xref: /freebsd/sys/kern/kern_synch.c (revision c99d0c5801ce22a9976e491d04ce7e1b8996dfdd)
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/blockcount.h>
48 #include <sys/condvar.h>
49 #include <sys/kdb.h>
50 #include <sys/kernel.h>
51 #include <sys/ktr.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resourcevar.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 const char 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(const 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 	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_TD(td)) {
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 		WITNESS_SAVE(lock, lock_witness);
190 		lock_state = class->lc_unlock(lock);
191 	} else
192 		/* GCC needs to follow the Yellow Brick Road */
193 		lock_state = -1;
194 
195 	/*
196 	 * We put ourselves on the sleep queue and start our timeout
197 	 * before calling thread_suspend_check, as we could stop there,
198 	 * and a wakeup or a SIGCONT (or both) could occur while we were
199 	 * stopped without resuming us.  Thus, we must be ready for sleep
200 	 * when cursig() is called.  If the wakeup happens while we're
201 	 * stopped, then td will no longer be on a sleep queue upon
202 	 * return from cursig().
203 	 */
204 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
205 	if (sbt != 0)
206 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
207 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
208 		sleepq_release(ident);
209 		WITNESS_SAVE(lock, lock_witness);
210 		lock_state = class->lc_unlock(lock);
211 		sleepq_lock(ident);
212 	}
213 	if (sbt != 0 && catch)
214 		rval = sleepq_timedwait_sig(ident, pri);
215 	else if (sbt != 0)
216 		rval = sleepq_timedwait(ident, pri);
217 	else if (catch)
218 		rval = sleepq_wait_sig(ident, pri);
219 	else {
220 		sleepq_wait(ident, pri);
221 		rval = 0;
222 	}
223 #ifdef KTRACE
224 	if (KTRPOINT(td, KTR_CSW))
225 		ktrcsw(0, 0, wmesg);
226 #endif
227 	PICKUP_GIANT();
228 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
229 		class->lc_lock(lock, lock_state);
230 		WITNESS_RESTORE(lock, lock_witness);
231 	}
232 	return (rval);
233 }
234 
235 int
236 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
237     sbintime_t sbt, sbintime_t pr, int flags)
238 {
239 	struct thread *td;
240 	int rval;
241 	WITNESS_SAVE_DECL(mtx);
242 
243 	td = curthread;
244 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
245 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
246 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
247 
248 	if (SCHEDULER_STOPPED_TD(td))
249 		return (0);
250 
251 	sleepq_lock(ident);
252 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
253 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
254 
255 	DROP_GIANT();
256 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
257 	WITNESS_SAVE(&mtx->lock_object, mtx);
258 	mtx_unlock_spin(mtx);
259 
260 	/*
261 	 * We put ourselves on the sleep queue and start our timeout.
262 	 */
263 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
264 	if (sbt != 0)
265 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
266 
267 	/*
268 	 * Can't call ktrace with any spin locks held so it can lock the
269 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
270 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
271 	 * we handle those requests.  This is safe since we have placed our
272 	 * thread on the sleep queue already.
273 	 */
274 #ifdef KTRACE
275 	if (KTRPOINT(td, KTR_CSW)) {
276 		sleepq_release(ident);
277 		ktrcsw(1, 0, wmesg);
278 		sleepq_lock(ident);
279 	}
280 #endif
281 #ifdef WITNESS
282 	sleepq_release(ident);
283 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
284 	    wmesg);
285 	sleepq_lock(ident);
286 #endif
287 	if (sbt != 0)
288 		rval = sleepq_timedwait(ident, 0);
289 	else {
290 		sleepq_wait(ident, 0);
291 		rval = 0;
292 	}
293 #ifdef KTRACE
294 	if (KTRPOINT(td, KTR_CSW))
295 		ktrcsw(0, 0, wmesg);
296 #endif
297 	PICKUP_GIANT();
298 	mtx_lock_spin(mtx);
299 	WITNESS_RESTORE(&mtx->lock_object, mtx);
300 	return (rval);
301 }
302 
303 /*
304  * pause_sbt() delays the calling thread by the given signed binary
305  * time. During cold bootup, pause_sbt() uses the DELAY() function
306  * instead of the _sleep() function to do the waiting. The "sbt"
307  * argument must be greater than or equal to zero. A "sbt" value of
308  * zero is equivalent to a "sbt" value of one tick.
309  */
310 int
311 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
312 {
313 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
314 
315 	/* silently convert invalid timeouts */
316 	if (sbt == 0)
317 		sbt = tick_sbt;
318 
319 	if ((cold && curthread == &thread0) || kdb_active ||
320 	    SCHEDULER_STOPPED()) {
321 		/*
322 		 * We delay one second at a time to avoid overflowing the
323 		 * system specific DELAY() function(s):
324 		 */
325 		while (sbt >= SBT_1S) {
326 			DELAY(1000000);
327 			sbt -= SBT_1S;
328 		}
329 		/* Do the delay remainder, if any */
330 		sbt = howmany(sbt, SBT_1US);
331 		if (sbt > 0)
332 			DELAY(sbt);
333 		return (EWOULDBLOCK);
334 	}
335 	return (_sleep(&pause_wchan[curcpu], NULL,
336 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
337 }
338 
339 /*
340  * Make all threads sleeping on the specified identifier runnable.
