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, 0,
145 "turnstile profiling");
146 static SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0,
147 "turnstile chain stats");
148 SYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD,
149 &turnstile_max_depth, 0, "maximum depth achieved of a single chain");
151 static struct mtx td_contested_lock;
152 static struct turnstile_chain turnstile_chains[TC_TABLESIZE];
153 static uma_zone_t turnstile_zone;
156 * Prototypes for non-exported routines.
158 static void init_turnstile0(void *dummy);
159 #ifdef TURNSTILE_PROFILING
160 static void init_turnstile_profiling(void *arg);
162 static void propagate_priority(struct thread *td);
163 static int turnstile_adjust_thread(struct turnstile *ts,
165 static struct thread *turnstile_first_waiter(struct turnstile *ts);
166 static void turnstile_setowner(struct turnstile *ts, struct thread *owner);
168 static void turnstile_dtor(void *mem, int size, void *arg);
170 static int turnstile_init(void *mem, int size, int flags);
171 static void turnstile_fini(void *mem, int size);
173 SDT_PROVIDER_DECLARE(sched);
174 SDT_PROBE_DEFINE(sched, , , sleep);
175 SDT_PROBE_DEFINE2(sched, , , wakeup, "struct thread *",
179 propagate_unlock_ts(struct turnstile *top, struct turnstile *ts)
183 mtx_unlock_spin(&ts->ts_lock);
187 propagate_unlock_td(struct turnstile *top, struct thread *td)
190 if (td->td_lock != &top->ts_lock)
195 * Walks the chain of turnstiles and their owners to propagate the priority
196 * of the thread being blocked to all the threads holding locks that have to
197 * release their locks before this thread can run again.
200 propagate_priority(struct thread *td)
202 struct turnstile *ts, *top;
205 THREAD_LOCK_ASSERT(td, MA_OWNED);
206 pri = td->td_priority;
207 top = ts = td->td_blocked;
208 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
211 * The original turnstile lock is held across the entire
212 * operation. We only ever lock down the chain so the lock
220 * This might be a read lock with no owner. There's
221 * not much we can do, so just bail.
223 propagate_unlock_ts(top, ts);
228 * Wait for the thread lock to be stable and then only
229 * acquire if it is not the turnstile lock.
231 thread_lock_block_wait(td);
232 if (td->td_lock != &ts->ts_lock) {
233 thread_lock_flags(td, MTX_DUPOK);
234 propagate_unlock_ts(top, ts);
236 MPASS(td->td_proc != NULL);
237 MPASS(td->td_proc->p_magic == P_MAGIC);
240 * If the thread is asleep, then we are probably about
241 * to deadlock. To make debugging this easier, show
242 * backtrace of misbehaving thread and panic to not
243 * leave the kernel deadlocked.
245 if (TD_IS_SLEEPING(td)) {
247 "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
248 td->td_tid, td->td_proc->p_pid);
249 kdb_backtrace_thread(td);
250 panic("sleeping thread");
254 * If this thread already has higher priority than the
255 * thread that is being blocked, we are finished.
257 if (td->td_priority <= pri) {
258 propagate_unlock_td(top, td);
263 * Bump this thread's priority.
265 sched_lend_prio(td, pri);
268 * If lock holder is actually running or on the run queue
271 if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) {
272 MPASS(td->td_blocked == NULL);
273 propagate_unlock_td(top, td);
279 * For UP, we check to see if td is curthread (this shouldn't
280 * ever happen however as it would mean we are in a deadlock.)
282 KASSERT(td != curthread, ("Deadlock detected"));
286 * If we aren't blocked on a lock, we should be.
288 KASSERT(TD_ON_LOCK(td), (
289 "thread %d(%s):%d holds %s but isn't blocked on a lock\n",
290 td->td_tid, td->td_name, td->td_state,
291 ts->ts_lockobj->lo_name));
294 * Pick up the lock that td is blocked on.
