xref: /freebsd/sys/kern/kern_synch.c (revision dafba19e42e78cd3d7c9264ece49ddd3d7d70da5)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
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.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
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.
23  *
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
34  * SUCH DAMAGE.
35  */
36 
37 #include <sys/cdefs.h>
38 #include "opt_ktrace.h"
39 #include "opt_sched.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/blockcount.h>
44 #include <sys/condvar.h>
45 #include <sys/kdb.h>
46 #include <sys/kernel.h>
47 #include <sys/ktr.h>
48 #include <sys/ktrace.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resourcevar.h>
53 #include <sys/sched.h>
54 #include <sys/sdt.h>
55 #include <sys/signalvar.h>
56 #include <sys/sleepqueue.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/vmmeter.h>
62 #ifdef KTRACE
63 #include <sys/uio.h>
64 #endif
65 #ifdef EPOCH_TRACE
66 #include <sys/epoch.h>
67 #endif
68 
69 #include <machine/cpu.h>
70 
71 static void synch_setup(void *dummy);
72 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
73     NULL);
74 
75 int	hogticks;
76 static const char pause_wchan[MAXCPU];
77 
78 static struct callout loadav_callout;
79 
80 struct loadavg averunnable =
81 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
82 /*
83  * Constants for averages over 1, 5, and 15 minutes
84  * when sampling at 5 second intervals.
85  */
86 static uint64_t cexp[3] = {
87 	0.9200444146293232 * FSCALE,	/* exp(-1/12) */
88 	0.9834714538216174 * FSCALE,	/* exp(-1/60) */
89 	0.9944598480048967 * FSCALE,	/* exp(-1/180) */
90 };
91 
92 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
93 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE,
94     "Fixed-point scale factor used for calculating load average values");
95 
96 static void	loadav(void *arg);
97 
98 SDT_PROVIDER_DECLARE(sched);
99 SDT_PROBE_DEFINE(sched, , , preempt);
100 
101 static void
102 sleepinit(void *unused)
103 {
104 
105 	hogticks = (hz / 10) * 2;	/* Default only. */
106 	init_sleepqueues();
107 }
108 
109 /*
110  * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
111  * it is available.
112  */
113 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
114 
115 /*
116  * General sleep call.  Suspends the current thread until a wakeup is
117  * performed on the specified identifier.  The thread will then be made
118  * runnable with the specified priority.  Sleeps at most sbt units of time
119  * (0 means no timeout).  If pri includes the PCATCH flag, let signals
120  * interrupt the sleep, otherwise ignore them while sleeping.  Returns 0 if
121  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
122  * signal becomes pending, ERESTART is returned if the current system
123  * call should be restarted if possible, and EINTR is returned if the system
124  * call should be interrupted by the signal (return EINTR).
125  *
126  * The lock argument is unlocked before the caller is suspended, and
127  * re-locked before _sleep() returns.  If priority includes the PDROP
128  * flag the lock is not re-locked before returning.
