btr0sea.c, buf0lru.c, buf0buf.c, ha0ha.c, hash0hash.h, ha0ha.h, buf0buf.h:

Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index memory overhead grows by 67 %, maybe we have to tune this later somehow


innobase/include/buf0buf.h:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/include/ha0ha.h:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/include/hash0hash.h:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/ha/ha0ha.c:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/buf/buf0buf.c:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/buf/buf0lru.c:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow
innobase/btr/btr0sea.c:
  Link adaptive hash index entries to the buffer page, so that we can remove them quickly without knowing the record structure on that page; this was requested by Marko for the compact InnoDB table format; note that the adaptive hash index overhead grows by 67 %, maybe we have to tune this later somehow

innobase/btr/btr0sea.c

@@ -915,17 +915,6 @@ btr_search_drop_page_hash_index(
 {
 	hash_table_t*	table;
 	buf_block_t*	block;
-	ulint		n_fields;
-	ulint		n_bytes;
-	rec_t*		rec;
-	rec_t*		sup;
-	ulint		fold;
-	ulint		prev_fold;
-	dulint		tree_id;
-	ulint		n_cached;
-	ulint		n_recs;
-	ulint*		folds;
-	ulint		i;
 
 #ifdef UNIV_SYNC_DEBUG
 	ut_ad(!rw_lock_own(&btr_search_latch, RW_LOCK_SHARED));
@@ -951,72 +940,17 @@ btr_search_drop_page_hash_index(
 	      			|| (block->buf_fix_count == 0));
 #endif /* UNIV_SYNC_DEBUG */
 
-	n_fields = block->curr_n_fields;
+	ut_a(block->curr_n_fields + block->curr_n_bytes > 0);
-	n_bytes = block->curr_n_bytes;
-
-	ut_a(n_fields + n_bytes > 0);
 
 	rw_lock_s_unlock(&btr_search_latch);
 	
-	n_recs = page_get_n_recs(page);
-
-	/* Calculate and cache fold values into an array for fast deletion
-	from the hash index */
-
-	folds = mem_alloc(n_recs * sizeof(ulint));
-
-	n_cached = 0;
-
-	sup = page_get_supremum_rec(page);
-
-	rec = page_get_infimum_rec(page);
-	rec = page_rec_get_next(rec);
-
-	if (rec != sup) {
-		ut_a(n_fields <= rec_get_n_fields(rec));
-
-		if (n_bytes > 0) {
-			ut_a(n_fields < rec_get_n_fields(rec));
-		}
-	}
-
-	tree_id = btr_page_get_index_id(page);
-	
-	prev_fold = 0;
-
-	while (rec != sup) {
-		/* FIXME: in a mixed tree, not all records may have enough
-		ordering fields: */
-		
-		fold = rec_fold(rec, n_fields, n_bytes, tree_id);
-
-		if (fold == prev_fold && prev_fold != 0) {
-
-			goto next_rec;
-		}
-
-		/* Remove all hash nodes pointing to this page from the
-		hash chain */
-
-		folds[n_cached] = fold;
-		n_cached++;
-next_rec:
-		rec = page_rec_get_next(rec);
-		prev_fold = fold;
-	}
-
 	rw_lock_x_lock(&btr_search_latch);
 
-	for (i = 0; i < n_cached; i++) {
+	ha_remove_all_nodes_to_page(table, page);
-
-		ha_remove_all_nodes_to_page(table, folds[i], page);
-	}
 
 	block->is_hashed = FALSE;
 
 	rw_lock_x_unlock(&btr_search_latch);
-
-	mem_free(folds);
 }
 
 /************************************************************************

innobase/buf/buf0buf.c

@@ -465,6 +465,7 @@ buf_block_init(
 	block->in_LRU_list = FALSE;
 
 	block->n_pointers = 0;
+	block->hash_nodes = NULL;
 
 	rw_lock_create(&(block->lock));
 	ut_ad(rw_lock_validate(&(block->lock)));

