~percona-dev/percona-server/bug693818 : contents of innodb_buffer_pool

~percona-dev/percona-server/bug693818 : (revision 182)
# name       : innodb_buffer_pool_shm.patch
# introduced : 12
# 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/innodb_plugin/buf/buf0buf.c b/storage/innodb_plugin/buf/buf0buf.c
--- a/storage/innodb_plugin/buf/buf0buf.c	2010-07-14 16:32:49.669501663 +0900
+++ b/storage/innodb_plugin/buf/buf0buf.c	2010-07-14 16:40:16.149438645 +0900
@@ -53,6 +53,10 @@
 #include "page0zip.h"
 #include "trx0trx.h"
 #include "srv0start.h"
+#include "que0que.h"
+#include "read0read.h"
+#include "row0row.h"
+#include "ha_prototypes.h"
 
 /* prototypes for new functions added to ha_innodb.cc */
 trx_t* innobase_get_trx();
@@ -310,6 +314,30 @@
 UNIV_INTERN ibool		buf_debug_prints = FALSE;
 #endif /* UNIV_DEBUG */
 
+/* Buffer pool shared memory segment information */
+typedef	struct buf_shm_info_struct	buf_shm_info_t;
+
+struct buf_shm_info_struct {
+	char	head_str[8];
+	ulint	binary_id;
+	ibool	is_new;		/* during initializing */
+	ibool	clean;		/* clean shutdowned and free */
+	ibool	reusable;	/* reusable */
+	ulint	buf_pool_size;	/* backup value */
+	ulint	page_size;	/* backup value */
+	ulint	frame_offset;	/* offset of the first frame based on chunk->mem */
+	ulint	zip_hash_offset;
+	ulint	zip_hash_n;
+
+	ulint	checksum;
+
+	buf_pool_t	buf_pool_backup;
+	buf_chunk_t	chunk_backup;
+
+	ib_uint64_t	dummy;
+};
+
+#define BUF_SHM_INFO_HEAD "XTRA_SHM"
 #endif /* !UNIV_HOTBACKUP */
 
 /********************************************************************//**
@@ -756,6 +784,45 @@
 #endif /* UNIV_SYNC_DEBUG */
 }
 
+static
+void
+buf_block_reuse(
+/*============*/
+	buf_block_t*	block,
+	ptrdiff_t	frame_offset)
+{
+	/* block_init */
+	block->frame += frame_offset;
+
+	UNIV_MEM_DESC(block->frame, UNIV_PAGE_SIZE, block);
+
+	block->index = NULL;
+
+#ifdef UNIV_DEBUG
+	/* recreate later */
+	block->page.in_page_hash = FALSE;
+	block->page.in_zip_hash = FALSE;
+#endif /* UNIV_DEBUG */
+
+#if defined UNIV_AHI_DEBUG || defined UNIV_DEBUG
+	block->n_pointers = 0;
+#endif /* UNIV_AHI_DEBUG || UNIV_DEBUG */
+
+	if (block->page.zip.data)
+		block->page.zip.data += frame_offset;
+
+	block->is_hashed = FALSE;
+
+	mutex_create(&block->mutex, SYNC_BUF_BLOCK);
+
+	rw_lock_create(&block->lock, SYNC_LEVEL_VARYING);
+	ut_ad(rw_lock_validate(&(block->lock)));
+
+#ifdef UNIV_SYNC_DEBUG
+	rw_lock_create(&block->debug_latch, SYNC_NO_ORDER_CHECK);
+#endif /* UNIV_SYNC_DEBUG */
+}
+
 /********************************************************************//**
 Allocates a chunk of buffer frames.
 @return	chunk, or NULL on failure */
@@ -768,26 +835,188 @@
 {
 	buf_block_t*	block;
 	byte*		frame;
+	ulint		zip_hash_n = 0;
+	ulint		zip_hash_mem_size = 0;
+	hash_table_t*	zip_hash_tmp = NULL;
 	ulint		i;
+	buf_shm_info_t*	shm_info = NULL;
 
 	/* Round down to a multiple of page size,
 	although it already should be. */
 	mem_size = ut_2pow_round(mem_size, UNIV_PAGE_SIZE);
+
+	srv_buffer_pool_shm_is_reused = FALSE;
+
+	if (srv_buffer_pool_shm_key) {
+		/* zip_hash size */
+		zip_hash_n = (mem_size / UNIV_PAGE_SIZE) * 2;
+		zip_hash_mem_size = ut_2pow_round(hash_create_needed(zip_hash_n)
+						  + (UNIV_PAGE_SIZE - 1), UNIV_PAGE_SIZE);
+	}
+
 	/* Reserve space for the block descriptors. */
 	mem_size += ut_2pow_round((mem_size / UNIV_PAGE_SIZE) * (sizeof *block)
 				  + (UNIV_PAGE_SIZE - 1), UNIV_PAGE_SIZE);
+	if (srv_buffer_pool_shm_key) {
+		 mem_size += ut_2pow_round(sizeof(buf_shm_info_t)
+					   + (UNIV_PAGE_SIZE - 1), UNIV_PAGE_SIZE);
+		 mem_size += zip_hash_mem_size;
+	}
 
