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/eventhandler.h>
53 #include <sys/gtaskqueue.h>
55 #include <sys/kernel.h>
56 #include <sys/kthread.h>
59 #include <sys/mutex.h>
61 #include <sys/resource.h>
62 #include <sys/resourcevar.h>
63 #include <sys/sched.h>
65 #include <sys/signalvar.h>
66 #include <sys/sleepqueue.h>
70 #include <vm/vm_map.h>
71 #include <sys/sysctl.h>
73 #include <sys/interrupt.h>
74 #include <sys/limits.h>
75 #include <sys/timetc.h>
82 #include <sys/pmckern.h>
83 PMC_SOFT_DEFINE( , , clock, hard);
84 PMC_SOFT_DEFINE( , , clock, stat);
85 PMC_SOFT_DEFINE_EX( , , clock, prof, \
86 cpu_startprofclock, cpu_stopprofclock);
90 extern void hardclock_device_poll(void);
91 #endif /* DEVICE_POLLING */
93 static void initclocks(void *dummy);
94 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
96 /* Spin-lock protecting profiling statistics. */
97 static struct mtx time_lock;
99 SDT_PROVIDER_DECLARE(sched);
100 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
103 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
106 long cp_time[CPUSTATES];
109 unsigned int cp_time32[CPUSTATES];
112 read_cpu_time(cp_time);
114 if (req->flags & SCTL_MASK32) {
116 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
117 for (i = 0; i < CPUSTATES; i++)
118 cp_time32[i] = (unsigned int)cp_time[i];
119 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
124 return SYSCTL_OUT(req, 0, sizeof(cp_time));
125 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
130 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
131 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
133 static long empty[CPUSTATES];
136 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
143 unsigned int cp_time32[CPUSTATES];
149 if (req->flags & SCTL_MASK32)
150 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
153 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
155 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
156 if (!CPU_ABSENT(c)) {
158 cp_time = pcpu->pc_cp_time;
163 if (req->flags & SCTL_MASK32) {
164 for (i = 0; i < CPUSTATES; i++)
165 cp_time32[i] = (unsigned int)cp_time[i];
166 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
169 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
174 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
175 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
178 static const char *blessed[] = {
184 static int slptime_threshold = 1800;
185 static int blktime_threshold = 900;
186 static int sleepfreq = 3;
189 deadlres_td_on_lock(struct proc *p, struct thread *td, int blkticks)
193 sx_assert(&allproc_lock, SX_LOCKED);
194 PROC_LOCK_ASSERT(p, MA_OWNED);
195 THREAD_LOCK_ASSERT(td, MA_OWNED);
197 * The thread should be blocked on a turnstile, simply check
198 * if the turnstile channel is in good state.
200 MPASS(td->td_blocked != NULL);
202 tticks = ticks - td->td_blktick;
203 if (tticks > blkticks)
205 * Accordingly with provided thresholds, this thread is stuck
206 * for too long on a turnstile.
208 panic("%s: possible deadlock detected for %p, "
209 "blocked for %d ticks\n", __func__, td, tticks);
213 deadlres_td_sleep_q(struct proc *p, struct thread *td, int slpticks)
216 int i, slptype, tticks;
218 sx_assert(&allproc_lock, SX_LOCKED);
219 PROC_LOCK_ASSERT(p, MA_OWNED);
220 THREAD_LOCK_ASSERT(td, MA_OWNED);
222 * Check if the thread is sleeping on a lock, otherwise skip the check.
223 * Drop the thread lock in order to avoid a LOR with the sleepqueue
226 wchan = td->td_wchan;
227 tticks = ticks - td->td_slptick;
228 slptype = sleepq_type(wchan);
229 if ((slptype == SLEEPQ_SX || slptype == SLEEPQ_LK) &&
233 * Accordingly with provided thresholds, this thread is stuck
234 * for too long on a sleepqueue.
