2 * Copyright (c) 2009 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Implement fast persistent locks based on atomic_cmpset_int() with
36 * semantics similar to lockmgr locks but faster and taking up much less
37 * space. Taken from HAMMER's lock implementation.
39 * These are meant to complement our LWKT tokens. Tokens are only held
40 * while the thread is running. Mutexes can be held across blocking
43 * Most of the support is in sys/mutex[2].h. We mostly provide backoff
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/kernel.h>
50 #include <sys/sysctl.h>
51 #include <sys/thread.h>
53 #include <machine/cpufunc.h>
55 #include <sys/thread2.h>
56 #include <sys/mutex2.h>
58 static __int64_t mtx_contention_count;
59 static __int64_t mtx_collision_count;
60 static __int64_t mtx_wakeup_count;
62 SYSCTL_QUAD(_kern, OID_AUTO, mtx_contention_count, CTLFLAG_RW,
63 &mtx_contention_count, 0, "");
64 SYSCTL_QUAD(_kern, OID_AUTO, mtx_collision_count, CTLFLAG_RW,
65 &mtx_collision_count, 0, "");
66 SYSCTL_QUAD(_kern, OID_AUTO, mtx_wakeup_count, CTLFLAG_RW,
67 &mtx_wakeup_count, 0, "");
69 static void mtx_chain_link(mtx_t mtx);
70 static void mtx_delete_link(mtx_t mtx, mtx_link_t link);
73 * Exclusive-lock a mutex, block until acquired. Recursion is allowed.
75 * Returns 0 on success, or the tsleep() return code on failure.
76 * An error can only be returned if PCATCH is specified in the flags.
79 __mtx_lock_ex(mtx_t mtx, mtx_link_t link, const char *ident, int flags, int to)
88 nlock = MTX_EXCLUSIVE | 1;
89 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
90 mtx->mtx_owner = curthread;
94 } else if ((lock & MTX_EXCLUSIVE) &&
95 mtx->mtx_owner == curthread) {
96 KKASSERT((lock & MTX_MASK) != MTX_MASK);
98 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
104 * Clearing MTX_EXLINK in lock causes us to loop until
105 * MTX_EXLINK is available. However, to avoid
106 * unnecessary cpu cache traffic we poll instead.
108 * Setting MTX_EXLINK in nlock causes us to loop until
109 * we can acquire MTX_EXLINK.
111 * Also set MTX_EXWANTED coincident with EXLINK, if
116 if (lock & MTX_EXLINK) {
118 ++mtx_collision_count;
122 /*lock &= ~MTX_EXLINK;*/
123 nlock = lock | MTX_EXWANTED | MTX_EXLINK;
125 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
127 * Check for early abort
129 if (link->state == MTX_LINK_ABORTED) {
130 atomic_clear_int(&mtx->mtx_lock,
134 if (mtx->mtx_link == NULL) {
135 atomic_clear_int(&mtx->mtx_lock,
142 * Success. Link in our structure then
143 * release EXLINK and sleep.
146 link->state = MTX_LINK_LINKED;
148 link->next = mtx->mtx_link;
149 link->prev = link->next->prev;
150 link->next->prev = link;
151 link->prev->next = link;
155 mtx->mtx_link = link;
157 tsleep_interlock(link, 0);
158 atomic_clear_int(&mtx->mtx_lock, MTX_EXLINK);
161 mycpu->gd_cnt.v_lock_name[0] = 'X';
162 strncpy(mycpu->gd_cnt.v_lock_name + 1,
164 sizeof(mycpu->gd_cnt.v_lock_name) - 2);
165 ++mycpu->gd_cnt.v_lock_colls;
167 error = tsleep(link, flags | PINTERLOCKED,
169 ++mtx_contention_count;
172 * Normal unlink, we should own the exclusive
175 if (link->state == MTX_LINK_LINKED)
176 mtx_delete_link(mtx, link);
177 if (link->state == MTX_LINK_ACQUIRED) {
178 KKASSERT(mtx->mtx_owner == link->owner);
184 * Aborted lock (mtx_abort_ex called).
186 if (link->state == MTX_LINK_ABORTED) {
192 * tsleep error, else retry.
