2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1990, 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_synch.c 8.9 (Berkeley) 5/19/95
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
42 #include "opt_ktrace.h"
43 #include "opt_sched.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/condvar.h>
49 #include <sys/kernel.h>
52 #include <sys/mutex.h>
54 #include <sys/resourcevar.h>
55 #include <sys/refcount.h>
56 #include <sys/sched.h>
58 #include <sys/signalvar.h>
59 #include <sys/sleepqueue.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysproto.h>
64 #include <sys/vmmeter.h>
67 #include <sys/ktrace.h>
70 #include <sys/epoch.h>
73 #include <machine/cpu.h>
75 static void synch_setup(void *dummy);
76 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
80 static uint8_t pause_wchan[MAXCPU];
82 static struct callout loadav_callout;
84 struct loadavg averunnable =
85 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
87 * Constants for averages over 1, 5, and 15 minutes
88 * when sampling at 5 second intervals.
90 static fixpt_t cexp[3] = {
91 0.9200444146293232 * FSCALE, /* exp(-1/12) */
92 0.9834714538216174 * FSCALE, /* exp(-1/60) */
93 0.9944598480048967 * FSCALE, /* exp(-1/180) */
96 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
97 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE, "");
99 static void loadav(void *arg);
101 SDT_PROVIDER_DECLARE(sched);
102 SDT_PROBE_DEFINE(sched, , , preempt);
105 sleepinit(void *unused)
108 hogticks = (hz / 10) * 2; /* Default only. */
113 * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
116 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
119 * General sleep call. Suspends the current thread until a wakeup is
120 * performed on the specified identifier. The thread will then be made
121 * runnable with the specified priority. Sleeps at most sbt units of time
122 * (0 means no timeout). If pri includes the PCATCH flag, let signals
123 * interrupt the sleep, otherwise ignore them while sleeping. Returns 0 if
124 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
125 * signal becomes pending, ERESTART is returned if the current system
126 * call should be restarted if possible, and EINTR is returned if the system
127 * call should be interrupted by the signal (return EINTR).
129 * The lock argument is unlocked before the caller is suspended, and
130 * re-locked before _sleep() returns. If priority includes the PDROP
131 * flag the lock is not re-locked before returning.
134 _sleep(void *ident, struct lock_object *lock, int priority,
135 const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
138 struct lock_class *class;
139 uintptr_t lock_state;
140 int catch, pri, rval, sleepq_flags;
141 WITNESS_SAVE_DECL(lock_witness);
145 if (KTRPOINT(td, KTR_CSW))
148 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
149 "Sleeping on \"%s\"", wmesg);
150 KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL,
151 ("sleeping without a lock"));
152 KASSERT(ident != NULL, ("_sleep: NULL ident"));
153 KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
155 if (__predict_false(curthread->td_epochnest > 0))
156 epoch_trace_list(curthread);
158 KASSERT(td->td_epochnest == 0, ("sleeping in an epoch section"));
159 if (priority & PDROP)
160 KASSERT(lock != NULL && lock != &Giant.lock_object,
161 ("PDROP requires a non-Giant lock"));
163 class = LOCK_CLASS(lock);
167 if (SCHEDULER_STOPPED_TD(td)) {
168 if (lock != NULL && priority & PDROP)
169 class->lc_unlock(lock);
172 catch = priority & PCATCH;
173 pri = priority & PRIMASK;
175 KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
177 if ((uint8_t *)ident >= &pause_wchan[0] &&
178 (uint8_t *)ident <= &pause_wchan[MAXCPU - 1])
179 sleepq_flags = SLEEPQ_PAUSE;
181 sleepq_flags = SLEEPQ_SLEEP;
183 sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
186 CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
187 td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
189 if (lock == &Giant.lock_object)
190 mtx_assert(&Giant, MA_OWNED);
192 if (lock != NULL && lock != &Giant.lock_object &&
193 !(class->lc_flags & LC_SLEEPABLE)) {
194 WITNESS_SAVE(lock, lock_witness);
195 lock_state = class->lc_unlock(lock);
197 /* GCC needs to follow the Yellow Brick Road */
201 * We put ourselves on the sleep queue and start our timeout
202 * before calling thread_suspend_check, as we could stop there,
203 * and a wakeup or a SIGCONT (or both) could occur while we were
204 * stopped without resuming us. Thus, we must be ready for sleep
205 * when cursig() is called. If the wakeup happens while we're
206 * stopped, then td will no longer be on a sleep queue upon
207 * return from cursig().
