xref: /freebsd/sys/kern/kern_synch.c (revision c1d255d3ffdbe447de3ab875bf4e7d7accc5bfc5)
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     "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 	td = curthread;
145 #ifdef KTRACE
146 	if (KTRPOINT(td, KTR_CSW))
147 		ktrcsw(1, 0, wmesg);
148 #endif
149 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
150 	    "Sleeping on \"%s\"", wmesg);
151 	KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL ||
152 	    (priority & PNOLOCK) != 0,
153 	    ("sleeping without a lock"));
154 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
155 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
156 	if (priority & PDROP)
157 		KASSERT(lock != NULL && lock != &Giant.lock_object,
158 		    ("PDROP requires a non-Giant lock"));
159 	if (lock != NULL)
160 		class = LOCK_CLASS(lock);
161 	else
162 		class = NULL;
163 
164 	if (SCHEDULER_STOPPED_TD(td)) {
165 		if (lock != NULL && priority & PDROP)
166 			class->lc_unlock(lock);
167 		return (0);
168 	}
169 	catch = priority & PCATCH;
170 	pri = priority & PRIMASK;
171 
172 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
173 
174 	if ((uintptr_t)ident >= (uintptr_t)&pause_wchan[0] &&
175 	    (uintptr_t)ident <= (uintptr_t)&pause_wchan[MAXCPU - 1])
176 		sleepq_flags = SLEEPQ_PAUSE;
177 	else
178 		sleepq_flags = SLEEPQ_SLEEP;
179 	if (catch)
180 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
181 
182 	sleepq_lock(ident);
183 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
184 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
185 
186 	if (lock == &Giant.lock_object)
187 		mtx_assert(&Giant, MA_OWNED);
188 	DROP_GIANT();
189 	if (lock != NULL && lock != &Giant.lock_object &&
190 	    !(class->lc_flags & LC_SLEEPABLE)) {
191 		KASSERT(!(class->lc_flags & LC_SPINLOCK),
192 		    ("spin locks can only use msleep_spin"));
193 		WITNESS_SAVE(lock, lock_witness);
194 		lock_state = class->lc_unlock(lock);
195 	} else
196 		/* GCC needs to follow the Yellow Brick Road */
197 		lock_state = -1;
198 
199 	/*
200 	 * We put ourselves on the sleep queue and start our timeout
201 	 * before calling thread_suspend_check, as we could stop there,
202 	 * and a wakeup or a SIGCONT (or both) could occur while we were
203 	 * stopped without resuming us.  Thus, we must be ready for sleep
204 	 * when cursig() is called.  If the wakeup happens while we're
205 	 * stopped, then td will no longer be on a sleep queue upon
206 	 * return from cursig().
207 	 */
208 	sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
209 	if (sbt != 0)
210 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
211 	if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
212 		sleepq_release(ident);
213 		WITNESS_SAVE(lock, lock_witness);
214 		lock_state = class->lc_unlock(lock);
215 		sleepq_lock(ident);
216 	}
217 	if (sbt != 0 && catch)
218 		rval = sleepq_timedwait_sig(ident, pri);
219 	else if (sbt != 0)
220 		rval = sleepq_timedwait(ident, pri);
221 	else if (catch)
222 		rval = sleepq_wait_sig(ident, pri);
223 	else {
224 		sleepq_wait(ident, pri);
225 		rval = 0;
226 	}
227 #ifdef KTRACE
228 	if (KTRPOINT(td, KTR_CSW))
229 		ktrcsw(0, 0, wmesg);
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 	return (rval);
237 }
238 
239 int
240 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
241     sbintime_t sbt, sbintime_t pr, int flags)
242 {
243 	struct thread *td;
244 	int rval;
245 	WITNESS_SAVE_DECL(mtx);
246 
247 	td = curthread;
248 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
249 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
250 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
251 
252 	if (SCHEDULER_STOPPED_TD(td))
253 		return (0);
254 
255 	sleepq_lock(ident);
256 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
257 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
258 
259 	DROP_GIANT();
260 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
261 	WITNESS_SAVE(&mtx->lock_object, mtx);
262 	mtx_unlock_spin(mtx);
263 
264 	/*
265 	 * We put ourselves on the sleep queue and start our timeout.
