branches/innodb+: Delete buffer port from branches/fts:r2283

btr/btr0btr.c

@@ -558,6 +558,7 @@ btr_page_get_father_node_ptr(
 				its page x-latched */
 	mtr_t*		mtr)	/* in: mtr */
 {
+	page_t*		page;
 	dtuple_t*	tuple;
 	rec_t*		user_rec;
 	rec_t*		node_ptr;
@@ -574,7 +575,19 @@ btr_page_get_father_node_ptr(
 	ut_ad(dict_index_get_page(index) != page_no);
 
 	level = btr_page_get_level(btr_cur_get_page(cursor), mtr);
-	user_rec = btr_cur_get_rec(cursor);
+
+	page = btr_cur_get_page(cursor);
+
+	if (UNIV_UNLIKELY(page_get_n_recs(page) == 0)) {
+		/* Empty pages can result from buffered delete operations.
+		The first record from the free list can be used to find the
+		father node. */
+		user_rec = page_header_get_ptr(page, PAGE_FREE);
+		ut_a(user_rec);
+	} else {
+		user_rec = btr_cur_get_rec(cursor);
+	}
+
 	ut_a(page_rec_is_user_rec(user_rec));
 	tuple = dict_index_build_node_ptr(index, user_rec, 0, heap, level);
 

buf/buf0buddy.c

@@ -213,6 +213,7 @@ buf_buddy_block_register(
 	buf_block_t*	block)	/* in: buffer frame to allocate */
 {
 	const ulint	fold = BUF_POOL_ZIP_FOLD(block);
+
 	ut_ad(buf_pool_mutex_own());
 	ut_ad(!mutex_own(&buf_pool_zip_mutex));
 
@@ -224,6 +225,7 @@ buf_buddy_block_register(
 	ut_ad(!block->page.in_page_hash);
 	ut_ad(!block->page.in_zip_hash);
 	ut_d(block->page.in_zip_hash = TRUE);
+
 	HASH_INSERT(buf_page_t, hash, buf_pool->zip_hash, fold, &block->page);
 
 	buf_buddy_n_frames++;
@@ -278,23 +280,21 @@ buf_buddy_alloc_clean(
 			TRUE if storage was allocated from the LRU list
 			and buf_pool_mutex was temporarily released */
 {
+	ulint		count;
 	buf_page_t*	bpage;
 
 	ut_ad(buf_pool_mutex_own());
 	ut_ad(!mutex_own(&buf_pool_zip_mutex));
 
-	if (buf_buddy_n_frames < buf_buddy_max_n_frames) {
-
-		goto free_LRU;
-	}
+	if (buf_buddy_n_frames >= buf_buddy_max_n_frames
+	    && ((BUF_BUDDY_LOW << i) >= PAGE_ZIP_MIN_SIZE
+		&& i < BUF_BUDDY_SIZES)) {
 
-	if (BUF_BUDDY_LOW << i >= PAGE_ZIP_MIN_SIZE
-	    && i < BUF_BUDDY_SIZES) {
 		/* Try to find a clean compressed-only page
 		of the same size. */
 
-		page_zip_des_t	dummy_zip;
 		ulint		j;
+		page_zip_des_t	dummy_zip;
 
 		page_zip_set_size(&dummy_zip, BUF_BUDDY_LOW << i);
 
@@ -335,9 +335,12 @@ buf_buddy_alloc_clean(
 	/* Free blocks from the end of the LRU list until enough space
 	is available. */
 
+	count = 0;
+
 free_LRU:
-	for (bpage = UT_LIST_GET_LAST(buf_pool->LRU); bpage;
-	     bpage = UT_LIST_GET_PREV(LRU, bpage)) {
+	for (bpage = UT_LIST_GET_LAST(buf_pool->LRU);
+	     bpage;
+	     bpage = UT_LIST_GET_PREV(LRU, bpage), ++count) {
 
 		void*		ret;
 		mutex_t*	block_mutex = buf_page_get_mutex(bpage);
@@ -440,20 +443,19 @@ buf_buddy_alloc_low(
 	}
 
 	/* Try replacing a clean page in the buffer pool. */
-
 	block = buf_buddy_alloc_clean(i, lru);
 
 	if (block) {
 
 		goto func_exit;
 	}
-
 	/* Try replacing an uncompressed page in the buffer pool. */
 	buf_pool_mutex_exit();
 	block = buf_LRU_get_free_block(0);
 	*lru = TRUE;
 	buf_pool_mutex_enter();
 
+
 alloc_big:
 	buf_buddy_block_register(block);
 

buf/buf0buf.c

@@ -1346,6 +1346,69 @@ buf_pool_resize(void)
 	buf_pool_page_hash_rebuild();
 }
 
+/********************************************************************
+Add watch for the given page to be read in. Caller must have the buffer pool
+mutex reserved. */
+static
+void
+buf_pool_add_watch(
+/*===============*/
+	ulint	space,		/* in: space id */
+	ulint	page_no)	/* in: page number */
+{
+	ut_ad(mutex_own(&buf_pool_mutex));
+
+	/* There can't be multiple watches at the same time. */
+	ut_a(!buf_pool->watch_active);
+
+	buf_pool->watch_active = TRUE;
+	buf_pool->watch_space = space;
+	buf_pool->watch_happened = FALSE;
+	buf_pool->watch_page_no = page_no;
+}
+
+/********************************************************************
+Stop watching if the marked page is read in. */
+UNIV_INTERN
+void
+buf_pool_remove_watch(void)
+/*=======================*/
+{
+	buf_pool_mutex_enter();
+
+	ut_ad(buf_pool->watch_active);
+
+	buf_pool->watch_active = FALSE;
+
+	buf_pool_mutex_exit();
+}
+
+/********************************************************************
+Check if the given page is being watched and has been read to the buffer
+pool. */
+UNIV_INTERN
+ibool
+buf_pool_watch_happened(
+/*====================*/
+				/* out: TRUE if the given page is being
+				watched and it has been read in */
+	ulint	space,		/* in: space id */
+	ulint	page_no)	/* in: page number */
+{
+	ulint	ret;
+
+	buf_pool_mutex_enter();
+
+	ret = buf_pool->watch_active
+		&& space == buf_pool->watch_space
+		&& page_no == buf_pool->watch_page_no
+		&& buf_pool->watch_happened;
+
+	buf_pool_mutex_exit();
+
+	return(ret);
+}
+
 /************************************************************************
 Moves to the block to the start of the LRU list if there is a danger
 that the block would drift out of the buffer pool. */
@@ -1763,7 +1826,8 @@ buf_page_get_gen(
 	ulint		rw_latch,/* in: RW_S_LATCH, RW_X_LATCH, RW_NO_LATCH */
 	buf_block_t*	guess,	/* in: guessed block or NULL */
 	ulint		mode,	/* in: BUF_GET, BUF_GET_IF_IN_POOL,
-				BUF_GET_NO_LATCH, BUF_GET_NOWAIT */
+				BUF_GET_NO_LATCH, BUF_GET_NOWAIT or
+				BUF_GET_IF_IN_POOL_OR_WATCH*/
 	const char*	file,	/* in: file name */
 	ulint		line,	/* in: line where called */
 	mtr_t*		mtr)	/* in: mini-transaction */
@@ -1778,11 +1842,17 @@ buf_page_get_gen(
 	      || (rw_latch == RW_X_LATCH)
 	      || (rw_latch == RW_NO_LATCH));
 	ut_ad((mode != BUF_GET_NO_LATCH) || (rw_latch == RW_NO_LATCH));
-	ut_ad((mode == BUF_GET) || (mode == BUF_GET_IF_IN_POOL)
-	      || (mode == BUF_GET_NO_LATCH) || (mode == BUF_GET_NOWAIT));
+
+	/* Check for acceptable modes. */
+	ut_ad(mode == BUF_GET
+	      || mode == BUF_GET_IF_IN_POOL
+	      || mode == BUF_GET_NO_LATCH
+	      || mode == BUF_GET_NOWAIT
+	      || mode == BUF_GET_IF_IN_POOL_OR_WATCH);
+
 	ut_ad(zip_size == fil_space_get_zip_size(space));
 #ifndef UNIV_LOG_DEBUG
-	ut_ad(!ibuf_inside() || ibuf_page(space, zip_size, offset));
+	ut_ad(!ibuf_inside() || ibuf_page(space, zip_size, offset, mtr));
 #endif
 	buf_pool->n_page_gets++;
 loop:
@@ -1818,9 +1888,14 @@ buf_page_get_gen(
 	if (block == NULL) {
 		/* Page not in buf_pool: needs to be read from file */
 
