xref: /freebsd/sys/kern/kern_synch.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 uint64_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     "Fixed-point scale factor used for calculating load average values");
99 
100 static void	loadav(void *arg);
101 
102 SDT_PROVIDER_DECLARE(sched);
103 SDT_PROBE_DEFINE(sched, , , preempt);
104 
105 static void
106 sleepinit(void *unused)
107 {
108 
109 	hogticks = (hz / 10) * 2;	/* Default only. */
110 	init_sleepqueues();
111 }
112 
113 /*
114  * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
115  * it is available.
116  */
117 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
118 
119 /*
120  * General sleep call.  Suspends the current thread until a wakeup is
121  * performed on the specified identifier.  The thread will then be made
122  * runnable with the specified priority.  Sleeps at most sbt units of time
123  * (0 means no timeout).  If pri includes the PCATCH flag, let signals
124  * interrupt the sleep, otherwise ignore them while sleeping.  Returns 0 if
125  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
126  * signal becomes pending, ERESTART is returned if the current system
127  * call should be restarted if possible, and EINTR is returned if the system
128  * call should be interrupted by the signal (return EINTR).
129  *
130  * The lock argument is unlocked before the caller is suspended, and
131  * re-locked before _sleep() returns.  If priority includes the PDROP
132  * flag the lock is not re-locked before returning.
133  */
134 int
135 _sleep(const void *ident, struct lock_object *lock, int priority,
136     const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
137 {
138 	struct thread *td;
139 	struct lock_class *class;
140 	uintptr_t lock_state;
141 	int catch, pri, rval, sleepq_flags;
142 	WITNESS_SAVE_DECL(lock_witness);
143 
144 	TSENTER();
145 	td = curthread;
146 #ifdef KTRACE
147 	if (KTRPOINT(td, KTR_CSW))
148 		ktrcsw(1, 0, wmesg);
149 #endif
150 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
151 	    "Sleeping on \"%s\"", wmesg);
152 	KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL ||
153 	    (priority & PNOLOCK) != 0,
154 	    ("sleeping without a lock"));
155 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
156 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
157 	if (priority & PDROP)
158 		KASSERT(lock != NULL && lock != &Giant.lock_object,
159 		    ("PDROP requires a non-Giant lock"));
160 	if (lock != NULL)
161 		class = LOCK_CLASS(lock);
162 	else
163 		class = NULL;
164 
165 	if (SCHEDULER_STOPPED_TD(td)) {
166 		if (lock != NULL && priority & PDROP)
167 			class->lc_unlock(lock);
168 		return (0);
169 	}
170 	catch = priority & PCATCH;
171 	pri = priority & PRIMASK;
172 
173 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
174 
175 	if ((uintptr_t)ident >= (uintptr_t)&pause_wchan[0] &&
176 	    (uintptr_t)ident <= (uintptr_t)&pause_wchan[MAXCPU - 1])
177 		sleepq_flags = SLEEPQ_PAUSE;
178 	else
179 		sleepq_flags = SLEEPQ_SLEEP;
180 	if (catch)
181 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
182 
183 	sleepq_lock(ident);
184 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
185 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
186 
187 	if (lock == &Giant.lock_object)
188 		mtx_assert(&Giant, MA_OWNED);
189 	DROP_GIANT();
190 	if (lock != NULL && lock != &Giant.lock_object &&
191 	    !(class->lc_flags & LC_SLEEPABLE)) {
192 		KASSERT(!(class->lc_flags & LC_SPINLOCK),
193 		    ("spin locks can only use msleep_spin"));
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 	TSEXIT();
238 	return (rval);
239 }
240 
241 int
242 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
243     sbintime_t sbt, sbintime_t pr, int flags)
244 {
245 	struct thread *td;
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_TD(td))
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 	/*
274 	 * Can't call ktrace with any spin locks held so it can lock the
275 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
276 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
277 	 * we handle those requests.  This is safe since we have placed our
278 	 * thread on the sleep queue already.
