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 /* Spin-lock protecting profiling statistics. */
94 static struct mtx time_lock;
96 SDT_PROVIDER_DECLARE(sched);
97 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
100 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
103 long cp_time[CPUSTATES];
106 unsigned int cp_time32[CPUSTATES];
109 read_cpu_time(cp_time);
111 if (req->flags & SCTL_MASK32) {
113 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
114 for (i = 0; i < CPUSTATES; i++)
115 cp_time32[i] = (unsigned int)cp_time[i];
116 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
121 return SYSCTL_OUT(req, 0, sizeof(cp_time));
122 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
127 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
128 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
130 static long empty[CPUSTATES];
133 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
140 unsigned int cp_time32[CPUSTATES];
146 if (req->flags & SCTL_MASK32)
147 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
150 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
152 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
153 if (!CPU_ABSENT(c)) {
155 cp_time = pcpu->pc_cp_time;
160 if (req->flags & SCTL_MASK32) {
161 for (i = 0; i < CPUSTATES; i++)
162 cp_time32[i] = (unsigned int)cp_time[i];
163 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
166 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
171 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
172 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
175 static const char *blessed[] = {
181 static int slptime_threshold = 1800;
182 static int blktime_threshold = 900;
183 static int sleepfreq = 3;
186 deadlres_td_on_lock(struct proc *p, struct thread *td, int blkticks)
190 sx_assert(&allproc_lock, SX_LOCKED);
191 PROC_LOCK_ASSERT(p, MA_OWNED);
192 THREAD_LOCK_ASSERT(td, MA_OWNED);
194 * The thread should be blocked on a turnstile, simply check
195 * if the turnstile channel is in good state.
197 MPASS(td->td_blocked != NULL);
199 tticks = ticks - td->td_blktick;
200 if (tticks > blkticks)
202 * Accordingly with provided thresholds, this thread is stuck
203 * for too long on a turnstile.
205 panic("%s: possible deadlock detected for %p (%s), "
206 "blocked for %d ticks\n", __func__,
207 td, sched_tdname(td), tticks);
211 deadlres_td_sleep_q(struct proc *p, struct thread *td, int slpticks)
214 int i, slptype, tticks;
216 sx_assert(&allproc_lock, SX_LOCKED);
217 PROC_LOCK_ASSERT(p, MA_OWNED);
218 THREAD_LOCK_ASSERT(td, MA_OWNED);
220 * Check if the thread is sleeping on a lock, otherwise skip the check.
221 * Drop the thread lock in order to avoid a LOR with the sleepqueue
224 wchan = td->td_wchan;
225 tticks = ticks - td->td_slptick;
226 slptype = sleepq_type(wchan);
227 if ((slptype == SLEEPQ_SX || slptype == SLEEPQ_LK) &&
230 * Accordingly with provided thresholds, this thread is stuck
231 * for too long on a sleepqueue.
232 * However, being on a sleepqueue, we might still check for the
235 for (i = 0; blessed[i] != NULL; i++)
236 if (!strcmp(blessed[i], td->td_wmesg))
239 panic("%s: possible deadlock detected for %p (%s), "
240 "blocked for %d ticks\n", __func__,
241 td, sched_tdname(td), tticks);
250 int blkticks, slpticks, tryl;
254 blkticks = blktime_threshold * hz;
255 slpticks = slptime_threshold * hz;
258 * Avoid to sleep on the sx_lock in order to avoid a
259 * possible priority inversion problem leading to
261 * If the lock can't be held after 100 tries, panic.
