branches/innodb+

Provide support for native AIO on Linux.

rb://46 approved by: Marko

fil/fil0fil.c

@@ -4436,11 +4436,14 @@ fil_aio_wait(
 
 	ut_ad(fil_validate());
 
-	if (os_aio_use_native_aio) {
+	if (srv_use_native_aio) {
 		srv_set_io_thread_op_info(segment, "native aio handle");
 #ifdef WIN_ASYNC_IO
 		ret = os_aio_windows_handle(segment, 0, &fil_node,
 					    &message, &type);
+#elif defined(LINUX_NATIVE_AIO)
+		ret = os_aio_linux_handle(segment, &fil_node,
+					  &message, &type);
 #else
 		ret = 0; /* Eliminate compiler warning */
 		ut_error;

handler/ha_innodb.cc

@@ -9573,6 +9573,11 @@ static MYSQL_SYSVAR_STR(version, innodb_version_str,
   PLUGIN_VAR_NOCMDOPT | PLUGIN_VAR_READONLY,
   "InnoDB version", NULL, NULL, INNODB_VERSION_STR);
 
+static MYSQL_SYSVAR_BOOL(use_native_aio, srv_use_native_aio,
+  PLUGIN_VAR_NOCMDARG | PLUGIN_VAR_READONLY,
+  "Use native AIO if supported on this platform.",
+  NULL, NULL, TRUE);
+
 static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(additional_mem_pool_size),
   MYSQL_SYSVAR(autoextend_increment),
@@ -9619,6 +9624,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(thread_sleep_delay),
   MYSQL_SYSVAR(autoinc_lock_mode),
   MYSQL_SYSVAR(version),
+  MYSQL_SYSVAR(use_native_aio),
   NULL
 };
 

include/os0file.h

@@ -51,12 +51,6 @@ typedef int	os_file_t;
 
 extern ulint	os_innodb_umask;
 
-/* If this flag is TRUE, then we will use the native aio of the
-OS (provided we compiled Innobase with it in), otherwise we will
-use simulated aio we build below with threads */
-
-extern ibool	os_aio_use_native_aio;
-
 #define OS_FILE_SECTOR_SIZE		512
 
 /* The next value should be smaller or equal to the smallest sector size used
@@ -98,6 +92,7 @@ log. */
 						to become available again */
 #define	OS_FILE_SHARING_VIOLATION	76
 #define	OS_FILE_ERROR_NOT_SPECIFIED	77
+#define	OS_FILE_AIO_INTERRUPTED		78
 
 /* Types for aio operations */
 #define OS_FILE_READ	10
@@ -556,9 +551,10 @@ in the three first aio arrays is the parameter n_segments given to the
 function. The caller must create an i/o handler thread for each segment in
 the four first arrays, but not for the sync aio array. */
 UNIV_INTERN
-void
+ibool
 os_aio_init(
 /*========*/
+				/* out: TRUE on success. */
 	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
@@ -737,4 +733,32 @@ innobase_mysql_tmpfile(void);
 			/* out: temporary file descriptor, or < 0 on error */
 #endif /* !UNIV_HOTBACKUP && !__NETWARE__ */
 
+
+#if defined(LINUX_NATIVE_AIO)
+/**************************************************************************
+This function is only used in Linux native asynchronous i/o.
+Waits for an aio operation to complete. This function is used to wait the
+for completed requests. The aio array of pending requests is divided
+into segments. The thread specifies which segment or slot it wants to wait
+for. NOTE: this function will also take care of freeing the aio slot,
+therefore no other thread is allowed to do the freeing! */
+UNIV_INTERN
+ibool
+os_aio_linux_handle(
+/*================*/
+				/* out: TRUE if the IO was successful */
+	ulint	global_seg,	/* in: segment number in the aio array
+				to wait for; segment 0 is the ibuf
+				i/o thread, segment 1 is log i/o thread,
+				then follow the non-ibuf read threads,
+				and the last are the non-ibuf write
+				threads. */
+	fil_node_t**message1,	/* out: the messages passed with the */
+	void**	message2,	/* aio request; note that in case the
+				aio operation failed, these output
+				parameters are valid and can be used to
+				restart the operation. */
+	ulint*	type);		/* out: OS_FILE_WRITE or ..._READ */
+#endif /* LINUX_NATIVE_AIO */
+
 #endif

include/srv0srv.h

@@ -68,6 +68,11 @@ extern ulint	srv_check_file_format_at_startup;
 on duplicate key checking and foreign key checking */
 extern ibool	srv_locks_unsafe_for_binlog;
 
+/* If this flag is TRUE, then we will use the native aio of the
+OS (provided we compiled Innobase with it in), otherwise we will
+use simulated aio we build below with threads.
+Currently we support native aio on windows and linux */
+extern my_bool	srv_use_native_aio;
 extern ulint	srv_n_data_files;
 extern char**	srv_data_file_names;
 extern ulint*	srv_data_file_sizes;

include/univ.i

@@ -162,6 +162,9 @@ operations (very slow); also UNIV_DEBUG must be defined */
 						for compressed pages */
 #define UNIV_ZIP_COPY				/* call page_zip_copy_recs()
 						more often */
+#define UNIV_AIO_DEBUG				/* prints info about
+						submitted and reaped AIO
+						requests to the log. */
 #endif
 
