2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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
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
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 $
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.
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.
49 * Each time a thread is created, a turnstile is allocated from a UMA zone
50 * and attached to that thread. When a thread blocks on a lock, if it is the
51 * first thread to block, it lends its turnstile to the lock. If the lock
52 * already has a turnstile, then it gives its turnstile to the lock's
53 * turnstile's free list. When a thread is woken up, it takes a turnstile from
54 * the free list if there are any other waiters. If it is the only thread
55 * blocked on the lock, then it reclaims the turnstile associated with the lock
56 * and removes it from the hash table.
59 #include <sys/cdefs.h>
60 __FBSDID("$FreeBSD$");
63 #include "opt_kdtrace.h"
64 #include "opt_turnstile_profiling.h"
65 #include "opt_sched.h"
67 #include <sys/param.h>
68 #include <sys/systm.h>
70 #include <sys/kernel.h>
73 #include <sys/mutex.h>
75 #include <sys/queue.h>
76 #include <sys/sched.h>
78 #include <sys/sysctl.h>
79 #include <sys/turnstile.h>
85 #include <sys/lockmgr.h>
90 * Constants for the hash table of turnstile chains. TC_SHIFT is a magic
91 * number chosen because the sleep queue's use the same value for the
92 * shift. Basically, we ignore the lower 8 bits of the address.
93 * TC_TABLESIZE must be a power of two for TC_MASK to work properly.
95 #define TC_TABLESIZE 128 /* Must be power of 2. */
96 #define TC_MASK (TC_TABLESIZE - 1)
98 #define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK)
99 #define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)]
102 * There are three different lists of turnstiles as follows. The list
103 * connected by ts_link entries is a per-thread list of all the turnstiles
104 * attached to locks that we own. This is used to fixup our priority when
105 * a lock is released. The other two lists use the ts_hash entries. The
106 * first of these two is the turnstile chain list that a turnstile is on
107 * when it is attached to a lock. The second list to use ts_hash is the
108 * free list hung off of a turnstile that is attached to a lock.
110 * Each turnstile contains three lists of threads. The two ts_blocked lists
111 * are linked list of threads blocked on the turnstile's lock. One list is
112 * for exclusive waiters, and the other is for shared waiters. The
113 * ts_pending list is a linked list of threads previously awakened by
114 * turnstile_signal() or turnstile_wait() that are waiting to be put on
118 * c - turnstile chain lock
119 * q - td_contested lock
122 struct mtx ts_lock; /* Spin lock for self. */
123 struct threadqueue ts_blocked[2]; /* (c + q) Blocked threads. */
124 struct threadqueue ts_pending; /* (c) Pending threads. */
125 LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */
126 LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */
127 LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */
128 struct lock_object *ts_lockobj; /* (c) Lock we reference. */
129 struct thread *ts_owner; /* (c + q) Who owns the lock. */
132 struct turnstile_chain {
133 LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */
134 struct mtx tc_lock; /* Spin lock for this chain. */
135 #ifdef TURNSTILE_PROFILING
136 u_int tc_depth; /* Length of tc_queues. */
137 u_int tc_max_depth; /* Max length of tc_queues. */
141 #ifdef TURNSTILE_PROFILING
142 u_int turnstile_max_depth;
143 static SYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD, 0,
144 "turnstile profiling");
145 static SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0,
146 "turnstile chain stats");
147 SYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD,
148 &turnstile_max_depth, 0, "maximum depth achieved of a single chain");
150 static struct mtx td_contested_lock;
151 static struct turnstile_chain turnstile_chains[TC_TABLESIZE];
152 static uma_zone_t turnstile_zone;
155 * Prototypes for non-exported routines.
157 static void init_turnstile0(void *dummy);
158 #ifdef TURNSTILE_PROFILING
159 static void init_turnstile_profiling(void *arg);
161 static void propagate_priority(struct thread *td);
162 static int turnstile_adjust_thread(struct turnstile *ts,
164 static struct thread *turnstile_first_waiter(struct turnstile *ts);
165 static void turnstile_setowner(struct turnstile *ts, struct thread *owner);
167 static void turnstile_dtor(void *mem, int size, void *arg);
169 static int turnstile_init(void *mem, int size, int flags);
170 static void turnstile_fini(void *mem, int size);
172 SDT_PROVIDER_DECLARE(sched);
173 SDT_PROBE_DEFINE(sched, , , sleep, sleep);
174 SDT_PROBE_DEFINE2(sched, , , wakeup, wakeup, "struct thread *",
178 * Walks the chain of turnstiles and their owners to propagate the priority
179 * of the thread being blocked to all the threads holding locks that have to
180 * release their locks before this thread can run again.