129  */
130 int
131 _sleep(const void *ident, struct lock_object *lock, int priority,
132     const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
133 {
134 	struct thread *td __ktrace_used;
135 	struct lock_class *class;
136 	struct timespec sw_out_tv __ktrace_used;
137 	uintptr_t lock_state;
138 	int catch, pri, rval, sleepq_flags;
139 	WITNESS_SAVE_DECL(lock_witness);
140 
141 	TSENTER();
142 	td = curthread;
143 #ifdef KTRACE
144 	if (KTRPOINT(td, KTR_CSW))
145 		nanotime(&sw_out_tv);
146 #endif
147 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
148 	    "Sleeping on \"%s\"", wmesg);
149 	KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL ||
150 	    (priority & PNOLOCK) != 0,
151 	    ("sleeping without a lock"));
152 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
153 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
154 	if (priority & PDROP)
155 		KASSERT(lock != NULL && lock != &Giant.lock_object,
156 		    ("PDROP requires a non-Giant lock"));
157 	if (lock != NULL)
158 		class = LOCK_CLASS(lock);
159 	else
160 		class = NULL;
161 
162 	if (SCHEDULER_STOPPED()) {
163 		if (lock != NULL && priority & PDROP)
164 			class->lc_unlock(lock);
165 		return (0);
166 	}
167 	catch = priority & PCATCH;
168 	pri = priority & PRIMASK;
169 
170 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
171 
172 	if ((uintptr_t)ident >= (uintptr_t)&pause_wchan[0] &&
173 	    (uintptr_t)ident <= (uintptr_t)&pause_wchan[MAXCPU - 1])
174 		sleepq_flags = SLEEPQ_PAUSE;
175 	else
176 		sleepq_flags = SLEEPQ_SLEEP;
177 	if (catch)
178 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
179 
180 	sleepq_lock(ident);
181 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
182 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
183 
184 	if (lock == &Giant.lock_object)
185 		mtx_assert(&Giant, MA_OWNED);
186 	DROP_GIANT();
187 	if (lock != NULL && lock != &Giant.lock_object &&
188 	    !(class->lc_flags & LC_SLEEPABLE)) {
189 		KASSERT(!(class->lc_flags & LC_SPINLOCK),
190 		    ("spin locks can only use msleep_spin"));
191 		WITNESS_SAVE(lock, lock_witness);
192 		lock_state = class->lc_unlock(lock);
193 	} else
194 		/* GCC needs to follow the Yellow Brick Road */
195 		lock_state = -1;
196 
197 	/*
198 	 * We put ourselves on the sleep queue and start our timeout
199 	 * before calling thread_suspend_check, as we could stop there,
200 	 * and a wakeup or a SIGCONT (or both) could occur while we were
201 	 * stopped without resuming us.  Thus, we must be ready for sleep
202 	 * when cursig() is called.  If the wakeup happens while we're
203 	 * stopped, then td will no longer be on a sleep queue upon
204 	 * return from cursig().
205 	 */
206 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
207 	if (sbt != 0)
208 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
209 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
210 		sleepq_release(ident);
211 		WITNESS_SAVE(lock, lock_witness);
212 		lock_state = class->lc_unlock(lock);
213 		sleepq_lock(ident);
214 	}
215 	if (sbt != 0 && catch)
216 		rval = sleepq_timedwait_sig(ident, pri);
217 	else if (sbt != 0)
218 		rval = sleepq_timedwait(ident, pri);
219 	else if (catch)
220 		rval = sleepq_wait_sig(ident, pri);
221 	else {
222 		sleepq_wait(ident, pri);
223 		rval = 0;
224 	}
225 #ifdef KTRACE
226 	if (KTRPOINT(td, KTR_CSW)) {
227 		ktrcsw_out(&sw_out_tv, wmesg);
228 		ktrcsw(0, 0, wmesg);
229 	}
230 #endif
231 	PICKUP_GIANT();
232 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
233 		class->lc_lock(lock, lock_state);
234 		WITNESS_RESTORE(lock, lock_witness);
235 	}
236 	TSEXIT();
237 	return (rval);
238 }
239 
240 int
241 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
242     sbintime_t sbt, sbintime_t pr, int flags)
243 {
244 	struct thread *td __ktrace_used;
245 	struct timespec sw_out_tv __ktrace_used;
246 	int rval;
247 	WITNESS_SAVE_DECL(mtx);
248 
249 	td = curthread;
250 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
251 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
252 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
253 
254 	if (SCHEDULER_STOPPED())
255 		return (0);
256 
257 	sleepq_lock(ident);
258 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
259 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
260 
261 	DROP_GIANT();
262 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
263 	WITNESS_SAVE(&mtx->lock_object, mtx);
264 	mtx_unlock_spin(mtx);
265 
266 	/*
267 	 * We put ourselves on the sleep queue and start our timeout.
