xref: /freebsd/sys/kern/kern_synch.c (revision 94cba8034ba53725c225c85e35724f0c2b13cea5)
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 	    ("sleeping without a lock"));
153 	KASSERT(ident != NULL, ("_sleep: NULL ident"));
154 	KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
155 	if (priority & PDROP)
156 		KASSERT(lock != NULL && lock != &Giant.lock_object,
157 		    ("PDROP requires a non-Giant lock"));
158 	if (lock != NULL)
159 		class = LOCK_CLASS(lock);
160 	else
161 		class = NULL;
162 
163 	if (SCHEDULER_STOPPED_TD(td)) {
164 		if (lock != NULL && priority & PDROP)
165 			class->lc_unlock(lock);
166 		return (0);
167 	}
168 	catch = priority & PCATCH;
169 	pri = priority & PRIMASK;
170 
171 	KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
172 
173 	if ((uintptr_t)ident >= (uintptr_t)&pause_wchan[0] &&
174 	    (uintptr_t)ident <= (uintptr_t)&pause_wchan[MAXCPU - 1])
175 		sleepq_flags = SLEEPQ_PAUSE;
176 	else
177 		sleepq_flags = SLEEPQ_SLEEP;
178 	if (catch)
179 		sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
180 
181 	sleepq_lock(ident);
182 	CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
183 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
184 
185 	if (lock == &Giant.lock_object)
186 		mtx_assert(&Giant, MA_OWNED);
187 	DROP_GIANT();
188 	if (lock != NULL && lock != &Giant.lock_object &&
189 	    !(class->lc_flags & LC_SLEEPABLE)) {
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 	return (rval);
234 }
235 
236 int
237 msleep_spin_sbt(const void *ident, struct mtx *mtx, const char *wmesg,
238     sbintime_t sbt, sbintime_t pr, int flags)
239 {
240 	struct thread *td;
241 	int rval;
242 	WITNESS_SAVE_DECL(mtx);
243 
244 	td = curthread;
245 	KASSERT(mtx != NULL, ("sleeping without a mutex"));
246 	KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
247 	KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
248 
249 	if (SCHEDULER_STOPPED_TD(td))
250 		return (0);
251 
252 	sleepq_lock(ident);
253 	CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
254 	    td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
255 
256 	DROP_GIANT();
257 	mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
258 	WITNESS_SAVE(&mtx->lock_object, mtx);
259 	mtx_unlock_spin(mtx);
260 
261 	/*
262 	 * We put ourselves on the sleep queue and start our timeout.
263 	 */
264 	sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
265 	if (sbt != 0)
266 		sleepq_set_timeout_sbt(ident, sbt, pr, flags);
267 
268 	/*
269 	 * Can't call ktrace with any spin locks held so it can lock the
270 	 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
271 	 * any spin lock.  Thus, we have to drop the sleepq spin lock while
272 	 * we handle those requests.  This is safe since we have placed our
273 	 * thread on the sleep queue already.
274 	 */
275 #ifdef KTRACE
276 	if (KTRPOINT(td, KTR_CSW)) {
277 		sleepq_release(ident);
278 		ktrcsw(1, 0, wmesg);
279 		sleepq_lock(ident);
280 	}
281 #endif
282 #ifdef WITNESS
283 	sleepq_release(ident);
284 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
285 	    wmesg);
286 	sleepq_lock(ident);
287 #endif
288 	if (sbt != 0)
289 		rval = sleepq_timedwait(ident, 0);
290 	else {
291 		sleepq_wait(ident, 0);
292 		rval = 0;
293 	}
294 #ifdef KTRACE
295 	if (KTRPOINT(td, KTR_CSW))
296 		ktrcsw(0, 0, wmesg);
297 #endif
298 	PICKUP_GIANT();
299 	mtx_lock_spin(mtx);
300 	WITNESS_RESTORE(&mtx->lock_object, mtx);
301 	return (rval);
302 }
303 
304 /*
305  * pause_sbt() delays the calling thread by the given signed binary
306  * time. During cold bootup, pause_sbt() uses the DELAY() function
307  * instead of the _sleep() function to do the waiting. The "sbt"
308  * argument must be greater than or equal to zero. A "sbt" value of
309  * zero is equivalent to a "sbt" value of one tick.
