd_sample_pc;
156 (int64_t)globaldata_find(cpu)->gd_sample_sp;
158 if ((error = SYSCTL_OUT(req, &tmp, size)) != 0)
162 if (root_error == 0) {
164 int n = sniff_target;
174 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
175 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
178 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
180 long cpu_states[CPUSTATES] = {0};
182 size_t size = sizeof(cpu_states);
184 for (cpu = 0; cpu < ncpus; ++cpu) {
185 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
186 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
187 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
188 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
189 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
192 error = SYSCTL_OUT(req, cpu_states, size);
197 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
198 sysctl_cp_time, "LU", "CPU time statistics");
201 sysctl_cp_times(SYSCTL_HANDLER_ARGS)
203 long cpu_states[CPUSTATES] = {0};
205 size_t size = sizeof(cpu_states);
207 for (error = 0, cpu = 0; error == 0 && cpu < ncpus; ++cpu) {
208 cpu_states[CP_USER] = cputime_percpu[cpu].cp_user;
209 cpu_states[CP_NICE] = cputime_percpu[cpu].cp_nice;
210 cpu_states[CP_SYS] = cputime_percpu[cpu].cp_sys;
211 cpu_states[CP_INTR] = cputime_percpu[cpu].cp_intr;
212 cpu_states[CP_IDLE] = cputime_percpu[cpu].cp_idle;
213 error = SYSCTL_OUT(req, cpu_states, size);
219 SYSCTL_PROC(_kern, OID_AUTO, cp_times, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
220 sysctl_cp_times, "LU", "per-CPU time statistics");
223 * boottime is used to calculate the 'real' uptime. Do not confuse this with
224 * microuptime(). microtime() is not drift compensated. The real uptime
225 * with compensation is nanotime() - bootime. boottime is recalculated
226 * whenever the real time is set based on the compensated elapsed time
227 * in seconds (gd->gd_time_seconds).
229 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
230 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
233 * WARNING! time_second can backstep on time corrections. Also, unlike
234 * time_second, time_uptime is not a "real" time_t (seconds
235 * since the Epoch) but seconds since booting.
237 struct timespec boottime; /* boot time (realtime) for reference only */
238 time_t time_second; /* read-only 'passive' realtime in seconds */
239 time_t time_uptime; /* read-only 'passive' uptime in seconds */
242 * basetime is used to calculate the compensated real time of day. The
243 * basetime can be modified on a per-tick basis by the adjtime(),
244 * ntp_adjtime(), and sysctl-based time correction APIs.
246 * Note that frequency corrections can also be made by adjusting
249 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
250 * used on both SMP and UP systems to avoid MP races between cpu's and
251 * interrupt races on UP systems.
254 __uint32_t time_second;
255 sysclock_t cpuclock_base;
258 #define BASETIME_ARYSIZE 16
259 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
260 static struct timespec basetime[BASETIME_ARYSIZE];
261 static struct hardtime hardtime[BASETIME_ARYSIZE];
262 static volatile int basetime_index;
265 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
272 * Because basetime data and index may be updated by another cpu,
273 * a load fence is required to ensure that the data we read has
274 * not been speculatively read relative to a possibly updated index.
276 index = basetime_index;
278 bt = &basetime[index];
279 error = SYSCTL_OUT(req, bt, sizeof(*bt));
283 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
284 &boottime, timespec, "System boottime");
285 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
286 sysctl_get_basetime, "S,timespec", "System basetime");
288 static void hardclock(systimer_t info, int, struct intrframe *frame);
289 static void statclock(systimer_t info, int, struct intrframe *frame);
290 static void schedclock(systimer_t info, int, struct intrframe *frame);
291 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
293 int ticks; /* system master ticks at hz */
294 int clocks_running; /* tsleep/timeout clocks operational */
295 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
296 int64_t nsec_acc; /* accumulator */
297 int sched_ticks; /* global schedule clock ticks */
299 /* NTPD time correction fields */
300 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
301 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
302 int64_t ntp_delta; /* one-time correction in nsec */
303 int64_t ntp_big_delta = 1000000000;
304 int32_t ntp_tick_delta; /* current adjustment rate */
305 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
306 time_t ntp_leap_second; /* time of next leap second */
307 int ntp_leap_insert; /* whether to insert or remove a second */
308 struct spinlock ntp_spin;
311 * Finish initializing clock frequencies and start all clocks running.
