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"
43 #include "opt_kdtrace.h"
45 #include "opt_watchdog.h"
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/callout.h>
51 #include <sys/kernel.h>
52 #include <sys/kthread.h>
55 #include <sys/mutex.h>
57 #include <sys/resource.h>
58 #include <sys/resourcevar.h>
59 #include <sys/sched.h>
61 #include <sys/signalvar.h>
62 #include <sys/sleepqueue.h>
66 #include <vm/vm_map.h>
67 #include <sys/sysctl.h>
69 #include <sys/interrupt.h>
70 #include <sys/limits.h>
71 #include <sys/timetc.h>
78 #include <sys/pmckern.h>
79 PMC_SOFT_DEFINE( , , clock, hard);
80 PMC_SOFT_DEFINE( , , clock, stat);
81 PMC_SOFT_DEFINE_EX( , , clock, prof, \
82 cpu_startprofclock, cpu_stopprofclock);
86 extern void hardclock_device_poll(void);
87 #endif /* DEVICE_POLLING */
89 static void initclocks(void *dummy);
90 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
92 /* Spin-lock protecting profiling statistics. */
93 static struct mtx time_lock;
95 SDT_PROVIDER_DECLARE(sched);
96 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
99 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
102 long cp_time[CPUSTATES];
105 unsigned int cp_time32[CPUSTATES];
108 read_cpu_time(cp_time);
110 if (req->flags & SCTL_MASK32) {
112 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
113 for (i = 0; i < CPUSTATES; i++)
114 cp_time32[i] = (unsigned int)cp_time[i];
115 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
120 return SYSCTL_OUT(req, 0, sizeof(cp_time));
121 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
126 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
127 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
129 static long empty[CPUSTATES];
132 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
139 unsigned int cp_time32[CPUSTATES];
145 if (req->flags & SCTL_MASK32)
146 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
149 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
151 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
152 if (!CPU_ABSENT(c)) {
154 cp_time = pcpu->pc_cp_time;
159 if (req->flags & SCTL_MASK32) {
160 for (i = 0; i < CPUSTATES; i++)
161 cp_time32[i] = (unsigned int)cp_time[i];
162 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
165 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
170 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
171 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
174 static const char *blessed[] = {
180 static int slptime_threshold = 1800;
181 static int blktime_threshold = 900;
182 static int sleepfreq = 3;
190 int blkticks, i, slpticks, slptype, tryl, tticks;
194 blkticks = blktime_threshold * hz;
195 slpticks = slptime_threshold * hz;
198 * Avoid to sleep on the sx_lock in order to avoid a possible
199 * priority inversion problem leading to starvation.
200 * If the lock can't be held after 100 tries, panic.
202 if (!sx_try_slock(&allproc_lock)) {
204 panic("%s: possible deadlock detected on allproc_lock\n",
207 pause("allproc", sleepfreq * hz);
211 FOREACH_PROC_IN_SYSTEM(p) {
213 if (p->p_state == PRS_NEW) {
217 FOREACH_THREAD_IN_PROC(p, td) {
220 if (TD_ON_LOCK(td)) {
223 * The thread should be blocked on a
224 * turnstile, simply check if the
225 * turnstile channel is in good state.
227 MPASS(td->td_blocked != NULL);
229 tticks = ticks - td->td_blktick;
231 if (tticks > blkticks) {
234 * Accordingly with provided
235 * thresholds, this thread is
236 * stuck for too long on a
240 sx_sunlock(&allproc_lock);
241 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
242 __func__, td, tticks);
244 } else if (TD_IS_SLEEPING(td) &&
248 * Check if the thread is sleeping on a
249 * lock, otherwise skip the check.
