buf0buf.c, buf0buf.ic, buf0buf.h:

Reduce memory usage of the buffer headers
Many files:
  Merge InnoDB-4.1 with AWE support


sql/mysqld.cc:
  Merge InnoDB-4.1 with AWE support
sql/set_var.cc:
  Merge InnoDB-4.1 with AWE support
sql/ha_innodb.h:
  Merge InnoDB-4.1 with AWE support
sql/ha_innodb.cc:
  Merge InnoDB-4.1 with AWE support
innobase/btr/btr0cur.c:
  Merge InnoDB-4.1 with AWE support
innobase/btr/btr0pcur.c:
  Merge InnoDB-4.1 with AWE support
innobase/buf/buf0flu.c:
  Merge InnoDB-4.1 with AWE support
innobase/buf/buf0lru.c:
  Merge InnoDB-4.1 with AWE support
innobase/buf/buf0rea.c:
  Merge InnoDB-4.1 with AWE support
innobase/include/btr0pcur.h:
  Merge InnoDB-4.1 with AWE support
innobase/include/buf0lru.h:
  Merge InnoDB-4.1 with AWE support
innobase/include/log0recv.h:
  Merge InnoDB-4.1 with AWE support
innobase/include/os0proc.h:
  Merge InnoDB-4.1 with AWE support
innobase/include/srv0srv.h:
  Merge InnoDB-4.1 with AWE support
innobase/log/log0log.c:
  Merge InnoDB-4.1 with AWE support
innobase/log/log0recv.c:
  Merge InnoDB-4.1 with AWE support
innobase/os/os0file.c:
  Merge InnoDB-4.1 with AWE support
innobase/os/os0proc.c:
  Merge InnoDB-4.1 with AWE support
innobase/srv/srv0srv.c:
  Merge InnoDB-4.1 with AWE support
innobase/srv/srv0start.c:
  Merge InnoDB-4.1 with AWE support
innobase/trx/trx0sys.c:
  Merge InnoDB-4.1 with AWE support
innobase/trx/trx0trx.c:
  Merge InnoDB-4.1 with AWE support
innobase/ut/ut0ut.c:
  Merge InnoDB-4.1 with AWE support
innobase/include/buf0buf.h:
  Reduce memory usage of the buffer headers
innobase/include/buf0buf.ic:
  Reduce memory usage of the buffer headers
innobase/buf/buf0buf.c:
  Reduce memory usage of the buffer headers

innobase/btr/btr0cur.c

@@ -291,6 +291,7 @@ btr_cur_search_to_nth_level(
 		&& latch_mode <= BTR_MODIFY_LEAF && info->last_hash_succ
 		&& !estimate
 		&& mode != PAGE_CUR_LE_OR_EXTENDS
+		&& srv_use_adaptive_hash_indexes
 	        && btr_search_guess_on_hash(index, info, tuple, mode,
 						latch_mode, cursor,
 						has_search_latch, mtr)) {
@@ -495,9 +496,11 @@ retry_page_get:
 		cursor->up_bytes = up_bytes;
 
 #ifdef BTR_CUR_ADAPT		
-		btr_search_info_update(index, cursor);
+		if (srv_use_adaptive_hash_indexes) {
-#endif
 
+			btr_search_info_update(index, cursor);
+		}
+#endif
 		ut_ad(cursor->up_match != ULINT_UNDEFINED
 						|| mode != PAGE_CUR_GE);
 		ut_ad(cursor->up_match != ULINT_UNDEFINED

innobase/btr/btr0pcur.c

@@ -95,7 +95,9 @@ btr_pcur_store_position(
 	ut_a(cursor->latch_mode != BTR_NO_LATCHES);
 
 	if (page_get_n_recs(page) == 0) {
-		/* It must be an empty index tree */
+		/* It must be an empty index tree; NOTE that in this case
+		we do not store the modify_clock, but always do a search
+		if we restore the cursor position */
 
 		ut_a(btr_page_get_next(page, mtr) == FIL_NULL
 		     && btr_page_get_prev(page, mtr) == FIL_NULL);
@@ -128,12 +130,13 @@ btr_pcur_store_position(
 	} else {
 		cursor->rel_pos = BTR_PCUR_ON;
 	}
-	
+
 	cursor->old_stored = BTR_PCUR_OLD_STORED;
 	cursor->old_rec = dict_tree_copy_rec_order_prefix(tree, rec,
 						&(cursor->old_rec_buf),
 						&(cursor->buf_size));
 									
+	cursor->block_when_stored = buf_block_align(page);	
 	cursor->modify_clock = buf_frame_get_modify_clock(page);
 }
 
@@ -205,6 +208,9 @@ btr_pcur_restore_position(
 	if (cursor->rel_pos == BTR_PCUR_AFTER_LAST_IN_TREE
 	    || cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE) {
 
+	    	/* In these cases we do not try an optimistic restoration,
+	    	but always do a search */
+
 	    	if (cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE) {
 	    		from_left = TRUE;
 	    	} else {
@@ -214,6 +220,10 @@ btr_pcur_restore_position(
 		btr_cur_open_at_index_side(from_left,
 			btr_pcur_get_btr_cur(cursor)->index, latch_mode,
 					btr_pcur_get_btr_cur(cursor), mtr);
+
+		cursor->block_when_stored =
+				buf_block_align(btr_pcur_get_page(cursor));
+
 		return(FALSE);
 	}
 	
@@ -224,8 +234,9 @@ btr_pcur_restore_position(
 	if (latch_mode == BTR_SEARCH_LEAF || latch_mode == BTR_MODIFY_LEAF) {
 		/* Try optimistic restoration */
 	    
-		if (buf_page_optimistic_get(latch_mode, page,
+		if (buf_page_optimistic_get(latch_mode,
-						cursor->modify_clock, mtr)) {
+					    cursor->block_when_stored, page,
+					    cursor->modify_clock, mtr)) {
 			cursor->pos_state = BTR_PCUR_IS_POSITIONED;
 
 			buf_page_dbg_add_level(page, SYNC_TREE_NODE);
@@ -270,8 +281,6 @@ btr_pcur_restore_position(
 
 	btr_pcur_open_with_no_init(btr_pcur_get_btr_cur(cursor)->index, tuple,
 					mode, latch_mode, cursor, 0, mtr);
-
-	cursor->old_stored = BTR_PCUR_OLD_STORED;
 	
 	/* Restore the old search mode */
 	cursor->search_mode = old_mode;
@@ -280,11 +289,18 @@ btr_pcur_restore_position(
 	    && btr_pcur_is_on_user_rec(cursor, mtr)
 	    && 0 == cmp_dtuple_rec(tuple, btr_pcur_get_rec(cursor))) {
 
-	        /* We have to store the NEW value for the modify clock, since
+		/* We have to store the NEW value for the modify clock, since
-	        the cursor can now be on a different page! */
+		the cursor can now be on a different page! But we can retain
+		the value of old_rec */
+
+		cursor->modify_clock =
+			buf_frame_get_modify_clock(btr_pcur_get_page(cursor));
+
+		cursor->block_when_stored =
+			buf_block_align(btr_pcur_get_page(cursor));
+
+		cursor->old_stored = BTR_PCUR_OLD_STORED;
 
-	        cursor->modify_clock = buf_frame_get_modify_clock(
-				    buf_frame_align(btr_pcur_get_rec(cursor)));
 		mem_heap_free(heap);
 
 		return(TRUE);
@@ -292,6 +308,12 @@ btr_pcur_restore_position(
 
 	mem_heap_free(heap);
 
+	/* We have to store new position information, modify_clock etc.,
+	to the cursor because it can now be on a different page, the record
+	under it may have been removed, etc. */
+	
+	btr_pcur_store_position(cursor, mtr);
+
 	return(FALSE);
 }
 

innobase/buf/buf0buf.c

@@ -196,7 +196,29 @@ If a new page is referenced in the buf_pool, and several pages
 of its random access area (for instance, 32 consecutive pages
 in a tablespace) have recently been referenced, we may predict
 that the whole area may be needed in the near future, and issue
-the read requests for the whole area. */
+the read requests for the whole area.
+
+		AWE implementation
+		------------------
+
+By a 'block' we mean the buffer header of type buf_block_t. By a 'page'
+we mean the physical 16 kB memory area allocated from RAM for that block.
+By a 'frame' we mean a 16 kB area in the virtual address space of the
+process, in the frame_mem of buf_pool.
+
+We can map pages to the frames of the buffer pool.
+
+1) A buffer block allocated to use as a non-data page, e.g., to the lock
+table, is always mapped to a frame.
+2) A bufferfixed or io-fixed data page is always mapped to a frame.
+3) When we need to map a block to frame, we look from the list
+awe_LRU_free_mapped and try to unmap its last block, but note that
+bufferfixed or io-fixed pages cannot be unmapped.
+4) For every frame in the buffer pool there is always a block whose page is
+mapped to it. When we create the buffer pool, we map the first elements
+in the free list to the frames.
+5) When we have AWE enabled, we disable adaptive hash indexes.
+*/
 
 buf_pool_t*	buf_pool = NULL; /* The buffer buf_pool of the database */
 
@@ -346,12 +368,15 @@ void
 buf_block_init(
 /*===========*/
 	buf_block_t*	block,	/* in: pointer to control block */
-	byte*		frame)	/* in: pointer to buffer frame */
+	byte*		frame)	/* in: pointer to buffer frame, or NULL if in
+				the case of AWE there is no frame */
 {
 	block->state = BUF_BLOCK_NOT_USED;
 	
 	block->frame = frame;
 
+	block->awe_info = NULL;
+
 	block->modify_clock = ut_dulint_zero;
 	
 	block->file_page_was_freed = FALSE;
@@ -364,29 +389,37 @@ buf_block_init(
 	rw_lock_create(&(block->read_lock));
 	rw_lock_set_level(&(block->read_lock), SYNC_NO_ORDER_CHECK);
 	
+#ifdef UNIV_SYNC_DEBUG
 	rw_lock_create(&(block->debug_latch));
 	rw_lock_set_level(&(block->debug_latch), SYNC_NO_ORDER_CHECK);
+#endif
 }
 
 /************************************************************************
-Creates a buffer buf_pool object. */
+Creates the buffer pool. */
-static
+
 buf_pool_t*
-buf_pool_create(
+buf_pool_init(
-/*============*/
+/*==========*/
 				/* out, own: buf_pool object, NULL if not
-				enough memory */
+				enough memory or error */
 	ulint	max_size,	/* in: maximum size of the buf_pool in
 				blocks */
-	ulint	curr_size)	/* in: current size to use, must be <=
+	ulint	curr_size,	/* in: current size to use, must be <=
 				max_size, currently must be equal to
 				max_size */
+	ulint	n_frames)	/* in: number of frames; if AWE is used,
+				this is the size of the address space window
+				where physical memory pages are mapped; if
+				AWE is not used then this must be the same
+				as max_size */
 {
 	byte*		frame;
 	ulint		i;
 	buf_block_t*	block;
 	
 	ut_a(max_size == curr_size);
+	ut_a(srv_use_awe || n_frames == max_size);
 	
 	buf_pool = mem_alloc(sizeof(buf_pool_t));
 