310  */
311 int
312 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
313 {
314 	KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
315 
316 	/* silently convert invalid timeouts */
317 	if (sbt == 0)
318 		sbt = tick_sbt;
319 
320 	if ((cold && curthread == &thread0) || kdb_active ||
321 	    SCHEDULER_STOPPED()) {
322 		/*
323 		 * We delay one second at a time to avoid overflowing the
324 		 * system specific DELAY() function(s):
325 		 */
326 		while (sbt >= SBT_1S) {
327 			DELAY(1000000);
328 			sbt -= SBT_1S;
329 		}
330 		/* Do the delay remainder, if any */
331 		sbt = howmany(sbt, SBT_1US);
332 		if (sbt > 0)
333 			DELAY(sbt);
334 		return (EWOULDBLOCK);
335 	}
336 	return (_sleep(&pause_wchan[curcpu], NULL,
337 	    (flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
338 }
339 
340 /*
341  * Make all threads sleeping on the specified identifier runnable.
342  */
343 void
344 wakeup(const void *ident)
345 {
346 	int wakeup_swapper;
347 
348 	sleepq_lock(ident);
349 	wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
350 	sleepq_release(ident);
351 	if (wakeup_swapper) {
352 		KASSERT(ident != &proc0,
353 		    ("wakeup and wakeup_swapper and proc0"));
354 		kick_proc0();
355 	}
356 }
357 
358 /*
359  * Make a thread sleeping on the specified identifier runnable.
360  * May wake more than one thread if a target thread is currently
361  * swapped out.
362  */
363 void
364 wakeup_one(const void *ident)
365 {
366 	int wakeup_swapper;
367 
368 	sleepq_lock(ident);
369 	wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
370 	sleepq_release(ident);
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 	    0, 0);
383 	sleepq_release(ident);
384 	if (wakeup_swapper)
385 		kick_proc0();
386 }
387 
388 /*
389  * Signal sleeping waiters after the counter has reached zero.
390  */
391 void
392 _blockcount_wakeup(blockcount_t *bc, u_int old)
393 {
394 
395 	KASSERT(_BLOCKCOUNT_WAITERS(old),
396 	    ("%s: no waiters on %p", __func__, bc));
397 
398 	if (atomic_cmpset_int(&bc->__count, _BLOCKCOUNT_WAITERS_FLAG, 0))
399 		wakeup(bc);
400 }
401 
402 /*
403  * Wait for a wakeup or a signal.  This does not guarantee that the count is
404  * still zero on return.  Callers wanting a precise answer should use
405  * blockcount_wait() with an interlock.
406  *
407  * If there is no work to wait for, return 0.  If the sleep was interrupted by a
408  * signal, return EINTR or ERESTART, and return EAGAIN otherwise.
409  */
410 int
411 _blockcount_sleep(blockcount_t *bc, struct lock_object *lock, const char *wmesg,
412     int prio)
413 {
414 	void *wchan;
415 	uintptr_t lock_state;
416 	u_int old;
417 	int ret;
418 	bool catch, drop;
419 
420 	KASSERT(lock != &Giant.lock_object,
421 	    ("%s: cannot use Giant as the interlock", __func__));
422 
423 	catch = (prio & PCATCH) != 0;
424 	drop = (prio & PDROP) != 0;
425 	prio &= PRIMASK;
426 
427 	/*
428 	 * Synchronize with the fence in blockcount_release().  If we end up
429 	 * waiting, the sleepqueue lock acquisition will provide the required
430 	 * side effects.
