xref: /freebsd/sys/kern/kern_synch.c (revision e9b148a3185f41e3a09e91ea75cae7828d908845)
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  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_ktrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/condvar.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/proc.h>
54 #include <sys/resourcevar.h>
55 #include <sys/refcount.h>
56 #include <sys/sched.h>
57 #include <sys/sdt.h>
58 #include <sys/signalvar.h>
59 #include <sys/sleepqueue.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysproto.h>
64 #include <sys/vmmeter.h>
65 #ifdef KTRACE
66 #include <sys/uio.h>
67 #include <sys/ktrace.h>
68 #endif
69 #ifdef EPOCH_TRACE
70 #include <sys/epoch.h>
71 #endif
72 
73 #include <machine/cpu.h>
74 
75 static void synch_setup(void *dummy);
76 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
77     NULL);
78 
79 int	hogticks;
80 static uint8_t pause_wchan[MAXCPU];
81 
82 static struct callout loadav_callout;
83 
84 struct loadavg averunnable =
85 	{ {0, 0, 0}, FSCALE };	/* load average, of runnable procs */
86 /*
87  * Constants for averages over 1, 5, and 15 minutes
88  * when sampling at 5 second intervals.
89  */
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) */
94 };
95 
96 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
97 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE, "");
98 
99 static void	loadav(void *arg);
100 
101 SDT_PROVIDER_DECLARE(sched);
102 SDT_PROBE_DEFINE(sched, , , preempt);
103 
104 static void
105 sleepinit(void *unused)
106 {
107 
108 	hogticks = (hz / 10) * 2;	/* Default only. */
109 	init_sleepqueues();
110 }
111 
112 /*
113  * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
114  * it is available.
115  */
116 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
117 
118 /*
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).
128  *
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.
132  */
133 int
134 _sleep(void *ident, struct lock_object *lock, int priority,
135     const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
136 {
137 	struct thread *td;
138 	struct lock_class *class;
139 	uintptr_t lock_state;
140 	int catch, pri, rval, sleepq_flags;
141 	WITNESS_SAVE_DECL(lock_witness);
142 
143 	td = curthread;
144 #ifdef KTRACE
145 	if (KTRPOINT(td, KTR_CSW))
146 		ktrcsw(1, 0, wmesg);
147 #endif
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"));
154 #ifdef EPOCH_TRACE
155 	if (__predict_false(curthread->td_epochnest > 0))
156 		epoch_trace_list(curthread);
157 #endif
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"));
162 	if (lock != NULL)
163 		class = LOCK_CLASS(lock);
164 	else
165 		class = NULL;
166 
167 	if (SCHEDULER_STOPPED_TD(td)) {
168 		if (lock != NULL && priority & PDROP)
169 			class->lc_unlock(lock);
170 		return (0);
171 	}
172 	catch = priority & PCATCH;
173 	pri = priority & PRIMASK;
174 
175 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
176 
177 	if ((uint8_t *)ident >= &pause_wchan[0] &&
178 	    (uint8_t *)ident <= &pause_wchan[MAXCPU - 1])
179 		sleepq_flags = SLEEPQ_PAUSE;
180 	else
181 		sleepq_flags = SLEEPQ_SLEEP;
182 	if (catch)
183 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
184 
185 	sleepq_lock(ident);
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);
188 
189 	if (lock == &Giant.lock_object)
190 		mtx_assert(&Giant, MA_OWNED);
191 	DROP_GIANT();
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);
196 	} else
197 		/* GCC needs to follow the Yellow Brick Road */
198 		lock_state = -1;
199 
200 	/*
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().
208 	 */
209 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
210 	if (sbt != 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);
216 		sleepq_lock(ident);
217 	}
218 	if (sbt != 0 && catch)
219 		rval = sleepq_timedwait_sig(ident, pri);
220 	else if (sbt != 0)
221 		rval = sleepq_timedwait(ident, pri);
222 	else if (catch)
223 		rval = sleepq_wait_sig(ident, pri);
224 	else {
225 		sleepq_wait(ident, pri);
226 		rval = 0;
227 	}
228 #ifdef KTRACE
229 	if (KTRPOINT(td, KTR_CSW))
230 		ktrcsw(0, 0, wmesg);
231 #endif
232 	PICKUP_GIANT();
233 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
234 		class->lc_lock(lock, lock_state);
235 		WITNESS_RESTORE(lock, lock_witness);
236 	}
237 	return (rval);
238 }
239 
240 int
241 msleep_spin_sbt(void *ident, struct mtx *mtx, const char *wmesg,
242     sbintime_t sbt, sbintime_t pr, int flags)
243 {
244 	struct thread *td;
245 	int rval;
246 	WITNESS_SAVE_DECL(mtx);
247 
248 	td = curthread;
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"));
252 
253 	if (SCHEDULER_STOPPED_TD(td))
254 		return (0);
255 
256 	sleepq_lock(ident);
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);
259 
260 	DROP_GIANT();
261 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
262 	WITNESS_SAVE(&mtx->lock_object, mtx);
263 	mtx_unlock_spin(mtx);
264 
265 	/*
266 	 * We put ourselves on the sleep queue and start our timeout.
