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);
174 SDT_PROBE_DEFINE2(sched, , , 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, just
219 * panic and tell the user which thread misbehaved so
220 * they can hopefully get a stack trace from the truly
221 * misbehaving thread.
223 if (TD_IS_SLEEPING(td)) {
225 "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
226 td->td_tid, td->td_proc->p_pid);
227 kdb_backtrace_thread(td);
228 panic("sleeping thread");
232 * If this thread already has higher priority than the
233 * thread that is being blocked, we are finished.
235 if (td->td_priority <= pri) {
241 * Bump this thread's priority.
243 sched_lend_prio(td, pri);
246 * If lock holder is actually running or on the run queue
249 if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) {
250 MPASS(td->td_blocked == NULL);
257 * For UP, we check to see if td is curthread (this shouldn't
258 * ever happen however as it would mean we are in a deadlock.)
260 KASSERT(td != curthread, ("Deadlock detected"));
264 * If we aren't blocked on a lock, we should be.
266 KASSERT(TD_ON_LOCK(td), (
267 "thread %d(%s):%d holds %s but isn't blocked on a lock\n",
268 td->td_tid, td->td_name, td->td_state,
269 ts->ts_lockobj->lo_name));
272 * Pick up the lock that td is blocked on.
276 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
277 /* Resort td on the list if needed. */
278 if (!turnstile_adjust_thread(ts, td)) {
279 mtx_unlock_spin(&ts->ts_lock);
282 /* The thread lock is released as ts lock above. */
287 * Adjust the thread's position on a turnstile after its priority has been
291 turnstile_adjust_thread(struct turnstile *ts, struct thread *td)
293 struct thread *td1, *td2;
296 THREAD_LOCK_ASSERT(td, MA_OWNED);
297 MPASS(TD_ON_LOCK(td));
300 * This thread may not be blocked on this turnstile anymore
301 * but instead might already be woken up on another CPU
302 * that is waiting on the thread lock in turnstile_unpend() to
303 * finish waking this thread up. We can detect this case
304 * by checking to see if this thread has been given a
305 * turnstile by either turnstile_signal() or
306 * turnstile_broadcast(). In this case, treat the thread as
307 * if it was already running.
309 if (td->td_turnstile != NULL)
313 * Check if the thread needs to be moved on the blocked chain.
314 * It needs to be moved if either its priority is lower than
315 * the previous thread or higher than the next thread.
317 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
318 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
319 td2 = TAILQ_NEXT(td, td_lockq);
320 if ((td1 != NULL && td->td_priority < td1->td_priority) ||
321 (td2 != NULL && td->td_priority > td2->td_priority)) {
324 * Remove thread from blocked chain and determine where
325 * it should be moved to.
327 queue = td->td_tsqueue;
328 MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE);
329 mtx_lock_spin(&td_contested_lock);
330 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
331 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) {
332 MPASS(td1->td_proc->p_magic == P_MAGIC);
333 if (td1->td_priority > td->td_priority)
338 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
340 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
341 mtx_unlock_spin(&td_contested_lock);
344 "turnstile_adjust_thread: td %d put at tail on [%p] %s",
345 td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name);
348 "turnstile_adjust_thread: td %d moved before %d on [%p] %s",
349 td->td_tid, td1->td_tid, ts->ts_lockobj,
350 ts->ts_lockobj->lo_name);
356 * Early initialization of turnstiles. This is not done via a SYSINIT()
357 * since this needs to be initialized very early when mutexes are first
361 init_turnstiles(void)
365 for (i = 0; i < TC_TABLESIZE; i++) {
366 LIST_INIT(&turnstile_chains[i].tc_turnstiles);
367 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
370 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
371 LIST_INIT(&thread0.td_contested);
372 thread0.td_turnstile = NULL;
375 #ifdef TURNSTILE_PROFILING
377 init_turnstile_profiling(void *arg)
379 struct sysctl_oid *chain_oid;
383 for (i = 0; i < TC_TABLESIZE; i++) {
384 snprintf(chain_name, sizeof(chain_name), "%d", i);
385 chain_oid = SYSCTL_ADD_NODE(NULL,
386 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO,
387 chain_name, CTLFLAG_RD, NULL, "turnstile chain stats");
388 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
389 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0,
391 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
392 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth,
396 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
397 init_turnstile_profiling, NULL);
401 init_turnstile0(void *dummy)
404 turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile),
411 turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE);
412 thread0.td_turnstile = turnstile_alloc();
414 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
417 * Update a thread on the turnstile list after it's priority has been changed.