250 * Drop the thread lock in order to
251 * avoid a LOR with the sleepqueue
254 wchan = td->td_wchan;
255 tticks = ticks - td->td_slptick;
257 slptype = sleepq_type(wchan);
258 if ((slptype == SLEEPQ_SX ||
259 slptype == SLEEPQ_LK) &&
263 * Accordingly with provided
264 * thresholds, this thread is
265 * stuck for too long on a
267 * However, being on a
268 * sleepqueue, we might still
269 * check for the blessed
273 for (i = 0; blessed[i] != NULL;
275 if (!strcmp(blessed[i],
286 sx_sunlock(&allproc_lock);
287 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
288 __func__, td, tticks);
295 sx_sunlock(&allproc_lock);
297 /* Sleep for sleepfreq seconds. */
298 pause("-", sleepfreq * hz);
302 static struct kthread_desc deadlkres_kd = {
305 (struct thread **)NULL
308 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
310 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0,
311 "Deadlock resolver");
312 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
313 &slptime_threshold, 0,
314 "Number of seconds within is valid to sleep on a sleepqueue");
315 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
316 &blktime_threshold, 0,
317 "Number of seconds within is valid to block on a turnstile");
318 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
319 "Number of seconds between any deadlock resolver thread run");
320 #endif /* DEADLKRES */
323 read_cpu_time(long *cp_time)
328 /* Sum up global cp_time[]. */
329 bzero(cp_time, sizeof(long) * CPUSTATES);
332 for (j = 0; j < CPUSTATES; j++)
333 cp_time[j] += pc->pc_cp_time[j];
338 #include <sys/watchdog.h>
340 static int watchdog_ticks;
341 static int watchdog_enabled;
342 static void watchdog_fire(void);
343 static void watchdog_config(void *, u_int, int *);
344 #endif /* SW_WATCHDOG */
347 * Clock handling routines.
349 * This code is written to operate with two timers that run independently of
352 * The main timer, running hz times per second, is used to trigger interval
353 * timers, timeouts and rescheduling as needed.
355 * The second timer handles kernel and user profiling,
356 * and does resource use estimation. If the second timer is programmable,
357 * it is randomized to avoid aliasing between the two clocks. For example,
358 * the randomization prevents an adversary from always giving up the cpu
359 * just before its quantum expires. Otherwise, it would never accumulate
360 * cpu ticks. The mean frequency of the second timer is stathz.
362 * If no second timer exists, stathz will be zero; in this case we drive
363 * profiling and statistics off the main clock. This WILL NOT be accurate;
364 * do not do it unless absolutely necessary.
366 * The statistics clock may (or may not) be run at a higher rate while
367 * profiling. This profile clock runs at profhz. We require that profhz
368 * be an integral multiple of stathz.
370 * If the statistics clock is running fast, it must be divided by the ratio
371 * profhz/stathz for statistics. (For profiling, every tick counts.)
373 * Time-of-day is maintained using a "timecounter", which may or may
374 * not be related to the hardware generating the above mentioned
384 static DPCPU_DEFINE(int, pcputicks); /* Per-CPU version of ticks. */
385 static int global_hardclock_run = 0;
388 * Initialize clock frequencies and start both clocks running.
398 * Set divisors to 1 (normal case) and let the machine-specific
401 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
405 * Compute profhz/stathz, and fix profhz if needed.
407 i = stathz ? stathz : hz;
410 psratio = profhz / i;
412 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
417 * Each time the real-time timer fires, this function is called on all CPUs.
418 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only
419 * the other CPUs in the system need to call this function.
422 hardclock_cpu(int usermode)
424 struct pstats *pstats;
425 struct thread *td = curthread;
426 struct proc *p = td->td_proc;
430 * Run current process's virtual and profile time, as needed.
435 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
437 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
438 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
441 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
443 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
444 flags |= TDF_PROFPEND | TDF_ASTPENDING;
449 td->td_flags |= flags;
453 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
454 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
455 if (td->td_intr_frame != NULL)
456 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
458 callout_process(sbinuptime());
462 * The real-time timer, interrupting hz times per second.
465 hardclock(int usermode, uintfptr_t pc)
468 atomic_add_int(&ticks, 1);
469 hardclock_cpu(usermode);
471 cpu_tick_calibration();
473 * If no separate statistics clock is available, run it from here.