315 initclocks(void *dummy)
317 /*psratio = profhz / stathz;*/
318 spin_init(&ntp_spin, "ntp");
322 kpmap->tsc_freq = (uint64_t)tsc_frequency;
323 kpmap->tick_freq = hz;
328 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
329 * during SMP initialization.
331 * This routine is called concurrently during low-level SMP initialization
332 * and may not block in any way. Meaning, among other things, we can't
333 * acquire any tokens.
336 initclocks_pcpu(void)
338 struct globaldata *gd = mycpu;
341 if (gd->gd_cpuid == 0) {
342 gd->gd_time_seconds = 1;
343 gd->gd_cpuclock_base = sys_cputimer->count();
344 hardtime[0].time_second = gd->gd_time_seconds;
345 hardtime[0].cpuclock_base = gd->gd_cpuclock_base;
347 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
348 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
351 systimer_intr_enable();
357 * This routine is called on just the BSP, just after SMP initialization
358 * completes to * finish initializing any clocks that might contend/block
359 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
360 * that function is called from the idle thread bootstrap for each cpu and
361 * not allowed to block at all.
365 initclocks_other(void *dummy)
367 struct globaldata *ogd = mycpu;
368 struct globaldata *gd;
371 for (n = 0; n < ncpus; ++n) {
372 lwkt_setcpu_self(globaldata_find(n));
376 * Use a non-queued periodic systimer to prevent multiple
377 * ticks from building up if the sysclock jumps forward
378 * (8254 gets reset). The sysclock will never jump backwards.
379 * Our time sync is based on the actual sysclock, not the
382 * Install statclock before hardclock to prevent statclock
383 * from misinterpreting gd_flags for tick assignment when
386 systimer_init_periodic_nq(&gd->gd_statclock, statclock,
388 systimer_init_periodic_nq(&gd->gd_hardclock, hardclock,
390 /* XXX correct the frequency for scheduler / estcpu tests */
391 systimer_init_periodic_nq(&gd->gd_schedclock, schedclock,
394 ifpoll_init_pcpu(gd->gd_cpuid);
397 lwkt_setcpu_self(ogd);
399 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL);
402 * This sets the current real time of day. Timespecs are in seconds and
403 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
404 * instead we adjust basetime so basetime + gd_* results in the current
405 * time of day. This way the gd_* fields are guaranteed to represent
406 * a monotonically increasing 'uptime' value.
408 * When set_timeofday() is called from userland, the system call forces it
409 * onto cpu #0 since only cpu #0 can update basetime_index.
412 set_timeofday(struct timespec *ts)
414 struct timespec *nbt;
418 * XXX SMP / non-atomic basetime updates
421 ni = (basetime_index + 1) & BASETIME_ARYMASK;
425 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
426 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
427 if (nbt->tv_nsec < 0) {
428 nbt->tv_nsec += 1000000000;
433 * Note that basetime diverges from boottime as the clock drift is
434 * compensated for, so we cannot do away with boottime. When setting
435 * the absolute time of day the drift is 0 (for an instant) and we
436 * can simply assign boottime to basetime.
438 * Note that nanouptime() is based on gd_time_seconds which is drift
439 * compensated up to a point (it is guaranteed to remain monotonically
440 * increasing). gd_time_seconds is thus our best uptime guess and
441 * suitable for use in the boottime calculation. It is already taken
442 * into account in the basetime calculation above.
444 spin_lock(&ntp_spin);
445 boottime.tv_sec = nbt->tv_sec;
449 * We now have a new basetime, make sure all other cpus have it,
450 * then update the index.
454 spin_unlock(&ntp_spin);
460 * Each cpu has its own hardclock, but we only increments ticks and softticks
463 * NOTE! systimer! the MP lock might not be held here. We can only safely
464 * manipulate objects owned by the current cpu.
467 hardclock(systimer_t info, int in_ipi, struct intrframe *frame)
471 struct globaldata *gd = mycpu;
473 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) {
474 /* Defer to doreti on passive IPIQ processing */
479 * We update the compensation base to calculate fine-grained time
480 * from the sys_cputimer on a per-cpu basis in order to avoid
481 * having to mess around with locks. sys_cputimer is assumed to
482 * be consistent across all cpus. CPU N copies the base state from
483 * CPU 0 using the same FIFO trick that we use for basetime (so we
484 * don't catch a CPU 0 update in the middle).
