path: root/percona/5.0.91-b22-20100522/innodb_rw_lock.patch

diff -ruN a/innobase/btr/btr0cur.c b/innobase/btr/btr0cur.c
--- a/innobase/btr/btr0cur.c	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/btr/btr0cur.c	2009-10-22 15:18:44.000000000 +0900
@@ -313,7 +313,7 @@
 #ifdef UNIV_SEARCH_PERF_STAT
 	info->n_searches++;
 #endif	
-	if (btr_search_latch.writer == RW_LOCK_NOT_LOCKED
+	if (rw_lock_get_writer(&btr_search_latch) == RW_LOCK_NOT_LOCKED
 		&& latch_mode <= BTR_MODIFY_LEAF && info->last_hash_succ
 		&& !estimate
 #ifdef PAGE_CUR_LE_OR_EXTENDS
diff -ruN a/innobase/btr/btr0sea.c b/innobase/btr/btr0sea.c
--- a/innobase/btr/btr0sea.c	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/btr/btr0sea.c	2009-10-22 15:18:44.000000000 +0900
@@ -773,8 +773,8 @@
 		rw_lock_s_lock(&btr_search_latch);
 	}
 
-	ut_ad(btr_search_latch.writer != RW_LOCK_EX);
-	ut_ad(btr_search_latch.reader_count > 0);
+	ut_ad(rw_lock_get_writer(&btr_search_latch) != RW_LOCK_EX);
+	ut_ad(rw_lock_get_reader_count(&btr_search_latch) > 0);
 
 	rec = ha_search_and_get_data(btr_search_sys->hash_index, fold);
 
diff -ruN a/innobase/buf/buf0buf.c b/innobase/buf/buf0buf.c
--- a/innobase/buf/buf0buf.c	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/buf/buf0buf.c	2009-10-22 15:18:44.000000000 +0900
@@ -1292,7 +1292,7 @@
 
 	if (mode == BUF_GET_NOWAIT) {
 		if (rw_latch == RW_S_LATCH) {
-			success = rw_lock_s_lock_func_nowait(&(block->lock),
+			success = rw_lock_s_lock_nowait(&(block->lock),
 								file, line);
 			fix_type = MTR_MEMO_PAGE_S_FIX;
 		} else {
@@ -1442,7 +1442,7 @@
 	ut_ad(!ibuf_inside() || ibuf_page(block->space, block->offset));
 
 	if (rw_latch == RW_S_LATCH) {
-		success = rw_lock_s_lock_func_nowait(&(block->lock),
+		success = rw_lock_s_lock_nowait(&(block->lock),
 								file, line);
 		fix_type = MTR_MEMO_PAGE_S_FIX;
 	} else {
@@ -1596,7 +1596,7 @@
 	ut_ad(!ibuf_inside() || (mode == BUF_KEEP_OLD));
 
 	if (rw_latch == RW_S_LATCH) {
-		success = rw_lock_s_lock_func_nowait(&(block->lock),
+		success = rw_lock_s_lock_nowait(&(block->lock),
 								file, line);
 		fix_type = MTR_MEMO_PAGE_S_FIX;
 	} else {
diff -ruN a/innobase/include/buf0buf.ic b/innobase/include/buf0buf.ic
--- a/innobase/include/buf0buf.ic	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/include/buf0buf.ic	2009-10-22 16:12:25.000000000 +0900
@@ -523,7 +523,7 @@
 #ifdef UNIV_SYNC_DEBUG	
 	ibool	ret;
 
-	ret = rw_lock_s_lock_func_nowait(&(block->debug_latch), file, line);
+	ret = rw_lock_s_lock_nowait(&(block->debug_latch), file, line);
 
 	ut_ad(ret == TRUE);
 	ut_ad(mutex_own(&block->mutex));
diff -ruN a/innobase/include/os0sync.h b/innobase/include/os0sync.h
--- a/innobase/include/os0sync.h	2009-09-10 04:02:59.000000000 +0900
+++ b/innobase/include/os0sync.h	2009-10-22 15:18:44.000000000 +0900
@@ -1,11 +1,35 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 The interface to the operating system
 synchronization primitives.
 
-(c) 1995 Innobase Oy
-
 Created 9/6/1995 Heikki Tuuri
 *******************************************************/
+
 #ifndef os0sync_h
 #define os0sync_h
 
@@ -261,6 +285,23 @@
 /*===============*/
 	os_fast_mutex_t*	fast_mutex);	/* in: mutex to free */
 	
+#ifdef HAVE_ATOMIC_BUILTINS
+/**************************************************************
+Atomic compare-and-swap for InnoDB. Currently requires GCC atomic builtins.
+Returns true if swapped, ptr is pointer to target, old_val is value to
+compare to, new_val is the value to swap in. */
+#define os_compare_and_swap(ptr, old_val, new_val) \
+	__sync_bool_compare_and_swap(ptr, old_val, new_val)
+
+/**************************************************************
+Atomic increment for InnoDB. Currently requires GCC atomic builtins.
+Returns the resulting value, ptr is pointer to target, amount is the
+amount of increment. */
+#define os_atomic_increment(ptr, amount) \
+	__sync_add_and_fetch(ptr, amount)
+
+#endif /* HAVE_ATOMIC_BUILTINS */
+
 #ifndef UNIV_NONINL
 #include "os0sync.ic"
 #endif
diff -ruN a/innobase/include/sync0rw.h b/innobase/include/sync0rw.h
--- a/innobase/include/sync0rw.h	2009-09-10 04:02:59.000000000 +0900
+++ b/innobase/include/sync0rw.h	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 The read-write lock (for threads, not for database transactions)
 
-(c) 1995 Innobase Oy
-
 Created 9/11/1995 Heikki Tuuri
 *******************************************************/
 
@@ -24,6 +47,12 @@
 #define	RW_X_LATCH	2
 #define	RW_NO_LATCH	3
 
+/* We decrement lock_word by this amount for each x_lock. It is also the
+start value for the lock_word, meaning that it limits the maximum number
+of concurrent read locks before the rw_lock breaks. The current value of
+0x00100000 allows 1,048,575 concurrent readers and 2047 recursive writers.*/
+#define X_LOCK_DECR		0x00100000
+
 typedef struct rw_lock_struct		rw_lock_t;
 #ifdef UNIV_SYNC_DEBUG
 typedef struct rw_lock_debug_struct	rw_lock_debug_t;
@@ -47,14 +76,14 @@
 					there may be waiters for the event */
 #endif /* UNIV_SYNC_DEBUG */
 
-extern	ulint	rw_s_system_call_count;
-extern	ulint	rw_s_spin_wait_count;
-extern	ulint	rw_s_exit_count;
-extern	ulint	rw_s_os_wait_count;
-extern	ulint	rw_x_system_call_count;
-extern	ulint	rw_x_spin_wait_count;
-extern	ulint	rw_x_os_wait_count;
-extern	ulint	rw_x_exit_count;
+extern	ib_longlong	rw_s_spin_wait_count;
+extern	ib_longlong	rw_s_spin_round_count;
+extern	ib_longlong	rw_s_exit_count;
+extern	ib_longlong	rw_s_os_wait_count;
+extern	ib_longlong	rw_x_spin_wait_count;
+extern	ib_longlong	rw_x_spin_round_count;
+extern	ib_longlong	rw_x_os_wait_count;
+extern	ib_longlong	rw_x_exit_count;
 
 /**********************************************************************
 Creates, or rather, initializes an rw-lock object in a specified memory
@@ -116,8 +145,22 @@
 NOTE! The following macros should be used in rw s-locking, not the
 corresponding function. */
 
-#define rw_lock_s_lock_nowait(M)    rw_lock_s_lock_func_nowait(\
-					     (M), __FILE__, __LINE__)
+#define rw_lock_s_lock_nowait(M, F, L)    rw_lock_s_lock_low(\
+					  (M), 0, (F), (L))
+/**********************************************************************
+Low-level function which tries to lock an rw-lock in s-mode. Performs no
+spinning. */
+UNIV_INLINE
+ibool
+rw_lock_s_lock_low(
+/*===============*/
+				/* out: TRUE if success */
+	rw_lock_t*	lock,	/* in: pointer to rw-lock */
+	ulint		pass __attribute__((unused)),
+				/* in: pass value; != 0, if the lock will be
+				passed to another thread to unlock */
+	const char*	file_name, /* in: file name where lock requested */
+	ulint		line);	/* in: line where requested */
 /**********************************************************************
 NOTE! Use the corresponding macro, not directly this function, except if
 you supply the file name and line number. Lock an rw-lock in shared mode
@@ -135,18 +178,6 @@
 	const char*	file_name,/* in: file name where lock requested */
 	ulint		line);	/* in: line where requested */
 /**********************************************************************
-NOTE! Use the corresponding macro, not directly this function, except if
-you supply the file name and line number. Lock an rw-lock in shared mode
-for the current thread if the lock can be acquired immediately. */
-UNIV_INLINE
-ibool
-rw_lock_s_lock_func_nowait(
-/*=======================*/
-				/* out: TRUE if success */
-        rw_lock_t*   	lock,  	/* in: pointer to rw-lock */
-	const char*	file_name,/* in: file name where lock requested */
-	ulint		line);	/* in: line where requested */
-/**********************************************************************
 NOTE! Use the corresponding macro, not directly this function! Lock an
 rw-lock in exclusive mode for the current thread if the lock can be
 obtained immediately. */
@@ -338,6 +369,41 @@
 rw_lock_get_reader_count(
 /*=====================*/
 	rw_lock_t*	lock);
+/**********************************************************************
+Decrements lock_word the specified amount if it is greater than 0.
+This is used by both s_lock and x_lock operations. */
+UNIV_INLINE
+ibool
+rw_lock_lock_word_decr(
+/*===================*/
+					/* out: TRUE if decr occurs */
+	rw_lock_t*	lock,		/* in: rw-lock */
+	ulint		amount);	/* in: amount to decrement */
+/**********************************************************************
+Increments lock_word the specified amount and returns new value. */
+UNIV_INLINE
+lint
+rw_lock_lock_word_incr(
+/*===================*/
+					/* out: TRUE if decr occurs */
+	rw_lock_t*	lock,
+	ulint		amount);	/* in: rw-lock */
+/**********************************************************************
+This function sets the lock->writer_thread and lock->recursive fields.
+For platforms where we are using atomic builtins instead of lock->mutex
+it sets the lock->writer_thread field using atomics to ensure memory
+ordering. Note that it is assumed that the caller of this function
+effectively owns the lock i.e.: nobody else is allowed to modify
+lock->writer_thread at this point in time.
+The protocol is that lock->writer_thread MUST be updated BEFORE the
+lock->recursive flag is set. */
+UNIV_INLINE
+void
+rw_lock_set_writer_id_and_recursion_flag(
+/*=====================================*/
+	rw_lock_t*	lock,		/* in/out: lock to work on */
+	ibool		recursive);	/* in: TRUE if recursion
+					allowed */
 #ifdef UNIV_SYNC_DEBUG
 /**********************************************************************
 Checks if the thread has locked the rw-lock in the specified mode, with
@@ -417,47 +483,33 @@
 field. Then no new readers are allowed in. */
 