 #define UNIV_BTR_DEBUG				/* check B-tree links */

plug.in

@@ -12,6 +12,14 @@ MYSQL_PLUGIN_ACTIONS(innobase,  [
   AC_C_BIGENDIAN
   case "$target_os" in
 	lin*)
+             AC_CHECK_HEADER(libaio.h,
+               AC_CHECK_LIB(aio, io_setup,
+                 LIBS="$LIBS -laio"
+                 AC_DEFINE(LINUX_NATIVE_AIO, [1],
+                           [Linux native async I/O support]),
+                 AC_MSG_WARN([No Linux native async I/O])),
+               AC_MSG_WARN([No Linux native async I/O]))
+
 		CFLAGS="$CFLAGS -DUNIV_LINUX";;
 	hpux10*)
 		CFLAGS="$CFLAGS -DUNIV_MUST_NOT_INLINE -DUNIV_HPUX -DUNIV_HPUX10";;

srv/srv0srv.c

@@ -102,6 +102,12 @@ UNIV_INTERN ulint	srv_check_file_format_at_startup = DICT_TF_FORMAT_MAX;
 on duplicate key checking and foreign key checking */
 UNIV_INTERN ibool	srv_locks_unsafe_for_binlog = FALSE;
 
+/* If this flag is TRUE, then we will use the native aio of the
+OS (provided we compiled Innobase with it in), otherwise we will
+use simulated aio we build below with threads.
+Currently we support native aio on windows and linux */
+UNIV_INTERN my_bool	srv_use_native_aio = TRUE;
+
 UNIV_INTERN ulint	srv_n_data_files = 0;
 UNIV_INTERN char**	srv_data_file_names = NULL;
 /* size in database pages */

srv/srv0start.c

@@ -969,6 +969,7 @@ innobase_start_or_create_for_mysql(void)
 	ibool		log_file_created;
 	ibool		log_created	= FALSE;
 	ibool		log_opened	= FALSE;
+	ibool		success;
 	ib_uint64_t	min_flushed_lsn;
 	ib_uint64_t	max_flushed_lsn;
 #ifdef UNIV_LOG_ARCHIVE
@@ -1071,7 +1072,6 @@ innobase_start_or_create_for_mysql(void)
 
 	srv_is_being_started = TRUE;
 	srv_startup_is_before_trx_rollback_phase = TRUE;
-	os_aio_use_native_aio = FALSE;
 
 #ifdef __WIN__
 	if (os_get_os_version() == OS_WIN95
@@ -1083,12 +1083,30 @@ innobase_start_or_create_for_mysql(void)
 		but when run in conjunction with InnoDB Hot Backup, it seemed
 		to corrupt the data files. */
 
-		os_aio_use_native_aio = FALSE;
+		srv_use_native_aio = FALSE;
 	} else {
 		/* On Win 2000 and XP use async i/o */
-		os_aio_use_native_aio = TRUE;
+		srv_use_native_aio = TRUE;
 	}
+
+#elif defined(LINUX_NATIVE_AIO)
+
+	if (srv_use_native_aio) {
+		ut_print_timestamp(stderr);
+		fprintf(stderr,
+			"  InnoDB: Using Linux native AIO\n");
+	}
+#else
+	/* Currently native AIO is supported only on windows and linux
+	and that also when the support is compiled in. In all other
+	cases, we ignore the setting of innodb_use_native_aio. */
+
+	/* TODO: comment this out after internal testing. */
+	fprintf(stderr, "Ignoring innodb_use_native_aio\n");
+	srv_use_native_aio = FALSE;
+
 #endif
+
 	if (srv_file_flush_method_str == NULL) {
 		/* These are the default options */
 
@@ -1113,11 +1131,11 @@ innobase_start_or_create_for_mysql(void)
 #else
 	} else if (0 == ut_strcmp(srv_file_flush_method_str, "normal")) {
 		srv_win_file_flush_method = SRV_WIN_IO_NORMAL;
-		os_aio_use_native_aio = FALSE;
+		srv_use_native_aio = FALSE;
 
 	} else if (0 == ut_strcmp(srv_file_flush_method_str, "unbuffered")) {
 		srv_win_file_flush_method = SRV_WIN_IO_UNBUFFERED;
-		os_aio_use_native_aio = FALSE;
+		srv_use_native_aio = FALSE;
 
 	} else if (0 == ut_strcmp(srv_file_flush_method_str,
 				  "async_unbuffered")) {
@@ -1210,19 +1228,38 @@ innobase_start_or_create_for_mysql(void)
 		srv_n_file_io_threads = SRV_MAX_N_IO_THREADS;
 	}
 
-	if (!os_aio_use_native_aio) {
+	if (!srv_use_native_aio) {
 		/* In simulated aio we currently have use only for 4 threads */
 		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_MAX_N_PENDING_SYNC_IOS);
+		success = os_aio_init(8 * SRV_N_PENDING_IOS_PER_THREAD *
+				      srv_n_file_io_threads,
+				      srv_n_file_io_threads,
+				      SRV_MAX_N_PENDING_SYNC_IOS);
+		if (!success) {
+			return(DB_ERROR);
+		}
 	} else {
-		os_aio_init(SRV_N_PENDING_IOS_PER_THREAD
-			    * srv_n_file_io_threads,
-			    srv_n_file_io_threads,
-			    SRV_MAX_N_PENDING_SYNC_IOS);
+		/* Windows has a pending IO per thread limit.
+		Linux does not have any such restriction.
+		The question of what should be segment size
+		is a trade off. The larger size means longer
+		linear searches through the array and a smaller
+		value can lead to array being full, causing
+		unnecessary delays. The following value
+		for Linux is fairly arbitrary and needs to be
+		tested and tuned. */
+		success = os_aio_init(
+#if defined(LINUX_NATIVE_AIO)
+				      8 * 
+#endif /* LINUX_NATIVE_AIO */
+				      SRV_N_PENDING_IOS_PER_THREAD *
+				      srv_n_file_io_threads,
+				      srv_n_file_io_threads,
+				      SRV_MAX_N_PENDING_SYNC_IOS);
+		if (!success) {
+			return(DB_ERROR);
+		}
 	}
 
 	fil_init(srv_max_n_open_files);