183 propagate_priority(struct thread *td)
185 struct turnstile *ts;
188 THREAD_LOCK_ASSERT(td, MA_OWNED);
189 pri = td->td_priority;
191 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
193 * Grab a recursive lock on this turnstile chain so it stays locked
194 * for the whole operation. The caller expects us to return with
195 * the original lock held. We only ever lock down the chain so
196 * the lock order is constant.
198 mtx_lock_spin(&ts->ts_lock);
204 * This might be a read lock with no owner. There's
205 * not much we can do, so just bail.
207 mtx_unlock_spin(&ts->ts_lock);
211 thread_lock_flags(td, MTX_DUPOK);
212 mtx_unlock_spin(&ts->ts_lock);
213 MPASS(td->td_proc != NULL);
214 MPASS(td->td_proc->p_magic == P_MAGIC);
217 * If the thread is asleep, then we are probably about
218 * to deadlock. To make debugging this easier, show
219 * backtrace of misbehaving thread and panic to not
220 * leave the kernel deadlocked.
222 if (TD_IS_SLEEPING(td)) {
224 "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
225 td->td_tid, td->td_proc->p_pid);
226 kdb_backtrace_thread(td);
227 panic("sleeping thread");
231 * If this thread already has higher priority than the
232 * thread that is being blocked, we are finished.
234 if (td->td_priority <= pri) {
240 * Bump this thread's priority.
242 sched_lend_prio(td, pri);
245 * If lock holder is actually running or on the run queue
248 if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) {
249 MPASS(td->td_blocked == NULL);
256 * For UP, we check to see if td is curthread (this shouldn't
257 * ever happen however as it would mean we are in a deadlock.)
259 KASSERT(td != curthread, ("Deadlock detected"));
263 * If we aren't blocked on a lock, we should be.
265 KASSERT(TD_ON_LOCK(td), (
266 "thread %d(%s):%d holds %s but isn't blocked on a lock\n",
267 td->td_tid, td->td_name, td->td_state,
268 ts->ts_lockobj->lo_name));
271 * Pick up the lock that td is blocked on.
275 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
276 /* Resort td on the list if needed. */
277 if (!turnstile_adjust_thread(ts, td)) {
278 mtx_unlock_spin(&ts->ts_lock);
281 /* The thread lock is released as ts lock above. */
286 * Adjust the thread's position on a turnstile after its priority has been
290 turnstile_adjust_thread(struct turnstile *ts, struct thread *td)
292 struct thread *td1, *td2;
295 THREAD_LOCK_ASSERT(td, MA_OWNED);
296 MPASS(TD_ON_LOCK(td));
299 * This thread may not be blocked on this turnstile anymore
300 * but instead might already be woken up on another CPU
301 * that is waiting on the thread lock in turnstile_unpend() to
302 * finish waking this thread up. We can detect this case
303 * by checking to see if this thread has been given a
304 * turnstile by either turnstile_signal() or
305 * turnstile_broadcast(). In this case, treat the thread as
306 * if it was already running.
308 if (td->td_turnstile != NULL)
312 * Check if the thread needs to be moved on the blocked chain.
313 * It needs to be moved if either its priority is lower than
314 * the previous thread or higher than the next thread.
316 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
317 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
318 td2 = TAILQ_NEXT(td, td_lockq);
319 if ((td1 != NULL && td->td_priority < td1->td_priority) ||
320 (td2 != NULL && td->td_priority > td2->td_priority)) {
323 * Remove thread from blocked chain and determine where
324 * it should be moved to.