418 * The old priority is passed in as an argument.
421 turnstile_adjust(struct thread *td, u_char oldpri)
423 struct turnstile *ts;
425 MPASS(TD_ON_LOCK(td));
428 * Pick up the lock that td is blocked on.
432 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
433 mtx_assert(&ts->ts_lock, MA_OWNED);
435 /* Resort the turnstile on the list. */
436 if (!turnstile_adjust_thread(ts, td))
439 * If our priority was lowered and we are at the head of the
440 * turnstile, then propagate our new priority up the chain.
441 * Note that we currently don't try to revoke lent priorities
442 * when our priority goes up.
444 MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE ||
445 td->td_tsqueue == TS_SHARED_QUEUE);
446 if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) &&
447 td->td_priority < oldpri) {
448 propagate_priority(td);
453 * Set the owner of the lock this turnstile is attached to.
456 turnstile_setowner(struct turnstile *ts, struct thread *owner)
459 mtx_assert(&td_contested_lock, MA_OWNED);
460 MPASS(ts->ts_owner == NULL);
462 /* A shared lock might not have an owner. */
466 MPASS(owner->td_proc->p_magic == P_MAGIC);
467 ts->ts_owner = owner;
468 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
473 * UMA zone item deallocator.
476 turnstile_dtor(void *mem, int size, void *arg)
478 struct turnstile *ts;
481 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]));
482 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
483 MPASS(TAILQ_EMPTY(&ts->ts_pending));
488 * UMA zone item initializer.
491 turnstile_init(void *mem, int size, int flags)
493 struct turnstile *ts;
497 TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
498 TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]);
499 TAILQ_INIT(&ts->ts_pending);
500 LIST_INIT(&ts->ts_free);
501 mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN | MTX_RECURSE);
506 turnstile_fini(void *mem, int size)
508 struct turnstile *ts;
511 mtx_destroy(&ts->ts_lock);
515 * Get a turnstile for a new thread.
518 turnstile_alloc(void)
521 return (uma_zalloc(turnstile_zone, M_WAITOK));
525 * Free a turnstile when a thread is destroyed.
528 turnstile_free(struct turnstile *ts)
531 uma_zfree(turnstile_zone, ts);
535 * Lock the turnstile chain associated with the specified lock.
538 turnstile_chain_lock(struct lock_object *lock)
540 struct turnstile_chain *tc;
542 tc = TC_LOOKUP(lock);
543 mtx_lock_spin(&tc->tc_lock);
547 turnstile_trywait(struct lock_object *lock)
549 struct turnstile_chain *tc;
550 struct turnstile *ts;
552 tc = TC_LOOKUP(lock);
553 mtx_lock_spin(&tc->tc_lock);
554 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
555 if (ts->ts_lockobj == lock) {
556 mtx_lock_spin(&ts->ts_lock);
560 ts = curthread->td_turnstile;
562 mtx_lock_spin(&ts->ts_lock);
563 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
564 ts->ts_lockobj = lock;
570 turnstile_cancel(struct turnstile *ts)
572 struct turnstile_chain *tc;
573 struct lock_object *lock;
575 mtx_assert(&ts->ts_lock, MA_OWNED);
577 mtx_unlock_spin(&ts->ts_lock);
578 lock = ts->ts_lockobj;
579 if (ts == curthread->td_turnstile)
580 ts->ts_lockobj = NULL;
581 tc = TC_LOOKUP(lock);
582 mtx_unlock_spin(&tc->tc_lock);
586 * Look up the turnstile for a lock in the hash table locking the associated
587 * turnstile chain along the way. If no turnstile is found in the hash
588 * table, NULL is returned.
591 turnstile_lookup(struct lock_object *lock)
593 struct turnstile_chain *tc;
594 struct turnstile *ts;
596 tc = TC_LOOKUP(lock);
597 mtx_assert(&tc->tc_lock, MA_OWNED);
598 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
599 if (ts->ts_lockobj == lock) {
600 mtx_lock_spin(&ts->ts_lock);
607 * Unlock the turnstile chain associated with a given lock.
