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 hardclock(int cnt, int usermode)
427 struct pstats *pstats;
428 struct thread *td = curthread;
429 struct proc *p = td->td_proc;
430 int *t = DPCPU_PTR(pcputicks);
431 int flags, global, newticks;
435 * Update per-CPU and possibly global ticks values.
440 newticks = *t - global;
447 } while (!atomic_cmpset_int(&ticks, global, *t));
450 * Run current process's virtual and profile time, as needed.
455 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
457 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
459 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
462 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
464 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
466 flags |= TDF_PROFPEND | TDF_ASTPENDING;
471 td->td_flags |= flags;
476 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
477 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
478 if (td->td_intr_frame != NULL)
479 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
481 /* We are in charge to handle this tick duty. */
483 tc_ticktock(newticks);
484 #ifdef DEVICE_POLLING
485 /* Dangerous and no need to call these things concurrently. */
486 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
487 /* This is very short and quick. */
488 hardclock_device_poll();
489 atomic_store_rel_int(&devpoll_run, 0);
491 #endif /* DEVICE_POLLING */
492 if (watchdog_enabled > 0) {
493 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
494 if (i > 0 && i <= newticks)
498 if (curcpu == CPU_FIRST())
499 cpu_tick_calibration();
500 if (__predict_false(DPCPU_GET(epoch_cb_count)))
501 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
505 hardclock_sync(int cpu)
508 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
509 t = DPCPU_ID_PTR(cpu, pcputicks);
515 * Compute number of ticks in the specified amount of time.
518 tvtohz(struct timeval *tv)
524 * If the number of usecs in the whole seconds part of the time
525 * difference fits in a long, then the total number of usecs will
526 * fit in an unsigned long. Compute the total and convert it to
527 * ticks, rounding up and adding 1 to allow for the current tick
528 * to expire. Rounding also depends on unsigned long arithmetic
531 * Otherwise, if the number of ticks in the whole seconds part of
532 * the time difference fits in a long, then convert the parts to
533 * ticks separately and add, using similar rounding methods and
534 * overflow avoidance. This method would work in the previous
535 * case but it is slightly slower and assumes that hz is integral.
537 * Otherwise, round the time difference down to the maximum
538 * representable value.
540 * If ints have 32 bits, then the maximum value for any timeout in
541 * 10ms ticks is 248 days.
555 printf("tvotohz: negative time difference %ld sec %ld usec\n",
559 } else if (sec <= LONG_MAX / 1000000)
560 ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1;
561 else if (sec <= LONG_MAX / hz)
563 + howmany((unsigned long)usec, tick) + 1;
572 * Start profiling on a process.
574 * Kernel profiling passes proc0 which never exits and hence
575 * keeps the profile clock running constantly.
578 startprofclock(struct proc *p)
581 PROC_LOCK_ASSERT(p, MA_OWNED);
582 if (p->p_flag & P_STOPPROF)
584 if ((p->p_flag & P_PROFIL) == 0) {
585 p->p_flag |= P_PROFIL;
586 mtx_lock(&time_lock);
587 if (++profprocs == 1)
588 cpu_startprofclock();
589 mtx_unlock(&time_lock);
594 * Stop profiling on a process.
597 stopprofclock(struct proc *p)
600 PROC_LOCK_ASSERT(p, MA_OWNED);
601 if (p->p_flag & P_PROFIL) {
602 if (p->p_profthreads != 0) {
603 while (p->p_profthreads != 0) {
604 p->p_flag |= P_STOPPROF;
605 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
609 if ((p->p_flag & P_PROFIL) == 0)
611 p->p_flag &= ~P_PROFIL;
612 mtx_lock(&time_lock);
613 if (--profprocs == 0)
615 mtx_unlock(&time_lock);
620 * Statistics clock. Updates rusage information and calls the scheduler
621 * to adjust priorities of the active thread.
623 * This should be called by all active processors.