267 	 */
268 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
269 	if (sbt != 0)
270 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
271 
272 	/*
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.
278 	 */
279 #ifdef KTRACE
280 	if (KTRPOINT(td, KTR_CSW)) {
281 		sleepq_release(ident);
282 		ktrcsw(1, 0, wmesg);
283 		sleepq_lock(ident);
284 	}
285 #endif
286 #ifdef WITNESS
287 	sleepq_release(ident);
288 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
289 	    wmesg);
290 	sleepq_lock(ident);
291 #endif
292 	if (sbt != 0)
293 		rval = sleepq_timedwait(ident, 0);
294 	else {
295 		sleepq_wait(ident, 0);
296 		rval = 0;
297 	}
298 #ifdef KTRACE
299 	if (KTRPOINT(td, KTR_CSW))
300 		ktrcsw(0, 0, wmesg);
301 #endif
302 	PICKUP_GIANT();
303 	mtx_lock_spin(mtx);
304 	WITNESS_RESTORE(&mtx->lock_object, mtx);
305 	return (rval);
306 }
307 
308 /*
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.
314  */
315 int
316 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
317 {
318 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
319 
320 	/* silently convert invalid timeouts */
321 	if (sbt == 0)
322 		sbt = tick_sbt;
323 
324 	if ((cold && curthread == &thread0) || kdb_active ||
325 	    SCHEDULER_STOPPED()) {
326 		/*
327 		 * We delay one second at a time to avoid overflowing the
328 		 * system specific DELAY() function(s):
329 		 */
330 		while (sbt >= SBT_1S) {
331 			DELAY(1000000);
332 			sbt -= SBT_1S;
333 		}
334 		/* Do the delay remainder, if any */
335 		sbt = howmany(sbt, SBT_1US);
336 		if (sbt > 0)
337 			DELAY(sbt);
338 		return (EWOULDBLOCK);
339 	}
340 	return (_sleep(&pause_wchan[curcpu], NULL,
341 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
342 }
343 
344 /*
345  * Potentially release the last reference for refcount.  Check for
346  * unlikely conditions and signal the caller as to whether it was
347  * the final ref.
348  */
349 bool
350 refcount_release_last(volatile u_int *count, u_int n, u_int old)
351 {
352 	u_int waiter;
353 
354 	waiter = old & REFCOUNT_WAITER;
355 	old = REFCOUNT_COUNT(old);
356 	if (__predict_false(n > old || REFCOUNT_SATURATED(old))) {
357 		/*
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.
361 		 */
362 		_refcount_update_saturated(count);
363 		return (false);
364 	}
365 
366 	/*
367 	 * Attempt to atomically clear the waiter bit.  Wakeup waiters
368 	 * if we are successful.
369 	 */
370 	if (waiter != 0 && atomic_cmpset_int(count, REFCOUNT_WAITER, 0))
371 		wakeup(__DEVOLATILE(u_int *, count));
372 
373 	/*
374 	 * Last reference.  Signal the user to call the destructor.
375 	 *
376 	 * Ensure that the destructor sees all updates.  The fence_rel
377 	 * at the start of refcount_releasen synchronizes with this fence.
378 	 */
379 	atomic_thread_fence_acq();
380 	return (true);
381 }
382 
383 /*
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().
387  */
388 void
389 refcount_sleep(volatile u_int *count, const char *wmesg, int pri)
390 {
391 	void *wchan;
392 	u_int old;
393 
394 	if (REFCOUNT_COUNT(*count) == 0)
395 		return;
396 	wchan = __DEVOLATILE(void *, count);
397 	sleepq_lock(wchan);
398 	old = *count;
399 	for (;;) {
400 		if (REFCOUNT_COUNT(old) == 0) {
401 			sleepq_release(wchan);
402 			return;
403 		}
404 		if (old & REFCOUNT_WAITER)
405 			break;
406 		if (atomic_fcmpset_int(count, &old, old | REFCOUNT_WAITER))
407 			break;
408 	}
409 	sleepq_add(wchan, NULL, wmesg, 0, 0);
410 	sleepq_wait(wchan, pri);
411 }
412 
413 /*
414  * Make all threads sleeping on the specified identifier runnable.
415  */
416 void
417 wakeup(void *ident)
418 {
419 	int wakeup_swapper;
420 
421 	sleepq_lock(ident);
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"));
427 		kick_proc0();
428 	}
429 }
430 
431 /*
432  * Make a thread sleeping on the specified identifier runnable.
433  * May wake more than one thread if a target thread is currently
434  * swapped out.