200 ++mtx_collision_count;
206 _mtx_lock_ex_link(mtx_t mtx, mtx_link_t link,
207 const char *ident, int flags, int to)
209 return(__mtx_lock_ex(mtx, link, ident, flags, to));
213 _mtx_lock_ex(mtx_t mtx, const char *ident, int flags, int to)
215 struct mtx_link link;
217 mtx_link_init(&link);
218 return(__mtx_lock_ex(mtx, &link, ident, flags, to));
222 _mtx_lock_ex_quick(mtx_t mtx, const char *ident)
224 struct mtx_link link;
226 mtx_link_init(&link);
227 return(__mtx_lock_ex(mtx, &link, ident, 0, 0));
231 * Share-lock a mutex, block until acquired. Recursion is allowed.
233 * Returns 0 on success, or the tsleep() return code on failure.
234 * An error can only be returned if PCATCH is specified in the flags.
236 * NOTE: Shared locks get a mass-wakeup so if the tsleep fails we
237 * do not have to chain the wakeup().
240 __mtx_lock_sh(mtx_t mtx, const char *ident, int flags, int to)
247 lock = mtx->mtx_lock;
248 if ((lock & MTX_EXCLUSIVE) == 0) {
249 KKASSERT((lock & MTX_MASK) != MTX_MASK);
251 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
256 nlock = lock | MTX_SHWANTED;
257 tsleep_interlock(mtx, 0);
258 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
260 mycpu->gd_cnt.v_lock_name[0] = 'S';
261 strncpy(mycpu->gd_cnt.v_lock_name + 1,
263 sizeof(mycpu->gd_cnt.v_lock_name) - 2);
264 ++mycpu->gd_cnt.v_lock_colls;
266 error = tsleep(mtx, flags | PINTERLOCKED,
270 ++mtx_contention_count;
274 tsleep_remove(curthread);
278 ++mtx_collision_count;
284 _mtx_lock_sh(mtx_t mtx, const char *ident, int flags, int to)
286 return (__mtx_lock_sh(mtx, ident, flags, to));
290 _mtx_lock_sh_quick(mtx_t mtx, const char *ident)
292 return (__mtx_lock_sh(mtx, ident, 0, 0));
296 * Get an exclusive spinlock the hard way.
299 _mtx_spinlock(mtx_t mtx)
307 lock = mtx->mtx_lock;
309 nlock = MTX_EXCLUSIVE | 1;
310 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
311 mtx->mtx_owner = curthread;
314 } else if ((lock & MTX_EXCLUSIVE) &&
315 mtx->mtx_owner == curthread) {
316 KKASSERT((lock & MTX_MASK) != MTX_MASK);
318 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
325 for (bo = 0; bo < bb; ++bo)
327 ++mtx_contention_count;
330 ++mtx_collision_count;
335 * Attempt to acquire a spinlock, if we fail we must undo the
336 * gd->gd_spinlocks/gd->gd_curthead->td_critcount predisposition.
338 * Returns 0 on success, EAGAIN on failure.
341 _mtx_spinlock_try(mtx_t mtx)
343 globaldata_t gd = mycpu;
349 lock = mtx->mtx_lock;
351 nlock = MTX_EXCLUSIVE | 1;
352 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
353 mtx->mtx_owner = gd->gd_curthread;
356 } else if ((lock & MTX_EXCLUSIVE) &&
357 mtx->mtx_owner == gd->gd_curthread) {
358 KKASSERT((lock & MTX_MASK) != MTX_MASK);
360 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
365 --gd->gd_curthread->td_critcount;
370 ++mtx_collision_count;
378 _mtx_spinlock_sh(mtx_t mtx)
386 lock = mtx->mtx_lock;
387 if ((lock & MTX_EXCLUSIVE) == 0) {
388 KKASSERT((lock & MTX_MASK) != MTX_MASK);
390 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
397 for (bo = 0; bo < bb; ++bo)
399 ++mtx_contention_count;
402 ++mtx_collision_count;
409 _mtx_lock_ex_try(mtx_t mtx)
416 lock = mtx->mtx_lock;
418 nlock = MTX_EXCLUSIVE | 1;
419 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
420 mtx->mtx_owner = curthread;
423 } else if ((lock & MTX_EXCLUSIVE) &&
424 mtx->mtx_owner == curthread) {
425 KKASSERT((lock & MTX_MASK) != MTX_MASK);
427 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
434 ++mtx_collision_count;
440 _mtx_lock_sh_try(mtx_t mtx)
447 lock = mtx->mtx_lock;
448 if ((lock & MTX_EXCLUSIVE) == 0) {
449 KKASSERT((lock & MTX_MASK) != MTX_MASK);
451 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
458 ++mtx_collision_count;
464 * If the lock is held exclusively it must be owned by the caller. If the
465 * lock is already a shared lock this operation is a NOP. A panic will
466 * occur if the lock is not held either shared or exclusive.