235 * However, being on a sleepqueue, we might still check for the
238 for (i = 0; blessed[i] != NULL; i++)
239 if (!strcmp(blessed[i], td->td_wmesg))
242 panic("%s: possible deadlock detected for %p, "
243 "blocked for %d ticks\n", __func__, td, tticks);
252 int blkticks, slpticks, tryl;
256 blkticks = blktime_threshold * hz;
257 slpticks = slptime_threshold * hz;
260 * Avoid to sleep on the sx_lock in order to avoid a
261 * possible priority inversion problem leading to
263 * If the lock can't be held after 100 tries, panic.
265 if (!sx_try_slock(&allproc_lock)) {
267 panic("%s: possible deadlock detected "
268 "on allproc_lock\n", __func__);
270 pause("allproc", sleepfreq * hz);
274 FOREACH_PROC_IN_SYSTEM(p) {
276 if (p->p_state == PRS_NEW) {
280 FOREACH_THREAD_IN_PROC(p, td) {
283 deadlres_td_on_lock(p, td,
285 else if (TD_IS_SLEEPING(td) &&
287 deadlres_td_sleep_q(p, td,
293 sx_sunlock(&allproc_lock);
295 /* Sleep for sleepfreq seconds. */
296 pause("-", sleepfreq * hz);
300 static struct kthread_desc deadlkres_kd = {
303 (struct thread **)NULL
306 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
308 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0,
309 "Deadlock resolver");
310 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
311 &slptime_threshold, 0,
312 "Number of seconds within is valid to sleep on a sleepqueue");
313 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
314 &blktime_threshold, 0,
315 "Number of seconds within is valid to block on a turnstile");
316 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
317 "Number of seconds between any deadlock resolver thread run");
318 #endif /* DEADLKRES */
321 read_cpu_time(long *cp_time)
326 /* Sum up global cp_time[]. */
327 bzero(cp_time, sizeof(long) * CPUSTATES);
330 for (j = 0; j < CPUSTATES; j++)
331 cp_time[j] += pc->pc_cp_time[j];
335 #include <sys/watchdog.h>
337 static int watchdog_ticks;
338 static int watchdog_enabled;
339 static void watchdog_fire(void);
340 static void watchdog_config(void *, u_int, int *);
343 watchdog_attach(void)
345 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
349 * Clock handling routines.
351 * This code is written to operate with two timers that run independently of
354 * The main timer, running hz times per second, is used to trigger interval
355 * timers, timeouts and rescheduling as needed.
357 * The second timer handles kernel and user profiling,
358 * and does resource use estimation. If the second timer is programmable,
359 * it is randomized to avoid aliasing between the two clocks. For example,
360 * the randomization prevents an adversary from always giving up the cpu
361 * just before its quantum expires. Otherwise, it would never accumulate
362 * cpu ticks. The mean frequency of the second timer is stathz.
364 * If no second timer exists, stathz will be zero; in this case we drive
365 * profiling and statistics off the main clock. This WILL NOT be accurate;
366 * do not do it unless absolutely necessary.
368 * The statistics clock may (or may not) be run at a higher rate while
369 * profiling. This profile clock runs at profhz. We require that profhz
370 * be an integral multiple of stathz.
372 * If the statistics clock is running fast, it must be divided by the ratio
373 * profhz/stathz for statistics. (For profiling, every tick counts.)
375 * Time-of-day is maintained using a "timecounter", which may or may
376 * not be related to the hardware generating the above mentioned
386 DPCPU_DEFINE_STATIC(int, pcputicks); /* Per-CPU version of ticks. */
387 #ifdef DEVICE_POLLING
388 static int devpoll_run = 0;
392 * Initialize clock frequencies and start both clocks running.
396 initclocks(void *dummy)
401 * Set divisors to 1 (normal case) and let the machine-specific
404 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
408 * Compute profhz/stathz, and fix profhz if needed.
