2 * Copyright (c) 1982, 1986, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
41 #include "opt_device_polling.h"
42 #include "opt_hwpmc_hooks.h"
44 #include "opt_watchdog.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/callout.h>
50 #include <sys/kernel.h>
51 #include <sys/kthread.h>
54 #include <sys/mutex.h>
56 #include <sys/resource.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sched.h>
60 #include <sys/signalvar.h>
61 #include <sys/sleepqueue.h>
65 #include <vm/vm_map.h>
66 #include <sys/sysctl.h>
68 #include <sys/interrupt.h>
69 #include <sys/limits.h>
70 #include <sys/timetc.h>
77 #include <sys/pmckern.h>
78 PMC_SOFT_DEFINE( , , clock, hard);
79 PMC_SOFT_DEFINE( , , clock, stat);
80 PMC_SOFT_DEFINE_EX( , , clock, prof, \
81 cpu_startprofclock, cpu_stopprofclock);
85 extern void hardclock_device_poll(void);
86 #endif /* DEVICE_POLLING */
88 static void initclocks(void *dummy);
89 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
91 /* Spin-lock protecting profiling statistics. */
92 static struct mtx time_lock;
94 SDT_PROVIDER_DECLARE(sched);
95 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
98 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
101 long cp_time[CPUSTATES];
104 unsigned int cp_time32[CPUSTATES];
107 read_cpu_time(cp_time);
109 if (req->flags & SCTL_MASK32) {
111 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
112 for (i = 0; i < CPUSTATES; i++)
113 cp_time32[i] = (unsigned int)cp_time[i];
114 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
119 return SYSCTL_OUT(req, 0, sizeof(cp_time));
120 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
125 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
126 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
128 static long empty[CPUSTATES];
131 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
138 unsigned int cp_time32[CPUSTATES];
144 if (req->flags & SCTL_MASK32)
145 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
148 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
150 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
151 if (!CPU_ABSENT(c)) {
153 cp_time = pcpu->pc_cp_time;
158 if (req->flags & SCTL_MASK32) {
159 for (i = 0; i < CPUSTATES; i++)
160 cp_time32[i] = (unsigned int)cp_time[i];
161 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
164 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
169 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
170 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
173 static const char *blessed[] = {
179 static int slptime_threshold = 1800;
180 static int blktime_threshold = 900;
181 static int sleepfreq = 3;
189 int blkticks, i, slpticks, slptype, tryl, tticks;
193 blkticks = blktime_threshold * hz;
194 slpticks = slptime_threshold * hz;
197 * Avoid to sleep on the sx_lock in order to avoid a possible
198 * priority inversion problem leading to starvation.
199 * If the lock can't be held after 100 tries, panic.
201 if (!sx_try_slock(&allproc_lock)) {
203 panic("%s: possible deadlock detected on allproc_lock\n",
206 pause("allproc", sleepfreq * hz);
210 FOREACH_PROC_IN_SYSTEM(p) {
212 if (p->p_state == PRS_NEW) {
216 FOREACH_THREAD_IN_PROC(p, td) {
219 if (TD_ON_LOCK(td)) {
222 * The thread should be blocked on a
223 * turnstile, simply check if the
224 * turnstile channel is in good state.
226 MPASS(td->td_blocked != NULL);
228 tticks = ticks - td->td_blktick;
230 if (tticks > blkticks) {
233 * Accordingly with provided
234 * thresholds, this thread is
235 * stuck for too long on a
239 sx_sunlock(&allproc_lock);
240 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
241 __func__, td, tticks);
243 } else if (TD_IS_SLEEPING(td) &&
247 * Check if the thread is sleeping on a
248 * lock, otherwise skip the check.
