There's magic code to handle overflow records that have been deleted but

still need to be visible to othread threads (see bt_ovfl.c for lots more
information).  We were using WT_RESTART returns from the overflow read
function to handle this, but that led to a bug where, if the read of an
overflow item returns restart, the cursor may have already moved past
the entry, and so restart potentially skips a row.

Row-store used to store cached "deleted overflow" items each row's
WT_UPDATE list, while column-store stored cached "deleted overflow"
items in the page's modify structure.  Change row-store to do the same
thing as column-store, and then change the system to pass enough
information into the overflow read function that it never needs to
return restart, it can simply look-aside into the cached entries as soon
as we realize we're reading a "deleted overflow" item.

This means restart is now only returned as a result of modification
failure, remove a bunch of restart handling code.

1 files changed, 1 insertions, 2 deletions

diff --git a/src/include/cursor.i b/src/include/cursor.i
index 3b07bd162f6..477b879cfd8 100644
--- a/src/include/cursor.i
+++ b/src/include/cursor.i
@@ -253,8 +253,7 @@ slow:			WT_RET(__wt_row_leaf_key_work(
 		vb->size = 0;
 	} else {
 		__wt_cell_unpack(cell, unpack);
-		WT_RET(__wt_cell_unpack_ref(
-		    session, WT_PAGE_ROW_LEAF, unpack, vb));
+		WT_RET(__wt_page_cell_data_ref(session, cbt->page, unpack, vb));
 	}
 
 	return (0);