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>
41 #include "opt_device_polling.h"
42 #include "opt_hwpmc_hooks.h"
44 #include "opt_watchdog.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/callout.h>
49 #include <sys/epoch.h>
50 #include <sys/eventhandler.h>
51 #include <sys/gtaskqueue.h>
53 #include <sys/kernel.h>
54 #include <sys/kthread.h>
57 #include <sys/mutex.h>
59 #include <sys/resource.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sched.h>
63 #include <sys/signalvar.h>
64 #include <sys/sleepqueue.h>
68 #include <vm/vm_map.h>
69 #include <sys/sysctl.h>
71 #include <sys/interrupt.h>
72 #include <sys/limits.h>
73 #include <sys/timetc.h>
76 #include <sys/pmckern.h>
77 PMC_SOFT_DEFINE( , , clock, hard);
78 PMC_SOFT_DEFINE( , , clock, stat);
79 PMC_SOFT_DEFINE_EX( , , clock, prof, \
80 cpu_startprofclock, cpu_stopprofclock);
84 extern void hardclock_device_poll(void);
85 #endif /* DEVICE_POLLING */
87 /* Spin-lock protecting profiling statistics. */
88 static struct mtx time_lock;
90 SDT_PROVIDER_DECLARE(sched);
91 SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *");
94 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
97 long cp_time[CPUSTATES];
100 unsigned int cp_time32[CPUSTATES];
103 read_cpu_time(cp_time);
105 if (req->flags & SCTL_MASK32) {
107 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
108 for (i = 0; i < CPUSTATES; i++)
109 cp_time32[i] = (unsigned int)cp_time[i];
110 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
115 return SYSCTL_OUT(req, 0, sizeof(cp_time));
116 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
121 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
122 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
124 static long empty[CPUSTATES];
127 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
134 unsigned int cp_time32[CPUSTATES];
140 if (req->flags & SCTL_MASK32)
141 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
144 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
146 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
147 if (!CPU_ABSENT(c)) {
149 cp_time = pcpu->pc_cp_time;
154 if (req->flags & SCTL_MASK32) {
155 for (i = 0; i < CPUSTATES; i++)
156 cp_time32[i] = (unsigned int)cp_time[i];
157 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
160 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
165 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
166 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
169 static const char *blessed[] = {
175 static int slptime_threshold = 1800;
176 static int blktime_threshold = 900;
177 static int sleepfreq = 3;
180 deadlres_td_on_lock(struct proc *p, struct thread *td, int blkticks)
184 sx_assert(&allproc_lock, SX_LOCKED);
185 PROC_LOCK_ASSERT(p, MA_OWNED);
186 THREAD_LOCK_ASSERT(td, MA_OWNED);
188 * The thread should be blocked on a turnstile, simply check
189 * if the turnstile channel is in good state.
191 MPASS(td->td_blocked != NULL);
193 tticks = ticks - td->td_blktick;
194 if (tticks > blkticks)
196 * Accordingly with provided thresholds, this thread is stuck
197 * for too long on a turnstile.
199 panic("%s: possible deadlock detected for %p (%s), "
200 "blocked for %d ticks\n", __func__,
201 td, sched_tdname(td), tticks);
205 deadlres_td_sleep_q(struct proc *p, struct thread *td, int slpticks)
208 int i, slptype, tticks;
210 sx_assert(&allproc_lock, SX_LOCKED);
211 PROC_LOCK_ASSERT(p, MA_OWNED);
212 THREAD_LOCK_ASSERT(td, MA_OWNED);
214 * Check if the thread is sleeping on a lock, otherwise skip the check.
215 * Drop the thread lock in order to avoid a LOR with the sleepqueue
218 wchan = td->td_wchan;
219 tticks = ticks - td->td_slptick;
220 slptype = sleepq_type(wchan);
221 if ((slptype == SLEEPQ_SX || slptype == SLEEPQ_LK) &&
224 * Accordingly with provided thresholds, this thread is stuck
225 * for too long on a sleepqueue.
226 * However, being on a sleepqueue, we might still check for the
229 for (i = 0; blessed[i] != NULL; i++)
230 if (!strcmp(blessed[i], td->td_wmesg))
233 panic("%s: possible deadlock detected for %p (%s), "
234 "blocked for %d ticks\n", __func__,
235 td, sched_tdname(td), tticks);
244 int blkticks, slpticks, tryl;
248 blkticks = blktime_threshold * hz;
249 slpticks = slptime_threshold * hz;
252 * Avoid to sleep on the sx_lock in order to avoid a
253 * possible priority inversion problem leading to
255 * If the lock can't be held after 100 tries, panic.