410 i = stathz ? stathz : hz;
413 psratio = profhz / i;
416 /* Enable hardclock watchdog now, even if a hardware watchdog exists. */
419 /* Volunteer to run a software watchdog. */
420 if (wdog_software_attach == NULL)
421 wdog_software_attach = watchdog_attach;
425 static __noinline void
426 hardclock_itimer(struct thread *td, struct pstats *pstats, int cnt, int usermode)
434 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
436 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
438 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
441 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
443 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
445 flags |= TDF_PROFPEND | TDF_ASTPENDING;
450 td->td_flags |= flags;
456 hardclock(int cnt, int usermode)
458 struct pstats *pstats;
459 struct thread *td = curthread;
460 struct proc *p = td->td_proc;
461 int *t = DPCPU_PTR(pcputicks);
462 int global, i, newticks;
465 * Update per-CPU and possibly global ticks values.
470 newticks = *t - global;
477 } while (!atomic_fcmpset_int(&ticks, &global, *t));
480 * Run current process's virtual and profile time, as needed.
484 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) ||
485 timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)))
486 hardclock_itimer(td, pstats, cnt, usermode);
489 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
490 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
491 if (td->td_intr_frame != NULL)
492 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
494 /* We are in charge to handle this tick duty. */
496 tc_ticktock(newticks);
497 #ifdef DEVICE_POLLING
498 /* Dangerous and no need to call these things concurrently. */
499 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
500 /* This is very short and quick. */
501 hardclock_device_poll();
502 atomic_store_rel_int(&devpoll_run, 0);
504 #endif /* DEVICE_POLLING */
505 if (watchdog_enabled > 0) {
506 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
507 if (i > 0 && i <= newticks)
511 if (curcpu == CPU_FIRST())
512 cpu_tick_calibration();
513 if (__predict_false(DPCPU_GET(epoch_cb_count)))
514 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
518 hardclock_sync(int cpu)
521 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
522 t = DPCPU_ID_PTR(cpu, pcputicks);
528 * Compute number of ticks in the specified amount of time.
531 tvtohz(struct timeval *tv)
537 * If the number of usecs in the whole seconds part of the time
538 * difference fits in a long, then the total number of usecs will
539 * fit in an unsigned long. Compute the total and convert it to
540 * ticks, rounding up and adding 1 to allow for the current tick
541 * to expire. Rounding also depends on unsigned long arithmetic
544 * Otherwise, if the number of ticks in the whole seconds part of
545 * the time difference fits in a long, then convert the parts to
546 * ticks separately and add, using similar rounding methods and
547 * overflow avoidance. This method would work in the previous
548 * case but it is slightly slower and assumes that hz is integral.
550 * Otherwise, round the time difference down to the maximum
551 * representable value.
553 * If ints have 32 bits, then the maximum value for any timeout in
554 * 10ms ticks is 248 days.
568 printf("tvotohz: negative time difference %ld sec %ld usec\n",
572 } else if (sec <= LONG_MAX / 1000000)
573 ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1;
574 else if (sec <= LONG_MAX / hz)
576 + howmany((unsigned long)usec, tick) + 1;
585 * Start profiling on a process.
587 * Kernel profiling passes proc0 which never exits and hence
588 * keeps the profile clock running constantly.
591 startprofclock(struct proc *p)
594 PROC_LOCK_ASSERT(p, MA_OWNED);
595 if (p->p_flag & P_STOPPROF)
597 if ((p->p_flag & P_PROFIL) == 0) {
598 p->p_flag |= P_PROFIL;
599 mtx_lock(&time_lock);
600 if (++profprocs == 1)
601 cpu_startprofclock();
602 mtx_unlock(&time_lock);
607 * Stop profiling on a process.
610 stopprofclock(struct proc *p)
613 PROC_LOCK_ASSERT(p, MA_OWNED);
614 if (p->p_flag & P_PROFIL) {
615 if (p->p_profthreads != 0) {
616 while (p->p_profthreads != 0) {
617 p->p_flag |= P_STOPPROF;
618 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
622 if ((p->p_flag & P_PROFIL) == 0)
624 p->p_flag &= ~P_PROFIL;
625 mtx_lock(&time_lock);
626 if (--profprocs == 0)
628 mtx_unlock(&time_lock);
633 * Statistics clock. Updates rusage information and calls the scheduler
634 * to adjust priorities of the active thread.
