subr_turnstile.c revision 122590
1/*-
2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 *    notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 *    notice, this list of conditions and the following disclaimer in the
11 *    documentation and/or other materials provided with the distribution.
12 * 3. Berkeley Software Design Inc's name may not be used to endorse or
13 *    promote products derived from this software without specific prior
14 *    written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 *
28 *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29 *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30 */
31
32/*
33 * Implementation of turnstiles used to hold queue of threads blocked on
34 * non-sleepable locks.  Sleepable locks use condition variables to
35 * implement their queues.  Turnstiles differ from a sleep queue in that
36 * turnstile queue's are assigned to a lock held by an owning thread.  Thus,
37 * when one thread is enqueued onto a turnstile, it can lend its priority
38 * to the owning thread.
39 *
40 * We wish to avoid bloating locks with an embedded turnstile and we do not
41 * want to use back-pointers in the locks for the same reason.  Thus, we
42 * use a similar approach to that of Solaris 7 as described in Solaris
43 * Internals by Jim Mauro and Richard McDougall.  Turnstiles are looked up
44 * in a hash table based on the address of the lock.  Each entry in the
45 * hash table is a linked-lists of turnstiles and is called a turnstile
46 * chain.  Each chain contains a spin mutex that protects all of the
47 * turnstiles in the chain.
48 *
49 * Each time a thread is created, a turnstile is malloc'd and attached to
50 * that thread.  When a thread blocks on a lock, if it is the first thread
51 * to block, it lends its turnstile to the lock.  If the lock already has
52 * a turnstile, then it gives its turnstile to the lock's turnstile's free
53 * list.  When a thread is woken up, it takes a turnstile from the free list
54 * if there are any other waiters.  If it is the only thread blocked on the
55 * lock, then it reclaims the turnstile associated with the lock and removes
56 * it from the hash table.
57 *
58 * XXX: We should probably implement some sort of sleep queue that condition
59 * variables and sleepqueue's share.  On Solaris condition variables are
60 * implemented using a hash table of sleep queues similar to our current
61 * sleep queues.  We might want to investigate doing that ourselves.
62 */
63
64#include <sys/cdefs.h>
65__FBSDID("$FreeBSD: head/sys/kern/subr_turnstile.c 122590 2003-11-12 23:48:42Z jhb $");
66
67#include <sys/param.h>
68#include <sys/systm.h>
69#include <sys/kernel.h>
70#include <sys/ktr.h>
71#include <sys/lock.h>
72#include <sys/malloc.h>
73#include <sys/mutex.h>
74#include <sys/proc.h>
75#include <sys/queue.h>
76#include <sys/resourcevar.h>
77#include <sys/turnstile.h>
78#include <sys/sched.h>
79
80/*
81 * Constants for the hash table of turnstile chains.  TC_SHIFT is a magic
82 * number chosen because the sleep queue's use the same value for the
83 * shift.  Basically, we ignore the lower 8 bits of the address.
84 * TC_TABLESIZE must be a power of two for TC_MASK to work properly.
85 */
86#define	TC_TABLESIZE	128			/* Must be power of 2. */
87#define	TC_MASK		(TC_TABLESIZE - 1)
88#define	TC_SHIFT	8
89#define	TC_HASH(lock)	(((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK)
90#define	TC_LOOKUP(lock)	&turnstile_chains[TC_HASH(lock)]
91
92/*
93 * There are three different lists of turnstiles as follows.  The list
94 * connected by ts_link entries is a per-thread list of all the turnstiles
95 * attached to locks that we own.  This is used to fixup our priority when
96 * a lock is released.  The other two lists use the ts_hash entries.  The
97 * first of these two is turnstile chain list that a turnstile is on when
98 * it is attached to a lock.  The second list to use ts_hash is the free
99 * list hung off a turnstile that is attached to a lock.
100 *
101 * Each turnstile contains two lists of threads.  The ts_blocked list is
102 * a linked list of threads blocked on the turnstile's lock.  The
103 * ts_pending list is a linked list of threads previously awoken by
104 * turnstile_signal() or turnstile_wait() that are waiting to be put on
105 * the run queue.