268 	 */
269 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
270 	if (sbt != 0)
271 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
272 
273 #ifdef KTRACE
274 	if (KTRPOINT(td, KTR_CSW))
275 		nanotime(&sw_out_tv);
276 #endif
277 #ifdef WITNESS
278 	sleepq_release(ident);
279 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
280 	    wmesg);
281 	sleepq_lock(ident);
282 #endif
283 	if (sbt != 0)
284 		rval = sleepq_timedwait(ident, 0);
285 	else {
286 		sleepq_wait(ident, 0);
287 		rval = 0;
288 	}
289 #ifdef KTRACE
290 	if (KTRPOINT(td, KTR_CSW)) {
291 		ktrcsw_out(&sw_out_tv, wmesg);
292 		ktrcsw(0, 0, wmesg);
293 	}
294 #endif
295 	PICKUP_GIANT();
296 	mtx_lock_spin(mtx);
297 	WITNESS_RESTORE(&mtx->lock_object, mtx);
298 	return (rval);
299 }
300 
301 /*
302  * pause_sbt() delays the calling thread by the given signed binary
303  * time. During cold bootup, pause_sbt() uses the DELAY() function
304  * instead of the _sleep() function to do the waiting. The "sbt"
305  * argument must be greater than or equal to zero. A "sbt" value of
306  * zero is equivalent to a "sbt" value of one tick.
307  */
308 int
309 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
310 {
311 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
312 
313 	/* silently convert invalid timeouts */
314 	if (sbt == 0)
315 		sbt = tick_sbt;
316 
317 	if ((cold && curthread == &thread0) || kdb_active ||
318 	    SCHEDULER_STOPPED()) {
319 		/*
320 		 * We delay one second at a time to avoid overflowing the
321 		 * system specific DELAY() function(s):
322 		 */
323 		while (sbt >= SBT_1S) {
324 			DELAY(1000000);
325 			sbt -= SBT_1S;
326 		}
327 		/* Do the delay remainder, if any */
328 		sbt = howmany(sbt, SBT_1US);
329 		if (sbt > 0)
330 			DELAY(sbt);
331 		return (EWOULDBLOCK);
332 	}
333 	return (_sleep(&pause_wchan[curcpu], NULL,
334 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
335 }
336 
337 /*
338  * Make all threads sleeping on the specified identifier runnable.
339  */
340 void
341 wakeup(const void *ident)
342 {
343 	sleepq_lock(ident);
344 	sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
345 	sleepq_release(ident);
346 }
347 
348 /*
349  * Make a thread sleeping on the specified identifier runnable.
350  * May wake more than one thread if a target thread is currently
351  * swapped out.
352  */
353 void
354 wakeup_one(const void *ident)
355 {
356 	sleepq_lock(ident);
357 	sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_DROP, 0, 0);
358 }
359 
360 void
361 wakeup_any(const void *ident)
362 {
363 	sleepq_lock(ident);
364 	sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR | SLEEPQ_DROP, 0, 0);
365 }
366 
367 /*
368  * Signal sleeping waiters after the counter has reached zero.
369  */
370 void
371 _blockcount_wakeup(blockcount_t *bc, u_int old)
372 {
373 
374 	KASSERT(_BLOCKCOUNT_WAITERS(old),
375 	    ("%s: no waiters on %p", __func__, bc));
376 
377 	if (atomic_cmpset_int(&bc->__count, _BLOCKCOUNT_WAITERS_FLAG, 0))
378 		wakeup(bc);
379 }
380 
381 /*
382  * Wait for a wakeup or a signal.  This does not guarantee that the count is
383  * still zero on return.  Callers wanting a precise answer should use
384  * blockcount_wait() with an interlock.
385  *
386  * If there is no work to wait for, return 0.  If the sleep was interrupted by a
387  * signal, return EINTR or ERESTART, and return EAGAIN otherwise.
388  */
389 int
390 _blockcount_sleep(blockcount_t *bc, struct lock_object *lock, const char *wmesg,
391     int prio)
392 {
393 	void *wchan;
394 	uintptr_t lock_state;
395 	u_int old;
396 	int ret;
397 	bool catch, drop;
398 
399 	KASSERT(lock != &Giant.lock_object,
400 	    ("%s: cannot use Giant as the interlock", __func__));
401 
402 	catch = (prio & PCATCH) != 0;
403 	drop = (prio & PDROP) != 0;
404 	prio &= PRIMASK;
405 
406 	/*
407 	 * Synchronize with the fence in blockcount_release().  If we end up
408 	 * waiting, the sleepqueue lock acquisition will provide the required
409 	 * side effects.
410 	 *
411 	 * If there is no work to wait for, but waiters are present, try to put
412 	 * ourselves to sleep to avoid jumping ahead.
