subr_turnstile.c revision 123364
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 123364 2003-12-09 21:14:31Z 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		 * This thread may not be blocked on this turnstile anymore
237		 * but instead might already be woken up on another CPU
238		 * that is waiting on sched_lock in turnstile_unpend() to
239		 * finish waking this thread up.  We can detect this case
240		 * by checking to see if this thread has been given a
241		 * turnstile by either turnstile_signal() or
242		 * turnstile_wakeup().  In this case, treat the thread as
243		 * if it was already running.
244		 */
245		if (td->td_turnstile != NULL) {
246			mtx_unlock_spin(&tc->tc_lock);
247			return;
248		}
249
250		/*
251		 * Check if the thread needs to be moved up on
252		 * the blocked chain.  It doesn't need to be moved
253		 * if it is already at the head of the list or if
254		 * the item in front of it still has a higher priority.
255		 */
256		if (td == TAILQ_FIRST(&ts->ts_blocked)) {
257			mtx_unlock_spin(&tc->tc_lock);
258			continue;
259		}
260
261		td1 = TAILQ_PREV(td, threadqueue, td_lockq);
262		if (td1->td_priority <= pri) {
263			mtx_unlock_spin(&tc->tc_lock);
264			continue;
265		}
266
267		/*
268		 * Remove thread from blocked chain and determine where
269		 * it should be moved up to.  Since we know that td1 has
270		 * a lower priority than td, we know that at least one
271		 * thread in the chain has a lower priority and that
272		 * td1 will thus not be NULL after the loop.
273		 */
274		mtx_lock_spin(&td_contested_lock);
275		TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq);
276		TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq) {
277			MPASS(td1->td_proc->p_magic == P_MAGIC);
278			if (td1->td_priority > pri)
279				break;
280		}
281
282		MPASS(td1 != NULL);
283		TAILQ_INSERT_BEFORE(td1, td, td_lockq);
284		mtx_unlock_spin(&td_contested_lock);
285		CTR4(KTR_LOCK,
286		    "propagate_priority: td %p moved before %p on [%p] %s",
287		    td, td1, ts->ts_lockobj, ts->ts_lockobj->lo_name);
288		mtx_unlock_spin(&tc->tc_lock);
289	}
290}
291
292/*
293 * Early initialization of turnstiles.  This is not done via a SYSINIT()
294 * since this needs to be initialized very early when mutexes are first
295 * initialized.
296 */
297void
298init_turnstiles(void)
299{
300	int i;
301
302	for (i = 0; i < TC_TABLESIZE; i++) {
303		LIST_INIT(&turnstile_chains[i].tc_turnstiles);
304		mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
305		    NULL, MTX_SPIN);
306	}
307	mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
308	thread0.td_turnstile = NULL;
309}
310
311static void
312init_turnstile0(void *dummy)
313{
314
315	thread0.td_turnstile = turnstile_alloc();
316}
317SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
318
319/*
320 * Set the owner of the lock this turnstile is attached to.
321 */
322static void
323turnstile_setowner(struct turnstile *ts, struct thread *owner)
324{
325
326	mtx_assert(&td_contested_lock, MA_OWNED);
327	MPASS(owner->td_proc->p_magic == P_MAGIC);
328	MPASS(ts->ts_owner == NULL);
329	ts->ts_owner = owner;
330	LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
331}
332
333/*
334 * Malloc a turnstile for a new thread, initialize it and return it.
335 */
336struct turnstile *
337turnstile_alloc(void)
338{
339	struct turnstile *ts;
340
341	ts = malloc(sizeof(struct turnstile), M_TURNSTILE, M_WAITOK | M_ZERO);
342	TAILQ_INIT(&ts->ts_blocked);
343	TAILQ_INIT(&ts->ts_pending);
344	LIST_INIT(&ts->ts_free);
345	return (ts);
346}
347
348/*
349 * Free a turnstile when a thread is destroyed.
350 */
351void
352turnstile_free(struct turnstile *ts)
353{
354
355	MPASS(ts != NULL);
356	MPASS(TAILQ_EMPTY(&ts->ts_blocked));
357	MPASS(TAILQ_EMPTY(&ts->ts_pending));
358	free(ts, M_TURNSTILE);
359}
360
361/*
362 * Look up the turnstile for a lock in the hash table locking the associated
363 * turnstile chain along the way.  Return with the turnstile chain locked.
364 * If no turnstile is found in the hash table, NULL is returned.