 struct rw_lock_struct {
+	volatile lint	lock_word;
+				/* Holds the state of the lock. */
+	volatile ulint	waiters;/* 1: there are waiters */
+	volatile ibool	recursive;/* Default value FALSE which means the lock
+				is non-recursive. The value is typically set
+				to TRUE making normal rw_locks recursive. In
+				case of asynchronous IO, when a non-zero
+				value of 'pass' is passed then we keep the
+				lock non-recursive.
+				This flag also tells us about the state of
+				writer_thread field. If this flag is set
+				then writer_thread MUST contain the thread
+				id of the current x-holder or wait-x thread.
+				This flag must be reset in x_unlock
+				functions before incrementing the lock_word */
+	volatile os_thread_id_t	writer_thread;
+				/* Thread id of writer thread. Is only
+				guaranteed to have sane and non-stale
+				value iff recursive flag is set. */
 	os_event_t	event;	/* Used by sync0arr.c for thread queueing */
-
-#ifdef __WIN__
-	os_event_t	wait_ex_event;	/* This windows specific event is
-				used by the thread which has set the
-				lock state to RW_LOCK_WAIT_EX. The
-				rw_lock design guarantees that this
-				thread will be the next one to proceed
-				once the current the event gets
-				signalled. See LEMMA 2 in sync0sync.c */
-#endif
-
-	ulint	reader_count;	/* Number of readers who have locked this
-				lock in the shared mode */
-	ulint	writer; 	/* This field is set to RW_LOCK_EX if there
-				is a writer owning the lock (in exclusive
-				mode), RW_LOCK_WAIT_EX if a writer is
-				queueing for the lock, and
-				RW_LOCK_NOT_LOCKED, otherwise. */
-	os_thread_id_t	writer_thread;
-				/* Thread id of a possible writer thread */
-	ulint	writer_count;	/* Number of times the same thread has
-				recursively locked the lock in the exclusive
-				mode */
+	os_event_t	wait_ex_event;
+				/* Event for next-writer to wait on. A thread
+				must decrement lock_word before waiting. */
+#ifndef HAVE_ATOMIC_BUILTINS
 	mutex_t	mutex;		/* The mutex protecting rw_lock_struct */
-	ulint	pass; 		/* Default value 0. This is set to some
-				value != 0 given by the caller of an x-lock
-				operation, if the x-lock is to be passed to
-				another thread to unlock (which happens in
-				asynchronous i/o). */
-	ulint	waiters;	/* This ulint is set to 1 if there are
-				waiters (readers or writers) in the global
-				wait array, waiting for this rw_lock.
-				Otherwise, == 0. */
-	ibool	writer_is_wait_ex;
-				/* This is TRUE if the writer field is
-				RW_LOCK_WAIT_EX; this field is located far
-				from the memory update hotspot fields which
-				are at the start of this struct, thus we can
-				peek this field without causing much memory
-				bus traffic */
+#endif /* HAVE_ATOMIC_BUILTINS */
+
 	UT_LIST_NODE_T(rw_lock_t) list;
 				/* All allocated rw locks are put into a
 				list */
@@ -465,15 +517,23 @@
 	UT_LIST_BASE_NODE_T(rw_lock_debug_t) debug_list;
 				/* In the debug version: pointer to the debug
 				info list of the lock */
+	ulint	level;		/* Level in the global latching order. */
 #endif /* UNIV_SYNC_DEBUG */
-	ulint	level;		/* Level in the global latching
-				order; default SYNC_LEVEL_NONE */
+	ulint count_os_wait;	/* Count of os_waits. May not be accurate */
 	const char*	cfile_name;/* File name where lock created */
-	ulint	cline;		/* Line where created */
+        /* last s-lock file/line is not guaranteed to be correct */
 	const char*	last_s_file_name;/* File name where last s-locked */
 	const char*	last_x_file_name;/* File name where last x-locked */
-	ulint	last_s_line;	/* Line number where last time s-locked */
-	ulint	last_x_line;	/* Line number where last time x-locked */
+	ibool		writer_is_wait_ex;
+				/* This is TRUE if the writer field is
+				RW_LOCK_WAIT_EX; this field is located far
+				from the memory update hotspot fields which
+				are at the start of this struct, thus we can
+				peek this field without causing much memory
+				bus traffic */
+	unsigned	cline:14;	/* Line where created */
+	unsigned	last_s_line:14;	/* Line number where last time s-locked */
+	unsigned	last_x_line:14;	/* Line number where last time x-locked */
 	ulint	magic_n;
 };
 
diff -ruN a/innobase/include/sync0rw.ic b/innobase/include/sync0rw.ic
--- a/innobase/include/sync0rw.ic	2009-09-10 04:02:59.000000000 +0900
+++ b/innobase/include/sync0rw.ic	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 The read-write lock (for threads)
 
-(c) 1995 Innobase Oy
-
 Created 9/11/1995 Heikki Tuuri
 *******************************************************/
 
@@ -49,53 +72,88 @@
 ulint
 rw_lock_get_waiters(
 /*================*/
-	rw_lock_t*	lock)
+				/* out: 1 if waiters, 0 otherwise */
+	rw_lock_t*	lock)	/* in: rw-lock */
 {
 	return(lock->waiters);
 }
+
+/************************************************************************
+Sets lock->waiters to 1. It is not an error if lock->waiters is already
+1. On platforms where ATOMIC builtins are used this function enforces a
+memory barrier. */
 UNIV_INLINE
 void
-rw_lock_set_waiters(
-/*================*/
-	rw_lock_t*	lock,
-	ulint		flag)
+rw_lock_set_waiter_flag(
+/*====================*/
+	rw_lock_t*	lock)	/* in: rw-lock */
 {
-	lock->waiters = flag;
+#ifdef HAVE_ATOMIC_BUILTINS
+	os_compare_and_swap(&lock->waiters, 0, 1);
+#else /* HAVE_ATOMIC_BUILTINS */
+	lock->waiters = 1;
+#endif /* HAVE_ATOMIC_BUILTINS */
 }
+
+/************************************************************************
+Resets lock->waiters to 0. It is not an error if lock->waiters is already
+0. On platforms where ATOMIC builtins are used this function enforces a
+memory barrier. */
 UNIV_INLINE
-ulint
-rw_lock_get_writer(
-/*===============*/
-	rw_lock_t*	lock)
+void
+rw_lock_reset_waiter_flag(
+/*======================*/
+	rw_lock_t*	lock)	/* in: rw-lock */
 {
-	return(lock->writer);
+#ifdef HAVE_ATOMIC_BUILTINS
+	os_compare_and_swap(&lock->waiters, 1, 0);
+#else /* HAVE_ATOMIC_BUILTINS */
+	lock->waiters = 0;
+#endif /* HAVE_ATOMIC_BUILTINS */
 }
+
+/**********************************************************************
+Returns the write-status of the lock - this function made more sense
+with the old rw_lock implementation. */
 UNIV_INLINE
-void
-rw_lock_set_writer(
+ulint
+rw_lock_get_writer(
 /*===============*/
-	rw_lock_t*	lock,
-	ulint		flag)
+	rw_lock_t*	lock)
 {
-	lock->writer = flag;
+	lint lock_word = lock->lock_word;
+	if(lock_word > 0) {
+		/* return NOT_LOCKED in s-lock state, like the writer
+		member of the old lock implementation. */
+		return(RW_LOCK_NOT_LOCKED);
+	} else if (((-lock_word) % X_LOCK_DECR) == 0) {
+		return(RW_LOCK_EX);
+	} else {
+                ut_ad(lock_word > -X_LOCK_DECR);
+		return(RW_LOCK_WAIT_EX);
+	}
 }
+
+/**********************************************************************
+Returns number of readers. */
 UNIV_INLINE
 ulint
 rw_lock_get_reader_count(
 /*=====================*/
 	rw_lock_t*	lock)
 {
-	return(lock->reader_count);
-}
-UNIV_INLINE
-void
-rw_lock_set_reader_count(
-/*=====================*/
-	rw_lock_t*	lock,
-	ulint		count)
-{
-	lock->reader_count = count;
+	lint lock_word = lock->lock_word;
+	if(lock_word > 0) {
+		/* s-locked, no x-waiters */
+		return(X_LOCK_DECR - lock_word);
+	} else if (lock_word < 0 && lock_word > -X_LOCK_DECR) {
+		/* s-locked, with x-waiters */
+		return((ulint)(-lock_word));
+	}
+	return(0);
 }
+
+#ifndef HAVE_ATOMIC_BUILTINS
 UNIV_INLINE
 mutex_t*
 rw_lock_get_mutex(
@@ -104,6 +162,7 @@
 {
 	return(&(lock->mutex));
 }
+#endif
 
