[MDEV-5582] Plugin 'MEMORY' has ref_count=1 after shutdown with materialization+semijoin Created: 2014-01-28  Updated: 2014-02-07  Resolved: 2014-02-07

Status: Closed
Project: MariaDB Server
Component/s: None
Affects Version/s: 5.3.12, 5.5.35, 10.0.7
Fix Version/s: 5.5.36, 10.0.9, 5.3.13

Type: Bug Priority: Major
Reporter: Elena Stepanova Assignee: Sergei Petrunia
Resolution: Fixed Votes: 0
Labels: None

Issue Links:
Relates

 Description    

SET optimizer_switch = 'materialization=on,semijoin=on';
 
 
 
CREATE TABLE t1 (a INT);
 
INSERT INTO t1 VALUES (8),(9);
 
 
 
CREATE TABLE t2 (b INT);
 
INSERT INTO t2 VALUES (2),(3);
 
 
 
CREATE TABLE t3 (c INT, INDEX(c));
 
INSERT INTO t2 VALUES (4),(5);
 
 
 
SELECT * FROM t1 WHERE 9 IN ( SELECT b FROM t2 WHERE 1 IN ( SELECT MIN(c) FROM t3 ) );

In 5.3 and 5.5, it produces the error on shutdown:

140128 19:10:55 [Warning] Plugin 'MEMORY' will be forced to shutdown
 
140128 19:10:55 [ERROR] Plugin 'MEMORY' has ref_count=1 after shutdown.

In 10.0, it causes assertion failure:

sql/sql_class.cc:1634: virtual THD::~THD(): Assertion `status_var.memory_used == 0' failed.
 
140128 19:13:28 [ERROR] mysqld got signal 6 ;
 

#6  0x00007f4e893ea621 in *__GI___assert_fail (assertion=0xfd028f "status_var.memory_used == 0", file=<optimized out>, line=1634, function=0xfd2850 "virtual THD::~THD()") at assert.c:81
 
#7  0x00000000006295a2 in THD::~THD (this=0x7f4e8333d070, __in_chrg=<optimized out>) at 10.0/sql/sql_class.cc:1634
 
#8  0x000000000062972a in THD::~THD (this=0x7f4e8333d070, __in_chrg=<optimized out>) at 10.0/sql/sql_class.cc:1639
 
#9  0x00000000005ab8f2 in unlink_thd (thd=0x7f4e8333d070) at 10.0/sql/mysqld.cc:2715
 
#10 0x00000000005abc52 in one_thread_per_connection_end (thd=0x7f4e8333d070, put_in_cache=true) at 10.0/sql/mysqld.cc:2826
 
#11 0x000000000077a563 in do_handle_one_connection (thd_arg=0x7f4e8333d070) at 10.0/sql/sql_connect.cc:1390
 
#12 0x000000000077a210 in handle_one_connection (arg=0x7f4e8333d070) at 10.0/sql/sql_connect.cc:1293
 
#13 0x0000000000a9d1f1 in pfs_spawn_thread (arg=0x7f4e83b75df0) at 10.0/storage/perfschema/pfs.cc:1853
 
#14 0x00007f4e8a94ab50 in start_thread (arg=<optimized out>) at pthread_create.c:304
 
#15 0x00007f4e89499a7d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:112

Stack trace from:

revision-id: psergey@askmonty.org-20140121100700-9g8hl8vx8nmk1e2w
 
revno: 3965
 
branch-nick: 10.0



 Comments   
Comment by Sergei Petrunia [ 2014-02-07 ]

The leak is because materialization table for non-merged semi-join is created and not freed.

The query plan is:
mysql> explain SELECT * FROM t1 WHERE 9 IN ( SELECT b FROM t2 WHERE 1 IN ( SELECT MIN(c) FROM t3 ) );
----------------------------------------------------------------------------------+

id select_type table type possible_keys key key_len ref rows Extra

----------------------------------------------------------------------------------+

1 PRIMARY NULL NULL NULL NULL NULL NULL NULL Impossible WHERE
3 MATERIALIZED NULL NULL NULL NULL NULL NULL NULL No matching min/max row

----------------------------------------------------------------------------------+
2 rows in set (0.00 sec)

Comment by Sergei Petrunia [ 2014-02-07 ]

The following happens:

> JOIN::optimize()

convert_join_subqueries_to_semijoins()

  • 3rd-level subquery is converted to non-merged semi-join
  • 2nd-level subquery is converted into a semijoin.

setup_jtbm_semi_joins()

  • finds that subquery with id=3 is degenerate, still creates a dummy temporary
    table $T for it.

constant table code finds that non-merged semi-join is an empty
constant table, based on that, runs:
tables=NULL;
goto setup_subq_exit;

<JOIN::optimize()

>JOIN::cleanup(full=true)
if (table)

{ ... }

^^ the branch is not taken, bevause JOIN::table=NULL (and
JOIN::join_tab=NULL, too).

As a consequence, the temp.table is never freed.

Comment by Sergei Petrunia [ 2014-02-07 ]

Upon closer examination: JOIN::cleanup( full=true) actually does call cleanup_empty_jtbm_semi_joins(JOIN *join). The problem is that that function doesn't recurse down into semi-join nests.

Generated at Thu Feb 08 07:05:28 UTC 2024 using Jira 8.20.16#820016-sha1:9d11dbea5f4be3d4cc21f03a88dd11d8c8687422.