2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
43 #include "opt_device_polling.h"
44 #include "opt_hwpmc_hooks.h"
46 #include "opt_watchdog.h"
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/callout.h>
51 #include <sys/epoch.h>
52 #include <sys/gtaskqueue.h>
54 #include <sys/kernel.h>
55 #include <sys/kthread.h>
58 #include <sys/mutex.h>
60 #include <sys/resource.h>
61 #include <sys/resourcevar.h>
62 #include <sys/sched.h>
64 #include <sys/signalvar.h>
65 #include <sys/sleepqueue.h>
69 #include <vm/vm_map.h>
70 #include <sys/sysctl.h>
72 #include <sys/interrupt.h>
73 #include <sys/limits.h>
74 #include <sys/timetc.h>
81 #include <sys/pmckern.h>
82 PMC_SOFT_DEFINE( , , clock, hard);
83 PMC_SOFT_DEFINE( , , clock, stat);
84 PMC_SOFT_DEFINE_EX( , , clock, prof, \
85 cpu_startprofclock, cpu_stopprofclock);
89 extern void hardclock_device_poll(void);
90 #endif /* DEVICE_POLLING */
92 static void initclocks(void *dummy);
93 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
95 /* Spin-lock protecting profiling statistics. */
96 static struct mtx time_lock;
98 SDT_PROVIDER_DECLARE(sched);
99 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
102 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
105 long cp_time[CPUSTATES];
108 unsigned int cp_time32[CPUSTATES];
111 read_cpu_time(cp_time);
113 if (req->flags & SCTL_MASK32) {
115 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
116 for (i = 0; i < CPUSTATES; i++)
117 cp_time32[i] = (unsigned int)cp_time[i];
118 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
123 return SYSCTL_OUT(req, 0, sizeof(cp_time));
124 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
129 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
130 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
132 static long empty[CPUSTATES];
135 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
142 unsigned int cp_time32[CPUSTATES];
148 if (req->flags & SCTL_MASK32)
149 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
152 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
154 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
155 if (!CPU_ABSENT(c)) {
157 cp_time = pcpu->pc_cp_time;
162 if (req->flags & SCTL_MASK32) {
163 for (i = 0; i < CPUSTATES; i++)
164 cp_time32[i] = (unsigned int)cp_time[i];
165 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
168 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
173 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
174 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
177 static const char *blessed[] = {
183 static int slptime_threshold = 1800;
184 static int blktime_threshold = 900;
185 static int sleepfreq = 3;
193 int blkticks, i, slpticks, slptype, tryl, tticks;
197 blkticks = blktime_threshold * hz;
198 slpticks = slptime_threshold * hz;
201 * Avoid to sleep on the sx_lock in order to avoid a possible
202 * priority inversion problem leading to starvation.
203 * If the lock can't be held after 100 tries, panic.
205 if (!sx_try_slock(&allproc_lock)) {
207 panic("%s: possible deadlock detected on allproc_lock\n",
210 pause("allproc", sleepfreq * hz);
214 FOREACH_PROC_IN_SYSTEM(p) {
216 if (p->p_state == PRS_NEW) {
220 FOREACH_THREAD_IN_PROC(p, td) {
223 if (TD_ON_LOCK(td)) {
226 * The thread should be blocked on a
227 * turnstile, simply check if the
228 * turnstile channel is in good state.
230 MPASS(td->td_blocked != NULL);
232 tticks = ticks - td->td_blktick;
234 if (tticks > blkticks) {
237 * Accordingly with provided
238 * thresholds, this thread is
239 * stuck for too long on a
243 sx_sunlock(&allproc_lock);
244 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
245 __func__, td, tticks);
247 } else if (TD_IS_SLEEPING(td) &&
251 * Check if the thread is sleeping on a
252 * lock, otherwise skip the check.