 /**********************************************************************
 Returns the value of writer_count for the lock. Does not reserve the lock
@@ -115,7 +174,126 @@
 				/* out: value of writer_count */
 	rw_lock_t*	lock)	/* in: rw-lock */
 {
-	return(lock->writer_count);
+	lint lock_copy = lock->lock_word;
+	/* If there is a reader, lock_word is not divisible by X_LOCK_DECR */
+	if(lock_copy > 0 || (-lock_copy) % X_LOCK_DECR != 0) {
+		return(0);
+	}
+	return(((-lock_copy) / X_LOCK_DECR) + 1);
+}
+
+/**********************************************************************
+Two different implementations for decrementing the lock_word of a rw_lock:
+one for systems supporting atomic operations, one for others. This does
+does not support recusive x-locks: they should be handled by the caller and
+need not be atomic since they are performed by the current lock holder.
+Returns true if the decrement was made, false if not. */
+UNIV_INLINE
+ibool
+rw_lock_lock_word_decr(
+/*===================*/
+				/* out: TRUE if decr occurs */
+	rw_lock_t*	lock,	/* in: rw-lock */
+	ulint		amount)	/* in: amount of decrement */
+{
+
+#ifdef HAVE_ATOMIC_BUILTINS
+
+        lint local_lock_word = lock->lock_word;
+	while (local_lock_word > 0) {
+		if(os_compare_and_swap(&(lock->lock_word),
+                                       local_lock_word,
+                                       local_lock_word - amount)) {
+			return(TRUE);
+		}
+		local_lock_word = lock->lock_word;
+	}
+	return(FALSE);
+
+#else /* HAVE_ATOMIC_BUILTINS */
+
+	ibool success = FALSE;
+	mutex_enter(&(lock->mutex));
+	if(lock->lock_word > 0) {
+		lock->lock_word -= amount;
+		success = TRUE;
+	}
+	mutex_exit(&(lock->mutex));
+	return(success);
+
+#endif /* HAVE_ATOMIC_BUILTINS */
+}
+
+/**********************************************************************
+Two different implementations for incrementing the lock_word of a rw_lock:
+one for systems supporting atomic operations, one for others.
+Returns the value of lock_word after increment. */
+UNIV_INLINE
+lint
+rw_lock_lock_word_incr(
+/*===================*/
+				/* out: lock->lock_word after increment */
+	rw_lock_t*	lock,	/* in: rw-lock */
+	ulint		amount)	/* in: amount of increment */
+{
+
+#ifdef HAVE_ATOMIC_BUILTINS
+
+	return(os_atomic_increment(&(lock->lock_word), amount));
+
+#else /* HAVE_ATOMIC_BUILTINS */
+
+	lint local_lock_word;
+
+	mutex_enter(&(lock->mutex));
+
+	lock->lock_word += amount;
+	local_lock_word = lock->lock_word;
+
+	mutex_exit(&(lock->mutex));
+
+        return(local_lock_word);
+
+#endif /* HAVE_ATOMIC_BUILTINS */
+}
+
+/**********************************************************************
+This function sets the lock->writer_thread and lock->recursive fields.
+For platforms where we are using atomic builtins instead of lock->mutex
+it sets the lock->writer_thread field using atomics to ensure memory
+ordering. Note that it is assumed that the caller of this function
+effectively owns the lock i.e.: nobody else is allowed to modify
+lock->writer_thread at this point in time.
+The protocol is that lock->writer_thread MUST be updated BEFORE the
+lock->recursive flag is set. */
+UNIV_INLINE
+void
+rw_lock_set_writer_id_and_recursion_flag(
+/*=====================================*/
+	rw_lock_t*	lock,		/* in/out: lock to work on */
+	ibool		recursive)	/* in: TRUE if recursion
+					allowed */
+{
+	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
+
+#ifdef HAVE_ATOMIC_BUILTINS
+	os_thread_id_t	local_thread;
+	ibool		success;
+
+	local_thread = lock->writer_thread;
+	success = os_compare_and_swap(&lock->writer_thread,
+				      local_thread, curr_thread);
+	ut_a(success);
+	lock->recursive = recursive;
+
+#else /* HAVE_ATOMIC_BUILTINS */
+
+	mutex_enter(&lock->mutex);
+	lock->writer_thread = curr_thread;
+	lock->recursive = recursive;
+	mutex_exit(&lock->mutex);
+
+#endif /* HAVE_ATOMIC_BUILTINS */
 }
 
 /**********************************************************************
@@ -133,26 +311,21 @@
 	const char*	file_name, /* in: file name where lock requested */
 	ulint		line)	/* in: line where requested */
 {
-#ifdef UNIV_SYNC_DEBUG
-	ut_ad(mutex_own(rw_lock_get_mutex(lock)));
-#endif /* UNIV_SYNC_DEBUG */
-	/* Check if the writer field is free */
-
-	if (UNIV_LIKELY(lock->writer == RW_LOCK_NOT_LOCKED)) {
-		/* Set the shared lock by incrementing the reader count */
-		lock->reader_count++;
+	/* TODO: study performance of UNIV_LIKELY branch prediction hints. */
+	if (!rw_lock_lock_word_decr(lock, 1)) {
+		/* Locking did not succeed */
+		return(FALSE);
+	}
 
 #ifdef UNIV_SYNC_DEBUG
-		rw_lock_add_debug_info(lock, pass, RW_LOCK_SHARED, file_name,
-									line);
+	rw_lock_add_debug_info(lock, pass, RW_LOCK_SHARED, file_name, line);
 #endif
-		lock->last_s_file_name = file_name;
-		lock->last_s_line = line;
-
-		return(TRUE);	/* locking succeeded */
-	}
+	/* These debugging values are not set safely: they may be incorrect
+        or even refer to a line that is invalid for the file name. */
+	lock->last_s_file_name = file_name;
+	lock->last_s_line = line;
 
-	return(FALSE);	/* locking did not succeed */
+	return(TRUE);	/* locking succeeded */
 }
 
 /**********************************************************************
@@ -167,11 +340,10 @@
 	const char*	file_name,	/* in: file name where requested */
 	ulint		line)		/* in: line where lock requested */
 {
-	ut_ad(lock->writer == RW_LOCK_NOT_LOCKED);
-	ut_ad(rw_lock_get_reader_count(lock) == 0);
+	ut_ad(lock->lock_word == X_LOCK_DECR);
 	
-	/* Set the shared lock by incrementing the reader count */
-	lock->reader_count++;
+	/* Indicate there is a new reader by decrementing lock_word */
+	lock->lock_word--;
 
 	lock->last_s_file_name = file_name;
 	lock->last_s_line = line;
@@ -194,13 +366,11 @@
 	ulint		line)		/* in: line where lock requested */
 {
         ut_ad(rw_lock_validate(lock));
-	ut_ad(rw_lock_get_reader_count(lock) == 0);
-	ut_ad(rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED);
+	ut_ad(lock->lock_word == X_LOCK_DECR);
 
-	rw_lock_set_writer(lock, RW_LOCK_EX);
+	lock->lock_word -= X_LOCK_DECR;
 	lock->writer_thread = os_thread_get_curr_id();
-	lock->writer_count++;
-	lock->pass = 0;
+	lock->recursive = TRUE;
 			
 	lock->last_x_file_name = file_name;
 	lock->last_x_line = line;
@@ -241,15 +411,12 @@
 	ut_ad(!rw_lock_own(lock, RW_LOCK_SHARED)); /* see NOTE above */
 #endif /* UNIV_SYNC_DEBUG */
 
-	mutex_enter(rw_lock_get_mutex(lock));
-
-	if (UNIV_LIKELY(rw_lock_s_lock_low(lock, pass, file_name, line))) {
-		mutex_exit(rw_lock_get_mutex(lock));
+	/* TODO: study performance of UNIV_LIKELY branch prediction hints. */
+	if (rw_lock_s_lock_low(lock, pass, file_name, line)) {
 
 		return; /* Success */
 	} else {
 		/* Did not succeed, try spin wait */
-		mutex_exit(rw_lock_get_mutex(lock));
 
 		rw_lock_s_lock_spin(lock, pass, file_name, line);
 
@@ -259,86 +426,60 @@
 
 /**********************************************************************
 NOTE! Use the corresponding macro, not directly this function! Lock an
-rw-lock in shared mode for the current thread if the lock can be acquired
-immediately. */
+rw-lock in exclusive mode for the current thread if the lock can be
+obtained immediately. */
 UNIV_INLINE
 ibool
-rw_lock_s_lock_func_nowait(
+rw_lock_x_lock_func_nowait(
 /*=======================*/
 				/* out: TRUE if success */
         rw_lock_t*   	lock,  	/* in: pointer to rw-lock */
 	const char*	file_name,/* in: file name where lock requested */
 	ulint		line)	/* in: line where requested */
 {
-	ibool	success	= FALSE;
-
-	mutex_enter(rw_lock_get_mutex(lock));
-
-	if (lock->writer == RW_LOCK_NOT_LOCKED) {
-		/* Set the shared lock by incrementing the reader count */
-		lock->reader_count++;
+	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
 
-#ifdef UNIV_SYNC_DEBUG
-		rw_lock_add_debug_info(lock, 0, RW_LOCK_SHARED, file_name,
-									line);
-#endif
+	ibool success;
 
-		lock->last_s_file_name = file_name;
-		lock->last_s_line = line;
+#ifdef HAVE_ATOMIC_BUILTINS
+	success = os_compare_and_swap(&(lock->lock_word), X_LOCK_DECR, 0);
+#else
 		
+	success = FALSE;
+	mutex_enter(&(lock->mutex));
+	if (lock->lock_word == X_LOCK_DECR) {
+		lock->lock_word = 0;
 		success = TRUE;
 	}
+	mutex_exit(&(lock->mutex));
 
-	mutex_exit(rw_lock_get_mutex(lock));
-
-	return(success);
-}
+#endif
+	if (success) {
+		rw_lock_set_writer_id_and_recursion_flag(lock, TRUE);
 
-/**********************************************************************
-NOTE! Use the corresponding macro, not directly this function! Lock an
-rw-lock in exclusive mode for the current thread if the lock can be
-obtained immediately. */
-UNIV_INLINE
-ibool
-rw_lock_x_lock_func_nowait(
-/*=======================*/
-				/* out: TRUE if success */
-        rw_lock_t*   	lock,  	/* in: pointer to rw-lock */
-	const char*	file_name,/* in: file name where lock requested */
-	ulint		line)	/* in: line where requested */
-{
-	ibool		success		= FALSE;
-	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
-	mutex_enter(rw_lock_get_mutex(lock));
+	} else if (lock->recursive
+		   && os_thread_eq(lock->writer_thread, curr_thread)) {
+		/* Relock: this lock_word modification is safe since no other
+		threads can modify (lock, unlock, or reserve) lock_word while
+		there is an exclusive writer and this is the writer thread. */
+		lock->lock_word -= X_LOCK_DECR;
 
-	if (UNIV_UNLIKELY(rw_lock_get_reader_count(lock) != 0)) {
-	} else if (UNIV_LIKELY(rw_lock_get_writer(lock)
-			       == RW_LOCK_NOT_LOCKED)) {
-		rw_lock_set_writer(lock, RW_LOCK_EX);
-		lock->writer_thread = curr_thread;
-		lock->pass = 0;
-	relock:
-		lock->writer_count++;
+		ut_ad(((-lock->lock_word) % X_LOCK_DECR) == 0);
 			
+	} else {
+		/* Failure */
+		return(FALSE);
+	}
 #ifdef UNIV_SYNC_DEBUG
-		rw_lock_add_debug_info(lock, 0, RW_LOCK_EX, file_name, line);
+	rw_lock_add_debug_info(lock, 0, RW_LOCK_EX, file_name, line);
 #endif
 
-		lock->last_x_file_name = file_name;
-		lock->last_x_line = line;
-
-		success = TRUE;
-	} else if (rw_lock_get_writer(lock) == RW_LOCK_EX
-			&& lock->pass == 0
-			&& os_thread_eq(lock->writer_thread, curr_thread)) {
-		goto relock;
-	}
-
-	mutex_exit(rw_lock_get_mutex(lock));
+	lock->last_x_file_name = file_name;
+	lock->last_x_line = line;
 
         ut_ad(rw_lock_validate(lock));
 
