Bug #37433 Deadlock between open_table, close_open_tables,

                get_table_share, drop_open_table
            
In the partition handler code, LOCK_open and share->LOCK_ha_data
are acquired in the wrong order in certain cases.  When doing a
multi-row INSERT (i.e a INSERT..SELECT) in a table with auto-
increment column(s). the increments must be in a monotonically
continuous increasing sequence (i.e it can't have "holes"). To
achieve this, a lock is held for the duration of the operation.
share->LOCK_ha_data was used for this purpose.
            
Whenever there was a need to open a view _during_ the operation
(views are not currently pre-opened the way tables are), and
LOCK_open was grabbed, a deadlock could occur.  share->LOCK_ha_data
is other places used _while_ holding LOCK_open.
            
A new mutex was introduced in the HA_DATA_PARTITION structure,
for exclusive use of the autoincrement data fields, so we don't
need to overload the use of LOCK_ha_data here.
            
A module test case has not been supplied, since the problem occurs
as a result of a race condition, and testing for this condition 
is thus not deterministic.   Testing for it could be done by
setting up a test case as described in the bug report.

1 files changed, 1 insertions, 0 deletions

diff --git a/sql/ha_partition.h b/sql/ha_partition.h
index ea1ef4d05fc..29509515d23 100644
--- a/sql/ha_partition.h
+++ b/sql/ha_partition.h
@@ -45,6 +45,7 @@ typedef struct st_ha_data_partition
 {
   ulonglong next_auto_inc_val;                 /**< first non reserved value */
   bool auto_inc_initialized;
+  pthread_mutex_t mutex;
 } HA_DATA_PARTITION;
 
 #define PARTITION_BYTES_IN_POS 2