326 queue = td->td_tsqueue;
327 MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE);
328 mtx_lock_spin(&td_contested_lock);
329 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
330 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) {
331 MPASS(td1->td_proc->p_magic == P_MAGIC);
332 if (td1->td_priority > td->td_priority)
337 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
339 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
340 mtx_unlock_spin(&td_contested_lock);
343 "turnstile_adjust_thread: td %d put at tail on [%p] %s",
344 td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name);
347 "turnstile_adjust_thread: td %d moved before %d on [%p] %s",
348 td->td_tid, td1->td_tid, ts->ts_lockobj,
349 ts->ts_lockobj->lo_name);
355 * Early initialization of turnstiles. This is not done via a SYSINIT()
356 * since this needs to be initialized very early when mutexes are first
360 init_turnstiles(void)
364 for (i = 0; i < TC_TABLESIZE; i++) {
365 LIST_INIT(&turnstile_chains[i].tc_turnstiles);
366 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
369 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
370 LIST_INIT(&thread0.td_contested);
371 thread0.td_turnstile = NULL;
374 #ifdef TURNSTILE_PROFILING
376 init_turnstile_profiling(void *arg)
378 struct sysctl_oid *chain_oid;
382 for (i = 0; i < TC_TABLESIZE; i++) {
383 snprintf(chain_name, sizeof(chain_name), "%d", i);
384 chain_oid = SYSCTL_ADD_NODE(NULL,
385 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO,
386 chain_name, CTLFLAG_RD, NULL, "turnstile chain stats");
387 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
388 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0,
390 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
391 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth,
395 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
396 init_turnstile_profiling, NULL);
400 init_turnstile0(void *dummy)
403 turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile),
410 turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE);
411 thread0.td_turnstile = turnstile_alloc();
413 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
416 * Update a thread on the turnstile list after it's priority has been changed.
417 * The old priority is passed in as an argument.
420 turnstile_adjust(struct thread *td, u_char oldpri)
422 struct turnstile *ts;
424 MPASS(TD_ON_LOCK(td));
427 * Pick up the lock that td is blocked on.
431 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
432 mtx_assert(&ts->ts_lock, MA_OWNED);
434 /* Resort the turnstile on the list. */
435 if (!turnstile_adjust_thread(ts, td))
438 * If our priority was lowered and we are at the head of the
439 * turnstile, then propagate our new priority up the chain.
440 * Note that we currently don't try to revoke lent priorities
441 * when our priority goes up.
443 MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE ||
444 td->td_tsqueue == TS_SHARED_QUEUE);
445 if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) &&
446 td->td_priority < oldpri) {
447 propagate_priority(td);
452 * Set the owner of the lock this turnstile is attached to.
455 turnstile_setowner(struct turnstile *ts, struct thread *owner)
458 mtx_assert(&td_contested_lock, MA_OWNED);
459 MPASS(ts->ts_owner == NULL);
461 /* A shared lock might not have an owner. */
465 MPASS(owner->td_proc->p_magic == P_MAGIC);
466 ts->ts_owner = owner;
467 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
472 * UMA zone item deallocator.
475 turnstile_dtor(void *mem, int size, void *arg)
477 struct turnstile *ts;
480 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]));
481 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
482 MPASS(TAILQ_EMPTY(&ts->ts_pending));
487 * UMA zone item initializer.
490 turnstile_init(void *mem, int size, int flags)
492 struct turnstile *ts;
496 TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
497 TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]);
498 TAILQ_INIT(&ts->ts_pending);
499 LIST_INIT(&ts->ts_free);
500 mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN | MTX_RECURSE);
505 turnstile_fini(void *mem, int size)
507 struct turnstile *ts;
510 mtx_destroy(&ts->ts_lock);
514 * Get a turnstile for a new thread.
517 turnstile_alloc(void)
520 return (uma_zalloc(turnstile_zone, M_WAITOK));
524 * Free a turnstile when a thread is destroyed.
527 turnstile_free(struct turnstile *ts)
530 uma_zfree(turnstile_zone, ts);
534 * Lock the turnstile chain associated with the specified lock.
537 turnstile_chain_lock(struct lock_object *lock)
539 struct turnstile_chain *tc;
541 tc = TC_LOOKUP(lock);
542 mtx_lock_spin(&tc->tc_lock);
546 turnstile_trywait(struct lock_object *lock)
548 struct turnstile_chain *tc;
549 struct turnstile *ts;
551 tc = TC_LOOKUP(lock);
552 mtx_lock_spin(&tc->tc_lock);
553 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
554 if (ts->ts_lockobj == lock) {
555 mtx_lock_spin(&ts->ts_lock);
559 ts = curthread->td_turnstile;
561 mtx_lock_spin(&ts->ts_lock);
562 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
563 ts->ts_lockobj = lock;
569 turnstile_cancel(struct turnstile *ts)
571 struct turnstile_chain *tc;
572 struct lock_object *lock;
574 mtx_assert(&ts->ts_lock, MA_OWNED);
576 mtx_unlock_spin(&ts->ts_lock);
577 lock = ts->ts_lockobj;
578 if (ts == curthread->td_turnstile)
579 ts->ts_lockobj = NULL;
580 tc = TC_LOOKUP(lock);
581 mtx_unlock_spin(&tc->tc_lock);
585 * Look up the turnstile for a lock in the hash table locking the associated
586 * turnstile chain along the way. If no turnstile is found in the hash
587 * table, NULL is returned.