@@ -396,8 +429,38 @@ buf_pool_create(
 	mutex_set_level(&(buf_pool->mutex), SYNC_BUF_POOL);
 
 	mutex_enter(&(buf_pool->mutex));
-	
+
-	buf_pool->frame_mem = ut_malloc(UNIV_PAGE_SIZE * (max_size + 1));
+	if (srv_use_awe) {
+		/*----------------------------------------*/
+		/* Allocate the virtual address space window, i.e., the
+		buffer pool frames */
+
+		buf_pool->frame_mem = os_awe_allocate_virtual_mem_window(
+					UNIV_PAGE_SIZE * (n_frames + 1));
+					
+		/* Allocate the physical memory for AWE and the AWE info array
+		for buf_pool */
+
+		if ((curr_size % ((1024 * 1024) / UNIV_PAGE_SIZE)) != 0) {
+
+		        fprintf(stderr,
+"InnoDB: AWE: Error: physical memory must be allocated in full megabytes.\n"
+"InnoDB: Trying to allocate %lu database pages.\n", 
+			  curr_size);
+
+		        return(NULL);
+		}
+
+		if (!os_awe_allocate_physical_mem(&(buf_pool->awe_info),
+			curr_size / ((1024 * 1024) / UNIV_PAGE_SIZE))) {
+
+			return(NULL);
+		}
+		/*----------------------------------------*/
+	} else {
+		buf_pool->frame_mem = ut_malloc(
+					UNIV_PAGE_SIZE * (n_frames + 1));
+	}
 
 	if (buf_pool->frame_mem == NULL) {
 
@@ -414,21 +477,60 @@ buf_pool_create(
 	buf_pool->max_size = max_size;
 	buf_pool->curr_size = curr_size;
 
+	buf_pool->n_frames = n_frames;
+
 	/* Align pointer to the first frame */
 
 	frame = ut_align(buf_pool->frame_mem, UNIV_PAGE_SIZE);
-	buf_pool->frame_zero = frame;
 
+	buf_pool->frame_zero = frame;
 	buf_pool->high_end = frame + UNIV_PAGE_SIZE * curr_size;
 
-	/* Init block structs and assign frames for them */
+	if (srv_use_awe) {
+		/*----------------------------------------*/
+		/* Map an initial part of the allocated physical memory to
+		the window */
+
+		os_awe_map_physical_mem_to_window(buf_pool->frame_zero,
+					n_frames *
+					(UNIV_PAGE_SIZE / OS_AWE_X86_PAGE_SIZE),
+					buf_pool->awe_info);
+		/*----------------------------------------*/
+	}
+
+	buf_pool->blocks_of_frames = ut_malloc(sizeof(void*) * n_frames);
+	
+	if (buf_pool->blocks_of_frames == NULL) {
+
+		return(NULL);
+	}
+
+	/* Init block structs and assign frames for them; in the case of
+	AWE there are less frames than blocks. Then we assign the frames
+	to the first blocks (we already mapped the memory above). We also
+	init the awe_info for every block. */
+
 	for (i = 0; i < max_size; i++) {
 
 		block = buf_pool_get_nth_block(buf_pool, i);
+
+		if (i < n_frames) {
+			frame = buf_pool->frame_zero + i * UNIV_PAGE_SIZE;
+			*(buf_pool->blocks_of_frames + i) = block;
+		} else {
+			frame = NULL;
+		}
+		
 		buf_block_init(block, frame);
-		frame = frame + UNIV_PAGE_SIZE;
+
+		if (srv_use_awe) {
+			/*----------------------------------------*/
+			block->awe_info = buf_pool->awe_info
+				+ i * (UNIV_PAGE_SIZE / OS_AWE_X86_PAGE_SIZE);
+			/*----------------------------------------*/
+		}
 	}
-	
+
 	buf_pool->page_hash = hash_create(2 * max_size);
 
 	buf_pool->n_pend_reads = 0;
@@ -438,12 +540,14 @@ buf_pool_create(
 	buf_pool->n_pages_read = 0;
 	buf_pool->n_pages_written = 0;
 	buf_pool->n_pages_created = 0;
-
+	buf_pool->n_pages_awe_remapped = 0;
+	
 	buf_pool->n_page_gets = 0;
 	buf_pool->n_page_gets_old = 0;
 	buf_pool->n_pages_read_old = 0;
 	buf_pool->n_pages_written_old = 0;
 	buf_pool->n_pages_created_old = 0;
+	buf_pool->n_pages_awe_remapped_old = 0;
 	
 	/* 2. Initialize flushing fields
 	   ---------------------------- */
@@ -466,40 +570,120 @@ buf_pool_create(
 
 	buf_pool->LRU_old = NULL;
 
+	UT_LIST_INIT(buf_pool->awe_LRU_free_mapped);
+
 	/* Add control blocks to the free list */
 	UT_LIST_INIT(buf_pool->free);
+
 	for (i = 0; i < curr_size; i++) {
 
 		block = buf_pool_get_nth_block(buf_pool, i);
 
-		/* Wipe contents of page to eliminate a Purify warning */
+		if (block->frame) {
-		memset(block->frame, '\0', UNIV_PAGE_SIZE);
+			/* Wipe contents of frame to eliminate a Purify
+			warning */
 
-		UT_LIST_ADD_FIRST(free, buf_pool->free, block);
+			memset(block->frame, '\0', UNIV_PAGE_SIZE);
+
+			if (srv_use_awe) {
+				/* Add to the list of blocks mapped to
+				frames */
+				
+				UT_LIST_ADD_LAST(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+			}
+		}
+
+		UT_LIST_ADD_LAST(free, buf_pool->free, block);
 	}
 
 	mutex_exit(&(buf_pool->mutex));
 
-	btr_search_sys_create(curr_size * UNIV_PAGE_SIZE / sizeof(void*) / 64);
+	if (srv_use_adaptive_hash_indexes) {
+	  	btr_search_sys_create(
+			  curr_size * UNIV_PAGE_SIZE / sizeof(void*) / 64);
+	} else {
+	        /* Create only a small dummy system */
+	        btr_search_sys_create(1000);
+	}
 
 	return(buf_pool);
 }	
 
 /************************************************************************
-Initializes the buffer buf_pool of the database. */
+Maps the page of block to a frame, if not mapped yet. Unmaps some page
+from the end of the awe_LRU_free_mapped. */
 
 void
-buf_pool_init(
+buf_awe_map_page_to_frame(
-/*==========*/
+/*======================*/
-	ulint	max_size,	/* in: maximum size of the buf_pool in blocks */
+	buf_block_t*	block,		/* in: block whose page should be
-	ulint	curr_size)	/* in: current size to use, must be <=
+					mapped to a frame */
-				max_size */
+	ibool		add_to_mapped_list) /* in: TRUE if we in the case
+					we need to map the page should also
+					add the block to the
+					awe_LRU_free_mapped list */
 {
-	ut_a(buf_pool == NULL);
+	buf_block_t*	bck;
 
-	buf_pool_create(max_size, curr_size);
+	ut_ad(mutex_own(&(buf_pool->mutex)));
+	ut_ad(block);
 
-	ut_ad(buf_validate());
+	if (block->frame) {
+
+		return;
+	}
+
+	/* Scan awe_LRU_free_mapped from the end and try to find a block
+	which is not bufferfixed or io-fixed */
+
+	bck = UT_LIST_GET_LAST(buf_pool->awe_LRU_free_mapped);
+
+	while (bck) {	
+		if (bck->state == BUF_BLOCK_FILE_PAGE
+	    	    && (bck->buf_fix_count != 0 || bck->io_fix != 0)) {
+
+			/* We have to skip this */
+			bck = UT_LIST_GET_PREV(awe_LRU_free_mapped, bck);
+		} else {
+			/* We can map block to the frame of bck */
+
+			os_awe_map_physical_mem_to_window(
+				bck->frame,
+				UNIV_PAGE_SIZE / OS_AWE_X86_PAGE_SIZE,
+				block->awe_info);
+
+			block->frame = bck->frame;
+
+			*(buf_pool->blocks_of_frames
+				+ (((ulint)(block->frame
+						- buf_pool->frame_zero))
+						>> UNIV_PAGE_SIZE_SHIFT))
+				= block;
+			
+			bck->frame = NULL;
+			UT_LIST_REMOVE(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped,
+					bck);
+
+			if (add_to_mapped_list) {
+				UT_LIST_ADD_FIRST(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped,
+					block);
+			}
+
+			buf_pool->n_pages_awe_remapped++;
+			
+			return;
+		}
+	}
+
+	fprintf(stderr,
+"InnoDB: AWE: Fatal error: cannot find a page to unmap\n"
+"InnoDB: awe_LRU_free_mapped list length %lu\n",
+		UT_LIST_GET_LEN(buf_pool->awe_LRU_free_mapped));
+
+	ut_a(0);
 }
 
 /************************************************************************
@@ -508,7 +692,9 @@ UNIV_INLINE
 buf_block_t*
 buf_block_alloc(void)
 /*=================*/
-				/* out, own: the allocated block */
+				/* out, own: the allocated block; also if AWE
+				is used it is guaranteed that the page is
+				mapped to a frame */
 {
 	buf_block_t*	block;
 
@@ -846,6 +1032,19 @@ loop:
 		}
 	}		
 
+	/* If AWE is enabled and the page is not mapped to a frame, then
+	map it */
+
+	if (block->frame == NULL) {
+		ut_a(srv_use_awe);
+
+		/* We set second parameter TRUE because the block is in the
+		LRU list and we must put it to awe_LRU_free_mapped list once
+		mapped to a frame */
+		
+		buf_awe_map_page_to_frame(block, TRUE);
+	}
+	
 #ifdef UNIV_SYNC_DEBUG
 	buf_block_buf_fix_inc_debug(block, file, line);
 #else
@@ -940,28 +1139,27 @@ buf_page_optimistic_get_func(
 /*=========================*/
 				/* out: TRUE if success */
 	ulint		rw_latch,/* in: RW_S_LATCH, RW_X_LATCH */
-	buf_frame_t*	guess,	/* in: guessed frame */
+	buf_block_t*	block,	/* in: guessed buffer block */
+	buf_frame_t*	guess,	/* in: guessed frame; note that AWE may move
+				frames */
 	dulint		modify_clock,/* in: modify clock value if mode is
 				..._GUESS_ON_CLOCK */
 	char*		file,	/* in: file name */
 	ulint		line,	/* in: line where called */
 	mtr_t*		mtr)	/* in: mini-transaction */
 {
-	buf_block_t*	block;
 	ibool		accessed;
 	ibool		success;
 	ulint		fix_type;
 
-	ut_ad(mtr && guess);
+	ut_ad(mtr && block);
 	ut_ad((rw_latch == RW_S_LATCH) || (rw_latch == RW_X_LATCH));
-
-	buf_pool->n_page_gets++;
-
-	block = buf_block_align(guess);
 	
 	mutex_enter(&(buf_pool->mutex));
 
-	if (block->state != BUF_BLOCK_FILE_PAGE) {
+	/* If AWE is used, block may have a different frame now, e.g., NULL */
+	
+	if (block->state != BUF_BLOCK_FILE_PAGE || block->frame != guess) {
 
 		mutex_exit(&(buf_pool->mutex));
 
@@ -1054,12 +1252,15 @@ buf_page_optimistic_get_func(
 #ifdef UNIV_IBUF_DEBUG
 	ut_a(ibuf_count_get(block->space, block->offset) == 0);
 #endif
+	buf_pool->n_page_gets++;
+
 	return(TRUE);
 }
 
 /************************************************************************
 This is used to get access to a known database page, when no waiting can be
-done. */
+done. For example, if a search in an adaptive hash index leads us to this
+frame. */
 
 ibool
 buf_page_get_known_nowait(
@@ -1078,13 +1279,11 @@ buf_page_get_known_nowait(
 
 	ut_ad(mtr);
 	ut_ad((rw_latch == RW_S_LATCH) || (rw_latch == RW_X_LATCH));
-
-	buf_pool->n_page_gets++;
-
-	block = buf_block_align(guess);
 	
 	mutex_enter(&(buf_pool->mutex));
 