 	chunk->mem_size = mem_size;
+
+	if (srv_buffer_pool_shm_key) {
+		ulint	binary_id;
+		ibool	is_new;
+
+		ut_a(buf_pool->n_chunks == 1);
+
+		fprintf(stderr,
+		"InnoDB: Warning: The innodb_buffer_pool_shm_key option has been specified.\n"
+		"InnoDB: Do not change the following between restarts of the server while this option is being used:\n"
+		"InnoDB:   * the mysqld executable between restarts of the server.\n"
+		"InnoDB:   * the value of innodb_buffer_pool_size.\n"
+		"InnoDB:   * the value of innodb_page_size.\n"
+		"InnoDB:   * datafiles created by InnoDB during this session.\n"
+		"InnoDB: Otherwise, data corruption in datafiles may result.\n");
+
+		/* FIXME: This is vague id still */
+		binary_id = (ulint) ((byte*)mtr_commit - (byte*)btr_root_get)
+			  + (ulint) ((byte*)os_get_os_version - (byte*)buf_calc_page_new_checksum)
+			  + (ulint) ((byte*)page_dir_find_owner_slot - (byte*)dfield_data_is_binary_equal)
+			  + (ulint) ((byte*)que_graph_publish - (byte*)dict_casedn_str)
+			  + (ulint) ((byte*)read_view_oldest_copy_or_open_new - (byte*)fil_space_get_version)
+			  + (ulint) ((byte*)rec_get_n_extern_new - (byte*)fsp_get_size_low)
+			  + (ulint) ((byte*)row_get_trx_id_offset - (byte*)ha_create_func)
+			  + (ulint) ((byte*)srv_set_io_thread_op_info - (byte*)thd_is_replication_slave_thread)
+			  + (ulint) ((byte*)mutex_create_func - (byte*)ibuf_inside)
+			  + (ulint) ((byte*)trx_set_detailed_error - (byte*)lock_check_trx_id_sanity)
+			  + (ulint) ((byte*)ut_time - (byte*)mem_heap_strdup);
+
+		chunk->mem = os_shm_alloc(&chunk->mem_size, srv_buffer_pool_shm_key, &is_new);
+
+		if (UNIV_UNLIKELY(chunk->mem == NULL)) {
+			return(NULL);
+		}
+init_again:
+#ifdef UNIV_SET_MEM_TO_ZERO
+		if (is_new) {
+			memset(chunk->mem, '\0', chunk->mem_size);
+		}
+#endif
+		/* for ut_fold_binary_32(), these values should be 32-bit aligned */
+		ut_a(sizeof(buf_shm_info_t) % 4 == 0);
+		ut_a((ulint)chunk->mem % 4 == 0);
+		ut_a(chunk->mem_size % 4 == 0);
+
+		shm_info = chunk->mem;
+
+		zip_hash_tmp = (hash_table_t*)((byte*)chunk->mem + chunk->mem_size - zip_hash_mem_size);
+
+		if (is_new) {
+			strncpy(shm_info->head_str, BUF_SHM_INFO_HEAD, 8);
+			shm_info->binary_id = binary_id;
+			shm_info->is_new = TRUE;	/* changed to FALSE when the initialization is finished */
+			shm_info->clean = FALSE;	/* changed to TRUE when free the segment. */
+			shm_info->reusable = FALSE;	/* changed to TRUE when validation is finished. */
+			shm_info->buf_pool_size = srv_buf_pool_size;
+			shm_info->page_size = srv_page_size;
+			shm_info->zip_hash_offset = chunk->mem_size - zip_hash_mem_size;
+			shm_info->zip_hash_n = zip_hash_n;
+		} else {
+			ulint	checksum;
+
+			if (strncmp(shm_info->head_str, BUF_SHM_INFO_HEAD, 8)) {
+				fprintf(stderr,
+				"InnoDB: Error: The shared memory segment seems not to be for buffer pool.\n");
+				return(NULL);
+			}
+			if (shm_info->binary_id != binary_id) {
+				fprintf(stderr,
+				"InnoDB: Error: The shared memory segment seems not to be for this binary.\n");
+				return(NULL);
+			}
+			if (shm_info->is_new) {
+				fprintf(stderr,
+				"InnoDB: Error: The shared memory was not initialized yet.\n");
+				return(NULL);
+			}
+			if (shm_info->buf_pool_size != srv_buf_pool_size) {
+				fprintf(stderr,
+				"InnoDB: Error: srv_buf_pool_size is different (shm=%lu current=%lu).\n",
+				shm_info->buf_pool_size, srv_buf_pool_size);
+				return(NULL);
+			}
+			if (shm_info->page_size != srv_page_size) {
+				fprintf(stderr,
+				"InnoDB: Error: srv_page_size is different (shm=%lu current=%lu).\n",
+				shm_info->page_size, srv_page_size);
+				return(NULL);
+			}
+			if (!shm_info->reusable) {
+				fprintf(stderr,
+				"InnoDB: Warning: The shared memory has unrecoverable contents.\n"
+				"InnoDB: The shared memory segment is initialized.\n");
+				is_new = TRUE;
+				goto init_again;
+			}
+			if (!shm_info->clean) {
+				fprintf(stderr,
+				"InnoDB: Warning: The shared memory was not shut down cleanly.\n"
+				"InnoDB: The shared memory segment is initialized.\n");
+				is_new = TRUE;
+				goto init_again;
+			}
+
+			ut_a(shm_info->zip_hash_offset == chunk->mem_size - zip_hash_mem_size);
+			ut_a(shm_info->zip_hash_n == zip_hash_n);
+
+			/* check checksum */
+			if (srv_buffer_pool_shm_checksum) {
+				checksum = ut_fold_binary_32((byte*)chunk->mem + sizeof(buf_shm_info_t),
+							     chunk->mem_size - sizeof(buf_shm_info_t));
+			} else {
+				checksum = BUF_NO_CHECKSUM_MAGIC;
+			}
+
+			if (shm_info->checksum != BUF_NO_CHECKSUM_MAGIC
+			    && shm_info->checksum != checksum) {
+				fprintf(stderr,
+				"InnoDB: Error: checksum of the shared memory is not match. "
+				"(stored=%lu calculated=%lu)\n",
+				shm_info->checksum, checksum);
+				return(NULL);
+			}
+
+			/* flag to use the segment. */
+			shm_info->clean = FALSE;	/* changed to TRUE when free the segment. */
+		}
+
+		/* init zip_hash contents */
+		if (is_new) {
+			hash_create_init(zip_hash_tmp, zip_hash_n);
+		} else {
+			/* adjust offset is done later */
+			hash_create_reuse(zip_hash_tmp);
+
+			srv_buffer_pool_shm_is_reused = TRUE;
+		}
+	} else {
 	chunk->mem = os_mem_alloc_large(&chunk->mem_size);
 
 	if (UNIV_UNLIKELY(chunk->mem == NULL)) {
 
 		return(NULL);
 	}
+	}
 
 	/* Allocate the block descriptors from
 	the start of the memory block. */
+	if (srv_buffer_pool_shm_key) {
+		chunk->blocks = (buf_block_t*)((byte*)chunk->mem + sizeof(buf_shm_info_t));
+	} else {
 	chunk->blocks = chunk->mem;
+	}
 