486 * Note that we never allow info->time (aka gd->gd_hardclock.time)
487 * to reverse index gd_cpuclock_base, but that it is possible for
488 * it to temporarily get behind in the seconds if something in the
489 * system locks interrupts for a long period of time. Since periodic
490 * timers count events, though everything should resynch again
493 if (gd->gd_cpuid == 0) {
496 cputicks = info->time - gd->gd_cpuclock_base;
497 if (cputicks >= sys_cputimer->freq) {
498 cputicks /= sys_cputimer->freq;
499 if (cputicks != 0 && cputicks != 1)
500 kprintf("Warning: hardclock missed > 1 sec\n");
501 gd->gd_time_seconds += cputicks;
502 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks;
503 /* uncorrected monotonic 1-sec gran */
504 time_uptime += cputicks;
506 ni = (basetime_index + 1) & BASETIME_ARYMASK;
507 hardtime[ni].time_second = gd->gd_time_seconds;
508 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base;
514 gd->gd_time_seconds = hardtime[ni].time_second;
515 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base;
519 * The system-wide ticks counter and NTP related timedelta/tickdelta
520 * adjustments only occur on cpu #0. NTP adjustments are accomplished
521 * by updating basetime.
523 if (gd->gd_cpuid == 0) {
524 struct timespec *nbt;
532 if (tco->tc_poll_pps)
533 tco->tc_poll_pps(tco);
537 * Calculate the new basetime index. We are in a critical section
538 * on cpu #0 and can safely play with basetime_index. Start
539 * with the current basetime and then make adjustments.
541 ni = (basetime_index + 1) & BASETIME_ARYMASK;
543 *nbt = basetime[basetime_index];
546 * ntp adjustments only occur on cpu 0 and are protected by
547 * ntp_spin. This spinlock virtually never conflicts.
549 spin_lock(&ntp_spin);
552 * Apply adjtime corrections. (adjtime() API)
554 * adjtime() only runs on cpu #0 so our critical section is
555 * sufficient to access these variables.
557 if (ntp_delta != 0) {
558 nbt->tv_nsec += ntp_tick_delta;
559 ntp_delta -= ntp_tick_delta;
560 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
561 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
562 ntp_tick_delta = ntp_delta;
567 * Apply permanent frequency corrections. (sysctl API)
569 if (ntp_tick_permanent != 0) {
570 ntp_tick_acc += ntp_tick_permanent;
571 if (ntp_tick_acc >= (1LL << 32)) {
572 nbt->tv_nsec += ntp_tick_acc >> 32;
573 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
574 } else if (ntp_tick_acc <= -(1LL << 32)) {
575 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
576 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
577 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
581 if (nbt->tv_nsec >= 1000000000) {
583 nbt->tv_nsec -= 1000000000;
584 } else if (nbt->tv_nsec < 0) {
586 nbt->tv_nsec += 1000000000;
590 * Another per-tick compensation. (for ntp_adjtime() API)
593 nsec_acc += nsec_adj;
594 if (nsec_acc >= 0x100000000LL) {
595 nbt->tv_nsec += nsec_acc >> 32;
596 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
597 } else if (nsec_acc <= -0x100000000LL) {
598 nbt->tv_nsec -= -nsec_acc >> 32;
599 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
601 if (nbt->tv_nsec >= 1000000000) {
602 nbt->tv_nsec -= 1000000000;
604 } else if (nbt->tv_nsec < 0) {
605 nbt->tv_nsec += 1000000000;
609 spin_unlock(&ntp_spin);
611 /************************************************************
612 * LEAP SECOND CORRECTION *
613 ************************************************************
615 * Taking into account all the corrections made above, figure
616 * out the new real time. If the seconds field has changed
617 * then apply any pending leap-second corrections.
619 getnanotime_nbt(nbt, &nts);
621 if (time_second != nts.tv_sec) {
623 * Apply leap second (sysctl API). Adjust nts for changes
624 * so we do not have to call getnanotime_nbt again.
626 if (ntp_leap_second) {
627 if (ntp_leap_second == nts.tv_sec) {
628 if (ntp_leap_insert) {
640 * Apply leap second (ntp_adjtime() API), calculate a new
641 * nsec_adj field. ntp_update_second() returns nsec_adj
642 * as a per-second value but we need it as a per-tick value.
644 leap = ntp_update_second(time_second, &nsec_adj);
650 * Update the time_second 'approximate time' global.
652 time_second = nts.tv_sec;
656 * Finally, our new basetime is ready to go live!
