- mention innodb_file_per_table

[packages/mysql.git] / innodb_separate_doublewrite.patch

# name       : innodb_separate_doublewrite.patch
# introduced : 11 or before
# maintainer : Yasufumi
#
#!!! notice !!!
# Any small change to this file in the main branch
# should be done or reviewed by the maintainer!
--- a/storage/innobase/buf/buf0buf.c
+++ b/storage/innobase/buf/buf0buf.c
@@ -3810,7 +3810,8 @@
 		read_space_id = mach_read_from_4(
 			frame + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID);
 
-		if (bpage->space == TRX_SYS_SPACE
+		if ((bpage->space == TRX_SYS_SPACE
+		     || (srv_doublewrite_file && bpage->space == TRX_DOUBLEWRITE_SPACE))
 		    && trx_doublewrite_page_inside(bpage->offset)) {
 
 			ut_print_timestamp(stderr);
--- a/storage/innobase/buf/buf0flu.c
+++ b/storage/innobase/buf/buf0flu.c
@@ -793,7 +793,8 @@
 	write_buf = trx_doublewrite->write_buf;
 	i = 0;
 
-	fil_io(OS_FILE_WRITE, TRUE, TRX_SYS_SPACE, 0,
+	fil_io(OS_FILE_WRITE, TRUE,
+	       (srv_doublewrite_file ? TRX_DOUBLEWRITE_SPACE : TRX_SYS_SPACE), 0,
 	       trx_doublewrite->block1, 0, len,
 	       (void*) write_buf, NULL);
 
@@ -830,7 +831,8 @@
 		+ TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * UNIV_PAGE_SIZE;
 	ut_ad(i == TRX_SYS_DOUBLEWRITE_BLOCK_SIZE);
 
-	fil_io(OS_FILE_WRITE, TRUE, TRX_SYS_SPACE, 0,
+	fil_io(OS_FILE_WRITE, TRUE,
+	       (srv_doublewrite_file ? TRX_DOUBLEWRITE_SPACE : TRX_SYS_SPACE), 0,
 	       trx_doublewrite->block2, 0, len,
 	       (void*) write_buf, NULL);
 
@@ -860,7 +862,7 @@
 flush:
 	/* Now flush the doublewrite buffer data to disk */
 
-	fil_flush(TRX_SYS_SPACE, FALSE);
+	fil_flush(srv_doublewrite_file ? TRX_DOUBLEWRITE_SPACE : TRX_SYS_SPACE, FALSE);
 
 	/* We know that the writes have been flushed to disk now
 	and in recovery we will find them in the doublewrite buffer
--- a/storage/innobase/buf/buf0rea.c
+++ b/storage/innobase/buf/buf0rea.c
@@ -88,7 +88,9 @@
 	wake_later = mode & OS_AIO_SIMULATED_WAKE_LATER;
 	mode = mode & ~OS_AIO_SIMULATED_WAKE_LATER;
 
-	if (trx_doublewrite && space == TRX_SYS_SPACE
+	if (trx_doublewrite
+	    && (space == TRX_SYS_SPACE
+		|| (srv_doublewrite_file && space == TRX_DOUBLEWRITE_SPACE))
 	    && (   (offset >= trx_doublewrite->block1
 		    && offset < trx_doublewrite->block1
 		    + TRX_SYS_DOUBLEWRITE_BLOCK_SIZE)
--- a/storage/innobase/dict/dict0load.c
+++ b/storage/innobase/dict/dict0load.c
@@ -41,6 +41,7 @@
 #include "srv0start.h"
 #include "srv0srv.h"
 #include "ha_prototypes.h" /* innobase_casedn_str() */
+#include "trx0sys.h"
 
 
 /** Following are six InnoDB system tables */
@@ -811,7 +812,7 @@
 
 		mtr_commit(&mtr);
 
-		if (space_id == 0) {
+		if (trx_sys_sys_space(space_id)) {
 			/* The system tablespace always exists. */
 		} else if (in_crash_recovery) {
 			/* Check that the tablespace (the .ibd file) really
@@ -1682,7 +1683,7 @@
 	space = mach_read_from_4(field);
 
 	/* Check if the tablespace exists and has the right name */
-	if (space != 0) {
+	if (!trx_sys_sys_space(space)) {
 		flags = dict_sys_tables_get_flags(rec);
 
 		if (UNIV_UNLIKELY(flags == ULINT_UNDEFINED)) {
@@ -1835,7 +1836,7 @@
 		goto err_exit;
 	}
 
-	if (table->space == 0) {
+	if (trx_sys_sys_space(table->space)) {
 		/* The system tablespace is always available. */
 	} else if (!fil_space_for_table_exists_in_mem(
 			   table->space, name,
--- a/storage/innobase/fil/fil0fil.c
+++ b/storage/innobase/fil/fil0fil.c
@@ -657,7 +657,7 @@
 
 	UT_LIST_ADD_LAST(chain, space->chain, node);
 
-	if (id < SRV_LOG_SPACE_FIRST_ID && fil_system->max_assigned_id < id) {
+	if (id < SRV_EXTRA_SYS_SPACE_FIRST_ID && fil_system->max_assigned_id < id) {
 
 		fil_system->max_assigned_id = id;
 	}
@@ -721,14 +721,14 @@
 		size_bytes = (((ib_int64_t)size_high) << 32)
 			+ (ib_int64_t)size_low;
 #ifdef UNIV_HOTBACKUP
-		if (space->id == 0) {
+		if (trx_sys_sys_space(space->id)) {
 			node->size = (ulint) (size_bytes / UNIV_PAGE_SIZE);
 			os_file_close(node->handle);
 			goto add_size;
 		}
 #endif /* UNIV_HOTBACKUP */
 		ut_a(space->purpose != FIL_LOG);
-		ut_a(space->id != 0);
+		ut_a(!trx_sys_sys_space(space->id));
 
 		if (size_bytes < FIL_IBD_FILE_INITIAL_SIZE * UNIV_PAGE_SIZE) {
 			fprintf(stderr,
@@ -774,7 +774,7 @@
 		}
 
 		if (UNIV_UNLIKELY(space_id == ULINT_UNDEFINED
-				  || space_id == 0)) {
+				  || trx_sys_sys_space(space_id))) {
 			fprintf(stderr,
 				"InnoDB: Error: tablespace id %lu"
 				" in file %s is not sensible\n",
@@ -842,7 +842,7 @@
 
 	system->n_open++;
 