+		if (mode == BUF_GET_IF_IN_POOL_OR_WATCH) {
+			buf_pool_add_watch(space, offset);
+		}
+
 		buf_pool_mutex_exit();
 
-		if (mode == BUF_GET_IF_IN_POOL) {
+		if (mode == BUF_GET_IF_IN_POOL
+		    || mode == BUF_GET_IF_IN_POOL_OR_WATCH) {
 
 			return(NULL);
 		}
@@ -1837,7 +1912,18 @@ buf_page_get_gen(
 
 	must_read = buf_block_get_io_fix(block) == BUF_IO_READ;
 
-	if (must_read && mode == BUF_GET_IF_IN_POOL) {
+	if (must_read
+	    && (mode == BUF_GET_IF_IN_POOL
+		|| mode == BUF_GET_IF_IN_POOL_OR_WATCH)) {
+
+		/* The page is being read to bufer pool,
+		but we can't wait around for the read to
+		complete. */
+
+		if (mode == BUF_GET_IF_IN_POOL_OR_WATCH) {
+			buf_pool_add_watch(space, offset);
+		}
+
 		/* The page is only being read to buffer */
 		buf_pool_mutex_exit();
 
@@ -2140,7 +2226,7 @@ buf_page_optimistic_get_func(
 	ut_ad(!ibuf_inside()
 	      || ibuf_page(buf_block_get_space(block),
 			   buf_block_get_zip_size(block),
-			   buf_block_get_page_no(block)));
+			   buf_block_get_page_no(block), mtr));
 
 	if (rw_latch == RW_S_LATCH) {
 		success = rw_lock_s_lock_func_nowait(&(block->lock),
@@ -2392,6 +2478,25 @@ buf_page_init_low(
 #endif /* UNIV_DEBUG_FILE_ACCESSES */
 }
 
+/************************************************************************
+Set watch happened flag. */
+UNIV_INLINE
+void
+buf_page_notify_watch(
+/*==================*/
+	ulint		space,	/* in: space id of page read in */
+	ulint		offset)	/* in: offset of page read in */
+{
+	ut_ad(buf_pool_mutex_own());
+
+	if (buf_pool->watch_active
+	    && space == buf_pool->watch_space
+	    && offset == buf_pool->watch_page_no) {
+
+		buf_pool->watch_happened = TRUE;
+	}
+}
+
 #ifdef UNIV_HOTBACKUP
 /************************************************************************
 Inits a page to the buffer buf_pool, for use in ibbackup --restore. */
@@ -2481,6 +2586,7 @@ buf_page_init(
 	}
 
 	buf_page_init_low(&block->page);
+	buf_page_notify_watch(space, offset);
 
 	ut_ad(!block->page.in_zip_hash);
 	ut_ad(!block->page.in_page_hash);
@@ -2531,7 +2637,8 @@ buf_page_init_for_read(
 
 		mtr_start(&mtr);
 
-		if (!ibuf_page_low(space, zip_size, offset, &mtr)) {
+		if (!recv_no_ibuf_operations
+		    && !ibuf_page(space, zip_size, offset, &mtr)) {
 
 			mtr_commit(&mtr);
 
@@ -2583,7 +2690,9 @@ buf_page_init_for_read(
 	if (block) {
 		bpage = &block->page;
 		mutex_enter(&block->mutex);
+
 		buf_page_init(space, offset, block);
+		buf_page_notify_watch(space, offset);
 
 		/* The block must be put to the LRU list, to the old blocks */
 		buf_LRU_add_block(bpage, TRUE/* to old blocks */);
@@ -2650,11 +2759,15 @@ buf_page_init_for_read(
 		mutex_enter(&buf_pool_zip_mutex);
 		UNIV_MEM_DESC(bpage->zip.data,
 			      page_zip_get_size(&bpage->zip), bpage);
+
 		buf_page_init_low(bpage);
+		buf_page_notify_watch(space, offset);
+
 		bpage->state	= BUF_BLOCK_ZIP_PAGE;
 		bpage->space	= space;
 		bpage->offset	= offset;
 
+
 #ifdef UNIV_DEBUG
 		bpage->in_page_hash = FALSE;
 		bpage->in_zip_hash = FALSE;
@@ -2748,6 +2861,7 @@ buf_page_create(
 	mutex_enter(&block->mutex);
 
 	buf_page_init(space, offset, block);
+	buf_page_notify_watch(space, offset);
 
 	/* The block must be put to the LRU list */
 	buf_LRU_add_block(&block->page, FALSE);
@@ -3539,7 +3653,7 @@ buf_print_io(
 
 	fprintf(file,
 		"Buffer pool size   %lu\n"
-		"Free buffers       %lu\n"
+		"Free buffers	   %lu\n"
 		"Database pages     %lu\n"
 		"Modified db pages  %lu\n"
 		"Pending reads %lu\n"

buf/buf0lru.c

@@ -497,7 +497,7 @@ buf_LRU_get_free_block(
 
 		if (!buf_lru_switched_on_innodb_mon) {
 
-	   		/* Over 67 % of the buffer pool is occupied by lock
+			/* Over 67 % of the buffer pool is occupied by lock
 			heaps or the adaptive hash index. This may be a memory
 			leak! */
 

fil/fil0fil.c

@@ -191,8 +191,6 @@ struct fil_space_struct {
 				currently in the list above */
 	UT_LIST_NODE_T(fil_space_t) space_list;
 				/* list of all spaces */
-	ibuf_data_t*	ibuf_data;
-				/* insert buffer data */
 	ulint		magic_n;
 };
 