+	block = buf_block_align(guess);
+
 	if (block->state == BUF_BLOCK_REMOVE_HASH) {
 	        /* Another thread is just freeing the block from the LRU list
 	        of the buffer pool: do not try to access this page; this
@@ -1152,6 +1351,8 @@ buf_page_get_known_nowait(
 	ut_a((mode == BUF_KEEP_OLD)
 		|| (ibuf_count_get(block->space, block->offset) == 0));
 #endif
+	buf_pool->n_page_gets++;
+
 	return(TRUE);
 }
 
@@ -1732,7 +1933,7 @@ buf_print(void)
 	
 	ut_ad(buf_pool);
 
-	size = buf_pool_get_curr_size() / UNIV_PAGE_SIZE;
+	size = buf_pool->curr_size;
 
 	index_ids = mem_alloc(sizeof(dulint) * size);
 	counts = mem_alloc(sizeof(ulint) * size);
@@ -1847,7 +2048,7 @@ buf_print_io(
 		return;
 	}
 
-	size = buf_pool_get_curr_size() / UNIV_PAGE_SIZE;
+	size = buf_pool->curr_size;
 
 	mutex_enter(&(buf_pool->mutex));
 	
@@ -1866,6 +2067,15 @@ buf_print_io(
 	buf += sprintf(buf,
 		"Modified db pages  %lu\n",
 				UT_LIST_GET_LEN(buf_pool->flush_list));
+	if (srv_use_awe) {
+		buf += sprintf(buf,
+		"AWE: Buffer pool memory frames                        %lu\n",
+				buf_pool->n_frames);
+		
+		buf += sprintf(buf,
+		"AWE: Database pages and free buffers mapped in frames %lu\n",
+				UT_LIST_GET_LEN(buf_pool->awe_LRU_free_mapped));
+	}
 
 	buf += sprintf(buf, "Pending reads %lu \n", buf_pool->n_pend_reads);
 
@@ -1891,6 +2101,13 @@ buf_print_io(
 		(buf_pool->n_pages_written - buf_pool->n_pages_written_old)
 		/ time_elapsed);
 
+	if (srv_use_awe) {
+		buf += sprintf(buf, "AWE: %.2f page remaps/s\n",
+		(buf_pool->n_pages_awe_remapped
+				- buf_pool->n_pages_awe_remapped_old)
+			/ time_elapsed);
+	}
+		
 	if (buf_pool->n_page_gets > buf_pool->n_page_gets_old) {
 		buf += sprintf(buf, "Buffer pool hit rate %lu / 1000\n",
 		1000
@@ -1906,6 +2123,7 @@ buf_print_io(
 	buf_pool->n_pages_read_old = buf_pool->n_pages_read;
 	buf_pool->n_pages_created_old = buf_pool->n_pages_created;
 	buf_pool->n_pages_written_old = buf_pool->n_pages_written;
+	buf_pool->n_pages_awe_remapped_old = buf_pool->n_pages_awe_remapped;
 
 	mutex_exit(&(buf_pool->mutex));
 }
@@ -1922,6 +2140,7 @@ buf_refresh_io_stats(void)
 	buf_pool->n_pages_read_old = buf_pool->n_pages_read;
 	buf_pool->n_pages_created_old = buf_pool->n_pages_created;
 	buf_pool->n_pages_written_old = buf_pool->n_pages_written;
+	buf_pool->n_pages_awe_remapped_old = buf_pool->n_pages_awe_remapped; 
 }
 
 /*************************************************************************

innobase/buf/buf0flu.c

@@ -24,6 +24,7 @@ Created 11/11/1995 Heikki Tuuri
 #include "log0log.h"
 #include "os0file.h"
 #include "trx0sys.h"
+#include "srv0srv.h"
 
 /* When flushed, dirty blocks are searched in neigborhoods of this size, and
 flushed along with the original page. */
@@ -103,7 +104,7 @@ buf_flush_ready_for_replace(
 /*========================*/
 				/* out: TRUE if can replace immediately */
 	buf_block_t*	block)	/* in: buffer control block, must be in state
-				BUF_BLOCK_FILE_PAGE and in the LRU list*/
+				BUF_BLOCK_FILE_PAGE and in the LRU list */
 {
 	ut_ad(mutex_own(&(buf_pool->mutex)));
 	ut_ad(block->state == BUF_BLOCK_FILE_PAGE);
@@ -134,7 +135,6 @@ buf_flush_ready_for_flush(
 
 	if ((ut_dulint_cmp(block->oldest_modification, ut_dulint_zero) > 0)
 	    					&& (block->io_fix == 0)) {
-
 	    	if (flush_type != BUF_FLUSH_LRU) {
 
 			return(TRUE);
@@ -436,6 +436,20 @@ buf_flush_try_page(
 	    && block && buf_flush_ready_for_flush(block, flush_type)) {
 	
 		block->io_fix = BUF_IO_WRITE;
+
+		/* If AWE is enabled and the page is not mapped to a frame,
+		then map it */
+
+		if (block->frame == NULL) {
+			ut_a(srv_use_awe);
+
+			/* We set second parameter TRUE because the block is
+			in the LRU list and we must put it to
+			awe_LRU_free_mapped list once mapped to a frame */
+		
+			buf_awe_map_page_to_frame(block, TRUE);
+		}
+
 		block->flush_type = flush_type;
 
 		if (buf_pool->n_flush[flush_type] == 0) {
@@ -486,6 +500,20 @@ buf_flush_try_page(
 		..._ready_for_flush). */
 
 		block->io_fix = BUF_IO_WRITE;
+
+		/* If AWE is enabled and the page is not mapped to a frame,
+		then map it */
+
+		if (block->frame == NULL) {
+			ut_a(srv_use_awe);
+
+			/* We set second parameter TRUE because the block is
+			in the LRU list and we must put it to
+			awe_LRU_free_mapped list once mapped to a frame */
+		
+			buf_awe_map_page_to_frame(block, TRUE);
+		}
+
 		block->flush_type = flush_type;
 
 		if (buf_pool->n_flush[flush_type] == 0) {
@@ -511,6 +539,20 @@ buf_flush_try_page(
 			&& buf_flush_ready_for_flush(block, flush_type)) {
 	
 		block->io_fix = BUF_IO_WRITE;
+
+		/* If AWE is enabled and the page is not mapped to a frame,
+		then map it */
+
+		if (block->frame == NULL) {
+			ut_a(srv_use_awe);
+
+			/* We set second parameter TRUE because the block is
+			in the LRU list and we must put it to
+			awe_LRU_free_mapped list once mapped to a frame */
+		
+			buf_awe_map_page_to_frame(block, TRUE);
+		}
+
 		block->flush_type = flush_type;
 
 		if (buf_pool->n_flush[block->flush_type] == 0) {

innobase/buf/buf0lru.c

@@ -132,7 +132,13 @@ buf_LRU_search_and_free_block(
 
 			mutex_exit(&(buf_pool->mutex));
 
-			btr_search_drop_page_hash_index(block->frame);
+			/* Remove possible adaptive hash index built on the
+			page; in the case of AWE the block may not have a
+			frame at all */
+			
+			if (block->frame) {
+				btr_search_drop_page_hash_index(block->frame);
+			}
 
 			mutex_enter(&(buf_pool->mutex));
 
@@ -196,7 +202,9 @@ list. */
 buf_block_t*
 buf_LRU_get_free_block(void)
 /*========================*/
-				/* out: the free control block */
+				/* out: the free control block; also if AWE is
+				used, it is guaranteed that the block has its
+				page mapped to a frame when we return */
 {
 	buf_block_t*	block		= NULL;
 	ibool		freed;
@@ -257,6 +265,22 @@ loop:
 		
 		block = UT_LIST_GET_FIRST(buf_pool->free);
 		UT_LIST_REMOVE(free, buf_pool->free, block);
+
+		if (srv_use_awe) {
+			if (block->frame) {
+				/* Remove from the list of mapped pages */
+		
+				UT_LIST_REMOVE(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+			} else {
+				/* We map the page to a frame; second param
+				FALSE below because we do not want it to be
+				added to the awe_LRU_free_mapped list */
+
+				buf_awe_map_page_to_frame(block, FALSE);
+			}
+		}
+		
 		block->state = BUF_BLOCK_READY_FOR_USE;
 
 		mutex_exit(&(buf_pool->mutex));
@@ -429,6 +453,13 @@ buf_LRU_remove_block(
 	/* Remove the block from the LRU list */
 	UT_LIST_REMOVE(LRU, buf_pool->LRU, block);
 
+	if (srv_use_awe && block->frame) {
+		/* Remove from the list of mapped pages */
+		
+		UT_LIST_REMOVE(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+	}	
+
 	/* If the LRU list is so short that LRU_old not defined, return */
 	if (UT_LIST_GET_LEN(buf_pool->LRU) < BUF_LRU_OLD_MIN_LEN) {
 
@@ -475,6 +506,13 @@ buf_LRU_add_block_to_end_low(
 
 	UT_LIST_ADD_LAST(LRU, buf_pool->LRU, block);
 
+	if (srv_use_awe && block->frame) {
+		/* Add to the list of mapped pages */
+		
+		UT_LIST_ADD_LAST(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+	}
+	
 	if (UT_LIST_GET_LEN(buf_pool->LRU) >= BUF_LRU_OLD_MIN_LEN) {
 
 		buf_pool->LRU_old_len++;
@@ -518,6 +556,15 @@ buf_LRU_add_block_low(
 	block->old = old;
 	cl = buf_pool_clock_tic();
 
+	if (srv_use_awe && block->frame) {
+		/* Add to the list of mapped pages; for simplicity we always
+		add to the start, even if the user would have set 'old'
+		TRUE */
+		
+		UT_LIST_ADD_FIRST(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+	}
+
 	if (!old || (UT_LIST_GET_LEN(buf_pool->LRU) < BUF_LRU_OLD_MIN_LEN)) {
 
 		UT_LIST_ADD_FIRST(LRU, buf_pool->LRU, block);
@@ -613,6 +660,13 @@ buf_LRU_block_free_non_file_page(
 	memset(block->frame, '\0', UNIV_PAGE_SIZE);
 #endif	
 	UT_LIST_ADD_FIRST(free, buf_pool->free, block);
+
+	if (srv_use_awe && block->frame) {
+		/* Add to the list of mapped pages */
+		
+		UT_LIST_ADD_FIRST(awe_LRU_free_mapped,
+					buf_pool->awe_LRU_free_mapped, block);
+	}
 }
 
 /**********************************************************************
@@ -639,7 +693,9 @@ buf_LRU_block_remove_hashed_page(
 
 	buf_pool->freed_page_clock += 1;
 
- 	buf_frame_modify_clock_inc(block->frame);
+	/* Note that if AWE is enabled the block may not have a frame at all */
+	
+ 	buf_block_modify_clock_inc(block);
 		
 	HASH_DELETE(buf_block_t, hash, buf_pool->page_hash,
 			buf_page_address_fold(block->space, block->offset),

innobase/buf/buf0rea.c

@@ -576,7 +576,7 @@ buf_read_recv_pages(
 
 		os_aio_print_debug = FALSE;
 
-		while (buf_pool->n_pend_reads >= RECV_POOL_N_FREE_BLOCKS / 2) {
+		while (buf_pool->n_pend_reads >= recv_n_pool_free_frames / 2) {
 
 			os_aio_simulated_wake_handler_threads();
 			os_thread_sleep(500000);

innobase/include/btr0pcur.h

@@ -466,6 +466,9 @@ struct btr_pcur_struct{
 					BTR_PCUR_AFTER, depending on whether
 					cursor was on, before, or after the
 					old_rec record */
+	buf_block_t*	block_when_stored;/* buffer block when the position was
+					stored; note that if AWE is on, frames
+					may move */
 	dulint		modify_clock;	/* the modify clock value of the
 					buffer block when the cursor position
 					was stored */

innobase/include/buf0buf.h

@@ -30,6 +30,7 @@ Created 11/5/1995 Heikki Tuuri
 #include "sync0rw.h"
 #include "hash0hash.h"
 #include "ut0byte.h"
+#include "os0proc.h"
 