 	/* Align a pointer to the first frame.  Note that when
 	os_large_page_size is smaller than UNIV_PAGE_SIZE,
@@ -795,8 +1024,13 @@
 	it is bigger, we may allocate more blocks than requested. */
 
 	frame = ut_align(chunk->mem, UNIV_PAGE_SIZE);
+	if (srv_buffer_pool_shm_key) {
+		/* reserve zip_hash space and always -1 for reproductibity */
+		chunk->size = (chunk->mem_size - zip_hash_mem_size) / UNIV_PAGE_SIZE - 1;
+	} else {
 	chunk->size = chunk->mem_size / UNIV_PAGE_SIZE
 		- (frame != chunk->mem);
+	}
 
 	/* Subtract the space needed for block descriptors. */
 	{
@@ -810,6 +1044,98 @@
 		chunk->size = size;
 	}
 
+	if (shm_info && !(shm_info->is_new)) {
+		/* convert the shared memory segment for reuse */
+		ptrdiff_t	phys_offset;
+		ptrdiff_t	logi_offset;
+		ptrdiff_t	blocks_offset;
+		void*		previous_frame_address;
+
+		if (chunk->size < shm_info->chunk_backup.size) {
+			fprintf(stderr,
+			"InnoDB: Error: The buffer pool became smaller because of allocated address.\n"
+			"InnoDB: Retrying may avoid this situation.\n");
+			shm_info->clean = TRUE; /* release the flag for retrying */
+			return(NULL);
+		}
+
+		chunk->size = shm_info->chunk_backup.size;
+		phys_offset = frame - ((byte*)chunk->mem + shm_info->frame_offset);
+		logi_offset = frame - chunk->blocks[0].frame;
+		previous_frame_address = chunk->blocks[0].frame;
+		blocks_offset = (byte*)chunk->blocks - (byte*)shm_info->chunk_backup.blocks;
+
+		if (phys_offset || logi_offset || blocks_offset) {
+			fprintf(stderr,
+			"InnoDB: Buffer pool in the shared memory segment should be converted.\n"
+			"InnoDB: Previous frames in address      : %p\n"
+			"InnoDB: Previous frames were located    : %p\n"
+			"InnoDB: Current frames should be located: %p\n"
+			"InnoDB: Pysical offset                  : %ld (%#lx)\n"
+			"InnoDB: Logical offset (frames)         : %ld (%#lx)\n"
+			"InnoDB: Logical offset (blocks)         : %ld (%#lx)\n",
+				(byte*)chunk->mem + shm_info->frame_offset,
+				chunk->blocks[0].frame, frame,
+				phys_offset, phys_offset, logi_offset, logi_offset,
+				blocks_offset, blocks_offset);
+		} else {
+			fprintf(stderr,
+			"InnoDB: Buffer pool in the shared memory segment can be used as it is.\n");
+		}
+
+		if (phys_offset) {
+			fprintf(stderr,
+			"InnoDB: Aligning physical offset...");
+
+			memmove(frame, (byte*)chunk->mem + shm_info->frame_offset,
+				chunk->size * UNIV_PAGE_SIZE);
+
+			fprintf(stderr,
+			" Done.\n");
+		}
+
+		/* buf_block_t */
+		block = chunk->blocks;
+		for (i = chunk->size; i--; ) {
+			buf_block_reuse(block, logi_offset);
+			block++;
+		}
+
+		if (logi_offset || blocks_offset) {
+			fprintf(stderr,
+			"InnoDB: Aligning logical offset...");
+
+
+			/* buf_pool_t buf_pool_backup */
+			UT_LIST_OFFSET(flush_list, buf_page_t, shm_info->buf_pool_backup.flush_list,
+					previous_frame_address, logi_offset, blocks_offset);
+			UT_LIST_OFFSET(free, buf_page_t, shm_info->buf_pool_backup.free,
+					previous_frame_address, logi_offset, blocks_offset);
+			UT_LIST_OFFSET(LRU, buf_page_t, shm_info->buf_pool_backup.LRU,
+					previous_frame_address, logi_offset, blocks_offset);
+			if (shm_info->buf_pool_backup.LRU_old)
+				shm_info->buf_pool_backup.LRU_old =
+					(buf_page_t*)((byte*)(shm_info->buf_pool_backup.LRU_old)
+						+ (((void*)shm_info->buf_pool_backup.LRU_old > previous_frame_address)
+						  ? logi_offset : blocks_offset));
+
+			UT_LIST_OFFSET(unzip_LRU, buf_block_t, shm_info->buf_pool_backup.unzip_LRU,
+					previous_frame_address, logi_offset, blocks_offset);
+
+			UT_LIST_OFFSET(zip_list, buf_page_t, shm_info->buf_pool_backup.zip_clean,
+					previous_frame_address, logi_offset, blocks_offset);
+			for (i = 0; i < BUF_BUDDY_SIZES_MAX; i++) {
+				UT_LIST_OFFSET(zip_list, buf_page_t, shm_info->buf_pool_backup.zip_free[i],
+					previous_frame_address, logi_offset, blocks_offset);
+			}
+
+			HASH_OFFSET(zip_hash_tmp, buf_page_t, hash,
+					previous_frame_address, logi_offset, blocks_offset);
+
+			fprintf(stderr,
+			" Done.\n");
+		}
+	} else {
 	/* Init block structs and assign frames for them. Then we
 	assign the frames to the first blocks (we already mapped the
 	memory above). */
@@ -833,6 +1159,11 @@
 		block++;
 		frame += UNIV_PAGE_SIZE;
 	}
+	}
+
+	if (shm_info) {
+		shm_info->frame_offset = chunk->blocks[0].frame - (byte*)chunk->mem;
+	}
 
 	return(chunk);
 }
@@ -1014,6 +1345,8 @@
 		UNIV_MEM_UNDESC(block);
 	}
 
+	ut_a(!srv_buffer_pool_shm_key);
+
 	os_mem_free_large(chunk->mem, chunk->mem_size);
 }
 
@@ -1063,7 +1396,10 @@
 	srv_buf_pool_curr_size = buf_pool->curr_size * UNIV_PAGE_SIZE;
 
 	buf_pool->page_hash = hash_create(2 * buf_pool->curr_size);
+	/* zip_hash is allocated to shm when srv_buffer_pool_shm_key is enabled */
+	if (!srv_buffer_pool_shm_key) {
 	buf_pool->zip_hash = hash_create(2 * buf_pool->curr_size);
+	}
 
 	buf_pool->last_printout_time = time(NULL);
 