298 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
299 /* Resort td on the list if needed. */
300 if (!turnstile_adjust_thread(ts, td)) {
301 propagate_unlock_ts(top, ts);
304 /* The thread lock is released as ts lock above. */
309 * Adjust the thread's position on a turnstile after its priority has been
313 turnstile_adjust_thread(struct turnstile *ts, struct thread *td)
315 struct thread *td1, *td2;
318 THREAD_LOCK_ASSERT(td, MA_OWNED);
319 MPASS(TD_ON_LOCK(td));
322 * This thread may not be blocked on this turnstile anymore
323 * but instead might already be woken up on another CPU
324 * that is waiting on the thread lock in turnstile_unpend() to
325 * finish waking this thread up. We can detect this case
326 * by checking to see if this thread has been given a
327 * turnstile by either turnstile_signal() or
328 * turnstile_broadcast(). In this case, treat the thread as
329 * if it was already running.
331 if (td->td_turnstile != NULL)
335 * Check if the thread needs to be moved on the blocked chain.
336 * It needs to be moved if either its priority is lower than
337 * the previous thread or higher than the next thread.
339 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
340 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
341 td2 = TAILQ_NEXT(td, td_lockq);
342 if ((td1 != NULL && td->td_priority < td1->td_priority) ||
343 (td2 != NULL && td->td_priority > td2->td_priority)) {
346 * Remove thread from blocked chain and determine where
347 * it should be moved to.
349 queue = td->td_tsqueue;
350 MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE);
351 mtx_lock_spin(&td_contested_lock);
352 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
353 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) {
354 MPASS(td1->td_proc->p_magic == P_MAGIC);
355 if (td1->td_priority > td->td_priority)
360 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
362 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
363 mtx_unlock_spin(&td_contested_lock);
366 "turnstile_adjust_thread: td %d put at tail on [%p] %s",
367 td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name);
370 "turnstile_adjust_thread: td %d moved before %d on [%p] %s",
371 td->td_tid, td1->td_tid, ts->ts_lockobj,
372 ts->ts_lockobj->lo_name);
378 * Early initialization of turnstiles. This is not done via a SYSINIT()
379 * since this needs to be initialized very early when mutexes are first
383 init_turnstiles(void)
387 for (i = 0; i < TC_TABLESIZE; i++) {
388 LIST_INIT(&turnstile_chains[i].tc_turnstiles);
389 mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain",
392 mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN);
393 LIST_INIT(&thread0.td_contested);
394 thread0.td_turnstile = NULL;
397 #ifdef TURNSTILE_PROFILING
399 init_turnstile_profiling(void *arg)
401 struct sysctl_oid *chain_oid;
405 for (i = 0; i < TC_TABLESIZE; i++) {
406 snprintf(chain_name, sizeof(chain_name), "%d", i);
407 chain_oid = SYSCTL_ADD_NODE(NULL,
408 SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO,
409 chain_name, CTLFLAG_RD, NULL, "turnstile chain stats");
410 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
411 "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0,
413 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
414 "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth,
418 SYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
419 init_turnstile_profiling, NULL);
423 init_turnstile0(void *dummy)
426 turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile),
433 turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE);
434 thread0.td_turnstile = turnstile_alloc();
436 SYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL);
439 * Update a thread on the turnstile list after it's priority has been changed.
440 * The old priority is passed in as an argument.
443 turnstile_adjust(struct thread *td, u_char oldpri)
445 struct turnstile *ts;
447 MPASS(TD_ON_LOCK(td));
450 * Pick up the lock that td is blocked on.
454 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
455 mtx_assert(&ts->ts_lock, MA_OWNED);
457 /* Resort the turnstile on the list. */
458 if (!turnstile_adjust_thread(ts, td))
461 * If our priority was lowered and we are at the head of the
462 * turnstile, then propagate our new priority up the chain.
463 * Note that we currently don't try to revoke lent priorities
464 * when our priority goes up.
466 MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE ||
467 td->td_tsqueue == TS_SHARED_QUEUE);
468 if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) &&
469 td->td_priority < oldpri) {
470 propagate_priority(td);
475 * Set the owner of the lock this turnstile is attached to.
478 turnstile_setowner(struct turnstile *ts, struct thread *owner)
481 mtx_assert(&td_contested_lock, MA_OWNED);
482 MPASS(ts->ts_owner == NULL);
484 /* A shared lock might not have an owner. */
488 MPASS(owner->td_proc->p_magic == P_MAGIC);
489 ts->ts_owner = owner;
490 LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link);
495 * UMA zone item deallocator.