590 turnstile_lookup(struct lock_object *lock)
592 struct turnstile_chain *tc;
593 struct turnstile *ts;
595 tc = TC_LOOKUP(lock);
596 mtx_assert(&tc->tc_lock, MA_OWNED);
597 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
598 if (ts->ts_lockobj == lock) {
599 mtx_lock_spin(&ts->ts_lock);
606 * Unlock the turnstile chain associated with a given lock.
609 turnstile_chain_unlock(struct lock_object *lock)
611 struct turnstile_chain *tc;
613 tc = TC_LOOKUP(lock);
614 mtx_unlock_spin(&tc->tc_lock);
618 * Return a pointer to the thread waiting on this turnstile with the
619 * most important priority or NULL if the turnstile has no waiters.
621 static struct thread *
622 turnstile_first_waiter(struct turnstile *ts)
624 struct thread *std, *xtd;
626 std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]);
627 xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
628 if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority))
634 * Take ownership of a turnstile and adjust the priority of the new
635 * owner appropriately.
638 turnstile_claim(struct turnstile *ts)
640 struct thread *td, *owner;
641 struct turnstile_chain *tc;
643 mtx_assert(&ts->ts_lock, MA_OWNED);
644 MPASS(ts != curthread->td_turnstile);
647 mtx_lock_spin(&td_contested_lock);
648 turnstile_setowner(ts, owner);
649 mtx_unlock_spin(&td_contested_lock);
651 td = turnstile_first_waiter(ts);
653 MPASS(td->td_proc->p_magic == P_MAGIC);
654 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
657 * Update the priority of the new owner if needed.
660 if (td->td_priority < owner->td_priority)
661 sched_lend_prio(owner, td->td_priority);
662 thread_unlock(owner);
663 tc = TC_LOOKUP(ts->ts_lockobj);
664 mtx_unlock_spin(&ts->ts_lock);
665 mtx_unlock_spin(&tc->tc_lock);
669 * Block the current thread on the turnstile assicated with 'lock'. This
670 * function will context switch and not return until this thread has been
671 * woken back up. This function must be called with the appropriate
672 * turnstile chain locked and will return with it unlocked.
675 turnstile_wait(struct turnstile *ts, struct thread *owner, int queue)
677 struct turnstile_chain *tc;
678 struct thread *td, *td1;
679 struct lock_object *lock;
682 mtx_assert(&ts->ts_lock, MA_OWNED);
684 MPASS(owner->td_proc->p_magic == P_MAGIC);
685 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
688 * If the lock does not already have a turnstile, use this thread's
689 * turnstile. Otherwise insert the current thread into the
690 * turnstile already in use by this lock.
692 tc = TC_LOOKUP(ts->ts_lockobj);
693 mtx_assert(&tc->tc_lock, MA_OWNED);
694 if (ts == td->td_turnstile) {
695 #ifdef TURNSTILE_PROFILING
697 if (tc->tc_depth > tc->tc_max_depth) {
698 tc->tc_max_depth = tc->tc_depth;
699 if (tc->tc_max_depth > turnstile_max_depth)
700 turnstile_max_depth = tc->tc_max_depth;
703 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
704 KASSERT(TAILQ_EMPTY(&ts->ts_pending),
705 ("thread's turnstile has pending threads"));
706 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]),
707 ("thread's turnstile has exclusive waiters"));
708 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]),
709 ("thread's turnstile has shared waiters"));
710 KASSERT(LIST_EMPTY(&ts->ts_free),
711 ("thread's turnstile has a non-empty free list"));
712 MPASS(ts->ts_lockobj != NULL);
713 mtx_lock_spin(&td_contested_lock);
714 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
715 turnstile_setowner(ts, owner);
716 mtx_unlock_spin(&td_contested_lock);
718 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq)
719 if (td1->td_priority > td->td_priority)
721 mtx_lock_spin(&td_contested_lock);
723 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
725 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
726 MPASS(owner == ts->ts_owner);
727 mtx_unlock_spin(&td_contested_lock);
728 MPASS(td->td_turnstile != NULL);
729 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
732 thread_lock_set(td, &ts->ts_lock);
733 td->td_turnstile = NULL;
735 /* Save who we are blocked on and switch. */
736 lock = ts->ts_lockobj;
737 td->td_tsqueue = queue;
739 td->td_lockname = lock->lo_name;
740 td->td_blktick = ticks;
742 mtx_unlock_spin(&tc->tc_lock);
743 propagate_priority(td);
745 if (LOCK_LOG_TEST(lock, 0))
746 CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__,
747 td->td_tid, lock, lock->lo_name);
749 SDT_PROBE0(sched, , , sleep);
751 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
752 mi_switch(SW_VOL | SWT_TURNSTILE, NULL);
754 if (LOCK_LOG_TEST(lock, 0))
755 CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s",
756 __func__, td->td_tid, lock, lock->lo_name);
761 * Pick the highest priority thread on this turnstile and put it on the
762 * pending list. This must be called with the turnstile chain locked.
