bug #57006 "Deadlock between HANDLER and FLUSH TABLES WITH READ
LOCK" and bug #54673 "It takes too long to get readlock for
'FLUSH TABLES WITH READ LOCK'".
The first bug manifested itself as a deadlock which occurred
when a connection, which had some table open through HANDLER
statement, tried to update some data through DML statement
while another connection tried to execute FLUSH TABLES WITH
READ LOCK concurrently.
What happened was that FTWRL in the second connection managed
to perform first step of GRL acquisition and thus blocked all
upcoming DML. After that it started to wait for table open
through HANDLER statement to be flushed. When the first connection
tried to execute DML it has started to wait for GRL/the second
connection creating deadlock.
The second bug manifested itself as starvation of FLUSH TABLES
WITH READ LOCK statements in cases when there was a constant
stream of concurrent DML statements (in two or more
connections).
This has happened because requests for protection against GRL
which were acquired by DML statements were ignoring presence of
pending GRL and thus the latter was starved.
This patch solves both these problems by re-implementing GRL
using metadata locks.
Similar to the old implementation acquisition of GRL in new
implementation is two-step. During the first step we block
all concurrent DML and DDL statements by acquiring global S
metadata lock (each DML and DDL statement acquires global IX
lock for its duration). During the second step we block commits
by acquiring global S lock in COMMIT namespace (commit code
acquires global IX lock in this namespace).
Note that unlike in old implementation acquisition of
protection against GRL in DML and DDL is semi-automatic.
We assume that any statement which should be blocked by GRL
will either open and acquires write-lock on tables or acquires
metadata locks on objects it is going to modify. For any such
statement global IX metadata lock is automatically acquired
for its duration.
The first problem is solved because waits for GRL become
visible to deadlock detector in metadata locking subsystem
and thus deadlocks like one in the first bug become impossible.
The second problem is solved because global S locks which
are used for GRL implementation are given preference over
IX locks which are acquired by concurrent DML (and we can
switch to fair scheduling in future if needed).
Important change:
FTWRL/GRL no longer blocks DML and DDL on temporary tables.
Before this patch behavior was not consistent in this respect:
in some cases DML/DDL statements on temporary tables were
blocked while in others they were not. Since the main use cases
for FTWRL are various forms of backups and temporary tables are
not preserved during backups we have opted for consistently
allowing DML/DDL on temporary tables during FTWRL/GRL.
Important change:
This patch changes thread state names which are used when
DML/DDL of FTWRL is waiting for global read lock. It is now
either "Waiting for global read lock" or "Waiting for commit
lock" depending on the stage on which FTWRL is.
Incompatible change:
To solve deadlock in events code which was exposed by this
patch we have to replace LOCK_event_metadata mutex with
metadata locks on events. As result we have to prohibit
DDL on events under LOCK TABLES.
This patch also adds extensive test coverage for interaction
of DML/DDL and FTWRL.
Performance of new and old global read lock implementations
in sysbench tests were compared. There were no significant
difference between new and old implementations.
tree for embedded server
Test case for bug #56251 "Deadlock with INSERT
DELAYED and MERGE tables" can't be run against
embedded server. Embedded server converts all
DELAYED INSERTs into ordinary INSERTs and this
test can't work properly if such conversion
happens.
Moved this test from merge.test to delayed.test
which is skipped if test suite is run with
--embedded-server option.
with two connections doing LOCK TABLE+INSERT DELAYED".
Disabled --ps-protocol for this part of the test as INSERT
DELAYED simply doesn't work with it under LOCK TABLES.
The problem was that deadlocks involving INSERT DELAYED were not detected.
The reason for this is that two threads are involved in INSERT DELAYED:
the connection thread and the handler thread. The connection thread would
wait while the handler thread acquired locks and opened the table.
In essence, this adds an edge to the wait-for-graph between the
connection thread and the handler thread that the deadlock detector is
unaware of. Therefore many deadlocks involving INSERT DELAYED were not
detected.