662 * Update kpmap on each tick. TS updates are integrated with
663 * fences and upticks allowing userland to read the data
669 w = (kpmap->upticks + 1) & 1;
670 getnanouptime(&kpmap->ts_uptime[w]);
671 getnanotime(&kpmap->ts_realtime[w]);
679 * lwkt thread scheduler fair queueing
681 lwkt_schedulerclock(curthread);
684 * softticks are handled for all cpus
686 hardclock_softtick(gd);
689 * Rollup accumulated vmstats, copy-back for critical path checks.
691 vmstats_rollup_cpu(gd);
692 mycpu->gd_vmstats = vmstats;
695 * ITimer handling is per-tick, per-cpu.
697 * We must acquire the per-process token in order for ksignal()
698 * to be non-blocking. For the moment this requires an AST fault,
699 * the ksignal() cannot be safely issued from this hard interrupt.
701 * XXX Even the trytoken here isn't right, and itimer operation in
702 * a multi threaded environment is going to be weird at the
705 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
708 ++p->p_upmap->runticks;
710 if (frame && CLKF_USERMODE(frame) &&
711 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
712 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
713 p->p_flags |= P_SIGVTALRM;
716 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
717 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
718 p->p_flags |= P_SIGPROF;
722 lwkt_reltoken(&p->p_token);
728 * The statistics clock typically runs at a 125Hz rate, and is intended
729 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
731 * NOTE! systimer! the MP lock might not be held here. We can only safely
732 * manipulate objects owned by the current cpu.
734 * The stats clock is responsible for grabbing a profiling sample.
735 * Most of the statistics are only used by user-level statistics programs.
736 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
739 * Like the other clocks, the stat clock is called from what is effectively
740 * a fast interrupt, so the context should be the thread/process that got
744 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
750 globaldata_t gd = mycpu;
758 * How big was our timeslice relative to the last time? Calculate
761 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
762 * during early boot. Just use the systimer count to be nice
763 * to e.g. qemu. The systimer has a better chance of being
764 * MPSAFE at early boot.
766 cv = sys_cputimer->count();
767 scv = gd->statint.gd_statcv;
771 bump = (sys_cputimer->freq64_usec * (cv - scv)) >> 32;
777 gd->statint.gd_statcv = cv;
780 stv = &gd->gd_stattv;
781 if (stv->tv_sec == 0) {
784 bump = tv.tv_usec - stv->tv_usec +
785 (tv.tv_sec - stv->tv_sec) * 1000000;
797 if (frame && CLKF_USERMODE(frame)) {
799 * Came from userland, handle user time and deal with
802 if (p && (p->p_flags & P_PROFIL))
803 addupc_intr(p, CLKF_PC(frame), 1);
804 td->td_uticks += bump;
807 * Charge the time as appropriate
809 if (p && p->p_nice > NZERO)
810 cpu_time.cp_nice += bump;
812 cpu_time.cp_user += bump;
814 int intr_nest = gd->gd_intr_nesting_level;
818 * IPI processing code will bump gd_intr_nesting_level
819 * up by one, which breaks following CLKF_INTR testing,
820 * so we subtract it by one here.
826 * Kernel statistics are just like addupc_intr, only easier.
829 if (g->state == GMON_PROF_ON && frame) {
830 i = CLKF_PC(frame) - g->lowpc;
831 if (i < g->textsize) {
832 i /= HISTFRACTION * sizeof(*g->kcount);
838 #define IS_INTR_RUNNING ((frame && CLKF_INTR(intr_nest)) || CLKF_INTR_TD(td))
841 * Came from kernel mode, so we were:
842 * - handling an interrupt,
843 * - doing syscall or trap work on behalf of the current
845 * - spinning in the idle loop.
846 * Whichever it is, charge the time as appropriate.
847 * Note that we charge interrupts to the current process,
848 * regardless of whether they are ``for'' that process,
849 * so that we know how much of its real time was spent
850 * in ``non-process'' (i.e., interrupt) work.
852 * XXX assume system if frame is NULL. A NULL frame
853 * can occur if ipi processing is done from a crit_exit().
855 if (IS_INTR_RUNNING) {
857 * If we interrupted an interrupt thread, well,
858 * count it as interrupt time.
860 td->td_iticks += bump;
863 do_pctrack(frame, PCTRACK_INT);
865 cpu_time.cp_intr += bump;
866 } else if (gd->gd_flags & GDF_VIRTUSER) {
868 * The vkernel doesn't do a good job providing trap
869 * frames that we can test. If the GDF_VIRTUSER
870 * flag is set we probably interrupted user mode.
