2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Berkeley Software Design Inc's name may not be used to endorse or
15 * promote products derived from this software without specific prior
18 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
31 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35 * Implementation of turnstiles used to hold queue of threads blocked on
36 * non-sleepable locks. Sleepable locks use condition variables to
37 * implement their queues. Turnstiles differ from a sleep queue in that
38 * turnstile queue's are assigned to a lock held by an owning thread. Thus,
39 * when one thread is enqueued onto a turnstile, it can lend its priority
40 * to the owning thread.
42 * We wish to avoid bloating locks with an embedded turnstile and we do not
43 * want to use back-pointers in the locks for the same reason. Thus, we
44 * use a similar approach to that of Solaris 7 as described in Solaris
45 * Internals by Jim Mauro and Richard McDougall. Turnstiles are looked up
46 * in a hash table based on the address of the lock. Each entry in the
47 * hash table is a linked-lists of turnstiles and is called a turnstile
48 * chain. Each chain contains a spin mutex that protects all of the
49 * turnstiles in the chain.
51 * Each time a thread is created, a turnstile is allocated from a UMA zone
52 * and attached to that thread. When a thread blocks on a lock, if it is the
53 * first thread to block, it lends its turnstile to the lock. If the lock
54 * already has a turnstile, then it gives its turnstile to the lock's
55 * turnstile's free list. When a thread is woken up, it takes a turnstile from
56 * the free list if there are any other waiters. If it is the only thread
57 * blocked on the lock, then it reclaims the turnstile associated with the lock
58 * and removes it from the hash table.
61 #include <sys/cdefs.h>
62 __FBSDID("$FreeBSD$");
65 #include "opt_turnstile_profiling.h"
66 #include "opt_sched.h"
68 #include <sys/param.h>
69 #include <sys/systm.h>
71 #include <sys/kernel.h>
74 #include <sys/mutex.h>
76 #include <sys/queue.h>
77 #include <sys/sched.h>
79 #include <sys/sysctl.h>
80 #include <sys/turnstile.h>
86 #include <sys/lockmgr.h>
91 * Constants for the hash table of turnstile chains. TC_SHIFT is a magic
92 * number chosen because the sleep queue's use the same value for the
93 * shift. Basically, we ignore the lower 8 bits of the address.
94 * TC_TABLESIZE must be a power of two for TC_MASK to work properly.
96 #define TC_TABLESIZE 128 /* Must be power of 2. */
97 #define TC_MASK (TC_TABLESIZE - 1)
99 #define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK)
100 #define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)]
103 * There are three different lists of turnstiles as follows. The list
104 * connected by ts_link entries is a per-thread list of all the turnstiles
105 * attached to locks that we own. This is used to fixup our priority when
106 * a lock is released. The other two lists use the ts_hash entries. The
107 * first of these two is the turnstile chain list that a turnstile is on
108 * when it is attached to a lock. The second list to use ts_hash is the
109 * free list hung off of a turnstile that is attached to a lock.
111 * Each turnstile contains three lists of threads. The two ts_blocked lists
112 * are linked list of threads blocked on the turnstile's lock. One list is
113 * for exclusive waiters, and the other is for shared waiters. The
114 * ts_pending list is a linked list of threads previously awakened by
115 * turnstile_signal() or turnstile_wait() that are waiting to be put on
119 * c - turnstile chain lock
120 * q - td_contested lock
123 struct mtx ts_lock; /* Spin lock for self. */
124 struct threadqueue ts_blocked[2]; /* (c + q) Blocked threads. */
125 struct threadqueue ts_pending; /* (c) Pending threads. */
126 LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */
127 LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */
128 LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */
129 struct lock_object *ts_lockobj; /* (c) Lock we reference. */
130 struct thread *ts_owner; /* (c + q) Who owns the lock. */
133 struct turnstile_chain {
134 LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */
135 struct mtx tc_lock; /* Spin lock for this chain. */
136 #ifdef TURNSTILE_PROFILING
137 u_int tc_depth; /* Length of tc_queues. */
138 u_int tc_max_depth; /* Max length of tc_queues. */
142 #ifdef TURNSTILE_PROFILING
143 u_int turnstile_max_depth;
144 static SYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
145 "turnstile profiling");
146 static SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains,
147 CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
148 "turnstile chain stats");
149 SYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD,
150 &turnstile_max_depth, 0, "maximum depth achieved of a single chain");
152 static struct mtx td_contested_lock;
153 static struct turnstile_chain turnstile_chains[TC_TABLESIZE];
154 static uma_zone_t turnstile_zone;
157 * Prototypes for non-exported routines.
