~percona-dev/percona-server/release-5.5.11-20.2-fix-bug-764138 : contents of innodb_expand

~percona-dev/percona-server/release-5.5.11-20.2-fix-bug-764138 : (revision 97)
# name       : innodb_expand_import.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!
diff -ruN a/storage/innobase/fil/fil0fil.c b/storage/innobase/fil/fil0fil.c
--- a/storage/innobase/fil/fil0fil.c	2010-12-03 15:09:51.274957577 +0900
+++ b/storage/innobase/fil/fil0fil.c	2010-12-03 15:52:23.553986552 +0900
@@ -40,6 +40,12 @@
 #include "dict0dict.h"
 #include "page0page.h"
 #include "page0zip.h"
+#include "trx0trx.h"
+#include "trx0sys.h"
+#include "pars0pars.h"
+#include "row0mysql.h"
+#include "row0row.h"
+#include "que0que.h"
 #ifndef UNIV_HOTBACKUP
 # include "buf0lru.h"
 # include "ibuf0ibuf.h"
@@ -3078,7 +3084,7 @@
 
 	file = os_file_create_simple_no_error_handling(
 		innodb_file_data_key, filepath, OS_FILE_OPEN,
-		OS_FILE_READ_ONLY, &success);
+		OS_FILE_READ_WRITE, &success);
 	if (!success) {
 		/* The following call prints an error message */
 		os_file_get_last_error(TRUE);
@@ -3125,6 +3131,466 @@
 	space_id = fsp_header_get_space_id(page);
 	space_flags = fsp_header_get_flags(page);
 
+	if (srv_expand_import) {
+
+		ibool		file_is_corrupt = FALSE;
+		byte*		buf3;
+		byte*		descr_page;
+		ibool		descr_is_corrupt = FALSE;
+		index_id_t	old_id[31];
+		index_id_t	new_id[31];
+		ulint		root_page[31];
+		ulint		n_index;
+		os_file_t	info_file = -1;
+		char*		info_file_path;
+		ulint	i;
+		int		len;
+		ib_uint64_t	current_lsn;
+		ulint		size_low, size_high, size, free_limit;
+		ib_int64_t	size_bytes, free_limit_bytes;
+		dict_table_t*	table;
+		dict_index_t*	index;
+		fil_system_t*	system;
+		fil_node_t*	node = NULL;
+		fil_space_t*	space;
+
+		buf3 = ut_malloc(2 * UNIV_PAGE_SIZE);
+		descr_page = ut_align(buf3, UNIV_PAGE_SIZE);
+
+		current_lsn = log_get_lsn();
+
+		/* check the header page's consistency */
+		if (buf_page_is_corrupted(page,
+					  dict_table_flags_to_zip_size(space_flags))) {
+			fprintf(stderr, "InnoDB: page 0 of %s seems corrupt.\n", filepath);
+			file_is_corrupt = TRUE;
+			descr_is_corrupt = TRUE;
+		}
+
+		/* store as first descr page */
+		memcpy(descr_page, page, UNIV_PAGE_SIZE);
+
+		/* get free limit (page number) of the table space */
+/* these should be same to the definition in fsp0fsp.c */
+#define FSP_HEADER_OFFSET	FIL_PAGE_DATA
+#define	FSP_FREE_LIMIT		12
+		free_limit = mach_read_from_4(FSP_HEADER_OFFSET + FSP_FREE_LIMIT + page);
+		free_limit_bytes = (ib_int64_t)free_limit * (ib_int64_t)UNIV_PAGE_SIZE;
+
+		/* overwrite fsp header */
+		fsp_header_init_fields(page, id, flags);
+		mach_write_to_4(page + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID, id);
+		space_id = id;
+		space_flags = flags;
+		if (mach_read_from_8(page + FIL_PAGE_FILE_FLUSH_LSN) > current_lsn)