@@ -476,33 +474,6 @@ fil_space_get_type(
 	return(space->purpose);
 }
 
-/***********************************************************************
-Returns the ibuf data of a file space. */
-UNIV_INTERN
-ibuf_data_t*
-fil_space_get_ibuf_data(
-/*====================*/
-			/* out: ibuf data for this space */
-	ulint	id)	/* in: space id */
-{
-	fil_system_t*	system		= fil_system;
-	fil_space_t*	space;
-
-	ut_ad(system);
-
-	ut_a(id == 0);
-
-	mutex_enter(&(system->mutex));
-
-	space = fil_space_get_by_id(id);
-
-	mutex_exit(&(system->mutex));
-
-	ut_a(space);
-
-	return(space->ibuf_data);
-}
-
 /**************************************************************************
 Checks if all the file nodes in a space are flushed. The caller must hold
 the fil_system mutex. */
@@ -1183,8 +1154,6 @@ fil_space_create(
 	UT_LIST_INIT(space->chain);
 	space->magic_n = FIL_SPACE_MAGIC_N;
 
-	space->ibuf_data = NULL;
-
 	rw_lock_create(&space->latch, SYNC_FSP);
 
 	HASH_INSERT(fil_space_t, hash, system->spaces, id, space);
@@ -1649,25 +1618,6 @@ fil_set_max_space_id_if_bigger(
 	mutex_exit(&(system->mutex));
 }
 
-/********************************************************************
-Initializes the ibuf data structure for space 0 == the system tablespace.
-This can be called after the file space headers have been created and the
-dictionary system has been initialized. */
-UNIV_INTERN
-void
-fil_ibuf_init_at_db_start(void)
-/*===========================*/
-{
-	fil_space_t*	space;
-
-	space = UT_LIST_GET_FIRST(fil_system->space_list);
-
-	ut_a(space);
-	ut_a(space->purpose == FIL_TABLESPACE);
-
-	space->ibuf_data = ibuf_data_init_for_space(space->id);
-}
-
 /********************************************************************
 Writes the flushed lsn and the latest archived log number to the page header
 of the first page of a data file of the system tablespace (space 0),
@@ -4266,13 +4216,13 @@ fil_io(
 	      || sync || is_log);
 #ifdef UNIV_SYNC_DEBUG
 	ut_ad(!ibuf_inside() || is_log || (type == OS_FILE_WRITE)
-	      || ibuf_page(space_id, zip_size, block_offset));
+	      || ibuf_page(space_id, zip_size, block_offset, NULL));
 #endif
 #endif
 	if (sync) {
 		mode = OS_AIO_SYNC;
 	} else if (type == OS_FILE_READ && !is_log
-		   && ibuf_page(space_id, zip_size, block_offset)) {
+		   && ibuf_page(space_id, zip_size, block_offset, NULL)) {
 		mode = OS_AIO_IBUF;
 	} else if (is_log) {
 		mode = OS_AIO_LOG;

fsp/fsp0fsp.c

@@ -2191,8 +2191,8 @@ fseg_create_general(
 		/* This thread did not own the latch before this call: free
 		excess pages from the insert buffer free list */
 
-		if (space == 0) {
-			ibuf_free_excess_pages(0);
+		if (space == IBUF_SPACE_ID) {
+			ibuf_free_excess_pages();
 		}
 	}
 
@@ -2759,8 +2759,8 @@ fseg_alloc_free_page_general(
 		/* This thread did not own the latch before this call: free
 		excess pages from the insert buffer free list */
 
-		if (space == 0) {
-			ibuf_free_excess_pages(0);
+		if (space == IBUF_SPACE_ID) {
+			ibuf_free_excess_pages();
 		}
 	}
 

include/btr0btr.h

@@ -42,6 +42,8 @@ failure. */
 #define	BTR_SEARCH_PREV		35
 #define	BTR_MODIFY_PREV		36
 
+/* BTR_INSERT, BTR_DELETE and BTR_DELETE_MARK are mutually exclusive. */
+
 /* If this is ORed to the latch mode, it means that the search tuple will be
 inserted to the index, at the searched position */
 #define BTR_INSERT		512
@@ -55,6 +57,19 @@ UNIQUE definition on secondary indexes when we decide if we can use the
 insert buffer to speed up inserts */
 #define BTR_IGNORE_SEC_UNIQUE	2048
 
+/* Try to delete mark the record at the searched position using the
+insert/delete buffer. */
+#define BTR_DELETE_MARK		4096
+
+/* Try to delete the record at the searched position using the insert/delete
+buffer. */
+#define BTR_DELETE		8192
+
+/* If the leaf page is not in the buffer pool: don't read it in, set
+cursor->leaf_in_buf_pool to FALSE, and set buf_pool_t::watch_* that
+watches for the page to get read in. */
+#define BTR_WATCH_LEAF		16384
+
 /******************************************************************
 Gets the root node of a tree and x-latches it. */
 UNIV_INTERN

include/btr0btr.ic

@@ -118,7 +118,7 @@ btr_page_get_level(
 /*===============*/
 				/* out: level, leaf level == 0 */
 	const page_t*	page,	/* in: index page */
-	mtr_t*		mtr __attribute__((unused)))
+	mtr_t*		mtr UNIV_UNUSED)
 				/* in: mini-transaction handle */
 {
 	ut_ad(page && mtr);
@@ -160,7 +160,7 @@ btr_page_get_next(
 /*==============*/
 				/* out: next page number */
 	const page_t*	page,	/* in: index page */
-	mtr_t*		mtr __attribute__((unused)))
+	mtr_t*		mtr UNIV_UNUSED)
 				/* in: mini-transaction handle */
 {
 	ut_ad(page && mtr);
@@ -200,7 +200,7 @@ btr_page_get_prev(
 /*==============*/
 				/* out: prev page number */
 	const page_t*	page,	/* in: index page */
-	mtr_t*	mtr __attribute__((unused))) /* in: mini-transaction handle */
+	mtr_t*	mtr     UNIV_UNUSED) /* in: mini-transaction handle */
 {
 	ut_ad(page && mtr);
 

include/btr0cur.h

@@ -312,8 +312,8 @@ btr_cur_del_mark_set_sec_rec(
 	que_thr_t*	thr,	/* in: query thread */
 	mtr_t*		mtr);	/* in: mtr */
 /***************************************************************
-Sets a secondary index record delete mark to FALSE. This function is
-only used by the insert buffer insert merge mechanism. */
+Sets a secondary index record delete mark to the given value. This
+function is only used by the insert buffer insert merge mechanism. */
 UNIV_INTERN
 void
 btr_cur_del_unmark_for_ibuf(
@@ -323,6 +323,7 @@ btr_cur_del_unmark_for_ibuf(
 					corresponding to rec, or NULL
 					when the tablespace is
 					uncompressed */
+	ibool		val,		/* value to set */
 	mtr_t*		mtr);		/* in: mtr */
 /*****************************************************************
 Tries to compress a page of the tree if it seems useful. It is assumed
@@ -572,7 +573,20 @@ btr_push_update_extern_fields(
 	const upd_t*	update,	/* in: update vector */
 	mem_heap_t*	heap)	/* in: memory heap */
 	__attribute__((nonnull));
+/***************************************************************
+Sets a secondary index record's delete mark to the given value. This
+function is only used by the insert buffer merge mechanism. */
 