@@ -1078,6 +1414,86 @@
 	--------------------------- */
 	/* All fields are initialized by mem_zalloc(). */
 
+	if (srv_buffer_pool_shm_key) {
+		buf_shm_info_t*	shm_info;
+
+		ut_a((byte*)chunk->blocks == (byte*)chunk->mem + sizeof(buf_shm_info_t));
+		shm_info = chunk->mem;
+
+		buf_pool->zip_hash = (hash_table_t*)((byte*)chunk->mem + shm_info->zip_hash_offset);
+
+		if(shm_info->is_new) {
+			shm_info->is_new = FALSE; /* initialization was finished */
+		} else {
+			buf_block_t*	block = chunk->blocks;
+			buf_page_t*	b;
+
+			/* shm_info->buf_pool_backup should be converted */
+			/* at buf_chunk_init(). So copy simply. */
+			buf_pool->flush_list 		= shm_info->buf_pool_backup.flush_list;
+			buf_pool->freed_page_clock 	= shm_info->buf_pool_backup.freed_page_clock;
+			buf_pool->free			= shm_info->buf_pool_backup.free;
+			buf_pool->LRU			= shm_info->buf_pool_backup.LRU;
+			buf_pool->LRU_old		= shm_info->buf_pool_backup.LRU_old;
+			buf_pool->LRU_old_len		= shm_info->buf_pool_backup.LRU_old_len;
+			buf_pool->unzip_LRU		= shm_info->buf_pool_backup.unzip_LRU;
+			buf_pool->zip_clean		= shm_info->buf_pool_backup.zip_clean;
+			for (i = 0; i < BUF_BUDDY_SIZES_MAX; i++) {
+				buf_pool->zip_free[i]	= shm_info->buf_pool_backup.zip_free[i];
+			}
+
+			for (i = 0; i < chunk->size; i++, block++) {
+				if (buf_block_get_state(block)
+				    == BUF_BLOCK_FILE_PAGE) {
+					ut_d(block->page.in_page_hash = TRUE);
+					HASH_INSERT(buf_page_t, hash, buf_pool->page_hash,
+						    buf_page_address_fold(
+							    block->page.space,
+							    block->page.offset),
+						    &block->page);
+				}
+			}
+
+			for (b = UT_LIST_GET_FIRST(buf_pool->zip_clean); b;
+			     b = UT_LIST_GET_NEXT(zip_list, b)) {
+				ut_ad(!b->in_flush_list);
+				ut_ad(b->in_LRU_list);
+
+				ut_d(b->in_page_hash = TRUE);
+				HASH_INSERT(buf_page_t, hash, buf_pool->page_hash,
+					    buf_page_address_fold(b->space, b->offset), b);
+			}
+
+			for (b = UT_LIST_GET_FIRST(buf_pool->flush_list); b;
+			     b = UT_LIST_GET_NEXT(flush_list, b)) {
+				ut_ad(b->in_flush_list);
+				ut_ad(b->in_LRU_list);
+
+				switch (buf_page_get_state(b)) {
+				case BUF_BLOCK_ZIP_DIRTY:
+					ut_d(b->in_page_hash = TRUE);
+					HASH_INSERT(buf_page_t, hash, buf_pool->page_hash,
+						    buf_page_address_fold(b->space,
+							    		  b->offset), b);
+					break;
+				case BUF_BLOCK_FILE_PAGE:
+					/* uncompressed page */
+					break;
+				case BUF_BLOCK_ZIP_FREE:
+				case BUF_BLOCK_ZIP_PAGE:
+				case BUF_BLOCK_NOT_USED:
+				case BUF_BLOCK_READY_FOR_USE:
+				case BUF_BLOCK_MEMORY:
+				case BUF_BLOCK_REMOVE_HASH:
+					ut_error;
+					break;
+				}
+			}
+
+
+		}
+	}
+
 	mutex_exit(&LRU_list_mutex);
 	rw_lock_x_unlock(&page_hash_latch);
 	buf_pool_mutex_exit();
@@ -1102,6 +1518,34 @@
 	buf_chunk_t*	chunk;
 	buf_chunk_t*	chunks;
 
+	if (srv_buffer_pool_shm_key) {
+		buf_shm_info_t*	shm_info;
+
+		ut_a(buf_pool->n_chunks == 1);
+
+		chunk = buf_pool->chunks;
+		shm_info = chunk->mem;
+		ut_a((byte*)chunk->blocks == (byte*)chunk->mem + sizeof(buf_shm_info_t));
+
+		/* validation the shared memory segment doesn't have unrecoverable contents. */
+		/* Currently, validation became not needed */
+		shm_info->reusable = TRUE;
+
+		memcpy(&(shm_info->buf_pool_backup), buf_pool, sizeof(buf_pool_t));
+		memcpy(&(shm_info->chunk_backup), chunk, sizeof(buf_chunk_t));
+
+		if (srv_fast_shutdown < 2) {
+			if (srv_buffer_pool_shm_checksum) {
+				shm_info->checksum = ut_fold_binary_32((byte*)chunk->mem + sizeof(buf_shm_info_t),
+								       chunk->mem_size - sizeof(buf_shm_info_t));
+			} else {
+				shm_info->checksum = BUF_NO_CHECKSUM_MAGIC;
+			}
+			shm_info->clean = TRUE;
+		}
+
+		os_shm_free(chunk->mem, chunk->mem_size);
+	} else {
 	chunks = buf_pool->chunks;
 	chunk = chunks + buf_pool->n_chunks;
 
@@ -1110,10 +1554,13 @@
 		would fail at shutdown. */
 		os_mem_free_large(chunk->mem, chunk->mem_size);
 	}
+	}
 
 	mem_free(buf_pool->chunks);
 	hash_table_free(buf_pool->page_hash);
+	if (!srv_buffer_pool_shm_key) {
 	hash_table_free(buf_pool->zip_hash);
+	}
 	mem_free(buf_pool);
 	buf_pool = NULL;
 }
@@ -1308,6 +1755,11 @@
 	//buf_pool_mutex_enter();
 	mutex_enter(&LRU_list_mutex);
 
+	if (srv_buffer_pool_shm_key) {
+		/* Cannot support shrink */
+		goto func_done;
+	}
+
 shrink_again:
 	if (buf_pool->n_chunks <= 1) {
 