 /* Flags for flush types */
 #define BUF_FLUSH_LRU		1
@@ -58,23 +59,34 @@ extern ibool		buf_debug_prints;/* If this is set TRUE, the program
 					occurs */
 
 /************************************************************************
-Initializes the buffer pool of the database. */
+Creates the buffer pool. */
 
-void
+buf_pool_t*
 buf_pool_init(
 /*==========*/
-	ulint	max_size,	/* in: maximum size of the pool in blocks */
+				/* out, own: buf_pool object, NULL if not
-	ulint	curr_size);	/* in: current size to use, must be <=
+				enough memory or error */
+	ulint	max_size,	/* in: maximum size of the buf_pool in
+				blocks */
+	ulint	curr_size,	/* in: current size to use, must be <=
+				max_size, currently must be equal to
 				max_size */
+	ulint	n_frames);	/* in: number of frames; if AWE is used,
+				this is the size of the address space window
+				where physical memory pages are mapped; if
+				AWE is not used then this must be the same
+				as max_size */
 /*************************************************************************
-Gets the current size of buffer pool in bytes. */
+Gets the current size of buffer buf_pool in bytes. In the case of AWE, the
+size of AWE window (= the frames). */
 UNIV_INLINE
 ulint
 buf_pool_get_curr_size(void);
 /*========================*/
 			/* out: size in bytes */
 /*************************************************************************
-Gets the maximum size of buffer pool in bytes. */
+Gets the maximum size of buffer pool in bytes. In the case of AWE, the
+size of AWE window (= the frames). */
 UNIV_INLINE
 ulint
 buf_pool_get_max_size(void);
@@ -138,8 +150,8 @@ improve debugging. Only values RW_S_LATCH and RW_X_LATCH are allowed as LA! */
 NOTE! The following macros should be used instead of
 buf_page_optimistic_get_func, to improve debugging. Only values RW_S_LATCH and
 RW_X_LATCH are allowed as LA! */
-#define buf_page_optimistic_get(LA, G, MC, MTR) buf_page_optimistic_get_func(\
+#define buf_page_optimistic_get(LA, BL, G, MC, MTR) buf_page_optimistic_get_func(\
-				LA, G, MC, IB__FILE__, __LINE__, MTR)
+				LA, BL, G, MC, IB__FILE__, __LINE__, MTR)
 /************************************************************************
 This is the general function used to get optimistic access to a database
 page. */
@@ -149,7 +161,9 @@ buf_page_optimistic_get_func(
 /*=========================*/
 				/* out: TRUE if success */
 	ulint		rw_latch,/* in: RW_S_LATCH, RW_X_LATCH */
-	buf_frame_t*	guess,	/* in: guessed frame */
+	buf_block_t*	block,	/* in: guessed block */
+	buf_frame_t*	guess,	/* in: guessed frame; note that AWE may move
+				frames */
 	dulint		modify_clock,/* in: modify clock value if mode is
 				..._GUESS_ON_CLOCK */
 	char*		file,	/* in: file name */
@@ -350,6 +364,16 @@ buf_frame_modify_clock_inc(
 				/* out: new value */
 	buf_frame_t*	frame);	/* in: pointer to a frame */
 /************************************************************************
+Increments the modify clock of a frame by 1. The caller must (1) own the
+buf_pool mutex and block bufferfix count has to be zero, (2) or own an x-lock
+on the block. */
+UNIV_INLINE
+dulint
+buf_block_modify_clock_inc(
+/*=======================*/
+				/* out: new value */
+	buf_block_t*	block);	/* in: block */
+/************************************************************************
 Returns the value of the modify clock. The caller must have an s-lock 
 or x-lock on the block. */
 UNIV_INLINE
@@ -428,7 +452,7 @@ UNIV_INLINE
 buf_frame_t*
 buf_frame_align(
 /*============*/
-			/* out: pointer to block */
+			/* out: pointer to frame */
 	byte*	ptr);	/* in: pointer to a frame */
 /***********************************************************************
 Checks if a pointer points to the block array of the buffer pool (blocks, not
@@ -505,6 +529,19 @@ buf_pool_invalidate(void);
 --------------------------- LOWER LEVEL ROUTINES -------------------------
 =========================================================================*/
 
+/************************************************************************
+Maps the page of block to a frame, if not mapped yet. Unmaps some page
+from the end of the awe_LRU_free_mapped. */
+
+void
+buf_awe_map_page_to_frame(
+/*======================*/
+	buf_block_t*	block,		/* in: block whose page should be
+					mapped to a frame */
+	ibool		add_to_mapped_list);/* in: TRUE if we in the case
+					we need to map the page should also
+					add the block to the
+					awe_LRU_free_mapped list */
 /*************************************************************************
 Adds latch level info for the rw-lock protecting the buffer frame. This
 should be called in the debug version after a successful latching of a
@@ -638,7 +675,16 @@ struct buf_block_struct{
 	byte*		frame;		/* pointer to buffer frame which
 					is of size UNIV_PAGE_SIZE, and
 					aligned to an address divisible by
-					UNIV_PAGE_SIZE */
+					UNIV_PAGE_SIZE; if AWE is used, this
+					will be NULL for the pages which are
+					currently not mapped into the virtual
+					address space window of the buffer
+					pool */
+	os_awe_t*	awe_info;	/* if AWE is used, then an array of
+					awe page infos for
+					UNIV_PAGE_SIZE / OS_AWE_X86_PAGE_SIZE
+					(normally = 4) physical memory
+					pages; otherwise NULL */
 	ulint		space;		/* space id of the page */
 	ulint		offset;		/* page number within the space */
 	ulint		lock_hash_val;	/* hashed value of the page address
@@ -691,6 +737,10 @@ struct buf_block_struct{
 					/* node of the free block list */
 	UT_LIST_NODE_T(buf_block_t) LRU;
 					/* node of the LRU list */
+	UT_LIST_NODE_T(buf_block_t) awe_LRU_free_mapped;
+					/* in the AWE version node in the
+					list of free and LRU blocks which are
+					mapped to a frame */
 	ulint		LRU_position;	/* value which monotonically
 					decreases (or may stay constant if
 					the block is in the old blocks) toward
@@ -758,11 +808,12 @@ struct buf_block_struct{
 					BTR_SEARCH_RIGHT_SIDE in hash
 					indexing */
 	/* 6. Debug fields */
-
+#ifdef UNIV_SYNC_DEBUG
 	rw_lock_t	debug_latch;	/* in the debug version, each thread
 					which bufferfixes the block acquires
 					an s-latch here; so we can use the
 					debug utilities in sync0rw */
+#endif
         ibool           file_page_was_freed;
                                         /* this is set to TRUE when fsp
                                         frees a page in buffer pool */
@@ -781,16 +832,36 @@ struct buf_pool_struct{
 					struct and control blocks, except the
 					read-write lock in them */
 	byte*		frame_mem;	/* pointer to the memory area which
-					was allocated for the frames */
+					was allocated for the frames; in AWE
+					this is the virtual address space
+					window where we map pages stored
+					in physical memory */
 	byte*		frame_zero;	/* pointer to the first buffer frame:
 					this may differ from frame_mem, because
 					this is aligned by the frame size */
-	byte*		high_end;	/* pointer to the end of the
+	byte*		high_end;	/* pointer to the end of the buffer
-					buffer pool */
+					frames */
+	ulint		n_frames;	/* number of frames */
 	buf_block_t*	blocks;		/* array of buffer control blocks */
+	buf_block_t**	blocks_of_frames;/* inverse mapping which can be used
+					to retrieve the buffer control block
+					of a frame; this is an array which
+					lists the blocks of frames in the
+					order frame_zero,
+					frame_zero + UNIV_PAGE_SIZE, ...
+					a control block is always assigned
+					for each frame, even if the frame does
+					not contain any data; note that in AWE
+					there are more control blocks than
+					buffer frames */
+	os_awe_t*	awe_info;	/* if AWE is used, AWE info for the
+					physical 4 kB memory pages associated
+					with buffer frames */
 	ulint		max_size;	/* number of control blocks ==
 					maximum pool size in pages */
-	ulint		curr_size;	/* current pool size in pages */
+	ulint		curr_size;	/* current pool size in pages;
+					currently always the same as
+					max_size */
 	hash_table_t*	page_hash;	/* hash table of the file pages */
 
 	ulint		n_pend_reads;	/* number of pending read operations */
@@ -802,11 +873,14 @@ struct buf_pool_struct{
 	ulint		n_pages_created;/* number of pages created in the pool
 					with no read */
 	ulint		n_page_gets;	/* number of page gets performed;
-					also successful seraches through
+					also successful searches through
 					the adaptive hash index are
 					counted as page gets; this field
 					is NOT protected by the buffer
 					pool mutex */
+	ulint		n_pages_awe_remapped; /* if AWE is enabled, the
+					number of remaps of blocks to
+					buffer frames */
 	ulint		n_page_gets_old;/* n_page_gets when buf_print was
 					last time called: used to calculate
 					hit rate */
@@ -815,6 +889,7 @@ struct buf_pool_struct{
 	ulint		n_pages_written_old;/* number write operations */
 	ulint		n_pages_created_old;/* number of pages created in
 					the pool with no read */
+	ulint		n_pages_awe_remapped_old;
 	/* 2. Page flushing algorithm fields */
 
 	UT_LIST_BASE_NODE_T(buf_block_t) flush_list;
@@ -847,7 +922,10 @@ struct buf_pool_struct{
 	/* 3. LRU replacement algorithm fields */
 
 	UT_LIST_BASE_NODE_T(buf_block_t) free;
-					/* base node of the free block list */
+					/* base node of the free block list;
+					in the case of AWE, at the start are
+					always free blocks for which the
+					physical memory is mapped to a frame */
 	UT_LIST_BASE_NODE_T(buf_block_t) LRU;
 					/* base node of the LRU list */
 	buf_block_t*	LRU_old; 	/* pointer to the about 3/8 oldest
@@ -859,6 +937,12 @@ struct buf_pool_struct{
 					see buf0lru.c for the restrictions
 					on this value; not defined if
 					LRU_old == NULL */
+	UT_LIST_BASE_NODE_T(buf_block_t) awe_LRU_free_mapped;
+					/* list of those blocks which are
+					in the LRU list or the free list, and
+					where the page is mapped to a frame;
+					thus, frames allocated, e.g., to the
+					locki table, are not in this list */
 };
 
 /* States of a control block */

innobase/include/buf0buf.ic

@@ -36,25 +36,27 @@ buf_block_peek_if_too_old(
 }
 
 /*************************************************************************
-Gets the current size of buffer buf_pool in bytes. */
+Gets the current size of buffer buf_pool in bytes. In the case of AWE, the
+size of AWE window (= the frames). */
 UNIV_INLINE
 ulint
 buf_pool_get_curr_size(void)
 /*========================*/
 			/* out: size in bytes */
 {
-	return((buf_pool->curr_size) * UNIV_PAGE_SIZE);
+	return((buf_pool->n_frames) * UNIV_PAGE_SIZE);
 }	
 
 /*************************************************************************
-Gets the maximum size of buffer buf_pool in bytes. */
+Gets the maximum size of buffer buf_pool in bytes. In the case of AWE, the
+size of AWE window (= the frames). */
 UNIV_INLINE
 ulint
 buf_pool_get_max_size(void)
 /*=======================*/
 			/* out: size in bytes */
 {
-	return((buf_pool->max_size) * UNIV_PAGE_SIZE);
+	return((buf_pool->n_frames) * UNIV_PAGE_SIZE);
 }	
 