innobase/buf/buf0lru.c

@@ -789,6 +789,7 @@ buf_LRU_block_free_non_file_page(
 	      || (block->state == BUF_BLOCK_READY_FOR_USE));
 
 	ut_a(block->n_pointers == 0);
+	ut_a(block->hash_nodes == NULL);
 	ut_a(!block->in_free_list);
 
 	block->state = BUF_BLOCK_NOT_USED;

innobase/ha/ha0ha.c

@@ -65,10 +65,53 @@ ha_create(
 	return(table);
 }
 
+/*****************************************************************
+Removes an adaptive hash index node from the doubly linked list of hash nodes
+for the buffer block. */
+UNIV_INLINE
+void
+ha_remove_buf_block_node(
+/*=====================*/
+	buf_block_t*	block,	/* in: buffer block */
+	ha_node_t*	node)	/* in: an adaptive hash index node */
+{
+	if (node == block->hash_nodes) {
+		block->hash_nodes = node->next_for_block;
+	}
+
+	if (node->prev_for_block != NULL) {
+		(node->prev_for_block)->next_for_block = node->next_for_block;
+	}
+
+	if (node->next_for_block != NULL) {
+		(node->next_for_block)->prev_for_block = node->prev_for_block;
+	}
+}
+
+/*****************************************************************
+Adds an adaptive hash index node to the start of the doubly linked list of
+hash nodes for the buffer block. */
+UNIV_INLINE
+void
+ha_add_buf_block_node(
+/*==================*/
+	buf_block_t*	block,	/* in: buffer block */
+	ha_node_t*	node)	/* in: an adaptive hash index node */
+{
+	node->next_for_block = block->hash_nodes;
+	node->prev_for_block = NULL;
+
+	block->hash_nodes = node;
+
+	if (node->next_for_block != NULL) {
+		(node->next_for_block)->prev_for_block = node;
+	}
+}
+
 /*****************************************************************
 Inserts an entry into a hash table. If an entry with the same fold number
 is found, its node is updated to point to the new data, and no new node
-is inserted. */
+is inserted. This function is only used in the adaptive hash index. */
 
 ibool
 ha_insert_for_fold(
@@ -84,6 +127,7 @@ ha_insert_for_fold(
 {
 	hash_cell_t*	cell;
 	ha_node_t*	node;
+	buf_block_t*	block;
 	ha_node_t*	prev_node;
 	buf_block_t*	prev_block;
 	ulint		hash;
@@ -92,6 +136,9 @@ ha_insert_for_fold(
 #ifdef UNIV_SYNC_DEBUG
 	ut_ad(!table->mutexes || mutex_own(hash_get_mutex(table, fold)));
 #endif /* UNIV_SYNC_DEBUG */
+
+	block = buf_block_align(data);
+
 	hash = hash_calc_hash(fold, table);
 
 	cell = hash_get_nth_cell(table, hash);
@@ -104,7 +151,15 @@ ha_insert_for_fold(
 				prev_block = buf_block_align(prev_node->data);
 				ut_a(prev_block->n_pointers > 0);
 				prev_block->n_pointers--;
-				buf_block_align(data)->n_pointers++;
+
+				block->n_pointers++;
+				
+				if (prev_block != block) {
+					ha_remove_buf_block_node(prev_block,
+								prev_node);
+					ha_add_buf_block_node(block,
+								prev_node);
+				}
 			}
 
 			prev_node->data = data;
@@ -131,7 +186,9 @@ ha_insert_for_fold(
 	ha_node_set_data(node, data);
 
 	if (table->adaptive) {
-		buf_block_align(data)->n_pointers++;
+		block->n_pointers++;
+
+		ha_add_buf_block_node(block, node);
 	}
 
 	node->fold = fold;
@@ -166,9 +223,15 @@ ha_delete_hash_node(
 	hash_table_t*	table,		/* in: hash table */
 	ha_node_t*	del_node)	/* in: node to be deleted */
 {
+	buf_block_t*	block;
+
 	if (table->adaptive) {
-		ut_a(buf_block_align(del_node->data)->n_pointers > 0);
+		block = buf_block_align(del_node->data);
-		buf_block_align(del_node->data)->n_pointers--;
+
+		ut_a(block->n_pointers > 0);
+
+		block->n_pointers--;
+		ha_remove_buf_block_node(block, del_node);
 	}
 