@@ -1551,6 +2003,11 @@
 buf_pool_resize(void)
 /*=================*/
 {
+	if (srv_buffer_pool_shm_key) {
+		/* Cannot support resize */
+		return;
+	}
+
 	//buf_pool_mutex_enter();
 	mutex_enter(&LRU_list_mutex);
 
diff -ruN a/storage/innodb_plugin/ha/hash0hash.c b/storage/innodb_plugin/ha/hash0hash.c
--- a/storage/innodb_plugin/ha/hash0hash.c	2010-06-04 00:49:59.000000000 +0900
+++ b/storage/innodb_plugin/ha/hash0hash.c	2010-07-14 16:40:16.150438366 +0900
@@ -128,6 +128,70 @@
 }
 
 /*************************************************************//**
+*/
+UNIV_INTERN
+ulint
+hash_create_needed(
+/*===============*/
+	ulint	n)
+{
+	ulint	prime;
+	ulint	offset;
+
+	prime = ut_find_prime(n);
+
+	offset = (sizeof(hash_table_t) + 7) / 8;
+	offset *= 8;
+
+	return(offset + sizeof(hash_cell_t) * prime);
+}
+
+UNIV_INTERN
+void
+hash_create_init(
+/*=============*/
+	hash_table_t*	table,
+	ulint		n)
+{
+	ulint	prime;
+	ulint	offset;
+
+	prime = ut_find_prime(n);
+
+	offset = (sizeof(hash_table_t) + 7) / 8;
+	offset *= 8;
+
+	table->array = (hash_cell_t*)(((byte*)table) + offset);
+	table->n_cells = prime;
+# if defined UNIV_AHI_DEBUG || defined UNIV_DEBUG
+	table->adaptive = FALSE;
+# endif /* UNIV_AHI_DEBUG || UNIV_DEBUG */
+	table->n_mutexes = 0;
+	table->mutexes = NULL;
+	table->heaps = NULL;
+	table->heap = NULL;
+	ut_d(table->magic_n = HASH_TABLE_MAGIC_N);
+
+	/* Initialize the cell array */
+	hash_table_clear(table);
+}
+
+UNIV_INTERN
+void
+hash_create_reuse(
+/*==============*/
+	hash_table_t*	table)
+{
+	ulint	offset;
+
+	offset = (sizeof(hash_table_t) + 7) / 8;
+	offset *= 8;
+
+	table->array = (hash_cell_t*)(((byte*)table) + offset);
+	ut_ad(table->magic_n == HASH_TABLE_MAGIC_N);
+}
+
+/*************************************************************//**
 Frees a hash table. */
 UNIV_INTERN
 void
diff -ruN a/storage/innodb_plugin/handler/ha_innodb.cc b/storage/innodb_plugin/handler/ha_innodb.cc
--- a/storage/innodb_plugin/handler/ha_innodb.cc	2010-07-14 16:34:18.597725479 +0900
+++ b/storage/innodb_plugin/handler/ha_innodb.cc	2010-07-14 16:40:16.159323612 +0900
@@ -198,6 +198,7 @@
 static my_bool	innobase_create_status_file		= FALSE;
 static my_bool	innobase_stats_on_metadata		= TRUE;
 static my_bool	innobase_use_sys_stats_table		= FALSE;
+static my_bool	innobase_buffer_pool_shm_checksum	= TRUE;
 
 static char*	internal_innobase_data_file_path	= NULL;
 
@@ -2457,6 +2458,7 @@
 	srv_use_doublewrite_buf = (ibool) innobase_use_doublewrite;
 	srv_use_checksums = (ibool) innobase_use_checksums;
 	srv_fast_checksum = (ibool) innobase_fast_checksum;
+	srv_buffer_pool_shm_checksum = (ibool) innobase_buffer_pool_shm_checksum;
 
 #ifdef HAVE_LARGE_PAGES
         if ((os_use_large_pages = (ibool) my_use_large_pages))
@@ -11405,6 +11407,16 @@
   "The size of the memory buffer InnoDB uses to cache data and indexes of its tables.",
   NULL, NULL, 128*1024*1024L, 32*1024*1024L, LONGLONG_MAX, 1024*1024L);
 
+static MYSQL_SYSVAR_UINT(buffer_pool_shm_key, srv_buffer_pool_shm_key,
+  PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
+  "[experimental] The key value of shared memory segment for the buffer pool. 0 (default) disables the feature.",
+  NULL, NULL, 0, 0, INT_MAX32, 0);
+
+static MYSQL_SYSVAR_BOOL(buffer_pool_shm_checksum, innobase_buffer_pool_shm_checksum,
+  PLUGIN_VAR_NOCMDARG | PLUGIN_VAR_READONLY,
+  "Enable buffer_pool_shm checksum validation (enabled by default).",
+  NULL, NULL, TRUE);
+
 static MYSQL_SYSVAR_ULONG(commit_concurrency, innobase_commit_concurrency,
   PLUGIN_VAR_RQCMDARG,
   "Helps in performance tuning in heavily concurrent environments.",
@@ -11680,6 +11692,8 @@
   MYSQL_SYSVAR(additional_mem_pool_size),
   MYSQL_SYSVAR(autoextend_increment),
   MYSQL_SYSVAR(buffer_pool_size),
+  MYSQL_SYSVAR(buffer_pool_shm_key),
+  MYSQL_SYSVAR(buffer_pool_shm_checksum),
   MYSQL_SYSVAR(checksums),
   MYSQL_SYSVAR(fast_checksum),
   MYSQL_SYSVAR(commit_concurrency),
diff -ruN a/storage/innodb_plugin/handler/innodb_patch_info.h b/storage/innodb_plugin/handler/innodb_patch_info.h
--- a/storage/innodb_plugin/handler/innodb_patch_info.h	2010-07-14 16:34:18.603733950 +0900
+++ b/storage/innodb_plugin/handler/innodb_patch_info.h	2010-07-14 16:40:16.164323927 +0900
@@ -47,5 +47,6 @@
 {"innodb_fast_checksum","Using the checksum on 32bit-unit calculation","incompatible for unpatched ver.","http://www.percona.com/docs/wiki/percona-xtradb"},
 {"innodb_files_extend","allow >4GB transaction log files, and can vary universal page size of datafiles","incompatible for unpatched ver.","http://www.percona.com/docs/wiki/percona-xtradb"},
 {"innodb_sys_tables_sys_indexes","Expose InnoDB SYS_TABLES and SYS_INDEXES schema tables","","http://www.percona.com/docs/wiki/percona-xtradb"},
+{"innodb_buffer_pool_shm","Put buffer pool contents to shared memory segment and reuse it at clean restart [experimental]","","http://www.percona.com/docs/wiki/percona-xtradb"},
 {NULL, NULL, NULL, NULL}
 };
diff -ruN a/storage/innodb_plugin/include/buf0buf.h b/storage/innodb_plugin/include/buf0buf.h
--- a/storage/innodb_plugin/include/buf0buf.h	2010-07-14 16:33:23.823323393 +0900
+++ b/storage/innodb_plugin/include/buf0buf.h	2010-07-14 16:40:16.166323436 +0900
@@ -36,6 +36,7 @@
 #include "ut0rbt.h"
 #ifndef UNIV_HOTBACKUP
 #include "os0proc.h"
+#include "srv0srv.h"
 