498 turnstile_dtor(void *mem, int size, void *arg)
500 struct turnstile *ts;
503 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]));
504 MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
505 MPASS(TAILQ_EMPTY(&ts->ts_pending));
510 * UMA zone item initializer.
513 turnstile_init(void *mem, int size, int flags)
515 struct turnstile *ts;
519 TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
520 TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]);
521 TAILQ_INIT(&ts->ts_pending);
522 LIST_INIT(&ts->ts_free);
523 mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN);
528 turnstile_fini(void *mem, int size)
530 struct turnstile *ts;
533 mtx_destroy(&ts->ts_lock);
537 * Get a turnstile for a new thread.
540 turnstile_alloc(void)
543 return (uma_zalloc(turnstile_zone, M_WAITOK));
547 * Free a turnstile when a thread is destroyed.
550 turnstile_free(struct turnstile *ts)
553 uma_zfree(turnstile_zone, ts);
557 * Lock the turnstile chain associated with the specified lock.
560 turnstile_chain_lock(struct lock_object *lock)
562 struct turnstile_chain *tc;
564 tc = TC_LOOKUP(lock);
565 mtx_lock_spin(&tc->tc_lock);
569 turnstile_trywait(struct lock_object *lock)
571 struct turnstile_chain *tc;
572 struct turnstile *ts;
574 tc = TC_LOOKUP(lock);
575 mtx_lock_spin(&tc->tc_lock);
576 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
577 if (ts->ts_lockobj == lock) {
578 mtx_lock_spin(&ts->ts_lock);
582 ts = curthread->td_turnstile;
584 mtx_lock_spin(&ts->ts_lock);
585 KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer"));
586 ts->ts_lockobj = lock;
592 turnstile_lock(struct turnstile *ts, struct lock_object **lockp,
595 struct turnstile_chain *tc;
596 struct lock_object *lock;
598 if ((lock = ts->ts_lockobj) == NULL)
600 tc = TC_LOOKUP(lock);
601 mtx_lock_spin(&tc->tc_lock);
602 mtx_lock_spin(&ts->ts_lock);
603 if (__predict_false(lock != ts->ts_lockobj)) {
604 mtx_unlock_spin(&tc->tc_lock);
605 mtx_unlock_spin(&ts->ts_lock);
614 turnstile_unlock(struct turnstile *ts, struct lock_object *lock)
616 struct turnstile_chain *tc;
618 mtx_assert(&ts->ts_lock, MA_OWNED);
619 mtx_unlock_spin(&ts->ts_lock);
620 if (ts == curthread->td_turnstile)
621 ts->ts_lockobj = NULL;
622 tc = TC_LOOKUP(lock);
623 mtx_unlock_spin(&tc->tc_lock);
627 turnstile_assert(struct turnstile *ts)
629 MPASS(ts->ts_lockobj == NULL);
633 turnstile_cancel(struct turnstile *ts)
635 struct turnstile_chain *tc;
636 struct lock_object *lock;
638 mtx_assert(&ts->ts_lock, MA_OWNED);
640 mtx_unlock_spin(&ts->ts_lock);
641 lock = ts->ts_lockobj;
642 if (ts == curthread->td_turnstile)
643 ts->ts_lockobj = NULL;
644 tc = TC_LOOKUP(lock);
645 mtx_unlock_spin(&tc->tc_lock);
649 * Look up the turnstile for a lock in the hash table locking the associated
650 * turnstile chain along the way. If no turnstile is found in the hash
651 * table, NULL is returned.
654 turnstile_lookup(struct lock_object *lock)
656 struct turnstile_chain *tc;
657 struct turnstile *ts;
659 tc = TC_LOOKUP(lock);
660 mtx_assert(&tc->tc_lock, MA_OWNED);
661 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
662 if (ts->ts_lockobj == lock) {
663 mtx_lock_spin(&ts->ts_lock);
670 * Unlock the turnstile chain associated with a given lock.
673 turnstile_chain_unlock(struct lock_object *lock)
675 struct turnstile_chain *tc;
677 tc = TC_LOOKUP(lock);
678 mtx_unlock_spin(&tc->tc_lock);
682 * Return a pointer to the thread waiting on this turnstile with the
683 * most important priority or NULL if the turnstile has no waiters.