636 * This should be called by all active processors.
639 statclock(int cnt, int usermode)
647 uint64_t runtime, new_switchtime;
652 cp_time = (long *)PCPU_PTR(cp_time);
655 * Charge the time as appropriate.
657 td->td_uticks += cnt;
658 if (p->p_nice > NZERO)
659 cp_time[CP_NICE] += cnt;
661 cp_time[CP_USER] += cnt;
664 * Came from kernel mode, so we were:
665 * - handling an interrupt,
666 * - doing syscall or trap work on behalf of the current
668 * - spinning in the idle loop.
669 * Whichever it is, charge the time as appropriate.
670 * Note that we charge interrupts to the current process,
671 * regardless of whether they are ``for'' that process,
672 * so that we know how much of its real time was spent
673 * in ``non-process'' (i.e., interrupt) work.
675 if ((td->td_pflags & TDP_ITHREAD) ||
676 td->td_intr_nesting_level >= 2) {
677 td->td_iticks += cnt;
678 cp_time[CP_INTR] += cnt;
680 td->td_pticks += cnt;
681 td->td_sticks += cnt;
682 if (!TD_IS_IDLETHREAD(td))
683 cp_time[CP_SYS] += cnt;
685 cp_time[CP_IDLE] += cnt;
689 /* Update resource usage integrals and maximums. */
690 MPASS(p->p_vmspace != NULL);
693 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
694 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
695 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
696 rss = pgtok(vmspace_resident_count(vm));
697 if (ru->ru_maxrss < rss)
699 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
700 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
701 SDT_PROBE2(sched, , , tick, td, td->td_proc);
702 thread_lock_flags(td, MTX_QUIET);
705 * Compute the amount of time during which the current
706 * thread was running, and add that to its total so far.
708 new_switchtime = cpu_ticks();
709 runtime = new_switchtime - PCPU_GET(switchtime);
710 td->td_runtime += runtime;
711 td->td_incruntime += runtime;
712 PCPU_SET(switchtime, new_switchtime);
714 for ( ; cnt > 0; cnt--)
718 if (td->td_intr_frame != NULL)
719 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
724 profclock(int cnt, int usermode, uintfptr_t pc)
735 * Came from user mode; CPU was in user state.
736 * If this process is being profiled, record the tick.
737 * if there is no related user location yet, don't
738 * bother trying to count it.
740 if (td->td_proc->p_flag & P_PROFIL)
741 addupc_intr(td, pc, cnt);
746 * Kernel statistics are just like addupc_intr, only easier.
749 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
751 if (i < g->textsize) {
758 if (td->td_intr_frame != NULL)
759 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
764 * Return information about system clocks.
767 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
769 struct clockinfo clkinfo;
771 * Construct clockinfo structure.
773 bzero(&clkinfo, sizeof(clkinfo));
776 clkinfo.profhz = profhz;
777 clkinfo.stathz = stathz ? stathz : hz;
778 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
781 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
782 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
783 0, 0, sysctl_kern_clockrate, "S,clockinfo",
784 "Rate and period of various kernel clocks");
787 watchdog_config(void *unused __unused, u_int cmd, int *error)
791 u = cmd & WD_INTERVAL;
792 if (u >= WD_TO_1SEC) {
793 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
794 watchdog_enabled = 1;
797 watchdog_enabled = 0;
802 * Handle a watchdog timeout by dumping interrupt information and
803 * then either dropping to DDB or panicking.
816 nintr = sintrcnt / sizeof(u_long);
818 printf("interrupt total\n");
819 while (--nintr >= 0) {
821 printf("%-12s %20lu\n", curname, *curintr);
822 curname += strlen(curname) + 1;
823 inttotal += *curintr++;
825 printf("Total %20ju\n", (uintmax_t)inttotal);
827 #if defined(KDB) && !defined(KDB_UNATTENDED)
829 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
831 panic("watchdog timeout");