 /** @name Modes for buf_page_get_gen */
 /* @{ */
@@ -1301,7 +1302,10 @@
 /**********************************************************************//**
 Compute the hash fold value for blocks in buf_pool->zip_hash. */
 /* @{ */
-#define BUF_POOL_ZIP_FOLD_PTR(ptr) ((ulint) (ptr) / UNIV_PAGE_SIZE)
+/* the fold should be relative when srv_buffer_pool_shm_key is enabled */
+#define BUF_POOL_ZIP_FOLD_PTR(ptr) (!srv_buffer_pool_shm_key\
+					?((ulint) (ptr) / UNIV_PAGE_SIZE)\
+					:((ulint) ((byte*)ptr - (byte*)(buf_pool->chunks->blocks->frame)) / UNIV_PAGE_SIZE))
 #define BUF_POOL_ZIP_FOLD(b) BUF_POOL_ZIP_FOLD_PTR((b)->frame)
 #define BUF_POOL_ZIP_FOLD_BPAGE(b) BUF_POOL_ZIP_FOLD((buf_block_t*) (b))
 /* @} */
diff -ruN a/storage/innodb_plugin/include/hash0hash.h b/storage/innodb_plugin/include/hash0hash.h
--- a/storage/innodb_plugin/include/hash0hash.h	2010-06-04 00:49:59.000000000 +0900
+++ b/storage/innodb_plugin/include/hash0hash.h	2010-07-14 16:40:16.168323262 +0900
@@ -49,6 +49,28 @@
 hash_create(
 /*========*/
 	ulint	n);	/*!< in: number of array cells */
+
+/*************************************************************//**
+*/
+UNIV_INTERN
+ulint
+hash_create_needed(
+/*===============*/
+	ulint	n);
+
+UNIV_INTERN
+void
+hash_create_init(
+/*=============*/
+	hash_table_t*	table,
+	ulint		n);
+
+UNIV_INTERN
+void
+hash_create_reuse(
+/*==============*/
+	hash_table_t*	table);
+
 #ifndef UNIV_HOTBACKUP
 /*************************************************************//**
 Creates a mutex array to protect a hash table. */
@@ -327,6 +349,33 @@
 		}\
 	}\
 } while (0)
+
+/********************************************************************//**
+Align nodes with moving location.*/
+#define HASH_OFFSET(TABLE, NODE_TYPE, PTR_NAME, FADDR, FOFFSET, BOFFSET) \
+do {\
+	ulint		i2222;\
+	ulint		cell_count2222;\
+\
+	cell_count2222 = hash_get_n_cells(TABLE);\
+\
+	for (i2222 = 0; i2222 < cell_count2222; i2222++) {\
+		NODE_TYPE*	node2222;\
+\
+		if ((TABLE)->array[i2222].node) \
+			(TABLE)->array[i2222].node = (void*)((byte*)(TABLE)->array[i2222].node \
+			+ (((TABLE)->array[i2222].node > (void*)FADDR)?FOFFSET:BOFFSET));\
+		node2222 = HASH_GET_FIRST((TABLE), i2222);\
+\
+		while (node2222) {\
+			if (node2222->PTR_NAME) \
+				node2222->PTR_NAME = (void*)((byte*)(node2222->PTR_NAME) \
+				+ ((((void*)node2222->PTR_NAME) > (void*)FADDR)?FOFFSET:BOFFSET));\
+\
+			node2222 = node2222->PTR_NAME;\
+		}\
+	}\
+} while (0)
 
 /************************************************************//**
 Gets the mutex index for a fold value in a hash table.
diff -ruN a/storage/innodb_plugin/include/os0proc.h b/storage/innodb_plugin/include/os0proc.h
--- a/storage/innodb_plugin/include/os0proc.h	2010-06-04 00:49:59.000000000 +0900
+++ b/storage/innodb_plugin/include/os0proc.h	2010-07-14 16:40:16.169321536 +0900
@@ -32,6 +32,11 @@
 #ifdef UNIV_LINUX
 #include <sys/ipc.h>
 #include <sys/shm.h>
+#else
+# if defined HAVE_SYS_IPC_H && HAVE_SYS_SHM_H
+#include <sys/ipc.h>
+#include <sys/shm.h>
+# endif
 #endif
 
 typedef void*			os_process_t;
@@ -70,6 +75,29 @@
 	ulint	size);			/*!< in: size returned by
 					os_mem_alloc_large() */
 
+
+/****************************************************************//**
+Allocates or attaches and reuses shared memory segment.
+The content is not cleared automatically.
+@return	allocated memory */
+UNIV_INTERN
+void*
+os_shm_alloc(
+/*=========*/
+	ulint*	n,			/*!< in/out: number of bytes */
+	uint	key,
+	ibool*	is_new);
+
+/****************************************************************//**
+Detach shared memory segment. */
+UNIV_INTERN
+void
+os_shm_free(
+/*========*/
+	void	*ptr,			/*!< in: pointer returned by
+					os_shm_alloc() */
+	ulint	size);			/*!< in: size returned by
+					os_shm_alloc() */
 #ifndef UNIV_NONINL
 #include "os0proc.ic"
 #endif
diff -ruN a/storage/innodb_plugin/include/srv0srv.h b/storage/innodb_plugin/include/srv0srv.h
--- a/storage/innodb_plugin/include/srv0srv.h	2010-07-14 16:32:49.695323045 +0900
+++ b/storage/innodb_plugin/include/srv0srv.h	2010-07-14 16:40:16.171325784 +0900
@@ -156,6 +156,10 @@
 extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 
+extern uint	srv_buffer_pool_shm_key;
+extern ibool	srv_buffer_pool_shm_is_reused;
+extern ibool	srv_buffer_pool_shm_checksum;
+
 extern ibool	srv_thread_concurrency_timer_based;
 
