Commit e61c4480 authored by heikki@hundin.mysql.fi's avatar heikki@hundin.mysql.fi
Browse files

ha_innobase.cc: 

  Since MySQL commits the stmt always at the end of an INSERT, it is enough to release auto-inc lock at innobase_commit and innobase_rollback; add also more comments to code
parent 4c7ac574

sql/ha_innobase.cc

+72 −12
Original line number Diff line number Diff line
@@ -130,8 +130,9 @@ static void innobase_print_error(const char* db_errpfx, char* buffer); 
/* General functions */



/**********************************************************************

Releases possible search latch, auto inc lock, and InnoDB thread FIFO ticket.

These should be released at each SQL statement end. */

Releases possible search latch and InnoDB thread FIFO ticket. These should

be released at each SQL statement end. It does no harm to release these

also in the middle of an SQL statement. */

static

void

innobase_release_stat_resources(
@@ -142,16 +143,6 @@ innobase_release_stat_resources( 
		trx_search_latch_release_if_reserved(trx);

	}



	if (trx->auto_inc_lock) {

		  	

		/* If we had reserved the auto-inc lock for

		some table in this SQL statement, we release it now */

		  	

		srv_conc_enter_innodb(trx);

		row_unlock_table_autoinc_for_mysql(trx);

		srv_conc_exit_innodb(trx);

	}



	if (trx->declared_to_be_inside_innodb) {

		/* Release our possible ticket in the FIFO */
@@ -635,6 +626,16 @@ innobase_commit( 


	trx = check_trx_exists(thd);



        if (trx->auto_inc_lock) {

		  	

		/* If we had reserved the auto-inc lock for

		some table in this SQL statement, we release it now */

		  	

		srv_conc_enter_innodb(trx);

		row_unlock_table_autoinc_for_mysql(trx);

		srv_conc_exit_innodb(trx);

	}



	if (trx_handle != (void*)&innodb_dummy_stmt_trx_handle) {

		innobase_commit_low(trx);

	}
@@ -704,6 +705,16 @@ innobase_rollback( 


	trx = check_trx_exists(thd);



        if (trx->auto_inc_lock) {

		  	

		/* If we had reserved the auto-inc lock for

		some table in this SQL statement, we release it now */

		  	

		srv_conc_enter_innodb(trx);

		row_unlock_table_autoinc_for_mysql(trx);

		srv_conc_exit_innodb(trx);

	}



	srv_conc_enter_innodb(trx);



	if (trx_handle != (void*)&innodb_dummy_stmt_trx_handle) {
@@ -1900,6 +1911,55 @@ convert_search_mode_to_innobase( 
	return(0);

}



/*

   BACKGROUND INFO: HOW A SELECT SQL QUERY IS EXECUTED

   ---------------------------------------------------

The following does not cover all the details, but explains how we determine

the start of a new SQL statement, and what is associated with it.



For each table in the database the MySQL interpreter may have several

table handle instances in use, also in a single SQL query. For each table

handle instance there is an InnoDB  'prebuilt' struct which contains most

of the InnoDB data associated with this table handle instance.



  A) if the user has not explicitly set any MySQL table level locks:



  1) MySQL calls ::external_lock to set an 'intention' table level lock on

the table of the handle instance. There we set

prebuilt->sql_stat_start = TRUE. The flag sql_stat_start should be set

true if we are taking this table handle instance to use in a new SQL

statement issued by the user. We also increment trx->n_mysql_tables_in_use.



  2) If prebuilt->sql_stat_start == TRUE we 'pre-compile' the MySQL search

instructions to prebuilt->template of the table handle instance in

::index_read. The template is used to save CPU time in large joins.



  3) In row_search_for_mysql, if prebuilt->sql_stat_start is true, we

allocate a new consistent read view for the trx if it does not yet have one,

or in the case of a locking read, set an InnoDB 'intention' table level

lock on the table.



  4) We do the SELECT. MySQL may repeatedly call ::index_read for the

same table handle instance, if it is a join.



  5) When the SELECT ends, MySQL removes its intention table level locks

in ::external_lock. When trx->n_mysql_tables_in_use drops to zero,

 (a) we execute a COMMIT there if the autocommit is on,

 (b) we also release possible 'SQL statement level resources' InnoDB may

have for this SQL statement. The MySQL interpreter does NOT execute

autocommit for pure read transactions, though it should. That is why the

table handler in that case has to execute the COMMIT in ::external_lock.



  B) If the user has explicitly set MySQL table level locks, then MySQL

does NOT call ::external_lock at the start of the statement. To determine

when we are at the start of a new SQL statement we at the start of

::index_read also compare the query id to the latest query id where the

table handle instance was used. If it has changed, we know we are at the

start of a new SQL statement. Since the query id can theoretically

overwrap, we use this test only as a secondary way of determining the

start of a new SQL statement. */





/**************************************************************************

Positions an index cursor to the index specified in the handle. Fetches the

row if any. */