mariadb

Tree: 1a60685cba

Commit Graph

Author	SHA1	Message	Date
unknown	1a60685cba	Patch changing how ALTER TABLE implementation handles table locking and invalidation in the most general case (non-temporary table and not simple RENAME or ENABLE/DISABLE KEYS or partitioning command). See comment for sql/sql_table.cc for more information. These changes are prerequisite for 5.1 version of fix for bug #23667 "CREATE TABLE LIKE is not isolated from alteration by other connections" mysql-test/include/mix1.inc: Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES, which should be consistent across all platforms and for all engines. mysql-test/r/alter_table-big.result: Changed test for bug #25044 to use @@debug and injected sleeps infrastructure. Extended test coverage for ALTER TABLE's behavior under concurrency. mysql-test/r/alter_table.result: Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES, which should be consistent across all platforms and for all engines. mysql-test/r/innodb_mysql.result: Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES, which should be consistent across all platforms and for all engines. mysql-test/t/alter_table-big.test: Changed test for bug #25044 to use @@debug and injected sleeps infrastructure. Extended test coverage for ALTER TABLE's behavior under concurrency. mysql-test/t/alter_table.test: Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES, which should be consistent across all platforms and for all engines. sql/mysql_priv.h: Made functions reopen_table() and close_handle_and_leave_table_as_lock() available outside of sql_base.cc file. Changed close_data_tables() in such way that after closing handler for the table it leaves TABLE object for it in table cache not as placeholder for ordinary name-lock but as placeholder for an exclusive name-lock. Renamed this routine to close_data_files_and_morph_locks(). sql/sql_base.cc: Made functions reopen_table() and close_handle_and_leave_table_as_lock() available outside of sql_base.cc file. Changed close_data_tables() in such way that after closing handler for the table it leaves TABLE object for it in table cache not as placeholder for ordinary name-lock but as placeholder for an exclusive name-lock. Renamed this routine to close_data_files_and_morph_locks(). Also adjusted it so it can work properly not only in LOCK TABLES mode. sql/sql_table.cc: Changed the way in which ALTER TABLE implementation handles table locking and invalidation in the most general case (non-temporary table and not simple RENAME or ENABLE/DISABLE KEYS or partitioning command) Now after preparing new version of the table we: 1) Wait until all other threads close old version of table. 2) Close instances of table open by this thread and replace them with exclusive name-locks. 3) Rename the old table to a temp name, rename the new one to the old name. 4) If we are under LOCK TABLES and don't do ALTER TABLE ... RENAME we reopen new version of table. 5) Write statement to the binary log. 6) If we are under LOCK TABLES and do ALTER TABLE ... RENAME we remove name-locks from list of open tables and table cache. 7) If we are not not under LOCK TABLES we rely on close_thread_tables() call to remove name-locks from table cache and list of open table. Such approach: a) Eliminates possibility for concurrent statement to sneak in and get access to the new version of the table before ALTER TABLE gets logged into binary log. b) Ensures that ALTER TABLE behaves under LOCK TABLES in the same way on all platforms and for all engines (in 5.0 this was not true) c) Preserves nice invariant that if table is open in some connection there is a guarantee that .FRM file for this table exists and is properly named.	19 years ago
unknown	c1a54d976a	Fixed test case for bug #25044 "ALTER TABLE ... ENABLE KEYS acquires global 'opening tables' lock." after merging it into 5.1 tree. mysql-test/r/alter_table-big.result: Fixed "show binlog events" usage as 5.1 and 5.0 sizes of binlog events differ. mysql-test/t/alter_table-big.test: This test needs statement based binlogging. Also fixed "show binlog events" usage as 5.1 and 5.0 sizes of binlog events differ.	19 years ago
unknown	ac5e6f60a8	Fix for bug #25044 "ALTER TABLE ... ENABLE KEYS acquires global 'opening tables' lock." Execution of ALTER TABLE ... ENABLE KEYS on a table (which can take rather long time) prevented concurrent execution of all statements using tables. The problem was caused by the fact that we were holding LOCK_open mutex during whole duration of this statement and particularly during call to handler::enable_indexes(). This behavior was introduced as part of the fix for bug 14262 "SP: DROP PROCEDURE\|VIEW (maybe more) write to binlog too late (race cond)" The patch simply restores old behavior. Note that we can safely do this as this operation takes exclusive lock (similar to name-lock) which blocks both DML and DDL on the table being altered. It also introduces mysql-test/include/wait_show_pattern.inc helper script which is used to make test-case for this bug robust enough. mysql-test/include/wait_slave_status.inc: Now wait_slave_status.inc reuses more generic wait_output_matches.inc script. sql/sql_table.cc: mysql_alter_table(): Changed ALTER TABLE ... ENABLE/DISABLE KEYS not to hold LOCK_open mutex during call to handler::enable_indexes() as the latter can take rather long time and therefore such ALTER would block execution of all other statements that use tables. We can safely do this as this operation takes exclusive lock (similar to name-lock) on the table which is altered. mysql-test/include/wait_show_pattern.inc: New BitKeeper file ``mysql-test/include/wait_show_pattern.inc'' mysql-test/r/alter_table-big.result: New BitKeeper file ``mysql-test/r/alter_table-big.result'' mysql-test/t/alter_table-big.test: New BitKeeper file ``mysql-test/t/alter_table-big.test''	19 years ago

Author

SHA1

Message

Date

unknown

1a60685cba

Patch changing how ALTER TABLE implementation handles table locking

and invalidation in the most general case (non-temporary table and
not simple RENAME or ENABLE/DISABLE KEYS or partitioning command).