 /***********************************************************************
@@ -207,54 +209,24 @@ buf_block_align(
 
 	frame_zero = buf_pool->frame_zero;
 
-	ut_ad((ulint)ptr >= (ulint)frame_zero);
+	if ((ulint)ptr < (ulint)frame_zero
-
+	    || (ulint)ptr > (ulint)(buf_pool->high_end)) {
-	block = buf_pool_get_nth_block(buf_pool, ((ulint)(ptr - frame_zero))
-						>> UNIV_PAGE_SIZE_SHIFT);
-	if (block < buf_pool->blocks
-	    || block >= buf_pool->blocks + buf_pool->max_size) {
 
+		ut_print_timestamp(stderr);	
 		fprintf(stderr,
-"InnoDB: Error: trying to access a stray pointer %lx\n"
+"  InnoDB: Error: trying to access a stray pointer %lx\n"
-"InnoDB: buf pool start is at %lx, number of pages %lu\n", (ulint)ptr,
+"InnoDB: buf pool start is at %lx, end at %lx\n"
-			(ulint)frame_zero, buf_pool->max_size);
+"InnoDB: Probable reason is database corruption or memory\n"
+"InnoDB: corruption. If this happens in an InnoDB database recovery,\n"
+"InnoDB: you can look from section 6.1 at http://www.innodb.com/ibman.html\n"
+"InnoDB: how to force recovery.\n",
+ 			(ulint)ptr, (ulint)frame_zero,
+					(ulint)(buf_pool->high_end));
 		ut_a(0);
 	}
-
+	
-	return(block);
+	block = *(buf_pool->blocks_of_frames + (((ulint)(ptr - frame_zero))
-}	
+						>> UNIV_PAGE_SIZE_SHIFT));
-
-/***********************************************************************
-Gets the block to whose frame the pointer is pointing to. Does not
-require a file page to be bufferfixed. */
-UNIV_INLINE
-buf_block_t*
-buf_block_align_low(
-/*================*/
-			/* out: pointer to block */
-	byte*	ptr)	/* in: pointer to a frame */
-{
-	buf_block_t*	block;
-	buf_frame_t*	frame_zero;
-
-	ut_ad(ptr);
-
-	frame_zero = buf_pool->frame_zero;
-
-	ut_ad((ulint)ptr >= (ulint)frame_zero);
-
-	block = buf_pool_get_nth_block(buf_pool, ((ulint)(ptr - frame_zero))
-						>> UNIV_PAGE_SIZE_SHIFT);
-	if (block < buf_pool->blocks
-	    || block >= buf_pool->blocks + buf_pool->max_size) {
-
-		fprintf(stderr,
-"InnoDB: Error: trying to access a stray pointer %lx\n"
-"InnoDB: buf pool start is at %lx, number of pages %lu\n", (ulint)ptr,
-			(ulint)frame_zero, buf_pool->max_size);
-		ut_a(0);
-	}
-
 	return(block);
 }	
 
@@ -264,7 +236,7 @@ UNIV_INLINE
 buf_frame_t*
 buf_frame_align(
 /*============*/
-			/* out: pointer to block */
+			/* out: pointer to frame */
 	byte*	ptr)	/* in: pointer to a frame */
 {
 	buf_frame_t*	frame;
@@ -273,14 +245,19 @@ buf_frame_align(
 
 	frame = ut_align_down(ptr, UNIV_PAGE_SIZE);
 
-	if (((ulint)frame
+	if (((ulint)frame < (ulint)(buf_pool->frame_zero))
-		< (ulint)(buf_pool->frame_zero))
+	    || (ulint)frame >= (ulint)(buf_pool->high_end)) {
-	    || ((ulint)frame > (ulint)(buf_pool_get_nth_block(buf_pool,
+
-					buf_pool->max_size - 1)->frame))) {
+		ut_print_timestamp(stderr);	
 		fprintf(stderr,
-"InnoDB: Error: trying to access a stray pointer %lx\n"
+"  InnoDB: Error: trying to access a stray pointer %lx\n"
-"InnoDB: buf pool start is at %lx, number of pages %lu\n", (ulint)ptr,
+"InnoDB: buf pool start is at %lx, end at %lx\n"
-			(ulint)(buf_pool->frame_zero), buf_pool->max_size);
+"InnoDB: Probable reason is database corruption or memory\n"
+"InnoDB: corruption. If this happens in an InnoDB database recovery,\n"
+"InnoDB: you can look from section 6.1 at http://www.innodb.com/ibman.html\n"
+"InnoDB: how to force recovery.\n",
+ 			(ulint)ptr, (ulint)(buf_pool->frame_zero),
+					(ulint)(buf_pool->high_end));
 		ut_a(0);
 	}
 
@@ -469,7 +446,7 @@ buf_frame_modify_clock_inc(
 
 	ut_ad(frame);
 
-	block = buf_block_align_low(frame);
+	block = buf_block_align(frame);
 
 	ut_ad((mutex_own(&(buf_pool->mutex)) && (block->buf_fix_count == 0))
 	      || rw_lock_own(&(block->lock), RW_LOCK_EXCLUSIVE));
@@ -479,6 +456,25 @@ buf_frame_modify_clock_inc(
 	return(block->modify_clock);
 }
 
+/************************************************************************
+Increments the modify clock of a frame by 1. The caller must (1) own the
+buf_pool mutex and block bufferfix count has to be zero, (2) or own an x-lock
+on the block. */
+UNIV_INLINE
+dulint
+buf_block_modify_clock_inc(
+/*=======================*/
+				/* out: new value */
+	buf_block_t*	block)	/* in: block */
+{
+	ut_ad((mutex_own(&(buf_pool->mutex)) && (block->buf_fix_count == 0))
+	      || rw_lock_own(&(block->lock), RW_LOCK_EXCLUSIVE));
+
+	UT_DULINT_INC(block->modify_clock);
+
+	return(block->modify_clock);
+}
+
 /************************************************************************
 Returns the value of the modify clock. The caller must have an s-lock 
 or x-lock on the block. */
@@ -508,15 +504,16 @@ void
 buf_block_buf_fix_inc_debug(
 /*========================*/
 	buf_block_t*	block,	/* in: block to bufferfix */
-	char*		file,	/* in: file name */
+	char*		file __attribute__ ((unused)),	/* in: file name */
-	ulint		line)	/* in: line */
+	ulint		line __attribute__ ((unused)))	/* in: line */
 {
+#ifdef UNIV_SYNC_DEBUG	
 	ibool	ret;
-	
+
 	ret = rw_lock_s_lock_func_nowait(&(block->debug_latch), file, line);
 
 	ut_ad(ret == TRUE);
-
+#endif
 	block->buf_fix_count++;
 }
 

innobase/include/buf0lru.h

@@ -53,7 +53,9 @@ LRU list to the free list. */
 buf_block_t*
 buf_LRU_get_free_block(void);
 /*=========================*/
-				/* out: the free control block */
+				/* out: the free control block; also if AWE is
+				used, it is guaranteed that the block has its
+				page mapped to a frame when we return */
 /**********************************************************************
 Puts a block back to the free list. */
 

innobase/include/log0recv.h

@@ -355,12 +355,7 @@ in the debug version: spaces with an odd number as the id are replicate
 spaces */
 #define RECV_REPLICA_SPACE_ADD	1
 
-/* This many blocks must be left free in the buffer pool when we scan
+extern ulint	recv_n_pool_free_frames;
-the log and store the scanned log records in the buffer pool: we will
-use these free blocks to read in pages when we start applying the
-log records to the database. */
-
-#define RECV_POOL_N_FREE_BLOCKS	 (ut_min(256, buf_pool_get_curr_size() / 8))
 
 #ifndef UNIV_NONINL
 #include "log0recv.ic"

innobase/include/os0proc.h

@@ -15,6 +15,76 @@ Created 9/30/1995 Heikki Tuuri
 typedef void*			os_process_t;
 typedef unsigned long int	os_process_id_t;
 
+/* The cell type in os_awe_allocate_mem page info */
+#ifdef __NT__
+typedef ULONG_PTR	os_awe_t;
+#else
+typedef ulint		os_awe_t;
+#endif
+
+/* Physical page size when Windows AWE is used. This is the normal
+page size of an Intel x86 processor. We cannot use AWE with 2 MB or 4 MB
+pages. */
+#define	OS_AWE_X86_PAGE_SIZE	4096
+
+/********************************************************************
+Windows AWE support. Tries to enable the "lock pages in memory" privilege for
+the current process so that the current process can allocate memory-locked
+virtual address space to act as the window where AWE maps physical memory. */
+
+ibool
+os_awe_enable_lock_pages_in_mem(void);
+/*=================================*/
+				/* out: TRUE if success, FALSE if error;
+				prints error info to stderr if no success */
+/********************************************************************
+Allocates physical RAM memory up to 64 GB in an Intel 32-bit x86
+processor. */
+
+ibool
+os_awe_allocate_physical_mem(
+/*=========================*/
+				/* out: TRUE if success */
+	os_awe_t** page_info,	/* out, own: array of opaque data containing
+				the info for allocated physical memory pages;
+				each allocated 4 kB physical memory page has
+				one slot of type os_awe_t in the array */
+	ulint	  n_megabytes);	/* in: number of megabytes to allocate */
+/********************************************************************
+Allocates a window in the virtual address space where we can map then
+pages of physical memory. */
+
+byte*
+os_awe_allocate_virtual_mem_window(
+/*===============================*/
+			/* out, own: allocated memory, or NULL if did not
+			succeed */
+	ulint	size);	/* in: virtual memory allocation size in bytes, must
+			be < 2 GB */
+/********************************************************************
+With this function you can map parts of physical memory allocated with
+the ..._allocate_physical_mem to the virtual address space allocated with
+the previous function. Intel implements this so that the process page
+tables are updated accordingly. A test on a 1.5 GHz AMD processor and XP
+showed that this takes < 1 microsecond, much better than the estimated 80 us
+for copying a 16 kB page memory to memory. But, the operation will at least
+partially invalidate the translation lookaside buffer (TLB) of all
+processors. Under a real-world load the performance hit may be bigger. */
+
+ibool
+os_awe_map_physical_mem_to_window(
+/*==============================*/
+					/* out: TRUE if success; the function
+					calls exit(1) in case of an error */
+	byte*		ptr,		/* in: a page-aligned pointer to
+					somewhere in the virtual address
+					space window; we map the physical mem
+					pages here */
+	ulint		n_mem_pages,	/* in: number of 4 kB mem pages to
+					map */
+	os_awe_t*	page_info);	/* in: array of page infos for those
+					pages; each page has one slot in the
+					array */
 /********************************************************************
 Converts the current process id to a number. It is not guaranteed that the
 number is unique. In Linux returns the 'process number' of the current

innobase/include/srv0srv.h

@@ -61,6 +61,7 @@ extern ulint	srv_flush_log_at_trx_commit;
 extern byte	srv_latin1_ordering[256];/* The sort order table of the latin1
 					character set */
 extern ulint	srv_pool_size;
+extern ulint	srv_awe_window_size;
 extern ulint	srv_mem_pool_size;
 extern ulint	srv_lock_table_size;
 
@@ -86,6 +87,8 @@ extern ibool	srv_use_doublewrite_buf;
 extern ibool    srv_set_thread_priorities;
 extern int      srv_query_thread_priority;
 