475 * XXX: this only works for UP
478 profclock(usermode, pc);
481 #ifdef DEVICE_POLLING
482 hardclock_device_poll(); /* this is very short and quick */
483 #endif /* DEVICE_POLLING */
485 if (watchdog_enabled > 0 && --watchdog_ticks <= 0)
487 #endif /* SW_WATCHDOG */
491 hardclock_cnt(int cnt, int usermode)
493 struct pstats *pstats;
494 struct thread *td = curthread;
495 struct proc *p = td->td_proc;
496 int *t = DPCPU_PTR(pcputicks);
497 int flags, global, newticks;
500 #endif /* SW_WATCHDOG */
503 * Update per-CPU and possibly global ticks values.
508 newticks = *t - global;
515 } while (!atomic_cmpset_int(&ticks, global, *t));
518 * Run current process's virtual and profile time, as needed.
523 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
525 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
527 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
530 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
532 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
534 flags |= TDF_PROFPEND | TDF_ASTPENDING;
539 td->td_flags |= flags;
543 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
544 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
545 if (td->td_intr_frame != NULL)
546 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
548 /* We are in charge to handle this tick duty. */
550 /* Dangerous and no need to call these things concurrently. */
551 if (atomic_cmpset_acq_int(&global_hardclock_run, 0, 1)) {
552 tc_ticktock(newticks);
553 #ifdef DEVICE_POLLING
554 /* This is very short and quick. */
555 hardclock_device_poll();
556 #endif /* DEVICE_POLLING */
557 atomic_store_rel_int(&global_hardclock_run, 0);
560 if (watchdog_enabled > 0) {
561 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
562 if (i > 0 && i <= newticks)
565 #endif /* SW_WATCHDOG */
567 if (curcpu == CPU_FIRST())
568 cpu_tick_calibration();
572 hardclock_sync(int cpu)
574 int *t = DPCPU_ID_PTR(cpu, pcputicks);
580 * Compute number of ticks in the specified amount of time.
586 register unsigned long ticks;
587 register long sec, usec;
590 * If the number of usecs in the whole seconds part of the time
591 * difference fits in a long, then the total number of usecs will
592 * fit in an unsigned long. Compute the total and convert it to
593 * ticks, rounding up and adding 1 to allow for the current tick
594 * to expire. Rounding also depends on unsigned long arithmetic
597 * Otherwise, if the number of ticks in the whole seconds part of
598 * the time difference fits in a long, then convert the parts to
599 * ticks separately and add, using similar rounding methods and
600 * overflow avoidance. This method would work in the previous
601 * case but it is slightly slower and assumes that hz is integral.
603 * Otherwise, round the time difference down to the maximum
604 * representable value.
606 * If ints have 32 bits, then the maximum value for any timeout in
607 * 10ms ticks is 248 days.
621 printf("tvotohz: negative time difference %ld sec %ld usec\n",
625 } else if (sec <= LONG_MAX / 1000000)
626 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
628 else if (sec <= LONG_MAX / hz)
630 + ((unsigned long)usec + (tick - 1)) / tick + 1;
639 * Start profiling on a process.
641 * Kernel profiling passes proc0 which never exits and hence
642 * keeps the profile clock running constantly.
646 register struct proc *p;
649 PROC_LOCK_ASSERT(p, MA_OWNED);
650 if (p->p_flag & P_STOPPROF)
652 if ((p->p_flag & P_PROFIL) == 0) {
653 p->p_flag |= P_PROFIL;
654 mtx_lock(&time_lock);
655 if (++profprocs == 1)
656 cpu_startprofclock();
657 mtx_unlock(&time_lock);
662 * Stop profiling on a process.
666 register struct proc *p;
669 PROC_LOCK_ASSERT(p, MA_OWNED);
670 if (p->p_flag & P_PROFIL) {
671 if (p->p_profthreads != 0) {
672 p->p_flag |= P_STOPPROF;
673 while (p->p_profthreads != 0)
674 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
676 p->p_flag &= ~P_STOPPROF;
678 if ((p->p_flag & P_PROFIL) == 0)
680 p->p_flag &= ~P_PROFIL;
681 mtx_lock(&time_lock);
682 if (--profprocs == 0)
684 mtx_unlock(&time_lock);
689 * Statistics clock. Updates rusage information and calls the scheduler
690 * to adjust priorities of the active thread.
