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;
188 deadlres_td_on_lock(struct proc *p, struct thread *td, int blkticks)
192 sx_assert(&allproc_lock, SX_LOCKED);
193 PROC_LOCK_ASSERT(p, MA_OWNED);
194 THREAD_LOCK_ASSERT(td, MA_OWNED);
196 * The thread should be blocked on a turnstile, simply check
197 * if the turnstile channel is in good state.
199 MPASS(td->td_blocked != NULL);
201 tticks = ticks - td->td_blktick;
202 if (tticks > blkticks)
204 * Accordingly with provided thresholds, this thread is stuck
205 * for too long on a turnstile.
207 panic("%s: possible deadlock detected for %p, "
208 "blocked for %d ticks\n", __func__, td, tticks);
212 deadlres_td_sleep_q(struct proc *p, struct thread *td, int slpticks)
215 int i, slptype, tticks;
217 sx_assert(&allproc_lock, SX_LOCKED);
218 PROC_LOCK_ASSERT(p, MA_OWNED);
219 THREAD_LOCK_ASSERT(td, MA_OWNED);
221 * Check if the thread is sleeping on a lock, otherwise skip the check.
222 * Drop the thread lock in order to avoid a LOR with the sleepqueue
225 wchan = td->td_wchan;
226 tticks = ticks - td->td_slptick;
227 slptype = sleepq_type(wchan);
228 if ((slptype == SLEEPQ_SX || slptype == SLEEPQ_LK) &&
232 * Accordingly with provided thresholds, this thread is stuck
233 * for too long on a sleepqueue.
234 * However, being on a sleepqueue, we might still check for the
237 for (i = 0; blessed[i] != NULL; i++)
238 if (!strcmp(blessed[i], td->td_wmesg))
241 panic("%s: possible deadlock detected for %p, "
242 "blocked for %d ticks\n", __func__, td, tticks);
251 int blkticks, slpticks, tryl;
255 blkticks = blktime_threshold * hz;
256 slpticks = slptime_threshold * hz;
259 * Avoid to sleep on the sx_lock in order to avoid a
260 * possible priority inversion problem leading to
262 * If the lock can't be held after 100 tries, panic.
264 if (!sx_try_slock(&allproc_lock)) {
266 panic("%s: possible deadlock detected "
267 "on allproc_lock\n", __func__);
269 pause("allproc", sleepfreq * hz);
273 FOREACH_PROC_IN_SYSTEM(p) {
275 if (p->p_state == PRS_NEW) {
279 FOREACH_THREAD_IN_PROC(p, td) {
282 deadlres_td_on_lock(p, td,
284 else if (TD_IS_SLEEPING(td) &&
286 deadlres_td_sleep_q(p, td,
292 sx_sunlock(&allproc_lock);
294 /* Sleep for sleepfreq seconds. */
295 pause("-", sleepfreq * hz);
299 static struct kthread_desc deadlkres_kd = {
302 (struct thread **)NULL
305 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
307 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0,
308 "Deadlock resolver");
309 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
310 &slptime_threshold, 0,
311 "Number of seconds within is valid to sleep on a sleepqueue");
312 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
313 &blktime_threshold, 0,
314 "Number of seconds within is valid to block on a turnstile");
315 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
316 "Number of seconds between any deadlock resolver thread run");
317 #endif /* DEADLKRES */
320 read_cpu_time(long *cp_time)
325 /* Sum up global cp_time[]. */
326 bzero(cp_time, sizeof(long) * CPUSTATES);
329 for (j = 0; j < CPUSTATES; j++)
330 cp_time[j] += pc->pc_cp_time[j];
334 #include <sys/watchdog.h>
336 static int watchdog_ticks;
337 static int watchdog_enabled;
338 static void watchdog_fire(void);
339 static void watchdog_config(void *, u_int, int *);
342 watchdog_attach(void)
344 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
348 * Clock handling routines.
350 * This code is written to operate with two timers that run independently of
353 * The main timer, running hz times per second, is used to trigger interval
354 * timers, timeouts and rescheduling as needed.
356 * The second timer handles kernel and user profiling,
357 * and does resource use estimation. If the second timer is programmable,
358 * it is randomized to avoid aliasing between the two clocks. For example,
359 * the randomization prevents an adversary from always giving up the cpu
360 * just before its quantum expires. Otherwise, it would never accumulate
361 * cpu ticks. The mean frequency of the second timer is stathz.
363 * If no second timer exists, stathz will be zero; in this case we drive
364 * profiling and statistics off the main clock. This WILL NOT be accurate;
365 * do not do it unless absolutely necessary.
367 * The statistics clock may (or may not) be run at a higher rate while
368 * profiling. This profile clock runs at profhz. We require that profhz
369 * be an integral multiple of stathz.
371 * If the statistics clock is running fast, it must be divided by the ratio
372 * profhz/stathz for statistics. (For profiling, every tick counts.)