365 */
366struct turnstile *
367turnstile_lookup(struct lock_object *lock)
368{
369	struct turnstile_chain *tc;
370	struct turnstile *ts;
371
372	tc = TC_LOOKUP(lock);
373	mtx_lock_spin(&tc->tc_lock);
374	LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
375		if (ts->ts_lockobj == lock)
376			return (ts);
377	return (NULL);
378}
379
380/*
381 * Unlock the turnstile chain associated with a given lock.
382 */
383void
384turnstile_release(struct lock_object *lock)
385{
386	struct turnstile_chain *tc;
387
388	tc = TC_LOOKUP(lock);
389	mtx_unlock_spin(&tc->tc_lock);
390}
391
392/*
393 * Take ownership of a turnstile and adjust the priority of the new
394 * owner appropriately.
395 */
396void
397turnstile_claim(struct turnstile *ts)
398{
399	struct turnstile_chain *tc;
400	struct thread *td, *owner;
401
402	tc = TC_LOOKUP(ts->ts_lockobj);
403	mtx_assert(&tc->tc_lock, MA_OWNED);
404
405	owner = curthread;
406	mtx_lock_spin(&td_contested_lock);
407	turnstile_setowner(ts, owner);
408	mtx_unlock_spin(&td_contested_lock);
409
410	td = TAILQ_FIRST(&ts->ts_blocked);
411	MPASS(td != NULL);
412	MPASS(td->td_proc->p_magic == P_MAGIC);
413	mtx_unlock_spin(&tc->tc_lock);
414
415	/*
416	 * Update the priority of the new owner if needed.
417	 */
418	mtx_lock_spin(&sched_lock);
419	if (td->td_priority < owner->td_priority)
420		owner->td_priority = td->td_priority;
421	mtx_unlock_spin(&sched_lock);
422}
423
424/*
425 * Block the current thread on the turnstile ts.  This function will context
426 * switch and not return until this thread has been woken back up.  This
427 * function must be called with the appropriate turnstile chain locked and
428 * will return with it unlocked.
429 */
430void
431turnstile_wait(struct turnstile *ts, struct lock_object *lock,
432    struct thread *owner)
433{
434	struct turnstile_chain *tc;
435	struct thread *td, *td1;
436
437	td = curthread;
438	tc = TC_LOOKUP(lock);
439	mtx_assert(&tc->tc_lock, MA_OWNED);
440	MPASS(td->td_turnstile != NULL);
441	MPASS(owner != NULL);
442	MPASS(owner->td_proc->p_magic == P_MAGIC);
443
444	/* If the passed in turnstile is NULL, use this thread's turnstile. */
445	if (ts == NULL) {
446		ts = td->td_turnstile;
447		LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
448		KASSERT(TAILQ_EMPTY(&ts->ts_pending),
449		    ("thread's turnstile has pending threads"));
450		KASSERT(TAILQ_EMPTY(&ts->ts_blocked),
451		    ("thread's turnstile has a non-empty queue"));
452		KASSERT(LIST_EMPTY(&ts->ts_free),
453		    ("thread's turnstile has a non-empty free list"));
454		KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
455		ts->ts_lockobj = lock;
456		mtx_lock_spin(&td_contested_lock);
457		TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq);
458		turnstile_setowner(ts, owner);
459		mtx_unlock_spin(&td_contested_lock);
460	} else {
461		TAILQ_FOREACH(td1, &ts->ts_blocked, td_lockq)
462			if (td1->td_priority > td->td_priority)
463				break;
464		mtx_lock_spin(&td_contested_lock);
465		if (td1 != NULL)
466			TAILQ_INSERT_BEFORE(td1, td, td_lockq);
467		else
468			TAILQ_INSERT_TAIL(&ts->ts_blocked, td, td_lockq);
469		mtx_unlock_spin(&td_contested_lock);
470		MPASS(td->td_turnstile != NULL);
471		LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
472		MPASS(owner == ts->ts_owner);
473	}
474	td->td_turnstile = NULL;
475	mtx_unlock_spin(&tc->tc_lock);
476
477	mtx_lock_spin(&sched_lock);
478	/*
479	 * Handle race condition where a thread on another CPU that owns
480	 * lock 'lock' could have woken us in between us dropping the
481	 * turnstile chain lock and acquiring the sched_lock.