+void
+btr_cur_set_deleted_flag_for_ibuf(
+/*==============================*/
+	rec_t*		rec,		/* in: record */
+	page_zip_des_t*	page_zip,	/* in/out: compressed page
+					corresponding to rec, or NULL
+					when the tablespace is
+					uncompressed */
+	ibool		val,		/* in: value to set */
+	mtr_t*		mtr);		/* in: mtr */
 /*######################################################################*/
 
 /* In the pessimistic delete, if the page data size drops below this
@@ -657,6 +671,28 @@ struct btr_cur_struct {
 					NULL */
 	ulint		fold;		/* fold value used in the search if
 					flag is BTR_CUR_HASH */
+	/*----- Delete buffering -------*/
+	ulint		ibuf_cnt;	/* in searches done on insert buffer
+					trees, this contains the "counter"
+					value (the first two bytes of the
+					fourth field) extracted from the
+					page above the leaf page, from the
+					father node pointer that pointed to
+					the leaf page. in other words, it
+					contains the minimum counter value
+					for records to be inserted on the
+					chosen leaf page. If for some reason
+					this can't be read, or if the search
+					ended on the leftmost leaf page in
+					the tree (in which case the father
+					node pointer had the 'minimum
+					record' flag set), this is
+					ULINT_UNDEFINED. */
+	ibool		leaf_in_buf_pool;
+					/* in: in searches done with
+					BTR_CHECK_LEAF, this is TRUE if the
+					leaf page is in the buffer pool,
+					FALSE otherwise. */
 	/*------------------------------*/
 	btr_path_t*	path_arr;	/* in estimating the number of
 					rows in range, we store in this array
@@ -675,6 +711,13 @@ struct btr_cur_struct {
 #define BTR_CUR_BINARY		3	/* success using the binary search */
 #define BTR_CUR_INSERT_TO_IBUF	4	/* performed the intended insert to
 					the insert buffer */
+#define BTR_CUR_DEL_MARK_IBUF	5	/* performed the intended delete
+					mark in the insert/delete buffer */
+#define BTR_CUR_DELETE_IBUF	6	/* performed the intended delete in
+					the insert/delete buffer */
+#define BTR_CUR_ABORTED		7	/* search with BTR_CHECK_LEAF
+					aborted due to leaf page not being
+					in buffer pool */
 
 /* If pessimistic delete fails because of lack of file space,
 there is still a good change of success a little later: try this many times,

include/btr0pcur.h

@@ -79,6 +79,16 @@ btr_pcur_open(
 	btr_pcur_t*	cursor, /* in: memory buffer for persistent cursor */
 	mtr_t*		mtr);	/* in: mtr */
 /******************************************************************
+Check if an operation was buffered. */
+UNIV_INLINE
+ibool
+btr_pcur_was_buffered(
+/*==================*/
+				/* out: TRUE if the operation was buffered
+				in the insert/delete buffer */
+	const btr_pcur_t*	cursor);
+				/* in: persistent cursor */
+/******************************************************************
 Opens an persistent cursor to an index tree without initializing the
 cursor. */
 UNIV_INLINE

include/btr0pcur.ic

@@ -506,6 +506,28 @@ btr_pcur_open(
 	cursor->trx_if_known = NULL;
 }
 
+/******************************************************************
+Check if an operation was buffered. */
+UNIV_INLINE
+ibool
+btr_pcur_was_buffered(
+/*==================*/
+				/* out: TRUE if the operation was buffered
+				in the insert/delete buffer */
+	const btr_pcur_t*	cursor)
+				/* in: persistent cursor */
+{
+	const btr_cur_t*	btr_cursor;
+
+	/* Look in the tree cursor */
+
+	btr_cursor = btr_pcur_get_btr_cur(cursor);
+
+	return((btr_cursor->flag == BTR_CUR_DEL_MARK_IBUF)
+	       || (btr_cursor->flag == BTR_CUR_DELETE_IBUF)
+	       || (btr_cursor->flag == BTR_CUR_INSERT_TO_IBUF));
+}
+
 /******************************************************************
 Opens an persistent cursor to an index tree without initializing the
 cursor. */

include/buf0buf.h

@@ -43,6 +43,10 @@ Created 11/5/1995 Heikki Tuuri
 					it is error-prone programming not to
 					set a latch, and it should be used
 					with care */
+#define BUF_GET_IF_IN_POOL_OR_WATCH	15
+					/* Get the page only if it's in the
+					buffer pool, if not then set a watch
+					on the page. */
 /* Modes for buf_page_get_known_nowait */
 #define BUF_MAKE_YOUNG	51
 #define BUF_KEEP_OLD	52
@@ -165,20 +169,22 @@ read the contents of the page unless you know it is safe. Do not modify
 the contents of the page! We have separated this case, because it is
 error-prone programming not to set a latch, and it should be used
 with care. */
-#define buf_page_get_with_no_latch(SP, ZS, OF, MTR)	   buf_page_get_gen(\
+#define buf_page_get_with_no_latch(SP, ZS, OF, MTR)  buf_page_get_gen(\
 				SP, ZS, OF, RW_NO_LATCH, NULL,\
-				BUF_GET_NO_LATCH, __FILE__, __LINE__, MTR)
+				BUF_GET_NO_LATCH, \
+				__FILE__, __LINE__, MTR)
 /******************************************************************
 NOTE! The following macros should be used instead of buf_page_get_gen, to
 improve debugging. Only values RW_S_LATCH and RW_X_LATCH are allowed as LA! */
 #define buf_page_get_nowait(SP, ZS, OF, LA, MTR)	buf_page_get_gen(\
 				SP, ZS, OF, LA, NULL,\
-				BUF_GET_NOWAIT, __FILE__, __LINE__, MTR)
+				BUF_GET_NOWAIT, \
+				__FILE__, __LINE__, MTR)
 /******************************************************************
 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, BL, MC, MTR)			     \
+#define buf_page_optimistic_get(LA, BL, MC, MTR)			\
 	buf_page_optimistic_get_func(LA, BL, MC, __FILE__, __LINE__, MTR)
 /************************************************************************
 This is the general function used to get optimistic access to a database
@@ -258,7 +264,8 @@ buf_page_get_gen(
 	ulint		rw_latch,/* in: RW_S_LATCH, RW_X_LATCH, RW_NO_LATCH */
 	buf_block_t*	guess,	/* in: guessed block or NULL */
 	ulint		mode,	/* in: BUF_GET, BUF_GET_IF_IN_POOL,
-				BUF_GET_NO_LATCH */
+				BUF_GET_NO_LATCH, BUF_GET_NOWAIT or
+				BUF_GET_IF_IN_POOL_WATCH*/
 	const char*	file,	/* in: file name */
 	ulint		line,	/* in: line where called */
 	mtr_t*		mtr);	/* in: mini-transaction */
@@ -952,8 +959,23 @@ UNIV_INTERN
 ulint
 buf_get_free_list_len(void);
 /*=======================*/
+/********************************************************************
+Stop watching if the marked page is read in. */
 
+void
+buf_pool_remove_watch(void);
+/*=======================*/
+/********************************************************************
+Check if the given page is being watched and has been read to the buffer
+pool. */
 
+ibool
+buf_pool_watch_happened(
+/*====================*/
+				/* out: TRUE if the given page is being
+				watched and it has been read in */
+	ulint	space,		/* in: space id */
+	ulint	page_no);	/* in: page number */
 