626 statclock(int cnt, int usermode)
634 uint64_t runtime, new_switchtime;
639 cp_time = (long *)PCPU_PTR(cp_time);
642 * Charge the time as appropriate.
644 td->td_uticks += cnt;
645 if (p->p_nice > NZERO)
646 cp_time[CP_NICE] += cnt;
648 cp_time[CP_USER] += cnt;
651 * Came from kernel mode, so we were:
652 * - handling an interrupt,
653 * - doing syscall or trap work on behalf of the current
655 * - spinning in the idle loop.
656 * Whichever it is, charge the time as appropriate.
657 * Note that we charge interrupts to the current process,
658 * regardless of whether they are ``for'' that process,
659 * so that we know how much of its real time was spent
660 * in ``non-process'' (i.e., interrupt) work.
662 if ((td->td_pflags & TDP_ITHREAD) ||
663 td->td_intr_nesting_level >= 2) {
664 td->td_iticks += cnt;
665 cp_time[CP_INTR] += cnt;
667 td->td_pticks += cnt;
668 td->td_sticks += cnt;
669 if (!TD_IS_IDLETHREAD(td))
670 cp_time[CP_SYS] += cnt;
672 cp_time[CP_IDLE] += cnt;
676 /* Update resource usage integrals and maximums. */
677 MPASS(p->p_vmspace != NULL);
680 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
681 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
682 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
683 rss = pgtok(vmspace_resident_count(vm));
684 if (ru->ru_maxrss < rss)
686 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
687 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
688 SDT_PROBE2(sched, , , tick, td, td->td_proc);
689 thread_lock_flags(td, MTX_QUIET);
692 * Compute the amount of time during which the current
693 * thread was running, and add that to its total so far.
695 new_switchtime = cpu_ticks();
696 runtime = new_switchtime - PCPU_GET(switchtime);
697 td->td_runtime += runtime;
698 td->td_incruntime += runtime;
699 PCPU_SET(switchtime, new_switchtime);
701 for ( ; cnt > 0; cnt--)
705 if (td->td_intr_frame != NULL)
706 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
711 profclock(int cnt, int usermode, uintfptr_t pc)
722 * Came from user mode; CPU was in user state.
723 * If this process is being profiled, record the tick.
724 * if there is no related user location yet, don't
725 * bother trying to count it.
727 if (td->td_proc->p_flag & P_PROFIL)
728 addupc_intr(td, pc, cnt);
733 * Kernel statistics are just like addupc_intr, only easier.
736 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
738 if (i < g->textsize) {
745 if (td->td_intr_frame != NULL)
746 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
751 * Return information about system clocks.
754 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
756 struct clockinfo clkinfo;
758 * Construct clockinfo structure.
760 bzero(&clkinfo, sizeof(clkinfo));
763 clkinfo.profhz = profhz;
764 clkinfo.stathz = stathz ? stathz : hz;
765 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
768 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
769 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
770 0, 0, sysctl_kern_clockrate, "S,clockinfo",
771 "Rate and period of various kernel clocks");
774 watchdog_config(void *unused __unused, u_int cmd, int *error)
778 u = cmd & WD_INTERVAL;
779 if (u >= WD_TO_1SEC) {
780 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
781 watchdog_enabled = 1;
784 watchdog_enabled = 0;
789 * Handle a watchdog timeout by dumping interrupt information and
790 * then either dropping to DDB or panicking.
803 nintr = sintrcnt / sizeof(u_long);
805 printf("interrupt total\n");
806 while (--nintr >= 0) {
808 printf("%-12s %20lu\n", curname, *curintr);
809 curname += strlen(curname) + 1;
810 inttotal += *curintr++;
812 printf("Total %20ju\n", (uintmax_t)inttotal);
814 #if defined(KDB) && !defined(KDB_UNATTENDED)
816 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
818 panic("watchdog timeout");