xref: /freebsd/sys/kern/kern_synch.c (revision 783d3ff6d7fae619db8a7990b8a6387de0c677b5)
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 	uintptr_t lock_state;
137 	int catch, pri, rval, sleepq_flags;
138 	WITNESS_SAVE_DECL(lock_witness);
139 
140 	TSENTER();
141 	td = curthread;
142 #ifdef KTRACE
143 	if (KTRPOINT(td, KTR_CSW))
144 		ktrcsw(1, 0, wmesg);
145 #endif
146 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
147 	    "Sleeping on \"%s\"", wmesg);
148 	KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL ||
149 	    (priority & PNOLOCK) != 0,
150 	    ("sleeping without a lock"));
151 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
152 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
153 	if (priority & PDROP)
154 		KASSERT(lock != NULL && lock != &Giant.lock_object,
155 		    ("PDROP requires a non-Giant lock"));
156 	if (lock != NULL)
157 		class = LOCK_CLASS(lock);
158 	else
159 		class = NULL;
160 
161 	if (SCHEDULER_STOPPED()) {
162 		if (lock != NULL && priority & PDROP)
163 			class->lc_unlock(lock);
164 		return (0);
165 	}
166 	catch = priority & PCATCH;
167 	pri = priority & PRIMASK;
168 
169 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
170 
171 	if ((uintptr_t)ident >= (uintptr_t)&pause_wchan[0] &&
172 	    (uintptr_t)ident <= (uintptr_t)&pause_wchan[MAXCPU - 1])
173 		sleepq_flags = SLEEPQ_PAUSE;
174 	else
175 		sleepq_flags = SLEEPQ_SLEEP;
176 	if (catch)
177 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
178 
179 	sleepq_lock(ident);
180 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
181 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
182 
183 	if (lock == &Giant.lock_object)
184 		mtx_assert(&Giant, MA_OWNED);
185 	DROP_GIANT();
186 	if (lock != NULL && lock != &Giant.lock_object &&
187 	    !(class->lc_flags & LC_SLEEPABLE)) {
188 		KASSERT(!(class->lc_flags & LC_SPINLOCK),
189 		    ("spin locks can only use msleep_spin"));
190 		WITNESS_SAVE(lock, lock_witness);
191 		lock_state = class->lc_unlock(lock);
192 	} else
193 		/* GCC needs to follow the Yellow Brick Road */
194 		lock_state = -1;
195 
196 	/*
197 	 * We put ourselves on the sleep queue and start our timeout
198 	 * before calling thread_suspend_check, as we could stop there,
199 	 * and a wakeup or a SIGCONT (or both) could occur while we were
200 	 * stopped without resuming us.  Thus, we must be ready for sleep
201 	 * when cursig() is called.  If the wakeup happens while we're
202 	 * stopped, then td will no longer be on a sleep queue upon
203 	 * return from cursig().
204 	 */
205 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
206 	if (sbt != 0)
207 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
208 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
209 		sleepq_release(ident);
210 		WITNESS_SAVE(lock, lock_witness);
211 		lock_state = class->lc_unlock(lock);
212 		sleepq_lock(ident);
213 	}
214 	if (sbt != 0 && catch)
215 		rval = sleepq_timedwait_sig(ident, pri);
216 	else if (sbt != 0)
217 		rval = sleepq_timedwait(ident, pri);
218 	else if (catch)
219 		rval = sleepq_wait_sig(ident, pri);
220 	else {
221 		sleepq_wait(ident, pri);
222 		rval = 0;
223 	}
224 #ifdef KTRACE
225 	if (KTRPOINT(td, KTR_CSW))
226 		ktrcsw(0, 0, wmesg);
227 #endif
228 	PICKUP_GIANT();
229 	if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
230 		class->lc_lock(lock, lock_state);
231 		WITNESS_RESTORE(lock, lock_witness);
232 	}
233 	TSEXIT();
234 	return (rval);
235 }
236 
237 int
238 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
239     sbintime_t sbt, sbintime_t pr, int flags)
240 {
241 	struct thread *td __ktrace_used;
242 	int rval;
243 	WITNESS_SAVE_DECL(mtx);
244 
245 	td = curthread;
246 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
247 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
248 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
249 
250 	if (SCHEDULER_STOPPED())
251 		return (0);
252 
253 	sleepq_lock(ident);
254 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
255 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
256 
257 	DROP_GIANT();
258 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
259 	WITNESS_SAVE(&mtx->lock_object, mtx);
260 	mtx_unlock_spin(mtx);
261 
262 	/*
263 	 * We put ourselves on the sleep queue and start our timeout.