279 	 */
280 #ifdef KTRACE
281 	if (KTRPOINT(td, KTR_CSW)) {
282 		sleepq_release(ident);
283 		ktrcsw(1, 0, wmesg);
284 		sleepq_lock(ident);
285 	}
286 #endif
287 #ifdef WITNESS
288 	sleepq_release(ident);
289 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
290 	    wmesg);
291 	sleepq_lock(ident);
292 #endif
293 	if (sbt != 0)
294 		rval = sleepq_timedwait(ident, 0);
295 	else {
296 		sleepq_wait(ident, 0);
297 		rval = 0;
298 	}
299 #ifdef KTRACE
300 	if (KTRPOINT(td, KTR_CSW))
301 		ktrcsw(0, 0, wmesg);
302 #endif
303 	PICKUP_GIANT();
304 	mtx_lock_spin(mtx);
305 	WITNESS_RESTORE(&mtx->lock_object, mtx);
306 	return (rval);
307 }
308 
309 /*
310  * pause_sbt() delays the calling thread by the given signed binary
311  * time. During cold bootup, pause_sbt() uses the DELAY() function
312  * instead of the _sleep() function to do the waiting. The "sbt"
313  * argument must be greater than or equal to zero. A "sbt" value of
314  * zero is equivalent to a "sbt" value of one tick.
315  */
316 int
317 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
318 {
319 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
320 
321 	/* silently convert invalid timeouts */
322 	if (sbt == 0)
323 		sbt = tick_sbt;
324 
325 	if ((cold && curthread == &thread0) || kdb_active ||
326 	    SCHEDULER_STOPPED()) {
327 		/*
328 		 * We delay one second at a time to avoid overflowing the
329 		 * system specific DELAY() function(s):
330 		 */
331 		while (sbt >= SBT_1S) {
332 			DELAY(1000000);
333 			sbt -= SBT_1S;
334 		}
335 		/* Do the delay remainder, if any */
336 		sbt = howmany(sbt, SBT_1US);
337 		if (sbt > 0)
338 			DELAY(sbt);
339 		return (EWOULDBLOCK);
340 	}
341 	return (_sleep(&pause_wchan[curcpu], NULL,
342 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
343 }
344 
345 /*
346  * Make all threads sleeping on the specified identifier runnable.
347  */
348 void
349 wakeup(const void *ident)
350 {
351 	int wakeup_swapper;
352 
353 	sleepq_lock(ident);
354 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
355 	sleepq_release(ident);
356 	if (wakeup_swapper) {
357 		KASSERT(ident != &proc0,
358 		    ("wakeup and wakeup_swapper and proc0"));
359 		kick_proc0();
360 	}
361 }
362 
363 /*
364  * Make a thread sleeping on the specified identifier runnable.
365  * May wake more than one thread if a target thread is currently
366  * swapped out.
367  */
368 void
369 wakeup_one(const void *ident)
370 {
371 	int wakeup_swapper;
372 
373 	sleepq_lock(ident);
374 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_DROP, 0, 0);
375 	if (wakeup_swapper)
376 		kick_proc0();
377 }
378 
379 void
380 wakeup_any(const void *ident)
381 {
382 	int wakeup_swapper;
383 
384 	sleepq_lock(ident);
385 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR |
386 	    SLEEPQ_DROP, 0, 0);
387 	if (wakeup_swapper)
388 		kick_proc0();
389 }
390 
391 /*
392  * Signal sleeping waiters after the counter has reached zero.
393  */
394 void
395 _blockcount_wakeup(blockcount_t *bc, u_int old)
396 {
397 
398 	KASSERT(_BLOCKCOUNT_WAITERS(old),
399 	    ("%s: no waiters on %p", __func__, bc));
400 
401 	if (atomic_cmpset_int(&bc->__count, _BLOCKCOUNT_WAITERS_FLAG, 0))
402 		wakeup(bc);
403 }
404 
405 /*
406  * Wait for a wakeup or a signal.  This does not guarantee that the count is
407  * still zero on return.  Callers wanting a precise answer should use
408  * blockcount_wait() with an interlock.
409  *
410  * If there is no work to wait for, return 0.  If the sleep was interrupted by a
411  * signal, return EINTR or ERESTART, and return EAGAIN otherwise.
412  */
413 int
414 _blockcount_sleep(blockcount_t *bc, struct lock_object *lock, const char *wmesg,
415     int prio)
416 {
417 	void *wchan;
418 	uintptr_t lock_state;
419 	u_int old;
420 	int ret;
421 	bool catch, drop;
422 
423 	KASSERT(lock != &Giant.lock_object,
424 	    ("%s: cannot use Giant as the interlock", __func__));
425 
426 	catch = (prio & PCATCH) != 0;
427 	drop = (prio & PDROP) != 0;
428 	prio &= PRIMASK;
429 
430 	/*
431 	 * Synchronize with the fence in blockcount_release().  If we end up
432 	 * waiting, the sleepqueue lock acquisition will provide the required
433 	 * side effects.