159 static void init_turnstile0(void *dummy);
160 #ifdef TURNSTILE_PROFILING
161 static void init_turnstile_profiling(void *arg);
163 static void propagate_priority(struct thread *td);
164 static int turnstile_adjust_thread(struct turnstile *ts,
166 static struct thread *turnstile_first_waiter(struct turnstile *ts);
167 static void turnstile_setowner(struct turnstile *ts, struct thread *owner);
169 static void turnstile_dtor(void *mem, int size, void *arg);
171 static int turnstile_init(void *mem, int size, int flags);
172 static void turnstile_fini(void *mem, int size);
174 SDT_PROVIDER_DECLARE(sched);
175 SDT_PROBE_DEFINE(sched, , , sleep);
176 SDT_PROBE_DEFINE2(sched, , , wakeup, "struct thread *",
180 propagate_unlock_ts(struct turnstile *top, struct turnstile *ts)
184 mtx_unlock_spin(&ts->ts_lock);
188 propagate_unlock_td(struct turnstile *top, struct thread *td)
191 if (td->td_lock != &top->ts_lock)
196 * Walks the chain of turnstiles and their owners to propagate the priority
197 * of the thread being blocked to all the threads holding locks that have to
198 * release their locks before this thread can run again.
201 propagate_priority(struct thread *td)
203 struct turnstile *ts, *top;
206 THREAD_LOCK_ASSERT(td, MA_OWNED);
207 pri = td->td_priority;
208 top = ts = td->td_blocked;
209 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
212 * The original turnstile lock is held across the entire
213 * operation. We only ever lock down the chain so the lock
221 * This might be a read lock with no owner. There's
222 * not much we can do, so just bail.
224 propagate_unlock_ts(top, ts);
229 * Wait for the thread lock to be stable and then only
230 * acquire if it is not the turnstile lock.
232 thread_lock_block_wait(td);
233 if (td->td_lock != &ts->ts_lock) {
234 thread_lock_flags(td, MTX_DUPOK);
235 propagate_unlock_ts(top, ts);
237 MPASS(td->td_proc != NULL);
238 MPASS(td->td_proc->p_magic == P_MAGIC);
241 * If the thread is asleep, then we are probably about
242 * to deadlock. To make debugging this easier, show
243 * backtrace of misbehaving thread and panic to not
244 * leave the kernel deadlocked.
246 if (TD_IS_SLEEPING(td)) {
248 "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
249 td->td_tid, td->td_proc->p_pid);
250 kdb_backtrace_thread(td);
251 panic("sleeping thread");
255 * If this thread already has higher priority than the
256 * thread that is being blocked, we are finished.
258 if (td->td_priority <= pri) {
259 propagate_unlock_td(top, td);
264 * Bump this thread's priority.
266 sched_lend_prio(td, pri);
269 * If lock holder is actually running or on the run queue
272 if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) {
273 MPASS(td->td_blocked == NULL);
274 propagate_unlock_td(top, td);
280 * For UP, we check to see if td is curthread (this shouldn't
281 * ever happen however as it would mean we are in a deadlock.)
283 KASSERT(td != curthread, ("Deadlock detected"));
287 * If we aren't blocked on a lock, we should be.
289 KASSERT(TD_ON_LOCK(td), (
290 "thread %d(%s):%d holds %s but isn't blocked on a lock\n",
291 td->td_tid, td->td_name, td->td_state,
292 ts->ts_lockobj->lo_name));
295 * Pick up the lock that td is blocked on.