264 	 */
265 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
266 	if (sbt != 0)
267 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
268 
269 	/*
270 	 * Can't call ktrace with any spin locks held so it can lock the
271 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
272 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
273 	 * we handle those requests.  This is safe since we have placed our
274 	 * thread on the sleep queue already.
275 	 */
276 #ifdef KTRACE
277 	if (KTRPOINT(td, KTR_CSW)) {
278 		sleepq_release(ident);
279 		ktrcsw(1, 0, wmesg);
280 		sleepq_lock(ident);
281 	}
282 #endif
283 #ifdef WITNESS
284 	sleepq_release(ident);
285 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
286 	    wmesg);
287 	sleepq_lock(ident);
288 #endif
289 	if (sbt != 0)
290 		rval = sleepq_timedwait(ident, 0);
291 	else {
292 		sleepq_wait(ident, 0);
293 		rval = 0;
294 	}
295 #ifdef KTRACE
296 	if (KTRPOINT(td, KTR_CSW))
297 		ktrcsw(0, 0, wmesg);
298 #endif
299 	PICKUP_GIANT();
300 	mtx_lock_spin(mtx);
301 	WITNESS_RESTORE(&mtx->lock_object, mtx);
302 	return (rval);
303 }
304 
305 /*
306  * pause_sbt() delays the calling thread by the given signed binary
307  * time. During cold bootup, pause_sbt() uses the DELAY() function
308  * instead of the _sleep() function to do the waiting. The "sbt"
309  * argument must be greater than or equal to zero. A "sbt" value of
310  * zero is equivalent to a "sbt" value of one tick.
311  */
312 int
313 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
314 {
315 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
316 
317 	/* silently convert invalid timeouts */
318 	if (sbt == 0)
319 		sbt = tick_sbt;
320 
321 	if ((cold && curthread == &thread0) || kdb_active ||
322 	    SCHEDULER_STOPPED()) {
323 		/*
324 		 * We delay one second at a time to avoid overflowing the
325 		 * system specific DELAY() function(s):
326 		 */
327 		while (sbt >= SBT_1S) {
328 			DELAY(1000000);
329 			sbt -= SBT_1S;
330 		}
331 		/* Do the delay remainder, if any */
332 		sbt = howmany(sbt, SBT_1US);
333 		if (sbt > 0)
334 			DELAY(sbt);
335 		return (EWOULDBLOCK);
336 	}
337 	return (_sleep(&pause_wchan[curcpu], NULL,
338 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
339 }
340 
341 /*
342  * Make all threads sleeping on the specified identifier runnable.
343  */
344 void
345 wakeup(const void *ident)
346 {
347 	int wakeup_swapper;
348 
349 	sleepq_lock(ident);
350 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
351 	sleepq_release(ident);
352 	if (wakeup_swapper) {
353 		KASSERT(ident != &proc0,
354 		    ("wakeup and wakeup_swapper and proc0"));
355 		kick_proc0();
356 	}
357 }
358 
359 /*
360  * Make a thread sleeping on the specified identifier runnable.
361  * May wake more than one thread if a target thread is currently
362  * swapped out.
363  */
364 void
365 wakeup_one(const void *ident)
366 {
367 	int wakeup_swapper;
368 
369 	sleepq_lock(ident);
370 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_DROP, 0, 0);
371 	if (wakeup_swapper)
372 		kick_proc0();
373 }
374 
375 void
376 wakeup_any(const void *ident)
377 {
378 	int wakeup_swapper;
379 
380 	sleepq_lock(ident);
381 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR |
382 	    SLEEPQ_DROP, 0, 0);
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 or a signal.  This does not guarantee that the count is
403  * still zero on return.  Callers wanting a precise answer should use
404  * blockcount_wait() with an interlock.
405  *
406  * If there is no work to wait for, return 0.  If the sleep was interrupted by a
407  * signal, return EINTR or ERESTART, and return EAGAIN otherwise.
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 	bool catch, drop;
418 
419 	KASSERT(lock != &Giant.lock_object,
420 	    ("%s: cannot use Giant as the interlock", __func__));
421 
422 	catch = (prio & PCATCH) != 0;
423 	drop = (prio & PDROP) != 0;
424 	prio &= PRIMASK;
425 
426 	/*
427 	 * Synchronize with the fence in blockcount_release().  If we end up
428 	 * waiting, the sleepqueue lock acquisition will provide the required
429 	 * side effects.