257 if (!sx_try_slock(&allproc_lock)) {
259 panic("%s: possible deadlock detected "
260 "on allproc_lock\n", __func__);
262 pause("allproc", sleepfreq * hz);
266 FOREACH_PROC_IN_SYSTEM(p) {
268 if (p->p_state == PRS_NEW) {
272 FOREACH_THREAD_IN_PROC(p, td) {
275 deadlres_td_on_lock(p, td,
277 else if (TD_IS_SLEEPING(td))
278 deadlres_td_sleep_q(p, td,
284 sx_sunlock(&allproc_lock);
286 /* Sleep for sleepfreq seconds. */
287 pause("-", sleepfreq * hz);
291 static struct kthread_desc deadlkres_kd = {
294 (struct thread **)NULL
297 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
299 static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
300 "Deadlock resolver");
301 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RWTUN,
302 &slptime_threshold, 0,
303 "Number of seconds within is valid to sleep on a sleepqueue");
304 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RWTUN,
305 &blktime_threshold, 0,
306 "Number of seconds within is valid to block on a turnstile");
307 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RWTUN, &sleepfreq, 0,
308 "Number of seconds between any deadlock resolver thread run");
309 #endif /* DEADLKRES */
312 read_cpu_time(long *cp_time)
317 /* Sum up global cp_time[]. */
318 bzero(cp_time, sizeof(long) * CPUSTATES);
321 for (j = 0; j < CPUSTATES; j++)
322 cp_time[j] += pc->pc_cp_time[j];
326 #include <sys/watchdog.h>
328 static int watchdog_ticks;
329 static int watchdog_enabled;
330 static void watchdog_fire(void);
331 static void watchdog_config(void *, u_int, int *);
334 watchdog_attach(void)
336 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
340 * Clock handling routines.
342 * This code is written to operate with two timers that run independently of
345 * The main timer, running hz times per second, is used to trigger interval
346 * timers, timeouts and rescheduling as needed.
348 * The second timer handles kernel and user profiling,
349 * and does resource use estimation. If the second timer is programmable,
350 * it is randomized to avoid aliasing between the two clocks. For example,
351 * the randomization prevents an adversary from always giving up the cpu
352 * just before its quantum expires. Otherwise, it would never accumulate
353 * cpu ticks. The mean frequency of the second timer is stathz.
355 * If no second timer exists, stathz will be zero; in this case we drive
356 * profiling and statistics off the main clock. This WILL NOT be accurate;
357 * do not do it unless absolutely necessary.
359 * The statistics clock may (or may not) be run at a higher rate while
360 * profiling. This profile clock runs at profhz. We require that profhz
361 * be an integral multiple of stathz.
363 * If the statistics clock is running fast, it must be divided by the ratio
364 * profhz/stathz for statistics. (For profiling, every tick counts.)
366 * Time-of-day is maintained using a "timecounter", which may or may
367 * not be related to the hardware generating the above mentioned
377 DPCPU_DEFINE_STATIC(int, pcputicks); /* Per-CPU version of ticks. */
378 #ifdef DEVICE_POLLING
379 static int devpoll_run = 0;
383 ast_oweupc(struct thread *td, int tda __unused)
385 if ((td->td_proc->p_flag & P_PROFIL) == 0)
387 addupc_task(td, td->td_profil_addr, td->td_profil_ticks);
388 td->td_profil_ticks = 0;
389 td->td_pflags &= ~TDP_OWEUPC;
393 ast_alrm(struct thread *td, int tda __unused)
399 kern_psignal(p, SIGVTALRM);
404 ast_prof(struct thread *td, int tda __unused)
410 kern_psignal(p, SIGPROF);
415 * Initialize clock frequencies and start both clocks running.
418 initclocks(void *dummy __unused)
423 * Set divisors to 1 (normal case) and let the machine-specific
426 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
430 * Compute profhz/stathz, and fix profhz if needed.
432 i = stathz ? stathz : hz;
435 psratio = profhz / i;
437 ast_register(TDA_OWEUPC, ASTR_ASTF_REQUIRED, 0, ast_oweupc);
438 ast_register(TDA_ALRM, ASTR_ASTF_REQUIRED, 0, ast_alrm);
439 ast_register(TDA_PROF, ASTR_ASTF_REQUIRED, 0, ast_prof);
442 /* Enable hardclock watchdog now, even if a hardware watchdog exists. */
445 /* Volunteer to run a software watchdog. */
446 if (wdog_software_attach == NULL)
447 wdog_software_attach = watchdog_attach;
450 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
452 static __noinline void
453 hardclock_itimer(struct thread *td, struct pstats *pstats, int cnt, int usermode)
461 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
463 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
465 ast |= TDAI(TDA_ALRM);
468 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
470 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
472 ast |= TDAI(TDA_PROF);
476 ast_sched_mask(td, ast);
480 hardclock(int cnt, int usermode)
482 struct pstats *pstats;
483 struct thread *td = curthread;
484 struct proc *p = td->td_proc;
485 int *t = DPCPU_PTR(pcputicks);
486 int global, i, newticks;
489 * Update per-CPU and possibly global ticks values.