+			mach_write_to_8(page + FIL_PAGE_FILE_FLUSH_LSN, current_lsn);
+		mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
+				srv_use_checksums
+				? buf_calc_page_new_checksum(page)
+						: BUF_NO_CHECKSUM_MAGIC);
+		mach_write_to_4(page + UNIV_PAGE_SIZE - FIL_PAGE_END_LSN_OLD_CHKSUM,
+				srv_use_checksums
+				? buf_calc_page_old_checksum(page)
+						: BUF_NO_CHECKSUM_MAGIC);
+		success = os_file_write(filepath, file, page, 0, 0, UNIV_PAGE_SIZE);
+
+		/* get file size */
+		os_file_get_size(file, &size_low, &size_high);
+		size_bytes = (((ib_int64_t)size_high) << 32)
+				+ (ib_int64_t)size_low;
+
+		if (size_bytes < free_limit_bytes) {
+			free_limit_bytes = size_bytes;
+			if (size_bytes >= FSP_EXTENT_SIZE * UNIV_PAGE_SIZE) {
+				fprintf(stderr, "InnoDB: free limit of %s is larger than its real size.\n", filepath);
+				file_is_corrupt = TRUE;
+			}
+		}
+
+		/* get cruster index information */
+		table = dict_table_get_low(name);
+		index = dict_table_get_first_index(table);
+		ut_a(index->page==3);
+
+		/* read metadata from .exp file */
+		n_index = 0;
+		memset(old_id, 0, sizeof(old_id));
+		memset(new_id, 0, sizeof(new_id));
+		memset(root_page, 0, sizeof(root_page));
+
+		info_file_path = fil_make_ibd_name(name, FALSE);
+		len = strlen(info_file_path);
+		info_file_path[len - 3] = 'e';
+		info_file_path[len - 2] = 'x';
+		info_file_path[len - 1] = 'p';
+
+		info_file = os_file_create_simple_no_error_handling(innodb_file_data_key,
+				info_file_path, OS_FILE_OPEN, OS_FILE_READ_ONLY, &success);
+		if (!success) {
+			fprintf(stderr, "InnoDB: Cannot open the file: %s\n", info_file_path);
+			file_is_corrupt = TRUE;
+			goto skip_info;
+		}
+		success = os_file_read(info_file, page, 0, 0, UNIV_PAGE_SIZE);
+		if (!success) {
+			fprintf(stderr, "InnoDB: Cannot read the file: %s\n", info_file_path);
+			file_is_corrupt = TRUE;
+			goto skip_info;
+		}
+		if (mach_read_from_4(page) != 0x78706f72UL
+		    || mach_read_from_4(page + 4) != 0x74696e66UL) {
+			fprintf(stderr, "InnoDB: %s seems to be an incorrect .exp file.\n", info_file_path);
+			file_is_corrupt = TRUE;
+			goto skip_info;
+		}
+
+		fprintf(stderr, "InnoDB: Import: The extended import of %s is being started.\n", name);
+
+		n_index = mach_read_from_4(page + 8);
+		fprintf(stderr, "InnoDB: Import: %lu indexes have been detected.\n", (ulong)n_index);
+		for (i = 0; i < n_index; i++) {
+			new_id[i] =
+				dict_table_get_index_on_name(table,
+						(char*)(page + (i + 1) * 512 + 12))->id;
+			old_id[i] = mach_read_from_8(page + (i + 1) * 512);
+			root_page[i] = mach_read_from_4(page + (i + 1) * 512 + 8);
+		}
+
+skip_info:
+		if (info_file != -1)
+			os_file_close(info_file);
+
+		/*
+		if (size_bytes >= 1024 * 1024) {
+			size_bytes = ut_2pow_round(size_bytes, 1024 * 1024);