872 * We also use this flag on the host when entering
875 td->td_uticks += bump;
878 * Charge the time as appropriate
880 if (p && p->p_nice > NZERO)
881 cpu_time.cp_nice += bump;
883 cpu_time.cp_user += bump;
885 td->td_sticks += bump;
886 if (td == &gd->gd_idlethread) {
888 * Token contention can cause us to mis-count
889 * a contended as idle, but it doesn't work
890 * properly for VKERNELs so just test on a
893 #ifdef _KERNEL_VIRTUAL
894 cpu_time.cp_idle += bump;
896 if (mycpu->gd_reqflags & RQF_IDLECHECK_WK_MASK)
897 cpu_time.cp_sys += bump;
899 cpu_time.cp_idle += bump;
903 * System thread was running.
907 do_pctrack(frame, PCTRACK_SYS);
909 cpu_time.cp_sys += bump;
913 #undef IS_INTR_RUNNING
919 * Sample the PC when in the kernel or in an interrupt. User code can
920 * retrieve the information and generate a histogram or other output.
924 do_pctrack(struct intrframe *frame, int which)
926 struct kinfo_pctrack *pctrack;
928 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
929 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
930 (void *)CLKF_PC(frame);
935 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
937 struct kinfo_pcheader head;
942 head.pc_ntrack = PCTRACK_SIZE;
943 head.pc_arysize = PCTRACK_ARYSIZE;
945 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
948 for (cpu = 0; cpu < ncpus; ++cpu) {
949 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
950 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
951 sizeof(struct kinfo_pctrack));
960 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
961 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
966 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
967 * the MP lock might not be held. We can safely manipulate parts of curproc
968 * but that's about it.
970 * Each cpu has its own scheduler clock.
973 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
980 if ((lp = lwkt_preempted_proc()) != NULL) {
982 * Account for cpu time used and hit the scheduler. Note
983 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
987 usched_schedulerclock(lp, info->periodic, info->time);
989 usched_schedulerclock(NULL, info->periodic, info->time);
991 if ((lp = curthread->td_lwp) != NULL) {
993 * Update resource usage integrals and maximums.
995 if ((ru = &lp->lwp_proc->p_ru) &&
996 (vm = lp->lwp_proc->p_vmspace) != NULL) {
997 ru->ru_ixrss += pgtok(vm->vm_tsize);
998 ru->ru_idrss += pgtok(vm->vm_dsize);
999 ru->ru_isrss += pgtok(vm->vm_ssize);
1000 if (lwkt_trytoken(&vm->vm_map.token)) {
1001 rss = pgtok(vmspace_resident_count(vm));
1002 if (ru->ru_maxrss < rss)
1003 ru->ru_maxrss = rss;
1004 lwkt_reltoken(&vm->vm_map.token);
1008 /* Increment the global sched_ticks */
1009 if (mycpu->gd_cpuid == 0)
1014 * Compute number of ticks for the specified amount of time. The
1015 * return value is intended to be used in a clock interrupt timed
1016 * operation and guaranteed to meet or exceed the requested time.
1017 * If the representation overflows, return INT_MAX. The minimum return
1018 * value is 1 ticks and the function will average the calculation up.
1019 * If any value greater then 0 microseconds is supplied, a value
1020 * of at least 2 will be returned to ensure that a near-term clock
1021 * interrupt does not cause the timeout to occur (degenerately) early.
1023 * Note that limit checks must take into account microseconds, which is
1024 * done simply by using the smaller signed long maximum instead of
1025 * the unsigned long maximum.
1027 * If ints have 32 bits, then the maximum value for any timeout in
1028 * 10ms ticks is 248 days.
1031 tvtohz_high(struct timeval *tv)
1048 kprintf("tvtohz_high: negative time difference "
1049 "%ld sec %ld usec\n",
1053 } else if (sec <= INT_MAX / hz) {
1054 ticks = (int)(sec * hz +
1055 ((u_long)usec + (ustick - 1)) / ustick) + 1;
1063 tstohz_high(struct timespec *ts)
1080 kprintf("tstohz_high: negative time difference "
1081 "%ld sec %ld nsec\n",
1085 } else if (sec <= INT_MAX / hz) {
1086 ticks = (int)(sec * hz +
1087 ((u_long)nsec + (nstick - 1)) / nstick) + 1;
1096 * Compute number of ticks for the specified amount of time, erroring on
1097 * the side of it being too low to ensure that sleeping the returned number
1098 * of ticks will not result in a late return.