266 	 */
267 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
268 	if (sbt != 0)
269 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
270 
271 	/*
272 	 * Can't call ktrace with any spin locks held so it can lock the
273 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
274 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
275 	 * we handle those requests.  This is safe since we have placed our
276 	 * thread on the sleep queue already.
277 	 */
278 #ifdef KTRACE
279 	if (KTRPOINT(td, KTR_CSW)) {
280 		sleepq_release(ident);
281 		ktrcsw(1, 0, wmesg);
282 		sleepq_lock(ident);
283 	}
284 #endif
285 #ifdef WITNESS
286 	sleepq_release(ident);
287 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
288 	    wmesg);
289 	sleepq_lock(ident);
290 #endif
291 	if (sbt != 0)
292 		rval = sleepq_timedwait(ident, 0);
293 	else {
294 		sleepq_wait(ident, 0);
295 		rval = 0;
296 	}
297 #ifdef KTRACE
298 	if (KTRPOINT(td, KTR_CSW))
299 		ktrcsw(0, 0, wmesg);
300 #endif
301 	PICKUP_GIANT();
302 	mtx_lock_spin(mtx);
303 	WITNESS_RESTORE(&mtx->lock_object, mtx);
304 	return (rval);
305 }
306 
307 /*
308  * pause_sbt() delays the calling thread by the given signed binary
309  * time. During cold bootup, pause_sbt() uses the DELAY() function
310  * instead of the _sleep() function to do the waiting. The "sbt"
311  * argument must be greater than or equal to zero. A "sbt" value of
312  * zero is equivalent to a "sbt" value of one tick.
313  */
314 int
315 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
316 {
317 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
318 
319 	/* silently convert invalid timeouts */
320 	if (sbt == 0)
321 		sbt = tick_sbt;
322 
323 	if ((cold && curthread == &thread0) || kdb_active ||
324 	    SCHEDULER_STOPPED()) {
325 		/*
326 		 * We delay one second at a time to avoid overflowing the
327 		 * system specific DELAY() function(s):
328 		 */
329 		while (sbt >= SBT_1S) {
330 			DELAY(1000000);
331 			sbt -= SBT_1S;
332 		}
333 		/* Do the delay remainder, if any */
334 		sbt = howmany(sbt, SBT_1US);
335 		if (sbt > 0)
336 			DELAY(sbt);
337 		return (EWOULDBLOCK);
338 	}
339 	return (_sleep(&pause_wchan[curcpu], NULL,
340 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
341 }
342 
343 /*
344  * Make all threads sleeping on the specified identifier runnable.
345  */
346 void
347 wakeup(const void *ident)
348 {
349 	int wakeup_swapper;
350 
351 	sleepq_lock(ident);
352 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
353 	sleepq_release(ident);
354 	if (wakeup_swapper) {
355 		KASSERT(ident != &proc0,
356 		    ("wakeup and wakeup_swapper and proc0"));
357 		kick_proc0();
358 	}
359 }
360 
361 /*
362  * Make a thread sleeping on the specified identifier runnable.
363  * May wake more than one thread if a target thread is currently
364  * swapped out.
365  */
366 void
367 wakeup_one(const void *ident)
368 {
369 	int wakeup_swapper;
370 
371 	sleepq_lock(ident);
372 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_DROP, 0, 0);
373 	if (wakeup_swapper)
374 		kick_proc0();
375 }
376 
377 void
378 wakeup_any(const void *ident)
379 {
380 	int wakeup_swapper;
381 
382 	sleepq_lock(ident);
383 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR |
384 	    SLEEPQ_DROP, 0, 0);
385 	if (wakeup_swapper)
386 		kick_proc0();
387 }
388 
389 /*
390  * Signal sleeping waiters after the counter has reached zero.
391  */
392 void
393 _blockcount_wakeup(blockcount_t *bc, u_int old)
394 {
395 
396 	KASSERT(_BLOCKCOUNT_WAITERS(old),
397 	    ("%s: no waiters on %p", __func__, bc));
398 
399 	if (atomic_cmpset_int(&bc->__count, _BLOCKCOUNT_WAITERS_FLAG, 0))
400 		wakeup(bc);
401 }
402 
403 /*
404  * Wait for a wakeup or a signal.  This does not guarantee that the count is
405  * still zero on return.  Callers wanting a precise answer should use
406  * blockcount_wait() with an interlock.