209 sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
211 sleepq_set_timeout_sbt(ident, sbt, pr, flags);
212 if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
213 sleepq_release(ident);
214 WITNESS_SAVE(lock, lock_witness);
215 lock_state = class->lc_unlock(lock);
218 if (sbt != 0 && catch)
219 rval = sleepq_timedwait_sig(ident, pri);
221 rval = sleepq_timedwait(ident, pri);
223 rval = sleepq_wait_sig(ident, pri);
225 sleepq_wait(ident, pri);
229 if (KTRPOINT(td, KTR_CSW))
233 if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
234 class->lc_lock(lock, lock_state);
235 WITNESS_RESTORE(lock, lock_witness);
241 msleep_spin_sbt(void *ident, struct mtx *mtx, const char *wmesg,
242 sbintime_t sbt, sbintime_t pr, int flags)
246 WITNESS_SAVE_DECL(mtx);
249 KASSERT(mtx != NULL, ("sleeping without a mutex"));
250 KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
251 KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
253 if (SCHEDULER_STOPPED_TD(td))
257 CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
258 td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
261 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
262 WITNESS_SAVE(&mtx->lock_object, mtx);
263 mtx_unlock_spin(mtx);
266 * We put ourselves on the sleep queue and start our timeout.
268 sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
270 sleepq_set_timeout_sbt(ident, sbt, pr, flags);
273 * Can't call ktrace with any spin locks held so it can lock the
274 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
275 * any spin lock. Thus, we have to drop the sleepq spin lock while
276 * we handle those requests. This is safe since we have placed our
277 * thread on the sleep queue already.
280 if (KTRPOINT(td, KTR_CSW)) {
281 sleepq_release(ident);
287 sleepq_release(ident);
288 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
293 rval = sleepq_timedwait(ident, 0);
295 sleepq_wait(ident, 0);
299 if (KTRPOINT(td, KTR_CSW))
304 WITNESS_RESTORE(&mtx->lock_object, mtx);
309 * pause_sbt() delays the calling thread by the given signed binary
310 * time. During cold bootup, pause_sbt() uses the DELAY() function
311 * instead of the _sleep() function to do the waiting. The "sbt"
312 * argument must be greater than or equal to zero. A "sbt" value of
313 * zero is equivalent to a "sbt" value of one tick.
316 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
318 KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
320 /* silently convert invalid timeouts */
324 if ((cold && curthread == &thread0) || kdb_active ||
325 SCHEDULER_STOPPED()) {
327 * We delay one second at a time to avoid overflowing the
328 * system specific DELAY() function(s):
330 while (sbt >= SBT_1S) {
334 /* Do the delay remainder, if any */
335 sbt = howmany(sbt, SBT_1US);
338 return (EWOULDBLOCK);
340 return (_sleep(&pause_wchan[curcpu], NULL,
341 (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
345 * Potentially release the last reference for refcount. Check for
346 * unlikely conditions and signal the caller as to whether it was
350 refcount_release_last(volatile u_int *count, u_int n, u_int old)
354 waiter = old & REFCOUNT_WAITER;
355 old = REFCOUNT_COUNT(old);
356 if (__predict_false(n > old || REFCOUNT_SATURATED(old))) {
358 * Avoid multiple destructor invocations if underflow occurred.
359 * This is not perfect since the memory backing the containing
360 * object may already have been reallocated.
362 _refcount_update_saturated(count);
367 * Attempt to atomically clear the waiter bit. Wakeup waiters
368 * if we are successful.
370 if (waiter != 0 && atomic_cmpset_int(count, REFCOUNT_WAITER, 0))
371 wakeup(__DEVOLATILE(u_int *, count));
374 * Last reference. Signal the user to call the destructor.
376 * Ensure that the destructor sees all updates. The fence_rel
377 * at the start of refcount_releasen synchronizes with this fence.
379 atomic_thread_fence_acq();
384 * Wait for a refcount wakeup. This does not guarantee that the ref is still
385 * zero on return and may be subject to transient wakeups. Callers wanting
386 * a precise answer should use refcount_wait().
