Hi, I made a test for MyISAM and a 5.0 patch (as you can see :-) ). You
were offline when I was done, so I just send you this email instead:
<gluh> 1. restore max_heap_table_size in the end of test case
done
gluh> 2. Please check this test case with valgrind and check also
another tab & read_record fields which maybe should be updated too.
No valgrind errors. This will be checked in pushbuild as well.
I don't think we should update any other fields in read_record nor
join_tab. The fields that are set up in read_record at this point in
time are in make_join_readinfo(). I find no other fields in there that
should need to be changed just because we change handler. The handler
change is meant to be transparent: just swapping out a memory table to disk.
OTOH I think that this bug is really a design problem. IMU a read_record
is part of the query execution plan, and a change of handlers should be
done before we start to compile a QEP. If you make a change in a C
program, you run gcc again, you don't go inside the binary and start
setting bytes are you? ;-) Alternatively, we should rerun the phase that
initialized read_record in the first place. But this is beyond a bug fix
IMHO. Maybe something to keep in mind if we do some bigger work on I_S?
gluh> 3. add + tab->read_record.file= table_list->table->file; into
<gluh> if (table_list->schema_table->fill_table(thd, table_list,
<gluh> tab->select_cond))
<gluh> {
<gluh> ..}
done.
I'm back online on Monday for more discussions.
Best Regards
Martin
mhansson@stripped wrote:
> Below is the list of changes that have just been committed into a local
> 5.0 repository of mhansson. When mhansson does a push these changes
> will be propagated to the main repository and, within 24 hours after the
> push, to the public repository.
> For information on how to access the public repository
> see http://dev.mysql.com/doc/mysql/en/installing-source-tree.html
>
> ChangeSet@stripped, 2008-03-19 14:32:28+01:00, mhansson@riffraff.(none) +3 -0
> Bug#34529: Crash on complex Falcon I_S select after ALTER .. PARTITION BY
>
> When swapping out heap I_S tables to disk, this is done after plan refinement.
> Thus, READ_RECORD::file will still point to the (deleted) heap handler at start
> of execution. This causes segmentation fault if join buffering is used and the
> query is a star query where the result is found to be empty before accessing
> some table. In this case that table has not been initialized (i.e. had its
> READ_RECORD re-initialized) before the cleanup routine tries to close the handler.
> Fixed by updating READ_RECORD::file when changing handler.
>
> mysql-test/r/information_schema.result@stripped, 2008-03-19 14:32:27+01:00,
> mhansson@riffraff.(none) +10 -0
> Bug#34529: Test result.
>
> mysql-test/t/information_schema.test@stripped, 2008-03-19 14:32:27+01:00,
> mhansson@riffraff.(none) +21 -0
> Bug#34529: Test case.
>
> sql/sql_show.cc@stripped, 2008-03-19 14:32:27+01:00, mhansson@riffraff.(none) +2 -0
> Bug#34529: The fix.
>
> diff -Nrup a/mysql-test/r/information_schema.result
> b/mysql-test/r/information_schema.result
> --- a/mysql-test/r/information_schema.result 2007-10-29 11:45:33 +01:00
> +++ b/mysql-test/r/information_schema.result 2008-03-19 14:32:27 +01:00
> @@ -1422,3 +1422,13 @@ show fields from information_schema.tabl
> ERROR 42S02: Unknown table 'table_names' in information_schema
> show keys from information_schema.table_names;
> ERROR 42S02: Unknown table 'table_names' in information_schema
> +USE information_schema;
> +SET max_heap_table_size = 16384;
> +CREATE TABLE test.t1( a INT );
> +SELECT *
> +FROM tables ta
> +JOIN collations co ON ( co.collation_name = ta.table_catalog )
> +JOIN character_sets cs ON ( cs.character_set_name = ta.table_catalog );
>
> +TABLE_CATALOG TABLE_SCHEMA TABLE_NAME TABLE_TYPE ENGINE VERSION ROW_FORMAT TABLE_ROWS AVG_ROW_LENGTH DATA_LENGTH MAX_DATA_LENGTH INDEX_LENGTH DATA_FREE AUTO_INCREMENT CREATE_TIME UPDATE_TIME CHECK_TIME TABLE_COLLATION CHECKSUM CREATE_OPTIONS TABLE_COMMENT COLLATION_NAME CHARACTER_SET_NAME ID IS_DEFAULT IS_COMPILED SORTLEN CHARACTER_SET_NAME DEFAULT_COLLATE_NAME DESCRIPTION MAXLEN
> +DROP TABLE test.t1;
> +SET max_heap_table_size = DEFAULT;
> diff -Nrup a/mysql-test/t/information_schema.test
> b/mysql-test/t/information_schema.test
> --- a/mysql-test/t/information_schema.test 2007-10-29 11:45:33 +01:00
> +++ b/mysql-test/t/information_schema.test 2008-03-19 14:32:27 +01:00
> @@ -1118,3 +1118,24 @@ show fields from information_schema.tabl
> --error 1109
> show keys from information_schema.table_names;
>
> +#
> +# Bug#34529: Crash on complex Falcon I_S select after ALTER .. PARTITION BY
> +#
> +USE information_schema;
> +SET max_heap_table_size = 16384;
> +
> +CREATE TABLE test.t1( a INT );
> +
> +# What we need to create here is a bit of a corner case:
> +# We need a star query with information_schema tables, where the first
> +# branch of the star join produces zero rows, so that reading of the
> +# second branch never happens. At the same time we have to make sure
> +# that data for at least the last table is swapped from MEMORY/HEAP to
> +# MyISAM. This and only this triggers the bug.
> +SELECT *
> +FROM tables ta
> +JOIN collations co ON ( co.collation_name = ta.table_catalog )
> +JOIN character_sets cs ON ( cs.character_set_name = ta.table_catalog );
> +
> +DROP TABLE test.t1;
> +SET max_heap_table_size = DEFAULT;
> diff -Nrup a/sql/sql_show.cc b/sql/sql_show.cc
> --- a/sql/sql_show.cc 2007-12-17 09:12:06 +01:00
> +++ b/sql/sql_show.cc 2008-03-19 14:32:27 +01:00
> @@ -4073,9 +4073,11 @@ bool get_schema_tables_result(JOIN *join
> {
> result= 1;
> join->error= 1;
> + tab->read_record.file= table_list->table->file;
> table_list->schema_table_state= executed_place;
> break;
> }
> + tab->read_record.file= table_list->table->file;
> table_list->schema_table_state= executed_place;
> }
> }
>
>