+extern ibool	srv_use_awe;
+extern ibool	srv_use_adaptive_hash_indexes;
 /*-------------------------------------------*/
 
 extern ulint	srv_n_rows_inserted;

innobase/log/log0log.c

@@ -437,25 +437,29 @@ log_group_calc_lsn_offset(
 	dulint		lsn,	/* in: lsn, must be within 4 GB of group->lsn */
 	log_group_t*	group)	/* in: log group */
 {
-	dulint	gr_lsn;
+        dulint	        gr_lsn;
-	ulint	gr_lsn_size_offset;
+       	ib_longlong     gr_lsn_size_offset;
-	ulint	difference;
+	ib_longlong	difference;
-	ulint	group_size;
+	ib_longlong	group_size;
-	ulint	offset;
+	ib_longlong	offset;
 	
 	ut_ad(mutex_own(&(log_sys->mutex)));
 
+	/* If total log file size is > 2 GB we can easily get overflows
+	with 32-bit integers. Use 64-bit integers instead. */
+
 	gr_lsn = group->lsn;
 
-	gr_lsn_size_offset = log_group_calc_size_offset(group->lsn_offset,
+	gr_lsn_size_offset = (ib_longlong)
-									group);
+	               log_group_calc_size_offset(group->lsn_offset, group);
-	group_size = log_group_get_capacity(group);
+
+	group_size = (ib_longlong) log_group_get_capacity(group);
 
 	if (ut_dulint_cmp(lsn, gr_lsn) >= 0) {
 			
-		difference = ut_dulint_minus(lsn, gr_lsn);
+		difference = (ib_longlong) ut_dulint_minus(lsn, gr_lsn);
 	} else {
-		difference = ut_dulint_minus(gr_lsn, lsn);
+		difference = (ib_longlong) ut_dulint_minus(gr_lsn, lsn);
 
 		difference = difference % group_size;
 
@@ -464,7 +468,13 @@ log_group_calc_lsn_offset(
 
 	offset = (gr_lsn_size_offset + difference) % group_size;
 
-	return(log_group_calc_real_offset(offset, group));
+	ut_a(offset <= 0xFFFFFFFF);
+
+	/* printf("Offset is %lu gr_lsn_offset is %lu difference is %lu\n",
+	       (ulint)offset,(ulint)gr_lsn_size_offset, (ulint)difference);
+	*/
+
+	return(log_group_calc_real_offset((ulint)offset, group));
 }
 
 /***********************************************************************
@@ -3054,8 +3064,8 @@ log_check_log_recs(
 	ut_memcpy(scan_buf, start, end - start);
 	
 	recv_scan_log_recs(TRUE,
-				buf_pool_get_curr_size() -
+				(buf_pool->n_frames -
-				RECV_POOL_N_FREE_BLOCKS * UNIV_PAGE_SIZE,	
+				recv_n_pool_free_frames) * UNIV_PAGE_SIZE,	
 				FALSE, scan_buf, end - start,
 				ut_dulint_align_down(buf_start_lsn,
 						OS_FILE_LOG_BLOCK_SIZE),

innobase/log/log0recv.c

@@ -71,6 +71,14 @@ ulint	recv_previous_parsed_rec_is_multi = 0;
 
 ulint	recv_max_parsed_page_no		= 0;
 
+/* This many frames must be left free in the buffer pool when we scan
+the log and store the scanned log records in the buffer pool: we will
+use these free frames to read in pages when we start applying the
+log records to the database. */
+
+ulint	recv_n_pool_free_frames		= 256;
+
+
 /************************************************************
 Creates the recovery system. */
 
@@ -1018,10 +1026,10 @@ recv_recover_page(
 		block = buf_block_align(page);
 
 		if (just_read_in) {
-		  /* Move the ownership of the x-latch on the page to this OS
+		  	/* Move the ownership of the x-latch on the page to
-		  thread, so that we can acquire a second x-latch on it. This
+			this OS thread, so that we can acquire a second
-		  is needed for the operations to the page to pass the debug
+			x-latch on it. This is needed for the operations to
-		  checks. */
+			the page to pass the debug checks. */
 
 			rw_lock_x_lock_move_ownership(&(block->lock));
 		}
@@ -2362,8 +2370,8 @@ recv_group_scan_log_recs(
 						group, start_lsn, end_lsn);
 
 		finished = recv_scan_log_recs(TRUE,
-				buf_pool_get_curr_size()
+				(buf_pool->n_frames
-				- RECV_POOL_N_FREE_BLOCKS * UNIV_PAGE_SIZE,
+				- recv_n_pool_free_frames) * UNIV_PAGE_SIZE,
 				TRUE, log_sys->buf,
 				RECV_SCAN_SIZE, start_lsn,
 				contiguous_lsn, group_scanned_lsn);
@@ -3001,8 +3009,8 @@ ask_again:
 			read_offset % UNIV_PAGE_SIZE, len, buf, NULL);
 
 		ret = recv_scan_log_recs(TRUE,
-				buf_pool_get_curr_size() -
+				(buf_pool->n_frames -
-				RECV_POOL_N_FREE_BLOCKS * UNIV_PAGE_SIZE,
+				recv_n_pool_free_frames) * UNIV_PAGE_SIZE,
 				TRUE, buf, len, start_lsn,
 				&dummy_lsn, &scanned_lsn);
 

innobase/os/os0file.c

@@ -2127,7 +2127,7 @@ os_aio_simulated_handle(
 	ulint		offs;
 	ulint		lowest_offset;
 	byte*		combined_buf;
-	byte*		combined_buf2= 0;	/* Remove warning */
+	byte*		combined_buf2;
 	ibool		ret;
 	ulint		n;
 	ulint		i;

innobase/os/os0proc.c

@@ -12,11 +12,469 @@ Created 9/30/1995 Heikki Tuuri
 #include "os0proc.ic"
 #endif
 
-#ifdef __WIN__
+#include "ut0mem.h"
-#include <windows.h>
+#include "ut0byte.h"
+
+
+/*
+How to get AWE to compile on Windows?
+-------------------------------------
+
+the Visual C++ has to be relatively recent and _WIN32_WINNT has to be
+defined to a value >= 0x0500 when windows.h is included. An easy way
+to accomplish that is to put
+
+#define _WIN32_WINNT	0x0500
+
+to the start of file \mysql\include\config-win.h
+
+Where does AWE work?
+-------------------
+
+See the error message in os_awe_allocate_physical_mem().
+
+How to assign privileges for mysqld to use AWE?
+-----------------------------------------------
+
+See the error message in os_awe_enable_lock_pages_in_mem().
+
+Use Windows AWE functions in this order
+---------------------------------------
+
+(1) os_awe_enable_lock_pages_in_mem();
+(2) os_awe_allocate_physical_mem();
+(3) os_awe_allocate_virtual_mem_window();
+(4) os_awe_map_physical_mem_to_window().
+
+To test 'AWE' in a computer which does not have the AWE API,
+you can compile with UNIV_SIMULATE_AWE defined in this file.
+*/
+
+#ifdef UNIV_SIMULATE_AWE
+/* If we simulate AWE, we allocate the 'physical memory' here */
+byte*		os_awe_simulate_mem;
+ulint		os_awe_simulate_mem_size;
+os_awe_t*	os_awe_simulate_page_info;
+byte*		os_awe_simulate_window;
+ulint		os_awe_simulate_window_size;
+/* In simulated AWE the following contains a NULL pointer or a pointer
+to a mapped 'physical page' for each 4 kB page in the AWE window */
+byte**		os_awe_simulate_map;
 #endif
 