-	if (space->purpose == FIL_TABLESPACE && space->id != 0) {
+	if (space->purpose == FIL_TABLESPACE && !trx_sys_sys_space(space->id)) {
 		/* Put the node to the LRU list */
 		UT_LIST_ADD_FIRST(LRU, system->LRU, node);
 	}
@@ -876,7 +876,7 @@
 	ut_a(system->n_open > 0);
 	system->n_open--;
 
-	if (node->space->purpose == FIL_TABLESPACE && node->space->id != 0) {
+	if (node->space->purpose == FIL_TABLESPACE && !trx_sys_sys_space(node->space->id)) {
 		ut_a(UT_LIST_GET_LEN(system->LRU) > 0);
 
 		/* The node is in the LRU list, remove it */
@@ -962,7 +962,7 @@
 retry:
 	mutex_enter(&fil_system->mutex);
 
-	if (space_id == 0 || space_id >= SRV_LOG_SPACE_FIRST_ID) {
+	if (trx_sys_sys_space(space_id) || space_id >= SRV_LOG_SPACE_FIRST_ID) {
 		/* We keep log files and system tablespace files always open;
 		this is important in preventing deadlocks in this module, as
 		a page read completion often performs another read from the
@@ -1193,7 +1193,7 @@
 			" tablespace memory cache!\n",
 			(ulong) space->id);
 
-		if (id == 0 || purpose != FIL_TABLESPACE) {
+		if (trx_sys_sys_space(id) || purpose != FIL_TABLESPACE) {
 
 			mutex_exit(&fil_system->mutex);
 
@@ -1255,6 +1255,7 @@
 	space->mark = FALSE;
 
 	if (UNIV_LIKELY(purpose == FIL_TABLESPACE && !recv_recovery_on)
+	    && UNIV_UNLIKELY(id < SRV_EXTRA_SYS_SPACE_FIRST_ID)
 	    && UNIV_UNLIKELY(id > fil_system->max_assigned_id)) {
 		if (!fil_system->space_id_reuse_warned) {
 			fil_system->space_id_reuse_warned = TRUE;
@@ -1338,7 +1339,7 @@
 			(ulong) SRV_LOG_SPACE_FIRST_ID);
 	}
 
-	success = (id < SRV_LOG_SPACE_FIRST_ID);
+	success = (id < SRV_EXTRA_SYS_SPACE_FIRST_ID);
 
 	if (success) {
 		*space_id = fil_system->max_assigned_id = id;
@@ -1601,6 +1602,8 @@
 	UT_LIST_INIT(fil_system->LRU);
 
 	fil_system->max_n_open = max_n_open;
+
+	fil_system->max_assigned_id = TRX_SYS_SPACE_MAX;
 }
 
 /*******************************************************************//**
@@ -1622,7 +1625,7 @@
 	space = UT_LIST_GET_FIRST(fil_system->space_list);
 
 	while (space != NULL) {
-		if (space->purpose != FIL_TABLESPACE || space->id == 0) {
+		if (space->purpose != FIL_TABLESPACE || trx_sys_sys_space(space->id)) {
 			node = UT_LIST_GET_FIRST(space->chain);
 
 			while (node != NULL) {
@@ -1712,6 +1715,10 @@
 		ut_error;
 	}
 
+	if (max_id >= SRV_EXTRA_SYS_SPACE_FIRST_ID) {
+		return;
+	}
+
 	mutex_enter(&fil_system->mutex);
 
 	if (fil_system->max_assigned_id < max_id) {
@@ -1730,6 +1737,7 @@
 ulint
 fil_write_lsn_and_arch_no_to_file(
 /*==============================*/
+	ulint		space_id,
 	ulint		sum_of_sizes,	/*!< in: combined size of previous files
 					in space, in database pages */
 	ib_uint64_t	lsn,		/*!< in: lsn to write */
@@ -1739,14 +1747,16 @@
 	byte*	buf1;
 	byte*	buf;
 
+	ut_a(trx_sys_sys_space(space_id));
+
 	buf1 = mem_alloc(2 * UNIV_PAGE_SIZE);
 	buf = ut_align(buf1, UNIV_PAGE_SIZE);
 
-	fil_read(TRUE, 0, 0, sum_of_sizes, 0, UNIV_PAGE_SIZE, buf, NULL);
+	fil_read(TRUE, space_id, 0, sum_of_sizes, 0, UNIV_PAGE_SIZE, buf, NULL);
 
 	mach_write_to_8(buf + FIL_PAGE_FILE_FLUSH_LSN, lsn);
 
-	fil_write(TRUE, 0, 0, sum_of_sizes, 0, UNIV_PAGE_SIZE, buf, NULL);
+	fil_write(TRUE, space_id, 0, sum_of_sizes, 0, UNIV_PAGE_SIZE, buf, NULL);
 
 	mem_free(buf1);
 