610 turnstile_chain_unlock(struct lock_object *lock)
612 struct turnstile_chain *tc;
614 tc = TC_LOOKUP(lock);
615 mtx_unlock_spin(&tc->tc_lock);
619 * Return a pointer to the thread waiting on this turnstile with the
620 * most important priority or NULL if the turnstile has no waiters.
622 static struct thread *
623 turnstile_first_waiter(struct turnstile *ts)
625 struct thread *std, *xtd;
627 std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]);
628 xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
629 if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority))
635 * Take ownership of a turnstile and adjust the priority of the new
636 * owner appropriately.
639 turnstile_claim(struct turnstile *ts)
641 struct thread *td, *owner;
642 struct turnstile_chain *tc;
644 mtx_assert(&ts->ts_lock, MA_OWNED);
645 MPASS(ts != curthread->td_turnstile);
648 mtx_lock_spin(&td_contested_lock);
649 turnstile_setowner(ts, owner);
650 mtx_unlock_spin(&td_contested_lock);
652 td = turnstile_first_waiter(ts);
654 MPASS(td->td_proc->p_magic == P_MAGIC);
655 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
658 * Update the priority of the new owner if needed.
661 if (td->td_priority < owner->td_priority)
662 sched_lend_prio(owner, td->td_priority);
663 thread_unlock(owner);
664 tc = TC_LOOKUP(ts->ts_lockobj);
665 mtx_unlock_spin(&ts->ts_lock);
666 mtx_unlock_spin(&tc->tc_lock);
670 * Block the current thread on the turnstile assicated with 'lock'. This
671 * function will context switch and not return until this thread has been
672 * woken back up. This function must be called with the appropriate
673 * turnstile chain locked and will return with it unlocked.
676 turnstile_wait(struct turnstile *ts, struct thread *owner, int queue)
678 struct turnstile_chain *tc;
679 struct thread *td, *td1;
680 struct lock_object *lock;
683 mtx_assert(&ts->ts_lock, MA_OWNED);
685 MPASS(owner->td_proc->p_magic == P_MAGIC);
686 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
689 * If the lock does not already have a turnstile, use this thread's
690 * turnstile. Otherwise insert the current thread into the
691 * turnstile already in use by this lock.
693 tc = TC_LOOKUP(ts->ts_lockobj);
694 mtx_assert(&tc->tc_lock, MA_OWNED);
695 if (ts == td->td_turnstile) {
696 #ifdef TURNSTILE_PROFILING
698 if (tc->tc_depth > tc->tc_max_depth) {
699 tc->tc_max_depth = tc->tc_depth;
700 if (tc->tc_max_depth > turnstile_max_depth)
701 turnstile_max_depth = tc->tc_max_depth;
704 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
705 KASSERT(TAILQ_EMPTY(&ts->ts_pending),
706 ("thread's turnstile has pending threads"));
707 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]),
708 ("thread's turnstile has exclusive waiters"));
709 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]),
710 ("thread's turnstile has shared waiters"));
711 KASSERT(LIST_EMPTY(&ts->ts_free),
712 ("thread's turnstile has a non-empty free list"));
713 MPASS(ts->ts_lockobj != NULL);
714 mtx_lock_spin(&td_contested_lock);
715 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
716 turnstile_setowner(ts, owner);
717 mtx_unlock_spin(&td_contested_lock);
719 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq)
720 if (td1->td_priority > td->td_priority)
722 mtx_lock_spin(&td_contested_lock);
724 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
726 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
727 MPASS(owner == ts->ts_owner);
728 mtx_unlock_spin(&td_contested_lock);
729 MPASS(td->td_turnstile != NULL);
730 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
733 thread_lock_set(td, &ts->ts_lock);
734 td->td_turnstile = NULL;
736 /* Save who we are blocked on and switch. */
737 lock = ts->ts_lockobj;
738 td->td_tsqueue = queue;
740 td->td_lockname = lock->lo_name;
741 td->td_blktick = ticks;
743 mtx_unlock_spin(&tc->tc_lock);
744 propagate_priority(td);
746 if (LOCK_LOG_TEST(lock, 0))
747 CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__,
748 td->td_tid, lock, lock->lo_name);
750 SDT_PROBE0(sched, , , sleep);
752 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
753 mi_switch(SW_VOL | SWT_TURNSTILE, NULL);
755 if (LOCK_LOG_TEST(lock, 0))
756 CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s",
757 __func__, td->td_tid, lock, lock->lo_name);
762 * Pick the highest priority thread on this turnstile and put it on the
763 * pending list. This must be called with the turnstile chain locked.
