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 error = tsleep(link, flags, ident, to);
162 ++mtx_contention_count;
165 * Normal unlink, we should own the exclusive
168 if (link->state == MTX_LINK_LINKED)
169 mtx_delete_link(mtx, link);
170 if (link->state == MTX_LINK_ACQUIRED) {
171 KKASSERT(mtx->mtx_owner == link->owner);
177 * Aborted lock (mtx_abort_ex called).
179 if (link->state == MTX_LINK_ABORTED) {
185 * tsleep error, else retry.
193 ++mtx_collision_count;
199 _mtx_lock_ex_link(mtx_t mtx, mtx_link_t link,
200 const char *ident, int flags, int to)
202 return(__mtx_lock_ex(mtx, link, ident, flags, to));
206 _mtx_lock_ex(mtx_t mtx, const char *ident, int flags, int to)
208 struct mtx_link link;
210 mtx_link_init(&link);
211 return(__mtx_lock_ex(mtx, &link, ident, flags, to));
215 _mtx_lock_ex_quick(mtx_t mtx, const char *ident)
217 struct mtx_link link;
219 mtx_link_init(&link);
220 return(__mtx_lock_ex(mtx, &link, ident, 0, 0));
224 * Share-lock a mutex, block until acquired. Recursion is allowed.
226 * Returns 0 on success, or the tsleep() return code on failure.
227 * An error can only be returned if PCATCH is specified in the flags.
229 * NOTE: Shared locks get a mass-wakeup so if the tsleep fails we
230 * do not have to chain the wakeup().
233 __mtx_lock_sh(mtx_t mtx, const char *ident, int flags, int to)
240 lock = mtx->mtx_lock;
241 if ((lock & MTX_EXCLUSIVE) == 0) {
242 KKASSERT((lock & MTX_MASK) != MTX_MASK);
244 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
249 nlock = lock | MTX_SHWANTED;
250 tsleep_interlock(mtx, 0);
251 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
252 error = tsleep(mtx, flags, ident, to);
255 ++mtx_contention_count;
258 tsleep_remove(curthread);
261 ++mtx_collision_count;
267 _mtx_lock_sh(mtx_t mtx, const char *ident, int flags, int to)
269 return (__mtx_lock_sh(mtx, ident, flags, to));
273 _mtx_lock_sh_quick(mtx_t mtx, const char *ident)
275 return (__mtx_lock_sh(mtx, ident, 0, 0));
279 _mtx_spinlock_ex(mtx_t mtx)
287 lock = mtx->mtx_lock;
289 nlock = MTX_EXCLUSIVE | 1;
290 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
291 mtx->mtx_owner = curthread;
294 } else if ((lock & MTX_EXCLUSIVE) &&
295 mtx->mtx_owner == curthread) {
296 KKASSERT((lock & MTX_MASK) != MTX_MASK);
298 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
305 for (bo = 0; bo < bb; ++bo)
307 ++mtx_contention_count;
310 ++mtx_collision_count;
315 _mtx_spinlock_sh(mtx_t mtx)
323 lock = mtx->mtx_lock;
324 if ((lock & MTX_EXCLUSIVE) == 0) {
325 KKASSERT((lock & MTX_MASK) != MTX_MASK);
327 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
334 for (bo = 0; bo < bb; ++bo)
336 ++mtx_contention_count;
339 ++mtx_collision_count;
344 _mtx_lock_ex_try(mtx_t mtx)
351 lock = mtx->mtx_lock;
353 nlock = MTX_EXCLUSIVE | 1;
354 if (atomic_cmpset_int(&mtx->mtx_lock, 0, nlock)) {
355 mtx->mtx_owner = curthread;
358 } else if ((lock & MTX_EXCLUSIVE) &&
359 mtx->mtx_owner == curthread) {
360 KKASSERT((lock & MTX_MASK) != MTX_MASK);
362 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
369 ++mtx_collision_count;
375 _mtx_lock_sh_try(mtx_t mtx)
382 lock = mtx->mtx_lock;
383 if ((lock & MTX_EXCLUSIVE) == 0) {
384 KKASSERT((lock & MTX_MASK) != MTX_MASK);
386 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
393 ++mtx_collision_count;
399 * If the lock is held exclusively it must be owned by the caller. If the
400 * lock is already a shared lock this operation is a NOP. A panic will
401 * occur if the lock is not held either shared or exclusive.