@@ -1782,7 +1792,7 @@
 		always open. */
 
 		if (space->purpose == FIL_TABLESPACE
-		    && space->id == 0) {
+		    && trx_sys_sys_space(space->id)) {
 			sum_of_sizes = 0;
 
 			node = UT_LIST_GET_FIRST(space->chain);
@@ -1790,7 +1800,7 @@
 				mutex_exit(&fil_system->mutex);
 
 				err = fil_write_lsn_and_arch_no_to_file(
-					sum_of_sizes, lsn, arch_log_no);
+					space->id, sum_of_sizes, lsn, arch_log_no);
 				if (err != DB_SUCCESS) {
 
 					return(err);
@@ -4159,7 +4169,7 @@
 	}
 
 #ifndef UNIV_HOTBACKUP
-	if (space_id == ULINT_UNDEFINED || space_id == 0) {
+	if (space_id == ULINT_UNDEFINED || trx_sys_sys_space(space_id)) {
 		fprintf(stderr,
 			"InnoDB: Error: tablespace id %lu in file %s"
 			" is not sensible\n",
@@ -4168,7 +4178,7 @@
 		goto func_exit;
 	}
 #else
-	if (space_id == ULINT_UNDEFINED || space_id == 0) {
+	if (space_id == ULINT_UNDEFINED || trx_sys_sys_space(space_id)) {
 		char*	new_path;
 
 		fprintf(stderr,
@@ -4989,7 +4999,7 @@
 	}
 
 	if (node->n_pending == 0 && space->purpose == FIL_TABLESPACE
-	    && space->id != 0) {
+	    && !trx_sys_sys_space(space->id)) {
 		/* The node is in the LRU list, remove it */
 
 		ut_a(UT_LIST_GET_LEN(system->LRU) > 0);
@@ -5035,7 +5045,7 @@
 	}
 
 	if (node->n_pending == 0 && node->space->purpose == FIL_TABLESPACE
-	    && node->space->id != 0) {
+	    && !trx_sys_sys_space(node->space->id)) {
 		/* The node must be put back to the LRU list */
 		UT_LIST_ADD_FIRST(LRU, system->LRU, node);
 	}
@@ -5646,7 +5656,7 @@
 		ut_a(fil_node->n_pending == 0);
 		ut_a(fil_node->open);
 		ut_a(fil_node->space->purpose == FIL_TABLESPACE);
-		ut_a(fil_node->space->id != 0);
+		ut_a(!trx_sys_sys_space(fil_node->space->id));
 
 		fil_node = UT_LIST_GET_NEXT(LRU, fil_node);
 	}
--- a/storage/innobase/fsp/fsp0fsp.c
+++ b/storage/innobase/fsp/fsp0fsp.c
@@ -48,7 +48,7 @@
 # include "log0log.h"
 #endif /* UNIV_HOTBACKUP */
 #include "dict0mem.h"
-
+#include "trx0sys.h"
 
 #define FSP_HEADER_OFFSET	FIL_PAGE_DATA	/* Offset of the space header
 						within a file page */
@@ -999,10 +999,10 @@
 	flst_init(header + FSP_SEG_INODES_FREE, mtr);
 
 	mlog_write_ull(header + FSP_SEG_ID, 1, mtr);
-	if (space == 0) {
+	if (space == TRX_SYS_SPACE || space == TRX_DOUBLEWRITE_SPACE) {
 		fsp_fill_free_list(FALSE, space, header, mtr);
 		btr_create(DICT_CLUSTERED | DICT_UNIVERSAL | DICT_IBUF,
-			   0, 0, DICT_IBUF_ID_MIN + space,
+			   space, 0, DICT_IBUF_ID_MIN + space,
 			   dict_ind_redundant, mtr);
 	} else {
 		fsp_fill_free_list(TRUE, space, header, mtr);
--- a/storage/innobase/handler/ha_innodb.cc
+++ b/storage/innobase/handler/ha_innodb.cc
@@ -163,6 +163,7 @@
 static char*	innobase_log_group_home_dir		= NULL;
 static char*	innobase_file_format_name		= NULL;
 static char*	innobase_change_buffering		= NULL;
+static char*	innobase_doublewrite_file		= NULL;
 
 /* The highest file format being used in the database. The value can be
 set by user, however, it will be adjusted to the newer file format if
@@ -2477,6 +2478,8 @@
 		goto error;
 	}
 
+	srv_doublewrite_file = innobase_doublewrite_file;
+
 	srv_use_sys_stats_table = (ibool) innobase_use_sys_stats_table;
 
 	/* -------------- Log files ---------------------------*/
@@ -11657,6 +11660,11 @@
   "Path to individual files and their sizes.",
   NULL, NULL, NULL);
 
+static MYSQL_SYSVAR_STR(doublewrite_file, innobase_doublewrite_file,
+  PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
+  "Path to special datafile for doublewrite buffer. (default is "": not used) ### ONLY FOR EXPERTS!!! ###",
+  NULL, NULL, NULL);
+
 static MYSQL_SYSVAR_LONG(autoinc_lock_mode, innobase_autoinc_lock_mode,
   PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
   "The AUTOINC lock modes supported by InnoDB:               "
@@ -11824,6 +11832,7 @@
   MYSQL_SYSVAR(commit_concurrency),
   MYSQL_SYSVAR(concurrency_tickets),
   MYSQL_SYSVAR(data_file_path),
+  MYSQL_SYSVAR(doublewrite_file),
   MYSQL_SYSVAR(data_home_dir),
   MYSQL_SYSVAR(doublewrite),
   MYSQL_SYSVAR(recovery_stats),
--- a/storage/innobase/include/mtr0log.ic
+++ b/storage/innobase/include/mtr0log.ic
@@ -27,8 +27,8 @@
 #include "ut0lst.h"
 #include "buf0buf.h"
 #include "fsp0types.h"
+#include "srv0srv.h"
 #include "trx0sys.h"
-
 /********************************************************//**
 Opens a buffer to mlog. It must be closed with mlog_close.
 @return	buffer, NULL if log mode MTR_LOG_NONE */
@@ -201,7 +201,8 @@
 	the doublewrite buffer is located in pages
 	FSP_EXTENT_SIZE, ..., 3 * FSP_EXTENT_SIZE - 1 in the
 	system tablespace */
-	if (space == TRX_SYS_SPACE
+	if ((space == TRX_SYS_SPACE
+	     || (srv_doublewrite_file && space == TRX_DOUBLEWRITE_SPACE))
 	    && offset >= FSP_EXTENT_SIZE && offset < 3 * FSP_EXTENT_SIZE) {
 		if (trx_doublewrite_buf_is_being_created) {
 			/* Do nothing: we only come to this branch in an
--- a/storage/innobase/include/srv0srv.h
+++ b/storage/innobase/include/srv0srv.h
@@ -129,6 +129,8 @@
 extern ulint*	srv_data_file_sizes;
 extern ulint*	srv_data_file_is_raw_partition;
 