 /* The common buffer control block structure
 for compressed and uncompressed frames */
@@ -1186,6 +1208,16 @@ struct buf_pool_struct{
 					buf_block_t file pages,
 					buf_page_in_file() == TRUE,
 					indexed by (space_id, offset) */
+	/*--------------------------*/	/* Delete buffering data */
+	ibool		watch_active;	/* if TRUE, set watch_happened to
+					TRUE when page watch_space/
+					watch_page_no is read in. */
+	ulint		watch_space;	/* space id of watched page */
+	ulint		watch_page_no;	/* page number of watched page */
+	ibool		watch_happened;	/* has watched page been read in */
+	/*--------------------------*/
+
+
 	hash_table_t*	zip_hash;	/* hash table of buf_block_t blocks
 					whose frames are allocated to the
 					zip buddy system,

include/fil0fil.h

@@ -158,14 +158,6 @@ fil_space_get_type(
 			/* out: FIL_TABLESPACE or FIL_LOG */
 	ulint	id);	/* in: space id */
 /***********************************************************************
-Returns the ibuf data of a file space. */
-UNIV_INTERN
-ibuf_data_t*
-fil_space_get_ibuf_data(
-/*====================*/
-			/* out: ibuf data for this space */
-	ulint	id);	/* in: space id */
-/***********************************************************************
 Appends a new file to the chain of files of a space. File must be closed. */
 UNIV_INTERN
 void
@@ -274,14 +266,6 @@ fil_set_max_space_id_if_bigger(
 /*===========================*/
 	ulint	max_id);/* in: maximum known id */
 /********************************************************************
-Initializes the ibuf data structure for space 0 == the system tablespace.
-This can be called after the file space headers have been created and the
-dictionary system has been initialized. */
-UNIV_INTERN
-void
-fil_ibuf_init_at_db_start(void);
-/*===========================*/
-/********************************************************************
 Writes the flushed lsn and the latest archived log number to the page
 header of the first page of each data file in the system tablespace. */
 UNIV_INTERN

include/ibuf0ibuf.h

@@ -18,23 +18,21 @@ Created 7/19/1997 Heikki Tuuri
 #include "ibuf0types.h"
 #include "fsp0fsp.h"
 
+/* Possible operations buffered in the insert/whatever buffer. See
+ibuf_insert(). DO NOT CHANGE THE VALUES OF THESE, THEY ARE STORED ON DISK. */
+typedef enum {
+	IBUF_OP_INSERT = 0,
+	IBUF_OP_DELETE_MARK = 1,
+	IBUF_OP_DELETE = 2,
+
+	/* Number of different operation types. */
+	IBUF_OP_COUNT = 3,
+} ibuf_op_t;
+
 extern ibuf_t*	ibuf;
 
 /**********************************************************************
-Creates the insert buffer data struct for a single tablespace. Reads the
-root page of the insert buffer tree in the tablespace. This function can
-be called only after the dictionary system has been initialized, as this
-creates also the insert buffer table and index for this tablespace. */
-UNIV_INTERN
-ibuf_data_t*
-ibuf_data_init_for_space(
-/*=====================*/
-			/* out, own: ibuf data struct, linked to the list
-			in ibuf control structure. */
-	ulint	space);	/* in: space id */
-/**********************************************************************
-Creates the insert buffer data structure at a database startup and
-initializes the data structures for the insert buffer of each tablespace. */
+Creates the insert buffer data structure at a database startup. */
 UNIV_INTERN
 void
 ibuf_init_at_db_start(void);
@@ -165,38 +163,29 @@ ibuf_page(
 			/* out: TRUE if level 2 or level 3 page */
 	ulint	space,	/* in: space id */
 	ulint	zip_size,/* in: compressed page size in bytes, or 0 */
-	ulint	page_no);/* in: page number */
-/***************************************************************************
-Checks if a page is a level 2 or 3 page in the ibuf hierarchy of pages. */
-UNIV_INTERN
-ibool
-ibuf_page_low(
-/*==========*/
-			/* out: TRUE if level 2 or level 3 page */
-	ulint	space,	/* in: space id */
-	ulint	zip_size,/* in: compressed page size in bytes, or 0 */
 	ulint	page_no,/* in: page number */
 	mtr_t*	mtr);	/* in: mtr which will contain an x-latch to the
 			bitmap page if the page is not one of the fixed
-			address ibuf pages */
+			address ibuf pages, or NULL, in which case a new
+			transaction is created. */
 /***************************************************************************
 Frees excess pages from the ibuf free list. This function is called when an OS
 thread calls fsp services to allocate a new file segment, or a new page to a
 file segment, and the thread did not own the fsp latch before this call. */
 UNIV_INTERN
 void
-ibuf_free_excess_pages(
-/*===================*/
-	ulint	space);		/* in: space id */
+ibuf_free_excess_pages(void);
+/*========================*/
 /*************************************************************************
-Makes an index insert to the insert buffer, instead of directly to the disk
-page, if this is possible. Does not do insert if the index is clustered
-or unique. */
+Buffer an operation in the insert/delete buffer, instead of doing it
+directly to the disk page, if this is possible. Does not do it if the index
+is clustered or unique. */
 UNIV_INTERN
 ibool
 ibuf_insert(
 /*========*/
 				/* out: TRUE if success */
+	ibuf_op_t	op,	/* in: operation type */
 	const dtuple_t*	entry,	/* in: index entry to insert */
 	dict_index_t*	index,	/* in: index where to insert */
 	ulint		space,	/* in: space id where to insert */
@@ -205,11 +194,11 @@ ibuf_insert(
 	que_thr_t*	thr);	/* in: query thread */
 /*************************************************************************
 When an index page is read from a disk to the buffer pool, this function
-inserts to the page the possible index entries buffered in the insert buffer.
-The entries are deleted from the insert buffer. If the page is not read, but
-created in the buffer pool, this function deletes its buffered entries from
-the insert buffer; there can exist entries for such a page if the page
-belonged to an index which subsequently was dropped. */
+applies any buffered operations to the page and deletes the entries from the
+insert buffer. If the page is not read, but created in the buffer pool, this
+function deletes its buffered entries from the insert buffer; there can
+exist entries for such a page if the page belonged to an index which
+subsequently was dropped. */
 UNIV_INTERN
 void
 ibuf_merge_or_delete_for_page(
@@ -300,6 +289,16 @@ void
 ibuf_print(
 /*=======*/
 	FILE*	file);	/* in: file where to print */
+/********************************************************************
+Read the first two bytes from a record's fourth field (counter field in new
+records; something else in older records). */
+
+ulint
+ibuf_rec_get_fake_counter(
+/*======================*/
+			/* out: "counter" field, or ULINT_UNDEFINED if for
+			some reason it can't be read*/
+	rec_t*	rec);	/* in: ibuf record */
 
 #define IBUF_HEADER_PAGE_NO	FSP_IBUF_HEADER_PAGE_NO
 #define IBUF_TREE_ROOT_PAGE_NO	FSP_IBUF_TREE_ROOT_PAGE_NO
@@ -309,6 +308,9 @@ for the file segment from which the pages for the ibuf tree are allocated */
 #define IBUF_HEADER		PAGE_DATA
 #define	IBUF_TREE_SEG_HEADER	0	/* fseg header for ibuf tree */
 