106 *
107 * Locking key:
108 *  c - turnstile chain lock
109 *  q - td_contested lock
110 */
111struct turnstile {
112	TAILQ_HEAD(, thread) ts_blocked;	/* (c + q) Blocked threads. */
113	TAILQ_HEAD(, thread) ts_pending;	/* (c) Pending threads. */
114	LIST_ENTRY(turnstile) ts_hash;		/* (c) Chain and free list. */
115	LIST_ENTRY(turnstile) ts_link;		/* (q) Contested locks. */
116	LIST_HEAD(, turnstile) ts_free;		/* (c) Free turnstiles. */
117	struct lock_object *ts_lockobj;		/* (c) Lock we reference. */
118	struct thread *ts_owner;		/* (c + q) Who owns the lock. */
119};
120
121struct turnstile_chain {
122	LIST_HEAD(, turnstile) tc_turnstiles;	/* List of turnstiles. */
123	struct mtx tc_lock;			/* Spin lock for this chain. */
124};
125
126static struct mtx td_contested_lock;
127static struct turnstile_chain turnstile_chains[TC_TABLESIZE];
128
129MALLOC_DEFINE(M_TURNSTILE, "turnstiles", "turnstiles");
130
131/*
132 * Prototypes for non-exported routines.
133 */
134static void	init_turnstile0(void *dummy);
135static void	propagate_priority(struct thread *);
136static void	turnstile_setowner(struct turnstile *ts, struct thread *owner);
137
138/*
139 * Walks the chain of turnstiles and their owners to propagate the priority
140 * of the thread being blocked to all the threads holding locks that have to
141 * release their locks before this thread can run again.
142 */
143static void
144propagate_priority(struct thread *td)
145{
146	struct turnstile_chain *tc;
147	struct turnstile *ts;
148	struct thread *td1;
149	int pri;
150
151	mtx_assert(&sched_lock, MA_OWNED);
152	pri = td->td_priority;
153	ts = td->td_blocked;
154	for (;;) {
155		td = ts->ts_owner;
156
157		if (td == NULL) {
158			/*
159			 * This really isn't quite right. Really
160			 * ought to bump priority of thread that
161			 * next acquires the lock.
162			 */
163			return;
164		}
165
166		MPASS(td->td_proc != NULL);
167		MPASS(td->td_proc->p_magic == P_MAGIC);
168
169		/*
170		 * XXX: The owner of a turnstile can be stale if it is the
171		 * first thread to grab a slock of a sx lock.  In that case
172		 * it is possible for us to be at SSLEEP or some other
173		 * weird state.  We should probably just return if the state
174		 * isn't SRUN or SLOCK.
175		 */
176		KASSERT(!TD_IS_SLEEPING(td),
177		    ("sleeping thread (pid %d) owns a non-sleepable lock",
178		    td->td_proc->p_pid));
179
180		/*
181		 * If this thread already has higher priority than the
182		 * thread that is being blocked, we are finished.
183		 */
184		if (td->td_priority <= pri)
185			return;
186
187		/*
188		 * If lock holder is actually running, just bump priority.
189		 */
190		if (TD_IS_RUNNING(td)) {
191			td->td_priority = pri;
192			return;
193		}
194
195#ifndef SMP
196		/*
197		 * For UP, we check to see if td is curthread (this shouldn't
198		 * ever happen however as it would mean we are in a deadlock.)
199		 */
200		KASSERT(td != curthread, ("Deadlock detected"));
201#endif
202
203		/*
204		 * If on run queue move to new run queue, and quit.
205		 * XXXKSE this gets a lot more complicated under threads
206		 * but try anyhow.
207		 */
208		if (TD_ON_RUNQ(td)) {
209			MPASS(td->td_blocked == NULL);
210			sched_prio(td, pri);
211			return;
212		}
213
214		/*
215		 * Bump this thread's priority.
216		 */
217		td->td_priority = pri;
218
219		/*
220		 * If we aren't blocked on a lock, we should be.
221		 */
222		KASSERT(TD_ON_LOCK(td), (
223		    "process %d(%s):%d holds %s but isn't blocked on a lock\n",
224		    td->td_proc->p_pid, td->td_proc->p_comm, td->td_state,
225		    ts->ts_lockobj->lo_name));
226
227		/*
228		 * Pick up the lock that td is blocked on.
