branches/zip: In r988, the function buf_block_align() was enclosed

within UNIV_DEBUG. The two remaining callers in non-debug builds,
btr_search_guess_on_hash() and btr_search_validate(), were rewritten
to call buf_page_hash_get().

To implement support for a resizeable buffer pool, the function
buf_block_align() had been rewritten to perform a page hash lookup in
the buffer pool. The caller was also made responsible for holding the
buffer pool mutex.

Because the page hash lookup is expensive and it has to be done while
holding the buffer pool mutex, implement buf_block_align() by pointer
arithmetics again, and make btr_search_guess_on_hash() call it. Note
that this will have to be adjusted if the interface to the resizeable
buffer pool is actually implemented.

rb://83 approved by Heikki Tuuri, to address Issue #161.

As a deviation from the approved patch, this patch also makes
btr_search_validate() (invoked by CHECK TABLE) check that
buf_pool->page_hash is consistent with buf_block_align().

ChangeLog

+2009-01-26	The InnoDB Team
+
+	* include/buf0buf.h, include/buf0buf.ic, buf/buf0buf.c,
+	include/mtr0log.ic, include/row0upd.ic, mtr/mtr0mtr.c,
+	btr/btr0sea.c:
+	Implement buf_block_align() with pointer arithmetics, as it is in
+	the built-in InnoDB distributed with MySQL.  Do not acquire the
+	buffer pool mutex before buf_block_align().  This removes a
+	scalability bottleneck in the adaptive hash index lookup.  In
+	CHECK TABLE, check that buf_pool->page_hash is consistent with
+	buf_block_align().
+
 2009-01-23	The InnoDB Team
 
 	* btr/btr0sea.c:

btr/btr0sea.c

@@ -758,7 +758,6 @@ btr_search_guess_on_hash(
 {
 	buf_block_t*	block;
 	rec_t*		rec;
-	const page_t*	page;
 	ulint		fold;
 	dulint		index_id;
 #ifdef notdefined
@@ -809,29 +808,7 @@ btr_search_guess_on_hash(
 		goto failure_unlock;
 	}
 
-	page = page_align(rec);
-	{
-		ulint	page_no		= page_get_page_no(page);
-		ulint	space_id	= page_get_space_id(page);
-
-		buf_pool_mutex_enter();
-		block = (buf_block_t*) buf_page_hash_get(space_id, page_no);
-		buf_pool_mutex_exit();
-	}
-
-	if (UNIV_UNLIKELY(!block)
-	    || UNIV_UNLIKELY(buf_block_get_state(block)
-			     != BUF_BLOCK_FILE_PAGE)) {
-
-		/* The block is most probably being freed.
-		The function buf_LRU_search_and_free_block()
-		first removes the block from buf_pool->page_hash
-		by calling buf_LRU_block_remove_hashed_page().
-		After that, it invokes btr_search_drop_page_hash_index().
-		Let us pretend that the block was also removed from
-		the adaptive hash index. */
-		goto failure_unlock;
-	}
+	block = buf_block_align(rec);
 
 	if (UNIV_LIKELY(!has_search_latch)) {
 
@@ -848,8 +825,9 @@ btr_search_guess_on_hash(
 		buf_block_dbg_add_level(block, SYNC_TREE_NODE_FROM_HASH);
 	}
 
-	if (UNIV_UNLIKELY(buf_block_get_state(block)
-			  == BUF_BLOCK_REMOVE_HASH)) {
+	if (UNIV_UNLIKELY(buf_block_get_state(block) != BUF_BLOCK_FILE_PAGE)) {
+		ut_ad(buf_block_get_state(block) == BUF_BLOCK_REMOVE_HASH);
+
 		if (UNIV_LIKELY(!has_search_latch)) {
 
 			btr_leaf_page_release(block, latch_mode, mtr);
@@ -858,7 +836,6 @@ btr_search_guess_on_hash(
 		goto failure;
 	}
 