+/* The insert buffer tree itself is always located in space 0. */
+#define IBUF_SPACE_ID		0
+
 #ifndef UNIV_NONINL
 #include "ibuf0ibuf.ic"
 #endif

include/ibuf0ibuf.ic

@@ -18,36 +18,37 @@ If there is this much of free space, the corresponding bits are set in the
 ibuf bitmap. */
 #define IBUF_PAGE_SIZE_PER_FREE_SPACE	32
 
-/* Insert buffer data struct for a single tablespace */
-struct ibuf_data_struct{
-	ulint		space;	/* space id */
-	ulint		seg_size;/* allocated pages if the file segment
-				containing ibuf header and tree */
-	ulint		size;	/* size of the insert buffer tree in pages */
-	ibool		empty;	/* after an insert to the ibuf tree is
-				performed, this is set to FALSE, and if a
-				contract operation finds the tree empty, this
-				is set to TRUE */
-	ulint		free_list_len;
-				/* length of the free list */
-	ulint		height;	/* tree height */
-	dict_index_t*	index;	/* insert buffer index */
-	UT_LIST_NODE_T(ibuf_data_t) data_list;
-				/* list of ibuf data structs */
-	ulint		n_inserts;/* number of inserts made to the insert
-				buffer */
-	ulint		n_merges;/* number of pages merged */
-	ulint		n_merged_recs;/* number of records merged */
-};
+/* Insert buffer struct */
 
 struct ibuf_struct{
 	ulint		size;		/* current size of the ibuf index
-					trees in pages */
-	ulint		max_size;	/* recommended maximum size in pages
-					for the ibuf index tree */
-	UT_LIST_BASE_NODE_T(ibuf_data_t) data_list;
-					/* list of ibuf data structs for
-					each tablespace */
+					tree, in pages */
+	ulint		max_size;	/* recommended maximum size of the
+					ibuf index tree, in pages */
+	ulint		seg_size;	/* allocated pages of the file
+					segment containing ibuf header and
+					tree */
+	ibool		empty;		/* after an insert to the ibuf tree
+					is performed, this is set to FALSE,
+					and if a contract operation finds
+					the tree empty, this is set to
+					TRUE */
+	ulint		free_list_len;	/* length of the free list */
+	ulint		height;		/* tree height */
+	dict_index_t*	index;		/* insert buffer index */
+
+	ulint		n_ops[IBUF_OP_COUNT];
+					/* number of operations of each type
+					done */
+	ulint		n_merges;	/* number of pages merged */
+	ulint		n_merged_ops[IBUF_OP_COUNT];
+					/* number of operations of each type
+					merged to index pages */
+	ulint		n_discarded_ops[IBUF_OP_COUNT];
+					/* number of operations of each type
+					discarded without merging due to the
+					tablespace being deleted or the
+					index being dropped */
 };
 
 /****************************************************************************

include/ibuf0types.h

@@ -9,7 +9,6 @@ Created 7/29/1997 Heikki Tuuri
 #ifndef ibuf0types_h
 #define ibuf0types_h
 
-typedef struct ibuf_data_struct	ibuf_data_t;
 typedef	struct ibuf_struct	ibuf_t;
 
 #endif

include/row0row.h

@@ -268,6 +268,9 @@ ibool
 row_search_index_entry(
 /*===================*/
 				/* out: TRUE if found */
+	ibool*		was_buffered,
+				/* out: TRUE if the operation was buffered
+				in the insert/delete buffer. Can be NULL. */
 	dict_index_t*	index,	/* in: index */
 	const dtuple_t*	entry,	/* in: index entry */
 	ulint		mode,	/* in: BTR_MODIFY_LEAF, ... */

include/univ.i

@@ -137,6 +137,9 @@ operations (very slow); also UNIV_DEBUG must be defined */
 						for compressed pages */
 #endif
 
+//#define UNIV_DEBUG
+//#define UNIV_SYNC_DEBUG
+//#define UNIV_IBUF_DEBUG
 #define UNIV_BTR_DEBUG				/* check B-tree links */
 #define UNIV_LIGHT_MEM_DEBUG			/* light memory debugging */
 
@@ -316,8 +319,11 @@ it is read. */
 /* Minimize cache-miss latency by moving data at addr into a cache before
 it is read or written. */
 # define UNIV_PREFETCH_RW(addr) __builtin_prefetch(addr, 1, 3)
+/* Tell the compiler that variable/function is unused. */
+# define UNIV_UNUSED    __attribute__ ((unused))
 #else
 /* Dummy versions of the macros */
+# define UNIV_UNUSED
 # define UNIV_EXPECT(expr,value) (expr)
 # define UNIV_LIKELY_NULL(expr) (expr)
 # define UNIV_PREFETCH_R(addr) ((void) 0)

row/row0purge.c

@@ -197,11 +197,12 @@ row_purge_remove_clust_if_poss(
 }
 
 /***************************************************************
-Removes a secondary index entry if possible. */
+Removes a secondary index entry if possible, without trying to use the
+insert/delete buffer. */
 static
 ibool
-row_purge_remove_sec_if_poss_low(
-/*=============================*/
+row_purge_remove_sec_if_poss_low_nonbuffered(
+/*=========================================*/
 				/* out: TRUE if success or if not found */
 	purge_node_t*	node,	/* in: row purge node */
 	dict_index_t*	index,	/* in: index */
@@ -212,7 +213,7 @@ row_purge_remove_sec_if_poss_low(
 	btr_pcur_t	pcur;
 	btr_cur_t*	btr_cur;
 	ibool		success;
-	ibool		old_has = 0; /* remove warning */
+	ibool		old_has = FALSE; /* remove warning */
 	ibool		found;
 	ulint		err;
 	mtr_t		mtr;
@@ -221,13 +222,13 @@ row_purge_remove_sec_if_poss_low(
 	log_free_check();
 	mtr_start(&mtr);
 
-	found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
+	found = row_search_index_entry(NULL, index, entry, mode, &pcur, &mtr);
 
 	if (!found) {
 		/* Not found */
 
 		/* fputs("PURGE:........sec entry not found\n", stderr); */
-		/* dtuple_print(stderr, entry); */
+		/* dtuple_print(entry); */
 
 		btr_pcur_close(&pcur);
 		mtr_commit(&mtr);
@@ -266,8 +267,13 @@ row_purge_remove_sec_if_poss_low(
 			ut_ad(mode == BTR_MODIFY_TREE);
 			btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
 						   FALSE, &mtr);
-			success = err == DB_SUCCESS;
-			ut_a(success || err == DB_OUT_OF_FILE_SPACE);
+			if (err == DB_SUCCESS) {
+				success = TRUE;
+			} else if (err == DB_OUT_OF_FILE_SPACE) {
+				success = FALSE;
+			} else {
+				ut_error;
+			}
 		}
 	}
 