403 * The exclusive count is converted to a shared count.
406 _mtx_downgrade(mtx_t mtx)
412 lock = mtx->mtx_lock;
413 if ((lock & MTX_EXCLUSIVE) == 0) {
414 KKASSERT((lock & MTX_MASK) > 0);
417 KKASSERT(mtx->mtx_owner == curthread);
418 nlock = lock & ~(MTX_EXCLUSIVE | MTX_SHWANTED);
419 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
420 if (lock & MTX_SHWANTED) {
427 ++mtx_collision_count;
432 * Upgrade a shared lock to an exclusive lock. The upgrade will fail if
433 * the shared lock has a count other then 1. Optimize the most likely case
434 * but note that a single cmpset can fail due to WANTED races.
436 * If the lock is held exclusively it must be owned by the caller and
437 * this function will simply return without doing anything. A panic will
438 * occur if the lock is held exclusively by someone other then the caller.
440 * Returns 0 on success, EDEADLK on failure.
443 _mtx_upgrade_try(mtx_t mtx)
450 lock = mtx->mtx_lock;
452 if ((lock & ~MTX_EXWANTED) == 1) {
453 nlock = lock | MTX_EXCLUSIVE;
454 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
455 mtx->mtx_owner = curthread;
458 } else if (lock & MTX_EXCLUSIVE) {
459 KKASSERT(mtx->mtx_owner == curthread);
466 ++mtx_collision_count;
472 * Unlock a lock. The caller must hold the lock either shared or exclusive.
474 * Any release which makes the lock available when others want an exclusive
475 * lock causes us to chain the owner to the next exclusive lock instead of
476 * releasing the lock.
479 _mtx_unlock(mtx_t mtx)
485 lock = mtx->mtx_lock;
486 nlock = lock & ~(MTX_SHWANTED | MTX_EXLINK);
490 * Last release, shared lock, no exclusive waiters.
492 nlock = lock & MTX_EXLINK;
493 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
495 } else if (nlock == (MTX_EXCLUSIVE | 1)) {
497 * Last release, exclusive lock, no exclusive waiters.
498 * Wake up any shared waiters.
500 mtx->mtx_owner = NULL;
501 nlock = lock & MTX_EXLINK;
502 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
503 if (lock & MTX_SHWANTED) {
509 } else if (nlock == (MTX_EXWANTED | 1)) {
511 * Last release, shared lock, with exclusive
514 * Wait for EXLINK to clear, then acquire it.
515 * We could use the cmpset for this but polling
516 * is better on the cpu caches.
518 * Acquire an exclusive lock leaving the lockcount
519 * set to 1, and get EXLINK for access to mtx_link.
523 if (lock & MTX_EXLINK) {
525 ++mtx_collision_count;
529 /*lock &= ~MTX_EXLINK;*/
530 nlock |= MTX_EXLINK | MTX_EXCLUSIVE;
531 nlock |= (lock & MTX_SHWANTED);
533 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
539 } else if (nlock == (MTX_EXCLUSIVE | MTX_EXWANTED | 1)) {
541 * Last release, exclusive lock, with exclusive
544 * leave the exclusive lock intact and the lockcount
545 * set to 1, and get EXLINK for access to mtx_link.
549 if (lock & MTX_EXLINK) {
551 ++mtx_collision_count;
555 /*lock &= ~MTX_EXLINK;*/
557 nlock |= (lock & MTX_SHWANTED);
559 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
567 * Not the last release (shared or exclusive)
570 KKASSERT((nlock & MTX_MASK) != MTX_MASK);
571 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
575 ++mtx_collision_count;
580 * Chain mtx_chain_link. Called with the lock held exclusively with a
581 * single ref count, and also with MTX_EXLINK held.