299 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
300 /* Resort td on the list if needed. */
301 if (!turnstile_adjust_thread(ts, td)) {
302 propagate_unlock_ts(top, ts);
305 /* The thread lock is released as ts lock above. */
310 * Adjust the thread's position on a turnstile after its priority has been
314 turnstile_adjust_thread(struct turnstile *ts, struct thread *td)
316 struct thread *td1, *td2;
319 THREAD_LOCK_ASSERT(td, MA_OWNED);
320 MPASS(TD_ON_LOCK(td));
323 * This thread may not be blocked on this turnstile anymore
324 * but instead might already be woken up on another CPU
325 * that is waiting on the thread lock in turnstile_unpend() to
326 * finish waking this thread up. We can detect this case
327 * by checking to see if this thread has been given a
328 * turnstile by either turnstile_signal() or
329 * turnstile_broadcast(). In this case, treat the thread as
330 * if it was already running.
332 if (td->td_turnstile != NULL)
336 * Check if the thread needs to be moved on the blocked chain.
337 * It needs to be moved if either its priority is lower than
338 * the previous thread or higher than the next thread.
340 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
341 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
342 td2 = TAILQ_NEXT(td, td_lockq);
343 if ((td1 != NULL && td->td_priority < td1->td_priority) ||
344 (td2 != NULL && td->td_priority > td2->td_priority)) {
347 * Remove thread from blocked chain and determine where
348 * it should be moved to.
350 queue = td->td_tsqueue;
351 MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE);
352 mtx_lock_spin(&td_contested_lock);
353 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
354 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) {
355 MPASS(td1->td_proc->p_magic == P_MAGIC);
356 if (td1->td_priority > td->td_priority)
361 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
363 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
364 mtx_unlock_spin(&td_contested_lock);
367 "turnstile_adjust_thread: td %d put at tail on [%p] %s",
368 td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name);
371 "turnstile_adjust_thread: td %d moved before %d on [%p] %s",
372 td->td_tid, td1->td_tid, ts->ts_lockobj,
373 ts->ts_lockobj->lo_name);
379 * Early initialization of turnstiles. This is not done via a SYSINIT()
380 * since this needs to be initialized very early when mutexes are first
384 init_turnstiles(void)
388 for (i = 0; i < TC_TABLESIZE; i++) {
389 LIST_INIT(&turnstile_chains[i].tc_turnstiles);
390 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
393 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
394 LIST_INIT(&thread0.td_contested);
395 thread0.td_turnstile = NULL;
398 #ifdef TURNSTILE_PROFILING
400 init_turnstile_profiling(void *arg)
402 struct sysctl_oid *chain_oid;
406 for (i = 0; i < TC_TABLESIZE; i++) {
407 snprintf(chain_name, sizeof(chain_name), "%d", i);
408 chain_oid = SYSCTL_ADD_NODE(NULL,
409 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO,
410 chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
411 "turnstile chain stats");
412 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
413 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0,
415 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
416 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth,
420 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
421 init_turnstile_profiling, NULL);
425 init_turnstile0(void *dummy)
428 turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile),
435 turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE);
436 thread0.td_turnstile = turnstile_alloc();
438 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
441 * Update a thread on the turnstile list after it's priority has been changed.
442 * The old priority is passed in as an argument.
445 turnstile_adjust(struct thread *td, u_char oldpri)
447 struct turnstile *ts;
449 MPASS(TD_ON_LOCK(td));
452 * Pick up the lock that td is blocked on.
456 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
457 mtx_assert(&ts->ts_lock, MA_OWNED);
459 /* Resort the turnstile on the list. */
460 if (!turnstile_adjust_thread(ts, td))
463 * If our priority was lowered and we are at the head of the
464 * turnstile, then propagate our new priority up the chain.
465 * Note that we currently don't try to revoke lent priorities
466 * when our priority goes up.
468 MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE ||
469 td->td_tsqueue == TS_SHARED_QUEUE);
470 if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) &&
471 td->td_priority < oldpri) {
472 propagate_priority(td);
477 * Set the owner of the lock this turnstile is attached to.
480 turnstile_setowner(struct turnstile *ts, struct thread *owner)
483 mtx_assert(&td_contested_lock, MA_OWNED);
484 MPASS(ts->ts_owner == NULL);
486 /* A shared lock might not have an owner. */
490 MPASS(owner->td_proc->p_magic == P_MAGIC);
491 ts->ts_owner = owner;
492 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
497 * UMA zone item deallocator.