-	return(success);
+	return(TRUE);
 }
 
 /**********************************************************************
@@ -354,39 +495,21 @@
 #endif
 	)
 {
-	mutex_t*	mutex	= &(lock->mutex);
-	ibool		sg 	= FALSE;
-
-        /* Acquire the mutex protecting the rw-lock fields */
-	mutex_enter(mutex);
-
-	/* Reset the shared lock by decrementing the reader count */
-
-	ut_a(lock->reader_count > 0);
-	lock->reader_count--;
+	ut_ad((lock->lock_word % X_LOCK_DECR) != 0);
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, pass, RW_LOCK_SHARED);
 #endif
 	
-	/* If there may be waiters and this was the last s-lock,
-	signal the object */
+	/* Increment lock_word to indicate 1 less reader */
+	if (rw_lock_lock_word_incr(lock, 1) == 0) {
 
-	if (UNIV_UNLIKELY(lock->waiters)
-			&& lock->reader_count == 0) {
-	       	sg = TRUE;
-
-		rw_lock_set_waiters(lock, 0);
-	}
-	
-	mutex_exit(mutex);
-
-	if (UNIV_UNLIKELY(sg)) {
-#ifdef __WIN__
+		/* wait_ex waiter exists. It may not be asleep, but we signal
+                anyway. We do not wake other waiters, because they can't
+                exist without wait_ex waiter and wait_ex waiter goes first.*/
 		os_event_set(lock->wait_ex_event);
-#endif
-		os_event_set(lock->event);
 		sync_array_object_signalled(sync_primary_wait_array);
+
 	}
 
         ut_ad(rw_lock_validate(lock));
@@ -405,16 +528,15 @@
 /*====================*/
 	rw_lock_t*	lock)	/* in: rw-lock */
 {
-	/* Reset the shared lock by decrementing the reader count */
-
-	ut_ad(lock->reader_count > 0);
-
-	lock->reader_count--;
+	ut_ad(lock->lock_word < X_LOCK_DECR);
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, 0, RW_LOCK_SHARED);
 #endif
 
+	/* Decrease reader count by incrementing lock_word */
+	lock->lock_word++;
+
 	ut_ad(!lock->waiters);
         ut_ad(rw_lock_validate(lock));
 #ifdef UNIV_SYNC_PERF_STAT
@@ -435,42 +557,32 @@
 #endif
 	)
 {
-	ibool	sg 	= FALSE;
-
-        /* Acquire the mutex protecting the rw-lock fields */
-	mutex_enter(&(lock->mutex));
-
-	/* Reset the exclusive lock if this thread no longer has an x-mode
-	lock */
-
-	ut_ad(lock->writer_count > 0);
+	ut_ad((lock->lock_word % X_LOCK_DECR) == 0);
 
-	lock->writer_count--;
-
-	if (lock->writer_count == 0) {
-		rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
+	/* lock->recursive flag also indicates if lock->writer_thread is
+	valid or stale. If we are the last of the recursive callers
+	then we must unset lock->recursive flag to indicate that the
+	lock->writer_thread is now stale.
+	Note that since we still hold the x-lock we can safely read the
+	lock_word. */
+	if (lock->lock_word == 0) {
+		/* Last caller in a possible recursive chain. */
+		lock->recursive = FALSE;
 	}
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, pass, RW_LOCK_EX);
 #endif
 	
-	/* If there may be waiters, signal the lock */
-	if (UNIV_UNLIKELY(lock->waiters)
-			&& lock->writer_count == 0) {
-
-	       	sg = TRUE;
-		rw_lock_set_waiters(lock, 0);
-	}
-	
-	mutex_exit(&(lock->mutex));
-
-	if (UNIV_UNLIKELY(sg)) {
-#ifdef __WIN__
-		os_event_set(lock->wait_ex_event);
-#endif
-		os_event_set(lock->event);
-		sync_array_object_signalled(sync_primary_wait_array);
+	if (rw_lock_lock_word_incr(lock, X_LOCK_DECR) == X_LOCK_DECR) {
+		/* Lock is now free. May have to signal read/write waiters.
+                We do not need to signal wait_ex waiters, since they cannot
+                exist when there is a writer. */
+		if (lock->waiters) {
+			rw_lock_reset_waiter_flag(lock);
+			os_event_set(lock->event);
+			sync_array_object_signalled(sync_primary_wait_array);
+		}
 	}
 
         ut_ad(rw_lock_validate(lock));
@@ -492,18 +604,18 @@
 	/* Reset the exclusive lock if this thread no longer has an x-mode
 	lock */
 
-	ut_ad(lock->writer_count > 0);
-
-	lock->writer_count--;
-
-	if (lock->writer_count == 0) {
-		rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
-	}
+	ut_ad((lock->lock_word % X_LOCK_DECR) == 0);
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, 0, RW_LOCK_EX);
 #endif
 
+	if (lock->lock_word == 0) {
+		lock->recursive = FALSE;
+	}
+
+	lock->lock_word += X_LOCK_DECR;
+
 	ut_ad(!lock->waiters);
         ut_ad(rw_lock_validate(lock));
 
diff -ruN a/innobase/include/sync0sync.h b/innobase/include/sync0sync.h
--- a/innobase/include/sync0sync.h	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/include/sync0sync.h	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 Mutex, the basic synchronization primitive
 
-(c) 1995 Innobase Oy
-
 Created 9/5/1995 Heikki Tuuri
 *******************************************************/
 
@@ -465,8 +488,11 @@
 struct mutex_struct {
 	os_event_t	event;	/* Used by sync0arr.c for the wait queue */
 	ulint	lock_word;	/* This ulint is the target of the atomic
-				test-and-set instruction in Win32 */
-#if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER)
+				test-and-set instruction in Win32 and
+				x86 32/64 with GCC 4.1.0 or later version */
+#if defined(_WIN32) && defined(UNIV_CAN_USE_X86_ASSEMBLER)
+#elif defined(HAVE_ATOMIC_BUILTINS)
+#else
 	os_fast_mutex_t
 		os_fast_mutex;	/* In other systems we use this OS mutex
 				in place of lock_word */
@@ -525,8 +551,7 @@
 /* The number of system calls made in this module. Intended for performance
 monitoring. */
 
-extern 	ulint	mutex_system_call_count;
-extern	ulint	mutex_exit_count;
+extern	ib_longlong	mutex_exit_count;
 
 /* Latching order checks start when this is set TRUE */
 extern ibool	sync_order_checks_on;
diff -ruN a/innobase/include/sync0sync.ic b/innobase/include/sync0sync.ic
--- a/innobase/include/sync0sync.ic	2009-09-10 04:02:59.000000000 +0900
+++ b/innobase/include/sync0sync.ic	2009-10-22 15:18:44.000000000 +0900
@@ -1,21 +1,34 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 Mutex, the basic synchronization primitive
 
-(c) 1995 Innobase Oy
-
 Created 9/5/1995 Heikki Tuuri
 *******************************************************/
 
-#if defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
-/* %z0: Use the size of operand %0 which in our case is *m to determine
-instruction size, it should end up as xchgl. "1" in the input constraint,
-says that "in" has to go in the same place as "out".*/
-#define TAS(m, in, out) \
-	asm volatile ("xchg%z0 %2, %0" \
-	: "=g" (*(m)), "=r" (out) \
-	: "1" (in))	/* Note: "1" here refers to "=r" (out) */
-#endif
-
 /**********************************************************************
 Sets the waiters field in a mutex. */
 
@@ -94,12 +107,8 @@
 	/* mutex_fence(); */
 
 	return(res);
-#elif defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
-	ulint	res;
-
-	TAS(&mutex->lock_word, 1, res);
-
-	return(res);
+#elif defined(HAVE_ATOMIC_BUILTINS)
+	return __sync_lock_test_and_set(&(mutex->lock_word), 1);
 #else
 	ibool	ret;
 
@@ -136,10 +145,11 @@
 	__asm   MOV     EDX, 0
         __asm   MOV     ECX, lw
         __asm   XCHG    EDX, DWORD PTR [ECX]                    
-#elif defined(not_defined) && defined(__GNUC__) && defined(UNIV_INTEL_X86)
-	ulint	res;
-
-	TAS(&mutex->lock_word, 0, res);
+#elif defined(HAVE_ATOMIC_BUILTINS)
+	/* In theory __sync_lock_release should be used to release the lock.
+	Unfortunately, it does not work properly alone. The workaround is
+	that more conservative __sync_lock_test_and_set is used instead. */
+	__sync_lock_test_and_set(&(mutex->lock_word), 0);
 #else
 	mutex->lock_word = 0;
 
diff -ruN a/innobase/row/row0sel.c b/innobase/row/row0sel.c
--- a/innobase/row/row0sel.c	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/row/row0sel.c	2009-10-22 15:18:44.000000000 +0900
@@ -1178,7 +1178,7 @@
 			rw_lock_s_lock(&btr_search_latch);
 
 			search_latch_locked = TRUE;
-		} else if (btr_search_latch.writer_is_wait_ex) {
+		} else if (rw_lock_get_writer(&btr_search_latch) == RW_LOCK_WAIT_EX) {
 
 			/* There is an x-latch request waiting: release the
 			s-latch for a moment; as an s-latch here is often
@@ -3123,7 +3123,7 @@
 	/* PHASE 0: Release a possible s-latch we are holding on the
 	adaptive hash index latch if there is someone waiting behind */
 
-	if (UNIV_UNLIKELY(btr_search_latch.writer != RW_LOCK_NOT_LOCKED)
+	if (UNIV_UNLIKELY(rw_lock_get_writer(&btr_search_latch) != RW_LOCK_NOT_LOCKED)
 	    && trx->has_search_latch) {
 
 		/* There is an x-latch request on the adaptive hash index:
diff -ruN a/innobase/sync/sync0arr.c b/innobase/sync/sync0arr.c
--- a/innobase/sync/sync0arr.c	2009-09-10 04:03:01.000000000 +0900
+++ b/innobase/sync/sync0arr.c	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 The wait array used in synchronization primitives
 
-(c) 1995 Innobase Oy
-
 Created 9/5/1995 Heikki Tuuri
 *******************************************************/
 
@@ -297,25 +320,21 @@
 }
 
 /***********************************************************************
-Puts the cell event in reset state. */
+Returns the event that the thread owning the cell waits for. */
 static
-ib_longlong
-sync_cell_event_reset(
-/*==================*/
-				/* out: value of signal_count
-				at the time of reset. */
-	ulint		type,	/* in: lock type mutex/rw_lock */
-	void*		object) /* in: the rw_lock/mutex object */
+os_event_t
+sync_cell_get_event(
+/*================*/
+	sync_cell_t*	cell) /* in: non-empty sync array cell */
 {
+	ulint type = cell->request_type;
+
 	if (type == SYNC_MUTEX) {
-		return(os_event_reset(((mutex_t *) object)->event));
-#ifdef __WIN__
+		return(((mutex_t *) cell->wait_object)->event);
 	} else if (type == RW_LOCK_WAIT_EX) {
-		return(os_event_reset(
-		       ((rw_lock_t *) object)->wait_ex_event));
-#endif
-	} else {
-		return(os_event_reset(((rw_lock_t *) object)->event));
+		return(((rw_lock_t *) cell->wait_object)->wait_ex_event);
+	} else { /* RW_LOCK_SHARED and RW_LOCK_EX wait on the same event */
+		return(((rw_lock_t *) cell->wait_object)->event);
 	}
 }		
 