584 mtx_chain_link(mtx_t mtx)
589 u_int clock; /* bits we own and want to clear */
592 * Chain the exclusive lock to the next link. The caller cleared
593 * SHWANTED so if there is no link we have to wake up any shared
597 if ((link = mtx->mtx_link) != NULL) {
598 KKASSERT(link->state == MTX_LINK_LINKED);
599 if (link->next == link) {
600 mtx->mtx_link = NULL;
601 clock |= MTX_EXWANTED;
603 mtx->mtx_link = link->next;
604 link->next->prev = link->prev;
605 link->prev->next = link->next;
607 link->state = MTX_LINK_ACQUIRED;
608 mtx->mtx_owner = link->owner;
611 * Chain was empty, release the exclusive lock's last count
612 * as well the bits shown.
614 clock |= MTX_EXCLUSIVE | MTX_EXWANTED | MTX_SHWANTED | 1;
618 * We have to uset cmpset here to deal with MTX_SHWANTED. If
619 * we just clear the bits we can miss a wakeup or, worse,
620 * leave mtx_lock unlocked with MTX_SHWANTED still set.
623 lock = mtx->mtx_lock;
624 nlock = lock & ~clock;
626 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock)) {
629 * Wakeup new exclusive holder. Leave
633 } else if (lock & MTX_SHWANTED) {
635 * Signal any shared waiters (and we also
638 mtx->mtx_owner = NULL;
645 ++mtx_collision_count;
650 * Delete a link structure after tsleep has failed. This code is not
651 * in the critical path as most exclusive waits are chained.
655 mtx_delete_link(mtx_t mtx, mtx_link_t link)
657 thread_t td = curthread;
662 * Acquire MTX_EXLINK.
664 * Do not use cmpxchg to wait for EXLINK to clear as this might
665 * result in too much cpu cache traffic.
669 lock = mtx->mtx_lock;
670 if (lock & MTX_EXLINK) {
672 ++mtx_collision_count;
675 /* lock &= ~MTX_EXLINK; */
676 nlock = lock | MTX_EXLINK;
677 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
680 ++mtx_collision_count;
684 * Delete the link and release EXLINK.
686 if (link->state == MTX_LINK_LINKED) {
687 if (link->next == link) {
688 mtx->mtx_link = NULL;
690 mtx->mtx_link = link->next;
691 link->next->prev = link->prev;
692 link->prev->next = link->next;
694 link->state = MTX_LINK_IDLE;
696 atomic_clear_int(&mtx->mtx_lock, MTX_EXLINK);
701 * Abort a mutex locking operation, causing mtx_lock_ex_link() to
702 * return ENOLCK. This may be called at any time after the
703 * mtx_link is initialized, including both before and after the call
704 * to mtx_lock_ex_link().
707 mtx_abort_ex_link(mtx_t mtx, mtx_link_t link)
709 thread_t td = curthread;
718 lock = mtx->mtx_lock;
719 if (lock & MTX_EXLINK) {
721 ++mtx_collision_count;
724 /* lock &= ~MTX_EXLINK; */
725 nlock = lock | MTX_EXLINK;
726 if (atomic_cmpset_int(&mtx->mtx_lock, lock, nlock))
729 ++mtx_collision_count;
735 switch(link->state) {
738 * Link not started yet
740 link->state = MTX_LINK_ABORTED;
742 case MTX_LINK_LINKED:
744 * de-link, mark aborted, and wakeup the thread.
746 if (link->next == link) {
747 mtx->mtx_link = NULL;
749 mtx->mtx_link = link->next;
750 link->next->prev = link->prev;
751 link->prev->next = link->next;
753 link->state = MTX_LINK_ABORTED;
756 case MTX_LINK_ACQUIRED:
758 * Too late, the lock was acquired. Let it complete.
763 * link already aborted, do nothing.
767 atomic_clear_int(&mtx->mtx_lock, MTX_EXLINK);