407  *
408  * If there is no work to wait for, return 0.  If the sleep was interrupted by a
409  * signal, return EINTR or ERESTART, and return EAGAIN otherwise.
410  */
411 int
412 _blockcount_sleep(blockcount_t *bc, struct lock_object *lock, const char *wmesg,
413     int prio)
414 {
415 	void *wchan;
416 	uintptr_t lock_state;
417 	u_int old;
418 	int ret;
419 	bool catch, drop;
420 
421 	KASSERT(lock != &Giant.lock_object,
422 	    ("%s: cannot use Giant as the interlock", __func__));
423 
424 	catch = (prio & PCATCH) != 0;
425 	drop = (prio & PDROP) != 0;
426 	prio &= PRIMASK;
427 
428 	/*
429 	 * Synchronize with the fence in blockcount_release().  If we end up
430 	 * waiting, the sleepqueue lock acquisition will provide the required
431 	 * side effects.
432 	 *
433 	 * If there is no work to wait for, but waiters are present, try to put
434 	 * ourselves to sleep to avoid jumping ahead.
435 	 */
436 	if (atomic_load_acq_int(&bc->__count) == 0) {
437 		if (lock != NULL && drop)
438 			LOCK_CLASS(lock)->lc_unlock(lock);
439 		return (0);
440 	}
441 	lock_state = 0;
442 	wchan = bc;
443 	sleepq_lock(wchan);
444 	DROP_GIANT();
445 	if (lock != NULL)
446 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
447 	old = blockcount_read(bc);
448 	ret = 0;
449 	do {
450 		if (_BLOCKCOUNT_COUNT(old) == 0) {
451 			sleepq_release(wchan);
452 			goto out;
453 		}
454 		if (_BLOCKCOUNT_WAITERS(old))
455 			break;
456 	} while (!atomic_fcmpset_int(&bc->__count, &old,
457 	    old | _BLOCKCOUNT_WAITERS_FLAG));
458 	sleepq_add(wchan, NULL, wmesg, catch ? SLEEPQ_INTERRUPTIBLE : 0, 0);
459 	if (catch)
460 		ret = sleepq_wait_sig(wchan, prio);
461 	else
462 		sleepq_wait(wchan, prio);
463 	if (ret == 0)
464 		ret = EAGAIN;
465 
466 out:
467 	PICKUP_GIANT();
468 	if (lock != NULL && !drop)
469 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
470 
471 	return (ret);
472 }
473 
474 static void
475 kdb_switch(void)
476 {
477 	thread_unlock(curthread);
478 	kdb_backtrace();
479 	kdb_reenter();
480 	panic("%s: did not reenter debugger", __func__);
481 }
482 
483 /*
484  * The machine independent parts of context switching.
485  *
486  * The thread lock is required on entry and is no longer held on return.
487  */
488 void
489 mi_switch(int flags)
490 {
491 	uint64_t runtime, new_switchtime;
492 	struct thread *td;
493 
494 	td = curthread;			/* XXX */
495 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
496 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
497 #ifdef INVARIANTS
498 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
499 		mtx_assert(&Giant, MA_NOTOWNED);
500 #endif
501 	KASSERT(td->td_critnest == 1 || KERNEL_PANICKED(),
502 		("mi_switch: switch in a critical section"));
503 	KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
504 	    ("mi_switch: switch must be voluntary or involuntary"));
505 
506 	/*
507 	 * Don't perform context switches from the debugger.
508 	 */
509 	if (kdb_active)
510 		kdb_switch();
511 	if (SCHEDULER_STOPPED_TD(td))
512 		return;
513 	if (flags & SW_VOL) {
514 		td->td_ru.ru_nvcsw++;
515 		td->td_swvoltick = ticks;
516 	} else {
517 		td->td_ru.ru_nivcsw++;
518 		td->td_swinvoltick = ticks;
519 	}
520 #ifdef SCHED_STATS
521 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
522 #endif
523 	/*
524 	 * Compute the amount of time during which the current
525 	 * thread was running, and add that to its total so far.