+		}
+		*/
+		if (!(flags & DICT_TF_ZSSIZE_MASK)) {
+			mem_heap_t*	heap = NULL;
+			ulint		offsets_[REC_OFFS_NORMAL_SIZE];
+			ulint*		offsets = offsets_;
+			ib_int64_t	offset;
+
+			size = (ulint) (size_bytes / UNIV_PAGE_SIZE);
+			/* over write space id of all pages */
+			rec_offs_init(offsets_);
+
+			fprintf(stderr, "InnoDB: Progress in %%:");
+
+			for (offset = 0; offset < free_limit_bytes; offset += UNIV_PAGE_SIZE) {
+				ulint		checksum_field;
+				ulint		old_checksum_field;
+				ibool		page_is_corrupt;
+
+				success = os_file_read(file, page,
+							(ulint)(offset & 0xFFFFFFFFUL),
+							(ulint)(offset >> 32), UNIV_PAGE_SIZE);
+
+				page_is_corrupt = FALSE;
+
+				/* check consistency */
+				if (memcmp(page + FIL_PAGE_LSN + 4,
+					   page + UNIV_PAGE_SIZE
+					   - FIL_PAGE_END_LSN_OLD_CHKSUM + 4, 4)) {
+
+					page_is_corrupt = TRUE;
+				}
+
+				if (mach_read_from_4(page + FIL_PAGE_OFFSET)
+				    != offset / UNIV_PAGE_SIZE) {
+
+					page_is_corrupt = TRUE;
+				}
+
+				checksum_field = mach_read_from_4(page
+								  + FIL_PAGE_SPACE_OR_CHKSUM);
+
+				old_checksum_field = mach_read_from_4(
+					page + UNIV_PAGE_SIZE
+					- FIL_PAGE_END_LSN_OLD_CHKSUM);
+
+				if (old_checksum_field != mach_read_from_4(page
+									   + FIL_PAGE_LSN)
+				    && old_checksum_field != BUF_NO_CHECKSUM_MAGIC
+				    && old_checksum_field
+				    != buf_calc_page_old_checksum(page)) {
+
+					page_is_corrupt = TRUE;
+				}
+
+				if (checksum_field != 0
+				    && checksum_field != BUF_NO_CHECKSUM_MAGIC
+				    && checksum_field
+				    != buf_calc_page_new_checksum(page)) {
+
+					page_is_corrupt = TRUE;
+				}
+
+				/* if it is free page, inconsistency is acceptable */
+				if (!offset) {
+					/* header page*/
+					/* it should be overwritten already */
+					ut_a(!page_is_corrupt);
+
+				} else if (!((offset / UNIV_PAGE_SIZE) % UNIV_PAGE_SIZE)) {
+					/* descr page (not header) */
+					if (page_is_corrupt) {
+						file_is_corrupt = TRUE;
+						descr_is_corrupt = TRUE;
+					} else {
+						ut_ad(fil_page_get_type(page) == FIL_PAGE_TYPE_XDES);
+						descr_is_corrupt = FALSE;
+					}
+
+					/* store as descr page */
+					memcpy(descr_page, page, UNIV_PAGE_SIZE);
+
+				} else if (descr_is_corrupt) {
+					/* unknown state of the page */
+					if (page_is_corrupt) {
+						file_is_corrupt = TRUE;
+					}
+
+				} else {
+					/* check free page or not */
+					/* These definitions should be same to fsp0fsp.c */
+#define	FSP_HEADER_SIZE		(32 + 5 * FLST_BASE_NODE_SIZE)
+
+#define	XDES_BITMAP		(FLST_NODE_SIZE + 12)
+#define	XDES_BITS_PER_PAGE	2
+#define	XDES_FREE_BIT		0
+#define	XDES_SIZE							\
+	(XDES_BITMAP + UT_BITS_IN_BYTES(FSP_EXTENT_SIZE * XDES_BITS_PER_PAGE))
+#define	XDES_ARR_OFFSET		(FSP_HEADER_OFFSET + FSP_HEADER_SIZE)