249 * Drop the thread lock in order to
250 * avoid a LOR with the sleepqueue
253 wchan = td->td_wchan;
254 tticks = ticks - td->td_slptick;
256 slptype = sleepq_type(wchan);
257 if ((slptype == SLEEPQ_SX ||
258 slptype == SLEEPQ_LK) &&
262 * Accordingly with provided
263 * thresholds, this thread is
264 * stuck for too long on a
266 * However, being on a
267 * sleepqueue, we might still
268 * check for the blessed
272 for (i = 0; blessed[i] != NULL;
274 if (!strcmp(blessed[i],
285 sx_sunlock(&allproc_lock);
286 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
287 __func__, td, tticks);
294 sx_sunlock(&allproc_lock);
296 /* Sleep for sleepfreq seconds. */
297 pause("-", sleepfreq * hz);
301 static struct kthread_desc deadlkres_kd = {
304 (struct thread **)NULL
307 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
309 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0,
310 "Deadlock resolver");
311 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
312 &slptime_threshold, 0,
313 "Number of seconds within is valid to sleep on a sleepqueue");
314 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
315 &blktime_threshold, 0,
316 "Number of seconds within is valid to block on a turnstile");
317 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
318 "Number of seconds between any deadlock resolver thread run");
319 #endif /* DEADLKRES */
322 read_cpu_time(long *cp_time)
327 /* Sum up global cp_time[]. */
328 bzero(cp_time, sizeof(long) * CPUSTATES);
331 for (j = 0; j < CPUSTATES; j++)
332 cp_time[j] += pc->pc_cp_time[j];
337 #include <sys/watchdog.h>
339 static int watchdog_ticks;
340 static int watchdog_enabled;
341 static void watchdog_fire(void);
342 static void watchdog_config(void *, u_int, int *);
343 #endif /* SW_WATCHDOG */
346 * Clock handling routines.
348 * This code is written to operate with two timers that run independently of
351 * The main timer, running hz times per second, is used to trigger interval
352 * timers, timeouts and rescheduling as needed.
354 * The second timer handles kernel and user profiling,
355 * and does resource use estimation. If the second timer is programmable,
356 * it is randomized to avoid aliasing between the two clocks. For example,
357 * the randomization prevents an adversary from always giving up the cpu
358 * just before its quantum expires. Otherwise, it would never accumulate
359 * cpu ticks. The mean frequency of the second timer is stathz.
361 * If no second timer exists, stathz will be zero; in this case we drive
362 * profiling and statistics off the main clock. This WILL NOT be accurate;
363 * do not do it unless absolutely necessary.
365 * The statistics clock may (or may not) be run at a higher rate while
366 * profiling. This profile clock runs at profhz. We require that profhz
367 * be an integral multiple of stathz.
369 * If the statistics clock is running fast, it must be divided by the ratio
370 * profhz/stathz for statistics. (For profiling, every tick counts.)
372 * Time-of-day is maintained using a "timecounter", which may or may
373 * not be related to the hardware generating the above mentioned
383 static DPCPU_DEFINE(int, pcputicks); /* Per-CPU version of ticks. */
384 #ifdef DEVICE_POLLING
385 static int devpoll_run = 0;
389 * Initialize clock frequencies and start both clocks running.
396 #ifdef EARLY_AP_STARTUP
403 * Set divisors to 1 (normal case) and let the machine-specific
406 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
410 * Compute profhz/stathz, and fix profhz if needed.
412 i = stathz ? stathz : hz;
415 psratio = profhz / i;
417 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
420 * Arrange for ticks to wrap 10 minutes after boot to help catch
421 * sign problems sooner.
423 ticks = INT_MAX - (hz * 10 * 60);
425 #ifdef EARLY_AP_STARTUP
427 * Fixup the tick counts in any blocked or sleeping threads to
428 * account for the jump above.
430 sx_slock(&allproc_lock);
431 FOREACH_PROC_IN_SYSTEM(p) {
433 if (p->p_state == PRS_NEW) {
437 FOREACH_THREAD_IN_PROC(p, td) {
439 if (TD_ON_LOCK(td)) {
440 MPASS(td->td_blktick == 0);
441 td->td_blktick = ticks;
443 if (TD_ON_SLEEPQ(td)) {
444 MPASS(td->td_slptick == 0);
445 td->td_slptick = ticks;
451 sx_sunlock(&allproc_lock);
456 * Each time the real-time timer fires, this function is called on all CPUs.
457 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only
458 * the other CPUs in the system need to call this function.
461 hardclock_cpu(int usermode)
463 struct pstats *pstats;
464 struct thread *td = curthread;
465 struct proc *p = td->td_proc;
469 * Run current process's virtual and profile time, as needed.
