real sync with patches rev12

buf/buf0buf.c

@@ -3595,13 +3595,15 @@ buf_print_io(
 	buf_pool_mutex_enter();
 
 	fprintf(file,
-		"Buffer pool size   %lu\n"
+		"Buffer pool size        %lu\n"
-		"Free buffers       %lu\n"
+		"Buffer pool size, bytes %lu\n"
-		"Database pages     %lu\n"
+		"Free buffers            %lu\n"
-		"Modified db pages  %lu\n"
+		"Database pages          %lu\n"
+		"Modified db pages       %lu\n"
 		"Pending reads %lu\n"
 		"Pending writes: LRU %lu, flush list %lu, single page %lu\n",
 		(ulong) size,
+		(ulong) size * UNIV_PAGE_SIZE,
 		(ulong) UT_LIST_GET_LEN(buf_pool->free),
 		(ulong) UT_LIST_GET_LEN(buf_pool->LRU),
 		(ulong) UT_LIST_GET_LEN(buf_pool->flush_list),

buf/buf0rea.c

@@ -187,6 +187,10 @@ buf_read_ahead_random(
 	ulint		i;
 	ulint		buf_read_ahead_random_area;
 
+	if (!(srv_read_ahead & 1)) {
+		return(0);
+	}
+
 	if (srv_startup_is_before_trx_rollback_phase) {
 		/* No read-ahead to avoid thread deadlocks */
 		return(0);
@@ -412,6 +416,10 @@ buf_read_ahead_linear(
 	const ulint	buf_read_ahead_linear_area
 		= BUF_READ_AHEAD_LINEAR_AREA;
 
+ 	if (!(srv_read_ahead & 2)) {
+ 		return(0);
+ 	}
+ 
 	if (UNIV_UNLIKELY(srv_startup_is_before_trx_rollback_phase)) {
 		/* No read-ahead to avoid thread deadlocks */
 		return(0);

fil/fil0fil.c

@@ -4816,3 +4816,30 @@ fil_page_get_type(
 
 	return(mach_read_from_2(page + FIL_PAGE_TYPE));
 }
+
+/*************************************************************************
+Return local hash table informations. */
+
+ulint
+fil_system_hash_cells(void)
+/*=======================*/
+{
+       if (fil_system) {
+               return (fil_system->spaces->n_cells
+                       + fil_system->name_hash->n_cells);
+       } else {
+               return 0;
+       }
+}
+
+ulint
+fil_system_hash_nodes(void)
+/*=======================*/
+{
+       if (fil_system) {
+               return (UT_LIST_GET_LEN(fil_system->space_list)
+                       * (sizeof(fil_space_t) + MEM_BLOCK_HEADER_SIZE));
+       } else {
+               return 0;
+       }
+}

handler/ha_innodb.cc

@@ -137,6 +137,7 @@ static long innobase_mirrored_log_groups, innobase_log_files_in_group,
 	innobase_force_recovery, innobase_open_files,
 	innobase_autoinc_lock_mode;
 
+static unsigned long innobase_read_io_threads, innobase_write_io_threads;
 static long long innobase_buffer_pool_size, innobase_log_file_size;
 
 /* The default values for the following char* start-up parameters
@@ -2066,6 +2067,8 @@ innobase_init(
 	srv_mem_pool_size = (ulint) innobase_additional_mem_pool_size;
 
 	srv_n_file_io_threads = (ulint) innobase_file_io_threads;
+	srv_n_read_io_threads = (ulint) innobase_read_io_threads;
+	srv_n_write_io_threads = (ulint) innobase_write_io_threads;
 
 	srv_force_recovery = (ulint) innobase_force_recovery;
 
@@ -9406,6 +9409,16 @@ static MYSQL_SYSVAR_BOOL(locks_unsafe_for_binlog, innobase_locks_unsafe_for_binl
   "Force InnoDB to not use next-key locking, to use only row-level locking.",
   NULL, NULL, FALSE);
 
+static MYSQL_SYSVAR_ULONG(show_verbose_locks, srv_show_verbose_locks,
+  PLUGIN_VAR_OPCMDARG,
+  "Whether to show records locked in SHOW INNODB STATUS.",
+  NULL, NULL, 0, 0, 1, 0);
+
+static MYSQL_SYSVAR_ULONG(show_locks_held, srv_show_locks_held,
+  PLUGIN_VAR_RQCMDARG,
+  "Number of locks held to print for each InnoDB transaction in SHOW INNODB STATUS.",
+  NULL, NULL, 10, 0, 1000, 0);
+
 #ifdef UNIV_LOG_ARCHIVE
 static MYSQL_SYSVAR_STR(log_arch_dir, innobase_log_arch_dir,
   PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
@@ -9556,7 +9569,32 @@ static MYSQL_SYSVAR_LONG(autoinc_lock_mode, innobase_autoinc_lock_mode,
 
 static MYSQL_SYSVAR_STR(version, innodb_version_str,
   PLUGIN_VAR_NOCMDOPT | PLUGIN_VAR_READONLY,
-  "InnoDB version", NULL, NULL, INNODB_VERSION_STR);
+  "Percona-InnoDB-plugin version", NULL, NULL, INNODB_VERSION_STR);
+
+static MYSQL_SYSVAR_ULONG(io_capacity, srv_io_capacity,
+  PLUGIN_VAR_RQCMDARG,
+  "Number of IO operations per second the server can do. Tunes background IO rate.",
+  NULL, NULL, 100, 100, 999999999, 0);
+
+static MYSQL_SYSVAR_ULONG(read_ahead, srv_read_ahead,
+  PLUGIN_VAR_RQCMDARG,
+  "Enable/Diasable read aheads bit0:random bit1:linear",
+  NULL, NULL, 3, 0, 3, 0);
+
+static MYSQL_SYSVAR_ULONG(adaptive_checkpoint, srv_adaptive_checkpoint,
+  PLUGIN_VAR_RQCMDARG,
+  "Enable/Disable flushing along modified age 0:disable 1:enable",
+  NULL, NULL, 0, 0, 1, 0);
+
+static MYSQL_SYSVAR_ULONG(read_io_threads, innobase_read_io_threads,
+  PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
+  "Number of background read I/O threads in InnoDB.",
+  NULL, NULL, 1, 1, 64, 0);
+
+static MYSQL_SYSVAR_ULONG(write_io_threads, innobase_write_io_threads,
+  PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
+  "Number of background write I/O threads in InnoDB.",
+  NULL, NULL, 1, 1, 64, 0);
 
 static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(additional_mem_pool_size),
@@ -9603,7 +9641,14 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(thread_concurrency),
   MYSQL_SYSVAR(thread_sleep_delay),
   MYSQL_SYSVAR(autoinc_lock_mode),
+  MYSQL_SYSVAR(show_verbose_locks),
+  MYSQL_SYSVAR(show_locks_held),
   MYSQL_SYSVAR(version),
+  MYSQL_SYSVAR(io_capacity),
+  MYSQL_SYSVAR(read_ahead),
+  MYSQL_SYSVAR(adaptive_checkpoint),
+  MYSQL_SYSVAR(read_io_threads),
+  MYSQL_SYSVAR(write_io_threads),
   NULL
 };
 

handler/innodb_patch_info.h

@@ -24,5 +24,8 @@ struct innodb_enhancement {
        const char *link;
 }innodb_enhancements[] = {
 {"innodb_show_enhancements","I_S.PERCONA_INNODB_ENHANCEMENTS","","http://www.percona.com/docs/wiki/percona-innodb-plugin:innodb_show_enhancements"},
-{NULL, NULL, NULL, NULL, NULL, NULL}
+{"innodb_show_status","Improvements to SHOW INNODB STATUS","Memory information and lock info fixes","http://www.percona.com/docs/wiki/percona-innodb-plugin:innodb_show_status"},
+{"innodb_io","Improvements to InnoDB IO","","http://www.percona.com/docs/wiki/percona-innodb-plugin:innodb_io"},
+{"innodb_rw_lock","InnoDB RW-lock fixes","Useful for 8+ cores SMP systems","http://www.percona.com/docs/wiki/percona-innodb-plugin:innodb_rw_lock"},
+{NULL, NULL, NULL, NULL}
 };

include/fil0fil.h

@@ -696,6 +696,16 @@ fil_page_get_type(
 				return value not defined */
 	const byte*	page);	/* in: file page */
 
+/*************************************************************************
+Return local hash table informations. */
+
+ulint
+fil_system_hash_cells(void);
+/*========================*/
+
+ulint
+fil_system_hash_nodes(void);
+/*========================*/
 
 typedef	struct fil_space_struct	fil_space_t;
 

include/os0file.h

@@ -561,8 +561,10 @@ os_aio_init(
 /*========*/
 	ulint	n,		/* in: maximum number of pending aio operations
 				allowed; n must be divisible by n_segments */
-	ulint	n_segments,	/* in: combined number of segments in the four
+//	ulint	n_segments,	/* in: combined number of segments in the four
-				first aio arrays; must be >= 4 */
+//				first aio arrays; must be >= 4 */
+	ulint	n_read_threads,  /* n_segments == 2 + n_read_threads + n_write_threads */
+	ulint	n_write_threads, /**/
 	ulint	n_slots_sync);	/* in: number of slots in the sync aio array */
 /***********************************************************************
 Requests an asynchronous i/o operation. */

include/srv0srv.h

@@ -90,6 +90,9 @@ extern ulint	srv_log_file_size;
 extern ulint	srv_log_buffer_size;
 extern ulong	srv_flush_log_at_trx_commit;
 