+
+					/*descr = descr_page + XDES_ARR_OFFSET + XDES_SIZE * xdes_calc_descriptor_index(zip_size, offset)*/
+					/*xdes_get_bit(descr, XDES_FREE_BIT, page % FSP_EXTENT_SIZE, mtr)*/
+					byte*	descr;
+					ulint	index;
+					ulint	byte_index;
+					ulint	bit_index;
+
+					descr = descr_page + XDES_ARR_OFFSET
+						+ XDES_SIZE * (ut_2pow_remainder((offset / UNIV_PAGE_SIZE), UNIV_PAGE_SIZE) / FSP_EXTENT_SIZE);
+
+					index = XDES_FREE_BIT + XDES_BITS_PER_PAGE * ((offset / UNIV_PAGE_SIZE) % FSP_EXTENT_SIZE);
+					byte_index = index / 8;
+					bit_index = index % 8;
+
+					if (ut_bit_get_nth(mach_read_from_1(descr + XDES_BITMAP + byte_index), bit_index)) {
+						/* free page */
+						if (page_is_corrupt) {
+							goto skip_write;
+						}
+					} else {
+						/* not free */
+						if (page_is_corrupt) {
+							file_is_corrupt = TRUE;
+						}
+					}
+				}
+
+				if (page_is_corrupt) {
+					fprintf(stderr, " [errp:%lld]", offset / UNIV_PAGE_SIZE);
+
+					/* cannot treat corrupt page */
+					goto skip_write;
+				}
+
+				if (mach_read_from_4(page + FIL_PAGE_OFFSET) || !offset) {
+					mach_write_to_4(page + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID, id);
+
+					for (i = 0; i < n_index; i++) {
+						if (offset / UNIV_PAGE_SIZE == root_page[i]) {
+							/* this is index root page */
+							mach_write_to_4(page + FIL_PAGE_DATA + PAGE_BTR_SEG_LEAF
+											+ FSEG_HDR_SPACE, id);
+							mach_write_to_4(page + FIL_PAGE_DATA + PAGE_BTR_SEG_TOP
+											+ FSEG_HDR_SPACE, id);
+							break;
+						}
+					}
+
+					if (fil_page_get_type(page) == FIL_PAGE_INDEX) {
+						index_id_t tmp = mach_read_from_8(page + (PAGE_HEADER + PAGE_INDEX_ID));
+
+						if (mach_read_from_2(page + PAGE_HEADER + PAGE_LEVEL) == 0
+						    && old_id[0] == tmp) {
+							/* leaf page of cluster index, reset trx_id of records */
+							rec_t*	rec;
+							rec_t*	supremum;
+							ulint	n_recs;
+
+							supremum = page_get_supremum_rec(page);
+							rec = page_rec_get_next(page_get_infimum_rec(page));
+							n_recs = page_get_n_recs(page);
+
+							while (rec && rec != supremum && n_recs > 0) {
+								ulint	n_fields;
+								ulint	i;
+								ulint	offset = index->trx_id_offset;
+								offsets = rec_get_offsets(rec, index, offsets,
+										ULINT_UNDEFINED, &heap);
+								n_fields = rec_offs_n_fields(offsets);
+								if (!offset) {
+									offset = row_get_trx_id_offset(rec, index, offsets);
+								}
+								trx_write_trx_id(rec + offset, 1);
+
+								for (i = 0; i < n_fields; i++) {
+									if (rec_offs_nth_extern(offsets, i)) {
+										ulint	local_len;
+										byte*	data;
+
+										data = rec_get_nth_field(rec, offsets, i, &local_len);
+
+										local_len -= BTR_EXTERN_FIELD_REF_SIZE;
+