685 static struct thread *
686 turnstile_first_waiter(struct turnstile *ts)
688 struct thread *std, *xtd;
690 std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]);
691 xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]);
692 if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority))
698 * Take ownership of a turnstile and adjust the priority of the new
699 * owner appropriately.
702 turnstile_claim(struct turnstile *ts)
704 struct thread *td, *owner;
705 struct turnstile_chain *tc;
707 mtx_assert(&ts->ts_lock, MA_OWNED);
708 MPASS(ts != curthread->td_turnstile);
711 mtx_lock_spin(&td_contested_lock);
712 turnstile_setowner(ts, owner);
713 mtx_unlock_spin(&td_contested_lock);
715 td = turnstile_first_waiter(ts);
717 MPASS(td->td_proc->p_magic == P_MAGIC);
718 THREAD_LOCKPTR_BLOCKED_ASSERT(td, &ts->ts_lock);
721 * Update the priority of the new owner if needed.
724 if (td->td_priority < owner->td_priority)
725 sched_lend_prio(owner, td->td_priority);
726 thread_unlock(owner);
727 tc = TC_LOOKUP(ts->ts_lockobj);
728 mtx_unlock_spin(&ts->ts_lock);
729 mtx_unlock_spin(&tc->tc_lock);
733 * Block the current thread on the turnstile assicated with 'lock'. This
734 * function will context switch and not return until this thread has been
735 * woken back up. This function must be called with the appropriate
736 * turnstile chain locked and will return with it unlocked.
739 turnstile_wait(struct turnstile *ts, struct thread *owner, int queue)
741 struct turnstile_chain *tc;
742 struct thread *td, *td1;
743 struct lock_object *lock;
746 mtx_assert(&ts->ts_lock, MA_OWNED);
748 MPASS(owner->td_proc->p_magic == P_MAGIC);
749 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
752 * If the lock does not already have a turnstile, use this thread's
753 * turnstile. Otherwise insert the current thread into the
754 * turnstile already in use by this lock.
756 tc = TC_LOOKUP(ts->ts_lockobj);
757 mtx_assert(&tc->tc_lock, MA_OWNED);
758 if (ts == td->td_turnstile) {
759 #ifdef TURNSTILE_PROFILING
761 if (tc->tc_depth > tc->tc_max_depth) {
762 tc->tc_max_depth = tc->tc_depth;
763 if (tc->tc_max_depth > turnstile_max_depth)
764 turnstile_max_depth = tc->tc_max_depth;
767 LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash);
768 KASSERT(TAILQ_EMPTY(&ts->ts_pending),
769 ("thread's turnstile has pending threads"));
770 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]),
771 ("thread's turnstile has exclusive waiters"));
772 KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]),
773 ("thread's turnstile has shared waiters"));
774 KASSERT(LIST_EMPTY(&ts->ts_free),
775 ("thread's turnstile has a non-empty free list"));
776 MPASS(ts->ts_lockobj != NULL);
777 mtx_lock_spin(&td_contested_lock);
778 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
779 turnstile_setowner(ts, owner);
780 mtx_unlock_spin(&td_contested_lock);
782 TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq)
783 if (td1->td_priority > td->td_priority)
785 mtx_lock_spin(&td_contested_lock);
787 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
789 TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq);
790 MPASS(owner == ts->ts_owner);
791 mtx_unlock_spin(&td_contested_lock);
792 MPASS(td->td_turnstile != NULL);
793 LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash);
796 thread_lock_set(td, &ts->ts_lock);
797 td->td_turnstile = NULL;
799 /* Save who we are blocked on and switch. */
800 lock = ts->ts_lockobj;
801 td->td_tsqueue = queue;
803 td->td_lockname = lock->lo_name;
804 td->td_blktick = ticks;
806 mtx_unlock_spin(&tc->tc_lock);
807 propagate_priority(td);
809 if (LOCK_LOG_TEST(lock, 0))
810 CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__,
811 td->td_tid, lock, lock->lo_name);
813 SDT_PROBE0(sched, , , sleep);
815 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
816 mi_switch(SW_VOL | SWT_TURNSTILE);
818 if (LOCK_LOG_TEST(lock, 0))
819 CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s",
820 __func__, td->td_tid, lock, lock->lo_name);
824 * Pick the highest priority thread on this turnstile and put it on the
825 * pending list. This must be called with the turnstile chain locked.