-#include "ut0mem.h"
+#ifdef __NT__
+os_awe_t*	os_awe_page_info;
+ulint		os_awe_n_pages;
+byte*		os_awe_window;
+ulint		os_awe_window_size;
+#endif
+
+/********************************************************************
+Windows AWE support. Tries to enable the "lock pages in memory" privilege for
+the current process so that the current process can allocate memory-locked
+virtual address space to act as the window where AWE maps physical memory. */
+
+ibool
+os_awe_enable_lock_pages_in_mem(void)
+/*=================================*/
+				/* out: TRUE if success, FALSE if error;
+				prints error info to stderr if no success */
+{
+#ifdef UNIV_SIMULATE_AWE
+
+	return(TRUE);
+
+#elif defined(__NT__)
+  	struct {
+    	DWORD 			Count;
+    	LUID_AND_ATTRIBUTES 	Privilege[1];
+  	} 	Info;
+	HANDLE	hProcess;
+  	HANDLE	Token;
+  	BOOL 	Result;
+
+	hProcess = GetCurrentProcess();
+
+  	/* Open the token of the current process */
+
+  	Result = OpenProcessToken(hProcess,
+                              TOKEN_ADJUST_PRIVILEGES,
+                              &Token);
+  	if (Result != TRUE) {
+    		fprintf(stderr,
+			"InnoDB: AWE: Cannot open process token, error %lu\n",
+			(ulint)GetLastError());
+    		return(FALSE);
+  	}
+
+  	Info.Count = 1;
+
+    	Info.Privilege[0].Attributes = SE_PRIVILEGE_ENABLED;
+
+  	/* Get the local unique identifier (LUID) of the SE_LOCK_MEMORY
+	privilege */
+
+  	Result = LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME,
+                                  &(Info.Privilege[0].Luid));
+  	if (Result != TRUE)  {
+    		fprintf(stderr,
+	"InnoDB: AWE: Cannot get local privilege value for %s, error %lu.\n",
+			SE_LOCK_MEMORY_NAME, (ulint)GetLastError());
+
+    		return(FALSE);
+  	}
+
+  	/* Try to adjust the privilege */
+
+  	Result = AdjustTokenPrivileges(Token, FALSE,
+                                   (PTOKEN_PRIVILEGES)&Info,
+                                   0, NULL, NULL);
+  	/* Check the result */
+
+  	if (Result != TRUE)  {
+    		fprintf(stderr,
+		"InnoDB: AWE: Cannot adjust process token privileges, error %u.\n",
+			GetLastError());
+    		return(FALSE);
+  	} else if (GetLastError() != ERROR_SUCCESS) {
+      		fprintf(stderr,
+"InnoDB: AWE: Cannot enable SE_LOCK_MEMORY privilege, error %lu.\n"
+"InnoDB: In Windows XP Home you cannot use AWE. In Windows 2000 and XP\n"
+"InnoDB: Professional you must go to the Control Panel, to\n"
+"InnoDB: Security Settings, to Local Policies, and enable\n"
+"InnoDB: the 'lock pages in memory' privilege for the user who runs\n"
+"InnoDB: the MySQL server.\n", GetLastError());
+
+		return(FALSE);
+	}
+
+	CloseHandle(Token);
+
+	return(TRUE);
+#else
+#ifdef __WIN__
+	fprintf(stderr,
+"InnoDB: AWE: Error: to use AWE you must use a ...-nt MySQL executable.\n");
+#endif	
+	return(FALSE);
+#endif
+}
+
+/********************************************************************
+Allocates physical RAM memory up to 64 GB in an Intel 32-bit x86
+processor. */
+
+ibool
+os_awe_allocate_physical_mem(
+/*=========================*/
+				/* out: TRUE if success */
+	os_awe_t** page_info,	/* out, own: array of opaque data containing
+				the info for allocated physical memory pages;
+				each allocated 4 kB physical memory page has
+				one slot of type os_awe_t in the array */
+	ulint	  n_megabytes)	/* in: number of megabytes to allocate */
+{
+#ifdef UNIV_SIMULATE_AWE
+	os_awe_simulate_page_info = ut_malloc(sizeof(os_awe_t) *
+		n_megabytes * ((1024 * 1024) / OS_AWE_X86_PAGE_SIZE));
+
+	os_awe_simulate_mem = ut_align(ut_malloc(
+					4096 + 1024 * 1024 * n_megabytes),
+					4096);
+	os_awe_simulate_mem_size = n_megabytes * 1024 * 1024;
+
+	*page_info = os_awe_simulate_page_info;
+
+	return(TRUE);
+
+#elif defined(__NT__)
+	BOOL		bResult;
+  	ULONG_PTR 	NumberOfPages;		/* Question: why does Windows
+  						use the name ULONG_PTR for
+  						a scalar integer type? Maybe
+  						because we may also refer to
+  						&NumberOfPages? */
+  	ULONG_PTR 	NumberOfPagesInitial;
+  	SYSTEM_INFO 	sSysInfo;
+  	int 		PFNArraySize;
+
+	if (n_megabytes > 64 * 1024) {
+
+		fprintf(stderr,
+"InnoDB: AWE: Error: tried to allocate %lu MB.\n"
+"InnoDB: AWE cannot allocate more than 64 GB in any computer.\n", n_megabytes);
+
+		return(FALSE);
+	}
+
+  	GetSystemInfo(&sSysInfo);  /* fill the system information structure */
+
+  	if ((ulint)OS_AWE_X86_PAGE_SIZE != (ulint)sSysInfo.dwPageSize) {
+		fprintf(stderr,
+"InnoDB: AWE: Error: this computer has a page size of %lu.\n"
+"InnoDB: Should be 4096 bytes for InnoDB AWE support to work.\n",
+			(ulint)sSysInfo.dwPageSize);
+
+		return(FALSE);
+	}
+
+  	/* Calculate the number of pages of memory to request */
+
+  	NumberOfPages = n_megabytes * ((1024 * 1024) / OS_AWE_X86_PAGE_SIZE);
+ 
+ 	/* Calculate the size of page_info for allocated physical pages */
+
+  	PFNArraySize = NumberOfPages * sizeof(ULONG_PTR);
+
+   	*page_info = (ULONG_PTR*)HeapAlloc(GetProcessHeap(), 0, PFNArraySize);
+
+	if (*page_info == NULL) {
+    		fprintf(stderr,
+"InnoDB: AWE: Failed to allocate page info array from process heap, error %lu\n",
+			(ulint)GetLastError());
+
+    		return(FALSE);
+  	}
+
+	ut_total_allocated_memory += PFNArraySize;
+
+  	/* Enable this process' privilege to lock pages to physical memory */
+
+	if (!os_awe_enable_lock_pages_in_mem()) {
+
+		return(FALSE);
+	}
+
+  	/* Allocate the physical memory */
+
+  	NumberOfPagesInitial = NumberOfPages;
+
+	os_awe_page_info = *page_info;
+	os_awe_n_pages = (ulint)NumberOfPages;
+
+	/* Compilation note: if the compiler complains the function is not
+	defined, see the note at the start of this file */
+
+ 	bResult = AllocateUserPhysicalPages(GetCurrentProcess(),
+                                       &NumberOfPages,
+                                       *page_info);
+  	if (bResult != TRUE) {
+    		fprintf(stderr,
+"InnoDB: AWE: Cannot allocate physical pages, error %lu.\n",
+			(ulint)GetLastError());
+
+    		return(FALSE);
+  	}
+
+  	if (NumberOfPagesInitial != NumberOfPages) {
+    		fprintf(stderr,
+"InnoDB: AWE: Error: allocated only %lu pages of %lu requested.\n"
+"InnoDB: Check that you have enough free RAM.\n"
+"InnoDB: In Windows XP Professional and 2000 Professional\n"
+"InnoDB: Windows PAE size is max 4 GB. In 2000 and .NET"
+"InnoDB: Advanced Servers and 2000 Datacenter Server it is 32 GB,\n"
+"InnoDB: and in .NET Datacenter Server it is 64 GB.\n"
+"InnoDB: A Microsoft web page said that the processor must be an Intel\n"
+"InnoDB: processor.",
+			(ulint)NumberOfPages,
+			(ulint)NumberOfPagesInitial);
+
+    		return(FALSE);
+  	}
+
+	fprintf(stderr,
+"InnoDB: Using Address Windowing Extensions (AWE); allocated %lu MB\n",
+		n_megabytes);
+
+	return(TRUE);	
+#else
+	return(FALSE);
+#endif
+}
+
+/********************************************************************
+Allocates a window in the virtual address space where we can map then
+pages of physical memory. */
+
+byte*
+os_awe_allocate_virtual_mem_window(
+/*===============================*/
+			/* out, own: allocated memory, or NULL if did not
+			succeed */
+	ulint	size)	/* in: virtual memory allocation size in bytes, must
+			be < 2 GB */
+{
+#ifdef UNIV_SIMULATE_AWE
+	ulint	i;
+
+	os_awe_simulate_window = ut_align(ut_malloc(4096 + size), 4096);
+	os_awe_simulate_window_size = size;
+
+	os_awe_simulate_map = ut_malloc(sizeof(byte*) * (size / 4096));
+
+	for (i = 0; i < (size / 4096); i++) {
+		*(os_awe_simulate_map + i) = NULL;
+	}
+
+	return(os_awe_simulate_window);
+	
+#elif defined(__NT__)
+	byte*	ptr;
+
+	if (size > 0x7FFFFFFFFF) {
+		fprintf(stderr,
+"InnoDB: AWE: Cannot allocate %lu bytes of virtual memory\n", size);
+
+		return(NULL);
+	}
+	
+	ptr = VirtualAlloc(NULL, (SIZE_T)size, MEM_RESERVE | MEM_PHYSICAL,
+							PAGE_READWRITE);
+	if (ptr == NULL) {
+		fprintf(stderr,
+"InnoDB: AWE: Cannot allocate %lu bytes of virtual memory, error %lu\n",
+			size, (ulint)GetLastError());
+
+		return(NULL);
+	}
+
+	os_awe_window = ptr;
+	os_awe_window_size = size;
+
+	ut_total_allocated_memory += size;
+
+	return(ptr);
+#else
+	return(NULL);
+#endif
+}
+
+/********************************************************************
+With this function you can map parts of physical memory allocated with
+the ..._allocate_physical_mem to the virtual address space allocated with
+the previous function. Intel implements this so that the process page
+tables are updated accordingly. A test on a 1.5 GHz AMD processor and XP
+showed that this takes < 1 microsecond, much better than the estimated 80 us
+for copying a 16 kB page memory to memory. But, the operation will at least
+partially invalidate the translation lookaside buffer (TLB) of all
+processors. Under a real-world load the performance hit may be bigger. */
+
+ibool
+os_awe_map_physical_mem_to_window(
+/*==============================*/
+					/* out: TRUE if success; the function
+					calls exit(1) in case of an error */
+	byte*		ptr,		/* in: a page-aligned pointer to
+					somewhere in the virtual address
+					space window; we map the physical mem
+					pages here */
+	ulint		n_mem_pages,	/* in: number of 4 kB mem pages to
+					map */
+	os_awe_t*	page_info)	/* in: array of page infos for those
+					pages; each page has one slot in the
+					array */
+{
+#ifdef UNIV_SIMULATE_AWE
+	ulint	i;
+	byte**	map;
+	byte*	page;
+	byte*	phys_page;
+
+	ut_a(ptr >= os_awe_simulate_window);
+	ut_a(ptr < os_awe_simulate_window + os_awe_simulate_window_size);
+	ut_a(page_info >= os_awe_simulate_page_info);
+	ut_a(page_info < os_awe_simulate_page_info +
+			 		(os_awe_simulate_mem_size / 4096));
+
+	/* First look if some other 'physical pages' are mapped at ptr,
+	and copy them back to where they were if yes */
+
+	map = os_awe_simulate_map
+			+ ((ulint)(ptr - os_awe_simulate_window)) / 4096;
+	page = ptr;
+		
+	for (i = 0; i < n_mem_pages; i++) {
+		if (*map != NULL) {
+			ut_memcpy(*map, page, 4096);
+		}
+		map++;
+		page += 4096;
+	}
+
+	/* Then copy to ptr the 'physical pages' determined by page_info; we
+	assume page_info is a segment of the array we created at the start */
+
+	phys_page = os_awe_simulate_mem
+			+ (ulint)(page_info - os_awe_simulate_page_info)
+			  * 4096;
+
+	ut_memcpy(ptr, phys_page, n_mem_pages * 4096);
+
+	/* Update the map */
+
+	map = os_awe_simulate_map
+			+ ((ulint)(ptr - os_awe_simulate_window)) / 4096;
+
+	for (i = 0; i < n_mem_pages; i++) {
+		*map = phys_page;
+
+		map++;
+		phys_page += 4096;
+	}
+
+	return(TRUE);
+	
+#elif defined(__NT__)
+	BOOL		bResult;
+	ULONG_PTR	n_pages;
+
+	n_pages = (ULONG_PTR)n_mem_pages;
+	
+	if (!(ptr >= os_awe_window)) {
+		fprintf(stderr,
+"InnoDB: AWE: Error: trying to map to address %lx but AWE window start %lx\n",
+		(ulint)ptr, (ulint)os_awe_window);
+		ut_a(0);
+	}
+
+	if (!(ptr <= os_awe_window + os_awe_window_size - UNIV_PAGE_SIZE)) {
+		fprintf(stderr,
+"InnoDB: AWE: Error: trying to map to address %lx but AWE window end %lx\n",
+		(ulint)ptr, (ulint)os_awe_window + os_awe_window_size);
+		ut_a(0);
+	}
+
+	if (!(page_info >= os_awe_page_info)) {
+		fprintf(stderr,
+"InnoDB: AWE: Error: trying to map page info at %lx but array start %lx\n",
+		(ulint)page_info, (ulint)os_awe_page_info);
+		ut_a(0);
+	}
+
+	if (!(page_info <= os_awe_page_info + (os_awe_n_pages - 4))) {
+		fprintf(stderr,
+"InnoDB: AWE: Error: trying to map page info at %lx but array end %lx\n",
+		(ulint)page_info, (ulint)(os_awe_page_info + os_awe_n_pages));
+		ut_a(0);
+	}
+
+	bResult = MapUserPhysicalPages((PVOID)ptr, n_pages, page_info);
+
+	if (bResult != TRUE) {
+		ut_print_timestamp(stderr);
+		fprintf(stderr,
+"  InnoDB: AWE: Mapping of %lu physical pages to address %lx failed,\n"
+"InnoDB: error %lu.\n"
+"InnoDB: Cannot continue operation.\n",
+			n_mem_pages, (ulint)ptr, (ulint)GetLastError());
+		exit(1);
+	}
+
+	return(TRUE);
+#else
+	return(FALSE);
+#endif
+}	
 
 /********************************************************************
 Converts the current process id to a number. It is not guaranteed that the

innobase/srv/srv0srv.c

@@ -140,9 +140,14 @@ byte	srv_latin1_ordering[256]	/* The sort order table of the latin1
 , 0xD8, 0x55, 0x55, 0x55, 0x59, 0x59, 0xDE, 0xFF
 };
 		
-ulint	srv_pool_size		= ULINT_MAX;	/* size in database pages;
+ulint	srv_pool_size		= ULINT_MAX;	/* size in pages; MySQL inits
-						MySQL originally sets this
+						this to size in kilobytes but
-						value in megabytes */ 
+						we normalize this to pages in
+						srv_boot() */
+ulint	srv_awe_window_size 	= 0;		/* size in pages; MySQL inits
+						this to bytes, but we
+						normalize it to pages in
+						srv_boot() */
 ulint	srv_mem_pool_size	= ULINT_MAX;	/* size in bytes */ 
 ulint	srv_lock_table_size	= ULINT_MAX;
 