374 * Time-of-day is maintained using a "timecounter", which may or may
375 * not be related to the hardware generating the above mentioned
385 DPCPU_DEFINE_STATIC(int, pcputicks); /* Per-CPU version of ticks. */
386 #ifdef DEVICE_POLLING
387 static int devpoll_run = 0;
391 * Initialize clock frequencies and start both clocks running.
395 initclocks(void *dummy)
400 * Set divisors to 1 (normal case) and let the machine-specific
403 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
407 * Compute profhz/stathz, and fix profhz if needed.
409 i = stathz ? stathz : hz;
412 psratio = profhz / i;
415 /* Enable hardclock watchdog now, even if a hardware watchdog exists. */
418 /* Volunteer to run a software watchdog. */
419 if (wdog_software_attach == NULL)
420 wdog_software_attach = watchdog_attach;
425 * Each time the real-time timer fires, this function is called on all CPUs.
426 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only
427 * the other CPUs in the system need to call this function.
430 hardclock_cpu(int usermode)
432 struct pstats *pstats;
433 struct thread *td = curthread;
434 struct proc *p = td->td_proc;
438 * Run current process's virtual and profile time, as needed.
443 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
445 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
446 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
449 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
451 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
452 flags |= TDF_PROFPEND | TDF_ASTPENDING;
456 td->td_flags |= flags;
460 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
461 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
462 if (td->td_intr_frame != NULL)
463 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
465 callout_process(sbinuptime());
466 if (__predict_false(DPCPU_GET(epoch_cb_count)))
467 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
471 * The real-time timer, interrupting hz times per second.
474 hardclock(int usermode, uintfptr_t pc)
477 atomic_add_int(&ticks, 1);
478 hardclock_cpu(usermode);
480 cpu_tick_calibration();
482 * If no separate statistics clock is available, run it from here.
484 * XXX: this only works for UP
487 profclock(usermode, pc);
490 #ifdef DEVICE_POLLING
491 hardclock_device_poll(); /* this is very short and quick */
492 #endif /* DEVICE_POLLING */
493 if (watchdog_enabled > 0 && --watchdog_ticks <= 0)
498 hardclock_cnt(int cnt, int usermode)
500 struct pstats *pstats;
501 struct thread *td = curthread;
502 struct proc *p = td->td_proc;
503 int *t = DPCPU_PTR(pcputicks);
504 int flags, global, newticks;
508 * Update per-CPU and possibly global ticks values.
513 newticks = *t - global;
520 } while (!atomic_cmpset_int(&ticks, global, *t));
523 * Run current process's virtual and profile time, as needed.
528 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
530 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
532 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
535 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
537 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
539 flags |= TDF_PROFPEND | TDF_ASTPENDING;
544 td->td_flags |= flags;
549 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
550 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
551 if (td->td_intr_frame != NULL)
552 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
554 /* We are in charge to handle this tick duty. */
556 tc_ticktock(newticks);
557 #ifdef DEVICE_POLLING
558 /* Dangerous and no need to call these things concurrently. */
559 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
560 /* This is very short and quick. */
561 hardclock_device_poll();
562 atomic_store_rel_int(&devpoll_run, 0);
564 #endif /* DEVICE_POLLING */
565 if (watchdog_enabled > 0) {
566 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
567 if (i > 0 && i <= newticks)
571 if (curcpu == CPU_FIRST())
572 cpu_tick_calibration();
573 if (__predict_false(DPCPU_GET(epoch_cb_count)))
574 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
578 hardclock_sync(int cpu)
581 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
582 t = DPCPU_ID_PTR(cpu, pcputicks);
588 * Compute number of ticks in the specified amount of time.
591 tvtohz(struct timeval *tv)
597 * If the number of usecs in the whole seconds part of the time
598 * difference fits in a long, then the total number of usecs will
599 * fit in an unsigned long. Compute the total and convert it to
600 * ticks, rounding up and adding 1 to allow for the current tick
601 * to expire. Rounding also depends on unsigned long arithmetic
604 * Otherwise, if the number of ticks in the whole seconds part of
605 * the time difference fits in a long, then convert the parts to
606 * ticks separately and add, using similar rounding methods and
607 * overflow avoidance. This method would work in the previous
608 * case but it is slightly slower and assumes that hz is integral.
610 * Otherwise, round the time difference down to the maximum
611 * representable value.
613 * If ints have 32 bits, then the maximum value for any timeout in
614 * 10ms ticks is 248 days.
628 printf("tvotohz: negative time difference %ld sec %ld usec\n",
632 } else if (sec <= LONG_MAX / 1000000)
633 ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1;
634 else if (sec <= LONG_MAX / hz)
636 + howmany((unsigned long)usec, tick) + 1;
645 * Start profiling on a process.
647 * Kernel profiling passes proc0 which never exits and hence
648 * keeps the profile clock running constantly.