This patch fixes the problem by having the connection thread acquire the
metadata lock the table before starting the handler thread. This allows the
deadlock detector to detect any possible deadlocks resulting from trying to
acquire a metadata lock the table. If a metadata lock is successfully acquired,
the handler thread is started and given a copy of the ticket representing the
metadata lock. When the handler thread then tries to lock and open the table,
it will find that it already has the metadata lock and therefore not acquire
any new metadata locks.
Test cases added to delayed.test.
FLUSH TABLES <list> WITH READ LOCK are incompatible" to
be pushed as separate patch.
Replaced thread state name "Waiting for table", which was
used by threads waiting for a metadata lock or table flush,
with a set of names which better reflect types of resources
being waited for.
Also replaced "Table lock" thread state name, which was used
by threads waiting on thr_lock.c table level lock, with more
elaborate "Waiting for table level lock", to make it
more consistent with other thread state names.
Updated test cases and their results according to these
changes.
Fixed sys_vars.query_cache_wlock_invalidate_func test to not
to wait for timeout of wait_condition.inc script.
------------------------------------------------------------
revno: 3035.4.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 39897-6.0
timestamp: Thu 2009-01-15 12:17:57 -0200
message:
Bug#39897: lock_multi fails in pushbuild: timeout waiting for processlist
The problem is that relying on the "Table lock" thread state in
its current position to detect that a thread is waiting on a lock
is race prone. The "Table lock" state change happens before the
thread actually tries to grab a lock on a table.
The solution is to move the "Table lock" state so that its set
only when a thread is actually going to wait for a lock. The state
change happens after the thread fails to grab the lock (because it
is owned by other thread) and proceeds to wait on a condition.
This is considered part of work related to WL#4284 "Transactional
DDL locking"
Warning: this patch contains an incompatible change.
When waiting on a lock in thr_lock.c, the server used to display "Locked"
processlist state. After this patch, the state is "Table lock".
The new state was actually intended to be display since year 2002,
when Monty added it. But up until removal of thd->locked boolean
member, this state was ignored by SHOW PROCESSLIST code.
Fixed a problem with the test case when executed with ps-protocol.
There the conflicing lock would be noticed during prepare, not
during execution of the insert - leading to a different (but
equally appropriate) error message.
The problem was a "self-deadlock" if the connection issuing INSERT DELAYED
had both the global read lock (FLUSH TABLES WITH READ LOCK) and LOCK TABLES
mode active. The table being inserted into had to be different from the
table(s) locked by LOCK TABLES.
For INSERT DELAYED, the connection thread waits until the handler thread has
opened and locked its table before returning. But since the global read lock
was active, the handler thread would be unable to lock and would wait for the
global read lock to go away.
So the handler thread would be waiting for the connection thread to release
the global read lock while the connection thread was waiting for the handler
thread to lock the table. This gave a "self-deadlock" (same connection,
different threads).
The deadlock would only happen if we also had LOCK TABLES mode since the
INSERT otherwise will try to get protection against global read lock before
starting the handler thread. It will then notice that the global read lock
is owned by the same connection and report ER_CANT_UPDATE_WITH_READLOCK.
This patch removes the deadlock by reporting ER_CANT_UPDATE_WITH_READLOCK
also if we are inside LOCK TABLES mode.
Test case added to delayed.test.
During insert, we are not reading the rows in a referring table but
instead using the last read row that happens to be in table->record[0].
Now INSERT into such view is denied.
Bug #47274 assert in open_table on CREATE TABLE <already existing>
The problem was an assertion during execution of CREATE TABLES.
This assertion would occur if INSERT DELAYED or REPLACE DELAYED
were used to update a table containing an AUTO_INCREMENT column
and if the inserted row had a user-supplied value for that column.
Any CREATE TABLE statement (including CREATE TABLE SELECT and
CREATE TABLE LIKE) trying to create the same table and
which followed the INSERT/REPLACED would cause the assertion.