@@ -334,6 +353,7 @@
         ulint*   	index)  /* out: index of the reserved cell */
 {
         sync_cell_t*   	cell;
+	os_event_t      event;
         ulint           i;
         
         ut_a(object);
@@ -372,8 +392,8 @@
 			/* Make sure the event is reset and also store
 			the value of signal_count at which the event
 			was reset. */
-			cell->signal_count = sync_cell_event_reset(type,
-								object);
+                        event = sync_cell_get_event(cell);
+			cell->signal_count = os_event_reset(event);
 
 			cell->reservation_time = time(NULL);
 
@@ -413,19 +433,7 @@
 	ut_a(!cell->waiting);
 	ut_ad(os_thread_get_curr_id() == cell->thread);
 
-	if (cell->request_type == SYNC_MUTEX) {
-		event = ((mutex_t*) cell->wait_object)->event;
-#ifdef __WIN__
-	/* On windows if the thread about to wait is the one which
-	has set the state of the rw_lock to RW_LOCK_WAIT_EX, then
-	it waits on a special event i.e.: wait_ex_event. */
-	} else if (cell->request_type == RW_LOCK_WAIT_EX) {
-		event = ((rw_lock_t*) cell->wait_object)->wait_ex_event;
-#endif
-	} else {	
-		event = ((rw_lock_t*) cell->wait_object)->event;
-	}
-
+	event = sync_cell_get_event(cell);
        	cell->waiting = TRUE;
 
 #ifdef UNIV_SYNC_DEBUG
@@ -464,6 +472,7 @@
 	mutex_t*	mutex;
 	rw_lock_t*	rwlock;
 	ulint		type;
+	ulint		writer;
 
 	type = cell->request_type;
 
@@ -492,9 +501,7 @@
 			(ulong) mutex->waiters);
 
 	} else if (type == RW_LOCK_EX
-#ifdef __WIN__
 		   || type == RW_LOCK_WAIT_EX
-#endif
 		   || type == RW_LOCK_SHARED) {
 
 		fputs(type == RW_LOCK_EX ? "X-lock on" : "S-lock on", file);
@@ -505,21 +512,24 @@
 			" RW-latch at %p created in file %s line %lu\n",
 			rwlock, rwlock->cfile_name,
 			(ulong) rwlock->cline);
-		if (rwlock->writer != RW_LOCK_NOT_LOCKED) {
+		writer = rw_lock_get_writer(rwlock);
+		if (writer != RW_LOCK_NOT_LOCKED) {
 			fprintf(file,
 			"a writer (thread id %lu) has reserved it in mode %s",
 				(ulong) os_thread_pf(rwlock->writer_thread),
-				rwlock->writer == RW_LOCK_EX
+				writer == RW_LOCK_EX
 				? " exclusive\n"
 				: " wait exclusive\n");
 		}
 		
 		fprintf(file,
-			"number of readers %lu, waiters flag %lu\n"
+			"number of readers %lu, waiters flag %lu, "
+                        "lock_word: %lx\n"
 			"Last time read locked in file %s line %lu\n"
 			"Last time write locked in file %s line %lu\n",
-			(ulong) rwlock->reader_count,
+			(ulong) rw_lock_get_reader_count(rwlock),
 			(ulong) rwlock->waiters,
+			rwlock->lock_word,
 			rwlock->last_s_file_name,
 			(ulong) rwlock->last_s_line,
 			rwlock->last_x_file_name,
@@ -773,28 +783,30 @@
 			return(TRUE);
 		}
 
-	} else if (cell->request_type == RW_LOCK_EX
-		   || cell->request_type == RW_LOCK_WAIT_EX) {
+	} else if (cell->request_type == RW_LOCK_EX) {
 
 	    	lock = cell->wait_object;
 
-	    	if (rw_lock_get_reader_count(lock) == 0
-		    && rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
+		if (lock->lock_word > 0) {
+		/* Either unlocked or only read locked. */
 
 			return(TRUE);
 		}
 
-	    	if (rw_lock_get_reader_count(lock) == 0
-		    && rw_lock_get_writer(lock) == RW_LOCK_WAIT_EX
-		    && os_thread_eq(lock->writer_thread, cell->thread)) {
+        } else if (cell->request_type == RW_LOCK_WAIT_EX) {
+
+		lock = cell->wait_object;
+
+                /* lock_word == 0 means all readers have left */
+		if (lock->lock_word == 0) {
 
 			return(TRUE);
 		}
-
 	} else if (cell->request_type == RW_LOCK_SHARED) {
 	    	lock = cell->wait_object;
 
-		if (rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
+                /* lock_word > 0 means no writer or reserved writer */
+		if (lock->lock_word > 0) {
 		
 			return(TRUE);
 		}
@@ -839,11 +851,15 @@
 /*========================*/
 	sync_array_t*	arr)	/* in: wait array */
 {
+#ifdef HAVE_ATOMIC_BUILTINS
+	(void) os_atomic_increment(&arr->sg_count, 1);
+#else
         sync_array_enter(arr);
 
 	arr->sg_count++;
 
         sync_array_exit(arr);
+#endif
 }
 
 /**************************************************************************
@@ -859,6 +875,7 @@
         sync_cell_t*   	cell;
         ulint           count;
         ulint           i;
+	os_event_t      event;
 
         sync_array_enter(arr);
 
@@ -868,36 +885,20 @@
         while (count < arr->n_reserved) {
 
         	cell = sync_array_get_nth_cell(arr, i);
+		i++;
 
-                if (cell->wait_object != NULL) {
-
+		if (cell->wait_object == NULL) {
+			continue;
+		}
                         count++;
 
                         if (sync_arr_cell_can_wake_up(cell)) {
 
-				if (cell->request_type == SYNC_MUTEX) {
-					mutex_t*	mutex;
+			event = sync_cell_get_event(cell);
 
-					mutex = cell->wait_object;
-					os_event_set(mutex->event);
-#ifdef __WIN__
-				} else if (cell->request_type
-					   == RW_LOCK_WAIT_EX) {
-					rw_lock_t*	lock;
-
-					lock = cell->wait_object;
-					os_event_set(lock->wait_ex_event);
-#endif
-				} else {
-					rw_lock_t*	lock;
-
-					lock = cell->wait_object;
-					os_event_set(lock->event);
-				}
-                        }
+			os_event_set(event);
                 }
 
-                i++;
         }
 
         sync_array_exit(arr);
@@ -1014,4 +1015,3 @@
         
         sync_array_exit(arr);
 }
-
diff -ruN a/innobase/sync/sync0rw.c b/innobase/sync/sync0rw.c
--- a/innobase/sync/sync0rw.c	2009-09-10 04:03:01.000000000 +0900
+++ b/innobase/sync/sync0rw.c	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 The read-write lock (for thread synchronization)
 
-(c) 1995 Innobase Oy
-
 Created 9/11/1995 Heikki Tuuri
 *******************************************************/
 
@@ -15,17 +38,110 @@
 #include "mem0mem.h"
 #include "srv0srv.h"
 
-ulint	rw_s_system_call_count	= 0;
-ulint	rw_s_spin_wait_count	= 0;
-ulint	rw_s_os_wait_count	= 0;
+/*
+	IMPLEMENTATION OF THE RW_LOCK
+	=============================
+The status of a rw_lock is held in lock_word. The initial value of lock_word is
+X_LOCK_DECR. lock_word is decremented by 1 for each s-lock and by X_LOCK_DECR
+for each x-lock. This describes the lock state for each value of lock_word:
+
+lock_word == X_LOCK_DECR:      Unlocked.
+0 < lock_word < X_LOCK_DECR:   Read locked, no waiting writers.
+			       (X_LOCK_DECR - lock_word) is the
+			       number of readers that hold the lock.
+lock_word == 0:		       Write locked
+-X_LOCK_DECR < lock_word < 0:  Read locked, with a waiting writer.
+			       (-lock_word) is the number of readers
+			       that hold the lock.
+lock_word <= -X_LOCK_DECR:     Recursively write locked. lock_word has been
+			       decremented by X_LOCK_DECR once for each lock,
+			       so the number of locks is:
+			       ((-lock_word) / X_LOCK_DECR) + 1
+When lock_word <= -X_LOCK_DECR, we also know that lock_word % X_LOCK_DECR == 0:
+other values of lock_word are invalid.
+
+The lock_word is always read and updated atomically and consistently, so that
+it always represents the state of the lock, and the state of the lock changes
+with a single atomic operation. This lock_word holds all of the information
+that a thread needs in order to determine if it is eligible to gain the lock
+or if it must spin or sleep. The one exception to this is that writer_thread
+must be verified before recursive write locks: to solve this scenario, we make
+writer_thread readable by all threads, but only writeable by the x-lock holder.
+
+The other members of the lock obey the following rules to remain consistent:
+
+recursive:	This and the writer_thread field together control the
+		behaviour of recursive x-locking.
+		lock->recursive must be FALSE in following states:
+			1) The writer_thread contains garbage i.e.: the
+			lock has just been initialized.
+			2) The lock is not x-held and there is no
+			x-waiter waiting on WAIT_EX event.
+			3) The lock is x-held or there is an x-waiter
+			waiting on WAIT_EX event but the 'pass' value
+			is non-zero.
+		lock->recursive is TRUE iff:
+			1) The lock is x-held or there is an x-waiter
+			waiting on WAIT_EX event and the 'pass' value
+			is zero.
+		This flag must be set after the writer_thread field
+		has been updated with a memory ordering barrier.
+		It is unset before the lock_word has been incremented.
+writer_thread:	Is used only in recursive x-locking. Can only be safely
+		read iff lock->recursive flag is TRUE.
+		This field is uninitialized at lock creation time and
+		is updated atomically when x-lock is acquired or when
+		move_ownership is called. A thread is only allowed to
+		set the value of this field to it's thread_id i.e.: a
+		thread cannot set writer_thread to some other thread's
+		id.
+waiters:	May be set to 1 anytime, but to avoid unnecessary wake-up
+		signals, it should only be set to 1 when there are threads
+		waiting on event. Must be 1 when a writer starts waiting to
+		ensure the current x-locking thread sends a wake-up signal
+		during unlock. May only be reset to 0 immediately before a
+		a wake-up signal is sent to event. On most platforms, a
+		memory barrier is required after waiters is set, and before
+		verifying lock_word is still held, to ensure some unlocker
+		really does see the flags new value.
+event:		Threads wait on event for read or writer lock when another
+		thread has an x-lock or an x-lock reservation (wait_ex). A
+		thread may only	wait on event after performing the following
+		actions in order:
+		   (1) Record the counter value of event (with os_event_reset).
+		   (2) Set waiters to 1.
+		   (3) Verify lock_word <= 0.
+		(1) must come before (2) to ensure signal is not missed.
+		(2) must come before (3) to ensure a signal is sent.
+		These restrictions force the above ordering.
+		Immediately before sending the wake-up signal, we should:
+		   (1) Verify lock_word == X_LOCK_DECR (unlocked)
+		   (2) Reset waiters to 0.
+wait_ex_event:	A thread may only wait on the wait_ex_event after it has
+		performed the following actions in order:
+		   (1) Decrement lock_word by X_LOCK_DECR.
+		   (2) Record counter value of wait_ex_event (os_event_reset,
+                       called from sync_array_reserve_cell).
+		   (3) Verify that lock_word < 0.
+		(1) must come first to ensures no other threads become reader
+                or next writer, and notifies unlocker that signal must be sent.
+                (2) must come before (3) to ensure the signal is not missed.
+		These restrictions force the above ordering.
+		Immediately before sending the wake-up signal, we should:
+		   Verify lock_word == 0 (waiting thread holds x_lock)
+*/
+
+ib_longlong	rw_s_spin_wait_count	= 0;
+ib_longlong	rw_s_spin_round_count	= 0;
+ib_longlong	rw_s_os_wait_count	= 0;
+
+ib_longlong	rw_s_exit_count		= 0;
+
+ib_longlong	rw_x_spin_wait_count	= 0;
+ib_longlong	rw_x_spin_round_count	= 0;
+ib_longlong	rw_x_os_wait_count	= 0;
 