+extern ulint    srv_show_locks_held;
+extern ulint    srv_show_verbose_locks;
+
 /* The sort order table of the MySQL latin1_swedish_ci character set
 collation */
 extern const byte*	srv_latin1_ordering;
@@ -100,6 +103,8 @@ extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 
 extern ulint	srv_n_file_io_threads;
+extern ulint	srv_n_read_io_threads;
+extern ulint	srv_n_write_io_threads;
 
 #ifdef UNIV_LOG_ARCHIVE
 extern ibool	srv_log_archive_on;
@@ -144,6 +149,11 @@ extern ulong	srv_max_buf_pool_modified_pct;
 extern ulong	srv_max_purge_lag;
 
 extern ulong	srv_replication_delay;
+
+extern ulint	srv_io_capacity;
+extern ulint	srv_read_ahead;
+extern ulint	srv_adaptive_checkpoint;
+
 /*-------------------------------------------*/
 
 extern ulint	srv_n_rows_inserted;

include/sync0rw.h

@@ -328,7 +328,17 @@ rw_lock_get_x_lock_count(
 Accessor functions for rw lock. */
 UNIV_INLINE
 ulint
-rw_lock_get_waiters(
+rw_lock_get_s_waiters(
+/*==================*/
+	rw_lock_t*	lock);
+UNIV_INLINE
+ulint
+rw_lock_get_x_waiters(
+/*==================*/
+	rw_lock_t*	lock);
+UNIV_INLINE
+ulint
+rw_lock_get_wx_waiters(
 /*================*/
 	rw_lock_t*	lock);
 UNIV_INLINE
@@ -412,6 +422,11 @@ rw_lock_debug_print(
 	rw_lock_debug_t*	info);	/* in: debug struct */
 #endif /* UNIV_SYNC_DEBUG */
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+/* This value means NOT_LOCKED */
+#define RW_LOCK_BIAS		0x00100000
+#endif
+
 /* NOTE! The structure appears here only for the compiler to know its size.
 Do not use its fields directly! The structure used in the spin lock
 implementation of a read-write lock. Several threads may have a shared lock
@@ -421,9 +436,9 @@ blocked by readers, a writer may queue for the lock by setting the writer
 field. Then no new readers are allowed in. */
 
 struct rw_lock_struct {
-	os_event_t	event;	/* Used by sync0arr.c for thread queueing */
+			/* Used by sync0arr.c for thread queueing */
-
+	os_event_t	s_event;	/* Used for s_lock */
-#ifdef __WIN__
+	os_event_t	x_event;	/* Used for x_lock */
 	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
@@ -431,30 +446,34 @@ struct rw_lock_struct {
 				thread will be the next one to proceed
 				once the current the event gets
 				signalled. See LEMMA 2 in sync0sync.c */
+
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	volatile lint	lock_word;	/* Used by using atomic builtin */
 #endif
 
-	ulint	reader_count;	/* Number of readers who have locked this
+	volatile 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
+	volatile 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;
+	volatile os_thread_id_t	writer_thread;
 				/* Thread id of a possible writer thread */
-	ulint	writer_count;	/* Number of times the same thread has
+	volatile ulint	writer_count;	/* Number of times the same thread has
 				recursively locked the lock in the exclusive
 				mode */
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_t	mutex;		/* The mutex protecting rw_lock_struct */
+#endif
 	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
+	volatile ulint	s_waiters;	/* 1: there are waiters (s_lock) */
-				waiters (readers or writers) in the global
+	volatile ulint	x_waiters;	/* 1: there are waiters (x_lock) */
-				wait array, waiting for this rw_lock.
+	volatile ulint	wait_ex_waiters;	/* 1: there are waiters (wait_ex) */
-				Otherwise, == 0. */
 	UT_LIST_NODE_T(rw_lock_t) list;
 				/* All allocated rw locks are put into a
 				list */
@@ -467,7 +486,7 @@ struct rw_lock_struct {
 	const char*	cfile_name;/* File name where lock created */
 	const char*	last_s_file_name;/* File name where last s-locked */
 	const char*	last_x_file_name;/* File name where last x-locked */
-	ibool		writer_is_wait_ex;
+	volatile 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

include/sync0rw.ic

@@ -47,20 +47,52 @@ rw_lock_remove_debug_info(
 Accessor functions for rw lock. */
 UNIV_INLINE
 ulint
-rw_lock_get_waiters(
+rw_lock_get_s_waiters(
 /*================*/
 	rw_lock_t*	lock)
 {
-	return(lock->waiters);
+	return(lock->s_waiters);
+}
+UNIV_INLINE
+ulint
+rw_lock_get_x_waiters(
+/*================*/
+	rw_lock_t*	lock)
+{
+	return(lock->x_waiters);
+}
+UNIV_INLINE
+ulint
+rw_lock_get_wx_waiters(
+/*================*/
+	rw_lock_t*      lock)
+{
+	return(lock->wait_ex_waiters);
 }
 UNIV_INLINE
 void
-rw_lock_set_waiters(
+rw_lock_set_s_waiters(
-/*================*/
 	rw_lock_t*	lock,
 	ulint		flag)
 {
-	lock->waiters = flag;
+	lock->s_waiters = flag;
+}
+UNIV_INLINE
+void
+rw_lock_set_x_waiters(
+	rw_lock_t*	lock,
+	ulint		flag)
+{
+	lock->x_waiters = flag;
+}
+UNIV_INLINE
+void
+rw_lock_set_wx_waiters(
+/*================*/
+	rw_lock_t*      lock,
+	ulint           flag)
+{
+	lock->wait_ex_waiters = flag;
 }
 UNIV_INLINE
 ulint
@@ -68,7 +100,19 @@ rw_lock_get_writer(
 /*===============*/
 	rw_lock_t*	lock)
 {
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (lock->writer == RW_LOCK_NOT_LOCKED) {
+		return(RW_LOCK_NOT_LOCKED);
+	}
+
+	if (lock->writer_is_wait_ex) {
+		return(RW_LOCK_WAIT_EX);
+	} else {
+		return(RW_LOCK_EX);
+	}
+#else
 	return(lock->writer);
+#endif
 }
 UNIV_INLINE
 void
@@ -96,6 +140,7 @@ rw_lock_set_reader_count(
 {
 	lock->reader_count = count;
 }
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 UNIV_INLINE
 mutex_t*
 rw_lock_get_mutex(
@@ -104,6 +149,7 @@ rw_lock_get_mutex(
 {
 	return(&(lock->mutex));
 }
+#endif
 
 /**********************************************************************
 Returns the value of writer_count for the lock. Does not reserve the lock
@@ -133,13 +179,27 @@ rw_lock_s_lock_low(
 	const char*	file_name, /* in: file name where lock requested */
 	ulint		line)	/* in: line where requested */
 {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	ut_ad(mutex_own(rw_lock_get_mutex(lock)));
+#endif
 
 	/* Check if the writer field is free */
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (UNIV_LIKELY(rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED)) {
+		/* try s-lock */
+		if(__sync_sub_and_fetch(&(lock->lock_word),1) <= 0) {
+			/* fail */
+			__sync_fetch_and_add(&(lock->lock_word),1);
+			return(FALSE);  /* locking did not succeed */
+		}
+		/* success */
+		__sync_fetch_and_add(&(lock->reader_count),1);
+#else
 	if (UNIV_LIKELY(lock->writer == RW_LOCK_NOT_LOCKED)) {
 		/* Set the shared lock by incrementing the reader count */
 		lock->reader_count++;
+#endif
 
 #ifdef UNIV_SYNC_DEBUG
 		rw_lock_add_debug_info(lock, pass, RW_LOCK_SHARED, file_name,
@@ -166,11 +226,15 @@ rw_lock_s_lock_direct(
 	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_writer(lock) == RW_LOCK_NOT_LOCKED);
 	ut_ad(rw_lock_get_reader_count(lock) == 0);
 
 	/* Set the shared lock by incrementing the reader count */
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	__sync_fetch_and_add(&(lock->reader_count),1);
+#else
 	lock->reader_count++;
+#endif
 
 	lock->last_s_file_name = file_name;
 	lock->last_s_line = line;
@@ -198,7 +262,11 @@ rw_lock_x_lock_direct(
 
 	rw_lock_set_writer(lock, RW_LOCK_EX);
 	lock->writer_thread = os_thread_get_curr_id();
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	__sync_fetch_and_add(&(lock->writer_count),1);
+#else
 	lock->writer_count++;
+#endif
 	lock->pass = 0;
 
 	lock->last_x_file_name = file_name;
@@ -240,15 +308,21 @@ rw_lock_s_lock_func(
 	ut_ad(!rw_lock_own(lock, RW_LOCK_SHARED)); /* see NOTE above */
 #endif /* UNIV_SYNC_DEBUG */
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(rw_lock_get_mutex(lock));
+#endif
 
 	if (UNIV_LIKELY(rw_lock_s_lock_low(lock, pass, file_name, line))) {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 		return; /* Success */
 	} else {
 		/* Did not succeed, try spin wait */
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 		rw_lock_s_lock_spin(lock, pass, file_name, line);
 