494 newticks = *t - global;
501 } while (!atomic_fcmpset_int(&ticks, &global, *t));
504 * Run current process's virtual and profile time, as needed.
508 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) ||
509 timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)))
510 hardclock_itimer(td, pstats, cnt, usermode);
513 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
514 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
515 if (td->td_intr_frame != NULL)
516 PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame);
518 /* We are in charge to handle this tick duty. */
520 tc_ticktock(newticks);
521 #ifdef DEVICE_POLLING
522 /* Dangerous and no need to call these things concurrently. */
523 if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) {
524 /* This is very short and quick. */
525 hardclock_device_poll();
526 atomic_store_rel_int(&devpoll_run, 0);
528 #endif /* DEVICE_POLLING */
529 if (watchdog_enabled > 0) {
530 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
531 if (i > 0 && i <= newticks)
534 intr_event_handle(clk_intr_event, NULL);
536 if (curcpu == CPU_FIRST())
537 cpu_tick_calibration();
538 if (__predict_false(DPCPU_GET(epoch_cb_count)))
539 GROUPTASK_ENQUEUE(DPCPU_PTR(epoch_cb_task));
543 hardclock_sync(int cpu)
546 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
547 t = DPCPU_ID_PTR(cpu, pcputicks);
553 * Regular integer scaling formula without losing precision:
555 #define TIME_INT_SCALE(value, mul, div) \
556 (((value) / (div)) * (mul) + (((value) % (div)) * (mul)) / (div))
559 * Macro for converting seconds and microseconds into actual ticks,
560 * based on the given hz value:
562 #define TIME_TO_TICKS(sec, usec, hz) \
563 ((sec) * (hz) + TIME_INT_SCALE(usec, hz, 1 << 6) / (1000000 >> 6))
565 #define TIME_ASSERT_VALID_HZ(hz) \
566 _Static_assert(TIME_TO_TICKS(INT_MAX / (hz) - 1, 999999, hz) >= 0 && \
567 TIME_TO_TICKS(INT_MAX / (hz) - 1, 999999, hz) < INT_MAX, \
568 "tvtohz() can overflow the regular integer type")
571 * Compile time assert the maximum and minimum values to fit into a
572 * regular integer when computing TIME_TO_TICKS():
574 TIME_ASSERT_VALID_HZ(HZ_MAXIMUM);
575 TIME_ASSERT_VALID_HZ(HZ_MINIMUM);
578 * The formula is mostly linear, but test some more common values just
581 TIME_ASSERT_VALID_HZ(1024);
582 TIME_ASSERT_VALID_HZ(1000);
583 TIME_ASSERT_VALID_HZ(128);
584 TIME_ASSERT_VALID_HZ(100);
587 * Compute number of ticks representing the specified amount of time.
588 * If the specified time is negative, a value of 1 is returned. This
589 * function returns a value from 1 up to and including INT_MAX.
592 tvtohz(struct timeval *tv)
597 * The values passed here may come from user-space and these
598 * checks ensure "tv_usec" is within its allowed range:
601 /* check for tv_usec underflow */
602 if (__predict_false(tv->tv_usec < 0)) {
603 tv->tv_sec += tv->tv_usec / 1000000;
604 tv->tv_usec = tv->tv_usec % 1000000;
605 /* convert tv_usec to a positive value */
606 if (__predict_true(tv->tv_usec < 0)) {
607 tv->tv_usec += 1000000;
610 /* check for tv_usec overflow */
611 } else if (__predict_false(tv->tv_usec >= 1000000)) {
612 tv->tv_sec += tv->tv_usec / 1000000;
613 tv->tv_usec = tv->tv_usec % 1000000;
616 /* check for tv_sec underflow */
617 if (__predict_false(tv->tv_sec < 0))
619 /* check for tv_sec overflow (including room for the tv_usec part) */
620 else if (__predict_false(tv->tv_sec >= tick_seconds_max))
623 /* cast to "int" to avoid platform differences */
624 retval = TIME_TO_TICKS((int)tv->tv_sec, (int)tv->tv_usec, hz);
626 /* add one additional tick */
631 * Start profiling on a process.
633 * Kernel profiling passes proc0 which never exits and hence
634 * keeps the profile clock running constantly.