 extern ulint	srv_n_file_io_threads;
diff -ruN a/storage/innodb_plugin/include/ut0lst.h b/storage/innodb_plugin/include/ut0lst.h
--- a/storage/innodb_plugin/include/ut0lst.h	2010-06-04 00:49:59.000000000 +0900
+++ b/storage/innodb_plugin/include/ut0lst.h	2010-07-14 16:40:16.172321547 +0900
@@ -257,5 +257,48 @@
 	ut_a(ut_list_node_313 == NULL);					\
 } while (0)
 
+/********************************************************************//**
+Align nodes with moving location.
+@param NAME		the name of the list
+@param TYPE		node type
+@param BASE		base node (not a pointer to it)
+@param OFFSET		offset moved */
+#define UT_LIST_OFFSET(NAME, TYPE, BASE, FADDR, FOFFSET, BOFFSET)	\
+do {									\
+	ulint	ut_list_i_313;						\
+	TYPE*	ut_list_node_313;					\
+									\
+	if ((BASE).start)						\
+		(BASE).start = (void*)((byte*)((BASE).start)			\
+			+ (((void*)((BASE).start) > (void*)FADDR)?FOFFSET:BOFFSET));\
+	if ((BASE).end)							\
+		(BASE).end   = (void*)((byte*)((BASE).end)			\
+			+ (((void*)((BASE).end) > (void*)FADDR)?FOFFSET:BOFFSET));\
+									\
+	ut_list_node_313 = (BASE).start;				\
+									\
+	for (ut_list_i_313 = (BASE).count; ut_list_i_313--; ) {		\
+		ut_a(ut_list_node_313);					\
+		if ((ut_list_node_313->NAME).prev)			\
+			(ut_list_node_313->NAME).prev = (void*)((byte*)((ut_list_node_313->NAME).prev)\
+				+ (((void*)((ut_list_node_313->NAME).prev) > (void*)FADDR)?FOFFSET:BOFFSET));\
+		if ((ut_list_node_313->NAME).next)			\
+			(ut_list_node_313->NAME).next =	(void*)((byte*)((ut_list_node_313->NAME).next)\
+				+ (((void*)((ut_list_node_313->NAME).next)> (void*)FADDR)?FOFFSET:BOFFSET));\
+		ut_list_node_313 = (ut_list_node_313->NAME).next;	\
+	}								\
+									\
+	ut_a(ut_list_node_313 == NULL);					\
+									\
+	ut_list_node_313 = (BASE).end;					\
+									\
+	for (ut_list_i_313 = (BASE).count; ut_list_i_313--; ) {		\
+		ut_a(ut_list_node_313);					\
+		ut_list_node_313 = (ut_list_node_313->NAME).prev;	\
+	}								\
+									\
+	ut_a(ut_list_node_313 == NULL);					\
+} while (0)
+
 #endif
 
diff -ruN a/storage/innodb_plugin/log/log0recv.c b/storage/innodb_plugin/log/log0recv.c
--- a/storage/innodb_plugin/log/log0recv.c	2010-10-01 15:25:27.106299166 +0900
+++ b/storage/innodb_plugin/log/log0recv.c	2010-10-01 15:26:33.689261436 +0900
@@ -2899,6 +2899,7 @@
 /*==========================*/
 {
 	ut_a(!recv_needed_recovery);
+	ut_a(!srv_buffer_pool_shm_is_reused);
 
 	recv_needed_recovery = TRUE;
 