434 	 *
435 	 * If there is no work to wait for, but waiters are present, try to put
436 	 * ourselves to sleep to avoid jumping ahead.
437 	 */
438 	if (atomic_load_acq_int(&bc->__count) == 0) {
439 		if (lock != NULL && drop)
440 			LOCK_CLASS(lock)->lc_unlock(lock);
441 		return (0);
442 	}
443 	lock_state = 0;
444 	wchan = bc;
445 	sleepq_lock(wchan);
446 	DROP_GIANT();
447 	if (lock != NULL)
448 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
449 	old = blockcount_read(bc);
450 	ret = 0;
451 	do {
452 		if (_BLOCKCOUNT_COUNT(old) == 0) {
453 			sleepq_release(wchan);
454 			goto out;
455 		}
456 		if (_BLOCKCOUNT_WAITERS(old))
457 			break;
458 	} while (!atomic_fcmpset_int(&bc->__count, &old,
459 	    old | _BLOCKCOUNT_WAITERS_FLAG));
460 	sleepq_add(wchan, NULL, wmesg, catch ? SLEEPQ_INTERRUPTIBLE : 0, 0);
461 	if (catch)
462 		ret = sleepq_wait_sig(wchan, prio);
463 	else
464 		sleepq_wait(wchan, prio);
465 	if (ret == 0)
466 		ret = EAGAIN;
467 
468 out:
469 	PICKUP_GIANT();
470 	if (lock != NULL && !drop)
471 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
472 
473 	return (ret);
474 }
475 
476 static void
477 kdb_switch(void)
478 {
479 	thread_unlock(curthread);
480 	kdb_backtrace();
481 	kdb_reenter();
482 	panic("%s: did not reenter debugger", __func__);
483 }
484 
485 /*
486  * The machine independent parts of context switching.
487  *
488  * The thread lock is required on entry and is no longer held on return.
489  */
490 void
491 mi_switch(int flags)
492 {
493 	uint64_t runtime, new_switchtime;
494 	struct thread *td;
495 
496 	td = curthread;			/* XXX */
497 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
498 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
499 #ifdef INVARIANTS
500 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
501 		mtx_assert(&Giant, MA_NOTOWNED);
502 #endif
503 	KASSERT(td->td_critnest == 1 || KERNEL_PANICKED(),
504 		("mi_switch: switch in a critical section"));
505 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
506 	    ("mi_switch: switch must be voluntary or involuntary"));
507 
508 	/*
509 	 * Don't perform context switches from the debugger.
510 	 */
511 	if (kdb_active)
512 		kdb_switch();
513 	if (SCHEDULER_STOPPED_TD(td))
514 		return;
515 	if (flags & SW_VOL) {
516 		td->td_ru.ru_nvcsw++;
517 		td->td_swvoltick = ticks;
518 	} else {
519 		td->td_ru.ru_nivcsw++;
520 		td->td_swinvoltick = ticks;
521 	}
522 #ifdef SCHED_STATS
523 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
524 #endif
525 	/*
526 	 * Compute the amount of time during which the current
527 	 * thread was running, and add that to its total so far.
528 	 */
529 	new_switchtime = cpu_ticks();
530 	runtime = new_switchtime - PCPU_GET(switchtime);
531 	td->td_runtime += runtime;
532 	td->td_incruntime += runtime;
533 	PCPU_SET(switchtime, new_switchtime);
534 	td->td_generation++;	/* bump preempt-detect counter */
535 	VM_CNT_INC(v_swtch);
536 	PCPU_SET(switchticks, ticks);
537 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
538 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
539 #ifdef KDTRACE_HOOKS
540 	if (SDT_PROBES_ENABLED() &&
541 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
542 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
543 		SDT_PROBE0(sched, , , preempt);
544 #endif
545 	sched_switch(td, flags);
546 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
547 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
548 
549 	/*
550 	 * If the last thread was exiting, finish cleaning it up.
551 	 */
552 	if ((td = PCPU_GET(deadthread))) {
553 		PCPU_SET(deadthread, NULL);
554 		thread_stash(td);
555 	}
556 	spinlock_exit();
557 }
558 
559 /*
560  * Change thread state to be runnable, placing it on the run queue if
561  * it is in memory.  If it is swapped out, return true so our caller
562  * will know to awaken the swapper.