1100 * The supplied timeval may not be negative and should be normalized. A
1101 * return value of 0 is possible if the timeval converts to less then
1104 * If ints have 32 bits, then the maximum value for any timeout in
1105 * 10ms ticks is 248 days.
1108 tvtohz_low(struct timeval *tv)
1114 if (sec <= INT_MAX / hz)
1115 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
1122 tstohz_low(struct timespec *ts)
1128 if (sec <= INT_MAX / hz)
1129 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1136 * Start profiling on a process.
1138 * Caller must hold p->p_token();
1140 * Kernel profiling passes proc0 which never exits and hence
1141 * keeps the profile clock running constantly.
1144 startprofclock(struct proc *p)
1146 if ((p->p_flags & P_PROFIL) == 0) {
1147 p->p_flags |= P_PROFIL;
1149 if (++profprocs == 1 && stathz != 0) {
1152 setstatclockrate(profhz);
1160 * Stop profiling on a process.
1162 * caller must hold p->p_token
1165 stopprofclock(struct proc *p)
1167 if (p->p_flags & P_PROFIL) {
1168 p->p_flags &= ~P_PROFIL;
1170 if (--profprocs == 0 && stathz != 0) {
1173 setstatclockrate(stathz);
1181 * Return information about system clocks.
1184 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1186 struct kinfo_clockinfo clkinfo;
1188 * Construct clockinfo structure.
1191 clkinfo.ci_tick = ustick;
1192 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1193 clkinfo.ci_profhz = profhz;
1194 clkinfo.ci_stathz = stathz ? stathz : hz;
1195 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1198 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1199 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1202 * We have eight functions for looking at the clock, four for
1203 * microseconds and four for nanoseconds. For each there is fast
1204 * but less precise version "get{nano|micro}[up]time" which will
1205 * return a time which is up to 1/HZ previous to the call, whereas
1206 * the raw version "{nano|micro}[up]time" will return a timestamp
1207 * which is as precise as possible. The "up" variants return the
1208 * time relative to system boot, these are well suited for time
1209 * interval measurements.
1211 * Each cpu independently maintains the current time of day, so all
1212 * we need to do to protect ourselves from changes is to do a loop
1213 * check on the seconds field changing out from under us.
1215 * The system timer maintains a 32 bit count and due to various issues
1216 * it is possible for the calculated delta to occasionally exceed
1217 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1218 * multiplication can easily overflow, so we deal with the case. For
1219 * uniformity we deal with the case in the usec case too.
1221 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1224 getmicrouptime(struct timeval *tvp)
1226 struct globaldata *gd = mycpu;
1230 tvp->tv_sec = gd->gd_time_seconds;
1231 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1232 } while (tvp->tv_sec != gd->gd_time_seconds);
1234 if (delta >= sys_cputimer->freq) {
1235 tvp->tv_sec += delta / sys_cputimer->freq;
1236 delta %= sys_cputimer->freq;
1238 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1239 if (tvp->tv_usec >= 1000000) {
1240 tvp->tv_usec -= 1000000;
1246 getnanouptime(struct timespec *tsp)
1248 struct globaldata *gd = mycpu;
1252 tsp->tv_sec = gd->gd_time_seconds;
1253 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1254 } while (tsp->tv_sec != gd->gd_time_seconds);
1256 if (delta >= sys_cputimer->freq) {
1257 tsp->tv_sec += delta / sys_cputimer->freq;
1258 delta %= sys_cputimer->freq;
1260 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1264 microuptime(struct timeval *tvp)
1266 struct globaldata *gd = mycpu;
1270 tvp->tv_sec = gd->gd_time_seconds;
1271 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1272 } while (tvp->tv_sec != gd->gd_time_seconds);
1274 if (delta >= sys_cputimer->freq) {
1275 tvp->tv_sec += delta / sys_cputimer->freq;
1276 delta %= sys_cputimer->freq;