+extern char*	srv_doublewrite_file;
+
 extern ibool	srv_recovery_stats;
 
 extern ibool	srv_auto_extend_last_data_file;
--- a/storage/innobase/include/srv0start.h
+++ b/storage/innobase/include/srv0start.h
@@ -127,4 +127,7 @@
 /** Log 'spaces' have id's >= this */
 #define SRV_LOG_SPACE_FIRST_ID		0xFFFFFFF0UL
 
+/** reserved for extra system tables */
+#define SRV_EXTRA_SYS_SPACE_FIRST_ID	0xFFFFFFE0UL
+
 #endif
--- a/storage/innobase/include/trx0sys.h
+++ b/storage/innobase/include/trx0sys.h
@@ -125,6 +125,22 @@
 /*=============*/
 	ulint	space,	/*!< in: space */
 	ulint	page_no);/*!< in: page number */
+/***************************************************************//**
+Checks if a space is the system tablespaces.
+@return TRUE if system tablespace */
+UNIV_INLINE
+ibool
+trx_sys_sys_space(
+/*==============*/
+	ulint	space);	/*!< in: space */
+/***************************************************************//**
+Checks if a space is the doublewrite tablespace.
+@return TRUE if doublewrite tablespace */
+UNIV_INLINE
+ibool
+trx_sys_doublewrite_space(
+/*======================*/
+	ulint	space);	/*!< in: space */
 /*****************************************************************//**
 Creates and initializes the central memory structures for the transaction
 system. This is called when the database is started. */
@@ -138,6 +154,13 @@
 void
 trx_sys_create(void);
 /*================*/
+/*****************************************************************//**
+Creates and initializes the dummy transaction system page for tablespace. */
+UNIV_INTERN
+void
+trx_sys_dummy_create(
+/*=================*/
+	ulint	space);
 /****************************************************************//**
 Looks for a free slot for a rollback segment in the trx system file copy.
 @return	slot index or ULINT_UNDEFINED if not found */
@@ -453,6 +476,8 @@
 
 /* Space id and page no where the trx system file copy resides */
 #define	TRX_SYS_SPACE	0	/* the SYSTEM tablespace */
+#define	TRX_DOUBLEWRITE_SPACE	0xFFFFFFE0UL	/* the doublewrite buffer tablespace if used */
+#define	TRX_SYS_SPACE_MAX	9	/* reserved max space id for system tablespaces */
 #include "fsp0fsp.h"
 #define	TRX_SYS_PAGE_NO	FSP_TRX_SYS_PAGE_NO
 
--- a/storage/innobase/include/trx0sys.ic
+++ b/storage/innobase/include/trx0sys.ic
@@ -71,6 +71,40 @@
 }
 
 /***************************************************************//**
+Checks if a space is the system tablespaces.
+@return TRUE if system tablespace */
+UNIV_INLINE
+ibool
+trx_sys_sys_space(
+/*==============*/
+	ulint	space)	/*!< in: space */
+{
+	if (srv_doublewrite_file) {
+		/* several spaces are reserved */
+		return((ibool)(space == TRX_SYS_SPACE || space == TRX_DOUBLEWRITE_SPACE));
+	} else {
+		return((ibool)(space == TRX_SYS_SPACE));
+	}
+}
+
+/***************************************************************//**
+Checks if a space is the doublewrite tablespace.
+@return TRUE if doublewrite tablespace */
+UNIV_INLINE
+ibool
+trx_sys_doublewrite_space(
+/*======================*/
+	ulint	space)	/*!< in: space */
+{
+	if (srv_doublewrite_file) {
+		/* doublewrite buffer is separated */
+		return((ibool)(space == TRX_DOUBLEWRITE_SPACE));
+	} else {
+		return((ibool)(space == TRX_SYS_SPACE));
+	}
+}
+
+/***************************************************************//**
 Gets the pointer in the nth slot of the rseg array.
 @return	pointer to rseg object, NULL if slot not in use */
 UNIV_INLINE
--- a/storage/innobase/row/row0mysql.c
+++ b/storage/innobase/row/row0mysql.c
@@ -3425,7 +3425,7 @@
 		/* Do not drop possible .ibd tablespace if something went
 		wrong: we do not want to delete valuable data of the user */
 
-		if (err == DB_SUCCESS && space_id > 0) {
+		if (err == DB_SUCCESS && !trx_sys_sys_space(space_id)) {
 			if (!fil_space_for_table_exists_in_mem(space_id,
 							       name_or_path,
 							       is_temp, FALSE,
--- a/storage/innobase/srv/srv0srv.c
+++ b/storage/innobase/srv/srv0srv.c
@@ -163,6 +163,8 @@
 /* size in database pages */
 UNIV_INTERN ulint*	srv_data_file_sizes = NULL;
 