828 turnstile_signal(struct turnstile *ts, int queue)
830 struct turnstile_chain *tc __unused;
835 mtx_assert(&ts->ts_lock, MA_OWNED);
836 MPASS(curthread->td_proc->p_magic == P_MAGIC);
837 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
838 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
841 * Pick the highest priority thread blocked on this lock and
842 * move it to the pending list.
844 td = TAILQ_FIRST(&ts->ts_blocked[queue]);
845 MPASS(td->td_proc->p_magic == P_MAGIC);
846 mtx_lock_spin(&td_contested_lock);
847 TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq);
848 mtx_unlock_spin(&td_contested_lock);
849 TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq);
852 * If the turnstile is now empty, remove it from its chain and
853 * give it to the about-to-be-woken thread. Otherwise take a
854 * turnstile from the free list and give it to the thread.
856 empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
857 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]);
859 tc = TC_LOOKUP(ts->ts_lockobj);
860 mtx_assert(&tc->tc_lock, MA_OWNED);
861 MPASS(LIST_EMPTY(&ts->ts_free));
862 #ifdef TURNSTILE_PROFILING
866 ts = LIST_FIRST(&ts->ts_free);
868 LIST_REMOVE(ts, ts_hash);
869 td->td_turnstile = ts;
875 * Put all blocked threads on the pending list. This must be called with
876 * the turnstile chain locked.
879 turnstile_broadcast(struct turnstile *ts, int queue)
881 struct turnstile_chain *tc __unused;
882 struct turnstile *ts1;
886 mtx_assert(&ts->ts_lock, MA_OWNED);
887 MPASS(curthread->td_proc->p_magic == P_MAGIC);
888 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
890 * We must have the chain locked so that we can remove the empty
891 * turnstile from the hash queue.
893 tc = TC_LOOKUP(ts->ts_lockobj);
894 mtx_assert(&tc->tc_lock, MA_OWNED);
895 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
898 * Transfer the blocked list to the pending list.
900 mtx_lock_spin(&td_contested_lock);
901 TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq);
902 mtx_unlock_spin(&td_contested_lock);
905 * Give a turnstile to each thread. The last thread gets
906 * this turnstile if the turnstile is empty.
908 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) {
909 if (LIST_EMPTY(&ts->ts_free)) {
910 MPASS(TAILQ_NEXT(td, td_lockq) == NULL);
912 #ifdef TURNSTILE_PROFILING
916 ts1 = LIST_FIRST(&ts->ts_free);
918 LIST_REMOVE(ts1, ts_hash);
919 td->td_turnstile = ts1;
924 turnstile_calc_unlend_prio_locked(struct thread *td)
926 struct turnstile *nts;
929 THREAD_LOCK_ASSERT(td, MA_OWNED);
930 mtx_assert(&td_contested_lock, MA_OWNED);
933 LIST_FOREACH(nts, &td->td_contested, ts_link) {
934 cp = turnstile_first_waiter(nts)->td_priority;
942 * Wakeup all threads on the pending list and adjust the priority of the
943 * current thread appropriately. This must be called with the turnstile
947 turnstile_unpend(struct turnstile *ts)
949 TAILQ_HEAD( ,thread) pending_threads;
954 mtx_assert(&ts->ts_lock, MA_OWNED);
955 MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL);
956 MPASS(!TAILQ_EMPTY(&ts->ts_pending));
959 * Move the list of pending threads out of the turnstile and
960 * into a local variable.