@@ -271,11 +345,23 @@ rw_lock_s_lock_func_nowait(
 {
 	ibool	success	= FALSE;
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
+		/* try s-lock */
+		if(__sync_sub_and_fetch(&(lock->lock_word),1) <= 0) {
+			/* fail */
+			__sync_fetch_and_add(&(lock->lock_word),1);
+			return(FALSE);	/* locking did not succeed */
+		}
+		/* success */
+		__sync_fetch_and_add(&(lock->reader_count),1);
+#else
 	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++;
+#endif
 
 #ifdef UNIV_SYNC_DEBUG
 		rw_lock_add_debug_info(lock, 0, RW_LOCK_SHARED, file_name,
@@ -288,7 +374,9 @@ rw_lock_s_lock_func_nowait(
 		success = TRUE;
 	}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 	return(success);
 }
@@ -308,6 +396,55 @@ rw_lock_x_lock_func_nowait(
 {
 	ibool		success		= FALSE;
 	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if ((lock->lock_word == RW_LOCK_BIAS)
+			&& rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
+		/* try x-lock */
+		if(__sync_sub_and_fetch(&(lock->lock_word),
+				RW_LOCK_BIAS) == 0) {
+			/* success */
+			/* try to lock writer */
+			if(__sync_lock_test_and_set(&(lock->writer),RW_LOCK_EX)
+					== RW_LOCK_NOT_LOCKED) {
+				/* success */
+				lock->writer_thread = curr_thread;
+				lock->pass = 0;
+				lock->writer_is_wait_ex = FALSE;
+				/* next function may work as memory barrier */
+			relock:
+				__sync_fetch_and_add(&(lock->writer_count),1);
+
+#ifdef UNIV_SYNC_DEBUG
+				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;
+
+				ut_ad(rw_lock_validate(lock));
+
+				return(TRUE);
+			} else {
+				/* x-unlock */
+				__sync_fetch_and_add(&(lock->lock_word),
+					RW_LOCK_BIAS);
+			}
+		} else {
+			/* fail (x-lock) */
+			__sync_fetch_and_add(&(lock->lock_word),RW_LOCK_BIAS);
+		}
+	}
+
+	if (lock->pass == 0
+			&& os_thread_eq(lock->writer_thread, curr_thread)
+			&& rw_lock_get_writer(lock) == RW_LOCK_EX) {
+		goto relock;
+	}
+
+	ut_ad(rw_lock_validate(lock));
+
+	return(FALSE);
+#else
 	mutex_enter(rw_lock_get_mutex(lock));
 
 	if (UNIV_UNLIKELY(rw_lock_get_reader_count(lock) != 0)) {
@@ -338,6 +475,7 @@ relock:
 	ut_ad(rw_lock_validate(lock));
 
 	return(success);
+#endif
 }
 
 /**********************************************************************
@@ -353,16 +491,33 @@ rw_lock_s_unlock_func(
 #endif
 	)
 {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_t*	mutex	= &(lock->mutex);
-	ibool		sg	= FALSE;
+#endif
+	ibool		x_sg	= FALSE;
+	ibool		wx_sg	= FALSE;
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	ibool		last	= FALSE;
+#endif
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	/* Acquire the mutex protecting the rw-lock fields */
 	mutex_enter(mutex);
+#endif
 
 	/* Reset the shared lock by decrementing the reader count */
 
 	ut_a(lock->reader_count > 0);
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	/* unlock lock_word */
+	__sync_fetch_and_add(&(lock->lock_word),1);
+
+	if(__sync_sub_and_fetch(&(lock->reader_count),1) == 0) {
+		last = TRUE;
+	}
+#else
 	lock->reader_count--;
+#endif
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, pass, RW_LOCK_SHARED);
@@ -371,20 +526,36 @@ rw_lock_s_unlock_func(
 	/* If there may be waiters and this was the last s-lock,
 	signal the object */
 
-	if (UNIV_UNLIKELY(lock->waiters)
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (UNIV_UNLIKELY(last && lock->wait_ex_waiters)) {
+#else
+	if (UNIV_UNLIKELY(lock->wait_ex_waiters)
 	    && lock->reader_count == 0) {
-		sg = TRUE;
+#endif
+		wx_sg = TRUE;
 
-		rw_lock_set_waiters(lock, 0);
+		rw_lock_set_wx_waiters(lock, 0);
+	}
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	else if (UNIV_UNLIKELY(last && lock->x_waiters)) {
+#else
+	else if (UNIV_UNLIKELY(lock->x_waiters)
+		 && lock->reader_count == 0) {
+#endif
+		x_sg = TRUE;
+
+		rw_lock_set_x_waiters(lock, 0);
 	}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(mutex);
-
-	if (UNIV_UNLIKELY(sg)) {
-#ifdef __WIN__
-		os_event_set(lock->wait_ex_event);
 #endif
-		os_event_set(lock->event);
+
+	if (UNIV_UNLIKELY(wx_sg)) {
+		os_event_set(lock->wait_ex_event);
+		sync_array_object_signalled(sync_primary_wait_array);
+	} else if (UNIV_UNLIKELY(x_sg)) {
+		os_event_set(lock->x_event);
 		sync_array_object_signalled(sync_primary_wait_array);
 	}
 
@@ -408,13 +579,19 @@ rw_lock_s_unlock_direct(
 
 	ut_ad(lock->reader_count > 0);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	__sync_sub_and_fetch(&(lock->reader_count),1);
+#else
 	lock->reader_count--;
+#endif
 
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_remove_debug_info(lock, 0, RW_LOCK_SHARED);
 #endif
 
-	ut_ad(!lock->waiters);
+	ut_ad(!lock->s_waiters);
+	ut_ad(!lock->x_waiters);
+	ut_ad(!lock->wait_ex_waiters);
 	ut_ad(rw_lock_validate(lock));
 #ifdef UNIV_SYNC_PERF_STAT
 	rw_s_exit_count++;
@@ -434,41 +611,81 @@ rw_lock_x_unlock_func(
 #endif
 	)
 {
-	ibool	sg	= FALSE;
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	ibool	last	= FALSE;
+#endif
+	ibool	s_sg	= FALSE;
+	ibool	x_sg	= FALSE;
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	/* Acquire the mutex protecting the rw-lock fields */
 	mutex_enter(&(lock->mutex));
+#endif
 
 	/* Reset the exclusive lock if this thread no longer has an x-mode
 	lock */
 
 	ut_ad(lock->writer_count > 0);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if(__sync_sub_and_fetch(&(lock->writer_count),1) == 0) {
+		last = TRUE;
+	}
+
+	if (last) {
+		/* unlock lock_word */
+		__sync_fetch_and_add(&(lock->lock_word),RW_LOCK_BIAS);
+
+		/* FIXME: It is a value of bad manners for pthread.
+			  But we shouldn't keep an ID of not-owner. */
+		lock->writer_thread = -1;
+
+		/* atomic operation may be safer about memory order. */
+		rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
+		__sync_synchronize();
+	}
+#else
 	lock->writer_count--;
 
 	if (lock->writer_count == 0) {
 		rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
 	}
+#endif
 
 #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)
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
-	    && lock->writer_count == 0) {
+	if (last) {
-
+#else
-		sg = TRUE;
+	if (lock->writer_count == 0) {
-		rw_lock_set_waiters(lock, 0);
+#endif
+		if(lock->s_waiters){
+			s_sg = TRUE;
+			rw_lock_set_s_waiters(lock, 0);
+		}
+		if(lock->x_waiters){
+			x_sg = TRUE;
+			rw_lock_set_x_waiters(lock, 0);
+		}
 	}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(&(lock->mutex));
+#endif
 
-	if (UNIV_UNLIKELY(sg)) {
+	if (UNIV_UNLIKELY(s_sg)) {
+		os_event_set(lock->s_event);
+		sync_array_object_signalled(sync_primary_wait_array);
+	}
+	if (UNIV_UNLIKELY(x_sg)) {
 #ifdef __WIN__
+		/* I doubt the necessity of it. */
 		os_event_set(lock->wait_ex_event);
 #endif
-		os_event_set(lock->event);
+		os_event_set(lock->x_event);
 		sync_array_object_signalled(sync_primary_wait_array);
 	}
 
@@ -493,9 +710,13 @@ rw_lock_x_unlock_direct(
 
 	ut_ad(lock->writer_count > 0);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if(__sync_sub_and_fetch(&(lock->writer_count),1) == 0) {
+#else
 	lock->writer_count--;
 
 	if (lock->writer_count == 0) {
+#endif
 		rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
 	}
 
@@ -503,7 +724,9 @@ rw_lock_x_unlock_direct(
 	rw_lock_remove_debug_info(lock, 0, RW_LOCK_EX);
 #endif
 
-	ut_ad(!lock->waiters);
+	ut_ad(!lock->s_waiters);
+	ut_ad(!lock->x_waiters);
+	ut_ad(!lock->wait_ex_waiters);
 	ut_ad(rw_lock_validate(lock));
 