+UNIV_INTERN char*	srv_doublewrite_file = NULL;
+
 UNIV_INTERN ibool	srv_recovery_stats = FALSE;
 
 /* if TRUE, then we auto-extend the last data file */
--- a/storage/innobase/srv/srv0start.c
+++ b/storage/innobase/srv/srv0start.c
@@ -714,6 +714,7 @@
 /*======================*/
 	ibool*		create_new_db,	/*!< out: TRUE if new database should be
 					created */
+	ibool*		create_new_doublewrite_file,
 #ifdef UNIV_LOG_ARCHIVE
 	ulint*		min_arch_log_no,/*!< out: min of archived log
 					numbers in data files */
@@ -746,6 +747,7 @@
 	*sum_of_new_sizes = 0;
 
 	*create_new_db = FALSE;
+	*create_new_doublewrite_file = FALSE;
 
 	srv_normalize_path_for_win(srv_data_home);
 
@@ -983,6 +985,142 @@
 				srv_data_file_is_raw_partition[i] != 0);
 	}
 
+	/* special file for doublewrite buffer */
+	if (srv_doublewrite_file)
+	{
+		srv_normalize_path_for_win(srv_doublewrite_file);
+
+		fprintf(stderr,
+			"InnoDB: Note: The innodb_doublewrite_file option has been specified.\n"
+			"InnoDB: This option is for experts only. Don't use it unless you understand WELL what it is.\n"
+			"InnoDB: ### Don't specify a file older than the last checkpoint. ###\n"
+			"InnoDB: Otherwise, the older doublewrite buffer will break your data during recovery!\n");
+
+		strcpy(name, srv_doublewrite_file);
+
+		/* First we try to create the file: if it already
+		exists, ret will get value FALSE */
+
+		files[i] = os_file_create(innodb_file_data_key, name, OS_FILE_CREATE,
+					  OS_FILE_NORMAL,
+					  OS_DATA_FILE, &ret);
+
+		if (ret == FALSE && os_file_get_last_error(FALSE)
+		    != OS_FILE_ALREADY_EXISTS
+#ifdef UNIV_AIX
+		    /* AIX 5.1 after security patch ML7 may have
+		    errno set to 0 here, which causes our function
+		    to return 100; work around that AIX problem */
+		    && os_file_get_last_error(FALSE) != 100
+#endif
+		    ) {
+			fprintf(stderr,
+				"InnoDB: Error in creating"
+				" or opening %s\n",
+				name);
+
+			return(DB_ERROR);
+		}
+
+		if (ret == FALSE) {
+			/* We open the data file */
+
+			files[i] = os_file_create(innodb_file_data_key,
+				name, OS_FILE_OPEN, OS_FILE_NORMAL,
+				OS_DATA_FILE, &ret);
+
+			if (!ret) {
+				fprintf(stderr,
+					"InnoDB: Error in opening %s\n", name);
+				os_file_get_last_error(TRUE);
+
+				return(DB_ERROR);
+			}
+
+			ret = os_file_get_size(files[i], &size, &size_high);
+			ut_a(ret);
+			/* Round size downward to megabytes */
+
+			rounded_size_pages
+				= (size / (1024 * 1024) + 4096 * size_high)
+					<< (20 - UNIV_PAGE_SIZE_SHIFT);
+
+			if (rounded_size_pages != TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9) {
+
+				fprintf(stderr,
+					"InnoDB: Warning: doublewrite buffer file %s"
+					" is of a different size\n"
+					"InnoDB: %lu pages"
+					" (rounded down to MB)\n"
+					"InnoDB: than intended size"
+					" %lu pages...\n",
+					name,
+					(ulong) rounded_size_pages,
+					(ulong) TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9);
+			}
+
+			fil_read_flushed_lsn_and_arch_log_no(
+				files[i], one_opened,
+#ifdef UNIV_LOG_ARCHIVE
+				min_arch_log_no, max_arch_log_no,
+#endif /* UNIV_LOG_ARCHIVE */
+				min_flushed_lsn, max_flushed_lsn);
+			one_opened = TRUE;
+		} else {
+			/* We created the data file and now write it full of
+			zeros */
+
+			*create_new_doublewrite_file = TRUE;
+
+			ut_print_timestamp(stderr);
+			fprintf(stderr,
+				"  InnoDB: Doublewrite buffer file %s did not"
+				" exist. It will be be created.\n",
+				name);
+
+			if (*create_new_db == FALSE) {
+				fprintf(stderr,
+					"InnoDB: Notice: Previous version's ibdata files may cause crash.\n"
+					"        If you use that, please use the ibdata files of this version.\n");
+			}
+
+			ut_print_timestamp(stderr);
+			fprintf(stderr,
+				"  InnoDB: Setting file %s size to %lu MB\n",
+				name,
+				(ulong) ((TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9)
+					 >> (20 - UNIV_PAGE_SIZE_SHIFT)));
+
+			fprintf(stderr,
+				"InnoDB: Database physically writes the"
+				" file full: wait...\n");
+
+			ret = os_file_set_size(
+				name, files[i],
+				srv_calc_low32(TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9),
+				srv_calc_high32(TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9));
+
+			if (!ret) {
+				fprintf(stderr,
+					"InnoDB: Error in creating %s:"
+					" probably out of disk space\n", name);
+
+				return(DB_ERROR);
+			}
+		}
+
+		ret = os_file_close(files[i]);
+		ut_a(ret);
+
+		fil_space_create(name, TRX_DOUBLEWRITE_SPACE, 0, FIL_TABLESPACE);
+
+		ut_a(fil_validate());
+
+		fil_node_create(name, TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9, TRX_DOUBLEWRITE_SPACE, FALSE);
+
+		i++;
+	}
+
 	return(DB_SUCCESS);
 }
 
