Microoptimize locking primitives by avoiding unnecessary atomic ops.

Inline version of primitives do an atomic op and if it fails they fallback to
actual primitives, which immediately retry the atomic op.

The obvious optimisation is to check if the lock is free and only then proceed
to do an atomic op.

Reviewed by:	jhb, vangyzen

7 files changed, 34 insertions, 16 deletions

diff --git a/sys/kern/kern_lock.c b/sys/kern/kern_lock.c
index 34d7b34..1b10623 100644
--- a/sys/kern/kern_lock.c
+++ b/sys/kern/kern_lock.c
@@ -787,8 +787,10 @@ __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 			break;
 		}
 
-		while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
-		    tid)) {
+		for (;;) {
+			if (lk->lk_lock == LK_UNLOCKED &&
+			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
+				break;
 #ifdef HWPMC_HOOKS
 			PMC_SOFT_CALL( , , lock, failed);
 #endif
@@ -1124,7 +1126,11 @@ __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
 			    __func__, iwmesg, file, line);
 		}
 
-		while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
+		for (;;) {
+			if (lk->lk_lock == LK_UNLOCKED &&
+			    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
+				break;
+
 #ifdef HWPMC_HOOKS
 			PMC_SOFT_CALL( , , lock, failed);
 #endif
diff --git a/sys/kern/kern_mutex.c b/sys/kern/kern_mutex.c
index 3f62d17..d167205 100644
--- a/sys/kern/kern_mutex.c
+++ b/sys/kern/kern_mutex.c
@@ -419,7 +419,9 @@ __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
 	all_time -= lockstat_nsecs(&m->lock_object);
 #endif
 
-	while (!_mtx_obtain_lock(m, tid)) {
+	for (;;) {
+		if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
+			break;
 #ifdef KDTRACE_HOOKS
 		spin_cnt++;
 #endif
@@ -602,8 +604,9 @@ _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
 #ifdef KDTRACE_HOOKS
 	spin_time -= lockstat_nsecs(&m->lock_object);
 #endif
-	while (!_mtx_obtain_lock(m, tid)) {
-
+	for (;;) {
+		if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
+			break;
 		/* Give interrupts a chance while we spin. */
 		spinlock_exit();
 		while (m->mtx_lock != MTX_UNOWNED) {
@@ -675,7 +678,9 @@ retry:
 			    m->lock_object.lo_name, file, line));
 		WITNESS_CHECKORDER(&m->lock_object,
 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
-		while (!_mtx_obtain_lock(m, tid)) {
+		for (;;) {
+			if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
+				break;
 			if (m->mtx_lock == tid) {
 				m->mtx_recurse++;
 				break;
diff --git a/sys/kern/kern_rwlock.c b/sys/kern/kern_rwlock.c
index 6541724..6a904d2 100644
--- a/sys/kern/kern_rwlock.c
+++ b/sys/kern/kern_rwlock.c
@@ -771,7 +771,9 @@ __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 	all_time -= lockstat_nsecs(&rw->lock_object);
 	state = rw->rw_lock;
 #endif
-	while (!_rw_write_lock(rw, tid)) {
+	for (;;) {
+		if (rw->rw_lock == RW_UNLOCKED && _rw_write_lock(rw, tid))
+			break;
 #ifdef KDTRACE_HOOKS
 		spin_cnt++;
 #endif
diff --git a/sys/kern/kern_sx.c b/sys/kern/kern_sx.c
index 96e117b..2a81c04 100644
--- a/sys/kern/kern_sx.c
+++ b/sys/kern/kern_sx.c
@@ -544,7 +544,10 @@ _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 	all_time -= lockstat_nsecs(&sx->lock_object);
 	state = sx->sx_lock;
 #endif
-	while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
+	for (;;) {
+		if (sx->sx_lock == SX_LOCK_UNLOCKED &&
+		    atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
+			break;
 #ifdef KDTRACE_HOOKS
 		spin_cnt++;
 #endif
diff --git a/sys/sys/mutex.h b/sys/sys/mutex.h
index a9ec072..0443922 100644
--- a/sys/sys/mutex.h
+++ b/sys/sys/mutex.h
@@ -185,7 +185,7 @@ void	thread_lock_flags_(struct thread *, int, const char *, int);
 #define __mtx_lock(mp, tid, opts, file, line) do {			\
 	uintptr_t _tid = (uintptr_t)(tid);				\
 									\
-	if (!_mtx_obtain_lock((mp), _tid))				\
+	if (((mp)->mtx_lock != MTX_UNOWNED || !_mtx_obtain_lock((mp), _tid)))\
 		_mtx_lock_sleep((mp), _tid, (opts), (file), (line));	\
 	else								\
 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,	\
@@ -203,7 +203,7 @@ void	thread_lock_flags_(struct thread *, int, const char *, int);
 	uintptr_t _tid = (uintptr_t)(tid);				\
 									\
 	spinlock_enter();						\
-	if (!_mtx_obtain_lock((mp), _tid)) {				\
+	if (((mp)->mtx_lock != MTX_UNOWNED || !_mtx_obtain_lock((mp), _tid))) {\
 		if ((mp)->mtx_lock == _tid)				\
 			(mp)->mtx_recurse++;				\
 		else							\
@@ -232,7 +232,7 @@ void	thread_lock_flags_(struct thread *, int, const char *, int);
 									\
 	if ((mp)->mtx_recurse == 0)					\
 		LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, mp);	\
-	if (!_mtx_release_lock((mp), _tid))				\
+	if ((mp)->mtx_lock != _tid || !_mtx_release_lock((mp), _tid))	\
 		_mtx_unlock_sleep((mp), (opts), (file), (line));	\
 } while (0)
 