229		 */
230		ts = td->td_blocked;
231		MPASS(ts != NULL);
232		tc = TC_LOOKUP(ts->ts_lockobj);
233		mtx_lock_spin(&tc->tc_lock);
234
235		/*
236		 * If this turnstile has no threads on its blocked queue
237		 * then it's possible that it was just woken up on another
238		 * CPU.  If so, we are done.
239		 */
240		if (TAILQ_EMPTY(&ts->ts_blocked)) {
241			mtx_unlock_spin(&tc->tc_lock);
242			return;
243		}
244
245		/*
246		 * Check if the thread needs to be moved up on
247		 * the blocked chain.  It doesn't need to be moved
248		 * if it is already at the head of the list or if
249		 * the item in front of it still has a higher priority.
250		 */
251		if (td == TAILQ_FIRST(&ts->ts_blocked)) {
252			mtx_unlock_spin(&tc->tc_lock);
253			continue;
254		}
255
256		td1 = TAILQ_PREV(td, threadqueue, td_lockq);
257		if (td1->td_priority <= pri) {
258			mtx_unlock_spin(&tc->tc_lock);
259			continue;
260		}
261
262		/*
263		 * Remove thread from blocked chain and determine where
264		 * it should be moved up to.  Since we know that td1 has
265		 * a lower priority than td, we know that at least one
266		 * thread in the chain has a lower priority and that
267		 * td1 will thus not be NULL after the loop.
268		 */
269		mtx_lock_spin(&td_contested_lock);
270		TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq);
271		TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) {
272			MPASS(td1->td_proc->p_magic == P_MAGIC);
273			if (td1->td_priority > pri)
274				break;
275		}
276
277		MPASS(td1 != NULL);
278		TAILQ_INSERT_BEFORE(td1, td, td_lockq);
279		mtx_unlock_spin(&td_contested_lock);
280		CTR4(KTR_LOCK,
281		    "propagate_priority: td %p moved before %p on [%p] %s",
282		    td, td1, ts->ts_lockobj, ts->ts_lockobj->lo_name);
283		mtx_unlock_spin(&tc->tc_lock);
284	}
285}
286
287/*
288 * Early initialization of turnstiles.  This is not done via a SYSINIT()
289 * since this needs to be initialized very early when mutexes are first
290 * initialized.
291 */
292void
293init_turnstiles(void)
294{
295	int i;
296
297	for (i = 0; i < TC_TABLESIZE; i++) {
298		LIST_INIT(&turnstile_chains[i].tc_turnstiles);
299		mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
300		    NULL, MTX_SPIN);
301	}
302	mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
303#ifdef INVARIANTS
304	thread0.td_turnstile = NULL;
305#endif
306}
307
308static void
309init_turnstile0(void *dummy)
310{
311
312	thread0.td_turnstile = turnstile_alloc();
313}
314SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
315
316/*
317 * Set the owner of the lock this turnstile is attached to.
318 */
319static void
320turnstile_setowner(struct turnstile *ts, struct thread *owner)
321{
322
323	mtx_assert(&td_contested_lock, MA_OWNED);
324	MPASS(owner->td_proc->p_magic == P_MAGIC);
325	MPASS(ts->ts_owner == NULL);
326	ts->ts_owner = owner;
327	LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
328}
329
330/*
331 * Malloc a turnstile for a new thread, initialize it and return it.
332 */
333struct turnstile *
334turnstile_alloc(void)
335{
336	struct turnstile *ts;
337
338	ts = malloc(sizeof(struct turnstile), M_TURNSTILE, M_WAITOK | M_ZERO);
339	TAILQ_INIT(&ts->ts_blocked);
340	TAILQ_INIT(&ts->ts_pending);
341	LIST_INIT(&ts->ts_free);
342	return (ts);
343}
344
345/*
346 * Free a turnstile when a thread is destroyed.
347 */
348void
349turnstile_free(struct turnstile *ts)
350{
351
352	MPASS(ts != NULL);
353	MPASS(TAILQ_EMPTY(&ts->ts_blocked));
354	MPASS(TAILQ_EMPTY(&ts->ts_pending));
355	free(ts, M_TURNSTILE);
356}
357
358/*
359 * Look up the turnstile for a lock in the hash table locking the associated
360 * turnstile chain along the way.  Return with the turnstile chain locked.
361 * If no turnstile is found in the hash table, NULL is returned.