766 turnstile_signal(struct turnstile *ts, int queue)
768 struct turnstile_chain *tc;
773 mtx_assert(&ts->ts_lock, MA_OWNED);
774 MPASS(curthread->td_proc->p_magic == P_MAGIC);
775 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
776 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
779 * Pick the highest priority thread blocked on this lock and
780 * move it to the pending list.
782 td = TAILQ_FIRST(&ts->ts_blocked[queue]);
783 MPASS(td->td_proc->p_magic == P_MAGIC);
784 mtx_lock_spin(&td_contested_lock);
785 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
786 mtx_unlock_spin(&td_contested_lock);
787 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
790 * If the turnstile is now empty, remove it from its chain and
791 * give it to the about-to-be-woken thread. Otherwise take a
792 * turnstile from the free list and give it to the thread.
794 empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
795 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]);
797 tc = TC_LOOKUP(ts->ts_lockobj);
798 mtx_assert(&tc->tc_lock, MA_OWNED);
799 MPASS(LIST_EMPTY(&ts->ts_free));
800 #ifdef TURNSTILE_PROFILING
804 ts = LIST_FIRST(&ts->ts_free);
806 LIST_REMOVE(ts, ts_hash);
807 td->td_turnstile = ts;
813 * Put all blocked threads on the pending list. This must be called with
814 * the turnstile chain locked.
817 turnstile_broadcast(struct turnstile *ts, int queue)
819 struct turnstile_chain *tc;
820 struct turnstile *ts1;
824 mtx_assert(&ts->ts_lock, MA_OWNED);
825 MPASS(curthread->td_proc->p_magic == P_MAGIC);
826 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
828 * We must have the chain locked so that we can remove the empty
829 * turnstile from the hash queue.
831 tc = TC_LOOKUP(ts->ts_lockobj);
832 mtx_assert(&tc->tc_lock, MA_OWNED);
833 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
836 * Transfer the blocked list to the pending list.
838 mtx_lock_spin(&td_contested_lock);
839 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq);
840 mtx_unlock_spin(&td_contested_lock);
843 * Give a turnstile to each thread. The last thread gets
844 * this turnstile if the turnstile is empty.
846 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
847 if (LIST_EMPTY(&ts->ts_free)) {
848 MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
850 #ifdef TURNSTILE_PROFILING
854 ts1 = LIST_FIRST(&ts->ts_free);
856 LIST_REMOVE(ts1, ts_hash);
857 td->td_turnstile = ts1;
862 * Wakeup all threads on the pending list and adjust the priority of the
863 * current thread appropriately. This must be called with the turnstile
867 turnstile_unpend(struct turnstile *ts, int owner_type)
869 TAILQ_HEAD( ,thread) pending_threads;
870 struct turnstile *nts;
875 mtx_assert(&ts->ts_lock, MA_OWNED);
876 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
877 MPASS(!TAILQ_EMPTY(&ts->ts_pending));
880 * Move the list of pending threads out of the turnstile and
881 * into a local variable.
883 TAILQ_INIT(&pending_threads);
884 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
886 if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
887 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]))
888 ts->ts_lockobj = NULL;
891 * Adjust the priority of curthread based on other contested
892 * locks it owns. Don't lower the priority below the base
898 mtx_lock_spin(&td_contested_lock);
900 * Remove the turnstile from this thread's list of contested locks
901 * since this thread doesn't own it anymore. New threads will
902 * not be blocking on the turnstile until it is claimed by a new
903 * owner. There might not be a current owner if this is a shared
906 if (ts->ts_owner != NULL) {
908 LIST_REMOVE(ts, ts_link);
910 LIST_FOREACH(nts, &td->td_contested, ts_link) {
911 cp = turnstile_first_waiter(nts)->td_priority;
915 mtx_unlock_spin(&td_contested_lock);
916 sched_unlend_prio(td, pri);
919 * Wake up all the pending threads. If a thread is not blocked
920 * on a lock, then it is currently executing on another CPU in
921 * turnstile_wait() or sitting on a run queue waiting to resume
922 * in turnstile_wait(). Set a flag to force it to try to acquire
923 * the lock again instead of blocking.