341  */
342 void
343 wakeup(const void *ident)
344 {
345 	int wakeup_swapper;
346 
347 	sleepq_lock(ident);
348 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
349 	sleepq_release(ident);
350 	if (wakeup_swapper) {
351 		KASSERT(ident != &proc0,
352 		    ("wakeup and wakeup_swapper and proc0"));
353 		kick_proc0();
354 	}
355 }
356 
357 /*
358  * Make a thread sleeping on the specified identifier runnable.
359  * May wake more than one thread if a target thread is currently
360  * swapped out.
361  */
362 void
363 wakeup_one(const void *ident)
364 {
365 	int wakeup_swapper;
366 
367 	sleepq_lock(ident);
368 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
369 	sleepq_release(ident);
370 	if (wakeup_swapper)
371 		kick_proc0();
372 }
373 
374 void
375 wakeup_any(const void *ident)
376 {
377 	int wakeup_swapper;
378 
379 	sleepq_lock(ident);
380 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR,
381 	    0, 0);
382 	sleepq_release(ident);
383 	if (wakeup_swapper)
384 		kick_proc0();
385 }
386 
387 /*
388  * Signal sleeping waiters after the counter has reached zero.
389  */
390 void
391 _blockcount_wakeup(blockcount_t *bc, u_int old)
392 {
393 
394 	KASSERT(_BLOCKCOUNT_WAITERS(old),
395 	    ("%s: no waiters on %p", __func__, bc));
396 
397 	if (atomic_cmpset_int(&bc->__count, _BLOCKCOUNT_WAITERS_FLAG, 0))
398 		wakeup(bc);
399 }
400 
401 /*
402  * Wait for a wakeup.  This does not guarantee that the count is still zero on
403  * return and may be subject to transient wakeups.  Callers wanting a precise
404  * answer should use blockcount_wait() with an interlock.
405  *
406  * Return 0 if there is no work to wait for, and 1 if we slept waiting for work
407  * to complete.  In the latter case the counter value must be re-read.
408  */
409 int
410 _blockcount_sleep(blockcount_t *bc, struct lock_object *lock, const char *wmesg,
411     int prio)
412 {
413 	void *wchan;
414 	uintptr_t lock_state;
415 	u_int old;
416 	int ret;
417 
418 	KASSERT(lock != &Giant.lock_object,
419 	    ("%s: cannot use Giant as the interlock", __func__));
420 
421 	/*
422 	 * Synchronize with the fence in blockcount_release().  If we end up
423 	 * waiting, the sleepqueue lock acquisition will provide the required
424 	 * side effects.
425 	 *
426 	 * If there is no work to wait for, but waiters are present, try to put
427 	 * ourselves to sleep to avoid jumping ahead.
428 	 */
429 	if (atomic_load_acq_int(&bc->__count) == 0) {
430 		if (lock != NULL && (prio & PDROP) != 0)
431 			LOCK_CLASS(lock)->lc_unlock(lock);
432 		return (0);
433 	}
434 	lock_state = 0;
435 	wchan = bc;
436 	sleepq_lock(wchan);
437 	DROP_GIANT();
438 	if (lock != NULL)
439 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
440 	old = blockcount_read(bc);
441 	do {
442 		if (_BLOCKCOUNT_COUNT(old) == 0) {
443 			sleepq_release(wchan);
444 			ret = 0;
445 			goto out;
446 		}
447 		if (_BLOCKCOUNT_WAITERS(old))
448 			break;
449 	} while (!atomic_fcmpset_int(&bc->__count, &old,
450 	    old | _BLOCKCOUNT_WAITERS_FLAG));
451 	sleepq_add(wchan, NULL, wmesg, 0, 0);
452 	sleepq_wait(wchan, prio);
453 	ret = 1;
454 
455 out:
456 	PICKUP_GIANT();
457 	if (lock != NULL && (prio & PDROP) == 0)
458 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
459 
460 	return (ret);
461 }
462 
463 static void
464 kdb_switch(void)
465 {
466 	thread_unlock(curthread);
467 	kdb_backtrace();
468 	kdb_reenter();
469 	panic("%s: did not reenter debugger", __func__);
470 }
471 
472 /*
473  * The machine independent parts of context switching.
474  *
475  * The thread lock is required on entry and is no longer held on return.
476  */
477 void
478 mi_switch(int flags)
479 {
480 	uint64_t runtime, new_switchtime;
481 	struct thread *td;
482 
483 	td = curthread;			/* XXX */
484 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
485 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
486 #ifdef INVARIANTS
487 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
488 		mtx_assert(&Giant, MA_NOTOWNED);
489 #endif
490 	KASSERT(td->td_critnest == 1 || KERNEL_PANICKED(),
491 		("mi_switch: switch in a critical section"));
492 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
493 	    ("mi_switch: switch must be voluntary or involuntary"));
494 
495 	/*
496 	 * Don't perform context switches from the debugger.