362 */
363struct turnstile *
364turnstile_lookup(struct lock_object *lock)
365{
366	struct turnstile_chain *tc;
367	struct turnstile *ts;
368
369	tc = TC_LOOKUP(lock);
370	mtx_lock_spin(&tc->tc_lock);
371	LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
372		if (ts->ts_lockobj == lock)
373			return (ts);
374	return (NULL);
375}
376
377/*
378 * Unlock the turnstile chain associated with a given lock.
379 */
380void
381turnstile_release(struct lock_object *lock)
382{
383	struct turnstile_chain *tc;
384
385	tc = TC_LOOKUP(lock);
386	mtx_unlock_spin(&tc->tc_lock);
387}
388
389/*
390 * Take ownership of a turnstile and adjust the priority of the new
391 * owner appropriately.
392 */
393void
394turnstile_claim(struct turnstile *ts)
395{
396	struct turnstile_chain *tc;
397	struct thread *td, *owner;
398
399	tc = TC_LOOKUP(ts->ts_lockobj);
400	mtx_assert(&tc->tc_lock, MA_OWNED);
401
402	owner = curthread;
403	mtx_lock_spin(&td_contested_lock);
404	turnstile_setowner(ts, owner);
405	mtx_unlock_spin(&td_contested_lock);
406
407	td = TAILQ_FIRST(&ts->ts_blocked);
408	MPASS(td != NULL);
409	MPASS(td->td_proc->p_magic == P_MAGIC);
410	mtx_unlock_spin(&tc->tc_lock);
411
412	/*
413	 * Update the priority of the new owner if needed.
414	 */
415	mtx_lock_spin(&sched_lock);
416	if (td->td_priority < owner->td_priority)
417		owner->td_priority = td->td_priority;
418	mtx_unlock_spin(&sched_lock);
419}
420
421/*
422 * Block the current thread on the turnstile ts.  This function will context
423 * switch and not return until this thread has been woken back up.  This
424 * function must be called with the appropriate turnstile chain locked and
425 * will return with it unlocked.
426 */
427void
428turnstile_wait(struct turnstile *ts, struct lock_object *lock,
429    struct thread *owner)
430{
431	struct turnstile_chain *tc;
432	struct thread *td, *td1;
433
434	td = curthread;
435	tc = TC_LOOKUP(lock);
436	mtx_assert(&tc->tc_lock, MA_OWNED);
437	MPASS(td->td_turnstile != NULL);
438	MPASS(owner != NULL);
439	MPASS(owner->td_proc->p_magic == P_MAGIC);
440
441	/* If the passed in turnstile is NULL, use this thread's turnstile. */
442	if (ts == NULL) {
443		ts = td->td_turnstile;
444		LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
445		KASSERT(TAILQ_EMPTY(&ts->ts_pending),
446		    ("thread's turnstile has pending threads"));
447		KASSERT(TAILQ_EMPTY(&ts->ts_blocked),
448		    ("thread's turnstile has a non-empty queue"));
449		KASSERT(LIST_EMPTY(&ts->ts_free),
450		    ("thread's turnstile has a non-empty free list"));
451		KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
452		ts->ts_lockobj = lock;
453		mtx_lock_spin(&td_contested_lock);
454		TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq);
455		turnstile_setowner(ts, owner);
456		mtx_unlock_spin(&td_contested_lock);
457	} else {
458		TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq)
459			if (td1->td_priority > td->td_priority)
460				break;
461		mtx_lock_spin(&td_contested_lock);
462		if (td1 != NULL)
463			TAILQ_INSERT_BEFORE(td1, td, td_lockq);
464		else
465			TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq);
466		mtx_unlock_spin(&td_contested_lock);
467		MPASS(td->td_turnstile != NULL);
468		LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
469		MPASS(owner == ts->ts_owner);
470	}
471#ifdef INVARIANTS
472	td->td_turnstile = NULL;
473#endif
474	mtx_unlock_spin(&tc->tc_lock);
475
476	mtx_lock_spin(&sched_lock);
477	/*
478	 * Handle race condition where a thread on another CPU that owns
479	 * lock 'lock' could have woken us in between us dropping the
480	 * turnstile chain lock and acquiring the sched_lock.