500 turnstile_dtor(void *mem, int size, void *arg)
502 struct turnstile *ts;
505 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]));
506 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
507 MPASS(TAILQ_EMPTY(&ts->ts_pending));
512 * UMA zone item initializer.
515 turnstile_init(void *mem, int size, int flags)
517 struct turnstile *ts;
521 TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
522 TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]);
523 TAILQ_INIT(&ts->ts_pending);
524 LIST_INIT(&ts->ts_free);
525 mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN);
530 turnstile_fini(void *mem, int size)
532 struct turnstile *ts;
535 mtx_destroy(&ts->ts_lock);
539 * Get a turnstile for a new thread.
542 turnstile_alloc(void)
545 return (uma_zalloc(turnstile_zone, M_WAITOK));
549 * Free a turnstile when a thread is destroyed.
552 turnstile_free(struct turnstile *ts)
555 uma_zfree(turnstile_zone, ts);
559 * Lock the turnstile chain associated with the specified lock.
562 turnstile_chain_lock(struct lock_object *lock)
564 struct turnstile_chain *tc;
566 tc = TC_LOOKUP(lock);
567 mtx_lock_spin(&tc->tc_lock);
571 turnstile_trywait(struct lock_object *lock)
573 struct turnstile_chain *tc;
574 struct turnstile *ts;
576 tc = TC_LOOKUP(lock);
577 mtx_lock_spin(&tc->tc_lock);
578 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
579 if (ts->ts_lockobj == lock) {
580 mtx_lock_spin(&ts->ts_lock);
584 ts = curthread->td_turnstile;
586 mtx_lock_spin(&ts->ts_lock);
587 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
588 ts->ts_lockobj = lock;
594 turnstile_lock(struct turnstile *ts, struct lock_object **lockp,
597 struct turnstile_chain *tc;
598 struct lock_object *lock;
600 if ((lock = ts->ts_lockobj) == NULL)
602 tc = TC_LOOKUP(lock);
603 mtx_lock_spin(&tc->tc_lock);
604 mtx_lock_spin(&ts->ts_lock);
605 if (__predict_false(lock != ts->ts_lockobj)) {
606 mtx_unlock_spin(&tc->tc_lock);
607 mtx_unlock_spin(&ts->ts_lock);
616 turnstile_unlock(struct turnstile *ts, struct lock_object *lock)
618 struct turnstile_chain *tc;
620 mtx_assert(&ts->ts_lock, MA_OWNED);
621 mtx_unlock_spin(&ts->ts_lock);
622 if (ts == curthread->td_turnstile)
623 ts->ts_lockobj = NULL;
624 tc = TC_LOOKUP(lock);
625 mtx_unlock_spin(&tc->tc_lock);
629 turnstile_assert(struct turnstile *ts)
631 MPASS(ts->ts_lockobj == NULL);
635 turnstile_cancel(struct turnstile *ts)
637 struct turnstile_chain *tc;
638 struct lock_object *lock;
640 mtx_assert(&ts->ts_lock, MA_OWNED);
642 mtx_unlock_spin(&ts->ts_lock);
643 lock = ts->ts_lockobj;
644 if (ts == curthread->td_turnstile)
645 ts->ts_lockobj = NULL;
646 tc = TC_LOOKUP(lock);
647 mtx_unlock_spin(&tc->tc_lock);
651 * Look up the turnstile for a lock in the hash table locking the associated
652 * turnstile chain along the way. If no turnstile is found in the hash
653 * table, NULL is returned.
656 turnstile_lookup(struct lock_object *lock)
658 struct turnstile_chain *tc;
659 struct turnstile *ts;
661 tc = TC_LOOKUP(lock);
662 mtx_assert(&tc->tc_lock, MA_OWNED);
663 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
664 if (ts->ts_lockobj == lock) {
665 mtx_lock_spin(&ts->ts_lock);
672 * Unlock the turnstile chain associated with a given lock.