497 	 */
498 	if (kdb_active)
499 		kdb_switch();
500 	if (SCHEDULER_STOPPED_TD(td))
501 		return;
502 	if (flags & SW_VOL) {
503 		td->td_ru.ru_nvcsw++;
504 		td->td_swvoltick = ticks;
505 	} else {
506 		td->td_ru.ru_nivcsw++;
507 		td->td_swinvoltick = ticks;
508 	}
509 #ifdef SCHED_STATS
510 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
511 #endif
512 	/*
513 	 * Compute the amount of time during which the current
514 	 * thread was running, and add that to its total so far.
515 	 */
516 	new_switchtime = cpu_ticks();
517 	runtime = new_switchtime - PCPU_GET(switchtime);
518 	td->td_runtime += runtime;
519 	td->td_incruntime += runtime;
520 	PCPU_SET(switchtime, new_switchtime);
521 	td->td_generation++;	/* bump preempt-detect counter */
522 	VM_CNT_INC(v_swtch);
523 	PCPU_SET(switchticks, ticks);
524 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
525 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
526 #ifdef KDTRACE_HOOKS
527 	if (SDT_PROBES_ENABLED() &&
528 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
529 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
530 		SDT_PROBE0(sched, , , preempt);
531 #endif
532 	sched_switch(td, flags);
533 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
534 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
535 
536 	/*
537 	 * If the last thread was exiting, finish cleaning it up.
538 	 */
539 	if ((td = PCPU_GET(deadthread))) {
540 		PCPU_SET(deadthread, NULL);
541 		thread_stash(td);
542 	}
543 	spinlock_exit();
544 }
545 
546 /*
547  * Change thread state to be runnable, placing it on the run queue if
548  * it is in memory.  If it is swapped out, return true so our caller
549  * will know to awaken the swapper.
550  *
551  * Requires the thread lock on entry, drops on exit.
552  */
553 int
554 setrunnable(struct thread *td, int srqflags)
555 {
556 	int swapin;
557 
558 	THREAD_LOCK_ASSERT(td, MA_OWNED);
559 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
560 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
561 
562 	swapin = 0;
563 	switch (td->td_state) {
564 	case TDS_RUNNING:
565 	case TDS_RUNQ:
566 		break;
567 	case TDS_CAN_RUN:
568 		KASSERT((td->td_flags & TDF_INMEM) != 0,
569 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
570 		    td, td->td_flags, td->td_inhibitors));
571 		/* unlocks thread lock according to flags */
572 		sched_wakeup(td, srqflags);
573 		return (0);
574 	case TDS_INHIBITED:
575 		/*
576 		 * If we are only inhibited because we are swapped out
577 		 * arrange to swap in this process.
578 		 */
579 		if (td->td_inhibitors == TDI_SWAPPED &&
580 		    (td->td_flags & TDF_SWAPINREQ) == 0) {
581 			td->td_flags |= TDF_SWAPINREQ;
582 			swapin = 1;
583 		}
584 		break;
585 	default:
586 		panic("setrunnable: state 0x%x", td->td_state);
587 	}
588 	if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
589 		thread_unlock(td);
590 
591 	return (swapin);
592 }
593 
594 /*
595  * Compute a tenex style load average of a quantity on
596  * 1, 5 and 15 minute intervals.
597  */
598 static void
599 loadav(void *arg)
600 {
601 	int i, nrun;
602 	struct loadavg *avg;
603 
604 	nrun = sched_load();
605 	avg = &averunnable;
606 
607 	for (i = 0; i < 3; i++)
608 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
609 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
610 
611 	/*
612 	 * Schedule the next update to occur after 5 seconds, but add a
613 	 * random variation to avoid synchronisation with processes that
614 	 * run at regular intervals.
615 	 */
616 	callout_reset_sbt(&loadav_callout,
617 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
618 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
619 }
620 
621 /* ARGSUSED */
622 static void
623 synch_setup(void *dummy)
624 {
625 	callout_init(&loadav_callout, 1);
626 
627 	/* Kick off timeout driven events by calling first time. */
628 	loadav(NULL);
629 }
630 
631 int
632 should_yield(void)
633 {
634 
635 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
636 }
637 
638 void
639 maybe_yield(void)
640 {
641 
642 	if (should_yield())
643 		kern_yield(PRI_USER);
644 }
645 
646 void
647 kern_yield(int prio)
648 {
649 	struct thread *td;
650 
651 	td = curthread;
652 	DROP_GIANT();
653 	thread_lock(td);
654 	if (prio == PRI_USER)
655 		prio = td->td_user_pri;
656 	if (prio >= 0)
657 		sched_prio(td, prio);
658 	mi_switch(SW_VOL | SWT_RELINQUISH);
659 	PICKUP_GIANT();
660 }
661 
662 /*
663  * General purpose yield system call.
664  */
665 int
666 sys_yield(struct thread *td, struct yield_args *uap)
667 {
668 
669 	thread_lock(td);
670 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
671 		sched_prio(td, PRI_MAX_TIMESHARE);
672 	mi_switch(SW_VOL | SWT_RELINQUISH);
673 	td->td_retval[0] = 0;
674 	return (0);
675 }
676