+										mach_write_to_4(data + local_len + BTR_EXTERN_SPACE_ID, id);
+									}
+								}
+
+								rec = page_rec_get_next(rec);
+								n_recs--;
+							}
+						}
+
+						for (i = 0; i < n_index; i++) {
+							if (old_id[i] == tmp) {
+								mach_write_to_8(page + (PAGE_HEADER + PAGE_INDEX_ID), new_id[i]);
+								break;
+							}
+						}
+					}
+
+					if (mach_read_from_8(page + FIL_PAGE_LSN) > current_lsn) {
+						mach_write_to_8(page + FIL_PAGE_LSN, current_lsn);
+						mach_write_to_8(page + UNIV_PAGE_SIZE - FIL_PAGE_END_LSN_OLD_CHKSUM,
+										current_lsn);
+					}
+
+					mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
+							srv_use_checksums
+							? buf_calc_page_new_checksum(page)
+									: BUF_NO_CHECKSUM_MAGIC);
+					mach_write_to_4(page + UNIV_PAGE_SIZE - FIL_PAGE_END_LSN_OLD_CHKSUM,
+							srv_use_checksums
+							? buf_calc_page_old_checksum(page)
+									: BUF_NO_CHECKSUM_MAGIC);
+
+					success = os_file_write(filepath, file, page,
+								(ulint)(offset & 0xFFFFFFFFUL),
+								(ulint)(offset >> 32), UNIV_PAGE_SIZE);
+				}
+
+skip_write:
+				if (free_limit_bytes
+				    && ((ib_int64_t)((offset + UNIV_PAGE_SIZE) * 100) / free_limit_bytes)
+					!= ((offset * 100) / free_limit_bytes)) {
+					fprintf(stderr, " %lu",
+						(ulong)((ib_int64_t)((offset + UNIV_PAGE_SIZE) * 100) / free_limit_bytes));
+				}
+			}
+
+			fprintf(stderr, " done.\n");
+
+			/* update SYS_INDEXES set root page */
+			index = dict_table_get_first_index(table);
+			while (index) {
+				for (i = 0; i < n_index; i++) {
+					if (new_id[i] == index->id) {
+						break;
+					}
+				}
+
+				if (i != n_index
+				    && root_page[i] != index->page) {
+					/* must update */
+					ulint	error;
+					trx_t*	trx;
+					pars_info_t*	info = NULL;
+
+					trx = trx_allocate_for_mysql();
+					trx->op_info = "extended import";
+
+					info = pars_info_create();
+
+					pars_info_add_ull_literal(info, "indexid", new_id[i]);
+					pars_info_add_int4_literal(info, "new_page", (lint) root_page[i]);
+
+					error = que_eval_sql(info,
+						"PROCEDURE UPDATE_INDEX_PAGE () IS\n"
+						"BEGIN\n"
+						"UPDATE SYS_INDEXES"
+						" SET PAGE_NO = :new_page"
+						" WHERE ID = :indexid;\n"
+						"COMMIT WORK;\n"
+						"END;\n",
+						FALSE, trx);
+
+					if (error != DB_SUCCESS) {
+						fprintf(stderr, "InnoDB: failed to update SYS_INDEXES\n");
+					}
+
+					trx_commit_for_mysql(trx);
+
+					trx_free_for_mysql(trx);
+
+					index->page = root_page[i];
+				}
+
+				index = dict_table_get_next_index(index);
+			}
+			if (UNIV_LIKELY_NULL(heap)) {
+				mem_heap_free(heap);
+			}
+		} else {
+			/* zip page? */
+			size = (ulint)
+			(size_bytes
+					/ dict_table_flags_to_zip_size(flags));
+			fprintf(stderr, "InnoDB: Import: The table %s seems to be in a newer format."