651 startprofclock(struct proc *p)
654 PROC_LOCK_ASSERT(p, MA_OWNED);
655 if (p->p_flag & P_STOPPROF)
657 if ((p->p_flag & P_PROFIL) == 0) {
658 p->p_flag |= P_PROFIL;
659 mtx_lock(&time_lock);
660 if (++profprocs == 1)
661 cpu_startprofclock();
662 mtx_unlock(&time_lock);
667 * Stop profiling on a process.
670 stopprofclock(struct proc *p)
673 PROC_LOCK_ASSERT(p, MA_OWNED);
674 if (p->p_flag & P_PROFIL) {
675 if (p->p_profthreads != 0) {
676 while (p->p_profthreads != 0) {
677 p->p_flag |= P_STOPPROF;
678 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
682 if ((p->p_flag & P_PROFIL) == 0)
684 p->p_flag &= ~P_PROFIL;
685 mtx_lock(&time_lock);
686 if (--profprocs == 0)
688 mtx_unlock(&time_lock);
693 * Statistics clock. Updates rusage information and calls the scheduler
694 * to adjust priorities of the active thread.
696 * This should be called by all active processors.
699 statclock(int usermode)
702 statclock_cnt(1, usermode);
706 statclock_cnt(int cnt, int usermode)
718 cp_time = (long *)PCPU_PTR(cp_time);
721 * Charge the time as appropriate.
723 td->td_uticks += cnt;
724 if (p->p_nice > NZERO)
725 cp_time[CP_NICE] += cnt;
727 cp_time[CP_USER] += cnt;
730 * Came from kernel mode, so we were:
731 * - handling an interrupt,
732 * - doing syscall or trap work on behalf of the current
734 * - spinning in the idle loop.
735 * Whichever it is, charge the time as appropriate.
736 * Note that we charge interrupts to the current process,
737 * regardless of whether they are ``for'' that process,
738 * so that we know how much of its real time was spent
739 * in ``non-process'' (i.e., interrupt) work.
741 if ((td->td_pflags & TDP_ITHREAD) ||
742 td->td_intr_nesting_level >= 2) {
743 td->td_iticks += cnt;
744 cp_time[CP_INTR] += cnt;
746 td->td_pticks += cnt;
747 td->td_sticks += cnt;
748 if (!TD_IS_IDLETHREAD(td))
749 cp_time[CP_SYS] += cnt;
751 cp_time[CP_IDLE] += cnt;
755 /* Update resource usage integrals and maximums. */
756 MPASS(p->p_vmspace != NULL);
759 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
760 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
761 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
762 rss = pgtok(vmspace_resident_count(vm));
763 if (ru->ru_maxrss < rss)
765 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
766 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
767 SDT_PROBE2(sched, , , tick, td, td->td_proc);
768 thread_lock_flags(td, MTX_QUIET);
769 for ( ; cnt > 0; cnt--)
773 if (td->td_intr_frame != NULL)
774 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
779 profclock(int usermode, uintfptr_t pc)
782 profclock_cnt(1, usermode, pc);
786 profclock_cnt(int cnt, int usermode, uintfptr_t pc)
797 * Came from user mode; CPU was in user state.
798 * If this process is being profiled, record the tick.
799 * if there is no related user location yet, don't
800 * bother trying to count it.
802 if (td->td_proc->p_flag & P_PROFIL)
803 addupc_intr(td, pc, cnt);
808 * Kernel statistics are just like addupc_intr, only easier.
811 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
813 if (i < g->textsize) {
820 if (td->td_intr_frame != NULL)
821 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
826 * Return information about system clocks.
829 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
831 struct clockinfo clkinfo;
833 * Construct clockinfo structure.
835 bzero(&clkinfo, sizeof(clkinfo));
838 clkinfo.profhz = profhz;
839 clkinfo.stathz = stathz ? stathz : hz;
840 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
843 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
844 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
845 0, 0, sysctl_kern_clockrate, "S,clockinfo",
846 "Rate and period of various kernel clocks");
849 watchdog_config(void *unused __unused, u_int cmd, int *error)
853 u = cmd & WD_INTERVAL;
854 if (u >= WD_TO_1SEC) {
855 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
856 watchdog_enabled = 1;
859 watchdog_enabled = 0;
864 * Handle a watchdog timeout by dumping interrupt information and
865 * then either dropping to DDB or panicking.
878 nintr = sintrcnt / sizeof(u_long);
880 printf("interrupt total\n");
881 while (--nintr >= 0) {
883 printf("%-12s %20lu\n", curname, *curintr);
884 curname += strlen(curname) + 1;
885 inttotal += *curintr++;
887 printf("Total %20ju\n", (uintmax_t)inttotal);
889 #if defined(KDB) && !defined(KDB_UNATTENDED)
891 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
893 panic("watchdog timeout");