962 TAILQ_INIT(&pending_threads);
963 TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq);
965 if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) &&
966 TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]))
967 ts->ts_lockobj = NULL;
970 * Adjust the priority of curthread based on other contested
971 * locks it owns. Don't lower the priority below the base
976 mtx_lock_spin(&td_contested_lock);
978 * Remove the turnstile from this thread's list of contested locks
979 * since this thread doesn't own it anymore. New threads will
980 * not be blocking on the turnstile until it is claimed by a new
981 * owner. There might not be a current owner if this is a shared
984 if (ts->ts_owner != NULL) {
986 LIST_REMOVE(ts, ts_link);
988 pri = turnstile_calc_unlend_prio_locked(td);
989 mtx_unlock_spin(&td_contested_lock);
990 sched_unlend_prio(td, pri);
993 * Wake up all the pending threads. If a thread is not blocked
994 * on a lock, then it is currently executing on another CPU in
995 * turnstile_wait() or sitting on a run queue waiting to resume
996 * in turnstile_wait(). Set a flag to force it to try to acquire
997 * the lock again instead of blocking.
999 while (!TAILQ_EMPTY(&pending_threads)) {
1000 td = TAILQ_FIRST(&pending_threads);
1001 TAILQ_REMOVE(&pending_threads, td, td_lockq);
1002 SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
1003 thread_lock_block_wait(td);
1004 THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock);
1005 MPASS(td->td_proc->p_magic == P_MAGIC);
1006 MPASS(TD_ON_LOCK(td));
1008 MPASS(TD_CAN_RUN(td));
1009 td->td_blocked = NULL;
1010 td->td_lockname = NULL;
1013 td->td_tsqueue = 0xff;
1015 sched_add(td, SRQ_HOLD | SRQ_BORING);
1017 mtx_unlock_spin(&ts->ts_lock);
1021 * Give up ownership of a turnstile. This must be called with the
1022 * turnstile chain locked.
1025 turnstile_disown(struct turnstile *ts)
1031 mtx_assert(&ts->ts_lock, MA_OWNED);
1032 MPASS(ts->ts_owner == curthread);
1033 MPASS(TAILQ_EMPTY(&ts->ts_pending));
1034 MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) ||
1035 !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]));
1038 * Remove the turnstile from this thread's list of contested locks
1039 * since this thread doesn't own it anymore. New threads will
1040 * not be blocking on the turnstile until it is claimed by a new
1043 mtx_lock_spin(&td_contested_lock);
1044 ts->ts_owner = NULL;
1045 LIST_REMOVE(ts, ts_link);
1046 mtx_unlock_spin(&td_contested_lock);
1049 * Adjust the priority of curthread based on other contested
1050 * locks it owns. Don't lower the priority below the base
1055 mtx_unlock_spin(&ts->ts_lock);
1056 mtx_lock_spin(&td_contested_lock);
1057 pri = turnstile_calc_unlend_prio_locked(td);
1058 mtx_unlock_spin(&td_contested_lock);
1059 sched_unlend_prio(td, pri);
1064 * Return the first thread in a turnstile.
1067 turnstile_head(struct turnstile *ts, int queue)
1072 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1073 mtx_assert(&ts->ts_lock, MA_OWNED);
1075 return (TAILQ_FIRST(&ts->ts_blocked[queue]));
1079 * Returns true if a sub-queue of a turnstile is empty.
1082 turnstile_empty(struct turnstile *ts, int queue)
1087 MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE);
1088 mtx_assert(&ts->ts_lock, MA_OWNED);
1090 return (TAILQ_EMPTY(&ts->ts_blocked[queue]));
1095 print_thread(struct thread *td, const char *prefix)
1098 db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid,
1099 td->td_proc->p_pid, td->td_name);
1103 print_queue(struct threadqueue *queue, const char *header, const char *prefix)
1107 db_printf("%s:\n", header);
1108 if (TAILQ_EMPTY(queue)) {
1109 db_printf("%sempty\n", prefix);
1112 TAILQ_FOREACH(td, queue, td_lockq) {
1113 print_thread(td, prefix);
1117 DB_SHOW_COMMAND(turnstile, db_show_turnstile)
1119 struct turnstile_chain *tc;
1120 struct turnstile *ts;
1121 struct lock_object *lock;
1128 * First, see if there is an active turnstile for the lock indicated
1131 lock = (struct lock_object *)addr;
1132 tc = TC_LOOKUP(lock);
1133 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1134 if (ts->ts_lockobj == lock)
1138 * Second, see if there is an active turnstile at the address
1141 for (i = 0; i < TC_TABLESIZE; i++)
1142 LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) {
1143 if (ts == (struct turnstile *)addr)
1147 db_printf("Unable to locate a turnstile via %p\n", (void *)addr);
1150 lock = ts->ts_lockobj;
1151 db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name,
1154 print_thread(ts->ts_owner, "Lock Owner: ");
1156 db_printf("Lock Owner: none\n");
1157 print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t");
1158 print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters",
1160 print_queue(&ts->ts_pending, "Pending Threads", "\t");
1165 * Show all the threads a particular thread is waiting on based on
1169 print_lockchain(struct thread *td, const char *prefix)
1171 struct lock_object *lock;
1172 struct lock_class *class;
1173 struct turnstile *ts;
1174 struct thread *owner;
1177 * Follow the chain. We keep walking as long as the thread is
1178 * blocked on a lock that has an owner.