435  */
436 void
437 wakeup_one(void *ident)
438 {
439 	int wakeup_swapper;
440 
441 	sleepq_lock(ident);
442 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
443 	sleepq_release(ident);
444 	if (wakeup_swapper)
445 		kick_proc0();
446 }
447 
448 void
449 wakeup_any(void *ident)
450 {
451 	int wakeup_swapper;
452 
453 	sleepq_lock(ident);
454 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR,
455 	    0, 0);
456 	sleepq_release(ident);
457 	if (wakeup_swapper)
458 		kick_proc0();
459 }
460 
461 static void
462 kdb_switch(void)
463 {
464 	thread_unlock(curthread);
465 	kdb_backtrace();
466 	kdb_reenter();
467 	panic("%s: did not reenter debugger", __func__);
468 }
469 
470 /*
471  * The machine independent parts of context switching.
472  */
473 void
474 mi_switch(int flags, struct thread *newtd)
475 {
476 	uint64_t runtime, new_switchtime;
477 	struct thread *td;
478 
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"));
482 #ifdef INVARIANTS
483 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
484 		mtx_assert(&Giant, MA_NOTOWNED);
485 #endif
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"));
491 
492 	/*
493 	 * Don't perform context switches from the debugger.
494 	 */
495 	if (kdb_active)
496 		kdb_switch();
497 	if (SCHEDULER_STOPPED_TD(td))
498 		return;
499 	if (flags & SW_VOL) {
500 		td->td_ru.ru_nvcsw++;
501 		td->td_swvoltick = ticks;
502 	} else {
503 		td->td_ru.ru_nivcsw++;
504 		td->td_swinvoltick = ticks;
505 	}
506 #ifdef SCHED_STATS
507 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
508 #endif
509 	/*
510 	 * Compute the amount of time during which the current
511 	 * thread was running, and add that to its total so far.
512 	 */
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 */
519 	VM_CNT_INC(v_swtch);
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);
523 #ifdef KDTRACE_HOOKS
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);
528 #endif
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);
532 
533 	/*
534 	 * If the last thread was exiting, finish cleaning it up.
535 	 */
536 	if ((td = PCPU_GET(deadthread))) {
537 		PCPU_SET(deadthread, NULL);
538 		thread_stash(td);
539 	}
540 }
541 
542 /*
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.
546  */
547 int
548 setrunnable(struct thread *td)
549 {
550 
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) {
555 	case TDS_RUNNING:
556 	case TDS_RUNQ:
557 		return (0);
558 	case TDS_INHIBITED:
559 		/*
560 		 * If we are only inhibited because we are swapped out
561 		 * then arange to swap in this process. Otherwise just return.
562 		 */
563 		if (td->td_inhibitors != TDI_SWAPPED)
564 			return (0);
565 		/* FALLTHROUGH */
566 	case TDS_CAN_RUN:
567 		break;
568 	default:
569 		printf("state is 0x%x", td->td_state);
570 		panic("setrunnable(2)");
571 	}
572 	if ((td->td_flags & TDF_INMEM) == 0) {
573 		if ((td->td_flags & TDF_SWAPINREQ) == 0) {
574 			td->td_flags |= TDF_SWAPINREQ;
575 			return (1);
576 		}
577 	} else
578 		sched_wakeup(td);
579 	return (0);
580 }
581 
582 /*
583  * Compute a tenex style load average of a quantity on
584  * 1, 5 and 15 minute intervals.
585  */
586 static void
587 loadav(void *arg)
588 {
589 	int i, nrun;
590 	struct loadavg *avg;
591 
592 	nrun = sched_load();
593 	avg = &averunnable;
594 
595 	for (i = 0; i < 3; i++)
596 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
597 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
598 
599 	/*
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.
603 	 */
604 	callout_reset_sbt(&loadav_callout,
605 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
606 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
607 }
608 
609 /* ARGSUSED */
610 static void
611 synch_setup(void *dummy)
612 {
613 	callout_init(&loadav_callout, 1);
614 
615 	/* Kick off timeout driven events by calling first time. */
616 	loadav(NULL);
617 }
618 
619 int
620 should_yield(void)
621 {
622 
623 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
624 }
625 
626 void
627 maybe_yield(void)
628 {
629 
630 	if (should_yield())
631 		kern_yield(PRI_USER);
632 }
633 
634 void
635 kern_yield(int prio)
636 {
637 	struct thread *td;
638 
639 	td = curthread;
640 	DROP_GIANT();
641 	thread_lock(td);
642 	if (prio == PRI_USER)
643 		prio = td->td_user_pri;
644 	if (prio >= 0)
645 		sched_prio(td, prio);
646 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
647 	thread_unlock(td);
648 	PICKUP_GIANT();
649 }
650 
651 /*
652  * General purpose yield system call.
653  */
654 int
655 sys_yield(struct thread *td, struct yield_args *uap)
656 {
657 
658 	thread_lock(td);
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);
662 	thread_unlock(td);
663 	td->td_retval[0] = 0;
664 	return (0);
665 }
666