@@ -277,6 +283,117 @@ row_purge_remove_sec_if_poss_low(
 	return(success);
 }
 
+/***************************************************************
+Removes a secondary index entry if possible. */
+static
+ibool
+row_purge_remove_sec_if_poss_low(
+/*=============================*/
+				/* out: TRUE if success or if not found */
+	purge_node_t*	node,	/* in: row purge node */
+	dict_index_t*	index,	/* in: index */
+	dtuple_t*	entry,	/* in: index entry */
+	ulint		mode)	/* in: latch mode BTR_MODIFY_LEAF or
+				BTR_MODIFY_TREE */
+{
+	mtr_t		mtr;
+	btr_pcur_t	pcur;
+	btr_cur_t*	btr_cur;
+	ibool		found;
+	ibool		success;
+	ibool		was_buffered;
+	ibool		old_has = FALSE;
+	ibool		leaf_in_buf_pool;
+
+	ut_a((mode == BTR_MODIFY_TREE) || (mode == BTR_MODIFY_LEAF));
+
+	if (mode == BTR_MODIFY_TREE) {
+		/* Can't use the insert/delete buffer if we potentially
+		need to split pages. */
+
+		return(row_purge_remove_sec_if_poss_low_nonbuffered(
+			       node, index, entry, mode));
+	}
+
+	log_free_check();
+
+	mtr_start(&mtr);
+
+	found = row_search_index_entry(
+		NULL, index, entry,
+		BTR_SEARCH_LEAF | BTR_WATCH_LEAF, &pcur, &mtr);
+
+	btr_cur = btr_pcur_get_btr_cur(&pcur);
+	leaf_in_buf_pool = btr_cur->leaf_in_buf_pool;
+
+	ut_a(!(found && !leaf_in_buf_pool));
+
+	btr_pcur_close(&pcur);
+	mtr_commit(&mtr);
+
+	if (leaf_in_buf_pool) {
+
+		if (found) {
+			/* Index entry exists and is in the buffer pool, no
+			need to use the insert/delete buffer. */
+
+			return(row_purge_remove_sec_if_poss_low_nonbuffered(
+				       node, index, entry, BTR_MODIFY_LEAF));
+		} else {
+			/* Index entry does not exist, nothing to do. */
+
+			return(TRUE);
+		}
+	}
+
+	/* We should remove the index record if no later version of the row,
+	which cannot be purged yet, requires its existence. If some
+	requires, we should do nothing. */
+
+	mtr_start(&mtr);
+
+	success = row_purge_reposition_pcur(BTR_SEARCH_LEAF, node, &mtr);
+
+	if (success) {
+		old_has = row_vers_old_has_index_entry(
+			TRUE, btr_pcur_get_rec(&node->pcur),
+			&mtr, index, entry);
+	}
+
+	btr_pcur_commit_specify_mtr(&node->pcur, &mtr);
+
+	if (success && old_has) {
+		/* Can't remove the index record yet. */
+
+		buf_pool_remove_watch();
+
+		return(TRUE);
+	}
+
+	mtr_start(&mtr);
+
+	btr_cur->thr = que_node_get_parent(node);
+
+	row_search_index_entry(&was_buffered, index, entry,
+			       BTR_MODIFY_LEAF | BTR_DELETE, &pcur,
+			       &mtr);
+
+	btr_pcur_close(&pcur);
+	mtr_commit(&mtr);
+
+	buf_pool_remove_watch();
+
+	if (!was_buffered) {
+		/* Page read into buffer pool or delete-buffering failed. */
+
+		return(row_purge_remove_sec_if_poss_low_nonbuffered(
+			       node, index, entry, BTR_MODIFY_LEAF));
+	}
+
+	return(TRUE);
+
+}
+
 /***************************************************************
 Removes a secondary index entry if possible. */
 UNIV_INLINE

row/row0row.c

@@ -789,6 +789,9 @@ ibool
 row_search_index_entry(
 /*===================*/
 				/* out: TRUE if found */
+	ibool*		was_buffered,
+				/* out: TRUE if the operation was buffered
+				in the insert/delete buffer. Can be NULL. */
 	dict_index_t*	index,	/* in: index */
 	const dtuple_t*	entry,	/* in: index entry */
 	ulint		mode,	/* in: BTR_MODIFY_LEAF, ... */
@@ -799,17 +802,48 @@ row_search_index_entry(
 	ulint	n_fields;
 	ulint	low_match;
 	rec_t*	rec;
+	ibool	ret;
 
 	ut_ad(dtuple_check_typed(entry));
 
 	btr_pcur_open(index, entry, PAGE_CUR_LE, mode, pcur, mtr);
+
+	ret = btr_pcur_was_buffered(pcur);
+
+	if (was_buffered) {
+		*was_buffered = ret;
+	}
+
+	if (ret) {
+		/* Operation was buffered in the insert/delete buffer;
+		pretend that we found the record. */
+
+		return(TRUE);
+	} else if ((mode & BTR_WATCH_LEAF)
+		   && !btr_pcur_get_btr_cur(pcur)->leaf_in_buf_pool) {
+
+		/* We did not read in the leaf page, thus we can't have
+		found anything. */
+
+		return(FALSE);
+	}
+
 	low_match = btr_pcur_get_low_match(pcur);
 
 	rec = btr_pcur_get_rec(pcur);
 
 	n_fields = dtuple_get_n_fields(entry);
 
-	return(!page_rec_is_infimum(rec) && low_match == n_fields);
+	if (page_rec_is_infimum(rec)) {
+
+		return(FALSE);
+	} else if (low_match != n_fields) {
+		/* Not found */
+
+		return(FALSE);
+	}
+
+	return(TRUE);
 }
 
 #ifndef UNIV_HOTBACKUP

row/row0uins.c

@@ -136,7 +136,7 @@ row_undo_ins_remove_sec_low(
 	log_free_check();
 	mtr_start(&mtr);
 
-	found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
+	found = row_search_index_entry(NULL, index, entry, mode, &pcur, &mtr);
 
 	btr_cur = btr_pcur_get_btr_cur(&pcur);
 

row/row0umod.c

@@ -307,7 +307,7 @@ row_undo_mod_del_mark_or_remove_sec_low(
 	log_free_check();
 	mtr_start(&mtr);
 
-	found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
+	found = row_search_index_entry(NULL, index, entry, mode, &pcur, &mtr);
 
 	btr_cur = btr_pcur_get_btr_cur(&pcur);
 
@@ -432,7 +432,7 @@ row_undo_mod_del_unmark_sec_and_undo_update(
 		return(DB_SUCCESS);
 	}
 
-	if (UNIV_UNLIKELY(!row_search_index_entry(index, entry,
+	if (UNIV_UNLIKELY(!row_search_index_entry(NULL, index, entry,
 						  mode, &pcur, &mtr))) {
 		fputs("InnoDB: error in sec index entry del undo in\n"
 		      "InnoDB: ", stderr);

row/row0upd.c

@@ -1451,21 +1451,23 @@ row_upd_sec_index_entry(
 	upd_node_t*	node,	/* in: row update node */
 	que_thr_t*	thr)	/* in: query thread */
 {
-	ibool		check_ref;
-	ibool		found;
-	dict_index_t*	index;
-	dtuple_t*	entry;
+	mtr_t		mtr;
+	rec_t*		rec;
 	btr_pcur_t	pcur;
-	btr_cur_t*	btr_cur;
 	mem_heap_t*	heap;
-	rec_t*		rec;
-	ulint		err	= DB_SUCCESS;
-	mtr_t		mtr;
-	trx_t*		trx	= thr_get_trx(thr);
+	dtuple_t*	entry;
+	dict_index_t*	index;
+	ibool		found;
+	btr_cur_t*	btr_cur;
+	ibool		referenced;
+	ibool		was_buffered;
+	ulint		err = DB_SUCCESS;
+	trx_t*		trx = thr_get_trx(thr);
+	ulint		mode = BTR_MODIFY_LEAF;
 
 	index = node->index;
 