468 * The exclusive count is converted to a shared count.
471 _mtx_downgrade(mtx_t mtx)
477 lock = mtx->mtx_lock;
478 if ((lock & MTX_EXCLUSIVE) == 0) {
479 KKASSERT((lock & MTX_MASK) > 0);
482 KKASSERT(mtx->mtx_owner == curthread);
483 nlock = lock & ~(MTX_EXCLUSIVE | MTX_SHWANTED);
484 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
485 if (lock & MTX_SHWANTED) {
492 ++mtx_collision_count;
497 * Upgrade a shared lock to an exclusive lock. The upgrade will fail if
498 * the shared lock has a count other then 1. Optimize the most likely case
499 * but note that a single cmpset can fail due to WANTED races.
501 * If the lock is held exclusively it must be owned by the caller and
502 * this function will simply return without doing anything. A panic will
503 * occur if the lock is held exclusively by someone other then the caller.
505 * Returns 0 on success, EDEADLK on failure.
508 _mtx_upgrade_try(mtx_t mtx)
515 lock = mtx->mtx_lock;
517 if ((lock & ~MTX_EXWANTED) == 1) {
518 nlock = lock | MTX_EXCLUSIVE;
519 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
520 mtx->mtx_owner = curthread;
523 } else if (lock & MTX_EXCLUSIVE) {
524 KKASSERT(mtx->mtx_owner == curthread);
531 ++mtx_collision_count;
537 * Unlock a lock. The caller must hold the lock either shared or exclusive.
539 * Any release which makes the lock available when others want an exclusive
540 * lock causes us to chain the owner to the next exclusive lock instead of
541 * releasing the lock.
544 _mtx_unlock(mtx_t mtx)
550 lock = mtx->mtx_lock;
551 nlock = lock & ~(MTX_SHWANTED | MTX_EXLINK);
555 * Last release, shared lock, no exclusive waiters.
557 nlock = lock & MTX_EXLINK;
558 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
560 } else if (nlock == (MTX_EXCLUSIVE | 1)) {
562 * Last release, exclusive lock, no exclusive waiters.
563 * Wake up any shared waiters.
565 mtx->mtx_owner = NULL;
566 nlock = lock & MTX_EXLINK;
567 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
568 if (lock & MTX_SHWANTED) {
574 } else if (nlock == (MTX_EXWANTED | 1)) {
576 * Last release, shared lock, with exclusive
579 * Wait for EXLINK to clear, then acquire it.
580 * We could use the cmpset for this but polling
581 * is better on the cpu caches.
583 * Acquire an exclusive lock leaving the lockcount
584 * set to 1, and get EXLINK for access to mtx_link.
588 if (lock & MTX_EXLINK) {
590 ++mtx_collision_count;
594 /*lock &= ~MTX_EXLINK;*/
595 nlock |= MTX_EXLINK | MTX_EXCLUSIVE;
596 nlock |= (lock & MTX_SHWANTED);
598 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
604 } else if (nlock == (MTX_EXCLUSIVE | MTX_EXWANTED | 1)) {
606 * Last release, exclusive lock, with exclusive
609 * leave the exclusive lock intact and the lockcount
610 * set to 1, and get EXLINK for access to mtx_link.