 #ifdef UNIV_SYNC_PERF_STAT

include/thr0loc.h

@@ -60,6 +60,17 @@ thr_local_get_in_ibuf_field(void);
 /*=============================*/
 			/* out: pointer to the in_ibuf field */
 
+/*************************************************************************
+Return local hash table informations. */
+
+ulint
+thr_local_hash_cells(void);
+/*=======================*/
+
+ulint
+thr_local_hash_nodes(void);
+/*=======================*/
+
 #ifndef UNIV_NONINL
 #include "thr0loc.ic"
 #endif

include/univ.i

@@ -12,6 +12,7 @@ Created 1/20/1994 Heikki Tuuri
 #define INNODB_VERSION_MAJOR	1
 #define INNODB_VERSION_MINOR	0
 #define INNODB_VERSION_BUGFIX	2
+#define PERCONA_INNODB_VERSION	1
 
 /* The following is the InnoDB version as shown in
 SELECT plugin_version FROM information_schema.plugins;
@@ -23,13 +24,14 @@ component, i.e. we show M.N.P as M.N */
 	(INNODB_VERSION_MAJOR << 8 | INNODB_VERSION_MINOR)
 
 /* auxiliary macros to help creating the version as string */
-#define __INNODB_VERSION(a, b, c)	(#a "." #b "." #c)
+#define __INNODB_VERSION(a, b, c, d)	(#a "." #b "." #c "-" #d)
-#define _INNODB_VERSION(a, b, c)	__INNODB_VERSION(a, b, c)
+#define _INNODB_VERSION(a, b, c, d)	__INNODB_VERSION(a, b, c, d)
 
 #define INNODB_VERSION_STR			\
 	_INNODB_VERSION(INNODB_VERSION_MAJOR,	\
 			INNODB_VERSION_MINOR,	\
-			INNODB_VERSION_BUGFIX)
+			INNODB_VERSION_BUGFIX,	\
+			PERCONA_INNODB_VERSION)
 
 #ifdef MYSQL_DYNAMIC_PLUGIN
 /* In the dynamic plugin, redefine some externally visible symbols

lock/lock0lock.c

@@ -4294,32 +4294,32 @@ lock_rec_print(
 
 	putc('\n', file);
 
-	block = buf_page_try_get(space, page_no, &mtr);
+	if ( srv_show_verbose_locks ) {
+		block = buf_page_try_get(space, page_no, &mtr);
+		if (block) {
+			for (i = 0; i < lock_rec_get_n_bits(lock); i++) {
 
-	if (block) {
+				if (lock_rec_get_nth_bit(lock, i)) {
-		for (i = 0; i < lock_rec_get_n_bits(lock); i++) {
 
-			if (lock_rec_get_nth_bit(lock, i)) {
+					const rec_t*	rec
+						= page_find_rec_with_heap_no(
+								buf_block_get_frame(block), i);
+					offsets = rec_get_offsets(
+							rec, lock->index, offsets,
+							ULINT_UNDEFINED, &heap);
 
-				const rec_t*	rec
+					fprintf(file, "Record lock, heap no %lu ",
-					= page_find_rec_with_heap_no(
+							(ulong) i);
-						buf_block_get_frame(block), i);
+					rec_print_new(file, rec, offsets);
-				offsets = rec_get_offsets(
+					putc('\n', file);
-					rec, lock->index, offsets,
+				}
-					ULINT_UNDEFINED, &heap);
+			}
-
+		} else {
-				fprintf(file, "Record lock, heap no %lu ",
+			for (i = 0; i < lock_rec_get_n_bits(lock); i++) {
-					(ulong) i);
+				fprintf(file, "Record lock, heap no %lu\n", (ulong) i);
-				rec_print_new(file, rec, offsets);
-				putc('\n', file);
 			}
 		}
-	} else {
-		for (i = 0; i < lock_rec_get_n_bits(lock); i++) {
-			fprintf(file, "Record lock, heap no %lu\n", (ulong) i);
-		}
 	}
-
 	mtr_commit(&mtr);
 	if (UNIV_LIKELY_NULL(heap)) {
 		mem_heap_free(heap);
@@ -4498,7 +4498,7 @@ loop:
 		}
 	}
 
-	if (!srv_print_innodb_lock_monitor) {
+        if (!srv_print_innodb_lock_monitor && !srv_show_locks_held) {
 		nth_trx++;
 		goto loop;
 	}
@@ -4557,8 +4557,8 @@ loop:
 
 	nth_lock++;
 
-	if (nth_lock >= 10) {
+	if (nth_lock >= srv_show_locks_held) {
-		fputs("10 LOCKS PRINTED FOR THIS TRX:"
+		fputs("TOO LOCKS PRINTED FOR THIS TRX:"
 		      " SUPPRESSING FURTHER PRINTS\n",
 		      file);
 

log/log0log.c

@@ -3258,6 +3258,15 @@ log_print(
 		log_sys->flushed_to_disk_lsn,
 		log_sys->last_checkpoint_lsn);
 
+	fprintf(file,
+		"Max checkpoint age  %lu\n"
+		"Modified age        %lu\n"
+		"Checkpoint age      %lu\n",
+			(ulong) log_sys->max_checkpoint_age,
+			(ulong) (log_sys->lsn -
+					log_buf_pool_get_oldest_modification()),
+			(ulong) (log_sys->lsn - log_sys->last_checkpoint_lsn));
+
 	current_time = time(NULL);
 
 	time_elapsed = 0.001 + difftime(current_time,

os/os0file.c

@@ -2920,8 +2920,10 @@ os_aio_init(
 /*========*/
 	ulint	n,		/* in: maximum number of pending aio operations
 				allowed; n must be divisible by n_segments */
-	ulint	n_segments,	/* in: combined number of segments in the four
+//	ulint	n_segments,	/* in: combined number of segments in the four
-				first aio arrays; must be >= 4 */
+//				first aio arrays; must be >= 4 */
+	ulint	n_read_threads,  /* n_segments == 2 + n_read_threads + n_write_threads*/
+	ulint	n_write_threads, /**/
 	ulint	n_slots_sync)	/* in: number of slots in the sync aio array */
 {
 	ulint	n_read_segs;
@@ -2929,6 +2931,8 @@ os_aio_init(
 	ulint	n_per_seg;
 	ulint	i;
 
+ 	ulint	n_segments = 2 + n_read_threads + n_write_threads;
+ 
 	ut_ad(n % n_segments == 0);
 	ut_ad(n_segments >= 4);
 
@@ -2939,8 +2943,8 @@ os_aio_init(
 	}
 
 	n_per_seg = n / n_segments;
-	n_write_segs = (n_segments - 2) / 2;
+	n_write_segs = n_write_threads;
-	n_read_segs = n_segments - 2 - n_write_segs;
+	n_read_segs = n_read_threads;
 
 	/* fprintf(stderr, "Array n per seg %lu\n", n_per_seg); */
 
@@ -3140,6 +3144,13 @@ os_aio_array_reserve_slot(
 	OVERLAPPED*	control;
 #endif
 	ulint		i;
+	ulint		prim_segment;
+	ulint		n;
+
+	n = array->n_slots / array->n_segments;
+	/* 64 blocks' striping ( aligning max(BUF_READ_AHEAD_AREA) ) */
+	prim_segment = ( offset >> (UNIV_PAGE_SIZE_SHIFT + 6) ) % (array->n_segments);
+
 loop:
 	os_mutex_enter(array->mutex);
 
@@ -3158,6 +3169,16 @@ loop:
 		goto loop;
 	}
 
+	for (i = prim_segment * n; i < array->n_slots; i++) {
+		slot = os_aio_array_get_nth_slot(array, i);
+
+		if (slot->reserved == FALSE) {
+			break;
+		}
+	}
+
+	if (slot->reserved == TRUE){
+		/* Not found after the intended segment. So we should search before. */
 	for (i = 0;; i++) {
 		slot = os_aio_array_get_nth_slot(array, i);
 
@@ -3165,6 +3186,7 @@ loop:
 			break;
 		}
 	}
+	}
 
 	array->n_reserved++;
 

srv/srv0srv.c

@@ -132,6 +132,10 @@ UNIV_INTERN ulint	srv_log_file_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_log_buffer_size	= ULINT_MAX;
 UNIV_INTERN ulong	srv_flush_log_at_trx_commit = 1;
 
+UNIV_INTERN ulint  srv_show_locks_held     = 10;
+UNIV_INTERN ulint  srv_show_verbose_locks  = 0;
+
+
 /* The sort order table of the MySQL latin1_swedish_ci character set
 collation */
 UNIV_INTERN const byte*	srv_latin1_ordering;
@@ -147,6 +151,8 @@ UNIV_INTERN ulint	srv_mem_pool_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_lock_table_size	= ULINT_MAX;
 
 UNIV_INTERN ulint	srv_n_file_io_threads	= ULINT_MAX;
+ulint	srv_n_read_io_threads	= 1;
+ulint	srv_n_write_io_threads	= 1;
 