@@ -996,6 +1134,7 @@
 /*====================================*/
 {
 	ibool		create_new_db;
+	ibool		create_new_doublewrite_file;
 	ibool		log_file_created;
 	ibool		log_created	= FALSE;
 	ibool		log_opened	= FALSE;
@@ -1461,6 +1600,7 @@
 	}
 
 	err = open_or_create_data_files(&create_new_db,
+					&create_new_doublewrite_file,
 #ifdef UNIV_LOG_ARCHIVE
 					&min_arch_log_no, &max_arch_log_no,
 #endif /* UNIV_LOG_ARCHIVE */
@@ -1628,6 +1768,14 @@
 		after the double write buffer has been created. */
 		trx_sys_create();
 
+		if (create_new_doublewrite_file) {
+			mtr_start(&mtr);
+			fsp_header_init(TRX_DOUBLEWRITE_SPACE, TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9, &mtr);
+			mtr_commit(&mtr);
+
+			trx_sys_dummy_create(TRX_DOUBLEWRITE_SPACE);
+		}
+
 		dict_create();
 
 		srv_startup_is_before_trx_rollback_phase = FALSE;
@@ -1661,6 +1809,13 @@
 		recv_recovery_from_archive_finish();
 #endif /* UNIV_LOG_ARCHIVE */
 	} else {
+		char*	save_srv_doublewrite_file = NULL;
+
+		if (create_new_doublewrite_file) {
+			/* doublewrite_file cannot be used for recovery yet. */
+			save_srv_doublewrite_file = srv_doublewrite_file;
+			srv_doublewrite_file = NULL;
+		}
 
 		/* Check if we support the max format that is stamped
 		on the system tablespace. 
@@ -1747,6 +1902,17 @@
 		we have finished the recovery process so that the
 		image of TRX_SYS_PAGE_NO is not stale. */
 		trx_sys_file_format_tag_init();
+
+		if (create_new_doublewrite_file) {
+			/* restore the value */
+			srv_doublewrite_file = save_srv_doublewrite_file;
+
+			mtr_start(&mtr);
+			fsp_header_init(TRX_DOUBLEWRITE_SPACE, TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * 9, &mtr);
+			mtr_commit(&mtr);
+
+			trx_sys_dummy_create(TRX_DOUBLEWRITE_SPACE);
+		}
 	}
 
 	if (!create_new_db && sum_of_new_sizes > 0) {
--- a/storage/innobase/trx/trx0sys.c
+++ b/storage/innobase/trx/trx0sys.c
@@ -415,6 +415,152 @@
 