637 startprofclock(struct proc *p)
640 PROC_LOCK_ASSERT(p, MA_OWNED);
641 if (p->p_flag & P_STOPPROF)
643 if ((p->p_flag & P_PROFIL) == 0) {
644 p->p_flag |= P_PROFIL;
645 mtx_lock(&time_lock);
646 if (++profprocs == 1)
647 cpu_startprofclock();
648 mtx_unlock(&time_lock);
653 * Stop profiling on a process.
656 stopprofclock(struct proc *p)
659 PROC_LOCK_ASSERT(p, MA_OWNED);
660 if (p->p_flag & P_PROFIL) {
661 if (p->p_profthreads != 0) {
662 while (p->p_profthreads != 0) {
663 p->p_flag |= P_STOPPROF;
664 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
668 if ((p->p_flag & P_PROFIL) == 0)
670 p->p_flag &= ~P_PROFIL;
671 mtx_lock(&time_lock);
672 if (--profprocs == 0)
674 mtx_unlock(&time_lock);
679 * Statistics clock. Updates rusage information and calls the scheduler
680 * to adjust priorities of the active thread.
682 * This should be called by all active processors.
685 statclock(int cnt, int usermode)
693 uint64_t runtime, new_switchtime;
698 cp_time = (long *)PCPU_PTR(cp_time);
701 * Charge the time as appropriate.
703 td->td_uticks += cnt;
704 if (p->p_nice > NZERO)
705 cp_time[CP_NICE] += cnt;
707 cp_time[CP_USER] += cnt;
710 * Came from kernel mode, so we were:
711 * - handling an interrupt,
712 * - doing syscall or trap work on behalf of the current
714 * - spinning in the idle loop.
715 * Whichever it is, charge the time as appropriate.
716 * Note that we charge interrupts to the current process,
717 * regardless of whether they are ``for'' that process,
718 * so that we know how much of its real time was spent
719 * in ``non-process'' (i.e., interrupt) work.
721 if ((td->td_pflags & TDP_ITHREAD) ||
722 td->td_intr_nesting_level >= 2) {
723 td->td_iticks += cnt;
724 cp_time[CP_INTR] += cnt;
726 td->td_pticks += cnt;
727 td->td_sticks += cnt;
728 if (!TD_IS_IDLETHREAD(td))
729 cp_time[CP_SYS] += cnt;
731 cp_time[CP_IDLE] += cnt;
735 /* Update resource usage integrals and maximums. */
736 MPASS(p->p_vmspace != NULL);
739 ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt;
740 ru->ru_idrss += pgtok(vm->vm_dsize) * cnt;
741 ru->ru_isrss += pgtok(vm->vm_ssize) * cnt;
742 rss = pgtok(vmspace_resident_count(vm));
743 if (ru->ru_maxrss < rss)
745 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
746 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
747 SDT_PROBE2(sched, , , tick, td, td->td_proc);
748 thread_lock_flags(td, MTX_QUIET);
751 * Compute the amount of time during which the current
752 * thread was running, and add that to its total so far.
754 new_switchtime = cpu_ticks();
755 runtime = new_switchtime - PCPU_GET(switchtime);
756 td->td_runtime += runtime;
757 td->td_incruntime += runtime;
758 PCPU_SET(switchtime, new_switchtime);
760 sched_clock(td, cnt);
763 if (td->td_intr_frame != NULL)
764 PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame);
769 profclock(int cnt, int usermode, uintfptr_t pc)
776 * Came from user mode; CPU was in user state.
777 * If this process is being profiled, record the tick.
778 * if there is no related user location yet, don't
779 * bother trying to count it.
781 if (td->td_proc->p_flag & P_PROFIL)
782 addupc_intr(td, pc, cnt);
785 if (td->td_intr_frame != NULL)
786 PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame);
791 * Return information about system clocks.
794 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
796 struct clockinfo clkinfo;
798 * Construct clockinfo structure.
800 bzero(&clkinfo, sizeof(clkinfo));
803 clkinfo.profhz = profhz;
804 clkinfo.stathz = stathz ? stathz : hz;
805 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
808 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
809 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
810 0, 0, sysctl_kern_clockrate, "S,clockinfo",
811 "Rate and period of various kernel clocks");
814 watchdog_config(void *unused __unused, u_int cmd, int *error)
818 u = cmd & WD_INTERVAL;
819 if (u >= WD_TO_1SEC) {
820 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
821 watchdog_enabled = 1;
824 watchdog_enabled = 0;
829 * Handle a watchdog timeout by dropping to DDB or panicking.
835 #if defined(KDB) && !defined(KDB_UNATTENDED)
837 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
839 panic("watchdog timeout");