-ulint	rw_s_exit_count		= 0;
-
-ulint	rw_x_system_call_count	= 0;
-ulint	rw_x_spin_wait_count	= 0;
-ulint	rw_x_os_wait_count	= 0;
-
-ulint	rw_x_exit_count		= 0;
+ib_longlong	rw_x_exit_count		= 0;
 
 /* The global list of rw-locks */
 rw_lock_list_t	rw_lock_list;
@@ -99,22 +215,30 @@
 	object is created, then the following call initializes
 	the sync system. */
 
+#ifndef HAVE_ATOMIC_BUILTINS
 	mutex_create(rw_lock_get_mutex(lock));
 	mutex_set_level(rw_lock_get_mutex(lock), SYNC_NO_ORDER_CHECK);
 
 	lock->mutex.cfile_name = cfile_name;
 	lock->mutex.cline = cline;
-#if defined UNIV_DEBUG && !defined UNIV_HOTBACKUP
+# if defined UNIV_DEBUG && !defined UNIV_HOTBACKUP
 	lock->mutex.cmutex_name = cmutex_name;
 	lock->mutex.mutex_type = 1;
-#endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
+# endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
   
-	rw_lock_set_waiters(lock, 0);
-	rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
-	lock->writer_count = 0;
-	rw_lock_set_reader_count(lock, 0);
-
-	lock->writer_is_wait_ex = FALSE;
+#else /* HAVE_ATOMIC_BUILTINS */
+# ifdef UNIV_DEBUG
+	UT_NOT_USED(cmutex_name);
+# endif
+#endif /* HAVE_ATOMIC_BUILTINS */
+
+	lock->lock_word = X_LOCK_DECR;
+	lock->waiters = 0;
+
+	/* We set this value to signify that lock->writer_thread
+	contains garbage at initialization and cannot be used for
+	recursive x-locking. */
+	lock->recursive = FALSE;
 
 #ifdef UNIV_SYNC_DEBUG
 	UT_LIST_INIT(lock->debug_list);
@@ -126,15 +250,13 @@
 	lock->cfile_name = cfile_name;
 	lock->cline = cline;
 
+	lock->count_os_wait = 0;
 	lock->last_s_file_name = "not yet reserved";
 	lock->last_x_file_name = "not yet reserved";
 	lock->last_s_line = 0;
 	lock->last_x_line = 0;
 	lock->event = os_event_create(NULL);
-
-#ifdef __WIN__
 	lock->wait_ex_event = os_event_create(NULL);
-#endif
 
 	mutex_enter(&rw_lock_list_mutex);
 	
@@ -158,23 +280,17 @@
 /*=========*/
 	rw_lock_t*	lock)	/* in: rw-lock */
 {
-#ifdef UNIV_DEBUG
 	ut_a(rw_lock_validate(lock));
-#endif /* UNIV_DEBUG */
-	ut_a(rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED);
-	ut_a(rw_lock_get_waiters(lock) == 0);
-	ut_a(rw_lock_get_reader_count(lock) == 0);
+	ut_a(lock->lock_word == X_LOCK_DECR);
 	
-	lock->magic_n = 0;
-
+#ifndef HAVE_ATOMIC_BUILTINS
 	mutex_free(rw_lock_get_mutex(lock));
+#endif /* HAVE_ATOMIC_BUILTINS */
 
 	mutex_enter(&rw_lock_list_mutex);
 	os_event_free(lock->event);
 
-#ifdef __WIN__
 	os_event_free(lock->wait_ex_event);
-#endif
 
 	if (UT_LIST_GET_PREV(list, lock)) {
 		ut_a(UT_LIST_GET_PREV(list, lock)->magic_n == RW_LOCK_MAGIC_N);
@@ -186,6 +302,8 @@
 	UT_LIST_REMOVE(list, rw_lock_list, lock);
 
 	mutex_exit(&rw_lock_list_mutex);
+
+	lock->magic_n = 0;
 }
 
 /**********************************************************************
@@ -199,19 +317,12 @@
 {
 	ut_a(lock);
 
-	mutex_enter(rw_lock_get_mutex(lock));
+	ulint waiters = rw_lock_get_waiters(lock);
+	lint lock_word = lock->lock_word;
 
 	ut_a(lock->magic_n == RW_LOCK_MAGIC_N);
-	ut_a((rw_lock_get_reader_count(lock) == 0)
-	     || (rw_lock_get_writer(lock) != RW_LOCK_EX));
-	ut_a((rw_lock_get_writer(lock) == RW_LOCK_EX)
-	     || (rw_lock_get_writer(lock) == RW_LOCK_WAIT_EX)
-	     || (rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED));
-	ut_a((rw_lock_get_waiters(lock) == 0)
-	     || (rw_lock_get_waiters(lock) == 1));
-	ut_a((lock->writer != RW_LOCK_EX) || (lock->writer_count > 0));
-	     
-	mutex_exit(rw_lock_get_mutex(lock));
+	ut_a(waiters == 0 || waiters == 1);
+	ut_a(lock_word > -X_LOCK_DECR ||(-lock_word) % X_LOCK_DECR == 0);
 
 	return(TRUE);
 }
@@ -232,18 +343,15 @@
 	ulint		line)	/* in: line where requested */
 {
         ulint    index;	/* index of the reserved wait cell */
-        ulint    i;   	/* spin round count */
+	ulint	 i = 0;	/* spin round count */
         
         ut_ad(rw_lock_validate(lock));
 
+	rw_s_spin_wait_count++;	/* Count calls to this function */
 lock_loop:
-	rw_s_spin_wait_count++;
 
 	/* Spin waiting for the writer field to become free */
-        i = 0;
-
-        while (rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED
-						&& i < SYNC_SPIN_ROUNDS) {
+	while (i < SYNC_SPIN_ROUNDS && lock->lock_word <= 0) {
         	if (srv_spin_wait_delay) {
         		ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
         	}
@@ -262,28 +370,32 @@
 		lock->cfile_name, (ulong) lock->cline, (ulong) i);
 	}
 
-	mutex_enter(rw_lock_get_mutex(lock));
-
         /* We try once again to obtain the lock */
-
 	if (TRUE == rw_lock_s_lock_low(lock, pass, file_name, line)) {
-		mutex_exit(rw_lock_get_mutex(lock));
+		rw_s_spin_round_count += i;
 
 		return; /* Success */
 	} else {
-		/* If we get here, locking did not succeed, we may
-		suspend the thread to wait in the wait array */
 
-		rw_s_system_call_count++;
+		if (i < SYNC_SPIN_ROUNDS) {
+			goto lock_loop;
+		}
+
+		rw_s_spin_round_count += i;
 
         	sync_array_reserve_cell(sync_primary_wait_array,
 				lock, RW_LOCK_SHARED,
 				file_name, line,
 				&index);
 
-		rw_lock_set_waiters(lock, 1);
-
-		mutex_exit(rw_lock_get_mutex(lock));
+		/* Set waiters before checking lock_word to ensure wake-up
+                signal is sent. This may lead to some unnecessary signals. */
+		rw_lock_set_waiter_flag(lock);
+
+		if (TRUE == rw_lock_s_lock_low(lock, pass, file_name, line)) {
+			sync_array_free_cell(sync_primary_wait_array, index);
+			return; /* Success */
+		}
 
 		if (srv_print_latch_waits) {
 			fprintf(stderr,
@@ -292,11 +404,13 @@
 		        lock, lock->cfile_name, (ulong) lock->cline);
 		}
 
-		rw_s_system_call_count++;
+		/* these stats may not be accurate */
+		lock->count_os_wait++;
 		rw_s_os_wait_count++;
 
        	 	sync_array_wait_event(sync_primary_wait_array, index);
 
+		i = 0;
         	goto lock_loop;
 	}        
 }
@@ -318,114 +432,130 @@
 {
 	ut_ad(rw_lock_is_locked(lock, RW_LOCK_EX));
 
-	mutex_enter(&(lock->mutex));
-
-	lock->writer_thread = os_thread_get_curr_id();
-
-	lock->pass = 0;
-
-	mutex_exit(&(lock->mutex));
+	rw_lock_set_writer_id_and_recursion_flag(lock, TRUE);
 }
 
 /**********************************************************************
-Low-level function for acquiring an exclusive lock. */
+Function for the next writer to call. Waits for readers to exit.
+The caller must have already decremented lock_word by X_LOCK_DECR.*/
 UNIV_INLINE
-ulint
-rw_lock_x_lock_low(
-/*===============*/
-				/* out: RW_LOCK_NOT_LOCKED if did
-				not succeed, RW_LOCK_EX if success,
-				RW_LOCK_WAIT_EX, if got wait reservation */
+void
+rw_lock_x_lock_wait(
+/*================*/
         rw_lock_t*   	lock,  	/* in: pointer to rw-lock */
+#ifdef UNIV_SYNC_DEBUG
 	ulint		pass,	/* in: pass value; != 0, if the lock will
 				be passed to another thread to unlock */
+#endif
 	const char*	file_name,/* in: file name where lock requested */
 	ulint		line)	/* in: line where requested */
 {
-#ifdef UNIV_SYNC_DEBUG
-	ut_ad(mutex_own(rw_lock_get_mutex(lock)));
-#endif /* UNIV_SYNC_DEBUG */
-	if (rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
+	ulint index;
+	ulint i = 0;
 