482	 */
483	if (td->td_flags & TDF_TSNOBLOCK) {
484		td->td_flags &= ~TDF_TSNOBLOCK;
485		mtx_unlock_spin(&sched_lock);
486		return;
487	}
488
489#ifdef notyet
490	/*
491	 * If we're borrowing an interrupted thread's VM context, we
492	 * must clean up before going to sleep.
493	 */
494	if (td->td_ithd != NULL) {
495		struct ithd *it = td->td_ithd;
496
497		if (it->it_interrupted) {
498			if (LOCK_LOG_TEST(lock, 0))
499				CTR3(KTR_LOCK, "%s: %p interrupted %p",
500				    __func__, it, it->it_interrupted);
501			intr_thd_fixup(it);
502		}
503	}
504#endif
505
506	/* Save who we are blocked on and switch. */
507	td->td_blocked = ts;
508	td->td_lockname = lock->lo_name;
509	TD_SET_LOCK(td);
510	propagate_priority(td);
511
512	if (LOCK_LOG_TEST(lock, 0))
513		CTR4(KTR_LOCK, "%s: td %p blocked on [%p] %s", __func__, td,
514		    lock, lock->lo_name);
515
516	td->td_proc->p_stats->p_ru.ru_nvcsw++;
517	mi_switch();
518
519	if (LOCK_LOG_TEST(lock, 0))
520		CTR4(KTR_LOCK, "%s: td %p free from blocked on [%p] %s",
521		    __func__, td, lock, lock->lo_name);
522
523	mtx_unlock_spin(&sched_lock);
524}
525
526/*
527 * Pick the highest priority thread on this turnstile and put it on the
528 * pending list.  This must be called with the turnstile chain locked.
529 */
530int
531turnstile_signal(struct turnstile *ts)
532{
533	struct turnstile_chain *tc;
534	struct thread *td;
535	int empty;
536
537	MPASS(ts != NULL);
538	MPASS(curthread->td_proc->p_magic == P_MAGIC);
539	MPASS(ts->ts_owner == curthread);
540	tc = TC_LOOKUP(ts->ts_lockobj);
541	mtx_assert(&tc->tc_lock, MA_OWNED);
542
543	/*
544	 * Pick the highest priority thread blocked on this lock and
545	 * move it to the pending list.
546	 */
547	td = TAILQ_FIRST(&ts->ts_blocked);
548	MPASS(td->td_proc->p_magic == P_MAGIC);
549	mtx_lock_spin(&td_contested_lock);
550	TAILQ_REMOVE(&ts->ts_blocked, td, td_lockq);
551	mtx_unlock_spin(&td_contested_lock);
552	TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
553
554	/*
555	 * If the turnstile is now empty, remove it from its chain and
556	 * give it to the about-to-be-woken thread.  Otherwise take a
557	 * turnstile from the free list and give it to the thread.
558	 */
559	empty = TAILQ_EMPTY(&ts->ts_blocked);
560	if (empty)
561		MPASS(LIST_EMPTY(&ts->ts_free));
562	else
563		ts = LIST_FIRST(&ts->ts_free);
564	MPASS(ts != NULL);
565	LIST_REMOVE(ts, ts_hash);
566	td->td_turnstile = ts;
567
568	return (empty);
569}
570
571/*
572 * Put all blocked threads on the pending list.  This must be called with
573 * the turnstile chain locked.
574 */
575void
576turnstile_wakeup(struct turnstile *ts)
577{
578	struct turnstile_chain *tc;
579	struct turnstile *ts1;
580	struct thread *td;
581
582	MPASS(ts != NULL);
583	MPASS(curthread->td_proc->p_magic == P_MAGIC);
584	MPASS(ts->ts_owner == curthread);
585	tc = TC_LOOKUP(ts->ts_lockobj);
586	mtx_assert(&tc->tc_lock, MA_OWNED);
587
588	/*
589	 * Transfer the blocked list to the pending list.
590	 */
591	mtx_lock_spin(&td_contested_lock);
592	TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked, td_lockq);
593	mtx_unlock_spin(&td_contested_lock);
594
595	/*
596	 * Give a turnstile to each thread.  The last thread gets
597	 * this turnstile.
598	 */
599	TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
600		if (LIST_EMPTY(&ts->ts_free)) {
601			MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
602			ts1 = ts;
603		} else
604			ts1 = LIST_FIRST(&ts->ts_free);
605		MPASS(ts1 != NULL);
606		LIST_REMOVE(ts1, ts_hash);
607		td->td_turnstile = ts1;
608	}
609}
610
611/*
612 * Wakeup all threads on the pending list and adjust the priority of the
613 * current thread appropriately.  This must be called with the turnstile
614 * chain locked.