413 	 */
414 	if (atomic_load_acq_int(&bc->__count) == 0) {
415 		if (lock != NULL && drop)
416 			LOCK_CLASS(lock)->lc_unlock(lock);
417 		return (0);
418 	}
419 	lock_state = 0;
420 	wchan = bc;
421 	sleepq_lock(wchan);
422 	DROP_GIANT();
423 	if (lock != NULL)
424 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
425 	old = blockcount_read(bc);
426 	ret = 0;
427 	do {
428 		if (_BLOCKCOUNT_COUNT(old) == 0) {
429 			sleepq_release(wchan);
430 			goto out;
431 		}
432 		if (_BLOCKCOUNT_WAITERS(old))
433 			break;
434 	} while (!atomic_fcmpset_int(&bc->__count, &old,
435 	    old | _BLOCKCOUNT_WAITERS_FLAG));
436 	sleepq_add(wchan, NULL, wmesg, catch ? SLEEPQ_INTERRUPTIBLE : 0, 0);
437 	if (catch)
438 		ret = sleepq_wait_sig(wchan, prio);
439 	else
440 		sleepq_wait(wchan, prio);
441 	if (ret == 0)
442 		ret = EAGAIN;
443 
444 out:
445 	PICKUP_GIANT();
446 	if (lock != NULL && !drop)
447 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
448 
449 	return (ret);
450 }
451 
452 static void
453 kdb_switch(void)
454 {
455 	thread_unlock(curthread);
456 	kdb_backtrace();
457 	kdb_reenter();
458 	panic("%s: did not reenter debugger", __func__);
459 }
460 
461 /*
462  * mi_switch(9): The machine-independent parts of context switching.
463  *
464  * The thread lock is required on entry and is no longer held on return.
465  */
466 void
467 mi_switch(int flags)
468 {
469 	uint64_t runtime, new_switchtime;
470 	struct thread *td;
471 
472 	td = curthread;			/* XXX */
473 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
474 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
475 #ifdef INVARIANTS
476 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
477 		mtx_assert(&Giant, MA_NOTOWNED);
478 #endif
479 	/* thread_lock() performs spinlock_enter(). */
480 	KASSERT(td->td_critnest == 1 || KERNEL_PANICKED(),
481 	    ("mi_switch: switch in a critical section"));
482 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
483 	    ("mi_switch: switch must be voluntary or involuntary"));
484 	KASSERT((flags & SW_TYPE_MASK) != 0,
485 	    ("mi_switch: a switch reason (type) must be specified"));
486 	KASSERT((flags & SW_TYPE_MASK) < SWT_COUNT,
487 	    ("mi_switch: invalid switch reason %d", (flags & SW_TYPE_MASK)));
488 
489 	/*
490 	 * Don't perform context switches from the debugger.
491 	 */
492 	if (kdb_active)
493 		kdb_switch();
494 	if (SCHEDULER_STOPPED())
495 		return;
496 	if (flags & SW_VOL) {
497 		td->td_ru.ru_nvcsw++;
498 		td->td_swvoltick = ticks;
499 	} else {
500 		td->td_ru.ru_nivcsw++;
501 		td->td_swinvoltick = ticks;
502 	}
503 #ifdef SCHED_STATS
504 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
505 #endif
506 	/*
507 	 * Compute the amount of time during which the current
508 	 * thread was running, and add that to its total so far.
509 	 */
510 	new_switchtime = cpu_ticks();
511 	runtime = new_switchtime - PCPU_GET(switchtime);
512 	td->td_runtime += runtime;
513 	td->td_incruntime += runtime;
514 	PCPU_SET(switchtime, new_switchtime);
515 	td->td_generation++;	/* bump preempt-detect counter */
516 	VM_CNT_INC(v_swtch);
517 	PCPU_SET(switchticks, ticks);
518 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
519 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
520 #ifdef KDTRACE_HOOKS
521 	if (SDT_PROBES_ENABLED() &&
522 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
523 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
524 		SDT_PROBE0(sched, , , preempt);
525 #endif
526 	sched_switch(td, flags);
527 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
528 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
529 
530 	/*
531 	 * If the last thread was exiting, finish cleaning it up.