1180 while (!db_pager_quit) {
1181 db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid,
1182 td->td_proc->p_pid, td->td_name);
1183 switch (td->td_state) {
1185 db_printf("is inactive\n");
1188 db_printf("can run\n");
1191 db_printf("is on a run queue\n");
1194 db_printf("running on CPU %d\n", td->td_oncpu);
1197 if (TD_ON_LOCK(td)) {
1198 ts = td->td_blocked;
1199 lock = ts->ts_lockobj;
1200 class = LOCK_CLASS(lock);
1201 db_printf("blocked on lock %p (%s) \"%s\"\n",
1202 lock, class->lc_name, lock->lo_name);
1203 if (ts->ts_owner == NULL)
1207 } else if (TD_ON_SLEEPQ(td)) {
1208 if (!lockmgr_chain(td, &owner) &&
1209 !sx_chain(td, &owner)) {
1210 db_printf("sleeping on %p \"%s\"\n",
1211 td->td_wchan, td->td_wmesg);
1219 db_printf("inhibited\n");
1222 db_printf("??? (%#x)\n", td->td_state);
1228 DB_SHOW_COMMAND(lockchain, db_show_lockchain)
1232 /* Figure out which thread to start with. */
1234 td = db_lookup_thread(addr, true);
1238 print_lockchain(td, "");
1240 DB_SHOW_ALIAS(sleepchain, db_show_lockchain);
1242 DB_SHOW_ALL_COMMAND(chains, db_show_allchains)
1249 FOREACH_PROC_IN_SYSTEM(p) {
1250 FOREACH_THREAD_IN_PROC(p, td) {
1251 if ((TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested))
1252 || (TD_IS_INHIBITED(td) && TD_ON_SLEEPQ(td))) {
1253 db_printf("chain %d:\n", i++);
1254 print_lockchain(td, " ");
1261 DB_SHOW_ALIAS(allchains, db_show_allchains)
1263 static void print_waiters(struct turnstile *ts, int indent);
1266 print_waiter(struct thread *td, int indent)
1268 struct turnstile *ts;
1273 for (i = 0; i < indent; i++)
1275 print_thread(td, "thread ");
1276 LIST_FOREACH(ts, &td->td_contested, ts_link)
1277 print_waiters(ts, indent + 1);
1281 print_waiters(struct turnstile *ts, int indent)
1283 struct lock_object *lock;
1284 struct lock_class *class;
1290 lock = ts->ts_lockobj;
1291 class = LOCK_CLASS(lock);
1292 for (i = 0; i < indent; i++)
1294 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name);
1295 TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq)
1296 print_waiter(td, indent + 1);
1297 TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq)
1298 print_waiter(td, indent + 1);
1299 TAILQ_FOREACH(td, &ts->ts_pending, td_lockq)
1300 print_waiter(td, indent + 1);
1303 DB_SHOW_COMMAND(locktree, db_show_locktree)
1305 struct lock_object *lock;
1306 struct lock_class *class;
1307 struct turnstile_chain *tc;
1308 struct turnstile *ts;
1312 lock = (struct lock_object *)addr;
1313 tc = TC_LOOKUP(lock);
1314 LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash)
1315 if (ts->ts_lockobj == lock)
1318 class = LOCK_CLASS(lock);
1319 db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name,
1322 print_waiters(ts, 0);