430 	 *
431 	 * If there is no work to wait for, but waiters are present, try to put
432 	 * ourselves to sleep to avoid jumping ahead.
433 	 */
434 	if (atomic_load_acq_int(&bc->__count) == 0) {
435 		if (lock != NULL && drop)
436 			LOCK_CLASS(lock)->lc_unlock(lock);
437 		return (0);
438 	}
439 	lock_state = 0;
440 	wchan = bc;
441 	sleepq_lock(wchan);
442 	DROP_GIANT();
443 	if (lock != NULL)
444 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
445 	old = blockcount_read(bc);
446 	ret = 0;
447 	do {
448 		if (_BLOCKCOUNT_COUNT(old) == 0) {
449 			sleepq_release(wchan);
450 			goto out;
451 		}
452 		if (_BLOCKCOUNT_WAITERS(old))
453 			break;
454 	} while (!atomic_fcmpset_int(&bc->__count, &old,
455 	    old | _BLOCKCOUNT_WAITERS_FLAG));
456 	sleepq_add(wchan, NULL, wmesg, catch ? SLEEPQ_INTERRUPTIBLE : 0, 0);
457 	if (catch)
458 		ret = sleepq_wait_sig(wchan, prio);
459 	else
460 		sleepq_wait(wchan, prio);
461 	if (ret == 0)
462 		ret = EAGAIN;
463 
464 out:
465 	PICKUP_GIANT();
466 	if (lock != NULL && !drop)
467 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
468 
469 	return (ret);
470 }
471 
472 static void
473 kdb_switch(void)
474 {
475 	thread_unlock(curthread);
476 	kdb_backtrace();
477 	kdb_reenter();
478 	panic("%s: did not reenter debugger", __func__);
479 }
480 
481 /*
482  * mi_switch(9): The machine-independent parts of context switching.
483  *
484  * The thread lock is required on entry and is no longer held on return.
485  */
486 void
487 mi_switch(int flags)
488 {
489 	uint64_t runtime, new_switchtime;
490 	struct thread *td;
491 
492 	td = curthread;			/* XXX */
493 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
494 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
495 #ifdef INVARIANTS
496 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
497 		mtx_assert(&Giant, MA_NOTOWNED);
498 #endif
499 	/* thread_lock() performs spinlock_enter(). */
500 	KASSERT(td->td_critnest == 1 || KERNEL_PANICKED(),
501 	    ("mi_switch: switch in a critical section"));
502 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
503 	    ("mi_switch: switch must be voluntary or involuntary"));
504 	KASSERT((flags & SW_TYPE_MASK) != 0,
505 	    ("mi_switch: a switch reason (type) must be specified"));
506 	KASSERT((flags & SW_TYPE_MASK) < SWT_COUNT,
507 	    ("mi_switch: invalid switch reason %d", (flags & SW_TYPE_MASK)));
508 
509 	/*
510 	 * Don't perform context switches from the debugger.
511 	 */
512 	if (kdb_active)
513 		kdb_switch();
514 	if (SCHEDULER_STOPPED())
515 		return;
516 	if (flags & SW_VOL) {
517 		td->td_ru.ru_nvcsw++;
518 		td->td_swvoltick = ticks;
519 	} else {
520 		td->td_ru.ru_nivcsw++;
521 		td->td_swinvoltick = ticks;
522 	}
523 #ifdef SCHED_STATS
524 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
525 #endif
526 	/*
527 	 * Compute the amount of time during which the current
528 	 * thread was running, and add that to its total so far.
529 	 */
530 	new_switchtime = cpu_ticks();
531 	runtime = new_switchtime - PCPU_GET(switchtime);
532 	td->td_runtime += runtime;
533 	td->td_incruntime += runtime;
534 	PCPU_SET(switchtime, new_switchtime);
535 	td->td_generation++;	/* bump preempt-detect counter */
536 	VM_CNT_INC(v_swtch);
537 	PCPU_SET(switchticks, ticks);
538 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
539 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
540 #ifdef KDTRACE_HOOKS
541 	if (SDT_PROBES_ENABLED() &&
542 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
543 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
544 		SDT_PROBE0(sched, , , preempt);
545 #endif
546 	sched_switch(td, flags);
547 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
548 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
549 
550 	/*
551 	 * If the last thread was exiting, finish cleaning it up.
552 	 */
553 	if ((td = PCPU_GET(deadthread))) {
554 		PCPU_SET(deadthread, NULL);
555 		thread_stash(td);
556 	}
557 	spinlock_exit();
558 }
559 
560 /*
561  * Change thread state to be runnable, placing it on the run queue if
562  * it is in memory.  If it is swapped out, return true so our caller
563  * will know to awaken the swapper.