1278 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1282 nanouptime(struct timespec *tsp)
1284 struct globaldata *gd = mycpu;
1288 tsp->tv_sec = gd->gd_time_seconds;
1289 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1290 } while (tsp->tv_sec != gd->gd_time_seconds);
1292 if (delta >= sys_cputimer->freq) {
1293 tsp->tv_sec += delta / sys_cputimer->freq;
1294 delta %= sys_cputimer->freq;
1296 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1303 getmicrotime(struct timeval *tvp)
1305 struct globaldata *gd = mycpu;
1306 struct timespec *bt;
1310 tvp->tv_sec = gd->gd_time_seconds;
1311 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1312 } while (tvp->tv_sec != gd->gd_time_seconds);
1314 if (delta >= sys_cputimer->freq) {
1315 tvp->tv_sec += delta / sys_cputimer->freq;
1316 delta %= sys_cputimer->freq;
1318 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1320 bt = &basetime[basetime_index];
1322 tvp->tv_sec += bt->tv_sec;
1323 tvp->tv_usec += bt->tv_nsec / 1000;
1324 while (tvp->tv_usec >= 1000000) {
1325 tvp->tv_usec -= 1000000;
1331 getnanotime(struct timespec *tsp)
1333 struct globaldata *gd = mycpu;
1334 struct timespec *bt;
1338 tsp->tv_sec = gd->gd_time_seconds;
1339 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1340 } while (tsp->tv_sec != gd->gd_time_seconds);
1342 if (delta >= sys_cputimer->freq) {
1343 tsp->tv_sec += delta / sys_cputimer->freq;
1344 delta %= sys_cputimer->freq;
1346 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1348 bt = &basetime[basetime_index];
1350 tsp->tv_sec += bt->tv_sec;
1351 tsp->tv_nsec += bt->tv_nsec;
1352 while (tsp->tv_nsec >= 1000000000) {
1353 tsp->tv_nsec -= 1000000000;
1359 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1361 struct globaldata *gd = mycpu;
1365 tsp->tv_sec = gd->gd_time_seconds;
1366 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1367 } while (tsp->tv_sec != gd->gd_time_seconds);
1369 if (delta >= sys_cputimer->freq) {
1370 tsp->tv_sec += delta / sys_cputimer->freq;
1371 delta %= sys_cputimer->freq;
1373 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1375 tsp->tv_sec += nbt->tv_sec;
1376 tsp->tv_nsec += nbt->tv_nsec;
1377 while (tsp->tv_nsec >= 1000000000) {
1378 tsp->tv_nsec -= 1000000000;
1385 microtime(struct timeval *tvp)
1387 struct globaldata *gd = mycpu;
1388 struct timespec *bt;
1392 tvp->tv_sec = gd->gd_time_seconds;
1393 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1394 } while (tvp->tv_sec != gd->gd_time_seconds);
1396 if (delta >= sys_cputimer->freq) {
1397 tvp->tv_sec += delta / sys_cputimer->freq;
1398 delta %= sys_cputimer->freq;
1400 tvp->tv_usec = (sys_cputimer->freq64_usec * delta) >> 32;
1402 bt = &basetime[basetime_index];
1404 tvp->tv_sec += bt->tv_sec;
1405 tvp->tv_usec += bt->tv_nsec / 1000;
1406 while (tvp->tv_usec >= 1000000) {
1407 tvp->tv_usec -= 1000000;
1413 nanotime(struct timespec *tsp)
1415 struct globaldata *gd = mycpu;
1416 struct timespec *bt;
1420 tsp->tv_sec = gd->gd_time_seconds;
1421 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1422 } while (tsp->tv_sec != gd->gd_time_seconds);
1424 if (delta >= sys_cputimer->freq) {
1425 tsp->tv_sec += delta / sys_cputimer->freq;
1426 delta %= sys_cputimer->freq;
1428 tsp->tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1430 bt = &basetime[basetime_index];
1432 tsp->tv_sec += bt->tv_sec;
1433 tsp->tv_nsec += bt->tv_nsec;
1434 while (tsp->tv_nsec >= 1000000000) {
1435 tsp->tv_nsec -= 1000000000;
1441 * Get an approximate time_t. It does not have to be accurate. This
1442 * function is called only from KTR and can be called with the system in
1443 * any state so do not use a critical section or other complex operation
1446 * NOTE: This is not exactly synchronized with real time. To do that we
1447 * would have to do what microtime does and check for a nanoseconds
1451 get_approximate_time_t(void)
1453 struct globaldata *gd = mycpu;
1454 struct timespec *bt;
1456 bt = &basetime[basetime_index];
1457 return(gd->gd_time_seconds + bt->tv_sec);