 #ifdef UNIV_LOG_ARCHIVE
 UNIV_INTERN ibool		srv_log_archive_on	= FALSE;
@@ -311,6 +317,15 @@ UNIV_INTERN int	srv_query_thread_priority = 0;
 
 UNIV_INTERN ulong	srv_replication_delay		= 0;
 
+ulint	srv_io_capacity = 100;
+
+/* Returns the number of IO operations that is X percent of the capacity.
+PCT_IO(5) -> returns the number of IO operations that is 5% of the max
+where max is srv_io_capacity. */
+#define PCT_IO(pct) ((ulint) (srv_io_capacity * ((double) pct / 100.0)))
+
+ulint	srv_read_ahead = 3; /* 1: random  2: linear  3: Both */
+ulint	srv_adaptive_checkpoint = 0; /* 0:disable 1:enable */
 /*-------------------------------------------*/
 UNIV_INTERN ulong	srv_n_spin_wait_rounds	= 20;
 UNIV_INTERN ulong	srv_n_free_tickets_to_enter = 500;
@@ -1614,6 +1629,14 @@ srv_printf_innodb_monitor(
 	time_t	current_time;
 	ulint	n_reserved;
 
+	ulint	btr_search_sys_subtotal;
+	ulint	lock_sys_subtotal;
+	ulint	recv_sys_subtotal;
+	ulint	io_counter_subtotal;
+
+	ulint	i;
+	trx_t*	trx;
+
 	mutex_enter(&srv_innodb_monitor_mutex);
 
 	current_time = time(NULL);
@@ -1657,24 +1680,6 @@ srv_printf_innodb_monitor(
 
 	mutex_exit(&dict_foreign_err_mutex);
 
-	lock_print_info_summary(file);
-	if (trx_start) {
-		long	t = ftell(file);
-		if (t < 0) {
-			*trx_start = ULINT_UNDEFINED;
-		} else {
-			*trx_start = (ulint) t;
-		}
-	}
-	lock_print_info_all_transactions(file);
-	if (trx_end) {
-		long	t = ftell(file);
-		if (t < 0) {
-			*trx_end = ULINT_UNDEFINED;
-		} else {
-			*trx_end = (ulint) t;
-		}
-	}
 	fputs("--------\n"
 	      "FILE I/O\n"
 	      "--------\n", file);
@@ -1705,10 +1710,84 @@ srv_printf_innodb_monitor(
 	      "BUFFER POOL AND MEMORY\n"
 	      "----------------------\n", file);
 	fprintf(file,
-		"Total memory allocated " ULINTPF
+			"Total memory allocated " ULINTPF
-		"; in additional pool allocated " ULINTPF "\n",
+			"; in additional pool allocated " ULINTPF "\n",
-		ut_total_allocated_memory,
+			ut_total_allocated_memory,
-		mem_pool_get_reserved(mem_comm_pool));
+			mem_pool_get_reserved(mem_comm_pool));
+	/* Calcurate reserved memories */
+	if (btr_search_sys && btr_search_sys->hash_index->heap) {
+		btr_search_sys_subtotal = mem_heap_get_size(btr_search_sys->hash_index->heap);
+	} else {
+		btr_search_sys_subtotal = 0;
+		for (i=0; i < btr_search_sys->hash_index->n_mutexes; i++) {
+			btr_search_sys_subtotal += mem_heap_get_size(btr_search_sys->hash_index->heaps[i]);
+		}
+	}
+
+	lock_sys_subtotal = 0;
+	if (trx_sys) {
+		mutex_enter(&kernel_mutex);
+		trx = UT_LIST_GET_FIRST(trx_sys->mysql_trx_list);
+		while (trx) {
+			lock_sys_subtotal += ((trx->lock_heap) ? mem_heap_get_size(trx->lock_heap) : 0);
+			trx = UT_LIST_GET_NEXT(mysql_trx_list, trx);
+		}
+		mutex_exit(&kernel_mutex);
+	}
+
+	recv_sys_subtotal = ((recv_sys && recv_sys->addr_hash)
+			? mem_heap_get_size(recv_sys->heap) : 0);
+
+	fprintf(file,
+			"Internal hash tables (constant factor + variable factor)\n"
+			"    Adaptive hash index %lu \t(%lu + %lu)\n"
+			"    Page hash           %lu\n"
+			"    Dictionary cache    %lu \t(%lu + %lu)\n"
+			"    File system         %lu \t(%lu + %lu)\n"
+			"    Lock system         %lu \t(%lu + %lu)\n"
+			"    Recovery system     %lu \t(%lu + %lu)\n"
+			"    Threads             %lu \t(%lu + %lu)\n",
+
+			(ulong) (btr_search_sys
+				? (btr_search_sys->hash_index->n_cells * sizeof(hash_cell_t)) : 0)
+			+ btr_search_sys_subtotal,
+			(ulong) (btr_search_sys
+				? (btr_search_sys->hash_index->n_cells * sizeof(hash_cell_t)) : 0),
+			(ulong) btr_search_sys_subtotal,
+
+			(ulong) (buf_pool->page_hash->n_cells * sizeof(hash_cell_t)),
+
+			(ulong) (dict_sys ? ((dict_sys->table_hash->n_cells
+						+ dict_sys->table_id_hash->n_cells
+						) * sizeof(hash_cell_t)
+					+ dict_sys->size) : 0),
+			(ulong) (dict_sys ? ((dict_sys->table_hash->n_cells
+							+ dict_sys->table_id_hash->n_cells
+							) * sizeof(hash_cell_t)) : 0),
+			(ulong) (dict_sys ? (dict_sys->size) : 0),
+
+			(ulong) (fil_system_hash_cells() * sizeof(hash_cell_t)
+					+ fil_system_hash_nodes()),
+			(ulong) (fil_system_hash_cells() * sizeof(hash_cell_t)),
+			(ulong) fil_system_hash_nodes(),
+
+			(ulong) ((lock_sys ? (lock_sys->rec_hash->n_cells * sizeof(hash_cell_t)) : 0)
+					+ lock_sys_subtotal),
+			(ulong) (lock_sys ? (lock_sys->rec_hash->n_cells * sizeof(hash_cell_t)) : 0),
+			(ulong) lock_sys_subtotal,
+
+			(ulong) (((recv_sys && recv_sys->addr_hash)
+						? (recv_sys->addr_hash->n_cells * sizeof(hash_cell_t)) : 0)
+					+ recv_sys_subtotal),
+			(ulong) ((recv_sys && recv_sys->addr_hash)
+					? (recv_sys->addr_hash->n_cells * sizeof(hash_cell_t)) : 0),
+			(ulong) recv_sys_subtotal,
+
+			(ulong) (thr_local_hash_cells() * sizeof(hash_cell_t)
+					+ thr_local_hash_nodes()),
+			(ulong) (thr_local_hash_cells() * sizeof(hash_cell_t)),
+			(ulong) thr_local_hash_nodes());
+
 	fprintf(file, "Dictionary memory allocated " ULINTPF "\n",
 		dict_sys->size);
 
@@ -1767,6 +1846,25 @@ srv_printf_innodb_monitor(
 	srv_n_rows_deleted_old = srv_n_rows_deleted;
 	srv_n_rows_read_old = srv_n_rows_read;
 
+	lock_print_info_summary(file);
+	if (trx_start) {
+		long	t = ftell(file);
+		if (t < 0) {
+			*trx_start = ULINT_UNDEFINED;
+		} else {
+			*trx_start = (ulint) t;
+		}
+	}
+	lock_print_info_all_transactions(file);
+	if (trx_end) {
+		long	t = ftell(file);
+		if (t < 0) {
+			*trx_end = ULINT_UNDEFINED;
+		} else {
+			*trx_end = (ulint) t;
+		}
+	}
+
 	fputs("----------------------------\n"
 	      "END OF INNODB MONITOR OUTPUT\n"
 	      "============================\n", file);
@@ -2203,6 +2301,8 @@ srv_master_thread(
 	ibool		skip_sleep	= FALSE;
 	ulint		i;
 
+	ib_uint64_t	oldest_lsn;
+	
 #ifdef UNIV_DEBUG_THREAD_CREATION
 	fprintf(stderr, "Master thread starts, id %lu\n",
 		os_thread_pf(os_thread_get_curr_id()));
@@ -2290,10 +2390,10 @@ loop:
 			+ log_sys->n_pending_writes;
 		n_ios = log_sys->n_log_ios + buf_pool->n_pages_read
 			+ buf_pool->n_pages_written;
-		if (n_pend_ios < 3 && (n_ios - n_ios_old < 5)) {
+		if (n_pend_ios < 3 && (n_ios - n_ios_old < PCT_IO(5))) {
 			srv_main_thread_op_info = "doing insert buffer merge";
 			ibuf_contract_for_n_pages(
-				TRUE, srv_insert_buffer_batch_size / 4);
+				TRUE, PCT_IO((srv_insert_buffer_batch_size / 4)));
 
 			srv_main_thread_op_info = "flushing log";
 
@@ -2306,7 +2406,7 @@ loop:
 			/* Try to keep the number of modified pages in the
 			buffer pool under the limit wished by the user */
 
-			n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, 100,
+ 			n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(100),
 							  IB_ULONGLONG_MAX);
 