615 */
616void
617turnstile_unpend(struct turnstile *ts)
618{
619	TAILQ_HEAD( ,thread) pending_threads;
620	struct turnstile_chain *tc;
621	struct thread *td;
622	int cp, pri;
623
624	MPASS(ts != NULL);
625	MPASS(ts->ts_owner == curthread);
626	tc = TC_LOOKUP(ts->ts_lockobj);
627	mtx_assert(&tc->tc_lock, MA_OWNED);
628	MPASS(!TAILQ_EMPTY(&ts->ts_pending));
629
630	/*
631	 * Move the list of pending threads out of the turnstile and
632	 * into a local variable.
633	 */
634	TAILQ_INIT(&pending_threads);
635	TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
636#ifdef INVARIANTS
637	if (TAILQ_EMPTY(&ts->ts_blocked))
638		ts->ts_lockobj = NULL;
639#endif
640
641	/*
642	 * Remove the turnstile from this thread's list of contested locks
643	 * since this thread doesn't own it anymore.  New threads will
644	 * not be blocking on the turnstile until it is claimed by a new
645	 * owner.
646	 */
647	mtx_lock_spin(&td_contested_lock);
648	ts->ts_owner = NULL;
649	LIST_REMOVE(ts, ts_link);
650	mtx_unlock_spin(&td_contested_lock);
651	mtx_unlock_spin(&tc->tc_lock);
652
653	/*
654	 * Adjust the priority of curthread based on other contested
655	 * locks it owns.  Don't lower the priority below the base
656	 * priority however.
657	 */
658	td = curthread;
659	pri = PRI_MAX;
660	mtx_lock_spin(&sched_lock);
661	mtx_lock_spin(&td_contested_lock);
662	LIST_FOREACH(ts, &td->td_contested, ts_link) {
663		cp = TAILQ_FIRST(&ts->ts_blocked)->td_priority;
664		if (cp < pri)
665			pri = cp;
666	}
667	mtx_unlock_spin(&td_contested_lock);
668	if (pri > td->td_base_pri)
669		pri = td->td_base_pri;
670	td->td_priority = pri;
671
672	/*
673	 * Wake up all the pending threads.  If a thread is not blocked
674	 * on a lock, then it is currently executing on another CPU in
675	 * turnstile_wait() or sitting on a run queue waiting to resume
676	 * in turnstile_wait().  Set a flag to force it to try to acquire
677	 * the lock again instead of blocking.
678	 */
679	while (!TAILQ_EMPTY(&pending_threads)) {
680		td = TAILQ_FIRST(&pending_threads);
681		TAILQ_REMOVE(&pending_threads, td, td_lockq);
682		MPASS(td->td_proc->p_magic == P_MAGIC);
683		if (TD_ON_LOCK(td)) {
684			td->td_blocked = NULL;
685			td->td_lockname = NULL;
686			TD_CLR_LOCK(td);
687			MPASS(TD_CAN_RUN(td));
688			setrunqueue(td);
689		} else {
690			td->td_flags |= TDF_TSNOBLOCK;
691			MPASS(TD_IS_RUNNING(td) || TD_ON_RUNQ(td));
692		}
693	}
694	mtx_unlock_spin(&sched_lock);
695}
696
697/*
698 * Return the first thread in a turnstile.
699 */
700struct thread *
701turnstile_head(struct turnstile *ts)
702{
703#ifdef INVARIANTS
704	struct turnstile_chain *tc;
705
706	MPASS(ts != NULL);
707	tc = TC_LOOKUP(ts->ts_lockobj);
708	mtx_assert(&tc->tc_lock, MA_OWNED);
709#endif
710	return (TAILQ_FIRST(&ts->ts_blocked));
711}
712
713/*
714 * Returns true if a turnstile is empty.
715 */
716int
717turnstile_empty(struct turnstile *ts)
718{
719#ifdef INVARIANTS
720	struct turnstile_chain *tc;
721
722	MPASS(ts != NULL);
723	tc = TC_LOOKUP(ts->ts_lockobj);
724	mtx_assert(&tc->tc_lock, MA_OWNED);
725#endif
726	return (TAILQ_EMPTY(&ts->ts_blocked));
727}
728