 	HASH_DELETE_AND_COMPACT(ha_node_t, next, table, del_node);
@@ -209,6 +272,8 @@ ha_search_and_update_if_found(
 	void*		data,	/* in: pointer to the data */
 	void*		new_data)/* in: new pointer to the data */
 {
+	buf_block_t*	old_block;
+	buf_block_t*	block;
 	ha_node_t*	node;
 
 #ifdef UNIV_SYNC_DEBUG
@@ -220,8 +285,15 @@ ha_search_and_update_if_found(
 	if (node) {
 		if (table->adaptive) {
 			ut_a(buf_block_align(node->data)->n_pointers > 0);
-			buf_block_align(node->data)->n_pointers--;
+
-			buf_block_align(new_data)->n_pointers++;
+			old_block = buf_block_align(node->data);
+			ut_a(old_block->n_pointers > 0);
+			old_block->n_pointers--;
+			ha_remove_buf_block_node(old_block, node);
+
+			block = buf_block_align(new_data);
+			block->n_pointers++;
+			ha_add_buf_block_node(block, node);
 		}
 
 		node->data = new_data;
@@ -236,43 +308,25 @@ void
 ha_remove_all_nodes_to_page(
 /*========================*/
 	hash_table_t*	table,	/* in: hash table */
-	ulint		fold,	/* in: fold value */
 	page_t*		page)	/* in: buffer page */
 {
+	buf_block_t*	block;
 	ha_node_t*	node;
 
-#ifdef UNIV_SYNC_DEBUG
+	block = buf_block_align(page);
-	ut_ad(!table->mutexes || mutex_own(hash_get_mutex(table, fold)));
+
-#endif /* UNIV_SYNC_DEBUG */
+	node = block->hash_nodes;
-	node = ha_chain_get_first(table, fold);
 
 	while (node) {
-		if (buf_frame_align(ha_node_get_data(node)) == page) {
-
 		/* Remove the hash node */
 
 		ha_delete_hash_node(table, node);
 
-			/* Start again from the first node in the chain
+		node = block->hash_nodes;
-			because the deletion may compact the heap of
-			nodes and move other nodes! */
-
-			node = ha_chain_get_first(table, fold);
-		} else {
-			node = ha_chain_get_next(node);
 	}
-	}
-#ifdef UNIV_DEBUG
-	/* Check that all nodes really got deleted */
 	
-	node = ha_chain_get_first(table, fold);
+	ut_a(block->n_pointers == 0);
-
+	ut_a(block->hash_nodes == NULL);
-	while (node) {
-		ut_a(buf_frame_align(ha_node_get_data(node)) != page);
-
-		node = ha_chain_get_next(node);
-	}
-#endif
 }
 
 /*****************************************************************
@@ -352,6 +406,7 @@ ha_print_info(
 			n_bufs++;
 		}
 				
-		fprintf(file, ", node heap has %lu buffer(s)\n", (ulong) n_bufs);
+		fprintf(file, ", node heap has %lu buffer(s)\n",
+							(ulong) n_bufs);
 	}
 }	

innobase/include/buf0buf.h

@@ -29,6 +29,7 @@ Created 11/5/1995 Heikki Tuuri
 #include "buf0types.h"
 #include "sync0rw.h"
 #include "hash0hash.h"
+#include "ha0ha.h"
 #include "ut0byte.h"
 #include "os0proc.h"
 