 			/* If we had to do the flush, it may have taken
@@ -2315,6 +2415,44 @@ loop:
 			iteration of this loop. */
 
 			skip_sleep = TRUE;
+		} else if (srv_adaptive_checkpoint) {
+
+			/* Try to keep modified age not to exceed
+			max_checkpoint_age * 7/8 line */
+
+			mutex_enter(&(log_sys->mutex));
+
+			oldest_lsn = buf_pool_get_oldest_modification();
+			if (oldest_lsn == 0) {
+
+				mutex_exit(&(log_sys->mutex));
+
+			} else {
+				if ((log_sys->lsn - oldest_lsn)
+				    > (log_sys->max_checkpoint_age) - ((log_sys->max_checkpoint_age) / 4)) {
+
+					/* 2nd defence line (max_checkpoint_age * 3/4) */
+
+					mutex_exit(&(log_sys->mutex));
+
+					n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(100),
+									  IB_ULONGLONG_MAX);
+					skip_sleep = TRUE;
+				} else if ((log_sys->lsn - oldest_lsn)
+					   > (log_sys->max_checkpoint_age)/2 ) {
+
+					/* 1st defence line (max_checkpoint_age * 1/2) */
+
+					mutex_exit(&(log_sys->mutex));
+
+					n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(10),
+									  IB_ULONGLONG_MAX);
+					skip_sleep = TRUE;
+				} else {
+					mutex_exit(&(log_sys->mutex));
+				}
+			}
+
 		}
 
 		if (srv_activity_count == old_activity_count) {
@@ -2341,10 +2479,10 @@ loop:
 	n_pend_ios = buf_get_n_pending_ios() + log_sys->n_pending_writes;
 	n_ios = log_sys->n_log_ios + buf_pool->n_pages_read
 		+ buf_pool->n_pages_written;
-	if (n_pend_ios < 3 && (n_ios - n_ios_very_old < 200)) {
+ 	if (n_pend_ios < 3 && (n_ios - n_ios_very_old < PCT_IO(200))) {
   
-		srv_main_thread_op_info = "flushing buffer pool pages";
+  		srv_main_thread_op_info = "flushing buffer pool pages";
-		buf_flush_batch(BUF_FLUSH_LIST, 100, IB_ULONGLONG_MAX);
+ 		buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(100), IB_ULONGLONG_MAX);
 
 		srv_main_thread_op_info = "flushing log";
 		log_buffer_flush_to_disk();
@@ -2354,7 +2492,7 @@ loop:
 	even if the server were active */
 
 	srv_main_thread_op_info = "doing insert buffer merge";
-	ibuf_contract_for_n_pages(TRUE, srv_insert_buffer_batch_size / 4);
+	ibuf_contract_for_n_pages(TRUE, PCT_IO((srv_insert_buffer_batch_size / 4)));
 
 	srv_main_thread_op_info = "flushing log";
 	log_buffer_flush_to_disk();
@@ -2394,14 +2532,14 @@ loop:
 		(> 70 %), we assume we can afford reserving the disk(s) for
 		the time it requires to flush 100 pages */
 
-		n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, 100,
+ 	        n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(100),
 						  IB_ULONGLONG_MAX);
 	} else {
 		/* Otherwise, we only flush a small number of pages so that
 		we do not unnecessarily use much disk i/o capacity from
 		other work */
 
-		n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, 10,
+ 	        n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(10),
 						  IB_ULONGLONG_MAX);
 	}
 
@@ -2489,7 +2627,7 @@ background_loop:
 		n_bytes_merged = 0;
 	} else {
 		n_bytes_merged = ibuf_contract_for_n_pages(
-			TRUE, srv_insert_buffer_batch_size);
+			TRUE, PCT_IO((srv_insert_buffer_batch_size * 5)));
 	}
 
 	srv_main_thread_op_info = "reserving kernel mutex";
@@ -2505,7 +2643,7 @@ flush_loop:
 	srv_main_thread_op_info = "flushing buffer pool pages";
 
 	if (srv_fast_shutdown < 2) {
-		n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, 100,
+ 	        n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, PCT_IO(100),
 						  IB_ULONGLONG_MAX);
 	} else {
 		/* In the fastest shutdown we do not flush the buffer pool

srv/srv0start.c

@@ -1204,24 +1204,28 @@ innobase_start_or_create_for_mysql(void)
 		return(DB_ERROR);
 	}
 
+	/* over write innodb_file_io_threads */
+	srv_n_file_io_threads = 2 + srv_n_read_io_threads + srv_n_write_io_threads;
+
 	/* Restrict the maximum number of file i/o threads */
 	if (srv_n_file_io_threads > SRV_MAX_N_IO_THREADS) {
 
 		srv_n_file_io_threads = SRV_MAX_N_IO_THREADS;
+		srv_n_read_io_threads = srv_n_write_io_threads = (SRV_MAX_N_IO_THREADS - 2) / 2;
 	}
 
 	if (!os_aio_use_native_aio) {
 		/* In simulated aio we currently have use only for 4 threads */
-		srv_n_file_io_threads = 4;
+		/*srv_n_file_io_threads = 4;*/
 
 		os_aio_init(8 * SRV_N_PENDING_IOS_PER_THREAD
 			    * srv_n_file_io_threads,
-			    srv_n_file_io_threads,
+			    srv_n_read_io_threads, srv_n_write_io_threads,
-			    SRV_MAX_N_PENDING_SYNC_IOS);
+			    SRV_MAX_N_PENDING_SYNC_IOS * 8);
 	} else {
 		os_aio_init(SRV_N_PENDING_IOS_PER_THREAD
 			    * srv_n_file_io_threads,
-			    srv_n_file_io_threads,
+			    srv_n_read_io_threads, srv_n_write_io_threads,
 			    SRV_MAX_N_PENDING_SYNC_IOS);
 	}
 

sync/sync0arr.c

@@ -307,13 +307,13 @@ sync_cell_event_reset(
 {
 	if (type == SYNC_MUTEX) {
 		return(os_event_reset(((mutex_t *) object)->event));
-#ifdef __WIN__
 	} else if (type == RW_LOCK_WAIT_EX) {
 		return(os_event_reset(
 		       ((rw_lock_t *) object)->wait_ex_event));
-#endif
+	} else if (type == RW_LOCK_SHARED) {
-	} else {
+		return(os_event_reset(((rw_lock_t *) object)->s_event));
-		return(os_event_reset(((rw_lock_t *) object)->event));
+	} else { /* RW_LOCK_EX */
+		return(os_event_reset(((rw_lock_t *) object)->x_event));
 	}
 }
 
@@ -413,15 +413,12 @@ sync_array_wait_event(
 
 	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 if (cell->request_type == RW_LOCK_SHARED) {
+		event = ((rw_lock_t*) cell->wait_object)->s_event;
 	} else {
-		event = ((rw_lock_t*) cell->wait_object)->event;
+		event = ((rw_lock_t*) cell->wait_object)->x_event;
 	}
 
 		cell->waiting = TRUE;
@@ -462,6 +459,7 @@ sync_array_cell_print(
 	mutex_t*	mutex;
 	rw_lock_t*	rwlock;
 	ulint		type;
+	ulint		writer;
 
 	type = cell->request_type;
 
@@ -491,12 +489,10 @@ sync_array_cell_print(
 			(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);
+		fputs(type == RW_LOCK_SHARED ? "S-lock on" : "X-lock on", file);
 
 		rwlock = cell->old_wait_rw_lock;
 
@@ -504,22 +500,24 @@ sync_array_cell_print(
 			" RW-latch at %p created in file %s line %lu\n",
 			(void*) 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, s_waiters flag %lu, x_waiters flag %lu\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) rwlock->waiters,
+			(ulong) rwlock->s_waiters,
+			(ulong) (rwlock->x_waiters || rwlock->wait_ex_waiters),
 			rwlock->last_s_file_name,
 			(ulong) rwlock->last_s_line,
 			rwlock->last_x_file_name,
@@ -844,11 +842,15 @@ sync_array_object_signalled(
 /*========================*/
 	sync_array_t*	arr)	/* in: wait array */
 {
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	__sync_fetch_and_add(&(arr->sg_count),1);
+#else
 	sync_array_enter(arr);
 
 	arr->sg_count++;
 
 	sync_array_exit(arr);
+#endif
 }
 
 /**************************************************************************
@@ -889,19 +891,23 @@ sync_arr_wake_threads_if_sema_free(void)
 
 					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 if (cell->request_type
-				} else {
+					   == RW_LOCK_SHARED) {
 					rw_lock_t*	lock;
 
 					lock = cell->wait_object;
-					os_event_set(lock->event);
+					os_event_set(lock->s_event);
+				} else {
+					rw_lock_t*      lock;
+
+					lock = cell->wait_object;
+					os_event_set(lock->x_event);
 				}
 			}
 		}

sync/sync0rw.c

@@ -119,6 +119,7 @@ rw_lock_create_func(
 	/* If this is the very first time a synchronization object is
 	created, then the following call initializes the sync system. */
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_create(rw_lock_get_mutex(lock), SYNC_NO_ORDER_CHECK);
 
 	lock->mutex.cfile_name = cfile_name;
@@ -128,8 +129,14 @@ rw_lock_create_func(
 	lock->mutex.cmutex_name = cmutex_name;
 	lock->mutex.mutex_type = 1;
 #endif /* UNIV_DEBUG && !UNIV_HOTBACKUP */
+#endif /* !HAVE_GCC_ATOMIC_BUILTINS */
 
