Details
-
Type:
Bug
-
Status: Open
-
Priority:
Major
-
Resolution: Unresolved
-
Affects Version/s: N/A
-
Fix Version/s: 5.5
-
Component/s: None
-
Labels:None
Description
MySQL 5.5.29 has a patch backported from MySQL 5.6, that supposedly fixes a bug.
The test case for this bugfix is wrong, it doesn't exercise the changed code. We need to fix the test case to see if there's a genuine bug, and apply the MySQL fix, if it's good.
The patch is attached.
Gliffy Diagrams
Attachments
Issue Links
Activity
- All
- Comments
- Work Log
- History
- Activity
- Transitions
If I take the cset "tor.didriksen@oracle.com-20110825105434-ll28zinu0nhazt3z", undo the changes in sql_select.cc (so that changes in other files remain, we get the added asserts and new testcases), and after that run the testsuite, I get:
mysqld: /home/psergey/dev2/mysql-5.6-r3404/sql/uniques.cc:587: bool Unique::get(TABLE*): Assertion `table->sort.record_pointers == __null' failed.
If I
- take the fix for mysql-5.5
- remove all changes except the assert statements
- add a testcase from the fix for 5.6
I get assertion failures
There is another bugfix that is related to MySQL BUG#12694872:
Bug#13340270: assertion table->sort.record_pointers == __null
bug#12694872 fixed a memory leak in Unique::get(), but did not expect
Unique::get() to be called twice before the end of the query, where
buffers are freed.
For a note, we have a 5.3 bug apparently with a similar valgrind warning: MDEV-5623.
The test case there is for Aria, but it can be tweaked to be valid for MyISAM as well, e.g.
SET optimizer_switch = 'index_merge=on,index_merge_sort_union=on,derived_merge=on,in_to_exists=on';
CREATE TABLE t1 (a INT, b INT, c VARCHAR(1), INDEX(a), INDEX(c)) ENGINE=MyISAM;
INSERT INTO t1 VALUES
(5,6,'z'),(2,6,'c'),(0,8,'a'),(1,2,'q'),(8,6,'y'),
(1,8,'y'),(1,3,'r'),(9,3,'v'),(1,9,'z'),(5,6,'r');
CREATE TABLE t2 (d VARCHAR(1), INDEX(d)) ENGINE=MyISAM;
INSERT INTO t2 VALUES ('x'),('a');
CREATE TABLE t3 (e VARCHAR(1)) ENGINE=MyISAM;
INSERT INTO t3 VALUES ('x'),('d');
PREPARE stmt FROM "
SELECT t1.* FROM t1 FORCE INDEX (c,a)
LEFT JOIN (
SELECT * FROM t2
WHERE EXISTS ( SELECT * FROM t3 WHERE e != d )
) AS sq ON (sq.d = c )
WHERE c < 'p' OR a > 43
ORDER BY t1.b
";
EXECUTE stmt;
EXECUTE stmt;
However, I am not getting the valgrind failure with it on a prefixed MySQL 5.5/5.6, so I guess it's not exactly the same.
"Changeset in 5.5 with the test case" is actually in 5.6