563  *
564  * Requires the thread lock on entry, drops on exit.
565  */
566 int
567 setrunnable(struct thread *td, int srqflags)
568 {
569 	int swapin;
570 
571 	THREAD_LOCK_ASSERT(td, MA_OWNED);
572 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
573 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
574 
575 	swapin = 0;
576 	switch (TD_GET_STATE(td)) {
577 	case TDS_RUNNING:
578 	case TDS_RUNQ:
579 		break;
580 	case TDS_CAN_RUN:
581 		KASSERT((td->td_flags & TDF_INMEM) != 0,
582 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
583 		    td, td->td_flags, td->td_inhibitors));
584 		/* unlocks thread lock according to flags */
585 		sched_wakeup(td, srqflags);
586 		return (0);
587 	case TDS_INHIBITED:
588 		/*
589 		 * If we are only inhibited because we are swapped out
590 		 * arrange to swap in this process.
591 		 */
592 		if (td->td_inhibitors == TDI_SWAPPED &&
593 		    (td->td_flags & TDF_SWAPINREQ) == 0) {
594 			td->td_flags |= TDF_SWAPINREQ;
595 			swapin = 1;
596 		}
597 		break;
598 	default:
599 		panic("setrunnable: state 0x%x", TD_GET_STATE(td));
600 	}
601 	if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
602 		thread_unlock(td);
603 
604 	return (swapin);
605 }
606 
607 /*
608  * Compute a tenex style load average of a quantity on
609  * 1, 5 and 15 minute intervals.
610  */
611 static void
612 loadav(void *arg)
613 {
614 	int i;
615 	uint64_t nrun;
616 	struct loadavg *avg;
617 
618 	nrun = (uint64_t)sched_load();
619 	avg = &averunnable;
620 
621 	for (i = 0; i < 3; i++)
622 		avg->ldavg[i] = (cexp[i] * (uint64_t)avg->ldavg[i] +
623 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
624 
625 	/*
626 	 * Schedule the next update to occur after 5 seconds, but add a
627 	 * random variation to avoid synchronisation with processes that
628 	 * run at regular intervals.
629 	 */
630 	callout_reset_sbt(&loadav_callout,
631 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
632 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
633 }
634 
635 static void
636 ast_scheduler(struct thread *td, int tda __unused)
637 {
638 #ifdef KTRACE
639 	if (KTRPOINT(td, KTR_CSW))
640 		ktrcsw(1, 1, __func__);
641 #endif
642 	thread_lock(td);
643 	sched_prio(td, td->td_user_pri);
644 	mi_switch(SW_INVOL | SWT_NEEDRESCHED);
645 #ifdef KTRACE
646 	if (KTRPOINT(td, KTR_CSW))
647 		ktrcsw(0, 1, __func__);
648 #endif
649 }
650 
651 static void
652 synch_setup(void *dummy __unused)
653 {
654 	callout_init(&loadav_callout, 1);
655 	ast_register(TDA_SCHED, ASTR_ASTF_REQUIRED, 0, ast_scheduler);
656 
657 	/* Kick off timeout driven events by calling first time. */
658 	loadav(NULL);
659 }
660 
661 int
662 should_yield(void)
663 {
664 
665 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
666 }
667 
668 void
669 maybe_yield(void)
670 {
671 
672 	if (should_yield())
673 		kern_yield(PRI_USER);
674 }
675 
676 void
677 kern_yield(int prio)
678 {
679 	struct thread *td;
680 
681 	td = curthread;
682 	DROP_GIANT();
683 	thread_lock(td);
684 	if (prio == PRI_USER)
685 		prio = td->td_user_pri;
686 	if (prio >= 0)
687 		sched_prio(td, prio);
688 	mi_switch(SW_VOL | SWT_RELINQUISH);
689 	PICKUP_GIANT();
690 }
691 
692 /*
693  * General purpose yield system call.
694  */
695 int
696 sys_yield(struct thread *td, struct yield_args *uap)
697 {
698 
699 	thread_lock(td);
700 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
701 		sched_prio(td, PRI_MAX_TIMESHARE);
702 	mi_switch(SW_VOL | SWT_RELINQUISH);
703 	td->td_retval[0] = 0;
704 	return (0);
705 }
706 
707 int
708 sys_sched_getcpu(struct thread *td, struct sched_getcpu_args *uap)
709 {
710 	td->td_retval[0] = td->td_oncpu;
711 	return (0);
712 }
713