-	rw_lock_set_waiters(lock, 0);
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	lock->lock_word = RW_LOCK_BIAS;
+#endif
+	rw_lock_set_s_waiters(lock, 0);
+	rw_lock_set_x_waiters(lock, 0);
+	rw_lock_set_wx_waiters(lock, 0);
 	rw_lock_set_writer(lock, RW_LOCK_NOT_LOCKED);
 	lock->writer_count = 0;
 	rw_lock_set_reader_count(lock, 0);
@@ -151,11 +158,9 @@ rw_lock_create_func(
 	lock->last_x_file_name = "not yet reserved";
 	lock->last_s_line = 0;
 	lock->last_x_line = 0;
-	lock->event = os_event_create(NULL);
+	lock->s_event = os_event_create(NULL);
-
+	lock->x_event = os_event_create(NULL);
-#ifdef __WIN__
 	lock->wait_ex_event = os_event_create(NULL);
-#endif
 
 	mutex_enter(&rw_lock_list_mutex);
 
@@ -181,19 +186,21 @@ rw_lock_free(
 {
 	ut_ad(rw_lock_validate(lock));
 	ut_a(rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED);
-	ut_a(rw_lock_get_waiters(lock) == 0);
+	ut_a(rw_lock_get_s_waiters(lock) == 0);
+	ut_a(rw_lock_get_x_waiters(lock) == 0);
+	ut_a(rw_lock_get_wx_waiters(lock) == 0);
 	ut_a(rw_lock_get_reader_count(lock) == 0);
 
 	lock->magic_n = 0;
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_free(rw_lock_get_mutex(lock));
+#endif
 
 	mutex_enter(&rw_lock_list_mutex);
-	os_event_free(lock->event);
+	os_event_free(lock->s_event);
-
+	os_event_free(lock->x_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);
@@ -211,6 +218,8 @@ rw_lock_free(
 /**********************************************************************
 Checks that the rw-lock has been initialized and that there are no
 simultaneous shared and exclusive locks. */
+/* MEMO: If HAVE_GCC_ATOMIC_BUILTINS, we should use this function statically. */
+ 
 UNIV_INTERN
 ibool
 rw_lock_validate(
@@ -219,7 +228,9 @@ rw_lock_validate(
 {
 	ut_a(lock);
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(rw_lock_get_mutex(lock));
+#endif
 
 	ut_a(lock->magic_n == RW_LOCK_MAGIC_N);
 	ut_a((rw_lock_get_reader_count(lock) == 0)
@@ -227,11 +238,17 @@ rw_lock_validate(
 	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)
+	ut_a((rw_lock_get_s_waiters(lock) == 0)
-	     || (rw_lock_get_waiters(lock) == 1));
+	     || (rw_lock_get_s_waiters(lock) == 1));
+	ut_a((rw_lock_get_x_waiters(lock) == 0)
+	     || (rw_lock_get_x_waiters(lock) == 1));
+	ut_a((rw_lock_get_wx_waiters(lock) == 0)
+	     || (rw_lock_get_wx_waiters(lock) == 1));
 	ut_a((lock->writer != RW_LOCK_EX) || (lock->writer_count > 0));
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 	return(TRUE);
 }
@@ -258,13 +275,14 @@ rw_lock_s_lock_spin(
 	ut_ad(rw_lock_validate(lock));
 
 lock_loop:
+	i = 0;
+spin_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
+	while (i < SYNC_SPIN_ROUNDS
-	       && i < SYNC_SPIN_ROUNDS) {
+	       && rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED) {
 		if (srv_spin_wait_delay) {
 			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
 		}
@@ -285,15 +303,27 @@ lock_loop:
 			lock->cfile_name, (ulong) lock->cline, (ulong) i);
 	}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(rw_lock_get_mutex(lock));
+#endif
 
 	/* We try once again to obtain the lock */
 
 	if (TRUE == rw_lock_s_lock_low(lock, pass, file_name, line)) {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 		return; /* Success */
 	} else {
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+		/* like sync0sync.c doing */
+		i++;
+
+		if (i < SYNC_SPIN_ROUNDS) {
+			goto spin_loop;
+		}
+#endif
 		/* If we get here, locking did not succeed, we may
 		suspend the thread to wait in the wait array */
 
@@ -304,9 +334,19 @@ lock_loop:
 					file_name, line,
 					&index);
 
-		rw_lock_set_waiters(lock, 1);
+		rw_lock_set_s_waiters(lock, 1);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+		/* like sync0sync.c doing */
+		for (i = 0; i < 4; i++) {
+			if (TRUE == rw_lock_s_lock_low(lock, pass, file_name, line)) {
+				sync_array_free_cell(sync_primary_wait_array, index);
+				return; /* Success */
+			}
+		}
+#else
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 		if (srv_print_latch_waits) {
 			fprintf(stderr,
@@ -343,13 +383,19 @@ rw_lock_x_lock_move_ownership(
 {
 	ut_ad(rw_lock_is_locked(lock, RW_LOCK_EX));
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(&(lock->mutex));
+#endif
 
 	lock->writer_thread = os_thread_get_curr_id();
 
 	lock->pass = 0;
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(&(lock->mutex));
+#else
+	__sync_synchronize();
+#endif
 }
 
 /**********************************************************************
@@ -367,6 +413,89 @@ rw_lock_x_lock_low(
 	const char*	file_name,/* in: file name where lock requested */
 	ulint		line)	/* in: line where requested */
 {
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	os_thread_id_t	curr_thread	= os_thread_get_curr_id();
+
+	/* try to lock writer */
+	if(__sync_lock_test_and_set(&(lock->writer),RW_LOCK_EX)
+			== RW_LOCK_NOT_LOCKED) {
+		/* success */
+		/* obtain RW_LOCK_WAIT_EX right */
+		lock->writer_thread = curr_thread;
+		lock->pass = pass;
+		lock->writer_is_wait_ex = TRUE;
+		/* atomic operation may be safer about memory order. */
+		__sync_synchronize();
+#ifdef UNIV_SYNC_DEBUG
+		rw_lock_add_debug_info(lock, pass, RW_LOCK_WAIT_EX,
+					file_name, line);
+#endif
+	}
+
+	if (!os_thread_eq(lock->writer_thread, curr_thread)) {
+		return(RW_LOCK_NOT_LOCKED);
+	}
+
+	switch(rw_lock_get_writer(lock)) {
+	    case RW_LOCK_WAIT_EX:
+		/* have right to try x-lock */
+		if (lock->lock_word == RW_LOCK_BIAS) {
+			/* try x-lock */
+			if(__sync_sub_and_fetch(&(lock->lock_word),
+					RW_LOCK_BIAS) == 0) {
+				/* success */
+				lock->pass = pass;
+				lock->writer_is_wait_ex = FALSE;
+				__sync_fetch_and_add(&(lock->writer_count),1);
+
+#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);
+#endif
+
+				lock->last_x_file_name = file_name;
+				lock->last_x_line = line;
+
+				/* Locking succeeded, we may return */
+				return(RW_LOCK_EX);
+			} else {
+				/* fail */
+				__sync_fetch_and_add(&(lock->lock_word),
+					RW_LOCK_BIAS);
+			}
+		}
+		/* There are readers, we have to wait */
+		return(RW_LOCK_WAIT_EX);
+
+		break;
+
+	    case RW_LOCK_EX:
+		/* already have x-lock */
+		if ((lock->pass == 0)&&(pass == 0)) {
+			__sync_fetch_and_add(&(lock->writer_count),1);
+
+#ifdef UNIV_SYNC_DEBUG
+			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 = line;
+
+			/* Locking succeeded, we may return */
+			return(RW_LOCK_EX);
+		}
+
+		return(RW_LOCK_NOT_LOCKED);
+
+		break;
+
+	    default: /* ??? */
+		return(RW_LOCK_NOT_LOCKED);
+	}
+#else /* HAVE_GCC_ATOMIC_BUILTINS */
+
 	ut_ad(mutex_own(rw_lock_get_mutex(lock)));
 
 	if (rw_lock_get_writer(lock) == RW_LOCK_NOT_LOCKED) {
@@ -447,6 +576,7 @@ rw_lock_x_lock_low(
 		/* Locking succeeded, we may return */
 		return(RW_LOCK_EX);
 	}
+#endif /* HAVE_GCC_ATOMIC_BUILTINS */
 