431 	 *
432 	 * If there is no work to wait for, but waiters are present, try to put
433 	 * ourselves to sleep to avoid jumping ahead.
434 	 */
435 	if (atomic_load_acq_int(&bc->__count) == 0) {
436 		if (lock != NULL && drop)
437 			LOCK_CLASS(lock)->lc_unlock(lock);
438 		return (0);
439 	}
440 	lock_state = 0;
441 	wchan = bc;
442 	sleepq_lock(wchan);
443 	DROP_GIANT();
444 	if (lock != NULL)
445 		lock_state = LOCK_CLASS(lock)->lc_unlock(lock);
446 	old = blockcount_read(bc);
447 	ret = 0;
448 	do {
449 		if (_BLOCKCOUNT_COUNT(old) == 0) {
450 			sleepq_release(wchan);
451 			goto out;
452 		}
453 		if (_BLOCKCOUNT_WAITERS(old))
454 			break;
455 	} while (!atomic_fcmpset_int(&bc->__count, &old,
456 	    old | _BLOCKCOUNT_WAITERS_FLAG));
457 	sleepq_add(wchan, NULL, wmesg, catch ? SLEEPQ_INTERRUPTIBLE : 0, 0);
458 	if (catch)
459 		ret = sleepq_wait_sig(wchan, prio);
460 	else
461 		sleepq_wait(wchan, prio);
462 	if (ret == 0)
463 		ret = EAGAIN;
464 
465 out:
466 	PICKUP_GIANT();
467 	if (lock != NULL && !drop)
468 		LOCK_CLASS(lock)->lc_lock(lock, lock_state);
469 
470 	return (ret);
471 }
472 
473 static void
474 kdb_switch(void)
475 {
476 	thread_unlock(curthread);
477 	kdb_backtrace();
478 	kdb_reenter();
479 	panic("%s: did not reenter debugger", __func__);
480 }
481 
482 /*
483  * The machine independent parts of context switching.
484  *
485  * The thread lock is required on entry and is no longer held on return.
486  */
487 void
488 mi_switch(int flags)
489 {
490 	uint64_t runtime, new_switchtime;
491 	struct thread *td;
492 
493 	td = curthread;			/* XXX */
494 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
495 	KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
496 #ifdef INVARIANTS
497 	if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
498 		mtx_assert(&Giant, MA_NOTOWNED);
499 #endif
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 
505 	/*
506 	 * Don't perform context switches from the debugger.
507 	 */
508 	if (kdb_active)
509 		kdb_switch();
510 	if (SCHEDULER_STOPPED_TD(td))
511 		return;
512 	if (flags & SW_VOL) {
513 		td->td_ru.ru_nvcsw++;
514 		td->td_swvoltick = ticks;
515 	} else {
516 		td->td_ru.ru_nivcsw++;
517 		td->td_swinvoltick = ticks;
518 	}
519 #ifdef SCHED_STATS
520 	SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
521 #endif
522 	/*
523 	 * Compute the amount of time during which the current
524 	 * thread was running, and add that to its total so far.
525 	 */
526 	new_switchtime = cpu_ticks();
527 	runtime = new_switchtime - PCPU_GET(switchtime);
528 	td->td_runtime += runtime;
529 	td->td_incruntime += runtime;
530 	PCPU_SET(switchtime, new_switchtime);
531 	td->td_generation++;	/* bump preempt-detect counter */
532 	VM_CNT_INC(v_swtch);
533 	PCPU_SET(switchticks, ticks);
534 	CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
535 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
536 #ifdef KDTRACE_HOOKS
537 	if (SDT_PROBES_ENABLED() &&
538 	    ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
539 	    (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
540 		SDT_PROBE0(sched, , , preempt);
541 #endif
542 	sched_switch(td, flags);
543 	CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
544 	    td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
545 
546 	/*
547 	 * If the last thread was exiting, finish cleaning it up.