See comment for sql/sql_table.cc for more information.

These changes are prerequisite for 5.1 version of fix for bug #23667
"CREATE TABLE LIKE is not isolated from alteration by other connections"


mysql-test/include/mix1.inc:
  Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES,
  which should be consistent across all platforms and for all engines.
mysql-test/r/alter_table-big.result:
  Changed test for bug #25044 to use @@debug and injected sleeps
  infrastructure. Extended test coverage for ALTER TABLE's behavior
  under concurrency.
mysql-test/r/alter_table.result:
  Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES,
  which should be consistent across all platforms and for all engines.
mysql-test/r/innodb_mysql.result:
  Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES,
  which should be consistent across all platforms and for all engines.
mysql-test/t/alter_table-big.test:
  Changed test for bug #25044 to use @@debug and injected sleeps
  infrastructure. Extended test coverage for ALTER TABLE's behavior
  under concurrency.
mysql-test/t/alter_table.test:
  Extended coverage for behavior of ALTER TABLE statement under LOCK TABLES,
  which should be consistent across all platforms and for all engines.
sql/mysql_priv.h:
  Made functions reopen_table() and close_handle_and_leave_table_as_lock()
  available outside of sql_base.cc file.
  Changed close_data_tables() in such way that after closing handler
  for the table it leaves TABLE object for it in table cache not as
  placeholder for ordinary name-lock but as placeholder for an exclusive
  name-lock. Renamed this routine to close_data_files_and_morph_locks().
sql/sql_base.cc:
  Made functions reopen_table() and close_handle_and_leave_table_as_lock()
  available outside of sql_base.cc file.
  Changed close_data_tables() in such way that after closing handler
  for the table it leaves TABLE object for it in table cache not as
  placeholder for ordinary name-lock but as placeholder for an exclusive
  name-lock. Renamed this routine to close_data_files_and_morph_locks().
  Also adjusted it so it can work properly not only in LOCK TABLES mode.
sql/sql_table.cc:
  Changed the way in which ALTER TABLE implementation handles table
  locking and invalidation in the most general case (non-temporary table
  and not simple RENAME or ENABLE/DISABLE KEYS or partitioning command)
  
  Now after preparing new version of the table we:
  1) Wait until all other threads close old version of table.
  2) Close instances of table open by this thread and replace them
     with exclusive name-locks.
  3) Rename the old table to a temp name, rename the new one to the
     old name.
  4) If we are under LOCK TABLES and don't do ALTER TABLE ... RENAME
     we reopen new version of table.
  5) Write statement to the binary log.
  6) If we are under LOCK TABLES and do ALTER TABLE ... RENAME we
     remove name-locks from list of open tables and table cache.
  7) If we are not not under LOCK TABLES we rely on close_thread_tables()
     call to remove name-locks from table cache and list of open table. 
  
  Such approach:
  a) Eliminates possibility for concurrent statement to sneak in and get
     access to the new version of the table before ALTER TABLE gets logged
     into binary log.
  b) Ensures that ALTER TABLE behaves under LOCK TABLES in the same way
     on all platforms and for all engines (in 5.0 this was not true) 
  c) Preserves nice invariant that if table is open in some connection
     there is a guarantee that .FRM file for this table exists and is
     properly named.

19 years ago

unknown

c1a54d976a

Fixed test case for bug #25044 "ALTER TABLE ... ENABLE KEYS acquires

global 'opening tables' lock." after merging it into 5.1 tree.


mysql-test/r/alter_table-big.result:
  Fixed "show binlog events" usage as 5.1 and 5.0 sizes of binlog
  events differ.
mysql-test/t/alter_table-big.test:
  This test needs statement based binlogging.
  Also fixed "show binlog events" usage as 5.1 and 5.0 sizes
  of binlog events differ.

19 years ago

unknown

ac5e6f60a8

Fix for bug #25044 "ALTER TABLE ... ENABLE KEYS acquires global 'opening

tables' lock."

Execution of ALTER TABLE ... ENABLE KEYS on a table (which can take rather
long time) prevented concurrent execution of all statements using tables.

The problem was caused by the fact that we were holding LOCK_open mutex
during whole duration of this statement and particularly during call
to handler::enable_indexes(). This behavior was introduced as part of the
fix for bug 14262 "SP: DROP PROCEDURE|VIEW (maybe more) write to binlog
too late (race cond)"

The patch simply restores old behavior. Note that we can safely do this as
this operation takes exclusive lock (similar to name-lock) which blocks both
DML and DDL on the table being altered.

It also introduces mysql-test/include/wait_show_pattern.inc helper script
which is used to make test-case for this bug robust enough.


mysql-test/include/wait_slave_status.inc:
  Now wait_slave_status.inc reuses more generic wait_output_matches.inc script.
sql/sql_table.cc:
  mysql_alter_table():
    Changed ALTER TABLE ... ENABLE/DISABLE KEYS not to hold LOCK_open mutex
    during call to handler::enable_indexes() as the latter can take rather
    long time and therefore such ALTER would block execution of all other
    statements that use tables. We can safely do this as this operation takes 
    exclusive lock (similar to name-lock) on the table which is altered.
mysql-test/include/wait_show_pattern.inc:
  New BitKeeper file ``mysql-test/include/wait_show_pattern.inc''
mysql-test/r/alter_table-big.result:
  New BitKeeper file ``mysql-test/r/alter_table-big.result''
mysql-test/t/alter_table-big.test:
  New BitKeeper file ``mysql-test/t/alter_table-big.test''

19 years ago

3 Commits (1a60685cbaf6bcd919189ac19f01f65c50d79b54)