diff -ruN a/storage/innodb_plugin/os/os0proc.c b/storage/innodb_plugin/os/os0proc.c
--- a/storage/innodb_plugin/os/os0proc.c	2010-06-04 00:49:59.000000000 +0900
+++ b/storage/innodb_plugin/os/os0proc.c	2010-07-14 16:40:16.174322953 +0900
@@ -229,3 +229,173 @@
 	}
 #endif
 }
+
+/****************************************************************//**
+Allocates or attaches and reuses shared memory segment.
+The content is not cleared automatically.
+@return	allocated memory */
+UNIV_INTERN
+void*
+os_shm_alloc(
+/*=========*/
+	ulint*	n,			/*!< in/out: number of bytes */
+	uint	key,
+	ibool*	is_new)
+{
+	void*	ptr;
+#if defined HAVE_SYS_IPC_H && HAVE_SYS_SHM_H
+	ulint	size;
+	int	shmid;
+
+	*is_new = FALSE;
+	fprintf(stderr,
+		"InnoDB: The shared memory segment containing the buffer pool is: key  %#x (%d).\n",
+		key, key);
+# if defined HAVE_LARGE_PAGES && defined UNIV_LINUX
+	if (!os_use_large_pages || !os_large_page_size) {
+		goto skip;
+	}
+
+	/* Align block size to os_large_page_size */
+	ut_ad(ut_is_2pow(os_large_page_size));
+	size = ut_2pow_round(*n + (os_large_page_size - 1),
+			     os_large_page_size);
+
+	shmid = shmget((key_t)key, (size_t)size,
+			IPC_CREAT | IPC_EXCL | SHM_HUGETLB | SHM_R | SHM_W);
+	if (shmid < 0) {
+		if (errno == EEXIST) {
+			fprintf(stderr,
+				"InnoDB: HugeTLB: The shared memory segment exists.\n");
+			shmid = shmget((key_t)key, (size_t)size,
+					SHM_HUGETLB | SHM_R | SHM_W);
+			if (shmid < 0) {
+				fprintf(stderr,
+					"InnoDB: HugeTLB: Warning: Failed to allocate %lu bytes. (reuse) errno %d\n",
+					size, errno);
+				goto skip;
+			} else {
+				fprintf(stderr,
+					"InnoDB: HugeTLB: The existent shared memory segment is used.\n");
+			}
+		} else {
+			fprintf(stderr,
+				"InnoDB: HugeTLB: Warning: Failed to allocate %lu bytes. (new) errno %d\n",
+				size, errno);
+			goto skip;
+		}
+	} else {
+		*is_new = TRUE;
+		fprintf(stderr,
+			"InnoDB: HugeTLB: A new shared memory segment has been created .\n");
+	}
+
+	ptr = shmat(shmid, NULL, 0);
+	if (ptr == (void *)-1) {
+		fprintf(stderr,
+			"InnoDB: HugeTLB: Warning: Failed to attach shared memory segment, errno %d\n",
+			errno);
+		ptr = NULL;
+	}
+
+	if (ptr) {
+		*n = size;
+		os_fast_mutex_lock(&ut_list_mutex);
+		ut_total_allocated_memory += size;
+		os_fast_mutex_unlock(&ut_list_mutex);
+		UNIV_MEM_ALLOC(ptr, size);
+		return(ptr);
+	}
+skip:
+	*is_new = FALSE;
+# endif /* HAVE_LARGE_PAGES && defined UNIV_LINUX */
+# ifdef HAVE_GETPAGESIZE
+	size = getpagesize();
+# else
+	size = UNIV_PAGE_SIZE;
+# endif
+	/* Align block size to system page size */
+	ut_ad(ut_is_2pow(size));
+	size = *n = ut_2pow_round(*n + (size - 1), size);
+
+	shmid = shmget((key_t)key, (size_t)size,
+			IPC_CREAT | IPC_EXCL | SHM_R | SHM_W);
+	if (shmid < 0) {
+		if (errno == EEXIST) {
+			fprintf(stderr,
+				"InnoDB: A shared memory segment containing the buffer pool seems to already exist.\n");
+			shmid = shmget((key_t)key, (size_t)size,
+					SHM_R | SHM_W);
+			if (shmid < 0) {
+				fprintf(stderr,
+					"InnoDB: Warning: Failed to allocate %lu bytes. (reuse) errno %d\n",
+					size, errno);
+				ptr = NULL;
+				goto end;
+			} else {
+				fprintf(stderr,
+					"InnoDB: The existent shared memory segment is used.\n");
+			}
+		} else {
+			fprintf(stderr,
+				"InnoDB: Warning: Failed to allocate %lu bytes. (new) errno %d\n",
+				size, errno);
+			ptr = NULL;
+			goto end;
+		}
+	} else {
+		*is_new = TRUE;
+		fprintf(stderr,
+			"InnoDB: A new shared memory segment has been created.\n");
+	}
+
+	ptr = shmat(shmid, NULL, 0);
+	if (ptr == (void *)-1) {
+		fprintf(stderr,
+			"InnoDB: Warning: Failed to attach shared memory segment, errno %d\n",
+			errno);
+		ptr = NULL;
+	}
+
+	if (ptr) {
+		*n = size;
+		os_fast_mutex_lock(&ut_list_mutex);
+		ut_total_allocated_memory += size;
+		os_fast_mutex_unlock(&ut_list_mutex);
+		UNIV_MEM_ALLOC(ptr, size);
+	}
+end:
+#else /* HAVE_SYS_IPC_H && HAVE_SYS_SHM_H */
+	fprintf(stderr, "InnoDB: shared memory segment is not supported.\n");
+	ptr = NULL;
+#endif /* HAVE_SYS_IPC_H && HAVE_SYS_SHM_H */
+	return(ptr);
+}
+
+/****************************************************************//**
+Detach shared memory segment. */
+UNIV_INTERN
+void
+os_shm_free(
+/*========*/
+	void	*ptr,			/*!< in: pointer returned by
+					os_shm_alloc() */
+	ulint	size)			/*!< in: size returned by
+					os_shm_alloc() */
+{
+	os_fast_mutex_lock(&ut_list_mutex);
+	ut_a(ut_total_allocated_memory >= size);
+	os_fast_mutex_unlock(&ut_list_mutex);
+
+#if defined HAVE_SYS_IPC_H && HAVE_SYS_SHM_H
+	if (!shmdt(ptr)) {
+		os_fast_mutex_lock(&ut_list_mutex);
+		ut_a(ut_total_allocated_memory >= size);
+		ut_total_allocated_memory -= size;
+		os_fast_mutex_unlock(&ut_list_mutex);
+		UNIV_MEM_FREE(ptr, size);
+	}
+#else /* HAVE_SYS_IPC_H && HAVE_SYS_SHM_H */
+	fprintf(stderr, "InnoDB: shared memory segment is not supported.\n");
+#endif /* HAVE_SYS_IPC_H && HAVE_SYS_SHM_H */
+}
diff -ruN a/storage/innodb_plugin/srv/srv0srv.c b/storage/innodb_plugin/srv/srv0srv.c
--- a/storage/innodb_plugin/srv/srv0srv.c	2010-07-14 16:33:23.848391648 +0900
+++ b/storage/innodb_plugin/srv/srv0srv.c	2010-07-14 16:40:16.177323553 +0900
@@ -211,6 +211,11 @@
 UNIV_INTERN ulint	srv_mem_pool_size	= ULINT_MAX;
 UNIV_INTERN ulint	srv_lock_table_size	= ULINT_MAX;
 
+/* key value for shm */
+UNIV_INTERN uint	srv_buffer_pool_shm_key	= 0;
+UNIV_INTERN ibool	srv_buffer_pool_shm_is_reused = FALSE;
+UNIV_INTERN ibool	srv_buffer_pool_shm_checksum = TRUE;
+
 /* This parameter is deprecated. Use srv_n_io_[read|write]_threads
 instead. */
 UNIV_INTERN ulint	srv_n_file_io_threads	= ULINT_MAX;
diff -ruN a/storage/innodb_plugin/srv/srv0start.c b/storage/innodb_plugin/srv/srv0start.c
--- a/storage/innodb_plugin/srv/srv0start.c	2010-07-14 16:33:23.851391514 +0900
+++ b/storage/innodb_plugin/srv/srv0start.c	2010-07-14 16:40:16.180321173 +0900
@@ -1744,6 +1744,8 @@
 		Note that this is not as heavy weight as it seems. At
 		this point there will be only ONE page in the buf_LRU
 		and there must be no page in the buf_flush list. */
+		/* buffer_pool_shm should not be reused when recovery was needed. */
+		if (!srv_buffer_pool_shm_is_reused)
 		buf_pool_invalidate();
 
 		/* We always try to do a recovery, even if the database had