389 refcount_sleep(volatile u_int *count, const char *wmesg, int pri)
394 if (REFCOUNT_COUNT(*count) == 0)
396 wchan = __DEVOLATILE(void *, count);
400 if (REFCOUNT_COUNT(old) == 0) {
401 sleepq_release(wchan);
404 if (old & REFCOUNT_WAITER)
406 if (atomic_fcmpset_int(count, &old, old | REFCOUNT_WAITER))
409 sleepq_add(wchan, NULL, wmesg, 0, 0);
410 sleepq_wait(wchan, pri);
414 * Make all threads sleeping on the specified identifier runnable.
422 wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
423 sleepq_release(ident);
424 if (wakeup_swapper) {
425 KASSERT(ident != &proc0,
426 ("wakeup and wakeup_swapper and proc0"));
432 * Make a thread sleeping on the specified identifier runnable.
433 * May wake more than one thread if a target thread is currently
437 wakeup_one(void *ident)
442 wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
443 sleepq_release(ident);
449 wakeup_any(void *ident)
454 wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR,
456 sleepq_release(ident);
464 thread_unlock(curthread);
467 panic("%s: did not reenter debugger", __func__);
471 * The machine independent parts of context switching.
474 mi_switch(int flags, struct thread *newtd)
476 uint64_t runtime, new_switchtime;
479 td = curthread; /* XXX */
480 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
481 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
483 if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
484 mtx_assert(&Giant, MA_NOTOWNED);
486 KASSERT(td->td_critnest == 1 || panicstr,
487 ("mi_switch: switch in a critical section"));
488 KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
489 ("mi_switch: switch must be voluntary or involuntary"));
490 KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
493 * Don't perform context switches from the debugger.
497 if (SCHEDULER_STOPPED_TD(td))
499 if (flags & SW_VOL) {
500 td->td_ru.ru_nvcsw++;
501 td->td_swvoltick = ticks;
503 td->td_ru.ru_nivcsw++;
504 td->td_swinvoltick = ticks;
507 SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
510 * Compute the amount of time during which the current
511 * thread was running, and add that to its total so far.
513 new_switchtime = cpu_ticks();
514 runtime = new_switchtime - PCPU_GET(switchtime);
515 td->td_runtime += runtime;
516 td->td_incruntime += runtime;
517 PCPU_SET(switchtime, new_switchtime);
518 td->td_generation++; /* bump preempt-detect counter */
520 PCPU_SET(switchticks, ticks);
521 CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
522 td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
524 if (SDT_PROBES_ENABLED() &&
525 ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
526 (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
527 SDT_PROBE0(sched, , , preempt);
529 sched_switch(td, newtd, flags);
530 CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
531 td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
534 * If the last thread was exiting, finish cleaning it up.
536 if ((td = PCPU_GET(deadthread))) {
537 PCPU_SET(deadthread, NULL);
543 * Change thread state to be runnable, placing it on the run queue if
544 * it is in memory. If it is swapped out, return true so our caller
545 * will know to awaken the swapper.
548 setrunnable(struct thread *td)
551 THREAD_LOCK_ASSERT(td, MA_OWNED);
552 KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
553 ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
554 switch (td->td_state) {
560 * If we are only inhibited because we are swapped out
561 * then arange to swap in this process. Otherwise just return.
563 if (td->td_inhibitors != TDI_SWAPPED)
569 printf("state is 0x%x", td->td_state);
570 panic("setrunnable(2)");
572 if ((td->td_flags & TDF_INMEM) == 0) {
573 if ((td->td_flags & TDF_SWAPINREQ) == 0) {
574 td->td_flags |= TDF_SWAPINREQ;
583 * Compute a tenex style load average of a quantity on
584 * 1, 5 and 15 minute intervals.
595 for (i = 0; i < 3; i++)
596 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
597 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
600 * Schedule the next update to occur after 5 seconds, but add a
601 * random variation to avoid synchronisation with processes that
602 * run at regular intervals.
604 callout_reset_sbt(&loadav_callout,
605 SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
606 loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
611 synch_setup(void *dummy)
613 callout_init(&loadav_callout, 1);
615 /* Kick off timeout driven events by calling first time. */
623 return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
631 kern_yield(PRI_USER);
642 if (prio == PRI_USER)
643 prio = td->td_user_pri;
645 sched_prio(td, prio);
646 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
652 * General purpose yield system call.
655 sys_yield(struct thread *td, struct yield_args *uap)
659 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
660 sched_prio(td, PRI_MAX_TIMESHARE);
661 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
663 td->td_retval[0] = 0;