481	 */
482	if (td->td_flags & TDF_TSNOBLOCK) {
483		td->td_flags &= ~TDF_TSNOBLOCK;
484		mtx_unlock_spin(&sched_lock);
485		return;
486	}
487
488#ifdef notyet
489	/*
490	 * If we're borrowing an interrupted thread's VM context, we
491	 * must clean up before going to sleep.
492	 */
493	if (td->td_ithd != NULL) {
494		struct ithd *it = td->td_ithd;
495
496		if (it->it_interrupted) {
497			if (LOCK_LOG_TEST(lock, 0))
498				CTR3(KTR_LOCK, "%s: %p interrupted %p",
499				    __func__, it, it->it_interrupted);
500			intr_thd_fixup(it);
501		}
502	}
503#endif
504
505	/* Save who we are blocked on and switch. */
506	td->td_blocked = ts;
507	td->td_lockname = lock->lo_name;
508	TD_SET_LOCK(td);
509	propagate_priority(td);
510
511	if (LOCK_LOG_TEST(lock, 0))
512		CTR4(KTR_LOCK, "%s: td %p blocked on [%p] %s", __func__, td,
513		    lock, lock->lo_name);
514
515	td->td_proc->p_stats->p_ru.ru_nvcsw++;
516	mi_switch();
517
518	if (LOCK_LOG_TEST(lock, 0))
519		CTR4(KTR_LOCK, "%s: td %p free from blocked on [%p] %s",
520		    __func__, td, lock, lock->lo_name);
521
522	mtx_unlock_spin(&sched_lock);
523}
524
525/*
526 * Pick the highest priority thread on this turnstile and put it on the
527 * pending list.  This must be called with the turnstile chain locked.
528 */
529int
530turnstile_signal(struct turnstile *ts)
531{
532	struct turnstile_chain *tc;
533	struct thread *td;
534	int empty;
535
536	MPASS(ts != NULL);
537	MPASS(curthread->td_proc->p_magic == P_MAGIC);
538	MPASS(ts->ts_owner == curthread);
539	tc = TC_LOOKUP(ts->ts_lockobj);
540	mtx_assert(&tc->tc_lock, MA_OWNED);
541
542	/*
543	 * Pick the highest priority thread blocked on this lock and
544	 * move it to the pending list.
545	 */
546	td = TAILQ_FIRST(&ts->ts_blocked);
547	MPASS(td->td_proc->p_magic == P_MAGIC);
548	mtx_lock_spin(&td_contested_lock);
549	TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq);
550	mtx_unlock_spin(&td_contested_lock);
551	TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
552
553	/*
554	 * If the turnstile is now empty, remove it from its chain and
555	 * give it to the about-to-be-woken thread.  Otherwise take a
556	 * turnstile from the free list and give it to the thread.
557	 */
558	empty = TAILQ_EMPTY(&ts->ts_blocked);
559	if (empty)
560		MPASS(LIST_EMPTY(&ts->ts_free));
561	else
562		ts = LIST_FIRST(&ts->ts_free);
563	LIST_REMOVE(ts, ts_hash);
564	td->td_turnstile = ts;
565
566	return (empty);
567}
568
569/*
570 * Put all blocked threads on the pending list.  This must be called with
571 * the turnstile chain locked.
572 */
573void
574turnstile_wakeup(struct turnstile *ts)
575{
576	struct turnstile_chain *tc;
577	struct turnstile *ts1;
578	struct thread *td;
579
580	MPASS(ts != NULL);
581	MPASS(curthread->td_proc->p_magic == P_MAGIC);
582	MPASS(ts->ts_owner == curthread);
583	tc = TC_LOOKUP(ts->ts_lockobj);
584	mtx_assert(&tc->tc_lock, MA_OWNED);
585
586	/*
587	 * Transfer the blocked list to the pending list.
588	 */
589	mtx_lock_spin(&td_contested_lock);
590	TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked, td_lockq);
591	mtx_unlock_spin(&td_contested_lock);
592
593	/*
594	 * Give a turnstile to each thread.  The last thread gets
595	 * this turnstile.
596	 */
597	TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
598		if (LIST_EMPTY(&ts->ts_free)) {
599			MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
600			ts1 = ts;
601		} else
602			ts1 = LIST_FIRST(&ts->ts_free);
603		LIST_REMOVE(ts1, ts_hash);
604		td->td_turnstile = ts1;
605	}
606}
607
608/*
609 * Wakeup all threads on the pending list and adjust the priority of the
610 * current thread appropriately.  This must be called with the turnstile
611 * chain locked.