564  *
565  * Requires the thread lock on entry, drops on exit.
566  */
567 int
568 setrunnable(struct thread *td, int srqflags)
569 {
570 	int swapin;
571 
572 	THREAD_LOCK_ASSERT(td, MA_OWNED);
573 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
574 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
575 
576 	swapin = 0;
577 	switch (TD_GET_STATE(td)) {
578 	case TDS_RUNNING:
579 	case TDS_RUNQ:
580 		break;
581 	case TDS_CAN_RUN:
582 		KASSERT((td->td_flags & TDF_INMEM) != 0,
583 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
584 		    td, td->td_flags, td->td_inhibitors));
585 		/* unlocks thread lock according to flags */
586 		sched_wakeup(td, srqflags);
587 		return (0);
588 	case TDS_INHIBITED:
589 		/*
590 		 * If we are only inhibited because we are swapped out
591 		 * arrange to swap in this process.
592 		 */
593 		if (td->td_inhibitors == TDI_SWAPPED &&
594 		    (td->td_flags & TDF_SWAPINREQ) == 0) {
595 			td->td_flags |= TDF_SWAPINREQ;
596 			swapin = 1;
597 		}
598 		break;
599 	default:
600 		panic("setrunnable: state 0x%x", TD_GET_STATE(td));
601 	}
602 	if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
603 		thread_unlock(td);
604 
605 	return (swapin);
606 }
607 
608 /*
609  * Compute a tenex style load average of a quantity on
610  * 1, 5 and 15 minute intervals.
611  */
612 static void
613 loadav(void *arg)
614 {
615 	int i;
616 	uint64_t nrun;
617 	struct loadavg *avg;
618 
619 	nrun = (uint64_t)sched_load();
620 	avg = &averunnable;
621 
622 	for (i = 0; i < 3; i++)
623 		avg->ldavg[i] = (cexp[i] * (uint64_t)avg->ldavg[i] +
624 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
625 
626 	/*
627 	 * Schedule the next update to occur after 5 seconds, but add a
628 	 * random variation to avoid synchronisation with processes that
629 	 * run at regular intervals.
630 	 */
631 	callout_reset_sbt(&loadav_callout,
632 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
633 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
634 }
635 
636 static void
637 ast_scheduler(struct thread *td, int tda __unused)
638 {
639 #ifdef KTRACE
640 	if (KTRPOINT(td, KTR_CSW))
641 		ktrcsw(1, 1, __func__);
642 #endif
643 	thread_lock(td);
644 	sched_prio(td, td->td_user_pri);
645 	mi_switch(SW_INVOL | SWT_NEEDRESCHED);
646 #ifdef KTRACE
647 	if (KTRPOINT(td, KTR_CSW))
648 		ktrcsw(0, 1, __func__);
649 #endif
650 }
651 
652 static void
653 synch_setup(void *dummy __unused)
654 {
655 	callout_init(&loadav_callout, 1);
656 	ast_register(TDA_SCHED, ASTR_ASTF_REQUIRED, 0, ast_scheduler);
657 
658 	/* Kick off timeout driven events by calling first time. */
659 	loadav(NULL);
660 }
661 
662 bool
663 should_yield(void)
664 {
665 
666 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
667 }
668 
669 void
670 maybe_yield(void)
671 {
672 
673 	if (should_yield())
674 		kern_yield(PRI_USER);
675 }
676 
677 void
678 kern_yield(int prio)
679 {
680 	struct thread *td;
681 
682 	td = curthread;
683 	DROP_GIANT();
684 	thread_lock(td);
685 	if (prio == PRI_USER)
686 		prio = td->td_user_pri;
687 	if (prio >= 0)
688 		sched_prio(td, prio);
689 	mi_switch(SW_VOL | SWT_RELINQUISH);
690 	PICKUP_GIANT();
691 }
692 
693 /*
694  * General purpose yield system call.
695  */
696 int
697 sys_yield(struct thread *td, struct yield_args *uap)
698 {
699 
700 	thread_lock(td);
701 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
702 		sched_prio(td, PRI_MAX_TIMESHARE);
703 	mi_switch(SW_VOL | SWT_RELINQUISH);
704 	td->td_retval[0] = 0;
705 	return (0);
706 }
707 
708 int
709 sys_sched_getcpu(struct thread *td, struct sched_getcpu_args *uap)
710 {
711 	td->td_retval[0] = td->td_oncpu;
712 	return (0);
713 }
714