The problem was only noticeable on debug builds of the server
and not present in the mysql-5.1 tree.
The cause of the problem was that the code for delayed insert did
not properly reset the TABLE->auto_increment_if_null flag after
The flag is used to indicate that a non-null value of an auto_increment field
has been provided by the user or retrieved from a current record.
Open_tables() contains an assertion that tests this flag, and this
was triggered by CREATE TABLE.
This patch fixes the problem by resetting the auto_increment_if_null
field to FALSE once INSERT/REPLACE DELAYED has updated the table,
similar to what is done already for regular INSERT statements.
Test case added to delayed.test.
upgrading lock, even with low_priority_updates
The problem is that there is no mechanism to control whether a
delayed insert takes a high or low priority lock on a table.
The solution is to modify the delayed insert thread ("handler")
to take into account the global value of low_priority_updates
when taking table locks. The value of low_priority_updates is
retrieved when the insert delayed thread is created and will
remain the same for the duration of the thread.
corrupts a MERGE table
Bug 26867 - LOCK TABLES + REPAIR + merge table result in
memory/cpu hogging
Bug 26377 - Deadlock with MERGE and FLUSH TABLE
Bug 25038 - Waiting TRUNCATE
Bug 25700 - merge base tables get corrupted by
optimize/analyze/repair table
Bug 30275 - Merge tables: flush tables or unlock tables
causes server to crash
Bug 19627 - temporary merge table locking
Bug 27660 - Falcon: merge table possible
Bug 30273 - merge tables: Can't lock file (errno: 155)
The problems were:
Bug 26379 - Combination of FLUSH TABLE and REPAIR TABLE
corrupts a MERGE table
1. A thread trying to lock a MERGE table performs busy waiting while
REPAIR TABLE or a similar table administration task is ongoing on
one or more of its MyISAM tables.
2. A thread trying to lock a MERGE table performs busy waiting until all
threads that did REPAIR TABLE or similar table administration tasks
on one or more of its MyISAM tables in LOCK TABLES segments do UNLOCK
TABLES. The difference against problem #1 is that the busy waiting
takes place *after* the administration task. It is terminated by
UNLOCK TABLES only.
3. Two FLUSH TABLES within a LOCK TABLES segment can invalidate the
lock. This does *not* require a MERGE table. The first FLUSH TABLES
can be replaced by any statement that requires other threads to
reopen the table. In 5.0 and 5.1 a single FLUSH TABLES can provoke
the problem.
Bug 26867 - LOCK TABLES + REPAIR + merge table result in
memory/cpu hogging
Trying DML on a MERGE table, which has a child locked and
repaired by another thread, made an infinite loop in the server.
Bug 26377 - Deadlock with MERGE and FLUSH TABLE
Locking a MERGE table and its children in parent-child order
and flushing the child deadlocked the server.
Bug 25038 - Waiting TRUNCATE
Truncating a MERGE child, while the MERGE table was in use,
let the truncate fail instead of waiting for the table to
become free.
Bug 25700 - merge base tables get corrupted by
optimize/analyze/repair table
Repairing a child of an open MERGE table corrupted the child.
It was necessary to FLUSH the child first.
Bug 30275 - Merge tables: flush tables or unlock tables
causes server to crash
Flushing and optimizing locked MERGE children crashed the server.
Bug 19627 - temporary merge table locking
Use of a temporary MERGE table with non-temporary children
could corrupt the children.
Temporary tables are never locked. So we do now prohibit
non-temporary chidlren of a temporary MERGE table.
Bug 27660 - Falcon: merge table possible
It was possible to create a MERGE table with non-MyISAM children.
Bug 30273 - merge tables: Can't lock file (errno: 155)
This was a Windows-only bug. Table administration statements
sometimes failed with "Can't lock file (errno: 155)".
These bugs are fixed by a new implementation of MERGE table open.
When opening a MERGE table in open_tables() we do now add the
child tables to the list of tables to be opened by open_tables()
(the "query_list"). The children are not opened in the handler at
this stage.