614 if (lock & MTX_EXLINK) {
616 ++mtx_collision_count;
620 /*lock &= ~MTX_EXLINK;*/
622 nlock |= (lock & MTX_SHWANTED);
624 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
632 * Not the last release (shared or exclusive)
635 KKASSERT((nlock & MTX_MASK) != MTX_MASK);
636 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
640 ++mtx_collision_count;
645 * Chain mtx_chain_link. Called with the lock held exclusively with a
646 * single ref count, and also with MTX_EXLINK held.
649 mtx_chain_link(mtx_t mtx)
654 u_int clock; /* bits we own and want to clear */
657 * Chain the exclusive lock to the next link. The caller cleared
658 * SHWANTED so if there is no link we have to wake up any shared
662 if ((link = mtx->mtx_link) != NULL) {
663 KKASSERT(link->state == MTX_LINK_LINKED);
664 if (link->next == link) {
665 mtx->mtx_link = NULL;
666 clock |= MTX_EXWANTED;
668 mtx->mtx_link = link->next;
669 link->next->prev = link->prev;
670 link->prev->next = link->next;
672 link->state = MTX_LINK_ACQUIRED;
673 mtx->mtx_owner = link->owner;
676 * Chain was empty, release the exclusive lock's last count
677 * as well the bits shown.
679 clock |= MTX_EXCLUSIVE | MTX_EXWANTED | MTX_SHWANTED | 1;
683 * We have to uset cmpset here to deal with MTX_SHWANTED. If
684 * we just clear the bits we can miss a wakeup or, worse,
685 * leave mtx_lock unlocked with MTX_SHWANTED still set.
688 lock = mtx->mtx_lock;
689 nlock = lock & ~clock;
691 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
694 * Wakeup new exclusive holder. Leave
698 } else if (lock & MTX_SHWANTED) {
700 * Signal any shared waiters (and we also
703 mtx->mtx_owner = NULL;
710 ++mtx_collision_count;
715 * Delete a link structure after tsleep has failed. This code is not
716 * in the critical path as most exclusive waits are chained.
720 mtx_delete_link(mtx_t mtx, mtx_link_t link)
722 thread_t td = curthread;
727 * Acquire MTX_EXLINK.
729 * Do not use cmpxchg to wait for EXLINK to clear as this might
730 * result in too much cpu cache traffic.
734 lock = mtx->mtx_lock;
735 if (lock & MTX_EXLINK) {
737 ++mtx_collision_count;
740 /* lock &= ~MTX_EXLINK; */
741 nlock = lock | MTX_EXLINK;
742 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
745 ++mtx_collision_count;
749 * Delete the link and release EXLINK.
751 if (link->state == MTX_LINK_LINKED) {
752 if (link->next == link) {
753 mtx->mtx_link = NULL;
755 mtx->mtx_link = link->next;
756 link->next->prev = link->prev;
757 link->prev->next = link->next;
759 link->state = MTX_LINK_IDLE;
761 atomic_clear_int(&mtx->mtx_lock, MTX_EXLINK);
766 * Abort a mutex locking operation, causing mtx_lock_ex_link() to
767 * return ENOLCK. This may be called at any time after the
768 * mtx_link is initialized, including both before and after the call
769 * to mtx_lock_ex_link().
772 mtx_abort_ex_link(mtx_t mtx, mtx_link_t link)
774 thread_t td = curthread;
783 lock = mtx->mtx_lock;
784 if (lock & MTX_EXLINK) {
786 ++mtx_collision_count;
789 /* lock &= ~MTX_EXLINK; */
790 nlock = lock | MTX_EXLINK;
791 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
794 ++mtx_collision_count;
800 switch(link->state) {
803 * Link not started yet
805 link->state = MTX_LINK_ABORTED;
807 case MTX_LINK_LINKED:
809 * de-link, mark aborted, and wakeup the thread.
811 if (link->next == link) {
812 mtx->mtx_link = NULL;
814 mtx->mtx_link = link->next;
815 link->next->prev = link->prev;
816 link->prev->next = link->next;
818 link->state = MTX_LINK_ABORTED;
821 case MTX_LINK_ACQUIRED:
823 * Too late, the lock was acquired. Let it complete.
828 * link already aborted, do nothing.
832 atomic_clear_int(&mtx->mtx_lock, MTX_EXLINK);