diff --git a/sys/sys/rwlock.h b/sys/sys/rwlock.h
index f8947c5..6b4f505 100644
--- a/sys/sys/rwlock.h
+++ b/sys/sys/rwlock.h
@@ -96,7 +96,7 @@
 #define	__rw_wlock(rw, tid, file, line) do {				\
 	uintptr_t _tid = (uintptr_t)(tid);				\
 									\
-	if (!_rw_write_lock((rw), _tid))				\
+	if ((rw)->rw_lock != RW_UNLOCKED || !_rw_write_lock((rw), _tid))\
 		_rw_wlock_hard((rw), _tid, (file), (line));		\
 	else 								\
 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,	\
@@ -112,7 +112,7 @@
 	else {								\
 		LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw,	\
 		    LOCKSTAT_WRITER);					\
-		if (!_rw_write_unlock((rw), _tid))			\
+		if ((rw)->rw_lock != _tid || !_rw_write_unlock((rw), _tid))\
 			_rw_wunlock_hard((rw), _tid, (file), (line));	\
 	}								\
 } while (0)
diff --git a/sys/sys/sx.h b/sys/sys/sx.h
index 03b51d3..57a31d9 100644
--- a/sys/sys/sx.h
+++ b/sys/sys/sx.h
@@ -150,7 +150,8 @@ __sx_xlock(struct sx *sx, struct thread *td, int opts, const char *file,
 	uintptr_t tid = (uintptr_t)td;
 	int error = 0;
 
-	if (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
+	if (sx->sx_lock != SX_LOCK_UNLOCKED ||
+	    !atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
 		error = _sx_xlock_hard(sx, tid, opts, file, line);
 	else 
 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
@@ -168,7 +169,8 @@ __sx_xunlock(struct sx *sx, struct thread *td, const char *file, int line)
 	if (sx->sx_recurse == 0)
 		LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx,
 		    LOCKSTAT_WRITER);
-	if (!atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
+	if (sx->sx_lock != tid ||
+	    !atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 		_sx_xunlock_hard(sx, tid, file, line);
 }