 		goto start_again;
 	}
+
+    if (srv_doublewrite_file) {
+	/* the same doublewrite buffer to TRX_SYS_SPACE should exist.
+	check and create if not exist.*/
+
+	mtr_start(&mtr);
+	trx_doublewrite_buf_is_being_created = TRUE;
+
+	block = buf_page_get(TRX_DOUBLEWRITE_SPACE, 0, TRX_SYS_PAGE_NO,
+			     RW_X_LATCH, &mtr);
+	buf_block_dbg_add_level(block, SYNC_NO_ORDER_CHECK);
+
+	doublewrite = buf_block_get_frame(block) + TRX_SYS_DOUBLEWRITE;
+
+	if (mach_read_from_4(doublewrite + TRX_SYS_DOUBLEWRITE_MAGIC)
+	    == TRX_SYS_DOUBLEWRITE_MAGIC_N) {
+		/* The doublewrite buffer has already been created:
+		just read in some numbers */
+
+		mtr_commit(&mtr);
+	} else {
+		fprintf(stderr,
+			"InnoDB: Doublewrite buffer not found in the doublewrite file:"
+			" creating new doublewrite buffer.\n");
+
+		if (buf_pool_get_curr_size()
+		    < ((2 * TRX_SYS_DOUBLEWRITE_BLOCK_SIZE
+			+ FSP_EXTENT_SIZE / 2 + 100)
+		       * UNIV_PAGE_SIZE)) {
+			fprintf(stderr,
+				"InnoDB: Cannot create the doublewrite buffer:"
+				" You must\n"
+				"InnoDB: increase your buffer pool size.\n"
+				"InnoDB: Cannot continue processing.\n");
+
+			exit(1);
+		}
+
+		block2 = fseg_create(TRX_DOUBLEWRITE_SPACE, TRX_SYS_PAGE_NO,
+				     TRX_SYS_DOUBLEWRITE
+				     + TRX_SYS_DOUBLEWRITE_FSEG, &mtr);
+
+		/* fseg_create acquires a second latch on the page,
+		therefore we must declare it: */
+
+		buf_block_dbg_add_level(block2, SYNC_NO_ORDER_CHECK);
+
+		if (block2 == NULL) {
+			fprintf(stderr,
+				"InnoDB: Cannot create the doublewrite buffer:"
+				" You must\n"
+				"InnoDB: increase your tablespace size.\n"
+				"InnoDB: Cannot continue processing.\n");
+
+			/* We exit without committing the mtr to prevent
+			its modifications to the database getting to disk */
+
+			exit(1);
+		}
+
+		fseg_header = buf_block_get_frame(block)
+			+ TRX_SYS_DOUBLEWRITE + TRX_SYS_DOUBLEWRITE_FSEG;
+		prev_page_no = 0;
+
+		for (i = 0; i < 2 * TRX_SYS_DOUBLEWRITE_BLOCK_SIZE
+			     + FSP_EXTENT_SIZE / 2; i++) {
+			page_no = fseg_alloc_free_page(fseg_header,
+						       prev_page_no + 1,
+						       FSP_UP, &mtr);
+			if (page_no == FIL_NULL) {
+				fprintf(stderr,
+					"InnoDB: Cannot create the doublewrite"
+					" buffer: You must\n"
+					"InnoDB: increase your"
+					" tablespace size.\n"
+					"InnoDB: Cannot continue operation.\n"
+					);
+
+				exit(1);
+			}
+
+			/* We read the allocated pages to the buffer pool;
+			when they are written to disk in a flush, the space
+			id and page number fields are also written to the
+			pages. When we at database startup read pages
+			from the doublewrite buffer, we know that if the
+			space id and page number in them are the same as
+			the page position in the tablespace, then the page
+			has not been written to in doublewrite. */
+
+#ifdef UNIV_SYNC_DEBUG
+			new_block =
+#endif /* UNIV_SYNC_DEBUG */
+			buf_page_get(TRX_DOUBLEWRITE_SPACE, 0, page_no,
+				     RW_X_LATCH, &mtr);
+			buf_block_dbg_add_level(new_block,
+						SYNC_NO_ORDER_CHECK);
+
+			if (i == FSP_EXTENT_SIZE / 2) {
+				ut_a(page_no == FSP_EXTENT_SIZE);
+				mlog_write_ulint(doublewrite
+						 + TRX_SYS_DOUBLEWRITE_BLOCK1,
+						 page_no, MLOG_4BYTES, &mtr);
+				mlog_write_ulint(doublewrite
+						 + TRX_SYS_DOUBLEWRITE_REPEAT
+						 + TRX_SYS_DOUBLEWRITE_BLOCK1,
+						 page_no, MLOG_4BYTES, &mtr);
+			} else if (i == FSP_EXTENT_SIZE / 2
+				   + TRX_SYS_DOUBLEWRITE_BLOCK_SIZE) {
+				ut_a(page_no == 2 * FSP_EXTENT_SIZE);
+				mlog_write_ulint(doublewrite
+						 + TRX_SYS_DOUBLEWRITE_BLOCK2,
+						 page_no, MLOG_4BYTES, &mtr);
+				mlog_write_ulint(doublewrite
+						 + TRX_SYS_DOUBLEWRITE_REPEAT
+						 + TRX_SYS_DOUBLEWRITE_BLOCK2,
+						 page_no, MLOG_4BYTES, &mtr);
+			} else if (i > FSP_EXTENT_SIZE / 2) {
+				ut_a(page_no == prev_page_no + 1);
+			}
+
+			prev_page_no = page_no;
+		}
+
+		mlog_write_ulint(doublewrite + TRX_SYS_DOUBLEWRITE_MAGIC,
+				 TRX_SYS_DOUBLEWRITE_MAGIC_N,
+				 MLOG_4BYTES, &mtr);
+		mlog_write_ulint(doublewrite + TRX_SYS_DOUBLEWRITE_MAGIC
+				 + TRX_SYS_DOUBLEWRITE_REPEAT,
+				 TRX_SYS_DOUBLEWRITE_MAGIC_N,
+				 MLOG_4BYTES, &mtr);
+
+		mlog_write_ulint(doublewrite
+				 + TRX_SYS_DOUBLEWRITE_SPACE_ID_STORED,
+				 TRX_SYS_DOUBLEWRITE_SPACE_ID_STORED_N,
+				 MLOG_4BYTES, &mtr);
+		mtr_commit(&mtr);
+
+		/* Flush the modified pages to disk and make a checkpoint */
+		log_make_checkpoint_at(IB_ULONGLONG_MAX, TRUE);
+
+		fprintf(stderr, "InnoDB: Doublewrite buffer created in the doublewrite file\n");
+		trx_sys_multiple_tablespace_format = TRUE;
+	}
+	trx_doublewrite_buf_is_being_created = FALSE;
+    }
 }
 
 /****************************************************************//**
@@ -438,10 +584,19 @@
 	ulint	source_page_no;
 	byte*	page;
 	byte*	doublewrite;
+	ulint	doublewrite_space_id;
 	ulint	space_id;
 	ulint	page_no;
 	ulint	i;
 
+	doublewrite_space_id = (srv_doublewrite_file ? TRX_DOUBLEWRITE_SPACE : TRX_SYS_SPACE);
+
+	if (srv_doublewrite_file) {
+		fprintf(stderr,
+			"InnoDB: doublewrite file '%s' is used.\n",
+			srv_doublewrite_file);
+	}
+
 	/* We do the file i/o past the buffer pool */
 
 	unaligned_read_buf = ut_malloc(2 * UNIV_PAGE_SIZE);
@@ -450,7 +605,7 @@
 	/* Read the trx sys header to check if we are using the doublewrite
 	buffer */
 
-	fil_io(OS_FILE_READ, TRUE, TRX_SYS_SPACE, 0, TRX_SYS_PAGE_NO, 0,
+	fil_io(OS_FILE_READ, TRUE, doublewrite_space_id, 0, TRX_SYS_PAGE_NO, 0,
 	       UNIV_PAGE_SIZE, read_buf, NULL);
 	doublewrite = read_buf + TRX_SYS_DOUBLEWRITE;
 
@@ -488,10 +643,10 @@
 
 	/* Read the pages from the doublewrite buffer to memory */
 
-	fil_io(OS_FILE_READ, TRUE, TRX_SYS_SPACE, 0, block1, 0,
+	fil_io(OS_FILE_READ, TRUE, doublewrite_space_id, 0, block1, 0,
 	       TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * UNIV_PAGE_SIZE,
 	       buf, NULL);
-	fil_io(OS_FILE_READ, TRUE, TRX_SYS_SPACE, 0, block2, 0,
+	fil_io(OS_FILE_READ, TRUE, doublewrite_space_id, 0, block2, 0,
 	       TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * UNIV_PAGE_SIZE,
 	       buf + TRX_SYS_DOUBLEWRITE_BLOCK_SIZE * UNIV_PAGE_SIZE,
 	       NULL);
@@ -547,7 +702,8 @@
 				" doublewrite buf.\n",
 				(ulong) space_id, (ulong) page_no, (ulong) i);
 