925 while (!TAILQ_EMPTY(&pending_threads)) {
926 td = TAILQ_FIRST(&pending_threads);
927 TAILQ_REMOVE(&pending_threads, td, td_lockq);
928 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
930 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
931 MPASS(td->td_proc->p_magic == P_MAGIC);
932 MPASS(TD_ON_LOCK(td));
934 MPASS(TD_CAN_RUN(td));
935 td->td_blocked = NULL;
936 td->td_lockname = NULL;
939 td->td_tsqueue = 0xff;
941 sched_add(td, SRQ_BORING);
944 mtx_unlock_spin(&ts->ts_lock);
948 * Give up ownership of a turnstile. This must be called with the
949 * turnstile chain locked.
952 turnstile_disown(struct turnstile *ts)
958 mtx_assert(&ts->ts_lock, MA_OWNED);
959 MPASS(ts->ts_owner == curthread);
960 MPASS(TAILQ_EMPTY(&ts->ts_pending));
961 MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) ||
962 !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
965 * Remove the turnstile from this thread's list of contested locks
966 * since this thread doesn't own it anymore. New threads will
967 * not be blocking on the turnstile until it is claimed by a new
970 mtx_lock_spin(&td_contested_lock);
972 LIST_REMOVE(ts, ts_link);
973 mtx_unlock_spin(&td_contested_lock);
976 * Adjust the priority of curthread based on other contested
977 * locks it owns. Don't lower the priority below the base
983 mtx_unlock_spin(&ts->ts_lock);
984 mtx_lock_spin(&td_contested_lock);
985 LIST_FOREACH(ts, &td->td_contested, ts_link) {
986 cp = turnstile_first_waiter(ts)->td_priority;
990 mtx_unlock_spin(&td_contested_lock);
991 sched_unlend_prio(td, pri);
996 * Return the first thread in a turnstile.
999 turnstile_head(struct turnstile *ts, int queue)
1004 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1005 mtx_assert(&ts->ts_lock, MA_OWNED);
1007 return (TAILQ_FIRST(&ts->ts_blocked[queue]));
1011 * Returns true if a sub-queue of a turnstile is empty.
1014 turnstile_empty(struct turnstile *ts, int queue)
1019 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1020 mtx_assert(&ts->ts_lock, MA_OWNED);
1022 return (TAILQ_EMPTY(&ts->ts_blocked[queue]));
1027 print_thread(struct thread *td, const char *prefix)
1030 db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid,
1031 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1036 print_queue(struct threadqueue *queue, const char *header, const char *prefix)
1040 db_printf("%s:\n", header);
1041 if (TAILQ_EMPTY(queue)) {
1042 db_printf("%sempty\n", prefix);
1045 TAILQ_FOREACH(td, queue, td_lockq) {
1046 print_thread(td, prefix);
1050 DB_SHOW_COMMAND(turnstile, db_show_turnstile)
1052 struct turnstile_chain *tc;
1053 struct turnstile *ts;
1054 struct lock_object *lock;
1061 * First, see if there is an active turnstile for the lock indicated
1064 lock = (struct lock_object *)addr;
1065 tc = TC_LOOKUP(lock);
1066 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1067 if (ts->ts_lockobj == lock)
1071 * Second, see if there is an active turnstile at the address
1074 for (i = 0; i < TC_TABLESIZE; i++)
1075 LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) {
1076 if (ts == (struct turnstile *)addr)
1080 db_printf("Unable to locate a turnstile via %p\n", (void *)addr);
1083 lock = ts->ts_lockobj;
1084 db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name,
1087 print_thread(ts->ts_owner, "Lock Owner: ");
1089 db_printf("Lock Owner: none\n");
1090 print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t");
1091 print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters",
1093 print_queue(&ts->ts_pending, "Pending Threads", "\t");
1098 * Show all the threads a particular thread is waiting on based on
1099 * non-sleepable and non-spin locks.