After opening the parent, open_tables() opens each child from the
now extended query_list. When the last child is opened, we remove
the children from the query_list again and attach the children to
the parent. This behaves similar to the old open. However it does
not open the MyISAM tables directly, but grabs them from the already
open children.
When closing a MERGE table in close_thread_table() we detach the
children only. Closing of the children is done implicitly because
they are in thd->open_tables.
For more detail see the comment at the top of ha_myisammrg.cc.
Changed from open_ltable() to open_and_lock_tables() in all places
that can be relevant for MERGE tables. The latter can handle tables
added to the list on the fly. When open_ltable() was used in a loop
over a list of tables, the list must be temporarily terminated
after every table for open_and_lock_tables().
table_list->required_type is set to FRMTYPE_TABLE to avoid open of
special tables. Handling of derived tables is suppressed.
These details are handled by the new function
open_n_lock_single_table(), which has nearly the same signature as
open_ltable() and can replace it in most cases.
In reopen_tables() some of the tables open by a thread can be
closed and reopened. When a MERGE child is affected, the parent
must be closed and reopened too. Closing of the parent is forced
before the first child is closed. Reopen happens in the order of
thd->open_tables. MERGE parents do not attach their children
automatically at open. This is done after all tables are reopened.
So all children are open when attaching them.
Special lock handling like mysql_lock_abort() or mysql_lock_remove()
needs to be suppressed for MERGE children or forwarded to the parent.
This depends on the situation. In loops over all open tables one
suppresses child lock handling. When a single table is touched,
forwarding is done.
Behavioral changes:
===================
This patch changes the behavior of temporary MERGE tables.
Temporary MERGE must have temporary children.
The old behavior was wrong. A temporary table is not locked. Hence
even non-temporary children were not locked. See
Bug 19627 - temporary merge table locking.
You cannot change the union list of a non-temporary MERGE table
when LOCK TABLES is in effect. The following does *not* work:
CREATE TABLE m1 ... ENGINE=MRG_MYISAM ...;
LOCK TABLES t1 WRITE, t2 WRITE, m1 WRITE;
ALTER TABLE m1 ... UNION=(t1,t2) ...;
However, you can do this with a temporary MERGE table.
You cannot create a MERGE table with CREATE ... SELECT, neither
as a temporary MERGE table, nor as a non-temporary MERGE table.
CREATE TABLE m1 ... ENGINE=MRG_MYISAM ... SELECT ...;
Gives error message: table is not BASE TABLE.
SQL_MODE was ignored when a client issued INSERT DELAYED.
Some system settings weren't copied as intended when a record was saved for
a delayed insert.
The patch for WL 1563 added a new duplicate key error message so that the
key name could be provided instead of the key number. But the error code
for the new message was used even though that did not need to change.
This could cause unnecessary problems for applications that used the old
ER_DUP_ENTRY error code to detect duplicate key errors.
INSERT DELAYED inserts garbage for BIT columns.
When delayed thread clones TABLE object, it didn't adjusted bit_ptr
to newly created record (though it correctly adjusts ptr and null_ptr).
This is fixed by correctly adjusting bit_ptr when performing a clone.
With this fix BIT values are stored correctly by INSERT DELAYED.
Bug#20627 - INSERT DELAYED does not honour auto_increment_* variables
INSERT DELAYED ignored an explicitly set INSERT_ID and session
specific auto_increment_* variables.
The problem was that the inserts are done by a system thread,
which does not have access to the session variables of the user
thread.
On a proposal of Guilhem I fixed it so that the variables are
copied to the data structure for every delayed row. The system
thread sets its session variables from these values.
The INSERT DELAYED should not maintain its own private auto-increment
counter, because this is assuming that other threads cannot insert
into the table while the INSERT DELAYED thread is inserting, which is
a wrong assumption.
So the start of processing of a batch of INSERT rows in the
INSERT DELAYED thread must be treated as a start of a new statement
and cached next_insert_id must be cleared.