532 	 */
533 	if ((td = PCPU_GET(deadthread))) {
534 		PCPU_SET(deadthread, NULL);
535 		thread_stash(td);
536 	}
537 	spinlock_exit();
538 }
539 
540 /*
541  * Change thread state to be runnable, placing it on the run queue.
542  *
543  * Requires the thread lock on entry, drops on exit.
544  */
545 void
546 setrunnable(struct thread *td, int srqflags)
547 {
548 	THREAD_LOCK_ASSERT(td, MA_OWNED);
549 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
550 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
551 
552 	switch (TD_GET_STATE(td)) {
553 	case TDS_RUNNING:
554 	case TDS_RUNQ:
555 	case TDS_INHIBITED:
556 		if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
557 			thread_unlock(td);
558 		break;
559 	case TDS_CAN_RUN:
560 		KASSERT((td->td_flags & TDF_INMEM) != 0,
561 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
562 		    td, td->td_flags, td->td_inhibitors));
563 		/* unlocks thread lock according to flags */
564 		sched_wakeup(td, srqflags);
565 		break;
566 	default:
567 		panic("setrunnable: state 0x%x", TD_GET_STATE(td));
568 	}
569 }
570 
571 /*
572  * Compute a tenex style load average of a quantity on
573  * 1, 5 and 15 minute intervals.
574  */
575 static void
576 loadav(void *arg)
577 {
578 	int i;
579 	uint64_t nrun;
580 	struct loadavg *avg;
581 
582 	nrun = (uint64_t)sched_load();
583 	avg = &averunnable;
584 
585 	for (i = 0; i < 3; i++)
586 		avg->ldavg[i] = (cexp[i] * (uint64_t)avg->ldavg[i] +
587 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
588 
589 	/*
590 	 * Schedule the next update to occur after 5 seconds, but add a
591 	 * random variation to avoid synchronisation with processes that
592 	 * run at regular intervals.
593 	 */
594 	callout_reset_sbt(&loadav_callout,
595 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
596 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
597 }
598 
599 void
600 ast_scheduler(struct thread *td, int tda __unused)
601 {
602 #ifdef KTRACE
603 	if (KTRPOINT(td, KTR_CSW))
604 		ktrcsw(1, 1, __func__);
605 #endif
606 	thread_lock(td);
607 	sched_prio(td, td->td_user_pri);
608 	mi_switch(SW_INVOL | SWT_NEEDRESCHED);
609 #ifdef KTRACE
610 	if (KTRPOINT(td, KTR_CSW))
611 		ktrcsw(0, 1, __func__);
612 #endif
613 }
614 
615 static void
616 synch_setup(void *dummy __unused)
617 {
618 	callout_init(&loadav_callout, 1);
619 	ast_register(TDA_SCHED, ASTR_ASTF_REQUIRED, 0, ast_scheduler);
620 
621 	/* Kick off timeout driven events by calling first time. */
622 	loadav(NULL);
623 }
624 
625 bool
626 should_yield(void)
627 {
628 
629 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
630 }
631 
632 void
633 maybe_yield(void)
634 {
635 
636 	if (should_yield())
637 		kern_yield(PRI_USER);
638 }
639 
640 void
641 kern_yield(int prio)
642 {
643 	struct thread *td;
644 
645 	td = curthread;
646 	DROP_GIANT();
647 	thread_lock(td);
648 	if (prio == PRI_USER)
649 		prio = td->td_user_pri;
650 	if (prio >= 0)
651 		sched_prio(td, prio);
652 	mi_switch(SW_VOL | SWT_RELINQUISH);
653 	PICKUP_GIANT();
654 }
655 
656 /*
657  * General purpose yield system call.
658  */
659 int
660 sys_yield(struct thread *td, struct yield_args *uap)
661 {
662 
663 	thread_lock(td);
664 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
665 		sched_prio(td, PRI_MAX_TIMESHARE);
666 	mi_switch(SW_VOL | SWT_RELINQUISH);
667 	td->td_retval[0] = 0;
668 	return (0);
669 }
670 
671 int
672 sys_sched_getcpu(struct thread *td, struct sched_getcpu_args *uap)
673 {
674 	td->td_retval[0] = td->td_oncpu;
675 	return (0);
676 }
677