675 turnstile_chain_unlock(struct lock_object *lock)
677 struct turnstile_chain *tc;
679 tc = TC_LOOKUP(lock);
680 mtx_unlock_spin(&tc->tc_lock);
684 * Return a pointer to the thread waiting on this turnstile with the
685 * most important priority or NULL if the turnstile has no waiters.
687 static struct thread *
688 turnstile_first_waiter(struct turnstile *ts)
690 struct thread *std, *xtd;
692 std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]);
693 xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
694 if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority))
700 * Take ownership of a turnstile and adjust the priority of the new
701 * owner appropriately.
704 turnstile_claim(struct turnstile *ts)
706 struct thread *td, *owner;
707 struct turnstile_chain *tc;
709 mtx_assert(&ts->ts_lock, MA_OWNED);
710 MPASS(ts != curthread->td_turnstile);
713 mtx_lock_spin(&td_contested_lock);
714 turnstile_setowner(ts, owner);
715 mtx_unlock_spin(&td_contested_lock);
717 td = turnstile_first_waiter(ts);
719 MPASS(td->td_proc->p_magic == P_MAGIC);
720 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
723 * Update the priority of the new owner if needed.
726 if (td->td_priority < owner->td_priority)
727 sched_lend_prio(owner, td->td_priority);
728 thread_unlock(owner);
729 tc = TC_LOOKUP(ts->ts_lockobj);
730 mtx_unlock_spin(&ts->ts_lock);
731 mtx_unlock_spin(&tc->tc_lock);
735 * Block the current thread on the turnstile assicated with 'lock'. This
736 * function will context switch and not return until this thread has been
737 * woken back up. This function must be called with the appropriate
738 * turnstile chain locked and will return with it unlocked.
741 turnstile_wait(struct turnstile *ts, struct thread *owner, int queue)
743 struct turnstile_chain *tc;
744 struct thread *td, *td1;
745 struct lock_object *lock;
748 mtx_assert(&ts->ts_lock, MA_OWNED);
750 MPASS(owner->td_proc->p_magic == P_MAGIC);
751 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
754 * If the lock does not already have a turnstile, use this thread's
755 * turnstile. Otherwise insert the current thread into the
756 * turnstile already in use by this lock.
758 tc = TC_LOOKUP(ts->ts_lockobj);
759 mtx_assert(&tc->tc_lock, MA_OWNED);
760 if (ts == td->td_turnstile) {
761 #ifdef TURNSTILE_PROFILING
763 if (tc->tc_depth > tc->tc_max_depth) {
764 tc->tc_max_depth = tc->tc_depth;
765 if (tc->tc_max_depth > turnstile_max_depth)
766 turnstile_max_depth = tc->tc_max_depth;
769 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
770 KASSERT(TAILQ_EMPTY(&ts->ts_pending),
771 ("thread's turnstile has pending threads"));
772 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]),
773 ("thread's turnstile has exclusive waiters"));
774 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]),
775 ("thread's turnstile has shared waiters"));
776 KASSERT(LIST_EMPTY(&ts->ts_free),
777 ("thread's turnstile has a non-empty free list"));
778 MPASS(ts->ts_lockobj != NULL);
779 mtx_lock_spin(&td_contested_lock);
780 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
781 turnstile_setowner(ts, owner);
782 mtx_unlock_spin(&td_contested_lock);
784 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq)
785 if (td1->td_priority > td->td_priority)
787 mtx_lock_spin(&td_contested_lock);
789 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
791 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
792 MPASS(owner == ts->ts_owner);
793 mtx_unlock_spin(&td_contested_lock);
794 MPASS(td->td_turnstile != NULL);
795 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
798 thread_lock_set(td, &ts->ts_lock);
799 td->td_turnstile = NULL;
801 /* Save who we are blocked on and switch. */
802 lock = ts->ts_lockobj;
803 td->td_tsqueue = queue;
805 td->td_lockname = lock->lo_name;
806 td->td_blktick = ticks;
808 mtx_unlock_spin(&tc->tc_lock);
809 propagate_priority(td);
811 if (LOCK_LOG_TEST(lock, 0))
812 CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__,
813 td->td_tid, lock, lock->lo_name);
815 SDT_PROBE0(sched, , , sleep);
817 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
818 mi_switch(SW_VOL | SWT_TURNSTILE);
820 if (LOCK_LOG_TEST(lock, 0))
821 CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s",
822 __func__, td->td_tid, lock, lock->lo_name);
826 * Pick the highest priority thread on this turnstile and put it on the
827 * pending list. This must be called with the turnstile chain locked.