612 */
613void
614turnstile_unpend(struct turnstile *ts)
615{
616	TAILQ_HEAD( ,thread) pending_threads;
617	struct turnstile_chain *tc;
618	struct thread *td;
619	int cp, pri;
620
621	MPASS(ts != NULL);
622	MPASS(ts->ts_owner == curthread);
623	tc = TC_LOOKUP(ts->ts_lockobj);
624	mtx_assert(&tc->tc_lock, MA_OWNED);
625	MPASS(!TAILQ_EMPTY(&ts->ts_pending));
626
627	/*
628	 * Move the list of pending threads out of the turnstile and
629	 * into a local variable.
630	 */
631	TAILQ_INIT(&pending_threads);
632	TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
633#ifdef INVARIANTS
634	if (TAILQ_EMPTY(&ts->ts_blocked))
635		ts->ts_lockobj = NULL;
636#endif
637
638	/*
639	 * Remove the turnstile from this thread's list of contested locks
640	 * since this thread doesn't own it anymore.  New threads will
641	 * not be blocking on the turnstile until it is claimed by a new
642	 * owner.
643	 */
644	mtx_lock_spin(&td_contested_lock);
645	ts->ts_owner = NULL;
646	LIST_REMOVE(ts, ts_link);
647	mtx_unlock_spin(&td_contested_lock);
648	mtx_unlock_spin(&tc->tc_lock);
649
650	/*
651	 * Adjust the priority of curthread based on other contested
652	 * locks it owns.  Don't lower the priority below the base
653	 * priority however.
654	 */
655	td = curthread;
656	pri = PRI_MAX;
657	mtx_lock_spin(&sched_lock);
658	mtx_lock_spin(&td_contested_lock);
659	LIST_FOREACH(ts, &td->td_contested, ts_link) {
660		cp = TAILQ_FIRST(&ts->ts_blocked)->td_priority;
661		if (cp < pri)
662			pri = cp;
663	}
664	mtx_unlock_spin(&td_contested_lock);
665	if (pri > td->td_base_pri)
666		pri = td->td_base_pri;
667	td->td_priority = pri;
668
669	/*
670	 * Wake up all the pending threads.  If a thread is not blocked
671	 * on a lock, then it is currently executing on another CPU in
672	 * turnstile_wait().  Set a flag to force it to try to acquire
673	 * the lock again instead of blocking.
674	 */
675	while (!TAILQ_EMPTY(&pending_threads)) {
676		td = TAILQ_FIRST(&pending_threads);
677		TAILQ_REMOVE(&pending_threads, td, td_lockq);
678		MPASS(td->td_proc->p_magic == P_MAGIC);
679		if (TD_ON_LOCK(td)) {
680			td->td_blocked = NULL;
681			td->td_lockname = NULL;
682			TD_CLR_LOCK(td);
683			MPASS(TD_CAN_RUN(td));
684			setrunqueue(td);
685		} else {
686			td->td_flags |= TDF_TSNOBLOCK;
687			MPASS(TD_IS_RUNNING(td));
688		}
689	}
690	mtx_unlock_spin(&sched_lock);
691}
692
693/*
694 * Return the first thread in a turnstile.
695 */
696struct thread *
697turnstile_head(struct turnstile *ts)
698{
699#ifdef INVARIANTS
700	struct turnstile_chain *tc;
701
702	MPASS(ts != NULL);
703	tc = TC_LOOKUP(ts->ts_lockobj);
704	mtx_assert(&tc->tc_lock, MA_OWNED);
705#endif
706	return (TAILQ_FIRST(&ts->ts_blocked));
707}
708
709/*
710 * Returns true if a turnstile is empty.
711 */
712int
713turnstile_empty(struct turnstile *ts)
714{
715#ifdef INVARIANTS
716	struct turnstile_chain *tc;
717
718	MPASS(ts != NULL);
719	tc = TC_LOOKUP(ts->ts_lockobj);
720	mtx_assert(&tc->tc_lock, MA_OWNED);
721#endif
722	return (TAILQ_EMPTY(&ts->ts_blocked));
723}
724