-	ut_ad(buf_block_get_state(block) == BUF_BLOCK_FILE_PAGE);
 	ut_ad(page_rec_is_user_rec(rec));
 
 	btr_cur_position(index, rec, block, cursor);
@@ -870,8 +847,8 @@ btr_search_guess_on_hash(
 	is positioned on. We cannot look at the next of the previous
 	record to determine if our guess for the cursor position is
 	right. */
-	if (UNIV_EXPECT(
-		    ut_dulint_cmp(index_id, btr_page_get_index_id(page)), 0)
+	if (UNIV_EXPECT
+	    (ut_dulint_cmp(index_id, btr_page_get_index_id(block->frame)), 0)
 	    || !btr_search_check_guess(cursor,
 				       has_search_latch,
 				       tuple, mode, mtr)) {
@@ -1690,7 +1667,6 @@ btr_search_validate(void)
 /*=====================*/
 				/* out: TRUE if ok */
 {
-	page_t*		page;
 	ha_node_t*	node;
 	ulint		n_page_dumps	= 0;
 	ibool		ok		= TRUE;
@@ -1725,28 +1701,30 @@ btr_search_validate(void)
 		node = hash_get_nth_cell(btr_search_sys->hash_index, i)->node;
 
 		for (; node != NULL; node = node->next) {
-			const buf_block_t*	block;
-
-			page = page_align(node->data);
-			{
-				ulint	page_no	= page_get_page_no(page);
-				ulint	space_id= page_get_space_id(page);
-
-				block = buf_block_hash_get(space_id, page_no);
-			}
-
-			if (UNIV_UNLIKELY(!block)) {
-
-				/* The block is most probably being freed.
-				The function buf_LRU_search_and_free_block()
-				first removes the block from
+			const buf_block_t*	block
+				= buf_block_align(node->data);
+			const buf_block_t*	hash_block
+				= buf_block_hash_get(
+					buf_block_get_space(block),
+					buf_block_get_page_no(block));
+			const page_t*		page
+				= buf_block_get_frame(block);
+
+			if (hash_block) {
+				ut_a(hash_block == block);
+			} else {
+				/* When a block is being freed,
+				buf_LRU_search_and_free_block() first
+				removes the block from
 				buf_pool->page_hash by calling
 				buf_LRU_block_remove_hashed_page().
 				After that, it invokes
-				btr_search_drop_page_hash_index().
-				Let us pretend that the block was also removed
-				from the adaptive hash index. */
-				continue;
+				btr_search_drop_page_hash_index() to
+				remove the block from
+				btr_search_sys->hash_index. */
+
+				ut_a(buf_block_get_state(block)
+				     == BUF_BLOCK_REMOVE_HASH);
 			}
 
 			ut_a(!dict_index_is_ibuf(block->index));

buf/buf0buf.c

@@ -1750,6 +1750,53 @@ buf_zip_decompress(
 	return(FALSE);
 }
 
+/***********************************************************************
+Gets the block to whose frame the pointer is pointing to. */
+UNIV_INTERN
+buf_block_t*
+buf_block_align(
+/*============*/
+				/* out: pointer to block, never NULL */
+	const byte*	ptr)	/* in: pointer to a frame */
+{
+	buf_chunk_t*	chunk;
+	ulint		i;
+
+	/* TODO: protect buf_pool->chunks with a mutex (it will
+	currently remain constant after buf_pool_init()) */
+	for (chunk = buf_pool->chunks, i = buf_pool->n_chunks; i--; chunk++) {
+		lint	offs = ptr - chunk->blocks->frame;
+
+		if (UNIV_UNLIKELY(offs < 0)) {
+
+			continue;
+		}
+
+		offs >>= UNIV_PAGE_SIZE_SHIFT;
+
+		if (UNIV_LIKELY((ulint) offs < chunk->size)) {
+			buf_block_t*	block = &chunk->blocks[offs];
+
+			/* The function buf_chunk_init() invokes
+			buf_block_init() so that block[n].frame ==
+			block->frame + n * UNIV_PAGE_SIZE.  Check it. */
+			ut_ad(block->frame == page_align(ptr));
+			/* The space id and page number should be
+			stamped on the page. */
+			ut_ad(block->page.space
+			      == page_get_space_id(page_align(ptr)));
+			ut_ad(block->page.offset
+			      == page_get_page_no(page_align(ptr)));
+
+			return(block);
+		}
+	}
+
+	/* The block should always be found. */
+	ut_error;
+	return(NULL);
+}
+
 /************************************************************************
 Find out if a buffer block was created by buf_chunk_init(). */
 static