263 if (!sx_try_slock(&allproc_lock)) {
265 panic("%s: possible deadlock detected "
266 "on allproc_lock\n", __func__);
268 pause("allproc", sleepfreq * hz);
272 FOREACH_PROC_IN_SYSTEM(p) {
274 if (p->p_state == PRS_NEW) {
278 FOREACH_THREAD_IN_PROC(p, td) {
281 deadlres_td_on_lock(p, td,
283 else if (TD_IS_SLEEPING(td))
284 deadlres_td_sleep_q(p, td,
290 sx_sunlock(&allproc_lock);
292 /* Sleep for sleepfreq seconds. */
293 pause("-", sleepfreq * hz);
297 static struct kthread_desc deadlkres_kd = {
300 (struct thread **)NULL
303 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
305 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
306 "Deadlock resolver");
307 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RWTUN,
308 &slptime_threshold, 0,
309 "Number of seconds within is valid to sleep on a sleepqueue");
310 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RWTUN,
311 &blktime_threshold, 0,
312 "Number of seconds within is valid to block on a turnstile");
313 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RWTUN, &sleepfreq, 0,
314 "Number of seconds between any deadlock resolver thread run");
315 #endif /* DEADLKRES */
318 read_cpu_time(long *cp_time)
323 /* Sum up global cp_time[]. */
324 bzero(cp_time, sizeof(long) * CPUSTATES);
327 for (j = 0; j < CPUSTATES; j++)
328 cp_time[j] += pc->pc_cp_time[j];
332 #include <sys/watchdog.h>
334 static int watchdog_ticks;
335 static int watchdog_enabled;
336 static void watchdog_fire(void);
337 static void watchdog_config(void *, u_int, int *);
340 watchdog_attach(void)
342 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
346 * Clock handling routines.
348 * This code is written to operate with two timers that run independently of
351 * The main timer, running hz times per second, is used to trigger interval
352 * timers, timeouts and rescheduling as needed.
354 * The second timer handles kernel and user profiling,
355 * and does resource use estimation. If the second timer is programmable,
356 * it is randomized to avoid aliasing between the two clocks. For example,
357 * the randomization prevents an adversary from always giving up the cpu
358 * just before its quantum expires. Otherwise, it would never accumulate
359 * cpu ticks. The mean frequency of the second timer is stathz.
361 * If no second timer exists, stathz will be zero; in this case we drive
362 * profiling and statistics off the main clock. This WILL NOT be accurate;
363 * do not do it unless absolutely necessary.
365 * The statistics clock may (or may not) be run at a higher rate while
366 * profiling. This profile clock runs at profhz. We require that profhz
367 * be an integral multiple of stathz.
369 * If the statistics clock is running fast, it must be divided by the ratio
370 * profhz/stathz for statistics. (For profiling, every tick counts.)
372 * Time-of-day is maintained using a "timecounter", which may or may
373 * not be related to the hardware generating the above mentioned
383 DPCPU_DEFINE_STATIC(int, pcputicks); /* Per-CPU version of ticks. */
384 #ifdef DEVICE_POLLING
385 static int devpoll_run = 0;
389 * Initialize clock frequencies and start both clocks running.
392 initclocks(void *dummy __unused)
397 * Set divisors to 1 (normal case) and let the machine-specific
400 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
404 * Compute profhz/stathz, and fix profhz if needed.
406 i = stathz ? stathz : hz;
409 psratio = profhz / i;
412 /* Enable hardclock watchdog now, even if a hardware watchdog exists. */
415 /* Volunteer to run a software watchdog. */
416 if (wdog_software_attach == NULL)
417 wdog_software_attach = watchdog_attach;
420 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
422 static __noinline void
423 hardclock_itimer(struct thread *td, struct pstats *pstats, int cnt, int usermode)
431 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
433 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
435 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
438 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
440 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
442 flags |= TDF_PROFPEND | TDF_ASTPENDING;
447 td->td_flags |= flags;
453 hardclock(int cnt, int usermode)
455 struct pstats *pstats;
456 struct thread *td = curthread;
457 struct proc *p = td->td_proc;
458 int *t = DPCPU_PTR(pcputicks);
459 int global, i, newticks;
462 * Update per-CPU and possibly global ticks values.
467 newticks = *t - global;
474 } while (!atomic_fcmpset_int(&ticks, &global, *t));
477 * Run current process's virtual and profile time, as needed.