692 * This should be called by all active processors.
695 statclock(int usermode)
698 statclock_cnt(1, usermode);
702 statclock_cnt(int cnt, int usermode)
714 cp_time = (long *)PCPU_PTR(cp_time);
717 * Charge the time as appropriate.
719 td->td_uticks += cnt;
720 if (p->p_nice > NZERO)
721 cp_time[CP_NICE] += cnt;
723 cp_time[CP_USER] += cnt;
726 * Came from kernel mode, so we were:
727 * - handling an interrupt,
728 * - doing syscall or trap work on behalf of the current
730 * - spinning in the idle loop.
731 * Whichever it is, charge the time as appropriate.
732 * Note that we charge interrupts to the current process,
733 * regardless of whether they are ``for'' that process,
734 * so that we know how much of its real time was spent
735 * in ``non-process'' (i.e., interrupt) work.
737 if ((td->td_pflags & TDP_ITHREAD) ||
738 td->td_intr_nesting_level >= 2) {
739 td->td_iticks += cnt;
740 cp_time[CP_INTR] += cnt;
742 td->td_pticks += cnt;
743 td->td_sticks += cnt;
744 if (!TD_IS_IDLETHREAD(td))
745 cp_time[CP_SYS] += cnt;
747 cp_time[CP_IDLE] += cnt;
751 /* Update resource usage integrals and maximums. */
752 MPASS(p->p_vmspace != NULL);
755 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
756 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
757 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
758 rss = pgtok(vmspace_resident_count(vm));
759 if (ru->ru_maxrss < rss)
761 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
762 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
763 SDT_PROBE2(sched, , , tick, td, td->td_proc);
764 thread_lock_flags(td, MTX_QUIET);
765 for ( ; cnt > 0; cnt--)
769 if (td->td_intr_frame != NULL)
770 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
775 profclock(int usermode, uintfptr_t pc)
778 profclock_cnt(1, usermode, pc);
782 profclock_cnt(int cnt, int usermode, uintfptr_t pc)
793 * Came from user mode; CPU was in user state.
794 * If this process is being profiled, record the tick.
795 * if there is no related user location yet, don't
796 * bother trying to count it.
798 if (td->td_proc->p_flag & P_PROFIL)
799 addupc_intr(td, pc, cnt);
804 * Kernel statistics are just like addupc_intr, only easier.
807 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
809 if (i < g->textsize) {
816 if (td->td_intr_frame != NULL)
817 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
822 * Return information about system clocks.
825 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
827 struct clockinfo clkinfo;
829 * Construct clockinfo structure.
831 bzero(&clkinfo, sizeof(clkinfo));
834 clkinfo.profhz = profhz;
835 clkinfo.stathz = stathz ? stathz : hz;
836 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
839 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
840 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
841 0, 0, sysctl_kern_clockrate, "S,clockinfo",
842 "Rate and period of various kernel clocks");
847 watchdog_config(void *unused __unused, u_int cmd, int *error)
851 u = cmd & WD_INTERVAL;
852 if (u >= WD_TO_1SEC) {
853 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
854 watchdog_enabled = 1;
857 watchdog_enabled = 0;
862 * Handle a watchdog timeout by dumping interrupt information and
863 * then either dropping to DDB or panicking.
876 nintr = sintrcnt / sizeof(u_long);
878 printf("interrupt total\n");
879 while (--nintr >= 0) {
881 printf("%-12s %20lu\n", curname, *curintr);
882 curname += strlen(curname) + 1;
883 inttotal += *curintr++;
885 printf("Total %20ju\n", (uintmax_t)inttotal);
887 #if defined(KDB) && !defined(KDB_UNATTENDED)
889 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
891 panic("watchdog timeout");
895 #endif /* SW_WATCHDOG */