-		if (rw_lock_get_reader_count(lock) == 0) {
+	ut_ad(lock->lock_word <= 0);
+
+	while (lock->lock_word < 0) {
+		if (srv_spin_wait_delay) {
+			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
+		}
+		if(i < SYNC_SPIN_ROUNDS) {
+			i++;
+			continue;
+		}
 			
-			rw_lock_set_writer(lock, RW_LOCK_EX);
-			lock->writer_thread = os_thread_get_curr_id();
-			lock->writer_count++;
-			lock->pass = pass;
+		/* If there is still a reader, then go to sleep.*/
+		rw_x_spin_round_count += i;
+		i = 0;
+		sync_array_reserve_cell(sync_primary_wait_array,
+					lock,
+					RW_LOCK_WAIT_EX,
+					file_name, line,
+					&index);
+		/* Check lock_word to ensure wake-up isn't missed.*/
+		if(lock->lock_word < 0) {
 			
+			/* these stats may not be accurate */
+			lock->count_os_wait++;
+			rw_x_os_wait_count++;
+
+                        /* Add debug info as it is needed to detect possible
+                        deadlock. We must add info for WAIT_EX thread for
+                        deadlock detection to work properly. */
 #ifdef UNIV_SYNC_DEBUG
-			rw_lock_add_debug_info(lock, pass, RW_LOCK_EX,
+			rw_lock_add_debug_info(lock, pass, RW_LOCK_WAIT_EX,
 							file_name, line);
 #endif
-			lock->last_x_file_name = file_name;
-			lock->last_x_line = line;
-		
-			/* Locking succeeded, we may return */
-			return(RW_LOCK_EX);
-		} else {
-			/* There are readers, we have to wait */
-			rw_lock_set_writer(lock, RW_LOCK_WAIT_EX);
-			lock->writer_thread = os_thread_get_curr_id();
-			lock->pass = pass;
-			lock->writer_is_wait_ex = TRUE;
 
+			sync_array_wait_event(sync_primary_wait_array,
+					      index);
 #ifdef UNIV_SYNC_DEBUG
-			rw_lock_add_debug_info(lock, pass, RW_LOCK_WAIT_EX,
-							file_name, line);
+			rw_lock_remove_debug_info(lock, pass,
+					       RW_LOCK_WAIT_EX);
 #endif
-
-			return(RW_LOCK_WAIT_EX);
+                        /* It is possible to wake when lock_word < 0.
+                        We must pass the while-loop check to proceed.*/
+		} else {
+			sync_array_free_cell(sync_primary_wait_array,
+					     index);
 		}
+	}
+	rw_x_spin_round_count += i;
+}
 
-	} else if ((rw_lock_get_writer(lock) == RW_LOCK_WAIT_EX)
-		   && os_thread_eq(lock->writer_thread,
-						os_thread_get_curr_id())) {
+/**********************************************************************
+Low-level function for acquiring an exclusive lock. */
+UNIV_INLINE
+ibool
+rw_lock_x_lock_low(
+/*===============*/
+				/* out: RW_LOCK_NOT_LOCKED if did
+				not succeed, RW_LOCK_EX if success. */
+	rw_lock_t*	lock,	/* in: pointer to rw-lock */
+	ulint		pass,	/* in: pass value; != 0, if the lock will
+				be passed to another thread to unlock */
+	const char*	file_name,/* in: file name where lock requested */
+	ulint		line)	/* in: line where requested */
+{
+	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
 
-		if (rw_lock_get_reader_count(lock) == 0) {
+	if (rw_lock_lock_word_decr(lock, X_LOCK_DECR)) {
 
-			rw_lock_set_writer(lock, RW_LOCK_EX);
-			lock->writer_count++;
-			lock->pass = pass;
-			lock->writer_is_wait_ex = FALSE;
+		/* lock->recursive also tells us if the writer_thread
+		field is stale or active. As we are going to write
+		our own thread id in that field it must be that the
+		current writer_thread value is not active. */
+		ut_a(!lock->recursive);
 
+		/* Decrement occurred: we are writer or next-writer. */
+		rw_lock_set_writer_id_and_recursion_flag(lock,
+						pass ? FALSE : TRUE);
+
+		rw_lock_x_lock_wait(lock,
 #ifdef UNIV_SYNC_DEBUG
-			rw_lock_remove_debug_info(lock, pass, RW_LOCK_WAIT_EX);
-			rw_lock_add_debug_info(lock, pass, RW_LOCK_EX,
-							file_name, line);
+				    pass,
 #endif
+                                    file_name, line);
 		
-			lock->last_x_file_name = file_name;
-			lock->last_x_line = line;
-
-			/* Locking succeeded, we may return */
-			return(RW_LOCK_EX);
+	} else {
+		/* Decrement failed: relock or failed lock */
+		if (!pass && lock->recursive
+		    && os_thread_eq(lock->writer_thread, curr_thread)) {
+			/* Relock */
+                        lock->lock_word -= X_LOCK_DECR;
+		} else {
+			/* Another thread locked before us */
+			return(FALSE);
 		}
-
-		return(RW_LOCK_WAIT_EX);
-
-	} else if ((rw_lock_get_writer(lock) == RW_LOCK_EX)
-		   && os_thread_eq(lock->writer_thread,
-						os_thread_get_curr_id())
-		   && (lock->pass == 0)
-		   && (pass == 0)) {
-
-		lock->writer_count++;
-
+	}
 #ifdef UNIV_SYNC_DEBUG
-		rw_lock_add_debug_info(lock, pass, RW_LOCK_EX, file_name,
-									line);
+	rw_lock_add_debug_info(lock, pass, RW_LOCK_EX,
+			       file_name, line);
 #endif
+	lock->last_x_file_name = file_name;
+	lock->last_x_line = (unsigned int) line;
 		
-		lock->last_x_file_name = file_name;
-		lock->last_x_line = line;
-
-		/* Locking succeeded, we may return */
-		return(RW_LOCK_EX);
-	}
-
-	/* Locking did not succeed */
-	return(RW_LOCK_NOT_LOCKED);
+	return(TRUE);
 }
 
 /**********************************************************************
@@ -448,47 +578,30 @@
 	ulint		line)	/* in: line where requested */
 {
         ulint	index;  /* index of the reserved wait cell */
-        ulint	state;	/* lock state acquired */
         ulint	i;	/* spin round count */
+	ibool   spinning = FALSE;
         
         ut_ad(rw_lock_validate(lock));
 
-lock_loop:
-        /* Acquire the mutex protecting the rw-lock fields */
-	mutex_enter_fast(&(lock->mutex));
-
-	state = rw_lock_x_lock_low(lock, pass, file_name, line);
+	i = 0;
 		
-	mutex_exit(&(lock->mutex));
+lock_loop:
         
-	if (state == RW_LOCK_EX) {
+	if (rw_lock_x_lock_low(lock, pass, file_name, line)) {
+		rw_x_spin_round_count += i;
 
 		return;	/* Locking succeeded */
 
-	} else if (state == RW_LOCK_NOT_LOCKED) {
-
- 		/* Spin waiting for the writer field to become free */
-		i = 0;
-
-        	while (rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED 
-               					&& i < SYNC_SPIN_ROUNDS) {
-        		if (srv_spin_wait_delay) {
-				ut_delay(ut_rnd_interval(0,
-							srv_spin_wait_delay));
-        		}
+	} else {
         		
-        		i++;
-        	}
-		if (i == SYNC_SPIN_ROUNDS) {
-			os_thread_yield();
+                if (!spinning) {
+                        spinning = TRUE;
+                        rw_x_spin_wait_count++;
 		}
-        } else if (state == RW_LOCK_WAIT_EX) {
 
- 		/* Spin waiting for the reader count field to become zero */
-		i = 0;
-
-        	while (rw_lock_get_reader_count(lock) != 0 
-               					&& i < SYNC_SPIN_ROUNDS) {
+		/* Spin waiting for the lock_word to become free */
+		while (i < SYNC_SPIN_ROUNDS
+		       && lock->lock_word <= 0) {
         		if (srv_spin_wait_delay) {
 				ut_delay(ut_rnd_interval(0,
 							srv_spin_wait_delay));
@@ -498,12 +611,13 @@
         	}
 		if (i == SYNC_SPIN_ROUNDS) {
 			os_thread_yield();
+		} else {
+			goto lock_loop;
 		}
-        } else {
-		i = 0; /* Eliminate a compiler warning */
-		ut_error;
 	}	
 
+	rw_x_spin_round_count += i;
+
 	if (srv_print_latch_waits) {
 		fprintf(stderr,
 	"Thread %lu spin wait rw-x-lock at %p cfile %s cline %lu rnds %lu\n",
@@ -511,39 +625,20 @@
 		lock->cfile_name, (ulong) lock->cline, (ulong) i);
 	}
 
-	rw_x_spin_wait_count++;
-
-        /* We try once again to obtain the lock. Acquire the mutex protecting
-	the rw-lock fields */
-
-	mutex_enter(rw_lock_get_mutex(lock));
-
-	state = rw_lock_x_lock_low(lock, pass, file_name, line);
-
-	if (state == RW_LOCK_EX) {
-		mutex_exit(rw_lock_get_mutex(lock));
-
-		return;	/* Locking succeeded */
-	}
-
-	rw_x_system_call_count++;
-
         sync_array_reserve_cell(sync_primary_wait_array,
 				lock,
-#ifdef __WIN__
-				/* On windows RW_LOCK_WAIT_EX signifies
-				that this thread should wait on the
-				special wait_ex_event. */
-				(state == RW_LOCK_WAIT_EX)
-				 ? RW_LOCK_WAIT_EX :
-#endif
 				RW_LOCK_EX,
 				file_name, line,
 				&index);
 
-	rw_lock_set_waiters(lock, 1);
-
-	mutex_exit(rw_lock_get_mutex(lock));
+	/* Waiters must be set before checking lock_word, to ensure signal
+	is sent. This could lead to a few unnecessary wake-up signals. */
+	rw_lock_set_waiter_flag(lock);
+
+	if (rw_lock_x_lock_low(lock, pass, file_name, line)) {
+		sync_array_free_cell(sync_primary_wait_array, index);
+		return; /* Locking succeeded */
+	}
 
 	if (srv_print_latch_waits) {
 		fprintf(stderr,
@@ -552,11 +647,13 @@
 		lock->cfile_name, (ulong) lock->cline);
 	}
 
-	rw_x_system_call_count++;
+	/* these stats may not be accurate */
+	lock->count_os_wait++;
 	rw_x_os_wait_count++;
 
         sync_array_wait_event(sync_primary_wait_array, index);
 