253 * Drop the thread lock in order to
254 * avoid a LOR with the sleepqueue
257 wchan = td->td_wchan;
258 tticks = ticks - td->td_slptick;
260 slptype = sleepq_type(wchan);
261 if ((slptype == SLEEPQ_SX ||
262 slptype == SLEEPQ_LK) &&
266 * Accordingly with provided
267 * thresholds, this thread is
268 * stuck for too long on a
270 * However, being on a
271 * sleepqueue, we might still
272 * check for the blessed
276 for (i = 0; blessed[i] != NULL;
278 if (!strcmp(blessed[i],
289 sx_sunlock(&allproc_lock);
290 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
291 __func__, td, tticks);
298 sx_sunlock(&allproc_lock);
300 /* Sleep for sleepfreq seconds. */
301 pause("-", sleepfreq * hz);
305 static struct kthread_desc deadlkres_kd = {
308 (struct thread **)NULL
311 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
313 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0,
314 "Deadlock resolver");
315 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
316 &slptime_threshold, 0,
317 "Number of seconds within is valid to sleep on a sleepqueue");
318 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
319 &blktime_threshold, 0,
320 "Number of seconds within is valid to block on a turnstile");
321 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
322 "Number of seconds between any deadlock resolver thread run");
323 #endif /* DEADLKRES */
326 read_cpu_time(long *cp_time)
331 /* Sum up global cp_time[]. */
332 bzero(cp_time, sizeof(long) * CPUSTATES);
335 for (j = 0; j < CPUSTATES; j++)
336 cp_time[j] += pc->pc_cp_time[j];
340 #include <sys/watchdog.h>
342 static int watchdog_ticks;
343 static int watchdog_enabled;
344 static void watchdog_fire(void);
345 static void watchdog_config(void *, u_int, int *);
348 watchdog_attach(void)
350 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
354 * Clock handling routines.
356 * This code is written to operate with two timers that run independently of
359 * The main timer, running hz times per second, is used to trigger interval
360 * timers, timeouts and rescheduling as needed.
362 * The second timer handles kernel and user profiling,
363 * and does resource use estimation. If the second timer is programmable,
364 * it is randomized to avoid aliasing between the two clocks. For example,
365 * the randomization prevents an adversary from always giving up the cpu
366 * just before its quantum expires. Otherwise, it would never accumulate
367 * cpu ticks. The mean frequency of the second timer is stathz.
369 * If no second timer exists, stathz will be zero; in this case we drive
370 * profiling and statistics off the main clock. This WILL NOT be accurate;
371 * do not do it unless absolutely necessary.
373 * The statistics clock may (or may not) be run at a higher rate while
374 * profiling. This profile clock runs at profhz. We require that profhz
375 * be an integral multiple of stathz.
377 * If the statistics clock is running fast, it must be divided by the ratio
378 * profhz/stathz for statistics. (For profiling, every tick counts.)
380 * Time-of-day is maintained using a "timecounter", which may or may
381 * not be related to the hardware generating the above mentioned
391 static DPCPU_DEFINE(int, pcputicks); /* Per-CPU version of ticks. */
392 #ifdef DEVICE_POLLING
393 static int devpoll_run = 0;
397 * Initialize clock frequencies and start both clocks running.
401 initclocks(void *dummy)
406 * Set divisors to 1 (normal case) and let the machine-specific
409 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
413 * Compute profhz/stathz, and fix profhz if needed.
415 i = stathz ? stathz : hz;
418 psratio = profhz / i;
421 /* Enable hardclock watchdog now, even if a hardware watchdog exists. */
424 /* Volunteer to run a software watchdog. */
425 if (wdog_software_attach == NULL)
426 wdog_software_attach = watchdog_attach;
431 * Each time the real-time timer fires, this function is called on all CPUs.
432 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only
433 * the other CPUs in the system need to call this function.
436 hardclock_cpu(int usermode)
438 struct pstats *pstats;
439 struct thread *td = curthread;
440 struct proc *p = td->td_proc;
444 * Run current process's virtual and profile time, as needed.
449 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
451 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
452 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
455 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
457 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
458 flags |= TDF_PROFPEND | TDF_ASTPENDING;
462 td->td_flags |= flags;
466 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
467 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
468 if (td->td_intr_frame != NULL)
469 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
471 callout_process(sbinuptime());
472 if (__predict_false(DPCPU_GET(epoch_cb_count)))
473 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
477 * The real-time timer, interrupting hz times per second.