 	/* Locking did not succeed */
 	return(RW_LOCK_NOT_LOCKED);
@@ -472,19 +602,33 @@ rw_lock_x_lock_func(
 	ulint		line)	/* in: line where requested */
 {
 	ulint	index;	/* index of the reserved wait cell */
-	ulint	state;	/* lock state acquired */
+	ulint	state = RW_LOCK_NOT_LOCKED;	/* lock state acquired */
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	ulint	prev_state = RW_LOCK_NOT_LOCKED;
+#endif
 	ulint	i;	/* spin round count */
 
 	ut_ad(rw_lock_validate(lock));
 
 lock_loop:
+	i = 0;
+
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	prev_state = state;
+#else
 	/* Acquire the mutex protecting the rw-lock fields */
 	mutex_enter_fast(&(lock->mutex));
+#endif
 
 	state = rw_lock_x_lock_low(lock, pass, file_name, line);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (state != prev_state) i=0; /* if progress, reset counter. */
+#else
 	mutex_exit(&(lock->mutex));
+#endif
 
+spin_loop:
 	if (state == RW_LOCK_EX) {
 
 		return;	/* Locking succeeded */
@@ -492,10 +636,9 @@ lock_loop:
 	} 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
+		while (i < SYNC_SPIN_ROUNDS
-		       && i < SYNC_SPIN_ROUNDS) {
+		       && rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED) {
 			if (srv_spin_wait_delay) {
 				ut_delay(ut_rnd_interval(0,
 							 srv_spin_wait_delay));
@@ -509,9 +652,12 @@ lock_loop:
 	} else if (state == RW_LOCK_WAIT_EX) {
 
 		/* Spin waiting for the reader count field to become zero */
-		i = 0;
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+		while (lock->lock_word != RW_LOCK_BIAS
+#else
 		while (rw_lock_get_reader_count(lock) != 0
+#endif
 		       && i < SYNC_SPIN_ROUNDS) {
 			if (srv_spin_wait_delay) {
 				ut_delay(ut_rnd_interval(0,
@@ -524,7 +670,6 @@ lock_loop:
 			os_thread_yield();
 		}
 	} else {
-		i = 0; /* Eliminate a compiler warning */
 		ut_error;
 	}
 
@@ -541,34 +686,69 @@ lock_loop:
 	/* We try once again to obtain the lock. Acquire the mutex protecting
 	the rw-lock fields */
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	prev_state = state;
+#else
 	mutex_enter(rw_lock_get_mutex(lock));
+#endif
 
 	state = rw_lock_x_lock_low(lock, pass, file_name, line);
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	if (state != prev_state) i=0; /* if progress, reset counter. */
+#endif
+
 	if (state == RW_LOCK_EX) {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 		return;	/* Locking succeeded */
 	}
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	/* like sync0sync.c doing */
+	i++;
+
+	if (i < SYNC_SPIN_ROUNDS) {
+		goto spin_loop;
+	}
+#endif
+
 	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);
+	if (state == RW_LOCK_WAIT_EX) {
+		rw_lock_set_wx_waiters(lock, 1);
+	} else {
+		rw_lock_set_x_waiters(lock, 1);
+	}
 
+#ifdef HAVE_GCC_ATOMIC_BUILTINS
+	/* like sync0sync.c doing */
+	for (i = 0; i < 4; i++) {
+		prev_state = state;
+		state = rw_lock_x_lock_low(lock, pass, file_name, line);
+		if (state == RW_LOCK_EX) {
+			sync_array_free_cell(sync_primary_wait_array, index);
+			return; /* Locking succeeded */
+		}
+		if (state != prev_state) {
+			/* retry! */
+			sync_array_free_cell(sync_primary_wait_array, index);
+			goto lock_loop;
+		}
+	}
+#else
 	mutex_exit(rw_lock_get_mutex(lock));
+#endif
 
 	if (srv_print_latch_waits) {
 		fprintf(stderr,
@@ -730,7 +910,9 @@ rw_lock_own(
 	ut_ad(lock);
 	ut_ad(rw_lock_validate(lock));
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(&(lock->mutex));
+#endif
 
 	info = UT_LIST_GET_FIRST(lock->debug_list);
 
@@ -740,7 +922,9 @@ rw_lock_own(
 		    && (info->pass == 0)
 		    && (info->lock_type == lock_type)) {
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 			mutex_exit(&(lock->mutex));
+#endif
 			/* Found! */
 
 			return(TRUE);
@@ -748,7 +932,9 @@ rw_lock_own(
 
 		info = UT_LIST_GET_NEXT(list, info);
 	}
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(&(lock->mutex));
+#endif
 
 	return(FALSE);
 }
@@ -770,21 +956,25 @@ rw_lock_is_locked(
 	ut_ad(lock);
 	ut_ad(rw_lock_validate(lock));
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_enter(&(lock->mutex));
+#endif
 
 	if (lock_type == RW_LOCK_SHARED) {
 		if (lock->reader_count > 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;
 	}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 	mutex_exit(&(lock->mutex));
+#endif
 
 	return(ret);
 }
@@ -814,16 +1004,26 @@ rw_lock_list_print_info(
 
 		count++;
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_enter(&(lock->mutex));
+#endif
 
 		if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
 		    || (rw_lock_get_reader_count(lock) != 0)
-		    || (rw_lock_get_waiters(lock) != 0)) {
+		    || (rw_lock_get_s_waiters(lock) != 0)
+		    || (rw_lock_get_x_waiters(lock) != 0)
+		    || (rw_lock_get_wx_waiters(lock) != 0)) {
 
 			fprintf(file, "RW-LOCK: %p ", (void*) lock);
 
-			if (rw_lock_get_waiters(lock)) {
+			if (rw_lock_get_s_waiters(lock)) {
-				fputs(" Waiters for the lock exist\n", file);
+				fputs(" s_waiters for the lock exist,", file);
+			}
+			if (rw_lock_get_x_waiters(lock)) {
+				fputs(" x_waiters for the lock exist\n", file);
+			}
+			if (rw_lock_get_wx_waiters(lock)) {
+				fputs(" wait_ex_waiters for the lock exist\n", file);
 			} else {
 				putc('\n', file);
 			}
@@ -835,7 +1035,9 @@ rw_lock_list_print_info(
 			}
 		}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(&(lock->mutex));
+#endif
 		lock = UT_LIST_GET_NEXT(list, lock);
 	}
 
@@ -860,10 +1062,18 @@ rw_lock_print(
 
 	if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
 	    || (rw_lock_get_reader_count(lock) != 0)
-	    || (rw_lock_get_waiters(lock) != 0)) {
+	    || (rw_lock_get_s_waiters(lock) != 0)
+	    || (rw_lock_get_x_waiters(lock) != 0)
+	    || (rw_lock_get_wx_waiters(lock) != 0)) {
 
-		if (rw_lock_get_waiters(lock)) {
+		if (rw_lock_get_s_waiters(lock)) {
-			fputs(" Waiters for the lock exist\n", stderr);
+			fputs(" s_waiters for the lock exist,", stderr);
+		}
+		if (rw_lock_get_x_waiters(lock)) {
+			fputs(" x_waiters for the lock exist\n", stderr);
+		}
+		if (rw_lock_get_wx_waiters(lock)) {
+			fputs(" wait_ex_waiters for the lock exist\n", stderr);
 		} else {
 			putc('\n', stderr);
 		}
@@ -922,14 +1132,18 @@ rw_lock_n_locked(void)
 	lock = UT_LIST_GET_FIRST(rw_lock_list);
 
 	while (lock != NULL) {
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_enter(rw_lock_get_mutex(lock));
+#endif
 
 		if ((rw_lock_get_writer(lock) != RW_LOCK_NOT_LOCKED)
 		    || (rw_lock_get_reader_count(lock) != 0)) {
 			count++;
 		}
 
+#ifndef HAVE_GCC_ATOMIC_BUILTINS
 		mutex_exit(rw_lock_get_mutex(lock));
+#endif
 		lock = UT_LIST_GET_NEXT(list, lock);
 	}
 

thr/thr0loc.c

@@ -32,6 +32,7 @@ static mutex_t		thr_local_mutex;
 
 /* The hash table. The module is not yet initialized when it is NULL. */
 static hash_table_t*	thr_local_hash	= NULL;
+ulint		thr_local_hash_n_nodes = 0;
 
 /* The private data for each thread should be put to
 the structure below and the accessor functions written
@@ -177,6 +178,7 @@ thr_local_create(void)
 		    os_thread_pf(os_thread_get_curr_id()),
 		    local);
 
+	thr_local_hash_n_nodes++;
 	mutex_exit(&thr_local_mutex);
 }
 
@@ -204,6 +206,7 @@ thr_local_free(
 
 	HASH_DELETE(thr_local_t, hash, thr_local_hash,
 		    os_thread_pf(id), local);
+	thr_local_hash_n_nodes--;
 
 	mutex_exit(&thr_local_mutex);
 
@@ -226,3 +229,29 @@ thr_local_init(void)
 
 	mutex_create(&thr_local_mutex, SYNC_THR_LOCAL);
 }
+
+/*************************************************************************
+Return local hash table informations. */
+
+ulint
+thr_local_hash_cells(void)
+/*======================*/
+{
+	if (thr_local_hash) {
+		return (thr_local_hash->n_cells);
+	} else {
+		return 0;
+	}
+}
+
+ulint
+thr_local_hash_nodes(void)
+/*======================*/
+{
+	if (thr_local_hash) {
+		return (thr_local_hash_n_nodes
+			* (sizeof(thr_local_t) + MEM_BLOCK_HEADER_SIZE));
+	} else {
+		return 0;
+	}
+}