481 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) ||
482 timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)))
483 hardclock_itimer(td, pstats, cnt, usermode);
486 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
487 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
488 if (td->td_intr_frame != NULL)
489 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
491 /* We are in charge to handle this tick duty. */
493 tc_ticktock(newticks);
494 #ifdef DEVICE_POLLING
495 /* Dangerous and no need to call these things concurrently. */
496 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
497 /* This is very short and quick. */
498 hardclock_device_poll();
499 atomic_store_rel_int(&devpoll_run, 0);
501 #endif /* DEVICE_POLLING */
502 if (watchdog_enabled > 0) {
503 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
504 if (i > 0 && i <= newticks)
507 intr_event_handle(clk_intr_event, NULL);
509 if (curcpu == CPU_FIRST())
510 cpu_tick_calibration();
511 if (__predict_false(DPCPU_GET(epoch_cb_count)))
512 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
516 hardclock_sync(int cpu)
519 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
520 t = DPCPU_ID_PTR(cpu, pcputicks);
526 * Regular integer scaling formula without losing precision:
528 #define TIME_INT_SCALE(value, mul, div) \
529 (((value) / (div)) * (mul) + (((value) % (div)) * (mul)) / (div))
532 * Macro for converting seconds and microseconds into actual ticks,
533 * based on the given hz value:
535 #define TIME_TO_TICKS(sec, usec, hz) \
536 ((sec) * (hz) + TIME_INT_SCALE(usec, hz, 1 << 6) / (1000000 >> 6))
538 #define TIME_ASSERT_VALID_HZ(hz) \
539 _Static_assert(TIME_TO_TICKS(INT_MAX / (hz) - 1, 999999, hz) >= 0 && \
540 TIME_TO_TICKS(INT_MAX / (hz) - 1, 999999, hz) < INT_MAX, \
541 "tvtohz() can overflow the regular integer type")
544 * Compile time assert the maximum and minimum values to fit into a
545 * regular integer when computing TIME_TO_TICKS():
547 TIME_ASSERT_VALID_HZ(HZ_MAXIMUM);
548 TIME_ASSERT_VALID_HZ(HZ_MINIMUM);
551 * The formula is mostly linear, but test some more common values just
554 TIME_ASSERT_VALID_HZ(1024);
555 TIME_ASSERT_VALID_HZ(1000);
556 TIME_ASSERT_VALID_HZ(128);
557 TIME_ASSERT_VALID_HZ(100);
560 * Compute number of ticks representing the specified amount of time.
561 * If the specified time is negative, a value of 1 is returned. This
562 * function returns a value from 1 up to and including INT_MAX.
565 tvtohz(struct timeval *tv)
570 * The values passed here may come from user-space and these
571 * checks ensure "tv_usec" is within its allowed range:
574 /* check for tv_usec underflow */
575 if (__predict_false(tv->tv_usec < 0)) {
576 tv->tv_sec += tv->tv_usec / 1000000;
577 tv->tv_usec = tv->tv_usec % 1000000;
578 /* convert tv_usec to a positive value */
579 if (__predict_true(tv->tv_usec < 0)) {
580 tv->tv_usec += 1000000;
583 /* check for tv_usec overflow */
584 } else if (__predict_false(tv->tv_usec >= 1000000)) {
585 tv->tv_sec += tv->tv_usec / 1000000;
586 tv->tv_usec = tv->tv_usec % 1000000;
589 /* check for tv_sec underflow */
590 if (__predict_false(tv->tv_sec < 0))
592 /* check for tv_sec overflow (including room for the tv_usec part) */
593 else if (__predict_false(tv->tv_sec >= tick_seconds_max))
596 /* cast to "int" to avoid platform differences */
597 retval = TIME_TO_TICKS((int)tv->tv_sec, (int)tv->tv_usec, hz);
599 /* add one additional tick */
604 * Start profiling on a process.
606 * Kernel profiling passes proc0 which never exits and hence
607 * keeps the profile clock running constantly.
610 startprofclock(struct proc *p)
613 PROC_LOCK_ASSERT(p, MA_OWNED);
614 if (p->p_flag & P_STOPPROF)
616 if ((p->p_flag & P_PROFIL) == 0) {
617 p->p_flag |= P_PROFIL;
618 mtx_lock(&time_lock);
619 if (++profprocs == 1)
620 cpu_startprofclock();
621 mtx_unlock(&time_lock);
626 * Stop profiling on a process.