548 	 */
549 	if ((td = PCPU_GET(deadthread))) {
550 		PCPU_SET(deadthread, NULL);
551 		thread_stash(td);
552 	}
553 	spinlock_exit();
554 }
555 
556 /*
557  * Change thread state to be runnable, placing it on the run queue if
558  * it is in memory.  If it is swapped out, return true so our caller
559  * will know to awaken the swapper.
560  *
561  * Requires the thread lock on entry, drops on exit.
562  */
563 int
564 setrunnable(struct thread *td, int srqflags)
565 {
566 	int swapin;
567 
568 	THREAD_LOCK_ASSERT(td, MA_OWNED);
569 	KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
570 	    ("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
571 
572 	swapin = 0;
573 	switch (td->td_state) {
574 	case TDS_RUNNING:
575 	case TDS_RUNQ:
576 		break;
577 	case TDS_CAN_RUN:
578 		KASSERT((td->td_flags & TDF_INMEM) != 0,
579 		    ("setrunnable: td %p not in mem, flags 0x%X inhibit 0x%X",
580 		    td, td->td_flags, td->td_inhibitors));
581 		/* unlocks thread lock according to flags */
582 		sched_wakeup(td, srqflags);
583 		return (0);
584 	case TDS_INHIBITED:
585 		/*
586 		 * If we are only inhibited because we are swapped out
587 		 * arrange to swap in this process.
588 		 */
589 		if (td->td_inhibitors == TDI_SWAPPED &&
590 		    (td->td_flags & TDF_SWAPINREQ) == 0) {
591 			td->td_flags |= TDF_SWAPINREQ;
592 			swapin = 1;
593 		}
594 		break;
595 	default:
596 		panic("setrunnable: state 0x%x", td->td_state);
597 	}
598 	if ((srqflags & (SRQ_HOLD | SRQ_HOLDTD)) == 0)
599 		thread_unlock(td);
600 
601 	return (swapin);
602 }
603 
604 /*
605  * Compute a tenex style load average of a quantity on
606  * 1, 5 and 15 minute intervals.
607  */
608 static void
609 loadav(void *arg)
610 {
611 	int i, nrun;
612 	struct loadavg *avg;
613 
614 	nrun = sched_load();
615 	avg = &averunnable;
616 
617 	for (i = 0; i < 3; i++)
618 		avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
619 		    nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
620 
621 	/*
622 	 * Schedule the next update to occur after 5 seconds, but add a
623 	 * random variation to avoid synchronisation with processes that
624 	 * run at regular intervals.
625 	 */
626 	callout_reset_sbt(&loadav_callout,
627 	    SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
628 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
629 }
630 
631 /* ARGSUSED */
632 static void
633 synch_setup(void *dummy)
634 {
635 	callout_init(&loadav_callout, 1);
636 
637 	/* Kick off timeout driven events by calling first time. */
638 	loadav(NULL);
639 }
640 
641 int
642 should_yield(void)
643 {
644 
645 	return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
646 }
647 
648 void
649 maybe_yield(void)
650 {
651 
652 	if (should_yield())
653 		kern_yield(PRI_USER);
654 }
655 
656 void
657 kern_yield(int prio)
658 {
659 	struct thread *td;
660 
661 	td = curthread;
662 	DROP_GIANT();
663 	thread_lock(td);
664 	if (prio == PRI_USER)
665 		prio = td->td_user_pri;
666 	if (prio >= 0)
667 		sched_prio(td, prio);
668 	mi_switch(SW_VOL | SWT_RELINQUISH);
669 	PICKUP_GIANT();
670 }
671 
672 /*
673  * General purpose yield system call.
674  */
675 int
676 sys_yield(struct thread *td, struct yield_args *uap)
677 {
678 
679 	thread_lock(td);
680 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
681 		sched_prio(td, PRI_MAX_TIMESHARE);
682 	mi_switch(SW_VOL | SWT_RELINQUISH);
683 	td->td_retval[0] = 0;
684 	return (0);
685 }
686