474 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
476 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
477 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
480 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
482 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
483 flags |= TDF_PROFPEND | TDF_ASTPENDING;
487 td->td_flags |= flags;
491 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
492 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
493 if (td->td_intr_frame != NULL)
494 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
496 callout_process(sbinuptime());
500 * The real-time timer, interrupting hz times per second.
503 hardclock(int usermode, uintfptr_t pc)
506 atomic_add_int(&ticks, 1);
507 hardclock_cpu(usermode);
509 cpu_tick_calibration();
511 * If no separate statistics clock is available, run it from here.
513 * XXX: this only works for UP
516 profclock(usermode, pc);
519 #ifdef DEVICE_POLLING
520 hardclock_device_poll(); /* this is very short and quick */
521 #endif /* DEVICE_POLLING */
523 if (watchdog_enabled > 0 && --watchdog_ticks <= 0)
525 #endif /* SW_WATCHDOG */
529 hardclock_cnt(int cnt, int usermode)
531 struct pstats *pstats;
532 struct thread *td = curthread;
533 struct proc *p = td->td_proc;
534 int *t = DPCPU_PTR(pcputicks);
535 int flags, global, newticks;
538 #endif /* SW_WATCHDOG */
541 * Update per-CPU and possibly global ticks values.
546 newticks = *t - global;
553 } while (!atomic_cmpset_int(&ticks, global, *t));
556 * Run current process's virtual and profile time, as needed.
561 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
563 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
565 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
568 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
570 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
572 flags |= TDF_PROFPEND | TDF_ASTPENDING;
577 td->td_flags |= flags;
582 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
583 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
584 if (td->td_intr_frame != NULL)
585 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
587 /* We are in charge to handle this tick duty. */
589 tc_ticktock(newticks);
590 #ifdef DEVICE_POLLING
591 /* Dangerous and no need to call these things concurrently. */
592 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
593 /* This is very short and quick. */
594 hardclock_device_poll();
595 atomic_store_rel_int(&devpoll_run, 0);
597 #endif /* DEVICE_POLLING */
599 if (watchdog_enabled > 0) {
600 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
601 if (i > 0 && i <= newticks)
604 #endif /* SW_WATCHDOG */
606 if (curcpu == CPU_FIRST())
607 cpu_tick_calibration();
611 hardclock_sync(int cpu)
613 int *t = DPCPU_ID_PTR(cpu, pcputicks);
619 * Compute number of ticks in the specified amount of time.
625 register unsigned long ticks;
626 register long sec, usec;
629 * If the number of usecs in the whole seconds part of the time
630 * difference fits in a long, then the total number of usecs will
631 * fit in an unsigned long. Compute the total and convert it to
632 * ticks, rounding up and adding 1 to allow for the current tick
633 * to expire. Rounding also depends on unsigned long arithmetic
636 * Otherwise, if the number of ticks in the whole seconds part of
637 * the time difference fits in a long, then convert the parts to
638 * ticks separately and add, using similar rounding methods and
639 * overflow avoidance. This method would work in the previous
640 * case but it is slightly slower and assumes that hz is integral.
642 * Otherwise, round the time difference down to the maximum
643 * representable value.
645 * If ints have 32 bits, then the maximum value for any timeout in
646 * 10ms ticks is 248 days.
660 printf("tvotohz: negative time difference %ld sec %ld usec\n",
664 } else if (sec <= LONG_MAX / 1000000)
665 ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1;
666 else if (sec <= LONG_MAX / hz)
668 + howmany((unsigned long)usec, tick) + 1;
677 * Start profiling on a process.
679 * Kernel profiling passes proc0 which never exits and hence
680 * keeps the profile clock running constantly.
684 register struct proc *p;
687 PROC_LOCK_ASSERT(p, MA_OWNED);
688 if (p->p_flag & P_STOPPROF)
690 if ((p->p_flag & P_PROFIL) == 0) {
691 p->p_flag |= P_PROFIL;
692 mtx_lock(&time_lock);
693 if (++profprocs == 1)
694 cpu_startprofclock();
695 mtx_unlock(&time_lock);
700 * Stop profiling on a process.