830 turnstile_signal(struct turnstile *ts, int queue)
832 struct turnstile_chain *tc __unused;
837 mtx_assert(&ts->ts_lock, MA_OWNED);
838 MPASS(curthread->td_proc->p_magic == P_MAGIC);
839 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
840 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
843 * Pick the highest priority thread blocked on this lock and
844 * move it to the pending list.
846 td = TAILQ_FIRST(&ts->ts_blocked[queue]);
847 MPASS(td->td_proc->p_magic == P_MAGIC);
848 mtx_lock_spin(&td_contested_lock);
849 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
850 mtx_unlock_spin(&td_contested_lock);
851 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
854 * If the turnstile is now empty, remove it from its chain and
855 * give it to the about-to-be-woken thread. Otherwise take a
856 * turnstile from the free list and give it to the thread.
858 empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
859 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]);
861 tc = TC_LOOKUP(ts->ts_lockobj);
862 mtx_assert(&tc->tc_lock, MA_OWNED);
863 MPASS(LIST_EMPTY(&ts->ts_free));
864 #ifdef TURNSTILE_PROFILING
868 ts = LIST_FIRST(&ts->ts_free);
870 LIST_REMOVE(ts, ts_hash);
871 td->td_turnstile = ts;
877 * Put all blocked threads on the pending list. This must be called with
878 * the turnstile chain locked.
881 turnstile_broadcast(struct turnstile *ts, int queue)
883 struct turnstile_chain *tc __unused;
884 struct turnstile *ts1;
888 mtx_assert(&ts->ts_lock, MA_OWNED);
889 MPASS(curthread->td_proc->p_magic == P_MAGIC);
890 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
892 * We must have the chain locked so that we can remove the empty
893 * turnstile from the hash queue.
895 tc = TC_LOOKUP(ts->ts_lockobj);
896 mtx_assert(&tc->tc_lock, MA_OWNED);
897 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
900 * Transfer the blocked list to the pending list.
902 mtx_lock_spin(&td_contested_lock);
903 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq);
904 mtx_unlock_spin(&td_contested_lock);
907 * Give a turnstile to each thread. The last thread gets
908 * this turnstile if the turnstile is empty.
910 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
911 if (LIST_EMPTY(&ts->ts_free)) {
912 MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
914 #ifdef TURNSTILE_PROFILING
918 ts1 = LIST_FIRST(&ts->ts_free);
920 LIST_REMOVE(ts1, ts_hash);
921 td->td_turnstile = ts1;
926 turnstile_calc_unlend_prio_locked(struct thread *td)
928 struct turnstile *nts;
931 THREAD_LOCK_ASSERT(td, MA_OWNED);
932 mtx_assert(&td_contested_lock, MA_OWNED);
935 LIST_FOREACH(nts, &td->td_contested, ts_link) {
936 cp = turnstile_first_waiter(nts)->td_priority;
944 * Wakeup all threads on the pending list and adjust the priority of the
945 * current thread appropriately. This must be called with the turnstile
949 turnstile_unpend(struct turnstile *ts)
951 TAILQ_HEAD( ,thread) pending_threads;
956 mtx_assert(&ts->ts_lock, MA_OWNED);
957 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
958 MPASS(!TAILQ_EMPTY(&ts->ts_pending));
961 * Move the list of pending threads out of the turnstile and
962 * into a local variable.