-	check_ref = row_upd_index_is_referenced(index, trx);
+	referenced = row_upd_index_is_referenced(index, trx);
 
 	heap = mem_heap_create(1024);
 
@@ -1476,8 +1478,24 @@ row_upd_sec_index_entry(
 	log_free_check();
 	mtr_start(&mtr);
 
-	found = row_search_index_entry(index, entry, BTR_MODIFY_LEAF, &pcur,
-				       &mtr);
+	btr_pcur_get_btr_cur(&pcur)->thr = thr;
+
+	/* We can only try to use the insert/delete buffer to buffer
+	delete-mark operations if the index we're modifying has no foreign
+	key constraints referring to it. */
+	if (!referenced) {
+		mode |= BTR_DELETE_MARK;
+	}
+
+	found = row_search_index_entry(
+		&was_buffered, index, entry, BTR_MODIFY_LEAF, &pcur, &mtr);
+
+	if (was_buffered) {
+		/* Entry was delete marked already. */
+
+		goto close_cur;
+	}
+
 	btr_cur = btr_pcur_get_btr_cur(&pcur);
 
 	rec = btr_cur_get_rec(btr_cur);
@@ -1504,15 +1522,20 @@ row_upd_sec_index_entry(
 		delete marked if we return after a lock wait in
 		row_ins_index_entry below */
 
-		if (!rec_get_deleted_flag(rec,
-					  dict_table_is_comp(index->table))) {
-			err = btr_cur_del_mark_set_sec_rec(0, btr_cur, TRUE,
-							   thr, &mtr);
-			if (err == DB_SUCCESS && check_ref) {
+		if (!rec_get_deleted_flag(
+			rec, dict_table_is_comp(index->table))) {
+
+			err = btr_cur_del_mark_set_sec_rec(
+				0, btr_cur, TRUE, thr, &mtr);
+
+			if (err == DB_SUCCESS && referenced) {
+
+				ulint*	offsets;
+
+				offsets = rec_get_offsets(
+					rec, index, NULL, ULINT_UNDEFINED,
+					&heap);
 
-				ulint*	offsets = rec_get_offsets(
-					rec, index, NULL,
-					ULINT_UNDEFINED, &heap);
 				/* NOTE that the following call loses
 				the position of pcur ! */
 				err = row_upd_check_references_constraints(
@@ -1522,6 +1545,7 @@ row_upd_sec_index_entry(
 		}
 	}
 
+close_cur:
 	btr_pcur_close(&pcur);
 	mtr_commit(&mtr);
 
@@ -1583,7 +1607,7 @@ row_upd_clust_rec_by_insert(
 	upd_node_t*	node,	/* in: row update node */
 	dict_index_t*	index,	/* in: clustered index of the record */
 	que_thr_t*	thr,	/* in: query thread */
-	ibool		check_ref,/* in: TRUE if index may be referenced in
+	ibool		referenced,/* in: TRUE if index may be referenced in
 				a foreign key constraint */
 	mtr_t*		mtr)	/* in: mtr; gets committed here */
 {
@@ -1629,16 +1653,21 @@ row_upd_clust_rec_by_insert(
 		btr_cur_mark_extern_inherited_fields(
 			btr_cur_get_page_zip(btr_cur),
 			rec, index, offsets, node->update, mtr);
-		if (check_ref) {
+		if (referenced) {
 			/* NOTE that the following call loses
 			the position of pcur ! */
+
 			err = row_upd_check_references_constraints(
 				node, pcur, table, index, offsets, thr, mtr);
+
 			if (err != DB_SUCCESS) {
+
 				mtr_commit(mtr);
+
 				if (UNIV_LIKELY_NULL(heap)) {
 					mem_heap_free(heap);
 				}
+
 				return(err);
 			}
 		}
@@ -1794,7 +1823,8 @@ row_upd_del_mark_clust_rec(
 	ulint*		offsets,/* in/out: rec_get_offsets() for the
 				record under the cursor */
 	que_thr_t*	thr,	/* in: query thread */
-	ibool		check_ref,/* in: TRUE if index may be referenced in
+	ibool		referenced,
+				/* in: TRUE if index may be referenced in
 				a foreign key constraint */
 	mtr_t*		mtr)	/* in: mtr; gets committed here */
 {
@@ -1819,13 +1849,11 @@ row_upd_del_mark_clust_rec(
 
 	err = btr_cur_del_mark_set_clust_rec(BTR_NO_LOCKING_FLAG,
 					     btr_cur, TRUE, thr, mtr);
-	if (err == DB_SUCCESS && check_ref) {
+	if (err == DB_SUCCESS && referenced) {
 		/* NOTE that the following call loses the position of pcur ! */
 
-		err = row_upd_check_references_constraints(node,
-							   pcur, index->table,
-							   index, offsets,
-							   thr, mtr);
+		err = row_upd_check_references_constraints(
+			node, pcur, index->table, index, offsets, thr, mtr);
 	}
 
 	mtr_commit(mtr);
@@ -1848,7 +1876,6 @@ row_upd_clust_step(
 	dict_index_t*	index;
 	btr_pcur_t*	pcur;
 	ibool		success;
-	ibool		check_ref;
 	ulint		err;
 	mtr_t*		mtr;
 	mtr_t		mtr_buf;
@@ -1856,11 +1883,12 @@ row_upd_clust_step(
 	mem_heap_t*	heap		= NULL;
 	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
 	ulint*		offsets;
+	ibool		referenced;
 	rec_offs_init(offsets_);
 
 	index = dict_table_get_first_index(node->table);
 
-	check_ref = row_upd_index_is_referenced(index, thr_get_trx(thr));
+	referenced = row_upd_index_is_referenced(index, thr_get_trx(thr));
 
 	pcur = node->pcur;
 
@@ -1930,8 +1958,9 @@ row_upd_clust_step(
 	/* NOTE: the following function calls will also commit mtr */
 
 	if (node->is_delete) {
-		err = row_upd_del_mark_clust_rec(node, index, offsets,
-						 thr, check_ref, mtr);
+		err = row_upd_del_mark_clust_rec(
+			node, index, offsets, thr, referenced, mtr);
+
 		if (err == DB_SUCCESS) {
 			node->state = UPD_NODE_UPDATE_ALL_SEC;
 			node->index = dict_table_get_next_index(index);
@@ -1979,8 +2008,9 @@ row_upd_clust_step(
 		choosing records to update. MySQL solves now the problem
 		externally! */
 
-		err = row_upd_clust_rec_by_insert(node, index, thr, check_ref,
-						  mtr);
+		err = row_upd_clust_rec_by_insert(
+			node, index, thr, referenced, mtr);
+
 		if (err != DB_SUCCESS) {
 
 			return(err);