1461 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1464 struct pps_fetch_args *fapi;
1466 struct pps_kcbind_args *kapi;
1470 case PPS_IOC_CREATE:
1472 case PPS_IOC_DESTROY:
1474 case PPS_IOC_SETPARAMS:
1475 app = (pps_params_t *)data;
1476 if (app->mode & ~pps->ppscap)
1478 pps->ppsparam = *app;
1480 case PPS_IOC_GETPARAMS:
1481 app = (pps_params_t *)data;
1482 *app = pps->ppsparam;
1483 app->api_version = PPS_API_VERS_1;
1485 case PPS_IOC_GETCAP:
1486 *(int*)data = pps->ppscap;
1489 fapi = (struct pps_fetch_args *)data;
1490 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1492 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
1493 return (EOPNOTSUPP);
1494 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1495 fapi->pps_info_buf = pps->ppsinfo;
1497 case PPS_IOC_KCBIND:
1499 kapi = (struct pps_kcbind_args *)data;
1500 /* XXX Only root should be able to do this */
1501 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1503 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1505 if (kapi->edge & ~pps->ppscap)
1507 pps->kcmode = kapi->edge;
1510 return (EOPNOTSUPP);
1518 pps_init(struct pps_state *pps)
1520 pps->ppscap |= PPS_TSFMT_TSPEC;
1521 if (pps->ppscap & PPS_CAPTUREASSERT)
1522 pps->ppscap |= PPS_OFFSETASSERT;
1523 if (pps->ppscap & PPS_CAPTURECLEAR)
1524 pps->ppscap |= PPS_OFFSETCLEAR;
1528 pps_event(struct pps_state *pps, sysclock_t count, int event)
1530 struct globaldata *gd;
1531 struct timespec *tsp;
1532 struct timespec *osp;
1533 struct timespec *bt;
1549 /* Things would be easier with arrays... */
1550 if (event == PPS_CAPTUREASSERT) {
1551 tsp = &pps->ppsinfo.assert_timestamp;
1552 osp = &pps->ppsparam.assert_offset;
1553 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1555 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1557 pcount = &pps->ppscount[0];
1558 pseq = &pps->ppsinfo.assert_sequence;
1560 tsp = &pps->ppsinfo.clear_timestamp;
1561 osp = &pps->ppsparam.clear_offset;
1562 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1564 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1566 pcount = &pps->ppscount[1];
1567 pseq = &pps->ppsinfo.clear_sequence;
1570 /* Nothing really happened */
1571 if (*pcount == count)
1577 ts.tv_sec = gd->gd_time_seconds;
1578 delta = count - gd->gd_cpuclock_base;
1579 } while (ts.tv_sec != gd->gd_time_seconds);
1581 if (delta >= sys_cputimer->freq) {
1582 ts.tv_sec += delta / sys_cputimer->freq;
1583 delta %= sys_cputimer->freq;
1585 ts.tv_nsec = (sys_cputimer->freq64_nsec * delta) >> 32;
1586 ni = basetime_index;
1589 ts.tv_sec += bt->tv_sec;
1590 ts.tv_nsec += bt->tv_nsec;
1591 while (ts.tv_nsec >= 1000000000) {
1592 ts.tv_nsec -= 1000000000;
1600 timespecadd(tsp, osp);
1601 if (tsp->tv_nsec < 0) {
1602 tsp->tv_nsec += 1000000000;
1608 /* magic, at its best... */
1609 tcount = count - pps->ppscount[2];
1610 pps->ppscount[2] = count;
1611 if (tcount >= sys_cputimer->freq) {
1612 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1613 sys_cputimer->freq64_nsec *
1614 (tcount % sys_cputimer->freq)) >> 32;
1616 delta = (sys_cputimer->freq64_nsec * tcount) >> 32;
1618 hardpps(tsp, delta);
1624 * Return the tsc target value for a delay of (ns).
1626 * Returns -1 if the TSC is not supported.
1629 tsc_get_target(int ns)
1631 #if defined(_RDTSC_SUPPORTED_)
1632 if (cpu_feature & CPUID_TSC) {
1633 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1640 * Compare the tsc against the passed target
1642 * Returns +1 if the target has been reached
1643 * Returns 0 if the target has not yet been reached
1644 * Returns -1 if the TSC is not supported.
1646 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1649 tsc_test_target(int64_t target)
1651 #if defined(_RDTSC_SUPPORTED_)
1652 if (cpu_feature & CPUID_TSC) {
1653 if ((int64_t)(target - rdtsc()) <= 0)
1662 * Delay the specified number of nanoseconds using the tsc. This function
1663 * returns immediately if the TSC is not supported. At least one cpu_pause()
1671 clk = tsc_get_target(ns);
1673 while (tsc_test_target(clk) == 0)