-		} else if (space_id == TRX_SYS_SPACE
+		} else if ((space_id == TRX_SYS_SPACE
+			    || (srv_doublewrite_file && space_id == TRX_DOUBLEWRITE_SPACE))
 			   && ((page_no >= block1
 				&& page_no
 				< block1 + TRX_SYS_DOUBLEWRITE_BLOCK_SIZE)
@@ -1016,6 +1172,83 @@
 }
 
 /*****************************************************************//**
+Creates dummy of the file page for the transaction system. */
+static
+void
+trx_sysf_dummy_create(
+/*==================*/
+	ulint	space,
+	mtr_t*	mtr)
+{
+	buf_block_t*	block;
+	page_t*		page;
+
+	ut_ad(mtr);
+
+	/* Note that below we first reserve the file space x-latch, and
+	then enter the kernel: we must do it in this order to conform
+	to the latching order rules. */
+
+	mtr_x_lock(fil_space_get_latch(space, NULL), mtr);
+	mutex_enter(&kernel_mutex);
+
+	/* Create the trx sys file block in a new allocated file segment */
+	block = fseg_create(space, 0, TRX_SYS + TRX_SYS_FSEG_HEADER,
+			    mtr);
+	buf_block_dbg_add_level(block, SYNC_TRX_SYS_HEADER);
+
+	fprintf(stderr, "%lu\n", buf_block_get_page_no(block));
+	ut_a(buf_block_get_page_no(block) == TRX_SYS_PAGE_NO);
+
+	page = buf_block_get_frame(block);
+
+	mlog_write_ulint(page + FIL_PAGE_TYPE, FIL_PAGE_TYPE_TRX_SYS,
+			 MLOG_2BYTES, mtr);
+
+	/* Reset the doublewrite buffer magic number to zero so that we
+	know that the doublewrite buffer has not yet been created (this
+	suppresses a Valgrind warning) */
+
+	mlog_write_ulint(page + TRX_SYS_DOUBLEWRITE
+			 + TRX_SYS_DOUBLEWRITE_MAGIC, 0, MLOG_4BYTES, mtr);
+
+#ifdef UNDEFINED
+	/* TODO: REMOVE IT: The bellow is not needed, I think */
+	sys_header = trx_sysf_get(mtr);
+
+	/* Start counting transaction ids from number 1 up */
+	mlog_write_dulint(sys_header + TRX_SYS_TRX_ID_STORE,
+			  ut_dulint_create(0, 1), mtr);
+
+	/* Reset the rollback segment slots */
+	for (i = 0; i < TRX_SYS_N_RSEGS; i++) {
+
+		trx_sysf_rseg_set_space(sys_header, i, ULINT_UNDEFINED, mtr);
+		trx_sysf_rseg_set_page_no(sys_header, i, FIL_NULL, mtr);
+	}
+
+	/* The remaining area (up to the page trailer) is uninitialized.
+	Silence Valgrind warnings about it. */
+	UNIV_MEM_VALID(sys_header + (TRX_SYS_RSEGS
+				     + TRX_SYS_N_RSEGS * TRX_SYS_RSEG_SLOT_SIZE
+				     + TRX_SYS_RSEG_SPACE),
+		       (UNIV_PAGE_SIZE - FIL_PAGE_DATA_END
+			- (TRX_SYS_RSEGS
+			   + TRX_SYS_N_RSEGS * TRX_SYS_RSEG_SLOT_SIZE
+			   + TRX_SYS_RSEG_SPACE))
+		       + page - sys_header);
+
+	/* Create the first rollback segment in the SYSTEM tablespace */
+	page_no = trx_rseg_header_create(space, 0, ULINT_MAX, &slot_no,
+					 mtr);
+	ut_a(slot_no == TRX_SYS_SYSTEM_RSEG_ID);
+	ut_a(page_no != FIL_NULL);
+#endif
+
+	mutex_exit(&kernel_mutex);
+}
+
+/*****************************************************************//**
 Creates and initializes the central memory structures for the transaction
 system. This is called when the database is started. */
 UNIV_INTERN
@@ -1387,6 +1620,26 @@
 	/* Does nothing at the moment */
 }
 
+/*****************************************************************//**
+Creates and initializes the dummy transaction system page for tablespace. */
+UNIV_INTERN
+void
+trx_sys_dummy_create(
+/*=================*/
+	ulint	space)
+{
+	mtr_t	mtr;
+
+	/* This function is only for doublewrite file for now */
+	ut_a(space == TRX_DOUBLEWRITE_SPACE);
+
+	mtr_start(&mtr);
+
+	trx_sysf_dummy_create(space, &mtr);
+
+	mtr_commit(&mtr);
+}
+
 /*********************************************************************
 Creates the rollback segments */
 UNIV_INTERN
--- /dev/null
+++ b/mysql-test/r/percona_innodb_doublewrite_file.result
@@ -0,0 +1,4 @@
+show variables like 'innodb_doublewrite%';
+Variable_name	Value
+innodb_doublewrite	ON
+innodb_doublewrite_file	ib_doublewrite
--- /dev/null
+++ b/mysql-test/t/percona_innodb_doublewrite_file-master.opt
@@ -0,0 +1 @@
+--innodb_doublewrite_file=ib_doublewrite
--- /dev/null
+++ b/mysql-test/t/percona_innodb_doublewrite_file.test
@@ -0,0 +1,2 @@
+--source include/have_innodb.inc
+show variables like 'innodb_doublewrite%';