480 hardclock(int usermode, uintfptr_t pc)
483 atomic_add_int(&ticks, 1);
484 hardclock_cpu(usermode);
486 cpu_tick_calibration();
488 * If no separate statistics clock is available, run it from here.
490 * XXX: this only works for UP
493 profclock(usermode, pc);
496 #ifdef DEVICE_POLLING
497 hardclock_device_poll(); /* this is very short and quick */
498 #endif /* DEVICE_POLLING */
499 if (watchdog_enabled > 0 && --watchdog_ticks <= 0)
504 hardclock_cnt(int cnt, int usermode)
506 struct pstats *pstats;
507 struct thread *td = curthread;
508 struct proc *p = td->td_proc;
509 int *t = DPCPU_PTR(pcputicks);
510 int flags, global, newticks;
514 * Update per-CPU and possibly global ticks values.
519 newticks = *t - global;
526 } while (!atomic_cmpset_int(&ticks, global, *t));
529 * Run current process's virtual and profile time, as needed.
534 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
536 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
538 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
541 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
543 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
545 flags |= TDF_PROFPEND | TDF_ASTPENDING;
550 td->td_flags |= flags;
555 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
556 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
557 if (td->td_intr_frame != NULL)
558 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
560 /* We are in charge to handle this tick duty. */
562 tc_ticktock(newticks);
563 #ifdef DEVICE_POLLING
564 /* Dangerous and no need to call these things concurrently. */
565 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
566 /* This is very short and quick. */
567 hardclock_device_poll();
568 atomic_store_rel_int(&devpoll_run, 0);
570 #endif /* DEVICE_POLLING */
571 if (watchdog_enabled > 0) {
572 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
573 if (i > 0 && i <= newticks)
577 if (curcpu == CPU_FIRST())
578 cpu_tick_calibration();
579 if (__predict_false(DPCPU_GET(epoch_cb_count)))
580 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
584 hardclock_sync(int cpu)
587 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
588 t = DPCPU_ID_PTR(cpu, pcputicks);
594 * Compute number of ticks in the specified amount of time.
597 tvtohz(struct timeval *tv)
603 * If the number of usecs in the whole seconds part of the time
604 * difference fits in a long, then the total number of usecs will
605 * fit in an unsigned long. Compute the total and convert it to
606 * ticks, rounding up and adding 1 to allow for the current tick
607 * to expire. Rounding also depends on unsigned long arithmetic
610 * Otherwise, if the number of ticks in the whole seconds part of
611 * the time difference fits in a long, then convert the parts to
612 * ticks separately and add, using similar rounding methods and
613 * overflow avoidance. This method would work in the previous
614 * case but it is slightly slower and assumes that hz is integral.
616 * Otherwise, round the time difference down to the maximum
617 * representable value.
619 * If ints have 32 bits, then the maximum value for any timeout in
620 * 10ms ticks is 248 days.
634 printf("tvotohz: negative time difference %ld sec %ld usec\n",
638 } else if (sec <= LONG_MAX / 1000000)
639 ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1;
640 else if (sec <= LONG_MAX / hz)
642 + howmany((unsigned long)usec, tick) + 1;
651 * Start profiling on a process.
653 * Kernel profiling passes proc0 which never exits and hence
654 * keeps the profile clock running constantly.
657 startprofclock(struct proc *p)
660 PROC_LOCK_ASSERT(p, MA_OWNED);
661 if (p->p_flag & P_STOPPROF)
663 if ((p->p_flag & P_PROFIL) == 0) {
664 p->p_flag |= P_PROFIL;
665 mtx_lock(&time_lock);
666 if (++profprocs == 1)
667 cpu_startprofclock();
668 mtx_unlock(&time_lock);
673 * Stop profiling on a process.