+	i = 0;
         goto lock_loop;
 }
 
@@ -697,7 +794,9 @@
 	rw_lock_t*	lock,	/* in: rw-lock */
 	ulint		level)	/* in: level */
 {
+#ifdef UNIV_SYNC_DEBUG
 	lock->level = level;
+#endif /* UNIV_SYNC_DEBUG */
 }
 
 #ifdef UNIV_SYNC_DEBUG
@@ -718,7 +817,7 @@
 	ut_ad(lock);
 	ut_ad(rw_lock_validate(lock));
 
-	mutex_enter(&(lock->mutex));
+	rw_lock_debug_mutex_enter();
 
 	info = UT_LIST_GET_FIRST(lock->debug_list);
 
@@ -728,7 +827,7 @@
 		    && (info->pass == 0)
 		    && (info->lock_type == lock_type)) {
 
-			mutex_exit(&(lock->mutex));
+			rw_lock_debug_mutex_exit();
 		    	/* Found! */
 
 		    	return(TRUE);
@@ -736,7 +835,7 @@
 
 		info = UT_LIST_GET_NEXT(list, info);
 	}
-	mutex_exit(&(lock->mutex));
+	rw_lock_debug_mutex_exit();
 
 	return(FALSE);
 }
@@ -758,22 +857,18 @@
 	ut_ad(lock);
 	ut_ad(rw_lock_validate(lock));
 	
-	mutex_enter(&(lock->mutex));
-
 	if (lock_type == RW_LOCK_SHARED) {
-		if (lock->reader_count > 0) {
+		if (rw_lock_get_reader_count(lock) > 0) {
 			ret = TRUE;
 		}
 	} else if (lock_type == RW_LOCK_EX) {
-		if (lock->writer == RW_LOCK_EX) {
+		if (rw_lock_get_writer(lock) == RW_LOCK_EX) {
 			ret = TRUE;
 		}
 	} else {
 		ut_error;
 	}
 
-	mutex_exit(&(lock->mutex));
-
 	return(ret);
 }
 
@@ -801,11 +896,10 @@
 
 		count++;
 
+#ifndef HAVE_ATOMIC_BUILTINS
 		mutex_enter(&(lock->mutex));
-
-		if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
-		    || (rw_lock_get_reader_count(lock) != 0)
-		    || (rw_lock_get_waiters(lock) != 0)) {
+#endif
+		if (lock->lock_word != X_LOCK_DECR) {
 
 			fprintf(stderr, "RW-LOCK: %p ", lock);
 
@@ -821,8 +915,10 @@
 				info = UT_LIST_GET_NEXT(list, info);
 			}
 		}
-
+#ifndef HAVE_ATOMIC_BUILTINS
 		mutex_exit(&(lock->mutex));
+#endif
+
 		lock = UT_LIST_GET_NEXT(list, lock);
 	}
 
@@ -845,9 +941,10 @@
 		"RW-LATCH INFO\n"
 		"RW-LATCH: %p ", lock);
 
-	if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
-	    || (rw_lock_get_reader_count(lock) != 0)
-	    || (rw_lock_get_waiters(lock) != 0)) {
+#ifndef HAVE_ATOMIC_BUILTINS
+	mutex_enter(&(lock->mutex));
+#endif
+	if (lock->lock_word != X_LOCK_DECR) {
 
 		if (rw_lock_get_waiters(lock)) {
 			fputs(" Waiters for the lock exist\n", stderr);
@@ -861,6 +958,9 @@
 			info = UT_LIST_GET_NEXT(list, info);
 		}
 	}
+#ifndef HAVE_ATOMIC_BUILTINS
+	mutex_exit(&(lock->mutex));
+#endif
 }
 
 /*************************************************************************
@@ -909,14 +1009,11 @@
 	lock = UT_LIST_GET_FIRST(rw_lock_list);
 
 	while (lock != NULL) {
-		mutex_enter(rw_lock_get_mutex(lock));
 
-		if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
-				|| (rw_lock_get_reader_count(lock) != 0)) {
+		if (lock->lock_word != X_LOCK_DECR) {
 			count++;
 		}
 
-		mutex_exit(rw_lock_get_mutex(lock));
 		lock = UT_LIST_GET_NEXT(list, lock);
 	}
 
diff -ruN a/innobase/sync/sync0sync.c b/innobase/sync/sync0sync.c
--- a/innobase/sync/sync0sync.c	2009-10-22 15:15:05.000000000 +0900
+++ b/innobase/sync/sync0sync.c	2009-10-22 15:18:44.000000000 +0900
@@ -1,8 +1,31 @@
+/*****************************************************************************
+
+Copyright (c) 1995, 2009, Innobase Oy. All Rights Reserved.
+Copyright (c) 2008, Google Inc.
+
+Portions of this file contain modifications contributed and copyrighted by
+Google, Inc. Those modifications are gratefully acknowledged and are described
+briefly in the InnoDB documentation. The contributions by Google are
+incorporated with their permission, and subject to the conditions contained in
+the file COPYING.Google.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+Place, Suite 330, Boston, MA 02111-1307 USA
+
+*****************************************************************************/
+
 /******************************************************
 Mutex, the basic synchronization primitive
 
-(c) 1995 Innobase Oy
-
 Created 9/5/1995 Heikki Tuuri
 *******************************************************/
 
@@ -140,17 +163,12 @@
 
 ulint	sync_dummy			= 0;
 
-/* The number of system calls made in this module. Intended for performance
-monitoring. */
-
-ulint	mutex_system_call_count		= 0;
-
 /* Number of spin waits on mutexes: for performance monitoring */
 
-ulint	mutex_spin_round_count		= 0;
-ulint	mutex_spin_wait_count		= 0;
-ulint	mutex_os_wait_count		= 0;
-ulint	mutex_exit_count		= 0;
+ib_longlong	mutex_spin_round_count		= 0;
+ib_longlong	mutex_spin_wait_count		= 0;
+ib_longlong	mutex_os_wait_count		= 0;
+ib_longlong	mutex_exit_count		= 0;
 
 /* The global array of wait cells for implementation of the database's own
 mutexes and read-write locks */
@@ -240,6 +258,8 @@
 {
 #if defined(_WIN32) && defined(UNIV_CAN_USE_X86_ASSEMBLER)
 	mutex_reset_lock_word(mutex);
+#elif defined(HAVE_ATOMIC_BUILTINS)
+	mutex_reset_lock_word(mutex);
 #else	
 	os_fast_mutex_init(&(mutex->os_fast_mutex));
 	mutex->lock_word = 0;
@@ -325,7 +345,9 @@
 
 	os_event_free(mutex->event);
 
-#if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER) 
+#if defined(_WIN32) && defined(UNIV_CAN_USE_X86_ASSEMBLER)
+#elif defined(HAVE_ATOMIC_BUILTINS)
+#else
 	os_fast_mutex_free(&(mutex->os_fast_mutex));
 #endif
 	/* If we free the mutex protecting the mutex list (freeing is
@@ -421,6 +443,12 @@
 #endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
   ut_ad(mutex);
 
+	/* This update is not thread safe, but we don't mind if the count
+	isn't exact. Moved out of ifdef that follows because we are willing
+	to sacrifice the cost of counting this as the data is valuable.
+	Count the number of calls to mutex_spin_wait. */
+	mutex_spin_wait_count++;
+
 mutex_loop:
 
   i = 0;
@@ -433,7 +461,6 @@
 
 spin_loop:
 #if defined UNIV_DEBUG && !defined UNIV_HOTBACKUP
-  mutex_spin_wait_count++;
   mutex->count_spin_loop++;
 #endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
 
@@ -502,8 +529,6 @@
   sync_array_reserve_cell(sync_primary_wait_array, mutex,
                           SYNC_MUTEX, file_name, line, &index);
 
-  mutex_system_call_count++;
-
   /* The memory order of the array reservation and the change in the
   waiters field is important: when we suspend a thread, we first
   reserve the cell and then set waiters field to 1. When threads are
@@ -551,7 +576,6 @@
             mutex->cfile_name, (ulong) mutex->cline, (ulong) i);
 #endif
 
-  mutex_system_call_count++;
   mutex_os_wait_count++;
 
 #ifndef UNIV_HOTBACKUP
@@ -1368,20 +1392,31 @@
 	FILE*	file)		/* in: file where to print */
 {
 #ifdef UNIV_SYNC_DEBUG
-	fprintf(stderr, "Mutex exits %lu, rws exits %lu, rwx exits %lu\n",
+	fprintf(file, "Mutex exits %llu, rws exits %llu, rwx exits %llu\n",
 		mutex_exit_count, rw_s_exit_count, rw_x_exit_count);
 #endif
 
 	fprintf(file,
-"Mutex spin waits %lu, rounds %lu, OS waits %lu\n"
-"RW-shared spins %lu, OS waits %lu; RW-excl spins %lu, OS waits %lu\n",
-			(ulong) mutex_spin_wait_count,
-		        (ulong) mutex_spin_round_count,
-			(ulong) mutex_os_wait_count,
-			(ulong) rw_s_spin_wait_count,
-		        (ulong) rw_s_os_wait_count,
-			(ulong) rw_x_spin_wait_count,
-		        (ulong) rw_x_os_wait_count);
+		"Mutex spin waits %llu, rounds %llu, OS waits %llu\n"
+		"RW-shared spins %llu, OS waits %llu;"
+		" RW-excl spins %llu, OS waits %llu\n",
+		mutex_spin_wait_count,
+		mutex_spin_round_count,
+		mutex_os_wait_count,
+		rw_s_spin_wait_count,
+		rw_s_os_wait_count,
+		rw_x_spin_wait_count,
+		rw_x_os_wait_count);
+
+	fprintf(file,
+		"Spin rounds per wait: %.2f mutex, %.2f RW-shared, "
+		"%.2f RW-excl\n",
+		(double) mutex_spin_round_count /
+		(mutex_spin_wait_count ? mutex_spin_wait_count : 1),
+		(double) rw_s_spin_round_count /
+		(rw_s_spin_wait_count ? rw_s_spin_wait_count : 1),
+		(double) rw_x_spin_round_count /
+		(rw_x_spin_wait_count ? rw_x_spin_wait_count : 1));
 }
 
 /***********************************************************************
diff -ruN a/patch_info/innodb_rw_lock.info b/patch_info/innodb_rw_lock.info
--- /dev/null	1970-01-01 09:00:00.000000000 +0900
+++ b/patch_info/innodb_rw_lock.info	2009-10-22 15:18:30.000000000 +0900
@@ -0,0 +1,6 @@
+File=innodb_rw_lock.patch
+Name=Fix of InnoDB rw_locks ported from InnoDB Plugin
+Version=1.0
+Author=InnoBase Oy.
+License=GPL
+Comment=