964 TAILQ_INIT(&pending_threads);
965 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
967 if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
968 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]))
969 ts->ts_lockobj = NULL;
972 * Adjust the priority of curthread based on other contested
973 * locks it owns. Don't lower the priority below the base
978 mtx_lock_spin(&td_contested_lock);
980 * Remove the turnstile from this thread's list of contested locks
981 * since this thread doesn't own it anymore. New threads will
982 * not be blocking on the turnstile until it is claimed by a new
983 * owner. There might not be a current owner if this is a shared
986 if (ts->ts_owner != NULL) {
988 LIST_REMOVE(ts, ts_link);
990 pri = turnstile_calc_unlend_prio_locked(td);
991 mtx_unlock_spin(&td_contested_lock);
992 sched_unlend_prio(td, pri);
995 * Wake up all the pending threads. If a thread is not blocked
996 * on a lock, then it is currently executing on another CPU in
997 * turnstile_wait() or sitting on a run queue waiting to resume
998 * in turnstile_wait(). Set a flag to force it to try to acquire
999 * the lock again instead of blocking.
1001 while (!TAILQ_EMPTY(&pending_threads)) {
1002 td = TAILQ_FIRST(&pending_threads);
1003 TAILQ_REMOVE(&pending_threads, td, td_lockq);
1004 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
1005 thread_lock_block_wait(td);
1006 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
1007 MPASS(td->td_proc->p_magic == P_MAGIC);
1008 MPASS(TD_ON_LOCK(td));
1010 MPASS(TD_CAN_RUN(td));
1011 td->td_blocked = NULL;
1012 td->td_lockname = NULL;
1015 td->td_tsqueue = 0xff;
1017 sched_add(td, SRQ_HOLD | SRQ_BORING);
1019 mtx_unlock_spin(&ts->ts_lock);
1023 * Give up ownership of a turnstile. This must be called with the
1024 * turnstile chain locked.
1027 turnstile_disown(struct turnstile *ts)
1033 mtx_assert(&ts->ts_lock, MA_OWNED);
1034 MPASS(ts->ts_owner == curthread);
1035 MPASS(TAILQ_EMPTY(&ts->ts_pending));
1036 MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) ||
1037 !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
1040 * Remove the turnstile from this thread's list of contested locks
1041 * since this thread doesn't own it anymore. New threads will
1042 * not be blocking on the turnstile until it is claimed by a new
1045 mtx_lock_spin(&td_contested_lock);
1046 ts->ts_owner = NULL;
1047 LIST_REMOVE(ts, ts_link);
1048 mtx_unlock_spin(&td_contested_lock);
1051 * Adjust the priority of curthread based on other contested
1052 * locks it owns. Don't lower the priority below the base
1057 mtx_unlock_spin(&ts->ts_lock);
1058 mtx_lock_spin(&td_contested_lock);
1059 pri = turnstile_calc_unlend_prio_locked(td);
1060 mtx_unlock_spin(&td_contested_lock);
1061 sched_unlend_prio(td, pri);
1066 * Return the first thread in a turnstile.
1069 turnstile_head(struct turnstile *ts, int queue)
1074 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1075 mtx_assert(&ts->ts_lock, MA_OWNED);
1077 return (TAILQ_FIRST(&ts->ts_blocked[queue]));
1081 * Returns true if a sub-queue of a turnstile is empty.
1084 turnstile_empty(struct turnstile *ts, int queue)
1089 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1090 mtx_assert(&ts->ts_lock, MA_OWNED);
1092 return (TAILQ_EMPTY(&ts->ts_blocked[queue]));
1097 print_thread(struct thread *td, const char *prefix)
1100 db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid,
1101 td->td_proc->p_pid, td->td_name);
1105 print_queue(struct threadqueue *queue, const char *header, const char *prefix)
1109 db_printf("%s:\n", header);
1110 if (TAILQ_EMPTY(queue)) {
1111 db_printf("%sempty\n", prefix);
1114 TAILQ_FOREACH(td, queue, td_lockq) {
1115 print_thread(td, prefix);
1119 DB_SHOW_COMMAND(turnstile, db_show_turnstile)
1121 struct turnstile_chain *tc;
1122 struct turnstile *ts;
1123 struct lock_object *lock;
1130 * First, see if there is an active turnstile for the lock indicated
1133 lock = (struct lock_object *)addr;
1134 tc = TC_LOOKUP(lock);
1135 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1136 if (ts->ts_lockobj == lock)
1140 * Second, see if there is an active turnstile at the address
1143 for (i = 0; i < TC_TABLESIZE; i++)
1144 LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) {
1145 if (ts == (struct turnstile *)addr)
1149 db_printf("Unable to locate a turnstile via %p\n", (void *)addr);
1152 lock = ts->ts_lockobj;
1153 db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name,
1156 print_thread(ts->ts_owner, "Lock Owner: ");
1158 db_printf("Lock Owner: none\n");
1159 print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t");
1160 print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters",
1162 print_queue(&ts->ts_pending, "Pending Threads", "\t");
1167 * Show all the threads a particular thread is waiting on based on
1171 print_lockchain(struct thread *td, const char *prefix)
1173 struct lock_object *lock;
1174 struct lock_class *class;
1175 struct turnstile *ts;
1176 struct thread *owner;
1179 * Follow the chain. We keep walking as long as the thread is
1180 * blocked on a lock that has an owner.