676 stopprofclock(struct proc *p)
679 PROC_LOCK_ASSERT(p, MA_OWNED);
680 if (p->p_flag & P_PROFIL) {
681 if (p->p_profthreads != 0) {
682 while (p->p_profthreads != 0) {
683 p->p_flag |= P_STOPPROF;
684 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
688 if ((p->p_flag & P_PROFIL) == 0)
690 p->p_flag &= ~P_PROFIL;
691 mtx_lock(&time_lock);
692 if (--profprocs == 0)
694 mtx_unlock(&time_lock);
699 * Statistics clock. Updates rusage information and calls the scheduler
700 * to adjust priorities of the active thread.
702 * This should be called by all active processors.
705 statclock(int usermode)
708 statclock_cnt(1, usermode);
712 statclock_cnt(int cnt, int usermode)
724 cp_time = (long *)PCPU_PTR(cp_time);
727 * Charge the time as appropriate.
729 td->td_uticks += cnt;
730 if (p->p_nice > NZERO)
731 cp_time[CP_NICE] += cnt;
733 cp_time[CP_USER] += cnt;
736 * Came from kernel mode, so we were:
737 * - handling an interrupt,
738 * - doing syscall or trap work on behalf of the current
740 * - spinning in the idle loop.
741 * Whichever it is, charge the time as appropriate.
742 * Note that we charge interrupts to the current process,
743 * regardless of whether they are ``for'' that process,
744 * so that we know how much of its real time was spent
745 * in ``non-process'' (i.e., interrupt) work.
747 if ((td->td_pflags & TDP_ITHREAD) ||
748 td->td_intr_nesting_level >= 2) {
749 td->td_iticks += cnt;
750 cp_time[CP_INTR] += cnt;
752 td->td_pticks += cnt;
753 td->td_sticks += cnt;
754 if (!TD_IS_IDLETHREAD(td))
755 cp_time[CP_SYS] += cnt;
757 cp_time[CP_IDLE] += cnt;
761 /* Update resource usage integrals and maximums. */
762 MPASS(p->p_vmspace != NULL);
765 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
766 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
767 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
768 rss = pgtok(vmspace_resident_count(vm));
769 if (ru->ru_maxrss < rss)
771 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
772 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
773 SDT_PROBE2(sched, , , tick, td, td->td_proc);
774 thread_lock_flags(td, MTX_QUIET);
775 for ( ; cnt > 0; cnt--)
779 if (td->td_intr_frame != NULL)
780 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
785 profclock(int usermode, uintfptr_t pc)
788 profclock_cnt(1, usermode, pc);
792 profclock_cnt(int cnt, int usermode, uintfptr_t pc)
803 * Came from user mode; CPU was in user state.
804 * If this process is being profiled, record the tick.
805 * if there is no related user location yet, don't
806 * bother trying to count it.
808 if (td->td_proc->p_flag & P_PROFIL)
809 addupc_intr(td, pc, cnt);
814 * Kernel statistics are just like addupc_intr, only easier.
817 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
819 if (i < g->textsize) {
826 if (td->td_intr_frame != NULL)
827 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
832 * Return information about system clocks.
835 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
837 struct clockinfo clkinfo;
839 * Construct clockinfo structure.
841 bzero(&clkinfo, sizeof(clkinfo));
844 clkinfo.profhz = profhz;
845 clkinfo.stathz = stathz ? stathz : hz;
846 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
849 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
850 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
851 0, 0, sysctl_kern_clockrate, "S,clockinfo",
852 "Rate and period of various kernel clocks");
855 watchdog_config(void *unused __unused, u_int cmd, int *error)
859 u = cmd & WD_INTERVAL;
860 if (u >= WD_TO_1SEC) {
861 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
862 watchdog_enabled = 1;
865 watchdog_enabled = 0;
870 * Handle a watchdog timeout by dumping interrupt information and
871 * then either dropping to DDB or panicking.
884 nintr = sintrcnt / sizeof(u_long);
886 printf("interrupt total\n");
887 while (--nintr >= 0) {
889 printf("%-12s %20lu\n", curname, *curintr);
890 curname += strlen(curname) + 1;
891 inttotal += *curintr++;
893 printf("Total %20ju\n", (uintmax_t)inttotal);
895 #if defined(KDB) && !defined(KDB_UNATTENDED)
897 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
899 panic("watchdog timeout");