@@ -218,6 +223,13 @@ ibool	srv_use_doublewrite_buf	= TRUE;
 
 ibool   srv_set_thread_priorities = TRUE;
 int     srv_query_thread_priority = 0;
+
+/* TRUE if the Address Windowing Extensions of Windows are used; then we must
+disable adaptive hash indexes */
+ibool	srv_use_awe			= FALSE;
+ibool	srv_use_adaptive_hash_indexes 	= TRUE;
+
+
 /*-------------------------------------------*/
 ulint	srv_n_spin_wait_rounds	= 20;
 ulint	srv_spin_wait_delay	= 5;
@@ -1956,9 +1968,19 @@ srv_normalize_init_values(void)
 
 	srv_log_buffer_size = srv_log_buffer_size / UNIV_PAGE_SIZE;
 
-	srv_pool_size = srv_pool_size / UNIV_PAGE_SIZE;
+	srv_pool_size = srv_pool_size / (UNIV_PAGE_SIZE / 1024);
+
+	srv_awe_window_size = srv_awe_window_size / UNIV_PAGE_SIZE;
 	
-	srv_lock_table_size = 20 * srv_pool_size;
+	if (srv_use_awe) {
+	        /* If we are using AWE we must save memory in the 32-bit
+		address space of the process, and cannot bind the lock
+		table size to the real buffer pool size. */
+
+	        srv_lock_table_size = 20 * srv_awe_window_size;
+	} else {
+	        srv_lock_table_size = 20 * srv_pool_size;
+	}
 
 	return(DB_SUCCESS);
 }
@@ -2323,6 +2345,12 @@ srv_sprintf_innodb_monitor(
 	"Total memory allocated %lu; in additional pool allocated %lu\n",
 				ut_total_allocated_memory,
 				mem_pool_get_reserved(mem_comm_pool));
+	if (srv_use_awe) {
+		buf += sprintf(buf,
+	"In addition to that %lu MB of AWE memory allocated\n",
+		srv_pool_size / ((1024 * 1024) / UNIV_PAGE_SIZE));
+	}
+	
 	buf_print_io(buf, buf_end);
 	buf = buf + strlen(buf);
 	ut_a(buf < buf_end + 1500);

innobase/srv/srv0start.c

@@ -935,6 +935,7 @@ innobase_start_or_create_for_mysql(void)
 /*====================================*/
 				/* out: DB_SUCCESS or error code */
 {
+	buf_pool_t*	ret;
 	ibool	create_new_db;
 	ibool	log_file_created;
 	ibool	log_created	= FALSE;
@@ -970,6 +971,11 @@ innobase_start_or_create_for_mysql(void)
 #ifdef UNIV_MEM_DEBUG
 	fprintf(stderr,
 "InnoDB: !!!!!!!!!!!!!! UNIV_MEM_DEBUG switched on !!!!!!!!!!!!!!!\n"); 
+#endif
+
+#ifdef UNIV_SIMULATE_AWE
+	fprintf(stderr,
+"InnoDB: !!!!!!!!!!!!!! UNIV_SIMULATE_AWE switched on !!!!!!!!!!!!!!!!!\n");
 #endif
 
         if (srv_sizeof_trx_t_in_ha_innodb_cc != (ulint)sizeof(trx_t)) {
@@ -1002,6 +1008,17 @@ innobase_start_or_create_for_mysql(void)
         srv_startup_is_before_trx_rollback_phase = TRUE;
 	os_aio_use_native_aio = FALSE;
 
+#if !defined(__NT__) && !defined(UNIV_SIMULATE_AWE)
+	if (srv_use_awe) {
+
+	        fprintf(stderr,
+"InnoDB: Error: You have specified innodb_buffer_pool_awe_mem_mb\n"
+"InnoDB: in my.cnf, but AWE can only be used in Windows 2000 and later.\n");
+
+	        return(DB_ERROR);
+	}
+#endif
+
 #ifdef __WIN__
 	if (os_get_os_version() == OS_WIN95
 	    || os_get_os_version() == OS_WIN31
@@ -1057,6 +1074,9 @@ innobase_start_or_create_for_mysql(void)
 	  	return(DB_ERROR);
 	}
 
+	/* Note that the call srv_boot() also changes the values of
+	srv_pool_size etc. to the units used by InnoDB internally */
+	
 	err = srv_boot();
 
 	if (err != DB_SUCCESS) {
@@ -1088,7 +1108,26 @@ innobase_start_or_create_for_mysql(void)
 	
 	fil_init(SRV_MAX_N_OPEN_FILES);
 
-	buf_pool_init(srv_pool_size, srv_pool_size);
+	if (srv_use_awe) {
+		fprintf(stderr,
+"InnoDB: Using AWE: Memory window is %lu MB and AWE memory is %lu MB\n",
+		srv_awe_window_size / ((1024 * 1024) / UNIV_PAGE_SIZE),
+		srv_pool_size / ((1024 * 1024) / UNIV_PAGE_SIZE));
+
+		/* We must disable adaptive hash indexes because they do not
+		tolerate remapping of pages in AWE */
+		
+		srv_use_adaptive_hash_indexes = FALSE;
+		ret = buf_pool_init(srv_pool_size, srv_pool_size,
+							srv_awe_window_size);
+	} else {
+		ret = buf_pool_init(srv_pool_size, srv_pool_size,
+							srv_pool_size);
+	}
+
+	if (ret == NULL) {
+		return(DB_ERROR);
+	}
 
 	fsp_init();
 	log_init();

innobase/trx/trx0sys.c

@@ -472,9 +472,9 @@ trx_sys_update_mysql_binlog_offset(
 	if (0 != ut_memcmp(sys_header + field + TRX_SYS_MYSQL_LOG_NAME,
 			file_name, 1 + ut_strlen(file_name))) {
 
-		mlog_write_string((byte*) (sys_header + field
+		mlog_write_string(sys_header + field
-					+ TRX_SYS_MYSQL_LOG_NAME),
+					+ TRX_SYS_MYSQL_LOG_NAME,
-			(byte*) file_name, 1 + ut_strlen(file_name), mtr);
+			file_name, 1 + ut_strlen(file_name), mtr);
 	}
 
 	if (mach_read_from_4(sys_header + field

innobase/trx/trx0trx.c

@@ -99,7 +99,7 @@ trx_create(
 
 	trx->mysql_log_file_name = NULL;
 	trx->mysql_log_offset = 0;
-	trx->mysql_master_log_file_name = (char*) "";
+	trx->mysql_master_log_file_name = "";
 	trx->mysql_master_log_pos = 0;
 	
 	trx->ignore_duplicates_in_insert = FALSE;

innobase/ut/ut0ut.c

@@ -197,6 +197,7 @@ ut_get_year_month_day(
   	*month = (ulint)cal_tm.wMonth;
   	*day = (ulint)cal_tm.wDay;
 #else
+  	struct tm  cal_tm;
   	struct tm* cal_tm_ptr;
   	time_t     tm;
 

sql/ha_innodb.cc

@@ -82,7 +82,8 @@ are declared in mysqld.cc: */
 
 long innobase_mirrored_log_groups, innobase_log_files_in_group,
      innobase_log_file_size, innobase_log_buffer_size,
-     innobase_buffer_pool_size, innobase_additional_mem_pool_size,
+     innobase_buffer_pool_size, innobase_buffer_pool_awe_mem_mb,
+     innobase_additional_mem_pool_size,
      innobase_file_io_threads, innobase_lock_wait_timeout,
      innobase_thread_concurrency, innobase_force_recovery;
 
@@ -753,7 +754,25 @@ innobase_init(void)
 	srv_log_buffer_size = (ulint) innobase_log_buffer_size;
 	srv_flush_log_at_trx_commit = (ulint) innobase_flush_log_at_trx_commit;
 
-	srv_pool_size = (ulint) innobase_buffer_pool_size;
+	/* We set srv_pool_size here in units of 1 kB. InnoDB internally
+	changes the value so that it becomes the number of database pages. */
+
+        if (innobase_buffer_pool_awe_mem_mb == 0) {
+	        /* Careful here: we first convert the signed long int to ulint
+	        and only after that divide */
+
+	        srv_pool_size = ((ulint) innobase_buffer_pool_size) / 1024;
+	} else {
+	        srv_use_awe = TRUE;
+	        srv_pool_size = (ulint)
+		                (1024 * innobase_buffer_pool_awe_mem_mb);
+	        srv_awe_window_size = (ulint) innobase_buffer_pool_size;
+
+		/* Note that what the user specified as
+		innodb_buffer_pool_size is actually the AWE memory window
+		size in this case, and the real buffer pool size is
+		determined by .._awe_mem_mb. */
+	}
 
 	srv_mem_pool_size = (ulint) innobase_additional_mem_pool_size;
 

sql/ha_innodb.h

@@ -178,7 +178,8 @@ extern char *innobase_home, *innobase_tmpdir, *innobase_logdir;
 extern long innobase_lock_scan_time;
 extern long innobase_mirrored_log_groups, innobase_log_files_in_group;
 extern long innobase_log_file_size, innobase_log_buffer_size;
-extern long innobase_buffer_pool_size, innobase_additional_mem_pool_size;
+extern long innobase_buffer_pool_size, innobase_buffer_pool_awe_mem_mb,
+  innobase_additional_mem_pool_size;
 extern long innobase_file_io_threads, innobase_lock_wait_timeout;
 extern long innobase_force_recovery, innobase_thread_concurrency;
 extern char *innobase_data_home_dir, *innobase_data_file_path;

sql/mysqld.cc

@@ -3194,6 +3194,7 @@ enum options {
   OPT_INNODB_LOG_FILE_SIZE,
   OPT_INNODB_LOG_BUFFER_SIZE,
   OPT_INNODB_BUFFER_POOL_SIZE,
+  OPT_INNODB_BUFFER_POOL_AWE_MEM_MB,
   OPT_INNODB_ADDITIONAL_MEM_POOL_SIZE,
   OPT_INNODB_FILE_IO_THREADS,
   OPT_INNODB_LOCK_WAIT_TIMEOUT,
@@ -3753,6 +3754,10 @@ struct my_option my_long_options[] =
    "The size of the memory buffer InnoDB uses to cache data and indexes of its tables.",
    (gptr*) &innobase_buffer_pool_size, (gptr*) &innobase_buffer_pool_size, 0,
    GET_LONG, REQUIRED_ARG, 8*1024*1024L, 1024*1024L, ~0L, 0, 1024*1024L, 0},
+  {"innodb_buffer_pool_awe_mem_mb", OPT_INNODB_BUFFER_POOL_AWE_MEM_MB,
+   "If Windows AWE is used, the size of InnoDB buffer pool allocated from the AWE memory.",
+   (gptr*) &innobase_buffer_pool_awe_mem_mb, (gptr*) &innobase_buffer_pool_awe_mem_mb, 0,
+   GET_LONG, REQUIRED_ARG, 0, 0, 63000, 0, 1, 0},
   {"innodb_additional_mem_pool_size", OPT_INNODB_ADDITIONAL_MEM_POOL_SIZE,
    "Size of a memory pool InnoDB uses to store data dictionary information and other internal data structures.",
    (gptr*) &innobase_additional_mem_pool_size,

sql/set_var.cc

@@ -449,6 +449,7 @@ struct show_var_st init_vars[]= {
 #ifdef HAVE_INNOBASE_DB
   {"innodb_additional_mem_pool_size", (char*) &innobase_additional_mem_pool_size, SHOW_LONG },
   {"innodb_buffer_pool_size", (char*) &innobase_buffer_pool_size, SHOW_LONG },
+  {"innodb_buffer_pool_awe_mem_mb", (char*) &innobase_buffer_pool_awe_mem_mb, SHOW_LONG },
   {"innodb_data_file_path", (char*) &innobase_data_file_path,	    SHOW_CHAR_PTR},
   {"innodb_data_home_dir",  (char*) &innobase_data_home_dir,	    SHOW_CHAR_PTR},
   {"innodb_file_io_threads", (char*) &innobase_file_io_threads, SHOW_LONG },