1102 print_lockchain(struct thread *td, const char *prefix)
1104 struct lock_object *lock;
1105 struct lock_class *class;
1106 struct turnstile *ts;
1109 * Follow the chain. We keep walking as long as the thread is
1110 * blocked on a turnstile that has an owner.
1112 while (!db_pager_quit) {
1113 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1114 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1116 switch (td->td_state) {
1118 db_printf("is inactive\n");
1121 db_printf("can run\n");
1124 db_printf("is on a run queue\n");
1127 db_printf("running on CPU %d\n", td->td_oncpu);
1130 if (TD_ON_LOCK(td)) {
1131 ts = td->td_blocked;
1132 lock = ts->ts_lockobj;
1133 class = LOCK_CLASS(lock);
1134 db_printf("blocked on lock %p (%s) \"%s\"\n",
1135 lock, class->lc_name, lock->lo_name);
1136 if (ts->ts_owner == NULL)
1141 db_printf("inhibited\n");
1144 db_printf("??? (%#x)\n", td->td_state);
1150 DB_SHOW_COMMAND(lockchain, db_show_lockchain)
1154 /* Figure out which thread to start with. */
1156 td = db_lookup_thread(addr, TRUE);
1160 print_lockchain(td, "");
1163 DB_SHOW_ALL_COMMAND(chains, db_show_allchains)
1170 FOREACH_PROC_IN_SYSTEM(p) {
1171 FOREACH_THREAD_IN_PROC(p, td) {
1172 if (TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested)) {
1173 db_printf("chain %d:\n", i++);
1174 print_lockchain(td, " ");
1181 DB_SHOW_ALIAS(allchains, db_show_allchains)
1184 * Show all the threads a particular thread is waiting on based on
1188 print_sleepchain(struct thread *td, const char *prefix)
1190 struct thread *owner;
1193 * Follow the chain. We keep walking as long as the thread is
1194 * blocked on a sleep lock that has an owner.
1196 while (!db_pager_quit) {
1197 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1198 td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name :
1200 switch (td->td_state) {
1202 db_printf("is inactive\n");
1205 db_printf("can run\n");
1208 db_printf("is on a run queue\n");
1211 db_printf("running on CPU %d\n", td->td_oncpu);
1214 if (TD_ON_SLEEPQ(td)) {
1215 if (lockmgr_chain(td, &owner) ||
1216 sx_chain(td, &owner)) {
1222 db_printf("sleeping on %p \"%s\"\n",
1223 td->td_wchan, td->td_wmesg);
1226 db_printf("inhibited\n");
1229 db_printf("??? (%#x)\n", td->td_state);
1235 DB_SHOW_COMMAND(sleepchain, db_show_sleepchain)
1239 /* Figure out which thread to start with. */
1241 td = db_lookup_thread(addr, TRUE);
1245 print_sleepchain(td, "");
1248 static void print_waiters(struct turnstile *ts, int indent);
1251 print_waiter(struct thread *td, int indent)
1253 struct turnstile *ts;
1258 for (i = 0; i < indent; i++)
1260 print_thread(td, "thread ");
1261 LIST_FOREACH(ts, &td->td_contested, ts_link)
1262 print_waiters(ts, indent + 1);
1266 print_waiters(struct turnstile *ts, int indent)
1268 struct lock_object *lock;
1269 struct lock_class *class;
1275 lock = ts->ts_lockobj;
1276 class = LOCK_CLASS(lock);
1277 for (i = 0; i < indent; i++)
1279 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name);
1280 TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq)
1281 print_waiter(td, indent + 1);
1282 TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq)
1283 print_waiter(td, indent + 1);
1284 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq)
1285 print_waiter(td, indent + 1);
1288 DB_SHOW_COMMAND(locktree, db_show_locktree)
1290 struct lock_object *lock;
1291 struct lock_class *class;
1292 struct turnstile_chain *tc;
1293 struct turnstile *ts;
1297 lock = (struct lock_object *)addr;
1298 tc = TC_LOOKUP(lock);
1299 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1300 if (ts->ts_lockobj == lock)
1303 class = LOCK_CLASS(lock);
1304 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name,
1307 print_waiters(ts, 0);