526 	 */
527 	new_switchtime = cpu_ticks();
528 	runtime = new_switchtime - PCPU_GET(switchtime);
529 	td->td_runtime += runtime;
530 	td->td_incruntime += runtime;
531 	PCPU_SET(switchtime, new_switchtime);
532 	td->td_generation++;	/* bump preempt-detect counter */
533 	VM_CNT_INC(v_swtch);
534 	PCPU_SET(switchticks, ticks);
535 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
536 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
537 #ifdef KDTRACE_HOOKS
538 	if (SDT_PROBES_ENABLED() &&
539 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
540 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
541 		SDT_PROBE0(sched, , , preempt);
542 #endif
543 	sched_switch(td, flags);
544 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
545 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
546 
547 	/*
548 	 * If the last thread was exiting, finish cleaning it up.
549 	 */
550 	if ((td = PCPU_GET(deadthread))) {
551 		PCPU_SET(deadthread, NULL);
552 		thread_stash(td);
553 	}
554 	spinlock_exit();
555 }
556 
557 /*
558  * Change thread state to be runnable, placing it on the run queue if
559  * it is in memory.  If it is swapped out, return true so our caller
560  * will know to awaken the swapper.
561  *
562  * Requires the thread lock on entry, drops on exit.
563  */
564 int
565 setrunnable(struct thread *td, int srqflags)
566 {
567 	int swapin;
568 
569 	THREAD_LOCK_ASSERT(td, MA_OWNED);
570 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
571 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
572 
573 	swapin = 0;
574 	switch (TD_GET_STATE(td)) {
575 	case TDS_RUNNING:
576 	case TDS_RUNQ:
577 		break;
578 	case TDS_CAN_RUN:
579 		KASSERT((td->td_flags & TDF_INMEM) != 0,
580 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
581 		    td, td->td_flags, td->td_inhibitors));
582 		/* unlocks thread lock according to flags */
583 		sched_wakeup(td, srqflags);
584 		return (0);
585 	case TDS_INHIBITED:
586 		/*
587 		 * If we are only inhibited because we are swapped out
588 		 * arrange to swap in this process.
589 		 */
590 		if (td->td_inhibitors == TDI_SWAPPED &&
591 		    (td->td_flags & TDF_SWAPINREQ) == 0) {
592 			td->td_flags |= TDF_SWAPINREQ;
593 			swapin = 1;
594 		}
595 		break;
596 	default:
597 		panic("setrunnable: state 0x%x", TD_GET_STATE(td));
598 	}
599 	if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
600 		thread_unlock(td);
601 
602 	return (swapin);
603 }
604 
605 /*
606  * Compute a tenex style load average of a quantity on
607  * 1, 5 and 15 minute intervals.
608  */
609 static void
610 loadav(void *arg)
611 {
612 	int i, nrun;
613 	struct loadavg *avg;
614 
615 	nrun = sched_load();
616 	avg = &averunnable;
617 
618 	for (i = 0; i < 3; i++)
619 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
620 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
621 
622 	/*
623 	 * Schedule the next update to occur after 5 seconds, but add a
624 	 * random variation to avoid synchronisation with processes that
625 	 * run at regular intervals.
626 	 */
627 	callout_reset_sbt(&loadav_callout,
628 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
629 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
630 }
631 
632 /* ARGSUSED */
633 static void
634 synch_setup(void *dummy)
635 {
636 	callout_init(&loadav_callout, 1);
637 
638 	/* Kick off timeout driven events by calling first time. */
639 	loadav(NULL);
640 }
641 
642 int
643 should_yield(void)
644 {
645 
646 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
647 }
648 
649 void
650 maybe_yield(void)
651 {
652 
653 	if (should_yield())
654 		kern_yield(PRI_USER);
655 }
656 
657 void
658 kern_yield(int prio)
659 {
660 	struct thread *td;
661 
662 	td = curthread;
663 	DROP_GIANT();
664 	thread_lock(td);
665 	if (prio == PRI_USER)
666 		prio = td->td_user_pri;
667 	if (prio >= 0)
668 		sched_prio(td, prio);
669 	mi_switch(SW_VOL | SWT_RELINQUISH);
670 	PICKUP_GIANT();
671 }
672 
673 /*
674  * General purpose yield system call.
675  */
676 int
677 sys_yield(struct thread *td, struct yield_args *uap)
678 {
679 
680 	thread_lock(td);
681 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
682 		sched_prio(td, PRI_MAX_TIMESHARE);
683 	mi_switch(SW_VOL | SWT_RELINQUISH);
684 	td->td_retval[0] = 0;
685 	return (0);
686 }
687