704 register struct proc *p;
707 PROC_LOCK_ASSERT(p, MA_OWNED);
708 if (p->p_flag & P_PROFIL) {
709 if (p->p_profthreads != 0) {
710 while (p->p_profthreads != 0) {
711 p->p_flag |= P_STOPPROF;
712 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
716 if ((p->p_flag & P_PROFIL) == 0)
718 p->p_flag &= ~P_PROFIL;
719 mtx_lock(&time_lock);
720 if (--profprocs == 0)
722 mtx_unlock(&time_lock);
727 * Statistics clock. Updates rusage information and calls the scheduler
728 * to adjust priorities of the active thread.
730 * This should be called by all active processors.
733 statclock(int usermode)
736 statclock_cnt(1, usermode);
740 statclock_cnt(int cnt, int usermode)
752 cp_time = (long *)PCPU_PTR(cp_time);
755 * Charge the time as appropriate.
757 td->td_uticks += cnt;
758 if (p->p_nice > NZERO)
759 cp_time[CP_NICE] += cnt;
761 cp_time[CP_USER] += cnt;
764 * Came from kernel mode, so we were:
765 * - handling an interrupt,
766 * - doing syscall or trap work on behalf of the current
768 * - spinning in the idle loop.
769 * Whichever it is, charge the time as appropriate.
770 * Note that we charge interrupts to the current process,
771 * regardless of whether they are ``for'' that process,
772 * so that we know how much of its real time was spent
773 * in ``non-process'' (i.e., interrupt) work.
775 if ((td->td_pflags & TDP_ITHREAD) ||
776 td->td_intr_nesting_level >= 2) {
777 td->td_iticks += cnt;
778 cp_time[CP_INTR] += cnt;
780 td->td_pticks += cnt;
781 td->td_sticks += cnt;
782 if (!TD_IS_IDLETHREAD(td))
783 cp_time[CP_SYS] += cnt;
785 cp_time[CP_IDLE] += cnt;
789 /* Update resource usage integrals and maximums. */
790 MPASS(p->p_vmspace != NULL);
793 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
794 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
795 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
796 rss = pgtok(vmspace_resident_count(vm));
797 if (ru->ru_maxrss < rss)
799 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
800 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
801 SDT_PROBE2(sched, , , tick, td, td->td_proc);
802 thread_lock_flags(td, MTX_QUIET);
803 for ( ; cnt > 0; cnt--)
807 if (td->td_intr_frame != NULL)
808 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
813 profclock(int usermode, uintfptr_t pc)
816 profclock_cnt(1, usermode, pc);
820 profclock_cnt(int cnt, int usermode, uintfptr_t pc)
831 * Came from user mode; CPU was in user state.
832 * If this process is being profiled, record the tick.
833 * if there is no related user location yet, don't
834 * bother trying to count it.
836 if (td->td_proc->p_flag & P_PROFIL)
837 addupc_intr(td, pc, cnt);
842 * Kernel statistics are just like addupc_intr, only easier.
845 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
847 if (i < g->textsize) {
854 if (td->td_intr_frame != NULL)
855 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
860 * Return information about system clocks.
863 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
865 struct clockinfo clkinfo;
867 * Construct clockinfo structure.
869 bzero(&clkinfo, sizeof(clkinfo));
872 clkinfo.profhz = profhz;
873 clkinfo.stathz = stathz ? stathz : hz;
874 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
877 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
878 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
879 0, 0, sysctl_kern_clockrate, "S,clockinfo",
880 "Rate and period of various kernel clocks");
885 watchdog_config(void *unused __unused, u_int cmd, int *error)
889 u = cmd & WD_INTERVAL;
890 if (u >= WD_TO_1SEC) {
891 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
892 watchdog_enabled = 1;
895 watchdog_enabled = 0;
900 * Handle a watchdog timeout by dumping interrupt information and
901 * then either dropping to DDB or panicking.
914 nintr = sintrcnt / sizeof(u_long);
916 printf("interrupt total\n");
917 while (--nintr >= 0) {
919 printf("%-12s %20lu\n", curname, *curintr);
920 curname += strlen(curname) + 1;
921 inttotal += *curintr++;
923 printf("Total %20ju\n", (uintmax_t)inttotal);
925 #if defined(KDB) && !defined(KDB_UNATTENDED)
927 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
929 panic("watchdog timeout");
933 #endif /* SW_WATCHDOG */