629 stopprofclock(struct proc *p)
632 PROC_LOCK_ASSERT(p, MA_OWNED);
633 if (p->p_flag & P_PROFIL) {
634 if (p->p_profthreads != 0) {
635 while (p->p_profthreads != 0) {
636 p->p_flag |= P_STOPPROF;
637 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
641 if ((p->p_flag & P_PROFIL) == 0)
643 p->p_flag &= ~P_PROFIL;
644 mtx_lock(&time_lock);
645 if (--profprocs == 0)
647 mtx_unlock(&time_lock);
652 * Statistics clock. Updates rusage information and calls the scheduler
653 * to adjust priorities of the active thread.
655 * This should be called by all active processors.
658 statclock(int cnt, int usermode)
666 uint64_t runtime, new_switchtime;
671 cp_time = (long *)PCPU_PTR(cp_time);
674 * Charge the time as appropriate.
676 td->td_uticks += cnt;
677 if (p->p_nice > NZERO)
678 cp_time[CP_NICE] += cnt;
680 cp_time[CP_USER] += cnt;
683 * Came from kernel mode, so we were:
684 * - handling an interrupt,
685 * - doing syscall or trap work on behalf of the current
687 * - spinning in the idle loop.
688 * Whichever it is, charge the time as appropriate.
689 * Note that we charge interrupts to the current process,
690 * regardless of whether they are ``for'' that process,
691 * so that we know how much of its real time was spent
692 * in ``non-process'' (i.e., interrupt) work.
694 if ((td->td_pflags & TDP_ITHREAD) ||
695 td->td_intr_nesting_level >= 2) {
696 td->td_iticks += cnt;
697 cp_time[CP_INTR] += cnt;
699 td->td_pticks += cnt;
700 td->td_sticks += cnt;
701 if (!TD_IS_IDLETHREAD(td))
702 cp_time[CP_SYS] += cnt;
704 cp_time[CP_IDLE] += cnt;
708 /* Update resource usage integrals and maximums. */
709 MPASS(p->p_vmspace != NULL);
712 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
713 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
714 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
715 rss = pgtok(vmspace_resident_count(vm));
716 if (ru->ru_maxrss < rss)
718 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
719 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
720 SDT_PROBE2(sched, , , tick, td, td->td_proc);
721 thread_lock_flags(td, MTX_QUIET);
724 * Compute the amount of time during which the current
725 * thread was running, and add that to its total so far.
727 new_switchtime = cpu_ticks();
728 runtime = new_switchtime - PCPU_GET(switchtime);
729 td->td_runtime += runtime;
730 td->td_incruntime += runtime;
731 PCPU_SET(switchtime, new_switchtime);
733 sched_clock(td, cnt);
736 if (td->td_intr_frame != NULL)
737 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
742 profclock(int cnt, int usermode, uintfptr_t pc)
753 * Came from user mode; CPU was in user state.
754 * If this process is being profiled, record the tick.
755 * if there is no related user location yet, don't
756 * bother trying to count it.
758 if (td->td_proc->p_flag & P_PROFIL)
759 addupc_intr(td, pc, cnt);
764 * Kernel statistics are just like addupc_intr, only easier.
767 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
769 if (i < g->textsize) {
776 if (td->td_intr_frame != NULL)
777 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
782 * Return information about system clocks.
785 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
787 struct clockinfo clkinfo;
789 * Construct clockinfo structure.
791 bzero(&clkinfo, sizeof(clkinfo));
794 clkinfo.profhz = profhz;
795 clkinfo.stathz = stathz ? stathz : hz;
796 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
799 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
800 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
801 0, 0, sysctl_kern_clockrate, "S,clockinfo",
802 "Rate and period of various kernel clocks");
805 watchdog_config(void *unused __unused, u_int cmd, int *error)
809 u = cmd & WD_INTERVAL;
810 if (u >= WD_TO_1SEC) {
811 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
812 watchdog_enabled = 1;
815 watchdog_enabled = 0;
820 * Handle a watchdog timeout by dumping interrupt information and
821 * then either dropping to DDB or panicking.
834 nintr = sintrcnt / sizeof(u_long);
836 printf("interrupt total\n");
837 while (--nintr >= 0) {
839 printf("%-12s %20lu\n", curname, *curintr);
840 curname += strlen(curname) + 1;
841 inttotal += *curintr++;
843 printf("Total %20ju\n", (uintmax_t)inttotal);
845 #if defined(KDB) && !defined(KDB_UNATTENDED)
847 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
849 panic("watchdog timeout");