include/buf0buf.h

@@ -858,15 +858,15 @@ Gets the compressed page descriptor corresponding to an uncompressed page
 if applicable. */
 #define buf_block_get_page_zip(block) \
 	(UNIV_LIKELY_NULL((block)->page.zip.data) ? &(block)->page.zip : NULL)
-#if defined UNIV_DEBUG || defined UNIV_ZIP_DEBUG
 /***********************************************************************
 Gets the block to whose frame the pointer is pointing to. */
-UNIV_INLINE
-const buf_block_t*
+UNIV_INTERN
+buf_block_t*
 buf_block_align(
 /*============*/
-				/* out: pointer to block */
+				/* out: pointer to block, never NULL */
 	const byte*	ptr);	/* in: pointer to a frame */
+#if defined UNIV_DEBUG || defined UNIV_ZIP_DEBUG
 /*************************************************************************
 Gets the compressed page descriptor corresponding to an uncompressed page
 if applicable. */

include/buf0buf.ic

@@ -631,29 +631,6 @@ buf_block_get_zip_size(
 }
 
 #if defined UNIV_DEBUG || defined UNIV_ZIP_DEBUG
-/***********************************************************************
-Gets the block to whose frame the pointer is pointing to. */
-UNIV_INLINE
-const buf_block_t*
-buf_block_align(
-/*============*/
-				/* out: pointer to block */
-	const byte*	ptr)	/* in: pointer to a frame */
-{
-	const buf_block_t*	block;
-	ulint			space_id, page_no;
-
-	ptr = (const byte*) ut_align_down(ptr, UNIV_PAGE_SIZE);
-	page_no = mach_read_from_4(ptr + FIL_PAGE_OFFSET);
-	space_id = mach_read_from_4(ptr + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID);
-
-	block = (const buf_block_t*) buf_page_hash_get(space_id, page_no);
-	ut_ad(block);
-	ut_ad(buf_block_get_state(block) == BUF_BLOCK_FILE_PAGE);
-	ut_ad(block->frame == ptr);
-	return(block);
-}
-
 /*************************************************************************
 Gets the compressed page descriptor corresponding to an uncompressed page
 if applicable. */
@@ -664,11 +641,7 @@ buf_frame_get_page_zip(
 				/* out: compressed page descriptor, or NULL */
 	const byte*	ptr)	/* in: pointer to the page */
 {
-	const page_zip_des_t*	page_zip;
-	buf_pool_mutex_enter();
-	page_zip = buf_block_get_page_zip(buf_block_align(ptr));
-	buf_pool_mutex_exit();
-	return(page_zip);
+	return(buf_block_get_page_zip(buf_block_align(ptr)));
 }
 #endif /* UNIV_DEBUG || UNIV_ZIP_DEBUG */
 

include/mtr0log.ic

@@ -191,10 +191,8 @@ mlog_write_initial_log_record_fast(
 #endif
 
 #ifdef UNIV_DEBUG
-	buf_pool_mutex_enter();
 	/* We now assume that all x-latched pages have been modified! */
 	block = (buf_block_t*) buf_block_align(ptr);
-	buf_pool_mutex_exit();
 
 	if (!mtr_memo_contains(mtr, block, MTR_MEMO_MODIFY)) {
 

include/row0upd.ic

@@ -139,9 +139,7 @@ row_upd_rec_sys_fields(
 	ut_ad(rec_offs_validate(rec, index, offsets));
 #ifdef UNIV_SYNC_DEBUG
 	if (!rw_lock_own(&btr_search_latch, RW_LOCK_EX)) {
-		buf_pool_mutex_enter();
 		ut_ad(!buf_block_align(rec)->is_hashed);
-		buf_pool_mutex_exit();
 	}
 #endif /* UNIV_SYNC_DEBUG */
 

mtr/mtr0mtr.c

@@ -309,12 +309,7 @@ mtr_memo_contains_page(
 	const byte*	ptr,	/* in: pointer to buffer frame */
 	ulint		type)	/* in: type of object */
 {
-	ibool	ret;
-
-	buf_pool_mutex_enter();
-	ret = mtr_memo_contains(mtr, buf_block_align(ptr), type);
-	buf_pool_mutex_exit();
-	return(ret);
+	return(mtr_memo_contains(mtr, buf_block_align(ptr), type));
 }
 
 /*************************************************************