765 turnstile_signal(struct turnstile *ts, int queue)
767 struct turnstile_chain *tc;
772 mtx_assert(&ts->ts_lock, MA_OWNED);
773 MPASS(curthread->td_proc->p_magic == P_MAGIC);
774 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
775 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
778 * Pick the highest priority thread blocked on this lock and
779 * move it to the pending list.
781 td = TAILQ_FIRST(&ts->ts_blocked[queue]);
782 MPASS(td->td_proc->p_magic == P_MAGIC);
783 mtx_lock_spin(&td_contested_lock);
784 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
785 mtx_unlock_spin(&td_contested_lock);
786 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
789 * If the turnstile is now empty, remove it from its chain and
790 * give it to the about-to-be-woken thread. Otherwise take a
791 * turnstile from the free list and give it to the thread.
793 empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
794 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]);
796 tc = TC_LOOKUP(ts->ts_lockobj);
797 mtx_assert(&tc->tc_lock, MA_OWNED);
798 MPASS(LIST_EMPTY(&ts->ts_free));
799 #ifdef TURNSTILE_PROFILING
803 ts = LIST_FIRST(&ts->ts_free);
805 LIST_REMOVE(ts, ts_hash);
806 td->td_turnstile = ts;
812 * Put all blocked threads on the pending list. This must be called with
813 * the turnstile chain locked.
816 turnstile_broadcast(struct turnstile *ts, int queue)
818 struct turnstile_chain *tc;
819 struct turnstile *ts1;
823 mtx_assert(&ts->ts_lock, MA_OWNED);
824 MPASS(curthread->td_proc->p_magic == P_MAGIC);
825 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
827 * We must have the chain locked so that we can remove the empty
828 * turnstile from the hash queue.
830 tc = TC_LOOKUP(ts->ts_lockobj);
831 mtx_assert(&tc->tc_lock, MA_OWNED);
832 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
835 * Transfer the blocked list to the pending list.
837 mtx_lock_spin(&td_contested_lock);
838 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq);
839 mtx_unlock_spin(&td_contested_lock);
842 * Give a turnstile to each thread. The last thread gets
843 * this turnstile if the turnstile is empty.
845 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
846 if (LIST_EMPTY(&ts->ts_free)) {
847 MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
849 #ifdef TURNSTILE_PROFILING
853 ts1 = LIST_FIRST(&ts->ts_free);
855 LIST_REMOVE(ts1, ts_hash);
856 td->td_turnstile = ts1;
861 * Wakeup all threads on the pending list and adjust the priority of the
862 * current thread appropriately. This must be called with the turnstile
866 turnstile_unpend(struct turnstile *ts, int owner_type)
868 TAILQ_HEAD( ,thread) pending_threads;
869 struct turnstile *nts;
874 mtx_assert(&ts->ts_lock, MA_OWNED);
875 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
876 MPASS(!TAILQ_EMPTY(&ts->ts_pending));
879 * Move the list of pending threads out of the turnstile and
880 * into a local variable.
