branches/zip issue# 33

This fix makes two basic changes in blob handling:
(The bug was introduced in r2252)

1) The blob prefixes are no longer stored in the undo if
  a) We are modifying a delete marked record and
  b) The record was delete marked by an already committed trx.

2) When building old row version to check if one of these versions
can hold an implicit lock on the record we stop our probe if
  a) The version is delete marked and
  b) The delete marking is done by a trx which is different from
     the current active trx on the record.

Reviewed by: Heikki

include/trx0rec.h

@@ -178,6 +178,9 @@ trx_undo_rec_get_partial_row(
 				record! */
 	dict_index_t*	index,	/* in: clustered index */
 	dtuple_t**	row,	/* out, own: partial row */
+	ibool		ignore_prefix, /* in: flag to indicate if we
+				expect blob prefixes in undo. Used
+				only in the assertion. */
 	mem_heap_t*	heap);	/* in: memory heap from which the memory
 				needed is allocated */
 /***************************************************************************

row/row0purge.c

@@ -559,8 +559,10 @@ err_exit:
 	/* Read to the partial row the fields that occur in indexes */
 
 	if (!(cmpl_info & UPD_NODE_NO_ORD_CHANGE)) {
-		ptr = trx_undo_rec_get_partial_row(ptr, clust_index,
-						   &node->row, node->heap);
+		ptr = trx_undo_rec_get_partial_row(
+			ptr, clust_index, &node->row,
+			type == TRX_UNDO_UPD_DEL_REC,
+			node->heap);
 	}
 
 	return(TRUE);

row/row0vers.c

@@ -163,6 +163,26 @@ row_vers_impl_x_locked_off_kernel(
 			clust_offsets = rec_get_offsets(
 				prev_version, clust_index, NULL,
 				ULINT_UNDEFINED, &heap);
+
+			vers_del = rec_get_deleted_flag(prev_version,
+							comp);
+			prev_trx_id = row_get_rec_trx_id(prev_version,
+							 clust_index,
+						 	 clust_offsets);
+
+			/* If the trx_id and prev_trx_id are
+			different and if the prev_version is marked
+			deleted then the prev_trx_id must have
+			already committed for the trx_id to be able to
+			modify the row. Therefore, prev_trx_id cannot
+			hold any implicit lock. */
+			if (0 != ut_dulint_cmp(trx_id, prev_trx_id)
+			    && vers_del) {
+	
+				mutex_enter(&kernel_mutex);
+				break;
+			}
+
 			/* The stack of versions is locked by mtr.
 			Thus, it is safe to fetch the prefixes for
 			externally stored columns. */
@@ -206,8 +226,6 @@ row_vers_impl_x_locked_off_kernel(
 		if prev_version would require rec to be in a different
 		state. */
 
-		vers_del = rec_get_deleted_flag(prev_version, comp);
-
 		/* We check if entry and rec are identified in the alphabetical
 		ordering */
 		if (0 == cmp_dtuple_rec(entry, rec, offsets)) {
@@ -243,9 +261,6 @@ row_vers_impl_x_locked_off_kernel(
 			break;
 		}
 
-		prev_trx_id = row_get_rec_trx_id(prev_version, clust_index,
-						 clust_offsets);
-
 		if (0 != ut_dulint_cmp(trx_id, prev_trx_id)) {
 			/* The versions modified by the trx_id transaction end
 			to prev_version: no implicit x-lock */

trx/trx0rec.c

@@ -533,6 +533,8 @@ trx_undo_page_report_modify(
 	ulint		type_cmpl;
 	byte*		type_cmpl_ptr;
 	ulint		i;
+	dulint		trx_id;
+	ibool		ignore_prefix = FALSE;
 	byte		ext_buf[REC_MAX_INDEX_COL_LEN
 				+ BTR_EXTERN_FIELD_REF_SIZE];
 
@@ -565,6 +567,11 @@ trx_undo_page_report_modify(
 		type_cmpl = TRX_UNDO_DEL_MARK_REC;
 	} else if (rec_get_deleted_flag(rec, dict_table_is_comp(table))) {
 		type_cmpl = TRX_UNDO_UPD_DEL_REC;
+		/* We are about to update a delete marked record.
+		We don't typically need the prefix in this case unless
+		the delete marking is done by the same transaction
+		(which we check below). */
+		ignore_prefix = TRUE;
 	} else {
 		type_cmpl = TRX_UNDO_UPD_EXIST_REC;
 	}
@@ -588,7 +595,16 @@ trx_undo_page_report_modify(
 					  index, DATA_TRX_ID), &flen);
 	ut_ad(flen == DATA_TRX_ID_LEN);
 
-	ptr += mach_dulint_write_compressed(ptr, trx_read_trx_id(field));
+	trx_id = trx_read_trx_id(field);
+
+	/* If it is an update of a delete marked record, then we are
+	allowed to ignore blob prefixes if the delete marking was done
+	by some other trx as it must have committed by now for us to
+	allow an over-write. */
+	if (ignore_prefix) {
+		ignore_prefix = ut_dulint_cmp(trx_id, trx->id) != 0;
+	}
+	ptr += mach_dulint_write_compressed(ptr, trx_id);
 
 	field = rec_get_nth_field(rec, offsets,
 				  dict_index_get_sys_col_pos(
@@ -663,6 +679,7 @@ trx_undo_page_report_modify(
 					ptr,
 					dict_index_get_nth_col(index, pos)
 					->ord_part
+					&& !ignore_prefix
 					&& flen < REC_MAX_INDEX_COL_LEN
 					? ext_buf : NULL,
 					dict_table_zip_size(table),
@@ -746,6 +763,7 @@ trx_undo_page_report_modify(
 					ptr = trx_undo_page_report_modify_ext(
 						ptr,
 						flen < REC_MAX_INDEX_COL_LEN
+						&& !ignore_prefix
 						? ext_buf : NULL,
 						dict_table_zip_size(table),
 						&field, &flen);
@@ -996,6 +1014,9 @@ trx_undo_rec_get_partial_row(
 				record! */
 	dict_index_t*	index,	/* in: clustered index */
 	dtuple_t**	row,	/* out, own: partial row */
+	ibool		ignore_prefix, /* in: flag to indicate if we
+				expect blob prefixes in undo. Used
+				only in the assertion. */
 	mem_heap_t*	heap)	/* in: memory heap from which the memory
 				needed is allocated */
 {
@@ -1045,7 +1066,8 @@ trx_undo_rec_get_partial_row(
 			/* If the prefix of this column is indexed,
 			ensure that enough prefix is stored in the
 			undo log record. */
-			ut_a(!col->ord_part
+			ut_a(ignore_prefix
+			     || !col->ord_part
 			     || dfield_get_len(dfield)
 			     >= REC_MAX_INDEX_COL_LEN
 			     + BTR_EXTERN_FIELD_REF_SIZE);