+					" It may not be possible to process it.\n", name);
+		}
+		/* .exp file should be removed */
+		success = os_file_delete(info_file_path);
+		if (!success) {
+			success = os_file_delete_if_exists(info_file_path);
+		}
+		mem_free(info_file_path);
+
+		system	= fil_system;
+		mutex_enter(&(system->mutex));
+		space = fil_space_get_by_id(id);
+		if (space)
+			node = UT_LIST_GET_FIRST(space->chain);
+		if (node && node->size < size) {
+			space->size += (size - node->size);
+			node->size = size;
+		}
+		mutex_exit(&(system->mutex));
+
+		ut_free(buf3);
+
+		if (file_is_corrupt) {
+			ut_print_timestamp(stderr);
+			fputs("  InnoDB: Error: file ",
+			      stderr);
+			ut_print_filename(stderr, filepath);
+			fprintf(stderr, " seems to be corrupt.\n"
+				"InnoDB: An attempt to convert and salvage all corrupt pages was not made.\n"
+				"InnoDB: ##### CAUTION #####\n"
+				"InnoDB: ## The .ibd file may cause InnoDB to crash, even though its re-import seems to have succeeded.\n"
+				"InnoDB: ## If you don't know how to salvage data from a .ibd, you should not use the file.\n"
+				"InnoDB: ###################\n");
+			success = FALSE;
+
+			ut_free(buf2);
+
+			goto func_exit;
+		}
+	}
+
 	ut_free(buf2);
 
 	if (UNIV_UNLIKELY(space_id != id
diff -ruN a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc
--- a/storage/innobase/handler/ha_innodb.cc	2010-12-03 15:49:59.195023983 +0900
+++ b/storage/innobase/handler/ha_innodb.cc	2010-12-03 15:52:23.555957062 +0900
@@ -7356,6 +7356,14 @@
 		err = row_discard_tablespace_for_mysql(dict_table->name, trx);
 	} else {
 		err = row_import_tablespace_for_mysql(dict_table->name, trx);
+
+		/* in expanded import mode re-initialize auto_increment again */
+		if ((err == DB_SUCCESS) && srv_expand_import &&
+		    (table->found_next_number_field != NULL)) {
+			dict_table_autoinc_lock(dict_table);
+			innobase_initialize_autoinc();
+			dict_table_autoinc_unlock(dict_table);
+		}
 	}
 
 	err = convert_error_code_to_mysql(err, dict_table->flags, NULL);
@@ -11618,6 +11626,11 @@
   "Choose method of innodb_adaptive_flushing. (native, [estimate], keep_average)",
   NULL, innodb_adaptive_flushing_method_update, 1, &adaptive_flushing_method_typelib);
 
+static MYSQL_SYSVAR_ULONG(import_table_from_xtrabackup, srv_expand_import,
+  PLUGIN_VAR_RQCMDARG,
+  "Enable/Disable converting automatically *.ibd files when import tablespace.",
+  NULL, NULL, 0, 0, 1, 0);
+
 static MYSQL_SYSVAR_ULONG(extra_rsegments, srv_extra_rsegments,
   PLUGIN_VAR_RQCMDARG,
   "Number of extra user rollback segments which are used in a round-robin fashion.",
@@ -11695,6 +11708,7 @@
   MYSQL_SYSVAR(flush_neighbor_pages),
   MYSQL_SYSVAR(read_ahead),
   MYSQL_SYSVAR(adaptive_flushing_method),
+  MYSQL_SYSVAR(import_table_from_xtrabackup),
   MYSQL_SYSVAR(extra_rsegments),
   MYSQL_SYSVAR(dict_size_limit),
   MYSQL_SYSVAR(use_sys_malloc),
diff -ruN a/storage/innobase/include/srv0srv.h b/storage/innobase/include/srv0srv.h
--- a/storage/innobase/include/srv0srv.h	2010-12-03 15:48:03.077954270 +0900
+++ b/storage/innobase/include/srv0srv.h	2010-12-03 15:52:23.561986996 +0900
@@ -232,6 +232,8 @@
 extern ulint	srv_read_ahead;
 extern ulint	srv_adaptive_flushing_method;
 
+extern ulint	srv_expand_import;
+
 extern ulint	srv_extra_rsegments;
 extern ulint	srv_dict_size_limit;
 /*-------------------------------------------*/
diff -ruN a/storage/innobase/srv/srv0srv.c b/storage/innobase/srv/srv0srv.c
--- a/storage/innobase/srv/srv0srv.c	2010-12-03 15:49:59.230956118 +0900
+++ b/storage/innobase/srv/srv0srv.c	2010-12-03 15:52:23.562954411 +0900
@@ -420,6 +420,8 @@
 UNIV_INTERN ulint	srv_read_ahead = 3; /* 1: random  2: linear  3: Both */
 UNIV_INTERN ulint	srv_adaptive_flushing_method = 0; /* 0: native  1: estimate  2: keep_average */
 
+UNIV_INTERN ulint	srv_expand_import = 0; /* 0:disable 1:enable */
+
 UNIV_INTERN ulint	srv_extra_rsegments = 127; /* extra rseg for users */
 UNIV_INTERN ulint	srv_dict_size_limit = 0;
 /*-------------------------------------------*/