1182 while (!db_pager_quit) {
1183 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1184 td->td_proc->p_pid, td->td_name);
1185 switch (td->td_state) {
1187 db_printf("is inactive\n");
1190 db_printf("can run\n");
1193 db_printf("is on a run queue\n");
1196 db_printf("running on CPU %d\n", td->td_oncpu);
1199 if (TD_ON_LOCK(td)) {
1200 ts = td->td_blocked;
1201 lock = ts->ts_lockobj;
1202 class = LOCK_CLASS(lock);
1203 db_printf("blocked on lock %p (%s) \"%s\"\n",
1204 lock, class->lc_name, lock->lo_name);
1205 if (ts->ts_owner == NULL)
1209 } else if (TD_ON_SLEEPQ(td)) {
1210 if (!lockmgr_chain(td, &owner) &&
1211 !sx_chain(td, &owner)) {
1212 db_printf("sleeping on %p \"%s\"\n",
1213 td->td_wchan, td->td_wmesg);
1221 db_printf("inhibited\n");
1224 db_printf("??? (%#x)\n", td->td_state);
1230 DB_SHOW_COMMAND(lockchain, db_show_lockchain)
1234 /* Figure out which thread to start with. */
1236 td = db_lookup_thread(addr, true);
1240 print_lockchain(td, "");
1242 DB_SHOW_ALIAS(sleepchain, db_show_lockchain);
1244 DB_SHOW_ALL_COMMAND(chains, db_show_allchains)
1251 FOREACH_PROC_IN_SYSTEM(p) {
1252 FOREACH_THREAD_IN_PROC(p, td) {
1253 if ((TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested))
1254 || (TD_IS_INHIBITED(td) && TD_ON_SLEEPQ(td))) {
1255 db_printf("chain %d:\n", i++);
1256 print_lockchain(td, " ");
1263 DB_SHOW_ALIAS(allchains, db_show_allchains)
1265 static void print_waiters(struct turnstile *ts, int indent);
1268 print_waiter(struct thread *td, int indent)
1270 struct turnstile *ts;
1275 for (i = 0; i < indent; i++)
1277 print_thread(td, "thread ");
1278 LIST_FOREACH(ts, &td->td_contested, ts_link)
1279 print_waiters(ts, indent + 1);
1283 print_waiters(struct turnstile *ts, int indent)
1285 struct lock_object *lock;
1286 struct lock_class *class;
1292 lock = ts->ts_lockobj;
1293 class = LOCK_CLASS(lock);
1294 for (i = 0; i < indent; i++)
1296 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name);
1297 TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq)
1298 print_waiter(td, indent + 1);
1299 TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq)
1300 print_waiter(td, indent + 1);
1301 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq)
1302 print_waiter(td, indent + 1);
1305 DB_SHOW_COMMAND(locktree, db_show_locktree)
1307 struct lock_object *lock;
1308 struct lock_class *class;
1309 struct turnstile_chain *tc;
1310 struct turnstile *ts;
1314 lock = (struct lock_object *)addr;
1315 tc = TC_LOOKUP(lock);
1316 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1317 if (ts->ts_lockobj == lock)
1320 class = LOCK_CLASS(lock);
1321 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name,
1324 print_waiters(ts, 0);