882 TAILQ_INIT(&pending_threads);
883 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
885 if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
886 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]))
887 ts->ts_lockobj = NULL;
890 * Adjust the priority of curthread based on other contested
891 * locks it owns. Don't lower the priority below the base
897 mtx_lock_spin(&td_contested_lock);
899 * Remove the turnstile from this thread's list of contested locks
900 * since this thread doesn't own it anymore. New threads will
901 * not be blocking on the turnstile until it is claimed by a new
902 * owner. There might not be a current owner if this is a shared
905 if (ts->ts_owner != NULL) {
907 LIST_REMOVE(ts, ts_link);
909 LIST_FOREACH(nts, &td->td_contested, ts_link) {
910 cp = turnstile_first_waiter(nts)->td_priority;
914 mtx_unlock_spin(&td_contested_lock);
915 sched_unlend_prio(td, pri);
918 * Wake up all the pending threads. If a thread is not blocked
919 * on a lock, then it is currently executing on another CPU in
920 * turnstile_wait() or sitting on a run queue waiting to resume
921 * in turnstile_wait(). Set a flag to force it to try to acquire
922 * the lock again instead of blocking.
924 while (!TAILQ_EMPTY(&pending_threads)) {
925 td = TAILQ_FIRST(&pending_threads);
926 TAILQ_REMOVE(&pending_threads, td, td_lockq);
927 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
929 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
930 MPASS(td->td_proc->p_magic == P_MAGIC);
931 MPASS(TD_ON_LOCK(td));
933 MPASS(TD_CAN_RUN(td));
934 td->td_blocked = NULL;
935 td->td_lockname = NULL;
938 td->td_tsqueue = 0xff;
940 sched_add(td, SRQ_BORING);
943 mtx_unlock_spin(&ts->ts_lock);
947 * Give up ownership of a turnstile. This must be called with the
948 * turnstile chain locked.
951 turnstile_disown(struct turnstile *ts)
957 mtx_assert(&ts->ts_lock, MA_OWNED);
958 MPASS(ts->ts_owner == curthread);
959 MPASS(TAILQ_EMPTY(&ts->ts_pending));
960 MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) ||
961 !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
964 * Remove the turnstile from this thread's list of contested locks
965 * since this thread doesn't own it anymore. New threads will
966 * not be blocking on the turnstile until it is claimed by a new
969 mtx_lock_spin(&td_contested_lock);
971 LIST_REMOVE(ts, ts_link);
972 mtx_unlock_spin(&td_contested_lock);
975 * Adjust the priority of curthread based on other contested
976 * locks it owns. Don't lower the priority below the base
982 mtx_unlock_spin(&ts->ts_lock);
983 mtx_lock_spin(&td_contested_lock);
984 LIST_FOREACH(ts, &td->td_contested, ts_link) {
985 cp = turnstile_first_waiter(ts)->td_priority;
989 mtx_unlock_spin(&td_contested_lock);
990 sched_unlend_prio(td, pri);
995 * Return the first thread in a turnstile.
998 turnstile_head(struct turnstile *ts, int queue)
1003 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1004 mtx_assert(&ts->ts_lock, MA_OWNED);
1006 return (TAILQ_FIRST(&ts->ts_blocked[queue]));
1010 * Returns true if a sub-queue of a turnstile is empty.
1013 turnstile_empty(struct turnstile *ts, int queue)
1018 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1019 mtx_assert(&ts->ts_lock, MA_OWNED);
1021 return (TAILQ_EMPTY(&ts->ts_blocked[queue]));
1026 print_thread(struct thread *td, const char *prefix)
1029 db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid,
1030 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1035 print_queue(struct threadqueue *queue, const char *header, const char *prefix)
1039 db_printf("%s:\n", header);
1040 if (TAILQ_EMPTY(queue)) {
1041 db_printf("%sempty\n", prefix);
1044 TAILQ_FOREACH(td, queue, td_lockq) {
1045 print_thread(td, prefix);
1049 DB_SHOW_COMMAND(turnstile, db_show_turnstile)
1051 struct turnstile_chain *tc;
1052 struct turnstile *ts;
1053 struct lock_object *lock;
1060 * First, see if there is an active turnstile for the lock indicated
1063 lock = (struct lock_object *)addr;
1064 tc = TC_LOOKUP(lock);
1065 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1066 if (ts->ts_lockobj == lock)
1070 * Second, see if there is an active turnstile at the address
1073 for (i = 0; i < TC_TABLESIZE; i++)
1074 LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) {
1075 if (ts == (struct turnstile *)addr)
1079 db_printf("Unable to locate a turnstile via %p\n", (void *)addr);
1082 lock = ts->ts_lockobj;
1083 db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name,
1086 print_thread(ts->ts_owner, "Lock Owner: ");
1088 db_printf("Lock Owner: none\n");
1089 print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t");
1090 print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters",
1092 print_queue(&ts->ts_pending, "Pending Threads", "\t");
1097 * Show all the threads a particular thread is waiting on based on
1098 * non-sleepable and non-spin locks.