@@ -817,7 +818,7 @@ struct buf_block_struct{
 					records with the same prefix should be
 					indexed in the hash index */
 					
-	/* The following 4 fields are protected by btr_search_latch: */
+	/* The following 6 fields are protected by btr_search_latch: */
 
 	ibool		is_hashed;	/* TRUE if hash index has already been
 					built on this page; note that it does
@@ -834,6 +835,11 @@ struct buf_block_struct{
 	ulint		curr_side;	/* BTR_SEARCH_LEFT_SIDE or
 					BTR_SEARCH_RIGHT_SIDE in hash
 					indexing */
+	ha_node_t* 	hash_nodes;	/* a doubly linked list of hash nodes
+					for this buffer block; this points to
+					the first node in the list if any;
+					note that we do not use UT_LST_ macros
+					to manipulate this list */
 	/* 6. Debug fields */
 #ifdef UNIV_SYNC_DEBUG
 	rw_lock_t	debug_latch;	/* in the debug version, each thread

innobase/include/ha0ha.h

@@ -54,7 +54,7 @@ ha_create(
 /*****************************************************************
 Inserts an entry into a hash table. If an entry with the same fold number
 is found, its node is updated to point to the new data, and no new node
-is inserted. */
+is inserted. This function is only used in the adaptive hash index. */
 
 ibool
 ha_insert_for_fold(
@@ -111,7 +111,6 @@ void
 ha_remove_all_nodes_to_page(
 /*========================*/
 	hash_table_t*	table,	/* in: hash table */
-	ulint		fold,	/* in: fold value */
 	page_t*		page);	/* in: buffer page */
 /*****************************************************************
 Validates a hash table. */
@@ -134,9 +133,18 @@ ha_print_info(
 
 typedef struct ha_node_struct ha_node_t;
 struct ha_node_struct {
-	ha_node_t* next; /* next chain node or NULL if none */
+	ha_node_t* 	next; 	/* next chain node; NULL if none */
 	void*		data;	/* pointer to the data */
 	ulint		fold;	/* fold value for the data */
+	ha_node_t* 	next_for_block;/* in an adaptive hash index
+				(btr0sea.c), a doubly linked list of hash
+				nodes for the buffer block; these nodes
+				contain pointers to index records on the
+				page; in the last node this field is NULL;
+				note that we do not use UT_LST_ macros
+				to manipulate this list */
+	ha_node_t* 	prev_for_block;/* pointer to the previous node; in the
+				first node NULL */
 };
 
 #ifndef UNIV_NONINL

innobase/include/hash0hash.h

@@ -166,7 +166,7 @@ hash_get_n_cells(
 /***********************************************************************
 Deletes a struct which is stored in the heap of the hash table, and compacts
 the heap. The fold value must be stored in the struct NODE in a field named
-'fold'. */
+'fold'. This macro is only used for the adaptive hash index. */
 
 #define HASH_DELETE_AND_COMPACT(TYPE, NAME, TABLE, NODE)\
 do {\
@@ -191,11 +191,23 @@ do {\
 		/* Copy the top node in place of NODE */\
 \
 		*(NODE) = *top_node111;\
+\
+		/* Update the adaptive hash list of the buffer block that\
+		corresponds to the top node */\
+		if (top_node111->next_for_block != NULL) {\
+			(top_node111->next_for_block)->prev_for_block = NODE;\
+		}\
+\
+		if (top_node111->prev_for_block != NULL) {\
+			(top_node111->prev_for_block)->next_for_block = NODE;\
+		} else {\
+			buf_block_align(top_node111->data)->hash_nodes = NODE;\
+		}\
+\
+		/* Look for the hash pointer to the top node, to update it */\
 \
 		cell111 = hash_get_nth_cell(TABLE,\
 				hash_calc_hash(top_node111->fold, TABLE));\
-\
-		/* Look for the pointer to the top node, to update it */\
 \
 		if (cell111->node == top_node111) {\
 			/* The top node is the first in the chain */\