1101 print_lockchain(struct thread *td, const char *prefix)
1103 struct lock_object *lock;
1104 struct lock_class *class;
1105 struct turnstile *ts;
1108 * Follow the chain. We keep walking as long as the thread is
1109 * blocked on a turnstile that has an owner.
1111 while (!db_pager_quit) {
1112 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1113 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1115 switch (td->td_state) {
1117 db_printf("is inactive\n");
1120 db_printf("can run\n");
1123 db_printf("is on a run queue\n");
1126 db_printf("running on CPU %d\n", td->td_oncpu);
1129 if (TD_ON_LOCK(td)) {
1130 ts = td->td_blocked;
1131 lock = ts->ts_lockobj;
1132 class = LOCK_CLASS(lock);
1133 db_printf("blocked on lock %p (%s) \"%s\"\n",
1134 lock, class->lc_name, lock->lo_name);
1135 if (ts->ts_owner == NULL)
1140 db_printf("inhibited\n");
1143 db_printf("??? (%#x)\n", td->td_state);
1149 DB_SHOW_COMMAND(lockchain, db_show_lockchain)
1153 /* Figure out which thread to start with. */
1155 td = db_lookup_thread(addr, TRUE);
1159 print_lockchain(td, "");
1162 DB_SHOW_ALL_COMMAND(chains, db_show_allchains)
1169 FOREACH_PROC_IN_SYSTEM(p) {
1170 FOREACH_THREAD_IN_PROC(p, td) {
1171 if (TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested)) {
1172 db_printf("chain %d:\n", i++);
1173 print_lockchain(td, " ");
1180 DB_SHOW_ALIAS(allchains, db_show_allchains)
1183 * Show all the threads a particular thread is waiting on based on
1187 print_sleepchain(struct thread *td, const char *prefix)
1189 struct thread *owner;
1192 * Follow the chain. We keep walking as long as the thread is
1193 * blocked on a sleep lock that has an owner.
1195 while (!db_pager_quit) {
1196 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1197 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1199 switch (td->td_state) {
1201 db_printf("is inactive\n");
1204 db_printf("can run\n");
1207 db_printf("is on a run queue\n");
1210 db_printf("running on CPU %d\n", td->td_oncpu);
1213 if (TD_ON_SLEEPQ(td)) {
1214 if (lockmgr_chain(td, &owner) ||
1215 sx_chain(td, &owner)) {
1221 db_printf("sleeping on %p \"%s\"\n",
1222 td->td_wchan, td->td_wmesg);
1225 db_printf("inhibited\n");
1228 db_printf("??? (%#x)\n", td->td_state);
1234 DB_SHOW_COMMAND(sleepchain, db_show_sleepchain)
1238 /* Figure out which thread to start with. */
1240 td = db_lookup_thread(addr, TRUE);
1244 print_sleepchain(td, "");
1247 static void print_waiters(struct turnstile *ts, int indent);
1250 print_waiter(struct thread *td, int indent)
1252 struct turnstile *ts;
1257 for (i = 0; i < indent; i++)
1259 print_thread(td, "thread ");
1260 LIST_FOREACH(ts, &td->td_contested, ts_link)
1261 print_waiters(ts, indent + 1);
1265 print_waiters(struct turnstile *ts, int indent)
1267 struct lock_object *lock;
1268 struct lock_class *class;
1274 lock = ts->ts_lockobj;
1275 class = LOCK_CLASS(lock);
1276 for (i = 0; i < indent; i++)
1278 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name);
1279 TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq)
1280 print_waiter(td, indent + 1);
1281 TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq)
1282 print_waiter(td, indent + 1);
1283 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq)
1284 print_waiter(td, indent + 1);
1287 DB_SHOW_COMMAND(locktree, db_show_locktree)
1289 struct lock_object *lock;
1290 struct lock_class *class;
1291 struct turnstile_chain *tc;
1292 struct turnstile *ts;
1296 lock = (struct lock_object *)addr;
1297 tc = TC_LOOKUP(lock);
1298 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1299 if (ts->ts_lockobj == lock)
1302 class = LOCK_CLASS(lock);
1303 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name,
1306 print_waiters(ts, 0);